aboutsummaryrefslogtreecommitdiff
path: root/src/core/CL/cl_kernels/direct_convolution1x1.cl
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/CL/cl_kernels/direct_convolution1x1.cl')
-rw-r--r--src/core/CL/cl_kernels/direct_convolution1x1.cl118
1 files changed, 1 insertions, 117 deletions
diff --git a/src/core/CL/cl_kernels/direct_convolution1x1.cl b/src/core/CL/cl_kernels/direct_convolution1x1.cl
index d0eea5bfb4..8ab2d1d4ea 100644
--- a/src/core/CL/cl_kernels/direct_convolution1x1.cl
+++ b/src/core/CL/cl_kernels/direct_convolution1x1.cl
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2016-2018 Arm Limited.
+ * Copyright (c) 2016-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -31,122 +31,6 @@
#if defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
-#if defined(DATA_LAYOUT_NHWC)
-
-#define PTR_TO_VALUE(PTR, DATA_TYPE) *((__global DATA_TYPE *)(PTR))
-
-/** This kernel performs a direct convolution to convolve the low three dimensions of a tensor with data layout NHWC
- *
- * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
- * @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32
- * @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1
- * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH
- * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
- *
- * @param[in] src_ptr Pointer to the source tensor. Supported data types: 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: 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 Z processed per workitem(in bytes)
- * @param[in] dst_stride_z Stride of the destination 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] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr
- * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
- * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
- * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes)
- * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
- * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
- * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr
- * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes)
- * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor
- * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension
- */
-__kernel void direct_convolution1x1_nhwc(
- TENSOR3D_DECLARATION(src),
- TENSOR3D_DECLARATION(dst),
- TENSOR3D_DECLARATION(weights),
-#ifdef HAS_BIAS
- VECTOR_DECLARATION(biases),
-#endif /* defined(HAS_BIAS) */
- unsigned int weights_stride_w)
-{
- Image src = CONVERT_TO_IMAGE_STRUCT(src);
- Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
- Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
-
-#ifdef HAS_BIAS
- Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
-#endif /* defined(HAS_BIAS) */
-
- VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)
- values = 0;
- const int id0 = get_global_id(0);
- const int id1 = get_global_id(1);
- const int id2 = get_global_id(2);
- weights.ptr += id0 * weights_stride_w;
- __global uchar *src_addr = (__global uchar *)offset(&src, 0, 0) - src_stride_x * id0 + id2 * STRIDE_Y * (int)src_stride_z;
-
- for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d)
- {
- DATA_TYPE weight = *(__global DATA_TYPE *)weights.ptr;
-#if STRIDE_X == 1
- VEC_DATA_TYPE(DATA_TYPE, 8)
- col0 = (VEC_DATA_TYPE(DATA_TYPE, 8))(
- PTR_TO_VALUE(src_addr + 0 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 1 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 2 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 3 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 4 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 5 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 6 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 7 * src_stride_y, DATA_TYPE));
-#elif STRIDE_X == 2 /* STRIDE_X == 1 */
- VEC_DATA_TYPE(DATA_TYPE, 8)
- col0 = (VEC_DATA_TYPE(DATA_TYPE, 8))(
- PTR_TO_VALUE(src_addr + 0 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 2 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 4 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 6 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 8 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 10 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 12 * src_stride_y, DATA_TYPE),
- PTR_TO_VALUE(src_addr + 14 * src_stride_y, DATA_TYPE));
-#else /* STRIDE_X not equals 1 or 2 */
-#error "STRIDE_X larger than 2 is not supported"
-#endif /* STRIDE_X == 2 */
- values = ADD_OP(values, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))weight, col0));
-
- src_addr += src_stride_x;
- weights.ptr += weights_stride_x;
- }
-
-#ifdef HAS_BIAS
- values = ADD_OP(values, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, id0))));
-#endif /* defined(HAS_BIAS) */
-
- *((__global DATA_TYPE *)dst.ptr) = values.s0;
- *((__global DATA_TYPE *)(dst.ptr + 1 * dst_stride_y)) = values.s1;
- *((__global DATA_TYPE *)(dst.ptr + 2 * dst_stride_y)) = values.s2;
- *((__global DATA_TYPE *)(dst.ptr + 3 * dst_stride_y)) = values.s3;
- *((__global DATA_TYPE *)(dst.ptr + 4 * dst_stride_y)) = values.s4;
- *((__global DATA_TYPE *)(dst.ptr + 5 * dst_stride_y)) = values.s5;
- *((__global DATA_TYPE *)(dst.ptr + 6 * dst_stride_y)) = values.s6;
- *((__global DATA_TYPE *)(dst.ptr + 7 * dst_stride_y)) = values.s7;
-}
-#endif // defined(DATA_LAYOUT_NHWC)
-
#if STRIDE_X == 3
#define INPUT_PIXEL_STR(data_size) extract_input_stride3_##data_size
#define INPUT_PIXEL(data_size) INPUT_PIXEL_STR(data_size)
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-17 22:02:20 +0000