Why is setting rounding mode invalid

[dataowner888]

All operators (such as conv2d) convert floating-point to integer using the RTNE mode of rounding by default. I set conv2d to TO_ ZERO mode, there is no difference in the calculation results. Please help check the cause of the issue. The code is as follows:

` TEST(Conv2d, mytest) { auto ctx = tim::vx::Context::Create(); auto graph = ctx->CreateGraph(); tim::vx::ShapeType input_shape({1846, 48, 3, 3}); tim::vx::ShapeType weight_shape({3, 3, 3, 16}); tim::vx::ShapeType bias_shape({16}); tim::vx::ShapeType output_shape({923, 24, 16, 3});

tim::vx::TensorSpec input_spec(tim::vx::DataType::INT8, input_shape, tim::vx::TensorAttribute::INPUT); // tim::vx::Quantization quantization_input;

std::vector scales_input = {0.007874};

quantization_input.SetType(tim::vx::QuantType::ASYMMETRIC); quantization_input.SetScales(scales_input); quantization_input.SetZeroPoints(std::vector(1, 0)); // input_spec.SetQuantization(quantization_input);

tim::vx::TensorSpec weight_spec(tim::vx::DataType::INT8, weight_shape, tim::vx::TensorAttribute::CONSTANT); // tim::vx::Quantization quantization_weight; std::vector scales_weight(16); readFromFile((unsigned char )(scales_weight.data()), 164, (char *)"quant_scale_1.bin"); quantization_weight.SetType(tim::vx::QuantType::SYMMETRIC_PER_CHANNEL); quantization_weight.SetChannelDim(3); quantization_weight.SetScales(scales_weight); quantization_weight.SetZeroPoints(std::vector(16, 0)); // weight_spec.SetQuantization(quantization_weight);

tim::vx::TensorSpec bias_spec(tim::vx::DataType::INT32, bias_shape, tim::vx::TensorAttribute::CONSTANT); // tim::vx::Quantization quantization_bias; std::vector scales_bias(16); readFromFile((unsigned char )(scales_bias.data()), 164, (char *)"quant_scale_2.bin"); quantization_bias.SetType(tim::vx::QuantType::SYMMETRIC_PER_CHANNEL); quantization_bias.SetChannelDim(0); quantization_bias.SetScales(scales_bias); quantization_bias.SetZeroPoints(std::vector(16, 0)); //

bias_spec.SetQuantization(quantization_bias);

tim::vx::TensorSpec output_spec(tim::vx::DataType::INT8, output_shape, tim::vx::TensorAttribute::OUTPUT); // tim::vx::Quantization quantization_output;

std::vector scales_output = {0.041692};

quantization_output.SetType(tim::vx::QuantType::ASYMMETRIC); quantization_output.SetScales(scales_output); quantization_output.SetZeroPoints(std::vector(1, 0)); // output_spec.SetQuantization(quantization_output);

printf("read data from file\n");

std::vector input_data(input_shape[0]input_shape[1]input_shape[2]input_shape[3]); readFromFile(input_data.data(), input_data.size(), (char )"conv2d_input.bin"); std::vector weight_data(weight_shape[0]weight_shape[1]weight_shape[2]weight_shape[3]); readFromFile(weight_data.data(), weight_data.size(), (char )"conv2d_weight.bin"); std::vector bias_data(bias_shape[0]); readFromFile((unsigned char )(bias_data.data()), bias_data.size()4, (char )"conv2d_bias.bin"); std::vector golden(output_shape[0]output_shape[1]output_shape[2]output_shape[3]);

auto input_tensor = graph->CreateTensor(input_spec); auto weight_tensor = graph->CreateTensor(weight_spec, weight_data.data()); auto bias_tensor = graph->CreateTensor(bias_spec, bias_data.data()); auto output_tensor = graph->CreateTensor(output_spec);

//auto padding = tim::vx::PadType::AUTO; std::array<uint32_t, 2> stride({2, 2}); std::array<uint32_t, 2> dilation({1, 1}); std::array<uint32_t, 4> padval({1,1,1,1}); auto conv2d = graph->CreateOperation( padval, stride, dilation);

(*conv2d) .BindInput(input_tensor) .BindInput(weight_tensor) .BindInput(bias_tensor) .BindOutput(output_tensor);

conv2d->SetRoundingPolicy(tim::vx::OverflowPolicy::SATURATE, tim::vx::RoundingPolicy::TO_ZERO, tim::vx::RoundType::FLOOR);

EXPECT_TRUE(graph->Compile());

input_tensor->CopyDataToTensor(input_data.data());

EXPECT_TRUE(graph->Run());

uint32_t output_size = 1; for (auto i : output_tensor->GetShape()) { output_size *= i; } std::vector output(output_size); EXPECT_TRUE(output_tensor->CopyDataFromTensor(output.data()));

writeToFile(output.data(), output.size(), (char *)"conv2d_output.bin"); printf("status: ok\n");

EXPECT_EQ(golden, output); }`

[dataowner888]

@sunshinemyson Can you check the issue list daily to respond to customer issues faster

[sunshinemyson]

@dataowner888 ,

I'll check it with team. Thanks for raise your question.

[sunshinemyson]

@dataowner888 ,

This configuration is not valid anymore, we can not change it from application level.

Thanks

[dataowner888]

@sunshinemyson But I do need to change the floating-point rounding mode, and the OpenVX SPEC requires vxConvolutionLayer to support changing the floating-point rounding mode. Please provide me with a method to change the floating-point rounding mode. Thank you

[dataowner888]

@sunshinemyson Are there any available methods?

[sunshinemyson]

@dataowner888 ,

We don't have this feature. sorry for the inconvenience.