Skip to content

Commit eb7a1bf

Browse files
edgchen1edgchen1
authored
[WebGPU EP] Reduce forward declaration boilerplate in kernel registration (#27977)
### Description <!-- Describe your changes. --> Update WebGPU EP kernel registration to remove redundant forward declarations. E.g., change from this: ```c++ class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, Attention); static const BuildKernelCreateInfoFn function_table[] = { ... BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, Attention)>, ``` to this: ```c++ static const BuildKernelCreateInfoFn function_table[] = { ... BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, Attention)>, ``` One subtlety is that the function table needs to be moved outside of the function now, since function-local (block scope) forward declarations would introduce names with no linkage. Those names should have external linkage. We already do something similar here: https://github.com/microsoft/onnxruntime/blob/aaa49440491a88f812ba79355b18457c8fa9ed7c/onnxruntime/test/autoep/library/example_plugin_ep_kernel_registry/ep_kernel_registration.cc#L12 ### Motivation and Context <!-- - Why is this change required? What problem does it solve? - If it fixes an open issue, please link to the issue here. --> Reduce boilerplate.
1 parent 0e73535 commit eb7a1bf

File tree

2 files changed

+388
-777
lines changed

2 files changed

+388
-777
lines changed

‎onnxruntime/contrib_ops/webgpu/webgpu_contrib_kernels.cc‎

Lines changed: 25 additions & 49 deletions
Original file line numberDiff line numberDiff line change
@@ -12,62 +12,38 @@ namespace onnxruntime {
1212
namespace contrib {
1313
namespace webgpu {
1414

15-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, Attention);
16-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, BiasAdd);
17-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, CausalConvWithState);
18-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, BiasGelu);
19-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, BiasSplitGelu);
20-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, FastGelu);
21-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, FusedConv);
22-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, GatherBlockQuantized);
23-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, Gelu);
24-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, GroupQueryAttention);
25-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, LinearAttention);
26-
// LayerNormalization used to be a contrib op that (incorrectly) used kOnnxDomain so we need to version it
27-
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kOnnxDomain, 1, 16, LayerNormalization);
28-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, MatMulNBits);
29-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, MultiHeadAttention);
30-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, QuickGelu);
31-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, RotaryEmbedding);
32-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kOnnxDomain, 1, SimplifiedLayerNormalization);
33-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, SkipLayerNormalization);
34-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kOnnxDomain, 1, SimplifiedLayerNormalization);
35-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, SkipSimplifiedLayerNormalization);
36-
// class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, MoE);
37-
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, QMoE);
38-
3915
template <>
4016
KernelCreateInfo BuildKernelCreateInfo<void>() {
4117
KernelCreateInfo info;
4218
return info;
4319
}
4420

45-
Status RegisterWebGpuContribKernels(KernelRegistry& kernel_registry, bool enable_graph_capture) {
46-
static const BuildKernelCreateInfoFn function_table[] = {
47-
BuildKernelCreateInfo<void>, // default entry to avoid the list become empty after ops-reducing
48-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, Attention)>,
49-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, BiasAdd)>,
50-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, CausalConvWithState)>,
51-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, BiasGelu)>,
52-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, BiasSplitGelu)>,
53-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, FastGelu)>,
54-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, GatherBlockQuantized)>,
55-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, FusedConv)>,
56-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, Gelu)>,
57-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, LinearAttention)>,
58-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, MatMulNBits)>,
59-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, MultiHeadAttention)>,
60-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, QuickGelu)>,
61-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, RotaryEmbedding)>,
62-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, SkipLayerNormalization)>,
63-
// LayerNormalization used to be a contrib op that (incorrectly) used kOnnxDomain so we need to version it
64-
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kOnnxDomain, 1, 16, LayerNormalization)>,
65-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kOnnxDomain, 1, SimplifiedLayerNormalization)>,
66-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, SkipSimplifiedLayerNormalization)>,
67-
// BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, MoE)>,
68-
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, QMoE)>};
21+
static const BuildKernelCreateInfoFn build_kernel_create_info_function_table[] = {
22+
BuildKernelCreateInfo<void>, // default entry to avoid the list become empty after ops-reducing
23+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, Attention)>,
24+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, BiasAdd)>,
25+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, CausalConvWithState)>,
26+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, BiasGelu)>,
27+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, BiasSplitGelu)>,
28+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, FastGelu)>,
29+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, GatherBlockQuantized)>,
30+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, FusedConv)>,
31+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, Gelu)>,
32+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, LinearAttention)>,
33+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, MatMulNBits)>,
34+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, MultiHeadAttention)>,
35+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, QuickGelu)>,
36+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, RotaryEmbedding)>,
37+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, SkipLayerNormalization)>,
38+
// LayerNormalization used to be a contrib op that (incorrectly) used kOnnxDomain so we need to version it
39+
BuildKernelCreateInfo<class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kOnnxDomain, 1, 16, LayerNormalization)>,
40+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kOnnxDomain, 1, SimplifiedLayerNormalization)>,
41+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, SkipSimplifiedLayerNormalization)>,
42+
// BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, MoE)>,
43+
BuildKernelCreateInfo<class ONNX_OPERATOR_KERNEL_CLASS_NAME(kWebGpuExecutionProvider, kMSDomain, 1, QMoE)>};
6944

70-
for (auto& function_table_entry : function_table) {
45+
Status RegisterWebGpuContribKernels(KernelRegistry& kernel_registry, bool enable_graph_capture) {
46+
for (auto& function_table_entry : build_kernel_create_info_function_table) {
7147
KernelCreateInfo info = function_table_entry();
7248
if (info.kernel_def != nullptr) { // filter disabled entries where type is void
7349
ORT_RETURN_IF_ERROR(kernel_registry.Register(std::move(info)));

0 commit comments

Comments
 (0)