diff options
Diffstat (limited to 'src/cpu/kernels/meanstddevnorm/generic/neon/qasymm8.cpp')
| -rw-r--r-- | src/cpu/kernels/meanstddevnorm/generic/neon/qasymm8.cpp | 124 |
1 files changed, 62 insertions, 62 deletions
diff --git a/src/cpu/kernels/meanstddevnorm/generic/neon/qasymm8.cpp b/src/cpu/kernels/meanstddevnorm/generic/neon/qasymm8.cpp index 53af1e4b16..32654df5dc 100644 --- a/src/cpu/kernels/meanstddevnorm/generic/neon/qasymm8.cpp +++ b/src/cpu/kernels/meanstddevnorm/generic/neon/qasymm8.cpp @@ -23,6 +23,7 @@ */ #include "arm_compute/core/Helpers.h" #include "arm_compute/core/Window.h" + #include "src/core/NEON/NEAsymm.h" #include "src/core/NEON/NEMath.h" #include "src/core/NEON/wrapper/wrapper.h" @@ -69,77 +70,76 @@ void neon_qasymm8_meanstddevnorm(ITensor *input, ITensor *output, float epsilon, const float32x4_t quant_min_vec = vdupq_n_f32(0.0f); execute_window_loop( - win, [&](const Coordinates &) - { - int x = window_start_x; - auto in_ptr = reinterpret_cast<const uint8_t *>(input_itr.ptr()); - auto out_ptr = reinterpret_cast<uint8_t *>(output_itr.ptr()); + win, + [&](const Coordinates &) + { + int x = window_start_x; + auto in_ptr = reinterpret_cast<const uint8_t *>(input_itr.ptr()); + auto out_ptr = reinterpret_cast<uint8_t *>(output_itr.ptr()); - uint32x4_t sum_vec = vdupq_n_u32(0); - uint32x4_t sum_sq_vec = vdupq_n_u32(0); + uint32x4_t sum_vec = vdupq_n_u32(0); + uint32x4_t sum_sq_vec = vdupq_n_u32(0); - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const uint8x16_t data = vld1q_u8(in_ptr + x); - sum_vec = vaddq_u32(sum_vec, vpaddlq_u16(vpaddlq_u8(data))); - const uint16x8_t squares_low = vmull_u8(vget_low_u8(data), vget_low_u8(data)); - const uint16x8_t squares_high = vmull_u8(vget_high_u8(data), vget_high_u8(data)); - sum_sq_vec = vaddq_u32(sum_sq_vec, vaddq_u32(vpaddlq_u16(squares_low), vpaddlq_u16(squares_high))); - } + for (; x <= (window_end_x - window_step_x); x += window_step_x) + { + const uint8x16_t data = vld1q_u8(in_ptr + x); + sum_vec = vaddq_u32(sum_vec, vpaddlq_u16(vpaddlq_u8(data))); + const uint16x8_t squares_low = vmull_u8(vget_low_u8(data), vget_low_u8(data)); + const uint16x8_t squares_high = vmull_u8(vget_high_u8(data), vget_high_u8(data)); + sum_sq_vec = vaddq_u32(sum_sq_vec, vaddq_u32(vpaddlq_u16(squares_low), vpaddlq_u16(squares_high))); + } #ifdef __aarch64__ - sum_vec = vpaddq_u32(sum_vec, sum_vec); - sum_vec = vpaddq_u32(sum_vec, sum_vec); - uint32_t sum = vgetq_lane_u32(sum_vec, 0); - sum_sq_vec = vpaddq_u32(sum_sq_vec, sum_sq_vec); - sum_sq_vec = vpaddq_u32(sum_sq_vec, sum_sq_vec); - uint32_t sum_sq = vgetq_lane_u32(sum_sq_vec, 0); + sum_vec = vpaddq_u32(sum_vec, sum_vec); + sum_vec = vpaddq_u32(sum_vec, sum_vec); + uint32_t sum = vgetq_lane_u32(sum_vec, 0); + sum_sq_vec = vpaddq_u32(sum_sq_vec, sum_sq_vec); + sum_sq_vec = vpaddq_u32(sum_sq_vec, sum_sq_vec); + uint32_t sum_sq = vgetq_lane_u32(sum_sq_vec, 0); #elif __arm__ // #ifdef __aarch64__ - uint32_t sum = vgetq_lane_u32(sum_vec, 0) + - vgetq_lane_u32(sum_vec, 1) + - vgetq_lane_u32(sum_vec, 2) + - vgetq_lane_u32(sum_vec, 3); + uint32_t sum = vgetq_lane_u32(sum_vec, 0) + vgetq_lane_u32(sum_vec, 1) + vgetq_lane_u32(sum_vec, 2) + + vgetq_lane_u32(sum_vec, 3); - uint32_t sum_sq = vgetq_lane_u32(sum_sq_vec, 0) + - vgetq_lane_u32(sum_sq_vec, 1) + - vgetq_lane_u32(sum_sq_vec, 2) + - vgetq_lane_u32(sum_sq_vec, 3); + uint32_t sum_sq = vgetq_lane_u32(sum_sq_vec, 0) + vgetq_lane_u32(sum_sq_vec, 1) + + vgetq_lane_u32(sum_sq_vec, 2) + vgetq_lane_u32(sum_sq_vec, 3); #endif // #ifdef __aarch64__ - for(; x < window_end_x; ++x) - { - auto data = static_cast<uint32_t>(*(in_ptr + x)); - sum += data; - sum_sq += (data * data); - } + for (; x < window_end_x; ++x) + { + auto data = static_cast<uint32_t>(*(in_ptr + x)); + sum += data; + sum_sq += (data * data); + } - const float mean = (static_cast<float>(sum) / static_cast<float>(input->info()->dimension(0))); - const float var = (static_cast<float>(sum_sq) / static_cast<float>(input->info()->dimension(0))) - (mean * mean); - const float stdev_inv = 1.0f / sqrtf(var + epsilon); - const float32x4_t v_scale = vdupq_n_f32(stdev_inv * output_inv_scale); - const float32x4_t v_offset = vdupq_n_f32(-mean * stdev_inv * output_inv_scale + output_offset); - for(x = window_start_x; x <= (window_end_x - window_step_x); x += window_step_x) - { - const uint8x16_t data = vld1q_u8(in_ptr + x); - float32x4_t db1 = vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(data))))); - float32x4_t db2 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(data))))); - float32x4_t db3 = vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(data))))); - float32x4_t db4 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(data))))); - db1 = clamp_v4f32(vaddq_f32(vmulq_f32(db1, v_scale), v_offset), quant_min_vec, quant_max_vec); - db2 = clamp_v4f32(vaddq_f32(vmulq_f32(db2, v_scale), v_offset), quant_min_vec, quant_max_vec); - db3 = clamp_v4f32(vaddq_f32(vmulq_f32(db3, v_scale), v_offset), quant_min_vec, quant_max_vec); - db4 = clamp_v4f32(vaddq_f32(vmulq_f32(db4, v_scale), v_offset), quant_min_vec, quant_max_vec); - const uint8x16_t out = fuse_shorts_u16(fuse_words_f32(db1, db2), fuse_words_f32(db3, db4)); - vst1q_u8(out_ptr + x, out); - } + const float mean = (static_cast<float>(sum) / static_cast<float>(input->info()->dimension(0))); + const float var = + (static_cast<float>(sum_sq) / static_cast<float>(input->info()->dimension(0))) - (mean * mean); + const float stdev_inv = 1.0f / sqrtf(var + epsilon); + const float32x4_t v_scale = vdupq_n_f32(stdev_inv * output_inv_scale); + const float32x4_t v_offset = vdupq_n_f32(-mean * stdev_inv * output_inv_scale + output_offset); + for (x = window_start_x; x <= (window_end_x - window_step_x); x += window_step_x) + { + const uint8x16_t data = vld1q_u8(in_ptr + x); + float32x4_t db1 = vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(data))))); + float32x4_t db2 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(data))))); + float32x4_t db3 = vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(data))))); + float32x4_t db4 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(data))))); + db1 = clamp_v4f32(vaddq_f32(vmulq_f32(db1, v_scale), v_offset), quant_min_vec, quant_max_vec); + db2 = clamp_v4f32(vaddq_f32(vmulq_f32(db2, v_scale), v_offset), quant_min_vec, quant_max_vec); + db3 = clamp_v4f32(vaddq_f32(vmulq_f32(db3, v_scale), v_offset), quant_min_vec, quant_max_vec); + db4 = clamp_v4f32(vaddq_f32(vmulq_f32(db4, v_scale), v_offset), quant_min_vec, quant_max_vec); + const uint8x16_t out = fuse_shorts_u16(fuse_words_f32(db1, db2), fuse_words_f32(db3, db4)); + vst1q_u8(out_ptr + x, out); + } - for(; x < window_end_x; ++x) - { - auto data = static_cast<float32_t>(*(in_ptr + x)); - const uint8_t res = data * (stdev_inv * output_inv_scale) + (-mean * stdev_inv * output_inv_scale + output_offset); - *(out_ptr + x) = res; - } - }, - input_itr, output_itr); + for (; x < window_end_x; ++x) + { + auto data = static_cast<float32_t>(*(in_ptr + x)); + const uint8_t res = + data * (stdev_inv * output_inv_scale) + (-mean * stdev_inv * output_inv_scale + output_offset); + *(out_ptr + x) = res; + } + }, + input_itr, output_itr); } } // namespace cpu } // namespace arm_compute |