nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 31 Packages Projects Releases Wiki Activity

Remove pipeline cache mutexes (#19195)

Browse Source 
* Remove mutex for pipeline caches, since they are now per-thread.

* Add comment

* Run clang-format

* Cleanup

* Run CI again

* Run CI once more

* Run clang-format
This commit is contained in:
Nikhil Jain
2026-02-01 18:47:29 -08:00
committed by GitHub
parent 3bc8d2cf23
commit 2dc3ce2166
1 changed files with 93 additions and 120 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-webgpu/ggml-webgpu.cpp
+11 -38
View File
@@ -146,8 +146,13 @@ struct webgpu_submission_futures {
struct webgpu_buf_pool {
std::vector<webgpu_pool_bufs> free;
// The pool must be synchronized because
// 1. The memset pool is shared globally by every ggml buffer,
// since allocating a pool per ggml buffer would consume too much memory.
// 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::condition_variable cv;
void init(wgpu::Device device,
@@ -266,7 +271,7 @@ struct webgpu_command {
#endif
};
struct webgpu_capabilities_base {
struct webgpu_capabilities {
wgpu::Limits limits;
bool supports_subgroup_matrix = false;
@@ -286,7 +291,7 @@ struct webgpu_global_context_struct {
wgpu::Device device;
wgpu::Queue queue;
webgpu_capabilities_base capabilities;
webgpu_capabilities capabilities;
// Shared buffer to move data from device to host
wgpu::Buffer get_tensor_staging_buf;
// Global mutex for pipeline and staging buffer, will be refactored to exclude pipeline caches.
@@ -361,7 +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;
@@ -384,7 +388,6 @@ 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;
};
@@ -861,21 +864,16 @@ static webgpu_command ggml_webgpu_pad(webgpu_context & ctx, ggml_tensor * src, g
};
webgpu_pipeline pipeline;
{
// TODO: remove guard once pipeline caches are per-thread
std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
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);
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);
@@ -944,9 +942,6 @@ static std::optional<webgpu_command> ggml_webgpu_set_rows(webgpu_context & ctx,
};
webgpu_pipeline pipeline;
// TODO: remove guard once pipeline caches are per-thread
{
std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
auto it = ctx->set_rows_pipelines.find(key);
if (it != ctx->set_rows_pipelines.end()) {
pipeline = it->second;
@@ -958,7 +953,6 @@ static std::optional<webgpu_command> ggml_webgpu_set_rows(webgpu_context & ctx,
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);
@@ -1261,9 +1255,6 @@ static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx,
};
webgpu_pipeline pipeline;
// TODO: remove guard once pipeline caches are per-thread
{
std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
auto it = ctx->flash_attn_pipelines.find(key);
if (it != ctx->flash_attn_pipelines.end()) {
pipeline = it->second;
@@ -1284,7 +1275,6 @@ static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx,
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);
@@ -1308,9 +1298,6 @@ static webgpu_command ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor * s
};
webgpu_pipeline pipeline;
{
// TODO: remove guard once pipeline caches are per-thread
std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
auto it = ctx->unary_pipelines.find(pipeline_key);
if (it != ctx->unary_pipelines.end()) {
pipeline = it->second;
@@ -1322,7 +1309,6 @@ static webgpu_command ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor * s
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);
@@ -1743,9 +1729,6 @@ static webgpu_command ggml_webgpu_argmax(webgpu_context & ctx, ggml_tensor * src
};
webgpu_pipeline pipeline;
{
// TODO: remove guard once pipeline caches are per-thread
std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
auto it = ctx->argmax_pipelines.find(shader_lib_ctx.vec4);
if (it != ctx->argmax_pipelines.end()) {
pipeline = it->second;
@@ -1756,7 +1739,6 @@ static webgpu_command ggml_webgpu_argmax(webgpu_context & ctx, ggml_tensor * src
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);
}
@@ -1772,7 +1754,6 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr
.order = order
};
std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
webgpu_pipeline argsort_pipeline;
auto it = ctx->argsort_pipelines.find(order);
if (it != ctx->argsort_pipelines.end()) {
@@ -1963,9 +1944,6 @@ static webgpu_command ggml_webgpu_cumsum(webgpu_context & ctx, ggml_tensor * src
.max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup,
};
webgpu_pipeline pipeline;
// TODO: remove guard once pipeline caches are per-thread
{
std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
auto it = ctx->cumsum_pipelines.find(1);
if (it != ctx->cumsum_pipelines.end()) {
pipeline = it->second;
@@ -1976,7 +1954,6 @@ static webgpu_command ggml_webgpu_cumsum(webgpu_context & ctx, ggml_tensor * src
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);
}
@@ -2009,9 +1986,6 @@ static webgpu_command ggml_webgpu_sum_rows(webgpu_context & ctx, ggml_tensor * s
};
webgpu_pipeline pipeline;
{
// TODO: remove guard once pipeline caches are per-thread
std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
auto it = ctx->sum_rows_pipelines.find(1);
if (it != ctx->sum_rows_pipelines.end()) {
pipeline = it->second;
@@ -2022,7 +1996,6 @@ static webgpu_command ggml_webgpu_sum_rows(webgpu_context & ctx, ggml_tensor * s
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);
}
@@ -3016,8 +2989,8 @@ 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,
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