diff options
Diffstat (limited to 'src/core/cpu/kernels/pooling/neon')
| -rw-r--r-- | src/core/cpu/kernels/pooling/neon/fp16.cpp | 4 | ||||
| -rw-r--r-- | src/core/cpu/kernels/pooling/neon/fp32.cpp | 4 | ||||
| -rw-r--r-- | src/core/cpu/kernels/pooling/neon/list.h | 4 | ||||
| -rw-r--r-- | src/core/cpu/kernels/pooling/neon/nchw/all.cpp | 2 |
4 files changed, 7 insertions, 7 deletions
diff --git a/src/core/cpu/kernels/pooling/neon/fp16.cpp b/src/core/cpu/kernels/pooling/neon/fp16.cpp index 314be3704e..1ecceafe86 100644 --- a/src/core/cpu/kernels/pooling/neon/fp16.cpp +++ b/src/core/cpu/kernels/pooling/neon/fp16.cpp @@ -93,7 +93,7 @@ void pooling2_f16_maxpool_indices(const ITensor *src, ITensor *dst0, ITensor *ds // Store result vst1q_f16(reinterpret_cast<float16_t *>(out.ptr()) + x_off, vres); - const uint32_t offset_base = offset_no_padding<float16_t>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y); + const uint32_t offset_base = offset_no_padding<float16_t>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y, DataLayout::NHWC); const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float16_t) + x_off; const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float16_t) - pad_right; const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float16_t) - pad_right * src->info()->tensor_shape()[1]; @@ -132,7 +132,7 @@ void pooling2_f16_maxpool_indices(const ITensor *src, ITensor *dst0, ITensor *ds // Store result *(reinterpret_cast<float16_t *>(out.ptr()) + x_off) = res; - const uint32_t offset_base = offset_no_padding<float16_t>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y); + const uint32_t offset_base = offset_no_padding<float16_t>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y, DataLayout::NHWC); const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float16_t) + x_off; const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float16_t) - pad_right; const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float16_t) - pad_right * src->info()->tensor_shape()[1]; diff --git a/src/core/cpu/kernels/pooling/neon/fp32.cpp b/src/core/cpu/kernels/pooling/neon/fp32.cpp index e319047d76..a2bd4a6bb3 100644 --- a/src/core/cpu/kernels/pooling/neon/fp32.cpp +++ b/src/core/cpu/kernels/pooling/neon/fp32.cpp @@ -95,7 +95,7 @@ void pooling2_f32_maxpool_indices(const ITensor *src, ITensor *dst0, ITensor *ds // Store result vst1q_f32(reinterpret_cast<float *>(out.ptr()) + x_off, vres); - const uint32_t offset_base = offset_no_padding<float>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y); + const uint32_t offset_base = offset_no_padding<float>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y, DataLayout::NHWC); const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float) + x_off; const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float) - pad_right; const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float) - pad_right * src->info()->tensor_shape()[1]; @@ -124,7 +124,7 @@ void pooling2_f32_maxpool_indices(const ITensor *src, ITensor *dst0, ITensor *ds // Store result *(reinterpret_cast<float *>(out.ptr()) + x_off) = res; - const uint32_t offset_base = offset_no_padding<float>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y); + const uint32_t offset_base = offset_no_padding<float>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y, DataLayout::NHWC); const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float) + x_off; const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float) - pad_right; const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float) - pad_right * src->info()->tensor_shape()[1]; diff --git a/src/core/cpu/kernels/pooling/neon/list.h b/src/core/cpu/kernels/pooling/neon/list.h index 3435ee6724..bec1536f61 100644 --- a/src/core/cpu/kernels/pooling/neon/list.h +++ b/src/core/cpu/kernels/pooling/neon/list.h @@ -59,7 +59,7 @@ DECLARE_POOLING_KERNEL(poolingMxN_fp32_neon_nchw); #undef DECLARE_POOLING_KERNEL template <typename T> -inline uint32_t offset_no_padding(uint32_t padded_offset, const Coordinates &id, const ITensorInfo &info, int pool_stride_x, int pool_stride_y) +inline uint32_t offset_no_padding(uint32_t padded_offset, const Coordinates &id, const ITensorInfo &info, int pool_stride_x, int pool_stride_y, DataLayout data_layout) { const int pad_left = info.padding().left; const int pad_right = info.padding().right; @@ -70,7 +70,7 @@ inline uint32_t offset_no_padding(uint32_t padded_offset, const Coordinates &id, const int pad_horiz = pad_left + pad_right; const int pad_vert = pad_top + pad_bottom; - if(info.data_layout() == DataLayout::NCHW) + if(data_layout == DataLayout::NCHW) { const uint32_t offset_base = padded_offset - sizeof(T) * pad_horiz * id.y() * pool_stride_y /* subtract padding elems per row */ diff --git a/src/core/cpu/kernels/pooling/neon/nchw/all.cpp b/src/core/cpu/kernels/pooling/neon/nchw/all.cpp index 47ac7b4f7f..80eac684aa 100644 --- a/src/core/cpu/kernels/pooling/neon/nchw/all.cpp +++ b/src/core/cpu/kernels/pooling/neon/nchw/all.cpp @@ -150,7 +150,7 @@ void pooling2_nchw_maxpool_indices(const ITensor *src, ITensor *dst0, ITensor *d *(reinterpret_cast<T *>(out.ptr())) = static_cast<T>(vget_lane_f32(max_data, 0)); // Calculate max data indice, which will be used in max unpool. - const uint32_t offset_base = offset_no_padding<T>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y); + const uint32_t offset_base = offset_no_padding<T>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y, DataLayout::NCHW); const uint32_t offset_top = (uint32_t)(offset_base / sizeof(T)); const uint32_t offset_bottom = offset_top + in_stride_y / sizeof(T) - pad_right - pad_left; const uint32x2_t voffset_top = { offset_top, offset_top + 1u }; |