[WASM] Add ClipByValue

[dsmilkov]

Call into XNNPACK clamp operator for doing clipping.

NOTE: we can't do operator cleanup upon disposal of a tensor, since there are no weights-like tensors for clamp

---

This change is 

[annxingyuan]

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

---

tfjs-backend-wasm/src/cc/kernels/ClipByValue.cc, line 60 at r1 (raw file):

  const int flags = 0;
  OperatorCacheKey cache_key = {
      static_cast<float>(channels),

Just wondering - what are the implications of not performing the cast?

[Maratyszcza]

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

---

tfjs-backend-wasm/src/cc/kernels/ClipByValue.cc, line 71 at r1 (raw file):

    xnn_status status = xnn_create_clamp_nc_f32(channels, strides, strides, min,
                                                max, flags, &clamp_op);
    if (status != xnn_status_success) {

The function should exit early if xnn_create_clamp_nc_f32 failed. It is not safe to call xnn_setup_clamp_nc_f32 and xnn_run_operator with nullptr for the op

[dsmilkov]

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

---

tfjs-backend-wasm/src/cc/kernels/ClipByValue.cc, line 60 at r1 (raw file):

Previously, annxingyuan (Ann Yuan) wrote…

Just wondering - what are the implications of not performing the cast?

compile error (can't auto-cast int to float)

---

tfjs-backend-wasm/src/cc/kernels/ClipByValue.cc, line 71 at r1 (raw file):

Previously, Maratyszcza (Marat Dukhan) wrote…

The function should exit early if xnn_create_clamp_nc_f32 failed. It is not safe to call xnn_setup_clamp_nc_f32 and xnn_run_operator with nullptr for the op

Thanks! #2367 will replace all util::warn calls with throw_js_exception which will result in terminating the program.

[Maratyszcza]

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

---

tfjs-backend-wasm/src/cc/kernels/ClipByValue.cc, line 60 at r1 (raw file):

Previously, dsmilkov (Daniel Smilkov) wrote…

compile error (can't auto-cast int to float)

Why do you use floats as keys? It isn't safe (conversion int->float is lossy, and any NaN would break everything, because (NaN == NaN) == false

[dsmilkov]

PTAL. I also added unit tests

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

---

tfjs-backend-wasm/src/cc/kernels/ClipByValue.cc, line 60 at r1 (raw file):

Previously, Maratyszcza (Marat Dukhan) wrote…

Why do you use floats as keys? It isn't safe (conversion int->float is lossy, and any NaN would break everything, because (NaN == NaN) == false

Because my cache key had a mixture of ints and floats and I didn't know about std::tuple. Changed to std::tuple and removed the casting.

[Maratyszcza]

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

---

tfjs-backend-wasm/src/cc/kernels/ClipByValue.cc, line 58 at r2 (raw file):

  xnn_operator_t clamp_op = nullptr;
  const int channels = x_info.size;

Since you don't use strides, you can set channels = strides = 1 and batch_size = x_info.size. This will get you much better reuse of xnn_operator_t objects, because channels is a part of the key, but batch_size is not.

[dsmilkov]

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

---

tfjs-backend-wasm/src/cc/kernels/ClipByValue.cc, line 58 at r2 (raw file):

Previously, Maratyszcza (Marat Dukhan) wrote…

Since you don't use strides, you can set channels = strides = 1 and batch_size = x_info.size. This will get you much better reuse of xnn_operator_t objects, because channels is a part of the key, but batch_size is not.

Nice! I reduced the cache key to 2 floats (min & max) since all the other arguments are hard-coded to constants.