Skip to content

[ExecuTorch][WebGPU] Add split_with_sizes_copy op #21281

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
JCNTH merged 3 commits into from
Jul 23, 2026
Merged

[ExecuTorch][WebGPU] Add split_with_sizes_copy op #21281

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
df815e8
[ExecuTorch][WebGPU] Add split_with_sizes_copy op
JCNTH Jul 23, 2026
96b81b0
[ExecuTorch][WebGPU] split_with_sizes_copy op tests
JCNTH Jul 23, 2026
1721a82
[ExecuTorch][WebGPU] Migrate 14 single-pipeline ops to make_compute_p…
JCNTH Jul 23, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
70 changes: 14 additions & 56 deletions backends/webgpu/runtime/ops/boolean_op/BooleanOp.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,61 +84,22 @@ void dispatch_bool_op(
utils::make_uniform(device, &params, sizeof(BoolOpParams));
graph.add_uniform_buffer_bytes(sizeof(BoolOpParams));

WGPUShaderSourceWGSL wgsl_desc = {};
wgsl_desc.chain.sType = WGPUSType_ShaderSourceWGSL;
wgsl_desc.code = {wgsl, WGPU_STRLEN};
WGPUShaderModuleDescriptor shader_desc = {};
shader_desc.nextInChain = &wgsl_desc.chain;
WGPUShaderModule shader = wgpuDeviceCreateShaderModule(device, &shader_desc);

WGPUBindGroupLayoutEntry entries[3] = {};
entries[0].buffer.type = WGPUBufferBindingType_ReadOnlyStorage;
entries[1].buffer.type = WGPUBufferBindingType_Storage;
entries[2].buffer.type = WGPUBufferBindingType_Uniform;
for (uint32_t i = 0; i < 3; i++) {
entries[i].binding = i;
entries[i].visibility = WGPUShaderStage_Compute;
}

WGPUBindGroupLayoutDescriptor bgl_desc = {};
bgl_desc.entryCount = 3;
bgl_desc.entries = entries;
WGPUBindGroupLayout bgl = wgpuDeviceCreateBindGroupLayout(device, &bgl_desc);

WGPUPipelineLayoutDescriptor pl_desc = {};
pl_desc.bindGroupLayoutCount = 1;
pl_desc.bindGroupLayouts = &bgl;
WGPUPipelineLayout pipeline_layout =
wgpuDeviceCreatePipelineLayout(device, &pl_desc);

WGPUComputePipelineDescriptor pipeline_desc = {};
pipeline_desc.layout = pipeline_layout;
pipeline_desc.compute.module = shader;
pipeline_desc.compute.entryPoint = {"main", WGPU_STRLEN};
pipeline_desc.compute.constantCount = 1;
pipeline_desc.compute.constants = &wg_size_constant;
WGPUComputePipeline pipeline =
wgpuDeviceCreateComputePipeline(device, &pipeline_desc);

WGPUBindGroupEntry bg_entries[3] = {};
bg_entries[0].binding = 0;
bg_entries[0].buffer = self_tensor.buffer;
bg_entries[0].size = self_bind_size;
bg_entries[1].binding = 1;
bg_entries[1].buffer = out_tensor.buffer;
bg_entries[1].size = out_bind_size;
bg_entries[2].binding = 2;
bg_entries[2].buffer = params_buf;
bg_entries[2].size = sizeof(BoolOpParams);

WGPUBindGroupDescriptor bg_desc = {};
bg_desc.layout = bgl;
bg_desc.entryCount = 3;
bg_desc.entries = bg_entries;
WGPUBindGroup bind_group = wgpuDeviceCreateBindGroup(device, &bg_desc);
utils::ComputePipelineBundle bundle = utils::make_compute_pipeline(
device,
wgsl,
{
{0,
WGPUBufferBindingType_ReadOnlyStorage,
self_tensor.buffer,
self_bind_size},
{1, WGPUBufferBindingType_Storage, out_tensor.buffer, out_bind_size},
{2, WGPUBufferBindingType_Uniform, params_buf, sizeof(BoolOpParams)},
},
&wg_size_constant,
1);

const size_t dispatch_idx =
graph.add_dispatch({pipeline, bind_group, workgroup_count});
graph.add_dispatch({bundle.pipeline, bundle.bind_group, workgroup_count});

WGPUBuffer p_buf = params_buf;
auto resize =
Expand All @@ -158,9 +119,6 @@ void dispatch_bool_op(
};
graph.add_tensor_resize_hook(self_id, resize);

wgpuShaderModuleRelease(shader);
wgpuBindGroupLayoutRelease(bgl);
wgpuPipelineLayoutRelease(pipeline_layout);
graph.own_uniform_buffer(params_buf);
}

Expand Down
116 changes: 27 additions & 89 deletions backends/webgpu/runtime/ops/div/BinaryOp.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,94 +78,35 @@ void div_impl(WebGPUGraph& graph, const std::vector<int>& args) {
utils::make_uniform(device, &in2_meta, sizeof(TensorMeta));
graph.add_uniform_buffer_bytes(3 * sizeof(TensorMeta));

WGPUShaderSourceWGSL wgsl_desc = {};
wgsl_desc.chain.sType = WGPUSType_ShaderSourceWGSL;
wgsl_desc.code = {kBinaryDivWGSL, WGPU_STRLEN};

WGPUShaderModuleDescriptor shader_desc = {};
shader_desc.nextInChain = &wgsl_desc.chain;
WGPUShaderModule shader = wgpuDeviceCreateShaderModule(device, &shader_desc);

WGPUBindGroupLayoutEntry entries[6] = {};

entries[0].binding = 0;
entries[0].visibility = WGPUShaderStage_Compute;
entries[0].buffer.type = WGPUBufferBindingType_ReadOnlyStorage;

entries[1].binding = 1;
entries[1].visibility = WGPUShaderStage_Compute;
entries[1].buffer.type = WGPUBufferBindingType_ReadOnlyStorage;

entries[2].binding = 2;
entries[2].visibility = WGPUShaderStage_Compute;
entries[2].buffer.type = WGPUBufferBindingType_Storage;

entries[3].binding = 3;
entries[3].visibility = WGPUShaderStage_Compute;
entries[3].buffer.type = WGPUBufferBindingType_Uniform;

entries[4].binding = 4;
entries[4].visibility = WGPUShaderStage_Compute;
entries[4].buffer.type = WGPUBufferBindingType_Uniform;

entries[5].binding = 5;
entries[5].visibility = WGPUShaderStage_Compute;
entries[5].buffer.type = WGPUBufferBindingType_Uniform;

WGPUBindGroupLayoutDescriptor bgl_desc = {};
bgl_desc.entryCount = 6;
bgl_desc.entries = entries;
WGPUBindGroupLayout bgl = wgpuDeviceCreateBindGroupLayout(device, &bgl_desc);

WGPUPipelineLayoutDescriptor pl_desc = {};
pl_desc.bindGroupLayoutCount = 1;
pl_desc.bindGroupLayouts = &bgl;
WGPUPipelineLayout pipeline_layout =
wgpuDeviceCreatePipelineLayout(device, &pl_desc);

WGPUComputePipelineDescriptor pipeline_desc = {};
pipeline_desc.layout = pipeline_layout;
pipeline_desc.compute.module = shader;
pipeline_desc.compute.entryPoint = {"main", WGPU_STRLEN};
pipeline_desc.compute.constantCount = 1;
pipeline_desc.compute.constants = &wg_size_constant;
WGPUComputePipeline pipeline =
wgpuDeviceCreateComputePipeline(device, &pipeline_desc);

WGPUBindGroupEntry bg_entries[6] = {};

bg_entries[0].binding = 0;
bg_entries[0].buffer = in1_tensor.buffer;
bg_entries[0].size = in1_tensor.nbytes;

bg_entries[1].binding = 1;
bg_entries[1].buffer = in2_tensor.buffer;
bg_entries[1].size = in2_tensor.nbytes;

bg_entries[2].binding = 2;
bg_entries[2].buffer = out_tensor.buffer;
bg_entries[2].size = out_tensor.nbytes;

bg_entries[3].binding = 3;
bg_entries[3].buffer = out_meta_buf;
bg_entries[3].size = sizeof(TensorMeta);

bg_entries[4].binding = 4;
bg_entries[4].buffer = in1_meta_buf;
bg_entries[4].size = sizeof(TensorMeta);

bg_entries[5].binding = 5;
bg_entries[5].buffer = in2_meta_buf;
bg_entries[5].size = sizeof(TensorMeta);

WGPUBindGroupDescriptor bg_desc = {};
bg_desc.layout = bgl;
bg_desc.entryCount = 6;
bg_desc.entries = bg_entries;
WGPUBindGroup bind_group = wgpuDeviceCreateBindGroup(device, &bg_desc);
utils::ComputePipelineBundle bundle = utils::make_compute_pipeline(
device,
kBinaryDivWGSL,
{
{0,
WGPUBufferBindingType_ReadOnlyStorage,
in1_tensor.buffer,
in1_tensor.nbytes},
{1,
WGPUBufferBindingType_ReadOnlyStorage,
in2_tensor.buffer,
in2_tensor.nbytes},
{2,
WGPUBufferBindingType_Storage,
out_tensor.buffer,
out_tensor.nbytes},
{3, WGPUBufferBindingType_Uniform, out_meta_buf, sizeof(TensorMeta)},
{4, WGPUBufferBindingType_Uniform, in1_meta_buf, sizeof(TensorMeta)},
{5, WGPUBufferBindingType_Uniform, in2_meta_buf, sizeof(TensorMeta)},
},
&wg_size_constant,
1);

const size_t dispatch_idx = graph.add_dispatch(
{pipeline, bind_group, workgroup_count.x, "div", workgroup_count.y});
{bundle.pipeline,
bundle.bind_group,
workgroup_count.x,
"div",
workgroup_count.y});

// Dynamic shapes: rebuild all 3 broadcast TensorMeta UBOs + dispatch.
WGPUBuffer o_buf = out_meta_buf, a_buf = in1_meta_buf, b_buf = in2_meta_buf;
Expand Down Expand Up @@ -206,9 +147,6 @@ void div_impl(WebGPUGraph& graph, const std::vector<int>& args) {
graph.add_tensor_resize_hook(in1_id, div_resize);
graph.add_tensor_resize_hook(in2_id, div_resize);

wgpuShaderModuleRelease(shader);
wgpuBindGroupLayoutRelease(bgl);
wgpuPipelineLayoutRelease(pipeline_layout);
// Graph owns them so a resize hook can rewrite them; freed in the dtor.
graph.own_uniform_buffer(out_meta_buf);
graph.own_uniform_buffer(in1_meta_buf);
Expand Down
86 changes: 20 additions & 66 deletions backends/webgpu/runtime/ops/expand_copy/ExpandCopy.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,72 +60,26 @@ void expand_copy_impl(WebGPUGraph& graph, const std::vector<int>& args) {
utils::make_uniform(device, &in_meta, sizeof(TensorMeta));
graph.add_uniform_buffer_bytes(2 * sizeof(TensorMeta));

WGPUShaderSourceWGSL wgsl_desc = {};
wgsl_desc.chain.sType = WGPUSType_ShaderSourceWGSL;
wgsl_desc.code = {kExpandCopyWGSL, WGPU_STRLEN};
WGPUShaderModuleDescriptor shader_desc = {};
shader_desc.nextInChain = &wgsl_desc.chain;
WGPUShaderModule shader = wgpuDeviceCreateShaderModule(device, &shader_desc);

WGPUBindGroupLayoutEntry entries[4] = {};
entries[0].binding = 0;
entries[0].visibility = WGPUShaderStage_Compute;
entries[0].buffer.type = WGPUBufferBindingType_ReadOnlyStorage;
entries[1].binding = 1;
entries[1].visibility = WGPUShaderStage_Compute;
entries[1].buffer.type = WGPUBufferBindingType_Storage;
entries[2].binding = 2;
entries[2].visibility = WGPUShaderStage_Compute;
entries[2].buffer.type = WGPUBufferBindingType_Uniform;
entries[3].binding = 3;
entries[3].visibility = WGPUShaderStage_Compute;
entries[3].buffer.type = WGPUBufferBindingType_Uniform;

WGPUBindGroupLayoutDescriptor bgl_desc = {};
bgl_desc.entryCount = 4;
bgl_desc.entries = entries;
WGPUBindGroupLayout bgl = wgpuDeviceCreateBindGroupLayout(device, &bgl_desc);

WGPUPipelineLayoutDescriptor pl_desc = {};
pl_desc.bindGroupLayoutCount = 1;
pl_desc.bindGroupLayouts = &bgl;
WGPUPipelineLayout pipeline_layout =
wgpuDeviceCreatePipelineLayout(device, &pl_desc);

WGPUComputePipelineDescriptor pipeline_desc = {};
pipeline_desc.layout = pipeline_layout;
pipeline_desc.compute.module = shader;
pipeline_desc.compute.entryPoint = {"main", WGPU_STRLEN};
pipeline_desc.compute.constantCount = 1;
pipeline_desc.compute.constants = &wg_size_constant;
WGPUComputePipeline pipeline =
wgpuDeviceCreateComputePipeline(device, &pipeline_desc);

WGPUBindGroupEntry bg_entries[4] = {};
bg_entries[0].binding = 0;
bg_entries[0].buffer = in_tensor.buffer;
bg_entries[0].size = in_tensor.nbytes;
bg_entries[1].binding = 1;
bg_entries[1].buffer = out_tensor.buffer;
bg_entries[1].size = out_tensor.nbytes;
bg_entries[2].binding = 2;
bg_entries[2].buffer = out_meta_buf;
bg_entries[2].size = sizeof(TensorMeta);
bg_entries[3].binding = 3;
bg_entries[3].buffer = in_meta_buf;
bg_entries[3].size = sizeof(TensorMeta);

WGPUBindGroupDescriptor bg_desc = {};
bg_desc.layout = bgl;
bg_desc.entryCount = 4;
bg_desc.entries = bg_entries;
WGPUBindGroup bind_group = wgpuDeviceCreateBindGroup(device, &bg_desc);

graph.add_dispatch({pipeline, bind_group, workgroup_count});

wgpuShaderModuleRelease(shader);
wgpuBindGroupLayoutRelease(bgl);
wgpuPipelineLayoutRelease(pipeline_layout);
utils::ComputePipelineBundle bundle = utils::make_compute_pipeline(
device,
kExpandCopyWGSL,
{
{0,
WGPUBufferBindingType_ReadOnlyStorage,
in_tensor.buffer,
in_tensor.nbytes},
{1,
WGPUBufferBindingType_Storage,
out_tensor.buffer,
out_tensor.nbytes},
{2, WGPUBufferBindingType_Uniform, out_meta_buf, sizeof(TensorMeta)},
{3, WGPUBufferBindingType_Uniform, in_meta_buf, sizeof(TensorMeta)},
},
&wg_size_constant,
1);

graph.add_dispatch({bundle.pipeline, bundle.bind_group, workgroup_count});

wgpuBufferRelease(out_meta_buf);
wgpuBufferRelease(in_meta_buf);
}
Expand Down
72 changes: 20 additions & 52 deletions backends/webgpu/runtime/ops/fill/Fill.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,57 +67,28 @@ void add_fill(
utils::make_uniform(device, &params, sizeof(FillParams));
graph.add_uniform_buffer_bytes(sizeof(FillParams));

WGPUShaderSourceWGSL wgsl_desc = {};
wgsl_desc.chain.sType = WGPUSType_ShaderSourceWGSL;
wgsl_desc.code = {kFillWGSL, WGPU_STRLEN};
WGPUShaderModuleDescriptor shader_desc = {};
shader_desc.nextInChain = &wgsl_desc.chain;
WGPUShaderModule shader = wgpuDeviceCreateShaderModule(device, &shader_desc);

WGPUBindGroupLayoutEntry entries[2] = {};
entries[0].binding = 0;
entries[0].visibility = WGPUShaderStage_Compute;
entries[0].buffer.type = WGPUBufferBindingType_Storage;
entries[1].binding = 1;
entries[1].visibility = WGPUShaderStage_Compute;
entries[1].buffer.type = WGPUBufferBindingType_Uniform;

WGPUBindGroupLayoutDescriptor bgl_desc = {};
bgl_desc.entryCount = 2;
bgl_desc.entries = entries;
WGPUBindGroupLayout bgl = wgpuDeviceCreateBindGroupLayout(device, &bgl_desc);

WGPUPipelineLayoutDescriptor pl_desc = {};
pl_desc.bindGroupLayoutCount = 1;
pl_desc.bindGroupLayouts = &bgl;
WGPUPipelineLayout pipeline_layout =
wgpuDeviceCreatePipelineLayout(device, &pl_desc);

WGPUComputePipelineDescriptor pipeline_desc = {};
pipeline_desc.layout = pipeline_layout;
pipeline_desc.compute.module = shader;
pipeline_desc.compute.entryPoint = {"main", WGPU_STRLEN};
pipeline_desc.compute.constantCount = 1;
pipeline_desc.compute.constants = &wg_size_constant;
WGPUComputePipeline pipeline =
wgpuDeviceCreateComputePipeline(device, &pipeline_desc);

WGPUBindGroupEntry bg_entries[2] = {};
bg_entries[0].binding = 0;
bg_entries[0].buffer = out_tensor.buffer;
bg_entries[0].size = out_tensor.nbytes;
bg_entries[1].binding = 1;
bg_entries[1].buffer = uniform_buffer;
bg_entries[1].size = sizeof(FillParams);

WGPUBindGroupDescriptor bg_desc = {};
bg_desc.layout = bgl;
bg_desc.entryCount = 2;
bg_desc.entries = bg_entries;
WGPUBindGroup bind_group = wgpuDeviceCreateBindGroup(device, &bg_desc);
utils::ComputePipelineBundle bundle = utils::make_compute_pipeline(
device,
kFillWGSL,
{
{0,
WGPUBufferBindingType_Storage,
out_tensor.buffer,
out_tensor.nbytes},
{1,
WGPUBufferBindingType_Uniform,
uniform_buffer,
sizeof(FillParams)},
},
&wg_size_constant,
1);

const size_t dispatch_idx = graph.add_dispatch(
{pipeline, bind_group, workgroup_count.x, "", workgroup_count.y});
{bundle.pipeline,
bundle.bind_group,
workgroup_count.x,
"",
workgroup_count.y});

// Dynamic shapes: recompute num_elements/dispatch from the live output dims.
WGPUBuffer params_buf = uniform_buffer;
Expand All @@ -136,9 +107,6 @@ void add_fill(
g.dispatch_at(dispatch_idx).workgroup_count_y = wgc.y;
});

wgpuShaderModuleRelease(shader);
wgpuBindGroupLayoutRelease(bgl);
wgpuPipelineLayoutRelease(pipeline_layout);
graph.own_uniform_buffer(uniform_buffer);
}

Expand Down
Loading
Loading