onednn-src 0.1.13

Source of oneAPI Deep Neural Network Library (oneDNN)
Documentation
/*******************************************************************************
* Copyright 2021 Intel Corporation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
*     http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*******************************************************************************/

#include "gpu/intel/include/dispatch.h"
#include "gpu/intel/include/io.h"
#include "gpu/intel/include/types.h"

#include "gpu/intel/concat/common.h"

#define SCALE_PTR(n) __global const float *scale##n

#define IS_IN_PART(x) (dst_dims[CONCAT_AXIS] < CONCAT3(SRC, x, _END))
#define WRITE_DST_X(x) write(dst + dst_off, x);
#if SCALES_MASK > 0
#define EXTRACT_SCALES_IF(n) \
    if (SCALES_MASK & (1 << n)) tmp_src_scale = *scale##n;
#define WRITE_DST WRITE_DST_X((float)src_val *tmp_src_scale)
#define INIT_FLOAT_SCALE float tmp_src_scale = 1.f;
#define SCALE_PTRS , REDUCE(NUM_INPUTS, JOIN_COMMA, SCALE_PTR)
#else
#define WRITE_DST WRITE_DST_X(src_val)
#define EXTRACT_SCALES_IF(n)
#define INIT_FLOAT_SCALE
#define SCALE_PTRS
#define WRITE_SCALE write(dst + dst_off, src_val);
#endif

#define SET_DIMS(x, y) \
    { \
        part = y; \
        EXTRACT_SCALES_IF(y); \
        if (y > 0) { \
            src_dims[CONCAT_AXIS] \
                    = dst_dims[CONCAT_AXIS] - CONCAT3(SRC, x, _END); \
        } \
        src_off = OFF_MD(CONCAT2(SRC, y), src_dims[0], src_dims[1], \
                src_dims[2], src_dims[3], src_dims[4], src_dims[5]); \
        src = CONCAT2(src, y); \
    }

#define SRC_DATA_T SRC0_DATA_T
#define DD(i) CONCAt2(DST_D, i)
#define NEEDS_PADDING(dim0, dim1, dim2, dim3, dim4, dim5) \
    dim0 >= DD(0) || dim1 >= DD(1) || dim2 >= DD(2) || dim3 >= DD(3) \
            || dim4 >= DD(4) || dim5 >= DD(5)

KERNEL_ATTR
__kernel void xe_concat(__global DST_DATA_T *dst, long dst_offset0,
        __global const SRC_DATA_T *src0, __global const SRC_DATA_T *src1,
        __global const SRC_DATA_T *src2, __global const SRC_DATA_T *src3,
        __global const SRC_DATA_T *src4, __global const SRC_DATA_T *src5,
        __global const SRC_DATA_T *src6, __global const SRC_DATA_T *src7,
        __global const SRC_DATA_T *src8, __global const SRC_DATA_T *src9,
        __global const SRC_DATA_T *src10, __global const SRC_DATA_T *src11,
        __global const SRC_DATA_T *src12, __global const SRC_DATA_T *src13,
        __global const SRC_DATA_T *src14,
        __global const SRC_DATA_T *src15 SCALE_PTRS) {
    dst += dst_offset0;
    off_t dst_dims[6], src_dims[6];
    src_dims[0] = dst_dims[0] = GWS_GET_D0();
    src_dims[1] = dst_dims[1] = GWS_GET_D1();
    src_dims[2] = dst_dims[2] = GWS_GET_D2();
    src_dims[3] = dst_dims[3] = GWS_GET_D3();
    src_dims[4] = dst_dims[4] = GWS_GET_D4();
    src_dims[5] = dst_dims[5] = GWS_GET_D5();

    const off_t iter_dim_end = min(
            dst_dims[ITER_DIM_IDX] + ITER_DIM_CHUNK, ITER_DIM_PADDED_SIZE);

    if (NEEDS_PADDING(dst_dims[0], dst_dims[1], dst_dims[2], dst_dims[3],
                dst_dims[4], dst_dims[5])) {
        for (; dst_dims[ITER_DIM_IDX] < iter_dim_end;
                dst_dims[ITER_DIM_IDX]++) {
            const off_t dst_off = OFF_MD(DST, dst_dims[0], dst_dims[1],
                    dst_dims[2], dst_dims[3], dst_dims[4], dst_dims[5]);
#if SUB_GROUP_SIZE > 1
            float zval = 0.0f;
            block_write(dst + dst_off, &zval, VECT_DT_N);
#else // SUB_GROUP_SIZE > 1
            write(dst + dst_off, 0.0f);
#endif // SUB_GROUP_SIZE > 1
        }
        return;
    }
    for (; dst_dims[ITER_DIM_IDX] < min(DD(ITER_DIM_IDX), iter_dim_end);
            dst_dims[ITER_DIM_IDX]++, src_dims[ITER_DIM_IDX]++) {
        int part;
        off_t src_off;
        const __global SRC_DATA_T *src;
        INIT_FLOAT_SCALE;

        if (IS_IN_PART(0)) SET_DIMS(0, 0)
#if NUM_INPUTS >= 2
        else if (IS_IN_PART(1))
            SET_DIMS(0, 1)
#endif
#if NUM_INPUTS >= 3
        else if (IS_IN_PART(2))
            SET_DIMS(1, 2)
#endif
#if NUM_INPUTS >= 4
        else if (IS_IN_PART(3))
            SET_DIMS(2, 3)
#endif
#if NUM_INPUTS >= 5
        else if (IS_IN_PART(4))
            SET_DIMS(3, 4)
#endif
#if NUM_INPUTS >= 6
        else if (IS_IN_PART(5))
            SET_DIMS(4, 5)
#endif
#if NUM_INPUTS >= 7
        else if (IS_IN_PART(6))
            SET_DIMS(5, 6)
#endif
#if NUM_INPUTS >= 8
        else if (IS_IN_PART(7))
            SET_DIMS(6, 7)
#endif
#if NUM_INPUTS >= 9
        else if (IS_IN_PART(8))
            SET_DIMS(7, 8)
#endif
#if NUM_INPUTS >= 10
        else if (IS_IN_PART(9))
            SET_DIMS(8, 9)
#endif
#if NUM_INPUTS >= 11
        else if (IS_IN_PART(10))
            SET_DIMS(9, 10)
#endif
#if NUM_INPUTS >= 12
        else if (IS_IN_PART(11))
            SET_DIMS(10, 11)
#endif
#if NUM_INPUTS >= 13
        else if (IS_IN_PART(12))
            SET_DIMS(11, 12)
#endif
#if NUM_INPUTS >= 14
        else if (IS_IN_PART(13))
            SET_DIMS(12, 13)
#endif
#if NUM_INPUTS >= 15
        else if (IS_IN_PART(14))
            SET_DIMS(13, 14)
#endif
#if NUM_INPUTS >= 16
        else if (IS_IN_PART(15))
            SET_DIMS(14, 15)
#endif

        const off_t dst_off = OFF_MD(DST, dst_dims[0], dst_dims[1], dst_dims[2],
                dst_dims[3], dst_dims[4], dst_dims[5]);

        FLT_ACC_DATA_T src_val;
#if SUB_GROUP_SIZE > 1
        block_load(&src_val, (__global SRC_DATA_T *)src + src_off);
#if SCALES_MASK > 0
        src_val = SCALES_MASK ? src_val * tmp_src_scale : src_val;
#endif
        block_write(dst + dst_off, &src_val, VECT_DT_N);
#else // SUB_GROUP_SIZE > 1
        load(&src_val, src + src_off);
        WRITE_DST
#endif // SUB_GROUP_SIZE > 1
    }
    for (; dst_dims[ITER_DIM_IDX] < iter_dim_end; dst_dims[ITER_DIM_IDX]++) {
        const off_t dst_off = OFF_MD(DST, dst_dims[0], dst_dims[1], dst_dims[2],
                dst_dims[3], dst_dims[4], dst_dims[5]);
#if SUB_GROUP_SIZE > 1
        FLT_ACC_DATA_T zval = 0.f;
        block_write(dst + dst_off, &zval, VECT_DT_N);
#else // SUB_GROUP_SIZE > 1
        write(dst + dst_off, 0.f);
#endif // SUB_GROUP_SIZE > 1
    }
}