[wasm] Add AvgPool kernel.

[annxingyuan]

Changes

• Add AvgPool kernel to wasm.
• Exclude worker_node_test from bundle in wasm.

To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.

---

This change is 

[Maratyszcza]

Reviewable status: 0 of 1 approvals obtained (waiting on @annxingyuan, @dsmilkov, and @Maratyszcza)

---

tfjs-backend-wasm/src/cc/kernels/AvgPool.cc, line 47 at r1 (raw file):

             const int filter_width, int pad_top, int pad_right, int pad_bottom,
             int pad_left, const int stride_height, const int stride_width,
             const int input_channels, const int output_channels,

input_channels == output_channels in pooling ops. Would be better to replace it with a single channels argument.

---

tfjs-backend-wasm/src/cc/kernels/AvgPool.cc, line 82 at r1 (raw file):

          "XNN status for xnn_create_average_pooling2d_nhwc_f32 is not "
          "successful. ",
          "Got status %d. Use -c dbg to see XNN logs.", status);

The function should return early in case of error. It is not safe to pass avg_pool_op = nullptr to xnn_setup_average_pooling2d_nhwc_f32.

[dsmilkov]

Reviewed 7 of 9 files at r1.
Reviewable status: complete! 1 of 1 approvals obtained (waiting on @annxingyuan and @dsmilkov)

---

tfjs-backend-wasm/src/cc/kernels/AvgPool.cc, line 63 at r1 (raw file):

                                pad_left,       filter_height,   filter_width,
                                stride_height,  stride_width,    channels,
                                input_channels, output_channels, flags};

See Marat's comment above. This means you can reduce the cache_key by keeping channels and dropping input_channels and output_channels

---

tfjs-core/src/ops/pool.ts, line 197 at r1 (raw file):

  if (convInfo.filterWidth === 1 && convInfo.filterHeight === 1 &&
      util.arraysEqual(convInfo.inShape, convInfo.outShape) &&
      convInfo.padInfo.type === 'VALID') {

maybe you don't need pad to be valid? not sure though, double check me on this.

[annxingyuan]

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @dsmilkov and @Maratyszcza)

---

tfjs-backend-wasm/src/cc/kernels/AvgPool.cc, line 47 at r1 (raw file):

Previously, Maratyszcza (Marat Dukhan) wrote…

input_channels == output_channels in pooling ops. Would be better to replace it with a single channels argument.

Done

---

tfjs-backend-wasm/src/cc/kernels/AvgPool.cc, line 63 at r1 (raw file):

Previously, dsmilkov (Daniel Smilkov) wrote…

See Marat's comment above. This means you can reduce the cache_key by keeping channels and dropping input_channels and output_channels

I changed input_channels / output_channels to input_pixel_stride / output_pixel_stride to match XNN pack argument names. Currently they're set to input_channels and so are identical, but i think it makes sense to key on all the arguments to the XNN operator.

---

tfjs-backend-wasm/src/cc/kernels/AvgPool.cc, line 82 at r1 (raw file):

Previously, Maratyszcza (Marat Dukhan) wrote…

The function should return early in case of error. It is not safe to pass avg_pool_op = nullptr to xnn_setup_average_pooling2d_nhwc_f32.

Done

---

tfjs-core/src/ops/pool.ts, line 197 at r1 (raw file):

Previously, dsmilkov (Daniel Smilkov) wrote…

maybe you don't need pad to be valid? not sure though, double check me on this.

Actually we don't need a padding condition at all since we're checking that input / output shapes match.

[dsmilkov]

Nice! Thanks for relaxing maxPool as well

Reviewed 6 of 6 files at r2.
Reviewable status: complete! 1 of 1 approvals obtained (waiting on @Maratyszcza)