Merge branch 'main' of https://github.com/microsoft/onnxruntime into …

…pengwa/mem_improvement

.github/workflows/codeql.yml

@@ -37,7 +37,7 @@ jobs:
 
     # Initializes the CodeQL tools for scanning.
     - name: Initialize CodeQL
-      uses: github/codeql-action/init@v2
+      uses: github/codeql-action/init@v3
       with:
         languages: ${{ matrix.language }}
         # If you wish to specify custom queries, you can do so here or in a config file.
@@ -51,7 +51,7 @@ jobs:
     # If this step fails, then you should remove it and run the build manually (see below)
     - if: ${{ matrix.language != 'cpp' }}
       name: Autobuild
-      uses: github/codeql-action/autobuild@v2
+      uses: github/codeql-action/autobuild@v3
 
     - name: Perform CodeQL Analysis
-      uses: github/codeql-action/analyze@v2
+      uses: github/codeql-action/analyze@v3

.github/workflows/lint.yml

@@ -36,7 +36,7 @@ jobs:
     steps:
       - uses: actions/checkout@v4
       - name: Setup Python
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           # Version range or exact version of Python to use, using SemVer's version range syntax. Reads from .python-version if unset.
           python-version: "3.10"
@@ -65,7 +65,7 @@ jobs:
       - name: Upload SARIF file
         if: always()
         continue-on-error: true
-        uses: github/codeql-action/upload-sarif@v2
+        uses: github/codeql-action/upload-sarif@v3
         with:
           # Path to SARIF file relative to the root of the repository
           sarif_file: lintrunner.sarif

.github/workflows/linux.yml

@@ -17,7 +17,7 @@ jobs:
        - uses: actions/checkout@v4
          with:
            submodules: true
-       - uses: actions/setup-python@v4
+       - uses: actions/setup-python@v5
          with:
            python-version: '3.8.x'
            architecture: 'x64'

.github/workflows/publish-csharp-apidocs.yml

@@ -26,7 +26,7 @@ jobs:
     steps:
     - uses: actions/checkout@v4
     - name: Setup .NET
-      uses: actions/setup-dotnet@v3
+      uses: actions/setup-dotnet@v4
       with:
         dotnet-version: 6.0.x
     - name: Restore dependencies

.github/workflows/windows.yml

@@ -21,7 +21,7 @@ jobs:
       - uses: actions/checkout@v4
         with:
           submodules: false
-      - uses: actions/setup-python@v4
+      - uses: actions/setup-python@v5
         with:
           python-version: '3.11.x'
           architecture: 'x64'

cmake/onnxruntime_providers_dml.cmake

@@ -62,7 +62,7 @@
   target_link_libraries(onnxruntime_providers_dml PRIVATE delayimp.lib)
 
   if (NOT GDK_PLATFORM)
-    set(onnxruntime_DELAYLOAD_FLAGS "${onnxruntime_DELAYLOAD_FLAGS} /DELAYLOAD:DirectML.dll /DELAYLOAD:d3d12.dll /DELAYLOAD:dxgi.dll /DELAYLOAD:api-ms-win-core-com-l1-1-0.dll /DELAYLOAD:shlwapi.dll /DELAYLOAD:oleaut32.dll /DELAYLOAD:ext-ms-win-dxcore-l1-*.dll /ignore:4199")
+    set(onnxruntime_DELAYLOAD_FLAGS "${onnxruntime_DELAYLOAD_FLAGS} /DELAYLOAD:DirectML.dll /DELAYLOAD:d3d12.dll /DELAYLOAD:dxgi.dll /DELAYLOAD:dxcore.dll /DELAYLOAD:api-ms-win-core-com-l1-1-0.dll /DELAYLOAD:shlwapi.dll /DELAYLOAD:oleaut32.dll /DELAYLOAD:ext-ms-win-dxcore-l1-*.dll /ignore:4199")
   endif()
 
   target_compile_definitions(onnxruntime_providers_dml

js/web/lib/wasm/jsep/init.ts

@@ -69,7 +69,7 @@ class ComputeContextImpl implements ComputeContext {
     const heapU32 = module.HEAPU32;
 
     // extract context data
-    let dataIndex = (contextDataOffset >> 2);
+    let dataIndex = (contextDataOffset >>> 2);
     this.opKernelContext = heapU32[dataIndex++];
     const inputCount = heapU32[dataIndex++];
     this.outputCount = heapU32[dataIndex++];

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

@@ -5,7 +5,7 @@ import {env} from 'onnxruntime-common';
 
 import {TensorView} from '../../tensor-view';
 import {PoolConvUtil, ShapeUtil} from '../../util';
-import {AttributeWithCacheKey, createAttributeWithCacheKey} from '../attribute-with-cache-key';
+import {AttributeWithCacheKey} from '../attribute-with-cache-key';
 import {ComputeContext, ProgramInfo, ProgramInputTensorInfoDependency, ProgramUniform} from '../types';
 
 import {createTensorShapeVariables, getElementAt, IndicesHelper, inputVariable, outputVariable, ShaderHelper, UniformsArrayType} from './common';
@@ -63,7 +63,7 @@ const getUniformAndPadInfo = <AttributeType extends AveragePoolAttributes|MaxPoo
     const sw = attributes.strides[attributes.strides.length - 1];
     const pwStart = attributes.pads[attributes.pads.length / 2 - 1];
     const pwEnd = attributes.pads[attributes.pads.length - 1];
-    const pwStartEnd = !!(pwStart + pwEnd);
+    const pwStartEndNotZero = !!(pwStart + pwEnd);
     programUniforms.push(
         {type: 'uint32', data: kw},
         {type: 'uint32', data: sw},
@@ -74,13 +74,13 @@ const getUniformAndPadInfo = <AttributeType extends AveragePoolAttributes|MaxPoo
         {name: 'kw', type: 'u32'}, {name: 'sw', type: 'u32'}, {name: 'pwStart', type: 'u32'},
         {name: 'pwEnd', type: 'u32'});
 
-    let phStartEnd = false;
+    let phStartEndNotZero = false;
     if (attributes.kernelShape.length === 2) {
       const kh = attributes.kernelShape[attributes.kernelShape.length - 2];
       const sh = attributes.strides[attributes.strides.length - 2];
       const phStart = attributes.pads[attributes.pads.length / 2 - 2];
       const phEnd = attributes.pads[attributes.pads.length - 2];
-      phStartEnd = !!(phStart + phEnd);
+      phStartEndNotZero = !!(phStart + phEnd);
       programUniforms.push(
           {type: 'uint32', data: kh}, {type: 'uint32', data: sh}, {type: 'uint32', data: phStart},
           {type: 'uint32', data: phEnd});
@@ -89,7 +89,7 @@ const getUniformAndPadInfo = <AttributeType extends AveragePoolAttributes|MaxPoo
           {name: 'kh', type: 'u32'}, {name: 'sh', type: 'u32'}, {name: 'phStart', type: 'u32'},
           {name: 'phEnd', type: 'u32'});
     }
-    return [programUniforms, uniforms, true, pwStartEnd, phStartEnd];
+    return [programUniforms, uniforms, true, pwStartEndNotZero, phStartEndNotZero];
   } else {
     if (isChannelsLast) {
       throw new Error('Pooling with kernelShape.length > 2 is not supported for NHWC format.');
@@ -110,8 +110,8 @@ const getUniformAndPadInfo = <AttributeType extends AveragePoolAttributes|MaxPoo
 
 const generatePoolingCode = <AttributeType extends AveragePoolAttributes|MaxPoolAttributes>(
     shaderHelper: ShaderHelper, x: IndicesHelper, rank: number, outputShapeRank: number, attributes: AttributeType,
-    op1: string, op2: string, start: number, uniforms: UniformsArrayType, hasPads: boolean, pwStartEnd: boolean,
-    phStartEnd: boolean): string => {
+    op1: string, op2: string, start: number, uniforms: UniformsArrayType, hasPads: boolean, pwStartEndNotZero: boolean,
+    phStartEndNotZero: boolean): string => {
   const isChannelsLast = attributes.format === 'NHWC';
   const dataType = x.type.value;
   const output = outputVariable('output', x.type.tensor, outputShapeRank);
@@ -121,7 +121,7 @@ const generatePoolingCode = <AttributeType extends AveragePoolAttributes|MaxPool
     let codeH = '';
     let codeHEnd = '';
     const dimIdxW = rank - (isChannelsLast ? 2 : 1);
-    if (pwStartEnd === true) {
+    if (pwStartEndNotZero) {
       codeW = `
                 for (var i: u32 = 0u; i < uniforms.kw; i++) {
                   xIndices[${dimIdxW}] = indices[${dimIdxW}] * uniforms.sw - uniforms.pwStart + i;
@@ -144,7 +144,7 @@ const generatePoolingCode = <AttributeType extends AveragePoolAttributes|MaxPool
 
     if (attributes.kernelShape.length === 2) {
       const dimIdxH = rank - (isChannelsLast ? 3 : 2);
-      if (phStartEnd === true) {
+      if (phStartEndNotZero) {
         codeH = `
                 for (var j: u32 = 0u; j < uniforms.kh; j++) {
                   xIndices[${dimIdxH}] = indices[${dimIdxH}] * uniforms.sh - uniforms.phStart + j;
@@ -258,6 +258,15 @@ export interface PoolCommonAttributes extends FormatAttributes {
   readonly pads: readonly number[];
 }
 
+const createShaderKeyFromAttributes = (attributes: PoolCommonAttributes): string =>
+    (`${attributes.format};${attributes.ceilMode};${attributes.autoPad};${attributes.kernelShape.length}`);
+
+const createAveragePoolShaderKeyFromAttributes = (attributes: AveragePoolAttributes): string =>
+    (`${createShaderKeyFromAttributes(attributes)};${attributes.countIncludePad}`);
+
+const createMaxPoolShaderKeyFromAttributes = (attributes: MaxPoolAttributes): string =>
+    (`${createShaderKeyFromAttributes(attributes)};${attributes.storageOrder};${attributes.dilations}`);
+
 const parsePoolCommonAttributes = (attributes: Record<string, unknown>): PoolCommonAttributes => ({
   format: attributes.format as FormatAttributes['format'],
   autoPad: ['NOTSET', 'VALID', 'SAME_UPPER', 'SAME_LOWER'][attributes.auto_pad as number],
@@ -285,25 +294,22 @@ const createAveragePoolProgramInfo =
       } else {
         op2 += `value /= ${dataType}(i32(uniforms.kernelSize) - pad);`;
       }
-      const [programUniforms, uniforms, hasPads, pwStartEnd, phStartEnd] =
+      const [programUniforms, uniforms, hasPads, pwStartEndNotZero, phStartEndNotZero] =
           getUniformAndPadInfo(outputShape, adjustedAttributes);
-      programUniforms.push(...createTensorShapeVariables(input.dims));
-      programUniforms.push(...createTensorShapeVariables(outputShape));
+      programUniforms.push(...createTensorShapeVariables(input.dims), ...createTensorShapeVariables(outputShape));
       const inputDependencies: ProgramInputTensorInfoDependency[] = ['rank'];
       return {
         name,
-        shaderCache: {
-          hint: attributes.cacheKey + hasPads + pwStartEnd + phStartEnd + adjustedAttributes.countIncludePad,
-          inputDependencies
-        },
+        shaderCache:
+            {hint: `${attributes.cacheKey};${hasPads};${pwStartEndNotZero};${phStartEndNotZero}`, inputDependencies},
         getRunData: () => ({
           outputs: [{dims: outputShape, dataType: input.dataType}],
           dispatchGroup: {x: Math.ceil(ShapeUtil.size(outputShape) / 64 /* workgroup size */)},
           programUniforms
         }),
         getShaderSource: shaderHelper => generatePoolingCode(
             shaderHelper, x, input.dims.length, outputShape.length, adjustedAttributes, op1, op2, 0.0, uniforms,
-            hasPads, pwStartEnd, phStartEnd),
+            hasPads, pwStartEndNotZero, phStartEndNotZero),
       };
     };
 
@@ -315,8 +321,8 @@ export const parseAveragePoolAttributes = (attributes: Record<string, unknown>):
   if (attr.ceilMode !== 0) {
     throw new Error('using ceil() in shape computation is not yet supported for AveragePool');
   }
-
-  return createAttributeWithCacheKey({countIncludePad, ...attr});
+  const averagePoolAttributes = {countIncludePad, ...attr, cacheKey: ''};
+  return {...averagePoolAttributes, cacheKey: createAveragePoolShaderKeyFromAttributes(averagePoolAttributes)};
 };
 
 export const averagePool = (context: ComputeContext, attributes: AveragePoolAttributes): void => {
@@ -332,8 +338,7 @@ const globalPoolAttributes = {
   strides: [],
   pads: [],
   storageOrder: 0,
-  dilations: [],
-  cacheKey: ''
+  dilations: []
 };
 
 export const parseGlobalAveragePoolAttributes = (attributes: Record<string, unknown>): AveragePoolAttributes => {
@@ -361,21 +366,21 @@ const createMaxPoolProgramInfo =
       const op2 = '';
       const x = inputVariable('x', input.dataType, input.dims.length);
       const inputDependencies: ProgramInputTensorInfoDependency[] = ['rank'];
-      const [programUniforms, uniforms, hasPads, pwStartEnd, phStartEnd] =
+      const [programUniforms, uniforms, hasPads, pwStartEndNotZero, phStartEndNotZero] =
           getUniformAndPadInfo(outputShape, adjustedAttributes);
-      programUniforms.push(...createTensorShapeVariables(input.dims));
-      programUniforms.push(...createTensorShapeVariables(outputShape));
+      programUniforms.push(...createTensorShapeVariables(input.dims), ...createTensorShapeVariables(outputShape));
       return {
         name,
-        shaderCache: {hint: attributes.cacheKey + hasPads, inputDependencies},
+        shaderCache:
+            {hint: `${attributes.cacheKey};${hasPads};${pwStartEndNotZero};${phStartEndNotZero}`, inputDependencies},
         getRunData: () => ({
           outputs: [{dims: outputShape, dataType: input.dataType}],
           dispatchGroup: {x: Math.ceil(ShapeUtil.size(outputShape) / 64 /* workgroup size */)},
           programUniforms
         }),
         getShaderSource: shaderHelper => generatePoolingCode(
             shaderHelper, x, input.dims.length, outputShape.length, adjustedAttributes, op1, op2, -1e5, uniforms,
-            hasPads, pwStartEnd, phStartEnd),
+            hasPads, pwStartEndNotZero, phStartEndNotZero),
       };
     };
 
@@ -396,8 +401,8 @@ export const parseMaxPoolAttributes = (attributes: Record<string, unknown>): Max
   if (attr.ceilMode !== 0) {
     throw new Error('using ceil() in shape computation is not yet supported for MaxPool');
   }
-
-  return createAttributeWithCacheKey({storageOrder, dilations, ...attr});
+  const maxPoolAttributes = {storageOrder, dilations, ...attr, cacheKey: ''};
+  return {...maxPoolAttributes, cacheKey: createMaxPoolShaderKeyFromAttributes(maxPoolAttributes)};
 };
 
 export const parseGlobalMaxPoolAttributes = (attributes: Record<string, unknown>): MaxPoolAttributes => {