diff options
Diffstat (limited to 'src/cpu/kernels/add/generic/neon/qsymm16.cpp')
-rw-r--r-- | src/cpu/kernels/add/generic/neon/qsymm16.cpp | 162 |
1 files changed, 84 insertions, 78 deletions
diff --git a/src/cpu/kernels/add/generic/neon/qsymm16.cpp b/src/cpu/kernels/add/generic/neon/qsymm16.cpp index e76e408d6e..ac2de0557a 100644 --- a/src/cpu/kernels/add/generic/neon/qsymm16.cpp +++ b/src/cpu/kernels/add/generic/neon/qsymm16.cpp @@ -25,14 +25,16 @@ #include "arm_compute/core/ITensor.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/Traits.h" -#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" + #include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" namespace arm_compute { namespace cpu { -void add_qsymm16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) +void add_qsymm16_neon( + const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) { ARM_COMPUTE_UNUSED(policy); @@ -57,7 +59,7 @@ void add_qsymm16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, co 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) + 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; @@ -65,7 +67,7 @@ void add_qsymm16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, co const ITensor *broadcast_tensor = is_broadcast_input_2 ? src1 : src0; const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0; const UniformQuantizationInfo broadcast_qinfo = broadcast_tensor->info()->quantization_info().uniform(); - const UniformQuantizationInfo non_broadcast_qinfo = non_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)); @@ -74,48 +76,50 @@ void add_qsymm16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, co Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); Iterator output(dst, win); - execute_window_loop(win, [&](const Coordinates &) - { - const auto non_broadcast_input_ptr = reinterpret_cast<const int16_t *>(non_broadcast_input.ptr()); - const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr()); - - const int16_t broadcast_value = *reinterpret_cast<const int16_t *>(broadcast_input.ptr()); - const int16x8_t broadcast_value_vec = vdupq_n_s16(broadcast_value); - - const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(broadcast_value_vec))), vscale2); - const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(broadcast_value_vec))), vscale2); - const float bfs = static_cast<int32_t>(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) + execute_window_loop( + win, + [&](const Coordinates &) { - const int16x8_t a = vld1q_s16(non_broadcast_input_ptr + x); - const auto af_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1); - const auto af_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1); - - int32x4_t rf_0{}; - int32x4_t rf_1{}; + const auto non_broadcast_input_ptr = reinterpret_cast<const int16_t *>(non_broadcast_input.ptr()); + const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr()); + + const int16_t broadcast_value = *reinterpret_cast<const int16_t *>(broadcast_input.ptr()); + const int16x8_t broadcast_value_vec = vdupq_n_s16(broadcast_value); + + const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(broadcast_value_vec))), vscale2); + const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(broadcast_value_vec))), vscale2); + const float bfs = static_cast<int32_t>(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 auto af_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1); + const auto af_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1); + + int32x4_t rf_0{}; + int32x4_t rf_1{}; #ifdef __aarch64__ - rf_0 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo)); - rf_1 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo)); + rf_0 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo)); #else //__aarch64__ - rf_0 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo)); - rf_1 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo)); + rf_0 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo)); #endif //__aarch64__ - const int16x8_t pa = vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)); - vst1q_s16(output_ptr + x, pa); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - const float afs = static_cast<int32_t>(*(non_broadcast_input_ptr + x)) * non_broadcast_qinfo.scale; - *(output_ptr + x) = quantize_qsymm16((afs + bfs), oq_info); - } - }, - broadcast_input, non_broadcast_input, output); + const int16x8_t pa = vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)); + vst1q_s16(output_ptr + x, pa); + } + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + const float afs = static_cast<int32_t>(*(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 { @@ -127,48 +131,50 @@ void add_qsymm16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, co Iterator input2(src1, input2_win); Iterator output(dst, win); - execute_window_loop(win, [&](const Coordinates &) - { - const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr()); - const auto input2_ptr = reinterpret_cast<const int16_t *>(input2.ptr()); - const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr()); - - // Compute S elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) + execute_window_loop( + win, + [&](const Coordinates &) { - const int16x8_t a = vld1q_s16(input1_ptr + x); - const int16x8_t b = vld1q_s16(input2_ptr + x); - - const auto af_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1); - const auto af_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1); - const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(b))), vscale2); - const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(b))), vscale2); - - int32x4_t rf_0{}; - int32x4_t rf_1{}; + const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr()); + const auto input2_ptr = reinterpret_cast<const int16_t *>(input2.ptr()); + const auto output_ptr = reinterpret_cast<int16_t *>(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 auto af_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1); + const auto af_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1); + const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(b))), vscale2); + const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(b))), vscale2); + + int32x4_t rf_0{}; + int32x4_t rf_1{}; #ifdef __aarch64__ - rf_0 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo)); - rf_1 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo)); + rf_0 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo)); #else //__aarch64__ - rf_0 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo)); - rf_1 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo)); + rf_0 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo)); + rf_1 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo)); #endif //__aarch64__ - const int16x8_t pa = vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)); - vst1q_s16(output_ptr + x, pa); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - const float afs = static_cast<int32_t>((*(input1_ptr + x))) * iq1_info.scale; - const float bfs = static_cast<int32_t>((*(input2_ptr + x))) * iq2_info.scale; - *(output_ptr + x) = quantize_qsymm16((afs + bfs), dst->info()->quantization_info()); - } - }, - input1, input2, output); + const int16x8_t pa = vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)); + vst1q_s16(output_ptr + x, pa); + } + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + const float afs = static_cast<int32_t>((*(input1_ptr + x))) * iq1_info.scale; + const float bfs = static_cast<int32_t>((*(input2_ptr + x))) * iq2_info.scale; + *(output_ptr + x) = quantize_qsymm16((afs + bfs), dst->info()->quantization_info()); + } + }, + input1, input2, output); } } } // namespace cpu -} // namespace arm_compute
\ No newline at end of file +} // namespace arm_compute |