Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

opencl: ragged-tile MoE prefill FP16 GEMM optimization (skip padded expert tiles) (#25433)

Browse Source 
* opencl: ragged-tile MoE prefill GEMM (skip padded expert tiles)

The MoE prefill GEMM groups tokens into TILESIZE_N=32 per-expert tiles; at low
tokens-per-expert most tiles are mostly padding. When a tile's upper 16 slots
are all padding (router index 0xFFFFFFFF), skip the second dotx16_reduce8 half.
Numerically identical (skipped lanes are padding). Applied to all eight *_f32_ns
MoE GEMMs; default on, opt out with GGML_OPENCL_MOE_RAGGED_FP16=0.

* opencl: quarter-granularity ragged MoE tile-skip (8-col skip-groups)

Replace the two half-tile dotx16_reduce8 calls in the 8 *_f32_ns MoE GEMMs with
four dotx8_reduce4 (8-column) calls, skipping each empty trailing skip-group
independently. Padding is always trailing, so the kernel rounds the valid count
up to the skip granularity and skips fully-padding groups. Byte-identical to the
non-skipped path. New env GGML_OPENCL_MOE_RAGGED_GRAN={8,16,32} (quarter/half/
off); default quarter.

* opencl: move ragged moe env var in cl_init

---------

Co-authored-by: Li He <lih@qti.qualcomm.com>
This commit is contained in:
Hongqiang Wang
2026-07-08 09:44:55 -07:00
committed by GitHub
co-authored by Li He
parent 1ee093937f
commit 167d057604
9 changed files with 618 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-opencl/ggml-opencl.cpp
+29
View File
@@ -517,6 +517,10 @@ struct ggml_backend_opencl_context {
bool has_qcom_subgroup_shuffle = false; // specifically cl_qcom_subgroup_shuffle
bool disable_fusion;
// ragged moe, use int to directly pass to kernel
cl_uint adreno_use_moe_ragged;
cl_uint adreno_moe_ragged_skip_gran;
bool adreno_has_large_buffer;
bool adreno_use_large_buffer;
bool adreno_use_bin_kernels;
@@ -5342,6 +5346,15 @@ static ggml_backend_opencl_context * ggml_cl_init(ggml_backend_dev_t dev) {
backend_ctx->adreno_use_large_buffer = getenv("GGML_OPENCL_ADRENO_USE_LARGE_BUFFER") != nullptr &&
backend_ctx->gpu_family == GPU_FAMILY::ADRENO;
// ragged moe, unspecified or non-zero means enabled, set to 0 to disable
static const char * ragged_fp16_env = getenv("GGML_OPENCL_MOE_RAGGED_FP16");
backend_ctx->adreno_use_moe_ragged = (ragged_fp16_env == NULL) ? 1 : (atoi(ragged_fp16_env) != 0);
// ragged moe, tile-skip granularity (columns per skip-group): 8 = quarter (default),
// 16 = half (legacy), 32 = disabled. Override with GGML_OPENCL_MOE_RAGGED_GRAN={8,16,32}
static const char * ragged_gran_env = getenv("GGML_OPENCL_MOE_RAGGED_GRAN");
backend_ctx->adreno_moe_ragged_skip_gran = (ragged_gran_env != NULL) ? atoi(ragged_gran_env) : 8;
#ifdef GGML_OPENCL_USE_ADRENO_BIN_KERNELS
// try loading adreno binary kernels if enabled
// if fails to load, builtin kernels will be used
@@ -19338,6 +19351,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &(backend_ctx->prealloc_total_tiles.buffer)));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_use_moe_ragged));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_moe_ragged_skip_gran));
// set thread grid
global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -19564,6 +19579,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &(backend_ctx->prealloc_total_tiles.buffer)));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_use_moe_ragged));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_moe_ragged_skip_gran));
// set thread grid
global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -19740,6 +19757,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &(backend_ctx->prealloc_total_tiles.buffer)));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_use_moe_ragged));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_moe_ragged_skip_gran));
// set thread grid
global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -19917,6 +19936,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &(backend_ctx->prealloc_total_tiles.buffer)));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_use_moe_ragged));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_moe_ragged_skip_gran));
// set thread grid
global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -20174,6 +20195,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &(backend_ctx->prealloc_total_tiles.buffer)));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_use_moe_ragged));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_moe_ragged_skip_gran));
// set thread grid
global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -20352,6 +20375,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &(backend_ctx->prealloc_total_tiles.buffer)));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_use_moe_ragged));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_moe_ragged_skip_gran));
// set thread grid
global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -20527,6 +20552,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &(backend_ctx->prealloc_total_tiles.buffer)));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_use_moe_ragged));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_moe_ragged_skip_gran));
// set thread grid
global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -20710,6 +20737,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &(backend_ctx->prealloc_total_tiles.buffer)));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_use_moe_ragged));
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint), &backend_ctx->adreno_moe_ragged_skip_gran));
// set thread grid
global_size[1] = static_cast<size_t>((ne01 + 63) / 64);