happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Run clang-format

This commit is contained in:
Nikhil Jain 2026-01-29 20:56:18 -08:00
parent edee0d6bc6
commit a962180095
No known key found for this signature in database
1 changed files with 91 additions and 93 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

184
ggml/src/ggml-webgpu/ggml-webgpu.cpp
View File

@ -152,7 +152,7 @@ struct webgpu_buf_pool {
// 2. For the per-thread buffer pools in webgpu_context,
// buffers are allocated and freed in Dawn callbacks,
// which can run on a different thread than the calling thread.
std::mutex mutex;
std::mutex mutex;
std::condition_variable cv;
void init(wgpu::Device device,
@ -366,7 +366,6 @@ struct webgpu_context_struct {
std::unordered_map<ggml_webgpu_pad_pipeline_key, webgpu_pipeline, ggml_webgpu_pad_pipeline_key_hash> pad_pipelines;
size_t memset_bytes_per_thread;
};
typedef std::shared_ptr<webgpu_context_struct> webgpu_context;
@ -388,9 +387,9 @@ struct ggml_backend_webgpu_device_context {
// Per-thread data required to actually run WebGPU operations in a backend instance
struct ggml_backend_webgpu_context {
webgpu_context webgpu_ctx;
std::once_flag init_once;
std::string name;
webgpu_context webgpu_ctx;
std::once_flag init_once;
std::string name;
};
// Per-thread data related to buffers
@ -866,17 +865,16 @@ static webgpu_command ggml_webgpu_pad(webgpu_context & ctx, ggml_tensor * src, g
};
webgpu_pipeline pipeline;
auto it = ctx->pad_pipelines.find(pipeline_key);
if (it != ctx->pad_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_pad_shader(ctx->p, wgsl_pad, shader_lib_ctx);
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
pipeline.context = processed.decisions;
ctx->pad_pipelines.emplace(pipeline_key, pipeline);
}
auto it = ctx->pad_pipelines.find(pipeline_key);
if (it != ctx->pad_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_processed_shader processed = ggml_webgpu_preprocess_pad_shader(ctx->p, wgsl_pad, shader_lib_ctx);
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
pipeline.context = processed.decisions;
ctx->pad_pipelines.emplace(pipeline_key, pipeline);
}
ggml_webgpu_generic_shader_decisions decisions =
*static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context);
@ -945,17 +943,17 @@ static std::optional<webgpu_command> ggml_webgpu_set_rows(webgpu_context & ctx,
};
webgpu_pipeline pipeline;
auto it = ctx->set_rows_pipelines.find(key);
if (it != ctx->set_rows_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_set_rows_shader(ctx->p, wgsl_set_rows, shader_lib_ctx);
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
pipeline.context = processed.decisions;
ctx->set_rows_pipelines.emplace(key, pipeline);
}
auto it = ctx->set_rows_pipelines.find(key);
if (it != ctx->set_rows_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_set_rows_shader(ctx->p, wgsl_set_rows, shader_lib_ctx);
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
pipeline.context = processed.decisions;
ctx->set_rows_pipelines.emplace(key, pipeline);
}
ggml_webgpu_generic_shader_decisions decisions =
*static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context);
@ -1258,26 +1256,26 @@ static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx,
};
webgpu_pipeline pipeline;
auto it = ctx->flash_attn_pipelines.find(key);
if (it != ctx->flash_attn_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_flash_attn_shader_lib_context shader_lib_ctx = {
.key = key,
.sg_mat_m = ctx->global_ctx->capabilities.sg_mat_m,
.sg_mat_n = ctx->global_ctx->capabilities.sg_mat_n,
.sg_mat_k = ctx->global_ctx->capabilities.sg_mat_k,
.wg_mem_limit_bytes = ctx->global_ctx->capabilities.limits.maxComputeWorkgroupStorageSize,
.max_subgroup_size = ctx->global_ctx->capabilities.max_subgroup_size
};
auto it = ctx->flash_attn_pipelines.find(key);
if (it != ctx->flash_attn_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_flash_attn_shader_lib_context shader_lib_ctx = {
.key = key,
.sg_mat_m = ctx->global_ctx->capabilities.sg_mat_m,
.sg_mat_n = ctx->global_ctx->capabilities.sg_mat_n,
.sg_mat_k = ctx->global_ctx->capabilities.sg_mat_k,
.wg_mem_limit_bytes = ctx->global_ctx->capabilities.limits.maxComputeWorkgroupStorageSize,
.max_subgroup_size = ctx->global_ctx->capabilities.max_subgroup_size
};
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_flash_attn_shader(ctx->p, wgsl_flash_attn, shader_lib_ctx);
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
pipeline.context = processed.decisions;
ctx->flash_attn_pipelines.emplace(key, pipeline);
}
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_flash_attn_shader(ctx->p, wgsl_flash_attn, shader_lib_ctx);
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
pipeline.context = processed.decisions;
ctx->flash_attn_pipelines.emplace(key, pipeline);
}
ggml_webgpu_flash_attn_shader_decisions decisions =
*static_cast<ggml_webgpu_flash_attn_shader_decisions *>(pipeline.context);
@ -1301,17 +1299,17 @@ static webgpu_command ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor * s
};
webgpu_pipeline pipeline;
auto it = ctx->unary_pipelines.find(pipeline_key);
if (it != ctx->unary_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_unary_shader(ctx->p, wgsl_unary, shader_lib_ctx);
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
pipeline.context = processed.decisions;
ctx->unary_pipelines.emplace(pipeline_key, pipeline);
}
auto it = ctx->unary_pipelines.find(pipeline_key);
if (it != ctx->unary_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_unary_shader(ctx->p, wgsl_unary, shader_lib_ctx);
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
pipeline.context = processed.decisions;
ctx->unary_pipelines.emplace(pipeline_key, pipeline);
}
ggml_webgpu_generic_shader_decisions decisions =
*static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context);
@ -1732,16 +1730,16 @@ static webgpu_command ggml_webgpu_argmax(webgpu_context & ctx, ggml_tensor * src
};
webgpu_pipeline pipeline;
auto it = ctx->argmax_pipelines.find(shader_lib_ctx.vec4);
if (it != ctx->argmax_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_generic_shader(ctx->p, wgsl_argmax, shader_lib_ctx, "argmax");
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
ctx->argmax_pipelines.emplace(shader_lib_ctx.vec4, pipeline);
}
auto it = ctx->argmax_pipelines.find(shader_lib_ctx.vec4);
if (it != ctx->argmax_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_generic_shader(ctx->p, wgsl_argmax, shader_lib_ctx, "argmax");
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
ctx->argmax_pipelines.emplace(shader_lib_ctx.vec4, pipeline);
}
uint32_t wg_x = ggml_nelements(dst);
return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
}
@ -1757,8 +1755,8 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr
.order = order
};
webgpu_pipeline argsort_pipeline;
auto it = ctx->argsort_pipelines.find(order);
webgpu_pipeline argsort_pipeline;
auto it = ctx->argsort_pipelines.find(order);
if (it != ctx->argsort_pipelines.end()) {
argsort_pipeline = it->second;
} else {
@ -1947,16 +1945,16 @@ static webgpu_command ggml_webgpu_cumsum(webgpu_context & ctx, ggml_tensor * src
.max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup,
};
webgpu_pipeline pipeline;
auto it = ctx->cumsum_pipelines.find(1);
if (it != ctx->cumsum_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_generic_shader(ctx->p, wgsl_cumsum, shader_lib_ctx, "cumsum");
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
ctx->cumsum_pipelines.emplace(1, pipeline);
}
auto it = ctx->cumsum_pipelines.find(1);
if (it != ctx->cumsum_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_generic_shader(ctx->p, wgsl_cumsum, shader_lib_ctx, "cumsum");
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
ctx->cumsum_pipelines.emplace(1, pipeline);
}
uint32_t wg_x = ggml_nrows(dst);
return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
}
@ -1989,16 +1987,16 @@ static webgpu_command ggml_webgpu_sum_rows(webgpu_context & ctx, ggml_tensor * s
};
webgpu_pipeline pipeline;
auto it = ctx->sum_rows_pipelines.find(1);
if (it != ctx->sum_rows_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_generic_shader(ctx->p, wgsl_sum_rows, shader_lib_ctx, "sum_rows");
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
ctx->sum_rows_pipelines.emplace(1, pipeline);
}
auto it = ctx->sum_rows_pipelines.find(1);
if (it != ctx->sum_rows_pipelines.end()) {
pipeline = it->second;
} else {
ggml_webgpu_processed_shader processed =
ggml_webgpu_preprocess_generic_shader(ctx->p, wgsl_sum_rows, shader_lib_ctx, "sum_rows");
pipeline =
ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
ctx->sum_rows_pipelines.emplace(1, pipeline);
}
uint32_t wg_x = total_sum ? 1 : ggml_nrows(dst);
return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
}
@ -2992,10 +2990,10 @@ static bool create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {
#ifdef GGML_WEBGPU_GPU_PROFILE
// Initialize buffer pool for timestamp queries, used for profiling
ctx->webgpu_global_ctx->timestamp_query_buf_pool.init(ctx->webgpu_global_ctx->device, WEBGPU_NUM_TIMESTAMP_QUERY_BUFS,
WEBGPU_TIMESTAMP_QUERY_BUF_SIZE_BYTES,
wgpu::BufferUsage::QueryResolve | wgpu::BufferUsage::CopySrc,
wgpu::BufferUsage::MapRead | wgpu::BufferUsage::CopyDst);
ctx->webgpu_global_ctx->timestamp_query_buf_pool.init(
ctx->webgpu_global_ctx->device, WEBGPU_NUM_TIMESTAMP_QUERY_BUFS, WEBGPU_TIMESTAMP_QUERY_BUF_SIZE_BYTES,
wgpu::BufferUsage::QueryResolve | wgpu::BufferUsage::CopySrc,
wgpu::BufferUsage::MapRead | wgpu::BufferUsage::CopyDst);
#endif
GGML_LOG_INFO(