From ac4e873dad6aa6291fc36aff62047a896db04f6a Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Wed, 5 Jul 2017 17:02:25 +0100 Subject: COMPMID-417: Port DepthConcatenate to QS8/QS16 for NEON/CL. Change-Id: I3dddae63043c7aa18d908a4fc8abacf3c64f98ca Reviewed-on: http://mpd-gerrit.cambridge.arm.com/80081 Tested-by: Kaizen Reviewed-by: Steven Niu --- src/core/CL/cl_kernels/concatenate.cl | 26 ++++++++++++++++---------- src/core/CL/cl_kernels/helpers.h | 6 ++++++ 2 files changed, 22 insertions(+), 10 deletions(-) (limited to 'src/core/CL/cl_kernels') diff --git a/src/core/CL/cl_kernels/concatenate.cl b/src/core/CL/cl_kernels/concatenate.cl index 00f5189508..a92ab5bdad 100644 --- a/src/core/CL/cl_kernels/concatenate.cl +++ b/src/core/CL/cl_kernels/concatenate.cl @@ -25,29 +25,35 @@ /** This kernel concatenates the input tensor into the output tensor along the third dimension * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8, QS16, F16, F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: F32 + * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] offset The offset to the first valid element of the output tensor in bytes + * @param[in] offsets The offsets to the first valid element of the output tensor in bytes */ __kernel void concatenate_depth( - IMAGE_DECLARATION(src), - IMAGE_DECLARATION(dst), - unsigned int offset) + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst), + int3 offsets) { - Image src = CONVERT_TO_IMAGE_STRUCT(src); - Image dst = CONVERT_TO_IMAGE_STRUCT(dst); + Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); + Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - float4 source_values = vload4(0, (__global float *)src.ptr); + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, -offsets.x, -offsets.y, 0)); - vstore4(source_values, 0, (__global float *)(dst.ptr + offset)); + VSTORE(VEC_SIZE) + (source_values, 0, (__global DATA_TYPE *)(dst.ptr + offsets.z)); } diff --git a/src/core/CL/cl_kernels/helpers.h b/src/core/CL/cl_kernels/helpers.h index 29a43f769b..0b6d92dfd0 100644 --- a/src/core/CL/cl_kernels/helpers.h +++ b/src/core/CL/cl_kernels/helpers.h @@ -30,6 +30,12 @@ #define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val) +#define VLOAD_STR(size) vload##size +#define VLOAD(size) VLOAD_STR(size) + +#define VSTORE_STR(size) vstore##size +#define VSTORE(size) VSTORE_STR(size) + #define VEC_DATA_TYPE_STR(type, size) type##size #define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size) -- cgit v1.2.1