From 3689fcd5915cd902cb4ea5f618f2a6e42f6dc4a1 Mon Sep 17 00:00:00 2001 From: Manuel Bottini Date: Fri, 14 Jun 2019 17:18:12 +0100 Subject: COMPMID-2408: Add QSYMM16 support for ElementwiseAddition for NEON Change-Id: I22991e9369ffba9b51a94522ff4977933e887b94 Signed-off-by: Manuel Bottini Reviewed-on: https://review.mlplatform.org/c/1352 Comments-Addressed: Arm Jenkins Reviewed-by: Michele Di Giorgio Tested-by: Arm Jenkins Reviewed-by: Giuseppe Rossini --- .../NEON/kernels/NEArithmeticAdditionKernel.cpp | 183 ++++++++++++++++++++- 1 file changed, 177 insertions(+), 6 deletions(-) (limited to 'src/core/NEON/kernels') diff --git a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp index 164026c1ab..733d1372a1 100644 --- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp +++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp @@ -335,6 +335,172 @@ void add_QASYMM8_QASYMM8_QASYMM8(const ITensor *in1, const ITensor *in2, ITensor } } +void add_QSYMM16_QSYMM16_QSYMM16(const ITensor *in1, const ITensor *in2, ITensor *out, ConvertPolicy policy, const Window &window) +{ + ARM_COMPUTE_UNUSED(policy); + + // Create input windows + Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()); + Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()); + + // Clear X Dimension on execution window as we handle manually + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + const int window_step_x = 8; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const bool is_broadcast_across_x = (input1_win.x().step() == 0) || (input2_win.x().step() == 0); + + const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform(); + const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = out->info()->quantization_info().uniform(); + + const float32x4_t vscale1 = vdupq_n_f32(iq1_info.scale); + const float32x4_t vscale2 = vdupq_n_f32(iq2_info.scale); + const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale); + + if(is_broadcast_across_x) + { + const bool is_broadcast_input_2 = input2_win.x().step() == 0; + Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win; + Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win; + const ITensor *broadcast_tensor = is_broadcast_input_2 ? in2 : in1; + const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1; + const UniformQuantizationInfo broadcast_qinfo = broadcast_tensor->info()->quantization_info().uniform(); + const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform(); + + // Clear X Dimension on execution window as we handle manually + non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator broadcast_input(broadcast_tensor, broadcast_win); + Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); + Iterator output(out, win); + + execute_window_loop(win, [&](const Coordinates &) + { + const auto non_broadcast_input_ptr = reinterpret_cast(non_broadcast_input.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + const int16_t broadcast_value = *reinterpret_cast(broadcast_input.ptr()); + const int16x8_t broadcast_value_vec = vdupq_n_s16(broadcast_value); + + const float32x4x2_t bf = + { + { + vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(broadcast_value_vec))), vscale2), + vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(broadcast_value_vec))), vscale2), + } + }; + const float bfs = static_cast(broadcast_value) * broadcast_qinfo.scale; + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const int16x8_t a = vld1q_s16(non_broadcast_input_ptr + x); + const float32x4x2_t af = + { + { + vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1), + vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1), + } + }; + + const int32x4x4_t rf = + { + { +#ifdef __aarch64__ + vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af.val[0], bf.val[0]), invvscaleo)), + vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af.val[1], bf.val[1]), invvscaleo)), +#else //__aarch64__ + vcvtq_s32_f32(vmulq_f32(vaddq_f32(af.val[0], bf.val[0]), invvscaleo)), + vcvtq_s32_f32(vmulq_f32(vaddq_f32(af.val[1], bf.val[1]), invvscaleo)), +#endif //__aarch64__ + } + }; + + const int16x8_t pa = vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])); + vst1q_s16(output_ptr + x, pa); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const float afs = static_cast(*(non_broadcast_input_ptr + x)) * non_broadcast_qinfo.scale; + *(output_ptr + x) = quantize_qsymm16((afs + bfs), oq_info); + } + }, + broadcast_input, non_broadcast_input, output); + } + else + { + // Clear X Dimension on execution window as we handle manually + input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input1(in1, input1_win); + Iterator input2(in2, input2_win); + Iterator output(out, win); + + execute_window_loop(win, [&](const Coordinates &) + { + const auto input1_ptr = reinterpret_cast(input1.ptr()); + const auto input2_ptr = reinterpret_cast(input2.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const int16x8_t a = vld1q_s16(input1_ptr + x); + const int16x8_t b = vld1q_s16(input2_ptr + x); + + const float32x4x2_t af = + { + { + vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1), + vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1), + } + }; + + const float32x4x2_t bf = + { + { + vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(b))), vscale2), + vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(b))), vscale2), + } + }; + + const int32x4x2_t rf = + { + { +#ifdef __aarch64__ + vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af.val[0], bf.val[0]), invvscaleo)), + vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af.val[1], bf.val[1]), invvscaleo)), +#else //__aarch64__ + vcvtq_s32_f32(vmulq_f32(vaddq_f32(af.val[0], bf.val[0]), invvscaleo)), + vcvtq_s32_f32(vmulq_f32(vaddq_f32(af.val[1], bf.val[1]), invvscaleo)), +#endif //__aarch64__ + } + }; + + const int16x8_t pa = vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])); + vst1q_s16(output_ptr + x, pa); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const float afs = static_cast((*(input1_ptr + x))) * iq1_info.scale; + const float bfs = static_cast((*(input2_ptr + x))) * iq2_info.scale; + *(output_ptr + x) = quantize_qsymm16((afs + bfs), out->info()->quantization_info()); + } + }, + input1, input2, output); + } +} + void add_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, ConvertPolicy policy, const Window &window) { // Create input windows @@ -475,8 +641,8 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, ARM_COMPUTE_UNUSED(policy); ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&input1); - 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_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, 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::QSYMM16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32); const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape()); @@ -496,7 +662,8 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, && !(input1.data_type() == DataType::S16 && input2.data_type() == DataType::S16 && output.data_type() == DataType::S16) && !(input1.data_type() == DataType::F32 && input2.data_type() == DataType::F32 && output.data_type() == DataType::F32) && !(input1.data_type() == DataType::F16 && input2.data_type() == DataType::F16 && output.data_type() == DataType::F16) - && !(input1.data_type() == DataType::QASYMM8 && input2.data_type() == DataType::QASYMM8 && output.data_type() == DataType::QASYMM8), + && !(input1.data_type() == DataType::QASYMM8 && input2.data_type() == DataType::QASYMM8 && output.data_type() == DataType::QASYMM8) + && !(input1.data_type() == DataType::QSYMM16 && input2.data_type() == DataType::QSYMM16 && output.data_type() == DataType::QSYMM16), "You called addition with the wrong image formats"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0), @@ -528,10 +695,14 @@ std::pair validate_and_configure_window(ITensorInfo &input1, ITe { set_format_if_unknown(output, Format::F32); } - else if(input1.data_type() == DataType::QASYMM8 || input2.data_type() == DataType::QASYMM8) + else if(input1.data_type() == DataType::QASYMM8) { set_data_type_if_unknown(output, DataType::QASYMM8); } + else if(input1.data_type() == DataType::QSYMM16) + { + set_data_type_if_unknown(output, DataType::QSYMM16); + } } Window win = calculate_max_window(valid_region, Steps()); @@ -540,9 +711,7 @@ std::pair validate_and_configure_window(ITensorInfo &input1, ITe Coordinates coord; coord.set_num_dimensions(output.num_dimensions()); output.set_valid_region(valid_region); - return std::make_pair(Status{}, win); - ; } } // namespace @@ -564,6 +733,8 @@ void NEArithmeticAdditionKernel::configure(const ITensor *input1, const ITensor { { "add_wrap_QASYMM8_QASYMM8_QASYMM8", &add_QASYMM8_QASYMM8_QASYMM8 }, { "add_saturate_QASYMM8_QASYMM8_QASYMM8", &add_QASYMM8_QASYMM8_QASYMM8 }, + { "add_wrap_QSYMM16_QSYMM16_QSYMM16", &add_QSYMM16_QSYMM16_QSYMM16 }, + { "add_saturate_QSYMM16_QSYMM16_QSYMM16", &add_QSYMM16_QSYMM16_QSYMM16 }, { "add_wrap_U8_U8_U8", &add_same }, { "add_saturate_U8_U8_U8", &add_same }, { "add_wrap_S16_U8_S16", &add_S16_U8_S16 }, -- cgit v1.2.1