From 13a51e11680aa24a9b841a4afe4079419bc8b80c Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Tue, 18 Sep 2018 13:09:30 +0100 Subject: COMPMID-1554 Implementing Space to Batch on OpenCL - NHWC Change-Id: Ifa37a6758f79d0a6ca771dcfb4c55a5d96b452d0 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/148892 Reviewed-by: Georgios Pinitas Tested-by: bsgcomp --- src/core/CL/cl_kernels/space_to_batch.cl | 125 ++++++++++++++++++++++++++++++- 1 file changed, 123 insertions(+), 2 deletions(-) (limited to 'src/core/CL/cl_kernels/space_to_batch.cl') diff --git a/src/core/CL/cl_kernels/space_to_batch.cl b/src/core/CL/cl_kernels/space_to_batch.cl index 1343695ed1..1d641cc0f4 100644 --- a/src/core/CL/cl_kernels/space_to_batch.cl +++ b/src/core/CL/cl_kernels/space_to_batch.cl @@ -59,7 +59,7 @@ * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image */ -__kernel void space_to_batch( +__kernel void space_to_batch_nchw( TENSOR4D_DECLARATION(input), IMAGE_DECLARATION(paddings), VECTOR_DECLARATION(block_shape), @@ -92,6 +92,74 @@ __kernel void space_to_batch( *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, in_x, in_y, z, w)); } } +/** Calculate the space to batch conversion. (NHWC) + * + * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float + * @note The block shape tensor rank must be passed at compile time using -DBLOCK_SHAPE_DIM. e.g. -DBLOCK_SHAPE_DIM=2 + * + * @param[in] input_ptr Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 + * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source image + * @param[in] paddings_ptr Pointer to the second source image. Supported data types: S32 + * @param[in] paddings_stride_x Stride of the paddinds tensor in X dimension (in bytes) + * @param[in] paddings_step_x paddings_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] paddings_stride_y Stride of the paddinds tensor in Y dimension (in bytes) + * @param[in] paddings_step_y paddings_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] paddingse_offset_first_element_in_bytes The offset of the first element in the second source image + * @param[in] block_shape_ptr Pointer to the block shape tensor. Supported data types: S32 + * @param[in] block_shape_stride_x Stride of the block shape tensor in X dimension (in bytes) + * @param[in] block_shape_step_x block_shape_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] block_shape_stride_y Stride of the block shape tensor in Y dimension (in bytes) + * @param[in] block_shape_step_y block_shape_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] block_shape_offset_first_element_in_bytes The offset of the first element in the block shapetensor + * @param[in] batch_id The output tensor batch id + * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr + * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image + */ +__kernel void space_to_batch_nhwc( + TENSOR4D_DECLARATION(input), + IMAGE_DECLARATION(paddings), + VECTOR_DECLARATION(block_shape), + const int batch_id, + TENSOR3D_DECLARATION(output)) +{ + Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0); + Image pad = CONVERT_TO_IMAGE_STRUCT_NO_STEP(paddings); + Vector block = CONVERT_TO_VECTOR_STRUCT_NO_STEP(block_shape); + Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output); + + const int PAD_LEFT_X = *((__global int *)offset(&pad, 0, 0)); + const int PAD_RIGHT_X = *((__global int *)offset(&pad, 1, 0)); + const int PAD_LEFT_Y = *((__global int *)offset(&pad, 0, 1)); + const int PAD_RIGHT_Y = *((__global int *)offset(&pad, 1, 1)); + + int block_x = *((__global int *)vector_offset(&block, 0)); + int block_y = *((__global int *)vector_offset(&block, 1)); + + const int out_x = get_global_id(1); + const int out_y = get_global_id(2); + const int z = get_global_id(0); + + if((out_x >= PAD_LEFT_X && out_x < WIDTH_OUT - PAD_RIGHT_X) && (out_y >= PAD_LEFT_Y && out_y < HEIGHT_OUT - PAD_RIGHT_Y)) + { + const int r = (BATCH_SIZE / (block_x * block_y)); + const int w = batch_id % r; + const int in_x = (out_x - PAD_LEFT_X) * block_x + (batch_id / r) % block_x; + const int in_y = (out_y - PAD_LEFT_Y) * block_y + (batch_id / r) / block_x; + *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, z, in_x, in_y, w)); + } +} #endif // defined(BATCH_SIZE) && defined(DATA_TYPE) #if defined(BATCH_SIZE) && defined(DATA_TYPE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y) && defined(PAD_LEFT_X) && defined(PAD_RIGHT_X) && defined(PAD_LEFT_Y) && defined(PAD_RIGHT_Y) @@ -124,7 +192,7 @@ __kernel void space_to_batch( * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image */ -__kernel void space_to_batch_static( +__kernel void space_to_batch_static_nchw( TENSOR4D_DECLARATION(input), const int batch_id, TENSOR3D_DECLARATION(output)) @@ -148,4 +216,57 @@ __kernel void space_to_batch_static( *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, in_x, in_y, z, w)); } } +/** Calculate the space to batch conversion. (NHWC) + * + * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float + * @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2 + * @note The block shape x must be passed at compile time using -DBLOCK_SHAPE_X. e.g. -DBLOCK_SHAPE_X=2 + * @note The block shape y must be passed at compile time using -DBLOCK_SHAPE_Y. e.g. -DBLOCK_SHAPE_Y=2 + * @note The starting pad value of x must be passed at compile time using -DPAD_LEFT_X. e.g. -DPAD_LEFT_X=2 + * @note The ending pad value of x must be passed at compile time using -DPAD_RIGHT_X. e.g. -DPAD_RIGHT_X=2 + * @note The starting pad value of y must be passed at compile time using -DPAD_LEFT_Y. e.g. -DPAD_LEFT_Y=2 + * @note The ending pad value of y must be passed at compile time using -DPAD_RIGHT_Y. e.g. -DPAD_RIGHT_X=2 + * + * @param[in] input_ptr Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 + * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source image + * @param[in] batch_id The output tensor batch id + * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr + * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image + */ +__kernel void space_to_batch_static_nhwc( + TENSOR4D_DECLARATION(input), + const int batch_id, + TENSOR3D_DECLARATION(output)) +{ + Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0); + Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output); + + int block_x = BLOCK_SHAPE_X; + int block_y = *((__global int *)vector_offset(&block, 1)); + + const int out_x = get_global_id(1); + const int out_y = get_global_id(2); + const int z = get_global_id(0); + + if((out_x >= PAD_LEFT_X && out_x < WIDTH_OUT - PAD_RIGHT_X) && (out_y >= PAD_LEFT_Y && out_y < HEIGHT_OUT - PAD_RIGHT_Y)) + { + const int r = (BATCH_SIZE / (block_x * block_y)); + const int w = batch_id % r; + const int in_x = (out_x - PAD_LEFT_X) * block_x + (batch_id / r) % block_x; + const int in_y = (out_y - PAD_LEFT_Y) * block_y + (batch_id / r) / block_x; + *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, z, in_x, in_y, w)); + } +} #endif // defined(BATCH_SIZE) && defined(DATA_TYPE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y) && defined(PAD_LEFT_X) && defined(PAD_RIGHT_X) && defined(PAD_LEFT_Y) && defined(PAD_RIGHT_Y) -- cgit v1.2.1