[js/webgpu] Add hardSigmoid activation for fusedConv

js/web/lib/wasm/jsep/webgpu/ops/3rd-party/conv2d_mm_webgpu.ts

@@ -24,7 +24,7 @@ import {TensorView} from '../../../tensor-view';
 import {ProgramInfo, ProgramInputTensorInfoDependency, ProgramUniform} from '../../types';
 import {createTensorShapeVariables, inputVariable, outputVariable, ShaderHelper, tensorTypeToWsglStorageType, UniformsArrayType} from '../common';
 import {ConvAttributes} from '../conv';
-import {getActivationSnippet} from '../fuse-utils';
+import {appendActivationUniforms, appendActivationUniformsData, getActivationSnippet} from '../fuse-utils';
 
 import {biasSnippet, typeSnippet} from './activation_util';
 import {utilFunctions} from './conv_util';
@@ -193,10 +193,7 @@ export const createConv2DMatMulProgramInfo =
         {type: 'int32', data: [attributes.pads[0], attributes.pads[1]]}, {type: 'int32', data: attributes.strides},
         {type: 'int32', data: attributes.dilations}
       ];
-      if (attributes.activation === 'Clip') {
-        programUniforms.push(
-            {type: 'float32', data: attributes.clipMax!}, {type: 'float32', data: attributes.clipMin!});
-      }
+      appendActivationUniformsData(attributes, programUniforms);
       programUniforms.push(
           ...createTensorShapeVariables(inputs[0].dims), ...createTensorShapeVariables(inputs[1].dims));
       const inputDependencies: ProgramInputTensorInfoDependency[] = ['rank', 'rank'];
@@ -212,9 +209,7 @@ export const createConv2DMatMulProgramInfo =
           {name: 'pad', type: 'i32', length: 2}, {name: 'stride', type: 'i32', length: 2},
           {name: 'dilation', type: 'i32', length: 2}
         ];
-        if (attributes.activation === 'Clip') {
-          uniforms.push({name: 'clip_max', type: 'f32'}, {name: 'clip_min', type: 'f32'});
-        }
+        appendActivationUniforms(attributes, uniforms);
 
         // TODO: support component 2, 3.
         const components = isVec4 ? 4 : 1;

js/web/lib/wasm/jsep/webgpu/ops/3rd-party/conv_backprop_mm_webgpu.ts

@@ -24,7 +24,7 @@ import {TensorView} from '../../../tensor-view';
 import {ProgramInfo, ProgramInputTensorInfoDependency, ProgramUniform} from '../../types';
 import {createTensorShapeVariables, inputVariable, outputVariable, ShaderHelper, UniformsArrayType} from '../common';
 import {ConvTransposeAttributes} from '../conv-transpose';
-import {getActivationSnippet} from '../fuse-utils';
+import {appendActivationUniforms, appendActivationUniformsData, getActivationSnippet} from '../fuse-utils';
 
 import {biasSnippet, typeSnippet} from './activation_util';
 import {utilFunctions} from './conv_util';
@@ -201,10 +201,7 @@ export const createConv2DTransposeMatMulProgramInfo =
         {type: 'int32', data: attributes.strides}, {type: 'int32', data: attributes.dilations},
         {type: 'int32', data: filterDims}, {type: 'int32', data: pads}
       ];
-      if (attributes.activation === 'Clip') {
-        programUniforms.push(
-            {type: 'float32', data: attributes.clipMax!}, {type: 'float32', data: attributes.clipMin!});
-      }
+      appendActivationUniformsData(attributes, programUniforms);
       programUniforms.push(
           ...createTensorShapeVariables(inputs[0].dims), ...createTensorShapeVariables(inputs[1].dims));
 
@@ -237,9 +234,7 @@ export const createConv2DTransposeMatMulProgramInfo =
           {name: 'filter_dims', type: 'i32', length: filterDims.length},
           {name: 'pads', type: 'i32', length: pads.length}
         ];
-        if (attributes.activation === 'Clip') {
-          uniforms.push({name: 'clip_max', type: 'f32'}, {name: 'clip_min', type: 'f32'});
-        }
+        appendActivationUniforms(attributes, uniforms);
         return `
         ${utilFunctions('uniforms.result_strides')}
         ${shaderHelper.registerUniforms(uniforms).declareVariables(...inputVariables, output)};

js/web/lib/wasm/jsep/webgpu/ops/3rd-party/matmul_packed_webgpu.ts

@@ -23,7 +23,7 @@ import {TensorView} from '../../../tensor-view';
 import {ShapeUtil} from '../../../util';
 import {ProgramInfo, ProgramInputTensorInfoDependency, ProgramUniform} from '../../types';
 import {createTensorShapeVariables, getBroadcastDims, IndicesHelper, inputVariable, internalVariable, outputVariable, ShaderHelper, tensorTypeToWsglStorageType, UniformsArrayType} from '../common';
-import {getActivationSnippet, InternalActivationAttributes} from '../fuse-utils';
+import {appendActivationUniforms, appendActivationUniformsData, getActivationSnippet, InternalActivationAttributes} from '../fuse-utils';
 
 import {typeSnippet} from './activation_util';
 
@@ -449,11 +449,7 @@ export const createMatmulProgramInfo =
       const outputShapeTemp = [batchSize, dimAOuter, dimBOuter / components];
       const programUniforms: ProgramUniform[] =
           [{type: 'int32', data: dimAOuter}, {type: 'int32', data: dimBOuter}, {type: 'int32', data: dimInner}];
-      if (activationAttributes.activation === 'Clip') {
-        programUniforms.push(
-            {type: 'float32', data: activationAttributes.clipMax!},
-            {type: 'float32', data: activationAttributes.clipMin!});
-      }
+      appendActivationUniformsData(activationAttributes, programUniforms);
       programUniforms.push(
           ...createTensorShapeVariables(outerDims), ...createTensorShapeVariables(aShapeTemp),
           ...createTensorShapeVariables(bShapeTemp));
@@ -481,9 +477,7 @@ export const createMatmulProgramInfo =
         }
         const uniforms: UniformsArrayType =
             [{name: 'dim_a_outer', type: 'i32'}, {name: 'dim_b_outer', type: 'i32'}, {name: 'dim_inner', type: 'i32'}];
-        if (activationAttributes.activation === 'Clip') {
-          uniforms.push({name: 'clip_max', type: 'f32'}, {name: 'clip_min', type: 'f32'});
-        }
+        appendActivationUniforms(activationAttributes, uniforms);
         const applyActivation = getActivationSnippet(activationAttributes, output.type.value);
         const declareFunctions = matMulReadWriteFnSource(
             components, hasBias, applyActivation, [batchDims, A, B, output], [outerDimsA, outerDimsB, outerDims],

js/web/lib/wasm/jsep/webgpu/ops/conv-grouped.ts

@@ -7,7 +7,7 @@ import {ProgramInfo, ProgramInputTensorInfoDependency, ProgramUniform} from '../
 
 import {createTensorShapeVariables, getMaxComponents, inputVariable, outputVariable, ShaderHelper, UniformsArrayType} from './common';
 import {calculateOutputShape, ConvAttributes} from './conv';
-import {getActivationSnippet} from './fuse-utils';
+import {appendActivationUniforms, appendActivationUniformsData, getActivationSnippet} from './fuse-utils';
 
 /**
  * naive grouped conv implementation, supports 1d/2d conv
@@ -32,10 +32,7 @@ export const createGroupedConvProgramInfo =
         {type: 'uint32', data: [attributes.strides[0], attributes.strides[1]]},
         {type: 'uint32', data: [attributes.pads[0], attributes.pads[1]]}, {type: 'uint32', data: outputChannelsPerGroup}
       ];
-      if (attributes.activation === 'Clip') {
-        programUniforms.push(
-            {type: 'float32', data: attributes.clipMax!}, {type: 'float32', data: attributes.clipMin!});
-      }
+      appendActivationUniformsData(attributes, programUniforms);
       programUniforms.push(
           ...createTensorShapeVariables(xShape), ...createTensorShapeVariables(wShape),
           ...createTensorShapeVariables(outputShape));
@@ -61,9 +58,7 @@ export const createGroupedConvProgramInfo =
           {name: 'strides', type: 'u32', length: 2}, {name: 'pads', type: 'u32', length: 2},
           {name: 'output_channels_per_group', type: 'u32'}
         ];
-        if (attributes.activation === 'Clip') {
-          uniforms.push({name: 'clip_max', type: 'f32'}, {name: 'clip_min', type: 'f32'});
-        }
+        appendActivationUniforms(attributes, uniforms);
         return `
   ${shaderHelper.registerUniforms(uniforms).declareVariables(...inputVars, output)}
 

js/web/lib/wasm/jsep/webgpu/ops/fuse-utils.ts

@@ -2,11 +2,16 @@
 // Licensed under the MIT License.
 
 import {MAX_CLIP, MIN_CLIP} from '../../util';
+import {ProgramUniform} from '../types';
+
+import {UniformsArrayType} from './common';
 
 export interface InternalActivationAttributes {
   readonly activation: string;
   readonly clipMin?: number;
   readonly clipMax?: number;
+  readonly alpha?: number;
+  readonly beta?: number;
 }
 
 export const getActivationSnippet = (attributes: InternalActivationAttributes, valueType: string): string => {
@@ -17,16 +22,39 @@ export const getActivationSnippet = (attributes: InternalActivationAttributes, v
       return `value = (${valueType}(1.0) / (${valueType}(1.0) + exp(-value)));`;
     case 'Clip':
       return `value = clamp(value, ${valueType}(uniforms.clip_min), ${valueType}(uniforms.clip_max));`;
+    case 'HardSigmoid':
+      return `value = max(${valueType}(0.0), min(${valueType}(1.0), ${valueType}(uniforms.alpha) * value + ${
+          valueType}(uniforms.beta)));`;
     // TODO: adding other activations that can be fused.
     default:
       return '';
   }
 };
 
+export const appendActivationUniformsData =
+    (attributes: InternalActivationAttributes, programUniform: ProgramUniform[]) => {
+      if (attributes.activation === 'Clip') {
+        programUniform.push({type: 'float32', data: attributes.clipMax!}, {type: 'float32', data: attributes.clipMin!});
+      } else if (attributes.activation === 'HardSigmoid') {
+        programUniform.push({type: 'float32', data: attributes.alpha!}, {type: 'float32', data: attributes.beta!});
+      }
+    };
+
+export const appendActivationUniforms = (attributes: InternalActivationAttributes, uniforms: UniformsArrayType) => {
+  if (attributes.activation === 'Clip') {
+    uniforms.push({name: 'clip_max', type: 'f32'}, {name: 'clip_min', type: 'f32'});
+  } else if (attributes.activation === 'HardSigmoid') {
+    uniforms.push({name: 'alpha', type: 'f32'}, {name: 'beta', type: 'f32'});
+  }
+};
+
 export const parseInternalActivationAttributes =
     (attributes: Record<string, unknown>|undefined): InternalActivationAttributes => {
       const activation = attributes?.activation as string || '';
-
+      if (activation === 'HardSigmoid') {
+        const [alpha, beta] = attributes?.activation_params as [number, number] || [0.2, 0.5];
+        return {activation, alpha, beta};
+      }
       if (activation === 'Clip') {
         const [clipMin, clipMax] = attributes?.activation_params as [number, number] || [MIN_CLIP, MAX_CLIP];
         return {activation, clipMax, clipMin};

js/web/lib/wasm/jsep/webgpu/ops/matmul.ts

@@ -7,7 +7,7 @@ import {ComputeContext, ProgramInfo, ProgramUniform} from '../types';
 
 import {createMatmulProgramInfo} from './3rd-party/matmul_packed_webgpu';
 import {createTensorShapeVariables, getBroadcastDims, getMaxComponents, IndicesHelper, inputVariable, internalVariable, outputVariable, ShaderHelper, UniformsArrayType,} from './common';
-import {getActivationSnippet, InternalActivationAttributes} from './fuse-utils';
+import {appendActivationUniforms, appendActivationUniformsData, getActivationSnippet, InternalActivationAttributes} from './fuse-utils';
 
 export const createNaiveMatmulProgramInfo =
     (inputs: readonly TensorView[], activationAttributes: InternalActivationAttributes, outputShape: readonly number[],
@@ -32,11 +32,7 @@ export const createNaiveMatmulProgramInfo =
         {type: 'uint32', data: outputSize}, {type: 'uint32', data: M}, {type: 'uint32', data: N},
         {type: 'uint32', data: K}
       ];
-      if (activationAttributes.activation === 'Clip') {
-        programUniforms.push(
-            {type: 'float32', data: activationAttributes.clipMax!},
-            {type: 'float32', data: activationAttributes.clipMin!});
-      }
+      appendActivationUniformsData(activationAttributes, programUniforms);
       programUniforms.push(
           ...createTensorShapeVariables(outerDims), ...createTensorShapeVariables(aShape),
           ...createTensorShapeVariables(bShape));
@@ -69,9 +65,7 @@ export const createNaiveMatmulProgramInfo =
           {name: 'output_size', type: 'u32'}, {name: 'M', type: 'u32'}, {name: 'N', type: 'u32'},
           {name: 'K', type: 'u32'}
         ];
-        if (activationAttributes.activation === 'Clip') {
-          uniforms.push({name: 'clip_max', type: 'f32'}, {name: 'clip_min', type: 'f32'});
-        }
+        appendActivationUniforms(activationAttributes, uniforms);
 
         const getIndices = (variable: IndicesHelper, broadCastDims: number[]) => {
           const rank = variable.rank;

js/web/test/data/ops/fused-conv.jsonc

@@ -142,5 +142,73 @@
         ]
       }
     ]
+  },
+  {
+    "name": "fused conv with HardSigmoid",
+    "operator": "FusedConv",
+    "attributes": [
+      { "name": "activation", "data": "HardSigmoid", "type": "string" },
+      { "name": "kernel_shape", "data": [2, 2], "type": "ints" },
+      { "name": "activation_params", "data": [2.0, 5.0], "type": "floats" }
+    ],
+    "opset": { "domain": "com.microsoft", "version": 1 },
+    "cases": [
+      {
+        "name": "T[0]",
+        "inputs": [
+          {
+            "data": [10, 20, -30, -40, -50, -60, 70, 80, 90],
+            "dims": [1, 1, 3, 3],
+            "type": "float32"
+          },
+          {
+            "data": [1, 2, 3, 4],
+            "dims": [1, 1, 2, 2],
+            "type": "float32"
+          }
+        ],
+        "outputs": [
+          {
+            "data": [0, 0, 1, 1],
+            "dims": [1, 1, 2, 2],
+            "type": "float32"
+          }
+        ]
+      }
+    ]
+  },
+  {
+    "name": "NHWC conv with HardSigmoid",
+    "operator": "Conv",
+    "attributes": [
+      { "name": "activation", "data": "HardSigmoid", "type": "string" },
+      { "name": "kernel_shape", "data": [2, 2], "type": "ints" },
+      { "name": "activation_params", "data": [2.0, 5.0], "type": "floats" }
+    ],
+    "opset": { "domain": "com.ms.internal.nhwc", "version": 1 },
+    "cases": [
+      {
+        "name": "T[0]",
+        "inputs": [
+          {
+            "data": [10, 20, -30, -40, -50, -60, 70, 80, 90],
+            "dims": [1, 3, 3, 1],
+            "type": "float32"
+          },
+          {
+            "data": [1, 2, 3, 4],
+            "dims": [1, 1, 2, 2],
+            "type": "float32"
+          }
+        ],
+        "outputs": [
+          {
+            "data": [0, 0, 1, 1],
+            "dims": [1, 2, 2, 1],
+            "type": "float32"
+          }
+        ]
+      }
+    ]
   }
 ]

onnxruntime/core/optimizer/conv_activation_fusion.cc

@@ -111,7 +111,7 @@ class ConvActivationSelector : public NodeSelector {
       if (!graph_utils::IsSupportedOptypeVersionAndDomain(*next_node, "Relu", {6, 13, 14})) {
         return std::nullopt;
       }
-    } else if (node_ep.empty() || node_ep == kCpuExecutionProvider) {
+    } else if (node_ep.empty() || node_ep == kCpuExecutionProvider || node_ep == kJsExecutionProvider) {
       if (!is_supported_non_cuda_rocm_ep_activation(*next_node) &&
           !graph_utils::IsSupportedOptypeVersionAndDomain(*next_node, "HardSigmoid", {6})) {
         return std::nullopt;