thek0tyara/llama-cpp-turboquant
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 4

opencl: allow large buffer for adreno (#20997)

Browse source
This commit is contained in:
lhez 2026-03-26 08:52:21 -07:00 • committed by GitHub
parent f8d4abae86
commit ded446b34c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 26 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

26
ggml/src/ggml-opencl/ggml-opencl.cpp
View file

@ -394,6 +394,9 @@ struct ggml_backend_opencl_context {
bool fp16_support; bool fp16_support;
bool has_vector_subgroup_broadcast; bool has_vector_subgroup_broadcast;
bool disable_fusion; bool disable_fusion;
bool adreno_has_large_buffer;
bool adreno_use_large_buffer;
ggml_cl_compiler_version adreno_cl_compiler_version; ggml_cl_compiler_version adreno_cl_compiler_version;
int adreno_wave_size; int adreno_wave_size;
@ -787,6 +790,10 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
" -cl-mad-enable -cl-unsafe-math-optimizations" " -cl-mad-enable -cl-unsafe-math-optimizations"
" -cl-finite-math-only -cl-fast-relaxed-math"; " -cl-finite-math-only -cl-fast-relaxed-math";
if (backend_ctx->adreno_use_large_buffer) {
compile_opts += " -qcom-enable-large-buffer ";
}
GGML_LOG_INFO("ggml_opencl: loading OpenCL kernels"); GGML_LOG_INFO("ggml_opencl: loading OpenCL kernels");
// add // add
@ -3020,6 +3027,8 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
// Check if ext_buffer contains cl_khr_fp16 // Check if ext_buffer contains cl_khr_fp16
backend_ctx->fp16_support = strstr(ext_buffer, "cl_khr_fp16") != NULL; backend_ctx->fp16_support = strstr(ext_buffer, "cl_khr_fp16") != NULL;
GGML_LOG_INFO("ggml_opencl: device FP16 support: %s\n", backend_ctx->fp16_support ? "true" : "false"); GGML_LOG_INFO("ggml_opencl: device FP16 support: %s\n", backend_ctx->fp16_support ? "true" : "false");
// check Adreno large buffer support
backend_ctx->adreno_has_large_buffer = strstr(ext_buffer, "cl_qcom_large_buffer") != NULL;
// fp16 is required // fp16 is required
if (!backend_ctx->fp16_support) { if (!backend_ctx->fp16_support) {
@ -3086,6 +3095,18 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
GGML_LOG_INFO("ggml_opencl: using kernels optimized for Adreno (GGML_OPENCL_USE_ADRENO_KERNELS)\n"); GGML_LOG_INFO("ggml_opencl: using kernels optimized for Adreno (GGML_OPENCL_USE_ADRENO_KERNELS)\n");
#endif // GGML_OPENCL_USE_ADRENO_KERNELS #endif // GGML_OPENCL_USE_ADRENO_KERNELS
// determine whether to use large buffer for Adreno
backend_ctx->adreno_use_large_buffer = getenv("GGML_OPENCL_ADRENO_USE_LARGE_BUFFER") != nullptr &&
backend_ctx->gpu_family == GPU_FAMILY::ADRENO;
if (backend_ctx->adreno_use_large_buffer) {
if (!backend_ctx->adreno_has_large_buffer) {
GGML_LOG_INFO("ggml_opencl: Adreno large buffer requested but not supported by driver, will use regular buffer\n");
backend_ctx->adreno_use_large_buffer = false;
} else {
GGML_LOG_INFO("ggml_opencl: Adreno large buffer enabled\n");
}
}
cl_int err; cl_int err;
// A local ref of cl_context for convenience // A local ref of cl_context for convenience
@ -5660,6 +5681,11 @@ static ggml_backend_buffer_t ggml_backend_opencl_buffer_type_alloc_buffer(ggml_b
cl_int err; cl_int err;
cl_mem mem = clCreateBuffer(backend_ctx->context, CL_MEM_READ_WRITE, size, NULL, &err); cl_mem mem = clCreateBuffer(backend_ctx->context, CL_MEM_READ_WRITE, size, NULL, &err);
if (err != CL_SUCCESS && backend_ctx->adreno_use_large_buffer) {
cl_mem_properties props[] = { 0x41A6 /* CL_LARGE_BUFFER_QCOM */, 1, 0 };
mem = clCreateBufferWithProperties(backend_ctx->context, props, CL_MEM_READ_WRITE, size, NULL, &err);
}
if (err != CL_SUCCESS) { if (err != CL_SUCCESS) {
GGML_LOG_INFO("%s: failed to allocate %.2f MiB\n", __func__, size / 1024.0 / 1024.0); GGML_LOG_INFO("%s: failed to allocate %.2f MiB\n", __func__, size / 1024.0 / 1024.0);
return nullptr; return nullptr;