diff options
author | Omar Al Khatib <omar.alkhatib@arm.com> | 2022-12-20 14:36:45 +0000 |
---|---|---|
committer | Omar Al Khatib <omar.alkhatib@arm.com> | 2022-12-29 09:14:15 +0000 |
commit | 939b21ad4b9ed15d43b4ee8b17484e57ed55a01f (patch) | |
tree | 6c90f78ad73161ffd0f54956a77fa50ef1916819 /src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp | |
parent | bb1ab0505452691d0e536921cd4c4e2ce32e40ed (diff) | |
download | ComputeLibrary-939b21ad4b9ed15d43b4ee8b17484e57ed55a01f.tar.gz |
Use CPU quantized addition kernel for quantized subtraction
Resolves : [COMPMID-5629]
Signed-off-by: Omar Al Khatib <omar.alkhatib@arm.com>
Change-Id: I061ea5bdafa3a01e66ff869d158f26a38d19e125
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8835
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp')
-rw-r--r-- | src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp | 182 |
1 files changed, 1 insertions, 181 deletions
diff --git a/src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp b/src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp index a285e483ed..dfdf8fe85b 100644 --- a/src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp +++ b/src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp @@ -21,12 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/core/Helpers.h" #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/cpu/kernels/add/generic/neon/impl.h" namespace arm_compute @@ -35,183 +31,7 @@ namespace cpu { void add_qasymm8_signed_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) { - ARM_COMPUTE_UNUSED(policy); - - // Create input windows - Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()); - Window input2_win = window.broadcast_if_dimension_le_one(src1->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)); - - constexpr 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 bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x(); - - const UniformQuantizationInfo iq1_info = src0->info()->quantization_info().uniform(); - const UniformQuantizationInfo iq2_info = src1->info()->quantization_info().uniform(); - const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform(); - - const auto scale1 = iq1_info.scale / oq_info.scale; - const auto scale2 = iq2_info.scale / oq_info.scale; - const auto offset = float(oq_info.offset) - scale1 * float(iq1_info.offset) - scale2 * float(iq2_info.offset); - - 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 ? src1 : src0; - const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0; - - const auto af_scale = is_broadcast_input_2 ? scale1 : scale2; - const auto bf_scale = is_broadcast_input_2 ? scale2 : scale1; - const auto vscale1 = vdupq_n_f32(af_scale); - - // 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(dst, win); - - execute_window_loop(win, [&](const Coordinates &) - { - const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr()); - const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr()); - - const auto broadcast_value = *reinterpret_cast<const int8_t *>(broadcast_input.ptr()); - const auto bf = vdupq_n_f32(float(broadcast_value) * scale2 + offset); - const auto bfs = float(broadcast_value) * bf_scale + offset; - - // Compute S elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const int8x16_t a = vld1q_s8(non_broadcast_input_ptr + x); - - const auto a_s16_0 = vmovl_s8(vget_low_s8(a)); - const auto a_s16_1 = vmovl_s8(vget_high_s8(a)); - - const auto af_0 = vmlaq_f32(bf, vcvtq_f32_s32(vmovl_s16(vget_low_s16(a_s16_0))), vscale1); - const auto af_1 = vmlaq_f32(bf, vcvtq_f32_s32(vmovl_s16(vget_high_s16(a_s16_0))), vscale1); - const auto af_2 = vmlaq_f32(bf, vcvtq_f32_s32(vmovl_s16(vget_low_s16(a_s16_1))), vscale1); - const auto af_3 = vmlaq_f32(bf, vcvtq_f32_s32(vmovl_s16(vget_high_s16(a_s16_1))), vscale1); - - int32x4_t rf_0{}; - int32x4_t rf_1{}; - int32x4_t rf_2{}; - int32x4_t rf_3{}; - -#ifdef __aarch64__ - rf_0 = vcvtnq_s32_f32(af_0); - rf_1 = vcvtnq_s32_f32(af_1); - rf_2 = vcvtnq_s32_f32(af_2); - rf_3 = vcvtnq_s32_f32(af_3); -#else //__aarch64__ - rf_0 = vcvtq_s32_f32(af_0); - rf_1 = vcvtq_s32_f32(af_1); - rf_2 = vcvtq_s32_f32(af_2); - rf_3 = vcvtq_s32_f32(af_3); -#endif //__aarch64__ - - const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1))); - const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3))); - vst1q_s8(output_ptr + x, vcombine_s8(pa, pb)); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - const auto result = float(non_broadcast_input_ptr[x]) * af_scale + bfs; -#ifdef __aarch64__ - output_ptr[x] = utility::clamp<int, int8_t>(support::cpp11::lround(result)); -#else // __aarch64__ - output_ptr[x] = utility::clamp<int, int8_t>(support::cpp11::trunc(result)); -#endif // __aarch64__ - } - }, - 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(src0, input1_win); - Iterator input2(src1, input2_win); - Iterator output(dst, win); - - const auto vscale1 = vdupq_n_f32(scale1); - const auto vscale2 = vdupq_n_f32(scale2); - const auto voffset = vdupq_n_f32(offset); - - execute_window_loop(win, [&](const Coordinates &) - { - const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.ptr()); - const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.ptr()); - const auto output_ptr = reinterpret_cast<int8_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 int8x16_t a = vld1q_s8(input1_ptr + x); - const int8x16_t b = vld1q_s8(input2_ptr + x); - - const auto a_s16_0 = vmovl_s8(vget_low_s8(a)); - const auto a_s16_1 = vmovl_s8(vget_high_s8(a)); - const auto b_s16_0 = vmovl_s8(vget_low_s8(b)); - const auto b_s16_1 = vmovl_s8(vget_high_s8(b)); - - const auto af_0 = vmlaq_f32(voffset, vcvtq_f32_s32(vmovl_s16(vget_low_s16(a_s16_0))), vscale1); - const auto af_1 = vmlaq_f32(voffset, vcvtq_f32_s32(vmovl_s16(vget_high_s16(a_s16_0))), vscale1); - const auto af_2 = vmlaq_f32(voffset, vcvtq_f32_s32(vmovl_s16(vget_low_s16(a_s16_1))), vscale1); - const auto af_3 = vmlaq_f32(voffset, vcvtq_f32_s32(vmovl_s16(vget_high_s16(a_s16_1))), vscale1); - - const auto bf_0 = vmlaq_f32(af_0, vcvtq_f32_s32(vmovl_s16(vget_low_s16(b_s16_0))), vscale2); - const auto bf_1 = vmlaq_f32(af_1, vcvtq_f32_s32(vmovl_s16(vget_high_s16(b_s16_0))), vscale2); - const auto bf_2 = vmlaq_f32(af_2, vcvtq_f32_s32(vmovl_s16(vget_low_s16(b_s16_1))), vscale2); - const auto bf_3 = vmlaq_f32(af_3, vcvtq_f32_s32(vmovl_s16(vget_high_s16(b_s16_1))), vscale2); - - int32x4_t rf_0{}; - int32x4_t rf_1{}; - int32x4_t rf_2{}; - int32x4_t rf_3{}; - -#ifdef __aarch64__ - rf_0 = vcvtnq_s32_f32(bf_0); - rf_1 = vcvtnq_s32_f32(bf_1); - rf_2 = vcvtnq_s32_f32(bf_2); - rf_3 = vcvtnq_s32_f32(bf_3); -#else //__aarch64__ - rf_0 = vcvtq_s32_f32(bf_0); - rf_1 = vcvtq_s32_f32(bf_1); - rf_2 = vcvtq_s32_f32(bf_2); - rf_3 = vcvtq_s32_f32(bf_3); -#endif //__aarch64__ - - const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1))); - const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3))); - vst1q_s8(output_ptr + x, vcombine_s8(pa, pb)); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - const auto result = float(input1_ptr[x]) * scale1 + float(input2_ptr[x]) * scale2 + offset; -#ifdef __aarch64__ - output_ptr[x] = utility::clamp<int, int8_t>(support::cpp11::lround(result)); -#else // __aarch64__ - output_ptr[x] = utility::clamp<int, int8_t>(support::cpp11::trunc(result)); -#endif // __aarch64__ - } - }, - input1, input2, output); - } + add_sub_qasymm8_signed_neon(src0, src1, dst, policy, window, true /*is_addition*/); } } // namespace cpu } // namespace arm_compute
\ No newline at end of file |