Merge branch 'master' into sync_msft_10_7_2026

Author: jatinwadhwa921

js/web/lib/onnxjs/util.ts

@@ -1260,6 +1260,8 @@ export class PoolConvUtil {
    * @param pads Padding for the beginning and ending along each axis.
    * @param autoPad DEPRECATED attribute supported for legacy models. Specifies how to implicitly calculate pads in each
    *     dimension. Can take values NOTSET, SAME_UPPER, SAME_LOWER, or VALID.
+   * @param ceilMode When set to 1, use ceil() instead of floor() to compute the output spatial size (and apply the
+   *     "shrink the last window if it starts entirely in padding" rule). Defaults to 0 (floor).
    */
   static computePoolOutputShape(
     isGlobalOperator: boolean,
@@ -1269,6 +1271,7 @@ export class PoolConvUtil {
     kernelShape: number[],
     pads: number[],
     autoPad?: string,
+    ceilMode = 0,
   ): number[] {
     if (inputDims.length <= 0) {
       throw new Error('input shape must be of size greater than 0');
@@ -1286,6 +1289,7 @@ export class PoolConvUtil {
       kernelShape,
       pads,
       autoPad,
+      ceilMode,
     );
     return outputDims;
   }
@@ -1332,6 +1336,7 @@ export class PoolConvUtil {
     kernelShape: readonly number[],
     pads: number[],
     autoPad?: string,
+    ceilMode = 0,
   ) {
     if (isGlobalOperator) {
       for (let dim = 0; dim < inputDims.length - 2; dim++) {
@@ -1349,12 +1354,44 @@ export class PoolConvUtil {
             dim,
             dim + inputDims.length - 2,
             autoPad,
+            ceilMode,
           ),
         );
       }
     }
   }
 
+  // Computes the output spatial size for a single dimension.
+  // Produces results identical to the C++ PoolAttributes::ComputeOutputSize
+  // (onnxruntime/core/providers/cpu/nn/pool_attributes.h), including the ceil_mode
+  // "shrink the last window if it starts entirely in the trailing padding" rule. The JS
+  // signature takes a pre-computed `numerator` (equal to `inSize + padHead + padTail - dkernel`,
+  // matching the C++ `in_size + pad_head + pad_tail - dilation * (kernel - 1) - 1`) instead of
+  // the raw pooling attributes, but the computed output size is the same.
+  // Keep in sync with the onnxjs/jsep copy.
+  // NOTE: In this onnxjs copy the ceilMode path exists for shape-test parity with the jsep copy;
+  // the onnxjs WebGL pooling caller (backends/webgl/ops/pool.ts) intentionally does NOT pass
+  // ceilMode (legacy path still throws on ceil_mode != 0), so it always uses the floor default.
+  private static computeOutputSize(
+    numerator: number,
+    stride: number,
+    inSize: number,
+    padHead: number,
+    ceilMode: number,
+  ): number {
+    let outSize = Math.floor(numerator / stride) + 1;
+    // Match C++ `ceil_mode == 1` exactly so out-of-spec ceil_mode values do not diverge.
+    if (ceilMode === 1) {
+      outSize = Math.ceil(numerator / stride) + 1;
+      // Ensure the last pooling window starts inside the image (ref: https://github.com/onnx/onnx/pull/5741).
+      // inSize and padHead are needed here to reconstruct the last window's start position.
+      if ((outSize - 1) * stride >= inSize + padHead) {
+        outSize -= 1;
+      }
+    }
+    return outSize;
+  }
+
   // helper for computeShapeHelper() and adjustPadsBasedOnAutoPad()
   // adjusts pad value for given 'autoPad' string and computes output shape along a particular dimension
   private static adjustPadAndReturnShape(
@@ -1366,30 +1403,44 @@ export class PoolConvUtil {
     padHeadIndex: number,
     padTailIndex: number,
     autoPad?: string,
+    ceilMode = 0,
   ): number {
     const dkernel = dilation * (kernel - 1) + 1;
     if (autoPad && autoPad !== 'NOTSET') {
       switch (autoPad) {
         case 'VALID':
           pads[padHeadIndex] = 0;
           pads[padTailIndex] = 0;
-          return Math.floor((inSize - dkernel) / stride + 1);
+          return PoolConvUtil.computeOutputSize(inSize - dkernel, stride, inSize, 0, ceilMode);
         case 'SAME_LOWER':
         case 'SAME_UPPER':
           if (dilation !== 1) {
             throw new Error('Dilation not supported for SAME_UPPER or SAME_LOWER');
           } else {
-            const legacyTargetSize = (inSize + stride - 1) / stride;
+            // Integer division to match C++ pool_attributes.h ComputeSizePadDilations; float division mis-rounds SAME_* pads.
+            const legacyTargetSize = Math.floor((inSize + stride - 1) / stride);
             const padNeeded = (legacyTargetSize - 1) * stride + kernel - inSize;
             pads[padHeadIndex] = autoPad === 'SAME_LOWER' ? Math.floor((padNeeded + 1) / 2) : Math.floor(padNeeded / 2);
             pads[padTailIndex] = padNeeded - pads[padHeadIndex];
-            return Math.floor((inSize + padNeeded - kernel) / stride + 1);
+            return PoolConvUtil.computeOutputSize(
+              inSize + pads[padHeadIndex] + pads[padTailIndex] - dkernel,
+              stride,
+              inSize,
+              pads[padHeadIndex],
+              ceilMode,
+            );
           }
         default:
           throw new Error('Unsupported AutoPad type');
       }
     } else {
-      return Math.floor((inSize + pads[padHeadIndex] + pads[padTailIndex] - dkernel) / stride + 1);
+      return PoolConvUtil.computeOutputSize(
+        inSize + pads[padHeadIndex] + pads[padTailIndex] - dkernel,
+        stride,
+        inSize,
+        pads[padHeadIndex],
+        ceilMode,
+      );
     }
   }
 }

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

@@ -366,6 +366,8 @@ export class PoolConvUtil {
    * @param pads Padding for the beginning and ending along each axis.
    * @param autoPad DEPRECATED attribute supported for legacy models. Specifies how to implicitly calculate pads in each
    *     dimension. Can take values NOTSET, SAME_UPPER, SAME_LOWER, or VALID.
+   * @param ceilMode When set to 1, use ceil() instead of floor() to compute the output spatial size (and apply the
+   *     "shrink the last window if it starts entirely in padding" rule). Defaults to 0 (floor).
    */
   static computePoolOutputShape(
     isGlobalOperator: boolean,
@@ -375,6 +377,7 @@ export class PoolConvUtil {
     kernelShape: number[],
     pads: number[],
     autoPad?: string,
+    ceilMode = 0,
   ): number[] {
     if (inputDims.length <= 0) {
       throw new Error('input shape must be of size greater than 0');
@@ -392,6 +395,7 @@ export class PoolConvUtil {
       kernelShape,
       pads,
       autoPad,
+      ceilMode,
     );
     return outputDims;
   }
@@ -438,6 +442,7 @@ export class PoolConvUtil {
     kernelShape: readonly number[],
     pads: number[],
     autoPad?: string,
+    ceilMode = 0,
   ) {
     if (isGlobalOperator) {
       for (let dim = 0; dim < inputDims.length - 2; dim++) {
@@ -455,12 +460,41 @@ export class PoolConvUtil {
             dim,
             dim + inputDims.length - 2,
             autoPad,
+            ceilMode,
           ),
         );
       }
     }
   }
 
+  // Computes the output spatial size for a single dimension.
+  // Produces results identical to the C++ PoolAttributes::ComputeOutputSize
+  // (onnxruntime/core/providers/cpu/nn/pool_attributes.h), including the ceil_mode
+  // "shrink the last window if it starts entirely in the trailing padding" rule. The JS
+  // signature takes a pre-computed `numerator` (equal to `inSize + padHead + padTail - dkernel`,
+  // matching the C++ `in_size + pad_head + pad_tail - dilation * (kernel - 1) - 1`) instead of
+  // the raw pooling attributes, but the computed output size is the same.
+  // Keep in sync with the onnxjs/jsep copy.
+  private static computeOutputSize(
+    numerator: number,
+    stride: number,
+    inSize: number,
+    padHead: number,
+    ceilMode: number,
+  ): number {
+    let outSize = Math.floor(numerator / stride) + 1;
+    // Match C++ `ceil_mode == 1` exactly so out-of-spec ceil_mode values do not diverge.
+    if (ceilMode === 1) {
+      outSize = Math.ceil(numerator / stride) + 1;
+      // Ensure the last pooling window starts inside the image (ref: https://github.com/onnx/onnx/pull/5741).
+      // inSize and padHead are needed here to reconstruct the last window's start position.
+      if ((outSize - 1) * stride >= inSize + padHead) {
+        outSize -= 1;
+      }
+    }
+    return outSize;
+  }
+
   // helper for computeShapeHelper() and adjustPadsBasedOnAutoPad()
   // adjusts pad value for given 'autoPad' string and computes output shape along a particular dimension
   private static adjustPadAndReturnShape(
@@ -472,30 +506,44 @@ export class PoolConvUtil {
     padHeadIndex: number,
     padTailIndex: number,
     autoPad?: string,
+    ceilMode = 0,
   ): number {
     const dkernel = dilation * (kernel - 1) + 1;
     if (autoPad && autoPad !== 'NOTSET') {
       switch (autoPad) {
         case 'VALID':
           pads[padHeadIndex] = 0;
           pads[padTailIndex] = 0;
-          return Math.floor((inSize - dkernel) / stride + 1);
+          return PoolConvUtil.computeOutputSize(inSize - dkernel, stride, inSize, 0, ceilMode);
         case 'SAME_LOWER':
         case 'SAME_UPPER':
           if (dilation !== 1) {
             throw new Error('Dilation not supported for SAME_UPPER or SAME_LOWER');
           } else {
-            const legacyTargetSize = (inSize + stride - 1) / stride;
+            // Integer division to match C++ pool_attributes.h ComputeSizePadDilations; float division mis-rounds SAME_* pads.
+            const legacyTargetSize = Math.floor((inSize + stride - 1) / stride);
             const padNeeded = (legacyTargetSize - 1) * stride + kernel - inSize;
             pads[padHeadIndex] = autoPad === 'SAME_LOWER' ? Math.floor((padNeeded + 1) / 2) : Math.floor(padNeeded / 2);
             pads[padTailIndex] = padNeeded - pads[padHeadIndex];
-            return Math.floor((inSize + padNeeded - kernel) / stride + 1);
+            return PoolConvUtil.computeOutputSize(
+              inSize + pads[padHeadIndex] + pads[padTailIndex] - dkernel,
+              stride,
+              inSize,
+              pads[padHeadIndex],
+              ceilMode,
+            );
           }
         default:
           throw new Error('Unsupported AutoPad type');
       }
     } else {
-      return Math.floor((inSize + pads[padHeadIndex] + pads[padTailIndex] - dkernel) / stride + 1);
+      return PoolConvUtil.computeOutputSize(
+        inSize + pads[padHeadIndex] + pads[padTailIndex] - dkernel,
+        stride,
+        inSize,
+        pads[padHeadIndex],
+        ceilMode,
+      );
     }
   }
 }

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

@@ -20,7 +20,8 @@ import {
 } from './common';
 
 // TODO: support:
-// - ceil_mode                 "test_maxpool_2d_ceil"
+// - ceil_mode kernel execution "test_maxpool_2d_ceil" (output SHAPE already honors ceil_mode
+//   via PoolConvUtil.computePoolOutputShape; the WebGPU kernel padding/divisor handling is pending)
 // - storage_order             "test_maxpool_with_argmax_2d_precomputed_strides"
 // - [MaxPool] dilations       "test_maxpool_2d_dilations"
 // - [MaxPool] output[1]       "test_maxpool_with_argmax_2d_precomputed_pads"
@@ -56,6 +57,7 @@ const getAdjustedPoolAttributesAndOutputShape = <AttributeType extends AveragePo
     kernelShape,
     pads,
     attributes.autoPad,
+    attributes.ceilMode,
   );
 
   const newAttributes = Object.assign({}, attributes);
@@ -391,9 +393,15 @@ export const parseAveragePoolAttributes = (attributes: Record<string, unknown>):
   const countIncludePad = (attributes.count_include_pad as number) === 0 ? false : true;
 
   const attr = parsePoolCommonAttributes(attributes);
-  // TODO: support attribute 'ceil_mode'
+  // ceil_mode is honored by the output-shape math (PoolConvUtil.computePoolOutputShape now
+  // threads ceilMode and matches the C++ reference), but end-to-end execution stays gated
+  // here: the WebGPU pooling kernel must also apply the ceil_mode trailing-padding handling
+  // (and, for AveragePool, the count_include_pad divisor) before this throw can be removed.
+  // Tracked follow-up: remove this guard + add kernel ceil_mode padding support.
   if (attr.ceilMode !== 0) {
-    throw new Error('using ceil() in shape computation is not yet supported for AveragePool');
+    throw new Error(
+      'ceil_mode output-shape is computed, but ceil_mode kernel execution (padding/divisor) is not yet implemented in the WebGPU AveragePool kernel',
+    );
   }
   const averagePoolAttributes = { countIncludePad, ...attr, cacheKey: '' };
   return { ...averagePoolAttributes, cacheKey: createAveragePoolShaderKeyFromAttributes(averagePoolAttributes) };
@@ -491,12 +499,19 @@ export const parseMaxPoolAttributes = (attributes: Record<string, unknown>): Max
   const dilations = attributes.dilations as [number, number];
 
   const attr = parsePoolCommonAttributes(attributes);
-  // TODO: support attribute 'ceil_mode' and 'storage_order'
+  // TODO: support attribute 'storage_order'
   if (storageOrder !== 0) {
     throw new Error('column major storage order is not yet supported for MaxPool');
   }
+  // ceil_mode is honored by the output-shape math (PoolConvUtil.computePoolOutputShape now
+  // threads ceilMode and matches the C++ reference), but end-to-end execution stays gated
+  // here: the WebGPU pooling kernel must also apply the ceil_mode trailing-padding handling
+  // before this throw can be removed. Tracked follow-up: remove this guard + add kernel
+  // ceil_mode padding support.
   if (attr.ceilMode !== 0) {
-    throw new Error('using ceil() in shape computation is not yet supported for MaxPool');
+    throw new Error(
+      'ceil_mode output-shape is computed, but ceil_mode kernel execution (padding) is not yet implemented in the WebGPU MaxPool kernel',
+    );
   }
   const maxPoolAttributes = { storageOrder, dilations, ...attr, cacheKey: '' };
   return { ...maxPoolAttributes, cacheKey: createMaxPoolShaderKeyFromAttributes(maxPoolAttributes) };

js/web/test/unittests/index.ts

@@ -13,4 +13,6 @@ if (typeof window !== 'undefined') {
 
 require('./backends/wasm/test-model-metadata');
 
+require('./pool-output-shape');
+
 require('./opset');