llama.cpp/ggml_vk_generate_shaders.py

2364 lines
89 KiB
Python
Raw Normal View History

ggml : add Vulkan backend (#2059) * Vulkan loader code * Fix matmul kernel, continue implementation * Continue implementation * Vulkan memory management * Vulkan development * Matmul call * Add aligned malloc and free for VMA * Continue implementation * First matmul success * GEMM Kernel optimization * 1D Blocktiling * 2D Blocktiling * Write coalescing * Continue vulkan implementation and optimization * First FP16 attempt, disabled for now * Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel * Enable device extensions properly, restore fp16 matmul op * Fix mulmat_f16 * Output FP32 in fp16 matmul shader * Fix f16_to_f32 kernel * dequant_q4_0 kernel * Add VMA library * Avoid requesting dedicated memory, VMA can decide that by itself * Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly * add cmake commands * Add 2d write operation, profiling code * Fix 2d write * Fix queue selection for AMD RADV * Fix trailing whitespace in vk_mem_alloc.h * Add WIP warp tile mat mul shaders * Disable glslc optimization * Disable glslc optimization for CMake * Optimize warptile matmul shader, replace blocktile with it * Add split-k optimization for small matrix multiplication Use semaphores for synchronization instead of fences or waitidle Rework async write/read for synchronization * Fix validation errors, improve compatibility with AMD GPUs * Rework command buffer handling * Variable matmul kernel using specialization constants * Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints * Reuse semaphores * Handle stage flags during command buffer submission properly * Increase matmul test runs for consistent results * Fix F32 matmul * Add vectorized loading and zeropadding for matrix multiplication * Use pinned memory for f16 preprocessing * Don't force aligned matmul * Don't free before queue done * Replace VMA library with native Vulkan buffer management * Basic offloading support with mul_f32 and dmmv for q4_0 * Run glslc commands in parallel * Unroll loops in dmmv shader * Reduce usage of waitIdle * Reuse pinned allocation for f16 conversion * Handle devices with only a single queue * Fix trailing whitespace in CMakeLists.txt * Allow parallel execution of kernels, parallelize third and fourth dimension calls * Add fallback for devices only supporting one DescriptorSet per DescriptorPool * Move to graph function similar to CUDA implementation * Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function * Add F32 dmmv shaders * Batch submissions * Add .spv to gitignore * Split off matrix vector multiplication for separate optimization * Use single command buffer for matrix vector multiplication ops * Reduce overhead of mul_f32 calls by using a single command buffer * Add submission batching to mul_f32 * Fix tests * Add missing barrier * Add further missing barrier * Add further ops * Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions * Remove unnecessary cblas link * Fix descriptor set pre-allocation assert * Add runtime shader compilation, start transferring shaders to this approach * Transfer remaining shaders to header and compile on runtime * Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16 * Add support for q4_1, q5_0, q5_1 and q8_0 * Remove unnecessary scalar layout extension * Parse graph early to pre-record command buffers * Add q6_k support * Add multi-submit for command buffers * Fix q6_k dequant shader for AMD * Fix q6_k for GPUs without fp16 support * Simplify q6_k fp16 fix * Minor fixes * Fix wg_denom of m-mulmat shaders * Add Python-based Vulkan shader generator * Replace shaderc dependency with precompiled shaders Fix python script to generate shaders * Clean up code * Fix shader generator script Windows compatibility Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com> * Close file before deletion * Fix vulkan shader fp32 name * Add q2_k and q3_k support Add validation check to compare shader results to cpu results * Add q4_k support * Add q5_k support * Bake SPIR-V bytecode into the library instead of loading shaders from file * Switch to signal semaphores for flexibility Prepare broadcasting support for mul mat * Finish broadcasting mul mat support for GQA * Clean up unused functions Add repeat op * Add further ops, not yet enabled. Improve semaphore code * Reduce number of used semaphores by utilizing timelines more properly * Remove queue information * Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations * Add Vulkan to llama-bench * Remove cblas dependency * Fix matmul k-split bug * Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader * Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug * Fix issues with float16 overflows in shaders * Fix issues with older Vulkan headers on Ubuntu 22.04 * Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers * Implement further ops, rework op_f32 calls, fix bugs * Finish full offloading support, add last remaining ops, fix bugs, remove redundant code * Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders * Merge upstream changes, fix conflicts, adapt soft_max op * Fix Python and shader header format * Free model gpu buffers on exit * Use single queue per device to simplify code * Add matmul shader support for running multiple calculations in parallel * Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible * Fix missing event cast * Replace uint64_t(-1) with UINT64_MAX, rename function for clarity * Fix warning about empty C function parameters * Fix compiler warnings * Properly implement Vulkan backend buffer handling * Fix oversized host staging buffers * Simplify barrier synchronization calls * Fix gcc warnings * Implement max_size for backend buffer types to limit the size of a single allocation * Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size * refactor multi buf * Disable unsupported ops to fix tests * Check for maintenance4 support before using it * Handle devices with only a single queue * Fix single queue logic * propagate buffer usage in multi buffers * Implement rope_neox op * Cleanup header and other files * Simplify gpu_extras by removing events and putting staging memcpys into contexts * Move queue into context Add not-yet-enabled async backend ops * Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization * Add get_max_size to SYCL backend. Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : fix trailing whitespace --------- Co-authored-by: Henri Vasserman <henv@hot.ee> Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 18:03:59 +01:00
#!/usr/bin/env python
import argparse
import asyncio
import os
import sys
from tempfile import gettempdir, NamedTemporaryFile
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_f16_defines = """
#define QUANT_K 1
#define QUANT_R 1
ggml : add Vulkan backend (#2059) * Vulkan loader code * Fix matmul kernel, continue implementation * Continue implementation * Vulkan memory management * Vulkan development * Matmul call * Add aligned malloc and free for VMA * Continue implementation * First matmul success * GEMM Kernel optimization * 1D Blocktiling * 2D Blocktiling * Write coalescing * Continue vulkan implementation and optimization * First FP16 attempt, disabled for now * Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel * Enable device extensions properly, restore fp16 matmul op * Fix mulmat_f16 * Output FP32 in fp16 matmul shader * Fix f16_to_f32 kernel * dequant_q4_0 kernel * Add VMA library * Avoid requesting dedicated memory, VMA can decide that by itself * Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly * add cmake commands * Add 2d write operation, profiling code * Fix 2d write * Fix queue selection for AMD RADV * Fix trailing whitespace in vk_mem_alloc.h * Add WIP warp tile mat mul shaders * Disable glslc optimization * Disable glslc optimization for CMake * Optimize warptile matmul shader, replace blocktile with it * Add split-k optimization for small matrix multiplication Use semaphores for synchronization instead of fences or waitidle Rework async write/read for synchronization * Fix validation errors, improve compatibility with AMD GPUs * Rework command buffer handling * Variable matmul kernel using specialization constants * Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints * Reuse semaphores * Handle stage flags during command buffer submission properly * Increase matmul test runs for consistent results * Fix F32 matmul * Add vectorized loading and zeropadding for matrix multiplication * Use pinned memory for f16 preprocessing * Don't force aligned matmul * Don't free before queue done * Replace VMA library with native Vulkan buffer management * Basic offloading support with mul_f32 and dmmv for q4_0 * Run glslc commands in parallel * Unroll loops in dmmv shader * Reduce usage of waitIdle * Reuse pinned allocation for f16 conversion * Handle devices with only a single queue * Fix trailing whitespace in CMakeLists.txt * Allow parallel execution of kernels, parallelize third and fourth dimension calls * Add fallback for devices only supporting one DescriptorSet per DescriptorPool * Move to graph function similar to CUDA implementation * Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function * Add F32 dmmv shaders * Batch submissions * Add .spv to gitignore * Split off matrix vector multiplication for separate optimization * Use single command buffer for matrix vector multiplication ops * Reduce overhead of mul_f32 calls by using a single command buffer * Add submission batching to mul_f32 * Fix tests * Add missing barrier * Add further missing barrier * Add further ops * Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions * Remove unnecessary cblas link * Fix descriptor set pre-allocation assert * Add runtime shader compilation, start transferring shaders to this approach * Transfer remaining shaders to header and compile on runtime * Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16 * Add support for q4_1, q5_0, q5_1 and q8_0 * Remove unnecessary scalar layout extension * Parse graph early to pre-record command buffers * Add q6_k support * Add multi-submit for command buffers * Fix q6_k dequant shader for AMD * Fix q6_k for GPUs without fp16 support * Simplify q6_k fp16 fix * Minor fixes * Fix wg_denom of m-mulmat shaders * Add Python-based Vulkan shader generator * Replace shaderc dependency with precompiled shaders Fix python script to generate shaders * Clean up code * Fix shader generator script Windows compatibility Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com> * Close file before deletion * Fix vulkan shader fp32 name * Add q2_k and q3_k support Add validation check to compare shader results to cpu results * Add q4_k support * Add q5_k support * Bake SPIR-V bytecode into the library instead of loading shaders from file * Switch to signal semaphores for flexibility Prepare broadcasting support for mul mat * Finish broadcasting mul mat support for GQA * Clean up unused functions Add repeat op * Add further ops, not yet enabled. Improve semaphore code * Reduce number of used semaphores by utilizing timelines more properly * Remove queue information * Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations * Add Vulkan to llama-bench * Remove cblas dependency * Fix matmul k-split bug * Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader * Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug * Fix issues with float16 overflows in shaders * Fix issues with older Vulkan headers on Ubuntu 22.04 * Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers * Implement further ops, rework op_f32 calls, fix bugs * Finish full offloading support, add last remaining ops, fix bugs, remove redundant code * Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders * Merge upstream changes, fix conflicts, adapt soft_max op * Fix Python and shader header format * Free model gpu buffers on exit * Use single queue per device to simplify code * Add matmul shader support for running multiple calculations in parallel * Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible * Fix missing event cast * Replace uint64_t(-1) with UINT64_MAX, rename function for clarity * Fix warning about empty C function parameters * Fix compiler warnings * Properly implement Vulkan backend buffer handling * Fix oversized host staging buffers * Simplify barrier synchronization calls * Fix gcc warnings * Implement max_size for backend buffer types to limit the size of a single allocation * Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size * refactor multi buf * Disable unsupported ops to fix tests * Check for maintenance4 support before using it * Handle devices with only a single queue * Fix single queue logic * propagate buffer usage in multi buffers * Implement rope_neox op * Cleanup header and other files * Simplify gpu_extras by removing events and putting staging memcpys into contexts * Move queue into context Add not-yet-enabled async backend ops * Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization * Add get_max_size to SYCL backend. Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : fix trailing whitespace --------- Co-authored-by: Henri Vasserman <henv@hot.ee> Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 18:03:59 +01:00
#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 = """
#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_f16_dequant_func = """
#define DEQUANT_FUNC f16vec2 v = f16vec2(data_a[ib + 0], data_a[ib + 1]);
"""
shader_f16_dequant_func_compat = """
#define DEQUANT_FUNC vec2 v = vec2(data_a[ib + 0], data_a[ib + 1]);
"""
shader_q4_0_dequant_func = """
#define DEQUANT_FUNC const float16_t d = data_a[ib].d; \
const uint8_t vui = data_a[ib].qs[iqs]; \
f16vec2 v = f16vec2(vui & 0xF, vui >> 4); \
v = (v - 8.0hf)*d;
"""
shader_q4_0_dequant_func_compat = """
#define DEQUANT_FUNC const float d = float(data_a[ib].d); \
const uint vui = uint(data_a[ib].qs[iqs]); \
vec2 v = vec2(vui & 0xF, vui >> 4); \
v = (v - 8.0f)*d;
"""
shader_q4_1_dequant_func = """
#define DEQUANT_FUNC const float16_t d = data_a[ib].d; \
const float16_t m = data_a[ib].m; \
const uint8_t vui = data_a[ib].qs[iqs]; \
f16vec2 v = f16vec2(vui & 0xF, vui >> 4); \
v = v*d + m;
"""
shader_q4_1_dequant_func_compat = """
#define DEQUANT_FUNC 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]); \
vec2 v = vec2(vui & 0xF, vui >> 4); \
v = v*d + m;
"""
shader_q5_0_dequant_func = """
#define DEQUANT_FUNC const float16_t d = 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 uint8_t vui = data_a[ib].qs[iqs]; \
f16vec2 v = f16vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y); \
v = (v - 16.0hf) * d;
"""
shader_q5_0_dequant_func_compat = """
#define DEQUANT_FUNC 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]); \
vec2 v = vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y); \
v = (v - 16.0f) * d;
"""
shader_q5_1_dequant_func = """
#define DEQUANT_FUNC const float16_t d = data_a[ib].d; \
const float16_t m = data_a[ib].m; \
const ivec2 qh = ivec2(((data_a[ib].qh >> iqs) << 4) & 0x10, (data_a[ib].qh >> (iqs + 12)) & 0x10); \
const uint8_t vui = data_a[ib].qs[iqs]; \
f16vec2 v = f16vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y); \
v = v*d + m;
"""
shader_q5_1_dequant_func_compat = """
#define DEQUANT_FUNC const float d = float(data_a[ib].d); \
const float m = float(data_a[ib].m); \
const ivec2 qh = ivec2(((data_a[ib].qh >> iqs) << 4) & 0x10, (data_a[ib].qh >> (iqs + 12)) & 0x10); \
const uint vui = uint(data_a[ib].qs[iqs]); \
vec2 v = vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y); \
v = v*d + m;
"""
shader_q8_0_dequant_func = """
#define DEQUANT_FUNC const float16_t d = data_a[ib].d; \
f16vec2 v = f16vec2(data_a[ib].qs[iqs], data_a[ib].qs[iqs + 1]); \
v = v * d;
"""
shader_q8_0_dequant_func_compat = """
#define DEQUANT_FUNC const float d = float(data_a[ib].d); \
vec2 v = vec2(int(data_a[ib].qs[iqs]), int(data_a[ib].qs[iqs + 1])); \
v = v * d;
"""
# MULMAT
mulmat_head = """#version 450
#extension GL_EXT_control_flow_attributes : enable
#extension GL_EXT_shader_16bit_storage : require
#ifndef LOAD_VEC
#define LOAD_VEC 1
#endif
"""
mulmat_body = """
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[];};
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;
} 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;
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)];
void main() {
const uint i13 = gl_GlobalInvocationID.z / p.ne12;
const uint i12 = gl_GlobalInvocationID.z % 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 = gl_LocalInvocationID.x % (BK / LOAD_VEC);
const uint loadc = gl_LocalInvocationID.x / (BK / LOAD_VEC);
const uint loadstride = gl_WorkGroupSize.x * LOAD_VEC / BK;
const uint start_k = ik * p.k_split;
const uint end_k = min(p.K, (ik + 1) * p.k_split);
uint pos_a = (batch_idx_a * p.batch_stride_a + ir * BM * p.stride_a + start_k) / LOAD_VEC;
uint pos_b = (gl_GlobalInvocationID.z * p.batch_stride_b + ic * BN * p.stride_b + start_k) / LOAD_VEC;
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) {
#if LOAD_VEC == 8
const uint idx = pos_a + (loadc + l) * p.stride_a / LOAD_VEC + loadr;
buf_a[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 0] = FLOAT_TYPE(data_a[idx][0].x);
buf_a[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 1] = FLOAT_TYPE(data_a[idx][0].y);
buf_a[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 2] = FLOAT_TYPE(data_a[idx][0].z);
buf_a[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 3] = FLOAT_TYPE(data_a[idx][0].w);
buf_a[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 4] = FLOAT_TYPE(data_a[idx][1].x);
buf_a[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 5] = FLOAT_TYPE(data_a[idx][1].y);
buf_a[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 6] = FLOAT_TYPE(data_a[idx][1].z);
buf_a[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 7] = FLOAT_TYPE(data_a[idx][1].w);
#elif LOAD_VEC == 4
const uint idx = pos_a + (loadc + l) * p.stride_a / LOAD_VEC + loadr;
buf_a[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 0] = FLOAT_TYPE(data_a[idx].x);
buf_a[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 1] = FLOAT_TYPE(data_a[idx].y);
buf_a[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 2] = FLOAT_TYPE(data_a[idx].z);
buf_a[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 3] = FLOAT_TYPE(data_a[idx].w);
#else
if (ir * BM + loadc + l < p.M && block + loadr < end_k) {
buf_a[(loadc + l) * (BK+1) + loadr] = FLOAT_TYPE(data_a[pos_a + (loadc + l) * p.stride_a + loadr]);
} else {
buf_a[(loadc + l) * (BK+1) + loadr] = FLOAT_TYPE(0.0f);
}
#endif
}
[[unroll]] for (uint l = 0; l < BN; l += loadstride) {
#if LOAD_VEC == 8
const uint idx = pos_b + (loadc + l) * p.stride_b / LOAD_VEC + loadr;
buf_b[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 0] = FLOAT_TYPE(data_b[idx][0].x);
buf_b[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 1] = FLOAT_TYPE(data_b[idx][0].y);
buf_b[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 2] = FLOAT_TYPE(data_b[idx][0].z);
buf_b[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 3] = FLOAT_TYPE(data_b[idx][0].w);
buf_b[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 4] = FLOAT_TYPE(data_b[idx][1].x);
buf_b[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 5] = FLOAT_TYPE(data_b[idx][1].y);
buf_b[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 6] = FLOAT_TYPE(data_b[idx][1].z);
buf_b[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 7] = FLOAT_TYPE(data_b[idx][1].w);
#elif LOAD_VEC == 4
const uint idx = pos_b + (loadc + l) * p.stride_b / LOAD_VEC + loadr;
buf_b[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 0] = FLOAT_TYPE(data_b[idx].x);
buf_b[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 1] = FLOAT_TYPE(data_b[idx].y);
buf_b[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 2] = FLOAT_TYPE(data_b[idx].z);
buf_b[(loadc + l) * (BK+1) + loadr * LOAD_VEC + 3] = FLOAT_TYPE(data_b[idx].w);
#else
if (ic * BN + loadc + l < p.N && block + loadr < end_k) {
buf_b[(loadc + l) * (BK+1) + loadr] = FLOAT_TYPE(data_b[pos_b + (loadc + l) * p.stride_b + loadr]);
} else {
buf_b[(loadc + l) * (BK+1) + loadr] = FLOAT_TYPE(0.0f);
}
#endif
}
barrier();
pos_a += BK / LOAD_VEC;
pos_b += BK / LOAD_VEC;
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 = gl_GlobalInvocationID.z * 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
"""
dequant_body = """
layout(local_size_x = 256, 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[];};
layout (push_constant) uniform parameter
{
int M;
int K;
int stride_a;
int stride_b;
} p;
void main() {
const int i = int(gl_GlobalInvocationID.x);
// Transposed
const int row = i % (p.K / QUANT_K);
const int col = i / (p.K / QUANT_K);
if (row * QUANT_K >= p.K || col >= p.M) {
return;
}
const int stride_a = p.stride_a / QUANT_K;
const int ib = col * stride_a + row;
const int y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
const int step = QUANT_R == 1 ? 2 : 1;
[[unroll]] for (int iqs = 0; iqs < QUANT_K/QUANT_R; iqs += step) {
DEQUANT_FUNC
data_b[col * p.stride_b + row*QUANT_K + iqs + 0 ] = D_TYPE(v.x);
data_b[col * p.stride_b + row*QUANT_K + iqs + y_offset] = D_TYPE(v.y);
}
}
"""
# 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[];};
layout (push_constant) uniform parameter
{
int M;
int K;
int stride_a;
int stride_b;
} p;
void main() {
[[unroll]] for (int wgy = 0; wgy < 256; wgy++) {
const int i = int(gl_WorkGroupID.x * 256 + wgy);
if (i >= p.M * p.K / QUANT_K) {
return;
}
const int tid = int(gl_LocalInvocationID.x);
const int ip = tid / 32;
const int il = tid - 32 * ip;
const int is = 8 * ip + il / 16;
const int y_idx = i * QUANT_K + 128 * ip + il;
const int 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[];};
layout (push_constant) uniform parameter
{
int M;
int K;
int stride_a;
int stride_b;
} p;
void main() {
[[unroll]] for (int wgy = 0; wgy < 256; wgy++) {
const int i = int(gl_WorkGroupID.x * 256 + wgy);
if (i >= p.M * p.K / QUANT_K) {
return;
}
const int r = int(gl_LocalInvocationID.x) / 4;
const int tid = r / 2;
const int is0 = r % 2;
const int l0 = 16 * is0 + 4 * (int(gl_LocalInvocationID.x) % 4);
const int n = tid / 4;
const int j = tid - 4*n;
const uint8_t m = uint8_t(1 << (4*n + j));
const int is = 8*n + 2*j + is0;
const int 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 int y_idx = i * QUANT_K + 128 * n + 32 * j;
const int qs_idx = 32*n;
for (int 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[];};
layout (push_constant) uniform parameter
{
int M;
int K;
int stride_a;
int stride_b;
} p;
void main() {
[[unroll]] for (int wgy = 0; wgy < 256; wgy++) {
const int i = int(gl_WorkGroupID.x * 256 + wgy);
if (i >= p.M * p.K / QUANT_K) {
return;
}
const int tid = int(gl_LocalInvocationID.x);
const int il = tid / 8;
const int ir = tid % 8;
const int is = 2 * il;
const int 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 int y_idx = i * QUANT_K + 64 * il + n * ir;
const int 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 (int 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[];};
layout (push_constant) uniform parameter
{
int M;
int K;
int stride_a;
int stride_b;
} p;
void main() {
[[unroll]] for (int wgy = 0; wgy < 256; wgy++) {
const int i = int(gl_WorkGroupID.x * 256 + wgy);
if (i >= p.M * p.K / QUANT_K) {
return;
}
const int tid = int(gl_LocalInvocationID.x);
const int il = tid / 16;
const int ir = tid % 16;
const int 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 int y_idx = i * QUANT_K + 64 * il + 2 * ir;
const int qs_idx = 32*il + 2 * ir;
const int 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[];};
layout (push_constant) uniform parameter
{
int M;
int K;
int stride_a;
int stride_b;
} p;
void main() {
[[unroll]] for (int wgy = 0; wgy < 256; wgy++) {
const int i = int(gl_WorkGroupID.x * 256 + wgy);
if (i >= p.M * p.K / QUANT_K) {
return;
}
const int tid = int(gl_LocalInvocationID.x);
const int ip = tid / 32;
const int il = tid - 32 * ip;
const int is = 8 * ip + il / 16;
const int y_idx = i * QUANT_K + 128 * ip + il;
const int 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
"""
mul_mat_vec_body = """
layout(local_size_x = QUANT_K, 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
{
int ncols;
int b_offset;
int d_offset;
} p;
shared FLOAT_TYPE tmp[QUANT_K];
void main() {
const int block_size = int(gl_WorkGroupSize.x);
const int row = int(gl_WorkGroupID.x);
const int tid = int(gl_LocalInvocationID.x);
const int y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
tmp[tid] = FLOAT_TYPE(0.0f);
[[unroll]] for (int i = 0; i < p.ncols/block_size; i += 2) {
const int col = i*block_size + 2*tid;
const int ib = (row*p.ncols + col)/QUANT_K; // block index
const int iqs = (col%QUANT_K)/QUANT_R; // quant index
const int iybs = col - col%QUANT_K; // y block start index
DEQUANT_FUNC
// matrix multiplication
tmp[tid] += FLOAT_TYPE(v.x) * FLOAT_TYPE(data_b[p.b_offset + iybs + iqs + 0]);
tmp[tid] += FLOAT_TYPE(v.y) * FLOAT_TYPE(data_b[p.b_offset + iybs + iqs + y_offset]);
}
// 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[p.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;
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
{
int ncols;
int b_offset;
int d_offset;
} p;
shared FLOAT_TYPE tmp[32];
void main() {
const int row = int(gl_WorkGroupID.x);
const int num_blocks_per_row = p.ncols / QUANT_K;
const int ib0 = row*num_blocks_per_row;
const int tid = int(gl_LocalInvocationID.x)/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = int(gl_LocalInvocationID.x)%K_QUANTS_PER_ITERATION; // 0 or 0, 1
const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8
const int v_im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int v_in = tid - step*v_im; // 0...15 or 0...7
const int l0 = K_QUANTS_PER_ITERATION*v_in; // 0...15
const int q_offset = 32*v_im + l0;
const int s_offset = 8*v_im;
const int y_offset = 128*v_im + l0;
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
[[unroll]] for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const int 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[p.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[p.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[p.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[p.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[p.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[p.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[p.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[p.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[p.b_offset + y_idx + l + 0]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[p.b_offset + y_idx + l + 16]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[p.b_offset + y_idx + l + 32]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[p.b_offset + y_idx + l + 48]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[p.b_offset + y_idx + l + 64]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[p.b_offset + y_idx + l + 80]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[p.b_offset + y_idx + l + 96]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF)
+ FLOAT_TYPE(data_b[p.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 (int s = 16; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
dst[p.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;
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
{
int ncols;
int b_offset;
int d_offset;
} p;
shared FLOAT_TYPE tmp[32];
void main() {
const int row = int(gl_WorkGroupID.x);
const int num_blocks_per_row = p.ncols / QUANT_K;
const int ib0 = row*num_blocks_per_row;
const int tid = int(gl_LocalInvocationID.x)/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = int(gl_LocalInvocationID.x)%K_QUANTS_PER_ITERATION; // 0 or 0, 1
const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8
const int v_im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int v_in = tid - step*v_im; // 0...15 or 0...7
const uint8_t m = uint8_t(1 << (4 * v_im));
const int l0 = K_QUANTS_PER_ITERATION*v_in; // 0...15
const int q_offset = 32*v_im + l0;
const int 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 (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const int 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[p.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[p.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[p.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[p.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[p.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[p.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[p.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[p.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 (int s = 16; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
dst[p.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;
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
{
int ncols;
int b_offset;
int d_offset;
} p;
shared FLOAT_TYPE tmp[32];
void main() {
const int row = int(gl_WorkGroupID.x);
const int num_blocks_per_row = p.ncols / QUANT_K;
const int ib0 = row*num_blocks_per_row;
const int tid = int(gl_LocalInvocationID.x)/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = int(gl_LocalInvocationID.x)%K_QUANTS_PER_ITERATION; // 0 or 0, 1
const int step = 8/K_QUANTS_PER_ITERATION; // 8 or 4
const int il = tid/step; // 0...3
const int ir = tid - step*il; // 0...7 or 0...3
const int n = 2 * K_QUANTS_PER_ITERATION; // 2 or 4
const int v_im = il / 2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const int v_in = il % 2;
const int l0 = n * (2 * ir + v_in); // 0...15
const int q_offset = 32*v_im + l0;
const int y_offset = 64*v_im + l0;
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
[[unroll]] for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const int y1_idx = i * QUANT_K + y_offset;
const int 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(data_b[p.b_offset + y1_idx] * q4_0 + data_b[p.b_offset + y1_idx + 1] * q4_1 + data_b[p.b_offset + y1_idx + 2] * q4_2 + data_b[p.b_offset + y1_idx + 3] * q4_3);
const FLOAT_TYPE sy = FLOAT_TYPE(data_b[p.b_offset + y1_idx + 32] * q4_4 + data_b[p.b_offset + y1_idx + 33] * q4_5 + data_b[p.b_offset + y1_idx + 34] * q4_6 + data_b[p.b_offset + y1_idx + 35] * q4_7);
const FLOAT_TYPE sz = FLOAT_TYPE(data_b[p.b_offset + y2_idx] * q4_8 + data_b[p.b_offset + y2_idx + 1] * q4_9 + data_b[p.b_offset + y2_idx + 2] * q4_10 + data_b[p.b_offset + y2_idx + 3] * q4_11);
const FLOAT_TYPE sw = FLOAT_TYPE(data_b[p.b_offset + y2_idx + 32] * q4_12 + data_b[p.b_offset + y2_idx + 33] * q4_13 + data_b[p.b_offset + y2_idx + 34] * q4_14 + data_b[p.b_offset + y2_idx + 35] * q4_15);
const FLOAT_TYPE smin = FLOAT_TYPE(
data_b[p.b_offset + y1_idx ] * sc2 + data_b[p.b_offset + y1_idx + 32] * sc3 + data_b[p.b_offset + y2_idx ] * sc6 + data_b[p.b_offset + y2_idx + 32] * sc7
+ data_b[p.b_offset + y1_idx + 1] * sc2 + data_b[p.b_offset + y1_idx + 33] * sc3 + data_b[p.b_offset + y2_idx + 1] * sc6 + data_b[p.b_offset + y2_idx + 33] * sc7
+ data_b[p.b_offset + y1_idx + 2] * sc2 + data_b[p.b_offset + y1_idx + 34] * sc3 + data_b[p.b_offset + y2_idx + 2] * sc6 + data_b[p.b_offset + y2_idx + 34] * sc7
+ data_b[p.b_offset + y1_idx + 3] * sc2 + data_b[p.b_offset + y1_idx + 35] * sc3 + data_b[p.b_offset + y2_idx + 3] * sc6 + data_b[p.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(data_b[p.b_offset + y1_idx ] * q4_0 + data_b[p.b_offset + y1_idx + 1] * q4_1);
const FLOAT_TYPE sy = FLOAT_TYPE(data_b[p.b_offset + y1_idx + 32] * q4_2 + data_b[p.b_offset + y1_idx + 33] * q4_3);
const FLOAT_TYPE sz = FLOAT_TYPE(data_b[p.b_offset + y2_idx ] * q4_4 + data_b[p.b_offset + y2_idx + 1] * q4_5);
const FLOAT_TYPE sw = FLOAT_TYPE(data_b[p.b_offset + y2_idx + 32] * q4_6 + data_b[p.b_offset + y2_idx + 33] * q4_7);
const FLOAT_TYPE smin = FLOAT_TYPE(
data_b[p.b_offset + y1_idx] * sc2 + data_b[p.b_offset + y1_idx + 32] * sc3 + data_b[p.b_offset + y2_idx] * sc6 + data_b[p.b_offset + y2_idx + 32] * sc7
+ data_b[p.b_offset + y1_idx + 1] * sc2 + data_b[p.b_offset + y1_idx + 33] * sc3 + data_b[p.b_offset + y2_idx + 1] * sc6 + data_b[p.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 (int s = 16; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
dst[p.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;
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
{
int ncols;
int b_offset;
int d_offset;
} p;
shared FLOAT_TYPE tmp[32];
void main() {
const int row = int(gl_WorkGroupID.x);
const int num_blocks_per_row = p.ncols / QUANT_K;
const int ib0 = row*num_blocks_per_row;
const int tid = int(gl_LocalInvocationID.x)/2; // 0...31 or 0...16
const int ix = int(gl_LocalInvocationID.x)%2; // 0 or 0, 1
const int il = tid/4; // 0...3
const int ir = tid - 4*il; // 0...7 or 0...3
const int v_im = il / 2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const int v_in = il % 2;
const int l0 = 4*ir + 2*v_in; // 0...15
const int q_offset = 32*v_im + l0;
const int 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 (int i = ix; i < num_blocks_per_row; i += 2) {
const int y1_idx = i * QUANT_K + y_offset;
const int 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(
data_b[p.b_offset + y1_idx ] * (q4_0 + (((data_a[ib0 + i].qh[l0 ] & hm1) != 0) ? 16 : 0))
+ data_b[p.b_offset + y1_idx + 1] * (q4_1 + (((data_a[ib0 + i].qh[l0 + 1] & hm1) != 0) ? 16 : 0))
+ data_b[p.b_offset + y1_idx + 16] * (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0))
+ data_b[p.b_offset + y1_idx + 17] * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0))
);
const FLOAT_TYPE sy = FLOAT_TYPE(
data_b[p.b_offset + y1_idx + 32] * (q4_4 + (((data_a[ib0 + i].qh[l0 ] & (hm1 << 1)) != 0) ? 16 : 0))
+ data_b[p.b_offset + y1_idx + 33] * (q4_5 + (((data_a[ib0 + i].qh[l0 + 1] & (hm1 << 1)) != 0) ? 16 : 0))
+ data_b[p.b_offset + y1_idx + 48] * (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0))
+ data_b[p.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(
data_b[p.b_offset + y2_idx ] * (q4_8 + (((data_a[ib0 + i].qh[l0 ] & hm2) != 0) ? 16 : 0))
+ data_b[p.b_offset + y2_idx + 1] * (q4_9 + (((data_a[ib0 + i].qh[l0 + 1] & hm2) != 0) ? 16 : 0))
+ data_b[p.b_offset + y2_idx + 16] * (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0))
+ data_b[p.b_offset + y2_idx + 17] * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0))
);
const FLOAT_TYPE sw = FLOAT_TYPE(
data_b[p.b_offset + y2_idx + 32] * (q4_12 + (((data_a[ib0 + i].qh[l0 ] & (hm2 << 1)) != 0) ? 16 : 0))
+ data_b[p.b_offset + y2_idx + 33] * (q4_13 + (((data_a[ib0 + i].qh[l0 + 1] & (hm2 << 1)) != 0) ? 16 : 0))
+ data_b[p.b_offset + y2_idx + 48] * (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0))
+ data_b[p.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(
(data_b[p.b_offset + y1_idx] + data_b[p.b_offset + y1_idx + 1] + data_b[p.b_offset + y1_idx + 16] + data_b[p.b_offset + y1_idx + 17]) * sc2 + (data_b[p.b_offset + y1_idx + 32] + data_b[p.b_offset + y1_idx + 33] + data_b[p.b_offset + y1_idx + 48] + data_b[p.b_offset + y1_idx + 49]) * sc3
+ (data_b[p.b_offset + y2_idx] + data_b[p.b_offset + y2_idx + 1] + data_b[p.b_offset + y2_idx + 16] + data_b[p.b_offset + y2_idx + 17]) * sc6 + (data_b[p.b_offset + y2_idx + 32] + data_b[p.b_offset + y2_idx + 33] + data_b[p.b_offset + y2_idx + 48] + data_b[p.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 (int s = 16; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
dst[p.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;
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
{
int ncols;
int b_offset;
int d_offset;
} p;
shared FLOAT_TYPE tmp[32];
void main() {
const int row = int(gl_WorkGroupID.x);
const int num_blocks_per_row = p.ncols / QUANT_K;
const int ib0 = row*num_blocks_per_row;
const int tid = int(gl_LocalInvocationID.x)/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = int(gl_LocalInvocationID.x)%K_QUANTS_PER_ITERATION; // 0 or 0, 1
const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8
const int v_im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int v_in = tid - step*v_im; // 0...15 or 0...7
#if K_QUANTS_PER_ITERATION == 1
const int l0 = K_QUANTS_PER_ITERATION*v_in; // 0...15
const int is = 0;
#else
const int l0 = 4 * v_in; // 0, 4, 8, ..., 28
const int is = v_in / 4;
#endif
const int ql_offset = 64*v_im + l0;
const int qh_offset = 32*v_im + l0;
const int s_offset = 8*v_im + is;
const int y_offset = 128*v_im + l0;
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
[[unroll]] for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const int 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[p.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[p.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[p.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[p.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[p.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[p.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[p.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[p.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[p.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[p.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[p.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[p.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 (int s = 16; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
dst[p.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];
}
}
"""
# F16 to F32
f32_to_f16_src = """#version 450
#extension GL_EXT_shader_16bit_storage : require
layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {float data_a[];};
layout (binding = 1) writeonly buffer D {float16_t data_b[];};
layout (push_constant) uniform parameter
{
int M;
int K;
int stride_a;
int stride_b;
} p;
void main() {
const int row = int(gl_GlobalInvocationID.x % p.K);
const int col = int(gl_GlobalInvocationID.x / p.K);
if (row < p.K && col < p.M) {
data_b[col * p.stride_b + row] = float16_t(data_a[col * p.stride_a + row]);
}
}
"""
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;
"""
# MUL F32
mul_body = """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) readonly buffer Y {B_TYPE data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint idx = gl_GlobalInvocationID.x;
if (idx >= p.KX) {
return;
}
data_d[idx] = D_TYPE(FLOAT_TYPE(data_a[idx]) * FLOAT_TYPE(data_b[idx % p.KY]));
}
"""
# ADD
add_body = """
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) readonly buffer Y {B_TYPE data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
void main() {
const uint idx = gl_GlobalInvocationID.x;
if (idx >= p.KX) {
return;
}
data_d[idx] = D_TYPE(FLOAT_TYPE(data_a[idx]) + FLOAT_TYPE(data_b[idx % p.KY]));
}
"""
# SCALE
scale_body = """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 idx = gl_GlobalInvocationID.x;
if (idx >= p.KX) {
return;
}
data_d[idx] = D_TYPE(FLOAT_TYPE(data_a[idx]) * FLOAT_TYPE(p.param1));
}
"""
# SQR
sqr_body = """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 idx = gl_GlobalInvocationID.x;
if (idx >= p.KX) {
return;
}
const FLOAT_TYPE val = FLOAT_TYPE(data_a[idx]);
data_d[idx] = D_TYPE(val * val);
}
"""
# CLAMP
clamp_body = """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 idx = gl_GlobalInvocationID.x;
if (idx >= p.KX) {
return;
}
const FLOAT_TYPE val = FLOAT_TYPE(data_a[idx]);
data_d[idx] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
}
"""
# CPY
cpy_src = """#version 450
#extension GL_EXT_shader_16bit_storage : require
layout (push_constant) uniform parameter
{
uint ne;
uint ne00; uint ne01; uint nb00; uint nb01; uint nb02;
uint ne10; uint ne11; uint nb10; uint nb11; uint nb12;
uint d_offset;
} p;
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[];};
void main() {
if (gl_GlobalInvocationID.x >= p.ne) {
return;
}
const uint i02 = gl_GlobalInvocationID.x / (p.ne00*p.ne01);
const uint i01 = (gl_GlobalInvocationID.x - i02*p.ne01*p.ne00) / p.ne00;
const uint i00 = gl_GlobalInvocationID.x - i02*p.ne01*p.ne00 - i01*p.ne00;
const uint a_idx = i00*p.nb00 + i01*p.nb01 + i02*p.nb02;
const uint i12 = gl_GlobalInvocationID.x / (p.ne10*p.ne11);
const uint i11 = (gl_GlobalInvocationID.x - i12*p.ne11*p.ne10) / p.ne10;
const uint i10 = gl_GlobalInvocationID.x - i12*p.ne11*p.ne10 - i11*p.ne10;
const uint d_idx = i10*p.nb10 + i11*p.nb11 + i12*p.nb12;
"""
cpy_end = """
data_d[p.d_offset + d_idx] = D_TYPE(data_a[a_idx]);
}
"""
# Causes an optimization error otherwise
cpy_f16_f16_end = """
data_d[p.d_offset + d_idx] = data_a[a_idx];
}
"""
# GET_ROWS
get_rows_body = """
#extension GL_EXT_control_flow_attributes : enable
#extension GL_EXT_shader_8bit_storage : require
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) readonly buffer Y {int data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
void main() {
const uint col = int(gl_GlobalInvocationID.x) * 2;
const uint row = int(gl_GlobalInvocationID.y);
if (col >= p.KY) {
return;
}
const uint r = uint(data_b[row]);
// copy data_a[r*p.KY + col] to dst[row*p.KX + col]
const uint xi = r*p.KY + col;
const uint di = row*p.KY + col;
const uint ib = xi/QUANT_K; // block index
const uint iqs = (xi%QUANT_K)/QUANT_R; // quant index
const uint iybs = di - di%QUANT_K; // y block start index
const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
DEQUANT_FUNC
dst[iybs + iqs + 0] = D_TYPE(v.x);
dst[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)));
ggml : add Vulkan backend (#2059) * Vulkan loader code * Fix matmul kernel, continue implementation * Continue implementation * Vulkan memory management * Vulkan development * Matmul call * Add aligned malloc and free for VMA * Continue implementation * First matmul success * GEMM Kernel optimization * 1D Blocktiling * 2D Blocktiling * Write coalescing * Continue vulkan implementation and optimization * First FP16 attempt, disabled for now * Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel * Enable device extensions properly, restore fp16 matmul op * Fix mulmat_f16 * Output FP32 in fp16 matmul shader * Fix f16_to_f32 kernel * dequant_q4_0 kernel * Add VMA library * Avoid requesting dedicated memory, VMA can decide that by itself * Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly * add cmake commands * Add 2d write operation, profiling code * Fix 2d write * Fix queue selection for AMD RADV * Fix trailing whitespace in vk_mem_alloc.h * Add WIP warp tile mat mul shaders * Disable glslc optimization * Disable glslc optimization for CMake * Optimize warptile matmul shader, replace blocktile with it * Add split-k optimization for small matrix multiplication Use semaphores for synchronization instead of fences or waitidle Rework async write/read for synchronization * Fix validation errors, improve compatibility with AMD GPUs * Rework command buffer handling * Variable matmul kernel using specialization constants * Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints * Reuse semaphores * Handle stage flags during command buffer submission properly * Increase matmul test runs for consistent results * Fix F32 matmul * Add vectorized loading and zeropadding for matrix multiplication * Use pinned memory for f16 preprocessing * Don't force aligned matmul * Don't free before queue done * Replace VMA library with native Vulkan buffer management * Basic offloading support with mul_f32 and dmmv for q4_0 * Run glslc commands in parallel * Unroll loops in dmmv shader * Reduce usage of waitIdle * Reuse pinned allocation for f16 conversion * Handle devices with only a single queue * Fix trailing whitespace in CMakeLists.txt * Allow parallel execution of kernels, parallelize third and fourth dimension calls * Add fallback for devices only supporting one DescriptorSet per DescriptorPool * Move to graph function similar to CUDA implementation * Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function * Add F32 dmmv shaders * Batch submissions * Add .spv to gitignore * Split off matrix vector multiplication for separate optimization * Use single command buffer for matrix vector multiplication ops * Reduce overhead of mul_f32 calls by using a single command buffer * Add submission batching to mul_f32 * Fix tests * Add missing barrier * Add further missing barrier * Add further ops * Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions * Remove unnecessary cblas link * Fix descriptor set pre-allocation assert * Add runtime shader compilation, start transferring shaders to this approach * Transfer remaining shaders to header and compile on runtime * Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16 * Add support for q4_1, q5_0, q5_1 and q8_0 * Remove unnecessary scalar layout extension * Parse graph early to pre-record command buffers * Add q6_k support * Add multi-submit for command buffers * Fix q6_k dequant shader for AMD * Fix q6_k for GPUs without fp16 support * Simplify q6_k fp16 fix * Minor fixes * Fix wg_denom of m-mulmat shaders * Add Python-based Vulkan shader generator * Replace shaderc dependency with precompiled shaders Fix python script to generate shaders * Clean up code * Fix shader generator script Windows compatibility Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com> * Close file before deletion * Fix vulkan shader fp32 name * Add q2_k and q3_k support Add validation check to compare shader results to cpu results * Add q4_k support * Add q5_k support * Bake SPIR-V bytecode into the library instead of loading shaders from file * Switch to signal semaphores for flexibility Prepare broadcasting support for mul mat * Finish broadcasting mul mat support for GQA * Clean up unused functions Add repeat op * Add further ops, not yet enabled. Improve semaphore code * Reduce number of used semaphores by utilizing timelines more properly * Remove queue information * Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations * Add Vulkan to llama-bench * Remove cblas dependency * Fix matmul k-split bug * Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader * Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug * Fix issues with float16 overflows in shaders * Fix issues with older Vulkan headers on Ubuntu 22.04 * Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers * Implement further ops, rework op_f32 calls, fix bugs * Finish full offloading support, add last remaining ops, fix bugs, remove redundant code * Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders * Merge upstream changes, fix conflicts, adapt soft_max op * Fix Python and shader header format * Free model gpu buffers on exit * Use single queue per device to simplify code * Add matmul shader support for running multiple calculations in parallel * Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible * Fix missing event cast * Replace uint64_t(-1) with UINT64_MAX, rename function for clarity * Fix warning about empty C function parameters * Fix compiler warnings * Properly implement Vulkan backend buffer handling * Fix oversized host staging buffers * Simplify barrier synchronization calls * Fix gcc warnings * Implement max_size for backend buffer types to limit the size of a single allocation * Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size * refactor multi buf * Disable unsupported ops to fix tests * Check for maintenance4 support before using it * Handle devices with only a single queue * Fix single queue logic * propagate buffer usage in multi buffers * Implement rope_neox op * Cleanup header and other files * Simplify gpu_extras by removing events and putting staging memcpys into contexts * Move queue into context Add not-yet-enabled async backend ops * Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization * Add get_max_size to SYCL backend. Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : fix trailing whitespace --------- Co-authored-by: Henri Vasserman <henv@hot.ee> Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 18:03:59 +01:00
}
"""
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;
data_d[i] = D_TYPE(data_a[i] - float(uint(col > p.n_past + row % p.rows_per_channel) * 0xFFFFFFFF));
}
"""
# 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;
const float eps = 1e-5f;
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 + 1e-5f);
[[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_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;
// Find max
vals[tid] = uintBitsToFloat(0xFF800000);
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
vals[tid] = max(vals[tid], FLOAT_TYPE(data_a[rowx * p.KX + col]) * p.param1 + (p.KY > 0 ? FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)));
}
barrier();
[[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
if (tid < s) {
vals[tid] = max(vals[tid], vals[tid + s]);
}
barrier();
}
const FLOAT_TYPE max_val = vals[0];
barrier();
// Sum up values
vals[tid] = FLOAT_TYPE(0.0f);
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
const uint i = rowx * p.KX + col;
const FLOAT_TYPE val = exp(FLOAT_TYPE(data_a[i]) * p.param1 + (p.KY > 0 ? 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 col = tid; col < p.KX; col += BLOCK_SIZE) {
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) 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;
} 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 float cur_rot = p.inv_ndims * ic - ib;
const int pos = data_b[i2];
const float theta_base = pos*p.freq_scale*pow(p.theta_scale, col/2.0f);
float cos_theta, sin_theta;
rope_yarn(theta_base, uint(cur_rot), 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);
}
"""
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
output_dir = gettempdir()
lock = asyncio.Lock()
shader_fnames = []
async def string_to_spv(name, code, defines, fp16):
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()
print(" ".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())])
print(f"ERROR compiling {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():
print("ggml_vulkan: Generating and compiling shaders to SPIR-V")
tasks = []
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 = []
stream.extend((mulmat_head, shader_float_type, mulmat_body))
tasks.append(string_to_spv("matmul_f32_l", "".join(stream), {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f32_m", "".join(stream), {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f32_s", "".join(stream), {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f32_aligned_l", "".join(stream), {"LOAD_VEC": load_vec, "A_TYPE": vec_type, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f32_aligned_m", "".join(stream), {"LOAD_VEC": load_vec, "A_TYPE": vec_type, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f32_aligned_s", "".join(stream), {"LOAD_VEC": load_vec, "A_TYPE": vec_type, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_l", "".join(stream), {"A_TYPE": "float16_t", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_m", "".join(stream), {"A_TYPE": "float16_t", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_s", "".join(stream), {"A_TYPE": "float16_t", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_aligned_l", "".join(stream), {"LOAD_VEC": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type_f16, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_aligned_m", "".join(stream), {"LOAD_VEC": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type_f16, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_aligned_s", "".join(stream), {"LOAD_VEC": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type_f16, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_f32_l", "".join(stream), {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_f32_m", "".join(stream), {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_f32_s", "".join(stream), {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_f32_aligned_l", "".join(stream), {"LOAD_VEC": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_f32_aligned_m", "".join(stream), {"LOAD_VEC": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_f16_f32_aligned_s", "".join(stream), {"LOAD_VEC": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# Build dequant shaders
tasks.append(string_to_spv("f32_to_f16", f32_to_f16_src, {}, fp16))
for i in range(0, VK_NUM_TYPES):
stream.clear()
stream.extend((dequant_head, shader_int8_ext, shader_float_type))
if i == GGML_TYPE_F16:
stream.extend((shader_f16_defines, shader_f16_dequant_func_compat if not fp16 else shader_f16_dequant_func, dequant_body))
elif i == GGML_TYPE_Q4_0:
stream.extend((shader_q4_0_defines, shader_q4_0_dequant_func_compat if not fp16 else shader_q4_0_dequant_func, dequant_body))
elif i == GGML_TYPE_Q4_1:
stream.extend((shader_q4_1_defines, shader_q4_1_dequant_func_compat if not fp16 else shader_q4_1_dequant_func, dequant_body))
elif i == GGML_TYPE_Q5_0:
stream.extend((shader_q5_0_defines, shader_q5_0_dequant_func_compat if not fp16 else shader_q5_0_dequant_func, dequant_body))
elif i == GGML_TYPE_Q5_1:
stream.extend((shader_q5_1_defines, shader_q5_1_dequant_func_compat if not fp16 else shader_q5_1_dequant_func, dequant_body))
elif i == GGML_TYPE_Q8_0:
stream.extend((shader_q8_0_defines, shader_q8_0_dequant_func_compat if not fp16 else shader_q8_0_dequant_func, dequant_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"}, fp16))
# get_rows
for i in range(0, VK_NUM_TYPES):
stream.clear()
stream.extend((generic_head, shader_int8_ext, shader_float_type))
if i == GGML_TYPE_F16:
stream.extend((shader_f16_defines, shader_f16_dequant_func_compat if not fp16 else shader_f16_dequant_func, get_rows_body))
elif i == GGML_TYPE_Q4_0:
stream.extend((shader_q4_0_defines, shader_q4_0_dequant_func_compat if not fp16 else shader_q4_0_dequant_func, get_rows_body))
elif i == GGML_TYPE_Q4_1:
stream.extend((shader_q4_1_defines, shader_q4_1_dequant_func_compat if not fp16 else shader_q4_1_dequant_func, get_rows_body))
elif i == GGML_TYPE_Q5_0:
stream.extend((shader_q5_0_defines, shader_q5_0_dequant_func_compat if not fp16 else shader_q5_0_dequant_func, get_rows_body))
elif i == GGML_TYPE_Q5_1:
stream.extend((shader_q5_1_defines, shader_q5_1_dequant_func_compat if not fp16 else shader_q5_1_dequant_func, get_rows_body))
elif i == GGML_TYPE_Q8_0:
stream.extend((shader_q8_0_defines, shader_q8_0_dequant_func_compat if not fp16 else shader_q8_0_dequant_func, get_rows_body))
else:
continue
tasks.append(string_to_spv(f"get_rows_{type_names[i]}", "".join(stream), {"B_TYPE": "float", "D_TYPE": "float16_t"}, fp16))
tasks.append(string_to_spv(f"get_rows_{type_names[i]}_f32", "".join(stream), {"B_TYPE": "float", "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, shader_f16_dequant_func_compat, mul_mat_vec_body))
elif i == GGML_TYPE_Q4_0:
stream.extend((shader_q4_0_defines, shader_q4_0_dequant_func_compat, mul_mat_vec_body))
elif i == GGML_TYPE_Q4_1:
stream.extend((shader_q4_1_defines, shader_q4_1_dequant_func_compat, mul_mat_vec_body))
elif i == GGML_TYPE_Q5_0:
stream.extend((shader_q5_0_defines, shader_q5_0_dequant_func_compat, mul_mat_vec_body))
elif i == GGML_TYPE_Q5_1:
stream.extend((shader_q5_1_defines, shader_q5_1_dequant_func_compat, mul_mat_vec_body))
elif i == GGML_TYPE_Q8_0:
stream.extend((shader_q8_0_defines, shader_q8_0_dequant_func_compat, mul_mat_vec_body))
elif i == GGML_TYPE_Q2_K:
stream.extend((shader_q2_K_defines, mul_mat_vec_q2_K_body))
elif i == GGML_TYPE_Q3_K:
stream.extend((shader_q3_K_defines, mul_mat_vec_q3_K_body))
elif i == GGML_TYPE_Q4_K:
stream.extend((shader_q4_K_defines, mul_mat_vec_q4_K_body))
elif i == GGML_TYPE_Q5_K:
stream.extend((shader_q5_K_defines, mul_mat_vec_q5_K_body))
elif i == GGML_TYPE_Q6_K:
stream.extend((shader_q6_K_defines, mul_mat_vec_q6_K_body))
else:
continue
tasks.append(string_to_spv(f"mul_mat_vec_{type_names[i]}_f32", "".join(stream), {"B_TYPE": "float", "D_TYPE": "float", "K_QUANTS_PER_ITERATION": K_QUANTS_PER_ITERATION}, fp16))
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"}, True))
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"}, True))
# Norms
tasks.append(string_to_spv("norm_f32", f"{generic_head}\n{shader_f32}\n{norm_body}", {"A_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("rms_norm_f32", f"{generic_head}\n{shader_f32}\n{rms_norm_body}", {"A_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("cpy_f32_f32", f"{cpy_src}\n{cpy_end}", {"A_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("cpy_f32_f16", f"{cpy_src}\n{cpy_end}", {"A_TYPE": "float", "D_TYPE": "float16_t"}, True))
tasks.append(string_to_spv("cpy_f16_f16", f"{cpy_src}\n{cpy_f16_f16_end}", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}, True))
tasks.append(string_to_spv("add_f32", f"{generic_head}\n{shader_f32}\n{add_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("split_k_reduce", mulmat_split_k_reduce_src, {}, True))
tasks.append(string_to_spv("mul_f32", f"{generic_head}\n{shader_f32}\n{mul_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("scale_f32", f"{generic_head}\n{shader_f32}\n{scale_body}", {"A_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("sqr_f32", f"{generic_head}\n{shader_f32}\n{sqr_body}", {"A_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("clamp_f32", f"{generic_head}\n{shader_f32}\n{clamp_body}", {"A_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("gelu_f32", f"{generic_head}\n{shader_f32}\n{gelu_body}", {"A_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("silu_f32", f"{generic_head}\n{shader_f32}\n{silu_body}", {"A_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("relu_f32", f"{generic_head}\n{shader_f32}\n{relu_body}", {"A_TYPE": "float", "D_TYPE": "float"}, True))
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"}, True))
tasks.append(string_to_spv("soft_max_f32", f"{generic_head}\n{shader_f32}\n{soft_max_body}", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("rope_f32", rope_src, {"A_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("rope_f16", rope_src, {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}, True))
tasks.append(string_to_spv("rope_neox_f32", rope_neox_src, {"A_TYPE": "float", "D_TYPE": "float"}, True))
tasks.append(string_to_spv("rope_neox_f16", rope_neox_src, {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}, True))
await asyncio.gather(*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")
args = parser.parse_args()
if args.glslc:
GLSLC = args.glslc
asyncio.run(main())