diff options
Diffstat (limited to 'src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp')
| -rw-r--r-- | src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp | 165 |
1 files changed, 164 insertions, 1 deletions
diff --git a/src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp b/src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp index 30caa4ebeb..2b23e46f40 100644 --- a/src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp +++ b/src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2022 Arm Limited. + * Copyright (c) 2018-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,6 +22,8 @@ * SOFTWARE. */ #include "src/cpu/kernels/elementwise_unary/generic/neon/impl.h" +#include "arm_compute/core/Helpers.h" +#include "src/core/NEON/NEAsymm.h" namespace arm_compute { @@ -111,5 +113,166 @@ template void elementwise_op<__fp16>(const ITensor *in, ITensor *out, const Wind template void elementwise_op<float>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op); template void elementwise_op<int32_t>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op); +template <> +void elementwise_op<int8_t>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op) +{ + const int window_step_x = 16; + const auto window_start_x = static_cast<int>(window.x().start()); + const auto window_end_x = static_cast<int>(window.x().end()); + const UniformQuantizationInfo qi_in = in->info()->quantization_info().uniform(); + const UniformQuantizationInfo qi_out = out->info()->quantization_info().uniform(); + const auto min_clamped_value = vdupq_n_f32((-128 - qi_out.offset) * qi_out.scale); + const auto max_clamped_value = vdupq_n_f32((127 - qi_out.offset) * qi_out.scale); + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(in, win); + Iterator output(out, win); + + execute_window_loop(win, [&](const Coordinates &) + { + int8x16_t vout; + auto output_ptr = reinterpret_cast<int8_t *>(output.ptr()); + const auto input_ptr = reinterpret_cast<const int8_t *>(input.ptr()); + const auto vconst_0_f32 = vdupq_n_f32(0); + auto clamped_value = (op == ElementWiseUnary::LOG) ? min_clamped_value : max_clamped_value; + + int x = window_start_x; + for(; x <= window_end_x - window_step_x; x += window_step_x) + { + const auto vin = wrapper::vloadq(input_ptr + x); + + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + + // Perform activation + float32x4x4_t vtmp_deq = + { + { + elementwise_op_imp<float>(op, vin_deq.val[0]), + elementwise_op_imp<float>(op, vin_deq.val[1]), + elementwise_op_imp<float>(op, vin_deq.val[2]), + elementwise_op_imp<float>(op, vin_deq.val[3]), + } + }; + + if((op == ElementWiseUnary::LOG) || (op == ElementWiseUnary::RSQRT)) + { + vtmp_deq.val[0] = vbslq_f32(vcleq_f32(vin_deq.val[0], vconst_0_f32), clamped_value, vtmp_deq.val[0]); + vtmp_deq.val[1] = vbslq_f32(vcleq_f32(vin_deq.val[1], vconst_0_f32), clamped_value, vtmp_deq.val[1]); + vtmp_deq.val[2] = vbslq_f32(vcleq_f32(vin_deq.val[2], vconst_0_f32), clamped_value, vtmp_deq.val[2]); + vtmp_deq.val[3] = vbslq_f32(vcleq_f32(vin_deq.val[3], vconst_0_f32), clamped_value, vtmp_deq.val[3]); + } + + // Re-quantize to new output space + vout = vquantize_signed(vtmp_deq, qi_out); + wrapper::vstore(output_ptr + x, vout); + } + for(; x < window_end_x; ++x) + { + qasymm8_signed_t in = *(reinterpret_cast<const qasymm8_signed_t *>(input_ptr + x)); + qasymm8_signed_t tmp = 0; + float tmp_f = dequantize_qasymm8_signed(in, qi_in); + if(tmp_f <= 0.0) + { + if(op == ElementWiseUnary::LOG) + { + tmp_f = (-128 - qi_out.offset) * qi_out.scale; + } + else if(op == ElementWiseUnary::RSQRT) + { + tmp_f = (127 - qi_out.offset) * qi_out.scale; + } + } + else + { + tmp_f = elementwise_op_scalar_imp<float>(op, tmp_f); + } + tmp = quantize_qasymm8_signed(tmp_f, qi_out, RoundingPolicy::TO_ZERO); // Set rounding policy TO_ZERO to be compatible with vquantize_signed() used above that follow same policy for armv7a. + // For aarch64 LUT is used and rounding to nearest is used + *(output_ptr + x) = tmp; + } + }, + input, output); +} +template <> +void elementwise_op<uint8_t>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op) +{ + const int window_step_x = 16; + const auto window_start_x = static_cast<int>(window.x().start()); + const auto window_end_x = static_cast<int>(window.x().end()); + const UniformQuantizationInfo qi_in = in->info()->quantization_info().uniform(); + const UniformQuantizationInfo qi_out = out->info()->quantization_info().uniform(); + const auto vconst_0_f32 = vdupq_n_f32(0); + const auto min_clamped_value = vdupq_n_f32((0 - qi_out.offset) * qi_out.scale); + const auto max_clamped_value = vdupq_n_f32((255 - qi_out.offset) * qi_out.scale); + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(in, win); + Iterator output(out, win); + + execute_window_loop(win, [&](const Coordinates &) + { + uint8x16_t vout; + auto clamped_value = (op == ElementWiseUnary::LOG) ? min_clamped_value : max_clamped_value; + auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr()); + const auto input_ptr = reinterpret_cast<const uint8_t *>(input.ptr()); + int x = window_start_x; + for(; x <= window_end_x - window_step_x; x += window_step_x) + { + const auto vin = wrapper::vloadq(input_ptr + x); + + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + + // Perform activation + float32x4x4_t vtmp_deq = + { + { + elementwise_op_imp<float>(op, vin_deq.val[0]), + elementwise_op_imp<float>(op, vin_deq.val[1]), + elementwise_op_imp<float>(op, vin_deq.val[2]), + elementwise_op_imp<float>(op, vin_deq.val[3]), + } + }; + if((op == ElementWiseUnary::LOG) || (op == ElementWiseUnary::RSQRT)) + { + vtmp_deq.val[0] = vbslq_f32(vcleq_f32(vin_deq.val[0], vconst_0_f32), clamped_value, vtmp_deq.val[0]); + vtmp_deq.val[1] = vbslq_f32(vcleq_f32(vin_deq.val[1], vconst_0_f32), clamped_value, vtmp_deq.val[1]); + vtmp_deq.val[2] = vbslq_f32(vcleq_f32(vin_deq.val[2], vconst_0_f32), clamped_value, vtmp_deq.val[2]); + vtmp_deq.val[3] = vbslq_f32(vcleq_f32(vin_deq.val[3], vconst_0_f32), clamped_value, vtmp_deq.val[3]); + } + + // Re-quantize to new output space + vout = vquantize(vtmp_deq, qi_out); + wrapper::vstore(output_ptr + x, vout); + } + for(; x < window_end_x; ++x) + { + qasymm8_t in = *(reinterpret_cast<const qasymm8_t *>(input_ptr + x)); + qasymm8_t tmp = 0; + float tmp_f = dequantize_qasymm8(in, qi_in); + if(tmp_f <= 0.0) + { + if(op == ElementWiseUnary::LOG) + { + tmp_f = (0 - qi_out.offset) * qi_out.scale; + } + else if(op == ElementWiseUnary::RSQRT) + { + tmp_f = (255 - qi_out.offset) * qi_out.scale; + } + } + else + { + tmp_f = elementwise_op_scalar_imp<float>(op, tmp_f); + } + tmp = quantize_qasymm8(tmp_f, qi_out, RoundingPolicy::TO_ZERO); + *(output_ptr + x) = tmp; + } + }, + input, output); +} } // namespace cpu } // namespace arm_compute |