[ExecuTorch][WebGPU] Adopt WebGPU layout:"auto" in make_compute_pipeline

backends/webgpu/runtime/WebGPUUtils.h

@@ -376,40 +376,34 @@ inline ComputePipelineBundle make_compute_pipeline(
   shader_desc.nextInChain = &wgsl_desc.chain;
   WGPUShaderModule shader = wgpuDeviceCreateShaderModule(device, &shader_desc);
 
-  std::vector<WGPUBindGroupLayoutEntry> layout_entries(bindings.size());
   std::vector<WGPUBindGroupEntry> bind_entries(bindings.size());
   for (size_t i = 0; i < bindings.size(); i++) {
-    layout_entries[i] = {};
-    layout_entries[i].binding = bindings[i].binding;
-    layout_entries[i].visibility = WGPUShaderStage_Compute;
-    layout_entries[i].buffer.type = bindings[i].type;
-
     bind_entries[i] = {};
     bind_entries[i].binding = bindings[i].binding;
     bind_entries[i].buffer = bindings[i].buffer;
     bind_entries[i].size = bindings[i].size;
   }
 
-  WGPUBindGroupLayoutDescriptor bgl_desc = {};
-  bgl_desc.entryCount = layout_entries.size();
-  bgl_desc.entries = layout_entries.data();
-  WGPUBindGroupLayout bgl = wgpuDeviceCreateBindGroupLayout(device, &bgl_desc);
-
-  WGPUPipelineLayoutDescriptor pl_desc = {};
-  pl_desc.bindGroupLayoutCount = 1;
-  pl_desc.bindGroupLayouts = &bgl;
-  WGPUPipelineLayout pipeline_layout =
-      wgpuDeviceCreatePipelineLayout(device, &pl_desc);
-
+  // layout = nullptr => WebGPU auto-derives the bind-group layout from the
+  // shader's statically-used @group/@binding declarations, replacing the
+  // hand-built bind-group layout + pipeline layout. Precondition: every
+  // declared binding must be statically referenced by the shader (optional
+  // bindings backed by a dummy buffer are, under a runtime guard) -- auto
+  // layout omits an unreferenced binding, whose bind-group entry would then
+  // mismatch. BindingSpec.type is unused under auto layout (kept so call sites
+  // need no change).
   WGPUComputePipelineDescriptor pipeline_desc = {};
-  pipeline_desc.layout = pipeline_layout;
+  pipeline_desc.layout = nullptr;
   pipeline_desc.compute.module = shader;
   pipeline_desc.compute.entryPoint = {entry_point, WGPU_STRLEN};
   pipeline_desc.compute.constantCount = constant_count;
   pipeline_desc.compute.constants = constants;
   WGPUComputePipeline pipeline =
       wgpuDeviceCreateComputePipeline(device, &pipeline_desc);
 
+  // Owned reference (must be released) -> handed to the bundle dtor.
+  WGPUBindGroupLayout bgl = wgpuComputePipelineGetBindGroupLayout(pipeline, 0);
+
   WGPUBindGroupDescriptor bg_desc = {};
   bg_desc.layout = bgl;
   bg_desc.entryCount = bind_entries.size();
@@ -419,7 +413,7 @@ inline ComputePipelineBundle make_compute_pipeline(
   ComputePipelineBundle bundle;
   bundle.shader = shader;
   bundle.bind_group_layout = bgl;
-  bundle.pipeline_layout = pipeline_layout;
+  bundle.pipeline_layout = nullptr; // none created under auto layout
   bundle.pipeline = pipeline;
   bundle.bind_group = bind_group;
   return bundle;

backends/webgpu/runtime/ops/max_pool2d/MaxPool2d.cpp

@@ -24,7 +24,7 @@ namespace {
 struct PoolParams {
   uint32_t N, C, IH, IW, OH, OW;
   uint32_t kH, kW, sH, sW, pH, pW, dH, dW;
-  uint32_t _p0, _p1;
+  uint32_t write_indices, _p1;
 };
 static_assert(sizeof(PoolParams) == 64, "PoolParams must be 64 bytes");
 
@@ -145,17 +145,13 @@ void max_pool2d_impl(WebGPUGraph& graph, const std::vector<int>& args) {
   params.pW = pW;
   params.dH = dH;
   params.dW = dW;
+  params.write_indices = has_indices ? 1u : 0u;
 
   WGPUBuffer uniform_buffer =
       utils::make_uniform(device, &params, sizeof(PoolParams));
   graph.add_uniform_buffer_bytes(sizeof(PoolParams));
 
   auto grid_constants = utils::make_grid_constants(grid);
-  WGPUConstantEntry constants[3] = {};
-  constants[0] = grid_constants[0];
-  constants[1] = grid_constants[1];
-  constants[2].key = {"write_indices", WGPU_STRLEN};
-  constants[2].value = has_indices ? 1.0 : 0.0;
 
   // write_indices==0 -> the shader never stores to out_idx, so a tiny dummy
   // buffer is safe (mirrors NativeLayerNorm.cpp's dummy_affine pattern).
@@ -177,8 +173,8 @@ void max_pool2d_impl(WebGPUGraph& graph, const std::vector<int>& args) {
            sizeof(PoolParams)},
           {3, WGPUBufferBindingType_Storage, idx.buffer, idx.nbytes},
       },
-      constants,
-      3);
+      grid_constants.data(),
+      grid_constants.size());
 
   graph.add_dispatch_2d(
       bundle.pipeline, bundle.bind_group, grid.count_x, grid.count_y);

backends/webgpu/runtime/ops/max_pool2d/max_pool2d.wgsl

@@ -13,7 +13,7 @@ struct Params {
   pW: u32,
   dH: u32,
   dW: u32,
-  _p0: u32,
+  write_indices: u32, // 1 = also write out_idx (runtime gate; keeps out_idx statically used)
   _p1: u32,
 }
 
@@ -24,7 +24,6 @@ struct Params {
 
 override wg_size: u32 = 256;
 override stride_x: u32 = 4294967295u; // = count_x * wg_size; set by host for 2D-spill
-override write_indices: u32 = 0u; // 1 = also write out_idx (compile-time gate, mirrors Vulkan)
 
 // max_pool2d (values [+ optional indices]), NCHW row-major, fp32. One thread per
 // output element (n, c, oh, ow); gather the window, take the max. General
@@ -73,7 +72,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     }
   }
   out_vals[i] = best;
-  if (write_indices != 0u) {
+  if (params.write_indices != 0u) {
     out_idx[i] = best_idx;
   }
 }

backends/webgpu/runtime/ops/max_pool2d/max_pool2d_wgsl.h

@@ -13,7 +13,7 @@
 namespace executorch::backends::webgpu {
 
 // @generated from max_pool2d.wgsl - DO NOT EDIT.
-// wgsl-sha256: 2215b7ba725face441baac5eca8e2133458b414377868354c3fda9046a88109b
+// wgsl-sha256: 205fd7013aaacb5e5d4bda7499d436da6ca76e02f910962223c0dbd65f95e6ac
 inline constexpr const char* kMaxPool2dWGSL = R"(
 struct Params {
   N: u32,
@@ -30,7 +30,7 @@ struct Params {
   pW: u32,
   dH: u32,
   dW: u32,
-  _p0: u32,
+  write_indices: u32, // 1 = also write out_idx (runtime gate; keeps out_idx statically used)
   _p1: u32,
 }
 
@@ -41,7 +41,6 @@ struct Params {
 
 override wg_size: u32 = 256;
 override stride_x: u32 = 4294967295u; // = count_x * wg_size; set by host for 2D-spill
-override write_indices: u32 = 0u; // 1 = also write out_idx (compile-time gate, mirrors Vulkan)
 
 // max_pool2d (values [+ optional indices]), NCHW row-major, fp32. One thread per
 // output element (n, c, oh, ow); gather the window, take the max. General
@@ -90,7 +89,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     }
   }
   out_vals[i] = best;
-  if (write_indices != 0u) {
+  if (params.write_indices != 0u) {
     out_idx[i] = best_idx;
   }
 }