diff options
Diffstat (limited to 'src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp')
| -rw-r--r-- | src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp | 45 |
1 files changed, 31 insertions, 14 deletions
diff --git a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp index 4ca2ef8aa3..286b94ebdc 100644 --- a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp +++ b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp @@ -51,9 +51,9 @@ Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, ARM_COMPUTE_UNUSED(rounding_policy); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input1); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input2); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(scale < 0, "Scale cannot be negative."); const TensorShape &out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape()); @@ -64,7 +64,7 @@ Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, if(output->total_size() > 0) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::U8 && (input1->data_type() != DataType::U8 || input2->data_type() != DataType::U8), "Output can only be U8 if both inputs are U8"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), "Wrong shape for output"); @@ -168,27 +168,44 @@ void CLPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const I data_type = "DATA_TYPE_INT"; } + const bool is_quantized = is_data_type_quantized_asymmetric(input1->info()->data_type()); + // Construct kernel name std::string kernel_name = "pixelwise_mul"; - kernel_name += (scale_int >= 0) ? "_int" : "_float"; + if(!is_data_type_quantized(output->info()->data_type())) + { + kernel_name += (scale_int >= 0) ? "_int" : "_float"; + } // Set kernel build options - std::set<std::string> build_opts; - build_opts.emplace((overflow_policy == ConvertPolicy::WRAP || is_data_type_float(output->info()->data_type())) ? "-DWRAP" : "-DSATURATE"); - build_opts.emplace((rounding_policy == RoundingPolicy::TO_ZERO) ? "-DROUND=_rtz" : "-DROUND=_rte"); - build_opts.emplace("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type())); - build_opts.emplace("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type())); - build_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); - build_opts.emplace("-DDATA_TYPE_RES=" + compute_type); - build_opts.emplace("-D" + data_type); + CLBuildOptions build_opts; + if(is_quantized) + { + build_opts.add_option("-DOFFSET_IN1=" + support::cpp11::to_string(input1->info()->quantization_info().offset)); + build_opts.add_option("-DOFFSET_IN2=" + support::cpp11::to_string(input2->info()->quantization_info().offset)); + build_opts.add_option("-DOFFSET_OUT=" + support::cpp11::to_string(output->info()->quantization_info().offset)); + build_opts.add_option("-DSCALE_IN1=" + support::cpp11::to_string(input1->info()->quantization_info().scale)); + build_opts.add_option("-DSCALE_IN2=" + support::cpp11::to_string(input2->info()->quantization_info().scale)); + build_opts.add_option("-DSCALE_OUT=" + support::cpp11::to_string(output->info()->quantization_info().scale)); + kernel_name += "_quantized"; + } + else + { + build_opts.add_option_if_else(overflow_policy == ConvertPolicy::WRAP || is_data_type_float(output->info()->data_type()), "-DWRAP", "-DSATURATE"); + build_opts.add_option_if_else(rounding_policy == RoundingPolicy::TO_ZERO, "-DROUND=_rtz", "-DROUND=_rte"); + build_opts.add_option("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type())); + build_opts.add_option("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type())); + build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); + build_opts.add_option("-DDATA_TYPE_RES=" + compute_type); + } // Create kernel - _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts)); + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); // Set scale argument unsigned int idx = 3 * num_arguments_per_3D_tensor(); //Skip the inputs and output parameters - if(scale_int >= 0) + if(scale_int >= 0 && !is_quantized) { _kernel.setArg(idx++, scale_int); } |