Skip to content

Commit b566811

Browse files
reeselevinereeselevine
committed
Add buffer label and enable dawn-specific toggles to turn off some checks
1 parent 74b8fc1 commit b566811

File tree

1 file changed

+40
-16
lines changed

1 file changed

+40
-16
lines changed

‎ggml/src/ggml-webgpu/ggml-webgpu.cpp‎

Lines changed: 40 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -309,10 +309,12 @@ struct ggml_backend_webgpu_context {
309309
struct ggml_backend_webgpu_buffer_context {
310310
webgpu_context webgpu_ctx;
311311
wgpu::Buffer buffer;
312+
std::string label;
312313

313-
ggml_backend_webgpu_buffer_context(webgpu_context ctx, wgpu::Buffer buf) :
314+
ggml_backend_webgpu_buffer_context(webgpu_context ctx, wgpu::Buffer buf, std::string lbl) :
314315
webgpu_ctx(std::move(ctx)),
315-
buffer(std::move(buf)) {}
316+
buffer(std::move(buf)),
317+
label(std::move(lbl)) {}
316318
};
317319

318320
/* End struct definitions */
@@ -1336,11 +1338,11 @@ static void ggml_backend_webgpu_buffer_memset_tensor(ggml_backend_buffer_t buffe
13361338

13371339
WEBGPU_CPU_PROFILE_TOTAL_START(memset_tensor);
13381340

1339-
WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_memset_tensor(" << buffer << ", " << tensor << ", " << value << ", "
1340-
<< offset << ", " << size << ")");
1341-
13421341
ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
13431342

1343+
WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_memset_tensor(" << buf_ctx->label << ", " << tensor << ", " << value
1344+
<< ", " << offset << ", " << size << ")");
1345+
13441346
size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
13451347

13461348
// This is a trick to set all bytes of a u32 to the same 1 byte value.
@@ -1354,12 +1356,13 @@ static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer,
13541356
const void * data,
13551357
size_t offset,
13561358
size_t size) {
1357-
WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", "
1358-
<< offset << ", " << size << ")");
13591359
WEBGPU_CPU_PROFILE_TOTAL_START(set_tensor);
13601360
ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
13611361
webgpu_context webgpu_ctx = buf_ctx->webgpu_ctx;
13621362

1363+
WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_set_tensor(" << buf_ctx->label << ", " << tensor << ", " << data
1364+
<< ", " << offset << ", " << size << ")");
1365+
13631366
size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
13641367

13651368
webgpu_ctx->queue.WriteBuffer(buf_ctx->buffer, total_offset, data, (size / 4) * 4);
@@ -1397,12 +1400,12 @@ static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer,
13971400
void * data,
13981401
size_t offset,
13991402
size_t size) {
1400-
WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", "
1401-
<< offset << ", " << size << ")");
14021403
WEBGPU_CPU_PROFILE_TOTAL_START(get_tensor);
1403-
ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
1404-
webgpu_context webgpu_ctx = buf_ctx->webgpu_ctx;
1405-
wgpu::Device device = webgpu_ctx->device;
1404+
ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
1405+
WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_get_tensor(" << buf_ctx->label << ", " << tensor << ", " << data
1406+
<< ", " << offset << ", " << size << ")");
1407+
webgpu_context webgpu_ctx = buf_ctx->webgpu_ctx;
1408+
wgpu::Device device = webgpu_ctx->device;
14061409

14071410
size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
14081411

@@ -1473,16 +1476,20 @@ static const char * ggml_backend_webgpu_buffer_type_get_name(ggml_backend_buffer
14731476

14741477
static ggml_backend_buffer_t ggml_backend_webgpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
14751478
size_t size) {
1476-
WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_type_alloc_buffer(" << size << ")");
1479+
static std::atomic<int> buffer_count;
1480+
int buffer_id = buffer_count++;
1481+
std::string buf_name = "tensor_buf" + std::to_string(buffer_id);
1482+
WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_type_alloc_buffer_" << buffer_id << ": " << size << " bytes");
14771483
ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
14781484

14791485
wgpu::Buffer buf;
14801486
ggml_webgpu_create_buffer(ctx->webgpu_ctx->device, buf,
14811487
(size + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) & ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1),
14821488
wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::CopyDst,
1483-
"allocated_buffer");
1489+
buf_name.c_str());
14841490

1485-
ggml_backend_webgpu_buffer_context * buf_ctx = new ggml_backend_webgpu_buffer_context(ctx->webgpu_ctx, buf);
1491+
ggml_backend_webgpu_buffer_context * buf_ctx =
1492+
new ggml_backend_webgpu_buffer_context(ctx->webgpu_ctx, buf, buf_name);
14861493

14871494
return ggml_backend_buffer_init(buft, ggml_backend_webgpu_buffer_interface, buf_ctx, size);
14881495
}
@@ -2129,6 +2136,15 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
21292136
required_features.push_back(wgpu::FeatureName::TimestampQuery);
21302137
#endif
21312138

2139+
const char * const deviceEnabledToggles[] = { "skip_validation", "disable_robustness", "disable_workgroup_init",
2140+
"disable_polyfills_on_integer_div_and_mod" };
2141+
const char * const deviceDisabledToggles[] = { "timestamp_quantization" };
2142+
wgpu::DawnTogglesDescriptor deviceTogglesDesc;
2143+
deviceTogglesDesc.enabledToggles = deviceEnabledToggles;
2144+
deviceTogglesDesc.enabledToggleCount = 4;
2145+
deviceTogglesDesc.disabledToggles = deviceDisabledToggles;
2146+
deviceTogglesDesc.disabledToggleCount = 1;
2147+
21322148
wgpu::DeviceDescriptor dev_desc;
21332149
dev_desc.requiredLimits = &ctx->limits;
21342150
dev_desc.requiredFeatures = required_features.data();
@@ -2146,6 +2162,7 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
21462162
GGML_ABORT("ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast<int>(reason),
21472163
std::string(message).c_str());
21482164
});
2165+
dev_desc.nextInChain = &deviceTogglesDesc;
21492166
ctx->instance.WaitAny(ctx->adapter.RequestDevice(
21502167
&dev_desc, wgpu::CallbackMode::AllowSpontaneous,
21512168
[ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) {
@@ -2243,11 +2260,18 @@ ggml_backend_reg_t ggml_backend_webgpu_reg() {
22432260
ctx.name = GGML_WEBGPU_NAME;
22442261
ctx.device_count = 1;
22452262

2263+
const char * const instanceEnabledToggles[] = { "allow_unsafe_apis" };
2264+
2265+
wgpu::DawnTogglesDescriptor instanceTogglesDesc;
2266+
instanceTogglesDesc.enabledToggles = instanceEnabledToggles;
2267+
instanceTogglesDesc.enabledToggleCount = 1;
22462268
wgpu::InstanceDescriptor instance_descriptor{};
22472269
std::vector<wgpu::InstanceFeatureName> instance_features = { wgpu::InstanceFeatureName::TimedWaitAny };
22482270
instance_descriptor.requiredFeatures = instance_features.data();
22492271
instance_descriptor.requiredFeatureCount = instance_features.size();
2250-
webgpu_ctx->instance = wgpu::CreateInstance(&instance_descriptor);
2272+
instance_descriptor.nextInChain = &instanceTogglesDesc;
2273+
2274+
webgpu_ctx->instance = wgpu::CreateInstance(&instance_descriptor);
22512275
GGML_ASSERT(webgpu_ctx->instance != nullptr);
22522276

22532277
static ggml_backend_reg reg = {

0 commit comments

Comments
 (0)