COMPMID-452 CL Generic Depthwise Convolution implementation.
Change-Id: I115e48fe6ce5e281f3791aa5d80fdc754cdd2b5e
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/85082
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
diff --git a/src/core/CL/cl_kernels/convolution_layer.cl b/src/core/CL/cl_kernels/convolution_layer.cl
index 162632b..9e9d0b0 100644
--- a/src/core/CL/cl_kernels/convolution_layer.cl
+++ b/src/core/CL/cl_kernels/convolution_layer.cl
@@ -117,6 +117,9 @@
* @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_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[in] filter_depth The depth of the used filter
+ * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes).
+ * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
*/
__kernel void im2col_generic(
TENSOR3D_DECLARATION(src),
@@ -192,6 +195,9 @@
* @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_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[in] filter_depth The depth of the used filter
+ * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes).
+ * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
*/
__kernel void im2col_kernel3x3_padx0_pady0(
TENSOR3D_DECLARATION(src),
@@ -279,6 +285,7 @@
*((__global DATA_TYPE *)(dst.ptr + idx)) = *((__global DATA_TYPE *)(src.ptr));
}
#endif // defined(WIDTH_OUTPUT)
+
/** This kernel reshapes the tensor's low three dimensions to single row for GEMM operation
*
* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl
index cbcdbf2..9c2c3a5 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution.cl
@@ -24,6 +24,8 @@
#include "helpers.h"
+#if defined(CONV_STRIDE_X)
+
#if CONV_STRIDE_X == 1
#define convolution1x3 convolution1x3_stride_1
#elif CONV_STRIDE_X == 2
@@ -186,4 +188,134 @@
weights_values2.s0, weights_values2.s1, weights_values2.s2);
vstore2(pixels, 0, (__global float *)dst.ptr);
-}
\ No newline at end of file
+}
+
+#endif //defined(CONV_STRIDE_X)
+
+#if defined(SRC_WIDTH) && defined(DATA_TYPE)
+/** This kernel reshapes each of the tensor's low three dimensions to single rows.
+ *
+ * @note Datatype and source width should be given as a preprocessor argument using -DDATA_TYPE=type and -DSRC_WIDTH=width. e.g. -DSRC_WIDTH=128
+ *
+ * @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 Y 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. 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_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void depthwise_weights_reshape(TENSOR3D_DECLARATION(src), IMAGE_DECLARATION(dst))
+{
+ Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+
+ __global DATA_TYPE *input_ptr = (__global DATA_TYPE *)src.ptr;
+ __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(1) * SRC_WIDTH * dst_stride_x + get_global_id(2) * dst_stride_y;
+
+ for(int i = 0; i < SRC_WIDTH; ++i, ++input_ptr)
+ {
+ *((__global DATA_TYPE *)(output_ptr + i * dst_stride_x)) = *input_ptr;
+ }
+}
+#endif //defined(SRC_WIDTH) && defined(DATA_TYPE)
+
+#if defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_X) && defined(PAD_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DATA_TYPE)
+/** This kernel performs a reshaping of the input tensor to a tensor used to perform depthwise convolution using vector to matrix multiplication.
+ *
+ * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
+ * @note The convolution information must be passed at compile time using -DSTRIDE_X, -DSTRIDE_Y, -DPAD_X, -DPAD_Y, -DKERNEL_WIDHT, -DKERNEL_HEIGHT, -DSRC_WIDTH, -DSRC_HEIGHT
+ *
+ * @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: 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 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
+ */
+
+__kernel void depthwise_im2col(TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst))
+{
+ Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+ const int src_pixel_linear = get_global_id(1) * STRIDE_X;
+ const int full_length = SRC_WIDTH + 2 * PAD_X;
+ const int max_initial_x = STRIDE_X * (((full_length - KERNEL_WIDTH) / STRIDE_X) + 1);
+
+ const int src_x = -PAD_X + src_pixel_linear % max_initial_x;
+ const int src_y = -PAD_Y + src_pixel_linear / max_initial_x * STRIDE_Y;
+ const int src_z = get_global_id(2);
+
+ __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + src_z * src_stride_z;
+ __global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst.ptr));
+
+ for(int y = src_y; y < src_y + KERNEL_HEIGHT; ++y)
+ {
+ for(int x = src_x; x < src_x + KERNEL_WIDTH; ++x, ++output_ptr)
+ {
+ if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT)
+ {
+ *output_ptr = 0;
+ }
+ else
+ {
+ *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y));
+ }
+ }
+ }
+}
+
+#endif //defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_X) && defined(PAD_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_WIDTH) && defined(DATA_TYPE)
+
+#if defined(CONV_WIDTH) && defined(CONV_HEIGHT) && defined(DATA_TYPE)
+
+/** This kernel performs a reshaping of the output of the depthwise generic convolution.
+ *
+ * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
+ * @note The convolution information must be passed at compile time using -DCONV_WIDTH, -DCONV_HEIGHT, e.g -DCONV_WIDTH=32, -DCONV_HEIGHT=42
+ *
+ * @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_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 Y 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
+ */
+__kernel void depthwise_vector_to_tensor(
+ VECTOR_DECLARATION(src),
+ TENSOR3D_DECLARATION(dst))
+{
+ Vector src = CONVERT_TO_VECTOR_STRUCT(src);
+
+ const int patch_size = CONV_WIDTH * CONV_HEIGHT;
+ const int id0 = get_global_id(0);
+ const int z = id0 / patch_size;
+ const int index2D = id0 - z * patch_size;
+
+ __global uchar *out_ptr = dst_ptr + dst_offset_first_element_in_bytes + index2D % CONV_WIDTH * dst_stride_x + index2D / CONV_WIDTH * dst_stride_y + z * dst_stride_z;
+ *((__global DATA_TYPE *)out_ptr) = *((__global DATA_TYPE *)src.ptr);
+}
+
+#endif //defined(CONV_WIDTH) && defined(CONV_HEIGHT) && defined(DATA_TYPE)
diff --git a/src/core/CL/cl_kernels/gemv.cl b/src/core/CL/cl_kernels/gemv.cl
new file mode 100644
index 0000000..76128f7
--- /dev/null
+++ b/src/core/CL/cl_kernels/gemv.cl
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "helpers.h"
+
+/** This kernel applies dot product to each plane on the input tensor and the corrispective column of the reshaped weight tensor.
+ *
+ * @note Datatype and source width and height should be given as a preprocessor argument using -DDATA_TYPE=type, -DSRC_WIDTH=width and -DSRC_HEIGHT=height. e.g. -DDATA_TYPE=short
+ *
+ * @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 Y 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] weights_ptr Pointer to the weights tensor. 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_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[out] dst_ptr Pointer to the destination tensor. 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_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void gemm_mv(TENSOR3D_DECLARATION(src), IMAGE_DECLARATION(weights), VECTOR_DECLARATION(dst))
+{
+ Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+
+ int y = get_global_id(1) * 4;
+ int z = get_global_id(2);
+
+ __global uchar *current_weights = weights_ptr + weights_offset_first_element_in_bytes + z * weights_stride_y;
+ __global uchar *input_ptr = src.ptr;
+
+ DATA_TYPE acc0 = (DATA_TYPE)0;
+ DATA_TYPE acc1 = (DATA_TYPE)0;
+ DATA_TYPE acc2 = (DATA_TYPE)0;
+ DATA_TYPE acc3 = (DATA_TYPE)0;
+
+ // This kernel handle 4 rows in per thread so that it can reuse the weights
+ for(int i = 0; i < SRC_WIDTH; i += 4)
+ {
+ VEC_DATA_TYPE(DATA_TYPE, 4)
+ weights = vload4(0, (__global DATA_TYPE *)(current_weights + i * weights_stride_x));
+
+ int4 offset = (int4)i * (int4)src_stride_x + (int4)(0, 1, 2, 3) * (int4)src_stride_y;
+
+ VEC_DATA_TYPE(DATA_TYPE, 4)
+ tmp0 = vload4(0, (__global DATA_TYPE *)(input_ptr + offset.s0));
+ VEC_DATA_TYPE(DATA_TYPE, 4)
+ tmp1 = vload4(0, (__global DATA_TYPE *)(input_ptr + offset.s1));
+ VEC_DATA_TYPE(DATA_TYPE, 4)
+ tmp2 = vload4(0, (__global DATA_TYPE *)(input_ptr + offset.s2));
+ VEC_DATA_TYPE(DATA_TYPE, 4)
+ tmp3 = vload4(0, (__global DATA_TYPE *)(input_ptr + offset.s3));
+
+ acc0 += dot(weights, tmp0);
+ acc1 += dot(weights, tmp1);
+ acc2 += dot(weights, tmp2);
+ acc3 += dot(weights, tmp3);
+ }
+
+ __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (y + z * SRC_HEIGHT) * dst_stride_x;
+
+ int rows_left = SRC_HEIGHT - (y + 4);
+
+ // This if check is used to handle the last few rows when it can't be divided by the four
+ if(rows_left >= 0)
+ {
+ VEC_DATA_TYPE(DATA_TYPE, 4)
+ out = (VEC_DATA_TYPE(DATA_TYPE, 4))(acc0, acc1, acc2, acc3);
+ vstore4(out, 0, (__global DATA_TYPE *)output_ptr);
+ }
+ else
+ {
+ switch(rows_left)
+ {
+ case -1: // three rows left; one is padding
+ *((__global DATA_TYPE *)(output_ptr + 2 * dst_stride_x)) = acc2;
+ case -2: // two rows left; two are padding
+ *((__global DATA_TYPE *)(output_ptr + 1 * dst_stride_x)) = acc1;
+ case -3: // one row left; three are padding
+ *((__global DATA_TYPE *)(output_ptr + 0 * dst_stride_x)) = acc0;
+ break;
+ }
+ }
+}