diff options
Diffstat (limited to 'src/cpu/kernels/scale/neon/fp16.cpp')
| -rw-r--r-- | src/cpu/kernels/scale/neon/fp16.cpp | 189 |
1 files changed, 118 insertions, 71 deletions
diff --git a/src/cpu/kernels/scale/neon/fp16.cpp b/src/cpu/kernels/scale/neon/fp16.cpp index 895f42215e..bd01569cc4 100644 --- a/src/cpu/kernels/scale/neon/fp16.cpp +++ b/src/cpu/kernels/scale/neon/fp16.cpp @@ -27,9 +27,10 @@ #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensorPack.h" #include "arm_compute/core/Window.h" + +#include "src/core/helpers/ScaleHelpers.h" #include "src/core/NEON/NEMath.h" #include "src/core/NEON/wrapper/wrapper.h" -#include "src/core/helpers/ScaleHelpers.h" #include "src/core/utils/ScaleUtils.h" #include "support/Rounding.h" @@ -41,8 +42,12 @@ namespace arm_compute { namespace { -void fp16_neon_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets, - float sampling_offset, bool align_corners, const Window &window) +void fp16_neon_scale_nearest(const ITensor *src, + ITensor *dst, + const ITensor *offsets, + float sampling_offset, + bool align_corners, + const Window &window) { const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right; const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom; @@ -62,33 +67,46 @@ void fp16_neon_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *of const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes(); const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3]; - execute_window_loop(win, [&](const Coordinates & id) - { - const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c; - const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr)); - const int offset_row = in_hi * in_stride_wc; - int32_t x = window_start_x; - const float16_t *in_ptr = reinterpret_cast<const float16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]); - - for(; x <= window_end_x - window_step_x; x += window_step_x) - { - wrapper::vstore(reinterpret_cast<float16_t *>(out.ptr()) + x, - wrapper::vloadq(in_ptr + offset + offset_row + x)); - } - for(; x < window_end_x; ++x) + execute_window_loop( + win, + [&](const Coordinates &id) { - *(reinterpret_cast<float16_t *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x); - } - }, - out); + const int32_t offset = + *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c; + const auto in_hi = static_cast<int>( + align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) + : std::floor((id.z() + sampling_offset) * hr)); + const int offset_row = in_hi * in_stride_wc; + int32_t x = window_start_x; + const float16_t *in_ptr = reinterpret_cast<const float16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]); + + for (; x <= window_end_x - window_step_x; x += window_step_x) + { + wrapper::vstore(reinterpret_cast<float16_t *>(out.ptr()) + x, + wrapper::vloadq(in_ptr + offset + offset_row + x)); + } + for (; x < window_end_x; ++x) + { + *(reinterpret_cast<float16_t *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x); + } + }, + out); } -void fp16_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, - BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, - bool align_corners, const Window &window) +void fp16_neon_scale_bilinear(const ITensor *src, + ITensor *dst, + const ITensor *offsets, + const ITensor *dx, + const ITensor *dy, + BorderMode border_mode, + PixelValue constant_border_value, + float sampling_offset, + bool align_corners, + const Window &window) { // Compute the ratio between source height and destination height - const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners); + const auto hr = + scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners); Iterator out(dst, window); const int in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right; @@ -103,68 +121,97 @@ void fp16_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *o win_in.set(Window::DimZ, Window::Dimension(0, 0, 0)); Iterator in(src, win_in); - if(border_mode == BorderMode::CONSTANT) + if (border_mode == BorderMode::CONSTANT) { using ConstType = typename std::conditional<std::is_same<float16_t, float16_t>::value, half, float16_t>::type; const float16_t const_border_value = static_cast<float16_t>(constant_border_value.get<ConstType>()); - execute_window_loop(window, [&](const Coordinates & id) - { - const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))); - const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z()))); - const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z()))); - const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset); - const float16_t *in_ptr = reinterpret_cast<const float16_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc; - - const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value; - const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value; - const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value; - const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value; - - *reinterpret_cast<float16_t *>(out.ptr()) = static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val)); - }, - in, out); + execute_window_loop( + window, + [&](const Coordinates &id) + { + const auto offset = + *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))); + const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z()))); + const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z()))); + const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset); + const float16_t *in_ptr = + reinterpret_cast<const float16_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc; + + const auto a00 = + (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value; + const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) + ? *(in_ptr + in_stride_c) + : const_border_value; + const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) + ? *(in_ptr + in_stride_wc) + : const_border_value; + const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) + ? *(in_ptr + in_stride_c + in_stride_wc) + : const_border_value; + + *reinterpret_cast<float16_t *>(out.ptr()) = + static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val)); + }, + in, out); } - else if(border_mode == BorderMode::REPLICATE) + else if (border_mode == BorderMode::REPLICATE) { - execute_window_loop(window, [&](const Coordinates & id) - { - const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))); - const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z()))); - const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z()))); - const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset); - - auto clamped_w = utility::clamp<int>(offset, 0, in_dim_w - 1); - auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1); - auto clamped_h = utility::clamp<int>(in_hi, 0, in_dim_h - 1); - auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1); - - const auto a00 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc); - const auto a01 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc); - const auto a10 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc); - const auto a11 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc); - - *reinterpret_cast<float16_t *>(out.ptr()) = static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val)); - }, - in, out); + execute_window_loop( + window, + [&](const Coordinates &id) + { + const auto offset = + *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))); + const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z()))); + const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z()))); + const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset); + + auto clamped_w = utility::clamp<int>(offset, 0, in_dim_w - 1); + auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1); + auto clamped_h = utility::clamp<int>(in_hi, 0, in_dim_h - 1); + auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1); + + const auto a00 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + + clamped_h * in_stride_wc); + const auto a01 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + + clamped_h * in_stride_wc); + const auto a10 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + + clamped_h1 * in_stride_wc); + const auto a11 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + + clamped_h1 * in_stride_wc); + + *reinterpret_cast<float16_t *>(out.ptr()) = + static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val)); + }, + in, out); } else { ARM_COMPUTE_ERROR("Not implemented"); } } -} +} // namespace namespace cpu { -void fp16_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, - InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, - bool align_corners, const Window &window) +void fp16_neon_scale(const ITensor *src, + ITensor *dst, + const ITensor *offsets, + const ITensor *dx, + const ITensor *dy, + InterpolationPolicy policy, + BorderMode border_mode, + PixelValue constant_border_value, + float sampling_offset, + bool align_corners, + const Window &window) { - if(policy == InterpolationPolicy::BILINEAR) + if (policy == InterpolationPolicy::BILINEAR) { - fp16_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window); + fp16_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, + align_corners, window); } - else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR) + else if (policy == InterpolationPolicy::NEAREST_NEIGHBOR) { fp16_neon_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window); } @@ -172,4 +219,4 @@ void fp16_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, c } // namespace cpu } // namespace arm_compute -#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
\ No newline at end of file +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ |