diff options
author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2022-11-17 11:03:39 +0000 |
---|---|---|
committer | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2022-12-09 11:06:23 +0000 |
commit | 76335eb8d8733b0bbc0110546797211540870c50 (patch) | |
tree | 812fc44de593c9e1e45ac8b534094511b06163bf /src/core/CL/cl_kernels/nhwc/indirect_convolution.cl | |
parent | f16973b8b4605f12608bffa9f0ca6ed590202d41 (diff) | |
download | ComputeLibrary-76335eb8d8733b0bbc0110546797211540870c50.tar.gz |
Implement the OpenCL kernel to compute the indirect convolution
- Implement indirect convolution kernel
- Add operator support
- Add test
Resolves COMPMID-5709
Change-Id: I9272304163471a5a40da7fdec204599f3c1d8e32
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8701
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/CL/cl_kernels/nhwc/indirect_convolution.cl')
-rw-r--r-- | src/core/CL/cl_kernels/nhwc/indirect_convolution.cl | 224 |
1 files changed, 221 insertions, 3 deletions
diff --git a/src/core/CL/cl_kernels/nhwc/indirect_convolution.cl b/src/core/CL/cl_kernels/nhwc/indirect_convolution.cl index 07c7212e77..c88f0034c5 100644 --- a/src/core/CL/cl_kernels/nhwc/indirect_convolution.cl +++ b/src/core/CL/cl_kernels/nhwc/indirect_convolution.cl @@ -21,13 +21,16 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ - +#include "activation_float_helpers.h" #include "helpers.h" #include "tile_helpers.h" +#if defined(INDIRECT_CONVOLUTION_ADDRESS_PRECALCULATION) //! @cond Doxygen_Suppress /** OpenCL kernel to compute the indirect convolution 2d indirect buffer. * + * @note This kernel only works for unit batch_size + * * @note The convolution padding (left and top) must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (e.g. -DPAD_LEFT=2, -DPAD_TOP=2) * @note The convolution strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y (e.g. -DSTRIDE_X=2, -DSTRIDE_Y=2) * @note The kernel width must be passed at compile time using -DWEI_CONV_WIDTH (e.g. -DWEI_CONV_WIDTH=9) @@ -38,7 +41,7 @@ * @note The number of M0 rows (width*height) to process must be passed at compile time using -DM0 (e.g. -DM0=2) * - M0 = 1, 2, 3, 4, 5, 6, 7, and 8 * - * @param[out] dst_img CLImage object to the destination tensor (DST_TENSOR_TYPE=IMAGE only) + * @param[out] dst_img (Not supported) CLImage object to the destination tensor (DST_TENSOR_TYPE=IMAGE only) * @param[out] dst_ptr Pointer to the destination tensor. Supported data type: INT32 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) @@ -90,4 +93,219 @@ __kernel void indirect_convolution_address_precalculation( VSTORE(1) (my[0].s[0], 0, (__global DST_DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(DST_DATA_TYPE) + y * dst_stride_y + z * dst_stride_z)); -}
\ No newline at end of file +} +#endif // defined(INDIRECT_CONVOLUTION_ADDRESS_PRECALCULATION) + +#if defined(INDIRECT_CONVOLUTION_NHWC) +//! @cond Doxygen_Suppress +/** OpenCL kernel to compute the indirect convolution. + * + * @note Data layout supported: NHWC + * @note Data type supported: F32/F16 + * @note The spatial dimensions of the weights must be passed at compile time using -DWEI_WIDTH and -DWEI_HEIGHT (e.g. -DWEI_WIDTH=9, -DWEI_HEIGHT=9) + * @note The spatial dimensions of the destination tensor must be passed at compile time using -DDST_WIDTH and -DDST_HEIGHT (e.g. -DDST_WIDTH=96, -DDST_HEIGHT=64) + * @note The channels of the source tensor must be passed at compile time using -DSRC_CHANNELS (e.g. -DSRC_CHANNELS=64) + * @note The tensor type ("BUFFER" or "IMAGE") of the source tensor must be passed at compile time using -DSRC_TENSOR_TYPE (e.g. -DSRC_TENSOR_TYPE=BUFFER) + * @note The tensor type ("BUFFER" or "IMAGE") of the weights tensor must be passed at compile time using -DWEI_TENSOR_TYPE (e.g. -DWEI_TENSOR_TYPE=BUFFER) + * @note The tensor type ("BUFFER" or "IMAGE") of the destination tensor must be passed at compile time using -DDST_TENSOR_TYPE (e.g. -DDST_TENSOR_TYPE=BUFFER) + * @note The data type of the source tensor must be passed at compile time using -DSRC_DATA_TYPE (e.g. -DSRC_DATA_TYPE=float) + * @note The data type of the weights tensor must be passed at compile time using -DWEI_DATA_TYPE (e.g. -DWEI_DATA_TYPE=float) + * @note The data type of the destination tensor must be passed at compile time using -DDST_DATA_TYPE (e.g. -DDST_DATA_TYPE=float) + * @note The number of M0 rows (width*height) to process must be passed at compile time using -DM0 (e.g. -DM0=2) + * @note The number of N0 output channels to process must be passed at compile time using -DN0 (e.g. -DN0=2) + * @note The number of K0 inner accumulations must be passed at compile time using -DK0 (e.g. -DK0=2) + * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_N0 (e.g. -DPARTIAL_N0=1) + * @note The vector length used for loading the values from the indirect buffer should be passed at compile time using -DIND_BUFF_VEC_SIZE (e.g. -DIND_BUFF_VEC_SIZE=4) + * @note The activation function to fuse and corresponding A and B values should be passed at compile time using -DACTIVATION_TYPE, -DA_VAL, and -DB_VAL + * (e.g. -DFUNCTION_TYPE=lu_brelu_op, -DA_VAL=3.0, and -DB_VAL=1.0) + * @note Only the following configurations of M0, N0 and K0 are currently supported: + * - M0 = 1, 2, 3, 4, 5, 6, and 8 + * - N0 = 2, 3, 4, 8, 16 + * - K0 = 2, 3, 4, 8, 16 (only 4, 8 and 16 if WEI_TENSOR_TYPE=IMAGE) + * + * @param[in] src_img (Not supported) CLImage object to the source tensor (SRC_TENSOR_TYPE=IMAGE only) + * @param[in] src_ptr Pointer to the source tensor. Supported data type: F16/F32 + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) + * @param[in] src_c The size of the channels dimension of the source tensor + * @param[in] src_w The size of the width dimension of the source tensor + * @param[in] src_h The size of the height dimension of the source tensor + * @param[in] src_n The size of the batches dimension of the source tensor + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] off_img (Not supported) CLImage object to the indirect buffer tensor (OFF_TENSOR_TYPE=IMAGE only) + * @param[in] off_ptr Pointer to the indirect buffer tensor. Supported data type: INT32 + * @param[in] off_stride_y Stride of the indirect buffer tensor in Y dimension (in bytes) + * @param[in] off_stride_z Stride of the indirect buffer tensor in Z dimension (in bytes) + * @param[in] off_stride_w Stride of the indirect buffer tensor in W dimension (in bytes) + * @param[in] off_c The size of the channels dimension of the indirect buffer tensor + * @param[in] off_w The size of the width dimension of the indirect buffer tensor + * @param[in] off_h The size of the height dimension of the indirect buffer tensor + * @param[in] off_n The size of the batches dimension of the indirect buffer tensor + * @param[in] off_offset_first_element_in_bytes The offset of the first element in the indirect buffer tensor + * @param[out] dst_img (Not supported) CLImage object to the destination tensor (DST_TENSOR_TYPE=IMAGE only) + * @param[out] dst_ptr Pointer to the destination tensor. Supported data type: same as the input tensor + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) + * @param[in] dst_c The size of the channels dimension of the destination tensor + * @param[in] dst_w The size of the width dimension of the destination tensor + * @param[in] dst_h The size of the height dimension of the destination tensor + * @param[in] dst_n The size of the batches dimension of the destination tensor + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[out] wei_img (Optional) CLImage object to the destination tensor (WEI_TENSOR_TYPE=IMAGE only) + * @param[out] wei_ptr Pointer to the weights tensor. Supported data type: same as the input tensor + * @param[in] wei_stride_y Stride of the weights tensor in Y dimension (in bytes) + * @param[in] wei_stride_z Stride of the weights tensor in Z dimension (in bytes) + * @param[in] wei_stride_w Stride of the weights tensor in W dimension (in bytes) + * @param[in] wei_c The size of the channels dimension of the weights tensor + * @param[in] wei_w The size of the width dimension of the weights tensor + * @param[in] wei_h The size of the height dimension of the weights tensor + * @param[in] wei_n The size of the batches dimension of the weights tensor + * @param[in] wei_offset_first_element_in_bytes The offset of the first element in the weights tensor + * @param[out] bia_img (Not supported) CLImage object to the destination tensor (BIA_TENSOR_TYPE=IMAGE only) + * @param[out] bia_ptr (Optional) Pointer to the bias tensor. Supported data type: same as the input tensor + * @param[in] bia_stride_y (Optional) Stride of the bias tensor in Y dimension (in bytes) + * @param[in] bia_stride_z (Optional) Stride of the bias tensor in Z dimension (in bytes) + * @param[in] bia_stride_w (Optional) Stride of the bias tensor in W dimension (in bytes) + * @param[in] bia_c (Optional) The size of the channels dimension of the bias tensor + * @param[in] bia_w (Optional) The size of the width dimension of the bias tensor + * @param[in] bia_h (Optional) The size of the height dimension of the bias tensor + * @param[in] bia_n (Optional) The size of the batches dimension of the bias tensor + * @param[in] bia_offset_first_element_in_bytes (Optional) The offset of the first element in the bias tensor + */ +//! @endcond +__kernel void indirect_convolution_nhwc( + TENSOR4D_T(src, SRC_TENSOR_TYPE), + TENSOR4D_T(off, OFF_TENSOR_TYPE), + TENSOR4D_T(dst, DST_TENSOR_TYPE), + TENSOR4D_T(wei, WEI_TENSOR_TYPE) +#if defined(HAS_BIAS) + , + VECTOR_DECLARATION(bia) +#endif // defined(HAS_BIAS) +) +{ + // All the tensor dimensions are passed at compile time. + // In case of dynamic tensor support, the following dimensions should be passed as function argument. +#define _IWEI_WIDTH WEI_WIDTH +#define _IWEI_HEIGHT WEI_HEIGHT +#define _ISRC_CHANNELS SRC_CHANNELS +#define _IDST_WIDTH DST_WIDTH +#define _IDST_HEIGHT DST_HEIGHT +#define _IY_MULTIPLIER (_IWEI_WIDTH * _IWEI_HEIGHT) + + const int cout = GET_SPATIAL_IDX(0, N0, PARTIAL_N0); // OFM + const int mout = GET_SPATIAL_IDX(1, M0, 0); // WIDTH x HEIGHT + const int bout = GET_SPATIAL_IDX(2, 1, 0); // BATCH SIZE IDX + + off_offset_first_element_in_bytes += get_global_id(1) * off_stride_y; + off_offset_first_element_in_bytes += bout * off_stride_z; + + // Initialize the accumulators + TILE(DST_DATA_TYPE, M0, N0, c); + + LOOP_UNROLLING(int, i, 0, 1, M0, + { + c[i].v = 0; + }) + + for(int i = 0; i < (_IWEI_WIDTH * _IWEI_HEIGHT); ++i) + { + TILE(int, 1, IND_BUFF_VEC_SIZE, my); + T_LOAD(int, 1, IND_BUFF_VEC_SIZE, OFF_TENSOR_TYPE, off, i * M0, 0, 1, 0, my); + + int ck = 0; + for(; ck <= (_ISRC_CHANNELS - K0); ck += K0) + { + TILE(SRC_DATA_TYPE, M0, K0, a); + TILE(WEI_DATA_TYPE, N0, K0, b); + + // Initialize tiles + LOOP_UNROLLING(int, i, 0, 1, M0, + { + a[i].v = 0.0; + }) + + LOOP_UNROLLING(int, i, 0, 1, N0, + { + b[i].v = 0.0; + }) + + // Load tile from the src tensor + T_LOAD2D_INDIRECT(SRC_DATA_TYPE, M0, K0, SRC_TENSOR_TYPE, src, ck, src_stride_y, my, a); + + // Load tile from the weights tensor + T_LOAD(WEI_DATA_TYPE, N0, K0, WEI_TENSOR_TYPE, wei, ck, cout * _IY_MULTIPLIER + i, _IY_MULTIPLIER, wei_stride_y, b); + + // Compute the matrix multiplication between two tiles + T_MMUL(SRC_DATA_TYPE, WEI_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, NT, T, a, b, c); + } + + // This #if directive should be removed in case of dynamic tensor support +#if defined(LEFTOVER_LOOP) + // Left-over accumulations + for(; ck < _ISRC_CHANNELS; ++ck) + { + TILE(SRC_DATA_TYPE, M0, 1, a); + TILE(WEI_DATA_TYPE, N0, 1, b); + + // Initialize tiles + LOOP_UNROLLING(int, i, 0, 1, M0, + { + a[i].v = 0.0; + }) + + LOOP_UNROLLING(int, i, 0, 1, N0, + { + b[i].v = 0.0; + }) + + // Load tile from the src tensor + T_LOAD2D_INDIRECT(SRC_DATA_TYPE, M0, 1, SRC_TENSOR_TYPE, src, ck, src_stride_y, my, a); + + // Load tile from the weights tensor + // The T_LOAD for the left-over elements can only use BUFFER because we load one element per iteration + T_LOAD(WEI_DATA_TYPE, N0, 1, BUFFER, wei, ck, cout * _IY_MULTIPLIER + i, _IY_MULTIPLIER, wei_stride_y, b); + + // Compute the matrix multiplication between two tiles + T_MMUL(SRC_DATA_TYPE, WEI_DATA_TYPE, DST_DATA_TYPE, M0, N0, 1, NT, T, a, b, c); + } +#endif // defined(LEFTOVER_LOOP) + } + +#if defined(HAS_BIAS) + TILE(BIA_DATA_TYPE, 1, N0, bias0); + + T_LOAD(BIA_DATA_TYPE, 1, N0, BUFFER, bia, cout, 0, 1, 0, bias0); + + // c = c + bias[broadcasted] + T_ELTWISE_BROADCAST_ADD_X(DST_DATA_TYPE, M0, N0, c, bias0, c); + +#endif // HAS_BIAS + + // Apply activation + T_ACTIVATION(DST_DATA_TYPE, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, c, c); + + TILE(uint, M0, 1, dst_indirect_y); + + // Calculate the destination indirect Y + LOOP_UNROLLING(int, i, 0, 1, M0, + { + dst_indirect_y[i].v = (uint)min(mout + i, (int)(_IDST_WIDTH * _IDST_HEIGHT) - 1); + dst_indirect_y[i].v += bout * (int)(_IDST_WIDTH * _IDST_HEIGHT); + }) + + const bool x_cond = PARTIAL_N0 != 0 && get_global_id(0) == 0; + + // Store the tile in reverse order so the invalid values are overwritten with the valid ones + T_STORE_INDIRECT_WIDTH_SELECT(DST_DATA_TYPE, M0, N0, PARTIAL_N0, DST_TENSOR_TYPE, dst, cout, dst_stride_y, x_cond, c, dst_indirect_y); + +#undef _IWEI_WIDTH +#undef _IWEI_HEIGHT +#undef _ISRC_CHANNELS +#undef _IDST_WIDTH +#undef _IDST_HEIGHT +#undef _IY_MULTIPLIER +} +#endif // defined(INDIRECT_CONVOLUTION_NHWC) |