Skip to content

GitLab

Unverified Commit 42998d79 authored by Bruce MacDonald's avatar Bruce MacDonald Committed by GitHub
Browse files

subprocess llama.cpp server (#401) 

* remove c code
* pack llama.cpp
* use request context for llama_cpp
* let llama_cpp decide the number of threads to use
* stop llama runner when app stops
* remove sample count and duration metrics
* use go generate to get libraries
* tmp dir for running llm
parent f4432e1d
Show whitespace changes
Inline Side-by-side
Showing with 23 additions and 33257 deletions
.gitignore
View file @ 42998d79
...@@ -5,4 +5,3 @@ ...@@ -5,4 +5,3 @@
.swp .swp
dist dist
ollama ollama
/ggml-metal.metal
.gitmodules 0 → 100644
View file @ 42998d79
[submodule "llm/llama.cpp/ggml"]
path = llm/llama.cpp/ggml
url = https://github.com/ggerganov/llama.cpp.git
api/types.go
View file @ 42998d79
...@@ -7,7 +7,6 @@ import ( ...@@ -7,7 +7,6 @@ import (
"math" "math"
"os" "os"
"reflect" "reflect"
"runtime"
"strings" "strings"
"time" "time"
) )
...@@ -113,8 +112,6 @@ type GenerateResponse struct { ...@@ -113,8 +112,6 @@ type GenerateResponse struct {
TotalDuration time.Duration `json:"total_duration,omitempty"` TotalDuration time.Duration `json:"total_duration,omitempty"`
LoadDuration time.Duration `json:"load_duration,omitempty"` LoadDuration time.Duration `json:"load_duration,omitempty"`
SampleCount int `json:"sample_count,omitempty"`
SampleDuration time.Duration `json:"sample_duration,omitempty"`
PromptEvalCount int `json:"prompt_eval_count,omitempty"` PromptEvalCount int `json:"prompt_eval_count,omitempty"`
PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"` PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
EvalCount int `json:"eval_count,omitempty"` EvalCount int `json:"eval_count,omitempty"`
...@@ -130,15 +127,6 @@ func (r *GenerateResponse) Summary() { ...@@ -130,15 +127,6 @@ func (r *GenerateResponse) Summary() {
fmt.Fprintf(os.Stderr, "load duration: %v\n", r.LoadDuration) fmt.Fprintf(os.Stderr, "load duration: %v\n", r.LoadDuration)
} }
if r.SampleCount > 0 {
fmt.Fprintf(os.Stderr, "sample count: %d token(s)\n", r.SampleCount)
}
if r.SampleDuration > 0 {
fmt.Fprintf(os.Stderr, "sample duration: %s\n", r.SampleDuration)
fmt.Fprintf(os.Stderr, "sample rate: %.2f tokens/s\n", float64(r.SampleCount)/r.SampleDuration.Seconds())
}
if r.PromptEvalCount > 0 { if r.PromptEvalCount > 0 {
fmt.Fprintf(os.Stderr, "prompt eval count: %d token(s)\n", r.PromptEvalCount) fmt.Fprintf(os.Stderr, "prompt eval count: %d token(s)\n", r.PromptEvalCount)
} }
...@@ -182,15 +170,16 @@ type Options struct { ...@@ -182,15 +170,16 @@ type Options struct {
RopeFrequencyScale float32 `json:"rope_frequency_scale,omitempty"` RopeFrequencyScale float32 `json:"rope_frequency_scale,omitempty"`
// Predict options // Predict options
RepeatLastN int `json:"repeat_last_n,omitempty"` NumPredict int `json:"num_predict,omitempty"`
RepeatPenalty float32 `json:"repeat_penalty,omitempty"`
FrequencyPenalty float32 `json:"frequency_penalty,omitempty"`
PresencePenalty float32 `json:"presence_penalty,omitempty"`
Temperature float32 `json:"temperature,omitempty"`
TopK int `json:"top_k,omitempty"` TopK int `json:"top_k,omitempty"`
TopP float32 `json:"top_p,omitempty"` TopP float32 `json:"top_p,omitempty"`
TFSZ float32 `json:"tfs_z,omitempty"` TFSZ float32 `json:"tfs_z,omitempty"`
TypicalP float32 `json:"typical_p,omitempty"` TypicalP float32 `json:"typical_p,omitempty"`
RepeatLastN int `json:"repeat_last_n,omitempty"`
Temperature float32 `json:"temperature,omitempty"`
RepeatPenalty float32 `json:"repeat_penalty,omitempty"`
PresencePenalty float32 `json:"presence_penalty,omitempty"`
FrequencyPenalty float32 `json:"frequency_penalty,omitempty"`
Mirostat int `json:"mirostat,omitempty"` Mirostat int `json:"mirostat,omitempty"`
MirostatTau float32 `json:"mirostat_tau,omitempty"` MirostatTau float32 `json:"mirostat_tau,omitempty"`
MirostatEta float32 `json:"mirostat_eta,omitempty"` MirostatEta float32 `json:"mirostat_eta,omitempty"`
...@@ -314,7 +303,7 @@ func DefaultOptions() Options { ...@@ -314,7 +303,7 @@ func DefaultOptions() Options {
MirostatEta: 0.1, MirostatEta: 0.1,
PenalizeNewline: true, PenalizeNewline: true,
NumThread: runtime.NumCPU(), NumThread: 0, // let the runtime decide
} }
} }
......
app/src/index.ts
View file @ 42998d79
...@@ -158,7 +158,7 @@ function restart() { ...@@ -158,7 +158,7 @@ function restart() {
app.on('before-quit', () => { app.on('before-quit', () => {
if (proc) { if (proc) {
proc.off('exit', restart) proc.off('exit', restart)
proc.kill() proc.kill('SIGINT') // send SIGINT signal to the server, which also stops any loaded llms
} }
}) })
......
docs/development.md
View file @ 42998d79
# Development # Development
- Install cmake or (optionally, required tools for GPUs)
- run `go generate ./...`
- run `go build .`
Install required tools: Install required tools:
``` ```
brew install go brew install go cmake gcc
``` ```
Enable CGO: Get the required libraries:
``` ```
export CGO_ENABLED=1 go generate ./...
``` ```
You will also need a C/C++ compiler such as GCC for MacOS and Linux or Mingw-w64 GCC for Windows.
Then build ollama: Then build ollama:
``` ```
......
go.mod
View file @ 42998d79
...@@ -38,9 +38,9 @@ require ( ...@@ -38,9 +38,9 @@ require (
github.com/twitchyliquid64/golang-asm v0.15.1 // indirect github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
github.com/ugorji/go/codec v1.2.11 // indirect github.com/ugorji/go/codec v1.2.11 // indirect
golang.org/x/arch v0.3.0 // indirect golang.org/x/arch v0.3.0 // indirect
golang.org/x/crypto v0.10.0 // indirect golang.org/x/crypto v0.10.0
golang.org/x/net v0.10.0 // indirect golang.org/x/net v0.10.0 // indirect
golang.org/x/sys v0.10.0 // indirect golang.org/x/sys v0.11.0 // indirect
golang.org/x/term v0.10.0 golang.org/x/term v0.10.0
golang.org/x/text v0.10.0 // indirect golang.org/x/text v0.10.0 // indirect
gonum.org/v1/gonum v0.13.0 gonum.org/v1/gonum v0.13.0
......
go.sum
View file @ 42998d79
...@@ -120,6 +120,7 @@ golang.org/x/arch v0.3.0/go.mod h1:5om86z9Hs0C8fWVUuoMHwpExlXzs5Tkyp9hOrfG7pp8= ...@@ -120,6 +120,7 @@ golang.org/x/arch v0.3.0/go.mod h1:5om86z9Hs0C8fWVUuoMHwpExlXzs5Tkyp9hOrfG7pp8=
golang.org/x/crypto v0.0.0-20210711020723-a769d52b0f97/go.mod h1:GvvjBRRGRdwPK5ydBHafDWAxML/pGHZbMvKqRZ5+Abc= golang.org/x/crypto v0.0.0-20210711020723-a769d52b0f97/go.mod h1:GvvjBRRGRdwPK5ydBHafDWAxML/pGHZbMvKqRZ5+Abc=
golang.org/x/crypto v0.10.0 h1:LKqV2xt9+kDzSTfOhx4FrkEBcMrAgHSYgzywV9zcGmM= golang.org/x/crypto v0.10.0 h1:LKqV2xt9+kDzSTfOhx4FrkEBcMrAgHSYgzywV9zcGmM=
golang.org/x/crypto v0.10.0/go.mod h1:o4eNf7Ede1fv+hwOwZsTHl9EsPFO6q6ZvYR8vYfY45I= golang.org/x/crypto v0.10.0/go.mod h1:o4eNf7Ede1fv+hwOwZsTHl9EsPFO6q6ZvYR8vYfY45I=
golang.org/x/exp v0.0.0-20230321023759-10a507213a29 h1:ooxPy7fPvB4kwsA2h+iBNHkAbp/4JxTSwCmvdjEYmug=
golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg= golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg=
golang.org/x/net v0.10.0 h1:X2//UzNDwYmtCLn7To6G58Wr6f5ahEAQgKNzv9Y951M= golang.org/x/net v0.10.0 h1:X2//UzNDwYmtCLn7To6G58Wr6f5ahEAQgKNzv9Y951M=
golang.org/x/net v0.10.0/go.mod h1:0qNGK6F8kojg2nk9dLZ2mShWaEBan6FAoqfSigmmuDg= golang.org/x/net v0.10.0/go.mod h1:0qNGK6F8kojg2nk9dLZ2mShWaEBan6FAoqfSigmmuDg=
...@@ -130,8 +131,8 @@ golang.org/x/sys v0.0.0-20210806184541-e5e7981a1069/go.mod h1:oPkhp1MJrh7nUepCBc ...@@ -130,8 +131,8 @@ golang.org/x/sys v0.0.0-20210806184541-e5e7981a1069/go.mod h1:oPkhp1MJrh7nUepCBc
golang.org/x/sys v0.0.0-20220310020820-b874c991c1a5/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg= golang.org/x/sys v0.0.0-20220310020820-b874c991c1a5/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
golang.org/x/sys v0.0.0-20220704084225-05e143d24a9e/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg= golang.org/x/sys v0.0.0-20220704084225-05e143d24a9e/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg= golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
golang.org/x/sys v0.10.0 h1:SqMFp9UcQJZa+pmYuAKjd9xq1f0j5rLcDIk0mj4qAsA= golang.org/x/sys v0.11.0 h1:eG7RXZHdqOJ1i+0lgLgCpSXAp6M3LYlAo6osgSi0xOM=
golang.org/x/sys v0.10.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg= golang.org/x/sys v0.11.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo= golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
golang.org/x/term v0.10.0 h1:3R7pNqamzBraeqj/Tj8qt1aQ2HpmlC+Cx/qL/7hn4/c= golang.org/x/term v0.10.0 h1:3R7pNqamzBraeqj/Tj8qt1aQ2HpmlC+Cx/qL/7hn4/c=
golang.org/x/term v0.10.0/go.mod h1:lpqdcUyK/oCiQxvxVrppt5ggO2KCZ5QblwqPnfZ6d5o= golang.org/x/term v0.10.0/go.mod h1:lpqdcUyK/oCiQxvxVrppt5ggO2KCZ5QblwqPnfZ6d5o=
......
llm/ggml-alloc.c deleted 100644 → 0
View file @ f4432e1d
/**
* llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
*
* MIT License
*
* Copyright (c) 2023 Georgi Gerganov
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "ggml-alloc.h"
#include "ggml.h"
#include <assert.h>
#include <stdarg.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#define UNUSED(x) (void)(x)
#define MAX(a, b) ((a) > (b) ? (a) : (b))
//#define GGML_ALLOCATOR_DEBUG
//#define AT_PRINTF printf
#define AT_PRINTF(...) ((void)0)
struct hash_node {
struct ggml_tensor * t;
int n_children;
int n_views;
};
static size_t hash(void * p) {
return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
}
static struct hash_node * hash_get(struct hash_node hash_table[], struct ggml_tensor * t) {
size_t h = hash(t);
// linear probing
size_t i = h;
while (hash_table[i].t != NULL) {
if (hash_table[i].t == t) {
return &hash_table[i];
}
i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
if (i == h) {
// hash table is full
GGML_ASSERT(false);
}
}
hash_table[i].t = t;
return &hash_table[i];
}
// TODO: GGML_PAD ?
static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
assert(alignment && !(alignment & (alignment - 1))); // power of 2
size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
return offset + align;
}
struct free_block {
void * addr;
size_t size;
};
#define MAX_FREE_BLOCKS 128
struct ggml_allocr {
void * data;
size_t size;
size_t alignment;
int n_free_blocks;
struct free_block free_blocks[MAX_FREE_BLOCKS];
struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE];
size_t max_size;
bool measure;
#ifdef GGML_ALLOCATOR_DEBUG
struct ggml_tensor * allocated_tensors[1024];
#endif
};
#ifdef GGML_ALLOCATOR_DEBUG
static void add_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
for (int i = 0; i < 1024; i++) {
if (alloc->allocated_tensors[i] == NULL) {
alloc->allocated_tensors[i] = tensor;
return;
}
}
GGML_ASSERT(!"out of allocated_tensors");
}
static void remove_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
for (int i = 0; i < 1024; i++) {
if (alloc->allocated_tensors[i] == tensor ||
(alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) {
alloc->allocated_tensors[i] = NULL;
return;
}
}
printf("tried to free tensor %s not found\n", tensor->name);
GGML_ASSERT(!"tensor not found");
}
#endif
static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
return ggml_nbytes(tensor);
UNUSED(alloc);
}
void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
size = aligned_offset(NULL, size, alloc->alignment);
AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
size_t max_avail = 0;
// find the best fitting free block
int best_fit_block = -1;
size_t best_fit_size = SIZE_MAX;
for (int i = 0; i < alloc->n_free_blocks; i++) {
struct free_block * block = &alloc->free_blocks[i];
max_avail = MAX(max_avail, block->size);
if (block->size >= size && block->size <= best_fit_size) {
best_fit_block = i;
best_fit_size = block->size;
}
}
AT_PRINTF("block %d\n", best_fit_block);
if (best_fit_block == -1) {
fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
__func__, size, max_avail);
GGML_ASSERT(!"not enough space in the buffer");
return;
}
struct free_block * block = &alloc->free_blocks[best_fit_block];
void * addr = block->addr;
block->addr = (char*)block->addr + size;
block->size -= size;
if (block->size == 0) {
// remove block if empty
alloc->n_free_blocks--;
for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
alloc->free_blocks[j] = alloc->free_blocks[j+1];
}
}
tensor->data = addr;
#ifdef GGML_ALLOCATOR_DEBUG
add_allocated_tensor(alloc, tensor);
size_t cur_max = (char*)addr - (char*)alloc->data + size;
if (cur_max > alloc->max_size) {
printf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
for (int i = 0; i < 1024; i++) {
if (alloc->allocated_tensors[i]) {
printf("%s (%.2f MB) ", alloc->allocated_tensors[i]->name, ggml_nbytes(alloc->allocated_tensors[i]) / 1024.0 / 1024.0);
}
}
printf("\n");
}
#endif
alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->data + size);
}
// this is a very naive implementation, but for our case the number of free blocks should be very small
static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
void * ptr = tensor->data;
if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
// the tensor was not allocated in this buffer
// this can happen because the graph allocator will try to free weights and other tensors from different buffers
// the easiest way to deal with this is just to ignore it
return;
}
size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
size = aligned_offset(NULL, size, alloc->alignment);
AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
#ifdef GGML_ALLOCATOR_DEBUG
remove_allocated_tensor(alloc, tensor);
#endif
// see if we can merge with an existing block
for (int i = 0; i < alloc->n_free_blocks; i++) {
struct free_block * block = &alloc->free_blocks[i];
// check if ptr is at the end of the block
if ((char*)block->addr + block->size == ptr) {
block->size += size;
// check if we can merge with the next block
if (i < alloc->n_free_blocks - 1 && (char*)block->addr + block->size == alloc->free_blocks[i+1].addr) {
block->size += alloc->free_blocks[i+1].size;
alloc->n_free_blocks--;
for (int j = i+1; j < alloc->n_free_blocks; j++) {
alloc->free_blocks[j] = alloc->free_blocks[j+1];
}
}
return;
}
// check if ptr is at the beginning of the block
if ((char*)ptr + size == block->addr) {
block->addr = ptr;
block->size += size;
// check if we can merge with the previous block
if (i > 0 && (char*)alloc->free_blocks[i-1].addr + alloc->free_blocks[i-1].size == block->addr) {
alloc->free_blocks[i-1].size += block->size;
alloc->n_free_blocks--;
for (int j = i; j < alloc->n_free_blocks; j++) {
alloc->free_blocks[j] = alloc->free_blocks[j+1];
}
}
return;
}
}
// otherwise, add a new block
GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
// insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
int insert_pos = 0;
while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].addr < ptr) {
insert_pos++;
}
// shift all blocks from insert_pos onward to make room for the new block
for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
alloc->free_blocks[i] = alloc->free_blocks[i-1];
}
// insert the new block
alloc->free_blocks[insert_pos].addr = ptr;
alloc->free_blocks[insert_pos].size = size;
alloc->n_free_blocks++;
}
void ggml_allocr_reset(struct ggml_allocr * alloc) {
alloc->n_free_blocks = 1;
size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
alloc->free_blocks[0].addr = (char *)alloc->data + align_offset;
alloc->free_blocks[0].size = alloc->size - align_offset;
}
struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) {
struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
*alloc = (struct ggml_allocr){
/*.data = */ data,
/*.size = */ size,
/*.alignment = */ alignment,
/*.n_free_blocks = */ 0,
/*.free_blocks = */ {{0}},
/*.hash_table = */ {{0}},
/*.max_size = */ 0,
/*.measure = */ false,
#ifdef GGML_ALLOCATOR_DEBUG
/*.allocated_tensors = */ = {0},
#endif
};
ggml_allocr_reset(alloc);
return alloc;
}
// address and size of the buffer when measuring
// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
static const size_t MEASURE_MAX_SIZE = 1ULL<<40; // 1 TB
struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
*alloc = (struct ggml_allocr){
/*.data = */ MEASURE_BASE_ADDR,
/*.size = */ MEASURE_MAX_SIZE,
/*.alignment = */ alignment,
/*.n_free_blocks = */ 0,
/*.free_blocks = */ {{0}},
/*.hash_table = */ {{0}},
/*.max_size = */ 0,
/*.measure = */ true,
#ifdef GGML_ALLOCATOR_DEBUG
/*.allocated_tensors = */ = {0},
#endif
};
ggml_allocr_reset(alloc);
return alloc;
}
void ggml_allocr_free(struct ggml_allocr * alloc) {
free(alloc);
}
bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
return alloc->measure;
}
//////////// compute graph allocator
static bool ggml_is_view(struct ggml_tensor * t) {
return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
}
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;
}
static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
switch (t->op) {
case GGML_OP_PERMUTE:
case GGML_OP_RESHAPE:
case GGML_OP_TRANSPOSE:
case GGML_OP_VIEW:
return t->src[0];
case GGML_OP_CPY:
return t->src[1];
default:
return NULL;
}
}
static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
struct ggml_tensor * parent = t;
do {
parent = get_view_parent(parent);
} while (ggml_is_view(parent));
return parent;
}
static bool ggml_op_can_inplace(enum ggml_op op) {
switch (op) {
case GGML_OP_SCALE:
case GGML_OP_DIAG_MASK_ZERO:
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_ADD:
case GGML_OP_ADD1:
case GGML_OP_ACC:
case GGML_OP_SUB:
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_SQR:
case GGML_OP_SQRT:
case GGML_OP_LOG:
case GGML_OP_UNARY:
case GGML_OP_ROPE:
case GGML_OP_RMS_NORM:
case GGML_OP_SET:
case GGML_OP_SOFT_MAX:
case GGML_OP_CONT:
return true;
default:
return false;
}
}
static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) {
struct hash_node * ht = alloc->hash_table;
if (node->data == NULL) {
if (ggml_is_view(node)) {
size_t offset;
switch(node->op) {
case GGML_OP_VIEW:
memcpy(&offset, node->op_params, sizeof(size_t));
node->data = (char *) node->src[0]->data + offset;
break;
case GGML_OP_PERMUTE:
case GGML_OP_RESHAPE:
case GGML_OP_TRANSPOSE:
node->data = node->src[0]->data;
break;
case GGML_OP_CPY:
node->data = node->src[1]->data;
break;
default:
GGML_ASSERT(!"unknown view op");
break;
}
} else {
// see if we can reuse a parent's buffer (inplace)
if (ggml_op_can_inplace(node->op)) {
for (int i = 0; i < GGML_MAX_SRC; i++) {
struct ggml_tensor * parent = node->src[i];
if (parent == NULL) {
break;
}
// if the node's data is external, then we cannot re-use it
if ((char *) parent->data < (char *) alloc->data ||
(char *) parent->data >= ((char *) alloc->data + alloc->size)) {
AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
continue;
}
struct hash_node * p_hn = hash_get(ht, parent);
if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
if (ggml_is_view(parent)) {
struct ggml_tensor * view_src = get_view_source(parent);
struct hash_node * view_src_hn = hash_get(ht, view_src);
if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
// TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
// the parent's data that it will need later (same layout requirement). the problem is that then
// we cannot free the tensor because the original address of the allocation is lost.
// adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
// for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
node->data = parent->data;
return;
}
}
else {
AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
node->data = parent->data;
}
return;
}
}
}
ggml_allocr_alloc(alloc, node);
}
}
}
static size_t ggml_allocator_alloc_graph_tensors_n(
struct ggml_allocr * alloc,
struct ggml_cgraph ** graphs, int n_graphs,
struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
// reset hash table
struct hash_node * ht = alloc->hash_table;
memset(ht, 0, sizeof(struct hash_node) * GGML_GRAPH_HASHTABLE_SIZE);
// count number of children and views
for (int g = 0; g < n_graphs; g++) {
struct ggml_cgraph * gf = graphs[g];
for (int i = 0; i < gf->n_nodes; i++) {
struct ggml_tensor * node = gf->nodes[i];
if (ggml_is_view(node)) {
struct ggml_tensor * view_src = get_view_source(node);
hash_get(ht, view_src)->n_views += 1;
}
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * parent = node->src[j];
if (parent == NULL) {
break;
}
hash_get(ht, parent)->n_children += 1;
}
}
}
// allocate tensors
for (int g = 0; g < n_graphs; g++) {
struct ggml_cgraph * gf = graphs[g];
AT_PRINTF("####### graph %d/%d\n", g, n_graphs);
// graph inputs are allocated first to ensure that they are not overwritten by each other
if (inputs != NULL && inputs[g] != NULL) {
for (int i = 0; inputs[g][i] != NULL; i++) {
struct ggml_tensor * input = inputs[g][i];
AT_PRINTF("input: %s\n", input->name);
allocate_node(alloc, input);
}
}
for (int i = 0; i < gf->n_nodes; i++) {
struct ggml_tensor * node = gf->nodes[i];
// allocate parents (leafs)
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * parent = node->src[j];
if (parent == NULL) {
break;
}
allocate_node(alloc, parent);
}
// allocate node
allocate_node(alloc, node);
AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * parent = node->src[j];
if (parent == NULL) {
break;
}
AT_PRINTF("%s", parent->name);
if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
AT_PRINTF(", ");
}
}
AT_PRINTF("\n");
// update parents
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * parent = node->src[j];
if (parent == NULL) {
break;
}
struct hash_node * p_hn = hash_get(ht, parent);
p_hn->n_children -= 1;
//AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
if (p_hn->n_children == 0 && p_hn->n_views == 0) {
if (ggml_is_view(parent)) {
struct ggml_tensor * view_src = get_view_source(parent);
struct hash_node * view_src_hn = hash_get(ht, view_src);
view_src_hn->n_views -= 1;
AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src->n_children, view_src->n_views);
if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
ggml_allocator_free_tensor(alloc, view_src);
}
}
else {
if (parent->data != node->data) {
ggml_allocator_free_tensor(alloc, parent);
}
}
}
}
AT_PRINTF("\n");
}
// free graph outputs here that wouldn't be freed otherwise because they have no children
if (outputs != NULL && outputs[g] != NULL) {
for (int i = 0; outputs[g][i] != NULL; i++) {
struct ggml_tensor * output = outputs[g][i];
AT_PRINTF("output: %s\n", output->name);
ggml_allocator_free_tensor(alloc, output);
}
}
}
return alloc->max_size;
}
size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
}
llm/ggml-alloc.h deleted 100644 → 0
View file @ f4432e1d
/**
* llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
*
* MIT License
*
* Copyright (c) 2023 Georgi Gerganov
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#pragma once
#include "ggml.h"
#ifdef __cplusplus
extern "C" {
#endif
GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);
GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
GGML_API void ggml_allocr_free(struct ggml_allocr * alloc);
GGML_API bool ggml_allocr_is_measure(struct ggml_allocr * alloc);
GGML_API void ggml_allocr_reset(struct ggml_allocr * alloc);
GGML_API void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
#ifdef __cplusplus
}
#endif
llm/ggml-cuda.cu deleted 100644 → 0
View file @ f4432e1d
This source diff could not be displayed because it is too large. You can view the blob instead.
llm/ggml-cuda.h deleted 100644 → 0
View file @ f4432e1d
/**
* llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
*
* MIT License
*
* Copyright (c) 2023 Georgi Gerganov
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#pragma once
#include "ggml.h"
#ifdef __cplusplus
extern "C" {
#endif
#define GGML_CUDA_MAX_DEVICES 16
void ggml_init_cublas(void);
void ggml_cuda_set_tensor_split(const float * tensor_split);
void ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
void ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
// TODO: export these with GGML_API
void * ggml_cuda_host_malloc(size_t size);
void ggml_cuda_host_free(void * ptr);
void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
void ggml_cuda_free_data(struct ggml_tensor * tensor);
void ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
void ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
void ggml_cuda_set_main_device(int main_device);
void ggml_cuda_set_mul_mat_q(bool mul_mat_q);
void ggml_cuda_set_scratch_size(size_t scratch_size);
void ggml_cuda_free_scratch(void);
bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
#ifdef __cplusplus
}
#endif
llm/ggml-metal.h deleted 100644 → 0
View file @ f4432e1d
//go:build darwin
/**
* llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
*
* MIT License
*
* Copyright (c) 2023 Georgi Gerganov
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
// An interface allowing to compute ggml_cgraph with Metal
//
// This is a fully functional interface that extends ggml with GPU support for Apple devices.
// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
//
// How it works?
//
// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
//
// You only need to make sure that all memory buffers that you used during the graph creation
// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
// used during the graph evaluation to determine the arguments of the compute kernels.
//
// Synchronization between device and host memory (for example for input and output tensors)
// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
//
#pragma once
#include <stddef.h>
#include <stdbool.h>
// max memory buffers that can be mapped to the device
#define GGML_METAL_MAX_BUFFERS 16
struct ggml_tensor;
struct ggml_cgraph;
#ifdef __cplusplus
extern "C" {
#endif
struct ggml_metal_context;
// number of command buffers to use
struct ggml_metal_context * ggml_metal_init(int n_cb);
void ggml_metal_free(struct ggml_metal_context * ctx);
// set the number of command buffers to use
void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
// creates a mapping between a host memory buffer and a device memory buffer
// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
// - the mapping is used during computation to determine the arguments of the compute kernels
// - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
// - max_size specifies the maximum size of a tensor and is used to create shared views such
// that it is guaranteed that the tensor will fit in at least one of the views
//
bool ggml_metal_add_buffer(
struct ggml_metal_context * ctx,
const char * name,
void * data,
size_t size,
size_t max_size);
// set data from host memory into the device
void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
// get data from the device into host memory
void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
// try to find operations that can be run concurrently in the graph
// you should run it again if the topology of your graph changes
void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
// if the graph has been optimized for concurrently dispatch
bool ggml_metal_if_optimized(struct ggml_metal_context * ctx);
// same as ggml_graph_compute but uses Metal
// creates gf->n_threads command buffers in parallel
void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
#ifdef __cplusplus
}
#endif
llm/ggml-metal.m deleted 100644 → 0
View file @ f4432e1d
//go:build darwin
/**
* llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
*
* MIT License
*
* Copyright (c) 2023 Georgi Gerganov
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#import "ggml-metal.h"
#import "ggml.h"
#import <Foundation/Foundation.h>
#import <Metal/Metal.h>
#import <MetalPerformanceShaders/MetalPerformanceShaders.h>
#undef MIN
#undef MAX
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#ifdef GGML_METAL_NDEBUG
#define metal_printf(...)
#else
#define metal_printf(...) fprintf(stderr, __VA_ARGS__)
#endif
#define UNUSED(x) (void)(x)
#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
struct ggml_metal_buffer {
const char * name;
void * data;
size_t size;
id<MTLBuffer> metal;
};
struct ggml_metal_context {
int n_cb;
float * logits;
id<MTLDevice> device;
id<MTLCommandQueue> queue;
id<MTLLibrary> library;
int n_buffers;
struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
int concur_list[GGML_MAX_CONCUR];
int concur_list_len;
// custom kernels
#define GGML_METAL_DECL_KERNEL(name) \
id<MTLFunction> function_##name; \
id<MTLComputePipelineState> pipeline_##name
GGML_METAL_DECL_KERNEL(add);
GGML_METAL_DECL_KERNEL(add_row); // TODO: avoid this extra kernel, instead extend the "add" kernel to support broadcast
GGML_METAL_DECL_KERNEL(mul);
GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
GGML_METAL_DECL_KERNEL(scale);
GGML_METAL_DECL_KERNEL(silu);
GGML_METAL_DECL_KERNEL(relu);
GGML_METAL_DECL_KERNEL(gelu);
GGML_METAL_DECL_KERNEL(soft_max);
GGML_METAL_DECL_KERNEL(diag_mask_inf);
GGML_METAL_DECL_KERNEL(get_rows_f16);
GGML_METAL_DECL_KERNEL(get_rows_q4_0);
GGML_METAL_DECL_KERNEL(get_rows_q4_1);
GGML_METAL_DECL_KERNEL(get_rows_q2_K);
GGML_METAL_DECL_KERNEL(get_rows_q3_K);
GGML_METAL_DECL_KERNEL(get_rows_q4_K);
GGML_METAL_DECL_KERNEL(get_rows_q5_K);
GGML_METAL_DECL_KERNEL(get_rows_q6_K);
GGML_METAL_DECL_KERNEL(rms_norm);
GGML_METAL_DECL_KERNEL(norm);
GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q2_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q3_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_DECL_KERNEL(rope);
GGML_METAL_DECL_KERNEL(alibi_f32);
GGML_METAL_DECL_KERNEL(cpy_f32_f16);
GGML_METAL_DECL_KERNEL(cpy_f32_f32);
GGML_METAL_DECL_KERNEL(cpy_f16_f16);
#undef GGML_METAL_DECL_KERNEL
};
// MSL code
// TODO: move the contents here when ready
// for now it is easier to work in a separate file
static NSString * const msl_library_source = @"see metal.metal";
// Here to assist with NSBundle Path Hack
@interface GGMLMetalClass : NSObject
@end
@implementation GGMLMetalClass
@end
struct ggml_metal_context * ggml_metal_init(int n_cb) {
fprintf(stderr, "%s: allocating\n", __func__);
struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
ctx->n_cb = n_cb;
ctx->device = MTLCreateSystemDefaultDevice();
ctx->queue = [ctx->device newCommandQueue];
ctx->n_buffers = 0;
ctx->concur_list_len = 0;
// determine if we can use MPS
if (MPSSupportsMTLDevice(ctx->device)) {
fprintf(stderr, "%s: using MPS\n", __func__);
} else {
fprintf(stderr, "%s: not using MPS\n", __func__);
GGML_ASSERT(false && "MPS not supported");
}
#if 0
// compile from source string and show compile log
{
NSError * error = nil;
ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
if (error) {
fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
}
#else
UNUSED(msl_library_source);
// read the source from "ggml-metal.metal" into a string and use newLibraryWithSource
{
NSError * error = nil;
//NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
fprintf(stderr, "%s: loading '%s'\n", __func__, [path UTF8String]);
NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
if (error) {
fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
#ifdef GGML_QKK_64
MTLCompileOptions* options = [MTLCompileOptions new];
options.preprocessorMacros = @{ @"QK_K" : @(64) };
ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
#else
ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error];
#endif
if (error) {
fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
}
#endif
// load kernels
{
#define GGML_METAL_ADD_KERNEL(name) \
ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:nil]; \
fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
GGML_METAL_ADD_KERNEL(add);
GGML_METAL_ADD_KERNEL(add_row);
GGML_METAL_ADD_KERNEL(mul);
GGML_METAL_ADD_KERNEL(mul_row);
GGML_METAL_ADD_KERNEL(scale);
GGML_METAL_ADD_KERNEL(silu);
GGML_METAL_ADD_KERNEL(relu);
GGML_METAL_ADD_KERNEL(gelu);
GGML_METAL_ADD_KERNEL(soft_max);
GGML_METAL_ADD_KERNEL(diag_mask_inf);
GGML_METAL_ADD_KERNEL(get_rows_f16);
GGML_METAL_ADD_KERNEL(get_rows_q4_0);
GGML_METAL_ADD_KERNEL(get_rows_q4_1);
GGML_METAL_ADD_KERNEL(get_rows_q2_K);
GGML_METAL_ADD_KERNEL(get_rows_q3_K);
GGML_METAL_ADD_KERNEL(get_rows_q4_K);
GGML_METAL_ADD_KERNEL(get_rows_q5_K);
GGML_METAL_ADD_KERNEL(get_rows_q6_K);
GGML_METAL_ADD_KERNEL(rms_norm);
GGML_METAL_ADD_KERNEL(norm);
GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q2_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q3_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_ADD_KERNEL(rope);
GGML_METAL_ADD_KERNEL(alibi_f32);
GGML_METAL_ADD_KERNEL(cpy_f32_f16);
GGML_METAL_ADD_KERNEL(cpy_f32_f32);
GGML_METAL_ADD_KERNEL(cpy_f16_f16);
#undef GGML_METAL_ADD_KERNEL
}
fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
fprintf(stderr, "%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
if (ctx->device.maxTransferRate != 0) {
fprintf(stderr, "%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
} else {
fprintf(stderr, "%s: maxTransferRate = built-in GPU\n", __func__);
}
return ctx;
}
void ggml_metal_free(struct ggml_metal_context * ctx) {
fprintf(stderr, "%s: deallocating\n", __func__);
for (int i = 0; i < ctx->n_buffers; ++i) {
[ctx->buffers[i].metal release];
}
free(ctx);
}
void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb) {
ctx->n_cb = n_cb;
}
bool ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
if (ctx->concur_list_len) {
return true;
}
return false;
}
// finds the Metal buffer that contains the tensor data on the GPU device
// the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
// Metal buffer based on the host memory pointer
//
static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
//fprintf(stderr, "%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
const int64_t tsize = ggml_nbytes(t);
// find the view that contains the tensor fully
for (int i = 0; i < ctx->n_buffers; ++i) {
const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
*offs = (size_t) ioffs;
//fprintf(stderr, "%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
return ctx->buffers[i].metal;
}
}
fprintf(stderr, "%s: error: buffer is nil\n", __func__);
return nil;
}
bool ggml_metal_add_buffer(
struct ggml_metal_context * ctx,
const char * name,
void * data,
size_t size,
size_t max_size) {
if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
fprintf(stderr, "%s: too many buffers\n", __func__);
return false;
}
if (data) {
// verify that the buffer does not overlap with any of the existing buffers
for (int i = 0; i < ctx->n_buffers; ++i) {
const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
fprintf(stderr, "%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
return false;
}
}
const size_t size_page = getpagesize();
size_t size_aligned = size;
if ((size_aligned % size_page) != 0) {
size_aligned += (size_page - (size_aligned % size_page));
}
// the buffer fits into the max buffer size allowed by the device
if (size_aligned <= ctx->device.maxBufferLength) {
ctx->buffers[ctx->n_buffers].name = name;
ctx->buffers[ctx->n_buffers].data = data;
ctx->buffers[ctx->n_buffers].size = size;
ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
if (ctx->buffers[ctx->n_buffers].metal == nil) {
fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
return false;
}
fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
++ctx->n_buffers;
} else {
// this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
// one of the views
const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
const size_t size_step = ctx->device.maxBufferLength - size_ovlp;
const size_t size_view = ctx->device.maxBufferLength;
for (size_t i = 0; i < size; i += size_step) {
const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
ctx->buffers[ctx->n_buffers].name = name;
ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
ctx->buffers[ctx->n_buffers].size = size_step_aligned;
ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
if (ctx->buffers[ctx->n_buffers].metal == nil) {
fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
return false;
}
fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
if (i + size_step < size) {
fprintf(stderr, "\n");
}
++ctx->n_buffers;
}
}
fprintf(stderr, ", (%8.2f / %8.2f)",
ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
fprintf(stderr, ", warning: current allocated size is greater than the recommended max working set size\n");
} else {
fprintf(stderr, "\n");
}
}
return true;
}
void ggml_metal_set_tensor(
struct ggml_metal_context * ctx,
struct ggml_tensor * t) {
metal_printf("%s: set input for tensor '%s'\n", __func__, t->name);
size_t offs;
id<MTLBuffer> id_dst = ggml_metal_get_buffer(ctx, t, &offs);
memcpy((void *) ((uint8_t *) id_dst.contents + offs), t->data, ggml_nbytes(t));
}
void ggml_metal_get_tensor(
struct ggml_metal_context * ctx,
struct ggml_tensor * t) {
metal_printf("%s: extract results for tensor '%s'\n", __func__, t->name);
size_t offs;
id<MTLBuffer> id_src = ggml_metal_get_buffer(ctx, t, &offs);
memcpy(t->data, (void *) ((uint8_t *) id_src.contents + offs), ggml_nbytes(t));
}
void ggml_metal_graph_find_concurrency(
struct ggml_metal_context * ctx,
struct ggml_cgraph * gf) {
int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time
int nodes_unused[GGML_MAX_CONCUR];
for (int i = 0; i < GGML_MAX_CONCUR; i++) { ctx->concur_list[i] = 0; }
for (int i = 0; i < gf->n_nodes; i++) { nodes_unused[i] = 1; }
ctx->concur_list_len = 0;
int n_left = gf->n_nodes;
int n_start = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
int level_pos = 0; // at ctx->concur_list, the last layer (level) ends at level_pos
while (n_left > 0) {
// number of nodes at a layer (that can be issued concurrently)
int concurrency = 0;
for (int i = n_start; i < ((n_start + search_depth > gf->n_nodes) ? gf->n_nodes : n_start + search_depth); i++) {
if (nodes_unused[i]) {
// if the requirements for gf->nodes[i] are satisfied
int exe_flag = 1;
// scan all srcs
for (int src_ind = 0; src_ind < GGML_MAX_SRC; src_ind++) {
struct ggml_tensor * src_cur = gf->nodes[i]->src[src_ind];
if (src_cur) {
// if is leaf nodes it's satisfied.
// TODO: ggml_is_leaf()
if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {
continue;
}
// otherwise this src should be the output from previous nodes.
int is_found = 0;
// scan 2*search_depth back because we inserted barrier.
//for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
for (int j = MAX(0, level_pos - 2*search_depth); j < level_pos; j++) {
if (ctx->concur_list[j] >= 0 && gf->nodes[ctx->concur_list[j]] == src_cur) {
is_found = 1;
break;
}
}
if (is_found == 0) {
exe_flag = 0;
break;
}
}
}
if (exe_flag) {
// check if nodes[i]'s data will be overwritten by a node before nodes[i].
// if node[5] and node[3] write to the same memory region, then we can't issue node[5] before node[3]
int64_t data_start = (int64_t) gf->nodes[i]->data;
int64_t length = (int64_t) ggml_nbytes(gf->nodes[i]);
for (int j = n_start; j < i; j++) {
if (nodes_unused[j] && gf->nodes[j]->op != GGML_OP_RESHAPE \
&& gf->nodes[j]->op != GGML_OP_VIEW \
&& gf->nodes[j]->op != GGML_OP_TRANSPOSE \
&& gf->nodes[j]->op != GGML_OP_PERMUTE) {
if (((int64_t)gf->nodes[j]->data) >= data_start + length || \
((int64_t)gf->nodes[j]->data) + (int64_t) ggml_nbytes(gf->nodes[j]) <= data_start) {
continue;
}
exe_flag = 0;
}
}
}
if (exe_flag) {
ctx->concur_list[level_pos + concurrency] = i;
nodes_unused[i] = 0;
concurrency++;
ctx->concur_list_len++;
}
}
}
n_left -= concurrency;
// adding a barrier different layer
ctx->concur_list[level_pos + concurrency] = -1;
ctx->concur_list_len++;
// jump all sorted nodes at nodes_bak
while (!nodes_unused[n_start]) {
n_start++;
}
level_pos += concurrency + 1;
}
if (ctx->concur_list_len > GGML_MAX_CONCUR) {
fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
}
}
void ggml_metal_graph_compute(
struct ggml_metal_context * ctx,
struct ggml_cgraph * gf) {
metal_printf("%s: evaluating graph\n", __func__);
// if there is ctx->concur_list, dispatch concurrently
// else fallback to serial dispatch
MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_CONCUR;
const int n_nodes = has_concur ? ctx->concur_list_len : gf->n_nodes;
edesc.dispatchType = has_concur ? MTLDispatchTypeConcurrent : MTLDispatchTypeSerial;
// create multiple command buffers and enqueue them
// then, we encode the graph into the command buffers in parallel
const int n_cb = ctx->n_cb;
NSMutableArray * command_buffers = [NSMutableArray arrayWithCapacity:n_cb];
for (int i = 0; i < n_cb; ++i) {
command_buffers[i] = [ctx->queue commandBuffer];
// enqueue the command buffers in order to specify their execution order
[command_buffers[i] enqueue];
}
// TODO: is this the best way to start threads?
dispatch_queue_t queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb;
dispatch_async(queue, ^{
size_t offs_src0 = 0;
size_t offs_src1 = 0;
size_t offs_dst = 0;
id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
id<MTLComputeCommandEncoder> encoder = nil;
const int node_start = (cb_idx + 0) * n_nodes_per_cb;
const int node_end = (cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb;
for (int ind = node_start; ind < node_end; ++ind) {
const int i = has_concur ? ctx->concur_list[ind] : ind;
if (i == -1) {
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
continue;
}
[encoder memoryBarrierWithScope:MTLBarrierScopeBuffers];
continue;
}
metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
struct ggml_tensor * src0 = gf->nodes[i]->src[0];
struct ggml_tensor * src1 = gf->nodes[i]->src[1];
struct ggml_tensor * dst = gf->nodes[i];
const int64_t ne00 = src0 ? src0->ne[0] : 0;
const int64_t ne01 = src0 ? src0->ne[1] : 0;
const int64_t ne02 = src0 ? src0->ne[2] : 0;
const int64_t ne03 = src0 ? src0->ne[3] : 0;
const uint64_t nb00 = src0 ? src0->nb[0] : 0;
const uint64_t nb01 = src0 ? src0->nb[1] : 0;
const uint64_t nb02 = src0 ? src0->nb[2] : 0;
const uint64_t nb03 = src0 ? src0->nb[3] : 0;
const int64_t ne10 = src1 ? src1->ne[0] : 0;
const int64_t ne11 = src1 ? src1->ne[1] : 0;
const int64_t ne12 = src1 ? src1->ne[2] : 0;
const int64_t ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
const uint64_t nb10 = src1 ? src1->nb[0] : 0;
const uint64_t nb11 = src1 ? src1->nb[1] : 0;
const uint64_t nb12 = src1 ? src1->nb[2] : 0;
const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
const int64_t ne0 = dst ? dst->ne[0] : 0;
const int64_t ne1 = dst ? dst->ne[1] : 0;
const int64_t ne2 = dst ? dst->ne[2] : 0;
const int64_t ne3 = dst ? dst->ne[3] : 0;
const uint64_t nb0 = dst ? dst->nb[0] : 0;
const uint64_t nb1 = dst ? dst->nb[1] : 0;
const uint64_t nb2 = dst ? dst->nb[2] : 0;
const uint64_t nb3 = dst ? dst->nb[3] : 0;
const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
id<MTLBuffer> id_dst = dst ? ggml_metal_get_buffer(ctx, dst, &offs_dst) : nil;
//metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
//if (src0) {
// metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
// ggml_is_contiguous(src0), src0->name);
//}
//if (src1) {
// metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
// ggml_is_contiguous(src1), src1->name);
//}
//if (dst) {
// metal_printf("%s: dst - %4s [%5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2,
// dst->name);
//}
switch (dst->op) {
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_TRANSPOSE:
case GGML_OP_PERMUTE:
{
// noop
} break;
case GGML_OP_ADD:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
if (ggml_nelements(src1) == ne10) {
// src1 is a row
[encoder setComputePipelineState:ctx->pipeline_add_row];
} else {
[encoder setComputePipelineState:ctx->pipeline_add];
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_MUL:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
if (ggml_nelements(src1) == ne10) {
// src1 is a row
[encoder setComputePipelineState:ctx->pipeline_mul_row];
} else {
[encoder setComputePipelineState:ctx->pipeline_mul];
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_SCALE:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
const float scale = *(const float *) src1->data;
[encoder setComputePipelineState:ctx->pipeline_scale];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&scale length:sizeof(scale) atIndex:2];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_UNARY:
switch (ggml_get_unary_op(gf->nodes[i])) {
case GGML_UNARY_OP_SILU:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
[encoder setComputePipelineState:ctx->pipeline_silu];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_RELU:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
[encoder setComputePipelineState:ctx->pipeline_relu];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_GELU:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
[encoder setComputePipelineState:ctx->pipeline_gelu];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
default:
{
fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
} break;
case GGML_OP_SOFT_MAX:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
const int nth = 32;
[encoder setComputePipelineState:ctx->pipeline_soft_max];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_DIAG_MASK_INF:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
const int n_past = ((int32_t *)(dst->op_params))[0];
[encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&n_past length:sizeof(int) atIndex:4];
[encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_MUL_MAT:
{
// TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
GGML_ASSERT(ne00 == ne10);
// GGML_ASSERT(ne02 == ne12); // Should be checked on individual data types until broadcast is implemented everywhere
GGML_ASSERT(ne03 == ne13);
if (ggml_is_contiguous(src0) &&
ggml_is_contiguous(src1) &&
(src0t == GGML_TYPE_F32 || src0t == GGML_TYPE_F16) && ne11 > 1) {
if (encoder != nil) {
[encoder endEncoding];
encoder = nil;
}
MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
// for F32 x F32 we use MPS
MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
MPSMatrixDescriptor * desc = [MPSMatrixDescriptor
matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc]
initWithDevice:ctx->device transposeLeft:false transposeRight:true
resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
// we need to do ne12 multiplications
// TODO: is there a way to do this in parallel - currently very slow ..
// TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
for (int64_t i02 = 0; i02 < ne12; ++i02) {
size_t offs_src0_cur = offs_src0 + i02/(ne12/ne02)*nb02; // gqa not used for now
size_t offs_src1_cur = offs_src1 + i02*nb12;
size_t offs_dst_cur = offs_dst + i02*nb2;
MPSMatrix * mat_src0 = [[MPSMatrix alloc] initWithBuffer:id_src0 offset:offs_src0_cur descriptor:desc0];
MPSMatrix * mat_src1 = [[MPSMatrix alloc] initWithBuffer:id_src1 offset:offs_src1_cur descriptor:desc1];
MPSMatrix * mat_dst = [[MPSMatrix alloc] initWithBuffer:id_dst offset:offs_dst_cur descriptor:desc ];
[mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
}
} else {
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
int nth0 = 32;
int nth1 = 1;
// use custom matrix x vector kernel
switch (src0t) {
case GGML_TYPE_F16:
{
nth0 = 64;
nth1 = 1;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
} break;
case GGML_TYPE_Q4_0:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 8;
nth1 = 8;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
} break;
case GGML_TYPE_Q4_1:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 8;
nth1 = 8;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
} break;
case GGML_TYPE_Q2_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_K_f32];
} break;
case GGML_TYPE_Q3_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_K_f32];
} break;
case GGML_TYPE_Q4_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
} break;
case GGML_TYPE_Q5_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_K_f32];
} break;
case GGML_TYPE_Q6_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 2;
nth1 = 32;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_K_f32];
} break;
default:
{
fprintf(stderr, "Asserting on type %d\n",(int)src0t);
GGML_ASSERT(false && "not implemented");
}
};
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:15];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:16];
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q3_K) {
#ifdef GGML_QKK_64
[encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#else
[encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#endif
}
else if (src0t == GGML_TYPE_Q5_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q6_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else {
[encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
}
} break;
case GGML_OP_GET_ROWS:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
switch (src0->type) {
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_K]; break;
case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_K]; break;
case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_K]; break;
case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_K]; break;
case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_K]; break;
default: GGML_ASSERT(false && "not implemented");
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
[encoder setBytes:&(dst->nb[1]) length:sizeof(uint64_t) atIndex:5];
const int64_t n = ggml_nelements(src1);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_RMS_NORM:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
const int nth = 512;
[encoder setComputePipelineState:ctx->pipeline_rms_norm];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
[encoder setBytes:&eps length:sizeof( float) atIndex:4];
[encoder setThreadgroupMemoryLength:nth/32*sizeof(float) atIndex:0];
const int64_t nrows = ggml_nrows(src0);
[encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_NORM:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
const float eps = 1e-5f;
const int nth = 256;
[encoder setComputePipelineState:ctx->pipeline_norm];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
[encoder setBytes:&eps length:sizeof( float) atIndex:4];
[encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
const int64_t nrows = ggml_nrows(src0);
[encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ALIBI:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
GGML_ASSERT((src0t == GGML_TYPE_F32));
const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
const int n_head = ((int32_t *) dst->op_params)[1];
float max_bias;
memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
if (__builtin_popcount(n_head) != 1) {
GGML_ASSERT(false && "only power-of-two n_head implemented");
}
const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
[encoder setComputePipelineState:ctx->pipeline_alibi_f32];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
[encoder setBytes:&m0 length:sizeof( float) atIndex:18];
const int nth = 32;
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ROPE:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
const int n_past = ((int32_t *) dst->op_params)[0];
const int n_dims = ((int32_t *) dst->op_params)[1];
const int mode = ((int32_t *) dst->op_params)[2];
float freq_base;
float freq_scale;
memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
[encoder setComputePipelineState:ctx->pipeline_rope];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
[encoder setBytes:&n_past length:sizeof( int) atIndex:18];
[encoder setBytes:&n_dims length:sizeof( int) atIndex:19];
[encoder setBytes:&mode length:sizeof( int) atIndex:20];
[encoder setBytes:&freq_base length:sizeof(float) atIndex:21];
[encoder setBytes:&freq_scale length:sizeof(float) atIndex:22];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_DUP:
case GGML_OP_CPY:
case GGML_OP_CONT:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
const int nth = 32;
switch (src0t) {
case GGML_TYPE_F32:
{
switch (dstt) {
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16]; break;
case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32]; break;
default: GGML_ASSERT(false && "not implemented");
};
} break;
case GGML_TYPE_F16:
{
switch (dstt) {
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f16_f16]; break;
case GGML_TYPE_F32: GGML_ASSERT(false && "cpy_f16_f32 not implemented"); break;
default: GGML_ASSERT(false && "not implemented");
};
} break;
default: GGML_ASSERT(false && "not implemented");
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
default:
{
fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
}
}
if (encoder != nil) {
[encoder endEncoding];
encoder = nil;
}
[command_buffer commit];
});
}
// wait for all threads to finish
dispatch_barrier_sync(queue, ^{});
[command_buffers[n_cb - 1] waitUntilCompleted];
// check status of command buffers
// needed to detect if the device ran out-of-memory for example (#1881)
for (int i = 0; i < n_cb; i++) {
MTLCommandBufferStatus status = (MTLCommandBufferStatus) [command_buffers[i] status];
if (status != MTLCommandBufferStatusCompleted) {
fprintf(stderr, "%s: command buffer %d failed with status %lu\n", __func__, i, status);
GGML_ASSERT(false);
}
}
}
llm/ggml-metal.metal deleted 100644 → 0
View file @ f4432e1d
//go:build darwin
/**
* llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
*
* MIT License
*
* Copyright (c) 2023 Georgi Gerganov
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include <metal_stdlib>
using namespace metal;
#define MAX(x, y) ((x) > (y) ? (x) : (y))
#define QK4_0 32
#define QR4_0 2
typedef struct {
half d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
#define QK4_1 32
typedef struct {
half d; // delta
half m; // min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static void dequantize_row_q4_0(device const block_q4_0 * x, device float * y, int k) {
const int qk = QK4_0;
assert(k % qk == 0);
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
const half d = x[i].d;
for (int j = 0; j < qk/2; ++j) {
const int x0 = (x[i].qs[j] & 0x0F) - 8;
const int x1 = (x[i].qs[j] >> 4) - 8;
y[i*qk + j + 0 ] = x0*d;
y[i*qk + j + qk/2] = x1*d;
}
}
}
static void dequantize_row_q4_1(device const block_q4_1 * x, device float * y, int k) {
const int qk = QK4_1;
assert(k % qk == 0);
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
const half d = x[i].d;
const half m = x[i].m;
for (int j = 0; j < qk/2; ++j) {
const int x0 = (x[i].qs[j] & 0x0F);
const int x1 = (x[i].qs[j] >> 4);
y[i*qk + j + 0 ] = x0*d + m;
y[i*qk + j + qk/2] = x1*d + m;
}
}
}
kernel void kernel_add(
device const float * src0,
device const float * src1,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] + src1[tpig];
}
// assumption: src1 is a row
// broadcast src1 into src0
kernel void kernel_add_row(
device const float * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] + src1[tpig % ne00];
}
kernel void kernel_mul(
device const float * src0,
device const float * src1,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] * src1[tpig];
}
// assumption: src1 is a row
// broadcast src1 into src0
kernel void kernel_mul_row(
device const float * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] * src1[tpig % ne00];
}
kernel void kernel_scale(
device const float * src0,
device float * dst,
constant float & scale,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] * scale;
}
kernel void kernel_silu(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
float x = src0[tpig];
dst[tpig] = x / (1.0f + exp(-x));
}
kernel void kernel_relu(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = max(0.0f, src0[tpig]);
}
constant float GELU_COEF_A = 0.044715f;
constant float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
kernel void kernel_gelu(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
float x = src0[tpig];
dst[tpig] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
}
kernel void kernel_soft_max(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
threadgroup float * buf [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
device const float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
device float * pdst = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
// parallel max
buf[tpitg[0]] = -INFINITY;
for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
buf[tpitg[0]] = MAX(buf[tpitg[0]], psrc0[i00]);
}
// reduce
threadgroup_barrier(mem_flags::mem_threadgroup);
for (uint i = ntg[0]/2; i > 0; i /= 2) {
if (tpitg[0] < i) {
buf[tpitg[0]] = MAX(buf[tpitg[0]], buf[tpitg[0] + i]);
}
threadgroup_barrier(mem_flags::mem_threadgroup);
}
// broadcast
if (tpitg[0] == 0) {
buf[0] = buf[0];
}
threadgroup_barrier(mem_flags::mem_threadgroup);
const float max = buf[0];
// parallel sum
buf[tpitg[0]] = 0.0f;
for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
buf[tpitg[0]] += exp(psrc0[i00] - max);
}
// reduce
threadgroup_barrier(mem_flags::mem_threadgroup);
for (uint i = ntg[0]/2; i > 0; i /= 2) {
if (tpitg[0] < i) {
buf[tpitg[0]] += buf[tpitg[0] + i];
}
threadgroup_barrier(mem_flags::mem_threadgroup);
}
// broadcast
if (tpitg[0] == 0) {
buf[0] = buf[0];
}
threadgroup_barrier(mem_flags::mem_threadgroup);
const float sum = buf[0];
for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
pdst[i00] = exp(psrc0[i00] - max) / sum;
}
}
kernel void kernel_diag_mask_inf(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int & n_past,
uint3 tpig[[thread_position_in_grid]]) {
const int64_t i02 = tpig[2];
const int64_t i01 = tpig[1];
const int64_t i00 = tpig[0];
if (i00 > n_past + i01) {
dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
} else {
dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
}
}
kernel void kernel_get_rows_f16(
device const void * src0,
device const int * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant uint64_t & nb1,
uint tpig[[thread_position_in_grid]]) {
const int i = tpig;
const int r = ((device int32_t *) src1)[i];
for (int j = 0; j < ne00; j++) {
dst[i*nb1 + j] = ((device half *) ((device char *) src0 + r*nb01))[j];
}
}
kernel void kernel_get_rows_q4_0(
device const void * src0,
device const int * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant uint64_t & nb1,
uint tpig[[thread_position_in_grid]]) {
const int i = tpig;
const int r = ((device int32_t *) src1)[i];
dequantize_row_q4_0(
(device const block_q4_0 *) ((device char *) src0 + r*nb01),
(device float *) ((device char *) dst + i*nb1), ne00);
}
kernel void kernel_get_rows_q4_1(
device const void * src0,
device const int * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant uint64_t & nb1,
uint tpig[[thread_position_in_grid]]) {
const int i = tpig;
const int r = ((device int32_t *) src1)[i];
dequantize_row_q4_1(
(device const block_q4_1 *) ((device char *) src0 + r*nb01),
(device float *) ((device char *) dst + i*nb1), ne00);
}
kernel void kernel_norm(
device const void * src0,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant float & eps,
threadgroup float * sum [[threadgroup(0)]],
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint ntg[[threads_per_threadgroup]]) {
device const float * x = (device const float *) ((device const char *) src0 + tgpig*nb01);
// MEAN
// parallel sum
sum[tpitg] = 0.0f;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
sum[tpitg] += x[i00];
}
// reduce
threadgroup_barrier(mem_flags::mem_threadgroup);
for (uint i = ntg/2; i > 0; i /= 2) {
if (tpitg < i) {
sum[tpitg] += sum[tpitg + i];
}
threadgroup_barrier(mem_flags::mem_threadgroup);
}
// broadcast
if (tpitg == 0) {
sum[0] /= ne00;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
const float mean = sum[0];
// recenter
device float * y = dst + tgpig*ne00;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
y[i00] = x[i00] - mean;
}
// VARIANCE
// parallel sum
sum[tpitg] = 0.0f;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
sum[tpitg] += y[i00] * y[i00];
}
// reduce
threadgroup_barrier(mem_flags::mem_threadgroup);
for (uint i = ntg/2; i > 0; i /= 2) {
if (tpitg < i) {
sum[tpitg] += sum[tpitg + i];
}
threadgroup_barrier(mem_flags::mem_threadgroup);
}
// broadcast
if (tpitg == 0) {
sum[0] /= ne00;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
const float variance = sum[0];
const float scale = 1.0f/sqrt(variance + eps);
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
y[i00] = y[i00] * scale;
}
}
kernel void kernel_rms_norm(
device const void * src0,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant float & eps,
threadgroup float * sum [[threadgroup(0)]],
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]],
uint tiisg[[thread_index_in_simdgroup]],
uint ntg[[threads_per_threadgroup]]) {
device const float4 * x = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
device const float * x_scalar = (device const float *) x;
float4 sumf=0;
float all_sum=0;
// parallel sum
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
sumf += x[i00] * x[i00];
}
all_sum = sumf[0] + sumf[1] + sumf[2] + sumf[3];
all_sum = simd_sum(all_sum);
if (tiisg == 0) {
sum[sgitg] = all_sum;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
// broadcast, simd group number is ntg / 32
for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
if (tpitg < i) {
sum[tpitg] += sum[tpitg + i];
}
}
if (tpitg == 0) {
for (int i = 4 * (ne00 / 4); i < ne00; i++) {sum[0] += x_scalar[i];}
sum[0] /= ne00;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
const float mean = sum[0];
const float scale = 1.0f/sqrt(mean + eps);
device float4 * y = (device float4 *) (dst + tgpig*ne00);
device float * y_scalar = (device float *) y;
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
y[i00] = x[i00] * scale;
}
if (tpitg == 0) {
for (int i00 = 4 * (ne00 / 4); i00 < ne00; i00++) {y_scalar[i00] = x_scalar[i00] * scale;}
}
}
// function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i])
// il indicates where the q4 quants begin (0 or QK4_0/4)
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 1 + il/2);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
+ yl[i + 1] * (qs[i / 2] & 0x0F00);
acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
+ yl[i + 9] * (qs[i / 2] & 0xF000);
}
return d * (sumy * -8.f + acc[0] + acc[1]);
}
// function for calculate inner product between half a q4_1 block and 16 floats (yl), sumy is SUM(yl[i])
// il indicates where the q4 quants begin (0 or QK4_0/4)
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float m = qb_curr->m;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
float2 acc = 0.f;
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
+ yl[i + 1] * (qs[i / 2] & 0x0F00);
acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
+ yl[i + 9] * (qs[i / 2] & 0xF000);
}
return d * (acc[0] + acc[1]) + sumy * m;
}
// putting them in the kernel cause a significant performance penalty
#define N_DST 4 // each SIMD group works on 4 rows
#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
//Note: This is a template, but strictly speaking it only applies to
// quantizations where the block size is 32. It also does not
// giard against the number of rows not being divisible by
// N_DST, so this is another explicit assumption of the implementation.
template<typename block_q_type, int nr, int nsg, int nw>
void mul_vec_q_n_f32(device const void * src0, device const float * src1, device float * dst,
int64_t ne00, int64_t ne10, int64_t ne0, int64_t ne01,
uint2 tgpig, uint tiisg, uint sgitg) {
const int nb = ne00/QK4_0;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int first_row = (r0 * nsg + sgitg) * nr;
device const block_q_type * x = (device const block_q_type *) src0 + first_row * nb;
device const float * y = (device const float *) src1 + r1*ne10;
float yl[16]; // src1 vector cache
float sumf[nr]={0.f};
const int ix = tiisg/2;
const int il = 8*(tiisg%2);
device const float * yb = y + ix * QK4_0 + il;
// each thread in a SIMD group deals with half a block.
for (int ib = ix; ib < nb; ib += nw/2) {
float sumy = 0;
for (int i = 0; i < 8; i += 2) {
sumy += yb[i] + yb[i+1];
yl[i+0] = yb[i+ 0];
yl[i+1] = yb[i+ 1]/256.f;
sumy += yb[i+16] + yb[i+17];
yl[i+8] = yb[i+16]/16.f;
yl[i+9] = yb[i+17]/4096.f;
}
for (int row = 0; row < nr; row++) {
sumf[row] += block_q_n_dot_y(x+ib+row*nb, sumy, yl, il);
}
yb += QK4_0 * 16;
}
for (int row = 0; row < nr; ++row) {
const float tot = simd_sum(sumf[row]);
if (tiisg == 0 && first_row + row < ne01) {
dst[r1*ne0 + first_row + row] = tot;
}
}
}
kernel void kernel_mul_mat_q4_0_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne10,
constant int64_t & ne0,
constant int64_t & ne01[[buffer(4)]],
uint2 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
mul_vec_q_n_f32<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne10,ne0,ne01,tgpig,tiisg,sgitg);
}
kernel void kernel_mul_mat_q4_1_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne10,
constant int64_t & ne0,
constant int64_t & ne01[[buffer(4)]],
uint2 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
mul_vec_q_n_f32<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne10,ne0,ne01,tgpig,tiisg,sgitg);
}
kernel void kernel_mul_mat_f16_f32(
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
threadgroup float * sum [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpig[[thread_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 tptg[[threads_per_threadgroup]]) {
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int64_t im = tgpig.z;
device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
sum[tpitg.x] = 0.0f;
for (int i = tpitg.x; i < ne00; i += tptg.x) {
sum[tpitg.x] += (float) x[i] * (float) y[i];
}
// accumulate the sum from all threads in the threadgroup
threadgroup_barrier(mem_flags::mem_threadgroup);
for (uint i = tptg.x/2; i > 0; i /= 2) {
if (tpitg.x < i) {
sum[tpitg.x] += sum[tpitg.x + i];
}
threadgroup_barrier(mem_flags::mem_threadgroup);
}
if (tpitg.x == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = sum[0];
}
}
kernel void kernel_alibi_f32(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
constant float & m0,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
float m_k = pow(m0, i2 + 1);
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
dst_data[i00] = src[0] + m_k * (i00 - ne00 + 1);
}
}
kernel void kernel_rope(
device const void * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
constant int & n_past,
constant int & n_dims,
constant int & mode,
constant float & freq_base,
constant float & freq_scale,
uint3 tpig[[thread_position_in_grid]]) {
const int64_t i3 = tpig[2];
const int64_t i2 = tpig[1];
const int64_t i1 = tpig[0];
const bool is_neox = mode & 2;
const float theta_scale = pow(freq_base, -2.0f/n_dims);
const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2);
float theta = freq_scale * (float)p;
if (!is_neox) {
for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
const float cos_theta = cos(theta);
const float sin_theta = sin(theta);
theta *= theta_scale;
device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
device float * dst_data = (device float *)((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
const float x0 = src[0];
const float x1 = src[1];
dst_data[0] = x0*cos_theta - x1*sin_theta;
dst_data[1] = x0*sin_theta + x1*cos_theta;
}
} else {
// TODO: implement
}
}
kernel void kernel_cpy_f16_f16(
device const half * src0,
device half * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
device half * dst_data = (device half *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const half * src = (device half *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
dst_data[i00] = src[0];
}
}
kernel void kernel_cpy_f32_f16(
device const float * src0,
device half * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
device half * dst_data = (device half *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
dst_data[i00] = src[0];
}
}
kernel void kernel_cpy_f32_f32(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
dst_data[i00] = src[0];
}
}
//============================================ k-quants ======================================================
#ifndef QK_K
#define QK_K 256
#else
static_assert(QK_K == 256 || QK_K == 64, "QK_K must be 256 or 64");
#endif
#if QK_K == 256
#define K_SCALE_SIZE 12
#else
#define K_SCALE_SIZE 4
#endif
typedef struct {
uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
uint8_t qs[QK_K/4]; // quants
half d; // super-block scale for quantized scales
half dmin; // super-block scale for quantized mins
} block_q2_K;
// 84 bytes / block
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
#if QK_K == 64
uint8_t scales[2];
#else
uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
#endif
half d; // super-block scale
} block_q3_K;
#if QK_K == 64
typedef struct {
half d[2]; // super-block scales/mins
uint8_t scales[2];
uint8_t qs[QK_K/2]; // 4-bit quants
} block_q4_K;
#else
typedef struct {
half d; // super-block scale for quantized scales
half dmin; // super-block scale for quantized mins
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;
#endif
#if QK_K == 64
typedef struct {
half d; // super-block scales/mins
int8_t scales[QK_K/16]; // 8-bit block scales
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
#else
typedef struct {
half d; // super-block scale for quantized scales
half dmin; // super-block scale for quantized mins
uint8_t scales[3*QK_K/64]; // 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;
// 176 bytes / block
#endif
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
half d; // super-block scale
} block_q6_K;
// 210 bytes / block
static inline uchar4 get_scale_min_k4(int j, device const uint8_t * q) {
uchar4 r;
if (j < 4) {
r[0] = q[j+0] & 63;
r[2] = q[j+1] & 63;
r[1] = q[j+4] & 63;
r[3] = q[j+5] & 63;
} else {
r[0] = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
r[2] = (q[j+5] & 0xF) | ((q[j-3] >> 6) << 4);
r[1] = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4);
r[3] = (q[j+5] >> 4) | ((q[j+1] >> 6) << 4);
}
return r;
}
//========================================== dequantization =============================
static void dequantize_row_q2_K(device const block_q2_K * x, device float * y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
for (int i = 0; i < nb; i++) {
const float d = x[i].d;
const float min = x[i].dmin;
device const uint8_t * q = x[i].qs;
#if QK_K == 256
int is = 0;
float dl, ml;
for (int n = 0; n < QK_K; n += 128) {
int shift = 0;
for (int j = 0; j < 4; ++j) {
uint8_t sc = x[i].scales[is++];
dl = d * (sc & 0xF); ml = min * (sc >> 4);
for (int l = 0; l < 16; ++l) *y++ = dl * ((int8_t)((q[l] >> shift) & 3)) - ml;
sc = x[i].scales[is++];
dl = d * (sc & 0xF); ml = min * (sc >> 4);
for (int l = 0; l < 16; ++l) *y++ = dl * ((int8_t)((q[l+16] >> shift) & 3)) - ml;
shift += 2;
}
q += 32;
}
#else
float dl1 = d * (x[i].scales[0] & 0xF), ml1 = min * (x[i].scales[0] >> 4);
float dl2 = d * (x[i].scales[1] & 0xF), ml2 = min * (x[i].scales[1] >> 4);
float dl3 = d * (x[i].scales[2] & 0xF), ml3 = min * (x[i].scales[2] >> 4);
float dl4 = d * (x[i].scales[3] & 0xF), ml4 = min * (x[i].scales[3] >> 4);
for (int l = 0; l < 16; ++l) {
y[l+ 0] = dl1 * ((q[l] >> 0) & 3) - ml1;
y[l+16] = dl2 * ((q[l] >> 2) & 3) - ml2;
y[l+32] = dl3 * ((q[l] >> 4) & 3) - ml3;
y[l+48] = dl4 * ((q[l] >> 6) & 3) - ml4;
}
y += QK_K;
#endif
}
}
static void dequantize_row_q3_K(device const block_q3_K * x, device float * y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
#if QK_K == 256
const uint16_t kmask1 = 0x0303;
const uint16_t kmask2 = 0x0f0f;
uint16_t aux[8];
thread const int8_t * scales = (thread const int8_t*)aux;
for (int i = 0; i < nb; i++) {
const float d_all = (float)(x[i].d);
device const uint8_t * q = x[i].qs;
device const uint8_t * h = x[i].hmask;
uint8_t m = 1;
device const uint16_t * a = (device const uint16_t *)x[i].scales;
aux[0] = (a[0] & kmask2) | (((a[4] >> 0) & kmask1) << 4);
aux[1] = (a[1] & kmask2) | (((a[5] >> 0) & kmask1) << 4);
aux[2] = (a[2] & kmask2) | (((a[4] >> 2) & kmask1) << 4);
aux[3] = (a[3] & kmask2) | (((a[5] >> 2) & kmask1) << 4);
aux[4] = ((a[0] >> 4) & kmask2) | (((a[4] >> 4) & kmask1) << 4);
aux[5] = ((a[1] >> 4) & kmask2) | (((a[5] >> 4) & kmask1) << 4);
aux[6] = ((a[2] >> 4) & kmask2) | (((a[4] >> 6) & kmask1) << 4);
aux[7] = ((a[3] >> 4) & kmask2) | (((a[5] >> 6) & kmask1) << 4);
int is = 0;
float dl;
for (int n = 0; n < QK_K; n += 128) {
int shift = 0;
for (int j = 0; j < 4; ++j) {
dl = d_all * (scales[is++] - 32);
for (int l = 0; l < 16; ++l) {
*y++ = dl * ((int8_t)((q[l+ 0] >> shift) & 3) - ((h[l+ 0] & m) ? 0 : 4));
}
dl = d_all * (scales[is++] - 32);
for (int l = 0; l < 16; ++l) {
*y++ = dl * ((int8_t)((q[l+16] >> shift) & 3) - ((h[l+16] & m) ? 0 : 4));
}
shift += 2;
m <<= 1;
}
q += 32;
}
}
#else
for (int i = 0; i < nb; i++) {
const float d_all = (float)(x[i].d);
device const uint8_t * q = x[i].qs;
device const uint8_t * hm = x[i].hmask;
const float d1 = d_all * ((x[i].scales[0] & 0xF) - 8);
const float d2 = d_all * ((x[i].scales[0] >> 4) - 8);
const float d3 = d_all * ((x[i].scales[1] & 0xF) - 8);
const float d4 = d_all * ((x[i].scales[1] >> 4) - 8);
for (int l = 0; l < 8; ++l) {
uint8_t h = hm[l];
y[l+ 0] = d1 * ((int8_t)((q[l+0] >> 0) & 3) - ((h & 0x01) ? 0 : 4));
y[l+ 8] = d1 * ((int8_t)((q[l+8] >> 0) & 3) - ((h & 0x02) ? 0 : 4));
y[l+16] = d2 * ((int8_t)((q[l+0] >> 2) & 3) - ((h & 0x04) ? 0 : 4));
y[l+24] = d2 * ((int8_t)((q[l+8] >> 2) & 3) - ((h & 0x08) ? 0 : 4));
y[l+32] = d3 * ((int8_t)((q[l+0] >> 4) & 3) - ((h & 0x10) ? 0 : 4));
y[l+40] = d3 * ((int8_t)((q[l+8] >> 4) & 3) - ((h & 0x20) ? 0 : 4));
y[l+48] = d4 * ((int8_t)((q[l+0] >> 6) & 3) - ((h & 0x40) ? 0 : 4));
y[l+56] = d4 * ((int8_t)((q[l+8] >> 6) & 3) - ((h & 0x80) ? 0 : 4));
}
y += QK_K;
}
#endif
}
static void dequantize_row_q4_K(device const block_q4_K * x, device float * y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
for (int i = 0; i < nb; i++) {
device const uint8_t * q = x[i].qs;
#if QK_K == 256
const float d = x[i].d;
const float min = x[i].dmin;
device const uint8_t * scales = x[i].scales;
int is = 0;
for (int j = 0; j < QK_K; j += 64) {
const uchar4 sc = get_scale_min_k4(is, scales);
const float d1 = d * sc[0]; const float m1 = min * sc[1];
const float d2 = d * sc[2]; const float m2 = min * sc[3];
for (int l = 0; l < 32; ++l) *y++ = d1 * (q[l] & 0xF) - m1;
for (int l = 0; l < 32; ++l) *y++ = d2 * (q[l] >> 4) - m2;
q += 32; is += 2;
}
#else
device const uint8_t * s = x[i].scales;
device const half2 * dh = (device const half2 *)x[i].d;
const float2 d = (float2)dh[0];
const float d1 = d[0] * (s[0] & 0xF);
const float d2 = d[0] * (s[1] & 0xF);
const float m1 = d[1] * (s[0] >> 4);
const float m2 = d[1] * (s[1] >> 4);
for (int l = 0; l < 32; ++l) {
y[l+ 0] = d1 * (q[l] & 0xF) - m1;
y[l+32] = d2 * (q[l] >> 4) - m2;
}
y += QK_K;
#endif
}
}
static void dequantize_row_q5_K(device const block_q5_K * x, device float * y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
#if QK_K == 256
for (int i = 0; i < nb; i++) {
const float d = (float)(x[i].d);
const float min = (float)(x[i].dmin);
device const uint8_t * ql = x[i].qs;
device const uint8_t * qh = x[i].qh;
int is = 0;
uint8_t u1 = 1, u2 = 2;
for (int j = 0; j < QK_K; j += 64) {
const uchar4 sc = get_scale_min_k4(is, x[i].scales);
const float d1 = d * sc[0]; const float m1 = min * sc[1];
const float d2 = d * sc[2]; const float m2 = min * sc[3];
for (int l = 0; l < 32; ++l) *y++ = d1 * ((ql[l] & 0xF) + (qh[l] & u1 ? 16 : 0)) - m1;
for (int l = 0; l < 32; ++l) *y++ = d2 * ((ql[l] >> 4) + (qh[l] & u2 ? 16 : 0)) - m2;
ql += 32; is += 2;
u1 <<= 2; u2 <<= 2;
}
}
#else
for (int i = 0; i < nb; i++) {
const float d = (float)x[i].d;
device const uint8_t * ql = x[i].qs;
device const uint8_t * qh = x[i].qh;
device const int8_t * sc = x[i].scales;
for (int l = 0; l < 8; ++l) {
y[l+ 0] = d * sc[0] * ((ql[l+ 0] & 0xF) - (qh[l] & 0x01 ? 0 : 16));
y[l+ 8] = d * sc[0] * ((ql[l+ 8] & 0xF) - (qh[l] & 0x02 ? 0 : 16));
y[l+16] = d * sc[1] * ((ql[l+16] & 0xF) - (qh[l] & 0x04 ? 0 : 16));
y[l+24] = d * sc[1] * ((ql[l+24] & 0xF) - (qh[l] & 0x08 ? 0 : 16));
y[l+32] = d * sc[2] * ((ql[l+ 0] >> 4) - (qh[l] & 0x10 ? 0 : 16));
y[l+40] = d * sc[2] * ((ql[l+ 8] >> 4) - (qh[l] & 0x20 ? 0 : 16));
y[l+48] = d * sc[3] * ((ql[l+16] >> 4) - (qh[l] & 0x40 ? 0 : 16));
y[l+56] = d * sc[3] * ((ql[l+24] >> 4) - (qh[l] & 0x80 ? 0 : 16));
}
y += QK_K;
}
#endif
}
static void dequantize_row_q6_K(device const block_q6_K * x, device float * y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
for (int i = 0; i < nb; i++) {
device const uint8_t * ql = x[i].ql;
device const uint8_t * qh = x[i].qh;
device const int8_t * sc = x[i].scales;
const float d = x[i].d;
#if QK_K == 256
for (int n = 0; n < QK_K; n += 128) {
for (int l = 0; l < 32; ++l) {
int is = l/16;
const int8_t q1 = (int8_t)((ql[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
const int8_t q2 = (int8_t)((ql[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
const int8_t q3 = (int8_t)((ql[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
const int8_t q4 = (int8_t)((ql[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
y[l + 0] = d * sc[is + 0] * q1;
y[l + 32] = d * sc[is + 2] * q2;
y[l + 64] = d * sc[is + 4] * q3;
y[l + 96] = d * sc[is + 6] * q4;
}
y += 128;
ql += 64;
qh += 32;
sc += 8;
}
#else
for (int l = 0; l < 16; ++l) {
const int8_t q1 = (int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
const int8_t q2 = (int8_t)((ql[l+16] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
const int8_t q3 = (int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
const int8_t q4 = (int8_t)((ql[l+16] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
y[l+ 0] = d * sc[0] * q1;
y[l+16] = d * sc[1] * q2;
y[l+32] = d * sc[2] * q3;
y[l+48] = d * sc[3] * q4;
}
y += 64;
#endif
}
}
kernel void kernel_get_rows_q2_K(
device const void * src0,
device const int * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant uint64_t & nb1,
uint tpig[[thread_position_in_grid]]) {
const int i = tpig;
const int r = ((device int32_t *) src1)[i];
dequantize_row_q2_K(
(device const block_q2_K *) ((device char *) src0 + r*nb01),
(device float *) ((device char *) dst + i*nb1), ne00);
}
kernel void kernel_get_rows_q3_K(
device const void * src0,
device const int * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant uint64_t & nb1,
uint tpig[[thread_position_in_grid]]) {
const int i = tpig;
const int r = ((device int32_t *) src1)[i];
dequantize_row_q3_K(
(device const block_q3_K *) ((device char *) src0 + r*nb01),
(device float *) ((device char *) dst + i*nb1), ne00);
}
kernel void kernel_get_rows_q4_K(
device const void * src0,
device const int * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant uint64_t & nb1,
uint tpig[[thread_position_in_grid]]) {
const int i = tpig;
const int r = ((device int32_t *) src1)[i];
dequantize_row_q4_K(
(device const block_q4_K *) ((device char *) src0 + r*nb01),
(device float *) ((device char *) dst + i*nb1), ne00);
}
kernel void kernel_get_rows_q5_K(
device const void * src0,
device const int * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant uint64_t & nb1,
uint tpig[[thread_position_in_grid]]) {
const int i = tpig;
const int r = ((device int32_t *) src1)[i];
dequantize_row_q5_K(
(device const block_q5_K *) ((device char *) src0 + r*nb01),
(device float *) ((device char *) dst + i*nb1), ne00);
}
kernel void kernel_get_rows_q6_K(
device const void * src0,
device const int * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant uint64_t & nb1,
uint tpig[[thread_position_in_grid]]) {
const int i = tpig;
const int r = ((device int32_t *) src1)[i];
dequantize_row_q6_K(
(device const block_q6_K *) ((device char *) src0 + r*nb01),
(device float *) ((device char *) dst + i*nb1), ne00);
}
//====================================== dot products =========================
kernel void kernel_mul_mat_q2_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne10,
constant int64_t & ne0,
constant int64_t & ne01[[buffer(4)]],
uint2 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
device const block_q2_K * x = (device const block_q2_K *) src0 + ib_row;
device const float * y = (device const float *) src1 + r1*ne10;
float yl[32];
float sumf[N_DST]={0.f}, all_sum;
const int step = sizeof(block_q2_K) * nb;
#if QK_K == 256
const int ix = tiisg/8; // 0...3
const int it = tiisg%8; // 0...7
const int im = it/4; // 0 or 1
const int ir = it%4; // 0...3
const int is = (8*ir)/16;// 0 or 1
device const float * y4 = y + ix * QK_K + 128 * im + 8 * ir;
for (int ib = ix; ib < nb; ib += 4) {
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
yl[i+ 8] = y4[i+32]; sumy[1] += yl[i+ 8];
yl[i+16] = y4[i+64]; sumy[2] += yl[i+16];
yl[i+24] = y4[i+96]; sumy[3] += yl[i+24];
}
device const uint8_t * sc = (device const uint8_t *)x[ib].scales + 8*im + is;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * im + 4 * ir;
device const half * dh = &x[ib].d;
for (int row = 0; row < N_DST; row++) {
float4 acc1 = {0.f, 0.f, 0.f, 0.f};
float4 acc2 = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
}
float dall = dh[0];
float dmin = dh[1] * 1.f/16.f;
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
(acc1[1] + 1.f/256.f * acc2[1]) * (sc[2] & 0xF) * 1.f/ 4.f +
(acc1[2] + 1.f/256.f * acc2[2]) * (sc[4] & 0xF) * 1.f/16.f +
(acc1[3] + 1.f/256.f * acc2[3]) * (sc[6] & 0xF) * 1.f/64.f) -
dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) + sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));
qs += step/2;
sc += step;
dh += step/2;
}
y4 += 4 * QK_K;
}
#else
const int ix = tiisg/2; // 0...15
const int it = tiisg%2; // 0...1
device const float * y4 = y + ix * QK_K + 8 * it;
for (int ib = ix; ib < nb; ib += 16) {
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
yl[i+ 8] = y4[i+16]; sumy[1] += yl[i+ 8];
yl[i+16] = y4[i+32]; sumy[2] += yl[i+16];
yl[i+24] = y4[i+48]; sumy[3] += yl[i+24];
}
device const uint8_t * sc = (device const uint8_t *)x[ib].scales;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
device const half * dh = &x[ib].d;
for (int row = 0; row < N_DST; row++) {
float4 acc1 = {0.f, 0.f, 0.f, 0.f};
float4 acc2 = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
}
float dall = dh[0];
float dmin = dh[1];
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
(acc1[1] + 1.f/256.f * acc2[1]) * (sc[1] & 0xF) * 1.f/ 4.f +
(acc1[2] + 1.f/256.f * acc2[2]) * (sc[2] & 0xF) * 1.f/16.f +
(acc1[3] + 1.f/256.f * acc2[3]) * (sc[3] & 0xF) * 1.f/64.f) -
dmin * (sumy[0] * (sc[0] >> 4) + sumy[1] * (sc[1] >> 4) + sumy[2] * (sc[2] >> 4) + sumy[3] * (sc[3] >> 4));
qs += step/2;
sc += step;
dh += step/2;
}
y4 += 16 * QK_K;
}
#endif
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + first_row + row] = all_sum;
}
}
}
#if QK_K == 256
kernel void kernel_mul_mat_q3_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne10,
constant int64_t & ne0,
constant int64_t & ne1,
uint2 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
device const block_q3_K * x = (device const block_q3_K *) src0 + first_row*nb;
device const float * yy = (device const float *) src1 + r1*ne10;
float yl[16];
const uint16_t kmask1 = 0x0303;
const uint16_t kmask2 = 0x0f0f;
const int tid = tiisg/2;
const int ix = tiisg%2;
const int ip = tid/8; // 0 or 1
const int il = tid/2 - 4*ip; // 0...3
const int ir = tid%2;
const int n = 8;
const int l0 = n*ir;
const uint16_t m1 = 1 << (4*ip + il);
const uint16_t m2 = m1 << 8;
const int shift = 2*il;
const uint16_t qm1 = 0x0003 << shift;
const uint16_t qm2 = 0x0300 << shift;
const int32_t v1 = 4 << shift;
const int32_t v2 = 1024 << shift;
const uint16_t s_shift1 = 4*ip;
const uint16_t s_shift2 = s_shift1 + 2*(il/2);
const int ik = 4 + (il%2);
const int q_offset = 32*ip + l0;
const int y_offset = 128*ip + 32*il + l0;
const int step = sizeof(block_q3_K) * nb / 2;
device const float * y1 = yy + ix*QK_K + y_offset;
float sumf1[2] = {0.f}, sumf2[2] = {0.f};
for (int i = ix; i < nb; i += 2) {
for (int l = 0; l < 8; ++l) {
yl[l+0] = y1[l+ 0];
yl[l+8] = y1[l+16];
}
device const uint16_t * q = (device const uint16_t *)(x[i].qs + q_offset);
device const uint16_t * h = (device const uint16_t *)(x[i].hmask + l0);
device const uint16_t * a = (device const uint16_t *)(x[i].scales);
device const half * dh = &x[i].d;
for (int row = 0; row < 2; ++row) {
const float d_all = (float)dh[0];
const char2 scales = as_type<char2>((uint16_t)(((a[il] >> s_shift1) & kmask2) | (((a[ik] >> s_shift2) & kmask1) << 4)));
float s1 = 0, s2 = 0;
for (int l = 0; l < n; l += 2) {
const uint16_t qs = q[l/2];
s1 += yl[l+0] * ((int32_t)(qs & qm1) - ((h[l/2] & m1) ? 0 : v1));
s2 += yl[l+1] * ((int32_t)(qs & qm2) - ((h[l/2] & m2) ? 0 : v2));
}
float d = d_all * (s1 + 1.f/256.f * s2);
sumf1[row] += d * scales[0];
sumf2[row] += d;
s1 = s2 = 0;
for (int l = 0; l < n; l += 2) {
const uint16_t qs = q[l/2+8];
s1 += yl[l+8] * ((int32_t)(qs & qm1) - ((h[l/2+8] & m1) ? 0 : v1));
s2 += yl[l+9] * ((int32_t)(qs & qm2) - ((h[l/2+8] & m2) ? 0 : v2));
}
d = d_all * (s1 + 1.f/256.f * s2);
sumf1[row] += d * scales[1];
sumf2[row] += d;
q += step;
h += step;
a += step;
dh += step;
}
y1 += 2 * QK_K;
}
for (int row = 0; row < 2; ++row) {
const float sumf = (sumf1[row] - 32.f*sumf2[row]) / (1 << shift);
const float tot = simd_sum(sumf);
if (tiisg == 0) {
dst[r1*ne0 + first_row + row] = tot;
}
}
}
#else
kernel void kernel_mul_mat_q3_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne10,
constant int64_t & ne0,
constant int64_t & ne1,
uint2 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int row = 2 * r0 + sgitg;
device const block_q3_K * x = (device const block_q3_K *) src0 + row*nb;
device const float * yy = (device const float *) src1 + r1*ne10;
const int ix = tiisg/4;
const int il = 4 * (tiisg%4);// 0, 4, 8, 12
const int im = il/8; // 0, 0, 1, 1
const int in = il%8; // 0, 4, 0, 4
float2 sum = {0.f, 0.f};
for (int i = ix; i < nb; i += 8) {
const float d_all = (float)(x[i].d);
device const uint16_t * q = (device const uint16_t *)(x[i].qs + il);
device const uint16_t * h = (device const uint16_t *)(x[i].hmask + in);
device const uint16_t * s = (device const uint16_t *)(x[i].scales);
device const float * y = yy + i * QK_K + il;
const float d1 = d_all * ((int32_t)(s[0] & 0x000F) - 8);
const float d2 = d_all * ((int32_t)(s[0] & 0x00F0) - 128) * 1.f/64.f;
const float d3 = d_all * ((int32_t)(s[0] & 0x0F00) - 2048) * 1.f/4096.f;
const float d4 = d_all * ((int32_t)(s[0] & 0xF000) - 32768) * 1.f/262144.f;
for (int l = 0; l < 4; l += 2) {
const uint16_t hm = h[l/2] >> im;
sum[0] += y[l+ 0] * d1 * ((int32_t)(q[l/2] & 0x0003) - ((hm & 0x0001) ? 0 : 4))
+ y[l+16] * d2 * ((int32_t)(q[l/2] & 0x000c) - ((hm & 0x0004) ? 0 : 16))
+ y[l+32] * d3 * ((int32_t)(q[l/2] & 0x0030) - ((hm & 0x0010) ? 0 : 64))
+ y[l+48] * d4 * ((int32_t)(q[l/2] & 0x00c0) - ((hm & 0x0040) ? 0 : 256));
sum[1] += y[l+ 1] * d1 * ((int32_t)(q[l/2] & 0x0300) - ((hm & 0x0100) ? 0 : 1024))
+ y[l+17] * d2 * ((int32_t)(q[l/2] & 0x0c00) - ((hm & 0x0400) ? 0 : 4096))
+ y[l+33] * d3 * ((int32_t)(q[l/2] & 0x3000) - ((hm & 0x1000) ? 0 : 16384))
+ y[l+49] * d4 * ((int32_t)(q[l/2] & 0xc000) - ((hm & 0x4000) ? 0 : 65536));
}
}
const float sumf = sum[0] + sum[1] * 1.f/256.f;
const float tot = simd_sum(sumf);
if (tiisg == 0) {
dst[r1*ne0 + row] = tot;
}
}
#endif
#if QK_K == 256
kernel void kernel_mul_mat_q4_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne10,
constant int64_t & ne0,
constant int64_t & ne01[[buffer(4)]],
uint2 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int ix = tiisg/8; // 0...3
const int it = tiisg%8; // 0...7
const int im = it/4; // 0 or 1
const int ir = it%4; // 0...3
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row;
device const float * y = (device const float *) src1 + r1*ne10;
float yl[16];
float yh[16];
float sumf[N_DST]={0.f}, all_sum;
const int step = sizeof(block_q4_K) * nb / 2;
device const float * y4 = y + ix * QK_K + 64 * im + 8 * ir;
uint16_t sc16[4];
thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
for (int ib = ix; ib < nb; ib += 4) {
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i+0] = y4[i+ 0]; sumy[0] += yl[i+0];
yl[i+8] = y4[i+ 32]; sumy[1] += yl[i+8];
yh[i+0] = y4[i+128]; sumy[2] += yh[i+0];
yh[i+8] = y4[i+160]; sumy[3] += yh[i+8];
}
device const uint16_t * sc = (device const uint16_t *)x[ib].scales + im;
device const uint16_t * q1 = (device const uint16_t *)x[ib].qs + 16 * im + 4 * ir;
device const half * dh = &x[ib].d;
for (int row = 0; row < N_DST; row++) {
sc16[0] = sc[0] & kmask1;
sc16[1] = sc[2] & kmask1;
sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);
device const uint16_t * q2 = q1 + 32;
float4 acc1 = {0.f, 0.f, 0.f, 0.f};
float4 acc2 = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+0] * (q1[i/2] & 0x000F);
acc1[1] += yl[i+1] * (q1[i/2] & 0x0F00);
acc1[2] += yl[i+8] * (q1[i/2] & 0x00F0);
acc1[3] += yl[i+9] * (q1[i/2] & 0xF000);
acc2[0] += yh[i+0] * (q2[i/2] & 0x000F);
acc2[1] += yh[i+1] * (q2[i/2] & 0x0F00);
acc2[2] += yh[i+8] * (q2[i/2] & 0x00F0);
acc2[3] += yh[i+9] * (q2[i/2] & 0xF000);
}
float dall = dh[0];
float dmin = dh[1];
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8[0] +
(acc1[2] + 1.f/256.f * acc1[3]) * sc8[1] * 1.f/16.f +
(acc2[0] + 1.f/256.f * acc2[1]) * sc8[4] +
(acc2[2] + 1.f/256.f * acc2[3]) * sc8[5] * 1.f/16.f) -
dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
q1 += step;
sc += step;
dh += step;
}
y4 += 4 * QK_K;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + first_row + row] = all_sum;
}
}
}
#else
kernel void kernel_mul_mat_q4_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne10,
constant int64_t & ne0,
constant int64_t & ne01[[buffer(4)]],
uint2 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int ix = tiisg/4; // 0...7
const int it = tiisg%4; // 0...3
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row;
device const float * y = (device const float *) src1 + r1*ne10;
float yl[8];
float yh[8];
float sumf[N_DST]={0.f}, all_sum;
const int step = sizeof(block_q4_K) * nb / 2;
device const float * y4 = y + ix * QK_K + 8 * it;
uint16_t sc16[4];
for (int ib = ix; ib < nb; ib += 8) {
float2 sumy = {0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i] = y4[i+ 0]; sumy[0] += yl[i];
yh[i] = y4[i+32]; sumy[1] += yh[i];
}
device const uint16_t * sc = (device const uint16_t *)x[ib].scales;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
device const half * dh = x[ib].d;
for (int row = 0; row < N_DST; row++) {
sc16[0] = sc[0] & 0x000f;
sc16[1] = sc[0] & 0x0f00;
sc16[2] = sc[0] & 0x00f0;
sc16[3] = sc[0] & 0xf000;
float2 acc1 = {0.f, 0.f};
float2 acc2 = {0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+0] * (qs[i/2] & 0x000F);
acc1[1] += yl[i+1] * (qs[i/2] & 0x0F00);
acc2[0] += yh[i+0] * (qs[i/2] & 0x00F0);
acc2[1] += yh[i+1] * (qs[i/2] & 0xF000);
}
float dall = dh[0];
float dmin = dh[1];
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc16[0] +
(acc2[0] + 1.f/256.f * acc2[1]) * sc16[1] * 1.f/4096.f) -
dmin * 1.f/16.f * (sumy[0] * sc16[2] + sumy[1] * sc16[3] * 1.f/256.f);
qs += step;
sc += step;
dh += step;
}
y4 += 8 * QK_K;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + first_row + row] = all_sum;
}
}
}
#endif
kernel void kernel_mul_mat_q5_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne10,
constant int64_t & ne0,
uint2 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
device const block_q5_K * x = (device const block_q5_K *) src0 + first_row*nb;
device const float * yy = (device const float *) src1 + r1*ne10;
float sumf[2]={0.f};
const int step = sizeof(block_q5_K) * nb;
#if QK_K == 256
#
float yl[16], yh[16];
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int tid = tiisg/4;
const int ix = tiisg%4;
const int im = tid/4;
const int ir = tid%4;
const int n = 8;
const int l0 = n*ir;
const int q_offset = 32*im + l0;
const int y_offset = 64*im + l0;
const uint8_t hm1 = 1u << (2*im);
const uint8_t hm2 = hm1 << 1;
const uint8_t hm3 = hm1 << 4;
const uint8_t hm4 = hm2 << 4;
uint16_t sc16[4];
thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
device const float * y1 = yy + ix*QK_K + y_offset;
for (int i = ix; i < nb; i += 4) {
device const uint8_t * q1 = x[i].qs + q_offset;
device const uint8_t * qh = x[i].qh + l0;
device const half * dh = &x[i].d;
device const uint16_t * a = (device const uint16_t *)x[i].scales + im;
device const float * y2 = y1 + 128;
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int l = 0; l < 8; ++l) {
yl[l+0] = y1[l+ 0]; sumy[0] += yl[l+0];
yl[l+8] = y1[l+32]; sumy[1] += yl[l+8];
yh[l+0] = y2[l+ 0]; sumy[2] += yh[l+0];
yh[l+8] = y2[l+32]; sumy[3] += yh[l+8];
}
for (int row = 0; row < 2; ++row) {
device const uint8_t * q2 = q1 + 64;
sc16[0] = a[0] & kmask1;
sc16[1] = a[2] & kmask1;
sc16[2] = ((a[4] >> 0) & kmask2) | ((a[0] & kmask3) >> 2);
sc16[3] = ((a[4] >> 4) & kmask2) | ((a[2] & kmask3) >> 2);
float4 acc = {0.f, 0.f, 0.f, 0.f};
for (int l = 0; l < n; ++l) {
uint8_t h = qh[l];
acc[0] += yl[l+0] * ((uint16_t)(q1[l] & 0x0F) + (h & hm1 ? 16 : 0));
acc[1] += yl[l+8] * ((uint16_t)(q1[l] & 0xF0) + (h & hm2 ? 256 : 0));
acc[2] += yh[l+0] * ((uint16_t)(q2[l] & 0x0F) + (h & hm3 ? 16 : 0));
acc[3] += yh[l+8] * ((uint16_t)(q2[l] & 0xF0) + (h & hm4 ? 256 : 0));
}
const float dall = dh[0];
const float dmin = dh[1];
sumf[row] += dall * (acc[0] * sc8[0] + acc[1] * sc8[1] * 1.f/16.f + acc[2] * sc8[4] + acc[3] * sc8[5] * 1.f/16.f) -
dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
q1 += step;
qh += step;
dh += step/2;
a += step/2;
}
y1 += 4 * QK_K;
}
#else
float yl[8], yh[8];
const int il = 4 * (tiisg/8); // 0, 4, 8, 12
const int ix = tiisg%8;
const int im = il/8; // 0, 0, 1, 1
const int in = il%8; // 0, 4, 0, 4
device const float * y = yy + ix*QK_K + il;
for (int i = ix; i < nb; i += 8) {
for (int l = 0; l < 4; ++l) {
yl[l+0] = y[l+ 0];
yl[l+4] = y[l+16];
yh[l+0] = y[l+32];
yh[l+4] = y[l+48];
}
device const half * dh = &x[i].d;
device const uint8_t * q = x[i].qs + il;
device const uint8_t * h = x[i].qh + in;
device const int8_t * s = x[i].scales;
for (int row = 0; row < 2; ++row) {
const float d = dh[0];
float2 acc = {0.f, 0.f};
for (int l = 0; l < 4; ++l) {
const uint8_t hl = h[l] >> im;
acc[0] += yl[l+0] * s[0] * ((int16_t)(q[l+ 0] & 0x0F) - (hl & 0x01 ? 0 : 16))
+ yl[l+4] * s[1] * ((int16_t)(q[l+16] & 0x0F) - (hl & 0x04 ? 0 : 16));
acc[1] += yh[l+0] * s[2] * ((int16_t)(q[l+ 0] & 0xF0) - (hl & 0x10 ? 0 : 256))
+ yh[l+4] * s[3] * ((int16_t)(q[l+16] & 0xF0) - (hl & 0x40 ? 0 : 256));
}
sumf[row] += d * (acc[0] + 1.f/16.f * acc[1]);
q += step;
h += step;
s += step;
dh += step/2;
}
y += 8 * QK_K;
}
#endif
for (int row = 0; row < 2; ++row) {
const float tot = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + first_row + row] = tot;
}
}
}
kernel void kernel_mul_mat_q6_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne10,
constant int64_t & ne0,
uint2 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const uint8_t kmask1 = 0x03;
const uint8_t kmask2 = 0x0C;
const uint8_t kmask3 = 0x30;
const uint8_t kmask4 = 0xC0;
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int row = 2 * r0 + sgitg;
device const block_q6_K * x = (device const block_q6_K *) src0 + row * nb; //r0*nb;
device const float * yy = (device const float *) src1 + r1*ne10;
float sumf = 0;
#if QK_K == 256
const int tid = tiisg/2;
const int ix = tiisg%2;
const int ip = tid/8; // 0 or 1
const int il = tid%8;
const int n = 4;
const int l0 = n*il;
const int is = 8*ip + l0/16;
const int y_offset = 128*ip + l0;
const int q_offset_l = 64*ip + l0;
const int q_offset_h = 32*ip + l0;
for (int i = ix; i < nb; i += 2) {
device const uint8_t * q1 = x[i].ql + q_offset_l;
device const uint8_t * q2 = q1 + 32;
device const uint8_t * qh = x[i].qh + q_offset_h;
device const int8_t * sc = x[i].scales + is;
device const float * y = yy + i * QK_K + y_offset;
const float dall = x[i].d;
float4 sums = {0.f, 0.f, 0.f, 0.f};
for (int l = 0; l < n; ++l) {
sums[0] += y[l+ 0] * ((int8_t)((q1[l] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
sums[1] += y[l+32] * ((int8_t)((q2[l] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
sums[2] += y[l+64] * ((int8_t)((q1[l] >> 4) | ((qh[l] & kmask3) << 0)) - 32);
sums[3] += y[l+96] * ((int8_t)((q2[l] >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
}
sumf += dall * (sums[0] * sc[0] + sums[1] * sc[2] + sums[2] * sc[4] + sums[3] * sc[6]);
}
#else
const int ix = tiisg/4;
const int il = 4*(tiisg%4);
for (int i = ix; i < nb; i += 8) {
device const float * y = yy + i * QK_K + il;
device const uint8_t * ql = x[i].ql + il;
device const uint8_t * qh = x[i].qh + il;
device const int8_t * s = x[i].scales;
const float d = x[i].d;
float4 sums = {0.f, 0.f, 0.f, 0.f};
for (int l = 0; l < 4; ++l) {
sums[0] += y[l+ 0] * ((int8_t)((ql[l+ 0] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
sums[1] += y[l+16] * ((int8_t)((ql[l+16] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
sums[2] += y[l+32] * ((int8_t)((ql[l+ 0] >> 4) | ((qh[l] & kmask3) >> 0)) - 32);
sums[3] += y[l+48] * ((int8_t)((ql[l+16] >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
}
sumf += d * (sums[0] * s[0] + sums[1] * s[1] + sums[2] * s[2] + sums[3] * s[3]);
}
#endif
const float tot = simd_sum(sumf);
if (tiisg == 0) {
dst[r1*ne0 + row] = tot;
}
}
llm/ggml-mpi.c deleted 100644 → 0
View file @ f4432e1d
//go:build mpi
/**
* llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
*
* MIT License
*
* Copyright (c) 2023 Georgi Gerganov
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "ggml-mpi.h"
#include "ggml.h"
#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define UNUSED GGML_UNUSED
struct ggml_mpi_context {
int rank;
int size;
};
void ggml_mpi_backend_init(void) {
MPI_Init(NULL, NULL);
}
void ggml_mpi_backend_free(void) {
MPI_Finalize();
}
struct ggml_mpi_context * ggml_mpi_init(void) {
struct ggml_mpi_context * ctx = calloc(1, sizeof(struct ggml_mpi_context));
MPI_Comm_rank(MPI_COMM_WORLD, &ctx->rank);
MPI_Comm_size(MPI_COMM_WORLD, &ctx->size);
return ctx;
}
void ggml_mpi_free(struct ggml_mpi_context * ctx) {
free(ctx);
}
int ggml_mpi_rank(struct ggml_mpi_context * ctx) {
return ctx->rank;
}
void ggml_mpi_eval_init(
struct ggml_mpi_context * ctx_mpi,
int * n_tokens,
int * n_past,
int * n_threads) {
UNUSED(ctx_mpi);
// synchronize the worker node parameters with the root node
MPI_Barrier(MPI_COMM_WORLD);
MPI_Bcast(n_tokens, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Bcast(n_past, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Bcast(n_threads, 1, MPI_INT, 0, MPI_COMM_WORLD);
}
static int ggml_graph_get_node_idx(struct ggml_cgraph * gf, const char * name) {
struct ggml_tensor * t = ggml_graph_get_tensor(gf, name);
if (t == NULL) {
fprintf(stderr, "%s: tensor %s not found\n", __func__, name);
return -1;
}
for (int i = 0; i < gf->n_nodes; i++) {
if (gf->nodes[i] == t) {
return i;
}
}
fprintf(stderr, "%s: tensor %s not found in graph (should not happen)\n", __func__, name);
return -1;
}
static void ggml_mpi_tensor_send(struct ggml_tensor * t, int mpi_rank_dst) {
MPI_Datatype mpi_type;
switch (t->type) {
case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
case GGML_TYPE_F32: mpi_type = MPI_FLOAT; break;
default: GGML_ASSERT(false && "not implemented");
}
const int retval = MPI_Send(t->data, ggml_nelements(t), mpi_type, mpi_rank_dst, 0, MPI_COMM_WORLD);
GGML_ASSERT(retval == MPI_SUCCESS);
}
static void ggml_mpi_tensor_recv(struct ggml_tensor * t, int mpi_rank_src) {
MPI_Datatype mpi_type;
switch (t->type) {
case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
case GGML_TYPE_F32: mpi_type = MPI_FLOAT; break;
default: GGML_ASSERT(false && "not implemented");
}
MPI_Status status; UNUSED(status);
const int retval = MPI_Recv(t->data, ggml_nelements(t), mpi_type, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
GGML_ASSERT(retval == MPI_SUCCESS);
}
// TODO: there are many improvements that can be done to this implementation
void ggml_mpi_graph_compute_pre(
struct ggml_mpi_context * ctx_mpi,
struct ggml_cgraph * gf,
int n_layers) {
const int mpi_rank = ctx_mpi->rank;
const int mpi_size = ctx_mpi->size;
struct ggml_tensor * inp_tokens = ggml_graph_get_tensor(gf, "inp_tokens");
if (inp_tokens == NULL) {
fprintf(stderr, "%s: tensor 'inp_tokens' not found\n", __func__);
return;
}
struct ggml_tensor * inp0 = ggml_graph_get_tensor(gf, "layer_inp_0");
if (inp0 == NULL) {
fprintf(stderr, "%s: tensor 'inp0' not found\n", __func__);
return;
}
GGML_ASSERT(inp0 == gf->nodes[0]);
// distribute the compute graph into slices across the MPI nodes
//
// the main node (0) processes the last layers + the remainder of the compute graph
// and is responsible to pass the input tokens to the first node (1)
//
// node 1: [( 0) * n_per_node, ( 1) * n_per_node)
// node 2: [( 1) * n_per_node, ( 2) * n_per_node)
// ...
// node n-1: [(n-2) * n_per_node, (n-1) * n_per_node)
// node 0: [(n-1) * n_per_node, n_nodes)
//
if (mpi_rank > 0) {
if (mpi_rank == 1) {
// the first node (1) receives the input tokens from the main node (0)
ggml_mpi_tensor_recv(inp_tokens, 0);
} else {
// recv input data for each node into the "inp0" tensor (i.e. the first node in the compute graph)
ggml_mpi_tensor_recv(inp0, mpi_rank - 1);
}
} else if (mpi_size > 1) {
// node 0 sends the input tokens to node 1
ggml_mpi_tensor_send(inp_tokens, 1);
// recv the output data from the last node
ggml_mpi_tensor_recv(inp0, mpi_size - 1);
}
{
const int n_per_node = (n_layers + (mpi_size - 1)) / mpi_size;
const int mpi_idx = mpi_rank > 0 ? mpi_rank - 1 : mpi_size - 1;
const int il0 = (mpi_idx + 0) * n_per_node;
const int il1 = MIN(n_layers, (mpi_idx + 1) * n_per_node);
char name_l0[GGML_MAX_NAME];
char name_l1[GGML_MAX_NAME];
snprintf(name_l0, sizeof(name_l0), "layer_inp_%d", il0);
snprintf(name_l1, sizeof(name_l1), "layer_inp_%d", il1);
const int idx_l0 = ggml_graph_get_node_idx(gf, name_l0);
const int idx_l1 = mpi_rank > 0 ? ggml_graph_get_node_idx(gf, name_l1) + 1 : gf->n_nodes;
if (idx_l0 < 0 || idx_l1 < 0) {
fprintf(stderr, "%s: layer input nodes not found\n", __func__);
return;
}
// attach the input data to all nodes that need it
// TODO: not great - should be able to do this without modifying the compute graph (see next TODO below)
for (int i = idx_l0; i < idx_l1; i++) {
if (gf->nodes[i]->src[0] == gf->nodes[idx_l0]) {
gf->nodes[i]->src[0] = inp0;
}
if (gf->nodes[i]->src[1] == gf->nodes[idx_l0]) {
gf->nodes[i]->src[1] = inp0;
}
}
// TODO: instead of rearranging the nodes, we should be able to execute a subset of the compute graph
for (int i = 1; i < idx_l1 - idx_l0; i++) {
gf->nodes[i] = gf->nodes[idx_l0 + i];
gf->grads[i] = gf->grads[idx_l0 + i];
}
// the first node performs the "get_rows" operation, the rest of the nodes get the data from the previous node
if (mpi_idx != 0) {
gf->nodes[0]->op = GGML_OP_NONE;
}
gf->n_nodes = idx_l1 - idx_l0;
//fprintf(stderr, "%s: node %d: processing %d nodes [%d, %d)\n", __func__, mpi_rank, gf->n_nodes, il0, il1);
}
}
void ggml_mpi_graph_compute_post(
struct ggml_mpi_context * ctx_mpi,
struct ggml_cgraph * gf,
int n_layers) {
UNUSED(n_layers);
const int mpi_rank = ctx_mpi->rank;
const int mpi_size = ctx_mpi->size;
// send the output data to the next node
if (mpi_rank > 0) {
ggml_mpi_tensor_send(gf->nodes[gf->n_nodes - 1], (mpi_rank + 1) % mpi_size);
}
}
llm/ggml-mpi.h deleted 100644 → 0
View file @ f4432e1d
//go:build mpi
/**
* llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
*
* MIT License
*
* Copyright (c) 2023 Georgi Gerganov
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#pragma once
struct ggml_context;
struct ggml_tensor;
struct ggml_cgraph;
#ifdef __cplusplus
extern "C" {
#endif
struct ggml_mpi_context;
void ggml_mpi_backend_init(void);
void ggml_mpi_backend_free(void);
struct ggml_mpi_context * ggml_mpi_init(void);
void ggml_mpi_free(struct ggml_mpi_context * ctx);
int ggml_mpi_rank(struct ggml_mpi_context * ctx);
void ggml_mpi_eval_init(
struct ggml_mpi_context * ctx_mpi,
int * n_tokens,
int * n_past,
int * n_threads);
void ggml_mpi_graph_compute_pre(
struct ggml_mpi_context * ctx_mpi,
struct ggml_cgraph * gf,
int n_layers);
void ggml_mpi_graph_compute_post(
struct ggml_mpi_context * ctx_mpi,
struct ggml_cgraph * gf,
int n_layers);
#ifdef __cplusplus
}
#endif
llm/ggml-opencl.cpp deleted 100644 → 0
View file @ f4432e1d
//go:build opencl
/**
* llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
*
* MIT License
*
* Copyright (c) 2023 Georgi Gerganov
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "ggml-opencl.h"
#include <array>
#include <atomic>
#include <sstream>
#include <vector>
#include <limits>
#define CL_TARGET_OPENCL_VERSION 110
#include <clblast.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include "ggml.h"
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
#define CL_DMMV_BLOCK_SIZE 32
#ifndef K_QUANTS_PER_ITERATION
#define K_QUANTS_PER_ITERATION 1
#else
static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
#endif
#define MULTILINE_QUOTE(...) #__VA_ARGS__
static std::string program_source = MULTILINE_QUOTE(
typedef char int8_t;
typedef uchar uint8_t;
typedef short int16_t;
typedef ushort uint16_t;
typedef int int32_t;
typedef uint uint32_t;
struct __attribute__ ((packed)) block_q4_0
{
half d;
uint8_t qs[QK4_0 / 2];
};
struct __attribute__ ((packed)) block_q4_1
{
half d;
half m;
uint8_t qs[QK4_1 / 2];
};
struct __attribute__ ((packed)) block_q5_0
{
half d;
uint32_t qh;
uint8_t qs[QK5_0 / 2];
};
struct __attribute__ ((packed)) block_q5_1
{
half d;
half m;
uint32_t qh;
uint8_t qs[QK5_1 / 2];
};
struct __attribute__ ((packed)) block_q8_0
{
half d;
int8_t qs[QK8_0];
};
struct __attribute__((packed)) block_q2_K
{
uint8_t scales[16];
uint8_t qs[64];
half d;
half dmin;
};
struct __attribute__((packed)) block_q3_K
{
uint8_t hmask[32];
uint8_t qs[64];
uint8_t scales[12];
half d;
};
struct __attribute__((packed)) block_q4_K
{
half d;
half dmin;
uint8_t scales[12];
uint8_t qs[128];
};
struct __attribute__((packed)) block_q5_K
{
half d;
half dmin;
uint8_t scales[12];
uint8_t qh[32];
uint8_t qs[128];
};
struct __attribute__((packed)) block_q6_K
{
uint8_t ql[128];
uint8_t qh[64];
int8_t scales[16];
half d;
};
__kernel void convert_fp16_to_fp32(__global half* x, __global float* y) {
const uint i = get_global_id(0);
y[i] = vload_half(0, &x[i]);
}
void dequantize_q4_0(__global const struct block_q4_0* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
const uint8_t vui = x[ib].qs[iqs];
const int8_t vi0 = vui & 0xF;
const int8_t vi1 = vui >> 4;
*v0 = (vi0 - 8)*d;
*v1 = (vi1 - 8)*d;
}
void dequantize_q4_1(__global const struct block_q4_1* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
const float m = vload_half(0, &x[ib].m);
const uint8_t vui = x[ib].qs[iqs];
const int8_t vi0 = vui & 0xF;
const int8_t vi1 = vui >> 4;
*v0 = vi0*d + m;
*v1 = vi1*d + m;
}
void dequantize_q5_0(__global const struct block_q5_0* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
uint32_t qh = x[ib].qh;
const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0) - 16;
const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1) - 16;
*v0 = x0*d;
*v1 = x1*d;
}
void dequantize_q5_1(__global const struct block_q5_1* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
const float m = vload_half(0, &x[ib].m);
uint32_t qh = x[ib].qh;
const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0);
const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1);
*v0 = x0*d + m;
*v1 = x1*d + m;
}
void dequantize_q8_0(__global const struct block_q8_0* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
const int8_t vi0 = x[ib].qs[iqs + 0];
const int8_t vi1 = x[ib].qs[iqs + 1];
*v0 = vi0*d;
*v1 = vi1*d;
}
void convert_f16(__global half* x, const int ib, const int iqs, float* v0, float* v1){
*v0 = vload_half(0, &x[ib + 0]);
*v1 = vload_half(0, &x[ib + 1]);
}
);
static std::string k_quants_source = MULTILINE_QUOTE(
inline void get_scale_min_k4(int j, const __global 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);
}
}
__kernel void dequantize_block_q2_K(__global const struct block_q2_K *x, __global float *yy)
{
const int i = get_group_id(0);
const int tid = get_local_id(0);
const int n = tid / 32;
const int l = tid - 32 * n;
const int is = 8 * n + l / 16;
const uint8_t q = x[i].qs[32 * n + l];
__global float *y = yy + i * QK_K + 128 * n;
const float dall = vload_half(0, &x[i].d);
const float dmin = vload_half(0, &x[i].dmin);
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);
}
__kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __global float *yy)
{
int r = get_local_id(0) / 4;
int i = get_group_id(0);
int tid = r / 2;
int is0 = r % 2;
int l0 = 16 * is0 + 4 * (get_local_id(0) % 4);
int n = tid / 4;
int j = tid - 4 * n;
uint8_t m = 1 << (4 * n + j);
int 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 = vload_half(0, &x[i].d);
float dl = d_all * (us - 32);
__global float *y = yy + i * QK_K + 128 * n + 32 * j;
const __global uint8_t *q = x[i].qs + 32 * n;
const __global 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));
}
__kernel void dequantize_block_q4_K(__global const struct block_q4_K *x, __global float *yy)
{
const int i = get_group_id(0);
const int tid = get_local_id(0);
const int il = tid / 8;
const int ir = tid % 8;
const int is = 2 * il;
const int n = 4;
__global float *y = yy + i * QK_K + 64 * il + n * ir;
const float dall = vload_half(0, &x[i].d);
const float dmin = vload_half(0, &x[i].dmin);
__global 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);
float d1 = dall * sc;
float m1 = dmin * m;
get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
float d2 = dall * sc;
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;
}
}
__kernel void dequantize_block_q5_K(__global const struct block_q5_K *x, __global float *yy)
{
const int i = get_group_id(0);
const int tid = get_local_id(0);
const int il = tid / 16;
const int ir = tid % 16;
const int is = 2 * il;
__global float *y = yy + i * QK_K + 64 * il + 2 * ir;
const float dall = vload_half(0, &x[i].d);
const float dmin = vload_half(0, &x[i].dmin);
__global const uint8_t *ql = x[i].qs + 32 * il + 2 * ir;
__global 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;
}
__kernel void dequantize_block_q6_K(__global const struct block_q6_K *x, __global float *yy)
{
const int i = get_group_id(0);
const int tid = get_local_id(0);
const int ip = tid / 32;
const int il = tid - 32 * ip;
const int is = 8 * ip + il / 16;
__global float *y = yy + i * QK_K + 128 * ip + il;
const float d = vload_half(0, &x[i].d);
__global const uint8_t *ql = x[i].ql + 64 * ip + il;
const uint8_t qh = x[i].qh[32 * ip + il];
__global 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);
}
__kernel void dequantize_mul_mat_vec_q2_K(__global const struct block_q2_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
const int row = get_group_id(0);
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
__global const struct block_q2_K * x = xx + ib0;
const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...31 or 0...15
const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; // 0 or 0,1
const int step = 16/K_QUANTS_PER_ITERATION;
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int in = tid - step*im; // 0...15 or 0...7
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 or 0...14 in steps of 2
const int q_offset = 32*im + l0;
const int s_offset = 8*im;
const int y_offset = 128*im + l0;
tmp[16 * ix + tid] = 0;
uint32_t aux[4];
const uint8_t * d = (const uint8_t *)aux;
const uint8_t * m = (const uint8_t *)(aux + 2);
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
__global const float * y = yy + i * QK_K + y_offset;
__global const uint8_t * q = x[i].qs + q_offset;
const float dall = vload_half(0, &x[i].d);
const float dmin = vload_half(0, &x[i].dmin);
__global const uint32_t * a = (__global const uint32_t *)(x[i].scales + s_offset);
aux[0] = a[0] & 0x0f0f0f0f;
aux[1] = a[1] & 0x0f0f0f0f;
aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
float sum1 = 0, sum2 = 0;
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
+ y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
+ y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
+ y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
+ y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
+ y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
+ y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
+y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
+ y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
}
tmp[16 * ix + tid] += dall * sum1 - dmin * sum2;
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=16; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
__kernel void dequantize_mul_mat_vec_q3_K(__global const struct block_q3_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
const uint16_t kmask1 = 0x0303;
const uint16_t kmask2 = 0x0f0f;
const int row = get_group_id(0);
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
__global const struct block_q3_K * x = xx + ib0;
const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; // 0 or 0,1
const int n = K_QUANTS_PER_ITERATION; // iterations in the inner loop
const int step = 16/K_QUANTS_PER_ITERATION;
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int in = tid - step*im; // 0....15 or 0...7
const uint8_t m = 1 << (4*im);
const int l0 = n*in; // 0...15 or 0...14 in steps of 2
const int q_offset = 32*im + l0;
const int y_offset = 128*im + l0;
uint16_t utmp[4];
const int8_t * s = (const int8_t *)utmp;
const uint16_t s_shift = 4*im;
tmp[16 * ix + tid] = 0;
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
__global const float * y = yy + i * QK_K + y_offset;
__global const uint8_t * q = x[i].qs + q_offset;
__global const uint8_t * h = x[i].hmask + l0;
__global const uint16_t * a = (__global const uint16_t *)x[i].scales;
utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
const float d = vload_half(0, &x[i].d);
float sum = 0;
for (int l = 0; l < n; ++l) {
sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
+ y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
+ y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
+ y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
+ y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
+ y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
+ y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
}
tmp[16 * ix + tid] += d * sum;
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=16; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
__kernel void dequantize_mul_mat_vec_q4_K(__global const struct block_q4_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
//to rename it later, just to test now
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int row = get_group_id(0);
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...15
const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION;
const int step = 8/K_QUANTS_PER_ITERATION;
const int il = tid/step; // 0...3
const int ir = tid - step*il;// 0...3
const int n = 2*K_QUANTS_PER_ITERATION;
const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const int in = il%2;
const int l0 = n*(2*ir + in);
const int q_offset = 32*im + l0;
const int y_offset = 64*im + l0;
uint16_t aux[4];
const uint8_t * sc = (const uint8_t *)aux;
__global const struct block_q4_K * x = xx + ib0;
tmp[16 * ix + tid] = 0;
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
__global const uint8_t * q1 = x[i].qs + q_offset;
__global const uint8_t * q2 = q1 + 64;
__global const float * y1 = yy + i*QK_K + y_offset;
__global const float * y2 = y1 + 128;
const float dall = vload_half(0, &x[i].d);
const float dmin = vload_half(0, &x[i].dmin);
__global const uint16_t * a = (__global const uint16_t *)x[i].scales;
aux[0] = a[im+0] & kmask1;
aux[1] = a[im+2] & kmask1;
aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
float4 s = (float4)(0.f);
float smin = 0;
for (int l = 0; l < n; ++l) {
s.x += y1[l] * (q1[l] & 0xF); s.y += y1[l+32] * (q1[l] >> 4);
s.z += y2[l] * (q2[l] & 0xF); s.w += y2[l+32] * (q2[l] >> 4);
smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
}
tmp[16 * ix + tid] += dall * (s.x * sc[0] + s.y * sc[1] + s.z * sc[4] + s.w * sc[5]) - dmin * smin;
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=16; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
__kernel void dequantize_mul_mat_vec_q5_K(__global const struct block_q5_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int row = get_group_id(0);
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
const int tid = get_local_id(0)/2; // 0...15
const int ix = get_local_id(0)%2;
const int il = tid/4; // 0...3
const int ir = tid - 4*il;// 0...3
const int n = 2;
const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const int in = il%2;
const int l0 = n*(2*ir + in);
const int q_offset = 32*im + l0;
const int y_offset = 64*im + l0;
const uint8_t hm1 = 1 << (2*im);
const uint8_t hm2 = hm1 << 4;
uint16_t aux[4];
const uint8_t * sc = (const uint8_t *)aux;
__global const struct block_q5_K * x = xx + ib0;
tmp[16 * ix + tid] = 0;
for (int i = ix; i < num_blocks_per_row; i += 2) {
__global const uint8_t * ql1 = x[i].qs + q_offset;
__global const uint8_t * ql2 = ql1 + 64;
__global const uint8_t * qh = x[i].qh + l0;
__global const float * y1 = yy + i*QK_K + y_offset;
__global const float * y2 = y1 + 128;
const float dall = vload_half(0, &x[i].d);
const float dmin = vload_half(0, &x[i].dmin);
__global const uint16_t * a = (__global const uint16_t *)x[i].scales;
aux[0] = a[im+0] & kmask1;
aux[1] = a[im+2] & kmask1;
aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
float4 sum = (float4)(0.f);
float smin = 0;
for (int l = 0; l < n; ++l) {
sum.x += y1[l+ 0] * ((ql1[l+ 0] & 0xF) + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
+ y1[l+16] * ((ql1[l+16] & 0xF) + (qh[l+16] & (hm1 << 0) ? 16 : 0));
sum.y += y1[l+32] * ((ql1[l+ 0] >> 4) + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
+ y1[l+48] * ((ql1[l+16] >> 4) + (qh[l+16] & (hm1 << 1) ? 16 : 0));
sum.z += y2[l+ 0] * ((ql2[l+ 0] & 0xF) + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
+ y2[l+16] * ((ql2[l+16] & 0xF) + (qh[l+16] & (hm2 << 0) ? 16 : 0));
sum.w += y2[l+32] * ((ql2[l+ 0] >> 4) + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
+ y2[l+48] * ((ql2[l+16] >> 4) + (qh[l+16] & (hm2 << 1) ? 16 : 0));
smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
+ (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
}
tmp[16 * ix + tid] += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=16; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
__kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx, __local float* tmp, __global const float * yy, __global float * dst, const int ncols) {
const int row = get_group_id(0);
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
__global const struct block_q6_K * x = xx + ib0;
const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; // 0 or 0, 1
const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int in = tid - step*im; // 0...15 or 0...7
\n#if K_QUANTS_PER_ITERATION == 1\n
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15
const int is = 0;
\n#else\n
const int l0 = 4 * in; // 0, 4, 8, ..., 28
const int is = in / 4;
\n#endif\n
const int ql_offset = 64*im + l0;
const int qh_offset = 32*im + l0;
const int s_offset = 8*im + is;
const int y_offset = 128*im + l0;
tmp[16 * ix + tid] = 0; // partial sum for thread in warp
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
__global const float * y = yy + i * QK_K + y_offset;
__global const uint8_t * ql = x[i].ql + ql_offset;
__global const uint8_t * qh = x[i].qh + qh_offset;
__global const int8_t * s = x[i].scales + s_offset;
const float d = vload_half(0, &x[i].d);
\n#if K_QUANTS_PER_ITERATION == 1\n
float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
+ y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
+ y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
+ y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
+ y[64] * s[4] * d * ((int8_t)((ql[ 0] >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
+ y[80] * s[5] * d * ((int8_t)((ql[16] >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
+ y[96] * s[6] * d * ((int8_t)((ql[32] >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
+y[112] * s[7] * d * ((int8_t)((ql[48] >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
tmp[16 * ix + tid] += sum;
\n#else\n
float sum = 0;
for (int l = 0; l < 4; ++l) {
sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
+ y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
+ y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
+ y[l+96] * s[6] * d * ((int8_t)((ql[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
}
tmp[16 * ix + tid] += sum;
\n#endif\n
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=16; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
);
std::string dequant_template = MULTILINE_QUOTE(
__kernel void KERNEL_NAME(__global X_TYPE* x, __global float* y) {
const int i = get_group_id(0)*get_local_size(0) + get_local_id(0)*2;
if (i >= get_global_size(0)) {
return;
}
const uint qk = QUANT_K;
const uint qr = QUANT_R;
const int ib = i/qk; // block index
const int iqs = (i%qk)/qr; // quant index
const int iybs = i - i%qk; // y block start index
const int y_offset = qr == 1 ? 1 : qk/2;
// dequantize
float v0, v1;
DEQUANT_FUNC(x, ib, iqs, &v0, &v1);
y[iybs + iqs + 0] = v0;
y[iybs + iqs + y_offset] = v1;
}
);
std::string dequant_mul_mat_vec_template = MULTILINE_QUOTE(
__kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float* y, __global float* dst, const int ncols) {
const int block_size = get_local_size(0);
const int row = get_group_id(0);
const int tid = get_local_id(0);
const uint qk = QUANT_K;
const uint qr = QUANT_R;
const int y_offset = qr == 1 ? 1 : qk/2;
tmp[tid] = 0;
for (int i = 0; i < ncols/block_size; i += 2) {
const int col = i*block_size + 2*tid;
const int ib = (row*ncols + col)/qk; // block index
const int iqs = (col%qk)/qr; // quant index
const int iybs = col - col%qk; // y block start index
// dequantize
float v0, v1;
DEQUANT_FUNC(x, ib, iqs, &v0, &v1);
// matrix multiplication
tmp[tid] += v0 * y[iybs + iqs + 0];
tmp[tid] += v1 * y[iybs + iqs + y_offset];
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=block_size/2; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
);
std::string mul_template = MULTILINE_QUOTE(
__kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y, const int y_offset, __global TYPE* dst, const int dst_offset, const int ky) {
const int i = get_group_id(0)*get_local_size(0) + get_local_id(0);
if (i >= get_global_size(0)) {
return;
}
dst[dst_offset + i] = x[x_offset + i] * y[y_offset + i%ky];
}
);
#define CL_CHECK(err) \
do { \
cl_int err_ = (err); \
if (err_ != CL_SUCCESS) { \
fprintf(stderr, "ggml_opencl: %s error %d at %s:%d\n", \
#err, err_, __FILE__, __LINE__); \
exit(1); \
} \
} while (0)
#define CLBLAST_CHECK(err) \
do { \
CLBlastStatusCode err_ = (err); \
if (err_ != CLBlastSuccess) { \
fprintf(stderr, "ggml_opencl: %s error %d at %s:%d\n", \
#err, err_, __FILE__, __LINE__); \
exit(1); \
} \
} while (0)
std::array<std::string, 5> dequant_str_keys = {
"KERNEL_NAME", "X_TYPE", "QUANT_K", "QUANT_R", "DEQUANT_FUNC"
};
std::array<std::string, 30> dequant_str_values = {
"dequantize_row_q4_0", "struct block_q4_0", "QK4_0", "QR4_0", "dequantize_q4_0",
"dequantize_row_q4_1", "struct block_q4_1", "QK4_1", "QR4_1", "dequantize_q4_1",
"dequantize_row_q5_0", "struct block_q5_0", "QK5_0", "QR5_0", "dequantize_q5_0",
"dequantize_row_q5_1", "struct block_q5_1", "QK5_1", "QR5_1", "dequantize_q5_1",
"dequantize_row_q8_0", "struct block_q8_0", "QK8_0", "QR8_0", "dequantize_q8_0",
"convert_row_f16", "half", "1", "1", "convert_f16"
};
std::array<std::string, 30> dequant_mul_mat_vec_str_values = {
"dequantize_mul_mat_vec_q4_0", "struct block_q4_0", "QK4_0", "QR4_0", "dequantize_q4_0",
"dequantize_mul_mat_vec_q4_1", "struct block_q4_1", "QK4_1", "QR4_1", "dequantize_q4_1",
"dequantize_mul_mat_vec_q5_0", "struct block_q5_0", "QK5_0", "QR5_0", "dequantize_q5_0",
"dequantize_mul_mat_vec_q5_1", "struct block_q5_1", "QK5_1", "QR5_1", "dequantize_q5_1",
"dequantize_mul_mat_vec_q8_0", "struct block_q8_0", "QK8_0", "QR8_0", "dequantize_q8_0",
"convert_mul_mat_vec_f16", "half", "1", "1", "convert_f16"
};
std::array<std::string, 2> mul_str_keys = {
"KERNEL_NAME", "TYPE"
};
std::array<std::string, 2> mul_str_values = {
"mul_f32", "float"
};
std::string& replace(std::string& s, const std::string& from, const std::string& to) {
size_t pos = 0;
while ((pos = s.find(from, pos)) != std::string::npos) {
s.replace(pos, from.length(), to);
pos += to.length();
}
return s;
}
std::string generate_kernels() {
std::stringstream src;
src << program_source << '\n';
src << k_quants_source << '\n';
for (size_t i = 0; i < dequant_str_values.size(); i += dequant_str_keys.size()) {
std::string dequant_kernel = dequant_template;
std::string dmmv_kernel = dequant_mul_mat_vec_template;
for (size_t j = 0; j < dequant_str_keys.size(); j++) {
replace(dequant_kernel, dequant_str_keys[j], dequant_str_values[i + j]);
replace(dmmv_kernel, dequant_str_keys[j], dequant_mul_mat_vec_str_values[i + j]);
}
src << dequant_kernel << '\n';
src << dmmv_kernel << '\n';
}
for (size_t i = 0; i < mul_str_values.size(); i += mul_str_keys.size()) {
std::string mul_kernel = mul_template;
for (size_t j = 0; j < mul_str_keys.size(); j++) {
replace(mul_kernel, mul_str_keys[j], mul_str_values[i + j]);
}
src << mul_kernel << '\n';
}
return src.str();
}
static cl_platform_id platform;
static cl_device_id device;
static cl_context context;
static cl_command_queue queue;
static cl_program program;
static cl_kernel convert_row_f16_cl;
static cl_kernel dequantize_row_q4_0_cl, dequantize_row_q4_1_cl, dequantize_row_q5_0_cl, dequantize_row_q5_1_cl, dequantize_row_q8_0_cl;
static cl_kernel dequantize_mul_mat_vec_q4_0_cl, dequantize_mul_mat_vec_q4_1_cl, dequantize_mul_mat_vec_q5_0_cl, dequantize_mul_mat_vec_q5_1_cl, dequantize_mul_mat_vec_q8_0_cl, convert_mul_mat_vec_f16_cl;
static cl_kernel dequantize_block_q2_k_cl, dequantize_block_q3_k_cl, dequantize_block_q4_k_cl, dequantize_block_q5_k_cl, dequantize_block_q6_k_cl;
static cl_kernel dequantize_mul_mat_vec_q2_K_cl, dequantize_mul_mat_vec_q3_K_cl, dequantize_mul_mat_vec_q4_K_cl, dequantize_mul_mat_vec_q5_K_cl, dequantize_mul_mat_vec_q6_K_cl;
static cl_kernel mul_f32_cl;
static bool fp16_support;
static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
cl_program p;
char *program_log;
size_t program_size;
size_t log_size;
int err;
program_size = strlen(program_buffer);
p = clCreateProgramWithSource(ctx, 1, (const char**)&program_buffer, &program_size, &err);
if(err < 0) {
fprintf(stderr, "OpenCL error creating program");
exit(1);
}
std::string compile_opts = "-cl-mad-enable -cl-unsafe-math-optimizations -cl-finite-math-only -cl-fast-relaxed-math "
"-DQK4_0=32 -DQR4_0=2 -DQK4_1=32 -DQR4_1=2 -DQK5_0=32 -DQR5_0=2 -DQK5_1=32 -DQR5_1=2 -DQK8_0=32 -DQR8_0=1 "
"-DQK_K=256 -DK_QUANTS_PER_ITERATION=" + std::to_string(K_QUANTS_PER_ITERATION);
err = clBuildProgram(p, 0, NULL, compile_opts.c_str(), NULL, NULL);
if(err < 0) {
clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
program_log = (char*) malloc(log_size + 1);
program_log[log_size] = '\0';
clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, log_size + 1, program_log, NULL);
fprintf(stderr, "ggml_opencl: kernel compile error:\n\n%s\n", program_log);
free(program_log);
exit(1);
}
return p;
}
void ggml_cl_init(void) {
cl_int err;
struct cl_device;
struct cl_platform {
cl_platform_id id;
unsigned number;
char name[128];
char vendor[128];
struct cl_device * devices;
unsigned n_devices;
struct cl_device * default_device;
};
struct cl_device {
struct cl_platform * platform;
cl_device_id id;
unsigned number;
cl_device_type type;
char name[128];
};
enum { NPLAT = 16, NDEV = 16 };
struct cl_platform platforms[NPLAT];
unsigned n_platforms = 0;
struct cl_device devices[NDEV];
unsigned n_devices = 0;
struct cl_device * default_device = NULL;
platform = NULL;
device = NULL;
cl_platform_id platform_ids[NPLAT];
CL_CHECK(clGetPlatformIDs(NPLAT, platform_ids, &n_platforms));
for (unsigned i = 0; i < n_platforms; i++) {
struct cl_platform * p = &platforms[i];
p->number = i;
p->id = platform_ids[i];
CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_NAME, sizeof(p->name), &p->name, NULL));
CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_VENDOR, sizeof(p->vendor), &p->vendor, NULL));
cl_device_id device_ids[NDEV];
cl_int clGetDeviceIDsError = clGetDeviceIDs(p->id, CL_DEVICE_TYPE_ALL, NDEV, device_ids, &p->n_devices);
if (clGetDeviceIDsError == CL_DEVICE_NOT_FOUND) {
p->n_devices = 0;
} else {
CL_CHECK(clGetDeviceIDsError);
}
p->devices = p->n_devices > 0 ? &devices[n_devices] : NULL;
p->default_device = NULL;
for (unsigned j = 0; j < p->n_devices; j++) {
struct cl_device * d = &devices[n_devices];
d->number = n_devices++;
d->id = device_ids[j];
d->platform = p;
CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_NAME, sizeof(d->name), &d->name, NULL));
CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_TYPE, sizeof(d->type), &d->type, NULL));
if (p->default_device == NULL && d->type == CL_DEVICE_TYPE_GPU) {
p->default_device = d;
}
}
if (default_device == NULL && p->default_device != NULL) {
default_device = p->default_device;
}
}
if (n_devices == 0) {
fprintf(stderr, "ggml_opencl: could find any OpenCL devices.\n");
exit(1);
}
char * user_platform_string = getenv("GGML_OPENCL_PLATFORM");
char * user_device_string = getenv("GGML_OPENCL_DEVICE");
int user_platform_number = -1;
int user_device_number = -1;
unsigned n;
if (user_platform_string != NULL && sscanf(user_platform_string, " %u", &n) == 1 && n < n_platforms) {
user_platform_number = (int)n;
}
if (user_device_string != NULL && sscanf(user_device_string, " %u", &n) == 1 && n < n_devices) {
user_device_number = (int)n;
}
if (user_platform_number != -1 && user_device_number != -1) {
cl_platform* platform = &platforms[user_platform_number];
if ((unsigned)user_device_number >= platform->n_devices) {
fprintf(stderr, "ggml_opencl: invalid device number %d\n", user_device_number);
exit(1);
}
default_device = &platform->devices[user_device_number];
} else {
struct cl_device * selected_devices = devices;
unsigned n_selected_devices = n_devices;
if (user_platform_number == -1 && user_platform_string != NULL && user_platform_string[0] != 0) {
for (unsigned i = 0; i < n_platforms; i++) {
struct cl_platform * p = &platforms[i];
if (strstr(p->name, user_platform_string) != NULL ||
strstr(p->vendor, user_platform_string) != NULL) {
user_platform_number = (int)i;
break;
}
}
if (user_platform_number == -1) {
fprintf(stderr, "ggml_opencl: no platform matching '%s' was found.\n", user_platform_string);
exit(1);
}
}
if (user_platform_number != -1) {
struct cl_platform * p = &platforms[user_platform_number];
selected_devices = p->devices;
n_selected_devices = p->n_devices;
default_device = p->default_device;
if (n_selected_devices == 0) {
fprintf(stderr, "ggml_opencl: selected platform '%s' does not have any devices.\n", p->name);
exit(1);
}
}
if (user_device_number == -1 && user_device_string != NULL && user_device_string[0] != 0) {
for (unsigned i = 0; i < n_selected_devices; i++) {
struct cl_device * d = &selected_devices[i];
if (strstr(d->name, user_device_string) != NULL) {
user_device_number = d->number;
break;
}
}
if (user_device_number == -1) {
fprintf(stderr, "ggml_opencl: no device matching '%s' was found.\n", user_device_string);
exit(1);
}
}
if (user_device_number != -1) {
selected_devices = &devices[user_device_number];
n_selected_devices = 1;
default_device = &selected_devices[0];
}
GGML_ASSERT(n_selected_devices > 0);
if (default_device == NULL) {
default_device = &selected_devices[0];
}
}
fprintf(stderr, "ggml_opencl: selecting platform: '%s'\n", default_device->platform->name);
fprintf(stderr, "ggml_opencl: selecting device: '%s'\n", default_device->name);
if (default_device->type != CL_DEVICE_TYPE_GPU) {
fprintf(stderr, "ggml_opencl: warning, not a GPU: '%s'.\n", default_device->name);
}
platform = default_device->platform->id;
device = default_device->id;
size_t ext_str_size;
clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 0, NULL, &ext_str_size);
char *ext_buffer = (char *)alloca(ext_str_size + 1);
clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL);
ext_buffer[ext_str_size] = '\0'; // ensure it is null terminated
// Check if ext_buffer contains cl_khr_fp16
fp16_support = strstr(ext_buffer, "cl_khr_fp16") != NULL;
fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
cl_context_properties properties[] = {
(intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)platform, 0
};
CL_CHECK((context = clCreateContext(properties, 1, &device, NULL, NULL, &err), err));
CL_CHECK((queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err),
(err != CL_INVALID_QUEUE_PROPERTIES && err != CL_INVALID_VALUE ? err :
(queue = clCreateCommandQueue(context, device, 0, &err), err)
)));
const std::string kernel_src = generate_kernels();
program = build_program_from_source(context, device, kernel_src.c_str());
// FP16 to FP32 kernel
CL_CHECK((convert_row_f16_cl = clCreateKernel(program, "convert_row_f16", &err), err));
// Dequantize kernels
CL_CHECK((dequantize_row_q4_0_cl = clCreateKernel(program, "dequantize_row_q4_0", &err), err));
CL_CHECK((dequantize_row_q4_1_cl = clCreateKernel(program, "dequantize_row_q4_1", &err), err));
CL_CHECK((dequantize_row_q5_0_cl = clCreateKernel(program, "dequantize_row_q5_0", &err), err));
CL_CHECK((dequantize_row_q5_1_cl = clCreateKernel(program, "dequantize_row_q5_1", &err), err));
CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
CL_CHECK((dequantize_block_q2_k_cl = clCreateKernel(program, "dequantize_block_q2_K", &err), err));
CL_CHECK((dequantize_block_q3_k_cl = clCreateKernel(program, "dequantize_block_q3_K", &err), err));
CL_CHECK((dequantize_block_q4_k_cl = clCreateKernel(program, "dequantize_block_q4_K", &err), err));
CL_CHECK((dequantize_block_q5_k_cl = clCreateKernel(program, "dequantize_block_q5_K", &err), err));
CL_CHECK((dequantize_block_q6_k_cl = clCreateKernel(program, "dequantize_block_q6_K", &err), err));
// dequant mul mat kernel
CL_CHECK((dequantize_mul_mat_vec_q4_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_0", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q4_1_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_1", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q5_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_0", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q5_1_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_1", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q8_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q8_0", &err), err));
CL_CHECK((convert_mul_mat_vec_f16_cl = clCreateKernel(program, "convert_mul_mat_vec_f16", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q2_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q2_K", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q3_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q3_K", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q4_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_K", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q5_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_K", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q6_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q6_K", &err), err));
// mul kernel
CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err));
}
static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return &dequantize_row_q4_0_cl;
case GGML_TYPE_Q4_1:
return &dequantize_row_q4_1_cl;
case GGML_TYPE_Q5_0:
return &dequantize_row_q5_0_cl;
case GGML_TYPE_Q5_1:
return &dequantize_row_q5_1_cl;
case GGML_TYPE_Q8_0:
return &dequantize_row_q8_0_cl;
case GGML_TYPE_Q2_K:
return &dequantize_block_q2_k_cl;
case GGML_TYPE_Q3_K:
return &dequantize_block_q3_k_cl;
case GGML_TYPE_Q4_K:
return &dequantize_block_q4_k_cl;
case GGML_TYPE_Q5_K:
return &dequantize_block_q5_k_cl;
case GGML_TYPE_Q6_K:
return &dequantize_block_q6_k_cl;
case GGML_TYPE_F16:
return &convert_row_f16_cl;
default:
return nullptr;
}
}
static size_t ggml_cl_global_denom(ggml_type type) {
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 1;
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
return 4;
case GGML_TYPE_Q4_K:
return 8;
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
return 4;
case GGML_TYPE_F16:
default:
return 1;
}
}
static size_t ggml_cl_local_size(ggml_type type) {
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 0;
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
return 64;
case GGML_TYPE_Q4_K:
return 32;
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
return 64;
case GGML_TYPE_F16:
default:
return 0;
}
}
static cl_kernel* ggml_get_dequantize_mul_mat_vec_cl(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return &dequantize_mul_mat_vec_q4_0_cl;
case GGML_TYPE_Q4_1:
return &dequantize_mul_mat_vec_q4_1_cl;
case GGML_TYPE_Q5_0:
return &dequantize_mul_mat_vec_q5_0_cl;
case GGML_TYPE_Q5_1:
return &dequantize_mul_mat_vec_q5_1_cl;
case GGML_TYPE_Q8_0:
return &dequantize_mul_mat_vec_q8_0_cl;
case GGML_TYPE_F16:
return &convert_mul_mat_vec_f16_cl;
case GGML_TYPE_Q2_K:
return &dequantize_mul_mat_vec_q2_K_cl;
case GGML_TYPE_Q3_K:
return &dequantize_mul_mat_vec_q3_K_cl;
case GGML_TYPE_Q4_K:
return &dequantize_mul_mat_vec_q4_K_cl;
case GGML_TYPE_Q5_K:
return &dequantize_mul_mat_vec_q5_K_cl;
case GGML_TYPE_Q6_K:
return &dequantize_mul_mat_vec_q6_K_cl;
default:
return nullptr;
}
}
// buffer pool for cl
#define MAX_CL_BUFFERS 256
struct scoped_spin_lock {
std::atomic_flag& lock;
scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
while (lock.test_and_set(std::memory_order_acquire)) {
; // spin
}
}
~scoped_spin_lock() {
lock.clear(std::memory_order_release);
}
scoped_spin_lock(const scoped_spin_lock&) = delete;
scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
};
struct cl_buffer {
cl_mem mem;
size_t size = 0;
};
static cl_buffer g_cl_buffer_pool[MAX_CL_BUFFERS];
static std::atomic_flag g_cl_pool_lock = ATOMIC_FLAG_INIT;
static cl_mem ggml_cl_pool_malloc(size_t size, size_t * actual_size) {
scoped_spin_lock lock(g_cl_pool_lock);
cl_int err;
int best_i = -1;
size_t best_size = std::numeric_limits<size_t>::max(); //smallest unused buffer that fits our needs
int worst_i = -1;
size_t worst_size = 0; //largest unused buffer seen so far
for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
cl_buffer &b = g_cl_buffer_pool[i];
if (b.size > 0 && b.size >= size && b.size < best_size)
{
best_i = i;
best_size = b.size;
}
if (b.size > 0 && b.size > worst_size)
{
worst_i = i;
worst_size = b.size;
}
}
if(best_i!=-1) //found the smallest buffer that fits our needs
{
cl_buffer& b = g_cl_buffer_pool[best_i];
cl_mem mem = b.mem;
*actual_size = b.size;
b.size = 0;
return mem;
}
if(worst_i!=-1) //no buffer that fits our needs, resize largest one to save memory
{
cl_buffer& b = g_cl_buffer_pool[worst_i];
cl_mem mem = b.mem;
b.size = 0;
clReleaseMemObject(mem);
}
cl_mem mem;
CL_CHECK((mem = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err), err));
*actual_size = size;
return mem;
}
static void ggml_cl_pool_free(cl_mem mem, size_t size) {
scoped_spin_lock lock(g_cl_pool_lock);
for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
cl_buffer& b = g_cl_buffer_pool[i];
if (b.size == 0) {
b.mem = mem;
b.size = size;
return;
}
}
fprintf(stderr, "WARNING: cl buffer pool full, increase MAX_CL_BUFFERS\n");
clReleaseMemObject(mem);
}
void ggml_cl_free_data(const struct ggml_tensor* tensor) {
if (tensor->backend != GGML_BACKEND_GPU) {
return;
}
cl_mem mem = (cl_mem)tensor->data;
clReleaseMemObject(mem);
}
static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t offset, const struct ggml_tensor * src, uint64_t i3, uint64_t i2, cl_event* ev) {
cl_int err;
const uint64_t ne0 = src->ne[0];
const uint64_t ne1 = src->ne[1];
const uint64_t nb0 = src->nb[0];
const uint64_t nb1 = src->nb[1];
const uint64_t nb2 = src->nb[2];
const uint64_t nb3 = src->nb[3];
const enum ggml_type type = src->type;
const size_t ts = ggml_type_size(type);
const size_t bs = ggml_blck_size(type);
const void * x = (const void *) ((const char *) src->data + i2*nb2 + i3*nb3);
if (nb0 == ts && nb1 == ts*ne0/bs) {
err = clEnqueueWriteBuffer(queue, dst, CL_FALSE, offset, ne1*nb1, x, 0, NULL, ev);
return err;
}
if (nb0 == ts) {
const size_t buffer_origin[3] = { offset, 0, 0 };
const size_t host_origin[3] = { 0, 0, 0 };
const size_t region[3] = { ts*ne0/bs, ne1, 1 };
err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, ts*ne0/bs, 0, nb1, 0, x, 0, NULL, ev);
return err;
}
for (uint64_t i1 = 0; i1 < ne1; i1++) {
// pretend the row is a matrix with cols=1
const size_t buffer_origin[3] = { offset, i1, 0 };
const size_t host_origin[3] = { 0, 0, 0 };
const size_t region[3] = { ts/bs, ne0, 1 };
err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, 0, 0, nb0, 0, ((const char *)x) + i1*nb0, 0, NULL, ev);
if (err != CL_SUCCESS) {
break;
}
}
return err;
}
static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
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 ne0 = ne00 * ne01 * ne02 * ne03;
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 nb10 = src1->nb[0];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
size_t x_size;
size_t d_size;
cl_mem d_X = ggml_cl_pool_malloc(ne0 * sizeof(float), &x_size); // src0
cl_mem d_Y = (cl_mem) src1->data; // src1 is already on device, broadcasted.
cl_mem d_D = ggml_cl_pool_malloc(ne0 * sizeof(float), &d_size); // dst
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
const int i0 = i03*ne02 + i02;
cl_event ev;
// copy src0 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, i0, src0, i03, i02, &ev));
if (nb10 == sizeof(float)) {
// Contiguous, avoid overhead from queueing many kernel runs
const int64_t i13 = i03%ne13;
const int64_t i12 = i02%ne12;
const int i1 = i13*ne12*ne11 + i12*ne11;
cl_int x_offset = 0;
cl_int y_offset = i1*ne10;
cl_int d_offset = 0;
size_t global = ne00 * ne01;
cl_int ky = ne10;
CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
} else {
for (int64_t i01 = 0; i01 < ne01; i01++) {
const int64_t i13 = i03%ne13;
const int64_t i12 = i02%ne12;
const int64_t i11 = i01%ne11;
const int i1 = i13*ne12*ne11 + i12*ne11 + i11;
cl_int x_offset = i01*ne00;
cl_int y_offset = i1*ne10;
cl_int d_offset = i01*ne00;
// compute
size_t global = ne00;
cl_int ky = ne10;
CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
}
}
CL_CHECK(clReleaseEvent(ev));
CL_CHECK(clFinish(queue));
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL));
}
}
ggml_cl_pool_free(d_X, x_size);
ggml_cl_pool_free(d_D, d_size);
}
void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
ggml_cl_mul_f32(src0, src1, dst);
}
static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
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 int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
size_t x_size;
size_t y_size;
size_t d_size;
cl_mem d_X;
if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
d_X = (cl_mem) src0->data;
} else {
d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
}
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
// copy data to device
if (src0->backend != GGML_BACKEND_GPU) {
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
}
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
CL_CHECK(clFinish(queue));
// compute
cl_event ev_sgemm;
clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
alpha,
d_X, 0, ne00,
d_Y, 0, ne10,
beta,
d_D, 0, ne01,
&queue, &ev_sgemm);
if (status != clblast::StatusCode::kSuccess) {
GGML_ASSERT(false);
}
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
}
}
if (src0->backend != GGML_BACKEND_GPU) {
ggml_cl_pool_free(d_X, x_size);
}
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
}
static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t /* wsize */) {
GGML_ASSERT(fp16_support);
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 int nb10 = src1->nb[0];
const int nb11 = src1->nb[1];
const int nb12 = src1->nb[2];
const int nb13 = src1->nb[3];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const ggml_fp16_t alpha = ggml_fp32_to_fp16(1.0f);
const ggml_fp16_t beta = ggml_fp32_to_fp16(0.0f);
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
size_t x_size;
size_t y_size;
size_t d_size;
cl_mem d_X;
if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
d_X = (cl_mem) src0->data;
} else {
d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
}
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * y_ne, &y_size);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * d_ne, &d_size);
bool src1_cont_rows = nb10 == sizeof(float);
bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
// copy src0 to device
if (src0->backend != GGML_BACKEND_GPU) {
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
}
// convert src1 to fp16
// TODO: use multiple threads
ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i03 * ne02 + i02);
char * src1i = (char *) src1->data + i03*nb13 + i02*nb12;
if (src1_cont_rows) {
if (src1_cont_cols) {
ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
}
else {
for (int64_t i01 = 0; i01 < ne11; i01++) {
ggml_fp32_to_fp16_row((float *) (src1i + i01*nb11), tmp + i01*ne10, ne10);
}
}
}
else {
for (int64_t i01 = 0; i01 < ne11; i01++) {
for (int64_t i00 = 0; i00 < ne10; i00++) {
// very slow due to no inlining
tmp[i01*ne10 + i00] = ggml_fp32_to_fp16(*(float *) (src1i + i01*nb11 + i00*nb10));
}
}
}
// copy src1 to device
CL_CHECK(clEnqueueWriteBuffer(queue, d_Y, false, 0, sizeof(ggml_fp16_t) * y_ne, tmp, 0, NULL, NULL));
CL_CHECK(clFinish(queue));
// compute
cl_event ev_sgemm;
clblast::StatusCode status = clblast::Gemm<cl_half>(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
alpha,
d_X, 0, ne00,
d_Y, 0, ne10,
beta,
d_D, 0, ne01,
&queue, &ev_sgemm);
if (status != clblast::StatusCode::kSuccess) {
GGML_ASSERT(false);
}
// copy dst to host, then convert to float
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
ggml_fp16_to_fp32_row(tmp, d, d_ne);
}
}
if (src0->backend != GGML_BACKEND_GPU) {
ggml_cl_pool_free(d_X, x_size);
}
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
}
static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
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 int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const ggml_type type = src0->type;
const bool mul_mat_vec = ne11 == 1;
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
const size_t q_sz = ggml_type_size(type) * x_ne / ggml_blck_size(type);
size_t x_size;
size_t y_size;
size_t d_size;
size_t q_size;
cl_mem d_X;
if (!mul_mat_vec) {
d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
}
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
cl_mem d_Q;
if (src0->backend == GGML_BACKEND_CPU) {
d_Q = ggml_cl_pool_malloc(q_sz, &q_size);
}
cl_kernel* to_fp32_cl = ggml_get_to_fp32_cl(type);
cl_kernel* dmmv = ggml_get_dequantize_mul_mat_vec_cl(type);
GGML_ASSERT(to_fp32_cl != nullptr);
const size_t global_denom = ggml_cl_global_denom(type);
const size_t local = ggml_cl_local_size(type);
size_t ev_idx = 0;
std::vector<cl_event> events;
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
// copy src0 to device if necessary
if (src0->backend == GGML_BACKEND_CPU) {
events.emplace_back();
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
} else if (src0->backend == GGML_BACKEND_GPU) {
d_Q = (cl_mem) src0->data;
} else {
GGML_ASSERT(false);
}
if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
// copy src1 to device
events.emplace_back();
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, events.data() + ev_idx++));
// compute
const size_t global = ne01 * CL_DMMV_BLOCK_SIZE;
const size_t local = CL_DMMV_BLOCK_SIZE;
const cl_int ncols = ne00;
events.emplace_back();
CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
CL_CHECK(clSetKernelArg(*dmmv, 1, sizeof(float) * local, NULL));
CL_CHECK(clSetKernelArg(*dmmv, 2, sizeof(cl_mem), &d_Y));
CL_CHECK(clSetKernelArg(*dmmv, 3, sizeof(cl_mem), &d_D));
CL_CHECK(clSetKernelArg(*dmmv, 4, sizeof(cl_int), &ncols));
CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, NULL, &global, &local, events.size() - 1, events.data(), events.data() + ev_idx++));
} else { // general dequantization kernel + CLBlast matrix matrix multiplication
// convert src0 to fp32 on device
const size_t global = x_ne / global_denom;
CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, NULL, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
// copy src1 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
events.emplace_back();
// wait for conversion
CL_CHECK(clFinish(queue));
// compute
clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
alpha,
d_X, 0, ne00,
d_Y, 0, ne10,
beta,
d_D, 0, ne01,
&queue, events.data() + ev_idx++);
if (status != clblast::StatusCode::kSuccess) {
GGML_ASSERT(false);
}
}
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &events[events.size() - 1], NULL));
for (auto *event : events) {
clReleaseEvent(event);
}
ev_idx = 0;
events.clear();
}
}
if (!mul_mat_vec) {
ggml_cl_pool_free(d_X, x_size);
}
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
if (src0->backend == GGML_BACKEND_CPU) {
ggml_cl_pool_free(d_Q, q_size);
}
}
bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
const int64_t ne10 = src1->ne[0];
const int64_t ne0 = dst->ne[0];
const int64_t ne1 = dst->ne[1];
// TODO: find the optimal values for these
if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
src1->type == GGML_TYPE_F32 &&
dst->type == GGML_TYPE_F32 &&
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU)) {
return true;
}
return false;
}
bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
// If device doesn't support FP16
if (!fp16_support) {
return false;
}
size_t src0_sz = ggml_nbytes(src0);
size_t src1_sz = ggml_nbytes(src1);
// mul_mat_q: src0 is converted to fp32 on device
size_t mul_mat_q_transfer = src0_sz + src1_sz;
// mul_mat_f16: src1 is converted to fp16 on cpu
size_t mul_mat_f16_transfer = src0_sz + sizeof(ggml_fp16_t) * ggml_nelements(src1);
// choose the smaller one to transfer to the device
// TODO: this is not always the best choice due to the overhead of converting to fp16
return mul_mat_f16_transfer < mul_mat_q_transfer;
}
void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize) {
GGML_ASSERT(ggml_cl_can_mul_mat(src0, src1, dst));
if (src0->type == GGML_TYPE_F32) {
ggml_cl_mul_mat_f32(src0, src1, dst);
}
else if (src0->type == GGML_TYPE_F16) {
if (ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
ggml_cl_mul_mat_f16(src0, src1, dst, wdata, wsize);
}
else {
ggml_cl_mul_mat_q_f32(src0, src1, dst);
}
}
else if (ggml_is_quantized(src0->type)) {
ggml_cl_mul_mat_q_f32(src0, src1, dst);
}
else {
GGML_ASSERT(false);
}
}
size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
if (ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
return ggml_nelements(src1) * sizeof(ggml_fp16_t);
}
return 0;
}
void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
const int64_t ne0 = tensor->ne[0];
const int64_t ne1 = tensor->ne[1];
const int64_t ne2 = tensor->ne[2];
const int64_t ne3 = tensor->ne[3];
const ggml_type type = tensor->type;
const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type);
size_t q_size;
cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size);
tensor->data = data;
// copy tensor to device
for (int64_t i3 = 0; i3 < ne3; i3++) {
for (int64_t i2 = 0; i2 < ne2; i2++) {
int i = i3*ne2 + i2;
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, dst, i*ne0*ne1, tensor, i3, i2, NULL));
}
}
CL_CHECK(clFinish(queue));
tensor->data = dst;
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
}
llm/ggml-opencl.h deleted 100644 → 0
View file @ f4432e1d
//go:build opencl
/**
* llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
*
* MIT License
*
* Copyright (c) 2023 Georgi Gerganov
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#pragma once
#include "ggml.h"
#ifdef __cplusplus
extern "C" {
#endif
void ggml_cl_init(void);
void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
void * ggml_cl_host_malloc(size_t size);
void ggml_cl_host_free(void * ptr);
void ggml_cl_free_data(const struct ggml_tensor* tensor);
void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
#ifdef __cplusplus
}
#endif
llm/ggml.c deleted 100644 → 0
View file @ f4432e1d
This source diff could not be displayed because it is too large. You can view the blob instead.
llm/ggml.h deleted 100644 → 0
View file @ f4432e1d
/**
* llama.cpp - git 3ebb00935f3f0522b75df49c2769ab1774b91380
*
* MIT License
*
* Copyright (c) 2023 Georgi Gerganov
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#pragma once
//
// GGML Tensor Library
//
// This documentation is still a work in progress.
// If you wish some specific topics to be covered, feel free to drop a comment:
//
// https://github.com/ggerganov/whisper.cpp/issues/40
//
// ## Overview
//
// This library implements:
//
// - a set of tensor operations
// - automatic differentiation
// - basic optimization algorithms
//
// The aim of this library is to provide a minimalistic approach for various machine learning tasks. This includes,
// but is not limited to, the following:
//
// - linear regression
// - support vector machines
// - neural networks
//
// The library allows the user to define a certain function using the available tensor operations. This function
// definition is represented internally via a computation graph. Each tensor operation in the function definition
// corresponds to a node in the graph. Having the computation graph defined, the user can choose to compute the
// function's value and/or its gradient with respect to the input variables. Optionally, the function can be optimized
// using one of the available optimization algorithms.
//
// For example, here we define the function: f(x) = a*x^2 + b
//
// {
// struct ggml_init_params params = {
// .mem_size = 16*1024*1024,
// .mem_buffer = NULL,
// };
//
// // memory allocation happens here
// struct ggml_context * ctx = ggml_init(params);
//
// struct ggml_tensor * x = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
//
// ggml_set_param(ctx, x); // x is an input variable
//
// struct ggml_tensor * a = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
// struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
// struct ggml_tensor * x2 = ggml_mul(ctx, x, x);
// struct ggml_tensor * f = ggml_add(ctx, ggml_mul(ctx, a, x2), b);
//
// ...
// }
//
// Notice that the function definition above does not involve any actual computation. The computation is performed only
// when the user explicitly requests it. For example, to compute the function's value at x = 2.0:
//
// {
// ...
//
// struct ggml_cgraph gf = ggml_build_forward(f);
//
// // set the input variable and parameter values
// ggml_set_f32(x, 2.0f);
// ggml_set_f32(a, 3.0f);
// ggml_set_f32(b, 4.0f);
//
// ggml_graph_compute_with_ctx(ctx, &gf, n_threads);
//
// printf("f = %f\n", ggml_get_f32_1d(f, 0));
//
// ...
// }
//
// The actual computation is performed in the ggml_graph_compute() function.
//
// The ggml_new_tensor_...() functions create new tensors. They are allocated in the memory buffer provided to the
// ggml_init() function. You have to be careful not to exceed the memory buffer size. Therefore, you have to know
// in advance how much memory you need for your computation. Alternatively, you can allocate a large enough memory
// and after defining the computation graph, call the ggml_used_mem() function to find out how much memory was
// actually needed.
//
// The ggml_set_param() function marks a tensor as an input variable. This is used by the automatic
// differentiation and optimization algorithms.
//
// The described approach allows to define the function graph once and then compute its forward or backward graphs
// multiple times. All computations will use the same memory buffer allocated in the ggml_init() function. This way
// the user can avoid the memory allocation overhead at runtime.
//
// The library supports multi-dimensional tensors - up to 4 dimensions. The FP16 and FP32 data types are first class
// citizens, but in theory the library can be extended to support FP8 and integer data types.
//
// Each tensor operation produces a new tensor. Initially the library was envisioned to support only the use of unary
// and binary operations. Most of the available operations fall into one of these two categories. With time, it became
// clear that the library needs to support more complex operations. The way to support these operations is not clear
// yet, but a few examples are demonstrated in the following operations:
//
// - ggml_permute()
// - ggml_conv_1d_1s()
// - ggml_conv_1d_2s()
//
// For each tensor operator, the library implements a forward and backward computation function. The forward function
// computes the output tensor value given the input tensor values. The backward function computes the adjoint of the
// input tensors given the adjoint of the output tensor. For a detailed explanation of what this means, take a
// calculus class, or watch the following video:
//
// What is Automatic Differentiation?
// https://www.youtube.com/watch?v=wG_nF1awSSY
//
//
// ## Tensor data (struct ggml_tensor)
//
// The tensors are stored in memory via the ggml_tensor struct. The structure provides information about the size of
// the tensor, the data type, and the memory buffer where the tensor data is stored. Additionally, it contains
// pointers to the "source" tensors - i.e. the tensors that were used to compute the current tensor. For example:
//
// {
// struct ggml_tensor * c = ggml_add(ctx, a, b);
//
// assert(c->src[0] == a);
// assert(c->src[1] == b);
// }
//
// The multi-dimensional tensors are stored in row-major order. The ggml_tensor struct contains fields for the
// number of elements in each dimension ("ne") as well as the number of bytes ("nb", a.k.a. stride). This allows
// to store tensors that are not contiguous in memory, which is useful for operations such as transposition and
// permutation. All tensor operations have to take the stride into account and not assume that the tensor is
// contiguous in memory.
//
// The data of the tensor is accessed via the "data" pointer. For example:
//
// {
// struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 2, 3);
//
// // a[2, 1] = 1.0f;
// *(float *) ((char *) a->data + 2*a->nb[1] + 1*a->nb[0]) = 1.0f;
//
// // a[0, 2] = 2.0f;
// *(float *) ((char *) a->data + 0*a->nb[1] + 2*a->nb[0]) = 2.0f;
//
// ...
// }
//
// Alternatively, there are helper functions, such as ggml_get_f32_1d() and ggml_set_f32_1d() that can be used.
//
// ## The matrix multiplication operator (ggml_mul_mat)
//
// TODO
//
//
// ## Multi-threading
//
// TODO
//
//
// ## Overview of ggml.c
//
// TODO
//
//
// ## SIMD optimizations
//
// TODO
//
//
// ## Debugging ggml
//
// TODO
//
//
#ifdef GGML_SHARED
# if defined(_WIN32) && !defined(__MINGW32__)
# ifdef GGML_BUILD
# define GGML_API __declspec(dllexport)
# else
# define GGML_API __declspec(dllimport)
# endif
# else
# define GGML_API __attribute__ ((visibility ("default")))
# endif
#else
# define GGML_API
#endif
// TODO: support for clang
#ifdef __GNUC__
# define GGML_DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
#elif defined(_MSC_VER)
# define GGML_DEPRECATED(func, hint) __declspec(deprecated(hint)) func
#else
# define GGML_DEPRECATED(func, hint) func
#endif
#include <stdint.h>
#include <stddef.h>
#include <stdbool.h>
#define GGML_FILE_MAGIC 0x67676d6c // "ggml"
#define GGML_FILE_VERSION 1
#define GGML_QNT_VERSION 2 // bump this on quantization format changes
#define GGML_QNT_VERSION_FACTOR 1000 // do not change this
#define GGML_MAX_DIMS 4
#define GGML_MAX_NODES 4096
#define GGML_MAX_PARAMS 256
#define GGML_MAX_CONTEXTS 64
#define GGML_MAX_SRC 6
#define GGML_MAX_NAME 48
#define GGML_MAX_OP_PARAMS 32
#define GGML_DEFAULT_N_THREADS 4
#define GGML_EXIT_SUCCESS 0
#define GGML_EXIT_ABORTED 1
#define GGML_UNUSED(x) (void)(x)
#define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
#define GGML_ASSERT(x) \
do { \
if (!(x)) { \
fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
abort(); \
} \
} while (0)
// used to copy the number of elements and stride in bytes of tensors into local variables.
// main purpose is to reduce code duplication and improve readability.
//
// example:
//
// GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
// GGML_TENSOR_LOCALS(size_t, nb1, src1, nb);
//
#define GGML_TENSOR_LOCALS_1(type, prefix, pointer, array) \
const type prefix##0 = (pointer)->array[0]; \
GGML_UNUSED(prefix##0);
#define GGML_TENSOR_LOCALS_2(type, prefix, pointer, array) \
GGML_TENSOR_LOCALS_1 (type, prefix, pointer, array) \
const type prefix##1 = (pointer)->array[1]; \
GGML_UNUSED(prefix##1);
#define GGML_TENSOR_LOCALS_3(type, prefix, pointer, array) \
GGML_TENSOR_LOCALS_2 (type, prefix, pointer, array) \
const type prefix##2 = (pointer)->array[2]; \
GGML_UNUSED(prefix##2);
#define GGML_TENSOR_LOCALS(type, prefix, pointer, array) \
GGML_TENSOR_LOCALS_3 (type, prefix, pointer, array) \
const type prefix##3 = (pointer)->array[3]; \
GGML_UNUSED(prefix##3);
#ifdef __cplusplus
extern "C" {
#endif
#ifdef __ARM_NEON
// we use the built-in 16-bit float type
typedef __fp16 ggml_fp16_t;
#else
typedef uint16_t ggml_fp16_t;
#endif
// convert FP16 <-> FP32
GGML_API float ggml_fp16_to_fp32(ggml_fp16_t x);
GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x);
GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int n);
GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n);
struct ggml_object;
struct ggml_context;
enum ggml_type {
GGML_TYPE_F32 = 0,
GGML_TYPE_F16 = 1,
GGML_TYPE_Q4_0 = 2,
GGML_TYPE_Q4_1 = 3,
// GGML_TYPE_Q4_2 = 4, support has been removed
// GGML_TYPE_Q4_3 (5) support has been removed
GGML_TYPE_Q5_0 = 6,
GGML_TYPE_Q5_1 = 7,
GGML_TYPE_Q8_0 = 8,
GGML_TYPE_Q8_1 = 9,
// k-quantizations
GGML_TYPE_Q2_K = 10,
GGML_TYPE_Q3_K = 11,
GGML_TYPE_Q4_K = 12,
GGML_TYPE_Q5_K = 13,
GGML_TYPE_Q6_K = 14,
GGML_TYPE_Q8_K = 15,
GGML_TYPE_I8,
GGML_TYPE_I16,
GGML_TYPE_I32,
GGML_TYPE_COUNT,
};
enum ggml_backend {
GGML_BACKEND_CPU = 0,
GGML_BACKEND_GPU = 10,
GGML_BACKEND_GPU_SPLIT = 20,
};
// model file types
enum ggml_ftype {
GGML_FTYPE_UNKNOWN = -1,
GGML_FTYPE_ALL_F32 = 0,
GGML_FTYPE_MOSTLY_F16 = 1, // except 1d tensors
GGML_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors
GGML_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors
GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
GGML_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors
GGML_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors
GGML_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors
GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors
GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors
GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors
GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors
GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors
};
// available tensor operations:
enum ggml_op {
GGML_OP_NONE = 0,
GGML_OP_DUP,
GGML_OP_ADD,
GGML_OP_ADD1,
GGML_OP_ACC,
GGML_OP_SUB,
GGML_OP_MUL,
GGML_OP_DIV,
GGML_OP_SQR,
GGML_OP_SQRT,
GGML_OP_LOG,
GGML_OP_SUM,
GGML_OP_SUM_ROWS,
GGML_OP_MEAN,
GGML_OP_ARGMAX,
GGML_OP_REPEAT,
GGML_OP_REPEAT_BACK,
GGML_OP_SILU_BACK,
GGML_OP_NORM, // normalize
GGML_OP_RMS_NORM,
GGML_OP_RMS_NORM_BACK,
GGML_OP_MUL_MAT,
GGML_OP_OUT_PROD,
GGML_OP_SCALE,
GGML_OP_SET,
GGML_OP_CPY,
GGML_OP_CONT,
GGML_OP_RESHAPE,
GGML_OP_VIEW,
GGML_OP_PERMUTE,
GGML_OP_TRANSPOSE,
GGML_OP_GET_ROWS,
GGML_OP_GET_ROWS_BACK,
GGML_OP_DIAG,
GGML_OP_DIAG_MASK_INF,
GGML_OP_DIAG_MASK_ZERO,
GGML_OP_SOFT_MAX,
GGML_OP_SOFT_MAX_BACK,
GGML_OP_ROPE,
GGML_OP_ROPE_BACK,
GGML_OP_ALIBI,
GGML_OP_CLAMP,
GGML_OP_CONV_1D,
GGML_OP_CONV_2D,
GGML_OP_POOL_1D,
GGML_OP_POOL_2D,
GGML_OP_FLASH_ATTN,
GGML_OP_FLASH_FF,
GGML_OP_FLASH_ATTN_BACK,
GGML_OP_WIN_PART,
GGML_OP_WIN_UNPART,
GGML_OP_UNARY,
GGML_OP_MAP_UNARY,
GGML_OP_MAP_BINARY,
GGML_OP_MAP_CUSTOM1_F32,
GGML_OP_MAP_CUSTOM2_F32,
GGML_OP_MAP_CUSTOM3_F32,
GGML_OP_MAP_CUSTOM1,
GGML_OP_MAP_CUSTOM2,
GGML_OP_MAP_CUSTOM3,
GGML_OP_CROSS_ENTROPY_LOSS,
GGML_OP_CROSS_ENTROPY_LOSS_BACK,
GGML_OP_COUNT,
};
enum ggml_unary_op {
GGML_UNARY_OP_ABS,
GGML_UNARY_OP_SGN,
GGML_UNARY_OP_NEG,
GGML_UNARY_OP_STEP,
GGML_UNARY_OP_TANH,
GGML_UNARY_OP_ELU,
GGML_UNARY_OP_RELU,
GGML_UNARY_OP_GELU,
GGML_UNARY_OP_GELU_QUICK,
GGML_UNARY_OP_SILU,
};
enum ggml_object_type {
GGML_OBJECT_TENSOR,
GGML_OBJECT_GRAPH,
GGML_OBJECT_WORK_BUFFER
};
// ggml object
struct ggml_object {
size_t offs;
size_t size;
struct ggml_object * next;
enum ggml_object_type type;
char padding[4];
};
static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
// n-dimensional tensor
struct ggml_tensor {
enum ggml_type type;
enum ggml_backend backend;
int n_dims;
int64_t ne[GGML_MAX_DIMS]; // number of elements
size_t nb[GGML_MAX_DIMS]; // stride in bytes:
// nb[0] = sizeof(type)
// nb[1] = nb[0] * ne[0] + padding
// nb[i] = nb[i-1] * ne[i-1]
// compute data
enum ggml_op op;
// op params - allocated as int32_t for alignment
int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
bool is_param;
struct ggml_tensor * grad;
struct ggml_tensor * src[GGML_MAX_SRC];
// performance
int perf_runs;
int64_t perf_cycles;
int64_t perf_time_us;
void * data;
char name[GGML_MAX_NAME];
void * extra; // extra things e.g. for ggml-cuda.cu
char padding[4];
};
static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
// the compute plan that needs to be prepared for ggml_graph_compute()
// since https://github.com/ggerganov/ggml/issues/287
struct ggml_cplan {
size_t work_size; // size of work buffer, calculated by `ggml_graph_plan()`
uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
int n_threads;
// the `n_tasks` of nodes, 1:1 mapping to cgraph nodes
int n_tasks[GGML_MAX_NODES];
// abort ggml_graph_compute when true
bool (*abort_callback)(void * data);
void * abort_callback_data;
};
// next prime after GGML_MAX_NODES
// #define GGML_GRAPH_HASHTABLE_SIZE 4099
// next prime after GGML_MAX_NODES * 2 (nodes + leafs)
#define GGML_GRAPH_HASHTABLE_SIZE 8273
// computation graph
struct ggml_cgraph {
int n_nodes;
int n_leafs;
struct ggml_tensor * nodes[GGML_MAX_NODES];
struct ggml_tensor * grads[GGML_MAX_NODES];
struct ggml_tensor * leafs[GGML_MAX_NODES];
void * visited_hash_table[GGML_GRAPH_HASHTABLE_SIZE];
// performance
int perf_runs;
int64_t perf_cycles;
int64_t perf_time_us;
};
static const size_t GGML_GRAPH_SIZE = sizeof(struct ggml_cgraph);
// scratch buffer
struct ggml_scratch {
size_t offs;
size_t size;
void * data;
};
struct ggml_init_params {
// memory pool
size_t mem_size; // bytes
void * mem_buffer; // if NULL, memory will be allocated internally
bool no_alloc; // don't allocate memory for the tensor data
};
// compute types
// NOTE: the INIT or FINALIZE pass is not scheduled unless explicitly enabled.
// This behavior was changed since https://github.com/ggerganov/llama.cpp/pull/1995.
enum ggml_task_type {
GGML_TASK_INIT = 0,
GGML_TASK_COMPUTE,
GGML_TASK_FINALIZE,
};
struct ggml_compute_params {
enum ggml_task_type type;
// ith = thread index, nth = number of threads
int ith, nth;
// work buffer for all threads
size_t wsize;
void * wdata;
};
// misc
GGML_API void ggml_time_init(void); // call this once at the beginning of the program
GGML_API int64_t ggml_time_ms(void);
GGML_API int64_t ggml_time_us(void);
GGML_API int64_t ggml_cycles(void);
GGML_API int64_t ggml_cycles_per_ms(void);
GGML_API void ggml_numa_init(void); // call once for better performance on NUMA systems
GGML_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node
GGML_API void ggml_print_object (const struct ggml_object * obj);
GGML_API void ggml_print_objects(const struct ggml_context * ctx);
GGML_API int64_t ggml_nelements (const struct ggml_tensor * tensor);
GGML_API int64_t ggml_nrows (const struct ggml_tensor * tensor);
GGML_API size_t ggml_nbytes (const struct ggml_tensor * tensor);
GGML_API size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split);
GGML_API int ggml_blck_size (enum ggml_type type);
GGML_API size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
GGML_API float ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
GGML_API const char * ggml_type_name(enum ggml_type type);
GGML_API const char * ggml_op_name (enum ggml_op op);
GGML_API const char * ggml_op_symbol(enum ggml_op op);
GGML_API size_t ggml_element_size(const struct ggml_tensor * tensor);
GGML_API bool ggml_is_quantized(enum ggml_type type);
// TODO: temporary until model loading of ggml examples is refactored
GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
GGML_API bool ggml_is_permuted (const struct ggml_tensor * tensor);
GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
// use this to compute the memory overhead of a tensor
GGML_API size_t ggml_tensor_overhead(void);
// main
GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
GGML_API void ggml_free(struct ggml_context * ctx);
GGML_API size_t ggml_used_mem(const struct ggml_context * ctx);
GGML_API size_t ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
GGML_API bool ggml_get_no_alloc(struct ggml_context * ctx);
GGML_API void ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
GGML_API void * ggml_get_mem_buffer (const struct ggml_context * ctx);
GGML_API size_t ggml_get_mem_size (const struct ggml_context * ctx);
GGML_API size_t ggml_get_max_tensor_size(const struct ggml_context * ctx);
GGML_API struct ggml_tensor * ggml_new_tensor(
struct ggml_context * ctx,
enum ggml_type type,
int n_dims,
const int64_t *ne);
GGML_API struct ggml_tensor * ggml_new_tensor_1d(
struct ggml_context * ctx,
enum ggml_type type,
int64_t ne0);
GGML_API struct ggml_tensor * ggml_new_tensor_2d(
struct ggml_context * ctx,
enum ggml_type type,
int64_t ne0,
int64_t ne1);
GGML_API struct ggml_tensor * ggml_new_tensor_3d(
struct ggml_context * ctx,
enum ggml_type type,
int64_t ne0,
int64_t ne1,
int64_t ne2);
GGML_API struct ggml_tensor * ggml_new_tensor_4d(
struct ggml_context * ctx,
enum ggml_type type,
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3);
GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src);
GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
GGML_API void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
GGML_API float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
GGML_API void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
GGML_API void * ggml_get_data (const struct ggml_tensor * tensor);
GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
GGML_API const char * ggml_get_name (const struct ggml_tensor * tensor);
GGML_API struct ggml_tensor * ggml_set_name ( struct ggml_tensor * tensor, const char * name);
GGML_API struct ggml_tensor * ggml_format_name( struct ggml_tensor * tensor, const char * fmt, ...);
//
// operations on tensors with backpropagation
//
GGML_API struct ggml_tensor * ggml_dup(
struct ggml_context * ctx,
struct ggml_tensor * a);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_dup_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_add(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_add_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_add1(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_add1_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_acc(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
size_t nb1,
size_t nb2,
size_t nb3,
size_t offset);
GGML_API struct ggml_tensor * ggml_acc_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
size_t nb1,
size_t nb2,
size_t nb3,
size_t offset);
GGML_API struct ggml_tensor * ggml_sub(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_sub_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_mul(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_mul_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_div(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_div_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_sqr(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_sqr_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_sqrt(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_sqrt_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_log(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_log_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
// return scalar
GGML_API struct ggml_tensor * ggml_sum(
struct ggml_context * ctx,
struct ggml_tensor * a);
// sums along rows, with input shape [a,b,c,d] return shape [1,b,c,d]
GGML_API struct ggml_tensor * ggml_sum_rows(
struct ggml_context * ctx,
struct ggml_tensor * a);
// mean along rows
GGML_API struct ggml_tensor * ggml_mean(
struct ggml_context * ctx,
struct ggml_tensor * a);
// argmax along rows
GGML_API struct ggml_tensor * ggml_argmax(
struct ggml_context * ctx,
struct ggml_tensor * a);
// if a is the same shape as b, and a is not parameter, return a
// otherwise, return a new tensor: repeat(a) to fit in b
GGML_API struct ggml_tensor * ggml_repeat(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_repeat_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_abs(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_abs_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_sgn(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_sgn_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_neg(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_neg_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_step(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_step_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_tanh(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_tanh_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_elu(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_elu_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_relu(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_relu_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
// TODO: double-check this computation is correct
GGML_API struct ggml_tensor * ggml_gelu(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_gelu_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_gelu_quick(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_gelu_quick_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_silu(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_silu_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
// a - x
// b - dy
GGML_API struct ggml_tensor * ggml_silu_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// normalize along rows
// TODO: eps is hardcoded to 1e-5 for now
GGML_API struct ggml_tensor * ggml_norm(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_norm_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_rms_norm(
struct ggml_context * ctx,
struct ggml_tensor * a,
float eps);
GGML_API struct ggml_tensor * ggml_rms_norm_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
float eps);
// a - x
// b - dy
// TODO: update with configurable eps
GGML_API struct ggml_tensor * ggml_rms_norm_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// A: n columns, m rows
// B: n columns, p rows (i.e. we transpose it internally)
// result is m columns, p rows
GGML_API struct ggml_tensor * ggml_mul_mat(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// A: m columns, n rows,
// B: p columns, n rows,
// result is m columns, p rows
GGML_API struct ggml_tensor * ggml_out_prod(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
//
// operations on tensors without backpropagation
//
GGML_API struct ggml_tensor * ggml_scale(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_scale_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// b -> view(a,offset,nb1,nb2,3), return modified a
GGML_API struct ggml_tensor * ggml_set(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
size_t nb1,
size_t nb2,
size_t nb3,
size_t offset);
// b -> view(a,offset,nb1,nb2,3), return view(a)
GGML_API struct ggml_tensor * ggml_set_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
size_t nb1,
size_t nb2,
size_t nb3,
size_t offset);
GGML_API struct ggml_tensor * ggml_set_1d(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
size_t offset);
GGML_API struct ggml_tensor * ggml_set_1d_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
size_t offset);
// b -> view(a,offset,nb1,nb2,3), return modified a
GGML_API struct ggml_tensor * ggml_set_2d(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
size_t nb1,
size_t offset);
// b -> view(a,offset,nb1,nb2,3), return view(a)
GGML_API struct ggml_tensor * ggml_set_2d_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
size_t nb1,
size_t offset);
// a -> b, return view(b)
GGML_API struct ggml_tensor * ggml_cpy(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// a -> b, in-place, return view(b)
GGML_API struct ggml_tensor * ggml_cpy_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// make contiguous
GGML_API struct ggml_tensor * ggml_cont(
struct ggml_context * ctx,
struct ggml_tensor * a);
// make contiguous, in-place
GGML_API struct ggml_tensor * ggml_cont_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
// return view(a), b specifies the new shape
// TODO: when we start computing gradient, make a copy instead of view
GGML_API struct ggml_tensor * ggml_reshape(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// return view(a)
// TODO: when we start computing gradient, make a copy instead of view
GGML_API struct ggml_tensor * ggml_reshape_1d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0);
GGML_API struct ggml_tensor * ggml_reshape_2d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1);
// return view(a)
// TODO: when we start computing gradient, make a copy instead of view
GGML_API struct ggml_tensor * ggml_reshape_3d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
int64_t ne2);
GGML_API struct ggml_tensor * ggml_reshape_4d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3);
// offset in bytes
GGML_API struct ggml_tensor * ggml_view_1d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
size_t offset);
GGML_API struct ggml_tensor * ggml_view_2d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
size_t nb1, // row stride in bytes
size_t offset);
GGML_API struct ggml_tensor * ggml_view_3d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
int64_t ne2,
size_t nb1, // row stride in bytes
size_t nb2, // slice stride in bytes
size_t offset);
GGML_API struct ggml_tensor * ggml_view_4d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3,
size_t nb1, // row stride in bytes
size_t nb2, // slice stride in bytes
size_t nb3,
size_t offset);
GGML_API struct ggml_tensor * ggml_permute(
struct ggml_context * ctx,
struct ggml_tensor * a,
int axis0,
int axis1,
int axis2,
int axis3);
// alias for ggml_permute(ctx, a, 1, 0, 2, 3)
GGML_API struct ggml_tensor * ggml_transpose(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_get_rows(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_get_rows_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c);
GGML_API struct ggml_tensor * ggml_diag(
struct ggml_context * ctx,
struct ggml_tensor * a);
// set elements above the diagonal to -INF
GGML_API struct ggml_tensor * ggml_diag_mask_inf(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_diag_mask_inf_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past);
// set elements above the diagonal to 0
GGML_API struct ggml_tensor * ggml_diag_mask_zero(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_diag_mask_zero_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past);
GGML_API struct ggml_tensor * ggml_soft_max(
struct ggml_context * ctx,
struct ggml_tensor * a);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_soft_max_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_soft_max_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_soft_max_back_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// rotary position embedding
// if mode & 1 == 1, skip n_past elements
// if mode & 2 == 1, GPT-NeoX style
// if mode & 4 == 1, ChatGLM style
// TODO: avoid creating a new tensor every time
GGML_API struct ggml_tensor * ggml_rope(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past,
int n_dims,
int mode,
int n_ctx);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_rope_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past,
int n_dims,
int mode,
int n_ctx);
// custom RoPE
GGML_API struct ggml_tensor * ggml_rope_custom(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past,
int n_dims,
int mode,
int n_ctx,
float freq_base,
float freq_scale);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past,
int n_dims,
int mode,
int n_ctx,
float freq_base,
float freq_scale);
// rotary position embedding backward, i.e compute dx from dy
// a - dy
GGML_API struct ggml_tensor * ggml_rope_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past,
int n_dims,
int mode,
int n_ctx);
// alibi position embedding
// in-place, returns view(a)
struct ggml_tensor * ggml_alibi(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past,
int n_head,
float bias_max);
// clamp
// in-place, returns view(a)
struct ggml_tensor * ggml_clamp(
struct ggml_context * ctx,
struct ggml_tensor * a,
float min,
float max);
GGML_API struct ggml_tensor * ggml_conv_1d(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int s0, // stride
int p0, // padding
int d0); // dilation
GGML_API struct ggml_tensor * ggml_conv_2d(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int s0,
int s1,
int p0,
int p1,
int d0,
int d1);
// conv_1d with padding = half
// alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
GGML_API struct ggml_tensor * ggml_conv_1d_ph(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int s,
int d);
enum ggml_op_pool {
GGML_OP_POOL_MAX,
GGML_OP_POOL_AVG,
GGML_OP_POOL_COUNT,
};
GGML_API struct ggml_tensor * ggml_pool_1d(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_op_pool op,
int k0, // kernel size
int s0, // stride
int p0); // padding
GGML_API struct ggml_tensor * ggml_pool_2d(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_op_pool op,
int k0,
int k1,
int s0,
int s1,
int p0,
int p1);
GGML_API struct ggml_tensor * ggml_flash_attn(
struct ggml_context * ctx,
struct ggml_tensor * q,
struct ggml_tensor * k,
struct ggml_tensor * v,
bool masked);
GGML_API struct ggml_tensor * ggml_flash_attn_back(
struct ggml_context * ctx,
struct ggml_tensor * q,
struct ggml_tensor * k,
struct ggml_tensor * v,
struct ggml_tensor * d,
bool masked);
GGML_API struct ggml_tensor * ggml_flash_ff(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b0,
struct ggml_tensor * b1,
struct ggml_tensor * c0,
struct ggml_tensor * c1);
// partition into non-overlapping windows with padding if needed
// example:
// a: 768 64 64 1
// w: 14
// res: 768 14 14 25
// used in sam
GGML_API struct ggml_tensor * ggml_win_part(
struct ggml_context * ctx,
struct ggml_tensor * a,
int w);
// reverse of ggml_win_part
// used in sam
GGML_API struct ggml_tensor * ggml_win_unpart(
struct ggml_context * ctx,
struct ggml_tensor * a,
int w0,
int h0,
int w);
GGML_API struct ggml_tensor * ggml_unary(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_unary_op op);
GGML_API struct ggml_tensor * ggml_unary_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_unary_op op);
// custom operators
typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_unary_op_f32_t fun),
"use ggml_map_custom1 instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_unary_op_f32_t fun),
"use ggml_map_custom1_inplace instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_binary_op_f32_t fun),
"use ggml_map_custom2 instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_binary_op_f32_t fun),
"use ggml_map_custom2_inplace instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_custom1_op_f32_t fun),
"use ggml_map_custom1 instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_custom1_op_f32_t fun),
"use ggml_map_custom1_inplace instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_custom2_op_f32_t fun),
"use ggml_map_custom2 instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_custom2_op_f32_t fun),
"use ggml_map_custom2_inplace instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
ggml_custom3_op_f32_t fun),
"use ggml_map_custom3 instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
ggml_custom3_op_f32_t fun),
"use ggml_map_custom3_inplace instead");
// custom operators v2
typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
#define GGML_N_TASKS_MAX -1
GGML_API struct ggml_tensor * ggml_map_custom1(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_custom1_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom1_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_custom1_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom2(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_custom2_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom2_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_custom2_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom3(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
ggml_custom3_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_map_custom3_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
ggml_custom3_op_t fun,
int n_tasks,
void * userdata);
// loss function
GGML_API struct ggml_tensor * ggml_cross_entropy_loss(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c);
//
// automatic differentiation
//
GGML_API void ggml_set_param(
struct ggml_context * ctx,
struct ggml_tensor * tensor);
GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor);
GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep);
// graph allocation in a context
GGML_API struct ggml_cgraph * ggml_new_graph (struct ggml_context * ctx);
GGML_API struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor);
GGML_API size_t ggml_graph_overhead(void);
// ggml_graph_plan() has to be called before ggml_graph_compute()
// when plan.work_size > 0, caller must allocate memory for plan.work_data
GGML_API struct ggml_cplan ggml_graph_plan (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
GGML_API int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph);
// same as ggml_graph_compute() but the work data is allocated as a part of the context
// note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
GGML_API void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
GGML_API void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
GGML_API struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
// print info and performance information for the graph
GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
// dump the graph into a file using the dot format
GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
//
// optimization
//
// optimization methods
enum ggml_opt_type {
GGML_OPT_ADAM,
GGML_OPT_LBFGS,
};
// linesearch methods
enum ggml_linesearch {
GGML_LINESEARCH_DEFAULT = 1,
GGML_LINESEARCH_BACKTRACKING_ARMIJO = 0,
GGML_LINESEARCH_BACKTRACKING_WOLFE = 1,
GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2,
};
// optimization return values
enum ggml_opt_result {
GGML_OPT_OK = 0,
GGML_OPT_DID_NOT_CONVERGE,
GGML_OPT_NO_CONTEXT,
GGML_OPT_INVALID_WOLFE,
GGML_OPT_FAIL,
GGML_LINESEARCH_FAIL = -128,
GGML_LINESEARCH_MINIMUM_STEP,
GGML_LINESEARCH_MAXIMUM_STEP,
GGML_LINESEARCH_MAXIMUM_ITERATIONS,
GGML_LINESEARCH_INVALID_PARAMETERS,
};
// optimization parameters
//
// see ggml.c (ggml_opt_default_params) for default values
//
struct ggml_opt_params {
enum ggml_opt_type type;
int n_threads;
// delta-based convergence test
//
// if past == 0 - disabled
// if past > 0:
// stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|)
//
int past;
float delta;
// maximum number of iterations without improvement
//
// if 0 - disabled
// if > 0:
// assume convergence if no cost improvement in this number of iterations
//
int max_no_improvement;
bool print_forward_graph;
bool print_backward_graph;
// ADAM parameters
struct {
int n_iter;
float sched; // schedule multiplier (fixed, decay or warmup)
float decay; // weight decay for AdamW, use 0.0f to disable
float alpha; // learning rate
float beta1;
float beta2;
float eps; // epsilon for numerical stability
float eps_f; // epsilon for convergence test
float eps_g; // epsilon for convergence test
} adam;
// LBFGS parameters
struct {
int m; // number of corrections to approximate the inv. Hessian
int n_iter;
int max_linesearch;
float eps; // convergence tolerance
float ftol; // line search tolerance
float wolfe;
float min_step;
float max_step;
enum ggml_linesearch linesearch;
} lbfgs;
};
struct ggml_opt_context {
struct ggml_context * ctx;
struct ggml_opt_params params;
int iter;
int64_t nx; // number of parameter elements
bool just_initialized;
struct {
struct ggml_tensor * x; // view of the parameters
struct ggml_tensor * g1; // gradient
struct ggml_tensor * g2; // gradient squared
struct ggml_tensor * m; // first moment
struct ggml_tensor * v; // second moment
struct ggml_tensor * mh; // first moment hat
struct ggml_tensor * vh; // second moment hat
struct ggml_tensor * pf; // past function values
float fx_best;
float fx_prev;
int n_no_improvement;
} adam;
struct {
struct ggml_tensor * x; // current parameters
struct ggml_tensor * xp; // previous parameters
struct ggml_tensor * g; // current gradient
struct ggml_tensor * gp; // previous gradient
struct ggml_tensor * d; // search direction
struct ggml_tensor * pf; // past function values
struct ggml_tensor * lmal; // the L-BFGS memory alpha
struct ggml_tensor * lmys; // the L-BFGS memory ys
struct ggml_tensor * lms; // the L-BFGS memory s
struct ggml_tensor * lmy; // the L-BFGS memory y
float fx_best;
float step;
int j;
int k;
int end;
int n_no_improvement;
} lbfgs;
};
GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
// optimize the function defined by the tensor f
GGML_API enum ggml_opt_result ggml_opt(
struct ggml_context * ctx,
struct ggml_opt_params params,
struct ggml_tensor * f);
// initialize optimizer context
GGML_API void ggml_opt_init(
struct ggml_context * ctx,
struct ggml_opt_context * opt,
struct ggml_opt_params params,
int64_t nx);
// continue optimizing the function defined by the tensor f
GGML_API enum ggml_opt_result ggml_opt_resume(
struct ggml_context * ctx,
struct ggml_opt_context * opt,
struct ggml_tensor * f);
// continue optimizing the function defined by the tensor f
GGML_API enum ggml_opt_result ggml_opt_resume_g(
struct ggml_context * ctx,
struct ggml_opt_context * opt,
struct ggml_tensor * f,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb);
//
// quantization
//
GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);
//
// system info
//
GGML_API int ggml_cpu_has_avx (void);
GGML_API int ggml_cpu_has_avx2 (void);
GGML_API int ggml_cpu_has_avx512 (void);
GGML_API int ggml_cpu_has_avx512_vbmi(void);
GGML_API int ggml_cpu_has_avx512_vnni(void);
GGML_API int ggml_cpu_has_fma (void);
GGML_API int ggml_cpu_has_neon (void);
GGML_API int ggml_cpu_has_arm_fma (void);
GGML_API int ggml_cpu_has_f16c (void);
GGML_API int ggml_cpu_has_fp16_va (void);
GGML_API int ggml_cpu_has_wasm_simd (void);
GGML_API int ggml_cpu_has_blas (void);
GGML_API int ggml_cpu_has_cublas (void);
GGML_API int ggml_cpu_has_clblast (void);
GGML_API int ggml_cpu_has_gpublas (void);
GGML_API int ggml_cpu_has_sse3 (void);
GGML_API int ggml_cpu_has_vsx (void);
//
// Internal types and functions exposed for tests and benchmarks
//
#ifdef __cplusplus
// restrict not standard in C++
#define GGML_RESTRICT
#else
#define GGML_RESTRICT restrict
#endif
typedef void (*ggml_to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
typedef void (*ggml_vec_dot_t) (const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
typedef struct {
ggml_to_float_t to_float;
ggml_from_float_t from_float;
ggml_from_float_t from_float_reference;
ggml_vec_dot_t vec_dot;
enum ggml_type vec_dot_type;
} ggml_type_traits_t;
ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type i);
#ifdef __cplusplus
}
#endif
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