Skip to content

GitLab

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

init

parent 97b02a89
Hide whitespace changes
Inline Side-by-side
Showing with 0 additions and 135463 deletions
llm/llama.cpp/ggml-sycl.cpp deleted 100644 → 0
View file @ 97b02a89
This source diff could not be displayed because it is too large. You can view the blob instead.
llm/llama.cpp/ggml-sycl.h deleted 100644 → 0
View file @ 97b02a89
//
// MIT license
// Copyright (C) 2024 Intel Corporation
// SPDX-License-Identifier: MIT
//
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#ifdef __cplusplus
extern "C" {
#endif
#define GGML_SYCL_MAX_DEVICES 48
#define GGML_SYCL_NAME "SYCL"
// backend API
GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
// devide buffer
GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
// split tensor buffer that splits matrices by rows across multiple devices
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
GGML_API void ggml_backend_sycl_print_sycl_devices(void);
GGML_API GGML_CALL void ggml_sycl_get_gpu_list(int *id_list, int max_len);
GGML_API GGML_CALL void ggml_sycl_get_device_description(int device, char *description, size_t description_size);
GGML_API GGML_CALL int ggml_backend_sycl_get_device_count();
GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);
// TODO: these are temporary
// ref: https://github.com/ggerganov/llama.cpp/pull/6022#issuecomment-1992615670
GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index);
GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id);
GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode();
// SYCL doesn't support registering host memory, keep here for reference
// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer);
#ifdef __cplusplus
}
#endif
llm/llama.cpp/ggml-vulkan-shaders.hpp deleted 100644 → 0
View file @ 97b02a89
This source diff could not be displayed because it is too large. You can view the blob instead.
llm/llama.cpp/ggml-vulkan.cpp deleted 100644 → 0
View file @ 97b02a89
This source diff could not be displayed because it is too large. You can view the blob instead.
llm/llama.cpp/ggml-vulkan.h deleted 100644 → 0
View file @ 97b02a89
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#ifdef __cplusplus
extern "C" {
#endif
#define GGML_VK_NAME "Vulkan"
#define GGML_VK_MAX_DEVICES 16
GGML_API void ggml_vk_instance_init(void);
// backend API
GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num);
GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend);
GGML_API GGML_CALL int ggml_backend_vk_get_device_count(void);
GGML_API GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
GGML_API GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
#ifdef __cplusplus
}
#endif
llm/llama.cpp/ggml.c deleted 100644 → 0
View file @ 97b02a89
This source diff could not be displayed because it is too large. You can view the blob instead.
llm/llama.cpp/ggml.h deleted 100644 → 0
View file @ 97b02a89
#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_new_graph(ctx);
// ggml_build_forward_expand(gf, 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:
//
// {
// const int nx = 2;
// const int ny = 3;
//
// struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nx, ny);
//
// for (int y = 0; y < ny; y++) {
// for (int x = 0; x < nx; x++) {
// *(float *) ((char *) a->data + y*a->nb[1] + x*a->nb[0]) = x + y;
// }
// }
//
// ...
// }
//
// 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
#ifdef GGML_MULTIPLATFORM
# if defined(_WIN32)
# define GGML_CALL
# else
# define GGML_CALL __attribute__((__ms_abi__))
# endif
#else
# define GGML_CALL
#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
#ifndef __GNUC__
# define GGML_ATTRIBUTE_FORMAT(...)
#elif defined(__MINGW32__)
# define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
#else
# define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
#endif
#include <stdbool.h>
#include <stddef.h>
#include <stdint.h>
#include <stdio.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_PARAMS 2048
#define GGML_MAX_CONTEXTS 64
#define GGML_MAX_SRC 10
#ifndef GGML_MAX_NAME
#define GGML_MAX_NAME 64
#endif
#define GGML_MAX_OP_PARAMS 64
#define GGML_DEFAULT_N_THREADS 4
#define GGML_DEFAULT_GRAPH_SIZE 2048
#if UINTPTR_MAX == 0xFFFFFFFF
#define GGML_MEM_ALIGN 4
#else
#define GGML_MEM_ALIGN 16
#endif
#define GGML_EXIT_SUCCESS 0
#define GGML_EXIT_ABORTED 1
#define GGUF_MAGIC "GGUF"
#define GGUF_VERSION 3
#define GGUF_DEFAULT_ALIGNMENT 32
#define GGML_UNUSED(x) (void)(x)
#define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
#define GGML_ASSERT(x) \
do { \
if (!(x)) { \
fflush(stdout); \
fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
ggml_print_backtrace(); \
abort(); \
} \
} while (0)
#ifndef NDEBUG
#define GGML_UNREACHABLE() GGML_ASSERT(!"statement should not be reached")
#elif defined(__GNUC__)
#define GGML_UNREACHABLE() __builtin_unreachable()
#elif defined(_MSC_VER)
#define GGML_UNREACHABLE() __assume(0)
#else
#define GGML_UNREACHABLE() ((void) 0)
#endif
// 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);
#define GGML_TENSOR_UNARY_OP_LOCALS \
GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) \
GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) \
GGML_TENSOR_LOCALS(size_t, nb, dst, nb)
#define GGML_TENSOR_BINARY_OP_LOCALS \
GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) \
GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
GGML_TENSOR_LOCALS(size_t, nb1, src1, nb) \
GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) \
GGML_TENSOR_LOCALS(size_t, nb, dst, nb)
#ifdef __cplusplus
extern "C" {
#endif
enum ggml_status {
GGML_STATUS_ALLOC_FAILED = -2,
GGML_STATUS_FAILED = -1,
GGML_STATUS_SUCCESS = 0,
GGML_STATUS_ABORTED = 1,
};
// get ggml_status name string
GGML_API GGML_CALL const char * ggml_status_to_string(enum ggml_status status);
// ieee 754-2008 half-precision float16
// todo: make this not an integral type
typedef uint16_t ggml_fp16_t;
GGML_API float ggml_fp16_to_fp32(ggml_fp16_t);
GGML_API ggml_fp16_t ggml_fp32_to_fp16(float);
GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t *, float *, int64_t);
GGML_API void ggml_fp32_to_fp16_row(const float *, ggml_fp16_t *, int64_t);
// google brain half-precision bfloat16
typedef struct { uint16_t bits; } ggml_bf16_t;
GGML_API ggml_bf16_t ggml_fp32_to_bf16(float);
GGML_API float ggml_bf16_to_fp32(ggml_bf16_t); // consider just doing << 16
GGML_API void ggml_bf16_to_fp32_row(const ggml_bf16_t *, float *, int64_t);
GGML_API void ggml_fp32_to_bf16_row(const float *, ggml_bf16_t *, int64_t);
struct ggml_object;
struct ggml_context;
// NOTE: always add types at the end of the enum to keep backward compatibility
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,
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_IQ2_XXS = 16,
GGML_TYPE_IQ2_XS = 17,
GGML_TYPE_IQ3_XXS = 18,
GGML_TYPE_IQ1_S = 19,
GGML_TYPE_IQ4_NL = 20,
GGML_TYPE_IQ3_S = 21,
GGML_TYPE_IQ2_S = 22,
GGML_TYPE_IQ4_XS = 23,
GGML_TYPE_I8 = 24,
GGML_TYPE_I16 = 25,
GGML_TYPE_I32 = 26,
GGML_TYPE_I64 = 27,
GGML_TYPE_F64 = 28,
GGML_TYPE_IQ1_M = 29,
GGML_TYPE_BF16 = 30,
GGML_TYPE_COUNT,
};
// precision
enum ggml_prec {
GGML_PREC_DEFAULT,
GGML_PREC_F32,
};
enum ggml_backend_type {
GGML_BACKEND_TYPE_CPU = 0,
GGML_BACKEND_TYPE_GPU = 10,
GGML_BACKEND_TYPE_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
GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ2_XS = 16, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ1_S = 18, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ4_NL = 19, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ3_S = 20, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ2_S = 21, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ4_XS = 22, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ1_M = 23, // except 1d tensors
GGML_FTYPE_MOSTLY_BF16 = 24, // 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_CONCAT,
GGML_OP_SILU_BACK,
GGML_OP_NORM, // normalize
GGML_OP_RMS_NORM,
GGML_OP_RMS_NORM_BACK,
GGML_OP_GROUP_NORM,
GGML_OP_MUL_MAT,
GGML_OP_MUL_MAT_ID,
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_CLAMP,
GGML_OP_CONV_TRANSPOSE_1D,
GGML_OP_IM2COL,
GGML_OP_CONV_TRANSPOSE_2D,
GGML_OP_POOL_1D,
GGML_OP_POOL_2D,
GGML_OP_UPSCALE, // nearest interpolate
GGML_OP_PAD,
GGML_OP_ARANGE,
GGML_OP_TIMESTEP_EMBEDDING,
GGML_OP_ARGSORT,
GGML_OP_LEAKY_RELU,
GGML_OP_FLASH_ATTN_EXT,
GGML_OP_FLASH_ATTN_BACK,
GGML_OP_SSM_CONV,
GGML_OP_SSM_SCAN,
GGML_OP_WIN_PART,
GGML_OP_WIN_UNPART,
GGML_OP_GET_REL_POS,
GGML_OP_ADD_REL_POS,
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_SIGMOID,
GGML_UNARY_OP_GELU,
GGML_UNARY_OP_GELU_QUICK,
GGML_UNARY_OP_SILU,
GGML_UNARY_OP_HARDSWISH,
GGML_UNARY_OP_HARDSIGMOID,
GGML_UNARY_OP_COUNT,
};
enum ggml_object_type {
GGML_OBJECT_TYPE_TENSOR,
GGML_OBJECT_TYPE_GRAPH,
GGML_OBJECT_TYPE_WORK_BUFFER
};
enum ggml_log_level {
GGML_LOG_LEVEL_ERROR = 2,
GGML_LOG_LEVEL_WARN = 3,
GGML_LOG_LEVEL_INFO = 4,
GGML_LOG_LEVEL_DEBUG = 5
};
enum ggml_tensor_flag {
GGML_TENSOR_FLAG_INPUT = 1,
GGML_TENSOR_FLAG_OUTPUT = 2,
GGML_TENSOR_FLAG_PARAM = 4,
};
// 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;
GGML_DEPRECATED(enum ggml_backend_type backend, "use the buffer type to find the storage location of the tensor");
struct ggml_backend_buffer * buffer;
int64_t ne[GGML_MAX_DIMS]; // number of elements
size_t nb[GGML_MAX_DIMS]; // stride in bytes:
// nb[0] = ggml_type_size(type)
// nb[1] = nb[0] * (ne[0] / ggml_blck_size(type)) + 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)];
int32_t flags;
struct ggml_tensor * grad;
struct ggml_tensor * src[GGML_MAX_SRC];
// performance
int perf_runs;
int64_t perf_cycles;
int64_t perf_time_us;
struct ggml_tensor * view_src;
size_t view_offs;
void * data;
char name[GGML_MAX_NAME];
void * extra; // extra things e.g. for ggml-cuda.cu
char padding[8];
};
static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
// Abort callback
// If not NULL, called before ggml computation
// If it returns true, the computation is aborted
typedef bool (*ggml_abort_callback)(void * data);
// 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;
// abort ggml_graph_compute when true
ggml_abort_callback abort_callback;
void * abort_callback_data;
};
enum ggml_cgraph_eval_order {
GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
GGML_CGRAPH_EVAL_ORDER_COUNT
};
struct ggml_hash_set {
size_t size;
struct ggml_tensor ** keys;
};
// computation graph
struct ggml_cgraph {
int size;
int n_nodes;
int n_leafs;
struct ggml_tensor ** nodes;
struct ggml_tensor ** grads;
struct ggml_tensor ** leafs;
struct ggml_hash_set visited_hash_table;
enum ggml_cgraph_eval_order order;
// performance
int perf_runs;
int64_t perf_cycles;
int64_t perf_time_us;
};
// 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_TYPE_INIT = 0,
GGML_TASK_TYPE_COMPUTE,
GGML_TASK_TYPE_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;
};
// numa strategies
enum ggml_numa_strategy {
GGML_NUMA_STRATEGY_DISABLED = 0,
GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
GGML_NUMA_STRATEGY_ISOLATE = 2,
GGML_NUMA_STRATEGY_NUMACTL = 3,
GGML_NUMA_STRATEGY_MIRROR = 4,
GGML_NUMA_STRATEGY_COUNT
};
//
// GUID
//
// GUID types
typedef uint8_t ggml_guid[16];
typedef ggml_guid * ggml_guid_t;
GGML_API bool ggml_guid_matches(ggml_guid_t guid_a, ggml_guid_t guid_b);
// 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_print_backtrace(void);
// accepts a UTF-8 path, even on Windows
GGML_API FILE * ggml_fopen(const char * fname, const char * mode);
GGML_API void ggml_numa_init(enum ggml_numa_strategy numa); // 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 GGML_CALL int64_t ggml_nelements (const struct ggml_tensor * tensor);
GGML_API GGML_CALL int64_t ggml_nrows (const struct ggml_tensor * tensor);
GGML_API GGML_CALL size_t ggml_nbytes (const struct ggml_tensor * tensor);
GGML_API size_t ggml_nbytes_pad (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN
GGML_API GGML_CALL int ggml_blck_size(enum ggml_type type);
GGML_API GGML_CALL size_t ggml_type_size(enum ggml_type type); // size in bytes for all elements in a block
GGML_API GGML_CALL size_t ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
GGML_DEPRECATED(
GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
"use ggml_row_size() instead");
GGML_API GGML_CALL const char * ggml_type_name(enum ggml_type type);
GGML_API GGML_CALL const char * ggml_op_name (enum ggml_op op);
GGML_API const char * ggml_op_symbol(enum ggml_op op);
GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
GGML_API GGML_CALL const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
GGML_API GGML_CALL size_t ggml_element_size(const struct ggml_tensor * tensor);
GGML_API GGML_CALL 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 GGML_CALL bool ggml_is_transposed(const struct ggml_tensor * tensor);
GGML_API GGML_CALL bool ggml_is_permuted (const struct ggml_tensor * tensor);
GGML_API GGML_CALL bool ggml_is_empty (const struct ggml_tensor * tensor);
GGML_API bool ggml_is_scalar (const struct ggml_tensor * tensor);
GGML_API bool ggml_is_vector (const struct ggml_tensor * tensor);
GGML_API bool ggml_is_matrix (const struct ggml_tensor * tensor);
GGML_API bool ggml_is_3d (const struct ggml_tensor * tensor);
GGML_API int ggml_n_dims (const struct ggml_tensor * tensor); // returns 1 for scalars
GGML_API GGML_CALL bool ggml_is_contiguous (const struct ggml_tensor * tensor);
GGML_API GGML_CALL bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous()
GGML_API GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
GGML_API GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
GGML_API bool ggml_are_same_stride(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);
GGML_API bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbytes);
// 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, struct ggml_tensor * src);
// Context tensor enumeration and lookup
GGML_API struct ggml_tensor * ggml_get_first_tensor(const struct ggml_context * ctx);
GGML_API struct ggml_tensor * ggml_get_next_tensor (const struct ggml_context * ctx, struct ggml_tensor * tensor);
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);
// Converts a flat index into coordinates
GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
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 int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_API void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, 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 float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_API void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, 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 GGML_CALL 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_ATTRIBUTE_FORMAT(2, 3)
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_add_cast(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
enum ggml_type type);
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);
// dst = a
// view(dst, nb1, nb2, nb3, offset) += b
// return dst
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);
// sums repetitions in a into shape of b
GGML_API struct ggml_tensor * ggml_repeat_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// concat a and b along dim
// used in stable-diffusion
GGML_API struct ggml_tensor * ggml_concat(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int dim);
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_leaky_relu(
struct ggml_context * ctx,
struct ggml_tensor * a, float negative_slope, bool inplace);
GGML_API struct ggml_tensor * ggml_relu_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_sigmoid(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_sigmoid_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
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);
// hardswish(x) = x * relu6(x + 3) / 6
GGML_API struct ggml_tensor * ggml_hardswish(
struct ggml_context * ctx,
struct ggml_tensor * a);
// hardsigmoid(x) = relu6(x + 3) / 6
GGML_API struct ggml_tensor * ggml_hardsigmoid(
struct ggml_context * ctx,
struct ggml_tensor * a);
// normalize along rows
GGML_API struct ggml_tensor * ggml_norm(
struct ggml_context * ctx,
struct ggml_tensor * a,
float eps);
GGML_API struct ggml_tensor * ggml_norm_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
float eps);
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);
// group normalize along ne0*ne1*n_groups
// used in stable-diffusion
// TODO: eps is hardcoded to 1e-6 for now
GGML_API struct ggml_tensor * ggml_group_norm(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_groups);
GGML_API struct ggml_tensor * ggml_group_norm_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_groups);
// a - x
// b - dy
GGML_API struct ggml_tensor * ggml_rms_norm_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
float eps);
// A: k columns, n rows => [ne03, ne02, n, k]
// B: k columns, m rows (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k]
// result is n columns, m rows => [ne03 * x, ne02 * y, m, n]
GGML_API struct ggml_tensor * ggml_mul_mat(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// change the precision of a matrix multiplication
// set to GGML_PREC_F32 for higher precision (useful for phi-2)
GGML_API void ggml_mul_mat_set_prec(
struct ggml_tensor * a,
enum ggml_prec prec);
// indirect matrix multiplication
GGML_API struct ggml_tensor * ggml_mul_mat_id(
struct ggml_context * ctx,
struct ggml_tensor * as,
struct ggml_tensor * b,
struct ggml_tensor * ids);
// 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,
float s);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_scale_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
float s);
// 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);
GGML_API struct ggml_tensor * ggml_cast(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_type type);
// make contiguous
GGML_API struct ggml_tensor * ggml_cont(
struct ggml_context * ctx,
struct ggml_tensor * a);
// make contiguous, with new shape
GGML_API struct ggml_tensor * ggml_cont_1d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0);
GGML_API struct ggml_tensor * ggml_cont_2d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1);
GGML_API struct ggml_tensor * ggml_cont_3d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
int64_t ne2);
GGML_API struct ggml_tensor * ggml_cont_4d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3);
// 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);
// supports 3D: a->ne[2] == b->ne[1]
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);
// fused soft_max(a*scale + mask*(ALiBi slope))
// mask is optional
// max_bias = 0.0f for no ALiBi
GGML_API struct ggml_tensor * ggml_soft_max_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * mask,
float scale,
float max_bias);
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 (NOT SUPPORTED)
// if mode & 2 == 1, GPT-NeoX style
// if mode & 4 == 1, ChatGLM style
//
// b is an int32 vector with size a->ne[2], it contains the positions
// c is freq factors (e.g. phi3-128k), (optional)
GGML_API struct ggml_tensor * ggml_rope(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
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,
struct ggml_tensor * b,
int n_dims,
int mode,
int n_ctx);
// custom RoPE
GGML_API struct ggml_tensor * ggml_rope_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims,
int mode,
int n_ctx,
int n_orig_ctx,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_rope_ext_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims,
int mode,
int n_ctx,
int n_orig_ctx,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow);
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int n_dims,
int mode,
int n_ctx,
int n_orig_ctx,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow),
"use ggml_rope_ext instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int n_dims,
int mode,
int n_ctx,
int n_orig_ctx,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow),
"use ggml_rope_ext_inplace instead");
struct ggml_tensor * ggml_rope_xpos_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int n_dims,
float base,
bool down);
// compute correction dims for YaRN RoPE scaling
GGML_CALL void ggml_rope_yarn_corr_dims(
int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
// 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,
struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims,
int mode,
int n_ctx,
int n_orig_ctx,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow,
float xpos_base,
bool xpos_down);
// clamp
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_clamp(
struct ggml_context * ctx,
struct ggml_tensor * a,
float min,
float max);
GGML_API struct ggml_tensor * ggml_im2col(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int s0,
int s1,
int p0,
int p1,
int d0,
int d1,
bool is_2D,
enum ggml_type dst_type);
GGML_API struct ggml_tensor * ggml_conv_depthwise_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);
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
// 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);
GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int s0,
int p0,
int d0);
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);
// kernel size is a->ne[0] x a->ne[1]
// stride is equal to kernel size
// padding is zero
// example:
// a: 16 16 3 768
// b: 1024 1024 3 1
// res: 64 64 768 1
// used in sam
GGML_API struct ggml_tensor * ggml_conv_2d_sk_p0(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// kernel size is a->ne[0] x a->ne[1]
// stride is 1
// padding is half
// example:
// a: 3 3 256 256
// b: 64 64 256 1
// res: 64 64 256 1
// used in sam
GGML_API struct ggml_tensor * ggml_conv_2d_s1_ph(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int stride);
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
// the result will have 2*p0 padding for the first dimension
// and 2*p1 padding for the second dimension
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,
float p0,
float p1);
// nearest interpolate
// multiplies ne0 and ne1 by scale factor
// used in stable-diffusion
GGML_API struct ggml_tensor * ggml_upscale(
struct ggml_context * ctx,
struct ggml_tensor * a,
int scale_factor);
// nearest interpolate
// nearest interpolate to specified dimensions
// used in tortoise.cpp
GGML_API struct ggml_tensor * ggml_upscale_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
int ne0,
int ne1,
int ne2,
int ne3);
// pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
GGML_API struct ggml_tensor * ggml_pad(
struct ggml_context * ctx,
struct ggml_tensor * a,
int p0,
int p1,
int p2,
int p3);
// Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
// timesteps: [N,]
// return: [N, dim]
GGML_API struct ggml_tensor * ggml_timestep_embedding(
struct ggml_context * ctx,
struct ggml_tensor * timesteps,
int dim,
int max_period);
// sort rows
enum ggml_sort_order {
GGML_SORT_ORDER_ASC,
GGML_SORT_ORDER_DESC,
};
GGML_API struct ggml_tensor * ggml_argsort(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_sort_order order);
GGML_API struct ggml_tensor * ggml_arange(
struct ggml_context * ctx,
float start,
float stop,
float step);
// top k elements per row
GGML_API struct ggml_tensor * ggml_top_k(
struct ggml_context * ctx,
struct ggml_tensor * a,
int k);
#define GGML_KQ_MASK_PAD 32
// q: [n_embd, n_batch, n_head, 1]
// k: [n_embd, n_kv, n_head_kv, 1]
// v: [n_embd, n_kv, n_head_kv, 1] !! not transposed !!
// mask: [n_kv, n_batch_pad, 1, 1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
// res: [n_embd, n_head, n_batch, 1] !! permuted !!
GGML_API struct ggml_tensor * ggml_flash_attn_ext(
struct ggml_context * ctx,
struct ggml_tensor * q,
struct ggml_tensor * k,
struct ggml_tensor * v,
struct ggml_tensor * mask,
float scale,
float max_bias);
GGML_API void ggml_flash_attn_ext_set_prec(
struct ggml_tensor * a,
enum ggml_prec prec);
// TODO: needs to be adapted to ggml_flash_attn_ext
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_ssm_conv(
struct ggml_context * ctx,
struct ggml_tensor * s,
struct ggml_tensor * x,
struct ggml_tensor * c,
struct ggml_tensor * sq);
GGML_API struct ggml_tensor * ggml_ssm_scan(
struct ggml_context * ctx,
struct ggml_tensor * s,
struct ggml_tensor * x,
struct ggml_tensor * dt,
struct ggml_tensor * A,
struct ggml_tensor * B,
struct ggml_tensor * C,
struct ggml_tensor * sq);
// 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);
// used in sam
GGML_API struct ggml_tensor * ggml_get_rel_pos(
struct ggml_context * ctx,
struct ggml_tensor * a,
int qh,
int kh);
// used in sam
GGML_API struct ggml_tensor * ggml_add_rel_pos(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * pw,
struct ggml_tensor * ph);
GGML_API struct ggml_tensor * ggml_add_rel_pos_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * pw,
struct ggml_tensor * ph);
// 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 void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);
// graph allocation in a context
GGML_API struct ggml_cgraph * ggml_new_graph (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
GGML_API struct ggml_cgraph * ggml_new_graph_custom (struct ggml_context * ctx, size_t size, bool grads);
GGML_API struct ggml_cgraph * ggml_graph_dup (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
GGML_API struct ggml_cgraph ggml_graph_view (struct ggml_cgraph * cgraph, int i0, int i1);
GGML_API void ggml_graph_cpy (struct ggml_cgraph * src, struct ggml_cgraph * dst);
GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph); // zero grads
GGML_API void ggml_graph_clear (struct ggml_cgraph * cgraph);
GGML_API size_t ggml_graph_overhead(void);
GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
// 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 (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
GGML_API enum ggml_status ggml_graph_compute ( struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
// 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 enum ggml_status 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);
// build gradient checkpointing backward graph gb for gf using provided checkpoints
// gb_tmp will contain original backward graph with rewritten backward process nodes,
// but without the second forward pass nodes.
GGML_API void ggml_build_backward_gradient_checkpointing(
struct ggml_context * ctx,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb,
struct ggml_cgraph * gb_tmp,
struct ggml_tensor * * checkpoints,
int n_checkpoints);
//
// optimization
//
// optimization methods
enum ggml_opt_type {
GGML_OPT_TYPE_ADAM,
GGML_OPT_TYPE_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_RESULT_OK = 0,
GGML_OPT_RESULT_DID_NOT_CONVERGE,
GGML_OPT_RESULT_NO_CONTEXT,
GGML_OPT_RESULT_INVALID_WOLFE,
GGML_OPT_RESULT_FAIL,
GGML_OPT_RESULT_CANCEL,
GGML_LINESEARCH_FAIL = -128,
GGML_LINESEARCH_MINIMUM_STEP,
GGML_LINESEARCH_MAXIMUM_STEP,
GGML_LINESEARCH_MAXIMUM_ITERATIONS,
GGML_LINESEARCH_INVALID_PARAMETERS,
};
typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);
// optimization parameters
//
// see ggml.c (ggml_opt_default_params) for default values
//
struct ggml_opt_params {
enum ggml_opt_type type;
size_t graph_size;
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;
int n_gradient_accumulation;
// 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
int decay_min_ndim; // minimum number of tensor dimension to apply weight decay
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
float gclip; // gradient clipping
} 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;
float loss_before;
float loss_after;
struct {
struct ggml_tensor * g; // current gradient
struct ggml_tensor * m; // first moment
struct ggml_tensor * v; // second moment
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,
ggml_opt_callback callback,
void * callback_data);
//
// tensor flags
//
GGML_API void ggml_set_input(struct ggml_tensor * tensor);
GGML_API void ggml_set_output(struct ggml_tensor * tensor);
//
// quantization
//
// - ggml_quantize_init can be called multiple times with the same type
// it will only initialize the quantization tables for the first call or after ggml_quantize_free
// automatically called by ggml_quantize_chunk for convenience
//
// - ggml_quantize_free will free any memory allocated by ggml_quantize_init
// call this at the end of the program to avoid memory leaks
//
// note: these are thread-safe
//
GGML_API void ggml_quantize_init(enum ggml_type type);
GGML_API void ggml_quantize_free(void);
// some quantization type cannot be used without an importance matrix
GGML_API bool ggml_quantize_requires_imatrix(enum ggml_type type);
// calls ggml_quantize_init internally (i.e. can allocate memory)
GGML_API size_t ggml_quantize_chunk(
enum ggml_type type,
const float * src,
void * dst,
int64_t start,
int64_t nrows,
int64_t n_per_row,
const float * imatrix);
//
// gguf
//
enum gguf_type {
GGUF_TYPE_UINT8 = 0,
GGUF_TYPE_INT8 = 1,
GGUF_TYPE_UINT16 = 2,
GGUF_TYPE_INT16 = 3,
GGUF_TYPE_UINT32 = 4,
GGUF_TYPE_INT32 = 5,
GGUF_TYPE_FLOAT32 = 6,
GGUF_TYPE_BOOL = 7,
GGUF_TYPE_STRING = 8,
GGUF_TYPE_ARRAY = 9,
GGUF_TYPE_UINT64 = 10,
GGUF_TYPE_INT64 = 11,
GGUF_TYPE_FLOAT64 = 12,
GGUF_TYPE_COUNT, // marks the end of the enum
};
struct gguf_context;
struct gguf_init_params {
bool no_alloc;
// if not NULL, create a ggml_context and allocate the tensor data in it
struct ggml_context ** ctx;
};
GGML_API struct gguf_context * gguf_init_empty(void);
GGML_API struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params);
//GGML_API struct gguf_context * gguf_init_from_buffer(..);
GGML_API void gguf_free(struct gguf_context * ctx);
GGML_API const char * gguf_type_name(enum gguf_type type);
GGML_API int gguf_get_version (const struct gguf_context * ctx);
GGML_API size_t gguf_get_alignment (const struct gguf_context * ctx);
GGML_API size_t gguf_get_data_offset(const struct gguf_context * ctx);
GGML_API void * gguf_get_data (const struct gguf_context * ctx);
GGML_API int gguf_get_n_kv(const struct gguf_context * ctx);
GGML_API int gguf_find_key(const struct gguf_context * ctx, const char * key);
GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int key_id);
GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int key_id);
GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id);
// will abort if the wrong type is used for the key
GGML_API uint8_t gguf_get_val_u8 (const struct gguf_context * ctx, int key_id);
GGML_API int8_t gguf_get_val_i8 (const struct gguf_context * ctx, int key_id);
GGML_API uint16_t gguf_get_val_u16 (const struct gguf_context * ctx, int key_id);
GGML_API int16_t gguf_get_val_i16 (const struct gguf_context * ctx, int key_id);
GGML_API uint32_t gguf_get_val_u32 (const struct gguf_context * ctx, int key_id);
GGML_API int32_t gguf_get_val_i32 (const struct gguf_context * ctx, int key_id);
GGML_API float gguf_get_val_f32 (const struct gguf_context * ctx, int key_id);
GGML_API uint64_t gguf_get_val_u64 (const struct gguf_context * ctx, int key_id);
GGML_API int64_t gguf_get_val_i64 (const struct gguf_context * ctx, int key_id);
GGML_API double gguf_get_val_f64 (const struct gguf_context * ctx, int key_id);
GGML_API bool gguf_get_val_bool(const struct gguf_context * ctx, int key_id);
GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int key_id);
GGML_API const void * gguf_get_val_data(const struct gguf_context * ctx, int key_id);
GGML_API int gguf_get_arr_n (const struct gguf_context * ctx, int key_id);
GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id);
GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);
GGML_API int gguf_get_n_tensors (const struct gguf_context * ctx);
GGML_API int gguf_find_tensor (const struct gguf_context * ctx, const char * name);
GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
GGML_API char * gguf_get_tensor_name (const struct gguf_context * ctx, int i);
GGML_API enum ggml_type gguf_get_tensor_type (const struct gguf_context * ctx, int i);
// removes key if it exists
GGML_API void gguf_remove_key(struct gguf_context * ctx, const char * key);
// overrides existing values or adds a new one
GGML_API void gguf_set_val_u8 (struct gguf_context * ctx, const char * key, uint8_t val);
GGML_API void gguf_set_val_i8 (struct gguf_context * ctx, const char * key, int8_t val);
GGML_API void gguf_set_val_u16 (struct gguf_context * ctx, const char * key, uint16_t val);
GGML_API void gguf_set_val_i16 (struct gguf_context * ctx, const char * key, int16_t val);
GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val);
GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t val);
GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float val);
GGML_API void gguf_set_val_u64 (struct gguf_context * ctx, const char * key, uint64_t val);
GGML_API void gguf_set_val_i64 (struct gguf_context * ctx, const char * key, int64_t val);
GGML_API void gguf_set_val_f64 (struct gguf_context * ctx, const char * key, double val);
GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool val);
GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val);
GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, int n);
GGML_API void gguf_set_arr_str (struct gguf_context * ctx, const char * key, const char ** data, int n);
// set or add KV pairs from another context
GGML_API void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src);
// manage tensor info
GGML_API void gguf_add_tensor(struct gguf_context * ctx, const struct ggml_tensor * tensor);
GGML_API void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type);
GGML_API void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data, size_t size);
// writing gguf files can be done in 2 ways:
//
// - write the entire gguf_context to a binary file in a single pass:
//
// gguf_write_to_file(ctx, fname);
//
// - first prepare a file with a placeholder for the meta data, write the tensor data, then write the meta data:
//
// FILE * f = fopen(fname, "wb");
// fseek(f, gguf_get_meta_size(ctx), SEEK_SET);
// fwrite(f, ...);
// void * data = gguf_meta_get_meta_data(ctx);
// fseek(f, 0, SEEK_SET);
// fwrite(f, data, gguf_get_meta_size(ctx));
// free(data);
// fclose(f);
//
// write the entire context to a binary file
GGML_API void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta);
// get the size in bytes of the meta data (header, kv pairs, tensor info) including padding
GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
GGML_API void gguf_get_meta_data(const struct gguf_context * ctx, void * data);
//
// system info
//
GGML_API int ggml_cpu_has_avx (void);
GGML_API int ggml_cpu_has_avx_vnni (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_avx512_bf16(void);
GGML_API int ggml_cpu_has_fma (void);
GGML_API int ggml_cpu_has_neon (void);
GGML_API int ggml_cpu_has_sve (void);
GGML_API int ggml_cpu_has_arm_fma (void);
GGML_API int ggml_cpu_has_metal (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_cuda (void);
GGML_API int ggml_cpu_has_clblast (void);
GGML_API int ggml_cpu_has_vulkan (void);
GGML_API int ggml_cpu_has_kompute (void);
GGML_API int ggml_cpu_has_gpublas (void);
GGML_API int ggml_cpu_has_sse3 (void);
GGML_API int ggml_cpu_has_ssse3 (void);
GGML_API int ggml_cpu_has_sycl (void);
GGML_API int ggml_cpu_has_rpc (void);
GGML_API int ggml_cpu_has_vsx (void);
GGML_API int ggml_cpu_has_matmul_int8(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, int64_t k);
typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
typedef void (*ggml_vec_dot_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
const void * GGML_RESTRICT y, size_t by, int nrc);
typedef struct {
const char * type_name;
int blck_size;
size_t type_size;
bool is_quantized;
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;
int64_t nrows; // number of rows to process simultaneously;
} ggml_type_traits_t;
GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
#ifdef __cplusplus
}
#endif
llm/llama.cpp/ggml_vk_generate_shaders.py deleted 100644 → 0
View file @ 97b02a89
#!/usr/bin/env python
import logging
import argparse
import asyncio
import os
import sys
from tempfile import gettempdir, NamedTemporaryFile
logger = logging.getLogger("ggml-vk-generate-shaders")
shader_f32 = """
#define FLOAT_TYPE float
"""
shader_f16 = """
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
#define FLOAT_TYPE float16_t
"""
shader_int8_ext = """
#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
"""
# Type-specific defines
shader_f32_defines = """
#define QUANT_K 1
#define QUANT_R 1
#define A_TYPE float
"""
shader_f16_defines = """
#define QUANT_K 1
#define QUANT_R 1
#define A_TYPE float16_t
"""
shader_q4_0_defines = """
#define QUANT_K 32
#define QUANT_R 2
struct block_q4_0
{
float16_t d;
uint8_t qs[16];
};
#define A_TYPE block_q4_0
"""
shader_q4_1_defines = """
#define QUANT_K 32
#define QUANT_R 2
struct block_q4_1
{
float16_t d;
float16_t m;
uint8_t qs[16];
};
#define A_TYPE block_q4_1
"""
shader_q5_0_defines = """
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#define QUANT_K 32
#define QUANT_R 2
struct block_q5_0
{
float16_t d;
uint16_t qh[2];
uint8_t qs[16];
};
#define A_TYPE block_q5_0
"""
shader_q5_1_defines = """
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#define QUANT_K 32
#define QUANT_R 2
struct block_q5_1
{
float16_t d;
float16_t m;
uint qh;
uint8_t qs[16];
};
#define A_TYPE block_q5_1
"""
shader_q8_0_defines = """
#define QUANT_K 32
#define QUANT_R 1
struct block_q8_0
{
float16_t d;
int8_t qs[32];
};
#define A_TYPE block_q8_0
"""
# K-quants
shader_q2_K_defines = """
#define QUANT_K 256
struct block_q2_K
{
uint8_t scales[QUANT_K/16];
uint8_t qs[QUANT_K/4];
f16vec2 d;
};
#define A_TYPE block_q2_K
"""
shader_q3_K_defines = """
#define QUANT_K 256
struct block_q3_K
{
uint8_t hmask[QUANT_K/8];
uint8_t qs[QUANT_K/4];
uint8_t scales[12];
float16_t d;
};
#define A_TYPE block_q3_K
"""
shader_q4_K_defines = """
#define QUANT_K 256
struct block_q4_K
{
f16vec2 d;
uint8_t scales[3*QUANT_K/64];
uint8_t qs[QUANT_K/2];
};
#define A_TYPE block_q4_K
"""
shader_q5_K_defines = """
#define QUANT_K 256
struct block_q5_K
{
f16vec2 d;
uint8_t scales[12];
uint8_t qh[QUANT_K/8];
uint8_t qs[QUANT_K/2];
};
#define A_TYPE block_q5_K
"""
shader_q6_K_defines = """
#define QUANT_K 256
struct block_q6_K
{
uint8_t ql[QUANT_K/2];
uint8_t qh[QUANT_K/4];
int8_t scales[QUANT_K/16];
float16_t d;
};
#define A_TYPE block_q6_K
"""
# Dequant functions
shader_float_dequant_func = """
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
}
"""
shader_q4_0_dequant_func = """
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const float d = float(data_a[a_offset + ib].d);
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
return (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
}
"""
shader_q4_1_dequant_func = """
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const float d = float(data_a[a_offset + ib].d);
const float m = float(data_a[a_offset + ib].m);
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
return vec2(vui & 0xF, vui >> 4) * d + m;
}
"""
shader_q5_0_dequant_func = """
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const float d = float(data_a[a_offset + ib].d);
const uint uint_qh = uint(data_a[a_offset + ib].qh[1]) << 16 | data_a[a_offset + ib].qh[0];
const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
}
"""
shader_q5_1_dequant_func = """
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const float d = float(data_a[a_offset + ib].d);
const float m = float(data_a[a_offset + ib].m);
const uint uint_qh = data_a[a_offset + ib].qh;
const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
}
"""
shader_q8_0_dequant_func = """
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const float d = float(data_a[a_offset + ib].d);
return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d;
}
"""
# MULMAT
mulmat_head = """#version 450
#extension GL_EXT_control_flow_attributes : enable
#extension GL_EXT_shader_16bit_storage : require
#ifdef MUL_MAT_ID
#extension GL_EXT_buffer_reference2 : require
#extension GL_EXT_nonuniform_qualifier : require
#extension GL_EXT_scalar_block_layout : require
#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
#define EXPERT_COUNT 8
#endif
#ifndef LOAD_VEC_A
#define LOAD_VEC_A 1
#endif
#ifndef LOAD_VEC_B
#define LOAD_VEC_B 1
#endif
"""
mulmat_body1 = """
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
#ifdef MUL_MAT_ID
layout (binding = 3) readonly buffer IDS {int data_ids[];};
#endif
layout (push_constant) uniform parameter
{
uint M;
uint N;
uint K;
uint stride_a;
uint stride_b;
uint stride_d;
uint k_split;
uint ne02;
uint ne12;
uint broadcast2;
uint broadcast3;
uint batch_stride_a;
uint batch_stride_b;
uint batch_stride_d;
#ifdef MUL_MAT_ID
uint expert_stride_a;
uint expert_stride_b0;
uint expert_stride_b1;
uint expert_stride_d;
uint ids_stride;
uint n_as;
uint nei0;
uint nei1;
uint nbi1;
uint ne11;
#endif
} p;
layout (constant_id = 1) const uint BM = 64;
layout (constant_id = 2) const uint BN = 64;
layout (constant_id = 3) const uint BK = 16; // Assumed to be 32 if working with a quant
layout (constant_id = 4) const uint WM = 32;
layout (constant_id = 5) const uint WN = 32;
layout (constant_id = 6) const uint WMITER = 2;
layout (constant_id = 7) const uint TM = 4;
layout (constant_id = 8) const uint TN = 2;
layout (constant_id = 9) const uint WARP = 32;
shared FLOAT_TYPE buf_a[BM * (BK+1)];
shared FLOAT_TYPE buf_b[BN * (BK+1)];
#ifdef MUL_MAT_ID
shared u8vec2 rowids[2048];
#endif
void main() {
#ifdef MUL_MAT_ID
const uint batch_idx = gl_GlobalInvocationID.z / p.n_as;
const uint expert_idx = gl_GlobalInvocationID.z % p.n_as;
#else
const uint batch_idx = gl_GlobalInvocationID.z;
#endif
const uint i13 = batch_idx / p.ne12;
const uint i12 = batch_idx % p.ne12;
const uint i03 = i13 / p.broadcast3;
const uint i02 = i12 / p.broadcast2;
const uint batch_idx_a = i03 * p.ne02 + i02;
const uint blocks_m = (p.M + BM - 1) / BM;
const uint ir = gl_WorkGroupID.x % blocks_m;
const uint ik = gl_WorkGroupID.x / blocks_m;
const uint ic = gl_WorkGroupID.y;
const uint warp_i = gl_LocalInvocationID.x / WARP;
const uint warp_r = warp_i % (BM / WM);
const uint warp_c = warp_i / (BM / WM);
const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
const uint WSUBM = WM / WMITER;
const uint WSUBN = WN / WNITER;
const uint tiw = gl_LocalInvocationID.x % WARP;
const uint tiwr = tiw % (WSUBM / TM);
const uint tiwc = tiw / (WSUBM / TM);
const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A);
const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A);
const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B);
const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B);
const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A / BK;
const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B / BK;
#ifdef MUL_MAT_ID
uint _ne1 = 0;
for (uint ii1 = 0; ii1 < p.nei1; ii1++) {
for (uint ii0 = 0; ii0 < p.nei0; ii0++) {
if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) {
rowids[_ne1] = u8vec2(ii0, ii1);
_ne1++;
}
}
}
const u8vec2 id = rowids[ir * BN + ic];
#endif
const uint start_k = ik * p.k_split;
const uint end_k = min(p.K, (ik + 1) * p.k_split);
uint pos_a = (
#ifdef MUL_MAT_ID
expert_idx * p.expert_stride_a +
#endif
batch_idx_a * p.batch_stride_a + ir * BM * p.stride_a + start_k) / LOAD_VEC_A;
uint pos_b = (
#ifdef MUL_MAT_ID
id.y * p.expert_stride_b1 +
(id.x % p.ne11) * p.expert_stride_b0 +
#endif
batch_idx * p.batch_stride_b +
ic * BN * p.stride_b + start_k) / LOAD_VEC_B;
float sums[WMITER * TM * WNITER * TN];
FLOAT_TYPE cache_a[WMITER * TM];
FLOAT_TYPE cache_b[WNITER * TN];
[[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
sums[i] = 0.0f;
}
[[unroll]] for (uint block = start_k; block < end_k; block += BK) {
[[unroll]] for (uint l = 0; l < BM; l += loadstride_a) {"""
mulmat_load_scalar = """
#if LOAD_VEC_A == 8
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
buf_a[buf_idx ] = FLOAT_TYPE(data_a[idx][0].x);
buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx][0].y);
buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx][0].z);
buf_a[buf_idx + 3] = FLOAT_TYPE(data_a[idx][0].w);
buf_a[buf_idx + 4] = FLOAT_TYPE(data_a[idx][1].x);
buf_a[buf_idx + 5] = FLOAT_TYPE(data_a[idx][1].y);
buf_a[buf_idx + 6] = FLOAT_TYPE(data_a[idx][1].z);
buf_a[buf_idx + 7] = FLOAT_TYPE(data_a[idx][1].w);
#elif LOAD_VEC_A == 4
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
buf_a[buf_idx ] = FLOAT_TYPE(data_a[idx].x);
buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx].y);
buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx].z);
buf_a[buf_idx + 3] = FLOAT_TYPE(data_a[idx].w);
#else
if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) {
buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]);
} else {
buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(0.0f);
}
#endif
"""
mulmat_load_q4_0 = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
const uint ib = idx / 16;
const uint iqs = idx & 0xF;
const float d = float(data_a[ib].d);
const uint vui = uint(data_a[ib].qs[iqs]);
const vec2 v = (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);"""
mulmat_load_q4_1 = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
const uint ib = idx / 16;
const uint iqs = idx & 0xF;
const float d = float(data_a[ib].d);
const float m = float(data_a[ib].m);
const uint vui = uint(data_a[ib].qs[iqs]);
const vec2 v = vec2(vui & 0xF, vui >> 4) * d + m;
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);"""
mulmat_load_q5_0 = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
const uint ib = idx / 16;
const uint iqs = idx & 0xF;
const float d = float(data_a[ib].d);
const uint uint_qh = uint(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0];
const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
const uint vui = uint(data_a[ib].qs[iqs]);
const vec2 v = (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);"""
mulmat_load_q5_1 = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
const uint ib = idx / 16;
const uint iqs = idx & 0xF;
const float d = float(data_a[ib].d);
const float m = float(data_a[ib].m);
const uint uint_qh = data_a[ib].qh;
const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
const uint vui = uint(data_a[ib].qs[iqs]);
const vec2 v = vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);"""
mulmat_load_q8_0 = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
const uint ib = idx / 16;
const uint iqs = (idx & 0xF) * 2;
const float d = float(data_a[ib].d);
const vec2 v = vec2(int(data_a[ib].qs[iqs]), int(data_a[ib].qs[iqs + 1])) * d;
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);"""
mulmat_load_q2_K = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
const uint ib = idx / 128; // 2 values per idx
const uint iqs = idx % 128; // 0..127
const uint qsi = (iqs / 64) * 32 + (iqs % 16) * 2; // 0,2,4..30
const uint scalesi = iqs / 8; // 0..15
const uint qsshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
const uvec2 qs = uvec2(data_a[ib].qs[qsi], data_a[ib].qs[qsi + 1]);
const uint scales = data_a[ib].scales[scalesi];
const vec2 d = vec2(data_a[ib].d);
const vec2 v = d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);"""
mulmat_load_q3_K = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
const uint ib = idx / 128; // 2 values per idx
const uint iqs = idx % 128; // 0..127
const uint n = iqs / 64; // 0,1
const uint qsi = n * 32 + (iqs % 16) * 2; // 0,2,4..62
const uint hmi = (iqs % 16) * 2; // 0,2,4..30
const uint j = (iqs % 64) / 4; // 0..3
const uint is = iqs / 8; // 0..15
const uint halfsplit = ((iqs % 64) / 16); // 0,1,2,3
const uint qsshift = halfsplit * 2; // 0,2,4,6
const uint m = 1 << (4 * n + halfsplit); // 1,2,4,8,16,32,64,128
const int8_t us = int8_t(is < 4 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+8] >> 0) & 3) << 4) :
is < 8 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+4] >> 2) & 3) << 4) :
is < 12 ? (data_a[ib].scales[is-8] >> 4) | (((data_a[ib].scales[is+0] >> 4) & 3) << 4) :
(data_a[ib].scales[is-8] >> 4) | (((data_a[ib].scales[is-4] >> 6) & 3) << 4));
const float dl = float(data_a[ib].d) * float(us - 32);
buf_a[buf_idx ] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi ] & m) != 0) ? 0 : 4)));
buf_a[buf_idx + 1] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));"""
mulmat_load_q4_K = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
const uint ib = idx / 128; // 2 values per idx
const uint iqs = idx % 128; // 0..127
const uint n = iqs / 32; // 0,1,2,3
const uint b = (iqs % 32) / 16; // 0,1
const uint is = 2 * n + b; // 0..7
const uint qsi = n * 32 + (iqs % 16) * 2; // 0,2,4..126
const vec2 loadd = vec2(data_a[ib].d);
uint8_t sc;
uint8_t mbyte;
if (is < 4) {
sc = uint8_t(data_a[ib].scales[is ] & 63);
mbyte = uint8_t(data_a[ib].scales[is + 4] & 63);
} else {
sc = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4));
mbyte = uint8_t((data_a[ib].scales[is + 4] >> 4) | ((data_a[ib].scales[is ] >> 6) << 4));
}
const float d = loadd.x * sc;
const float m = loadd.y * mbyte;
buf_a[buf_idx ] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi ] >> (b * 4)) & 0xF) - m);
buf_a[buf_idx + 1] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) - m);"""
mulmat_load_q5_K = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
const uint ib = idx / 128; // 2 values per idx
const uint iqs = idx % 128; // 0..127
const uint n = iqs / 32; // 0,1,2,3
const uint b = (iqs % 32) / 16; // 0,1
const uint is = 2 * n + b; // 0..7
const uint qsi = n * 32 + (iqs % 16) * 2; // 0,2,4..126
const uint qhi = (iqs % 16) * 2; // 0,2,4..30
const uint8_t hm = uint8_t(1 << (iqs / 16));
const vec2 loadd = vec2(data_a[ib].d);
uint8_t sc;
uint8_t mbyte;
if (is < 4) {
sc = uint8_t(data_a[ib].scales[is ] & 63);
mbyte = uint8_t(data_a[ib].scales[is + 4] & 63);
} else {
sc = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4));
mbyte = uint8_t((data_a[ib].scales[is + 4] >> 4) | ((data_a[ib].scales[is ] >> 6) << 4));
}
const float d = loadd.x * sc;
const float m = loadd.y * mbyte;
buf_a[buf_idx ] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi ] & hm) != 0 ? 16 : 0)) - m);
buf_a[buf_idx + 1] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0)) - m);"""
mulmat_load_q6_K = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
const uint ib = idx / 128; // 2 values per idx
const uint iqs = idx % 128; // 0..127
const uint n = iqs / 64; // 0,1
const uint b = (iqs % 64) / 32; // 0,1
const uint is_b = (iqs % 16) / 8; // 0,1
const uint qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
const uint is = 8 * n + qhshift + is_b; // 0..15
const uint qsi = n * 64 + (iqs % 32) * 2; // 0,2,4..126
const uint qhi = n * 32 + (iqs % 16) * 2; // 0,2,4..62
const float dscale = float(data_a[ib].d) * float(data_a[ib].scales[is]);
buf_a[buf_idx ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi ] >> qhshift) & 3) << 4)) - 32));
buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));"""
mulmat_body2 = """
}
[[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {
#if LOAD_VEC_B == 8
const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B;
buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx][0].x);
buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx][0].y);
buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx][0].z);
buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx][0].w);
buf_b[buf_idx + 4] = FLOAT_TYPE(data_b[idx][1].x);
buf_b[buf_idx + 5] = FLOAT_TYPE(data_b[idx][1].y);
buf_b[buf_idx + 6] = FLOAT_TYPE(data_b[idx][1].z);
buf_b[buf_idx + 7] = FLOAT_TYPE(data_b[idx][1].w);
#elif LOAD_VEC_B == 4
const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B;
buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx].x);
buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx].y);
buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx].z);
buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx].w);
#else
if (ic * BN + loadc_b + l < p.N && block + loadr_b < end_k) {
buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]);
} else {
buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f);
}
#endif
}
barrier();
pos_a += BK / LOAD_VEC_A;
pos_b += BK / LOAD_VEC_B;
for (uint i = 0; i < BK; i++) {
// Load from shared into cache
[[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
[[unroll]] for (uint j = 0; j < TM; j++) {
cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * (BK+1) + i];
}
}
[[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
[[unroll]] for (uint j = 0; j < TN; j++) {
cache_b[wsic * TN + j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * (BK+1) + i];
}
}
[[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
[[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
[[unroll]] for (uint cc = 0; cc < TN; cc++) {
[[unroll]] for (uint cr = 0; cr < TM; cr++) {
sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr] += float(cache_a[wsir * TM + cr]) * float(cache_b[wsic * TN + cc]);
}
}
}
}
}
barrier();
}
const uint dr = ir * BM + warp_r * WM;
const uint dc = ic * BN + warp_c * WN;
const uint offsets =
#ifdef MUL_MAT_ID
expert_idx * p.expert_stride_d +
#endif
batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
[[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
[[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
const uint dr_warp = dr + wsir * WSUBM + tiwr * TM;
const uint dc_warp = dc + wsic * WSUBN + tiwc * TN;
[[unroll]] for (uint cc = 0; cc < TN; cc++) {
[[unroll]] for (uint cr = 0; cr < TM; cr++) {
if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
}
}
}
}
}
}
"""
mulmat_split_k_reduce_src = """#version 450
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {float data_a[];};
layout (binding = 1) writeonly buffer D {float data_d[];};
layout (push_constant) uniform parameter {
uint ne;
uint k_num;
} p;
void main() {
const uint idx = gl_GlobalInvocationID.x;
if (idx >= p.ne) {
return;
}
float result = 0.0f;
[[unroll]] for (uint i = 0; i < p.k_num; i++) {
result += data_a[i * p.ne + idx];
}
data_d[idx] = result;
}
"""
# DEQUANT SHADER
dequant_head = """#version 450
#extension GL_EXT_control_flow_attributes : require
#extension GL_EXT_shader_16bit_storage : require
layout (push_constant) uniform parameter
{
uint M;
uint K;
uint stride_a;
uint stride_b;
uint nel;
} p;
"""
dequant_f32_body = """
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {float data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
const uint i = gl_GlobalInvocationID.x * 16;
if (i >= p.nel) {
return;
}
[[unroll]] for (uint l = 0; l < 16; l++) {
data_b[i + l] = D_TYPE(data_a[i + l]);
}
}
"""
dequant_q4_0_body = """
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {block_q4_0 data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
const uint tid = gl_LocalInvocationID.x % 64;
const uint il = tid/32;
const uint ir = tid%32;
const uint ib = 32*i + ir;
if (ib >= p.nel / 32) {
return;
}
const uint b_idx = 1024*i + 32*ir + 8*il;
const float d = float(data_a[ib].d);
const float dm = -8.0f * d;
const uint q_idx = 8*il;
[[unroll]] for (uint l = 0; l < 8; ++l) {
data_b[b_idx + l + 0] = D_TYPE(d * (data_a[ib].qs[q_idx + l] & 0xF) + dm);
data_b[b_idx + l + 16] = D_TYPE(d * (data_a[ib].qs[q_idx + l] >> 4) + dm);
}
}
"""
dequant_q4_1_body = """
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {block_q4_1 data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
const uint tid = gl_LocalInvocationID.x % 64;
const uint il = tid/32;
const uint ir = tid%32;
const uint ib = 32*i + ir;
if (ib >= p.nel / 32) {
return;
}
const uint b_idx = 1024*i + 32*ir + 8*il;
const float d = float(data_a[ib].d);
const float m = float(data_a[ib].m);
const uint q_idx = 8*il;
[[unroll]] for (uint l = 0; l < 8; ++l) {
data_b[b_idx + l + 0] = D_TYPE(d * (data_a[ib].qs[q_idx + l] & 0xF) + m);
data_b[b_idx + l + 16] = D_TYPE(d * (data_a[ib].qs[q_idx + l] >> 4) + m);
}
}
"""
dequant_q5_0_body = """
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {block_q5_0 data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
const uint tid = gl_LocalInvocationID.x % 64;
const uint il = tid/32;
const uint ir = tid%32;
const uint ib = 32*i + ir;
if (ib >= p.nel / 32) {
return;
}
const uint b_idx = 1024*i + 32*ir + 8*il;
const float d = float(data_a[ib].d);
const uint qh = uint(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0];
const uint q_idx = 8*il;
[[unroll]] for (uint l = 0; l < 8; ++l) {
const uint iqs = q_idx + l;
const uint vui = uint(data_a[ib].qs[iqs]);
data_b[b_idx + l + 0] = D_TYPE(d * (((vui & 0xF) | (((qh >> iqs) << 4) & 0x10)) - 16.0f));
data_b[b_idx + l + 16] = D_TYPE(d * (((vui >> 4) | ((qh >> (iqs + 12)) & 0x10)) - 16.0f));
}
}
"""
dequant_q5_1_body = """
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {block_q5_1 data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
const uint tid = gl_LocalInvocationID.x % 64;
const uint il = tid/32;
const uint ir = tid%32;
const uint ib = 32*i + ir;
if (ib >= p.nel / 32) {
return;
}
const uint b_idx = 1024*i + 32*ir + 8*il;
const float d = float(data_a[ib].d);
const float m = float(data_a[ib].m);
const uint qh = data_a[ib].qh;
const uint q_idx = 8*il;
[[unroll]] for (uint l = 0; l < 8; ++l) {
const uint iqs = q_idx + l;
const uint vui = uint(data_a[ib].qs[iqs]);
data_b[b_idx + l + 0] = D_TYPE(d * (((vui & 0xF) | (((qh >> iqs) << 4) & 0x10))) + m);
data_b[b_idx + l + 16] = D_TYPE(d * (((vui >> 4) | ((qh >> (iqs + 12)) & 0x10))) + m);
}
}
"""
dequant_q8_0_body = """
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {block_q8_0 data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
const uint tid = gl_LocalInvocationID.x % 64;
const uint il = tid/32;
const uint ir = tid%32;
const uint ib = 32*i + ir;
if (ib >= p.nel / 32) {
return;
}
const uint b_idx = 1024*i + 32*ir + 16*il;
const float d = float(data_a[ib].d);
const uint q_idx = 16*il;
[[unroll]] for (uint l = 0; l < 16; l += 2) {
data_b[b_idx + l ] = D_TYPE(d * data_a[ib].qs[q_idx + l ]);
data_b[b_idx + l + 1] = D_TYPE(d * data_a[ib].qs[q_idx + l + 1]);
}
}
"""
# K-quants
dequant_q2_K_body = """
layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
[[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
const uint i = gl_WorkGroupID.x * 256 + wgy;
if (i >= p.M * p.K / QUANT_K) {
return;
}
const uint tid = gl_LocalInvocationID.x;
const uint ip = tid / 32;
const uint il = tid - 32 * ip;
const uint is = 8 * ip + il / 16;
const uint y_idx = i * QUANT_K + 128 * ip + il;
const uint ql_idx = 32 * ip + il;
const uint8_t qs = data_a[i].qs[32 * ip + il];
FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
data_b[y_idx + 0] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+0] & 0xF) * ((qs >> 0) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+0] >> 4));
data_b[y_idx + 32] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+2] & 0xF) * ((qs >> 2) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+2] >> 4));
data_b[y_idx + 64] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+4] & 0xF) * ((qs >> 4) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+4] >> 4));
data_b[y_idx + 96] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+6] & 0xF) * ((qs >> 6) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+6] >> 4));
}
}
"""
dequant_q3_K_body = """
layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
[[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
const uint i = uint(gl_WorkGroupID.x * 256 + wgy);
if (i >= p.M * p.K / QUANT_K) {
return;
}
const uint r = gl_LocalInvocationID.x / 4;
const uint tid = r / 2;
const uint is0 = r % 2;
const uint l0 = 16 * is0 + 4 * (gl_LocalInvocationID.x % 4);
const uint n = tid / 4;
const uint j = tid - 4*n;
const uint8_t m = uint8_t(1 << (4*n + j));
const uint is = 8*n + 2*j + is0;
const uint shift = 2*j;
const int8_t us = int8_t(is < 4 ? (data_a[i].scales[is-0] & 0xF) | (((data_a[i].scales[is+8] >> 0) & 3) << 4) :
is < 8 ? (data_a[i].scales[is-0] & 0xF) | (((data_a[i].scales[is+4] >> 2) & 3) << 4) :
is < 12 ? (data_a[i].scales[is-8] >> 4) | (((data_a[i].scales[is+0] >> 4) & 3) << 4) :
(data_a[i].scales[is-8] >> 4) | (((data_a[i].scales[is-4] >> 6) & 3) << 4));
const FLOAT_TYPE d_all = FLOAT_TYPE(data_a[i].d);
const FLOAT_TYPE dl = d_all * FLOAT_TYPE(us - 32);
const uint y_idx = i * QUANT_K + 128 * n + 32 * j;
const uint qs_idx = 32*n;
for (uint l = l0; l < l0 + 4; ++l) {
data_b[y_idx + l] = D_TYPE(dl * FLOAT_TYPE(int8_t((data_a[i].qs[qs_idx + l] >> shift) & 3) - (((data_a[i].hmask[l] & m) != 0) ? 0 : 4)));
}
}
}
"""
dequant_q4_K_body = """
layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
[[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
const uint i = gl_WorkGroupID.x * 256 + wgy;
if (i >= p.M * p.K / QUANT_K) {
return;
}
const uint tid = gl_LocalInvocationID.x;
const uint il = tid / 8;
const uint ir = tid % 8;
const uint is = 2 * il;
const uint n = 4;
const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
const uint y_idx = i * QUANT_K + 64 * il + n * ir;
const uint qs_idx = 32*il + n * ir;
uint8_t sc;
uint8_t m;
if (is < 4) {
sc = uint8_t(data_a[i].scales[is] & 63);
m = uint8_t(data_a[i].scales[is + 4] & 63);
} else {
sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
m = uint8_t((data_a[i].scales[is + 4] >> 4) | ((data_a[i].scales[is ] >> 6) << 4));
}
const FLOAT_TYPE d1 = dall * sc;
const FLOAT_TYPE m1 = dmin * m;
if (is < 4) {
sc = uint8_t(data_a[i].scales[is + 1] & 63);
m = uint8_t(data_a[i].scales[is + 5] & 63);
} else {
sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
m = uint8_t((data_a[i].scales[is + 5] >> 4) | ((data_a[i].scales[is + 1] >> 6) << 4));
}
const FLOAT_TYPE d2 = dall * sc;
const FLOAT_TYPE m2 = dmin * m;
[[unroll]] for (uint l = 0; l < n; ++l) {
data_b[y_idx + l ] = D_TYPE(d1 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] & 0xF) - m1);
data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] >> 4) - m2);
}
}
}
"""
dequant_q5_K_body = """
layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
[[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
const uint i = gl_WorkGroupID.x * 256 + wgy;
if (i >= p.M * p.K / QUANT_K) {
return;
}
const uint tid = gl_LocalInvocationID.x;
const uint il = tid / 16;
const uint ir = tid % 16;
const uint is = 2 * il;
const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
const uint y_idx = i * QUANT_K + 64 * il + 2 * ir;
const uint qs_idx = 32*il + 2 * ir;
const uint qh_idx = 2 * ir;
uint8_t sc;
uint8_t m;
if (is < 4) {
sc = uint8_t(data_a[i].scales[is] & 63);
m = uint8_t(data_a[i].scales[is + 4] & 63);
} else {
sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
m = uint8_t((data_a[i].scales[is + 4] >> 4) | ((data_a[i].scales[is ] >> 6) << 4));
}
const FLOAT_TYPE d1 = dall * sc;
const FLOAT_TYPE m1 = dmin * m;
if (is < 4) {
sc = uint8_t(data_a[i].scales[is + 1] & 63);
m = uint8_t(data_a[i].scales[is + 5] & 63);
} else {
sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
m = uint8_t((data_a[i].scales[is + 5] >> 4) | ((data_a[i].scales[is + 1] >> 6) << 4));
}
const FLOAT_TYPE d2 = dall * sc;
const FLOAT_TYPE m2 = dmin * m;
const uint8_t hm1 = uint8_t(1 << (2 * il ));
const uint8_t hm2 = uint8_t(1 << (2 * il + 1));
data_b[y_idx ] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx ] & 0xF) + (((data_a[i].qh[qh_idx ] & hm1) != 0) ? 16 : 0)) - m1);
data_b[y_idx + 1] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1] & 0xF) + (((data_a[i].qh[qh_idx + 1] & hm1) != 0) ? 16 : 0)) - m1);
data_b[y_idx + 32] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx ] >> 4) + (((data_a[i].qh[qh_idx ] & hm2) != 0) ? 16 : 0)) - m2);
data_b[y_idx + 33] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1] >> 4) + (((data_a[i].qh[qh_idx + 1] & hm2) != 0) ? 16 : 0)) - m2);
}
}
"""
dequant_q6_K_body = """
layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
[[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
const uint i = gl_WorkGroupID.x * 256 + wgy;
if (i >= p.M * p.K / QUANT_K) {
return;
}
const uint tid = gl_LocalInvocationID.x;
const uint ip = tid / 32;
const uint il = tid - 32 * ip;
const uint is = 8 * ip + il / 16;
const uint y_idx = i * QUANT_K + 128 * ip + il;
const uint ql_idx = 64 * ip + il;
const uint8_t qh = data_a[i].qh[32 * ip + il];
const FLOAT_TYPE d = FLOAT_TYPE(data_a[i].d);
data_b[y_idx + 0] = D_TYPE(d * FLOAT_TYPE(data_a[i].scales[is + 0] * (int8_t((data_a[i].ql[ql_idx + 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32)));
data_b[y_idx + 32] = D_TYPE(d * FLOAT_TYPE(data_a[i].scales[is + 2] * (int8_t((data_a[i].ql[ql_idx + 32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32)));
data_b[y_idx + 64] = D_TYPE(d * FLOAT_TYPE(data_a[i].scales[is + 4] * (int8_t((data_a[i].ql[ql_idx + 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32)));
data_b[y_idx + 96] = D_TYPE(d * FLOAT_TYPE(data_a[i].scales[is + 6] * (int8_t((data_a[i].ql[ql_idx + 32] >> 4) | (((qh >> 6) & 3) << 4)) - 32)));
}
}
"""
# Mul Mat Vec
mul_mat_vec_head = """#version 450
#extension GL_EXT_control_flow_attributes : enable
#extension GL_EXT_shader_16bit_storage : require
#extension GL_EXT_shader_8bit_storage : require
#ifdef MUL_MAT_ID
#define EXPERT_COUNT 8
#endif
"""
mul_mat_vec_layout = """
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
#ifdef MUL_MAT_ID
layout (binding = 3) readonly buffer IDS {int data_ids[];};
#endif
layout (push_constant) uniform parameter
{
uint ncols;
uint stride_a;
uint stride_b;
uint stride_d;
uint ne02;
uint ne12;
uint broadcast2;
uint broadcast3;
uint batch_stride_a;
uint batch_stride_b;
uint batch_stride_d;
#ifdef MUL_MAT_ID
uint expert_stride_a;
uint expert_stride_b0;
uint expert_stride_b1;
uint expert_stride_d0;
uint expert_stride_d1;
uint ne11;
uint nei0;
uint nbi1;
uint n_as;
#endif
} p;
"""
mul_mat_vec_body = """
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
shared FLOAT_TYPE tmp[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.x;
const uint tid = gl_LocalInvocationID.x;
const uint batch_idx = gl_GlobalInvocationID.y;
#ifdef MUL_MAT_ID
const uint expert_idx1 = gl_GlobalInvocationID.z / p.nei0;
const uint expert_idx0 = gl_GlobalInvocationID.z % p.nei0;
#endif
const uint i13 = batch_idx / p.ne12;
const uint i12 = batch_idx % p.ne12;
const uint i03 = i13 / p.broadcast3;
const uint i02 = i12 / p.broadcast2;
const uint batch_idx_a = i03 * p.ne02 + i02;
#ifdef MUL_MAT_ID
const uint expert_id = data_ids[expert_idx1 * p.nbi1 + expert_idx0];
#endif
const uint a_offset =
#ifdef MUL_MAT_ID
expert_id * p.expert_stride_a +
#endif
batch_idx_a * p.batch_stride_a;
const uint b_offset =
#ifdef MUL_MAT_ID
(expert_idx0 % p.ne11) * p.expert_stride_b0 +
expert_idx1 * p.expert_stride_b1 +
#endif
batch_idx * p.batch_stride_b;
const uint d_offset =
#ifdef MUL_MAT_ID
expert_idx0 * p.expert_stride_b0 +
expert_idx1 * p.expert_stride_b1 +
#endif
batch_idx * p.batch_stride_d;
const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
tmp[tid] = FLOAT_TYPE(0.0f);
[[unroll]] for (uint i = 0; i < p.ncols/BLOCK_SIZE; i += 2) {
const uint col = i*BLOCK_SIZE + 2*tid;
const uint ib = (row*p.ncols + col)/QUANT_K; // block index
const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
const uint iybs = col - col%QUANT_K; // y block start index
vec2 v = dequantize(ib, iqs, a_offset / QUANT_K);
// matrix multiplication
tmp[tid] += FLOAT_TYPE(v.x) * FLOAT_TYPE(data_b[b_offset + iybs + iqs]) +
FLOAT_TYPE(v.y) * FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
}
// sum up partial sums and write back result
barrier();
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
}
}
"""
# K-quants
mul_mat_vec_q2_K_body = """
layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
shared FLOAT_TYPE tmp[32];
void main() {
const uint row = gl_WorkGroupID.x;
const uint batch_idx = gl_GlobalInvocationID.y;
#ifdef MUL_MAT_ID
const uint expert_idx1 = gl_GlobalInvocationID.z / p.nei0;
const uint expert_idx0 = gl_GlobalInvocationID.z % p.nei0;
#endif
const uint i13 = batch_idx / p.ne12;
const uint i12 = batch_idx % p.ne12;
const uint i03 = i13 / p.broadcast3;
const uint i02 = i12 / p.broadcast2;
const uint batch_idx_a = i03 * p.ne02 + i02;
#ifdef MUL_MAT_ID
const uint expert_id = data_ids[expert_idx1 * p.nbi1 + expert_idx0];
#endif
const uint a_offset =
#ifdef MUL_MAT_ID
expert_id * p.expert_stride_a +
#endif
batch_idx_a * p.batch_stride_a;
const uint b_offset =
#ifdef MUL_MAT_ID
(expert_idx0 % p.ne11) * p.expert_stride_b0 +
expert_idx1 * p.expert_stride_b1 +
#endif
batch_idx * p.batch_stride_b;
const uint d_offset =
#ifdef MUL_MAT_ID
expert_idx0 * p.expert_stride_b0 +
expert_idx1 * p.expert_stride_b1 +
#endif
batch_idx * p.batch_stride_d;
const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const uint ix = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1
const uint step = 16/K_QUANTS_PER_ITERATION; // 16 or 8
const uint v_im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const uint v_in = tid - step*v_im; // 0...15 or 0...7
const uint l0 = K_QUANTS_PER_ITERATION*v_in; // 0...15
const uint q_offset = 32*v_im + l0;
const uint s_offset = 8*v_im;
const uint y_offset = 128*v_im + l0;
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const uint y_idx = i * QUANT_K + y_offset;
const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
sum1 += FLOAT_TYPE(data_b[b_offset + y_idx + l + 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3);
sum2 += FLOAT_TYPE(data_b[b_offset + y_idx + l + 0]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF);
}
tmp[16 * ix + tid] += dall * sum1 - dmin * sum2;
}
// sum up partial sums and write back result
barrier();
[[unroll]] for (uint s = 16; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
}
}
"""
mul_mat_vec_q3_K_body = """
layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
shared FLOAT_TYPE tmp[32];
void main() {
const uint row = gl_WorkGroupID.x;
const uint batch_idx = gl_GlobalInvocationID.y;
#ifdef MUL_MAT_ID
const uint expert_idx1 = gl_GlobalInvocationID.z / p.nei0;
const uint expert_idx0 = gl_GlobalInvocationID.z % p.nei0;
#endif
const uint i13 = batch_idx / p.ne12;
const uint i12 = batch_idx % p.ne12;
const uint i03 = i13 / p.broadcast3;
const uint i02 = i12 / p.broadcast2;
const uint batch_idx_a = i03 * p.ne02 + i02;
#ifdef MUL_MAT_ID
const uint expert_id = data_ids[expert_idx1 * p.nbi1 + expert_idx0];
#endif
const uint a_offset =
#ifdef MUL_MAT_ID
expert_id * p.expert_stride_a +
#endif
batch_idx_a * p.batch_stride_a;
const uint b_offset =
#ifdef MUL_MAT_ID
(expert_idx0 % p.ne11) * p.expert_stride_b0 +
expert_idx1 * p.expert_stride_b1 +
#endif
batch_idx * p.batch_stride_b;
const uint d_offset =
#ifdef MUL_MAT_ID
expert_idx0 * p.expert_stride_b0 +
expert_idx1 * p.expert_stride_b1 +
#endif
batch_idx * p.batch_stride_d;
const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const uint ix = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1
const uint step = 16/K_QUANTS_PER_ITERATION; // 16 or 8
const uint v_im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const uint v_in = tid - step*v_im; // 0...15 or 0...7
const uint8_t m = uint8_t(1 << (4 * v_im));
const uint l0 = K_QUANTS_PER_ITERATION*v_in; // 0...15
const uint q_offset = 32*v_im + l0;
const uint y_offset = 128*v_im + l0;
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
const uint s_shift = 4 * v_im;
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const uint y_idx = i * QUANT_K + y_offset;
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
sum += FLOAT_TYPE(data_b[b_offset + y_idx + l + 0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4))
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4))
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4))
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4))
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4))
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4))
+ FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4))
+ FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4));
}
tmp[16 * ix + tid] += d * sum;
}
// sum up partial sums and write back result
barrier();
[[unroll]] for (uint s = 16; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
}
}
"""
mul_mat_vec_q4_K_body = """
layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
shared FLOAT_TYPE tmp[32];
void main() {
const uint row = gl_WorkGroupID.x;
const uint batch_idx = gl_GlobalInvocationID.y;
#ifdef MUL_MAT_ID
const uint expert_idx1 = gl_GlobalInvocationID.z / p.nei0;
const uint expert_idx0 = gl_GlobalInvocationID.z % p.nei0;
#endif
const uint i13 = batch_idx / p.ne12;
const uint i12 = batch_idx % p.ne12;
const uint i03 = i13 / p.broadcast3;
const uint i02 = i12 / p.broadcast2;
const uint batch_idx_a = i03 * p.ne02 + i02;
#ifdef MUL_MAT_ID
const uint expert_id = data_ids[expert_idx1 * p.nbi1 + expert_idx0];
#endif
const uint a_offset =
#ifdef MUL_MAT_ID
expert_id * p.expert_stride_a +
#endif
batch_idx_a * p.batch_stride_a;
const uint b_offset =
#ifdef MUL_MAT_ID
(expert_idx0 % p.ne11) * p.expert_stride_b0 +
expert_idx1 * p.expert_stride_b1 +
#endif
batch_idx * p.batch_stride_b;
const uint d_offset =
#ifdef MUL_MAT_ID
expert_idx0 * p.expert_stride_b0 +
expert_idx1 * p.expert_stride_b1 +
#endif
batch_idx * p.batch_stride_d;
const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const uint ix = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1
const uint step = 8/K_QUANTS_PER_ITERATION; // 8 or 4
const uint il = tid/step; // 0...3
const uint ir = tid - step*il; // 0...7 or 0...3
const uint n = 2 * K_QUANTS_PER_ITERATION; // 2 or 4
const uint v_im = il / 2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const uint v_in = il % 2;
const uint l0 = n * (2 * ir + v_in); // 0...15
const uint q_offset = 32*v_im + l0;
const uint y_offset = 64*v_im + l0;
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const uint y1_idx = i * QUANT_K + y_offset;
const uint y2_idx = y1_idx + 128;
const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
const uint8_t sc0 = uint8_t( data_a[ib0 + i].scales[v_im * 2 ] & 0x3f);
const uint8_t sc1 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 1] & 0x3f);
const uint8_t sc2 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 4] & 0x3f);
const uint8_t sc3 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 5] & 0x3f);
const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 ] & 0xc0) >> 2));
const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2));
const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2));
const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2));
#if K_QUANTS_PER_ITERATION == 2
const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset ] & 0xf);
const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] & 0xf);
const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset + 2] & 0xf);
const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset + 3] & 0xf);
const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset ] >> 4);
const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] >> 4);
const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 2] >> 4);
const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 3] >> 4);
const uint8_t q4_8 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
const uint8_t q4_9 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 66] & 0xf);
const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 67] & 0xf);
const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] >> 4);
const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] >> 4);
const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 66] >> 4);
const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 67] >> 4);
const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx]) * q4_0 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * q4_1 + FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * q4_2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * q4_3);
const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_4 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_5 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7);
const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx]) * q4_8 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * q4_9 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * q4_10 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * q4_11);
const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_12 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_13 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * q4_14 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15);
const FLOAT_TYPE smin = FLOAT_TYPE(
FLOAT_TYPE(data_b[b_offset + y1_idx ]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx ]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * sc7
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * sc7
);
tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * sc0 + sy * sc1 + sz * sc4 + sw * sc5) - dmin * smin);
#else
const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset ] & 0xf);
const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] & 0xf);
const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset ] >> 4);
const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] >> 4);
const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] >> 4);
const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] >> 4);
const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx ]) * q4_0 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * q4_1);
const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3);
const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx ]) * q4_4 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * q4_5);
const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7);
const FLOAT_TYPE smin = FLOAT_TYPE(
FLOAT_TYPE(data_b[b_offset + y1_idx]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7
);
tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) + sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin);
#endif
}
// sum up partial sums and write back result
barrier();
[[unroll]] for (uint s = 16; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
}
}
"""
mul_mat_vec_q5_K_body = """
layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
shared FLOAT_TYPE tmp[32];
void main() {
const uint row = gl_WorkGroupID.x;
const uint batch_idx = gl_GlobalInvocationID.y;
#ifdef MUL_MAT_ID
const uint expert_idx1 = gl_GlobalInvocationID.z / p.nei0;
const uint expert_idx0 = gl_GlobalInvocationID.z % p.nei0;
#endif
const uint i13 = batch_idx / p.ne12;
const uint i12 = batch_idx % p.ne12;
const uint i03 = i13 / p.broadcast3;
const uint i02 = i12 / p.broadcast2;
const uint batch_idx_a = i03 * p.ne02 + i02;
#ifdef MUL_MAT_ID
const uint expert_id = data_ids[expert_idx1 * p.nbi1 + expert_idx0];
#endif
const uint a_offset =
#ifdef MUL_MAT_ID
expert_id * p.expert_stride_a +
#endif
batch_idx_a * p.batch_stride_a;
const uint b_offset =
#ifdef MUL_MAT_ID
(expert_idx0 % p.ne11) * p.expert_stride_b0 +
expert_idx1 * p.expert_stride_b1 +
#endif
batch_idx * p.batch_stride_b;
const uint d_offset =
#ifdef MUL_MAT_ID
expert_idx0 * p.expert_stride_b0 +
expert_idx1 * p.expert_stride_b1 +
#endif
batch_idx * p.batch_stride_d;
const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
const uint tid = gl_LocalInvocationID.x/2; // 0...31 or 0...16
const uint ix = gl_LocalInvocationID.x%2; // 0 or 0, 1
const uint il = tid/4; // 0...3
const uint ir = tid - 4*il; // 0...7 or 0...3
const uint v_im = il / 2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const uint v_in = il % 2;
const uint l0 = 4*ir + 2*v_in; // 0...15
const uint q_offset = 32*v_im + l0;
const uint y_offset = 64*v_im + l0;
const uint8_t hm1 = uint8_t(1 << (2*v_im));
const uint8_t hm2 = uint8_t(hm1 << 4);
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += 2) {
const uint y1_idx = i * QUANT_K + y_offset;
const uint y2_idx = y1_idx + 128;
const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
const uint8_t sc0 = uint8_t( data_a[ib0 + i].scales[v_im * 2 ] & 0x3f);
const uint8_t sc1 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 1] & 0x3f);
const uint8_t sc2 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 4] & 0x3f);
const uint8_t sc3 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 5] & 0x3f);
const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 ] & 0xc0) >> 2));
const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2));
const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2));
const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2));
const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset ] & 0xf);
const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] & 0xf);
const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset + 16] & 0xf);
const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset + 17] & 0xf);
const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset ] >> 4);
const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] >> 4);
const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 16] >> 4);
const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 17] >> 4);
const uint8_t q4_8 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
const uint8_t q4_9 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] & 0xf);
const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] & 0xf);
const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] >> 4);
const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] >> 4);
const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] >> 4);
const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] >> 4);
const FLOAT_TYPE sx = FLOAT_TYPE(
FLOAT_TYPE(data_b[b_offset + y1_idx ]) * (q4_0 + (((data_a[ib0 + i].qh[l0 ] & hm1) != 0) ? 16 : 0))
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * (q4_1 + (((data_a[ib0 + i].qh[l0 + 1] & hm1) != 0) ? 16 : 0))
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) * (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0))
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0))
);
const FLOAT_TYPE sy = FLOAT_TYPE(
FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * (q4_4 + (((data_a[ib0 + i].qh[l0 ] & (hm1 << 1)) != 0) ? 16 : 0))
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * (q4_5 + (((data_a[ib0 + i].qh[l0 + 1] & (hm1 << 1)) != 0) ? 16 : 0))
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) * (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0))
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0))
);
const FLOAT_TYPE sz = FLOAT_TYPE(
FLOAT_TYPE(data_b[b_offset + y2_idx ]) * (q4_8 + (((data_a[ib0 + i].qh[l0 ] & hm2) != 0) ? 16 : 0))
+ FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * (q4_9 + (((data_a[ib0 + i].qh[l0 + 1] & hm2) != 0) ? 16 : 0))
+ FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) * (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0))
+ FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0))
);
const FLOAT_TYPE sw = FLOAT_TYPE(
FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * (q4_12 + (((data_a[ib0 + i].qh[l0 ] & (hm2 << 1)) != 0) ? 16 : 0))
+ FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * (q4_13 + (((data_a[ib0 + i].qh[l0 + 1] & (hm2 << 1)) != 0) ? 16 : 0))
+ FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) * (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0))
+ FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0))
);
const FLOAT_TYPE smin = FLOAT_TYPE(
(FLOAT_TYPE(data_b[b_offset + y1_idx]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17])) * sc2 + (FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49])) * sc3
+ (FLOAT_TYPE(data_b[b_offset + y2_idx]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17])) * sc6 + (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7
);
tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * sc0 + sy * sc1 + sz * sc4 + sw * sc5) - dmin * smin);
}
// sum up partial sums and write back result
barrier();
[[unroll]] for (uint s = 16; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
}
}
"""
mul_mat_vec_q6_K_body = """
layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
shared FLOAT_TYPE tmp[32];
void main() {
const uint row = gl_WorkGroupID.x;
const uint batch_idx = gl_GlobalInvocationID.y;
#ifdef MUL_MAT_ID
const uint expert_idx1 = gl_GlobalInvocationID.z / p.nei0;
const uint expert_idx0 = gl_GlobalInvocationID.z % p.nei0;
#endif
const uint i13 = batch_idx / p.ne12;
const uint i12 = batch_idx % p.ne12;
const uint i03 = i13 / p.broadcast3;
const uint i02 = i12 / p.broadcast2;
const uint batch_idx_a = i03 * p.ne02 + i02;
#ifdef MUL_MAT_ID
const uint expert_id = data_ids[expert_idx1 * p.nbi1 + expert_idx0];
#endif
const uint a_offset =
#ifdef MUL_MAT_ID
expert_id * p.expert_stride_a +
#endif
batch_idx_a * p.batch_stride_a;
const uint b_offset =
#ifdef MUL_MAT_ID
(expert_idx0 % p.ne11) * p.expert_stride_b0 +
expert_idx1 * p.expert_stride_b1 +
#endif
batch_idx * p.batch_stride_b;
const uint d_offset =
#ifdef MUL_MAT_ID
expert_idx0 * p.expert_stride_b0 +
expert_idx1 * p.expert_stride_b1 +
#endif
batch_idx * p.batch_stride_d;
const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const uint ix = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1
const uint step = 16/K_QUANTS_PER_ITERATION; // 16 or 8
const uint v_im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const uint v_in = tid - step*v_im; // 0...15 or 0...7
#if K_QUANTS_PER_ITERATION == 1
const uint l0 = v_in; // 0...15
const uint is = 0;
#else
const uint l0 = 4 * v_in; // 0, 4, 8, ..., 28
const uint is = v_in / 4;
#endif
const uint ql_offset = 64*v_im + l0;
const uint qh_offset = 32*v_im + l0;
const uint s_offset = 8*v_im + is;
const uint y_offset = 128*v_im + l0;
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const uint y_idx = i * QUANT_K + y_offset;
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
#if K_QUANTS_PER_ITERATION == 1
FLOAT_TYPE sum = FLOAT_TYPE(data_b[b_offset + y_idx + 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x03) << 4)) - 32)
+ FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32)
+ FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x0c) << 2)) - 32)
+ FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32)
+ FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 0] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x30) >> 0)) - 32)
+ FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32)
+ FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0xc0) >> 2)) - 32)
+ FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32);
tmp[16 * ix + tid] += sum;
#else
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 4; ++l) {
sum += FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32)
+ FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32);
}
tmp[16 * ix + tid] += sum;
#endif
}
// sum up partial sums and write back result
barrier();
[[unroll]] for (uint s = 16; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
}
}
"""
mul_mat_p021_src = """#version 450
#extension GL_EXT_control_flow_attributes : enable
#extension GL_EXT_shader_16bit_storage : require
#define BLOCK_SIZE 32
#define FLOAT_TYPE float
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
layout (push_constant) uniform parameter
{
uint ncols_x;
uint nrows_x;
uint nchannels_x;
uint nchannels_y;
uint b_offset;
uint d_offset;
} p;
shared FLOAT_TYPE tmp[BLOCK_SIZE];
void main() {
const uint tid = gl_LocalInvocationID.x;
const uint row_x = gl_GlobalInvocationID.y;
const uint channel = gl_GlobalInvocationID.z;
const uint channel_x = channel / (p.nchannels_y / p.nchannels_x);
const uint nrows_y = p.ncols_x;
const uint nrows_dst = p.nrows_x;
const uint row_dst = row_x;
tmp[tid] = FLOAT_TYPE(0.0f);
for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) {
const uint col_x = col_x0 + tid;
if (col_x >= p.ncols_x) {
break;
}
// x is transposed and permuted
const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x;
const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
const uint row_y = col_x;
// y is not transposed but permuted
const uint iy = channel*nrows_y + row_y;
tmp[tid] += xi * FLOAT_TYPE(data_b[iy]);
}
// dst is not transposed and not permuted
const uint idst = channel*nrows_dst + row_dst;
// sum up partial sums and write back result
barrier();
[[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
dst[idst] = tmp[0];
}
}
"""
mul_mat_nc_src = """#version 450
#extension GL_EXT_control_flow_attributes : enable
#extension GL_EXT_shader_16bit_storage : require
#define BLOCK_SIZE 32
#define FLOAT_TYPE float
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
layout (push_constant) uniform parameter
{
uint ncols_x;
uint nrows_x;
uint row_stride_x;
uint channel_stride_x;
uint channel_x_divisor;
uint b_offset;
uint d_offset;
} p;
shared FLOAT_TYPE tmp[BLOCK_SIZE];
void main() {
const uint tid = gl_LocalInvocationID.x;
const uint row_x = gl_GlobalInvocationID.y;
const uint channel = gl_GlobalInvocationID.z;
const uint channel_x = channel / p.channel_x_divisor;
const uint nrows_y = p.ncols_x;
const uint nrows_dst = p.nrows_x;
const uint row_dst = row_x;
const uint idst = channel*nrows_dst + row_dst;
tmp[tid] = 0.0f;
for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) {
const uint col_x = col_x0 + tid;
if (col_x >= p.ncols_x) {
break;
}
const uint row_y = col_x;
const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
const uint iy = channel*nrows_y + row_y;
const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
tmp[tid] += xi * FLOAT_TYPE(data_b[iy]);
}
// sum up partial sums and write back result
barrier();
[[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
dst[idst] = tmp[0];
}
}
"""
generic_head = """
#version 450
#extension GL_EXT_shader_16bit_storage : require
layout (push_constant) uniform parameter
{
uint KX;
uint KY;
float param1;
float param2;
} p;
"""
generic_unary_op_head = """#version 450
#extension GL_EXT_shader_16bit_storage : require
layout (push_constant) uniform parameter
{
uint ne;
uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
uint d_offset;
float param1; float param2;
} p;"""
generic_unary_op_layout = """
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};"""
generic_unary_op_funcs = """
uint src0_idx(uint idx) {
const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
const uint i02_offset = i02*p.ne01*p.ne00;
const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
}
uint dst_idx(uint idx) {
const uint i13 = idx / (p.ne12*p.ne11*p.ne10);
const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
const uint i12 = (idx - i13_offset) / (p.ne11*p.ne10);
const uint i12_offset = i12*p.ne11*p.ne10;
const uint i11 = (idx - i13_offset - i12_offset) / p.ne10;
const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
return i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10;
}"""
generic_unary_op_main = """
void main() {
if (gl_GlobalInvocationID.x >= p.ne) {
return;
}
"""
generic_unary_op_combined = f"{generic_unary_op_head}\n{generic_unary_op_layout}\n{generic_unary_op_funcs}\n{generic_unary_op_main}"
generic_binary_op_head = """#version 450
#extension GL_EXT_shader_16bit_storage : require
layout (push_constant) uniform parameter
{
uint ne;
uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
uint d_offset;
float param1; float param2;
} p;"""
generic_binary_op_layout = """
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};"""
generic_binary_op_funcs = """
uint src0_idx(uint idx) {
const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
const uint i02_offset = i02*p.ne01*p.ne00;
const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
}
uint src1_idx(uint idx) {
const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
const uint i02_offset = i02*p.ne01*p.ne00;
const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
return (i03 % p.ne13)*p.nb13 + (i02 % p.ne12)*p.nb12 + (i01 % p.ne11)*p.nb11 + (i00 % p.ne10)*p.nb10;
}
uint dst_idx(uint idx) {
const uint i23 = idx / (p.ne22*p.ne21*p.ne20);
const uint i23_offset = i23 * p.ne22*p.ne21*p.ne20;
const uint i22 = (idx - i23_offset) / (p.ne21*p.ne20);
const uint i22_offset = i22*p.ne21*p.ne20;
const uint i21 = (idx - i23_offset - i22_offset) / p.ne20;
const uint i20 = idx - i23_offset - i22_offset - i21*p.ne20;
return i23*p.nb23 + i22*p.nb22 + i21*p.nb21 + i20*p.nb20;
}"""
generic_binary_op_main = """
void main() {
if (gl_GlobalInvocationID.x >= p.ne) {
return;
}
"""
generic_binary_op_combined = f"{generic_binary_op_head}\n{generic_binary_op_layout}\n{generic_binary_op_funcs}\n{generic_binary_op_main}"
# MUL F32
mul_body = """
data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) * FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
}
"""
# ADD
add_body = """
data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) + FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
}
"""
# SCALE
scale_body = """
data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) * FLOAT_TYPE(p.param1));
}
"""
# SQR
sqr_body = """
const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(val * val);
}
"""
# CLAMP
clamp_body = """
const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
}
"""
# CPY
cpy_end = """
data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
}
"""
# Causes an optimization error otherwise
cpy_f16_f16_end = """
data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = data_a[src0_idx(gl_GlobalInvocationID.x)];
}
"""
# GET_ROWS
get_rows_float_body = """
void main() {
const uint i00 = gl_GlobalInvocationID.x;
const uint i10 = gl_GlobalInvocationID.y;
const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
if (i00 >= p.ne00) {
return;
}
const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
#ifndef OPTIMIZATION_ERROR_WORKAROUND
data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]);
#else
data_d[d_offset + i00] = data_a[a_offset + i00];
#endif
}
"""
get_rows_body = """
void main() {
const uint i00 = (gl_GlobalInvocationID.x)*2;
const uint i10 = gl_GlobalInvocationID.y;
const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
if (i00 >= p.ne00) {
return;
}
const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
const uint ib = a_offset + i00/QUANT_K; // block index
const uint iqs = (i00%QUANT_K)/QUANT_R; // quant index
const uint iybs = i00 - i00%QUANT_K; // dst block start index
const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
vec2 v = dequantize(ib, iqs, 0);
data_d[d_offset + iybs + iqs ] = D_TYPE(v.x);
data_d[d_offset + iybs + iqs + y_offset] = D_TYPE(v.y);
}
"""
# UNARY
gelu_body = """
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
const float GELU_COEF_A = 0.044715f;
const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
const uint i = gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
const float xi = float(data_a[i]);
const float val = SQRT_2_OVER_PI*xi*(1.0f + GELU_COEF_A*xi*xi);
data_d[i] = D_TYPE(0.5f*xi*(2.0f - 2.0f / (exp(2 * val) + 1)));
}
"""
silu_body = """
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint i = gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
const float xi = float(data_a[i]);
data_d[i] = D_TYPE(xi / (1.0f + exp(-xi)));
}
"""
relu_body = """
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint i = gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
data_d[i] = max(float(data_a[i]), 0);
}
"""
# DIAG_MASK_INF
diag_mask_inf_head = """#version 450
#extension GL_EXT_shader_16bit_storage : require
layout (push_constant) uniform parameter
{
uint ncols;
uint rows_per_channel;
uint n_past;
} p;
"""
diag_mask_inf_body = """
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint col = gl_GlobalInvocationID.y;
const uint row = gl_GlobalInvocationID.x;
if (col >= p.ncols) {
return;
}
const uint i = row*p.ncols + col;
if (col > p.n_past + row % p.rows_per_channel) {
data_d[i] = D_TYPE(uintBitsToFloat(0xFF800000));
} else {
data_d[i] = D_TYPE(data_a[i]);
}
}
"""
# NORMS
norm_body = """
#extension GL_EXT_control_flow_attributes : enable
#define BLOCK_SIZE 512
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
shared vec2 sum[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.x;
const uint tid = gl_LocalInvocationID.x;
sum[tid] = vec2(0.0f, 0.0f);
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
const float xi = float(data_a[row*p.KX + col]);
sum[tid].x += xi;
sum[tid].y += xi * xi;
}
// sum up partial sums and write back result
barrier();
[[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
if (tid < s) {
sum[tid] += sum[tid + s];
}
barrier();
}
const float mean = sum[0].x / p.KX;
const float var = sum[0].y / p.KX - mean * mean;
const float inv_std = inversesqrt(var + p.param1);
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
data_d[row*p.KX + col] = D_TYPE((float(data_a[row*p.KX + col]) - mean) * inv_std);
}
}
"""
rms_norm_body = """
#extension GL_EXT_control_flow_attributes : enable
#define BLOCK_SIZE 512
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
shared FLOAT_TYPE sum[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.x;
const uint tid = gl_LocalInvocationID.x;
sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]);
sum[tid] += xi * xi;
}
// sum up partial sums and write back result
barrier();
[[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
if (tid < s) {
sum[tid] += sum[tid + s];
}
barrier();
}
const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(p.KX);
const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col]));
}
}
"""
# SOFT_MAX
soft_max_head = """
#version 450
#extension GL_EXT_shader_16bit_storage : require
layout (push_constant) uniform parameter
{
uint KX;
uint KY;
float scale;
float max_bias;
float m0;
float m1;
uint n_head_log2;
} p;
"""
soft_max_body = """
#extension GL_EXT_control_flow_attributes : enable
#define BLOCK_SIZE 512
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
layout (binding = 2) buffer D {D_TYPE data_d[];};
shared FLOAT_TYPE vals[BLOCK_SIZE];
void main() {
const uint tid = gl_LocalInvocationID.x;
const uint rowx = gl_WorkGroupID.x;
const uint rowy = rowx % p.KY;
float slope = 1.0f;
// ALiBi
if (p.max_bias > 0.0f) {
const uint h = rowx/p.KY; // head index
const float base = h < p.n_head_log2 ? p.m0 : p.m1;
const uint exp = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
slope = pow(base, exp);
}
// Find max
FLOAT_TYPE max_val = uintBitsToFloat(0xFF800000);
[[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
const uint col = col0 + tid;
if (col >= p.KX) {
break;
}
max_val = max(max_val, FLOAT_TYPE(data_a[rowx * p.KX + col]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)));
}
vals[tid] = max_val;
barrier();
[[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
if (tid < s) {
vals[tid] = max(vals[tid], vals[tid + s]);
}
barrier();
}
max_val = vals[0];
barrier();
// Sum up values
vals[tid] = FLOAT_TYPE(0.0f);
[[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
const uint col = col0 + tid;
if (col >= p.KX) {
break;
}
const uint i = rowx * p.KX + col;
const FLOAT_TYPE val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
vals[tid] += val;
data_d[i] = D_TYPE(val);
}
barrier();
[[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
if (tid < s) {
vals[tid] += vals[tid + s];
}
barrier();
}
const D_TYPE divisor = D_TYPE(vals[0]);
[[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
const uint col = col0 + tid;
if (col >= p.KX) {
break;
}
data_d[rowx*p.KX + col] /= divisor;
}
}
"""
# ROPE
rope_src = """
#version 450
#extension GL_EXT_shader_16bit_storage : require
layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) readonly buffer Y {int data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
layout (push_constant) uniform parameter {
uint ncols;
float freq_scale;
uint p_delta_rows;
float freq_base;
float ext_factor;
float attn_factor;
float corr_dims[4];
} p;
float rope_yarn_ramp(const float low, const float high, const uint i0) {
const float y = (i0 / 2 - low) / max(0.001f, high - low);
return 1.0f - min(1.0f, max(0.0f, y));
}
void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out float sin_theta) {
float mscale = p.attn_factor;
// Get n-d rotational scaling corrected for extrapolation
float theta_interp = p.freq_scale * theta_extrap;
float theta = theta_interp;
if (p.ext_factor != 0.0f) {
float ramp_mix = rope_yarn_ramp(p.corr_dims[0], p.corr_dims[1], i0) * p.ext_factor;
theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
// Get n-d magnitude scaling corrected for interpolation
mscale *= 1.0f + 0.1f * log(1.0f / p.freq_scale);
}
cos_theta = cos(theta) * mscale;
sin_theta = sin(theta) * mscale;
}
void main() {
const uint col = gl_GlobalInvocationID.y * 2;
const uint row = gl_GlobalInvocationID.x;
if (col >= p.ncols) {
return;
}
const uint i = row*p.ncols + col;
const uint i2 = row/p.p_delta_rows;
const int pos = data_b[i2];
const float theta_base = pos * pow(p.freq_base, -float(col)/p.ncols);
float cos_theta, sin_theta;
rope_yarn(theta_base, col, cos_theta, sin_theta);
const float x0 = float(data_a[i + 0]);
const float x1 = float(data_a[i + 1]);
data_d[i + 0] = D_TYPE(x0*cos_theta - x1*sin_theta);
data_d[i + 1] = D_TYPE(x0*sin_theta + x1*cos_theta);
}
"""
rope_neox_src = """
#version 450
#extension GL_EXT_shader_16bit_storage : require
layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) readonly buffer Y {int data_b[];};
layout (binding = 2) readonly buffer Z {float data_freq_factors[];};
layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
layout (push_constant) uniform parameter {
uint ncols;
uint ndims;
float freq_scale;
uint p_delta_rows;
float freq_base;
float ext_factor;
float attn_factor;
float corr_dims[4];
float theta_scale;
float inv_ndims;
uint has_freq_facs;
} p;
float rope_yarn_ramp(const float low, const float high, const uint i0) {
const float y = (i0 / 2 - low) / max(0.001f, high - low);
return 1.0f - min(1.0f, max(0.0f, y));
}
void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out float sin_theta) {
float mscale = p.attn_factor;
// Get n-d rotational scaling corrected for extrapolation
float theta_interp = p.freq_scale * theta_extrap;
float theta = theta_interp;
if (p.ext_factor != 0.0f) {
float ramp_mix = rope_yarn_ramp(p.corr_dims[0], p.corr_dims[1], i0) * p.ext_factor;
theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
// Get n-d magnitude scaling corrected for interpolation
mscale *= 1.0f + 0.1f * log(1.0f / p.freq_scale);
}
cos_theta = cos(theta) * mscale;
sin_theta = sin(theta) * mscale;
}
void main() {
const uint col = gl_GlobalInvocationID.y * 2;
const uint row = gl_GlobalInvocationID.x;
if (col >= p.ncols) {
return;
}
const uint ib = col / p.ndims;
const uint ic = col % p.ndims;
if (ib > 0) {
const uint i = row*p.ncols + ib*p.ndims + ic;
data_d[i + 0] = data_a[i + 0];
data_d[i + 1] = data_a[i + 1];
return;
}
const uint i = row*p.ncols + ib*p.ndims + ic/2;
const uint i2 = row/p.p_delta_rows;
const int pos = data_b[i2];
const float freq_factor = p.has_freq_facs != 0 ? data_freq_factors[ic/2] : 1.0f;
const float theta_base = pos*p.freq_scale*pow(p.theta_scale, col/2.0f) / freq_factor;
float cos_theta, sin_theta;
rope_yarn(theta_base, ic, cos_theta, sin_theta);
const float x0 = float(data_a[i + 0]);
const float x1 = float(data_a[i + p.ndims/2]);
data_d[i + 0] = D_TYPE(x0*cos_theta - x1*sin_theta);
data_d[i + p.ndims/2] = D_TYPE(x0*sin_theta + x1*cos_theta);
}
"""
argsort_src = """
#version 450
#extension GL_EXT_shader_16bit_storage : require
#define BLOCK_SIZE 1024
#define ASC 0
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) buffer D {int data_d[];};
layout (push_constant) uniform parameter {
uint ncols;
uint ncols_pad;
uint order;
} p;
shared int dst_row[BLOCK_SIZE];
void swap(uint idx0, uint idx1) {
int tmp = dst_row[idx0];
dst_row[idx0] = dst_row[idx1];
dst_row[idx1] = tmp;
}
void main() {
// bitonic sort
const int col = int(gl_LocalInvocationID.x);
const uint row = gl_WorkGroupID.y;
if (col >= p.ncols_pad) {
return;
}
const uint row_offset = row * p.ncols;
// initialize indices
dst_row[col] = col;
barrier();
for (uint k = 2; k <= p.ncols_pad; k *= 2) {
for (uint j = k / 2; j > 0; j /= 2) {
const uint ixj = col ^ j;
if (ixj > col) {
if ((col & k) == 0) {
if (dst_row[col] >= p.ncols ||
(dst_row[ixj] < p.ncols && (p.order == ASC ?
data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]] :
data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]]))
) {
swap(col, ixj);
}
} else {
if (dst_row[ixj] >= p.ncols ||
(dst_row[col] < p.ncols && (p.order == ASC ?
data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]] :
data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]]))
) {
swap(col, ixj);
}
}
}
barrier();
}
}
if (col < p.ncols) {
data_d[row_offset + col] = dst_row[col];
}
}
"""
GLSLC = "glslc"
VK_NUM_TYPES = 16
GGML_TYPE_F32 = 0
GGML_TYPE_F16 = 1
GGML_TYPE_Q4_0 = 2
GGML_TYPE_Q4_1 = 3
GGML_TYPE_Q5_0 = 6
GGML_TYPE_Q5_1 = 7
GGML_TYPE_Q8_0 = 8
GGML_TYPE_Q8_1 = 9
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
type_names = {
GGML_TYPE_F32: "f32",
GGML_TYPE_F16: "f16",
GGML_TYPE_Q4_0: "q4_0",
GGML_TYPE_Q4_1: "q4_1",
GGML_TYPE_Q5_0: "q5_0",
GGML_TYPE_Q5_1: "q5_1",
GGML_TYPE_Q8_0: "q8_0",
GGML_TYPE_Q8_1: "q8_1",
GGML_TYPE_Q2_K: "q2_K",
GGML_TYPE_Q3_K: "q3_K",
GGML_TYPE_Q4_K: "q4_K",
GGML_TYPE_Q5_K: "q5_K",
GGML_TYPE_Q6_K: "q6_K",
GGML_TYPE_Q8_K: "q8_K",
}
K_QUANTS_PER_ITERATION = 2
ASYNCIO_CONCURRENCY = 64
output_dir = gettempdir()
lock = asyncio.Lock()
shader_fnames = []
async def string_to_spv(name, code, defines, fp16=True):
f = NamedTemporaryFile(mode="w", delete=False)
f.write(code)
f.flush()
name = f"{name}{'_fp32' if not fp16 else ''}"
fname = os.path.join(output_dir, f"{name}.comp")
cmd = [GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", f.name, "-o", fname]
cmd.extend([f"-D{key}={value}" for key, value in defines.items()])
proc = await asyncio.create_subprocess_exec(*cmd, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE)
stdout, stderr = await proc.communicate()
stdout = stdout.decode()
error = stderr.decode()
if proc.returncode:
# Generate preprocessed code
cmd = [GLSLC, "-E", f.name]
cmd.extend([f"-D{key}={value}" for key, value in defines.items()])
proc = await asyncio.create_subprocess_exec(*cmd, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE)
stdout, stderr = await proc.communicate()
logger.info(" ".join(cmd))
if proc.returncode:
raise RuntimeError(f"{name=} {f.name=} {stdout=} {stderr=}")
preprocessed_code = stdout.decode()
cmd.extend([f"-D{key}={value}" for key, value in defines.items()])
code_with_lines = "\n".join([f"{i + 1}: {line}" for i, line in enumerate(preprocessed_code.splitlines())])
logger.error(f"cannot compile {name}\n\n{code_with_lines}\n\n{error}")
f.close()
os.remove(f.name)
sys.exit(proc.returncode)
f.close()
os.remove(f.name)
async with lock:
shader_fnames.append((name, fname))
async def main():
logger.info("ggml_vulkan: Generating and compiling shaders to SPIR-V")
tasks = []
stream = []
for fp16 in (False, True):
# mulmat
if fp16:
shader_float_type = shader_f16
load_vec = "8"
vec_type_f16 = "f16mat2x4"
vec_type = "mat2x4"
else:
shader_float_type = shader_f32
load_vec = "4"
vec_type_f16 = "f16vec4"
vec_type = "vec4"
stream.clear()
stream.extend((mulmat_head, shader_float_type, mulmat_body1, mulmat_load_scalar, mulmat_body2))
tasks.append(string_to_spv("matmul_f32", "".join(stream), {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f32_aligned", "".join(stream), {"LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f32_f16", "".join(stream), {"A_TYPE": "float", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f32_f16_aligned", "".join(stream), {"LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type, "B_TYPE": vec_type_f16, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16", "".join(stream), {"A_TYPE": "float16_t", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_aligned", "".join(stream), {"LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type_f16, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_f32", "".join(stream), {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_f32_aligned", "".join(stream), {"LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q4_0_defines, mulmat_body1, mulmat_load_q4_0, mulmat_body2))
tasks.append(string_to_spv("matmul_q4_0_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q4_0", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q4_0_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q4_0", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q4_1_defines, mulmat_body1, mulmat_load_q4_1, mulmat_body2))
tasks.append(string_to_spv("matmul_q4_1_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q4_1", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q4_1_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q4_1", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q5_0_defines, mulmat_body1, mulmat_load_q5_0, mulmat_body2))
tasks.append(string_to_spv("matmul_q5_0_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q5_0", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q5_0_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q5_0", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q5_1_defines, mulmat_body1, mulmat_load_q5_1, mulmat_body2))
tasks.append(string_to_spv("matmul_q5_1_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q5_1", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q5_1_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q5_1", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q8_0_defines, mulmat_body1, mulmat_load_q8_0, mulmat_body2))
tasks.append(string_to_spv("matmul_q8_0_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q8_0", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q8_0_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q8_0", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q2_K_defines, mulmat_body1, mulmat_load_q2_K, mulmat_body2))
tasks.append(string_to_spv("matmul_q2_k_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q2_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q2_k_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q2_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q3_K_defines, mulmat_body1, mulmat_load_q3_K, mulmat_body2))
tasks.append(string_to_spv("matmul_q3_k_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q3_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q3_k_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q3_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q4_K_defines, mulmat_body1, mulmat_load_q4_K, mulmat_body2))
tasks.append(string_to_spv("matmul_q4_k_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q4_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q4_k_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q4_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q5_K_defines, mulmat_body1, mulmat_load_q5_K, mulmat_body2))
tasks.append(string_to_spv("matmul_q5_k_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q5_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q5_k_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q5_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q6_K_defines, mulmat_body1, mulmat_load_q6_K, mulmat_body2))
tasks.append(string_to_spv("matmul_q6_k_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q6_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q6_k_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q6_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# MUL_MAT_ID
# stream.clear()
# stream.extend((mulmat_head, shader_float_type, mulmat_body1, mulmat_load_scalar, mulmat_body2))
# tasks.append(string_to_spv("matmul_id_f32", "".join(stream), {"MUL_MAT_ID": "1", "A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_f32_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_f16", "".join(stream), {"MUL_MAT_ID": "1", "A_TYPE": "float16_t", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_f16_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type_f16, "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_f16_f32", "".join(stream), {"MUL_MAT_ID": "1", "A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_f16_f32_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# stream.clear()
# stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q4_0_defines, mulmat_body1, mulmat_load_q4_0, mulmat_body2))
# tasks.append(string_to_spv("matmul_id_q4_0_f32", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "A_TYPE": "block_q4_0", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_q4_0_f32_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q4_0", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# stream.clear()
# stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q4_1_defines, mulmat_body1, mulmat_load_q4_1, mulmat_body2))
# tasks.append(string_to_spv("matmul_id_q4_1_f32", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "A_TYPE": "block_q4_1", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_q4_1_f32_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q4_1", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# stream.clear()
# stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q5_0_defines, mulmat_body1, mulmat_load_q5_0, mulmat_body2))
# tasks.append(string_to_spv("matmul_id_q5_0_f32", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "A_TYPE": "block_q5_0", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_q5_0_f32_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q5_0", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# stream.clear()
# stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q5_1_defines, mulmat_body1, mulmat_load_q5_1, mulmat_body2))
# tasks.append(string_to_spv("matmul_id_q5_1_f32", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "A_TYPE": "block_q5_1", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_q5_1_f32_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q5_1", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# stream.clear()
# stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q8_0_defines, mulmat_body1, mulmat_load_q8_0, mulmat_body2))
# tasks.append(string_to_spv("matmul_id_q8_0_f32", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "A_TYPE": "block_q8_0", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_q8_0_f32_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q8_0", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# stream.clear()
# stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q2_K_defines, mulmat_body1, mulmat_load_q2_K, mulmat_body2))
# tasks.append(string_to_spv("matmul_id_q2_k_f32", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "A_TYPE": "block_q2_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_q2_k_f32_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q2_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# stream.clear()
# stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q3_K_defines, mulmat_body1, mulmat_load_q3_K, mulmat_body2))
# tasks.append(string_to_spv("matmul_id_q3_k_f32", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "A_TYPE": "block_q3_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_q3_k_f32_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q3_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# stream.clear()
# stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q4_K_defines, mulmat_body1, mulmat_load_q4_K, mulmat_body2))
# tasks.append(string_to_spv("matmul_id_q4_k_f32", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "A_TYPE": "block_q4_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_q4_k_f32_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q4_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# stream.clear()
# stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q5_K_defines, mulmat_body1, mulmat_load_q5_K, mulmat_body2))
# tasks.append(string_to_spv("matmul_id_q5_k_f32", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "A_TYPE": "block_q5_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_q5_k_f32_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q5_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# stream.clear()
# stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q6_K_defines, mulmat_body1, mulmat_load_q6_K, mulmat_body2))
# tasks.append(string_to_spv("matmul_id_q6_k_f32", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "A_TYPE": "block_q6_K", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
# tasks.append(string_to_spv("matmul_id_q6_k_f32_aligned", "".join(stream), {"MUL_MAT_ID": "1", "LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q6_K", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# Shaders where precision is needed, so no fp16 version
# mul mat vec
for i in range(0, VK_NUM_TYPES):
stream.clear()
stream.extend((mul_mat_vec_head, shader_int8_ext, shader_f32))
if i == GGML_TYPE_F16:
stream.extend((shader_f16_defines, mul_mat_vec_layout, shader_float_dequant_func, mul_mat_vec_body))
elif i == GGML_TYPE_Q4_0:
stream.extend((shader_q4_0_defines, mul_mat_vec_layout, shader_q4_0_dequant_func, mul_mat_vec_body))
elif i == GGML_TYPE_Q4_1:
stream.extend((shader_q4_1_defines, mul_mat_vec_layout, shader_q4_1_dequant_func, mul_mat_vec_body))
elif i == GGML_TYPE_Q5_0:
stream.extend((shader_q5_0_defines, mul_mat_vec_layout, shader_q5_0_dequant_func, mul_mat_vec_body))
elif i == GGML_TYPE_Q5_1:
stream.extend((shader_q5_1_defines, mul_mat_vec_layout, shader_q5_1_dequant_func, mul_mat_vec_body))
elif i == GGML_TYPE_Q8_0:
stream.extend((shader_q8_0_defines, mul_mat_vec_layout, shader_q8_0_dequant_func, mul_mat_vec_body))
elif i == GGML_TYPE_Q2_K:
stream.extend((shader_q2_K_defines, mul_mat_vec_layout, mul_mat_vec_q2_K_body))
elif i == GGML_TYPE_Q3_K:
stream.extend((shader_q3_K_defines, mul_mat_vec_layout, mul_mat_vec_q3_K_body))
elif i == GGML_TYPE_Q4_K:
stream.extend((shader_q4_K_defines, mul_mat_vec_layout, mul_mat_vec_q4_K_body))
elif i == GGML_TYPE_Q5_K:
stream.extend((shader_q5_K_defines, mul_mat_vec_layout, mul_mat_vec_q5_K_body))
elif i == GGML_TYPE_Q6_K:
stream.extend((shader_q6_K_defines, mul_mat_vec_layout, mul_mat_vec_q6_K_body))
else:
continue
tasks.append(string_to_spv(f"mul_mat_vec_{type_names[i]}_f32_f32", "".join(stream), {"B_TYPE": "float", "D_TYPE": "float", "K_QUANTS_PER_ITERATION": K_QUANTS_PER_ITERATION}))
tasks.append(string_to_spv(f"mul_mat_vec_{type_names[i]}_f16_f32", "".join(stream), {"B_TYPE": "float16_t", "D_TYPE": "float", "K_QUANTS_PER_ITERATION": K_QUANTS_PER_ITERATION}))
# tasks.append(string_to_spv(f"mul_mat_vec_id_{type_names[i]}_f32", "".join(stream), {"MUL_MAT_ID": "1", "B_TYPE": "float", "D_TYPE": "float", "K_QUANTS_PER_ITERATION": K_QUANTS_PER_ITERATION}))
# Dequant shaders
for i in range(0, VK_NUM_TYPES):
stream.clear()
stream.extend((dequant_head, shader_int8_ext, shader_f32))
if i == GGML_TYPE_F32:
stream.append(dequant_f32_body)
elif i == GGML_TYPE_Q4_0:
stream.extend((shader_q4_0_defines, dequant_q4_0_body))
elif i == GGML_TYPE_Q4_1:
stream.extend((shader_q4_1_defines, dequant_q4_1_body))
elif i == GGML_TYPE_Q5_0:
stream.extend((shader_q5_0_defines, dequant_q5_0_body))
elif i == GGML_TYPE_Q5_1:
stream.extend((shader_q5_1_defines, dequant_q5_1_body))
elif i == GGML_TYPE_Q8_0:
stream.extend((shader_q8_0_defines, dequant_q8_0_body))
elif i == GGML_TYPE_Q2_K:
stream.extend((shader_q2_K_defines, dequant_q2_K_body))
elif i == GGML_TYPE_Q3_K:
stream.extend((shader_q3_K_defines, dequant_q3_K_body))
elif i == GGML_TYPE_Q4_K:
stream.extend((shader_q4_K_defines, dequant_q4_K_body))
elif i == GGML_TYPE_Q5_K:
stream.extend((shader_q5_K_defines, dequant_q5_K_body))
elif i == GGML_TYPE_Q6_K:
stream.extend((shader_q6_K_defines, dequant_q6_K_body))
else:
continue
tasks.append(string_to_spv(f"dequant_{type_names[i]}", "".join(stream), {"D_TYPE": "float16_t"}))
# get_rows
for i in range(0, VK_NUM_TYPES):
stream.clear()
stream.extend((generic_binary_op_head, shader_int8_ext, shader_f32))
optimization_workaround = False
if i == GGML_TYPE_F32:
stream.extend((shader_f32_defines, generic_binary_op_layout, generic_binary_op_funcs, get_rows_float_body))
elif i == GGML_TYPE_F16:
stream.extend((shader_f16_defines, generic_binary_op_layout, generic_binary_op_funcs, get_rows_float_body))
optimization_workaround = True
elif i == GGML_TYPE_Q4_0:
stream.extend((shader_q4_0_defines, generic_binary_op_layout, shader_q4_0_dequant_func, generic_binary_op_funcs, get_rows_body))
elif i == GGML_TYPE_Q4_1:
stream.extend((shader_q4_1_defines, generic_binary_op_layout, shader_q4_1_dequant_func, generic_binary_op_funcs, get_rows_body))
elif i == GGML_TYPE_Q5_0:
stream.extend((shader_q5_0_defines, generic_binary_op_layout, shader_q5_0_dequant_func, generic_binary_op_funcs, get_rows_body))
elif i == GGML_TYPE_Q5_1:
stream.extend((shader_q5_1_defines, generic_binary_op_layout, shader_q5_1_dequant_func, generic_binary_op_funcs, get_rows_body))
elif i == GGML_TYPE_Q8_0:
stream.extend((shader_q8_0_defines, generic_binary_op_layout, shader_q8_0_dequant_func, generic_binary_op_funcs, get_rows_body))
else:
continue
if optimization_workaround:
tasks.append(string_to_spv(f"get_rows_{type_names[i]}", "".join(stream), {"B_TYPE": "int", "D_TYPE": "float16_t", "OPTIMIZATION_ERROR_WORKAROUND": "1"}))
else:
tasks.append(string_to_spv(f"get_rows_{type_names[i]}", "".join(stream), {"B_TYPE": "int", "D_TYPE": "float16_t"}))
tasks.append(string_to_spv(f"get_rows_{type_names[i]}_f32", "".join(stream), {"B_TYPE": "int", "D_TYPE": "float"}))
tasks.append(string_to_spv("mul_mat_vec_p021_f16_f32", mul_mat_p021_src, {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
tasks.append(string_to_spv("mul_mat_vec_nc_f16_f32", mul_mat_nc_src, {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
# Norms
tasks.append(string_to_spv("norm_f32", f"{generic_head}\n{shader_f32}\n{norm_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
tasks.append(string_to_spv("rms_norm_f32", f"{generic_head}\n{shader_f32}\n{rms_norm_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
tasks.append(string_to_spv("cpy_f32_f32", f"{generic_unary_op_combined}\n{cpy_end}", {"A_TYPE": "float", "D_TYPE": "float"}))
tasks.append(string_to_spv("cpy_f32_f16", f"{generic_unary_op_combined}\n{cpy_end}", {"A_TYPE": "float", "D_TYPE": "float16_t"}))
tasks.append(string_to_spv("cpy_f16_f16", f"{generic_unary_op_combined}\n{cpy_f16_f16_end}", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
tasks.append(string_to_spv("add_f32", f"{generic_binary_op_combined}\n{add_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
tasks.append(string_to_spv("split_k_reduce", mulmat_split_k_reduce_src, {}))
tasks.append(string_to_spv("mul_f32", f"{generic_binary_op_combined}\n{mul_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
tasks.append(string_to_spv("scale_f32", f"{generic_unary_op_combined}\n{scale_body}", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
tasks.append(string_to_spv("sqr_f32", f"{generic_unary_op_combined}\n{sqr_body}", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
tasks.append(string_to_spv("clamp_f32", f"{generic_unary_op_combined}\n{clamp_body}", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
tasks.append(string_to_spv("gelu_f32", f"{generic_head}\n{shader_f32}\n{gelu_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
tasks.append(string_to_spv("silu_f32", f"{generic_head}\n{shader_f32}\n{silu_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
tasks.append(string_to_spv("relu_f32", f"{generic_head}\n{shader_f32}\n{relu_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
tasks.append(string_to_spv("diag_mask_inf_f32", f"{diag_mask_inf_head}\n{shader_f32}\n{diag_mask_inf_body}", {"A_TYPE": "float", "D_TYPE": "float"}))
tasks.append(string_to_spv("soft_max_f32", f"{soft_max_head}\n{shader_f32}\n{soft_max_body}", {"A_TYPE": "float", "B_TYPE": "float", "C_TYPE": "float", "D_TYPE": "float"}))
tasks.append(string_to_spv("soft_max_f32_f16", f"{soft_max_head}\n{shader_f32}\n{soft_max_body}", {"A_TYPE": "float", "B_TYPE": "float16_t", "C_TYPE": "float16_t", "D_TYPE": "float"}))
tasks.append(string_to_spv("rope_f32", rope_src, {"A_TYPE": "float", "D_TYPE": "float"}))
tasks.append(string_to_spv("rope_f16", rope_src, {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
tasks.append(string_to_spv("rope_neox_f32", rope_neox_src, {"A_TYPE": "float", "D_TYPE": "float"}))
tasks.append(string_to_spv("rope_neox_f16", rope_neox_src, {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
tasks.append(string_to_spv("argsort_f32", argsort_src, {"A_TYPE": "float"}))
# Helper to decorate tasks with semaphore acquisition.
async def withSemaphore(sem, task):
async with sem:
return await task
# Run tasks concurrently guarded by a concurrency limit.
sem = asyncio.Semaphore(ASYNCIO_CONCURRENCY)
await asyncio.gather(*(withSemaphore(sem, task) for task in tasks))
with open("ggml-vulkan-shaders.hpp", "w") as f:
f.write("#include <cstdint>\n\n")
for name, path in sorted(shader_fnames):
with open(path, "rb") as spv:
counter = 0
newline_counter = 0
f.write(f"unsigned char {name}_data[] = {{\n")
for val in spv.read():
f.write(f"0x{val:02x},")
newline_counter += 1
counter += 1
if newline_counter >= 12:
newline_counter = 0
f.write("\n")
f.write("\n};\n")
f.write(f"const uint64_t {name}_len = {counter};\n\n")
os.remove(path)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="GGML Vulkan Shader Generator")
parser.add_argument("--glslc", help="Path to glslc")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args()
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
if args.glslc:
GLSLC = args.glslc
asyncio.run(main())
llm/llama.cpp/gguf-py/scripts/gguf-convert-endian.py deleted 100644 → 0
View file @ 97b02a89
#!/usr/bin/env python3
from __future__ import annotations
import logging
import argparse
import os
import sys
from tqdm import tqdm
from pathlib import Path
import numpy as np
# Necessary to load the local gguf package
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent))
import gguf
logger = logging.getLogger("gguf-convert-endian")
def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None:
if np.uint32(1) == np.uint32(1).newbyteorder("<"):
# Host is little endian
host_endian = "little"
swapped_endian = "big"
else:
# Sorry PDP or other weird systems that don't use BE or LE.
host_endian = "big"
swapped_endian = "little"
if reader.byte_order == "S":
file_endian = swapped_endian
else:
file_endian = host_endian
order = host_endian if args.order == "native" else args.order
logger.info(f"* Host is {host_endian.upper()} endian, GGUF file seems to be {file_endian.upper()} endian")
if file_endian == order:
logger.info(f"* File is already {order.upper()} endian. Nothing to do.")
sys.exit(0)
logger.info("* Checking tensors for conversion compatibility")
for tensor in reader.tensors:
if tensor.tensor_type not in (
gguf.GGMLQuantizationType.F32,
gguf.GGMLQuantizationType.F16,
gguf.GGMLQuantizationType.Q8_0,
):
raise ValueError(f"Cannot handle type {tensor.tensor_type.name} for tensor {repr(tensor.name)}")
logger.info(f"* Preparing to convert from {file_endian.upper()} to {order.upper()}")
if args.dry_run:
return
logger.warning("*** Warning *** Warning *** Warning **")
logger.warning("* This conversion process may damage the file. Ensure you have a backup.")
if order != host_endian:
logger.warning("* Requested endian differs from host, you will not be able to load the model on this machine.")
logger.warning("* The file will be modified immediately, so if conversion fails or is interrupted")
logger.warning("* the file will be corrupted. Enter exactly YES if you are positive you want to proceed:")
response = input("YES, I am sure> ")
if response != "YES":
logger.warning("You didn't enter YES. Okay then, see ya!")
sys.exit(0)
logger.info(f"* Converting fields ({len(reader.fields)})")
for idx, field in enumerate(reader.fields.values()):
logger.info(f"- {idx:4}: Converting field {repr(field.name)}, part count: {len(field.parts)}")
for part in field.parts:
part.byteswap(inplace=True)
logger.info(f"* Converting tensors ({len(reader.tensors)})")
for idx, tensor in enumerate(pbar := tqdm(reader.tensors, desc="Converting tensor")):
log_message = (
f"Converting tensor {repr(tensor.name)}, "
f"type={tensor.tensor_type.name}, "
f"elements={tensor.n_elements} "
)
# Byte-swap each part of the tensor's field
for part in tensor.field.parts:
part.byteswap(inplace=True)
# Byte-swap tensor data if necessary
if tensor.tensor_type == gguf.GGMLQuantizationType.Q8_0:
# Handle Q8_0 tensor blocks (block_q8_0)
# Specific handling of block_q8_0 is required.
# Each block_q8_0 consists of an f16 delta (scaling factor) followed by 32 int8 quantizations.
block_size = 34 # 34 bytes = <f16 delta scaling factor> + 32 * <int8 quant>
n_blocks = len(tensor.data) // block_size
for block_num in (inner_pbar := tqdm(range(n_blocks), desc="Byte-swapping Blocks", leave=False)):
block_offs = block_num * block_size
# Byte-Swap f16 sized delta field
delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16)
delta.byteswap(inplace=True)
# Byte-Swap Q8 weights
if block_num % 100000 == 0:
inner_pbar.set_description(f"Byte-swapping Blocks [{(n_blocks - block_num) // n_blocks}]")
else:
# Handle other tensor types
tensor.data.byteswap(inplace=True)
pbar.set_description(log_message)
logger.info("* Completion")
def main() -> None:
parser = argparse.ArgumentParser(description="Convert GGUF file byte order")
parser.add_argument(
"model", type=str,
help="GGUF format model filename",
)
parser.add_argument(
"order", type=str, choices=['big', 'little', 'native'],
help="Requested byte order",
)
parser.add_argument(
"--dry-run", action="store_true",
help="Don't actually change anything",
)
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
logger.info(f'* Loading: {args.model}')
reader = gguf.GGUFReader(args.model, 'r' if args.dry_run else 'r+')
convert_byteorder(reader, args)
if __name__ == "__main__":
main()
llm/llama.cpp/gguf-py/scripts/gguf-dump.py deleted 100644 → 0
View file @ 97b02a89
#!/usr/bin/env python3
from __future__ import annotations
import logging
import argparse
import os
import sys
from pathlib import Path
from typing import Any
import numpy as np
# Necessary to load the local gguf package
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent))
from gguf import GGUFReader, GGUFValueType # noqa: E402
logger = logging.getLogger("gguf-dump")
def get_file_host_endian(reader: GGUFReader) -> tuple[str, str]:
host_endian = 'LITTLE' if np.uint32(1) == np.uint32(1).newbyteorder("<") else 'BIG'
if reader.byte_order == 'S':
file_endian = 'BIG' if host_endian == 'LITTLE' else 'LITTLE'
else:
file_endian = host_endian
return (host_endian, file_endian)
# For more information about what field.parts and field.data represent,
# please see the comments in the modify_gguf.py example.
def dump_metadata(reader: GGUFReader, args: argparse.Namespace) -> None:
host_endian, file_endian = get_file_host_endian(reader)
print(f'* File is {file_endian} endian, script is running on a {host_endian} endian host.') # noqa: NP100
print(f'* Dumping {len(reader.fields)} key/value pair(s)') # noqa: NP100
for n, field in enumerate(reader.fields.values(), 1):
if not field.types:
pretty_type = 'N/A'
elif field.types[0] == GGUFValueType.ARRAY:
nest_count = len(field.types) - 1
pretty_type = '[' * nest_count + str(field.types[-1].name) + ']' * nest_count
else:
pretty_type = str(field.types[-1].name)
log_message = f' {n:5}: {pretty_type:10} | {len(field.data):8} | {field.name}'
if len(field.types) == 1:
curr_type = field.types[0]
if curr_type == GGUFValueType.STRING:
log_message += ' = {0}'.format(repr(str(bytes(field.parts[-1]), encoding='utf-8')[:60]))
elif field.types[0] in reader.gguf_scalar_to_np:
log_message += ' = {0}'.format(field.parts[-1][0])
print(log_message) # noqa: NP100
if args.no_tensors:
return
print(f'* Dumping {len(reader.tensors)} tensor(s)') # noqa: NP100
for n, tensor in enumerate(reader.tensors, 1):
prettydims = ', '.join('{0:5}'.format(d) for d in list(tensor.shape) + [1] * (4 - len(tensor.shape)))
print(f' {n:5}: {tensor.n_elements:10} | {prettydims} | {tensor.tensor_type.name:7} | {tensor.name}') # noqa: NP100
def dump_metadata_json(reader: GGUFReader, args: argparse.Namespace) -> None:
import json
host_endian, file_endian = get_file_host_endian(reader)
metadata: dict[str, Any] = {}
tensors: dict[str, Any] = {}
result = {
"filename": args.model,
"endian": file_endian,
"metadata": metadata,
"tensors": tensors,
}
for idx, field in enumerate(reader.fields.values()):
curr: dict[str, Any] = {
"index": idx,
"type": field.types[0].name if field.types else 'UNKNOWN',
"offset": field.offset,
}
metadata[field.name] = curr
if field.types[:1] == [GGUFValueType.ARRAY]:
curr["array_types"] = [t.name for t in field.types][1:]
if not args.json_array:
continue
itype = field.types[-1]
if itype == GGUFValueType.STRING:
curr["value"] = [str(bytes(field.parts[idx]), encoding="utf-8") for idx in field.data]
else:
curr["value"] = [pv for idx in field.data for pv in field.parts[idx].tolist()]
elif field.types[0] == GGUFValueType.STRING:
curr["value"] = str(bytes(field.parts[-1]), encoding="utf-8")
else:
curr["value"] = field.parts[-1].tolist()[0]
if not args.no_tensors:
for idx, tensor in enumerate(reader.tensors):
tensors[tensor.name] = {
"index": idx,
"shape": tensor.shape.tolist(),
"type": tensor.tensor_type.name,
"offset": tensor.field.offset,
}
json.dump(result, sys.stdout)
def main() -> None:
parser = argparse.ArgumentParser(description="Dump GGUF file metadata")
parser.add_argument("model", type=str, help="GGUF format model filename")
parser.add_argument("--no-tensors", action="store_true", help="Don't dump tensor metadata")
parser.add_argument("--json", action="store_true", help="Produce JSON output")
parser.add_argument("--json-array", action="store_true", help="Include full array values in JSON output (long)")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
if not args.json:
logger.info(f'* Loading: {args.model}')
reader = GGUFReader(args.model, 'r')
if args.json:
dump_metadata_json(reader, args)
else:
dump_metadata(reader, args)
if __name__ == '__main__':
main()
llm/llama.cpp/gguf-py/scripts/gguf-new-metadata.py deleted 100644 → 0
View file @ 97b02a89
#!/usr/bin/env python3
import logging
import argparse
import os
import sys
import json
from pathlib import Path
import numpy as np
from tqdm import tqdm
from typing import Any, Sequence, NamedTuple
# Necessary to load the local gguf package
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent))
import gguf
logger = logging.getLogger("gguf-new-metadata")
class MetadataDetails(NamedTuple):
type: gguf.GGUFValueType
value: Any
description: str = ''
def get_byteorder(reader: gguf.GGUFReader) -> gguf.GGUFEndian:
if np.uint32(1) == np.uint32(1).newbyteorder("<"):
# Host is little endian
host_endian = gguf.GGUFEndian.LITTLE
swapped_endian = gguf.GGUFEndian.BIG
else:
# Sorry PDP or other weird systems that don't use BE or LE.
host_endian = gguf.GGUFEndian.BIG
swapped_endian = gguf.GGUFEndian.LITTLE
if reader.byte_order == "S":
return swapped_endian
else:
return host_endian
def decode_field(field: gguf.ReaderField | None) -> Any:
if field and field.types:
main_type = field.types[0]
if main_type == gguf.GGUFValueType.ARRAY:
sub_type = field.types[-1]
if sub_type == gguf.GGUFValueType.STRING:
return [str(bytes(field.parts[idx]), encoding='utf-8') for idx in field.data]
else:
return [pv for idx in field.data for pv in field.parts[idx].tolist()]
if main_type == gguf.GGUFValueType.STRING:
return str(bytes(field.parts[-1]), encoding='utf-8')
else:
return field.parts[-1][0]
return None
def get_field_data(reader: gguf.GGUFReader, key: str) -> Any:
field = reader.get_field(key)
return decode_field(field)
def find_token(token_list: Sequence[int], token: str) -> Sequence[int]:
token_ids = [index for index, value in enumerate(token_list) if value == token]
if len(token_ids) == 0:
raise LookupError(f'Unable to find "{token}" in token list!')
return token_ids
def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new_metadata: dict[str, MetadataDetails], remove_metadata: Sequence[str]) -> None:
for field in reader.fields.values():
# Suppress virtual fields and fields written by GGUFWriter
if field.name == gguf.Keys.General.ARCHITECTURE or field.name.startswith('GGUF.'):
logger.debug(f'Suppressing {field.name}')
continue
# Skip old chat templates if we have new ones
if field.name.startswith(gguf.Keys.Tokenizer.CHAT_TEMPLATE) and gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata:
logger.debug(f'Skipping {field.name}')
continue
if field.name in remove_metadata:
logger.debug(f'Removing {field.name}')
continue
old_val = MetadataDetails(field.types[0], decode_field(field))
val = new_metadata.get(field.name, old_val)
if field.name in new_metadata:
logger.debug(f'Modifying {field.name}: "{old_val.value}" -> "{val.value}" {val.description}')
del new_metadata[field.name]
elif val.value is not None:
logger.debug(f'Copying {field.name}')
if val.value is not None:
writer.add_key(field.name)
writer.add_val(val.value, val.type)
if gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata:
logger.debug('Adding chat template(s)')
writer.add_chat_template(new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE].value)
del new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE]
for key, val in new_metadata.items():
logger.debug(f'Adding {key}: "{val.value}" {val.description}')
writer.add_key(key)
writer.add_val(val.value, val.type)
total_bytes = 0
for tensor in reader.tensors:
total_bytes += tensor.n_bytes
writer.add_tensor_info(tensor.name, tensor.data.shape, tensor.data.dtype, tensor.data.nbytes, tensor.tensor_type)
bar = tqdm(desc="Writing", total=total_bytes, unit="byte", unit_scale=True)
writer.write_header_to_file()
writer.write_kv_data_to_file()
writer.write_ti_data_to_file()
for tensor in reader.tensors:
writer.write_tensor_data(tensor.data)
bar.update(tensor.n_bytes)
writer.close()
def main() -> None:
tokenizer_metadata = (getattr(gguf.Keys.Tokenizer, n) for n in gguf.Keys.Tokenizer.__dict__.keys() if not n.startswith('_'))
token_names = dict((n.split('.')[-1][:-len('_token_id')], n) for n in tokenizer_metadata if n.endswith('_token_id'))
parser = argparse.ArgumentParser(description="Make a copy of a GGUF file with new metadata")
parser.add_argument("input", type=Path, help="GGUF format model input filename")
parser.add_argument("output", type=Path, help="GGUF format model output filename")
parser.add_argument("--general-name", type=str, help="The models general.name", metavar='"name"')
parser.add_argument("--general-description", type=str, help="The models general.description", metavar='"Description ..."')
parser.add_argument("--chat-template", type=str, help="Chat template string (or JSON string containing templates)", metavar='"{% ... %} ..."')
parser.add_argument("--chat-template-config", type=Path, help="Config file containing chat template(s)", metavar='tokenizer_config.json')
parser.add_argument("--pre-tokenizer", type=str, help="The models tokenizer.ggml.pre", metavar='"pre tokenizer"')
parser.add_argument("--remove-metadata", action="append", type=str, help="Remove metadata (by key name) from output model", metavar='general.url')
parser.add_argument("--special-token", action="append", type=str, help="Special token by value", nargs=2, metavar=(' | '.join(token_names.keys()), '"<token>"'))
parser.add_argument("--special-token-by-id", action="append", type=str, help="Special token by id", nargs=2, metavar=(' | '.join(token_names.keys()), '0'))
parser.add_argument("--force", action="store_true", help="Bypass warnings without confirmation")
parser.add_argument("--verbose", action="store_true", help="Increase output verbosity")
args = parser.parse_args(None if len(sys.argv) > 2 else ["--help"])
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
new_metadata = {}
remove_metadata = args.remove_metadata or []
if args.general_name:
new_metadata[gguf.Keys.General.NAME] = MetadataDetails(gguf.GGUFValueType.STRING, args.general_name)
if args.general_description:
new_metadata[gguf.Keys.General.DESCRIPTION] = MetadataDetails(gguf.GGUFValueType.STRING, args.general_description)
if args.chat_template:
new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, json.loads(args.chat_template) if args.chat_template.startswith('[') else args.chat_template)
if args.chat_template_config:
with open(args.chat_template_config, 'r') as fp:
config = json.load(fp)
template = config.get('chat_template')
if template:
new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template)
if args.pre_tokenizer:
new_metadata[gguf.Keys.Tokenizer.PRE] = MetadataDetails(gguf.GGUFValueType.STRING, args.pre_tokenizer)
if remove_metadata:
logger.warning('*** Warning *** Warning *** Warning **')
logger.warning('* Most metadata is required for a fully functional GGUF file,')
logger.warning('* removing crucial metadata may result in a corrupt output file!')
if not args.force:
logger.warning('* Enter exactly YES if you are positive you want to proceed:')
response = input('YES, I am sure> ')
if response != 'YES':
logger.info("You didn't enter YES. Okay then, see ya!")
sys.exit(0)
logger.info(f'* Loading: {args.input}')
reader = gguf.GGUFReader(args.input, 'r')
arch = get_field_data(reader, gguf.Keys.General.ARCHITECTURE)
endianess = get_byteorder(reader)
token_list = get_field_data(reader, gguf.Keys.Tokenizer.LIST) or []
for name, token in args.special_token or []:
if name not in token_names:
logger.warning(f'Unknown special token "{name}", ignoring...')
else:
ids = find_token(token_list, token)
new_metadata[token_names[name]] = MetadataDetails(gguf.GGUFValueType.UINT32, ids[0], f'= {token}')
if len(ids) > 1:
logger.warning(f'Multiple "{token}" tokens found, choosing ID {ids[0]}, use --special-token-by-id if you want another:')
logger.warning(', '.join(str(i) for i in ids))
for name, id_string in args.special_token_by_id or []:
if name not in token_names:
logger.warning(f'Unknown special token "{name}", ignoring...')
elif not id_string.isdecimal():
raise LookupError(f'Token ID "{id_string}" is not a valid ID!')
else:
id_int = int(id_string)
if id_int >= 0 and id_int < len(token_list):
new_metadata[token_names[name]] = MetadataDetails(gguf.GGUFValueType.UINT32, id_int, f'= {token_list[id_int]}')
else:
raise LookupError(f'Token ID {id_int} is not within token list!')
if os.path.isfile(args.output) and not args.force:
logger.warning('*** Warning *** Warning *** Warning **')
logger.warning(f'* The "{args.output}" GGUF file already exists, it will be overwritten!')
logger.warning('* Enter exactly YES if you are positive you want to proceed:')
response = input('YES, I am sure> ')
if response != 'YES':
logger.info("You didn't enter YES. Okay then, see ya!")
sys.exit(0)
logger.info(f'* Writing: {args.output}')
writer = gguf.GGUFWriter(args.output, arch=arch, endianess=endianess)
alignment = get_field_data(reader, gguf.Keys.General.ALIGNMENT)
if alignment is not None:
logger.debug(f'Setting custom alignment: {alignment}')
writer.data_alignment = alignment
copy_with_new_metadata(reader, writer, new_metadata, remove_metadata)
if __name__ == '__main__':
main()
llm/llama.cpp/gguf-py/scripts/gguf-set-metadata.py deleted 100644 → 0
View file @ 97b02a89
#!/usr/bin/env python3
import logging
import argparse
import os
import sys
from pathlib import Path
# Necessary to load the local gguf package
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent))
from gguf import GGUFReader # noqa: E402
logger = logging.getLogger("gguf-set-metadata")
def minimal_example(filename: str) -> None:
reader = GGUFReader(filename, 'r+')
field = reader.fields['tokenizer.ggml.bos_token_id']
if field is None:
return
part_index = field.data[0]
field.parts[part_index][0] = 2 # Set tokenizer.ggml.bos_token_id to 2
#
# So what's this field.data thing? It's helpful because field.parts contains
# _every_ part of the GGUF field. For example, tokenizer.ggml.bos_token_id consists
# of:
#
# Part index 0: Key length (27)
# Part index 1: Key data ("tokenizer.ggml.bos_token_id")
# Part index 2: Field type (4, the id for GGUFValueType.UINT32)
# Part index 3: Field value
#
# Note also that each part is an NDArray slice, so even a part that
# is only a single value like the key length will be a NDArray of
# the key length type (numpy.uint32).
#
# The .data attribute in the Field is a list of relevant part indexes
# and doesn't contain internal GGUF details like the key length part.
# In this case, .data will be [3] - just the part index of the
# field value itself.
def set_metadata(reader: GGUFReader, args: argparse.Namespace) -> None:
field = reader.get_field(args.key)
if field is None:
logger.error(f'! Field {repr(args.key)} not found')
sys.exit(1)
# Note that field.types is a list of types. This is because the GGUF
# format supports arrays. For example, an array of UINT32 would
# look like [GGUFValueType.ARRAY, GGUFValueType.UINT32]
handler = reader.gguf_scalar_to_np.get(field.types[0]) if field.types else None
if handler is None:
logger.error(f'! This tool only supports changing simple values, {repr(args.key)} has unsupported type {field.types}')
sys.exit(1)
current_value = field.parts[field.data[0]][0]
new_value = handler(args.value)
logger.info(f'* Preparing to change field {repr(args.key)} from {current_value} to {new_value}')
if current_value == new_value:
logger.info(f'- Key {repr(args.key)} already set to requested value {current_value}')
sys.exit(0)
if args.dry_run:
sys.exit(0)
if not args.force:
logger.warning('*** Warning *** Warning *** Warning **')
logger.warning('* Changing fields in a GGUF file can make it unusable. Proceed at your own risk.')
logger.warning('* Enter exactly YES if you are positive you want to proceed:')
response = input('YES, I am sure> ')
if response != 'YES':
logger.info("You didn't enter YES. Okay then, see ya!")
sys.exit(0)
field.parts[field.data[0]][0] = new_value
logger.info('* Field changed. Successful completion.')
def main() -> None:
parser = argparse.ArgumentParser(description="Set a simple value in GGUF file metadata")
parser.add_argument("model", type=str, help="GGUF format model filename")
parser.add_argument("key", type=str, help="Metadata key to set")
parser.add_argument("value", type=str, help="Metadata value to set")
parser.add_argument("--dry-run", action="store_true", help="Don't actually change anything")
parser.add_argument("--force", action="store_true", help="Change the field without confirmation")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
logger.info(f'* Loading: {args.model}')
reader = GGUFReader(args.model, 'r' if args.dry_run else 'r+')
set_metadata(reader, args)
if __name__ == '__main__':
main()
llm/llama.cpp/gguf-py/tests/test_gguf.py deleted 100644 → 0
View file @ 97b02a89
import gguf # noqa: F401
# TODO: add tests
def test_write_gguf() -> None:
pass
llm/llama.cpp/kompute-shaders/common.comp deleted 100644 → 0
View file @ 97b02a89
#extension GL_EXT_shader_16bit_storage: require
#extension GL_EXT_shader_8bit_storage: require
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
#extension GL_EXT_control_flow_attributes: enable
#extension GL_KHR_shader_subgroup_arithmetic : require
#extension GL_EXT_debug_printf : enable
#define QK4_0 32
#define QK4_1 32
#define GELU_COEF_A 0.044715
#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
#define TWOPI_F 6.283185307179586f
#define QK_K 256
#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
#define sizeof_block_q4_0 0x12
struct block_q4_0 {
float16_t d;
uint8_t qs[QK4_0 / 2];
};
mat4 dequantize_q4_0(const block_q4_0 xb, uint il) {
const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
const float d2 = d1 / 256.f;
const float md = -8.f * xb.d;
const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
const uint16_t mask1 = mask0 << 8;
mat4 reg;
for (int i=0;i<8;i++) {
uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
reg[i/2][2*(i%2)+0] = d1 * (b & mask0) + md;
reg[i/2][2*(i%2)+1] = d2 * (b & mask1) + md;
}
return reg;
}
#define sizeof_block_q4_1 0x14
struct block_q4_1 {
float16_t d;
float16_t m;
uint8_t qs[QK4_1 / 2];
};
mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
const float d2 = d1 / 256.f;
const float m = xb.m;
const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
const uint16_t mask1 = mask0 << 8;
mat4 reg;
for (int i=0;i<8;i++) {
uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
reg[i/2][2*(i%2)+0] = ((b & mask0) * d1) + m;
reg[i/2][2*(i%2)+1] = ((b & mask1) * d2) + m;
}
return reg;
}
#define sizeof_block_q6_k 210
struct block_q6_k {
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
float16_t d; // super-block scale
};
mat4 dequantize_q6_k(const block_q6_k xb, uint il) {
const float16_t d_all = xb.d;
const uint qlIndex = 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
const uint qhIndex = 32*(il/8) + 16*(il&1);
float16_t sc = xb.scales[(il%2) + 2 * ((il/2))];
il = (il/2) & 3;
const uint16_t kmask1 = il>1 ? uint16_t(il>2 ? 192 : 48) : uint16_t(il>0 ? 12 : 3);
const uint16_t kmask2 = il>1 ? uint8_t(0xF0) : uint8_t(0x0F);
const float16_t coef = il>1 ? float16_t(1.f/16.f) : float16_t(1.f);
const float16_t ml = float16_t(d_all * sc * 32.f);
const float16_t dl = float16_t(d_all * sc * coef);
mat4 reg;
for (int i = 0; i < 16; ++i) {
const float16_t q = (il&1) != 0 ? ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 2))
: ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 4));
reg[i/4][i%4] = dl * q - ml;
}
return reg;
}
#define QK8_0 32
// struct block_q8_0 {
// float16_t d; // delta
// int8_t qs[QK8_0]; // quants
// };
#define sizeof_block_q8_0 34
llm/llama.cpp/kompute-shaders/op_add.comp deleted 100644 → 0
View file @ 97b02a89
#version 450
#include "common.comp"
layout(local_size_x = 1024) in;
layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb00;
int nb01;
int nb02;
int nb03;
int ne10;
int ne11;
int ne12;
int ne13;
int nb10;
int nb11;
int nb12;
int nb13;
int ne0;
int nb0;
int nb1;
int nb2;
int nb3;
//int offs; // TODO: needed for GGML_OP_ACC, see metal code
} pcs;
// general-purpose kernel for addition of two tensors
// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
// cons: not very efficient
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const uint i13 = i03 % pcs.ne13;
const uint i12 = i02 % pcs.ne12;
const uint i11 = i01 % pcs.ne11;
int offs = 0; // TMP (see above)
uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + offs) / 4);
uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11 ) / 4);
uint dst_off = uint((i03*pcs.nb3 + i02*pcs.nb2 + i01*pcs.nb1 + offs) / 4);
for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
const uint i10 = i0 % pcs.ne10;
out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] + inB[pcs.inBOff + src1_off + i10];
}
}
llm/llama.cpp/kompute-shaders/op_addrow.comp deleted 100644 → 0
View file @ 97b02a89
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inAOff;
uint inBOff;
uint outOff;
uint row;
} pcs;
void main() {
const uint baseIndex = gl_WorkGroupID.x * 4;
for (uint x = 0; x < 4; x++) {
const uint i = baseIndex + x;
out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i % pcs.row) + pcs.inBOff];
}
}
llm/llama.cpp/kompute-shaders/op_cpy_f16_f16.comp deleted 100644 → 0
View file @ 97b02a89
#version 450
#include "common.comp"
#define IN_TYPE float16_t
#define IN_TYPE_SIZE 2
#define OUT_TYPE float16_t
#define OUT_TYPE_SIZE 2
layout(local_size_x = 1024) in;
layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
layout (push_constant) uniform parameter {
uint inOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
int ne1;
int ne2;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
out_[dst_data+i00] = OUT_TYPE(in_[src]);
}
}
llm/llama.cpp/kompute-shaders/op_cpy_f16_f32.comp deleted 100644 → 0
View file @ 97b02a89
#version 450
#include "common.comp"
#define IN_TYPE float16_t
#define IN_TYPE_SIZE 2
#define OUT_TYPE float
#define OUT_TYPE_SIZE 4
layout(local_size_x = 1024) in;
layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
layout (push_constant) uniform parameter {
uint inOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
int ne1;
int ne2;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
out_[dst_data+i00] = OUT_TYPE(in_[src]);
}
}
llm/llama.cpp/kompute-shaders/op_cpy_f32_f16.comp deleted 100644 → 0
View file @ 97b02a89
#version 450
#include "common.comp"
#define IN_TYPE float
#define IN_TYPE_SIZE 4
#define OUT_TYPE float16_t
#define OUT_TYPE_SIZE 2
layout(local_size_x = 1024) in;
layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
layout (push_constant) uniform parameter {
uint inOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
int ne1;
int ne2;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
out_[dst_data+i00] = OUT_TYPE(in_[src]);
}
}
llm/llama.cpp/kompute-shaders/op_cpy_f32_f32.comp deleted 100644 → 0
View file @ 97b02a89
#version 450
#include "common.comp"
#define IN_TYPE float
#define IN_TYPE_SIZE 4
#define OUT_TYPE float
#define OUT_TYPE_SIZE 4
layout(local_size_x = 1024) in;
layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
layout (push_constant) uniform parameter {
uint inOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
int ne1;
int ne2;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
out_[dst_data+i00] = OUT_TYPE(in_[src]);
}
}
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