mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 32 Packages Projects Releases Wiki Activity

opencl: add flattened Q4_K mv and general Q4_K mm (#20773)

Browse Source
This commit is contained in:
shaofeiqi
2026-03-22 22:45:11 -07:00
committed by GitHub
parent ec2b787ebe
commit 84ffd0c192
5 changed files with 733 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-opencl/ggml-opencl.cpp
+289
View File
@@ -534,11 +534,13 @@ struct ggml_backend_opencl_context {
cl_kernel kernel_restore_block_q4_0_noshuffle;
cl_kernel kernel_convert_block_q4_1_noshuffle;
cl_kernel kernel_restore_block_q4_1_noshuffle;
cl_kernel kernel_convert_block_q4_K, kernel_restore_block_q4_K;
cl_kernel kernel_convert_block_q6_K, kernel_restore_block_q6_K;
cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
cl_kernel kernel_mul_mv_q4_1_f32;
cl_kernel kernel_mul_mv_q4_1_f32_flat;
cl_kernel kernel_mul_mv_q4_K_f32;
cl_kernel kernel_mul_mv_q4_K_f32_flat;
cl_kernel kernel_mul_mv_q6_K_f32;
cl_kernel kernel_mul_mv_q6_K_f32_flat;
cl_kernel kernel_mul_mv_mxfp4_f32, kernel_mul_mv_mxfp4_f32_flat;
@@ -578,6 +580,7 @@ struct ggml_backend_opencl_context {
cl_kernel kernel_mul_mm_q4_0_f32_l4_lm;
cl_kernel kernel_mul_mm_q4_1_f32_l4_lm;
cl_kernel kernel_mul_mm_q8_0_f32_l4_lm;
cl_kernel kernel_mul_mm_q4_k_f32_l4_lm;
cl_kernel kernel_mul_mm_q6_k_f32_l4_lm;
std::vector<ProfilingInfo> profiling_info;
@@ -917,6 +920,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
CL_CHECK((backend_ctx->kernel_convert_block_q8_0 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q8_0", &err), err));
CL_CHECK((backend_ctx->kernel_restore_block_q8_0 = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0", &err), err));
CL_CHECK((backend_ctx->kernel_restore_block_q8_0_trans = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0_trans", &err), err));
CL_CHECK((backend_ctx->kernel_convert_block_q4_K = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_K", &err), err));
CL_CHECK((backend_ctx->kernel_restore_block_q4_K = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_K", &err), err));
CL_CHECK((backend_ctx->kernel_convert_block_q6_K = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K", &err), err));
CL_CHECK((backend_ctx->kernel_restore_block_q6_K = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q6_K", &err), err));
GGML_LOG_CONT(".");
@@ -1209,6 +1214,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
GGML_LOG_CONT(".");
}
// mul_mv_q4_k_f32_flat
{
#ifdef GGML_OPENCL_EMBED_KERNELS
const std::string kernel_src {
#include "mul_mv_q4_k_f32_flat.cl.h"
};
#else
const std::string kernel_src = read_file("mul_mv_q4_k_f32_flat.cl");
#endif
cl_program prog =
build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
CL_CHECK((backend_ctx->kernel_mul_mv_q4_K_f32_flat = clCreateKernel(prog, "kernel_mul_mv_q4_K_f32_flat", &err), err));
CL_CHECK(clReleaseProgram(prog));
GGML_LOG_CONT(".");
}
// mul_mv_q6_k_f32
{
#ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1482,6 +1504,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
GGML_LOG_CONT(".");
}
// mul_mm_q4_k_f32_l4_lm
{
#ifdef GGML_OPENCL_EMBED_KERNELS
const std::string kernel_src {
#include "mul_mm_q4_k_f32_l4_lm.cl.h"
};
#else
const std::string kernel_src = read_file("mul_mm_q4_k_f32_l4_lm.cl");
#endif
cl_program prog =
build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
CL_CHECK((backend_ctx->kernel_mul_mm_q4_k_f32_l4_lm = clCreateKernel(prog, "kernel_mul_mm_q4_k_f32_l4_lm", &err), err));
CL_CHECK(clReleaseProgram(prog));
GGML_LOG_CONT(".");
}
// mul_mm_q6_k_f32_l4_lm
{
#ifdef GGML_OPENCL_EMBED_KERNELS
@@ -3347,6 +3386,40 @@ struct ggml_tensor_extra_cl_q8_0 {
}
};
struct ggml_tensor_extra_cl_q4_K {
// Quantized values
cl_mem q = nullptr;
// Scales for each super block.
cl_mem s = nullptr;
// Scales
cl_mem d = nullptr;
// Min
cl_mem dm = nullptr;
~ggml_tensor_extra_cl_q4_K() {
reset();
}
void reset() {
if (q != nullptr) {
CL_CHECK(clReleaseMemObject(q));
q = nullptr;
}
if (s != nullptr) {
CL_CHECK(clReleaseMemObject(s));
s = nullptr;
}
if (d != nullptr) {
CL_CHECK(clReleaseMemObject(d));
d = nullptr;
}
if (dm != nullptr) {
CL_CHECK(clReleaseMemObject(dm));
dm = nullptr;
}
}
};
struct ggml_tensor_extra_cl_q6_K {
// Lower 4 bits of quantized weights.
cl_mem ql = nullptr;
@@ -3956,6 +4029,12 @@ struct ggml_backend_opencl_buffer_context {
for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0_in_use) {
delete e;
}
for (ggml_tensor_extra_cl_q4_K * e : temp_tensor_extras_q4_K) {
delete e;
}
for (ggml_tensor_extra_cl_q4_K * e : temp_tensor_extras_q4_K_in_use) {
delete e;
}
for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K) {
delete e;
}
@@ -4039,6 +4118,21 @@ struct ggml_backend_opencl_buffer_context {
return extra;
}
ggml_tensor_extra_cl_q4_K * ggml_opencl_alloc_temp_tensor_extra_q4_K() {
ggml_tensor_extra_cl_q4_K * extra;
if (temp_tensor_extras_q4_K.empty()) {
extra = new ggml_tensor_extra_cl_q4_K();
} else {
extra = temp_tensor_extras_q4_K.back();
temp_tensor_extras_q4_K.pop_back();
}
temp_tensor_extras_q4_K_in_use.push_back(extra);
extra->reset();
return extra;
}
ggml_tensor_extra_cl_q6_K * ggml_opencl_alloc_temp_tensor_extra_q6_K() {
ggml_tensor_extra_cl_q6_K * extra;
if (temp_tensor_extras_q6_K.empty()) {
@@ -4080,6 +4174,11 @@ struct ggml_backend_opencl_buffer_context {
}
temp_tensor_extras_q8_0_in_use.clear();
for (ggml_tensor_extra_cl_q4_K * e : temp_tensor_extras_q4_K_in_use) {
temp_tensor_extras_q4_K.push_back(e);
}
temp_tensor_extras_q4_K_in_use.clear();
for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K_in_use) {
temp_tensor_extras_q6_K.push_back(e);
}
@@ -4101,6 +4200,8 @@ struct ggml_backend_opencl_buffer_context {
std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4_in_use;
std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0;
std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0_in_use;
std::vector<ggml_tensor_extra_cl_q4_K *> temp_tensor_extras_q4_K;
std::vector<ggml_tensor_extra_cl_q4_K *> temp_tensor_extras_q4_K_in_use;
std::vector<ggml_tensor_extra_cl_q6_K *> temp_tensor_extras_q6_K;
std::vector<ggml_tensor_extra_cl_q6_K *> temp_tensor_extras_q6_K_in_use;
@@ -4835,6 +4936,83 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
return;
}
if (tensor->type == GGML_TYPE_Q4_K) {
ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
// Allocate the new extra and create aliases from the original.
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
ggml_tensor_extra_cl_q4_K * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_q4_K();
size_t size_d = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
size_t size_dm = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
size_t size_s = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*(3 * ggml_blck_size(tensor->type) / 64);
size_t size_q = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
GGML_ASSERT(size_d + size_dm + size_s + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
cl_int err;
cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
ggml_nbytes(tensor), NULL, &err);
CL_CHECK(err);
CL_CHECK(clEnqueueWriteBuffer(
queue, data_device, CL_TRUE, 0,
ggml_nbytes(tensor), data, 0, NULL, NULL));
cl_buffer_region region;
// Create subbuffer for d.
region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
region.size = size_d;
extra->d = clCreateSubBuffer(
extra_orig->data_device, CL_MEM_READ_WRITE,
CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
CL_CHECK(err);
auto previous_origin = region.origin;
// Create subbuffer for mins.
region.origin = align_to(previous_origin + size_d, backend_ctx->alignment);
region.size = size_dm;
extra->dm = clCreateSubBuffer(
extra_orig->data_device, CL_MEM_READ_WRITE,
CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
CL_CHECK(err);
previous_origin = region.origin;
// Create subbuffer for s.
region.origin = align_to(previous_origin + size_dm, backend_ctx->alignment);
region.size = size_s;
extra->s = clCreateSubBuffer(
extra_orig->data_device, CL_MEM_READ_WRITE,
CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
CL_CHECK(err);
previous_origin = region.origin;
// Create subbuffer for quants.
region.origin = align_to(previous_origin + size_s, backend_ctx->alignment);
region.size = size_q;
extra->q = clCreateSubBuffer(
extra_orig->data_device, CL_MEM_READ_WRITE,
CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
CL_CHECK(err);
cl_kernel kernel = backend_ctx->kernel_convert_block_q4_K;
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->s));
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->d));
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra->dm));
size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
size_t local_work_size[] = {64, 1, 1};
cl_event evt;
CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
CL_CHECK(clWaitForEvents(1, &evt));
CL_CHECK(clReleaseMemObject(data_device));
tensor->extra = extra;
return;
}
if (tensor->type == GGML_TYPE_Q6_K) {
ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
@@ -5245,6 +5423,34 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
CL_CHECK(clReleaseMemObject(data_device));
return;
}
if (tensor->type == GGML_TYPE_Q4_K) {
ggml_tensor_extra_cl_q4_K * extra = (ggml_tensor_extra_cl_q4_K *)tensor->extra;
cl_int err;
cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
ggml_nbytes(tensor), NULL, &err);
CL_CHECK(err);
cl_kernel kernel = backend_ctx->kernel_restore_block_q4_K;
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->s));
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->d));
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->dm));
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &data_device));
size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
size_t local_work_size[] = {1, 1, 1};
cl_event evt;
CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
global_work_size, local_work_size, 0, NULL, &evt));
CL_CHECK(clWaitForEvents(1, &evt));
CL_CHECK(clEnqueueReadBuffer(
queue, data_device, CL_TRUE, offset,
size, data, 0, NULL, NULL));
CL_CHECK(clReleaseMemObject(data_device));
return;
}
if (tensor->type == GGML_TYPE_Q6_K) {
ggml_tensor_extra_cl_q6_K * extra = (ggml_tensor_extra_cl_q6_K *)tensor->extra;
@@ -9357,6 +9563,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
ggml_tensor_extra_cl_q4_1 * extra0_q4_1 = (ggml_tensor_extra_cl_q4_1 *)src0->extra;
ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra;
ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
ggml_tensor_extra_cl_q4_K * extra0_q4_K = (ggml_tensor_extra_cl_q4_K *)src0->extra;
ggml_tensor_extra_cl_q6_K * extra0_q6_K = (ggml_tensor_extra_cl_q6_K *)src0->extra;
#endif
@@ -10005,6 +10212,50 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
return;
}
case GGML_TYPE_Q4_K: {
if (ne11 < 32) {
break;
}
if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
break;
}
kernel = backend_ctx->kernel_mul_mm_q4_k_f32_l4_lm;
nth0 = 128; // calculated as (BM*BN)/(TM*TN)
int batch_stride_a = ne00*ne01;
int batch_stride_b = ne10*ne11;
int batch_stride_d = ne0*ne1;
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0_q4_K->q));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra0_q4_K->s));
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra0_q4_K->d));
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra0_q4_K->dm));
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra1->data_device));
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset1));
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem), &extrad->data_device));
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offsetd));
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne02));
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne11));
CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int), &ne12));
CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int), &ne10)); // stride_a
CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int), &ne10)); // stride_b
CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int), &ne01)); // stride_d
CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int), &batch_stride_a));
CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int), &batch_stride_b));
CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int), &batch_stride_d));
CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int), &r2));
CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int), &r3));
// 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
size_t local_work_size[] = {(size_t)nth0, 1, 1};
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
return;
}
case GGML_TYPE_Q6_K: {
if (ne11 < 32) {
break;
@@ -10449,6 +10700,43 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K: {
#ifdef GGML_OPENCL_SOA_Q
kernel = backend_ctx->kernel_mul_mv_q4_K_f32_flat;
if (backend_ctx->gpu_family == INTEL) {
nth0 = 16;
nth1 = 1;
ndst = 4;
} else if (backend_ctx->gpu_family == ADRENO) {
nth0 = 64;
nth1 = 2;
ndst = 16;
} else {
GGML_ASSERT(false && "TODO: Unknown GPU");
}
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0_q4_K->q));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra0_q4_K->s));
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra0_q4_K->d));
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra0_q4_K->dm));
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra1->data_device));
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &offset1));
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem), &extrad->data_device));
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &offsetd));
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb01));
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb02));
CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb03));
CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int), &ne12));
CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int), &ne0));
CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int), &ne1));
CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int), &r2));
CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int), &r3));
#else
kernel = backend_ctx->kernel_mul_mv_q4_K_f32;
if (backend_ctx->gpu_family == INTEL) {
@@ -10482,6 +10770,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int), &ne1));
CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int), &r2));
CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int), &r3));
#endif // GGML_OPENCL_SOA_Q
break;
}
case GGML_TYPE_Q5_K: