[wasm] Add AvgPool3D and MaxPool3D kernels

[chunnienc]

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

---

This change is 

[pyu10055]

Reviewable status: 0 of 1 approvals obtained (waiting on @chunnienc and @mattsoulanille)

---

tfjs-backend-wasm/src/cc/pool3d_impl.h line 1 at r2 (raw file):

#ifdef __EMSCRIPTEN__

missing copyright header?

---

tfjs-backend-wasm/src/cc/pool3d_impl.cc line 1 at r1 (raw file):

#ifdef __EMSCRIPTEN__

missing copyright header

---

tfjs-backend-wasm/src/cc/pool3d_impl.cc line 45 at r1 (raw file):

};
template <typename OUT, typename IN, typename T>
inline void Pool3DImpl(

is this method used? seems to be a simplified version of NDHWCPool3DImpl?

[chunnienc]

Added license header. I should have deleted the file pool3d_impl.cc.

[pyu10055]

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @chunnienc and @mattsoulanille)

---

tfjs-backend-wasm/src/cc/pool3d_impl.h line 32 at r3 (raw file):

    return v;
  }
  return (v % d + v) % d;

this does not guarantee non negative, right?
maybe the method name need to be updated to better reflect the logic?

---

tfjs-backend-wasm/src/cc/pool3d_impl.h line 78 at r3 (raw file):

};
template <typename IN, typename OUT, typename FI, typename FAP, typename FAG>
inline void NDHWCPool3DImpl(const IN* x_buf, OUT* out_buf,

like the code sharing across ops.

[chunnienc]

this does not guarantee non negative, right?
maybe the method name need to be updated to better reflect the logic?

This method is intended to get the non-negative modulo. But you're right, the implementation was wrong. Fixed, thanks for the review and pointing out.

[mattsoulanille]

Reviewed 8 of 9 files at r2, 1 of 1 files at r3, 1 of 1 files at r4.
Reviewable status: complete! 2 of 1 approvals obtained (waiting on @chunnienc)

---

tfjs-backend-wasm/src/cc/pool3d_impl.h line 120 at r3 (raw file):

            int out_offset =
                info.out_offset(batch, y_depth, y_row, y_col, channel);
            out_buf[out_offset] = filter_aggregate(filter_data);

Thinking about future improvements, addition (in AvgPool3D) is associative, so I wonder if this could be parallelized or vectorized by xnnpack. That might not matter if this op is not the main source of slowness in models, though.

Code quote:

                  filter_apply(filter_data, x_buf[x_offset]);
                }
              }
            }
            int out_offset =
                info.out_offset(batch, y_depth, y_row, y_col, channel);
            out_buf[out_offset] = filter_aggregate(filter_data);

[pyu10055]

dition (in AvgPool3D) is associative, so I wonder if this could be parallelized or vectorized by xnnpack. That might not mat
LGTM

Reviewed 8 of 9 files at r2, 1 of 1 files at r4, all commit messages.
Reviewable status: complete! 2 of 1 approvals obtained (waiting on @chunnienc)

---

tfjs-backend-wasm/src/cc/pool3d_impl.h line 120 at r3 (raw file):

Previously, mattsoulanille (Matthew Soulanille) wrote…

Thinking about future improvements, addition (in AvgPool3D) is associative, so I wonder if this could be parallelized or vectorized by xnnpack. That might not matter if this op is not the main source of slowness in models, though.

does Xnnpack have pooling op?

[mattsoulanille]

Reviewable status: complete! 2 of 1 approvals obtained (waiting on @chunnienc)

---

tfjs-backend-wasm/src/cc/pool3d_impl.h line 120 at r3 (raw file):

Previously, pyu10055 (Ping Yu) wrote…

does Xnnpack have pooling op?

XNNPACK has 2D Pooling ops. I don't know if they could be used as part of a 3D pooling implementation. Also, I don't know if they have a general pooling utility like the one in this PR.

https://github.com/google/XNNPACK#operator-coverage

[chunnienc]

tflite has its own implementation of pooling3d (/lite/kernels/pooling3d.cc), and there is no xnnpack delegate for it. But I think it's more because MaxPool3D and AvgPool3D are not yet built-in ops there.

Similarly, conv3d is a built-in op in tflite but xnnpack delegate does not have it either. So I assume 3d implementation with xnnpack 2d is non-trivial. I will revisit this topic (using xnnpack for 3d ops) later if there is a need.

[mattsoulanille]

tflite has its own implementation of pooling3d (/lite/kernels/pooling3d.cc), and there is no xnnpack delegate for it. But I think it's more because MaxPool3D and AvgPool3D are not yet built-in ops there.

Similarly, conv3d is a built-in op in tflite but xnnpack delegate does not have it either. So I assume 3d implementation with xnnpack 2d is non-trivial. I will revisit this topic (using xnnpack for 3d ops) later if there is a need.

SGTM. Feel free to merge this whenever. I didn't mean to block it.