opencl: add q6_K gemm and gemv kernels for Adreno (#20089)
* opencl: add q6_K noshuffle kernels, initial q6_K gemv, some host code * opencl: add q6_K transpose * opencl: fix cvt kernel name * opencl: add call to q6_K gemv * opencl: fix q6_K scale transpose * opencl: fix loading for gemv q6_K, refactor * opencl: fix transpose_8_buf kernel assignment, refactor * opencl: refactor q6_K transpose * opencl: add gemm_noshuffle_q6_k_f32 * opencl: fix qh loading * opencl: refactor q6_K gemv host side, release bufs and imgs * opencl: refactor * opencl: fix q6_K dequant and scale selection * opencl: workaround compiler bug, fix dump_tensor * opencl: refactor q6_K convert kernels * opencl: unpack transformed q6_K in get_tensor * opencl: refactor, handle non-uniform workgroups * opencl: support non-vector subgroup bcast
This commit is contained in:
lhez
@@ -529,6 +529,7 @@ struct ggml_backend_opencl_context {
|
||||
cl_kernel kernel_convert_block_q4_1, kernel_restore_block_q4_1;
|
||||
cl_kernel kernel_convert_block_mxfp4, kernel_convert_block_mxfp4_trans, kernel_restore_block_mxfp4, kernel_restore_block_mxfp4_trans;
|
||||
cl_kernel kernel_convert_block_q8_0, kernel_restore_block_q8_0, kernel_restore_block_q8_0_trans;
|
||||
cl_kernel kernel_convert_block_q6_K_noshuffle, kernel_restore_block_q6_K_noshuffle;
|
||||
cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
|
||||
cl_kernel kernel_convert_block_q4_0_noshuffle;
|
||||
cl_kernel kernel_restore_block_q4_0_noshuffle;
|
||||
@@ -716,6 +717,8 @@ struct ggml_backend_opencl_context {
|
||||
cl_kernel kernel_gemm_noshuffle_q4_1_f32;
|
||||
cl_kernel kernel_mul_mm_q8_0_f32_8x4;
|
||||
cl_kernel CL_mul_mat_vec_q8_0_f32;
|
||||
cl_kernel kernel_gemv_noshuffle_q6_K_f32;
|
||||
cl_kernel kernel_gemm_noshuffle_q6_K_f32;
|
||||
#endif // GGML_OPENCL_USE_ADRENO_KERNELS
|
||||
|
||||
void free() {
|
||||
@@ -924,6 +927,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
|
||||
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));
|
||||
CL_CHECK((backend_ctx->kernel_convert_block_q6_K_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K_noshuffle", &err), err));
|
||||
CL_CHECK((backend_ctx->kernel_restore_block_q6_K_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q6_K_noshuffle", &err), err));
|
||||
GGML_LOG_CONT(".");
|
||||
}
|
||||
|
||||
@@ -2642,6 +2647,45 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
|
||||
CL_CHECK((backend_ctx->kernel_gemm_moe_mxfp4_f32 = clCreateKernel(backend_ctx->program_gemm_moe_mxfp4_f32, "kernel_gemm_moe_mxfp4_f32", &err), err));
|
||||
GGML_LOG_CONT(".");
|
||||
}
|
||||
|
||||
// gemv_noshuffle_q6_k_f32
|
||||
{
|
||||
#ifdef GGML_OPENCL_EMBED_KERNELS
|
||||
const std::string kernel_src {
|
||||
#include "gemv_noshuffle_q6_k_f32.cl.h"
|
||||
};
|
||||
#else
|
||||
const std::string kernel_src = read_file("gemv_noshuffle_q6_k_f32.cl");
|
||||
#endif
|
||||
|
||||
std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
|
||||
" -cl-mad-enable ";
|
||||
if (backend_ctx->has_vector_subgroup_broadcast) {
|
||||
CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT ";
|
||||
}
|
||||
|
||||
cl_program prog =
|
||||
build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_gemv_compile_opts);
|
||||
|
||||
CL_CHECK((backend_ctx->kernel_gemv_noshuffle_q6_K_f32 = clCreateKernel(prog, "kernel_gemv_noshuffle_q6_K_f32", &err), err));
|
||||
GGML_LOG_CONT(".");
|
||||
}
|
||||
|
||||
// gemm_noshuffle_q6_k_f32
|
||||
{
|
||||
#ifdef GGML_OPENCL_EMBED_KERNELS
|
||||
const std::string kernel_src {
|
||||
#include "gemm_noshuffle_q6_k_f32.cl.h"
|
||||
};
|
||||
#else
|
||||
const std::string kernel_src = read_file("gemm_noshuffle_q6_k_f32.cl");
|
||||
#endif
|
||||
cl_program prog =
|
||||
build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_moe_compile_opts);
|
||||
|
||||
CL_CHECK((backend_ctx->kernel_gemm_noshuffle_q6_K_f32 = clCreateKernel(prog, "kernel_gemm_noshuffle_q6_K_f32", &err), err));
|
||||
GGML_LOG_CONT(".");
|
||||
}
|
||||
#endif // GGML_OPENCL_USE_ADRENO_KERNELS
|
||||
GGML_LOG_CONT("\n");
|
||||
}
|
||||
@@ -5029,61 +5073,58 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
|
||||
"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_mem data_device;
|
||||
CL_CHECK((data_device = clCreateBuffer(context, CL_MEM_READ_WRITE, ggml_nbytes(tensor), NULL, &err), err));
|
||||
CL_CHECK(clEnqueueWriteBuffer(queue, data_device, CL_TRUE, 0, ggml_nbytes(tensor), data, 0, NULL, NULL));
|
||||
|
||||
cl_buffer_region region;
|
||||
|
||||
// Subbuffer for ql
|
||||
region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
|
||||
region.size = size_ql;
|
||||
extra->ql = clCreateSubBuffer(
|
||||
extra_orig->data_device, CL_MEM_READ_WRITE,
|
||||
CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err);
|
||||
CL_CHECK(err);
|
||||
CL_CHECK((extra->ql = clCreateSubBuffer(extra_orig->data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err), err));
|
||||
auto previous_origin = region.origin;
|
||||
|
||||
// Subbuffer for qh
|
||||
region.origin = align_to(previous_origin + size_ql, backend_ctx->alignment);
|
||||
region.size = size_qh;
|
||||
extra->qh = clCreateSubBuffer(
|
||||
extra_orig->data_device, CL_MEM_READ_WRITE,
|
||||
CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err);
|
||||
CL_CHECK(err);
|
||||
CL_CHECK((extra->qh = clCreateSubBuffer(extra_orig->data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err), err));
|
||||
previous_origin = region.origin;
|
||||
|
||||
// Subbuffer for scales
|
||||
region.origin = align_to(previous_origin + size_qh, backend_ctx->alignment);
|
||||
region.size = size_s;
|
||||
extra->s = clCreateSubBuffer(
|
||||
extra_orig->data_device, CL_MEM_READ_WRITE,
|
||||
CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err);
|
||||
CL_CHECK(err);
|
||||
CL_CHECK((extra->s = clCreateSubBuffer(extra_orig->data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err), err));
|
||||
previous_origin = region.origin;
|
||||
|
||||
// Create subbuffer for d.
|
||||
region.origin = align_to(previous_origin + size_s, backend_ctx->alignment);
|
||||
region.size = size_d;
|
||||
extra->d = clCreateSubBuffer(
|
||||
extra_orig->data_device, CL_MEM_READ_WRITE,
|
||||
CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err);
|
||||
CL_CHECK(err);
|
||||
CL_CHECK((extra->d = clCreateSubBuffer(extra_orig->data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err), err));
|
||||
previous_origin = region.origin;
|
||||
|
||||
// Flatten the weights
|
||||
cl_kernel kernel = backend_ctx->kernel_convert_block_q6_K;
|
||||
cl_kernel kernel;
|
||||
#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
|
||||
kernel = backend_ctx->kernel_convert_block_q6_K;
|
||||
if (use_adreno_kernels(backend_ctx, tensor)) {
|
||||
kernel = backend_ctx->kernel_convert_block_q6_K_noshuffle;
|
||||
}
|
||||
#else
|
||||
kernel = backend_ctx->kernel_convert_block_q6_K;
|
||||
#endif // GGML_OPENCL_USE_ADRENO_KERNELS
|
||||
|
||||
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
|
||||
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->ql));
|
||||
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->qh));
|
||||
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->s));
|
||||
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra->d));
|
||||
cl_uchar mask = 0xff;
|
||||
cl_ulong n_blk = ggml_nelements(tensor)/ggml_blck_size(tensor->type);
|
||||
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
|
||||
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->ql));
|
||||
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->qh));
|
||||
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->s));
|
||||
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra->d));
|
||||
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_uchar), &mask));
|
||||
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &n_blk));
|
||||
|
||||
size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
|
||||
size_t global_work_size[] = {(size_t)CEIL_DIV(n_blk, 64)*64, 1, 1};
|
||||
size_t local_work_size[] = {64, 1, 1};
|
||||
|
||||
cl_event evt;
|
||||
@@ -5097,6 +5138,29 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
|
||||
extra->size_d = size_d;
|
||||
|
||||
tensor->extra = extra;
|
||||
|
||||
#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
|
||||
if (use_adreno_kernels(backend_ctx, tensor)) {
|
||||
cl_int M = tensor->ne[1]; // ne01
|
||||
cl_int K = tensor->ne[0]; // ne00
|
||||
|
||||
// Transpose ql as ushort
|
||||
transpose_2d_as_16b(backend_ctx,
|
||||
extra->ql, extra->ql, size_ql, K/4, M);
|
||||
|
||||
// Transpose qh as uchar
|
||||
transpose_2d_as_8b(backend_ctx,
|
||||
extra->qh, extra->qh, size_qh, K/4, M);
|
||||
|
||||
// Transpose s as ushort
|
||||
transpose_2d_as_16b(backend_ctx,
|
||||
extra->s, extra->s, size_s, K/16/2, M);
|
||||
|
||||
// Transpose d as ushort
|
||||
transpose_2d_as_16b(backend_ctx,
|
||||
extra->d, extra->d, size_d, K/256, M);
|
||||
}
|
||||
#endif // GGML_OPENCL_USE_ADRENO_KERNELS
|
||||
return;
|
||||
}
|
||||
#endif // GGML_OPENCL_SOA_Q
|
||||
@@ -5454,19 +5518,78 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
|
||||
if (tensor->type == GGML_TYPE_Q6_K) {
|
||||
ggml_tensor_extra_cl_q6_K * extra = (ggml_tensor_extra_cl_q6_K *)tensor->extra;
|
||||
|
||||
#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
|
||||
if (use_adreno_kernels(backend_ctx, tensor)) {
|
||||
static ggml_cl_buffer buf_trans_ql;
|
||||
static ggml_cl_buffer buf_trans_qh;
|
||||
static ggml_cl_buffer buf_trans_s;
|
||||
static ggml_cl_buffer buf_trans_d;
|
||||
static ggml_cl_buffer buf_unpacked;
|
||||
|
||||
cl_int M = tensor->ne[1]; // ne01
|
||||
cl_int K = tensor->ne[0]; // ne00
|
||||
|
||||
GGML_ASSERT(K % ggml_blck_size(tensor->type) == 0);
|
||||
|
||||
size_t size_ql = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
|
||||
size_t size_qh = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/4;
|
||||
size_t size_s = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/16;
|
||||
size_t size_d = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
|
||||
GGML_ASSERT(size_ql + size_qh + size_s + size_d == ggml_nbytes(tensor) && "Incorrect tensor size");
|
||||
|
||||
buf_trans_ql.allocate(backend_ctx->context, size_ql);
|
||||
buf_trans_qh.allocate(backend_ctx->context, size_qh);
|
||||
buf_trans_s.allocate(backend_ctx->context, size_s);
|
||||
buf_trans_d.allocate(backend_ctx->context, size_d);
|
||||
buf_unpacked.allocate(backend_ctx->context, ggml_nbytes(tensor));
|
||||
|
||||
// transpose ql, qh, s and d back
|
||||
transpose_2d_as_16b(backend_ctx, extra->ql, buf_trans_ql.buffer, size_ql, M, K/4);
|
||||
transpose_2d_as_8b(backend_ctx, extra->qh, buf_trans_qh.buffer, size_qh, M, K/4);
|
||||
transpose_2d_as_16b(backend_ctx, extra->s, buf_trans_s.buffer, size_s, M, K/16/2);
|
||||
transpose_2d_as_16b(backend_ctx, extra->d, buf_trans_d.buffer, size_d, M, K/256);
|
||||
|
||||
// unpack
|
||||
cl_uchar mask = 0xFF;
|
||||
cl_ulong n_blk = ggml_nelements(tensor)/ggml_blck_size(tensor->type);
|
||||
cl_kernel kernel = backend_ctx->kernel_restore_block_q6_K_noshuffle;
|
||||
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &buf_trans_ql.buffer));
|
||||
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &buf_trans_qh.buffer));
|
||||
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &buf_trans_s.buffer));
|
||||
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &buf_trans_d.buffer));
|
||||
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &buf_unpacked.buffer));
|
||||
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_uchar), &mask));
|
||||
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &n_blk));
|
||||
|
||||
size_t global_work_size[] = {(size_t)n_blk, 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, buf_unpacked.buffer, CL_TRUE, offset, size, data, 0, NULL, NULL));
|
||||
|
||||
return;
|
||||
}
|
||||
#endif // GGML_OPENCL_USE_ADRENO_KERNELS
|
||||
|
||||
cl_int err;
|
||||
cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
|
||||
ggml_nbytes(tensor), NULL, &err);
|
||||
CL_CHECK(err);
|
||||
|
||||
cl_uchar mask = 0xFF;
|
||||
cl_ulong n_blk = ggml_nelements(tensor)/ggml_blck_size(tensor->type);
|
||||
cl_kernel kernel = backend_ctx->kernel_restore_block_q6_K;
|
||||
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->ql));
|
||||
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->qh));
|
||||
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), &data_device));
|
||||
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->ql));
|
||||
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->qh));
|
||||
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), &data_device));
|
||||
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_uchar), &mask));
|
||||
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &n_blk));
|
||||
|
||||
size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
|
||||
size_t global_work_size[] = {(size_t)n_blk, 1, 1};
|
||||
size_t local_work_size[] = {1, 1, 1};
|
||||
|
||||
cl_event evt;
|
||||
@@ -5759,6 +5882,8 @@ typedef struct {
|
||||
static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2,
|
||||
"wrong q4_0 block size/padding");
|
||||
|
||||
#define QK_MXFP4 32
|
||||
|
||||
#include <math.h>
|
||||
#ifdef __cplusplus
|
||||
#include "half.hpp"
|
||||
@@ -5802,7 +5927,7 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso
|
||||
buf_d = malloc(size_e);
|
||||
|
||||
CL_CHECK(clEnqueueReadBuffer(queue, extra->q, CL_TRUE, 0, size_q, buf_q, 0, NULL, NULL));
|
||||
CL_CHECK(clEnqueueReadBuffer(queue, extra->d, CL_TRUE, 0, size_e, buf_d, 0, NULL, NULL));
|
||||
CL_CHECK(clEnqueueReadBuffer(queue, extra->e, CL_TRUE, 0, size_e, buf_d, 0, NULL, NULL));
|
||||
CL_CHECK(clFinish(queue));
|
||||
} else {
|
||||
// Read out the tensor from GPU memory.
|
||||
@@ -9537,6 +9662,196 @@ static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_t
|
||||
#endif
|
||||
}
|
||||
|
||||
static void ggml_cl_mul_mat_q6_K_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
||||
#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
|
||||
GGML_ASSERT(src0);
|
||||
GGML_ASSERT(src0->extra);
|
||||
GGML_ASSERT(src1);
|
||||
GGML_ASSERT(src1->extra);
|
||||
GGML_ASSERT(dst);
|
||||
GGML_ASSERT(dst->extra);
|
||||
|
||||
ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
|
||||
|
||||
ggml_tensor_extra_cl_q6_K * extra0_q6_K = (ggml_tensor_extra_cl_q6_K *)src0->extra;
|
||||
ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
|
||||
ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
|
||||
|
||||
cl_ulong offset1 = extra1->offset + src1->view_offs;
|
||||
cl_ulong offsetd = extrad->offset + dst->view_offs;
|
||||
|
||||
const int ne00 = src0->ne[0];
|
||||
const int ne01 = src0->ne[1];
|
||||
|
||||
const int ne1 = dst->ne[1];
|
||||
|
||||
GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
|
||||
|
||||
cl_context context = backend_ctx->context;
|
||||
cl_kernel kernel;
|
||||
|
||||
cl_int err;
|
||||
cl_buffer_region region;
|
||||
cl_image_format img_fmt;
|
||||
cl_image_desc img_desc;
|
||||
|
||||
// subbuffer and image for activation
|
||||
if (ne1 == 1) {
|
||||
cl_mem ql_img = nullptr;
|
||||
cl_mem qh_img = nullptr;
|
||||
cl_mem b_sub_buffer = nullptr;
|
||||
cl_mem b_img = nullptr;
|
||||
|
||||
// image for ql
|
||||
img_fmt.image_channel_order = CL_R;
|
||||
img_fmt.image_channel_data_type = CL_FLOAT;
|
||||
memset(&img_desc, 0, sizeof(img_desc));
|
||||
img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
|
||||
img_desc.image_width = ne01 * ne00 / 8;
|
||||
img_desc.buffer = extra0_q6_K->ql;
|
||||
CL_CHECK((ql_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
|
||||
|
||||
// image for qh
|
||||
img_fmt.image_channel_order = CL_R;
|
||||
img_fmt.image_channel_data_type = CL_HALF_FLOAT;
|
||||
memset(&img_desc, 0, sizeof(img_desc));
|
||||
img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
|
||||
img_desc.image_width = ne01 * ne00 / 8;
|
||||
img_desc.buffer = extra0_q6_K->qh;
|
||||
CL_CHECK((qh_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
|
||||
|
||||
region.origin = offset1;
|
||||
region.size = ne00 * ne1 * sizeof(float);
|
||||
CL_CHECK((b_sub_buffer = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err), err));
|
||||
|
||||
img_fmt.image_channel_order = CL_RGBA;
|
||||
img_fmt.image_channel_data_type = CL_FLOAT;
|
||||
memset(&img_desc, 0, sizeof(img_desc));
|
||||
img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
|
||||
img_desc.image_width = ne00 * ne1 / 4;
|
||||
img_desc.buffer = b_sub_buffer;
|
||||
CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
|
||||
|
||||
kernel = backend_ctx->kernel_gemv_noshuffle_q6_K_f32;
|
||||
|
||||
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &ql_img));
|
||||
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &qh_img));
|
||||
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), &b_img));
|
||||
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem), &extrad->data_device));
|
||||
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &offsetd));
|
||||
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_int), &ne00));
|
||||
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_int), &ne01));
|
||||
|
||||
size_t local_work_size[3] = {64, 4, 1};
|
||||
size_t global_work_size[3] = {(size_t)CEIL_DIV(ne01/2, 64)*64, 4, 1};
|
||||
|
||||
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
|
||||
|
||||
CL_CHECK(clReleaseMemObject(ql_img));
|
||||
CL_CHECK(clReleaseMemObject(qh_img));
|
||||
CL_CHECK(clReleaseMemObject(b_sub_buffer));
|
||||
CL_CHECK(clReleaseMemObject(b_img));
|
||||
} else {
|
||||
cl_mem b_sub_buf;
|
||||
cl_mem b_buf_trans;
|
||||
cl_mem b_img;
|
||||
cl_mem b_img_trans;
|
||||
|
||||
// subbuffer for activation
|
||||
region.origin = offset1;
|
||||
region.size = ne00 * ne1 * sizeof(float);
|
||||
CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err), err));
|
||||
|
||||
// image for activation
|
||||
img_fmt.image_channel_order = CL_RGBA;
|
||||
img_fmt.image_channel_data_type = CL_FLOAT;
|
||||
memset(&img_desc, 0, sizeof(img_desc));
|
||||
img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
|
||||
img_desc.image_width = ne00 * ne1 / 4;
|
||||
img_desc.buffer = b_sub_buf;
|
||||
CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
|
||||
|
||||
// pad N to multiple of 8
|
||||
int extra_elements = ne1 % 8;
|
||||
int padding = 0;
|
||||
if (extra_elements > 0){
|
||||
padding = 8 - extra_elements;
|
||||
}
|
||||
|
||||
// subbuffer for transposed activation
|
||||
region.origin = 0;
|
||||
region.size = ne00 * (ne1 + padding) * sizeof(float)/2;
|
||||
backend_ctx->prealloc_act_trans.allocate(context, region.size);
|
||||
CL_CHECK((b_buf_trans = clCreateSubBuffer(backend_ctx->prealloc_act_trans.buffer, 0, CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err), err));
|
||||
|
||||
// image for transposed activation
|
||||
img_fmt.image_channel_order = CL_RGBA;
|
||||
img_fmt.image_channel_data_type = CL_HALF_FLOAT;
|
||||
memset(&img_desc, 0, sizeof(img_desc));
|
||||
img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
|
||||
img_desc.image_width = ne00 * (ne1 + padding) / 4;
|
||||
img_desc.buffer = b_buf_trans;
|
||||
CL_CHECK((b_img_trans = clCreateImage(context, 0, &img_fmt, &img_desc, NULL, &err), err));
|
||||
|
||||
// transpose activation
|
||||
int height_B = ne1/4;
|
||||
if (height_B == 0) {
|
||||
height_B = 1;
|
||||
}
|
||||
int width_B = ne00/4;
|
||||
int padded_height_B = (ne1 + padding) / 4;
|
||||
|
||||
kernel = backend_ctx->kernel_transpose_32_16;
|
||||
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &b_img));
|
||||
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &b_img_trans));
|
||||
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int), &height_B));
|
||||
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int), &width_B));
|
||||
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &padded_height_B));
|
||||
|
||||
size_t local_size_t[2] = { 1, 16 };
|
||||
size_t global_size_t[2] = { (size_t)width_B, (size_t)padded_height_B };
|
||||
backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_size_t, local_size_t, dst);
|
||||
|
||||
// gemm
|
||||
kernel = backend_ctx->kernel_gemm_noshuffle_q6_K_f32;
|
||||
int padded_N = ne1 + padding;
|
||||
|
||||
cl_ushort mask_f000 = 0xF000;
|
||||
cl_uchar mask_c0 = 0xC0;
|
||||
|
||||
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), &b_img_trans));
|
||||
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem), &extrad->data_device));
|
||||
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &offsetd));
|
||||
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne01));
|
||||
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &padded_N));
|
||||
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne00));
|
||||
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne1));
|
||||
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ushort),&mask_f000));
|
||||
CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_uchar), &mask_c0));
|
||||
|
||||
size_t global_work_size[3] = {(size_t)CEIL_DIV(ne1, 8), (size_t)CEIL_DIV(ne01, 4), 1};
|
||||
size_t local_work_size[3] = {2, 128, 1};
|
||||
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
|
||||
|
||||
CL_CHECK(clReleaseMemObject(b_sub_buf));
|
||||
CL_CHECK(clReleaseMemObject(b_img));
|
||||
CL_CHECK(clReleaseMemObject(b_buf_trans));
|
||||
CL_CHECK(clReleaseMemObject(b_img_trans));
|
||||
}
|
||||
#else
|
||||
GGML_UNUSED(backend);
|
||||
GGML_UNUSED(src0);
|
||||
GGML_UNUSED(src1);
|
||||
GGML_UNUSED(dst);
|
||||
#endif
|
||||
}
|
||||
|
||||
static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
||||
GGML_ASSERT(src0);
|
||||
GGML_ASSERT(src0->extra);
|
||||
@@ -9673,6 +9988,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
|
||||
return;
|
||||
}
|
||||
|
||||
// q6_K x fp32
|
||||
if (src0t == GGML_TYPE_Q6_K && src1t == GGML_TYPE_F32) {
|
||||
ggml_cl_mul_mat_q6_K_f32_adreno(backend, src0, src1, dst);
|
||||
return;
|
||||
}
|
||||
|
||||
// q4_0 x fp32
|
||||
if(src0t == GGML_TYPE_Q4_0 && src1t == GGML_TYPE_F32) {
|
||||
// TODO: remove duplicate definitions of image description + format -- move to top
|
||||
|
||||
Reference in New Issue
Block a user