sleepy/llama.cpp
1
0
Fork 0
Code Issues 16 Pull Requests Actions 31 Packages Projects Releases Wiki Activity

opencl: add general Q6_K mm and Q4_K mv (#19347)

Browse Source 
* opencl: add general q6_k mm

* opencl: refine condition for q6_K mm

* opencl: add general q4_K mv

* opencl: fix whitespace
This commit is contained in:
lhez
2026-02-11 10:33:13 -08:00
committed by GitHub
parent 914dde72ba
commit 4d3daf80f8
4 changed files with 461 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-opencl/ggml-opencl.cpp
+121 -2
View File
@@ -532,6 +532,7 @@ struct ggml_backend_opencl_context {
cl_kernel kernel_restore_block_q4_0_noshuffle;
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_K_f32;
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;
@@ -564,6 +565,7 @@ struct ggml_backend_opencl_context {
cl_kernel kernel_mul_mm_f32_f32_l4_lm;
cl_kernel kernel_mul_mm_f16_f32_l4_lm;
cl_kernel kernel_mul_mm_q8_0_f32_l4_lm;
cl_kernel kernel_mul_mm_q6_k_f32_l4_lm;
std::vector<ProfilingInfo> profiling_info;
@@ -1117,6 +1119,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
GGML_LOG_CONT(".");
}
// mul_mv_q4_k_f32
{
#ifdef GGML_OPENCL_EMBED_KERNELS
const std::string kernel_src {
#include "mul_mv_q4_k_f32.cl.h"
};
#else
const std::string kernel_src = read_file("mul_mv_q4_k_f32.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 = clCreateKernel(prog, "kernel_mul_mv_q4_K_f32", &err), err));
CL_CHECK(clReleaseProgram(prog));
GGML_LOG_CONT(".");
}
// mul_mv_q6_k_f32
{
#ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1358,6 +1377,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
GGML_LOG_CONT(".");
}
// mul_mm_q6_k_f32_l4_lm
{
#ifdef GGML_OPENCL_EMBED_KERNELS
const std::string kernel_src {
#include "mul_mm_q6_k_f32_l4_lm.cl.h"
};
#else
const std::string kernel_src = read_file("mul_mm_q6_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_q6_k_f32_l4_lm = clCreateKernel(prog, "kernel_mul_mm_q6_k_f32_l4_lm", &err), err));
CL_CHECK(clReleaseProgram(prog));
GGML_LOG_CONT(".");
}
// mul_mm_f16_f32_kq_kqv
{
#ifdef GGML_OPENCL_EMBED_KERNELS
@@ -3364,6 +3400,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
} else if (op->src[0]->type == GGML_TYPE_F32) {
return op->src[1]->type == GGML_TYPE_F32;
} else if (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_MXFP4 ||
op->src[0]->type == GGML_TYPE_Q4_K ||
op->src[0]->type == GGML_TYPE_Q6_K) {
return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
} else if (op->src[0]->type == GGML_TYPE_Q8_0) {
@@ -8927,6 +8964,50 @@ 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_Q6_K: {
if (ne11 < 32) {
break;
}
if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
break;
}
kernel = backend_ctx->kernel_mul_mm_q6_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_q6_K->ql));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra0_q6_K->qh));
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra0_q6_K->s));
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra0_q6_K->d));
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;
}
default:
break;
}
@@ -9262,7 +9343,42 @@ 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:
case GGML_TYPE_Q4_K: {
kernel = backend_ctx->kernel_mul_mv_q4_K_f32;
if (backend_ctx->gpu_family == INTEL) {
nth0 = 16;
nth1 = 1;
ndst = 4;
} else if (backend_ctx->gpu_family == ADRENO) {
nth0 = 64;
nth1 = 1;
ndst = 4;
} else {
GGML_ASSERT(false && "TODO: Unknown GPU");
}
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(int), &offset0));
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra1->data_device));
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int), &offset1));
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extrad->data_device));
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &offsetd));
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb01));
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb02));
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb03));
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne12));
CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb11));
CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb12));
CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb13));
CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int), &ne0));
CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int), &ne1));
CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int), &r2));
CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int), &r3));
break;
}
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
#ifdef GGML_OPENCL_SOA_Q
@@ -9424,7 +9540,10 @@ 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);
} else if (src0t == GGML_TYPE_Q4_K) {
GGML_ASSERT(false && "not implemented");
size_t global_work_size[] = {(size_t)(ne01+ndst*nth1-1)/(ndst*nth1)*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13};
size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1};
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
} else if (src0t == GGML_TYPE_Q3_K) {
GGML_ASSERT(false && "not implemented");
} else if (src0t == GGML_TYPE_Q5_K) {