diff options
Diffstat (limited to 'src/core/CL/kernels/CLQuantizationLayerKernel.cpp')
| -rw-r--r-- | src/core/CL/kernels/CLQuantizationLayerKernel.cpp | 52 |
1 files changed, 47 insertions, 5 deletions
diff --git a/src/core/CL/kernels/CLQuantizationLayerKernel.cpp b/src/core/CL/kernels/CLQuantizationLayerKernel.cpp index 3d7aff0712..ab3b5d271d 100644 --- a/src/core/CL/kernels/CLQuantizationLayerKernel.cpp +++ b/src/core/CL/kernels/CLQuantizationLayerKernel.cpp @@ -41,7 +41,7 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F32, DataType::F16); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); // Output must always be initialized @@ -62,8 +62,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen const bool multi_access_x = (input_width_x / vec_size_x > 0); if(multi_access_x) { - win.set(Window::DimX, - Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x)); + win.set(Window::DimX, Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x)); } Coordinates coord; @@ -99,10 +98,53 @@ void CLQuantizationLayerKernel::configure(const ICLTensor *input, ICLTensor *out const UniformQuantizationInfo qinfo = output->info()->quantization_info().uniform(); const DataType output_data_type = output->info()->data_type(); + float scale_to_apply = qinfo.scale; + int32_t offset_to_apply = qinfo.offset; + if(is_data_type_quantized_asymmetric(_input->info()->data_type())) + { + /* + * In case of requantization of a quantized input tensor to an output tensor with another quantization + * instead of of apply dequantization and then a quantization functions, we just compute new scale and + * offset to apply. + * + * Assuming: + * - q_i as input quantized value + * - q_o as output quantized value + * - z_i as input quantization offset value + * - z_o as output quantization offset value + * - s_i as input quantization scale value + * - s_o as output quantization scale value + * - z_n as new quantization offset value + * - s_n as new quantization scale value + * + * q_o = ( q_i - z_i ) * s_i / s_o + z_o + * + * We can rewrite the formula as: + * + * q_o = ( q_i * s_i / s_o ) - z_i * s_i / s_o + z_o + * + * q_o = q_i / s_n + z_n + * + * Where: + * + * s_n = s_o / s_i + * + * z_n = - z_i * s_i / s_o + z_o + * + */ + const UniformQuantizationInfo qinfo_in = _input->info()->quantization_info().uniform(); + scale_to_apply /= qinfo_in.scale; + // In order to minimize flooring we convert the offset to a float, + // then compute the new offset in the float domain, + // finally we convert it back as int32_t + offset_to_apply -= static_cast<int32_t>(static_cast<float>(qinfo_in.offset) * qinfo_in.scale / qinfo.scale); + } + // Create kernel CLBuildOptions build_opts; - build_opts.add_option("-DSCALE=" + float_to_string_with_full_precision(qinfo.scale)); - build_opts.add_option("-DOFFSET=" + support::cpp11::to_string(qinfo.offset)); + build_opts.add_option_if(is_data_type_float(_input->info()->data_type()), "-DIS_FLOAT"); + build_opts.add_option("-DSCALE=" + float_to_string_with_full_precision(scale_to_apply)); + build_opts.add_option("-DOFFSET=" + support::cpp11::to_string(offset_to_apply)); build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(input->info()->data_type())); build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output_data_type)); |