diff options
Diffstat (limited to 'src/cpu/kernels/sub/neon')
| -rw-r--r-- | src/cpu/kernels/sub/neon/list.h | 119 | ||||
| -rw-r--r-- | src/cpu/kernels/sub/neon/qasymm8.cpp | 9 | ||||
| -rw-r--r-- | src/cpu/kernels/sub/neon/qasymm8_signed.cpp | 9 | ||||
| -rw-r--r-- | src/cpu/kernels/sub/neon/qsymm16.cpp | 166 |
4 files changed, 156 insertions, 147 deletions
diff --git a/src/cpu/kernels/sub/neon/list.h b/src/cpu/kernels/sub/neon/list.h index f7e1a040bd..9f6c92271f 100644 --- a/src/cpu/kernels/sub/neon/list.h +++ b/src/cpu/kernels/sub/neon/list.h @@ -26,14 +26,16 @@ #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/Traits.h" + #include "src/core/NEON/wrapper/wrapper.h" namespace arm_compute { namespace cpu { -#define DECLARE_SUB_KERNEL(func_name) \ - void func_name(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) +#define DECLARE_SUB_KERNEL(func_name) \ + void func_name(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, \ + const Window &window) DECLARE_SUB_KERNEL(sub_qasymm8_neon_fixedpoint); DECLARE_SUB_KERNEL(sub_qasymm8_signed_neon_fixedpoint); @@ -44,7 +46,8 @@ DECLARE_SUB_KERNEL(sub_qsymm16_neon); #undef DECLARE_SUB_KERNEL template <typename T> -void sub_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) +void sub_same_neon( + const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) { /** SIMD vector tag type. */ using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>; @@ -68,7 +71,7 @@ void sub_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const Iterator input2(src1, window.broadcast_if_dimension_le_one(src1->info()->tensor_shape())); Iterator output(dst, window); - 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; @@ -84,41 +87,44 @@ void sub_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const Iterator output(dst, win); execute_window_loop( - win, [&](const Coordinates &) - { - const auto non_broadcast_input_ptr = reinterpret_cast<const T *>(non_broadcast_input.ptr()); - const auto output_ptr = reinterpret_cast<T *>(output.ptr()); + win, + [&](const Coordinates &) + { + const auto non_broadcast_input_ptr = reinterpret_cast<const T *>(non_broadcast_input.ptr()); + const auto output_ptr = reinterpret_cast<T *>(output.ptr()); - const T broadcast_value = *reinterpret_cast<const T *>(broadcast_input.ptr()); - const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{}); + const T broadcast_value = *reinterpret_cast<const T *>(broadcast_input.ptr()); + const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{}); - // Compute S elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x); - auto res = is_sat ? wrapper::vqsub(broadcast_value_vec, non_broadcast_v) : wrapper::vsub(broadcast_value_vec, non_broadcast_v); - if(is_broadcast_input_2) + // Compute S elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) { - res = wrapper::vmul(res, wrapper::vdup_n(static_cast<T>(-1), ExactTagType{})); + const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x); + auto res = is_sat ? wrapper::vqsub(broadcast_value_vec, non_broadcast_v) + : wrapper::vsub(broadcast_value_vec, non_broadcast_v); + if (is_broadcast_input_2) + { + res = wrapper::vmul(res, wrapper::vdup_n(static_cast<T>(-1), ExactTagType{})); + } + wrapper::vstore(output_ptr + x, res); } - wrapper::vstore(output_ptr + x, res); - } - // Compute left-over elements - for(; x < window_end_x; ++x) - { - const auto non_broadcast_v = *(non_broadcast_input_ptr + x); - auto res = is_sat ? wrapper::sub_sat(broadcast_value, non_broadcast_v) : broadcast_value - non_broadcast_v; - if(is_broadcast_input_2) + // Compute left-over elements + for (; x < window_end_x; ++x) { - res = static_cast<T>(-1) * res; + const auto non_broadcast_v = *(non_broadcast_input_ptr + x); + auto res = + is_sat ? wrapper::sub_sat(broadcast_value, non_broadcast_v) : broadcast_value - non_broadcast_v; + if (is_broadcast_input_2) + { + res = static_cast<T>(-1) * res; + } + + *(output_ptr + x) = res; } - - *(output_ptr + x) = res; - } - }, - broadcast_input, non_broadcast_input, output); + }, + broadcast_input, non_broadcast_input, output); } else { @@ -131,31 +137,32 @@ void sub_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const Iterator output(dst, win); execute_window_loop( - win, [&](const Coordinates &) - { - const auto input1_ptr = reinterpret_cast<const T *>(input1.ptr()); - const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr()); - const auto output_ptr = reinterpret_cast<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 auto val1 = wrapper::vloadq(input1_ptr + x); - const auto val2 = wrapper::vloadq(input2_ptr + x); - const auto res = is_sat ? wrapper::vqsub(val1, val2) : wrapper::vsub(val1, val2); - wrapper::vstore(output_ptr + x, res); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) + win, + [&](const Coordinates &) { - const auto val1 = *(input1_ptr + x); - const auto val2 = *(input2_ptr + x); - *(output_ptr + x) = is_sat ? wrapper::sub_sat(val1, val2) : val1 - val2; - } - }, - input1, input2, output); + const auto input1_ptr = reinterpret_cast<const T *>(input1.ptr()); + const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr()); + const auto output_ptr = reinterpret_cast<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 auto val1 = wrapper::vloadq(input1_ptr + x); + const auto val2 = wrapper::vloadq(input2_ptr + x); + const auto res = is_sat ? wrapper::vqsub(val1, val2) : wrapper::vsub(val1, val2); + wrapper::vstore(output_ptr + x, res); + } + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + const auto val1 = *(input1_ptr + x); + const auto val2 = *(input2_ptr + x); + *(output_ptr + x) = is_sat ? wrapper::sub_sat(val1, val2) : val1 - val2; + } + }, + input1, input2, output); } } } // namespace cpu diff --git a/src/cpu/kernels/sub/neon/qasymm8.cpp b/src/cpu/kernels/sub/neon/qasymm8.cpp index ea6e5826dd..b750afce6e 100644 --- a/src/cpu/kernels/sub/neon/qasymm8.cpp +++ b/src/cpu/kernels/sub/neon/qasymm8.cpp @@ -23,21 +23,24 @@ */ #include "arm_compute/core/ITensor.h" #include "arm_compute/core/Types.h" + #include "src/cpu/kernels/add/generic/neon/impl.h" namespace arm_compute { namespace cpu { -void sub_qasymm8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) +void sub_qasymm8_neon_fixedpoint( + const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) { add_sub_q8_neon_fixedpoint<uint8_t>(src0, src1, dst, policy, window, false /*is_addition*/); } -void sub_qasymm8_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) +void sub_qasymm8_neon( + const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) { add_sub_qasymm8_neon(src0, src1, dst, policy, window, false /*is_addition*/); } } // namespace cpu -} // namespace arm_compute
\ No newline at end of file +} // namespace arm_compute diff --git a/src/cpu/kernels/sub/neon/qasymm8_signed.cpp b/src/cpu/kernels/sub/neon/qasymm8_signed.cpp index a86c7f22f6..fb0bb62682 100644 --- a/src/cpu/kernels/sub/neon/qasymm8_signed.cpp +++ b/src/cpu/kernels/sub/neon/qasymm8_signed.cpp @@ -24,21 +24,24 @@ #include "arm_compute/core/ITensor.h" #include "arm_compute/core/Types.h" + #include "src/cpu/kernels/add/generic/neon/impl.h" namespace arm_compute { namespace cpu { -void sub_qasymm8_signed_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) +void sub_qasymm8_signed_neon_fixedpoint( + const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) { add_sub_q8_neon_fixedpoint<int8_t>(src0, src1, dst, policy, window, false /*is_addition*/); } -void sub_qasymm8_signed_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) +void sub_qasymm8_signed_neon( + const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) { add_sub_qasymm8_signed_neon(src0, src1, dst, policy, window, false /*is_addition*/); } } // namespace cpu -} // namespace arm_compute
\ No newline at end of file +} // namespace arm_compute diff --git a/src/cpu/kernels/sub/neon/qsymm16.cpp b/src/cpu/kernels/sub/neon/qsymm16.cpp index 4dfdc0e78c..23e4b03843 100644 --- a/src/cpu/kernels/sub/neon/qsymm16.cpp +++ b/src/cpu/kernels/sub/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 sub_qsymm16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) +void sub_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 sub_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 sub_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)); @@ -74,61 +76,62 @@ void sub_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()); + 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 int16_t broadcast_value = *reinterpret_cast<const int16_t *>(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<int32_t>(broadcast_value) * broadcast_qinfo.scale; + 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<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 float32x4x2_t af = + // 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 = - { - { + const int32x4x4_t rf = {{ #ifdef __aarch64__ - vcvtnq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[0], bf.val[0]), invvscaleo)), - vcvtnq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[1], bf.val[1]), invvscaleo)), + vcvtnq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) + : vsubq_f32(af.val[0], bf.val[0]), + invvscaleo)), + vcvtnq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) + : vsubq_f32(af.val[1], bf.val[1]), + invvscaleo)), #else //__aarch64__ - vcvtq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[0], bf.val[0]), invvscaleo)), - vcvtq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[1], bf.val[1]), invvscaleo)), + vcvtq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) + : vsubq_f32(af.val[0], bf.val[0]), + invvscaleo)), + vcvtq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) + : vsubq_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); - } + 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<int32_t>(*(non_broadcast_input_ptr + x)) * non_broadcast_qinfo.scale; - *(output_ptr + x) = quantize_qsymm16(is_broadcast_input_2 ? (bfs - afs) : (afs - bfs), oq_info); - } - }, - broadcast_input, non_broadcast_input, output); + // 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(is_broadcast_input_2 ? (bfs - afs) : (afs - bfs), oq_info); + } + }, + broadcast_input, non_broadcast_input, output); } else { @@ -140,38 +143,32 @@ void sub_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 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()); - const float32x4x2_t af = + // 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 = - { - { + 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 = - { - { + const int32x4x2_t rf = {{ #ifdef __aarch64__ vcvtnq_s32_f32(vmulq_f32(vsubq_f32(af.val[0], bf.val[0]), invvscaleo)), vcvtnq_s32_f32(vmulq_f32(vsubq_f32(af.val[1], bf.val[1]), invvscaleo)), @@ -179,23 +176,22 @@ void sub_qsymm16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, co vcvtq_s32_f32(vmulq_f32(vsubq_f32(af.val[0], bf.val[0]), invvscaleo)), vcvtq_s32_f32(vmulq_f32(vsubq_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); - } + 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<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); + // 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 |