Commit d3ad6274 authored by xuxzh1's avatar xuxzh1 🎱
Browse files

init

parent 97b02a89
#include "ggml-backend-impl.h"
#include "ggml-alloc.h"
#include "ggml-impl.h"
#include <assert.h>
#include <limits.h>
#include <stdarg.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#define MAX(a, b) ((a) > (b) ? (a) : (b))
// backend buffer type
const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
return buft->iface.get_name(buft);
}
GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
return buft->iface.alloc_buffer(buft, size);
}
size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
return buft->iface.get_alignment(buft);
}
size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
// get_max_size is optional, defaults to SIZE_MAX
if (buft->iface.get_max_size) {
return buft->iface.get_max_size(buft);
}
return SIZE_MAX;
}
GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
// get_alloc_size is optional, defaults to ggml_nbytes
if (buft->iface.get_alloc_size) {
size_t size = buft->iface.get_alloc_size(buft, tensor);
assert(size >= ggml_nbytes(tensor));
return size;
}
return ggml_nbytes(tensor);
}
bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
return buft->iface.supports_backend(buft, backend);
}
bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
if (buft->iface.is_host) {
return buft->iface.is_host(buft);
}
return false;
}
// backend buffer
GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
ggml_backend_buffer_type_t buft,
struct ggml_backend_buffer_i iface,
ggml_backend_buffer_context_t context,
size_t size) {
ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer));
(*buffer) = (struct ggml_backend_buffer) {
/* .interface = */ iface,
/* .buft = */ buft,
/* .context = */ context,
/* .size = */ size,
/* .usage = */ GGML_BACKEND_BUFFER_USAGE_ANY
};
return buffer;
}
const char * ggml_backend_buffer_name(ggml_backend_buffer_t buffer) {
return buffer->iface.get_name(buffer);
}
void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
if (buffer == NULL) {
return;
}
if (buffer->iface.free_buffer != NULL) {
buffer->iface.free_buffer(buffer);
}
free(buffer);
}
size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
return buffer->size;
}
void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
void * base = buffer->iface.get_base(buffer);
GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
return base;
}
GGML_CALL void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
// init_tensor is optional
if (buffer->iface.init_tensor) {
buffer->iface.init_tensor(buffer, tensor);
}
}
size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer) {
return ggml_backend_buft_get_alignment(ggml_backend_buffer_get_type(buffer));
}
size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) {
return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer));
}
size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor);
}
void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
buffer->iface.clear(buffer, value);
}
bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
return ggml_backend_buft_is_host(ggml_backend_buffer_get_type(buffer));
}
void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
buffer->usage = usage;
// FIXME: add a generic callback to the buffer interface
if (ggml_backend_buffer_is_multi_buffer(buffer)) {
ggml_backend_multi_buffer_set_usage(buffer, usage);
}
}
ggml_backend_buffer_type_t ggml_backend_buffer_get_type(ggml_backend_buffer_t buffer) {
return buffer->buft;
}
void ggml_backend_buffer_reset(ggml_backend_buffer_t buffer) {
if (buffer->iface.reset) {
buffer->iface.reset(buffer);
}
}
bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst) {
ggml_backend_buffer_t dst_buf = dst->view_src ? dst->view_src->buffer : dst->buffer;
if (dst_buf->iface.cpy_tensor) {
return src->buffer->iface.cpy_tensor(dst_buf, src, dst);
}
return false;
}
// backend
ggml_guid_t ggml_backend_guid(ggml_backend_t backend) {
if (backend == NULL) {
return NULL;
}
return backend->guid;
}
const char * ggml_backend_name(ggml_backend_t backend) {
if (backend == NULL) {
return "NULL";
}
return backend->iface.get_name(backend);
}
void ggml_backend_free(ggml_backend_t backend) {
if (backend == NULL) {
return;
}
backend->iface.free(backend);
}
ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend) {
return backend->iface.get_default_buffer_type(backend);
}
ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
return ggml_backend_buft_alloc_buffer(ggml_backend_get_default_buffer_type(backend), size);
}
size_t ggml_backend_get_alignment(ggml_backend_t backend) {
return ggml_backend_buft_get_alignment(ggml_backend_get_default_buffer_type(backend));
}
size_t ggml_backend_get_max_size(ggml_backend_t backend) {
return ggml_backend_buft_get_max_size(ggml_backend_get_default_buffer_type(backend));
}
void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
if (backend->iface.set_tensor_async == NULL) {
ggml_backend_tensor_set(tensor, data, offset, size);
} else {
backend->iface.set_tensor_async(backend, tensor, data, offset, size);
}
}
void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
if (backend->iface.get_tensor_async == NULL) {
ggml_backend_tensor_get(tensor, data, offset, size);
} else {
backend->iface.get_tensor_async(backend, tensor, data, offset, size);
}
}
GGML_CALL void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
GGML_ASSERT(buf != NULL && "tensor buffer not set");
GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
if (!size) {
return;
}
buf->iface.set_tensor(buf, tensor, data, offset, size);
}
GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
GGML_ASSERT(buf != NULL && "tensor buffer not set");
GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
if (!size) {
return;
}
buf->iface.get_tensor(buf, tensor, data, offset, size);
}
void ggml_backend_synchronize(ggml_backend_t backend) {
if (backend->iface.synchronize == NULL) {
return;
}
backend->iface.synchronize(backend);
}
ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
GGML_ASSERT(backend->iface.graph_plan_create != NULL);
return backend->iface.graph_plan_create(backend, cgraph);
}
void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
GGML_ASSERT(backend->iface.graph_plan_free != NULL);
backend->iface.graph_plan_free(backend, plan);
}
enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
GGML_ASSERT(backend->iface.graph_plan_compute != NULL);
return backend->iface.graph_plan_compute(backend, plan);
}
enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
enum ggml_status err = ggml_backend_graph_compute_async(backend, cgraph);
ggml_backend_synchronize(backend);
return err;
}
enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
return backend->iface.graph_compute(backend, cgraph);
}
bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
return backend->iface.supports_op(backend, op);
}
bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) {
if (backend->iface.offload_op != NULL) {
return backend->iface.offload_op(backend, op);
}
return false;
}
// backend copy
static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
if (a->type != b->type) {
return false;
}
for (int i = 0; i < GGML_MAX_DIMS; i++) {
if (a->ne[i] != b->ne[i]) {
return false;
}
if (a->nb[i] != b->nb[i]) {
return false;
}
}
return true;
}
void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
if (src == dst) {
return;
}
if (ggml_backend_buffer_is_host(src->buffer)) {
ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
} else if (ggml_backend_buffer_is_host(dst->buffer)) {
ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
} else if (!ggml_backend_buffer_copy_tensor(src, dst)) {
#ifndef NDEBUG
fprintf(stderr, "%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
#endif
size_t nbytes = ggml_nbytes(src);
void * data = malloc(nbytes);
ggml_backend_tensor_get(src, data, 0, nbytes);
ggml_backend_tensor_set(dst, data, 0, nbytes);
free(data);
}
}
void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst) {
GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
if (src == dst) {
return;
}
if (backend_dst->iface.cpy_tensor_async != NULL) {
if (backend_dst->iface.cpy_tensor_async(backend_src, backend_dst, src, dst)) {
return;
}
}
// an async copy would normally happen after all the queued operations on both backends are completed
// sync src, set_async dst
if (ggml_backend_buffer_is_host(src->buffer)) {
ggml_backend_synchronize(backend_src);
ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src));
} else {
ggml_backend_synchronize(backend_src);
ggml_backend_tensor_copy(src, dst);
ggml_backend_synchronize(backend_dst);
}
}
// events
ggml_backend_event_t ggml_backend_event_new(ggml_backend_t backend) {
if (backend->iface.event_new == NULL) {
return NULL;
}
return backend->iface.event_new(backend);
}
void ggml_backend_event_free(ggml_backend_event_t event) {
if (event == NULL) {
return;
}
event->backend->iface.event_free(event);
}
void ggml_backend_event_record(ggml_backend_event_t event) {
GGML_ASSERT(event->backend->iface.event_record != NULL);
event->backend->iface.event_record(event);
}
void ggml_backend_event_synchronize(ggml_backend_event_t event) {
GGML_ASSERT(event->backend->iface.event_synchronize != NULL);
event->backend->iface.event_synchronize(event);
}
void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
GGML_ASSERT(backend->iface.event_wait != NULL);
backend->iface.event_wait(backend, event);
}
// backend registry
#define GGML_REG_MAX_BACKENDS 16
struct ggml_backend_reg {
char name[128];
ggml_backend_init_fn init_fn;
ggml_backend_buffer_type_t default_buffer_type;
void * user_data;
};
static struct ggml_backend_reg ggml_backend_registry[GGML_REG_MAX_BACKENDS];
static size_t ggml_backend_registry_count = 0;
GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data);
GGML_CALL static void ggml_backend_registry_init(void) {
static bool initialized = false;
if (initialized) {
return;
}
initialized = true;
ggml_backend_register("CPU", ggml_backend_reg_cpu_init, ggml_backend_cpu_buffer_type(), NULL);
// add forward decls here to avoid including the backend headers
#ifdef GGML_USE_CUDA
extern GGML_CALL void ggml_backend_cuda_reg_devices(void);
ggml_backend_cuda_reg_devices();
#endif
#ifdef GGML_USE_SYCL
extern void ggml_backend_sycl_reg_devices(void);
ggml_backend_sycl_reg_devices();
#endif
#ifdef GGML_USE_METAL
extern GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data);
extern GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
ggml_backend_register("Metal", ggml_backend_reg_metal_init, ggml_backend_metal_buffer_type(), NULL);
#endif
#ifdef GGML_USE_VULKAN
extern GGML_CALL int ggml_backend_vk_reg_devices(void);
ggml_backend_vk_reg_devices();
#endif
#ifdef GGML_USE_KOMPUTE
extern GGML_CALL void ggml_backend_kompute_reg_devices(void);
ggml_backend_kompute_reg_devices();
#endif
}
GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
GGML_ASSERT(ggml_backend_registry_count < GGML_REG_MAX_BACKENDS);
size_t id = ggml_backend_registry_count;
ggml_backend_registry[id] = (struct ggml_backend_reg) {
/* .name = */ {0},
/* .fn = */ init_fn,
/* .default_buffer_type = */ default_buffer_type,
/* .user_data = */ user_data,
};
snprintf(ggml_backend_registry[id].name, sizeof(ggml_backend_registry[id].name), "%s", name);
#ifndef NDEBUG
fprintf(stderr, "%s: registered backend %s\n", __func__, name);
#endif
ggml_backend_registry_count++;
}
size_t ggml_backend_reg_get_count(void) {
ggml_backend_registry_init();
return ggml_backend_registry_count;
}
size_t ggml_backend_reg_find_by_name(const char * name) {
ggml_backend_registry_init();
for (size_t i = 0; i < ggml_backend_registry_count; i++) {
// TODO: case insensitive in a portable way
if (strcmp(ggml_backend_registry[i].name, name) == 0) {
return i;
}
}
// not found
return SIZE_MAX;
}
// init from backend:params string
ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str) {
ggml_backend_registry_init();
const char * params = strchr(backend_str, ':');
char backend_name[128];
if (params == NULL) {
snprintf(backend_name, sizeof(backend_name), "%s", backend_str);
params = "";
} else {
snprintf(backend_name, sizeof(backend_name), "%.*s", (int)(params - backend_str), backend_str);
params++;
}
size_t backend_i = ggml_backend_reg_find_by_name(backend_name);
if (backend_i == SIZE_MAX) {
fprintf(stderr, "%s: backend %s not found\n", __func__, backend_name);
return NULL;
}
return ggml_backend_reg_init_backend(backend_i, params);
}
const char * ggml_backend_reg_get_name(size_t i) {
ggml_backend_registry_init();
GGML_ASSERT(i < ggml_backend_registry_count);
return ggml_backend_registry[i].name;
}
ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params) {
ggml_backend_registry_init();
GGML_ASSERT(i < ggml_backend_registry_count);
return ggml_backend_registry[i].init_fn(params, ggml_backend_registry[i].user_data);
}
ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i) {
ggml_backend_registry_init();
GGML_ASSERT(i < ggml_backend_registry_count);
return ggml_backend_registry[i].default_buffer_type;
}
ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size) {
ggml_backend_registry_init();
GGML_ASSERT(i < ggml_backend_registry_count);
return ggml_backend_buft_alloc_buffer(ggml_backend_registry[i].default_buffer_type, size);
}
// backend CPU
static const size_t TENSOR_ALIGNMENT = 32; // required for mmap as gguf only guarantees 32-byte alignment
GGML_CALL static const char * ggml_backend_cpu_buffer_name(ggml_backend_buffer_t buffer) {
return "CPU";
GGML_UNUSED(buffer);
}
GGML_CALL static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
uintptr_t data = (uintptr_t)buffer->context;
// align the buffer
if (data % TENSOR_ALIGNMENT != 0) {
data = GGML_PAD(data, TENSOR_ALIGNMENT);
}
return (void *)data;
}
GGML_CALL static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
free(buffer->context);
}
GGML_CALL static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
memcpy((char *)tensor->data + offset, data, size);
GGML_UNUSED(buffer);
}
GGML_CALL static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
memcpy(data, (const char *)tensor->data + offset, size);
GGML_UNUSED(buffer);
}
GGML_CALL static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
if (ggml_backend_buffer_is_host(src->buffer)) {
memcpy(dst->data, src->data, ggml_nbytes(src));
return true;
}
return false;
GGML_UNUSED(buffer);
}
GGML_CALL static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
memset(buffer->context, value, buffer->size);
}
static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
/* .get_name = */ ggml_backend_cpu_buffer_name,
/* .free_buffer = */ ggml_backend_cpu_buffer_free_buffer,
/* .get_base = */ ggml_backend_cpu_buffer_get_base,
/* .init_tensor = */ NULL, // no initialization required
/* .set_tensor = */ ggml_backend_cpu_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_cpu_buffer_get_tensor,
/* .cpy_tensor = */ ggml_backend_cpu_buffer_cpy_tensor,
/* .clear = */ ggml_backend_cpu_buffer_clear,
/* .reset = */ NULL,
};
// for buffers from ptr, free is not called
static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
/* .get_name = */ ggml_backend_cpu_buffer_name,
/* .free_buffer = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
/* .get_base = */ ggml_backend_cpu_buffer_get_base,
/* .init_tensor = */ NULL, // no initialization required
/* .set_tensor = */ ggml_backend_cpu_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_cpu_buffer_get_tensor,
/* .cpy_tensor = */ ggml_backend_cpu_buffer_cpy_tensor,
/* .clear = */ ggml_backend_cpu_buffer_clear,
/* .reset = */ NULL,
};
GGML_CALL static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU";
GGML_UNUSED(buft);
}
GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
size += TENSOR_ALIGNMENT; // malloc may return an address that is not aligned
void * data = malloc(size); // TODO: use GGML_ALIGNED_MALLOC (move to ggml-impl.h)
if (data == NULL) {
fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
return NULL;
}
return ggml_backend_buffer_init(buft, cpu_backend_buffer_i, data, size);
}
GGML_CALL static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return TENSOR_ALIGNMENT;
GGML_UNUSED(buft);
}
GGML_CALL static bool ggml_backend_cpu_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
return ggml_backend_is_cpu(backend);
GGML_UNUSED(buft);
}
GGML_CALL static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
return true;
GGML_UNUSED(buft);
}
GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type;
}
#ifdef GGML_USE_CPU_HBM
// buffer type HBM
#include <hbwmalloc.h>
GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_HBM";
GGML_UNUSED(buft);
}
GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_get_name(ggml_backend_buffer_t buf) {
return "CPU_HBM";
GGML_UNUSED(buf);
}
GGML_CALL static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
hbw_free(buffer->context);
}
GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
//void * ptr = hbw_malloc(size);
void * ptr;
int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
if (result != 0) {
fprintf(stderr, "failed to allocate HBM buffer of size %zu\n", size);
return NULL;
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.get_name = ggml_backend_cpu_hbm_buffer_get_name;
buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type_hbm;
}
#endif
struct ggml_backend_cpu_context {
int n_threads;
void * work_data;
size_t work_size;
ggml_abort_callback abort_callback;
void * abort_callback_data;
};
GGML_CALL static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
return "CPU";
GGML_UNUSED(backend);
}
GGML_CALL static void ggml_backend_cpu_free(ggml_backend_t backend) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
free(cpu_ctx->work_data);
free(cpu_ctx);
free(backend);
}
GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) {
return ggml_backend_cpu_buffer_type();
GGML_UNUSED(backend);
}
struct ggml_backend_plan_cpu {
struct ggml_cplan cplan;
struct ggml_cgraph cgraph;
};
GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
cpu_plan->cgraph = *cgraph; // FIXME: deep copy
if (cpu_plan->cplan.work_size > 0) {
cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
if (cpu_plan->cplan.work_data == NULL) {
free(cpu_plan);
return NULL;
}
}
cpu_plan->cplan.abort_callback = cpu_ctx->abort_callback;
cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return cpu_plan;
}
GGML_CALL static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
free(cpu_plan->cplan.work_data);
free(cpu_plan);
GGML_UNUSED(backend);
}
GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
GGML_UNUSED(backend);
}
GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
if (cpu_ctx->work_size < cplan.work_size) {
free(cpu_ctx->work_data);
cpu_ctx->work_data = malloc(cplan.work_size);
if (cpu_ctx->work_data == NULL) {
cpu_ctx->work_size = 0;
return GGML_STATUS_ALLOC_FAILED;
}
cpu_ctx->work_size = cplan.work_size;
}
cplan.work_data = cpu_ctx->work_data;
cplan.abort_callback = cpu_ctx->abort_callback;
cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return ggml_graph_compute(cgraph, &cplan);
}
GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
switch (op->op) {
case GGML_OP_CPY:
return
op->type != GGML_TYPE_IQ2_XXS &&
op->type != GGML_TYPE_IQ2_XS &&
op->type != GGML_TYPE_IQ1_S &&
op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
case GGML_OP_MUL_MAT:
return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
default:
return true;
}
GGML_UNUSED(backend);
}
static struct ggml_backend_i cpu_backend_i = {
/* .get_name = */ ggml_backend_cpu_name,
/* .free = */ ggml_backend_cpu_free,
/* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create,
/* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free,
/* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute,
/* .graph_compute = */ ggml_backend_cpu_graph_compute,
/* .supports_op = */ ggml_backend_cpu_supports_op,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
/* .event_synchronize = */ NULL,
};
static ggml_guid_t ggml_backend_cpu_guid(void) {
static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
return &guid;
}
ggml_backend_t ggml_backend_cpu_init(void) {
struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
if (ctx == NULL) {
return NULL;
}
ctx->n_threads = GGML_DEFAULT_N_THREADS;
ctx->work_data = NULL;
ctx->work_size = 0;
ctx->abort_callback = NULL;
ctx->abort_callback_data = NULL;
ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
if (cpu_backend == NULL) {
free(ctx);
return NULL;
}
*cpu_backend = (struct ggml_backend) {
/* .guid = */ ggml_backend_cpu_guid(),
/* .interface = */ cpu_backend_i,
/* .context = */ ctx
};
return cpu_backend;
}
GGML_CALL bool ggml_backend_is_cpu(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
}
void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->n_threads = n_threads;
}
void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->abort_callback = abort_callback;
ctx->abort_callback_data = abort_callback_data;
}
GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), cpu_backend_buffer_i_from_ptr, ptr, size);
}
GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) {
return ggml_backend_cpu_init();
GGML_UNUSED(params);
GGML_UNUSED(user_data);
}
// multi-buffer buffer
struct ggml_backend_multi_buffer_context {
ggml_backend_buffer_t * buffers;
size_t n_buffers;
};
typedef struct ggml_backend_multi_buffer_context * ggml_backend_multi_buffer_context_t;
GGML_CALL static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) {
ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
return ctx->buffers[0]->iface.get_name(ctx->buffers[0]);
}
GGML_CALL static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
for (size_t i = 0; i < ctx->n_buffers; i++) {
ggml_backend_buffer_free(ctx->buffers[i]);
}
free(ctx->buffers);
free(ctx);
}
GGML_CALL static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
for (size_t i = 0; i < ctx->n_buffers; i++) {
ggml_backend_buffer_clear(ctx->buffers[i], value);
}
}
static struct ggml_backend_buffer_i ggml_backend_multi_buffer_context_interface(void) {
static struct ggml_backend_buffer_i multi_backend_buffer_i = {
/* .get_name = */ ggml_backend_multi_buffer_get_name,
/* .free_buffer = */ ggml_backend_multi_buffer_free_buffer,
/* .get_base = */ NULL,
/* .init_tensor = */ NULL,
/* .set_tensor = */ NULL,
/* .get_tensor = */ NULL,
/* .cpy_tensor = */ NULL,
/* .clear = */ ggml_backend_multi_buffer_clear,
/* .reset = */ NULL,
};
return multi_backend_buffer_i;
}
GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) {
ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) malloc(sizeof(struct ggml_backend_multi_buffer_context));
ctx->n_buffers = n_buffers;
ctx->buffers = (ggml_backend_buffer_t *) malloc(n_buffers * sizeof(ggml_backend_buffer_t));
GGML_ASSERT(ctx->buffers != NULL);
size_t total_size = 0;
for (size_t i = 0; i < n_buffers; i++) {
ctx->buffers[i] = buffers[i];
total_size += ggml_backend_buffer_get_size(buffers[i]);
}
return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_context_interface(), ctx, total_size);
}
GGML_CALL bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) {
return buffer->iface.get_name == ggml_backend_multi_buffer_get_name;
}
GGML_CALL void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
GGML_ASSERT(ggml_backend_buffer_is_multi_buffer(buffer));
ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
for (size_t i = 0; i < ctx->n_buffers; i++) {
ggml_backend_buffer_set_usage(ctx->buffers[i], usage);
}
}
// creates a copy of the tensor with the same memory layout
static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, const struct ggml_tensor * tensor) {
struct ggml_tensor * dup = ggml_dup_tensor(ctx, tensor);
for (int i = 0; i < GGML_MAX_DIMS; i++) {
dup->nb[i] = tensor->nb[i];
}
return dup;
}
static bool ggml_is_view_op(enum ggml_op op) {
return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE;
}
// scheduler
#ifndef GGML_SCHED_MAX_BACKENDS
#define GGML_SCHED_MAX_BACKENDS 16
#endif
#ifndef GGML_SCHED_MAX_SPLITS
#define GGML_SCHED_MAX_SPLITS 2048
#endif
#ifndef GGML_SCHED_MAX_SPLIT_INPUTS
#define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC
#endif
#ifndef GGML_SCHED_MAX_COPIES
#define GGML_SCHED_MAX_COPIES 4
#endif
struct ggml_backend_sched_split {
int backend_id;
int i_start;
int i_end;
struct ggml_tensor * inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
int n_inputs;
// graph view of this split
struct ggml_cgraph graph;
};
struct ggml_backend_sched {
bool is_reset; // true if the scheduler has been reset since the last graph split
bool is_alloc;
int n_backends;
ggml_backend_t backends[GGML_SCHED_MAX_BACKENDS];
ggml_backend_buffer_type_t bufts[GGML_SCHED_MAX_BACKENDS];
ggml_gallocr_t galloc;
// hash keys of the nodes in the graph
struct ggml_hash_set hash_set;
// hash values
int * tensor_backend_id;
struct ggml_tensor * (* tensor_copies)[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
int * node_backend_ids; // [graph_size]
int * leaf_backend_ids; // [graph_size]
// copy of the graph with modified inputs
struct ggml_cgraph * graph;
// graph splits
struct ggml_backend_sched_split * splits;
int n_splits;
int splits_capacity;
// pipeline parallelism support
int n_copies;
int cur_copy;
ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
int n_graph_inputs;
struct ggml_context * ctx;
ggml_backend_sched_eval_callback callback_eval;
void * callback_eval_user_data;
// align context_buffer to GGML_MEM_ALIGN
#ifdef _MSC_VER
__declspec(align(GGML_MEM_ALIGN))
#else
__attribute__((aligned(GGML_MEM_ALIGN)))
#endif
char context_buffer[GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)];
};
#define hash_id(tensor) ggml_hash_find_or_insert(sched->hash_set, tensor)
#define tensor_backend_id(tensor) sched->tensor_backend_id[hash_id(tensor)]
// returns the priority of the backend, lower id is higher priority
static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) {
for (int i = 0; i < sched->n_backends; i++) {
if (sched->backends[i] == backend) {
return i;
}
}
return -1;
}
static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor) {
ggml_backend_buffer_t buffer = tensor->buffer;
if (buffer == NULL) {
return -1;
}
// find highest prio backend that supports the buffer type
for (int i = 0; i < sched->n_backends; i++) {
if (ggml_backend_buft_supports_backend(buffer->buft, sched->backends[i])) {
return i;
}
}
fprintf(stderr, "%s: error: no backend supports buffer type %s used in tensor %s\n",
__func__, ggml_backend_buffer_name(buffer), tensor->name);
GGML_ASSERT(false);
return -1;
}
#if 0
static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
#define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
#define GET_CAUSE(node) causes[hash_id(node)]
#else
#define SET_CAUSE(node, ...)
#define GET_CAUSE(node) ""
#endif
// returns the backend that should be used for the node based on the current locations
static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) {
// TODO: use supports_op to check if the backend supports the op
// assign pre-allocated nodes to their backend
int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor);
if (cur_backend_id != -1) {
SET_CAUSE(tensor, "1.dst");
return cur_backend_id;
}
// view_src
if (tensor->view_src != NULL) {
cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src);
if (cur_backend_id != -1) {
SET_CAUSE(tensor, "1.vsrc");
return cur_backend_id;
}
}
// graph input
if (tensor->flags & GGML_TENSOR_FLAG_INPUT) {
cur_backend_id = sched->n_backends - 1; // last backend (assumed CPU)
SET_CAUSE(tensor, "1.inp");
return cur_backend_id;
}
// assign nodes that use weights to the backend of the weights
// operations with weights are preferably run on the same backend as the weights
for (int i = 0; i < GGML_MAX_SRC; i++) {
const struct ggml_tensor * src = tensor->src[i];
if (src == NULL) {
continue;
}
if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src);
// check if a backend with higher prio wants to offload the op
if (src_backend_id == sched->n_backends - 1) {
for (int b = 0; b < src_backend_id; b++) {
if (ggml_backend_offload_op(sched->backends[b], tensor)) {
SET_CAUSE(tensor, "1.off");
return b;
}
}
}
SET_CAUSE(tensor, "1.wgt%d", i);
return src_backend_id;
}
}
return -1;
}
static char * fmt_size(size_t size) {
static char buffer[128];
if (size >= 1024*1024) {
snprintf(buffer, sizeof(buffer), "%zuM", size/1024/1024);
} else {
snprintf(buffer, sizeof(buffer), "%zuK", size/1024);
}
return buffer;
}
static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
int cur_split = 0;
for (int i = 0; i < graph->n_nodes; i++) {
if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
ggml_backend_t split_backend = sched->backends[sched->splits[cur_split].backend_id];
fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend),
sched->splits[cur_split].n_inputs);
for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
}
fprintf(stderr, "\n");
cur_split++;
}
struct ggml_tensor * node = graph->nodes[i];
if (ggml_is_view_op(node->op)) {
continue;
}
ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
fprintf(stderr, "node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
continue;
}
ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
fprintf(stderr, " %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
}
fprintf(stderr, "\n");
}
}
//#define DEBUG_PASS1
//#define DEBUG_PASS2
//#define DEBUG_PASS3
//#define DEBUG_PASS4
// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
// reset splits
sched->n_splits = 0;
sched->n_graph_inputs = 0;
sched->is_reset = false;
struct ggml_init_params params = {
/* .mem_size = */ sizeof(sched->context_buffer),
/* .mem_buffer = */ sched->context_buffer,
/* .no_alloc = */ true
};
ggml_free(sched->ctx);
sched->ctx = ggml_init(params);
if (sched->ctx == NULL) {
fprintf(stderr, "%s: failed to initialize context\n", __func__);
GGML_ASSERT(false);
}
// pass 1: assign backends to ops with pre-allocated inputs
for (int i = 0; i < graph->n_leafs; i++) {
struct ggml_tensor * leaf = graph->leafs[i];
int * leaf_backend_id = &tensor_backend_id(leaf);
if (*leaf_backend_id != -1) {
// do not overwrite user assignments
continue;
}
*leaf_backend_id = ggml_backend_sched_backend_id_from_cur(sched, leaf);
}
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
int * node_backend_id = &tensor_backend_id(node);
if (*node_backend_id != -1) {
// do not overwrite user assignments
continue;
}
*node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
// src
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
continue;
}
int * src_backend_id = &tensor_backend_id(src);
if (*src_backend_id == -1) {
*src_backend_id = ggml_backend_sched_backend_id_from_cur(sched, src);
}
}
}
#ifdef DEBUG_PASS1
fprintf(stderr, "PASS 1 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
#endif
// pass 2: expand current backend assignments
// assign the same backend to adjacent nodes
// expand gpu backends (i.e. non last prio) up and down, ignoring cpu (the lowest priority backend)
// thus, cpu will never be used unless weights are on cpu, or there are no gpu ops between cpu ops
// pass 2.2 expand gpu down
{
int cur_backend_id = -1;
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
if (ggml_is_view_op(node->op)) {
continue;
}
int * node_backend_id = &tensor_backend_id(node);
if (*node_backend_id != -1) {
if (*node_backend_id == sched->n_backends - 1) {
// skip cpu (lowest prio backend)
cur_backend_id = -1;
} else {
cur_backend_id = *node_backend_id;
}
} else {
*node_backend_id = cur_backend_id;
SET_CAUSE(node, "2.2");
}
}
}
// pass 2.1 expand gpu up
{
int cur_backend_id = -1;
for (int i = graph->n_nodes - 1; i >= 0; i--) {
struct ggml_tensor * node = graph->nodes[i];
if (ggml_is_view_op(node->op)) {
continue;
}
int * node_backend_id = &tensor_backend_id(node);
if (*node_backend_id != -1) {
if (*node_backend_id == sched->n_backends - 1) {
// skip cpu (lowest prio backend)
cur_backend_id = -1;
} else {
cur_backend_id = *node_backend_id;
}
} else {
*node_backend_id = cur_backend_id;
SET_CAUSE(node, "2.1");
}
}
}
// pass 2.4 expand rest down
{
int cur_backend_id = -1;
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
if (ggml_is_view_op(node->op)) {
continue;
}
int * node_backend_id = &tensor_backend_id(node);
if (*node_backend_id != -1) {
cur_backend_id = *node_backend_id;
} else {
*node_backend_id = cur_backend_id;
SET_CAUSE(node, "2.4");
}
}
}
// pass 2.3 expand rest up
{
int cur_backend_id = -1;
for (int i = graph->n_nodes - 1; i >= 0; i--) {
struct ggml_tensor * node = graph->nodes[i];
if (ggml_is_view_op(node->op)) {
continue;
}
int * node_backend_id = &tensor_backend_id(node);
if (*node_backend_id != -1) {
cur_backend_id = *node_backend_id;
} else {
*node_backend_id = cur_backend_id;
SET_CAUSE(node, "2.3");
}
}
}
#ifdef DEBUG_PASS2
fprintf(stderr, "PASS 2 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
#endif
// pass 3: assign backends to remaining src from dst and view_src
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
int * cur_backend_id = &tensor_backend_id(node);
if (node->view_src != NULL && *cur_backend_id == -1) {
*cur_backend_id = tensor_backend_id(node->view_src);
SET_CAUSE(node, "3.vsrc");
}
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
continue;
}
int * src_backend_id = &tensor_backend_id(src);
if (*src_backend_id == -1) {
if (src->view_src != NULL) {
// views are always on the same backend as the source
*src_backend_id = tensor_backend_id(src->view_src);
SET_CAUSE(src, "3.vsrc");
} else {
*src_backend_id = *cur_backend_id;
SET_CAUSE(src, "3.cur");
}
}
}
}
#ifdef DEBUG_PASS3
fprintf(stderr, "PASS 3 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
#endif
// pass 4: split graph, find tensors that need to be copied
{
int i_split = 0;
struct ggml_backend_sched_split * split = &sched->splits[0];
// find the backend of the first split, skipping view ops
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
if (!ggml_is_view_op(node->op)) {
split->backend_id = tensor_backend_id(node);
break;
}
}
split->i_start = 0;
split->n_inputs = 0;
memset(split->inputs, 0, sizeof(split->inputs)); //HACK
int cur_backend_id = split->backend_id;
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
if (ggml_is_view_op(node->op)) {
continue;
}
const int node_backend_id = tensor_backend_id(node);
GGML_ASSERT(node_backend_id != -1); // all nodes should be assigned by now
// check if we should start a new split based on the sources of the current node
bool need_new_split = false;
if (node_backend_id == cur_backend_id && split->n_inputs > 0) {
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
continue;
}
// check if a weight is on a different backend
// by starting a new split, the memory of the previously offloaded weights can be reused
if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
int src_backend_id = tensor_backend_id(src);
if (src_backend_id != -1 && src_backend_id != cur_backend_id) {
need_new_split = true;
break;
}
}
// check if the split has too many inputs
if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
const size_t id = hash_id(src);
int src_backend_id = sched->tensor_backend_id[id];
if (src_backend_id != cur_backend_id && sched->tensor_copies[hash_id(src)][cur_backend_id][0] == NULL) {
//printf("starting new split because of too many inputs: node %s, input %s\n", node->name, src->name);
need_new_split = true;
break;
}
}
}
}
if (node_backend_id != cur_backend_id || need_new_split) {
split->i_end = i;
i_split++;
if (i_split >= sched->splits_capacity) {
sched->splits_capacity *= 2;
sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
GGML_ASSERT(sched->splits != NULL);
}
GGML_ASSERT(i_split < GGML_SCHED_MAX_SPLITS);
split = &sched->splits[i_split];
split->backend_id = node_backend_id;
split->i_start = i;
split->n_inputs = 0;
cur_backend_id = node_backend_id;
}
// find inputs that are not on the same backend
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
continue;
}
const int src_backend_id = tensor_backend_id(src);
assert(src_backend_id != -1); // all inputs should be assigned by now
if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1) {
size_t id = hash_id(src);
if (sched->tensor_copies[id][src_backend_id][0] == NULL) {
ggml_backend_t backend = sched->backends[src_backend_id];
for (int c = 0; c < sched->n_copies; c++) {
struct ggml_tensor * tensor_copy;
if (c == sched->cur_copy) {
tensor_copy = src; // use the original tensor as the current copy
} else {
tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
}
if (sched->n_copies > 1) {
ggml_set_input(tensor_copy);
ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
}
sched->tensor_copies[id][src_backend_id][c] = tensor_copy;
SET_CAUSE(tensor_copy, "4.cpy");
}
int n_graph_inputs = sched->n_graph_inputs++;
GGML_ASSERT(n_graph_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
sched->graph_inputs[n_graph_inputs] = src;
}
}
if (src_backend_id != node_backend_id) {
// create a copy of the input in the split's backend
const size_t id = hash_id(src);
if (sched->tensor_copies[id][cur_backend_id][0] == NULL) {
ggml_backend_t backend = sched->backends[cur_backend_id];
for (int c = 0; c < sched->n_copies; c++) {
struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
if (sched->n_copies > 1) {
ggml_set_input(tensor_copy);
ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
}
sched->tensor_copies[id][cur_backend_id][c] = tensor_copy;
SET_CAUSE(tensor_copy, "4.cpy");
}
int n_inputs = split->n_inputs++;
GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
split->inputs[n_inputs] = src;
}
node->src[j] = sched->tensor_copies[id][cur_backend_id][sched->cur_copy];
}
}
}
split->i_end = graph->n_nodes;
sched->n_splits = i_split + 1;
}
#ifdef DEBUG_PASS4
fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
#endif
// create copies of the graph for each split
// TODO: avoid this copy
struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2, false);
for (int i = 0; i < sched->n_splits; i++) {
struct ggml_backend_sched_split * split = &sched->splits[i];
split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
// add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
for (int j = 0; j < split->n_inputs; j++) {
assert(graph_copy->size > (graph_copy->n_nodes + 1));
struct ggml_tensor * input = split->inputs[j];
const size_t input_id = hash_id(input);
struct ggml_tensor * input_cpy = sched->tensor_copies[input_id][split->backend_id][sched->cur_copy];
// add a dependency to the input source so that it is not freed before the copy is done
struct ggml_tensor * input_dep = ggml_view_tensor(sched->ctx, input);
input_dep->src[0] = input;
sched->node_backend_ids[graph_copy->n_nodes] = sched->tensor_backend_id[input_id];
graph_copy->nodes[graph_copy->n_nodes++] = input_dep;
// add a dependency to the input copy so that it is allocated at the start of the split
sched->node_backend_ids[graph_copy->n_nodes] = split->backend_id;
graph_copy->nodes[graph_copy->n_nodes++] = input_cpy;
}
for (int j = split->i_start; j < split->i_end; j++) {
assert(graph_copy->size > graph_copy->n_nodes);
sched->node_backend_ids[graph_copy->n_nodes] = tensor_backend_id(graph->nodes[j]);
graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
}
}
if (sched->n_copies > 1) {
// add input copies as leafs so that they are allocated first
for (int i = 0; i < sched->n_graph_inputs; i++) {
struct ggml_tensor * input = sched->graph_inputs[i];
size_t id = hash_id(input);
int backend_id = tensor_backend_id(input);
for (int c = 0; c < sched->n_copies; c++) {
struct ggml_tensor * input_cpy = sched->tensor_copies[id][backend_id][c];
sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
}
}
for (int i = 0; i < sched->n_splits; i++) {
struct ggml_backend_sched_split * split = &sched->splits[i];
int backend_id = split->backend_id;
for (int j = 0; j < split->n_inputs; j++) {
struct ggml_tensor * input = split->inputs[j];
size_t id = hash_id(input);
for (int c = 0; c < sched->n_copies; c++) {
struct ggml_tensor * input_cpy = sched->tensor_copies[id][backend_id][c];
sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
}
}
}
}
// add leafs from the original graph
for (int i = 0; i < graph->n_leafs; i++) {
struct ggml_tensor * leaf = graph->leafs[i];
sched->leaf_backend_ids[graph_copy->n_leafs] = tensor_backend_id(leaf);
graph_copy->leafs[graph_copy->n_leafs++] = leaf;
}
sched->graph = graph_copy;
}
static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
// allocate graph
if (!ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) {
// the re-allocation may cause the split inputs to be moved to a different address
ggml_backend_sched_synchronize(sched);
#ifndef NDEBUG
fprintf(stderr, "%s: failed to allocate graph, reserving\n", __func__);
#endif
ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
if (!ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) {
fprintf(stderr, "%s: failed to allocate graph\n", __func__);
return false;
}
}
return true;
}
static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
struct ggml_backend_sched_split * splits = sched->splits;
for (int i = 0; i < sched->n_splits; i++) {
struct ggml_backend_sched_split * split = &splits[i];
int split_backend_id = split->backend_id;
ggml_backend_t split_backend = sched->backends[split_backend_id];
// copy the input tensors to the split backend
for (int j = 0; j < split->n_inputs; j++) {
ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]);
struct ggml_tensor * input = split->inputs[j];
struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split_backend_id][sched->cur_copy];
if (input->flags & GGML_TENSOR_FLAG_INPUT) {
// inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
} else {
ggml_backend_synchronize(split_backend);
}
ggml_backend_tensor_copy(input, input_cpy);
} else {
// wait for the split backend to finish using the input before overwriting it
if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
} else {
ggml_backend_synchronize(split_backend);
}
ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
}
}
if (!sched->callback_eval) {
enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
if (ec != GGML_STATUS_SUCCESS) {
return ec;
}
} else {
// similar to ggml_backend_compare_graph_backend
for (int j0 = 0; j0 < split->graph.n_nodes; j0++) {
struct ggml_tensor * t = split->graph.nodes[j0];
// check if the user needs data from this node
bool need = sched->callback_eval(t, true, sched->callback_eval_user_data);
int j1 = j0;
// determine the range [j0, j1] of nodes that can be computed together
while (!need && j1 < split->graph.n_nodes - 1) {
t = split->graph.nodes[++j1];
need = sched->callback_eval(t, true, sched->callback_eval_user_data);
}
struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1);
enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &gv);
if (ec != GGML_STATUS_SUCCESS) {
return ec;
}
// TODO: pass backend to the callback, then the user can decide if they want to synchronize
ggml_backend_synchronize(split_backend);
if (need && !sched->callback_eval(t, false, sched->callback_eval_user_data)) {
break;
}
j0 = j1;
}
}
// record the event of this copy
if (split->n_inputs > 0) {
if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy]);
}
}
}
sched->cur_copy = (sched->cur_copy + 1) % sched->n_copies;
return GGML_STATUS_SUCCESS;
}
ggml_backend_sched_t ggml_backend_sched_new(
ggml_backend_t * backends,
ggml_backend_buffer_type_t * bufts,
int n_backends,
size_t graph_size,
bool parallel) {
GGML_ASSERT(n_backends > 0);
GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched));
// initialize hash table
sched->hash_set = ggml_hash_set_new(graph_size);
sched->tensor_backend_id = calloc(sched->hash_set.size, sizeof(sched->tensor_backend_id[0]));
sched->tensor_copies = calloc(sched->hash_set.size, sizeof(sched->tensor_copies[0]));
const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
sched->node_backend_ids = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
sched->leaf_backend_ids = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
sched->n_backends = n_backends;
sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
const int initial_splits_capacity = 16;
sched->splits = calloc(initial_splits_capacity, sizeof(sched->splits[0]));
sched->splits_capacity = initial_splits_capacity;
for (int b = 0; b < n_backends; b++) {
sched->backends[b] = backends[b];
sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
GGML_ASSERT(ggml_backend_buft_supports_backend(sched->bufts[b], backends[b]));
if (sched->n_copies > 1) {
for (int c = 0; c < sched->n_copies; c++) {
sched->events[b][c] = ggml_backend_event_new(backends[b]);
}
}
}
sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
ggml_backend_sched_reset(sched);
return sched;
}
void ggml_backend_sched_free(ggml_backend_sched_t sched) {
if (sched == NULL) {
return;
}
for (int b = 0; b < sched->n_backends; b++) {
for (int c = 0; c < sched->n_copies; c++) {
ggml_backend_event_free(sched->events[b][c]);
}
}
ggml_gallocr_free(sched->galloc);
ggml_free(sched->ctx);
free(sched->splits);
free(sched->hash_set.keys);
free(sched->tensor_backend_id);
free(sched->tensor_copies);
free(sched->node_backend_ids);
free(sched->leaf_backend_ids);
free(sched);
}
void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
// reset state for the next run
if (!sched->is_reset) {
size_t hash_size = sched->hash_set.size;
memset(sched->hash_set.keys, 0, sizeof(sched->hash_set.keys[0]) * hash_size); // NOLINT
memset(sched->tensor_backend_id, -1, sizeof(sched->tensor_backend_id[0]) * hash_size);
memset(sched->tensor_copies, 0, sizeof(sched->tensor_copies[0]) * hash_size);
sched->is_reset = true;
}
sched->is_alloc = false;
}
bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes);
ggml_backend_sched_split_graph(sched, measure_graph);
// TODO: extract this to a separate function
if (!ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
return false;
}
ggml_backend_sched_reset(sched);
ggml_backend_sched_synchronize(sched);
return true;
}
bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes);
ggml_backend_sched_split_graph(sched, graph);
if (!ggml_backend_sched_alloc_splits(sched)) {
return false;
}
sched->is_alloc = true;
return true;
}
enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
enum ggml_status err = ggml_backend_sched_graph_compute_async(sched, graph);
ggml_backend_sched_synchronize(sched);
return err;
}
enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
if (!sched->is_reset && !sched->is_alloc) {
ggml_backend_sched_reset(sched);
}
if (!sched->is_alloc) {
if (!ggml_backend_sched_alloc_graph(sched, graph)) {
return GGML_STATUS_ALLOC_FAILED;
}
}
return ggml_backend_sched_compute_splits(sched);
}
void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
for (int i = 0; i < sched->n_backends; i++) {
ggml_backend_synchronize(sched->backends[i]);
}
}
void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {
sched->callback_eval = callback;
sched->callback_eval_user_data = user_data;
}
int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched) {
return sched->n_splits;
}
int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched) {
return sched->n_copies;
}
size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
int backend_index = ggml_backend_sched_backend_id(sched, backend);
GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
return ggml_gallocr_get_buffer_size(sched->galloc, backend_index);
}
void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
int backend_index = ggml_backend_sched_backend_id(sched, backend);
GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
tensor_backend_id(node) = backend_index;
}
ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) {
int backend_index = tensor_backend_id(node);
if (backend_index == -1) {
return NULL;
}
return sched->backends[backend_index];
}
// utils
void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
GGML_ASSERT(tensor->buffer == NULL);
GGML_ASSERT(tensor->view_src != NULL);
GGML_ASSERT(tensor->view_src->buffer != NULL);
GGML_ASSERT(tensor->view_src->data != NULL);
tensor->buffer = buffer;
tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
ggml_backend_buffer_init_tensor(buffer, tensor);
}
void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) {
GGML_ASSERT(tensor->buffer == NULL);
GGML_ASSERT(tensor->data == NULL);
GGML_ASSERT(tensor->view_src == NULL);
GGML_ASSERT(addr >= ggml_backend_buffer_get_base(buffer));
GGML_ASSERT((char *)addr + ggml_backend_buffer_get_alloc_size(buffer, tensor) <=
(char *)ggml_backend_buffer_get_base(buffer) + ggml_backend_buffer_get_size(buffer));
tensor->buffer = buffer;
tensor->data = addr;
ggml_backend_buffer_init_tensor(buffer, tensor);
}
static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies,
struct ggml_context * ctx_allocated, struct ggml_context * ctx_unallocated, struct ggml_tensor * src) {
GGML_ASSERT(src != NULL);
GGML_ASSERT(src->data && "graph must be allocated");
size_t id = ggml_hash_insert(hash_set, src);
if (id == GGML_HASHTABLE_ALREADY_EXISTS) {
return node_copies[ggml_hash_find(hash_set, src)];
}
struct ggml_tensor * dst = ggml_dup_tensor_layout(src->data && !src->view_src ? ctx_allocated : ctx_unallocated, src);
if (src->view_src != NULL) {
dst->view_src = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, src->view_src);
dst->view_offs = src->view_offs;
}
dst->op = src->op;
memcpy(dst->op_params, src->op_params, sizeof(dst->op_params));
ggml_set_name(dst, src->name);
// copy src
for (int i = 0; i < GGML_MAX_SRC; i++) {
struct ggml_tensor * s = src->src[i];
if (s == NULL) {
continue;
}
dst->src[i] = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, s);
}
node_copies[id] = dst;
return dst;
}
static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies, bool * node_init, struct ggml_tensor * src) {
size_t id = ggml_hash_find(hash_set, src);
if (node_init[id]) {
return;
}
node_init[id] = true;
struct ggml_tensor * dst = node_copies[id];
if (dst->view_src != NULL) {
graph_copy_init_tensor(hash_set, node_copies, node_init, src->view_src);
ggml_backend_view_init(dst->view_src->buffer, dst);
}
else {
ggml_backend_tensor_copy(src, dst);
}
// init src
for (int i = 0; i < GGML_MAX_SRC; i++) {
struct ggml_tensor * s = src->src[i];
if (s == NULL) {
continue;
}
graph_copy_init_tensor(hash_set, node_copies, node_init, s);
}
}
struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
struct ggml_hash_set hash_set = {
/* .size = */ graph->visited_hash_table.size,
/* .keys = */ calloc(graph->visited_hash_table.size, sizeof(hash_set.keys[0])) // NOLINT
};
struct ggml_tensor ** node_copies = calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
bool * node_init = calloc(hash_set.size, sizeof(node_init[0]));
struct ggml_init_params params = {
/* .mem_size = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),
/* .mem_buffer = */ NULL,
/* .no_alloc = */ true
};
struct ggml_context * ctx_allocated = ggml_init(params);
struct ggml_context * ctx_unallocated = ggml_init(params);
if (ctx_allocated == NULL || ctx_unallocated == NULL) {
fprintf(stderr, "failed to allocate context for graph copy\n");
free(hash_set.keys);
free(node_copies);
free(node_init);
ggml_free(ctx_allocated);
ggml_free(ctx_unallocated);
return (struct ggml_backend_graph_copy) {
/* .buffer = */ NULL,
/* .ctx_allocated = */ NULL,
/* .ctx_unallocated = */ NULL,
/* .graph = */ NULL,
};
}
// dup nodes
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, node);
}
// allocate nodes
ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx_allocated, backend);
if (buffer == NULL) {
fprintf(stderr, "failed to allocate buffer for graph copy\n");
free(hash_set.keys);
free(node_copies);
free(node_init);
ggml_free(ctx_allocated);
ggml_free(ctx_unallocated);
return (struct ggml_backend_graph_copy) {
/* .buffer = */ NULL,
/* .ctx_allocated = */ NULL,
/* .ctx_unallocated = */ NULL,
/* .graph = */ NULL,
};
}
//printf("copy buffer size: %zu MB\n", ggml_backend_buffer_get_size(buffer) / 1024 / 1024);
// copy data and init views
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
graph_copy_init_tensor(hash_set, node_copies, node_init, node);
}
// build graph copy
struct ggml_cgraph * graph_copy = ggml_new_graph_custom(ctx_allocated, graph->size, false);
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
struct ggml_tensor * node_copy = node_copies[ggml_hash_find(hash_set, node)];
graph_copy->nodes[i] = node_copy;
}
graph_copy->n_nodes = graph->n_nodes;
free(hash_set.keys);
free(node_copies);
free(node_init);
return (struct ggml_backend_graph_copy) {
/* .buffer = */ buffer,
/* .ctx_allocated = */ ctx_allocated,
/* .ctx_unallocated = */ ctx_unallocated,
/* .graph = */ graph_copy,
};
}
void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
ggml_backend_buffer_free(copy.buffer);
ggml_free(copy.ctx_allocated);
ggml_free(copy.ctx_unallocated);
}
bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
if (copy.buffer == NULL) {
return false;
}
struct ggml_cgraph * g1 = graph;
struct ggml_cgraph * g2 = copy.graph;
assert(g1->n_nodes == g2->n_nodes);
for (int i = 0; i < g1->n_nodes; i++) {
//printf("eval %d/%d\n", i, g1->n_nodes);
struct ggml_tensor * t1 = g1->nodes[i];
struct ggml_tensor * t2 = g2->nodes[i];
assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
ggml_backend_graph_compute(backend1, &g1v);
ggml_backend_graph_compute(backend2, &g2v);
if (ggml_is_view_op(t1->op)) {
continue;
}
// compare results, calculate rms etc
if (!callback(i, t1, t2, user_data)) {
break;
}
}
ggml_backend_graph_copy_free(copy);
return true;
}
#pragma once
#include "ggml.h"
#include "ggml-alloc.h"
#ifdef __cplusplus
extern "C" {
#endif
typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
typedef struct ggml_backend_event * ggml_backend_event_t;
typedef struct ggml_backend * ggml_backend_t;
typedef void * ggml_backend_graph_plan_t;
//
// Backend buffer
//
// buffer type
GGML_API const char * ggml_backend_buft_name (ggml_backend_buffer_type_t buft);
GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer (ggml_backend_buffer_type_t buft, size_t size);
GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
GGML_API size_t ggml_backend_buft_get_max_size (ggml_backend_buffer_type_t buft);
GGML_API GGML_CALL size_t ggml_backend_buft_get_alloc_size (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
GGML_API bool ggml_backend_buft_is_host (ggml_backend_buffer_type_t buft);
// buffer
enum ggml_backend_buffer_usage {
GGML_BACKEND_BUFFER_USAGE_ANY = 0,
GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
};
GGML_API const char * ggml_backend_buffer_name (ggml_backend_buffer_t buffer);
GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer);
GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_size (ggml_backend_buffer_t buffer);
GGML_API GGML_CALL void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_max_size (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
GGML_API void ggml_backend_buffer_clear (ggml_backend_buffer_t buffer, uint8_t value);
GGML_API bool ggml_backend_buffer_is_host (ggml_backend_buffer_t buffer);
GGML_API void ggml_backend_buffer_set_usage (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_get_type (ggml_backend_buffer_t buffer);
GGML_API void ggml_backend_buffer_reset (ggml_backend_buffer_t buffer);
//
// Backend
//
GGML_API ggml_guid_t ggml_backend_guid(ggml_backend_t backend);
GGML_API const char * ggml_backend_name(ggml_backend_t backend);
GGML_API void ggml_backend_free(ggml_backend_t backend);
GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
GGML_API size_t ggml_backend_get_max_size(ggml_backend_t backend);
GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
GGML_API GGML_CALL void ggml_backend_tensor_set( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
GGML_API void ggml_backend_graph_plan_free (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
GGML_API enum ggml_status ggml_backend_graph_compute (ggml_backend_t backend, struct ggml_cgraph * cgraph);
GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
// tensor copy between different backends
GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
// asynchronous copy
// the copy is performed after all the currently queued operations in backend_src
// backend_dst will wait for the copy to complete before performing other operations
// automatic fallback to sync copy if async is not supported
GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
// events
GGML_API ggml_backend_event_t ggml_backend_event_new (ggml_backend_t backend);
GGML_API void ggml_backend_event_free (ggml_backend_event_t event);
GGML_API void ggml_backend_event_record (ggml_backend_event_t event);
GGML_API void ggml_backend_event_synchronize(ggml_backend_event_t event);
GGML_API void ggml_backend_event_wait (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event
//
// CPU backend
//
GGML_API ggml_backend_t ggml_backend_cpu_init(void);
GGML_API GGML_CALL bool ggml_backend_is_cpu (ggml_backend_t backend);
GGML_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads);
GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
// Create a backend buffer from an existing pointer
GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
#ifdef GGML_USE_CPU_HBM
GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
#endif
//
// Backend registry
//
// The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
GGML_API size_t ggml_backend_reg_get_count(void);
GGML_API size_t ggml_backend_reg_find_by_name(const char * name);
GGML_API ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
GGML_API const char * ggml_backend_reg_get_name(size_t i);
GGML_API ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
GGML_API ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size);
//
// Backend scheduler
//
// The backend scheduler allows for multiple backends to be used together
// Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
// The backends are selected based on:
// - the backend that supports the operation
// - the location of the pre-allocated tensors (e.g. the weights)
/*
Example usage:
// operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
// preferrably to run on the same backend as the buffer
ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
// initialize buffers from a max size graph (optional)
reserve_graph = build_graph(sched, max_batch_size);
// manually assign nodes to a backend (optional, should not be needed in most cases)
struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);
ggml_backend_sched_reserve(sched, reserve_graph);
// compute
graph = build_graph(sched);
ggml_backend_sched_graph_compute(sched, graph);
// if there are graph inputs:
ggml_backend_sched_reset(sched);
ggml_backend_sched_alloc_graph(sched, graph);
ggml_backend_tensor_set(input_tensor, ...);
ggml_backend_sched_graph_compute(sched, graph);
}
*/
struct ggml_backend_sched;
typedef struct ggml_backend_sched * ggml_backend_sched_t;
// when ask == true, the scheduler wants to know if the user wants to observe this node
// this allows the scheduler to batch nodes together in order to evaluate them in a single call
//
// when ask == false, the scheduler is passing the node tensor to the user for observation
// if the user returns false, the scheduler will cancel the graph compute
//
typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
// Initialize a backend scheduler
GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
// Initialize backend buffers from a measure graph
GGML_API bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
// Get the number of splits of the last graph
GGML_API int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
GGML_API int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
GGML_API size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
GGML_API void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
GGML_API ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
// Allocate and compute graph on the backend scheduler
GGML_API bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
GGML_API enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
GGML_API enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
GGML_API void ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
// Reset all assignments and allocators - must be called before changing the node backends
GGML_API void ggml_backend_sched_reset(ggml_backend_sched_t sched);
// Set a callback to be called for each resulting node during graph compute
GGML_API void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
//
// Utils
//
struct ggml_backend_graph_copy {
ggml_backend_buffer_t buffer;
struct ggml_context * ctx_allocated;
struct ggml_context * ctx_unallocated;
struct ggml_cgraph * graph;
};
// Copy a graph to a different backend
GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
GGML_API void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
// Compare the output of two backends
GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
// Tensor initialization
GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
#ifdef __cplusplus
}
#endif
#ifndef GGML_COMMON_DECL
#if defined(GGML_COMMON_DECL_C)
#include <stdint.h>
typedef uint16_t ggml_half;
typedef uint32_t ggml_half2;
#define GGML_COMMON_AGGR
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_METAL)
#include <metal_stdlib>
typedef half ggml_half;
typedef half2 ggml_half2;
#define GGML_COMMON_AGGR
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_CUDA)
#include <cuda_fp16.h>
#include <cstdint>
typedef half ggml_half;
typedef half2 ggml_half2;
#define GGML_COMMON_AGGR data
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_HIP)
#include <hip/hip_fp16.h>
#include <cstdint>
typedef half ggml_half;
typedef half2 ggml_half2;
#define GGML_COMMON_AGGR data
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_SYCL)
#include <sycl/half_type.hpp>
#include <cstdint>
typedef sycl::half ggml_half;
typedef sycl::half2 ggml_half2;
#define GGML_COMMON_AGGR data
#define GGML_COMMON_DECL
#endif
#if defined(GGML_COMMON_DECL)
#ifndef __cplusplus
#ifndef static_assert
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
#define static_assert(cond, msg) _Static_assert(cond, msg)
#else
#define static_assert(cond, msg) struct global_scope_noop_trick
#endif
#endif
#endif // __cplusplus
// QK = number of values after dequantization
// QK_K = super-block size
#define QK_K 256
#define K_SCALE_SIZE 12
#if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
// QR = QK / number of values before dequantization
// QI = number of 32 bit integers before dequantization
#define QI4_0 (QK4_0 / (4 * QR4_0))
#define QR4_0 2
#define QI4_1 (QK4_1 / (4 * QR4_1))
#define QR4_1 2
#define QI5_0 (QK5_0 / (4 * QR5_0))
#define QR5_0 2
#define QI5_1 (QK5_1 / (4 * QR5_1))
#define QR5_1 2
#define QI8_0 (QK8_0 / (4 * QR8_0))
#define QR8_0 1
#define QI8_1 (QK8_1 / (4 * QR8_1))
#define QR8_1 1
#define QI2_K (QK_K / (4*QR2_K))
#define QR2_K 4
#define QI3_K (QK_K / (4*QR3_K))
#define QR3_K 4
#define QI4_K (QK_K / (4*QR4_K))
#define QR4_K 2
#define QI5_K (QK_K / (4*QR5_K))
#define QR5_K 2
#define QI6_K (QK_K / (4*QR6_K))
#define QR6_K 2
#define QI2_XXS (QK_K / (4*QR2_XXS))
#define QR2_XXS 8
#define QI2_XS (QK_K / (4*QR2_XS))
#define QR2_XS 8
#define QI2_S (QK_K / (4*QR2_S))
#define QR2_S 8
#define QI3_XXS (QK_K / (4*QR3_XXS))
#define QR3_XXS 8
#define QI3_XS (QK_K / (4*QR3_XS))
#define QR3_XS 8
#define QI1_S (QK_K / (4*QR1_S))
#define QR1_S 8
#define QI4_NL (QK4_NL / (4*QR4_NL))
#define QR4_NL 2
#define QI4_XS (QK_K / (4*QR4_XS))
#define QR4_XS 8
#endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
#define QK4_0 32
typedef struct {
ggml_half d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(ggml_half) + QK4_0 / 2, "wrong q4_0 block size/padding");
#define QK4_1 32
typedef struct {
union {
struct {
ggml_half d; // delta
ggml_half m; // min
} GGML_COMMON_AGGR;
ggml_half2 dm;
};
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK5_0 32
typedef struct {
ggml_half d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
#define QK5_1 32
typedef struct {
union {
struct {
ggml_half d; // delta
ggml_half m; // min
} GGML_COMMON_AGGR;
ggml_half2 dm;
};
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_half) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
#define QK8_0 32
typedef struct {
ggml_half d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(ggml_half) + QK8_0, "wrong q8_0 block size/padding");
#define QK8_1 32
typedef struct {
union {
struct {
ggml_half d; // delta
ggml_half s; // d * sum(qs[i])
} GGML_COMMON_AGGR;
ggml_half2 ds;
};
int8_t qs[QK8_1]; // quants
} block_q8_1;
static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
//
// Super-block quantization structures
//
// 2-bit quantization
// weight is represented as x = a * q + b
// 16 blocks of 16 elements each
// Effectively 2.625 bits per weight
typedef struct {
uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
uint8_t qs[QK_K/4]; // quants
union {
struct {
ggml_half d; // super-block scale for quantized scales
ggml_half dmin; // super-block scale for quantized mins
} GGML_COMMON_AGGR;
ggml_half2 dm;
};
} block_q2_K;
static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
// 3-bit quantization
// weight is represented as x = a * q
// 16 blocks of 16 elements each
// Effectively 3.4375 bits per weight
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
uint8_t scales[12]; // scales, quantized with 6 bits
ggml_half d; // super-block scale
} block_q3_K;
static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
// 4-bit quantization
// 8 blocks of 32 elements each
// weight is represented as x = a * q + b
// Effectively 4.5 bits per weight
typedef struct {
union {
struct {
ggml_half d; // super-block scale for quantized scales
ggml_half dmin; // super-block scale for quantized mins
} GGML_COMMON_AGGR;
ggml_half2 dm;
};
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
// 5-bit quantization
// 8 blocks of 32 elements each
// weight is represented as x = a * q + b
// Effectively 5.5 bits per weight
typedef struct {
union {
struct {
ggml_half d; // super-block scale for quantized scales
ggml_half dmin; // super-block scale for quantized mins
} GGML_COMMON_AGGR;
ggml_half2 dm;
};
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
// 6-bit quantization
// weight is represented as x = a * q
// 16 blocks of 16 elements each
// Effectively 6.5625 bits per weight
typedef struct {
uint8_t ql[QK_K/2]; // quants, lower 4 bits
uint8_t qh[QK_K/4]; // quants, upper 2 bits
int8_t scales[QK_K/16]; // scales, quantized with 8 bits
ggml_half d; // super-block scale
} block_q6_K;
static_assert(sizeof(block_q6_K) == sizeof(ggml_half) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
// This is only used for intermediate quantization and dot products
typedef struct {
float d; // delta
int8_t qs[QK_K]; // quants
int16_t bsums[QK_K/16]; // sum of quants in groups of 16
} block_q8_K;
static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
// (Almost) "true" 2-bit quantization.
// Due to the need to use blocks as per ggml design, it ends up using
// 2.0625 bpw because of the 16-bit scale for each block of 256.
typedef struct {
ggml_half d;
uint16_t qs[QK_K/8];
} block_iq2_xxs;
static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
// 2.3125 bpw quants
typedef struct {
ggml_half d;
uint16_t qs[QK_K/8];
uint8_t scales[QK_K/32];
} block_iq2_xs;
static_assert(sizeof(block_iq2_xs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
// 2.5625 bpw quants
typedef struct {
ggml_half d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t scales[QK_K/32];
} block_iq2_s;
static_assert(sizeof(block_iq2_s) == sizeof(ggml_half) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
// (Almost) "true" 3-bit quantization.
// Due to the need to use blocks as per ggml design, it ends up using
// 3.0625 bpw because of the 16-bit scale for each block of 256.
typedef struct {
ggml_half d;
uint8_t qs[3*QK_K/8];
} block_iq3_xxs;
static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
// 3.4375 bpw
#define IQ3S_N_SCALE QK_K/64
typedef struct {
ggml_half d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t signs[QK_K/8];
uint8_t scales[IQ3S_N_SCALE];
} block_iq3_s;
static_assert(sizeof(block_iq3_s) == sizeof(ggml_half) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
typedef struct {
ggml_half d;
uint8_t qs[QK_K/8];
uint16_t qh[QK_K/32];
} block_iq1_s;
static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
// 1.75 bpw
typedef struct {
uint8_t qs[QK_K/8]; // grid index, low 8 bits
uint8_t qh[QK_K/16]; // grid index, high 3 bits + grid shift bit (for two groups of 8)
uint8_t scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
} block_iq1_m;
static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
// Used by IQ1_M quants
typedef union {
ggml_half f16;
uint16_t u16;
} iq1m_scale_t;
// Non-linear quants
#define QK4_NL 32
typedef struct {
ggml_half d;
uint8_t qs[QK4_NL/2];
} block_iq4_nl;
static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");
typedef struct {
ggml_half d;
uint16_t scales_h;
uint8_t scales_l[QK_K/64];
uint8_t qs[QK_K/2];
} block_iq4_xs;
static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
#endif // GGML_COMMON_DECL
#endif // GGML_COMMON_DECL
////////////////////////////////////////////////////////////////////////////////
#ifndef GGML_COMMON_IMPL
#if defined(GGML_COMMON_IMPL_C)
#include <stdint.h>
#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
#define GGML_TABLE_END() };
#define GGML_COMMON_IMPL
#elif defined(GGML_COMMON_IMPL_METAL)
#include <metal_stdlib>
#define GGML_TABLE_BEGIN(type, name, size) static const constant type name[size] = {
#define GGML_TABLE_END() };
#define GGML_COMMON_IMPL
#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP)
#include <cstdint>
#define GGML_TABLE_BEGIN(type, name, size) static const __device__ type name[size] = {
#define GGML_TABLE_END() };
#define GGML_COMMON_IMPL
#elif defined(GGML_COMMON_IMPL_SYCL)
#include <cstdint>
#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
#define GGML_TABLE_END() };
#define GGML_COMMON_IMPL
#endif
#if defined(GGML_COMMON_IMPL)
GGML_TABLE_BEGIN(uint8_t, kmask_iq2xs, 8)
1, 2, 4, 8, 16, 32, 64, 128
GGML_TABLE_END()
GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128)
0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15,
144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159,
160, 33, 34, 163, 36, 165, 166, 39, 40, 169, 170, 43, 172, 45, 46, 175,
48, 177, 178, 51, 180, 53, 54, 183, 184, 57, 58, 187, 60, 189, 190, 63,
192, 65, 66, 195, 68, 197, 198, 71, 72, 201, 202, 75, 204, 77, 78, 207,
80, 209, 210, 83, 212, 85, 86, 215, 216, 89, 90, 219, 92, 221, 222, 95,
96, 225, 226, 99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
GGML_TABLE_END()
//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,
0xff000000ff000000, 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff,
0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, 0x00000000ffffffff,
0xff0000ff00000000, 0x000000ff000000ff, 0x000000ff0000ff00, 0xff0000ff0000ffff,
0x000000ff00ff0000, 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff,
0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, 0x000000ffff00ffff,
0xff0000ffffff0000, 0x000000ffffff00ff, 0x000000ffffffff00, 0xff0000ffffffffff,
0xff00ff0000000000, 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff,
0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, 0x0000ff0000ffffff,
0x0000ff00ff000000, 0xff00ff00ff0000ff, 0xff00ff00ff00ff00, 0x0000ff00ff00ffff,
0xff00ff00ffff0000, 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff,
0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, 0x0000ffff0000ffff,
0xff00ffff00ff0000, 0x0000ffff00ff00ff, 0x0000ffff00ffff00, 0xff00ffff00ffffff,
0xff00ffffff000000, 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff,
0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, 0x0000ffffffffffff,
0xffff000000000000, 0x00ff0000000000ff, 0x00ff00000000ff00, 0xffff00000000ffff,
0x00ff000000ff0000, 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff,
0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, 0x00ff0000ff00ffff,
0xffff0000ffff0000, 0x00ff0000ffff00ff, 0x00ff0000ffffff00, 0xffff0000ffffffff,
0x00ff00ff00000000, 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff,
0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, 0xffff00ff00ffffff,
0xffff00ffff000000, 0x00ff00ffff0000ff, 0x00ff00ffff00ff00, 0xffff00ffff00ffff,
0x00ff00ffffff0000, 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff,
0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, 0x00ffff000000ffff,
0xffffff0000ff0000, 0x00ffff0000ff00ff, 0x00ffff0000ffff00, 0xffffff0000ffffff,
0xffffff00ff000000, 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff,
0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, 0x00ffff00ffffffff,
0xffffffff00000000, 0x00ffffff000000ff, 0x00ffffff0000ff00, 0xffffffff0000ffff,
0x00ffffff00ff0000, 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff,
0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff,
0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff,
GGML_TABLE_END()
//#endif
GGML_TABLE_BEGIN(uint64_t, iq2xxs_grid, 256)
0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
GGML_TABLE_END()
GGML_TABLE_BEGIN(uint64_t, iq2xs_grid, 512)
0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808,
0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819,
0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819,
0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808,
0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b,
0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b,
0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908,
0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908,
0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919,
0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808,
0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919,
0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908,
0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b,
0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908,
0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08,
0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808,
0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808,
0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819,
0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908,
0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819,
0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808,
0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b,
0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819,
0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819,
0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808,
0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908,
0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19,
0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b,
0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b,
0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919,
0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808,
0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819,
0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819,
0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b,
0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908,
0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808,
0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819,
0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808,
0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919,
0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808,
0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808,
0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908,
0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908,
0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808,
0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b,
0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819,
0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919,
0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908,
0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808,
0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908,
0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919,
0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08,
0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19,
0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b,
0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b,
0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808,
0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08,
0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b,
0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908,
0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b,
0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908,
0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08,
0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808,
0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808,
0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08,
0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819,
0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919,
0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808,
0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808,
0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819,
0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819,
0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908,
0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908,
0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b,
0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908,
0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908,
0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908,
0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808,
0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819,
0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819,
0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819,
0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808,
0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b,
0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819,
0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819,
0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08,
0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808,
0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19,
0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919,
0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808,
0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19,
0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b,
0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808,
0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b,
0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b,
0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08,
0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b,
0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808,
0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819,
0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808,
0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808,
0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08,
0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b,
0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19,
0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08,
0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919,
0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08,
0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08,
0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908,
0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908,
0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b,
0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908,
0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808,
0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b,
0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808,
0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808,
0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19,
0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08,
0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808,
0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b,
0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808,
0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b,
0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
GGML_TABLE_END()
GGML_TABLE_BEGIN(uint64_t, iq2s_grid, 1024)
0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b,
0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919,
0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808,
0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908,
0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b,
0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908,
0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08,
0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19,
0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819,
0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919,
0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b,
0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908,
0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908,
0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919,
0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908,
0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b,
0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919,
0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b,
0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919,
0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908,
0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b,
0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b,
0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08,
0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b,
0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819,
0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808,
0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908,
0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b,
0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908,
0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08,
0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819,
0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808,
0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08,
0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819,
0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b,
0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908,
0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919,
0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b,
0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919,
0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808,
0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819,
0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919,
0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919,
0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808,
0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819,
0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b,
0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908,
0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b,
0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908,
0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919,
0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b,
0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919,
0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b,
0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819,
0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919,
0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908,
0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b,
0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908,
0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b,
0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908,
0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08,
0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908,
0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819,
0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819,
0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808,
0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08,
0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19,
0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819,
0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808,
0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819,
0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919,
0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808,
0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19,
0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08,
0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b,
0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908,
0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808,
0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819,
0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908,
0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819,
0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808,
0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808,
0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819,
0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908,
0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08,
0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819,
0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b,
0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b,
0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08,
0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19,
0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819,
0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919,
0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908,
0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808,
0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808,
0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908,
0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808,
0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08,
0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08,
0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908,
0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919,
0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808,
0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819,
0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908,
0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08,
0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819,
0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808,
0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808,
0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819,
0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808,
0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908,
0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b,
0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819,
0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808,
0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b,
0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808,
0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b,
0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19,
0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819,
0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08,
0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b,
0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908,
0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b,
0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b,
0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919,
0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808,
0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819,
0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908,
0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08,
0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08,
0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819,
0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919,
0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908,
0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b,
0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908,
0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b,
0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908,
0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08,
0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819,
0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808,
0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819,
0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919,
0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808,
0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808,
0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08,
0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819,
0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919,
0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808,
0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819,
0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919,
0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808,
0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b,
0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908,
0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808,
0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908,
0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b,
0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908,
0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b,
0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908,
0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b,
0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908,
0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08,
0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908,
0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b,
0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908,
0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08,
0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819,
0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919,
0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808,
0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19,
0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b,
0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919,
0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808,
0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819,
0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908,
0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919,
0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808,
0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808,
0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b,
0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919,
0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808,
0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b,
0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808,
0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919,
0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b,
0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08,
0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919,
0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808,
0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b,
0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908,
0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808,
0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808,
0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808,
0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908,
0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808,
0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808,
0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b,
0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908,
0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808,
0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808,
0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819,
0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919,
0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b,
0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808,
0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819,
0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b,
0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908,
0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08,
0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908,
0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919,
0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819,
0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908,
0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b,
0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808,
0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819,
0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908,
0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919,
0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808,
0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808,
0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808,
0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919,
0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908,
0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908,
0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08,
0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819,
0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b,
0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808,
0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819,
0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908,
0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819,
0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808,
0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808,
0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b,
0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908,
0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808,
0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908,
0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819,
0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819,
0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808,
0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b,
0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b,
0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819,
0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b,
0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b,
0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b,
0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819,
0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19,
0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819,
0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908,
0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808,
0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b,
GGML_TABLE_END()
GGML_TABLE_BEGIN(uint32_t, iq3xxs_grid, 256)
0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
GGML_TABLE_END()
GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305,
0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905,
0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09,
0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b,
0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b,
0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d,
0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03,
0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505,
0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03,
0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901,
0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d,
0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303,
0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501,
0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105,
0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505,
0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101,
0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707,
0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b,
0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01,
0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f,
0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305,
0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103,
0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509,
0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503,
0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b,
0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f,
0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f,
0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f,
0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109,
0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f,
0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509,
0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501,
0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303,
0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f,
0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907,
0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703,
0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03,
0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01,
0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01,
0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903,
0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505,
0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b,
0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107,
0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509,
0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303,
0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103,
0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05,
0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b,
0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f,
0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701,
0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909,
0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305,
0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d,
0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b,
0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d,
0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307,
0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09,
0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309,
0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709,
0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f,
0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303,
0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503,
0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b,
0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
GGML_TABLE_END()
#define NGRID_IQ1S 2048
#define IQ1S_DELTA 0.125f
#define IQ1M_DELTA 0.125f
#if defined(GGML_COMMON_IMPL_C)
GGML_TABLE_BEGIN(uint64_t, iq1s_grid, NGRID_IQ1S)
0xffffffffffffffff, 0xffffffffffffff01, 0xffffffffffff0000, 0xffffffffffff01ff,
0xffffffffffff0101, 0xffffffffff00ff00, 0xffffffffff000000, 0xffffffffff01ffff,
0xffffffffff01ff01, 0xffffffffff0101ff, 0xffffffffff010101, 0xffffffff00ff0000,
0xffffffff0000ff00, 0xffffffff000000ff, 0xffffffff00000001, 0xffffffff00010000,
0xffffffff01ffffff, 0xffffffff01ffff01, 0xffffffff01ff01ff, 0xffffffff01ff0101,
0xffffffff01000000, 0xffffffff0101ffff, 0xffffffff0101ff01, 0xffffffff010101ff,
0xffffffff01010101, 0xffffff00ffff00ff, 0xffffff00ffff0000, 0xffffff00ff00ff00,
0xffffff00ff0000ff, 0xffffff00ff000001, 0xffffff00ff000100, 0xffffff00ff000101,
0xffffff00ff010000, 0xffffff0000ffff00, 0xffffff0000ff0001, 0xffffff0000ff0100,
0xffffff000000ff01, 0xffffff0000000000, 0xffffff0000000101, 0xffffff000001ff00,
0xffffff00000100ff, 0xffffff0000010001, 0xffffff00000101ff, 0xffffff0001ff0000,
0xffffff000100ff00, 0xffffff00010000ff, 0xffffff0001000001, 0xffffff0001010000,
0xffffff01ffffffff, 0xffffff01ffffff01, 0xffffff01ffff01ff, 0xffffff01ffff0101,
0xffffff01ff000000, 0xffffff01ff01ffff, 0xffffff01ff01ff01, 0xffffff01ff0101ff,
0xffffff01ff010101, 0xffffff0100ff0000, 0xffffff010000ff00, 0xffffff0100000100,
0xffffff01000100ff, 0xffffff0100010100, 0xffffff0101ffffff, 0xffffff0101ffff01,
0xffffff0101ff01ff, 0xffffff0101ff0101, 0xffffff010100ff00, 0xffffff0101000000,
0xffffff0101000100, 0xffffff010101ffff, 0xffffff010101ff01, 0xffffff01010101ff,
0xffffff0101010101, 0xffff00ffff00ff00, 0xffff00ffff0000ff, 0xffff00ffff000001,
0xffff00ffff010000, 0xffff00ff00ffff00, 0xffff00ff00ff0100, 0xffff00ff00000000,
0xffff00ff00000101, 0xffff00ff000100ff, 0xffff00ff00010000, 0xffff00ff0100ff00,
0xffff00ff01000100, 0xffff00ff01010000, 0xffff0000ffffff00, 0xffff0000ffff00ff,
0xffff0000ffff0000, 0xffff0000ffff0001, 0xffff0000ff000000, 0xffff0000ff0001ff,
0xffff0000ff000101, 0xffff0000ff010100, 0xffff000000ffffff, 0xffff000000ff0000,
0xffff000000ff0101, 0xffff00000000ffff, 0xffff00000000ff00, 0xffff0000000000ff,
0xffff000000000000, 0xffff000000000001, 0xffff000000000100, 0xffff00000001ffff,
0xffff00000001ff01, 0xffff000000010000, 0xffff0000000101ff, 0xffff000000010101,
0xffff000001ffff00, 0xffff00000100ff00, 0xffff000001000000, 0xffff0000010001ff,
0xffff000001000101, 0xffff00000101ff00, 0xffff0000010100ff, 0xffff000001010000,
0xffff000001010001, 0xffff000001010100, 0xffff0001ff0000ff, 0xffff0001ff000100,
0xffff000100ffff00, 0xffff000100ff00ff, 0xffff00010000ffff, 0xffff00010000ff01,
0xffff000100000000, 0xffff0001000001ff, 0xffff00010001ffff, 0xffff00010001ff00,
0xffff000100010001, 0xffff000100010100, 0xffff000101ff0000, 0xffff00010100ff00,
0xffff0001010000ff, 0xffff000101000100, 0xffff01ffffffffff, 0xffff01ffffffff01,
0xffff01ffffff01ff, 0xffff01ffffff0101, 0xffff01ffff000000, 0xffff01ffff01ffff,
0xffff01ffff01ff01, 0xffff01ffff0101ff, 0xffff01ffff010101, 0xffff01ff00ff0000,
0xffff01ff0000ff00, 0xffff01ff00000001, 0xffff01ff00010000, 0xffff01ff01ffffff,
0xffff01ff01ffff01, 0xffff01ff01ff01ff, 0xffff01ff01ff0101, 0xffff01ff01000000,
0xffff01ff0101ffff, 0xffff01ff0101ff01, 0xffff01ff010101ff, 0xffff01ff01010101,
0xffff0100ffff0000, 0xffff0100ff00ff00, 0xffff0100ff0000ff, 0xffff0100ff000100,
0xffff0100ff0100ff, 0xffff0100ff010000, 0xffff010000ffff00, 0xffff01000000ffff,
0xffff01000000ff00, 0xffff010000000000, 0xffff01000001ff00, 0xffff0100000100ff,
0xffff010000010100, 0xffff01000100ff00, 0xffff0100010000ff, 0xffff010001000001,
0xffff010001000100, 0xffff010001010000, 0xffff0101ffffffff, 0xffff0101ffffff01,
0xffff0101ffff01ff, 0xffff0101ffff0101, 0xffff0101ff000000, 0xffff0101ff01ffff,
0xffff0101ff01ff01, 0xffff0101ff0101ff, 0xffff0101ff010101, 0xffff010100ff0000,
0xffff01010000ff00, 0xffff010100000100, 0xffff01010001ff00, 0xffff010100010000,
0xffff010101ffffff, 0xffff010101ffff01, 0xffff010101ff0000, 0xffff010101ff01ff,
0xffff010101ff0101, 0xffff010101000000, 0xffff01010101ffff, 0xffff01010101ff01,
0xffff0101010101ff, 0xffff010101010101, 0xff00ffffff00ffff, 0xff00ffffff00ff00,
0xff00ffffff0000ff, 0xff00ffffff000100, 0xff00ffffff0100ff, 0xff00ffffff010000,
0xff00ffff00ffff00, 0xff00ffff00ff00ff, 0xff00ffff0000ffff, 0xff00ffff00000000,
0xff00ffff000001ff, 0xff00ffff0001ff00, 0xff00ffff000100ff, 0xff00ffff00010000,
0xff00ffff00010100, 0xff00ffff0100ff00, 0xff00ffff010000ff, 0xff00ffff01000001,
0xff00ffff0101ff00, 0xff00ffff01010000, 0xff00ff00ffffff00, 0xff00ff00ffff00ff,
0xff00ff00ffff0001, 0xff00ff00ffff0100, 0xff00ff00ff00ffff, 0xff00ff00ff00ff01,
0xff00ff00ff000000, 0xff00ff00ff0001ff, 0xff00ff00ff01ff00, 0xff00ff00ff0100ff,
0xff00ff00ff010100, 0xff00ff0000ff0000, 0xff00ff0000ff0101, 0xff00ff000000ffff,
0xff00ff000000ff00, 0xff00ff000000ff01, 0xff00ff00000000ff, 0xff00ff0000000000,
0xff00ff0000000001, 0xff00ff0000000100, 0xff00ff000001ffff, 0xff00ff0000010000,
0xff00ff0001ff00ff, 0xff00ff000100ff01, 0xff00ff0001000000, 0xff00ff000101ff00,
0xff00ff00010100ff, 0xff00ff01ff00ff00, 0xff00ff01ff0000ff, 0xff00ff01ff000001,
0xff00ff01ff010000, 0xff00ff0100ffffff, 0xff00ff0100ff0001, 0xff00ff0100ff0100,
0xff00ff010000ff01, 0xff00ff0100000000, 0xff00ff01000001ff, 0xff00ff0100000101,
0xff00ff01000100ff, 0xff00ff0100010001, 0xff00ff0101ff0000, 0xff00ff010100ff00,
0xff00ff01010000ff, 0xff00ff0101000001, 0xff00ff0101010000, 0xff0000ffffffff00,
0xff0000ffffff0001, 0xff0000ffffff0100, 0xff0000ffff0000ff, 0xff0000ffff000000,
0xff0000ffff0001ff, 0xff0000ffff000100, 0xff0000ffff01ff00, 0xff0000ffff010001,
0xff0000ff00ffff00, 0xff0000ff00ff0000, 0xff0000ff00ff0001, 0xff0000ff00ff01ff,
0xff0000ff00ff0101, 0xff0000ff0000ff00, 0xff0000ff000000ff, 0xff0000ff00000000,
0xff0000ff00000001, 0xff0000ff00000100, 0xff0000ff0001ff01, 0xff0000ff00010000,
0xff0000ff000101ff, 0xff0000ff01ff00ff, 0xff0000ff01ff0100, 0xff0000ff0100ffff,
0xff0000ff010000ff, 0xff0000ff01000000, 0xff0000ff010001ff, 0xff0000ff01000100,
0xff0000ff01000101, 0xff0000ff0101ff00, 0xff0000ff010100ff, 0xff0000ff01010000,
0xff0000ff01010100, 0xff000000ffffff01, 0xff000000ffff0000, 0xff000000ffff0101,
0xff000000ff00ff00, 0xff000000ff0000ff, 0xff000000ff000000, 0xff000000ff000001,
0xff000000ff000100, 0xff000000ff01ffff, 0xff000000ff01ff01, 0xff000000ff010000,
0xff000000ff0101ff, 0xff000000ff010101, 0xff00000000ffff00, 0xff00000000ff00ff,
0xff00000000ff0000, 0xff00000000ff0001, 0xff0000000000ff00, 0xff0000000000ff01,
0xff000000000000ff, 0xff00000000000000, 0xff00000000000001, 0xff00000000000100,
0xff00000000000101, 0xff0000000001ff00, 0xff000000000100ff, 0xff00000000010000,
0xff00000000010001, 0xff00000000010100, 0xff00000001ffffff, 0xff00000001ffff01,
0xff00000001ff00ff, 0xff00000001ff0000, 0xff00000001ff01ff, 0xff00000001ff0101,
0xff0000000100ffff, 0xff0000000100ff00, 0xff000000010000ff, 0xff00000001000000,
0xff00000001000001, 0xff00000001000100, 0xff00000001000101, 0xff0000000101ffff,
0xff0000000101ff01, 0xff00000001010000, 0xff000001ffffff00, 0xff000001ffff00ff,
0xff000001ffff0000, 0xff000001ffff0001, 0xff000001ff000000, 0xff000001ff000001,
0xff000001ff0001ff, 0xff000001ff000101, 0xff000001ff01ff00, 0xff000001ff010001,
0xff00000100ffffff, 0xff00000100ffff01, 0xff00000100ff00ff, 0xff00000100ff0000,
0xff00000100ff01ff, 0xff00000100ff0101, 0xff0000010000ff00, 0xff00000100000000,
0xff00000100000001, 0xff000001000001ff, 0xff00000100000100, 0xff0000010001ff00,
0xff000001000100ff, 0xff00000100010000, 0xff000001000101ff, 0xff00000100010100,
0xff00000100010101, 0xff00000101ff0001, 0xff00000101ff0101, 0xff0000010100ff01,
0xff00000101000000, 0xff000001010100ff, 0xff00000101010100, 0xff0001ffff00ff00,
0xff0001ffff000001, 0xff0001ffff010000, 0xff0001ff00ffff00, 0xff0001ff00ff00ff,
0xff0001ff00ff0001, 0xff0001ff00ff0100, 0xff0001ff0000ffff, 0xff0001ff00000000,
0xff0001ff000001ff, 0xff0001ff00000101, 0xff0001ff0001ffff, 0xff0001ff0001ff00,
0xff0001ff000100ff, 0xff0001ff00010001, 0xff0001ff00010100, 0xff0001ff01ff0000,
0xff0001ff0100ff00, 0xff0001ff010000ff, 0xff0001ff01010000, 0xff000100ff00ffff,
0xff000100ff00ff01, 0xff000100ff000000, 0xff000100ff000101, 0xff000100ff01ff00,
0xff000100ff010000, 0xff00010000ffff01, 0xff00010000ff00ff, 0xff00010000ff0000,
0xff00010000ff01ff, 0xff0001000000ff00, 0xff000100000000ff, 0xff00010000000000,
0xff00010000000001, 0xff00010000000100, 0xff00010000000101, 0xff0001000001ffff,
0xff00010000010000, 0xff00010000010101, 0xff00010001ff0100, 0xff0001000100ff00,
0xff0001000100ff01, 0xff00010001000000, 0xff000100010001ff, 0xff0001000101ff00,
0xff00010001010001, 0xff00010001010100, 0xff000101ffff0100, 0xff000101ff000001,
0xff000101ff0100ff, 0xff000101ff010001, 0xff00010100ff00ff, 0xff00010100ff0001,
0xff00010100ff0100, 0xff0001010000ffff, 0xff0001010000ff01, 0xff00010100000000,
0xff000101000001ff, 0xff0001010001ff00, 0xff00010100010001, 0xff00010100010100,
0xff00010101ff0000, 0xff0001010100ff00, 0xff00010101000001, 0xff00010101000101,
0xff01ffffffffffff, 0xff01ffffffffff01, 0xff01ffffffff01ff, 0xff01ffffffff0101,
0xff01ffffff000000, 0xff01ffffff01ffff, 0xff01ffffff01ff01, 0xff01ffffff010000,
0xff01ffffff0101ff, 0xff01ffffff010101, 0xff01ffff00ff0000, 0xff01ffff0000ff00,
0xff01ffff00000100, 0xff01ffff0001ff00, 0xff01ffff00010000, 0xff01ffff01ffffff,
0xff01ffff01ffff01, 0xff01ffff01ff01ff, 0xff01ffff01ff0101, 0xff01ffff01000000,
0xff01ffff0101ffff, 0xff01ffff0101ff01, 0xff01ffff01010000, 0xff01ffff010101ff,
0xff01ffff01010101, 0xff01ff00ffff0000, 0xff01ff00ff00ff00, 0xff01ff00ff0000ff,
0xff01ff00ff000100, 0xff01ff00ff010000, 0xff01ff0000ffff01, 0xff01ff0000ff00ff,
0xff01ff0000ff0100, 0xff01ff0000000000, 0xff01ff00000001ff, 0xff01ff0000000101,
0xff01ff000001ff00, 0xff01ff00000100ff, 0xff01ff0000010000, 0xff01ff0000010001,
0xff01ff0001ff0000, 0xff01ff000100ffff, 0xff01ff0001000001, 0xff01ff0001000100,
0xff01ff0001010000, 0xff01ff01ffffff00, 0xff01ff01ffff01ff, 0xff01ff01ffff0101,
0xff01ff01ff00ff00, 0xff01ff01ff000000, 0xff01ff01ff01ffff, 0xff01ff01ff01ff01,
0xff01ff01ff0101ff, 0xff01ff01ff010101, 0xff01ff0100ff0000, 0xff01ff010000ff00,
0xff01ff0100000001, 0xff01ff0100000100, 0xff01ff0100010000, 0xff01ff0101ffff00,
0xff01ff0101ff01ff, 0xff01ff0101ff0101, 0xff01ff010100ff00, 0xff01ff0101000000,
0xff01ff010101ffff, 0xff01ff010101ff01, 0xff01ff01010101ff, 0xff01ff0101010101,
0xff0100ffffff0000, 0xff0100ffff0000ff, 0xff0100ffff000001, 0xff0100ffff000100,
0xff0100ffff010000, 0xff0100ff00ff00ff, 0xff0100ff00ff0000, 0xff0100ff00ff0001,
0xff0100ff00ff0100, 0xff0100ff0000ff01, 0xff0100ff00000000, 0xff0100ff000001ff,
0xff0100ff00000101, 0xff0100ff00010001, 0xff0100ff01ff0000, 0xff0100ff0100ff00,
0xff0100ff010000ff, 0xff0100ff01000100, 0xff0100ff0101ff00, 0xff0100ff01010000,
0xff010000ffff0100, 0xff010000ff000000, 0xff010000ff01ff00, 0xff010000ff010100,
0xff01000000ffffff, 0xff01000000ff0000, 0xff01000000ff01ff, 0xff0100000000ff00,
0xff010000000000ff, 0xff01000000000000, 0xff01000000000100, 0xff0100000001ff01,
0xff01000000010000, 0xff010000000101ff, 0xff01000001ff0100, 0xff0100000100ffff,
0xff010000010000ff, 0xff01000001000000, 0xff010000010001ff, 0xff01000001000101,
0xff0100000101ff00, 0xff010000010100ff, 0xff01000001010001, 0xff01000001010100,
0xff010001ffff0000, 0xff010001ff00ffff, 0xff010001ff00ff01, 0xff010001ff000100,
0xff010001ff010000, 0xff01000100ffff00, 0xff01000100ff0100, 0xff01000100000000,
0xff0100010001ffff, 0xff0100010001ff00, 0xff01000100010100, 0xff01000101ff00ff,
0xff01000101ff0001, 0xff0100010100ffff, 0xff01000101000101, 0xff0101ffffffffff,
0xff0101ffffffff01, 0xff0101ffffff01ff, 0xff0101ffffff0101, 0xff0101ffff000000,
0xff0101ffff01ffff, 0xff0101ffff01ff01, 0xff0101ffff0101ff, 0xff0101ffff010101,
0xff0101ff00ff0000, 0xff0101ff0000ff00, 0xff0101ff000000ff, 0xff0101ff00010000,
0xff0101ff01ffffff, 0xff0101ff01ffff01, 0xff0101ff01ff01ff, 0xff0101ff01ff0101,
0xff0101ff0101ffff, 0xff0101ff0101ff01, 0xff0101ff010101ff, 0xff0101ff01010101,
0xff010100ffff0100, 0xff010100ff00ff00, 0xff010100ff0000ff, 0xff010100ff000100,
0xff010100ff010000, 0xff01010000ff0001, 0xff01010000ff0100, 0xff0101000000ff01,
0xff01010000000000, 0xff0101000001ff00, 0xff010100000100ff, 0xff01010000010001,
0xff01010000010100, 0xff01010001ff0000, 0xff0101000100ffff, 0xff01010001000001,
0xff01010001000100, 0xff010100010100ff, 0xff01010001010000, 0xff010101ffffffff,
0xff010101ffffff01, 0xff010101ffff01ff, 0xff010101ffff0101, 0xff010101ff01ffff,
0xff010101ff01ff01, 0xff010101ff0101ff, 0xff010101ff010101, 0xff01010100ff0000,
0xff0101010000ff00, 0xff01010100000001, 0xff01010100000100, 0xff01010100010000,
0xff01010101ffffff, 0xff01010101ffff01, 0xff01010101ff01ff, 0xff01010101ff0101,
0xff01010101000000, 0xff0101010101ffff, 0xff0101010101ff01, 0xff010101010101ff,
0xff01010101010101, 0x00ffffffffff0000, 0x00ffffffff00ff00, 0x00ffffffff000001,
0x00ffffffff010000, 0x00ffffff00ff0100, 0x00ffffff0000ff01, 0x00ffffff00000000,
0x00ffffff000001ff, 0x00ffffff00000101, 0x00ffffff0001ff00, 0x00ffffff000100ff,
0x00ffffff00010001, 0x00ffffff010000ff, 0x00ffffff01000100, 0x00ffffff0101ff00,
0x00ffffff01010001, 0x00ffff00ffffffff, 0x00ffff00ffffff00, 0x00ffff00ffff00ff,
0x00ffff00ffff0001, 0x00ffff00ffff0100, 0x00ffff00ff00ff01, 0x00ffff00ff000000,
0x00ffff00ff000001, 0x00ffff00ff0001ff, 0x00ffff00ff000101, 0x00ffff00ff01ff00,
0x00ffff00ff010001, 0x00ffff00ff010100, 0x00ffff0000ff0000, 0x00ffff0000ff01ff,
0x00ffff0000ff0101, 0x00ffff000000ff00, 0x00ffff00000000ff, 0x00ffff0000000000,
0x00ffff0000000001, 0x00ffff0000000100, 0x00ffff0000000101, 0x00ffff0000010000,
0x00ffff00000101ff, 0x00ffff0000010101, 0x00ffff0001ffff00, 0x00ffff0001ff00ff,
0x00ffff0001ff0001, 0x00ffff000100ffff, 0x00ffff000100ff01, 0x00ffff0001000000,
0x00ffff000101ffff, 0x00ffff000101ff00, 0x00ffff000101ff01, 0x00ffff01ffff0000,
0x00ffff01ff00ff00, 0x00ffff01ff0000ff, 0x00ffff01ff000001, 0x00ffff01ff010000,
0x00ffff0100ffff00, 0x00ffff010000ff01, 0x00ffff0100000000, 0x00ffff0100000101,
0x00ffff01000100ff, 0x00ffff0100010100, 0x00ffff0101ff0100, 0x00ffff01010000ff,
0x00ffff0101010000, 0x00ff00ffffffff00, 0x00ff00ffff000000, 0x00ff00ffff000100,
0x00ff00ffff010100, 0x00ff00ff00ff0000, 0x00ff00ff00ff01ff, 0x00ff00ff00ff0101,
0x00ff00ff0000ff00, 0x00ff00ff000000ff, 0x00ff00ff00000000, 0x00ff00ff00000001,
0x00ff00ff0001ff00, 0x00ff00ff0001ff01, 0x00ff00ff00010000, 0x00ff00ff000101ff,
0x00ff00ff00010101, 0x00ff00ff01ffff00, 0x00ff00ff01ff0001, 0x00ff00ff01ff0100,
0x00ff00ff0100ffff, 0x00ff00ff0100ff01, 0x00ff00ff01000000, 0x00ff00ff0101ffff,
0x00ff00ff0101ff00, 0x00ff00ff01010100, 0x00ff0000ffffff00, 0x00ff0000ffffff01,
0x00ff0000ffff0000, 0x00ff0000ffff0101, 0x00ff0000ff00ff00, 0x00ff0000ff0000ff,
0x00ff0000ff000000, 0x00ff0000ff000001, 0x00ff0000ff000100, 0x00ff0000ff01ffff,
0x00ff0000ff010000, 0x00ff0000ff010101, 0x00ff000000ffff00, 0x00ff000000ff00ff,
0x00ff000000ff0000, 0x00ff000000ff0001, 0x00ff000000ff0100, 0x00ff00000000ffff,
0x00ff00000000ff00, 0x00ff0000000000ff, 0x00ff000000000000, 0x00ff000000000001,
0x00ff0000000001ff, 0x00ff000000000100, 0x00ff00000001ff00, 0x00ff0000000100ff,
0x00ff000000010000, 0x00ff000000010001, 0x00ff000000010100, 0x00ff000001ffff01,
0x00ff000001ff00ff, 0x00ff000001ff0000, 0x00ff000001ff01ff, 0x00ff00000100ff00,
0x00ff0000010000ff, 0x00ff000001000000, 0x00ff000001000001, 0x00ff000001000100,
0x00ff000001000101, 0x00ff000001010000, 0x00ff0000010101ff, 0x00ff000001010101,
0x00ff0001ffffff00, 0x00ff0001ffff0000, 0x00ff0001ffff0100, 0x00ff0001ff0000ff,
0x00ff0001ff000000, 0x00ff0001ff0001ff, 0x00ff0001ff000101, 0x00ff0001ff01ff00,
0x00ff0001ff0100ff, 0x00ff0001ff010100, 0x00ff000100ffffff, 0x00ff000100ffff01,
0x00ff000100ff0000, 0x00ff000100ff01ff, 0x00ff00010000ffff, 0x00ff00010000ff00,
0x00ff00010000ff01, 0x00ff000100000000, 0x00ff000100000001, 0x00ff000100000100,
0x00ff00010001ff01, 0x00ff000100010000, 0x00ff0001000101ff, 0x00ff000101ffff00,
0x00ff000101ff0000, 0x00ff000101ff0101, 0x00ff0001010000ff, 0x00ff000101000000,
0x00ff00010101ff00, 0x00ff0001010100ff, 0x00ff000101010001, 0x00ff01ffffff0000,
0x00ff01ffff00ff00, 0x00ff01ffff000000, 0x00ff01ffff000101, 0x00ff01ffff010000,
0x00ff01ff00ffff01, 0x00ff01ff00ff0100, 0x00ff01ff0000ffff, 0x00ff01ff00000000,
0x00ff01ff000001ff, 0x00ff01ff0001ff00, 0x00ff01ff000100ff, 0x00ff01ff00010001,
0x00ff01ff00010100, 0x00ff01ff01ff0000, 0x00ff01ff0100ff00, 0x00ff01ff010000ff,
0x00ff01ff01000001, 0x00ff01ff01000100, 0x00ff01ff01010000, 0x00ff0100ffffff00,
0x00ff0100ffff0000, 0x00ff0100ffff0001, 0x00ff0100ffff0101, 0x00ff0100ff00ffff,
0x00ff0100ff0000ff, 0x00ff0100ff000000, 0x00ff0100ff0001ff, 0x00ff0100ff01ff00,
0x00ff0100ff0100ff, 0x00ff0100ff010001, 0x00ff010000ffffff, 0x00ff010000ff0000,
0x00ff010000ff0101, 0x00ff01000000ff00, 0x00ff01000000ff01, 0x00ff0100000000ff,
0x00ff010000000000, 0x00ff010000000001, 0x00ff010000000100, 0x00ff01000001ffff,
0x00ff01000001ff01, 0x00ff010000010000, 0x00ff010000010001, 0x00ff010000010101,
0x00ff010001ff0001, 0x00ff010001ff0100, 0x00ff01000100ff01, 0x00ff010001000000,
0x00ff010001000001, 0x00ff0100010001ff, 0x00ff01000101ff00, 0x00ff0100010100ff,
0x00ff010001010001, 0x00ff010001010100, 0x00ff0101ff000001, 0x00ff010100ff00ff,
0x00ff010100ff0001, 0x00ff010100ff0100, 0x00ff010100000000, 0x00ff0101000001ff,
0x00ff010100000101, 0x00ff0101000100ff, 0x00ff010100010100, 0x00ff0101010000ff,
0x00ff010101010000, 0x0000ffffffffff00, 0x0000ffffffff00ff, 0x0000ffffffff0000,
0x0000ffffffff0001, 0x0000ffffffff0100, 0x0000ffffff00ff01, 0x0000ffffff000000,
0x0000ffffff000101, 0x0000ffffff01ff00, 0x0000ffffff0100ff, 0x0000ffffff010100,
0x0000ffff00ffffff, 0x0000ffff00ff0000, 0x0000ffff00ff01ff, 0x0000ffff0000ff00,
0x0000ffff000000ff, 0x0000ffff00000000, 0x0000ffff00000001, 0x0000ffff00000100,
0x0000ffff00010000, 0x0000ffff000101ff, 0x0000ffff01ff0001, 0x0000ffff01ff0100,
0x0000ffff01000000, 0x0000ffff010001ff, 0x0000ffff0101ffff, 0x0000ffff0101ff00,
0x0000ffff01010001, 0x0000ffff01010100, 0x0000ff00ffff0000, 0x0000ff00ffff01ff,
0x0000ff00ffff0100, 0x0000ff00ffff0101, 0x0000ff00ff00ff00, 0x0000ff00ff0000ff,
0x0000ff00ff000000, 0x0000ff00ff000001, 0x0000ff00ff0001ff, 0x0000ff00ff000100,
0x0000ff00ff01ffff, 0x0000ff00ff010000, 0x0000ff00ff010001, 0x0000ff00ff0101ff,
0x0000ff00ff010101, 0x0000ff0000ffff00, 0x0000ff0000ff00ff, 0x0000ff0000ff0000,
0x0000ff0000ff0001, 0x0000ff0000ff0100, 0x0000ff000000ffff, 0x0000ff000000ff00,
0x0000ff000000ff01, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
0x0000ff00000001ff, 0x0000ff0000000100, 0x0000ff0000000101, 0x0000ff000001ff00,
0x0000ff00000100ff, 0x0000ff0000010000, 0x0000ff0000010001, 0x0000ff0000010100,
0x0000ff0001ffff01, 0x0000ff0001ff0000, 0x0000ff000100ff00, 0x0000ff00010000ff,
0x0000ff0001000000, 0x0000ff0001000001, 0x0000ff0001000100, 0x0000ff000101ffff,
0x0000ff0001010000, 0x0000ff0001010101, 0x0000ff01ffffff00, 0x0000ff01ffff0001,
0x0000ff01ff00ff01, 0x0000ff01ff000000, 0x0000ff01ff000101, 0x0000ff01ff01ff00,
0x0000ff01ff0100ff, 0x0000ff0100ffff01, 0x0000ff0100ff0000, 0x0000ff0100ff0101,
0x0000ff010000ff00, 0x0000ff01000000ff, 0x0000ff0100000000, 0x0000ff0100000001,
0x0000ff0100000100, 0x0000ff010001ff01, 0x0000ff0100010000, 0x0000ff0101ff0000,
0x0000ff010100ffff, 0x0000ff010100ff01, 0x0000ff0101000000, 0x0000ff0101000100,
0x0000ff0101000101, 0x0000ff01010100ff, 0x000000ffffff00ff, 0x000000ffffff0000,
0x000000ffff00ff00, 0x000000ffff0000ff, 0x000000ffff000000, 0x000000ffff000001,
0x000000ffff0001ff, 0x000000ffff000100, 0x000000ffff01ff00, 0x000000ffff010000,
0x000000ffff0101ff, 0x000000ffff010101, 0x000000ff00ffff00, 0x000000ff00ff00ff,
0x000000ff00ff0000, 0x000000ff00ff0001, 0x000000ff00ff0100, 0x000000ff00ff0101,
0x000000ff0000ffff, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
0x000000ff00000001, 0x000000ff000001ff, 0x000000ff00000100, 0x000000ff00000101,
0x000000ff0001ff00, 0x000000ff0001ff01, 0x000000ff000100ff, 0x000000ff00010000,
0x000000ff00010001, 0x000000ff00010100, 0x000000ff01ffffff, 0x000000ff01ff01ff,
0x000000ff01ff0101, 0x000000ff0100ff00, 0x000000ff010000ff, 0x000000ff01000000,
0x000000ff01000001, 0x000000ff01000100, 0x000000ff0101ff00, 0x000000ff010100ff,
0x000000ff01010000, 0x000000ff01010101, 0x00000000ffffff00, 0x00000000ffffff01,
0x00000000ffff00ff, 0x00000000ffff0000, 0x00000000ffff0001, 0x00000000ffff0100,
0x00000000ff00ffff, 0x00000000ff00ff00, 0x00000000ff00ff01, 0x00000000ff0000ff,
0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff000101,
0x00000000ff01ff00, 0x00000000ff0100ff, 0x00000000ff010000, 0x00000000ff010001,
0x00000000ff010100, 0x0000000000ffffff, 0x0000000000ffff00, 0x0000000000ffff01,
0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001, 0x0000000000ff01ff,
0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
0x00000000000000ff, 0x0000000000000000, 0x0000000000000001, 0x00000000000001ff,
0x0000000000000100, 0x0000000000000101, 0x000000000001ffff, 0x000000000001ff00,
0x00000000000100ff, 0x0000000000010000, 0x0000000000010001, 0x00000000000101ff,
0x0000000000010100, 0x0000000000010101, 0x0000000001ffff00, 0x0000000001ff00ff,
0x0000000001ff0000, 0x0000000001ff0100, 0x0000000001ff0101, 0x000000000100ffff,
0x000000000100ff00, 0x00000000010000ff, 0x0000000001000000, 0x0000000001000001,
0x00000000010001ff, 0x0000000001000100, 0x000000000101ff00, 0x00000000010100ff,
0x0000000001010000, 0x0000000001010001, 0x0000000001010100, 0x00000001ffffffff,
0x00000001ffffff00, 0x00000001ffffff01, 0x00000001ffff00ff, 0x00000001ffff0001,
0x00000001ffff01ff, 0x00000001ffff0100, 0x00000001ff00ff00, 0x00000001ff0000ff,
0x00000001ff000000, 0x00000001ff0001ff, 0x00000001ff000100, 0x00000001ff01ffff,
0x00000001ff01ff00, 0x00000001ff01ff01, 0x00000001ff0100ff, 0x00000001ff010000,
0x00000001ff010001, 0x00000001ff0101ff, 0x00000001ff010100, 0x0000000100ffff00,
0x0000000100ff0000, 0x0000000100ff0001, 0x0000000100ff01ff, 0x0000000100ff0100,
0x0000000100ff0101, 0x000000010000ffff, 0x000000010000ff00, 0x000000010000ff01,
0x00000001000000ff, 0x0000000100000000, 0x0000000100000001, 0x00000001000001ff,
0x0000000100000100, 0x0000000100000101, 0x000000010001ff00, 0x00000001000100ff,
0x0000000100010000, 0x0000000100010100, 0x0000000101ffff01, 0x0000000101ff0000,
0x0000000101ff0001, 0x0000000101ff01ff, 0x0000000101ff0100, 0x0000000101ff0101,
0x000000010100ff00, 0x0000000101000000, 0x0000000101000101, 0x000000010101ff01,
0x0000000101010000, 0x0000000101010001, 0x00000001010101ff, 0x0000000101010100,
0x000001ffffff00ff, 0x000001ffffff0000, 0x000001ffffff0001, 0x000001ffffff0100,
0x000001ffff00ffff, 0x000001ffff000000, 0x000001ffff0001ff, 0x000001ffff01ff00,
0x000001ffff010101, 0x000001ff00ff0000, 0x000001ff00ff01ff, 0x000001ff00ff0101,
0x000001ff0000ff00, 0x000001ff000000ff, 0x000001ff00000000, 0x000001ff00000001,
0x000001ff000001ff, 0x000001ff00000100, 0x000001ff0001ffff, 0x000001ff0001ff01,
0x000001ff000100ff, 0x000001ff00010000, 0x000001ff01ffff01, 0x000001ff01ff0100,
0x000001ff0100ffff, 0x000001ff0100ff01, 0x000001ff01000000, 0x000001ff010001ff,
0x000001ff0101ff00, 0x000001ff01010100, 0x00000100ffffff00, 0x00000100ffffff01,
0x00000100ffff0000, 0x00000100ffff0101, 0x00000100ff00ff00, 0x00000100ff0000ff,
0x00000100ff000000, 0x00000100ff000001, 0x00000100ff000100, 0x00000100ff010000,
0x0000010000ffff00, 0x0000010000ff00ff, 0x0000010000ff0000, 0x0000010000ff0001,
0x0000010000ff0100, 0x000001000000ffff, 0x000001000000ff00, 0x000001000000ff01,
0x00000100000000ff, 0x0000010000000000, 0x0000010000000001, 0x00000100000001ff,
0x0000010000000100, 0x0000010000000101, 0x000001000001ff00, 0x00000100000100ff,
0x0000010000010000, 0x0000010000010001, 0x0000010000010100, 0x0000010001ffff00,
0x0000010001ff0000, 0x0000010001ff0100, 0x000001000100ff00, 0x00000100010000ff,
0x0000010001000000, 0x0000010001000001, 0x00000100010001ff, 0x0000010001000100,
0x0000010001010000, 0x00000101ffff00ff, 0x00000101ffff01ff, 0x00000101ff000000,
0x00000101ff000101, 0x00000101ff01ffff, 0x00000101ff010000, 0x00000101ff010001,
0x00000101ff010100, 0x0000010100ff0000, 0x0000010100ff01ff, 0x0000010100ff0100,
0x000001010000ff00, 0x0000010100000000, 0x0000010100000001, 0x00000101000001ff,
0x0000010100000100, 0x000001010001ff01, 0x0000010100010000, 0x00000101000101ff,
0x0000010100010101, 0x0000010101ffff00, 0x0000010101ff0101, 0x000001010100ff01,
0x0000010101000000, 0x0000010101000001, 0x00000101010001ff, 0x0000010101000101,
0x000001010101ff00, 0x0001ffffffff0000, 0x0001ffffff0000ff, 0x0001ffffff000001,
0x0001ffffff000100, 0x0001ffffff010000, 0x0001ffff00ff00ff, 0x0001ffff0000ffff,
0x0001ffff00000000, 0x0001ffff00000001, 0x0001ffff000001ff, 0x0001ffff00000101,
0x0001ffff0001ff00, 0x0001ffff000100ff, 0x0001ffff00010001, 0x0001ffff00010100,
0x0001ffff01ffff00, 0x0001ffff01000001, 0x0001ffff01010000, 0x0001ff00ffffff00,
0x0001ff00ffff00ff, 0x0001ff00ffff0001, 0x0001ff00ffff0100, 0x0001ff00ff00ff01,
0x0001ff00ff000000, 0x0001ff00ff01ff00, 0x0001ff00ff01ff01, 0x0001ff00ff010001,
0x0001ff00ff010100, 0x0001ff0000ff0000, 0x0001ff0000ff0100, 0x0001ff000000ff00,
0x0001ff0000000000, 0x0001ff0000000001, 0x0001ff0000000100, 0x0001ff0000010000,
0x0001ff0000010001, 0x0001ff0000010101, 0x0001ff0001ff00ff, 0x0001ff0001ff0101,
0x0001ff000100ff01, 0x0001ff0001000000, 0x0001ff000101ff00, 0x0001ff0001010001,
0x0001ff0001010100, 0x0001ff01ff00ff00, 0x0001ff01ff000001, 0x0001ff01ff000100,
0x0001ff0100ffffff, 0x0001ff0100ffff00, 0x0001ff0100ff0001, 0x0001ff0100000000,
0x0001ff0100000001, 0x0001ff01000001ff, 0x0001ff010001ffff, 0x0001ff0101ff0000,
0x0001ff010100ff00, 0x0001ff0101000001, 0x0001ff0101010000, 0x000100ffff00ff00,
0x000100ffff00ff01, 0x000100ffff000000, 0x000100ffff000001, 0x000100ffff000101,
0x000100ffff01ff00, 0x000100ffff010001, 0x000100ffff010100, 0x000100ff00ffffff,
0x000100ff00ffff01, 0x000100ff00ff0000, 0x000100ff00ff01ff, 0x000100ff00ff0101,
0x000100ff0000ff00, 0x000100ff000000ff, 0x000100ff00000000, 0x000100ff00000001,
0x000100ff00000100, 0x000100ff00000101, 0x000100ff0001ffff, 0x000100ff0001ff01,
0x000100ff00010000, 0x000100ff01ff00ff, 0x000100ff01ff0000, 0x000100ff01ff0100,
0x000100ff0100ffff, 0x000100ff0100ff01, 0x000100ff010000ff, 0x000100ff01000000,
0x000100ff01000001, 0x000100ff010001ff, 0x000100ff01000101, 0x000100ff0101ff00,
0x000100ff010100ff, 0x000100ff01010100, 0x00010000ffff0000, 0x00010000ffff01ff,
0x00010000ffff0101, 0x00010000ff00ff00, 0x00010000ff000000, 0x00010000ff000001,
0x00010000ff000100, 0x0001000000ff00ff, 0x0001000000ff0000, 0x0001000000ff0001,
0x0001000000ff0100, 0x000100000000ffff, 0x000100000000ff00, 0x00010000000000ff,
0x0001000000000000, 0x0001000000000001, 0x0001000000000100, 0x000100000001ff00,
0x00010000000100ff, 0x0001000000010000, 0x0001000000010001, 0x0001000000010100,
0x0001000001ff0001, 0x0001000001ff0100, 0x0001000001ff0101, 0x000100000100ff00,
0x0001000001000000, 0x0001000001000001, 0x0001000001000100, 0x0001000001000101,
0x000100000101ff01, 0x0001000001010000, 0x0001000001010001, 0x00010000010101ff,
0x00010001ffffff01, 0x00010001ffff0100, 0x00010001ff000000, 0x00010001ff01ffff,
0x00010001ff010001, 0x00010001ff0101ff, 0x00010001ff010100, 0x0001000100ffffff,
0x0001000100ff0000, 0x0001000100ff01ff, 0x0001000100ff0101, 0x000100010000ff00,
0x00010001000000ff, 0x0001000100000000, 0x0001000100000001, 0x00010001000001ff,
0x0001000100000101, 0x000100010001ffff, 0x0001000100010000, 0x00010001000101ff,
0x0001000101ffffff, 0x0001000101ffff01, 0x0001000101ff0000, 0x0001000101ff0101,
0x00010001010000ff, 0x0001000101000001, 0x00010001010001ff, 0x0001000101000100,
0x000100010101ffff, 0x00010001010100ff, 0x0001000101010001, 0x0001000101010101,
0x000101ffff000001, 0x000101ffff000100, 0x000101ffff010000, 0x000101ff00ffff00,
0x000101ff0000ff01, 0x000101ff00000000, 0x000101ff00000101, 0x000101ff0001ff00,
0x000101ff00010100, 0x000101ff01ff0000, 0x000101ff0100ff00, 0x000101ff010001ff,
0x000101ff01010001, 0x00010100ffffff00, 0x00010100ffff00ff, 0x00010100ff00ffff,
0x00010100ff000000, 0x00010100ff01ff00, 0x00010100ff0100ff, 0x00010100ff010001,
0x00010100ff010100, 0x0001010000ffffff, 0x0001010000ffff00, 0x0001010000ff0000,
0x0001010000ff0001, 0x0001010000ff01ff, 0x000101000000ff00, 0x00010100000000ff,
0x0001010000000000, 0x0001010000000001, 0x0001010000000100, 0x000101000001ffff,
0x0001010000010000, 0x0001010000010101, 0x0001010001ffff01, 0x0001010001ff00ff,
0x0001010001ff0101, 0x0001010001000000, 0x000101000101ff00, 0x00010100010100ff,
0x0001010001010000, 0x0001010001010100, 0x00010101ff00ff00, 0x00010101ff000001,
0x00010101ff0001ff, 0x0001010100ffff00, 0x0001010100ff00ff, 0x0001010100ff0100,
0x000101010000ffff, 0x0001010100000000, 0x00010101000001ff, 0x0001010100000101,
0x00010101000100ff, 0x0001010100010000, 0x0001010100010100, 0x0001010101ff0001,
0x00010101010000ff, 0x00010101010001ff, 0x0001010101000101, 0x0001010101010001,
0x01ffffffffffffff, 0x01ffffffffffff01, 0x01ffffffffff01ff, 0x01ffffffffff0101,
0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff, 0x01ffffffff010101,
0x01ffffff00ff0000, 0x01ffffff0000ffff, 0x01ffffff0000ff00, 0x01ffffff000000ff,
0x01ffffff00000001, 0x01ffffff00000100, 0x01ffffff00010000, 0x01ffffff01ffffff,
0x01ffffff01ffff01, 0x01ffffff01ff01ff, 0x01ffffff01ff0101, 0x01ffffff01000000,
0x01ffffff0101ffff, 0x01ffffff0101ff01, 0x01ffffff010101ff, 0x01ffffff01010101,
0x01ffff00ffff0000, 0x01ffff00ff00ff00, 0x01ffff00ff0000ff, 0x01ffff00ff000001,
0x01ffff00ff000100, 0x01ffff00ff010000, 0x01ffff0000ffff00, 0x01ffff0000ff00ff,
0x01ffff0000ff0100, 0x01ffff000000ffff, 0x01ffff000000ff01, 0x01ffff0000000000,
0x01ffff0000000001, 0x01ffff00000001ff, 0x01ffff0000000100, 0x01ffff00000100ff,
0x01ffff0000010001, 0x01ffff0000010100, 0x01ffff0001ff0000, 0x01ffff0001ff0100,
0x01ffff00010000ff, 0x01ffff0001000001, 0x01ffff0001000100, 0x01ffff0001010000,
0x01ffff01ffffffff, 0x01ffff01ffffff01, 0x01ffff01ffff01ff, 0x01ffff01ffff0101,
0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff01ff01, 0x01ffff01ff0101ff,
0x01ffff01ff010101, 0x01ffff010000ff00, 0x01ffff01000000ff, 0x01ffff0100000100,
0x01ffff0100010000, 0x01ffff0101ffffff, 0x01ffff0101ffff01, 0x01ffff0101ff01ff,
0x01ffff0101ff0101, 0x01ffff0101000000, 0x01ffff010101ffff, 0x01ffff010101ff01,
0x01ffff01010101ff, 0x01ffff0101010101, 0x01ff00ffff0000ff, 0x01ff00ffff000100,
0x01ff00ff00ffff00, 0x01ff00ff00ff00ff, 0x01ff00ff0000ff00, 0x01ff00ff00000000,
0x01ff00ff00000101, 0x01ff00ff0001ff00, 0x01ff00ff000100ff, 0x01ff00ff00010100,
0x01ff00ff010000ff, 0x01ff00ff01000100, 0x01ff0000ffffff00, 0x01ff0000ffff0100,
0x01ff0000ff00ff01, 0x01ff0000ff000000, 0x01ff0000ff000101, 0x01ff0000ff010001,
0x01ff0000ff010100, 0x01ff000000ffffff, 0x01ff000000ffff00, 0x01ff000000ff0000,
0x01ff000000ff01ff, 0x01ff00000000ff00, 0x01ff0000000000ff, 0x01ff000000000000,
0x01ff000000000001, 0x01ff000000000100, 0x01ff000000000101, 0x01ff000000010000,
0x01ff000000010001, 0x01ff0000000101ff, 0x01ff000000010101, 0x01ff000001ffff00,
0x01ff000001ff00ff, 0x01ff000001ff0001, 0x01ff000001ff0100, 0x01ff00000100ffff,
0x01ff00000100ff01, 0x01ff000001000000, 0x01ff0000010001ff, 0x01ff000001010001,
0x01ff0001ff00ff00, 0x01ff0001ff000001, 0x01ff0001ff000100, 0x01ff0001ff010000,
0x01ff000100ffff00, 0x01ff000100ff00ff, 0x01ff000100ff0100, 0x01ff000100ff0101,
0x01ff00010000ffff, 0x01ff000100000000, 0x01ff000100000100, 0x01ff000100000101,
0x01ff00010001ff00, 0x01ff000100010001, 0x01ff000100010101, 0x01ff000101ff0000,
0x01ff00010100ff00, 0x01ff000101000101, 0x01ff0001010100ff, 0x01ff01ffffffffff,
0x01ff01ffffffff01, 0x01ff01ffffff01ff, 0x01ff01ffffff0101, 0x01ff01ffff000000,
0x01ff01ffff01ffff, 0x01ff01ffff01ff01, 0x01ff01ffff0101ff, 0x01ff01ffff010101,
0x01ff01ff00ffff00, 0x01ff01ff00ff0000, 0x01ff01ff0000ff00, 0x01ff01ff000000ff,
0x01ff01ff00000100, 0x01ff01ff00010000, 0x01ff01ff00010100, 0x01ff01ff01ffffff,
0x01ff01ff01ffff01, 0x01ff01ff01ff01ff, 0x01ff01ff01ff0101, 0x01ff01ff01000000,
0x01ff01ff0101ffff, 0x01ff01ff0101ff01, 0x01ff01ff010101ff, 0x01ff01ff01010101,
0x01ff0100ffff0000, 0x01ff0100ffff0001, 0x01ff0100ff00ff00, 0x01ff0100ff0000ff,
0x01ff0100ff000001, 0x01ff0100ff010000, 0x01ff010000ffff00, 0x01ff010000ff00ff,
0x01ff010000ff0001, 0x01ff010000ff0100, 0x01ff01000000ffff, 0x01ff01000000ff01,
0x01ff010000000000, 0x01ff010000000101, 0x01ff01000001ff00, 0x01ff0100000100ff,
0x01ff010001ff0000, 0x01ff010001000001, 0x01ff010001000100, 0x01ff010001010000,
0x01ff0101ffffffff, 0x01ff0101ffffff01, 0x01ff0101ffff01ff, 0x01ff0101ffff0101,
0x01ff0101ff000000, 0x01ff0101ff01ffff, 0x01ff0101ff01ff01, 0x01ff0101ff0101ff,
0x01ff0101ff010101, 0x01ff010100ff0000, 0x01ff01010000ff00, 0x01ff0101000000ff,
0x01ff010100000001, 0x01ff010101ffffff, 0x01ff010101ffff01, 0x01ff010101ff01ff,
0x01ff010101ff0101, 0x01ff010101000000, 0x01ff01010101ffff, 0x01ff01010101ff01,
0x01ff0101010101ff, 0x01ff010101010101, 0x0100ffffffff0000, 0x0100ffffff00ff00,
0x0100ffffff000001, 0x0100ffffff0001ff, 0x0100ffffff000100, 0x0100ffffff010000,
0x0100ffff00ffff00, 0x0100ffff00ff0001, 0x0100ffff00ff0100, 0x0100ffff00000000,
0x0100ffff000001ff, 0x0100ffff00000101, 0x0100ffff00010100, 0x0100ffff00010101,
0x0100ffff01ff0000, 0x0100ffff0100ff00, 0x0100ffff010000ff, 0x0100ffff01000001,
0x0100ffff01000100, 0x0100ffff01010000, 0x0100ff00ffffff00, 0x0100ff00ffff00ff,
0x0100ff00ffff0001, 0x0100ff00ffff0100, 0x0100ff00ff00ffff, 0x0100ff00ff000000,
0x0100ff00ff0001ff, 0x0100ff00ff000101, 0x0100ff00ff01ff00, 0x0100ff00ff0100ff,
0x0100ff00ff010001, 0x0100ff00ff010100, 0x0100ff0000ffffff, 0x0100ff0000ff0000,
0x0100ff000000ffff, 0x0100ff000000ff00, 0x0100ff00000000ff, 0x0100ff0000000000,
0x0100ff0000000001, 0x0100ff0000000100, 0x0100ff000001ff01, 0x0100ff0000010000,
0x0100ff0001ff00ff, 0x0100ff0001ff0001, 0x0100ff000100ff01, 0x0100ff0001000000,
0x0100ff00010001ff, 0x0100ff000101ff00, 0x0100ff00010100ff, 0x0100ff0001010001,
0x0100ff0001010100, 0x0100ff01ffff0000, 0x0100ff01ff00ff00, 0x0100ff01ff0000ff,
0x0100ff01ff000100, 0x0100ff01ff010000, 0x0100ff0100ff00ff, 0x0100ff0100ff0001,
0x0100ff0100ff0100, 0x0100ff010000ffff, 0x0100ff010000ff01, 0x0100ff0100000000,
0x0100ff01000001ff, 0x0100ff0100010001, 0x0100ff0100010100, 0x0100ff0101ff0000,
0x0100ff01010000ff, 0x0100ff0101000001, 0x0100ff0101010100, 0x010000ffffffff00,
0x010000ffffff00ff, 0x010000ffffff0001, 0x010000ffff00ffff, 0x010000ffff000000,
0x010000ffff0001ff, 0x010000ffff010001, 0x010000ff00ffffff, 0x010000ff00ff0101,
0x010000ff0000ff00, 0x010000ff000000ff, 0x010000ff00000000, 0x010000ff00000001,
0x010000ff000001ff, 0x010000ff00000100, 0x010000ff0001ffff, 0x010000ff0001ff00,
0x010000ff0001ff01, 0x010000ff00010000, 0x010000ff01ff00ff, 0x010000ff01ff0001,
0x010000ff0100ff01, 0x010000ff010000ff, 0x010000ff01000000, 0x010000ff010001ff,
0x010000ff0101ff00, 0x010000ff01010100, 0x01000000ffffffff, 0x01000000ffff0000,
0x01000000ffff01ff, 0x01000000ffff0101, 0x01000000ff00ffff, 0x01000000ff00ff00,
0x01000000ff0000ff, 0x01000000ff000000, 0x01000000ff000001, 0x01000000ff000100,
0x01000000ff01ff00, 0x01000000ff010000, 0x01000000ff010100, 0x01000000ff010101,
0x0100000000ffff00, 0x0100000000ff00ff, 0x0100000000ff0000, 0x0100000000ff0001,
0x0100000000ff0100, 0x010000000000ffff, 0x010000000000ff00, 0x010000000000ff01,
0x01000000000000ff, 0x0100000000000000, 0x0100000000000001, 0x01000000000001ff,
0x0100000000000100, 0x0100000000000101, 0x010000000001ff00, 0x01000000000100ff,
0x0100000000010000, 0x0100000000010001, 0x0100000000010100, 0x0100000001ffff00,
0x0100000001ff0000, 0x0100000001ff01ff, 0x010000000100ff00, 0x010000000100ff01,
0x01000000010000ff, 0x0100000001000000, 0x0100000001000001, 0x0100000001000100,
0x0100000001000101, 0x010000000101ffff, 0x010000000101ff01, 0x0100000001010000,
0x01000000010101ff, 0x0100000001010101, 0x01000001ffffff00, 0x01000001ffff00ff,
0x01000001ff00ffff, 0x01000001ff000000, 0x01000001ff000100, 0x01000001ff01ffff,
0x01000001ff010001, 0x01000001ff010100, 0x0100000100ff0000, 0x0100000100ff01ff,
0x0100000100ff0100, 0x010000010000ff00, 0x010000010000ff01, 0x0100000100000000,
0x0100000100000001, 0x0100000100000100, 0x0100000100010000, 0x01000001000101ff,
0x0100000101ffff01, 0x0100000101ff00ff, 0x0100000101ff0100, 0x0100000101ff0101,
0x010000010100ff01, 0x01000001010000ff, 0x0100000101000000, 0x01000001010100ff,
0x0100000101010001, 0x0100000101010100, 0x010001ffffff0000, 0x010001ffff000001,
0x010001ffff000100, 0x010001ffff010000, 0x010001ff00ffff00, 0x010001ff00ff0001,
0x010001ff0000ffff, 0x010001ff0000ff01, 0x010001ff00000000, 0x010001ff00000001,
0x010001ff00000101, 0x010001ff000100ff, 0x010001ff00010000, 0x010001ff01ff0000,
0x010001ff0100ff00, 0x010001ff01000001, 0x010001ff01000100, 0x010001ff01010000,
0x01000100ffff00ff, 0x01000100ffff0001, 0x01000100ffff0100, 0x01000100ff00ffff,
0x01000100ff00ff01, 0x01000100ff000000, 0x01000100ff0001ff, 0x01000100ff000101,
0x01000100ff01ffff, 0x01000100ff01ff00, 0x01000100ff0100ff, 0x01000100ff010001,
0x0100010000ffffff, 0x0100010000ffff01, 0x0100010000ff0000, 0x0100010000ff01ff,
0x0100010000ff0101, 0x010001000000ff00, 0x01000100000000ff, 0x0100010000000000,
0x0100010000000001, 0x0100010000000100, 0x010001000001ff01, 0x0100010000010000,
0x0100010000010001, 0x0100010000010101, 0x0100010001ffff00, 0x0100010001ff00ff,
0x010001000100ffff, 0x010001000100ff01, 0x0100010001000000, 0x0100010001000101,
0x010001000101ff00, 0x0100010001010001, 0x01000101ffff0000, 0x01000101ff000000,
0x01000101ff010000, 0x0100010100ff00ff, 0x0100010100ff0001, 0x0100010100ff0100,
0x010001010000ffff, 0x0100010100000000, 0x01000101000001ff, 0x010001010001ff00,
0x0100010101ff0000, 0x010001010100ff00, 0x01000101010000ff, 0x0100010101000000,
0x0100010101000001, 0x0101ffffffffffff, 0x0101ffffffffff01, 0x0101ffffffff01ff,
0x0101ffffffff0101, 0x0101ffffff000000, 0x0101ffffff01ffff, 0x0101ffffff01ff01,
0x0101ffffff0101ff, 0x0101ffffff010101, 0x0101ffff00ff0000, 0x0101ffff0000ff00,
0x0101ffff000000ff, 0x0101ffff00000001, 0x0101ffff00000100, 0x0101ffff01ffffff,
0x0101ffff01ffff01, 0x0101ffff01ff01ff, 0x0101ffff01ff0101, 0x0101ffff01000000,
0x0101ffff0101ffff, 0x0101ffff0101ff01, 0x0101ffff010101ff, 0x0101ffff01010101,
0x0101ff00ffff0000, 0x0101ff00ffff0100, 0x0101ff00ff00ff00, 0x0101ff00ff0000ff,
0x0101ff00ff000001, 0x0101ff00ff000100, 0x0101ff00ff000101, 0x0101ff0000ff0001,
0x0101ff0000ff0100, 0x0101ff000000ff00, 0x0101ff0000000000, 0x0101ff00000001ff,
0x0101ff0000000101, 0x0101ff000001ff00, 0x0101ff00000100ff, 0x0101ff0001ff0000,
0x0101ff000100ffff, 0x0101ff000100ff01, 0x0101ff0001000001, 0x0101ff0001000100,
0x0101ff01ffffff01, 0x0101ff01ffff01ff, 0x0101ff01ffff0101, 0x0101ff01ff00ffff,
0x0101ff01ff000100, 0x0101ff01ff01ff01, 0x0101ff01ff0101ff, 0x0101ff01ff010101,
0x0101ff0100ff0000, 0x0101ff010000ff00, 0x0101ff0100000001, 0x0101ff0100000100,
0x0101ff0100010000, 0x0101ff0101ffffff, 0x0101ff0101ffff01, 0x0101ff0101ff01ff,
0x0101ff0101ff0101, 0x0101ff0101000000, 0x0101ff010101ffff, 0x0101ff010101ff01,
0x0101ff01010101ff, 0x0101ff0101010101, 0x010100ffff000100, 0x010100ffff010000,
0x010100ff00ffff00, 0x010100ff00ff00ff, 0x010100ff0000ffff, 0x010100ff000000ff,
0x010100ff00000000, 0x010100ff000001ff, 0x010100ff00000101, 0x010100ff0001ff00,
0x010100ff00010000, 0x010100ff00010001, 0x010100ff000101ff, 0x010100ff00010100,
0x010100ff01ff0000, 0x01010000ffff0001, 0x01010000ffff0100, 0x01010000ff00ffff,
0x01010000ff00ff01, 0x01010000ff000000, 0x01010000ff0001ff, 0x01010000ff010001,
0x01010000ff010100, 0x0101000000ffff01, 0x0101000000ff0000, 0x010100000000ff00,
0x01010000000000ff, 0x0101000000000000, 0x0101000000000001, 0x0101000000000100,
0x0101000000010000, 0x0101000000010101, 0x0101000001ffff00, 0x0101000001ff00ff,
0x0101000001ff0000, 0x0101000001ff0001, 0x0101000001ff0100, 0x010100000100ff01,
0x0101000001000000, 0x01010000010001ff, 0x01010001ffff0000, 0x01010001ff00ff00,
0x01010001ff000001, 0x01010001ff000101, 0x01010001ff01ff00, 0x01010001ff010000,
0x0101000100ff00ff, 0x0101000100ff0001, 0x0101000100ff0101, 0x010100010000ff01,
0x0101000100000000, 0x0101000100000001, 0x01010001000001ff, 0x010100010001ffff,
0x010100010001ff01, 0x0101000101ff0001, 0x010100010100ffff, 0x0101000101000000,
0x0101000101000001, 0x0101000101000100, 0x010100010101ff00, 0x01010001010100ff,
0x0101000101010001, 0x010101ffffffffff, 0x010101ffffffff01, 0x010101ffffff01ff,
0x010101ffffff0101, 0x010101ffff01ffff, 0x010101ffff01ff01, 0x010101ffff0101ff,
0x010101ffff010101, 0x010101ff0000ff00, 0x010101ff000000ff, 0x010101ff00000001,
0x010101ff00000100, 0x010101ff01ffffff, 0x010101ff01ffff01, 0x010101ff01ff01ff,
0x010101ff01ff0101, 0x010101ff01000000, 0x010101ff0101ffff, 0x010101ff0101ff01,
0x010101ff010101ff, 0x010101ff01010101, 0x01010100ffff0000, 0x01010100ff0000ff,
0x01010100ff000100, 0x01010100ff01ff00, 0x01010100ff010000, 0x0101010000ffff00,
0x010101000000ffff, 0x0101010000000000, 0x0101010000000101, 0x010101000001ff00,
0x0101010000010001, 0x0101010000010100, 0x010101000100ffff, 0x0101010001000001,
0x01010101ffffffff, 0x01010101ffffff01, 0x01010101ffff01ff, 0x01010101ffff0101,
0x01010101ff01ffff, 0x01010101ff01ff01, 0x01010101ff0101ff, 0x01010101ff010101,
0x010101010000ff00, 0x01010101000000ff, 0x0101010100000001, 0x0101010101ffffff,
0x0101010101ffff01, 0x0101010101ff01ff, 0x0101010101ff0101, 0x0101010101000000,
0x010101010101ffff, 0x010101010101ff01, 0x01010101010101ff, 0x0101010101010101,
GGML_TABLE_END()
#else
GGML_TABLE_BEGIN(uint32_t, iq1s_grid_gpu, NGRID_IQ1S)
0x00000000, 0x00000002, 0x00000101, 0x00000200, 0x00000202, 0x00010001, 0x00010101, 0x00020000,
0x00020002, 0x00020200, 0x00020202, 0x01000101, 0x01010001, 0x01010100, 0x01010102, 0x01020101,
0x02000000, 0x02000002, 0x02000200, 0x02000202, 0x02010101, 0x02020000, 0x02020002, 0x02020200,
0x02020202, 0x00000110, 0x00000111, 0x00010011, 0x00010110, 0x00010112, 0x00010211, 0x00010212,
0x00020111, 0x01000011, 0x01000112, 0x01000211, 0x01010012, 0x01010111, 0x01010212, 0x01020011,
0x01020110, 0x01020112, 0x01020210, 0x02000111, 0x02010011, 0x02010110, 0x02010112, 0x02020111,
0x00000020, 0x00000022, 0x00000220, 0x00000222, 0x00010121, 0x00020020, 0x00020022, 0x00020220,
0x00020222, 0x01000121, 0x01010021, 0x01010221, 0x01020120, 0x01020221, 0x02000020, 0x02000022,
0x02000220, 0x02000222, 0x02010021, 0x02010121, 0x02010221, 0x02020020, 0x02020022, 0x02020220,
0x02020222, 0x00011001, 0x00011100, 0x00011102, 0x00021101, 0x01001001, 0x01001201, 0x01011101,
0x01011202, 0x01021100, 0x01021101, 0x02011001, 0x02011201, 0x02021101, 0x00001011, 0x00001110,
0x00001111, 0x00001112, 0x00011111, 0x00011210, 0x00011212, 0x00021211, 0x01001010, 0x01001111,
0x01001212, 0x01011010, 0x01011011, 0x01011110, 0x01011111, 0x01011112, 0x01011211, 0x01021010,
0x01021012, 0x01021111, 0x01021210, 0x01021212, 0x02001011, 0x02011011, 0x02011111, 0x02011210,
0x02011212, 0x02021011, 0x02021110, 0x02021111, 0x02021112, 0x02021211, 0x00011120, 0x00011221,
0x01001021, 0x01001120, 0x01011020, 0x01011022, 0x01011121, 0x01011220, 0x01021020, 0x01021021,
0x01021122, 0x01021221, 0x02001121, 0x02011021, 0x02011120, 0x02011221, 0x00002000, 0x00002002,
0x00002200, 0x00002202, 0x00012101, 0x00022000, 0x00022002, 0x00022200, 0x00022202, 0x01002101,
0x01012001, 0x01012102, 0x01022101, 0x02002000, 0x02002002, 0x02002200, 0x02002202, 0x02012101,
0x02022000, 0x02022002, 0x02022200, 0x02022202, 0x00002111, 0x00012011, 0x00012110, 0x00012211,
0x00022110, 0x00022111, 0x01002011, 0x01012010, 0x01012011, 0x01012111, 0x01022011, 0x01022110,
0x01022211, 0x02012011, 0x02012110, 0x02012112, 0x02012211, 0x02022111, 0x00002020, 0x00002022,
0x00002220, 0x00002222, 0x00012121, 0x00022020, 0x00022022, 0x00022220, 0x00022222, 0x01002121,
0x01012021, 0x01012221, 0x01022021, 0x01022121, 0x02002020, 0x02002022, 0x02002121, 0x02002220,
0x02002222, 0x02012121, 0x02022020, 0x02022022, 0x02022220, 0x02022222, 0x00110000, 0x00110001,
0x00110100, 0x00110201, 0x00120100, 0x00120101, 0x01100001, 0x01100100, 0x01110000, 0x01110101,
0x01110200, 0x01120001, 0x01120100, 0x01120101, 0x01120201, 0x02110001, 0x02110100, 0x02110102,
0x02120001, 0x02120101, 0x00100011, 0x00100110, 0x00100112, 0x00100211, 0x00110010, 0x00110012,
0x00110111, 0x00110210, 0x00120011, 0x00120110, 0x00120211, 0x01100111, 0x01100212, 0x01110010,
0x01110011, 0x01110012, 0x01110110, 0x01110111, 0x01110112, 0x01110211, 0x01120010, 0x01120111,
0x02100110, 0x02110012, 0x02110111, 0x02120011, 0x02120110, 0x00110021, 0x00110120, 0x00110122,
0x00120121, 0x01100020, 0x01100122, 0x01100221, 0x01110022, 0x01110121, 0x01110220, 0x01110222,
0x01120120, 0x01120122, 0x02100121, 0x02110021, 0x02110120, 0x02110122, 0x02120121, 0x00101001,
0x00101102, 0x00101201, 0x00111100, 0x00111101, 0x00111200, 0x00111201, 0x00121001, 0x00121102,
0x01101001, 0x01101101, 0x01101102, 0x01101200, 0x01101202, 0x01111001, 0x01111100, 0x01111101,
0x01111102, 0x01111201, 0x01121002, 0x01121101, 0x01121200, 0x02101100, 0x02101201, 0x02111000,
0x02111100, 0x02111101, 0x02111200, 0x02111201, 0x02111202, 0x02121001, 0x02121100, 0x02121101,
0x02121201, 0x00101012, 0x00101111, 0x00101212, 0x00111011, 0x00111110, 0x00111111, 0x00111112,
0x00111211, 0x00121010, 0x00121012, 0x00121111, 0x00121210, 0x00121212, 0x01101011, 0x01101110,
0x01101111, 0x01101112, 0x01111011, 0x01111012, 0x01111110, 0x01111111, 0x01111112, 0x01111211,
0x01111212, 0x01121011, 0x01121110, 0x01121111, 0x01121112, 0x01121211, 0x02101010, 0x02101012,
0x02101110, 0x02101111, 0x02101210, 0x02101212, 0x02111010, 0x02111011, 0x02111110, 0x02111111,
0x02111112, 0x02111211, 0x02111212, 0x02121010, 0x02121012, 0x02121111, 0x00101021, 0x00101120,
0x00101121, 0x00101122, 0x00111121, 0x00111122, 0x00111220, 0x00111222, 0x00121021, 0x00121122,
0x01101020, 0x01101022, 0x01101120, 0x01101121, 0x01101220, 0x01101222, 0x01111021, 0x01111121,
0x01111122, 0x01111220, 0x01111221, 0x01121021, 0x01121120, 0x01121121, 0x01121220, 0x01121221,
0x01121222, 0x02101122, 0x02101222, 0x02111022, 0x02111121, 0x02121120, 0x02121221, 0x00112001,
0x00112102, 0x00122101, 0x01102001, 0x01102100, 0x01102102, 0x01102201, 0x01112000, 0x01112101,
0x01112200, 0x01112202, 0x01122000, 0x01122001, 0x01122100, 0x01122102, 0x01122201, 0x02102101,
0x02112001, 0x02112100, 0x02122101, 0x00112010, 0x00112012, 0x00112111, 0x00112212, 0x00122011,
0x00122111, 0x01102012, 0x01102110, 0x01102111, 0x01102210, 0x01112011, 0x01112110, 0x01112111,
0x01112112, 0x01112211, 0x01112212, 0x01122010, 0x01122111, 0x01122212, 0x02102211, 0x02112011,
0x02112012, 0x02112111, 0x02112210, 0x02122011, 0x02122112, 0x02122211, 0x00102221, 0x00112122,
0x00122120, 0x00122122, 0x01102120, 0x01102122, 0x01102221, 0x01112020, 0x01112022, 0x01112121,
0x01112220, 0x01122021, 0x01122122, 0x01122221, 0x02102121, 0x02112021, 0x02112122, 0x02112222,
0x00200000, 0x00200002, 0x00200200, 0x00200202, 0x00210101, 0x00220000, 0x00220002, 0x00220101,
0x00220200, 0x00220202, 0x01200101, 0x01210001, 0x01210201, 0x01220001, 0x01220101, 0x02200000,
0x02200002, 0x02200200, 0x02200202, 0x02210101, 0x02220000, 0x02220002, 0x02220101, 0x02220200,
0x02220202, 0x00200111, 0x00210011, 0x00210110, 0x00210211, 0x00220111, 0x01200012, 0x01200110,
0x01200211, 0x01210111, 0x01210210, 0x01210212, 0x01220011, 0x01220110, 0x01220111, 0x01220112,
0x02200111, 0x02210010, 0x02210112, 0x02210211, 0x02220111, 0x00200021, 0x00200220, 0x00200222,
0x00210021, 0x00210121, 0x00220020, 0x00220022, 0x00220220, 0x00220222, 0x01200121, 0x01210021,
0x01210122, 0x01210221, 0x01220121, 0x02200021, 0x02200220, 0x02200222, 0x02210021, 0x02210121,
0x02220020, 0x02220022, 0x02220220, 0x02220222, 0x00201101, 0x00211100, 0x00211102, 0x00211201,
0x00221101, 0x01201100, 0x01201101, 0x01201102, 0x01201201, 0x01211002, 0x01211101, 0x01211200,
0x01211202, 0x01221102, 0x02201101, 0x02211001, 0x02211100, 0x02211201, 0x02221001, 0x02221101,
0x00201211, 0x00211111, 0x00221011, 0x00221211, 0x01201010, 0x01201111, 0x01201210, 0x01211011,
0x01211110, 0x01211111, 0x01211211, 0x01221012, 0x01221111, 0x01221210, 0x02201211, 0x02211010,
0x02211110, 0x02211111, 0x02211210, 0x02211212, 0x02221011, 0x02221110, 0x02221112, 0x02221211,
0x00201121, 0x00211020, 0x00211022, 0x00211221, 0x00221121, 0x01201021, 0x01201221, 0x01211121,
0x01221020, 0x01221021, 0x01221221, 0x02201120, 0x02201122, 0x02211020, 0x02211222, 0x00202000,
0x00202002, 0x00202200, 0x00202202, 0x00212101, 0x00222000, 0x00222002, 0x00222200, 0x00222202,
0x01202101, 0x01212001, 0x01212100, 0x01222101, 0x02202000, 0x02202002, 0x02202200, 0x02202202,
0x02222000, 0x02222002, 0x02222200, 0x02222202, 0x00202211, 0x00212011, 0x00212110, 0x00212211,
0x00222111, 0x01202112, 0x01202211, 0x01212012, 0x01212111, 0x01222011, 0x01222110, 0x01222112,
0x01222211, 0x02202111, 0x02212010, 0x02212112, 0x02212211, 0x02222110, 0x02222111, 0x00202020,
0x00202022, 0x00202220, 0x00202222, 0x00222020, 0x00222022, 0x00222220, 0x00222222, 0x01202121,
0x01212021, 0x01212122, 0x01212221, 0x01222121, 0x02202020, 0x02202022, 0x02202220, 0x02202222,
0x02212121, 0x02222020, 0x02222022, 0x02222220, 0x02222222, 0x10000101, 0x10010001, 0x10010102,
0x10020101, 0x11000201, 0x11010002, 0x11010101, 0x11010200, 0x11010202, 0x11020001, 0x11020100,
0x11020102, 0x12010100, 0x12010201, 0x12020001, 0x12020102, 0x10000010, 0x10000011, 0x10000110,
0x10000112, 0x10000211, 0x10010012, 0x10010111, 0x10010112, 0x10010210, 0x10010212, 0x10020011,
0x10020112, 0x10020211, 0x11000111, 0x11000210, 0x11000212, 0x11010011, 0x11010110, 0x11010111,
0x11010112, 0x11010211, 0x11010212, 0x11020111, 0x11020210, 0x11020212, 0x12000011, 0x12000110,
0x12000112, 0x12010010, 0x12010012, 0x12010111, 0x12020010, 0x12020011, 0x12020012, 0x10000121,
0x10010021, 0x10010120, 0x10010122, 0x10020121, 0x11000021, 0x11010022, 0x11010121, 0x11010222,
0x11020120, 0x11020221, 0x12000221, 0x12010120, 0x12020121, 0x10001001, 0x10011101, 0x10011201,
0x10021201, 0x11001101, 0x11001200, 0x11001202, 0x11011001, 0x11011100, 0x11011101, 0x11011102,
0x11021001, 0x11021002, 0x11021101, 0x11021200, 0x11021202, 0x12001001, 0x12001102, 0x12001201,
0x12011000, 0x12011002, 0x12011101, 0x12021000, 0x12021001, 0x12021201, 0x10001011, 0x10001012,
0x10001111, 0x10001212, 0x10011011, 0x10011110, 0x10011111, 0x10011112, 0x10011211, 0x10021010,
0x10021111, 0x10021212, 0x11001011, 0x11001110, 0x11001111, 0x11001112, 0x11001211, 0x11011010,
0x11011011, 0x11011110, 0x11011111, 0x11011112, 0x11011210, 0x11011211, 0x11021011, 0x11021110,
0x11021111, 0x11021112, 0x11021211, 0x12001012, 0x12001110, 0x12001111, 0x12001210, 0x12011011,
0x12011110, 0x12011111, 0x12011112, 0x12011211, 0x12011212, 0x12021111, 0x12021210, 0x12021212,
0x10001021, 0x10001121, 0x10001221, 0x10011120, 0x10011121, 0x10011220, 0x10011222, 0x10021021,
0x10021120, 0x10021221, 0x11001020, 0x11001022, 0x11001121, 0x11001220, 0x11011020, 0x11011021,
0x11011022, 0x11011121, 0x11011122, 0x11011221, 0x11021022, 0x11021121, 0x11021220, 0x12001021,
0x12001121, 0x12001222, 0x12011120, 0x12011121, 0x12021021, 0x12021120, 0x12021122, 0x10002101,
0x10012001, 0x10012101, 0x10012202, 0x10022101, 0x11002002, 0x11002201, 0x11012000, 0x11012101,
0x11012200, 0x11022001, 0x11022100, 0x11022102, 0x11022201, 0x12002101, 0x12012001, 0x12012100,
0x12012102, 0x12012201, 0x12022101, 0x10002011, 0x10002111, 0x10002112, 0x10002212, 0x10012010,
0x10012110, 0x10012111, 0x10012210, 0x10022011, 0x10022110, 0x10022112, 0x11002010, 0x11002111,
0x11002212, 0x11012011, 0x11012012, 0x11012110, 0x11012111, 0x11012112, 0x11012211, 0x11022010,
0x11022012, 0x11022111, 0x11022112, 0x11022212, 0x12002112, 0x12002211, 0x12012012, 0x12012111,
0x12012112, 0x12012210, 0x12022011, 0x12022110, 0x12022112, 0x12022211, 0x10012122, 0x11002120,
0x11002122, 0x11002221, 0x11012121, 0x11012220, 0x11012222, 0x11022120, 0x11022221, 0x12012120,
0x12022121, 0x10100001, 0x10100100, 0x10100101, 0x10100102, 0x10100201, 0x10110002, 0x10110101,
0x10110202, 0x10120001, 0x10120100, 0x10120201, 0x11100000, 0x11100101, 0x11100200, 0x11110001,
0x11110100, 0x11110101, 0x11110102, 0x11110201, 0x11120101, 0x11120200, 0x12100102, 0x12100201,
0x12110101, 0x12110200, 0x12120000, 0x12120001, 0x12120102, 0x12120201, 0x10100111, 0x10100210,
0x10100211, 0x10100212, 0x10110011, 0x10110110, 0x10110111, 0x10110112, 0x10110210, 0x10110211,
0x10120010, 0x10120111, 0x10120112, 0x10120210, 0x10120212, 0x11100011, 0x11100110, 0x11100111,
0x11100112, 0x11100211, 0x11110010, 0x11110011, 0x11110012, 0x11110110, 0x11110111, 0x11110112,
0x11110210, 0x11110211, 0x11110212, 0x11120011, 0x11120110, 0x11120111, 0x11120112, 0x11120211,
0x12100012, 0x12100111, 0x12110011, 0x12110110, 0x12110111, 0x12110112, 0x12110211, 0x12120010,
0x12120111, 0x12120212, 0x10100021, 0x10100122, 0x10110022, 0x10110121, 0x10110222, 0x10120021,
0x10120120, 0x11100022, 0x11100121, 0x11100222, 0x11110021, 0x11110120, 0x11110121, 0x11110122,
0x11110221, 0x11120022, 0x11120121, 0x12100121, 0x12110020, 0x12110022, 0x12110121, 0x12110221,
0x12110222, 0x12120120, 0x10101100, 0x10101101, 0x10111001, 0x10111100, 0x10111101, 0x10111102,
0x10111200, 0x10111201, 0x10121001, 0x10121101, 0x10121200, 0x10121202, 0x11101001, 0x11101100,
0x11101101, 0x11101102, 0x11101201, 0x11101202, 0x11111000, 0x11111001, 0x11111100, 0x11111101,
0x11111102, 0x11111200, 0x11111201, 0x11111202, 0x11121001, 0x11121002, 0x11121100, 0x11121101,
0x11121102, 0x11121201, 0x12101000, 0x12101200, 0x12101202, 0x12111001, 0x12111100, 0x12111101,
0x12111102, 0x12111201, 0x12121001, 0x12121100, 0x12121101, 0x12121202, 0x10101011, 0x10101012,
0x10101110, 0x10101111, 0x10101112, 0x10101211, 0x10111010, 0x10111011, 0x10111012, 0x10111110,
0x10111111, 0x10111112, 0x10111211, 0x10111212, 0x10121011, 0x10121110, 0x10121111, 0x10121112,
0x10121211, 0x11101010, 0x11101011, 0x11101012, 0x11101110, 0x11101111, 0x11101112, 0x11101210,
0x11101211, 0x11111010, 0x11111011, 0x11111012, 0x11111110, 0x11111111, 0x11111112, 0x11111210,
0x11111211, 0x11111212, 0x11121010, 0x11121011, 0x11121110, 0x11121111, 0x11121112, 0x11121210,
0x11121211, 0x11121212, 0x12101011, 0x12101110, 0x12101111, 0x12101211, 0x12101212, 0x12111010,
0x12111011, 0x12111110, 0x12111111, 0x12111112, 0x12111210, 0x12111211, 0x12121011, 0x12121110,
0x12121111, 0x12121112, 0x12121211, 0x10101020, 0x10101021, 0x10101022, 0x10101120, 0x10101122,
0x10101220, 0x10101221, 0x10111021, 0x10111120, 0x10111121, 0x10111220, 0x10111221, 0x10121020,
0x10121021, 0x10121022, 0x10121120, 0x10121121, 0x10121122, 0x10121220, 0x10121221, 0x11101021,
0x11101121, 0x11101122, 0x11101220, 0x11101221, 0x11101222, 0x11111020, 0x11111021, 0x11111022,
0x11111120, 0x11111121, 0x11111122, 0x11111220, 0x11111221, 0x11111222, 0x11121021, 0x11121120,
0x11121121, 0x11121221, 0x12101022, 0x12101121, 0x12101122, 0x12101220, 0x12101221, 0x12101222,
0x12111021, 0x12111121, 0x12111222, 0x12121022, 0x12121121, 0x12121122, 0x12121220, 0x12121221,
0x10102100, 0x10102101, 0x10102102, 0x10102201, 0x10112000, 0x10112101, 0x10112200, 0x10122001,
0x10122202, 0x11102101, 0x11102200, 0x11102202, 0x11112001, 0x11112100, 0x11112101, 0x11112102,
0x11112200, 0x11112201, 0x11122000, 0x11122002, 0x11122100, 0x11122101, 0x12102002, 0x12102201,
0x12112000, 0x12112002, 0x12112101, 0x12112200, 0x12122001, 0x12122201, 0x10102011, 0x10102012,
0x10102111, 0x10102212, 0x10112011, 0x10112110, 0x10112111, 0x10112112, 0x10112211, 0x10122111,
0x11102011, 0x11102110, 0x11102111, 0x11102112, 0x11102211, 0x11112010, 0x11112011, 0x11112012,
0x11112110, 0x11112111, 0x11112112, 0x11112210, 0x11112211, 0x11112212, 0x11122011, 0x11122110,
0x11122111, 0x11122112, 0x11122211, 0x12102011, 0x12102111, 0x12102211, 0x12112011, 0x12112110,
0x12112111, 0x12112112, 0x12112210, 0x12112211, 0x12122111, 0x10102120, 0x10102220, 0x10112121,
0x10112222, 0x10122020, 0x10122121, 0x10122122, 0x10122221, 0x11102121, 0x11102220, 0x11102221,
0x11112021, 0x11112121, 0x11112122, 0x11112220, 0x11112221, 0x11122022, 0x11122121, 0x11122220,
0x11122222, 0x12102021, 0x12102222, 0x12112022, 0x12112121, 0x12112122, 0x12112220, 0x12112222,
0x12122021, 0x10200101, 0x10210100, 0x10210102, 0x10210201, 0x10220101, 0x11200100, 0x11210000,
0x11210101, 0x11210102, 0x11210200, 0x11210202, 0x11220001, 0x11220100, 0x11220102, 0x11220201,
0x12200001, 0x12210102, 0x12220101, 0x10200011, 0x10200110, 0x10200112, 0x10200211, 0x10210012,
0x10210111, 0x10220011, 0x10220012, 0x10220112, 0x10220211, 0x11200111, 0x11200211, 0x11210011,
0x11210111, 0x11210112, 0x11210211, 0x11220111, 0x11220112, 0x11220212, 0x12200110, 0x12200212,
0x12210012, 0x12210111, 0x12220011, 0x12220112, 0x12220211, 0x10210021, 0x10210122, 0x10210221,
0x11200020, 0x11200021, 0x11200122, 0x11210121, 0x11210122, 0x11210220, 0x11220020, 0x12200121,
0x12210021, 0x12210122, 0x12220121, 0x10211001, 0x10211002, 0x10211101, 0x10211102, 0x10211202,
0x10221001, 0x10221102, 0x10221201, 0x11201000, 0x11201002, 0x11201101, 0x11201200, 0x11201202,
0x11211001, 0x11211100, 0x11211101, 0x11211102, 0x11211201, 0x11211202, 0x11221000, 0x11221002,
0x11221101, 0x12201100, 0x12201101, 0x12201201, 0x12211000, 0x12211002, 0x12211100, 0x12211101,
0x12211102, 0x12211200, 0x12211202, 0x12221001, 0x12221100, 0x12221201, 0x10201111, 0x10201210,
0x10201212, 0x10211011, 0x10211111, 0x10211112, 0x10211211, 0x11201110, 0x11201111, 0x11201112,
0x11201211, 0x11211010, 0x11211011, 0x11211110, 0x11211111, 0x11211112, 0x11211211, 0x11221011,
0x11221110, 0x11221111, 0x11221112, 0x11221211, 0x12201112, 0x12201211, 0x12201212, 0x12211011,
0x12211111, 0x12211112, 0x12211211, 0x12211212, 0x12221012, 0x12221111, 0x12221112, 0x12221210,
0x10201022, 0x10201221, 0x10211121, 0x10221020, 0x10221122, 0x10221220, 0x10221221, 0x11201020,
0x11201121, 0x11201220, 0x11201222, 0x11211021, 0x11211120, 0x11211121, 0x11211122, 0x11211220,
0x11211222, 0x11221020, 0x11221121, 0x11221220, 0x12201020, 0x12201022, 0x12201121, 0x12201222,
0x12211120, 0x12211122, 0x12211220, 0x12211221, 0x12221020, 0x12221120, 0x12221122, 0x12221222,
0x10212102, 0x10212201, 0x10222101, 0x11202001, 0x11212002, 0x11212101, 0x11212202, 0x11222001,
0x11222201, 0x12202101, 0x12212001, 0x12212200, 0x12222102, 0x10202011, 0x10202110, 0x10212010,
0x10212111, 0x10222011, 0x10222110, 0x10222112, 0x10222211, 0x11202010, 0x11202011, 0x11202111,
0x11202112, 0x11202210, 0x11212011, 0x11212110, 0x11212111, 0x11212112, 0x11212211, 0x11222010,
0x11222111, 0x11222212, 0x12202012, 0x12202110, 0x12202212, 0x12212111, 0x12222011, 0x12222110,
0x12222111, 0x12222211, 0x10212021, 0x10212122, 0x10212220, 0x11202021, 0x11202120, 0x11202221,
0x11212020, 0x11212121, 0x11212220, 0x11212222, 0x11222120, 0x11222121, 0x11222221, 0x12202122,
0x12212120, 0x12212220, 0x12212222, 0x12222122, 0x20000000, 0x20000002, 0x20000200, 0x20000202,
0x20020000, 0x20020002, 0x20020200, 0x20020202, 0x21000101, 0x21010000, 0x21010001, 0x21010100,
0x21010102, 0x21010201, 0x21020101, 0x22000000, 0x22000002, 0x22000200, 0x22000202, 0x22010101,
0x22020000, 0x22020002, 0x22020200, 0x22020202, 0x20000111, 0x20010011, 0x20010110, 0x20010112,
0x20010211, 0x20020111, 0x21000011, 0x21000110, 0x21000211, 0x21010010, 0x21010012, 0x21010111,
0x21010112, 0x21010210, 0x21010211, 0x21020110, 0x21020112, 0x21020211, 0x22000111, 0x22000211,
0x22010110, 0x22010112, 0x22010211, 0x22020111, 0x20000020, 0x20000022, 0x20000220, 0x20000222,
0x20010121, 0x20020020, 0x20020022, 0x20020220, 0x20020222, 0x21010021, 0x21010120, 0x21010221,
0x21020121, 0x22000020, 0x22000022, 0x22000220, 0x22000222, 0x22010121, 0x22020020, 0x22020022,
0x22020220, 0x22020222, 0x20011100, 0x20011201, 0x21001001, 0x21001100, 0x21011001, 0x21011101,
0x21011202, 0x21021001, 0x21021100, 0x21021201, 0x22011100, 0x22011201, 0x20001011, 0x20001211,
0x20011012, 0x20011111, 0x20011212, 0x20021112, 0x20021211, 0x21001010, 0x21001011, 0x21001111,
0x21001210, 0x21011011, 0x21011110, 0x21011111, 0x21011112, 0x21011211, 0x21011212, 0x21021111,
0x21021112, 0x21021210, 0x21021212, 0x22001011, 0x22001110, 0x22001112, 0x22001211, 0x22011010,
0x22011012, 0x22011111, 0x22011210, 0x22021112, 0x20011021, 0x20011122, 0x20011221, 0x20021121,
0x21001021, 0x21001120, 0x21001221, 0x21001222, 0x21011020, 0x21011121, 0x21011221, 0x21011222,
0x21021021, 0x21021122, 0x21021222, 0x22001121, 0x22011021, 0x22011222, 0x22021120, 0x20002000,
0x20002002, 0x20002200, 0x20002202, 0x20012101, 0x20022000, 0x20022002, 0x20022200, 0x20022202,
0x21002001, 0x21002101, 0x21012001, 0x21012100, 0x21012201, 0x21022101, 0x21022201, 0x22002000,
0x22002002, 0x22002200, 0x22002202, 0x22012101, 0x22022000, 0x22022002, 0x22022200, 0x22022202,
0x20002111, 0x20002112, 0x20012011, 0x20012110, 0x20012112, 0x20022111, 0x21002011, 0x21002110,
0x21002112, 0x21002211, 0x21012010, 0x21012012, 0x21012111, 0x21012212, 0x21022011, 0x21022110,
0x22002111, 0x22012112, 0x22012211, 0x22022111, 0x20002020, 0x20002022, 0x20002220, 0x20002222,
0x20012121, 0x20022020, 0x20022022, 0x20022220, 0x20022222, 0x21002121, 0x21012021, 0x21012120,
0x21012122, 0x22002020, 0x22002022, 0x22002220, 0x22002222, 0x22012121, 0x22022020, 0x22022022,
0x22022220, 0x22022222, 0x20100101, 0x20110001, 0x20110102, 0x20110200, 0x20110201, 0x20120101,
0x21100001, 0x21100102, 0x21100201, 0x21110101, 0x21110200, 0x21110202, 0x21120201, 0x21120202,
0x22100101, 0x22110001, 0x22110100, 0x22110102, 0x22110201, 0x22120101, 0x20100011, 0x20100110,
0x20100112, 0x20100211, 0x20110010, 0x20110111, 0x20110210, 0x20110212, 0x20120011, 0x20120110,
0x20120112, 0x20120211, 0x21100010, 0x21100111, 0x21110010, 0x21110011, 0x21110110, 0x21110111,
0x21110112, 0x21110211, 0x21120012, 0x21120111, 0x22100110, 0x22100112, 0x22110012, 0x22110111,
0x22110210, 0x22120011, 0x22120110, 0x22120112, 0x22120211, 0x20100121, 0x20110021, 0x20110120,
0x20110221, 0x20120121, 0x21100120, 0x21100122, 0x21100221, 0x21110020, 0x21110022, 0x21110121,
0x21110220, 0x21120122, 0x21120221, 0x22100121, 0x22110120, 0x22110122, 0x22120221, 0x20101001,
0x20101100, 0x20101102, 0x20111000, 0x20111101, 0x20111200, 0x20121102, 0x21101000, 0x21101202,
0x21111001, 0x21111100, 0x21111101, 0x21111102, 0x21111200, 0x21111201, 0x21121000, 0x21121001,
0x21121002, 0x21121101, 0x22101100, 0x22101102, 0x22111002, 0x22111100, 0x22111101, 0x22111200,
0x22121001, 0x22121201, 0x20101010, 0x20101111, 0x20101210, 0x20101212, 0x20111010, 0x20111011,
0x20111110, 0x20111111, 0x20111112, 0x20111211, 0x20121011, 0x20121111, 0x20121211, 0x20121212,
0x21101011, 0x21101110, 0x21101111, 0x21101112, 0x21101211, 0x21111010, 0x21111011, 0x21111012,
0x21111110, 0x21111111, 0x21111112, 0x21111210, 0x21111211, 0x21111212, 0x21121011, 0x21121110,
0x21121111, 0x21121112, 0x21121211, 0x22101011, 0x22101111, 0x22101210, 0x22111011, 0x22111012,
0x22111110, 0x22111111, 0x22111112, 0x22111211, 0x22111212, 0x22121010, 0x22121012, 0x22121111,
0x22121210, 0x22121212, 0x20101021, 0x20101120, 0x20111020, 0x20111121, 0x20111221, 0x20121020,
0x20121122, 0x20121221, 0x21101121, 0x21101220, 0x21101221, 0x21111021, 0x21111022, 0x21111121,
0x21111122, 0x21111221, 0x21121121, 0x21121220, 0x22101022, 0x22101120, 0x22101221, 0x22101222,
0x22111022, 0x22111120, 0x22111121, 0x22121120, 0x22121122, 0x22121221, 0x20102101, 0x20112102,
0x20112201, 0x20122101, 0x21102001, 0x21102102, 0x21112000, 0x21112002, 0x21112101, 0x21112102,
0x21112202, 0x21122100, 0x21122101, 0x22102101, 0x22112001, 0x22112102, 0x22112201, 0x22122101,
0x20102110, 0x20102112, 0x20102211, 0x20112010, 0x20112012, 0x20112111, 0x20112210, 0x20112212,
0x20122010, 0x20122011, 0x20122110, 0x20122112, 0x21102010, 0x21102012, 0x21102111, 0x21102210,
0x21102212, 0x21112011, 0x21112110, 0x21112111, 0x21112112, 0x21112211, 0x21122012, 0x21122111,
0x21122112, 0x21122212, 0x22102011, 0x22102110, 0x22112010, 0x22112012, 0x22112111, 0x22112212,
0x22122011, 0x22122112, 0x20102121, 0x20112121, 0x20122121, 0x21102120, 0x21102122, 0x21102221,
0x21112020, 0x21112121, 0x21112220, 0x21122021, 0x22102121, 0x22112021, 0x22112120, 0x22112121,
0x22112122, 0x20200000, 0x20200002, 0x20200200, 0x20200202, 0x20210101, 0x20220000, 0x20220002,
0x20220200, 0x20220202, 0x21200101, 0x21210001, 0x21210100, 0x21210102, 0x21210201, 0x22200000,
0x22200002, 0x22200200, 0x22200202, 0x22210101, 0x22220000, 0x22220002, 0x22220200, 0x22220202,
0x20200111, 0x20200211, 0x20210011, 0x20210110, 0x20210112, 0x20210211, 0x20210212, 0x21200112,
0x21200211, 0x21210011, 0x21210111, 0x21210210, 0x21210212, 0x21220011, 0x21220110, 0x22200111,
0x22210010, 0x22210012, 0x22210112, 0x22210211, 0x20200022, 0x20200220, 0x20200222, 0x20210020,
0x20210221, 0x20220022, 0x20220220, 0x20220222, 0x21200121, 0x21210021, 0x21210122, 0x21210221,
0x21220121, 0x22200020, 0x22200022, 0x22200220, 0x22200222, 0x22210121, 0x22220020, 0x22220022,
0x22220220, 0x22220222, 0x20211201, 0x20221101, 0x21201001, 0x21201100, 0x21211000, 0x21211100,
0x21211101, 0x21211200, 0x21211202, 0x21221001, 0x21221101, 0x21221102, 0x21221200, 0x21221201,
0x22201101, 0x20201112, 0x20201211, 0x20211010, 0x20211012, 0x20211111, 0x20211210, 0x20221112,
0x20221211, 0x21201012, 0x21201111, 0x21211011, 0x21211110, 0x21211111, 0x21211112, 0x21211211,
0x21221111, 0x21221212, 0x22201011, 0x22201110, 0x22201111, 0x22201112, 0x22201211, 0x22211012,
0x22211111, 0x22211210, 0x20201121, 0x20211021, 0x20211122, 0x20211222, 0x20221021, 0x20221121,
0x21201120, 0x21201122, 0x21201222, 0x21211022, 0x21211121, 0x21211122, 0x21211220, 0x21221020,
0x21221022, 0x22201122, 0x22211020, 0x22211121, 0x22211122, 0x22211221, 0x22221021, 0x22221120,
0x22221122, 0x20202000, 0x20202002, 0x20202200, 0x20202202, 0x20222000, 0x20222002, 0x20222200,
0x20222202, 0x21212001, 0x21212100, 0x21212102, 0x21212201, 0x22202000, 0x22202002, 0x22202200,
0x22202202, 0x22212101, 0x22222000, 0x22222002, 0x22222200, 0x22222202, 0x20202111, 0x20212110,
0x20212211, 0x20222011, 0x20222111, 0x21202011, 0x21212010, 0x21212111, 0x21212212, 0x21222011,
0x21222112, 0x21222211, 0x22212010, 0x22212112, 0x20202020, 0x20202022, 0x20202220, 0x20202222,
0x20222020, 0x20222022, 0x20222220, 0x20222222, 0x21212021, 0x21212120, 0x21212122, 0x22202020,
0x22202022, 0x22202220, 0x22202222, 0x22212121, 0x22222020, 0x22222022, 0x22222220, 0x22222222,
GGML_TABLE_END()
#endif
#endif // GGML_COMMON_IMPL
#endif // GGML_COMMON_IMPL
#include "ggml-cuda.h"
#include "ggml.h"
#include "ggml-backend-impl.h"
#include "ggml-cuda/common.cuh"
#include "ggml-cuda/acc.cuh"
#include "ggml-cuda/arange.cuh"
#include "ggml-cuda/argsort.cuh"
#include "ggml-cuda/binbcast.cuh"
#include "ggml-cuda/clamp.cuh"
#include "ggml-cuda/concat.cuh"
#include "ggml-cuda/convert.cuh"
#include "ggml-cuda/cpy.cuh"
#include "ggml-cuda/diagmask.cuh"
#include "ggml-cuda/dmmv.cuh"
#include "ggml-cuda/fattn.cuh"
#include "ggml-cuda/getrows.cuh"
#include "ggml-cuda/im2col.cuh"
#include "ggml-cuda/mmq.cuh"
#include "ggml-cuda/mmvq.cuh"
#include "ggml-cuda/norm.cuh"
#include "ggml-cuda/pad.cuh"
#include "ggml-cuda/pool2d.cuh"
#include "ggml-cuda/quantize.cuh"
#include "ggml-cuda/rope.cuh"
#include "ggml-cuda/scale.cuh"
#include "ggml-cuda/softmax.cuh"
#include "ggml-cuda/sumrows.cuh"
#include "ggml-cuda/tsembd.cuh"
#include "ggml-cuda/unary.cuh"
#include "ggml-cuda/upscale.cuh"
#include <algorithm>
#include <array>
#include <atomic>
#include <cinttypes>
#include <cstddef>
#include <cstdint>
#include <float.h>
#include <limits>
#include <map>
#include <memory>
#include <mutex>
#include <stdint.h>
#include <stdio.h>
#include <stdarg.h>
#include <stdlib.h>
#include <string>
#include <vector>
static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
static void ggml_cuda_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
GGML_UNUSED(level);
GGML_UNUSED(user_data);
fprintf(stderr, "%s", msg);
}
ggml_log_callback ggml_cuda_log_callback = ggml_cuda_default_log_callback;
void * ggml_cuda_log_user_data = NULL;
GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data) {
ggml_cuda_log_callback = log_callback;
ggml_cuda_log_user_data = user_data;
}
#define GGML_CUDA_LOG_INFO(...) ggml_cuda_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
#define GGML_CUDA_LOG_WARN(...) ggml_cuda_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
#define GGML_CUDA_LOG_ERROR(...) ggml_cuda_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
GGML_ATTRIBUTE_FORMAT(2, 3)
static void ggml_cuda_log(enum ggml_log_level level, const char * format, ...) {
if (ggml_cuda_log_callback != NULL) {
va_list args;
va_start(args, format);
char buffer[128];
int len = vsnprintf(buffer, 128, format, args);
if (len < 128) {
ggml_cuda_log_callback(level, buffer, ggml_cuda_log_user_data);
} else {
std::vector<char> buffer2(len + 1); // vsnprintf adds a null terminator
va_end(args);
va_start(args, format);
vsnprintf(&buffer2[0], buffer2.size(), format, args);
ggml_cuda_log_callback(level, buffer2.data(), ggml_cuda_log_user_data);
}
va_end(args);
}
}
[[noreturn]]
void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg) {
int id = -1; // in case cudaGetDevice fails
cudaGetDevice(&id);
GGML_CUDA_LOG_ERROR("CUDA error: %s\n", msg);
GGML_CUDA_LOG_ERROR(" current device: %d, in function %s at %s:%d\n", id, func, file, line);
GGML_CUDA_LOG_ERROR(" %s\n", stmt);
// abort with GGML_ASSERT to get a stack trace
GGML_ASSERT(!"CUDA error");
}
// this is faster on Windows
// probably because the Windows CUDA libraries forget to make this check before invoking the drivers
void ggml_cuda_set_device(int device) {
int current_device;
CUDA_CHECK(cudaGetDevice(&current_device));
if (device == current_device) {
return;
}
CUDA_CHECK(cudaSetDevice(device));
}
int ggml_cuda_get_device() {
int id;
CUDA_CHECK(cudaGetDevice(&id));
return id;
}
static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
ggml_cuda_set_device(device);
#if defined(GGML_USE_HIPBLAS) && defined(GGML_HIP_UMA)
auto res = hipMallocManaged(ptr, size);
if (res == hipSuccess) {
// if error we "need" to know why...
CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
}
return res;
#else
return cudaMalloc(ptr, size);
#endif
}
static ggml_cuda_device_info ggml_cuda_init() {
#ifdef __HIP_PLATFORM_AMD__
// Workaround for a rocBLAS bug when using multiple graphics cards:
// https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
rocblas_initialize();
CUDA_CHECK(cudaDeviceSynchronize());
#endif
ggml_cuda_device_info info = {};
cudaError_t err = cudaGetDeviceCount(&info.device_count);
if (err != cudaSuccess) {
GGML_CUDA_LOG_ERROR("%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err));
return info;
}
GGML_ASSERT(info.device_count <= GGML_CUDA_MAX_DEVICES);
int64_t total_vram = 0;
#if defined(GGML_CUDA_FORCE_MMQ)
GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ: yes\n", __func__);
#else
GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ: no\n", __func__);
#endif
#if defined(CUDA_USE_TENSOR_CORES)
GGML_CUDA_LOG_INFO("%s: CUDA_USE_TENSOR_CORES: yes\n", __func__);
#else
GGML_CUDA_LOG_INFO("%s: CUDA_USE_TENSOR_CORES: no\n", __func__);
#endif
GGML_CUDA_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
for (int id = 0; id < info.device_count; ++id) {
int device_vmm = 0;
#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
CUdevice device;
CU_CHECK(cuDeviceGet(&device, id));
CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
if (device_vmm) {
CUmemAllocationProp alloc_prop = {};
alloc_prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
alloc_prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
alloc_prop.location.id = id;
CU_CHECK(cuMemGetAllocationGranularity(&info.devices[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
}
#endif // !defined(GGML_USE_HIPBLAS)
info.devices[id].vmm = !!device_vmm;
cudaDeviceProp prop;
CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
GGML_CUDA_LOG_INFO(" Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
info.default_tensor_split[id] = total_vram;
total_vram += prop.totalGlobalMem;
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
info.devices[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD;
#else
info.devices[id].cc = 100*prop.major + 10*prop.minor;
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
info.devices[id].smpb = prop.sharedMemPerBlock;
info.devices[id].nsm = prop.multiProcessorCount;
}
for (int id = 0; id < info.device_count; ++id) {
info.default_tensor_split[id] /= total_vram;
}
// configure logging to stdout
// CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
return info;
}
const ggml_cuda_device_info & ggml_cuda_info() {
static ggml_cuda_device_info info = ggml_cuda_init();
return info;
}
// #define DEBUG_CUDA_MALLOC
// buffer pool for cuda (legacy)
struct ggml_cuda_pool_leg : public ggml_cuda_pool {
static const int MAX_BUFFERS = 256;
int device;
struct ggml_cuda_buffer {
void * ptr = nullptr;
size_t size = 0;
};
ggml_cuda_buffer buffer_pool[MAX_BUFFERS] = {};
size_t pool_size = 0;
explicit ggml_cuda_pool_leg(int device) :
device(device) {
}
~ggml_cuda_pool_leg() {
ggml_cuda_set_device(device);
for (int i = 0; i < MAX_BUFFERS; ++i) {
ggml_cuda_buffer & b = buffer_pool[i];
if (b.ptr != nullptr) {
CUDA_CHECK(cudaFree(b.ptr));
pool_size -= b.size;
}
}
GGML_ASSERT(pool_size == 0);
}
void * alloc(size_t size, size_t * actual_size) override {
#ifdef DEBUG_CUDA_MALLOC
int nnz = 0;
size_t max_size = 0;
#endif
size_t best_diff = 1ull << 36;
int ibest = -1;
for (int i = 0; i < MAX_BUFFERS; ++i) {
ggml_cuda_buffer& b = buffer_pool[i];
if (b.ptr != nullptr) {
#ifdef DEBUG_CUDA_MALLOC
++nnz;
if (b.size > max_size) max_size = b.size;
#endif
if (b.size >= size) {
size_t diff = b.size - size;
if (diff < best_diff) {
best_diff = diff;
ibest = i;
if (!best_diff) {
void * ptr = b.ptr;
*actual_size = b.size;
b.ptr = nullptr;
b.size = 0;
return ptr;
}
}
}
}
}
if (ibest >= 0) {
ggml_cuda_buffer& b = buffer_pool[ibest];
void * ptr = b.ptr;
*actual_size = b.size;
b.ptr = nullptr;
b.size = 0;
return ptr;
}
void * ptr;
size_t look_ahead_size = (size_t) (1.05 * size);
look_ahead_size = 256 * ((look_ahead_size + 255)/256);
ggml_cuda_set_device(device);
CUDA_CHECK(ggml_cuda_device_malloc(&ptr, look_ahead_size, device));
*actual_size = look_ahead_size;
pool_size += look_ahead_size;
#ifdef DEBUG_CUDA_MALLOC
GGML_CUDA_LOG_INFO("%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
(uint32_t)(max_size / 1024 / 1024), (uint32_t)(pool_size / 1024 / 1024), (uint32_t)(size / 1024 / 1024));
#endif
return ptr;
}
void free(void * ptr, size_t size) override {
for (int i = 0; i < MAX_BUFFERS; ++i) {
ggml_cuda_buffer& b = buffer_pool[i];
if (b.ptr == nullptr) {
b.ptr = ptr;
b.size = size;
return;
}
}
GGML_CUDA_LOG_WARN("Cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
ggml_cuda_set_device(device);
CUDA_CHECK(cudaFree(ptr));
pool_size -= size;
}
};
// pool with virtual memory
#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
int device;
CUdeviceptr pool_addr = 0;
size_t pool_used = 0;
size_t pool_size = 0;
size_t granularity;
explicit ggml_cuda_pool_vmm(int device) :
device(device),
granularity(ggml_cuda_info().devices[device].vmm_granularity) {
}
~ggml_cuda_pool_vmm() {
if (pool_addr != 0) {
CU_CHECK(cuMemUnmap(pool_addr, pool_size));
CU_CHECK(cuMemAddressFree(pool_addr, CUDA_POOL_VMM_MAX_SIZE));
}
}
void * alloc(size_t size, size_t * actual_size) override {
// round up the allocation size to the alignment to ensure that all allocations are aligned for all data types
const size_t alignment = 128;
size = alignment * ((size + alignment - 1) / alignment);
size_t avail = pool_size - pool_used;
if (size > avail) {
// round up to the next multiple of the granularity
size_t reserve_size = size - avail;
reserve_size = granularity * ((reserve_size + granularity - 1) / granularity);
GGML_ASSERT(pool_size + reserve_size <= CUDA_POOL_VMM_MAX_SIZE);
// allocate more physical memory
CUmemAllocationProp prop = {};
prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
prop.location.id = device;
CUmemGenericAllocationHandle handle;
CU_CHECK(cuMemCreate(&handle, reserve_size, &prop, 0));
// reserve virtual address space (if not already reserved)
if (pool_addr == 0) {
CU_CHECK(cuMemAddressReserve(&pool_addr, CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0));
}
// map at the end of the pool
CU_CHECK(cuMemMap(pool_addr + pool_size, reserve_size, 0, handle, 0));
// the memory allocation handle is no longer needed after mapping
CU_CHECK(cuMemRelease(handle));
// set access
CUmemAccessDesc access = {};
access.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
access.location.id = device;
access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
CU_CHECK(cuMemSetAccess(pool_addr + pool_size, reserve_size, &access, 1));
// add to the pool
pool_size += reserve_size;
//printf("cuda pool[%d]: size increased to %llu MB (reserved %llu MB)\n",
// device, (unsigned long long) (pool_size/1024/1024),
// (unsigned long long) (reserve_size/1024/1024));
}
GGML_ASSERT(pool_addr != 0);
void * ptr = (void *) (pool_addr + pool_used);
*actual_size = size;
pool_used += size;
#ifdef DEBUG_CUDA_MALLOC
printf("cuda pool[%d]: allocated %llu bytes at %llx\n", device, (unsigned long long) size, ptr);
#endif
return ptr;
}
void free(void * ptr, size_t size) override {
#ifdef DEBUG_CUDA_MALLOC
printf("cuda pool[%d]: freed %llu bytes at %llx\n", device, (unsigned long long) size, ptr);
#endif
pool_used -= size;
// all deallocations must be in reverse order of the allocations
GGML_ASSERT(ptr == (void *) (pool_addr + pool_used));
}
};
#endif // !defined(GGML_USE_HIPBLAS)
std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
if (ggml_cuda_info().devices[device].vmm) {
return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
}
#endif
return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
}
// cuda buffer
struct ggml_backend_cuda_buffer_context {
int device;
void * dev_ptr = nullptr;
std::string name;
ggml_backend_cuda_buffer_context(int device, void * dev_ptr) :
device(device), dev_ptr(dev_ptr),
name(GGML_CUDA_NAME + std::to_string(device)) {
}
~ggml_backend_cuda_buffer_context() {
CUDA_CHECK(cudaFree(dev_ptr));
}
};
GGML_CALL static const char * ggml_backend_cuda_buffer_get_name(ggml_backend_buffer_t buffer) {
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
return ctx->name.c_str();
}
GGML_CALL static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
return buffer->iface.get_name == ggml_backend_cuda_buffer_get_name;
}
GGML_CALL static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
delete ctx;
}
GGML_CALL static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) {
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
return ctx->dev_ptr;
}
GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
if (tensor->view_src != NULL) {
assert(tensor->view_src->buffer->buft == buffer->buft);
return;
}
if (ggml_is_quantized(tensor->type)) {
// initialize padding to 0 to avoid possible NaN values
size_t original_size = ggml_nbytes(tensor);
size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
if (padded_size > original_size && tensor->view_src == nullptr) {
ggml_cuda_set_device(ctx->device);
CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size));
}
}
}
GGML_CALL static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
ggml_cuda_set_device(ctx->device);
CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cudaStreamPerThread));
CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
}
GGML_CALL static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
ggml_cuda_set_device(ctx->device);
CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cudaStreamPerThread));
CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
}
GGML_CALL static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
if (ggml_backend_buffer_is_cuda(src->buffer)) {
ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context;
ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)dst->buffer->context;
if (src_ctx->device == dst_ctx->device) {
CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(src), cudaMemcpyDeviceToDevice, cudaStreamPerThread));
} else {
#ifdef GGML_CUDA_NO_PEER_COPY
return false;
#else
CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, dst_ctx->device, src->data, src_ctx->device, ggml_nbytes(src), cudaStreamPerThread));
#endif
}
CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
return true;
}
return false;
GGML_UNUSED(buffer);
}
GGML_CALL static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
ggml_cuda_set_device(ctx->device);
CUDA_CHECK(cudaDeviceSynchronize());
CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size));
CUDA_CHECK(cudaDeviceSynchronize());
}
static ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
/* .get_name = */ ggml_backend_cuda_buffer_get_name,
/* .free_buffer = */ ggml_backend_cuda_buffer_free_buffer,
/* .get_base = */ ggml_backend_cuda_buffer_get_base,
/* .init_tensor = */ ggml_backend_cuda_buffer_init_tensor,
/* .set_tensor = */ ggml_backend_cuda_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_cuda_buffer_get_tensor,
/* .cpy_tensor = */ ggml_backend_cuda_buffer_cpy_tensor,
/* .clear = */ ggml_backend_cuda_buffer_clear,
/* .reset = */ NULL,
};
// cuda buffer type
struct ggml_backend_cuda_buffer_type_context {
int device;
std::string name;
};
GGML_CALL static const char * ggml_backend_cuda_buffer_type_name(ggml_backend_buffer_type_t buft) {
ggml_backend_cuda_buffer_type_context * ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
return ctx->name.c_str();
}
GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
ggml_cuda_set_device(buft_ctx->device);
size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0
void * dev_ptr;
cudaError_t err = ggml_cuda_device_malloc(&dev_ptr, size, buft_ctx->device);
if (err != cudaSuccess) {
// clear the error
cudaGetLastError();
GGML_CUDA_LOG_ERROR("%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size / 1024.0 / 1024.0, buft_ctx->device, cudaGetErrorString(err));
return nullptr;
}
ggml_backend_cuda_buffer_context * ctx = new ggml_backend_cuda_buffer_context(buft_ctx->device, dev_ptr);
return ggml_backend_buffer_init(buft, ggml_backend_cuda_buffer_interface, ctx, size);
}
GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return 128;
GGML_UNUSED(buft);
}
GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
size_t size = ggml_nbytes(tensor);
int64_t ne0 = tensor->ne[0];
if (ggml_is_quantized(tensor->type)) {
if (ne0 % MATRIX_ROW_PADDING != 0) {
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
}
}
return size;
GGML_UNUSED(buft);
}
GGML_CALL static bool ggml_backend_cuda_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
if (!ggml_backend_is_cuda(backend)) {
return false;
}
ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
return buft_ctx->device == cuda_ctx->device;
}
static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
/* .get_name = */ ggml_backend_cuda_buffer_type_name,
/* .alloc_buffer = */ ggml_backend_cuda_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cuda_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_cuda_buffer_type_get_alloc_size,
/* .supports_backend = */ ggml_backend_cuda_buffer_type_supports_backend,
/* .is_host = */ NULL,
};
GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
if (device >= ggml_backend_cuda_get_device_count()) {
return nullptr;
}
static ggml_backend_buffer_type ggml_backend_cuda_buffer_types[GGML_CUDA_MAX_DEVICES];
static bool ggml_backend_cuda_buffer_type_initialized = false;
if (!ggml_backend_cuda_buffer_type_initialized) {
for (int i = 0; i < GGML_CUDA_MAX_DEVICES; i++) {
ggml_backend_cuda_buffer_types[i] = {
/* .iface = */ ggml_backend_cuda_buffer_type_interface,
/* .context = */ new ggml_backend_cuda_buffer_type_context{i, GGML_CUDA_NAME + std::to_string(i)},
};
}
ggml_backend_cuda_buffer_type_initialized = true;
}
return &ggml_backend_cuda_buffer_types[device];
}
// cuda split buffer
static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split) {
int64_t min_compute_capability = INT_MAX;
int64_t max_compute_capability = INT_MIN;
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
if (tensor_split[id] < (id + 1 < ggml_backend_cuda_get_device_count() ? tensor_split[id + 1] : 1.0f)) {
if (min_compute_capability > ggml_cuda_info().devices[id].cc) {
min_compute_capability = ggml_cuda_info().devices[id].cc;
}
if (max_compute_capability < ggml_cuda_info().devices[id].cc) {
max_compute_capability = ggml_cuda_info().devices[id].cc;
}
}
}
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
switch(type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
return max_compute_capability >= CC_RDNA2 ? 128 : 64;
case GGML_TYPE_F16:
case GGML_TYPE_F32:
return 1;
case GGML_TYPE_Q2_K:
return max_compute_capability >= CC_RDNA2 ? 128 : 32;
case GGML_TYPE_Q3_K:
return min_compute_capability < CC_RDNA2 ? 128 : 64;
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
return max_compute_capability >= CC_RDNA2 ? 128 : 64;
default:
GGML_ASSERT(false);
}
#else
switch(type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
return max_compute_capability >= CC_VOLTA ? 128 : 64;
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
return 64;
case GGML_TYPE_F16:
case GGML_TYPE_F32:
return 1;
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
return max_compute_capability >= CC_VOLTA ? 128 : 64;
case GGML_TYPE_Q6_K:
return 64;
default:
GGML_ASSERT(false);
}
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
}
static void get_row_split(int64_t * row_low, int64_t * row_high, const ggml_tensor * tensor, const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split, int id) {
const int64_t nrows = ggml_nrows(tensor);
const int64_t rounding = get_row_rounding(tensor->type, tensor_split);
*row_low = id == 0 ? 0 : nrows*tensor_split[id];
*row_low -= *row_low % rounding;
if (id == ggml_backend_cuda_get_device_count() - 1) {
*row_high = nrows;
} else {
*row_high = nrows*tensor_split[id + 1];
*row_high -= *row_high % rounding;
}
}
static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]);
}
struct ggml_backend_cuda_split_buffer_type_context {
std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split;
};
struct ggml_backend_cuda_split_buffer_context {
~ggml_backend_cuda_split_buffer_context() {
for (ggml_tensor_extra_gpu * extra : tensor_extras) {
for (int id = 0; id < GGML_CUDA_MAX_DEVICES; ++id) {
for (int64_t is = 0; is < GGML_CUDA_MAX_STREAMS; ++is) {
if (extra->events[id][is] != nullptr) {
CUDA_CHECK(cudaEventDestroy(extra->events[id][is]));
}
}
if (extra->data_device[id] != nullptr) {
CUDA_CHECK(cudaFree(extra->data_device[id]));
}
}
delete extra;
}
}
std::vector<ggml_tensor_extra_gpu *> tensor_extras;
};
GGML_CALL static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_t buffer) {
return GGML_CUDA_NAME "_Split";
GGML_UNUSED(buffer);
}
static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) {
return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name;
GGML_UNUSED(ggml_backend_buffer_is_cuda_split); // only used in debug builds currently, avoid unused function warning in release builds
}
GGML_CALL static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
delete ctx;
}
GGML_CALL static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
// the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced
return (void *)0x1000;
GGML_UNUSED(buffer);
}
GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
const int64_t ne0 = tensor->ne[0];
ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
ctx->tensor_extras.push_back(extra);
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
int64_t row_low, row_high;
get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id);
int64_t nrows_split = row_high - row_low;
if (nrows_split == 0) {
continue;
}
size_t size = ggml_nbytes_split(tensor, nrows_split);
const size_t original_size = size;
// pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
if (ne0 % MATRIX_ROW_PADDING != 0) {
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
}
// FIXME: do not crash if cudaMalloc fails
// currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first
ggml_cuda_set_device(id);
char * buf;
CUDA_CHECK(ggml_cuda_device_malloc((void**)&buf, size, id));
// set padding to 0 to avoid possible NaN values
if (size > original_size) {
CUDA_CHECK(cudaMemset(buf + original_size, 0, size - original_size));
}
extra->data_device[id] = buf;
for (int64_t is = 0; is < GGML_CUDA_MAX_STREAMS; ++is) {
CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id][is], cudaEventDisableTiming));
}
}
tensor->extra = extra;
}
GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
// split tensors must always be set in their entirety at once
GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor));
ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
const int64_t ne0 = tensor->ne[0];
const size_t nb1 = tensor->nb[1];
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra;
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
int64_t row_low, row_high;
get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id);
int64_t nrows_split = row_high - row_low;
if (nrows_split == 0) {
continue;
}
const size_t offset_split = row_low*nb1;
size_t size = ggml_nbytes_split(tensor, nrows_split);
const size_t original_size = size;
// pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
if (ne0 % MATRIX_ROW_PADDING != 0) {
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
}
const char * buf_host = (const char *)data + offset_split;
CUDA_CHECK(cudaMemcpyAsync(extra->data_device[id], buf_host, original_size, cudaMemcpyHostToDevice, cudaStreamPerThread));
}
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
}
}
GGML_CALL static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
// split tensors must always be set in their entirety at once
GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor));
ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
const int64_t ne0 = tensor->ne[0];
const size_t nb1 = tensor->nb[1];
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra;
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
int64_t row_low, row_high;
get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id);
int64_t nrows_split = row_high - row_low;
if (nrows_split == 0) {
continue;
}
const size_t offset_split = row_low*nb1;
size_t size = ggml_nbytes_split(tensor, nrows_split);
const size_t original_size = size;
// pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
if (ne0 % MATRIX_ROW_PADDING != 0) {
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
}
char * buf_host = (char *)data + offset_split;
CUDA_CHECK(cudaMemcpyAsync(buf_host, extra->data_device[id], original_size, cudaMemcpyDeviceToHost, cudaStreamPerThread));
}
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
}
}
GGML_CALL static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
GGML_UNUSED(buffer);
GGML_UNUSED(value);
}
static struct ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = {
/* .get_name = */ ggml_backend_cuda_split_buffer_get_name,
/* .free_buffer = */ ggml_backend_cuda_split_buffer_free_buffer,
/* .get_base = */ ggml_backend_cuda_split_buffer_get_base,
/* .init_tensor = */ ggml_backend_cuda_split_buffer_init_tensor,
/* .set_tensor = */ ggml_backend_cuda_split_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_cuda_split_buffer_get_tensor,
/* .cpy_tensor = */ NULL,
/* .clear = */ ggml_backend_cuda_split_buffer_clear,
/* .reset = */ NULL,
};
// cuda split buffer type
GGML_CALL static const char * ggml_backend_cuda_split_buffer_type_name(ggml_backend_buffer_type_t buft) {
return GGML_CUDA_NAME "_Split";
GGML_UNUSED(buft);
}
GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
// since we don't know the exact split after rounding, we cannot allocate the device buffers at this point
// instead, we allocate them for each tensor separately in init_tensor
// however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated,
// as returned by get_alloc_size. this limit is enforced during tensor allocation by ggml-alloc, so it must be correct.
ggml_backend_cuda_split_buffer_context * ctx = new ggml_backend_cuda_split_buffer_context();
return ggml_backend_buffer_init(buft, ggml_backend_cuda_split_buffer_interface, ctx, size);
}
GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return 128;
GGML_UNUSED(buft);
}
GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context;
size_t total_size = 0;
const int64_t ne0 = tensor->ne[0];
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
int64_t row_low, row_high;
get_row_split(&row_low, &row_high, tensor, ctx->tensor_split, id);
int64_t nrows_split = row_high - row_low;
if (nrows_split == 0) {
continue;
}
total_size += ggml_nbytes_split(tensor, nrows_split);
// pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
if (ne0 % MATRIX_ROW_PADDING != 0) {
total_size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
}
}
return total_size;
}
GGML_CALL static bool ggml_backend_cuda_split_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
return ggml_backend_is_cuda(backend);
GGML_UNUSED(buft);
}
GGML_CALL static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
return false;
GGML_UNUSED(buft);
}
static ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface = {
/* .get_name = */ ggml_backend_cuda_split_buffer_type_name,
/* .alloc_buffer = */ ggml_backend_cuda_split_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cuda_split_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_cuda_split_buffer_type_get_alloc_size,
/* .supports_backend = */ ggml_backend_cuda_split_buffer_type_supports_backend,
/* .is_host = */ ggml_backend_cuda_split_buffer_type_is_host,
};
GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split) {
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
static std::map<std::array<float, GGML_CUDA_MAX_DEVICES>, struct ggml_backend_buffer_type> buft_map;
std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split_arr = {};
bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + GGML_CUDA_MAX_DEVICES, [](float x) { return x == 0.0f; });
if (all_zero) {
tensor_split_arr = ggml_cuda_info().default_tensor_split;
} else {
float split_sum = 0.0f;
for (int i = 0; i < ggml_backend_cuda_get_device_count(); ++i) {
tensor_split_arr[i] = split_sum;
split_sum += tensor_split[i];
}
for (int i = 0; i < ggml_backend_cuda_get_device_count(); ++i) {
tensor_split_arr[i] /= split_sum;
}
}
auto it = buft_map.find(tensor_split_arr);
if (it != buft_map.end()) {
return &it->second;
}
struct ggml_backend_buffer_type buft {
/* .iface = */ ggml_backend_cuda_split_buffer_type_interface,
/* .context = */ new ggml_backend_cuda_split_buffer_type_context{tensor_split_arr},
};
auto result = buft_map.emplace(tensor_split_arr, buft);
return &result.first->second;
}
// host buffer type
GGML_CALL static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
return GGML_CUDA_NAME "_Host";
GGML_UNUSED(buft);
}
GGML_CALL static const char * ggml_backend_cuda_host_buffer_name(ggml_backend_buffer_t buffer) {
return GGML_CUDA_NAME "_Host";
GGML_UNUSED(buffer);
}
GGML_CALL static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
CUDA_CHECK(cudaFreeHost(buffer->context));
}
static void * ggml_cuda_host_malloc(size_t size) {
if (getenv("GGML_CUDA_NO_PINNED") != nullptr) {
return nullptr;
}
void * ptr = nullptr;
cudaError_t err = cudaMallocHost((void **) &ptr, size);
if (err != cudaSuccess) {
// clear the error
cudaGetLastError();
GGML_CUDA_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
size / 1024.0 / 1024.0, cudaGetErrorString(err));
return nullptr;
}
return ptr;
}
GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
void * ptr = ggml_cuda_host_malloc(size);
if (ptr == nullptr) {
// fallback to cpu buffer
return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.get_name = ggml_backend_cuda_host_buffer_name;
buffer->iface.free_buffer = ggml_backend_cuda_host_buffer_free_buffer;
return buffer;
}
GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_type_host = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cuda_host_buffer_type_name,
/* .alloc_buffer = */ ggml_backend_cuda_host_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
/* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
/* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host,
},
/* .context = */ nullptr,
};
return &ggml_backend_cuda_buffer_type_host;
}
//static bool ggml_backend_buffer_is_cuda_host(ggml_backend_buffer_t buffer) {
// return buffer->buft->iface.get_name == ggml_backend_cuda_host_buffer_type_name;
//}
/// kernels
typedef void (*ggml_cuda_op_mul_mat_t)(
ggml_backend_cuda_context & ctx,
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
const int64_t src1_padded_row_size, cudaStream_t stream);
#ifndef GGML_CUDA_PEER_MAX_BATCH_SIZE
#define GGML_CUDA_PEER_MAX_BATCH_SIZE 128
#endif // GGML_CUDA_PEER_MAX_BATCH_SIZE
#define MUL_MAT_SRC1_COL_STRIDE 128
static __global__ __launch_bounds__(1024) void mul_mat_p021_f16_f32(
const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y) {
const half * x = (const half *) vx;
const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
const int channel = blockDim.z*blockIdx.z + threadIdx.z;
const int channel_x = channel / (nchannels_y / nchannels_x);
const int nrows_y = ncols_x;
const int nrows_dst = nrows_x;
const int row_dst = row_x;
float tmp = 0.0f;
for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
const int col_x = col_x0 + threadIdx.x;
if (col_x >= ncols_x) {
break;
}
// x is transposed and permuted
const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x;
const float xi = __half2float(x[ix]);
const int row_y = col_x;
// y is not transposed but permuted
const int iy = channel*nrows_y + row_y;
tmp += xi * y[iy];
}
// dst is not transposed and not permuted
const int idst = channel*nrows_dst + row_dst;
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
if (threadIdx.x == 0) {
dst[idst] = tmp;
}
}
static __global__ __launch_bounds__(1024) void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
const int row_stride_x, const int channel_stride_x, const int channel_x_divisor) {
const half * x = (const half *) vx;
const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
const int channel = blockDim.z*blockIdx.z + threadIdx.z;
const int channel_x = channel / channel_x_divisor;
const int nrows_y = ncols_x;
const int nrows_dst = nrows_x;
const int row_dst = row_x;
const int idst = channel*nrows_dst + row_dst;
float tmp = 0.0f;
for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
const int col_x = col_x0 + threadIdx.x;
if (col_x >= ncols_x) {
break;
}
const int row_y = col_x;
const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
const int iy = channel*nrows_y + row_y;
const float xi = __half2float(x[ix]);
tmp += xi * y[iy];
}
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
if (threadIdx.x == 0) {
dst[idst] = tmp;
}
}
static void ggml_mul_mat_p021_f16_f32_cuda(
const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
const int nchannels_x, const int nchannels_y, cudaStream_t stream) {
const dim3 block_nums(1, nrows_x, nchannels_y);
const dim3 block_dims(WARP_SIZE, 1, 1);
mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y);
}
static void ggml_mul_mat_vec_nc_f16_f32_cuda(
const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int row_stride_x,
const int nchannels_x, const int nchannels_y, const int channel_stride_x, cudaStream_t stream) {
const dim3 block_nums(1, nrows_x, nchannels_y);
const dim3 block_dims(WARP_SIZE, 1, 1);
mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
(vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y/nchannels_x);
}
static cudaError_t ggml_cuda_cpy_tensor_2d(
void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
GGML_ASSERT(ggml_backend_buffer_is_cuda(src->buffer));
char * src_ptr = (char *) src->data;
char * dst_ptr = (char *) dst;
const int64_t ne0 = src->ne[0];
const int64_t nb0 = src->nb[0];
const int64_t nb1 = src->nb[1];
const int64_t nb2 = src->nb[2];
const int64_t nb3 = src->nb[3];
const enum ggml_type type = src->type;
const int64_t ts = ggml_type_size(type);
const int64_t bs = ggml_blck_size(type);
int64_t i1_diff = i1_high - i1_low;
const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3;
if (nb0 == ts && nb1 == ts*ne0/bs) {
return cudaMemcpyAsync(dst_ptr, x, i1_diff*nb1, cudaMemcpyDeviceToDevice, stream);
} else if (nb0 == ts) {
return cudaMemcpy2DAsync(dst_ptr, ts*ne0/bs, x, nb1, ts*ne0/bs, i1_diff, cudaMemcpyDeviceToDevice, stream);
} else {
for (int64_t i1 = 0; i1 < i1_diff; i1++) {
const void * rx = (const void *) ((const char *) x + i1*nb1);
void * rd = (void *) (dst_ptr + i1*ts*ne0/bs);
// pretend the row is a matrix with cols=1
cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, cudaMemcpyDeviceToDevice, stream);
if (r != cudaSuccess) {
return r;
}
}
return cudaSuccess;
}
}
static void ggml_cuda_op_mul_mat_cublas(
ggml_backend_cuda_context & ctx,
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
const int64_t src1_padded_row_size, cudaStream_t stream) {
GGML_ASSERT(src0_dd_i != nullptr);
GGML_ASSERT(src1_ddf_i != nullptr);
GGML_ASSERT(dst_dd_i != nullptr);
const int64_t ne00 = src0->ne[0];
const int64_t ne10 = src1->ne[0];
const int64_t ne0 = dst->ne[0];
const int64_t row_diff = row_high - row_low;
int id = ggml_cuda_get_device();
// the main device has a larger memory buffer to hold the results from all GPUs
// ldc == nrows of the matrix that cuBLAS writes into
int64_t ldc = id == ctx.device ? ne0 : row_diff;
const int compute_capability = ggml_cuda_info().devices[id].cc;
if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) {
// convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
if (src0->type != GGML_TYPE_F16) {
const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src0->type);
GGML_ASSERT(to_fp16_cuda != nullptr);
size_t ne = row_diff*ne00;
src0_as_f16.alloc(ne);
to_fp16_cuda(src0_dd_i, src0_as_f16.get(), ne, stream);
}
const half * src0_ptr = src0->type == GGML_TYPE_F16 ? (const half *) src0_dd_i : src0_as_f16.get();
ggml_cuda_pool_alloc<half> src1_as_f16(ctx.pool(id));
if (src1->type != GGML_TYPE_F16) {
const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
GGML_ASSERT(to_fp16_cuda != nullptr);
size_t ne = src1_ncols*ne10;
src1_as_f16.alloc(ne);
to_fp16_cuda(src1_ddf_i, src1_as_f16.get(), ne, stream);
}
const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16.get();
ggml_cuda_pool_alloc<half> dst_f16(ctx.pool(id), row_diff*src1_ncols);
const half alpha_f16 = 1.0f;
const half beta_f16 = 0.0f;
CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
CUBLAS_CHECK(
cublasGemmEx(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N,
row_diff, src1_ncols, ne10,
&alpha_f16, src0_ptr, CUDA_R_16F, ne00,
src1_ptr, CUDA_R_16F, ne10,
&beta_f16, dst_f16.get(), CUDA_R_16F, ldc,
CUBLAS_COMPUTE_16F,
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
to_fp32_cuda(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
} else {
ggml_cuda_pool_alloc<float> src0_ddq_as_f32(ctx.pool(id));
ggml_cuda_pool_alloc<float> src1_ddq_as_f32(ctx.pool(id));
if (src0->type != GGML_TYPE_F32) {
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type);
GGML_ASSERT(to_fp32_cuda != nullptr);
src0_ddq_as_f32.alloc(row_diff*ne00);
to_fp32_cuda(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream);
}
if (src1->type != GGML_TYPE_F32) {
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src1->type);
GGML_ASSERT(to_fp32_cuda != nullptr);
src1_ddq_as_f32.alloc(src1_ncols*ne10);
to_fp32_cuda(src1_ddf_i, src1_ddq_as_f32.get(), src1_ncols*ne10, stream);
}
const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();
const float alpha = 1.0f;
const float beta = 0.0f;
CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
CUBLAS_CHECK(
cublasSgemm(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N,
row_diff, src1_ncols, ne10,
&alpha, src0_ddf_i, ne00,
src1_ddf1_i, ne10,
&beta, dst_dd_i, ldc));
}
GGML_UNUSED(dst);
GGML_UNUSED(src1_ddq_i);
GGML_UNUSED(src1_padded_row_size);
}
static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
static bool peer_access_enabled = false;
const bool enable_peer_access = n_tokens <= GGML_CUDA_PEER_MAX_BATCH_SIZE;
if (peer_access_enabled == enable_peer_access) {
return;
}
#ifdef NDEBUG
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
ggml_cuda_set_device(id);
CUDA_CHECK(cudaDeviceSynchronize());
}
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
ggml_cuda_set_device(id);
for (int id_other = 0; id_other < ggml_backend_cuda_get_device_count(); ++id_other) {
if (id == id_other) {
continue;
}
if (id != main_device && id_other != main_device) {
continue;
}
int can_access_peer;
CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
if (can_access_peer) {
if (enable_peer_access) {
cudaError_t err = cudaDeviceEnablePeerAccess(id_other, 0);
if (err != cudaErrorPeerAccessAlreadyEnabled) {
CUDA_CHECK(err);
}
} else {
cudaError_t err = cudaDeviceDisablePeerAccess(id_other);
if (err != cudaErrorPeerAccessNotEnabled) {
CUDA_CHECK(err);
}
}
}
}
}
ggml_cuda_set_device(main_device);
#endif // NDEBUG
peer_access_enabled = enable_peer_access;
GGML_UNUSED(main_device);
}
static void ggml_cuda_op_mul_mat(
ggml_backend_cuda_context & ctx,
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_cuda_op_mul_mat_t op,
const bool convert_src1_to_q8_1) {
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int64_t ne12 = src1->ne[2];
const int64_t ne13 = src1->ne[3];
const int64_t nrows1 = ggml_nrows(src1);
GGML_ASSERT(ne03 == ne13);
const int64_t ne0 = dst->ne[0];
const int64_t ne1 = dst->ne[1];
const int64_t nb2 = dst->nb[2];
const int64_t nb3 = dst->nb[3];
GGML_ASSERT(ggml_backend_buffer_is_cuda(dst->buffer));
GGML_ASSERT(ggml_backend_buffer_is_cuda(src1->buffer));
ggml_backend_cuda_buffer_context * src1_ctx = (ggml_backend_cuda_buffer_context *) src1->buffer->context;
ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *) dst->buffer->context;
GGML_ASSERT(src1->type == GGML_TYPE_F32 || (src1->ne[2] == 1 && src1->ne[3] == 1));
GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0);
const int64_t i02_divisor = ne12 / ne02;
const size_t src0_ts = ggml_type_size(src0->type);
const size_t src0_bs = ggml_blck_size(src0->type);
const size_t q8_1_ts = sizeof(block_q8_1);
const size_t q8_1_bs = QK8_1;
const bool src0_is_contiguous = ggml_is_contiguous(src0);
const bool src1_is_contiguous = ggml_is_contiguous(src1);
const int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);
const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer);
GGML_ASSERT(!(split && ne02 > 1));
GGML_ASSERT(!(split && ne03 > 1));
GGML_ASSERT(!(split && ne02 < ne12));
ggml_tensor_extra_gpu * src0_extra = split ? (ggml_tensor_extra_gpu *) src0->extra : nullptr;
std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split;
if (split) {
ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
tensor_split = buft_ctx->tensor_split;
}
struct dev_data {
ggml_cuda_pool_alloc<char> src0_dd_alloc;
ggml_cuda_pool_alloc<float> src1_ddf_alloc;
ggml_cuda_pool_alloc<char> src1_ddq_alloc;
ggml_cuda_pool_alloc<float> dst_dd_alloc;
char * src0_dd = nullptr;
float * src1_ddf = nullptr; // float
char * src1_ddq = nullptr; // q8_1
float * dst_dd = nullptr;
int64_t row_low;
int64_t row_high;
};
dev_data dev[GGML_CUDA_MAX_DEVICES];
int used_devices = 0;
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
// by default, use all rows
dev[id].row_low = 0;
dev[id].row_high = ne01;
// for multi GPU, get the row boundaries from tensor split
// and round to mul_mat_q tile sizes
if (split) {
const int64_t rounding = get_row_rounding(src0->type, tensor_split);
if (id != 0) {
dev[id].row_low = ne01*tensor_split[id];
if (dev[id].row_low < ne01) {
dev[id].row_low -= dev[id].row_low % rounding;
}
}
if (id != ggml_backend_cuda_get_device_count() - 1) {
dev[id].row_high = ne01*tensor_split[id + 1];
if (dev[id].row_high < ne01) {
dev[id].row_high -= dev[id].row_high % rounding;
}
}
}
}
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
if ((!split && id != ctx.device) || dev[id].row_low == dev[id].row_high) {
continue;
}
used_devices++;
const bool src1_on_device = id == src1_ctx->device;
const bool dst_on_device = id == dst_ctx->device;
ggml_cuda_set_device(id);
cudaStream_t stream = ctx.stream(id, 0);
if (src0_is_contiguous) {
dev[id].src0_dd = split ? (char *) src0_extra->data_device[id] : (char *) src0->data;
} else {
dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), ggml_nbytes(src0));
}
if (src1_on_device && src1_is_contiguous) {
dev[id].src1_ddf = (float *) src1->data;
} else {
dev[id].src1_ddf = dev[id].src1_ddf_alloc.alloc(ctx.pool(id), ggml_nelements(src1));
}
if (convert_src1_to_q8_1) {
dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(ctx.pool(id), nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs);
if (src1_on_device && src1_is_contiguous) {
quantize_row_q8_1_cuda(dev[id].src1_ddf, dev[id].src1_ddq, ne10, nrows1, src1_padded_col_size, stream);
CUDA_CHECK(cudaGetLastError());
}
}
if (dst_on_device) {
dev[id].dst_dd = (float *) dst->data;
} else {
const size_t size_dst_ddf = split ? (dev[id].row_high - dev[id].row_low)*ne1 : ggml_nelements(dst);
dev[id].dst_dd = dev[id].dst_dd_alloc.alloc(ctx.pool(id), size_dst_ddf);
}
}
// if multiple devices are used they need to wait for the main device
// here an event is recorded that signals that the main device has finished calculating the input data
if (split && used_devices > 1) {
ggml_cuda_set_device(ctx.device);
CUDA_CHECK(cudaEventRecord(src0_extra->events[ctx.device][0], ctx.stream()));
}
const int64_t src1_col_stride = split && used_devices > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11;
for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) {
const int64_t is = split ? (src1_col_0/src1_col_stride) % GGML_CUDA_MAX_STREAMS : 0;
const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride;
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
if ((!split && id != ctx.device) || dev[id].row_low == dev[id].row_high) {
continue;
}
const bool src1_on_device = id == src1_ctx->device;
const bool dst_on_device = id == dst_ctx->device;
const int64_t row_diff = dev[id].row_high - dev[id].row_low;
ggml_cuda_set_device(id);
cudaStream_t stream = ctx.stream(id, is);
// wait for main GPU data if necessary
if (split && (id != ctx.device || is != 0)) {
CUDA_CHECK(cudaStreamWaitEvent(stream, src0_extra->events[ctx.device][0], 0));
}
for (int64_t i0 = 0; i0 < ne13*ne12; ++i0) {
const int64_t i03 = i0 / ne12;
const int64_t i02 = i0 % ne12;
const size_t src1_ddq_i_offset = (i0*ne11 + src1_col_0) * src1_padded_col_size*q8_1_ts/q8_1_bs;
// for split tensors the data begins at i0 == i0_offset_low
char * src0_dd_i = dev[id].src0_dd + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs;
float * src1_ddf_i = dev[id].src1_ddf + (i0*ne11 + src1_col_0) * ne10;
char * src1_ddq_i = dev[id].src1_ddq + src1_ddq_i_offset;
float * dst_dd_i = dev[id].dst_dd + (i0*ne1 + src1_col_0) * (dst_on_device ? ne0 : row_diff);
// the main device memory buffer can be on VRAM scratch, with space for all partial results
// in that case an offset on dst_ddf_i is needed
if (id == ctx.device) {
dst_dd_i += dev[id].row_low; // offset is 0 if no tensor split
}
// copy src0, src1 to device if necessary
if (src1_is_contiguous) {
if (id != ctx.device) {
if (convert_src1_to_q8_1) {
char * src1_ddq_i_source = dev[ctx.device].src1_ddq + src1_ddq_i_offset;
CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddq_i, id, src1_ddq_i_source, ctx.device,
src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, stream));
} else {
float * src1_ddf_i_source = (float *) src1->data;
src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10;
CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddf_i, id, src1_ddf_i_source, ctx.device,
src1_ncols*ne10*sizeof(float), stream));
}
}
} else if (src1_on_device && !src1_is_contiguous) {
CUDA_CHECK(ggml_cuda_cpy_tensor_2d(
src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
} else {
GGML_ASSERT(false);
}
if (convert_src1_to_q8_1 && !src1_is_contiguous) {
quantize_row_q8_1_cuda(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
CUDA_CHECK(cudaGetLastError());
}
if (src1_col_0 == 0 && !src0_is_contiguous && i02 % i02_divisor == 0) {
CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_dd_i, src0, i03, i02/i02_divisor, dev[id].row_low, dev[id].row_high, stream));
}
// do the computation
op(ctx, src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i,
dev[id].row_low, dev[id].row_high, src1_ncols, src1_padded_col_size, stream);
CUDA_CHECK(cudaGetLastError());
// copy dst to host or other device if necessary
if (!dst_on_device) {
void * dst_off_device = dst->data;
if (split) {
// src0 = weight matrix is saved as a transposed matrix for better memory layout.
// dst is NOT transposed.
// The outputs of matrix matrix multiplications can therefore NOT simply be concatenated for >1 GPU.
// Instead they need to be copied to the correct slice in ne0 = dst row index.
// If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results.
float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
dhf_dst_i += src1_col_0*ne0 + dev[id].row_low;
#if !defined(GGML_USE_HIPBLAS)
// cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
cudaMemcpy3DPeerParms p = {};
p.dstDevice = ctx.device;
p.dstPtr = make_cudaPitchedPtr(dhf_dst_i, ne0*sizeof(float), row_diff, src1_ncols);
p.srcDevice = id;
p.srcPtr = make_cudaPitchedPtr(dst_dd_i, row_diff*sizeof(float), row_diff, src1_ncols);
p.extent = make_cudaExtent(row_diff*sizeof(float), src1_ncols, 1);
CUDA_CHECK(cudaMemcpy3DPeerAsync(&p, stream));
#else
// HIP does not support cudaMemcpy3DPeerAsync or vmm pools
CUDA_CHECK(cudaMemcpy2DAsync(dhf_dst_i, ne0*sizeof(float),
dst_dd_i, row_diff*sizeof(float),
row_diff*sizeof(float), src1_ncols,
cudaMemcpyDeviceToDevice, stream));
#endif
} else {
float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
dhf_dst_i += src1_col_0*ne0;
CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_dd_i, src1_ncols*ne0*sizeof(float), cudaMemcpyDeviceToDevice, stream));
}
}
// add event for the main device to wait on until other device is done
if (split && (id != ctx.device || is != 0)) {
CUDA_CHECK(cudaEventRecord(src0_extra->events[id][is], stream));
}
}
}
}
// main device waits for all other devices to be finished
if (split && ggml_backend_cuda_get_device_count() > 1) {
int64_t is_max = (ne11 + MUL_MAT_SRC1_COL_STRIDE - 1) / MUL_MAT_SRC1_COL_STRIDE;
is_max = is_max <= GGML_CUDA_MAX_STREAMS ? is_max : GGML_CUDA_MAX_STREAMS;
ggml_cuda_set_device(ctx.device);
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
if (dev[id].row_low == dev[id].row_high) {
continue;
}
for (int64_t is = 0; is < is_max; ++is) {
CUDA_CHECK(cudaStreamWaitEvent(ctx.stream(), src0_extra->events[id][is], 0));
}
}
}
}
static void ggml_cuda_mul_mat_vec_p021(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne12 = src1->ne[2];
cudaStream_t main_stream = ctx.stream();
void * src0_ddq = src0->data;
float * src1_ddf = (float *) src1->data;
float * dst_ddf = (float *) dst->data;
ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
}
static void ggml_cuda_mul_mat_vec_nc(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(!ggml_is_transposed(src0));
GGML_ASSERT(!ggml_is_transposed(src1));
GGML_ASSERT(!ggml_is_permuted(src0));
GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t nb01 = src0->nb[1];
const int64_t nb02 = src0->nb[2];
const int64_t ne12 = src1->ne[2];
cudaStream_t main_stream = ctx.stream();
void * src0_ddq = src0->data;
float * src1_ddf = (float *) src1->data;
float * dst_ddf = (float *) dst->data;
const int64_t row_stride_x = nb01 / sizeof(half);
const int64_t channel_stride_x = nb02 / sizeof(half);
ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
}
static __global__ __launch_bounds__(1024) void k_compute_batched_ptrs(
const half * src0_as_f16, const half * src1_as_f16, char * dst,
const void ** ptrs_src, void ** ptrs_dst,
int64_t ne12, int64_t ne13,
int64_t ne23,
size_t nb02, size_t nb03,
size_t nb12, size_t nb13,
size_t nbd2, size_t nbd3,
int64_t r2, int64_t r3) {
int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
if (i13 >= ne13 || i12 >= ne12) {
return;
}
int64_t i03 = i13 / r3;
int64_t i02 = i12 / r2;
ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
ptrs_dst[0*ne23 + i12 + i13*ne12] = ( char *) dst + i12*nbd2 + i13*nbd3;
}
static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(!ggml_is_transposed(src0));
GGML_ASSERT(!ggml_is_transposed(src1));
GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_TENSOR_BINARY_OP_LOCALS
const int64_t ne_dst = ggml_nelements(dst);
cudaStream_t main_stream = ctx.stream();
CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
void * src0_ddq = src0->data;
half * src0_f16 = (half *) src0_ddq;
float * src1_ddf = (float *) src1->data;
float * dst_ddf = (float *) dst->data;
// convert src1 to fp16
ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
if (src1->type != GGML_TYPE_F16) {
const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
const int64_t ne_src1 = ggml_nelements(src1);
src1_f16_alloc.alloc(ne_src1);
GGML_ASSERT(to_fp16_cuda != nullptr);
to_fp16_cuda(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
}
half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get();
ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
char * dst_t;
cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
cudaDataType_t cu_data_type = CUDA_R_16F;
// dst strides
size_t nbd2 = dst->nb[2];
size_t nbd3 = dst->nb[3];
const half alpha_f16 = 1.0f;
const half beta_f16 = 0.0f;
const float alpha_f32 = 1.0f;
const float beta_f32 = 0.0f;
const void * alpha = &alpha_f16;
const void * beta = &beta_f16;
if (dst->op_params[0] == GGML_PREC_DEFAULT) {
dst_t = (char *) dst_f16.alloc(ne_dst);
nbd2 /= sizeof(float) / sizeof(half);
nbd3 /= sizeof(float) / sizeof(half);
} else {
dst_t = (char *) dst_ddf;
cu_compute_type = CUBLAS_COMPUTE_32F;
cu_data_type = CUDA_R_32F;
alpha = &alpha_f32;
beta = &beta_f32;
}
GGML_ASSERT(ne12 % ne02 == 0);
GGML_ASSERT(ne13 % ne03 == 0);
// broadcast factors
const int64_t r2 = ne12/ne02;
const int64_t r3 = ne13/ne03;
#if 0
// use cublasGemmEx
{
for (int i13 = 0; i13 < ne13; ++i13) {
for (int i12 = 0; i12 < ne12; ++i12) {
int i03 = i13 / r3;
int i02 = i12 / r2;
CUBLAS_CHECK(
cublasGemmEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
ne01, ne11, ne10,
alpha, (const char *) src0_as_f16 + i02*src0->nb[2] + i03*src0->nb[3] , CUDA_R_16F, nb01/sizeof(half),
(const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F, nb11/sizeof(float),
beta, ( char *) dst_t + i12*nbd2 + i13*nbd3, cu_data_type, ne01,
cu_compute_type,
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
}
}
}
#else
if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
// there is no broadcast and src0, src1 are contiguous across dims 2, 3
// use cublasGemmStridedBatchedEx
CUBLAS_CHECK(
cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
ne01, ne11, ne10,
alpha, (const char *) src0_f16, CUDA_R_16F, nb01/nb00, nb02/nb00, // strideA
(const char *) src1_f16, CUDA_R_16F, nb11/nb10, nb12/nb10, // strideB
beta, ( char *) dst_t, cu_data_type, ne01, nb2/nb0, // strideC
ne12*ne13,
cu_compute_type,
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
} else {
// use cublasGemmBatchedEx
const int ne23 = ne12*ne13;
ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
ggml_cuda_pool_alloc< void *> ptrs_dst(ctx.pool(), 1*ne23);
dim3 block_dims(ne13, ne12);
k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
src0_f16, src1_f16, dst_t,
ptrs_src.get(), ptrs_dst.get(),
ne12, ne13,
ne23,
nb02, nb03,
src1->type == GGML_TYPE_F16 ? nb12 : nb12/2,
src1->type == GGML_TYPE_F16 ? nb13 : nb13/2,
nbd2, nbd3,
r2, r3);
CUDA_CHECK(cudaGetLastError());
CUBLAS_CHECK(
cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
ne01, ne11, ne10,
alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F, nb01/nb00,
(const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F, nb11/nb10,
beta, ( void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01,
ne23,
cu_compute_type,
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
}
#endif
if (dst->op_params[0] == GGML_PREC_DEFAULT) {
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream);
}
}
static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer);
int64_t min_compute_capability = INT_MAX;
bool any_pascal_with_slow_fp16 = false;
if (split) {
ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
auto & tensor_split = buft_ctx->tensor_split;
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
// skip devices that are not going to do any work:
if (tensor_split[id] >= (id + 1 < ggml_backend_cuda_get_device_count() ? tensor_split[id + 1] : 1.0f)) {
continue;
}
if (min_compute_capability > ggml_cuda_info().devices[id].cc) {
min_compute_capability = ggml_cuda_info().devices[id].cc;
}
if (ggml_cuda_info().devices[id].cc == 610) {
any_pascal_with_slow_fp16 = true;
}
}
} else {
min_compute_capability = ggml_cuda_info().devices[ctx.device].cc;
any_pascal_with_slow_fp16 = ggml_cuda_info().devices[ctx.device].cc == 610;
}
// check data types and tensor shapes for custom matrix multiplication kernels:
bool use_dequantize_mul_mat_vec = (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
&& src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src1->ne[1] == 1;
bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
&& src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
bool use_mul_mat_q = ggml_cuda_supports_mmq(src0->type)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
const bool fp16_performance_good = min_compute_capability >= CC_RDNA1;
#ifdef CUDA_USE_TENSOR_CORES
use_mul_mat_q = use_mul_mat_q && min_compute_capability < CC_RDNA3;
#endif // CUDA_USE_TENSOR_CORES
#else
// fp16 performance is good on Volta or newer and on P100 (compute capability 6.0)
const bool fp16_performance_good = min_compute_capability >= CC_PASCAL && !any_pascal_with_slow_fp16;
// mmvq and mmq need the __dp4a instruction which on NVIDIA is only available for CC >= 6.1
use_mul_mat_vec_q = use_mul_mat_vec_q && min_compute_capability >= MIN_CC_DP4A;
use_mul_mat_q = use_mul_mat_q && min_compute_capability >= MIN_CC_DP4A;
#ifdef CUDA_USE_TENSOR_CORES
// when tensor cores are available, use them for large batch size
// ref: https://github.com/ggerganov/llama.cpp/pull/3776
use_mul_mat_q = use_mul_mat_q && (!fp16_performance_good || src1->ne[1] <= MMQ_MAX_BATCH_SIZE);
#endif // CUDA_USE_TENSOR_CORES
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
// if mmvq is available it's a better choice than dmmv:
#ifndef GGML_CUDA_FORCE_DMMV
use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
#endif // GGML_CUDA_FORCE_DMMV
// debug helpers
//printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
//printf(" %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
//printf("src1: %8d %8d %8d %8d\n", src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]);
//printf(" %8d %8d %8d %8d\n", src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]);
//printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
//printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
if (!split && !fp16_performance_good && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
// KQ single-batch
ggml_cuda_mul_mat_vec_p021(ctx, src0, src1, dst);
} else if (!split && !fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
// KQV single-batch
ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst);
} else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || fp16_performance_good) && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
// KQ + KQV multi-batch
ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
} else if (use_dequantize_mul_mat_vec) {
ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
} else if (use_mul_mat_vec_q) {
ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
} else if (use_mul_mat_q) {
ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
} else {
ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
}
}
struct mmid_row_mapping {
int32_t i1;
int32_t i2;
};
static __global__ __launch_bounds__(1024) void k_copy_src1_to_contiguous(const char * __restrict__ src1_original, char * __restrict__ src1_contiguous,
int * __restrict__ cur_src1_row, mmid_row_mapping * __restrict__ row_mapping,
const char * __restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0,
int64_t ne11, int64_t ne10,
size_t nb11, size_t nb12) {
int32_t iid1 = blockIdx.x;
int32_t id = blockIdx.y;
const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0);
if (row_id_i != i02) {
return;
}
const int64_t i11 = id % ne11;
const int64_t i12 = iid1;
__shared__ int src1_row;
if (threadIdx.x == 0) {
src1_row = atomicAdd(cur_src1_row, 1);
row_mapping[src1_row] = {id, iid1};
}
__syncthreads();
const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12);
float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11);
for (int i = threadIdx.x; i < ne10; i += blockDim.x) {
src1_row_contiguous[i] = src1_row_original[i];
}
}
static __global__ __launch_bounds__(1024) void k_copy_dst_from_contiguous(char * __restrict__ dst_original, const char * __restrict__ dst_contiguous,
const mmid_row_mapping * __restrict__ row_mapping,
int64_t ne0,
size_t nb1, size_t nb2) {
int32_t i = blockIdx.x;
const int32_t i1 = row_mapping[i].i1;
const int32_t i2 = row_mapping[i].i2;
const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1);
float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2);
for (int j = threadIdx.x; j < ne0; j += blockDim.x) {
dst_row_original[j] = dst_row_contiguous[j];
}
}
static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * ids = dst->src[2];
GGML_TENSOR_BINARY_OP_LOCALS
GGML_ASSERT(!ggml_backend_buffer_is_cuda_split(src0->buffer) && "mul_mat_id does not support split buffers");
cudaStream_t stream = ctx.stream();
const int64_t n_as = ne02;
const int64_t n_ids = ids->ne[0];
std::vector<char> ids_host(ggml_nbytes(ids));
const char * ids_dev = (const char *) ids->data;
CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
CUDA_CHECK(cudaStreamSynchronize(stream));
ggml_tensor src0_row = *src0;
ggml_tensor src1_row = *src1;
ggml_tensor dst_row = *dst;
char * src0_original = (char *) src0->data;
char * src1_original = (char *) src1->data;
char * dst_original = (char *) dst->data;
src0_row.ne[2] = 1;
src0_row.ne[3] = 1;
src0_row.nb[3] = nb02;
src1_row.ne[1] = 1;
src1_row.ne[2] = 1;
src1_row.ne[3] = 1;
src1_row.nb[2] = nb11;
src1_row.nb[3] = nb11;
dst_row.ne[1] = 1;
dst_row.ne[2] = 1;
dst_row.ne[3] = 1;
dst_row.nb[2] = nb1;
dst_row.nb[3] = nb1;
if (ne12 == 1) {
for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
for (int64_t id = 0; id < n_ids; id++) {
const int32_t i02 = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
GGML_ASSERT(i02 >= 0 && i02 < n_as);
const int64_t i11 = id % ne11;
const int64_t i12 = iid1;
const int64_t i1 = id;
const int64_t i2 = i12;
src0_row.data = src0_original + i02*nb02;
src1_row.data = src1_original + i11*nb11 + i12*nb12;
dst_row.data = dst_original + i1*nb1 + i2*nb2;
ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
}
}
} else {
ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
ggml_cuda_pool_alloc<char> dst_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst));
src1_row.data = src1_contiguous.get();
dst_row.data = dst_contiguous.get();
for (int64_t i02 = 0; i02 < n_as; i02++) {
int64_t num_src1_rows = 0;
for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
for (int64_t id = 0; id < n_ids; id++) {
const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
if (row_id_i != i02) {
continue;
}
num_src1_rows++;
}
}
if (num_src1_rows == 0) {
continue;
}
ggml_cuda_pool_alloc<int> dev_cur_src1_row(ctx.pool(), 1);
ggml_cuda_pool_alloc<mmid_row_mapping> dev_row_mapping(ctx.pool(), num_src1_rows);
CUDA_CHECK(cudaMemsetAsync(dev_cur_src1_row.get(), 0, sizeof(int), stream));
{
dim3 block_dims(std::min((unsigned int)ne10, 768u));
dim3 grid_dims(ids->ne[1], n_ids);
k_copy_src1_to_contiguous<<<grid_dims, block_dims, 0, stream>>>(
src1_original, src1_contiguous.get(),
dev_cur_src1_row.get(), dev_row_mapping.get(),
ids_dev, i02, ids->nb[1], ids->nb[0],
ne11, ne10,
nb11, nb12);
CUDA_CHECK(cudaGetLastError());
}
src0_row.data = src0_original + i02*nb02;
GGML_ASSERT(nb11 == sizeof(float)*ne10);
GGML_ASSERT(nb1 == sizeof(float)*ne0);
src1_row.ne[1] = num_src1_rows;
src1_row.nb[1] = nb11;
src1_row.nb[2] = num_src1_rows*nb11;
src1_row.nb[3] = num_src1_rows*nb11;
dst_row.ne[1] = num_src1_rows;
dst_row.nb[1] = nb1;
dst_row.nb[2] = num_src1_rows*nb1;
dst_row.nb[3] = num_src1_rows*nb1;
ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
{
dim3 block_dims(std::min((unsigned int)ne0, 768u));
dim3 grid_dims(num_src1_rows);
k_copy_dst_from_contiguous<<<grid_dims, block_dims, 0, stream>>>(
dst_original, dst_contiguous.get(),
dev_row_mapping.get(),
ne0,
nb1, nb2);
CUDA_CHECK(cudaGetLastError());
}
}
}
}
static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) {
// why is this here instead of mul_mat?
if (dst->src[0] != nullptr && ggml_backend_buffer_is_cuda_split(dst->src[0]->buffer)) {
ggml_cuda_set_peer_access(dst->src[1]->ne[1], ctx.device);
}
switch (dst->op) {
case GGML_OP_REPEAT:
ggml_cuda_op_repeat(ctx, dst);
break;
case GGML_OP_GET_ROWS:
ggml_cuda_op_get_rows(ctx, dst);
break;
case GGML_OP_DUP:
ggml_cuda_dup(ctx, dst);
break;
case GGML_OP_CPY:
ggml_cuda_cpy(ctx, dst->src[0], dst->src[1]);
break;
case GGML_OP_CONT:
ggml_cuda_dup(ctx, dst);
break;
case GGML_OP_ADD:
ggml_cuda_op_add(ctx, dst);
break;
case GGML_OP_ACC:
ggml_cuda_op_acc(ctx, dst);
break;
case GGML_OP_MUL:
ggml_cuda_op_mul(ctx, dst);
break;
case GGML_OP_DIV:
ggml_cuda_op_div(ctx, dst);
break;
case GGML_OP_UNARY:
switch (ggml_get_unary_op(dst)) {
case GGML_UNARY_OP_GELU:
ggml_cuda_op_gelu(ctx, dst);
break;
case GGML_UNARY_OP_SILU:
ggml_cuda_op_silu(ctx, dst);
break;
case GGML_UNARY_OP_GELU_QUICK:
ggml_cuda_op_gelu_quick(ctx, dst);
break;
case GGML_UNARY_OP_TANH:
ggml_cuda_op_tanh(ctx, dst);
break;
case GGML_UNARY_OP_RELU:
ggml_cuda_op_relu(ctx, dst);
break;
case GGML_UNARY_OP_SIGMOID:
ggml_cuda_op_sigmoid(ctx, dst);
break;
case GGML_UNARY_OP_HARDSIGMOID:
ggml_cuda_op_hardsigmoid(ctx, dst);
break;
case GGML_UNARY_OP_HARDSWISH:
ggml_cuda_op_hardswish(ctx, dst);
break;
default:
return false;
}
break;
case GGML_OP_NORM:
ggml_cuda_op_norm(ctx, dst);
break;
case GGML_OP_GROUP_NORM:
ggml_cuda_op_group_norm(ctx, dst);
break;
case GGML_OP_CONCAT:
ggml_cuda_op_concat(ctx, dst);
break;
case GGML_OP_UPSCALE:
ggml_cuda_op_upscale(ctx, dst);
break;
case GGML_OP_PAD:
ggml_cuda_op_pad(ctx, dst);
break;
case GGML_OP_ARANGE:
ggml_cuda_op_arange(ctx, dst);
break;
case GGML_OP_TIMESTEP_EMBEDDING:
ggml_cuda_op_timestep_embedding(ctx, dst);
break;
case GGML_OP_LEAKY_RELU:
ggml_cuda_op_leaky_relu(ctx, dst);
break;
case GGML_OP_RMS_NORM:
ggml_cuda_op_rms_norm(ctx, dst);
break;
case GGML_OP_MUL_MAT:
if (dst->src[0]->ne[3] != dst->src[1]->ne[3]) {
GGML_CUDA_LOG_ERROR("%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]);
return false;
} else {
ggml_cuda_mul_mat(ctx, dst->src[0], dst->src[1], dst);
}
break;
case GGML_OP_MUL_MAT_ID:
ggml_cuda_mul_mat_id(ctx, dst);
break;
case GGML_OP_SCALE:
ggml_cuda_op_scale(ctx, dst);
break;
case GGML_OP_SQR:
ggml_cuda_op_sqr(ctx, dst);
break;
case GGML_OP_CLAMP:
ggml_cuda_op_clamp(ctx, dst);
break;
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_PERMUTE:
case GGML_OP_TRANSPOSE:
break;
case GGML_OP_DIAG_MASK_INF:
ggml_cuda_op_diag_mask_inf(ctx, dst);
break;
case GGML_OP_SOFT_MAX:
ggml_cuda_op_soft_max(ctx, dst);
break;
case GGML_OP_ROPE:
ggml_cuda_op_rope(ctx, dst);
break;
case GGML_OP_IM2COL:
ggml_cuda_op_im2col(ctx, dst);
break;
case GGML_OP_POOL_2D:
ggml_cuda_op_pool2d(ctx, dst);
break;
case GGML_OP_SUM_ROWS:
ggml_cuda_op_sum_rows(ctx, dst);
break;
case GGML_OP_ARGSORT:
ggml_cuda_op_argsort(ctx, dst);
break;
case GGML_OP_FLASH_ATTN_EXT:
ggml_cuda_flash_attn_ext(ctx, dst);
break;
default:
return false;
}
cudaError_t err = cudaGetLastError();
if (err != cudaSuccess) {
GGML_CUDA_LOG_ERROR("%s: %s failed\n", __func__, ggml_op_desc(dst));
CUDA_CHECK(err);
}
return true;
}
////////////////////////////////////////////////////////////////////////////////
// backend
GGML_CALL static const char * ggml_backend_cuda_name(ggml_backend_t backend) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
return cuda_ctx->name.c_str();
}
GGML_CALL static void ggml_backend_cuda_free(ggml_backend_t backend) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
delete cuda_ctx;
delete backend;
}
GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer_type(ggml_backend_t backend) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
return ggml_backend_cuda_buffer_type(cuda_ctx->device);
}
GGML_CALL static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cuda_ctx->stream()));
}
GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cuda_ctx->stream()));
}
GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
GGML_ASSERT(ggml_backend_is_cuda(backend_src) || ggml_backend_is_cuda(backend_dst));
ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
if (!ggml_backend_buffer_is_cuda(src->buffer)) {
return false;
}
if (!ggml_backend_buffer_is_cuda(dst->buffer)) {
return false;
}
// device -> device
ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;
if (backend_src != backend_dst) {
ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
GGML_ASSERT(cuda_ctx_src->device == buf_ctx_src->device);
GGML_ASSERT(cuda_ctx_dst->device == buf_ctx_dst->device);
// copy on src stream
if (cuda_ctx_src->device == cuda_ctx_dst->device) {
CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
} else {
#ifdef GGML_CUDA_NO_PEER_COPY
return false;
#else
CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, cuda_ctx_dst->device, src->data, cuda_ctx_src->device, ggml_nbytes(dst), cuda_ctx_src->stream()));
#endif
}
// record event on src stream
if (!cuda_ctx_src->copy_event) {
ggml_cuda_set_device(cuda_ctx_src->device);
CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming));
}
CUDA_CHECK(cudaEventRecord(cuda_ctx_src->copy_event, cuda_ctx_src->stream()));
// wait on dst stream for the copy to complete
CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_dst->stream(), cuda_ctx_src->copy_event, 0));
} else {
// src and dst are on the same backend
CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
}
return true;
}
GGML_CALL static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
CUDA_CHECK(cudaStreamSynchronize(cuda_ctx->stream()));
GGML_UNUSED(backend);
}
static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
graph_node_properties->node_address = node->data;
graph_node_properties->node_op = node->op;
for (int i = 0; i < GGML_MAX_DIMS; i++) {
graph_node_properties->ne[i] = node->ne[i];
graph_node_properties->nb[i] = node->nb[i];
}
for (int i = 0; i < GGML_MAX_SRC; i++) {
graph_node_properties->src_address[i] = node->src[i] ? node->src[i]->data : nullptr;
}
}
static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
if (node->data != graph_node_properties->node_address &&
node->op != GGML_OP_CPY &&
node->op != GGML_OP_VIEW) {
return false;
}
if (node->op != graph_node_properties->node_op) {
return false;
}
for (int i = 0; i < GGML_MAX_DIMS; i++) {
if (node->ne[i] != graph_node_properties->ne[i]) {
return false;
}
if (node->nb[i] != graph_node_properties->nb[i]) {
return false;
}
}
for (int i = 0; i < GGML_MAX_SRC; i++) {
if (node->src[i] &&
node->src[i]->data != graph_node_properties->src_address[i] &&
node->op != GGML_OP_CPY &&
node->op != GGML_OP_VIEW
) {
return false;
}
}
return true;
}
GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
ggml_cuda_set_device(cuda_ctx->device);
#ifdef USE_CUDA_GRAPH
static const bool disable_cuda_graphs_due_to_env = (getenv("GGML_CUDA_DISABLE_GRAPHS") != nullptr);
// Objects required for CUDA Graph
if (cuda_ctx->cuda_graph == nullptr) {
cuda_ctx->cuda_graph.reset(new ggml_cuda_graph());
}
bool use_cuda_graph = true;
bool cuda_graph_update_required = false;
// vector of pointers to CUDA cpy kernels, which are required to identify
// kernel parameters which need updated in the graph for each token
std::vector<void *> ggml_cuda_cpy_fn_ptrs;
if (cuda_ctx->cuda_graph->graph == nullptr) {
if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
#ifndef NDEBUG
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
#endif
}
}
// Disable CUDA graphs in presence of env var, old GPU, use-case which is changing too rapidly,
// or previous graph capture failure.
// Also disable for multi-gpu for now. TO DO investigate
if (disable_cuda_graphs_due_to_env
|| cuda_ctx->cuda_graph->disable_due_to_gpu_arch
|| cuda_ctx->cuda_graph->disable_due_to_too_many_updates
|| cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture) {
use_cuda_graph = false;
}
if (use_cuda_graph) {
if (cuda_ctx->cuda_graph->instance == nullptr) {
cuda_graph_update_required = true;
}
// Check if the graph size has changed
if (cuda_ctx->cuda_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
cuda_graph_update_required = true;
cuda_ctx->cuda_graph->ggml_graph_properties.resize(cgraph->n_nodes);
}
// Loop over nodes in GGML graph to determine if CUDA graph update is required
// and store properties to allow this comparison for the next token
for (int i = 0; i < cgraph->n_nodes; i++) {
bool has_matching_properties = true;
if (!cuda_graph_update_required) {
has_matching_properties = ggml_graph_node_has_matching_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
}
if (!has_matching_properties) {
cuda_graph_update_required = true;
}
set_ggml_graph_node_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
}
// Loop over nodes in GGML graph to obtain info needed for CUDA graph
cuda_ctx->cuda_graph->updated_kernel_arg.clear();
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
if (node->src[0] && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
#ifndef NDEBUG
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to split buffer\n", __func__);
#endif
}
if (node->op == GGML_OP_MUL_MAT_ID) {
use_cuda_graph = false; // This node type is not supported by CUDA graph capture
#ifndef NDEBUG
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
#endif
}
if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {
// disable CUDA graphs for batch size > 1 for now.
// Changes in batch size or context size can cause changes to the grid size of some kernels.
use_cuda_graph = false;
#ifndef NDEBUG
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
#endif
}
if (node->op == GGML_OP_CPY) {
// store the copy op parameter which changes with each token.
cuda_ctx->cuda_graph->updated_kernel_arg.push_back((char **) &(node->src[1]->data));
// store a pointer to each copy op CUDA kernel to identify it later
void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) {
ggml_cuda_cpy_fn_ptrs.push_back(ptr);
}
}
if (!use_cuda_graph) {
break;
}
}
// Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
if (use_cuda_graph && cuda_graph_update_required) {
cuda_ctx->cuda_graph->number_consecutive_updates++;
} else {
cuda_ctx->cuda_graph->number_consecutive_updates = 0;
}
if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) {
cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true;
#ifndef NDEBUG
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
#endif
}
}
if (use_cuda_graph && cuda_graph_update_required) { // Start CUDA graph capture
CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
}
#else
bool use_cuda_graph = false;
bool cuda_graph_update_required = false;
#endif // USE_CUDA_GRAPH
bool graph_evaluated_or_captured = false;
while (!graph_evaluated_or_captured) {
// Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
// With the use of CUDA graphs, the execution will be performed by the graph launch.
if (!use_cuda_graph || cuda_graph_update_required) {
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
continue;
}
#ifndef NDEBUG
assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
for (int j = 0; j < GGML_MAX_SRC; j++) {
if (node->src[j] != nullptr) {
assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer));
}
}
#endif
bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
if (!ok) {
GGML_CUDA_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
}
GGML_ASSERT(ok);
}
}
#ifdef USE_CUDA_GRAPH
if (use_cuda_graph && cuda_graph_update_required) { // End CUDA graph capture
if (cuda_ctx->cuda_graph->graph != nullptr) {
CUDA_CHECK(cudaGraphDestroy(cuda_ctx->cuda_graph->graph));
cuda_ctx->cuda_graph->graph = nullptr;
}
CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
#if 0
if (disable_cuda_graphs_due_to_failed_capture) {
use_cuda_graph = false;
cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture = true;
#ifndef NDEBUG
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to failed graph capture\n", __func__);
#endif
} else {
graph_evaluated_or_captured = true; // CUDA graph has been captured
}
#endif
graph_evaluated_or_captured = true; // CUDA graph has been captured
} else {
graph_evaluated_or_captured = true; // ggml graph has been directly evaluated
}
}
if (use_cuda_graph) {
if (cuda_ctx->cuda_graph->instance == nullptr) { // Create executable graph from captured graph.
CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
}
// Perform update to graph (if required for this token), and change copy parameter (required for every token)
if (cuda_graph_update_required) {
// Extract nodes from graph
if (cuda_ctx->cuda_graph->num_nodes == 0) {
// First call with null argument gets number of nodes in graph
CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
}
// Subsequent call with non-null argument gets nodes
cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes);
cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes);
if (cuda_ctx->cuda_graph->num_nodes > 0) {
CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, cuda_ctx->cuda_graph->nodes.data(), &cuda_ctx->cuda_graph->num_nodes));
// Loop over nodes, and extract kernel parameters from each node
for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
cudaGraphNodeType node_type;
CUDA_CHECK(cudaGraphNodeGetType(cuda_ctx->cuda_graph->nodes[i], &node_type));
if (node_type == cudaGraphNodeTypeKernel) {
cudaError_t stat = cudaGraphKernelNodeGetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]); // Get params using runtime
if (stat == cudaErrorInvalidDeviceFunction) {
// Fails due to incorrect handling by CUDA runtime of CUDA BLAS node.
// We don't need to update blas nodes, so clear error and move on.
cudaGetLastError();
} else {
GGML_ASSERT(stat == cudaSuccess);
}
}
}
}
}
// One of the arguments to the copy kernel is updated for each token, hence we need to
// replace that argument with the updated value in the CUDA graph
if (!cuda_graph_update_required) { // on update steps, the live parameters will already be captured
int k = 0;
for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) {
char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++);
cuda_ctx->cuda_graph->params[i].kernelParams[1] = updated_kernel_arg_ptr;
CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]));
}
}
}
// Update graph executable
cudaGraphExecUpdateResultInfo result_info;
cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
if (stat == cudaErrorGraphExecUpdateFailure) {
#ifndef NDEBUG
GGML_CUDA_LOG_ERROR("%s: CUDA graph update failed\n", __func__);
#endif
// The pre-existing graph exec cannot be updated due to violated constraints
// so instead clear error and re-instantiate
cudaGetLastError();
CUDA_CHECK(cudaGraphExecDestroy(cuda_ctx->cuda_graph->instance));
cuda_ctx->cuda_graph->instance = nullptr;
CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
} else {
GGML_ASSERT(stat == cudaSuccess);
}
// Launch graph
CUDA_CHECK(cudaGraphLaunch(cuda_ctx->cuda_graph->instance, cuda_ctx->stream()));
#else
graph_evaluated_or_captured = true;
#endif // USE_CUDA_GRAPH
}
return GGML_STATUS_SUCCESS;
}
GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
switch (op->op) {
case GGML_OP_UNARY:
switch (ggml_get_unary_op(op)) {
case GGML_UNARY_OP_GELU:
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_RELU:
case GGML_UNARY_OP_SIGMOID:
case GGML_UNARY_OP_HARDSIGMOID:
case GGML_UNARY_OP_HARDSWISH:
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_TANH:
return true;
default:
return false;
}
break;
case GGML_OP_MUL_MAT:
case GGML_OP_MUL_MAT_ID:
{
struct ggml_tensor * a;
struct ggml_tensor * b;
if (op->op == GGML_OP_MUL_MAT) {
a = op->src[0];
b = op->src[1];
} else {
a = op->src[2];
b = op->src[1];
}
if (a->ne[3] != b->ne[3]) {
return false;
}
ggml_type a_type = a->type;
if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS ||
a_type == GGML_TYPE_IQ1_S || a_type == GGML_TYPE_IQ4_NL || a_type == GGML_TYPE_IQ3_S ||
a_type == GGML_TYPE_IQ1_M || a_type == GGML_TYPE_IQ2_S || a_type == GGML_TYPE_IQ4_XS) {
if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
return false;
}
}
return true;
} break;
case GGML_OP_GET_ROWS:
{
switch (op->src[0]->type) {
case GGML_TYPE_F16:
case GGML_TYPE_F32:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
return true;
default:
return false;
}
} break;
case GGML_OP_CPY:
{
ggml_type src0_type = op->src[0]->type;
ggml_type src1_type = op->src[1]->type;
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
return true;
}
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
return true;
}
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
return true;
}
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_0) {
return true;
}
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_1) {
return true;
}
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_0) {
return true;
}
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_1) {
return true;
}
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) {
return true;
}
if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
return true;
}
if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
return true;
}
return false;
} break;
case GGML_OP_DUP:
case GGML_OP_REPEAT:
case GGML_OP_CONCAT:
{
ggml_type src0_type = op->src[0]->type;
return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
} break;
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_PERMUTE:
case GGML_OP_TRANSPOSE:
case GGML_OP_NORM:
case GGML_OP_ADD:
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_RMS_NORM:
case GGML_OP_SCALE:
case GGML_OP_SQR:
case GGML_OP_CLAMP:
case GGML_OP_CONT:
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_SOFT_MAX:
return true;
case GGML_OP_ROPE:
return ggml_is_contiguous(op->src[0]);
case GGML_OP_IM2COL:
case GGML_OP_POOL_2D:
case GGML_OP_SUM_ROWS:
case GGML_OP_ARGSORT:
case GGML_OP_ACC:
case GGML_OP_GROUP_NORM:
case GGML_OP_UPSCALE:
case GGML_OP_PAD:
case GGML_OP_ARANGE:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_LEAKY_RELU:
return true;
case GGML_OP_FLASH_ATTN_EXT:
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
#else
if (op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128) {
return true;
}
return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA;
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
default:
return false;
}
GGML_UNUSED(backend);
}
GGML_CALL static bool ggml_backend_cuda_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
const int min_batch_size = 32;
return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
(op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
GGML_UNUSED(backend);
}
static ggml_backend_event_t ggml_backend_cuda_event_new(ggml_backend_t backend) {
#ifdef GGML_CUDA_NO_PEER_COPY
return nullptr;
#else
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
ggml_cuda_set_device(cuda_ctx->device);
cudaEvent_t event;
CUDA_CHECK(cudaEventCreateWithFlags(&event, cudaEventDisableTiming));
return new ggml_backend_event {
/* .backend = */ backend,
/* .context = */ event,
};
#endif
}
static void ggml_backend_cuda_event_free(ggml_backend_event_t event) {
CUDA_CHECK(cudaEventDestroy((cudaEvent_t)event->context));
delete event;
}
static void ggml_backend_cuda_event_record(ggml_backend_event_t event) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)event->backend->context;
CUDA_CHECK(cudaEventRecord((cudaEvent_t)event->context, cuda_ctx->stream()));
}
static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
if (ggml_backend_is_cuda(event->backend)) {
CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx->stream(), (cudaEvent_t)event->context, 0));
} else {
#if 0
// untested
auto wait_fn = [](void * user_data) {
ggml_backend_event_t event = (ggml_backend_event_t)user_data;
ggml_backend_event_synchronize(event);
};
CUDA_CHECK(cudaLaunchHostFunc(cuda_ctx->stream(), wait_fn, event));
#endif
GGML_ASSERT(false);
}
}
static void ggml_backend_cuda_event_synchronize(ggml_backend_event_t event) {
CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
}
static ggml_backend_i ggml_backend_cuda_interface = {
/* .get_name = */ ggml_backend_cuda_name,
/* .free = */ ggml_backend_cuda_free,
/* .get_default_buffer_type = */ ggml_backend_cuda_get_default_buffer_type,
/* .set_tensor_async = */ ggml_backend_cuda_set_tensor_async,
/* .get_tensor_async = */ ggml_backend_cuda_get_tensor_async,
/* .cpy_tensor_async = */ ggml_backend_cuda_cpy_tensor_async,
/* .synchronize = */ ggml_backend_cuda_synchronize,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_cuda_graph_compute,
/* .supports_op = */ ggml_backend_cuda_supports_op,
/* .offload_op = */ ggml_backend_cuda_offload_op,
/* .event_new = */ ggml_backend_cuda_event_new,
/* .event_free = */ ggml_backend_cuda_event_free,
/* .event_record = */ ggml_backend_cuda_event_record,
/* .event_wait = */ ggml_backend_cuda_event_wait,
/* .event_synchronize = */ ggml_backend_cuda_event_synchronize,
};
static ggml_guid_t ggml_backend_cuda_guid() {
static ggml_guid guid = { 0x2c, 0xdd, 0xe8, 0x1c, 0x65, 0xb3, 0x65, 0x73, 0x6a, 0x12, 0x88, 0x61, 0x1c, 0xc9, 0xdc, 0x25 };
return &guid;
}
GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
if (device < 0 || device >= ggml_backend_cuda_get_device_count()) {
GGML_CUDA_LOG_ERROR("%s: invalid device %d\n", __func__, device);
return nullptr;
}
ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context(device);
if (ctx == nullptr) {
GGML_CUDA_LOG_ERROR("%s: failed to allocate context\n", __func__);
return nullptr;
}
ggml_backend_t cuda_backend = new ggml_backend {
/* .guid = */ ggml_backend_cuda_guid(),
/* .interface = */ ggml_backend_cuda_interface,
/* .context = */ ctx
};
return cuda_backend;
}
GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cuda_guid());
}
GGML_CALL int ggml_backend_cuda_get_device_count() {
return ggml_cuda_info().device_count;
}
GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size) {
cudaDeviceProp prop;
CUDA_CHECK(cudaGetDeviceProperties(&prop, device));
snprintf(description, description_size, "%s", prop.name);
}
GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total) {
ggml_cuda_set_device(device);
CUDA_CHECK(cudaMemGetInfo(free, total));
}
GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) {
if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) {
return false;
}
#if CUDART_VERSION >= 11100
cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly);
if (err != cudaSuccess) {
// clear the error
cudaGetLastError();
GGML_CUDA_LOG_WARN("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__,
size / 1024.0 / 1024.0, cudaGetErrorString(err));
return false;
}
return true;
#else
return false;
#endif
}
GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer) {
if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) {
return;
}
cudaError_t err = cudaHostUnregister(buffer);
if (err != cudaSuccess) {
// clear the error
cudaGetLastError();
}
}
// backend registry
GGML_CALL static ggml_backend_t ggml_backend_reg_cuda_init(const char * params, void * user_data) {
ggml_backend_t cuda_backend = ggml_backend_cuda_init((int) (intptr_t) user_data);
return cuda_backend;
GGML_UNUSED(params);
}
extern "C" GGML_CALL int ggml_backend_cuda_reg_devices();
GGML_CALL int ggml_backend_cuda_reg_devices() {
int device_count = ggml_backend_cuda_get_device_count();
//int device_count = 1; // DEBUG: some tools require delaying CUDA initialization
for (int i = 0; i < device_count; i++) {
char name[128];
snprintf(name, sizeof(name), "%s%d", GGML_CUDA_NAME, i);
ggml_backend_register(name, ggml_backend_reg_cuda_init, ggml_backend_cuda_buffer_type(i), (void *) (intptr_t) i);
}
return device_count;
}
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#ifdef GGML_USE_HIPBLAS
#define GGML_CUDA_NAME "ROCm"
#define GGML_CUBLAS_NAME "hipBLAS"
#else
#define GGML_CUDA_NAME "CUDA"
#define GGML_CUBLAS_NAME "cuBLAS"
#endif
#ifdef __cplusplus
extern "C" {
#endif
#define GGML_CUDA_MAX_DEVICES 16
// backend API
GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
// device buffer
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
// split tensor buffer that splits matrices by rows across multiple devices
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
GGML_API GGML_CALL int ggml_backend_cuda_get_device_count(void);
GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data);
#ifdef __cplusplus
}
#endif
#include "acc.cuh"
static __global__ __launch_bounds__(1024) void acc_f32(const float * x, const float * y, float * dst, const int ne,
const int ne10, const int ne11, const int ne12,
const int nb1, const int nb2, int offset) {
const int i = blockDim.x * blockIdx.x + threadIdx.x;
if (i >= ne) {
return;
}
int src1_idx = i - offset;
int oz = src1_idx / nb2;
int oy = (src1_idx - (oz * nb2)) / nb1;
int ox = src1_idx % nb1;
if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
} else {
dst[i] = x[i];
}
}
static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
const int ne10, const int ne11, const int ne12,
const int nb1, const int nb2, const int offset, cudaStream_t stream) {
int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
}
void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const float * src0_d = (const float *)src0->data;
const float * src1_d = (const float *)src1->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
// int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
int offset = dst->op_params[3] / 4; // offset in bytes
acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, stream);
}
#include "common.cuh"
#define CUDA_ACC_BLOCK_SIZE 256
void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
#include "arange.cuh"
static __global__ __launch_bounds__(1024) void arange_f32(float * dst, const int ne0, const float start, const float step) {
// blockIDx.x: idx of ne0 / BLOCK_SIZE
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
if (nidx >= ne0) {
return;
}
dst[nidx] = start + step * nidx;
}
static void arange_f32_cuda(float * dst, const int ne0, const float start, const float step, cudaStream_t stream) {
int num_blocks = (ne0 + CUDA_ARANGE_BLOCK_SIZE - 1) / CUDA_ARANGE_BLOCK_SIZE;
arange_f32<<<num_blocks, CUDA_ARANGE_BLOCK_SIZE, 0, stream>>>(dst, ne0, start, step);
}
void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(dst->type == GGML_TYPE_F32);
float start;
float stop;
float step;
memcpy(&start, (float *)dst->op_params + 0, sizeof(float));
memcpy(&stop, (float *)dst->op_params + 1, sizeof(float));
memcpy(&step, (float *)dst->op_params + 2, sizeof(float));
int64_t steps = (int64_t)ceil((stop - start) / step);
GGML_ASSERT(ggml_nelements(dst) == steps);
arange_f32_cuda(dst_d, dst->ne[0], start, step, stream);
}
#include "common.cuh"
#define CUDA_ARANGE_BLOCK_SIZE 256
void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
#include "argsort.cuh"
template<typename T>
static inline __device__ void ggml_cuda_swap(T & a, T & b) {
T tmp = a;
a = b;
b = tmp;
}
template<ggml_sort_order order>
static __global__ __launch_bounds__(1024) void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
// bitonic sort
int col = threadIdx.x;
int row = blockIdx.y;
if (col >= ncols_pad) {
return;
}
const float * x_row = x + row * ncols;
extern __shared__ int dst_row[];
// initialize indices
dst_row[col] = col;
__syncthreads();
for (int k = 2; k <= ncols_pad; k *= 2) {
for (int j = k / 2; j > 0; j /= 2) {
int ixj = col ^ j;
if (ixj > col) {
if ((col & k) == 0) {
if (dst_row[col] >= ncols ||
(dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
x_row[dst_row[col]] > x_row[dst_row[ixj]] :
x_row[dst_row[col]] < x_row[dst_row[ixj]]))
) {
ggml_cuda_swap(dst_row[col], dst_row[ixj]);
}
} else {
if (dst_row[ixj] >= ncols ||
(dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
x_row[dst_row[col]] < x_row[dst_row[ixj]] :
x_row[dst_row[col]] > x_row[dst_row[ixj]]))
) {
ggml_cuda_swap(dst_row[col], dst_row[ixj]);
}
}
}
__syncthreads();
}
}
// copy the result to dst without the padding
if (col < ncols) {
dst[row * ncols + col] = dst_row[col];
}
}
static int next_power_of_2(int x) {
int n = 1;
while (n < x) {
n *= 2;
}
return n;
}
static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
// bitonic sort requires ncols to be power of 2
const int ncols_pad = next_power_of_2(ncols);
const dim3 block_dims(ncols_pad, 1, 1);
const dim3 block_nums(1, nrows, 1);
const size_t shared_mem = ncols_pad * sizeof(int);
GGML_ASSERT(shared_mem <= ggml_cuda_info().devices[ggml_cuda_get_device()].smpb);
if (order == GGML_SORT_ORDER_ASC) {
k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
} else if (order == GGML_SORT_ORDER_DESC) {
k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
} else {
GGML_ASSERT(false);
}
}
void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_I32);
GGML_ASSERT(ggml_is_contiguous(src0));
const int64_t ncols = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
argsort_f32_i32_cuda(src0_d, (int *)dst_d, ncols, nrows, order, stream);
}
#include "common.cuh"
void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
#include "binbcast.cuh"
static __device__ __forceinline__ float op_repeat(const float a, const float b) {
return b;
GGML_UNUSED(a);
}
static __device__ __forceinline__ float op_add(const float a, const float b) {
return a + b;
}
static __device__ __forceinline__ float op_mul(const float a, const float b) {
return a * b;
}
static __device__ __forceinline__ float op_div(const float a, const float b) {
return a / b;
}
template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
static __global__ __launch_bounds__(1024) void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
int ne0, int ne1, int ne2, int ne3,
int ne10, int ne11, int ne12, int ne13,
/*int s0, */ int s1, int s2, int s3,
/*int s00,*/ int s01, int s02, int s03,
/*int s10,*/ int s11, int s12, int s13) {
const int i0s = blockDim.x*blockIdx.x + threadIdx.x;
const int i1 = (blockDim.y*blockIdx.y + threadIdx.y);
const int i2 = (blockDim.z*blockIdx.z + threadIdx.z) / ne3;
const int i3 = (blockDim.z*blockIdx.z + threadIdx.z) % ne3;
if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
return;
}
const int i11 = i1 % ne11;
const int i12 = i2 % ne12;
const int i13 = i3 % ne13;
const size_t i_src0 = i3*s03 + i2*s02 + i1*s01;
const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
const size_t i_dst = i3*s3 + i2*s2 + i1*s1;
const src0_t * src0_row = src0 + i_src0;
const src1_t * src1_row = src1 + i_src1;
dst_t * dst_row = dst + i_dst;
for (int i0 = i0s; i0 < ne0; i0 += blockDim.x*gridDim.x) {
const int i10 = i0 % ne10;
dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
}
}
template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
static __global__ __launch_bounds__(1024) void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
int ne0, int ne1, int ne2, int ne3,
int ne10, int ne11, int ne12, int ne13,
/*int s0, */ int s1, int s2, int s3,
/*int s00,*/ int s01, int s02, int s03,
/*int s10,*/ int s11, int s12, int s13) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
const int i3 = i/(ne2*ne1*ne0);
const int i2 = (i/(ne1*ne0)) % ne2;
const int i1 = (i/ne0) % ne1;
const int i0 = i % ne0;
if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
return;
}
const int i11 = i1 % ne11;
const int i12 = i2 % ne12;
const int i13 = i3 % ne13;
const size_t i_src0 = i3*s03 + i2*s02 + i1*s01;
const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
const size_t i_dst = i3*s3 + i2*s2 + i1*s1;
const src0_t * src0_row = src0 + i_src0;
const src1_t * src1_row = src1 + i_src1;
dst_t * dst_row = dst + i_dst;
const int i10 = i0 % ne10;
dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
}
template<float (*bin_op)(const float, const float)>
struct bin_bcast_cuda {
template<typename src0_t, typename src1_t, typename dst_t>
void operator()(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst,
const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd,
cudaStream_t stream) {
GGML_TENSOR_BINARY_OP_LOCALS
int nr0 = ne10/ne0;
int nr1 = ne11/ne1;
int nr2 = ne12/ne2;
int nr3 = ne13/ne3;
int nr[4] = { nr0, nr1, nr2, nr3 };
// collapse dimensions until first broadcast dimension
int64_t cne[] = {ne0, ne1, ne2, ne3};
int64_t cne0[] = {ne00, ne01, ne02, ne03};
int64_t cne1[] = {ne10, ne11, ne12, ne13};
size_t cnb[] = {nb0, nb1, nb2, nb3};
size_t cnb0[] = {nb00, nb01, nb02, nb03};
size_t cnb1[] = {nb10, nb11, nb12, nb13};
auto collapse = [](int64_t cne[]) {
cne[0] *= cne[1];
cne[1] = cne[2];
cne[2] = cne[3];
cne[3] = 1;
};
auto collapse_nb = [](size_t cnb[], const int64_t cne[]) {
cnb[1] *= cne[1];
cnb[2] *= cne[2];
cnb[3] *= cne[3];
};
if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
for (int i = 0; i < 4; i++) {
if (nr[i] != 1) {
break;
}
if (i > 0) {
collapse_nb(cnb, cne);
collapse_nb(cnb0, cne0);
collapse_nb(cnb1, cne1);
collapse(cne);
collapse(cne0);
collapse(cne1);
}
}
}
{
int64_t ne0 = cne[0];
int64_t ne1 = cne[1];
int64_t ne2 = cne[2];
int64_t ne3 = cne[3];
//int64_t ne00 = cne0[0]; GGML_UNUSED(ne00);
//int64_t ne01 = cne0[1]; GGML_UNUSED(ne01);
//int64_t ne02 = cne0[2]; GGML_UNUSED(ne02);
//int64_t ne03 = cne0[3]; GGML_UNUSED(ne03);
int64_t ne10 = cne1[0];
int64_t ne11 = cne1[1];
int64_t ne12 = cne1[2];
int64_t ne13 = cne1[3];
size_t nb0 = cnb[0];
size_t nb1 = cnb[1];
size_t nb2 = cnb[2];
size_t nb3 = cnb[3];
size_t nb00 = cnb0[0];
size_t nb01 = cnb0[1];
size_t nb02 = cnb0[2];
size_t nb03 = cnb0[3];
size_t nb10 = cnb1[0];
size_t nb11 = cnb1[1];
size_t nb12 = cnb1[2];
size_t nb13 = cnb1[3];
size_t s0 = nb0 / sizeof(dst_t);
size_t s1 = nb1 / sizeof(dst_t);
size_t s2 = nb2 / sizeof(dst_t);
size_t s3 = nb3 / sizeof(dst_t);
size_t s10 = nb10 / sizeof(src1_t);
size_t s11 = nb11 / sizeof(src1_t);
size_t s12 = nb12 / sizeof(src1_t);
size_t s13 = nb13 / sizeof(src1_t);
size_t s00 = nb00 / sizeof(src0_t);
size_t s01 = nb01 / sizeof(src0_t);
size_t s02 = nb02 / sizeof(src0_t);
size_t s03 = nb03 / sizeof(src0_t);
GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
GGML_ASSERT(s0 == 1);
GGML_ASSERT(s00 == 1);
GGML_ASSERT(s10 == 1);
const int block_size = 128;
int64_t hne0 = std::max(ne0/2LL, 1LL);
dim3 block_dims;
block_dims.x = std::min<unsigned int>(hne0, block_size);
block_dims.y = std::min<unsigned int>(ne1, block_size / block_dims.x);
block_dims.z = std::min(std::min<unsigned int>(ne2*ne3, block_size / block_dims.x / block_dims.y), 64U);
dim3 block_nums(
(hne0 + block_dims.x - 1) / block_dims.x,
(ne1 + block_dims.y - 1) / block_dims.y,
(ne2*ne3 + block_dims.z - 1) / block_dims.z
);
if (block_nums.z > 65535) {
// this is the maximum number of blocks in z dimension, fallback to 1D grid kernel
int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
k_bin_bcast_unravel<bin_op><<<block_num, block_size, 0, stream>>>(
src0_dd, src1_dd, dst_dd,
ne0, ne1, ne2, ne3,
ne10, ne11, ne12, ne13,
/* s0, */ s1, s2, s3,
/* s00, */ s01, s02, s03,
/* s10, */ s11, s12, s13);
} else {
k_bin_bcast<bin_op><<<block_nums, block_dims, 0, stream>>>(
src0_dd, src1_dd, dst_dd,
ne0, ne1, ne2, ne3,
ne10, ne11, ne12, ne13,
/* s0, */ s1, s2, s3,
/* s00, */ s01, s02, s03,
/* s10, */ s11, s12, s13);
}
}
}
};
template<class op>
static void ggml_cuda_op_bin_bcast(
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
const void * src0_dd, const void * src1_dd, void * dst_dd, cudaStream_t stream) {
GGML_ASSERT(src1->type == GGML_TYPE_F32);
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
op()(src0, src1, dst, (const float *)src0_dd, (const float *)src1_dd, (float *)dst_dd, stream);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (half *) dst_dd, stream);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (float *)dst_dd, stream);
} else {
fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
GGML_ASSERT(false);
}
}
void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_repeat>>(dst, dst->src[0], dst, nullptr, dst->src[0]->data, dst->data, ctx.stream());
}
void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_add>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
}
void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_mul>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
}
void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_div>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
}
#include "common.cuh"
void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
#include "clamp.cuh"
static __global__ __launch_bounds__(1024) void clamp_f32(const float * x, float * dst, const float min, const float max, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
}
static void clamp_f32_cuda(const float * x, float * dst, const float min, const float max, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE;
clamp_f32<<<num_blocks, CUDA_CLAMP_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
}
void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
float min;
float max;
memcpy(&min, dst->op_params, sizeof(float));
memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
clamp_f32_cuda(src0_d, dst_d, min, max, ggml_nelements(src0), stream);
}
#include "common.cuh"
#define CUDA_CLAMP_BLOCK_SIZE 256
void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
#pragma once
#include "ggml.h"
#include "ggml-cuda.h"
#include <memory>
#if defined(GGML_USE_HIPBLAS)
#define GGML_COMMON_DECL_HIP
#define GGML_COMMON_IMPL_HIP
#else
#define GGML_COMMON_DECL_CUDA
#define GGML_COMMON_IMPL_CUDA
#endif
#include "ggml-common.h"
#include <cstdio>
#include <array>
#include <cassert>
#include <cfloat>
#include <string>
#include <vector>
#if defined(GGML_USE_HIPBLAS)
#include <hip/hip_runtime.h>
#include <hipblas/hipblas.h>
#include <hip/hip_fp16.h>
#ifdef __HIP_PLATFORM_AMD__
// for rocblas_initialize()
#include "rocblas/rocblas.h"
#endif // __HIP_PLATFORM_AMD__
#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
#define CUBLAS_OP_N HIPBLAS_OP_N
#define CUBLAS_OP_T HIPBLAS_OP_T
#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
#define CUBLAS_TF32_TENSOR_OP_MATH 0
#define CUDA_R_16F HIPBLAS_R_16F
#define CUDA_R_32F HIPBLAS_R_32F
#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
#define cublasCreate hipblasCreate
#define cublasDestroy hipblasDestroy
#define cublasGemmEx hipblasGemmEx
#define cublasGemmBatchedEx hipblasGemmBatchedEx
#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
#define cublasHandle_t hipblasHandle_t
#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
#define cublasSetStream hipblasSetStream
#define cublasSgemm hipblasSgemm
#define cublasStatus_t hipblasStatus_t
#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
#define cudaDeviceProp hipDeviceProp_t
#define cudaDeviceSynchronize hipDeviceSynchronize
#define cudaError_t hipError_t
#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
#define cudaEventCreateWithFlags hipEventCreateWithFlags
#define cudaEventDisableTiming hipEventDisableTiming
#define cudaEventRecord hipEventRecord
#define cudaEventSynchronize hipEventSynchronize
#define cudaEvent_t hipEvent_t
#define cudaEventDestroy hipEventDestroy
#define cudaFree hipFree
#define cudaFreeHost hipHostFree
#define cudaGetDevice hipGetDevice
#define cudaGetDeviceCount hipGetDeviceCount
#define cudaGetDeviceProperties hipGetDeviceProperties
#define cudaGetErrorString hipGetErrorString
#define cudaGetLastError hipGetLastError
#define cudaHostRegister hipHostRegister
#define cudaHostRegisterPortable hipHostRegisterPortable
#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
#define cudaHostUnregister hipHostUnregister
#define cudaLaunchHostFunc hipLaunchHostFunc
#define cudaMalloc hipMalloc
#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
#define cudaMemcpy hipMemcpy
#define cudaMemcpyAsync hipMemcpyAsync
#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
#define cudaMemcpy2DAsync hipMemcpy2DAsync
#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
#define cudaMemcpyKind hipMemcpyKind
#define cudaMemset hipMemset
#define cudaMemsetAsync hipMemsetAsync
#define cudaMemGetInfo hipMemGetInfo
#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
#define cudaSetDevice hipSetDevice
#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
#define cudaStreamDestroy hipStreamDestroy
#define cudaStreamFireAndForget hipStreamFireAndForget
#define cudaStreamNonBlocking hipStreamNonBlocking
#define cudaStreamPerThread hipStreamPerThread
#define cudaStreamSynchronize hipStreamSynchronize
#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
#define cudaStream_t hipStream_t
#define cudaSuccess hipSuccess
#define __trap abort
#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED
#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE
#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH
#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR
#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED
#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
#else
#include <cuda_runtime.h>
#include <cuda.h>
#include <cublas_v2.h>
#include <cuda_fp16.h>
#if CUDART_VERSION < 11020
#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
#define CUBLAS_COMPUTE_16F CUDA_R_16F
#define CUBLAS_COMPUTE_32F CUDA_R_32F
#define cublasComputeType_t cudaDataType_t
#endif // CUDART_VERSION < 11020
#endif // defined(GGML_USE_HIPBLAS)
#define STRINGIZE_IMPL(...) #__VA_ARGS__
#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
#define WARP_SIZE 32
#define CUDART_HMAX 11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
#define CUDART_HMASK 12000 // CUDA 12.0, min. ver. for half2 -> uint mask comparisons
#define CC_PASCAL 600
#define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
#define CC_VOLTA 700
#define CC_AMPERE 800
#define CC_OFFSET_AMD 1000000
#define CC_RDNA1 (CC_OFFSET_AMD + 1010)
#define CC_RDNA2 (CC_OFFSET_AMD + 1030)
#define CC_RDNA3 (CC_OFFSET_AMD + 1100)
// define this if you want to always fallback to MMQ kernels and not use cuBLAS for matrix multiplication
// on modern hardware, using cuBLAS is recommended as it utilizes F16 tensor cores which are very performant
// for large computational tasks. the drawback is that this requires some extra amount of VRAM:
// - 7B quantum model: +100-200 MB
// - 13B quantum model: +200-400 MB
//
//#define GGML_CUDA_FORCE_MMQ
// TODO: improve this to be correct for more hardware
// for example, currently fails for GeForce GTX 1660 which is TURING arch (> VOLTA) but does not have tensor cores
#if !defined(GGML_CUDA_FORCE_MMQ)
#define CUDA_USE_TENSOR_CORES
#endif
#define MMVQ_MAX_BATCH_SIZE 8 // max batch size to use MMVQ kernels
#define MMQ_MAX_BATCH_SIZE 32 // max batch size to use MMQ kernels when tensor cores are available
#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
#define GGML_CUDA_MAX_STREAMS 8
[[noreturn]]
void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg);
#define CUDA_CHECK_GEN(err, success, error_fn) \
do { \
auto err_ = (err); \
if (err_ != (success)) { \
ggml_cuda_error(#err, __func__, __FILE__, __LINE__, error_fn(err_)); \
} \
} while (0)
#define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString)
#if CUDART_VERSION >= 12000
static const char * cublas_get_error_str(const cublasStatus_t err) {
return cublasGetStatusString(err);
}
#else
static const char * cublas_get_error_str(const cublasStatus_t err) {
switch (err) {
case CUBLAS_STATUS_SUCCESS: return "CUBLAS_STATUS_SUCCESS";
case CUBLAS_STATUS_NOT_INITIALIZED: return "CUBLAS_STATUS_NOT_INITIALIZED";
case CUBLAS_STATUS_ALLOC_FAILED: return "CUBLAS_STATUS_ALLOC_FAILED";
case CUBLAS_STATUS_INVALID_VALUE: return "CUBLAS_STATUS_INVALID_VALUE";
case CUBLAS_STATUS_ARCH_MISMATCH: return "CUBLAS_STATUS_ARCH_MISMATCH";
case CUBLAS_STATUS_MAPPING_ERROR: return "CUBLAS_STATUS_MAPPING_ERROR";
case CUBLAS_STATUS_EXECUTION_FAILED: return "CUBLAS_STATUS_EXECUTION_FAILED";
case CUBLAS_STATUS_INTERNAL_ERROR: return "CUBLAS_STATUS_INTERNAL_ERROR";
case CUBLAS_STATUS_NOT_SUPPORTED: return "CUBLAS_STATUS_NOT_SUPPORTED";
default: return "unknown error";
}
}
#endif // CUDART_VERSION >= 12000
#define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str)
#if !defined(GGML_USE_HIPBLAS)
static const char * cu_get_error_str(CUresult err) {
const char * err_str;
cuGetErrorString(err, &err_str);
return err_str;
}
#define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
#endif
#if CUDART_VERSION >= 11100
#define GGML_CUDA_ASSUME(x) __builtin_assume(x)
#else
#define GGML_CUDA_ASSUME(x)
#endif // CUDART_VERSION >= 11100
#ifdef GGML_CUDA_F16
typedef half dfloat; // dequantize float
typedef half2 dfloat2;
#else
typedef float dfloat; // dequantize float
typedef float2 dfloat2;
#endif //GGML_CUDA_F16
#if defined(GGML_USE_HIPBLAS)
#define __CUDA_ARCH__ 1300
#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
defined(__gfx1150__) || defined(__gfx1151__)
#define RDNA3
#endif
#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
#define RDNA2
#endif
#ifndef __has_builtin
#define __has_builtin(x) 0
#endif
typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
#if __has_builtin(__builtin_elementwise_sub_sat)
const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
return reinterpret_cast<const int &>(c);
#else
int8x4_t c;
int16_t tmp;
#pragma unroll
for (int i = 0; i < 4; i++) {
tmp = va[i] - vb[i];
if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
c[i] = tmp;
}
return reinterpret_cast<int &>(c);
#endif // __has_builtin(__builtin_elementwise_sub_sat)
}
static __device__ __forceinline__ int __vsub4(const int a, const int b) {
return __vsubss4(a, b);
}
static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) {
const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
unsigned int c;
uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
#pragma unroll
for (int i = 0; i < 4; ++i) {
vc[i] = va[i] == vb[i] ? 0xff : 0x00;
}
return c;
}
static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__)
c = __builtin_amdgcn_sdot4(a, b, c, false);
#elif defined(RDNA3)
c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
#elif defined(__gfx1010__) || defined(__gfx900__)
int tmp1;
int tmp2;
asm("\n \
v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \
v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \
v_add3_u32 %0, %1, %2, %0 \n \
v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \
v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \
v_add3_u32 %0, %1, %2, %0 \n \
"
: "+v"(c), "=&v"(tmp1), "=&v"(tmp2)
: "v"(a), "v"(b)
);
#else
const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3];
#endif
return c;
}
#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
// __shfl_xor() for half2 was added in ROCm 5.6
static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
typedef union half2_b32 {
half2 val;
int b32;
} half2_b32_t;
half2_b32_t tmp;
tmp.val = var;
tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
return tmp.val;
}
#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
#endif // defined(GGML_USE_HIPBLAS)
// #define FP16_AVAILABLE (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
#define FP16_AVAILABLE ((defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || (defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= CC_PASCAL))
#define FP16_MMA_AVAILABLE !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
static bool fast_fp16_available(const int cc) {
return cc >= CC_PASCAL && cc != 610;
}
static bool fp16_mma_available(const int cc) {
return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
}
[[noreturn]]
static __device__ void no_device_code(
const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
file_name, line, function_name, arch);
GGML_UNUSED(arch_list);
#else
printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
file_name, line, function_name, arch, arch_list);
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
__trap();
GGML_UNUSED(no_device_code); // suppress unused function warning
}
#ifdef __CUDA_ARCH__
#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__))
#else
#define NO_DEVICE_CODE //GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.")
#endif // __CUDA_ARCH__
static __device__ __forceinline__ float warp_reduce_sum(float x) {
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
x += __shfl_xor_sync(0xffffffff, x, mask, 32);
}
return x;
}
static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
}
return a;
}
static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
#if FP16_AVAILABLE
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
reinterpret_cast<half&>(a.x) += __low2half(a_other);
reinterpret_cast<half&>(a.y) += __high2half(a_other);
}
return a;
#else
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
}
return a;
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#else
NO_DEVICE_CODE;
return a;
#endif // FP16_AVAILABLE
}
static __device__ __forceinline__ float warp_reduce_max(float x) {
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
}
return x;
}
/*
static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
#if FP16_AVAILABLE
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
return __float2half(fmaxf(__half2float(a), __half2float(b)));
#else
return __hmax(a, b);
// return __fmax(a, b);
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
#else
NO_DEVICE_CODE;
GGML_UNUSED(b);
return a;
#endif // FP16_AVAILABLE
}
*/
static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
#if FP16_AVAILABLE
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
return __float2half(fmaxf(__half2float(a), __half2float(b)));
#else
// 使用标准C++函数 std::max
return __float2half(std::max(__half2float(a), __half2float(b)));
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
#else
NO_DEVICE_CODE;
GGML_UNUSED(b);
return a;
#endif // FP16_AVAILABLE
}
static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if CUDART_VERSION >= CUDART_HMAX
return __hmax2(a, b);
#else
half2 ret;
reinterpret_cast<half&>(ret.x) = __float2half(fmaxf( __low2float(a), __low2float(b)));
reinterpret_cast<half&>(ret.y) = __float2half(fmaxf(__high2float(a), __high2float(b)));
return ret;
#endif // CUDART_VERSION >= CUDART_HMAX
#else
GGML_UNUSED(a);
GGML_UNUSED(b);
NO_DEVICE_CODE;
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
}
static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
}
return x;
#else
GGML_UNUSED(x);
NO_DEVICE_CODE;
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
}
#if CUDART_VERSION < CUDART_HMASK
static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) {
const uint32_t mask_low = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b)));
const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
return mask_low | mask_high;
}
#endif // CUDART_VERSION < 12000
// TODO: move to ggml-common.h
static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
static __device__ __forceinline__ float get_alibi_slope(
const float max_bias, const uint32_t h, const uint32_t n_head_log2, const float m0, const float m1
) {
if (max_bias <= 0.0f) {
return 1.0f;
}
const float base = h < n_head_log2 ? m0 : m1;
const int exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
return powf(base, exph);
}
//////////////////////
struct ggml_cuda_device_info {
int device_count;
struct cuda_device_info {
int cc; // compute capability
int nsm; // number of streaming multiprocessors
size_t smpb; // max. shared memory per block
bool vmm; // virtual memory support
size_t vmm_granularity; // granularity of virtual memory
size_t total_vram;
};
cuda_device_info devices[GGML_CUDA_MAX_DEVICES] = {};
std::array<float, GGML_CUDA_MAX_DEVICES> default_tensor_split = {};
};
const ggml_cuda_device_info & ggml_cuda_info();
void ggml_cuda_set_device(int device);
int ggml_cuda_get_device();
struct ggml_cuda_pool {
virtual ~ggml_cuda_pool() = default;
virtual void * alloc(size_t size, size_t * actual_size) = 0;
virtual void free(void * ptr, size_t size) = 0;
};
template<typename T>
struct ggml_cuda_pool_alloc {
ggml_cuda_pool * pool = nullptr;
T * ptr = nullptr;
size_t actual_size = 0;
ggml_cuda_pool_alloc() = default;
explicit ggml_cuda_pool_alloc(ggml_cuda_pool & pool) : pool(&pool) {
}
ggml_cuda_pool_alloc(ggml_cuda_pool & pool, size_t size) : pool(&pool) {
alloc(size);
}
~ggml_cuda_pool_alloc() {
if (ptr != nullptr) {
pool->free(ptr, actual_size);
}
}
// size is in number of elements
T * alloc(size_t size) {
GGML_ASSERT(pool != nullptr);
GGML_ASSERT(ptr == nullptr);
ptr = (T *) pool->alloc(size * sizeof(T), &this->actual_size);
return ptr;
}
T * alloc(ggml_cuda_pool & pool, size_t size) {
this->pool = &pool;
return alloc(size);
}
T * get() {
return ptr;
}
ggml_cuda_pool_alloc(const ggml_cuda_pool_alloc &) = delete;
ggml_cuda_pool_alloc(ggml_cuda_pool_alloc &&) = delete;
ggml_cuda_pool_alloc& operator=(const ggml_cuda_pool_alloc &) = delete;
ggml_cuda_pool_alloc& operator=(ggml_cuda_pool_alloc &&) = delete;
};
// backend interface
struct ggml_tensor_extra_gpu {
void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors
cudaEvent_t events[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS]; // events for synchronizing multiple GPUs
};
#if (CUDART_VERSION >= 12000) && defined(GGML_CUDA_USE_GRAPHS)
#define USE_CUDA_GRAPH
#endif
struct ggml_graph_node_properties {
void * node_address;
ggml_op node_op;
int64_t ne[GGML_MAX_DIMS];
size_t nb[GGML_MAX_DIMS];
void * src_address[GGML_MAX_SRC];
};
struct ggml_cuda_graph {
#ifdef USE_CUDA_GRAPH
~ggml_cuda_graph() {
if (instance != nullptr) {
CUDA_CHECK(cudaGraphExecDestroy(instance));
}
if (graph != nullptr) {
CUDA_CHECK(cudaGraphDestroy(graph));
}
}
cudaGraph_t graph = nullptr;
cudaGraphExec_t instance = nullptr;
size_t num_nodes = 0;
std::vector<cudaGraphNode_t> nodes;
std::vector<cudaKernelNodeParams> params;
bool disable_due_to_gpu_arch = false;
bool disable_due_to_too_many_updates = false;
bool disable_due_to_failed_graph_capture = false;
int number_consecutive_updates = 0;
std::vector<ggml_graph_node_properties> ggml_graph_properties;
std::vector<char **> updated_kernel_arg;
#endif
};
struct ggml_backend_cuda_context {
int device;
std::string name;
cudaEvent_t copy_event = nullptr;
cudaStream_t streams[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { { nullptr } };
cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
std::unique_ptr<ggml_cuda_graph> cuda_graph;
explicit ggml_backend_cuda_context(int device) :
device(device),
name(GGML_CUDA_NAME + std::to_string(device)) {
}
~ggml_backend_cuda_context() {
if (copy_event != nullptr) {
CUDA_CHECK(cudaEventDestroy(copy_event));
}
for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
if (streams[i][j] != nullptr) {
CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
}
}
if (cublas_handles[i] != nullptr) {
CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
}
}
}
cudaStream_t stream(int device, int stream) {
if (streams[device][stream] == nullptr) {
ggml_cuda_set_device(device);
CUDA_CHECK(cudaStreamCreateWithFlags(&streams[device][stream], cudaStreamNonBlocking));
}
return streams[device][stream];
}
cudaStream_t stream() {
return stream(device, 0);
}
cublasHandle_t cublas_handle(int device) {
if (cublas_handles[device] == nullptr) {
ggml_cuda_set_device(device);
CUBLAS_CHECK(cublasCreate(&cublas_handles[device]));
CUBLAS_CHECK(cublasSetMathMode(cublas_handles[device], CUBLAS_TF32_TENSOR_OP_MATH));
}
return cublas_handles[device];
}
cublasHandle_t cublas_handle() {
return cublas_handle(device);
}
// pool
std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES];
static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device);
ggml_cuda_pool & pool(int device) {
if (pools[device] == nullptr) {
pools[device] = new_pool_for_device(device);
}
return *pools[device];
}
ggml_cuda_pool & pool() {
return pool(device);
}
};
#include "concat.cuh"
// contiguous kernels
static __global__ __launch_bounds__(1024) void concat_f32_dim0(const float * x, const float * y, float * dst, const int ne0, const int ne00) {
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
if (nidx >= ne0) {
return;
}
int offset_dst =
nidx +
blockIdx.y * ne0 +
blockIdx.z * ne0 * gridDim.y;
if (nidx < ne00) { // src0
int offset_src =
nidx +
blockIdx.y * ne00 +
blockIdx.z * ne00 * gridDim.y;
dst[offset_dst] = x[offset_src];
} else {
int offset_src =
(nidx - ne00) +
blockIdx.y * (ne0 - ne00) +
blockIdx.z * (ne0 - ne00) * gridDim.y;
dst[offset_dst] = y[offset_src];
}
}
static __global__ __launch_bounds__(1024) void concat_f32_dim1(const float * x, const float * y, float * dst, const int ne0, const int ne01) {
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
if (nidx >= ne0) {
return;
}
int offset_dst =
nidx +
blockIdx.y * ne0 +
blockIdx.z * ne0 * gridDim.y;
if (blockIdx.y < ne01) { // src0
int offset_src =
nidx +
blockIdx.y * ne0 +
blockIdx.z * ne0 * ne01;
dst[offset_dst] = x[offset_src];
} else {
int offset_src =
nidx +
(blockIdx.y - ne01) * ne0 +
blockIdx.z * ne0 * (gridDim.y - ne01);
dst[offset_dst] = y[offset_src];
}
}
static __global__ __launch_bounds__(1024) void concat_f32_dim2(const float * x, const float * y, float * dst, const int ne0, const int ne02) {
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
if (nidx >= ne0) {
return;
}
int offset_dst =
nidx +
blockIdx.y * ne0 +
blockIdx.z * ne0 * gridDim.y;
if (blockIdx.z < ne02) { // src0
int offset_src =
nidx +
blockIdx.y * ne0 +
blockIdx.z * ne0 * gridDim.y;
dst[offset_dst] = x[offset_src];
} else {
int offset_src =
nidx +
blockIdx.y * ne0 +
(blockIdx.z - ne02) * ne0 * gridDim.y;
dst[offset_dst] = y[offset_src];
}
}
static void concat_f32_cuda(const float * x, const float * y, float * dst, int ne00, int ne01, int ne02, int ne0, int ne1, int ne2, int dim, cudaStream_t stream) {
int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
dim3 gridDim(num_blocks, ne1, ne2);
if (dim == 0) {
concat_f32_dim0<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne00);
return;
}
if (dim == 1) {
concat_f32_dim1<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne01);
return;
}
concat_f32_dim2<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
}
// non-contiguous kernel (slow)
static __global__ __launch_bounds__(1024) void concat_f32_non_cont(
const char * src0,
const char * src1,
char * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne03,
uint64_t nb00,
uint64_t nb01,
uint64_t nb02,
uint64_t nb03,
int64_t /*ne10*/,
int64_t /*ne11*/,
int64_t /*ne12*/,
int64_t /*ne13*/,
uint64_t nb10,
uint64_t nb11,
uint64_t nb12,
uint64_t nb13,
int64_t ne0,
int64_t /*ne1*/,
int64_t /*ne2*/,
int64_t /*ne3*/,
uint64_t nb0,
uint64_t nb1,
uint64_t nb2,
uint64_t nb3,
int32_t dim) {
const int64_t i3 = blockIdx.z;
const int64_t i2 = blockIdx.y;
const int64_t i1 = blockIdx.x;
int64_t o[4] = {0, 0, 0, 0};
o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
const float * x;
for (int i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
x = (const float *)(src0 + (i3 )*nb03 + (i2 )*nb02 + (i1 )*nb01 + (i0 )*nb00);
} else {
x = (const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
}
float * y = (float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
*y = *x;
}
}
void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
cudaStream_t stream = ctx.stream();
const int32_t dim = ((int32_t *) dst->op_params)[0];
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
const float * src0_d = (const float *)src0->data;
const float * src1_d = (const float *)src1->data;
float * dst_d = (float *)dst->data;
if (dim != 3) {
for (int i3 = 0; i3 < dst->ne[3]; i3++) {
concat_f32_cuda(
src0_d + i3 * (src0->nb[3] / 4),
src1_d + i3 * (src1->nb[3] / 4),
dst_d + i3 * ( dst->nb[3] / 4),
src0->ne[0], src0->ne[1], src0->ne[2],
dst->ne[0], dst->ne[1], dst->ne[2], dim, stream);
}
} else {
const size_t size0 = ggml_nbytes(src0);
const size_t size1 = ggml_nbytes(src1);
CUDA_CHECK(cudaMemcpyAsync(dst_d, src0_d, size0, cudaMemcpyDeviceToDevice, stream));
CUDA_CHECK(cudaMemcpyAsync(dst_d + size0/4, src1_d, size1, cudaMemcpyDeviceToDevice, stream));
}
} else {
dim3 grid_dim(dst->ne[1], dst->ne[2], dst->ne[3]);
concat_f32_non_cont<<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
(const char *)src0->data,
(const char *)src1->data,
( char *)dst->data,
src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], dim);
}
}
#include "common.cuh"
#define CUDA_CONCAT_BLOCK_SIZE 256
void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
#include "convert.cuh"
#include "dequantize.cuh"
#define CUDA_Q8_0_NE_ALIGN 2048
template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
const int64_t i = (int64_t)2*(blockDim.x*blockIdx.x + threadIdx.x);
if (i >= k) {
return;
}
const int64_t ib = i/qk; // block index
const int64_t iqs = (i%qk)/qr; // quant index
const int64_t iybs = i - i%qk; // y block start index
const int64_t y_offset = qr == 1 ? 1 : qk/2;
// dequantize
dfloat2 v;
dequantize_kernel(vx, ib, iqs, v);
y[iybs + iqs + 0] = v.x;
y[iybs + iqs + y_offset] = v.y;
}
template <bool need_check>
static __global__ __launch_bounds__(1024) void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int64_t k) {
#if __CUDA_ARCH__ >= CC_PASCAL
constexpr int nint = CUDA_Q8_0_NE_ALIGN/sizeof(int) + WARP_SIZE;
const int64_t i0 = CUDA_Q8_0_NE_ALIGN*blockIdx.x;
const int * x0 = ((int *) vx) + blockIdx.x * nint;
half2 * y2 = (half2 *) (y + i0);
__shared__ int vals[nint];
#pragma unroll
for (int ix0 = 0; ix0 < nint; ix0 += WARP_SIZE) {
if (need_check && i0*sizeof(block_q8_0)/QK8_0 + sizeof(int)*(ix0 + threadIdx.x) >= k*sizeof(block_q8_0)/QK8_0) {
break;
}
const int ix = ix0 + threadIdx.x;
vals[ix] = x0[ix];
}
__syncthreads();
#pragma unroll
for (int iy = 0; iy < CUDA_Q8_0_NE_ALIGN; iy += 2*WARP_SIZE) {
if (need_check && i0 + iy + 2*threadIdx.x >= k) {
return;
}
const half * b0 = ((const half *) vals) + (sizeof(block_q8_0)/sizeof(half)) * ((iy + 2*threadIdx.x)/QK8_0);
const half d = *b0;
const char2 qs = ((const char2 *) (b0 + 1))[threadIdx.x % (QK8_0/2)];
y2[iy/2 + threadIdx.x] = __hmul2(make_half2(qs.x, qs.y), __half2half2(d));
}
#else
GGML_UNUSED(vx);
GGML_UNUSED(y);
GGML_UNUSED(k);
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= CC_PASCAL
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) {
const int64_t i = blockIdx.x;
// assume 32 threads
const int64_t tid = threadIdx.x;
const int64_t il = tid/8;
const int64_t ir = tid%8;
const int64_t ib = 8*i + ir;
if (ib >= nb32) {
return;
}
dst_t * y = yy + 256*i + 32*ir + 4*il;
const block_q4_0 * x = (const block_q4_0 *)vx + ib;
const float d = __half2float(x->d);
const float dm = -8*d;
const uint8_t * q = x->qs + 4*il;
for (int l = 0; l < 4; ++l) {
y[l+ 0] = d * (q[l] & 0xF) + dm;
y[l+16] = d * (q[l] >> 4) + dm;
}
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) {
const int64_t i = blockIdx.x;
// assume 32 threads
const int64_t tid = threadIdx.x;
const int64_t il = tid/8;
const int64_t ir = tid%8;
const int64_t ib = 8*i + ir;
if (ib >= nb32) {
return;
}
dst_t * y = yy + 256*i + 32*ir + 4*il;
const block_q4_1 * x = (const block_q4_1 *)vx + ib;
const float2 d = __half22float2(x->dm);
const uint8_t * q = x->qs + 4*il;
for (int l = 0; l < 4; ++l) {
y[l+ 0] = d.x * (q[l] & 0xF) + d.y;
y[l+16] = d.x * (q[l] >> 4) + d.y;
}
}
//================================== k-quants
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int64_t i = blockIdx.x;
const block_q2_K * x = (const block_q2_K *) vx;
const int64_t tid = threadIdx.x;
const int64_t n = tid/32;
const int64_t l = tid - 32*n;
const int64_t is = 8*n + l/16;
const uint8_t q = x[i].qs[32*n + l];
dst_t * y = yy + i*QK_K + 128*n;
float dall = __low2half(x[i].dm);
float dmin = __high2half(x[i].dm);
y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int64_t i = blockIdx.x;
const block_q3_K * x = (const block_q3_K *) vx;
const int64_t r = threadIdx.x/4;
const int64_t tid = r/2;
const int64_t is0 = r%2;
const int64_t l0 = 16*is0 + 4*(threadIdx.x%4);
const int64_t n = tid / 4;
const int64_t j = tid - 4*n;
uint8_t m = 1 << (4*n + j);
int64_t is = 8*n + 2*j + is0;
int shift = 2*j;
int8_t us = is < 4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
is < 8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
is < 12 ? (x[i].scales[is-8] >> 4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
(x[i].scales[is-8] >> 4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
float d_all = x[i].d;
float dl = d_all * (us - 32);
dst_t * y = yy + i*QK_K + 128*n + 32*j;
const uint8_t * q = x[i].qs + 32*n;
const uint8_t * hm = x[i].hmask;
for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
}
static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
if (j < 4) {
d = q[j] & 63; m = q[j + 4] & 63;
} else {
d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4);
}
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const block_q4_K * x = (const block_q4_K *) vx;
const int64_t i = blockIdx.x;
// assume 32 threads
const int64_t tid = threadIdx.x;
const int64_t il = tid/8;
const int64_t ir = tid%8;
const int64_t is = 2*il;
const int64_t n = 4;
dst_t * y = yy + i*QK_K + 64*il + n*ir;
const float dall = __low2half(x[i].dm);
const float dmin = __high2half(x[i].dm);
const uint8_t * q = x[i].qs + 32*il + n*ir;
uint8_t sc, m;
get_scale_min_k4(is + 0, x[i].scales, sc, m);
const float d1 = dall * sc; const float m1 = dmin * m;
get_scale_min_k4(is + 1, x[i].scales, sc, m);
const float d2 = dall * sc; const float m2 = dmin * m;
for (int l = 0; l < n; ++l) {
y[l + 0] = d1 * (q[l] & 0xF) - m1;
y[l +32] = d2 * (q[l] >> 4) - m2;
}
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const block_q5_K * x = (const block_q5_K *) vx;
const int64_t i = blockIdx.x;
// assume 64 threads - this is very slightly better than the one below
const int64_t tid = threadIdx.x;
const int64_t il = tid/16; // il is in 0...3
const int64_t ir = tid%16; // ir is in 0...15
const int64_t is = 2*il; // is is in 0...6
dst_t * y = yy + i*QK_K + 64*il + 2*ir;
const float dall = __low2half(x[i].dm);
const float dmin = __high2half(x[i].dm);
const uint8_t * ql = x[i].qs + 32*il + 2*ir;
const uint8_t * qh = x[i].qh + 2*ir;
uint8_t sc, m;
get_scale_min_k4(is + 0, x[i].scales, sc, m);
const float d1 = dall * sc; const float m1 = dmin * m;
get_scale_min_k4(is + 1, x[i].scales, sc, m);
const float d2 = dall * sc; const float m2 = dmin * m;
uint8_t hm = 1 << (2*il);
y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1;
y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1;
hm <<= 1;
y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2;
y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2;
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const block_q6_K * x = (const block_q6_K *) vx;
const int64_t i = blockIdx.x;
// assume 64 threads - this is very slightly better than the one below
const int64_t tid = threadIdx.x;
const int64_t ip = tid/32; // ip is 0 or 1
const int64_t il = tid - 32*ip; // 0...32
const int64_t is = 8*ip + il/16;
dst_t * y = yy + i*QK_K + 128*ip + il;
const float d = x[i].d;
const uint8_t * ql = x[i].ql + 64*ip + il;
const uint8_t qh = x[i].qh[32*ip + il];
const int8_t * sc = x[i].scales + is;
y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
y[64] = d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32);
y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int64_t i = blockIdx.x;
const block_iq2_xxs * x = (const block_iq2_xxs *) vx;
const int64_t tid = threadIdx.x;
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
const uint16_t * q2 = x[i].qs + 4*ib;
const uint8_t * aux8 = (const uint8_t *)q2;
const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]);
const uint32_t aux32 = q2[2] | (q2[3] << 16);
const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int64_t i = blockIdx.x;
const block_iq2_xs * x = (const block_iq2_xs *) vx;
const int64_t tid = threadIdx.x;
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
const uint16_t * q2 = x[i].qs + 4*ib;
const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int64_t i = blockIdx.x;
const block_iq2_s * x = (const block_iq2_s *) vx;
const int64_t tid = threadIdx.x;
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int64_t i = blockIdx.x;
const block_iq3_xxs * x = (const block_iq3_xxs *) vx;
const int64_t tid = threadIdx.x;
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
const uint8_t * q3 = x[i].qs + 8*ib;
const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib;
const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]);
const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]);
const uint32_t aux32 = gas[0] | (gas[1] << 16);
const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f;
const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
for (int j = 0; j < 4; ++j) {
y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
}
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int64_t i = blockIdx.x;
const block_iq3_s * x = (const block_iq3_s *) vx;
const int64_t tid = threadIdx.x;
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
const uint8_t * qs = x[i].qs + 8*ib;
const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)));
const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)));
const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf));
const uint8_t signs = x[i].signs[4*ib + il];
for (int j = 0; j < 4; ++j) {
y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
}
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int64_t i = blockIdx.x;
const block_iq1_s * x = (const block_iq1_s *) vx;
const int64_t tid = threadIdx.x;
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA;
const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1);
uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)];
grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
grid32[0] &= 0x0f0f0f0f;
for (int j = 0; j < 8; ++j) {
y[j] = d * (q[j] + delta);
}
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_iq1_m(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int64_t i = blockIdx.x;
const block_iq1_m * x = (const block_iq1_m *) vx;
const int64_t tid = threadIdx.x;
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
const uint16_t * sc = (const uint16_t *)x[i].scales;
iq1m_scale_t scale;
scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
const int64_t ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4);
const float d = (float)scale.f16 * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1);
const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA;
uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)];
grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
grid32[0] &= 0x0f0f0f0f;
for (int j = 0; j < 8; ++j) {
y[j] = d * (q[j] + delta);
}
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int64_t i = blockIdx.x;
const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
const int64_t tid = threadIdx.x;
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 4*il;
const uint8_t * q4 = x[ib].qs + 4*il;
const float d = (float)x[ib].d;
for (int j = 0; j < 4; ++j) {
y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
y[j+16] = d * kvalues_iq4nl[q4[j] >> 4];
}
}
template<typename dst_t>
static __global__ __launch_bounds__(1024) void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int64_t i = blockIdx.x;
const block_iq4_xs * x = (const block_iq4_xs *)vx;
const int64_t tid = threadIdx.x;
const int64_t il = tid/8; // 0...3
const int64_t ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 4*il;
const uint8_t * q4 = x[i].qs + 16*ib + 4*il;
const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
for (int j = 0; j < 4; ++j) {
y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
y[j+16] = d * kvalues_iq4nl[q4[j] >> 4];
}
}
template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE);
dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}
static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int64_t k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_Q8_0_NE_ALIGN - 1) / CUDA_Q8_0_NE_ALIGN;
if (k % CUDA_Q8_0_NE_ALIGN == 0) {
const bool need_check = false;
dequantize_block_q8_0_f16<need_check><<<num_blocks, WARP_SIZE, 0, stream>>>(vx, y, k);
} else {
const bool need_check = true;
dequantize_block_q8_0_f16<need_check><<<num_blocks, WARP_SIZE, 0, stream>>>(vx, y, k);
}
}
template<typename dst_t>
static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb32 = k / 32;
const int nb = (k + 255) / 256;
dequantize_block_q4_0<<<nb, 32, 0, stream>>>(vx, y, nb32);
}
template<typename dst_t>
static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb32 = k / 32;
const int nb = (k + 255) / 256;
dequantize_block_q4_1<<<nb, 32, 0, stream>>>(vx, y, nb32);
}
template<typename dst_t>
static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_q4_K<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq2_xxs<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq2_xs<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq2_s<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq3_s<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq1_s<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = (k + QK_K - 1) / QK_K;
dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq1_m<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = (k + QK_K - 1) / QK_K;
dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
}
template <typename src_t, typename dst_t>
static __global__ __launch_bounds__(1024) void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
const src_t * x = (src_t *) vx;
y[i] = x[i];
}
template <typename src_t, typename dst_t>
static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}
to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return dequantize_row_q4_0_cuda;
case GGML_TYPE_Q4_1:
return dequantize_row_q4_1_cuda;
case GGML_TYPE_Q5_0:
return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
case GGML_TYPE_Q5_1:
return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
case GGML_TYPE_Q8_0:
if (ggml_cuda_info().devices[ggml_cuda_get_device()].cc >= CC_PASCAL) {
return dequantize_block_q8_0_f16_cuda;
}
return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
case GGML_TYPE_Q2_K:
return dequantize_row_q2_K_cuda;
case GGML_TYPE_Q3_K:
return dequantize_row_q3_K_cuda;
case GGML_TYPE_Q4_K:
return dequantize_row_q4_K_cuda;
case GGML_TYPE_Q5_K:
return dequantize_row_q5_K_cuda;
case GGML_TYPE_Q6_K:
return dequantize_row_q6_K_cuda;
case GGML_TYPE_IQ2_XXS:
return dequantize_row_iq2_xxs_cuda;
case GGML_TYPE_IQ2_XS:
return dequantize_row_iq2_xs_cuda;
case GGML_TYPE_IQ2_S:
return dequantize_row_iq2_s_cuda;
case GGML_TYPE_IQ3_XXS:
return dequantize_row_iq3_xxs_cuda;
case GGML_TYPE_IQ1_S:
return dequantize_row_iq1_s_cuda;
case GGML_TYPE_IQ1_M:
return dequantize_row_iq1_m_cuda;
case GGML_TYPE_IQ4_NL:
return dequantize_row_iq4_nl_cuda;
case GGML_TYPE_IQ4_XS:
return dequantize_row_iq4_xs_cuda;
case GGML_TYPE_IQ3_S:
return dequantize_row_iq3_s_cuda;
case GGML_TYPE_F32:
return convert_unary_cuda<float>;
default:
return nullptr;
}
}
to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return dequantize_row_q4_0_cuda;
case GGML_TYPE_Q4_1:
return dequantize_row_q4_1_cuda;
case GGML_TYPE_Q5_0:
return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
case GGML_TYPE_Q5_1:
return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
case GGML_TYPE_Q8_0:
return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
case GGML_TYPE_Q2_K:
return dequantize_row_q2_K_cuda;
case GGML_TYPE_Q3_K:
return dequantize_row_q3_K_cuda;
case GGML_TYPE_Q4_K:
return dequantize_row_q4_K_cuda;
case GGML_TYPE_Q5_K:
return dequantize_row_q5_K_cuda;
case GGML_TYPE_Q6_K:
return dequantize_row_q6_K_cuda;
case GGML_TYPE_IQ2_XXS:
return dequantize_row_iq2_xxs_cuda;
case GGML_TYPE_IQ2_XS:
return dequantize_row_iq2_xs_cuda;
case GGML_TYPE_IQ2_S:
return dequantize_row_iq2_s_cuda;
case GGML_TYPE_IQ3_XXS:
return dequantize_row_iq3_xxs_cuda;
case GGML_TYPE_IQ1_S:
return dequantize_row_iq1_s_cuda;
case GGML_TYPE_IQ1_M:
return dequantize_row_iq1_m_cuda;
case GGML_TYPE_IQ4_NL:
return dequantize_row_iq4_nl_cuda;
case GGML_TYPE_IQ4_XS:
return dequantize_row_iq4_xs_cuda;
case GGML_TYPE_IQ3_S:
return dequantize_row_iq3_s_cuda;
case GGML_TYPE_F16:
return convert_unary_cuda<half>;
default:
return nullptr;
}
}
#include "common.cuh"
#define CUDA_DEQUANTIZE_BLOCK_SIZE 256
template<typename T>
using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int64_t k, cudaStream_t stream);
typedef to_t_cuda_t<float> to_fp32_cuda_t;
typedef to_t_cuda_t<half> to_fp16_cuda_t;
to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type);
to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type);
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment