CUDA Resize-18 implementation

[yuslepukhin]

Description

Implement Resize-18 on CUDA.

Motivation and Context

Performance

[gedoensmax]

@yuslepukhin or @tianleiwu can you elaborate on the check used here:

onnxruntime/onnxruntime/core/providers/cuda/tensor/upsample.cc

Lines 180 to 184 in 5ee62a6

	if (!is_3D) {
	if (!(scales[0] == 1.0f && scales[1] == 1.0f)) {
	return ORT_MAKE_STATUS(ONNXRUNTIME, NOT_IMPLEMENTED, "Resize", ": NDHWC is not supported yet");
	}
	}

I am not clear on why this is a safe check for NDHWC vs NCDHW as the scales for D and C are often both == 1.0f as suggested by some unit tests:

onnxruntime/onnxruntime/test/providers/cpu/tensor/resize_op_test.cc

Line 779 in 8293aa1

std::vector<float> scales{1.0f, 1.0f, 2.0f, 2.0f, 1.0f};

I am looking at this to support the operators fully with channel last.

[gedoensmax]

I also noticed that I believe the resize kernel is not sufficiently tested for NCHW + int8/uint8 cases in CUDA EP:

onnxruntime/onnxruntime/core/providers/cuda/tensor/resize_impl.cu

Lines 328 to 334 in 5ee62a6

	float y_offset_1 = 1.0f - y_offset_0;
	float x_offset_1 = 1.0f - x_offset_0;
	output_data[id] =
	x00 * static_cast<T>(y_offset_1 * x_offset_1) +
	x01 * static_cast<T>(y_offset_0 * x_offset_1) +
	x10 * static_cast<T>(y_offset_1 * x_offset_0) +
	x11 * static_cast<T>(y_offset_0 * x_offset_0);

Judging by this line the result for int kernels will alway be 0. I verified this by adding the following unites which is the same as: NhwcResizeOpLinearDownSampleTest_4DBilinear_uint8 Please let me know if this sounds sensible then I will go ahead a file an issue for this case.

TEST(ResizeOpTest, ResizeOpLinearDownSampleTest_4DBilinear_uint8) {
  OpTester test("Resize", 13);
  std::vector<float> roi{};
  std::vector<float> scales{1.0f, 1.0f, 0.6f, 0.6f};

  test.AddAttribute("mode", "linear");

  constexpr int64_t N = 1, H = 2, W = 4, C = 1;
  std::vector<uint8_t> X = {
      1, 2, 3, 4,
      5, 6, 7, 8};

  test.AddInput<uint8_t>("X", {N, C, H, W}, X);
  test.AddInput<float>("roi", {0}, roi);
  test.AddInput<float>("scales", {4}, scales);

  std::vector<uint8_t> Y = {2, 4};

  test.AddOutput<uint8_t>("Y", {N, C, static_cast<int64_t>(H * scales[2]), static_cast<int64_t>(W * scales[3])}, Y);
  // ROCm: results mismatch
  test.Run(OpTester::ExpectResult::kExpectSuccess, "",
           {kRocmExecutionProvider});
}

[yuslepukhin]

Go ahead and file an issue, and we will look into it. If you have a suggested change, do not hesitate to propose.