Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 17b0f8ba60ec9db4b96471f9406843bee6a43a4f / . / src / core / CL / cl_kernels / gemmlowp.cl
blob: 8c1fa548e4a4754d37e6dd504c100ae1659e1184 [file] [log] [blame]
/*
* Copyright (c) 2017-2018 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"
#include "helpers_asymm.h"
#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
#define ARM_DOT(x, y, val) val = arm_dot_acc((x), (y), (val));
#else // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
#define ARM_DOT(x, y, val) val += arm_dot((x), (y));
#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#if defined(COLS_B) && defined(MULT_INTERLEAVE4X4_HEIGHT) && defined(TRANSPOSE1XW_WIDTH_STEP)
/** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1)
* Matrix A and matrix B must be reshaped respectively with @ref CLGEMMInterleave4x4Kernel and @ref CLGEMMTranspose1xWKernel before running the matrix multiplication
*
* @note The number of matrix B columns needs to be passed at compile time using -DCOLS_B: e.g. -DCOLS_B=1024
* @note The transposition width step (mult_transpose1xW_width * 4) must be passed at compile time using -DTRANSPOSE1XW_WIDTH_STEP (i.e. -DTRANSPOSE1XW_WIDTH_STEP=2)
* @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2)
*
* @note In case the output has to be reinterpreted as a 3D tensor (i.e. output of convolution layer), the following information must be passed at compile time:
* -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
* -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
* -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
* (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
*
* @param[in] src0_ptr Pointer to the source matrix. Supported data type: QASYMM8
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported data type: same as @p src0_ptr
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data type: S32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_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 matrix
* @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
*/
__kernel void gemmlowp_mm_interleaved_transposed_midgard(IMAGE_DECLARATION(src0),
IMAGE_DECLARATION(src1),
IMAGE_DECLARATION(dst),
uint src0_stride_z,
uint src1_stride_z,
uint dst_stride_z
#if defined(REINTERPRET_OUTPUT_AS_3D)
,
uint cross_plane_pad
#endif // REINTERPRET_OUTPUT_AS_3D
)
{
const int x = get_global_id(0) / TRANSPOSE1XW_WIDTH_STEP;
const int y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;
const int z = get_global_id(2);
// Offset
const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4;
const int offset_row_b = (get_global_id(0) % TRANSPOSE1XW_WIDTH_STEP) * 4;
// src_addr_a = address of matrix A
// src_addr_b = address of matrix B
__global uchar *src_addr_a = (__global uchar *)(src0_ptr + z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes);
__global uchar *src_addr_b = (__global uchar *)(src1_ptr + x * src1_stride_y + src1_offset_first_element_in_bytes);
#if defined(MATRIX_B_DEPTH)
// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
src_addr_b += (z % MATRIX_B_DEPTH) * src1_stride_z;
#else // defined(MATRIX_B_DEPTH)
src_addr_b += z * src1_stride_z;
#endif // defined(MATRIX_B_DEPTH)
// Compute end row address for matrix B
__global uchar *src_end_addr_b = src_addr_b + COLS_B;
src_addr_a += offset_row_a;
src_addr_b += offset_row_b;
// Reset accumulators
int4 c00 = 0;
int4 c10 = 0;
int4 c20 = 0;
int4 c30 = 0;
for(; src_addr_b <= (src_end_addr_b - (int)(8 * TRANSPOSE1XW_WIDTH_STEP)); src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * TRANSPOSE1XW_WIDTH_STEP)
{
// Load values from matrix A (interleaved) and matrix B (transposed)
int4 a0 = convert_int4(vload4(0, src_addr_a));
int4 b0 = convert_int4(vload4(0, src_addr_b));
c00 += (int4)a0.s0 * b0;
c10 += (int4)a0.s1 * b0;
c20 += (int4)a0.s2 * b0;
c30 += (int4)a0.s3 * b0;
a0 = convert_int4(vload4(0, src_addr_a + 4 * MULT_INTERLEAVE4X4_HEIGHT));
b0 = convert_int4(vload4(0, src_addr_b + 4 * TRANSPOSE1XW_WIDTH_STEP));
c00 += (int4)a0.s0 * b0;
c10 += (int4)a0.s1 * b0;
c20 += (int4)a0.s2 * b0;
c30 += (int4)a0.s3 * b0;
}
for(; src_addr_b < src_end_addr_b; src_addr_a += (4 * MULT_INTERLEAVE4X4_HEIGHT), src_addr_b += (4 * TRANSPOSE1XW_WIDTH_STEP))
{
// Load values from matrix A (interleaved) and matrix B (transposed)
int4 a0 = convert_int4(vload4(0, src_addr_a));
int4 b0 = convert_int4(vload4(0, src_addr_b));
c00 += (int4)a0.s0 * b0;
c10 += (int4)a0.s1 * b0;
c20 += (int4)a0.s2 * b0;
c30 += (int4)a0.s3 * b0;
}
// Compute destination address
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
#if defined(REINTERPRET_OUTPUT_AS_3D)
// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
// in order to take into account the presence of possible cross plane paddings
//
// | |
// | plane0 |
// | |
// |__________________|
// |******************|
// | cross_plane_pad |
// |******************|
// | |
// | plane1 |
// | |
// |__________________|
// The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D
uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;
zout = min(DEPTH_GEMM3D - 1, zout);
// Add offset due to the cross plane paddings
zout *= (cross_plane_pad * dst_stride_y);
// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
// multiply dst_stride_z by DEPTH_GEMM3D
dst.ptr += z * dst_stride_z * DEPTH_GEMM3D;
// Store 4x4 block
vstore4(c00, 0, (__global int *)(dst.ptr + 0 * dst_stride_y + zout.s0));
vstore4(c10, 0, (__global int *)(dst.ptr + 1 * dst_stride_y + zout.s1));
vstore4(c20, 0, (__global int *)(dst.ptr + 2 * dst_stride_y + zout.s2));
vstore4(c30, 0, (__global int *)(dst.ptr + 3 * dst_stride_y + zout.s3));
#else // defined(REINTERPRET_OUTPUT_AS_3D)
// Add offset for batched GEMM
dst.ptr += z * dst_stride_z;
// Store 4x4 block
vstore4(c00, 0, (__global int *)(dst.ptr + 0 * dst_stride_y));
vstore4(c10, 0, (__global int *)(dst.ptr + 1 * dst_stride_y));
vstore4(c20, 0, (__global int *)(dst.ptr + 2 * dst_stride_y));
vstore4(c30, 0, (__global int *)(dst.ptr + 3 * dst_stride_y));
#endif // defined(REINTERPRET_OUTPUT_AS_3D)
}
/** This OpenCL kernel is optimized for Bifrost and computes the matrix multiplication between matrix A (src0) and matrix B (src1)
* Matrix A and matrix B must be reshaped respectively with @ref CLGEMMInterleave4x4Kernel and @ref CLGEMMTranspose1xWKernel before running the matrix multiplication
*
* @attention The number of matrix B columns needs to be passed at compile time using -DCOLS_B
* @note The transposition width step (mult_transpose1xW_width * 4) must be passed at compile time using -DTRANSPOSE1XW_WIDTH_STEP (i.e. -DTRANSPOSE1XW_WIDTH_STEP=2)
* @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2)
*
* @note In case the output has to be reinterpreted as a 3D tensor (i.e. output of convolution layer), the following information must be passed at compile time:
* -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
* -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
* -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
* (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
*
* @param[in] src0_ptr Pointer to the source matrix. Supported data type: QASYMM8
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported data type: same as @p src0_ptr
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data type: S32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_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 matrix
* @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
*/
__kernel void gemmlowp_mm_interleaved_transposed_bifrost(IMAGE_DECLARATION(src0),
IMAGE_DECLARATION(src1),
IMAGE_DECLARATION(dst),
uint src0_stride_z,
uint src1_stride_z,
uint dst_stride_z
#if defined(REINTERPRET_OUTPUT_AS_3D)
,
uint cross_plane_pad
#endif // REINTERPRET_OUTPUT_AS_3D
)
{
const int x = get_global_id(0) / TRANSPOSE1XW_WIDTH_STEP;
const int y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;
const int z = get_global_id(2);
// Offset
const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4;
const int offset_row_b = (get_global_id(0) % TRANSPOSE1XW_WIDTH_STEP) * 4;
// src_addr_a = address of matrix A
// src_addr_b = address of matrix B
__global uchar *src_addr_a = (__global uchar *)(src0_ptr + z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes);
__global uchar *src_addr_b = (__global uchar *)(src1_ptr + x * src1_stride_y + src1_offset_first_element_in_bytes);
#if defined(MATRIX_B_DEPTH)
// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
src_addr_b += (z % MATRIX_B_DEPTH) * src1_stride_z;
#else // defined(MATRIX_B_DEPTH)
src_addr_b += z * src1_stride_z;
#endif // defined(MATRIX_B_DEPTH)
// Compute end row address for matrix B
__global uchar *src_end_addr_b = src_addr_b + COLS_B;
src_addr_a += offset_row_a;
src_addr_b += offset_row_b;
// Reset accumulators
uint c00 = 0;
uint c01 = 0;
uint c02 = 0;
uint c03 = 0;
uint c10 = 0;
uint c11 = 0;
uint c12 = 0;
uint c13 = 0;
uint c20 = 0;
uint c21 = 0;
uint c22 = 0;
uint c23 = 0;
uint c30 = 0;
uint c31 = 0;
uint c32 = 0;
uint c33 = 0;
#if MULT_INTERLEAVE4X4_HEIGHT == 1
for(; src_addr_b <= (src_end_addr_b - (int)(32 * TRANSPOSE1XW_WIDTH_STEP)); src_addr_a += (32 * MULT_INTERLEAVE4X4_HEIGHT), src_addr_b += (32 * TRANSPOSE1XW_WIDTH_STEP))
{
// Load values from matrix A (interleaved) and matrix B (transposed)
uchar16 a0 = vload16(0, src_addr_a);
uchar4 b0 = vload4(0, src_addr_b);
c00 += (ushort)a0.s0 * b0.s0;
c01 += (ushort)a0.s0 * b0.s1;
c02 += (ushort)a0.s0 * b0.s2;
c03 += (ushort)a0.s0 * b0.s3;
c10 += (ushort)a0.s1 * b0.s0;
c11 += (ushort)a0.s1 * b0.s1;
c12 += (ushort)a0.s1 * b0.s2;
c13 += (ushort)a0.s1 * b0.s3;
c20 += (ushort)a0.s2 * b0.s0;
c21 += (ushort)a0.s2 * b0.s1;
c22 += (ushort)a0.s2 * b0.s2;
c23 += (ushort)a0.s2 * b0.s3;
c30 += (ushort)a0.s3 * b0.s0;
c31 += (ushort)a0.s3 * b0.s1;
c32 += (ushort)a0.s3 * b0.s2;
c33 += (ushort)a0.s3 * b0.s3;
// Load values from matrix B (transposed)
b0 = vload4(0, src_addr_b + 4 * TRANSPOSE1XW_WIDTH_STEP);
c00 += (ushort)a0.s4 * b0.s0;
c01 += (ushort)a0.s4 * b0.s1;
c02 += (ushort)a0.s4 * b0.s2;
c03 += (ushort)a0.s4 * b0.s3;
c10 += (ushort)a0.s5 * b0.s0;
c11 += (ushort)a0.s5 * b0.s1;
c12 += (ushort)a0.s5 * b0.s2;
c13 += (ushort)a0.s5 * b0.s3;
c20 += (ushort)a0.s6 * b0.s0;
c21 += (ushort)a0.s6 * b0.s1;
c22 += (ushort)a0.s6 * b0.s2;
c23 += (ushort)a0.s6 * b0.s3;
c30 += (ushort)a0.s7 * b0.s0;
c31 += (ushort)a0.s7 * b0.s1;
c32 += (ushort)a0.s7 * b0.s2;
c33 += (ushort)a0.s7 * b0.s3;
// Load values from matrix B (transposed)
b0 = vload4(0, src_addr_b + 8 * TRANSPOSE1XW_WIDTH_STEP);
c00 += (ushort)a0.s8 * b0.s0;
c01 += (ushort)a0.s8 * b0.s1;
c02 += (ushort)a0.s8 * b0.s2;
c03 += (ushort)a0.s8 * b0.s3;
c10 += (ushort)a0.s9 * b0.s0;
c11 += (ushort)a0.s9 * b0.s1;
c12 += (ushort)a0.s9 * b0.s2;
c13 += (ushort)a0.s9 * b0.s3;
c20 += (ushort)a0.sA * b0.s0;
c21 += (ushort)a0.sA * b0.s1;
c22 += (ushort)a0.sA * b0.s2;
c23 += (ushort)a0.sA * b0.s3;
c30 += (ushort)a0.sB * b0.s0;
c31 += (ushort)a0.sB * b0.s1;
c32 += (ushort)a0.sB * b0.s2;
c33 += (ushort)a0.sB * b0.s3;
// Load values from matrix B (transposed)
b0 = vload4(0, src_addr_b + 12 * TRANSPOSE1XW_WIDTH_STEP);
c00 += (ushort)a0.sC * b0.s0;
c01 += (ushort)a0.sC * b0.s1;
c02 += (ushort)a0.sC * b0.s2;
c03 += (ushort)a0.sC * b0.s3;
c10 += (ushort)a0.sD * b0.s0;
c11 += (ushort)a0.sD * b0.s1;
c12 += (ushort)a0.sD * b0.s2;
c13 += (ushort)a0.sD * b0.s3;
c20 += (ushort)a0.sE * b0.s0;
c21 += (ushort)a0.sE * b0.s1;
c22 += (ushort)a0.sE * b0.s2;
c23 += (ushort)a0.sE * b0.s3;
c30 += (ushort)a0.sF * b0.s0;
c31 += (ushort)a0.sF * b0.s1;
c32 += (ushort)a0.sF * b0.s2;
c33 += (ushort)a0.sF * b0.s3;
// Load values from matrix A (interleaved) and matrix B (transposed)
a0 = vload16(0, src_addr_a + 16);
b0 = vload4(0, src_addr_b + 16 * TRANSPOSE1XW_WIDTH_STEP);
c00 += (ushort)a0.s0 * b0.s0;
c01 += (ushort)a0.s0 * b0.s1;
c02 += (ushort)a0.s0 * b0.s2;
c03 += (ushort)a0.s0 * b0.s3;
c10 += (ushort)a0.s1 * b0.s0;
c11 += (ushort)a0.s1 * b0.s1;
c12 += (ushort)a0.s1 * b0.s2;
c13 += (ushort)a0.s1 * b0.s3;
c20 += (ushort)a0.s2 * b0.s0;
c21 += (ushort)a0.s2 * b0.s1;
c22 += (ushort)a0.s2 * b0.s2;
c23 += (ushort)a0.s2 * b0.s3;
c30 += (ushort)a0.s3 * b0.s0;
c31 += (ushort)a0.s3 * b0.s1;
c32 += (ushort)a0.s3 * b0.s2;
c33 += (ushort)a0.s3 * b0.s3;
// Load values from matrix B (transposed)
b0 = vload4(0, src_addr_b + 20 * TRANSPOSE1XW_WIDTH_STEP);
c00 += (ushort)a0.s4 * b0.s0;
c01 += (ushort)a0.s4 * b0.s1;
c02 += (ushort)a0.s4 * b0.s2;
c03 += (ushort)a0.s4 * b0.s3;
c10 += (ushort)a0.s5 * b0.s0;
c11 += (ushort)a0.s5 * b0.s1;
c12 += (ushort)a0.s5 * b0.s2;
c13 += (ushort)a0.s5 * b0.s3;
c20 += (ushort)a0.s6 * b0.s0;
c21 += (ushort)a0.s6 * b0.s1;
c22 += (ushort)a0.s6 * b0.s2;
c23 += (ushort)a0.s6 * b0.s3;
c30 += (ushort)a0.s7 * b0.s0;
c31 += (ushort)a0.s7 * b0.s1;
c32 += (ushort)a0.s7 * b0.s2;
c33 += (ushort)a0.s7 * b0.s3;
// Load values from matrix B (transposed)
b0 = vload4(0, src_addr_b + 24 * TRANSPOSE1XW_WIDTH_STEP);
c00 += (ushort)a0.s8 * b0.s0;
c01 += (ushort)a0.s8 * b0.s1;
c02 += (ushort)a0.s8 * b0.s2;
c03 += (ushort)a0.s8 * b0.s3;
c10 += (ushort)a0.s9 * b0.s0;
c11 += (ushort)a0.s9 * b0.s1;
c12 += (ushort)a0.s9 * b0.s2;
c13 += (ushort)a0.s9 * b0.s3;
c20 += (ushort)a0.sA * b0.s0;
c21 += (ushort)a0.sA * b0.s1;
c22 += (ushort)a0.sA * b0.s2;
c23 += (ushort)a0.sA * b0.s3;
c30 += (ushort)a0.sB * b0.s0;
c31 += (ushort)a0.sB * b0.s1;
c32 += (ushort)a0.sB * b0.s2;
c33 += (ushort)a0.sB * b0.s3;
// Load values from matrix B (transposed)
b0 = vload4(0, src_addr_b + 28 * TRANSPOSE1XW_WIDTH_STEP);
c00 += (ushort)a0.sC * b0.s0;
c01 += (ushort)a0.sC * b0.s1;
c02 += (ushort)a0.sC * b0.s2;
c03 += (ushort)a0.sC * b0.s3;
c10 += (ushort)a0.sD * b0.s0;
c11 += (ushort)a0.sD * b0.s1;
c12 += (ushort)a0.sD * b0.s2;
c13 += (ushort)a0.sD * b0.s3;
c20 += (ushort)a0.sE * b0.s0;
c21 += (ushort)a0.sE * b0.s1;
c22 += (ushort)a0.sE * b0.s2;
c23 += (ushort)a0.sE * b0.s3;
c30 += (ushort)a0.sF * b0.s0;
c31 += (ushort)a0.sF * b0.s1;
c32 += (ushort)a0.sF * b0.s2;
c33 += (ushort)a0.sF * b0.s3;
}
#endif // MULT_INTERLEAVE4X4_HEIGHT == 1
for(; src_addr_b < src_end_addr_b; src_addr_a += (4 * MULT_INTERLEAVE4X4_HEIGHT), src_addr_b += (4 * TRANSPOSE1XW_WIDTH_STEP))
{
// Load values from matrix A (interleaved) and matrix B (transposed)
uchar4 a0 = vload4(0, src_addr_a);
uchar4 b0 = vload4(0, src_addr_b);
c00 += (ushort)a0.s0 * b0.s0;
c01 += (ushort)a0.s0 * b0.s1;
c02 += (ushort)a0.s0 * b0.s2;
c03 += (ushort)a0.s0 * b0.s3;
c10 += (ushort)a0.s1 * b0.s0;
c11 += (ushort)a0.s1 * b0.s1;
c12 += (ushort)a0.s1 * b0.s2;
c13 += (ushort)a0.s1 * b0.s3;
c20 += (ushort)a0.s2 * b0.s0;
c21 += (ushort)a0.s2 * b0.s1;
c22 += (ushort)a0.s2 * b0.s2;
c23 += (ushort)a0.s2 * b0.s3;
c30 += (ushort)a0.s3 * b0.s0;
c31 += (ushort)a0.s3 * b0.s1;
c32 += (ushort)a0.s3 * b0.s2;
c33 += (ushort)a0.s3 * b0.s3;
}
// Compute destination address
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
#if defined(REINTERPRET_OUTPUT_AS_3D)
// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
// in order to take into account the presence of possible cross plane paddings
//
// | |
// | plane0 |
// | |
// |__________________|
// |******************|
// | cross_plane_pad |
// |******************|
// | |
// | plane1 |
// | |
// |__________________|
// The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D
uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;
zout = min(DEPTH_GEMM3D - 1, zout);
// Add offset due to the cross plane paddings
zout *= (cross_plane_pad * dst_stride_y);
// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
// multiply dst_stride_z by DEPTH_GEMM3D
dst.ptr += z * dst_stride_z * DEPTH_GEMM3D;
// Store 4x4 block
vstore4((int4)(c00, c01, c02, c03), 0, (__global int *)(dst.ptr + 0 * dst_stride_y + zout.s0));
vstore4((int4)(c10, c11, c12, c13), 0, (__global int *)(dst.ptr + 1 * dst_stride_y + zout.s1));
vstore4((int4)(c20, c21, c22, c23), 0, (__global int *)(dst.ptr + 2 * dst_stride_y + zout.s2));
vstore4((int4)(c30, c31, c32, c33), 0, (__global int *)(dst.ptr + 3 * dst_stride_y + zout.s3));
#else // defined(REINTERPRET_OUTPUT_AS_3D)
// Add offset for batched GEMM
dst.ptr += z * dst_stride_z;
// Store 4x4 block
vstore4((int4)(c00, c01, c02, c03), 0, (__global int *)(dst.ptr + 0 * dst_stride_y));
vstore4((int4)(c10, c11, c12, c13), 0, (__global int *)(dst.ptr + 1 * dst_stride_y));
vstore4((int4)(c20, c21, c22, c23), 0, (__global int *)(dst.ptr + 2 * dst_stride_y));
vstore4((int4)(c30, c31, c32, c33), 0, (__global int *)(dst.ptr + 3 * dst_stride_y));
#endif // defined(REINTERPRET_OUTPUT_AS_3D)
}
#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
/** This OpenCL kernel is optimized for Bifrost and computes the matrix multiplication between matrix A (src0) and matrix B (src1)
* Matrix A and matrix B must be reshaped respectively with @ref CLGEMMInterleave4x4Kernel and @ref CLGEMMTranspose1xWKernel before running the matrix multiplication
*
* @attention The number of matrix B columns needs to be passed at compile time using -DCOLS_B
* @note The transposition width step (mult_transpose1xW_width * 4) must be passed at compile time using -DTRANSPOSE1XW_WIDTH_STEP (i.e. -DTRANSPOSE1XW_WIDTH_STEP=2)
* @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2)
*
* @note In case the output has to be reinterpreted as a 3D tensor (i.e. output of convolution layer), the following information must be passed at compile time:
* -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
* -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
* -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
* (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
*
* @param[in] src0_ptr Pointer to the source matrix. Supported data type: QASYMM8
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported data type: same as @p src0_ptr
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data type: S32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_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 matrix
* @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
*/
__kernel void gemmlowp_mm_interleaved_transposed_bifrost_dot8(IMAGE_DECLARATION(src0),
IMAGE_DECLARATION(src1),
IMAGE_DECLARATION(dst),
uint src0_stride_z,
uint src1_stride_z,
uint dst_stride_z
#if defined(REINTERPRET_OUTPUT_AS_3D)
,
uint cross_plane_pad
#endif // REINTERPRET_OUTPUT_AS_3D
)
{
// Offset
const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4;
const int offset_row_b = (get_global_id(0) % TRANSPOSE1XW_WIDTH_STEP) * 4;
// src_addr_a = address of matrix A
// src_addr_b = address of matrix B
__global uchar *src_addr_a = (__global uchar *)(src0_ptr + (get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT) * src0_stride_y + get_global_id(2) * src0_stride_z + src0_offset_first_element_in_bytes);
__global uchar *src_addr_b = (__global uchar *)(src1_ptr + (get_global_id(0) / TRANSPOSE1XW_WIDTH_STEP) * src1_stride_y + src1_offset_first_element_in_bytes);
#if defined(MATRIX_B_DEPTH)
// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
src_addr_b += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
#else // defined(MATRIX_B_DEPTH)
src_addr_b += get_global_id(2) * src1_stride_z;
#endif // defined(MATRIX_B_DEPTH)
src_addr_a += offset_row_a;
src_addr_b += offset_row_b;
// Reset accumulators
uint c00 = 0;
uint c01 = 0;
uint c02 = 0;
uint c03 = 0;
uint c10 = 0;
uint c11 = 0;
uint c12 = 0;
uint c13 = 0;
uint c20 = 0;
uint c21 = 0;
uint c22 = 0;
uint c23 = 0;
uint c30 = 0;
uint c31 = 0;
uint c32 = 0;
uint c33 = 0;
#define COLS_MTX_B (COLS_B / (16 * MULT_TRANSPOSE1XW_WIDTH))
#if MULT_INTERLEAVE4X4_HEIGHT == 1
int i = 0;
for(; i <= (int)(COLS_MTX_B - 8); i += 8)
{
// Load values from matrix A (interleaved) and matrix B (transposed)
uchar16 a0 = vload16(0, src_addr_a);
uchar4 b0 = vload4(0, src_addr_b);
uchar4 b1 = vload4(0, src_addr_b + 4 * TRANSPOSE1XW_WIDTH_STEP);
uchar4 b2 = vload4(0, src_addr_b + 8 * TRANSPOSE1XW_WIDTH_STEP);
uchar4 b3 = vload4(0, src_addr_b + 12 * TRANSPOSE1XW_WIDTH_STEP);
uchar4 b4 = vload4(0, src_addr_b + 16 * TRANSPOSE1XW_WIDTH_STEP);
uchar4 b5 = vload4(0, src_addr_b + 20 * TRANSPOSE1XW_WIDTH_STEP);
uchar4 b6 = vload4(0, src_addr_b + 24 * TRANSPOSE1XW_WIDTH_STEP);
uchar4 b7 = vload4(0, src_addr_b + 28 * TRANSPOSE1XW_WIDTH_STEP);
// Accumulate
ARM_DOT((uchar4)(a0.s0123), (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), c00);
ARM_DOT((uchar4)(a0.s0123), (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), c01);
ARM_DOT((uchar4)(a0.s0123), (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), c02);
ARM_DOT((uchar4)(a0.s0123), (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), c03);
ARM_DOT((uchar4)(a0.s4567), (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), c10);
ARM_DOT((uchar4)(a0.s4567), (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), c11);
ARM_DOT((uchar4)(a0.s4567), (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), c12);
ARM_DOT((uchar4)(a0.s4567), (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), c13);
ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), c20);
ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), c21);
ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), c22);
ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), c23);
ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), c30);
ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), c31);
ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), c32);
ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), c33);
// Accumulate
a0 = vload16(0, src_addr_a + 16);
ARM_DOT((uchar4)(a0.s0123), (uchar4)(b4.s0, b5.s0, b6.s0, b7.s0), c00);
ARM_DOT((uchar4)(a0.s0123), (uchar4)(b4.s1, b5.s1, b6.s1, b7.s1), c01);
ARM_DOT((uchar4)(a0.s0123), (uchar4)(b4.s2, b5.s2, b6.s2, b7.s2), c02);
ARM_DOT((uchar4)(a0.s0123), (uchar4)(b4.s3, b5.s3, b6.s3, b7.s3), c03);
ARM_DOT((uchar4)(a0.s4567), (uchar4)(b4.s0, b5.s0, b6.s0, b7.s0), c10);
ARM_DOT((uchar4)(a0.s4567), (uchar4)(b4.s1, b5.s1, b6.s1, b7.s1), c11);
ARM_DOT((uchar4)(a0.s4567), (uchar4)(b4.s2, b5.s2, b6.s2, b7.s2), c12);
ARM_DOT((uchar4)(a0.s4567), (uchar4)(b4.s3, b5.s3, b6.s3, b7.s3), c13);
ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b4.s0, b5.s0, b6.s0, b7.s0), c20);
ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b4.s1, b5.s1, b6.s1, b7.s1), c21);
ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b4.s2, b5.s2, b6.s2, b7.s2), c22);
ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b4.s3, b5.s3, b6.s3, b7.s3), c23);
ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b4.s0, b5.s0, b6.s0, b7.s0), c30);
ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b4.s1, b5.s1, b6.s1, b7.s1), c31);
ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b4.s2, b5.s2, b6.s2, b7.s2), c32);
ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b4.s3, b5.s3, b6.s3, b7.s3), c33);
src_addr_a += 32;
src_addr_b += 32 * TRANSPOSE1XW_WIDTH_STEP;
}
#endif // MULT_INTERLEAVE4X4_HEIGHT == 1
int i_left_over = 0;
for(; i < (int)(COLS_MTX_B); ++i)
{
// Load values from matrix A (interleaved) and matrix B (transposed)
uchar16 a0 = vload16(0, src_addr_a + (i_left_over % 4) + ((i_left_over / 4) * 16));
uchar4 b0 = vload4(0, src_addr_b);
c00 += a0.s0 * b0.s0;
c01 += a0.s0 * b0.s1;
c02 += a0.s0 * b0.s2;
c03 += a0.s0 * b0.s3;
c10 += a0.s4 * b0.s0;
c11 += a0.s4 * b0.s1;
c12 += a0.s4 * b0.s2;
c13 += a0.s4 * b0.s3;
c20 += a0.s8 * b0.s0;
c21 += a0.s8 * b0.s1;
c22 += a0.s8 * b0.s2;
c23 += a0.s8 * b0.s3;
c30 += a0.sC * b0.s0;
c31 += a0.sC * b0.s1;
c32 += a0.sC * b0.s2;
c33 += a0.sC * b0.s3;
i_left_over++;
src_addr_b += 4 * TRANSPOSE1XW_WIDTH_STEP;
}
// Compute destination address
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
#if defined(REINTERPRET_OUTPUT_AS_3D)
// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
// in order to take into account the presence of possible cross plane paddings
//
// | |
// | plane0 |
// | |
// |__________________|
// |******************|
// | cross_plane_pad |
// |******************|
// | |
// | plane1 |
// | |
// |__________________|
// The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D
uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;
zout = min(DEPTH_GEMM3D - 1, zout);
// Add offset due to the cross plane paddings
zout *= (cross_plane_pad * dst_stride_y);
// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
// multiply dst_stride_z by DEPTH_GEMM3D
dst.ptr += get_global_id(2) * dst_stride_z * DEPTH_GEMM3D;
// Store 4x4 block
vstore4((int4)(c00, c01, c02, c03), 0, (__global int *)(dst.ptr + 0 * dst_stride_y + zout.s0));
vstore4((int4)(c10, c11, c12, c13), 0, (__global int *)(dst.ptr + 1 * dst_stride_y + zout.s1));
vstore4((int4)(c20, c21, c22, c23), 0, (__global int *)(dst.ptr + 2 * dst_stride_y + zout.s2));
vstore4((int4)(c30, c31, c32, c33), 0, (__global int *)(dst.ptr + 3 * dst_stride_y + zout.s3));
#else // defined(REINTERPRET_OUTPUT_AS_3D)
// Add offset for batched GEMM
dst.ptr += get_global_id(2) * dst_stride_z;
// Store 4x4 block
vstore4((int4)(c00, c01, c02, c03), 0, (__global int *)(dst.ptr + 0 * dst_stride_y));
vstore4((int4)(c10, c11, c12, c13), 0, (__global int *)(dst.ptr + 1 * dst_stride_y));
vstore4((int4)(c20, c21, c22, c23), 0, (__global int *)(dst.ptr + 2 * dst_stride_y));
vstore4((int4)(c30, c31, c32, c33), 0, (__global int *)(dst.ptr + 3 * dst_stride_y));
#endif // defined(REINTERPRET_OUTPUT_AS_3D)
}
#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#endif // defined(COLS_B) && defined(MULT_INTERLEAVE4X4_HEIGHT) && defined(TRANSPOSE1XW_WIDTH_STEP)
#if defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_Y) && defined(COLS_A)
#define VECTOR_UCHAR VEC_DATA_TYPE(uchar, NUM_ELEMS_PROCESSED_PER_THREAD_X)
#define VECTOR_UINT VEC_DATA_TYPE(uint, NUM_ELEMS_PROCESSED_PER_THREAD_X)
#define VECTOR_INT VEC_DATA_TYPE(int, NUM_ELEMS_PROCESSED_PER_THREAD_X)
/** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped
*
* @attention The number of matrix A columns needs to be passed at compile time using -DCOLS_A
*
* @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
* -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
* -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
* -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
* -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
* (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
*
* @param[in] src0_ptr Pointer to the source matrix. Supported data type: QASYMM8
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported data type: same as @p src0_ptr
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data type: S32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_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 matrix
* @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D)
* @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements for the output tensor (only if defined REINTERPRET_OUTPUT_AS_3D)
*/
__kernel void gemmlowp_mm_midgard(IMAGE_DECLARATION(src0),
IMAGE_DECLARATION(src1),
IMAGE_DECLARATION(dst),
uint src0_stride_z,
uint src1_stride_z,
uint dst_stride_z
#if defined(REINTERPRET_INPUT_AS_3D)
,
uint src_cross_plane_pad
#endif // REINTERPRET_INPUT_AS_3D
#if defined(REINTERPRET_OUTPUT_AS_3D)
,
uint dst_cross_plane_pad
#endif // REINTERPRET_OUTPUT_AS_3D
)
{
int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;
// Compute starting address for matrix A and Matrix B
int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
// Update address for the matrix A
src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;
// Update address for the matrix B
src_addr.s1 += idx;
#if defined(REINTERPRET_INPUT_AS_3D)
// Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension
// in order to take into account the presence of possible cross plane paddings
//
// | |
// | plane0 |
// | |
// |__________________|
// |******************|
// | cross_plane_pad |
// |******************|
// | |
// | plane1 |
// | |
// |__________________|
// The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D
uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;
zin = min(DEPTH_GEMM3D - 1, zin);
// Add offset due to the cross plane paddings
zin *= (src_cross_plane_pad * src0_stride_y);
// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
// multiply src0_stride_z by DEPTH_GEMM3D
src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;
#else // defined(REINTERPRET_INPUT_AS_3D)
// Add offset for batched GEMM
src_addr.s0 += get_global_id(2) * src0_stride_z;
#endif // defined(REINTERPRET_INPUT_AS_3D)
#if defined(MATRIX_B_DEPTH)
// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
#else // defined(MATRIX_B_DEPTH)
src_addr.s1 += get_global_id(2) * src1_stride_z;
#endif // defined(MATRIX_B_DEPTH)
int end_row_vec_a = src_addr.s0 + COLS_A;
VECTOR_UINT acc0 = 0;
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
VECTOR_UINT acc1 = 0;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
VECTOR_UINT acc2 = 0;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
VECTOR_UINT acc3 = 0;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
VECTOR_UINT acc4 = 0;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
for(; src_addr.s0 <= (end_row_vec_a - 2); src_addr += (int2)(2, 2 * src1_stride_y))
{
// Load values from matrix A
uchar2 a0 = vload2(0, src0_ptr + src_addr.s0 + 0 * src0_stride_y);
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
uchar2 a1 = vload2(0, src0_ptr + src_addr.s0 + 1 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
uchar2 a2 = vload2(0, src0_ptr + src_addr.s0 + 2 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
uchar2 a3 = vload2(0, src0_ptr + src_addr.s0 + 3 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
uchar2 a4 = vload2(0, src0_ptr + src_addr.s0 + 4 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
// Load values from matrix B
VECTOR_UCHAR b0 = VLOAD(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, src1_ptr + src_addr.s1);
VECTOR_UCHAR b1 = VLOAD(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, src1_ptr + src_addr.s1 + src1_stride_y);
// Accumulate
acc0 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a0.s0;
acc0 += CONVERT(b1, VECTOR_UINT) * (VECTOR_UINT)a0.s1;
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
acc1 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a1.s0;
acc1 += CONVERT(b1, VECTOR_UINT) * (VECTOR_UINT)a1.s1;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
acc2 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a2.s0;
acc2 += CONVERT(b1, VECTOR_UINT) * (VECTOR_UINT)a2.s1;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
acc3 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a3.s0;
acc3 += CONVERT(b1, VECTOR_UINT) * (VECTOR_UINT)a3.s1;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
acc4 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a4.s0;
acc4 += CONVERT(b1, VECTOR_UINT) * (VECTOR_UINT)a4.s1;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
}
for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(1, src1_stride_y))
{
// Load values from matrix A
uchar a0 = *(src0_ptr + src_addr.s0 + 0 * src0_stride_y);
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
uchar a1 = *(src0_ptr + src_addr.s0 + 1 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
uchar a2 = *(src0_ptr + src_addr.s0 + 2 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
uchar a3 = *(src0_ptr + src_addr.s0 + 3 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
uchar a4 = *(src0_ptr + src_addr.s0 + 4 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
// Load values from matrix B
VECTOR_UCHAR b0 = VLOAD(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, src1_ptr + src_addr.s1);
// Accumulate
acc0 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a0;
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
acc1 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a1;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
acc2 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a2;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
acc3 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a3;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
acc4 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a4;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
}
const int z = get_global_id(2);
// Compute destination address
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
#if defined(REINTERPRET_OUTPUT_AS_3D)
// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
// in order to take into account the presence of possible cross plane paddings
//
// | |
// | plane0 |
// | |
// |__________________|
// |******************|
// | cross_plane_pad |
// |******************|
// | |
// | plane1 |
// | |
// |__________________|
// The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D
uint8 zout = ((uint8)(0, 1, 2, 3, 4, 5, 6, 7) + (uint8)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint8)HEIGHT_GEMM3D;
zout = min(DEPTH_GEMM3D - 1, zout);
// Add offset due to the cross plane paddings
zout *= (dst_cross_plane_pad * dst_stride_y);
// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
// multiply dst_stride_z by DEPTH_GEMM3D
dst.ptr += z * dst_stride_z * DEPTH_GEMM3D;
// Store the result
VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X)
(CONVERT(acc0, VECTOR_INT), 0, (__global int *)(dst.ptr + 0 * dst_stride_y + zout.s0));
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X)
(CONVERT(acc1, VECTOR_INT), 0, (__global int *)(dst.ptr + 1 * dst_stride_y + zout.s1));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X)
(CONVERT(acc2, VECTOR_INT), 0, (__global int *)(dst.ptr + 2 * dst_stride_y + zout.s2));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X)
(CONVERT(acc3, VECTOR_INT), 0, (__global int *)(dst.ptr + 3 * dst_stride_y + zout.s3));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X)
(CONVERT(acc4, VECTOR_INT), 0, (__global int *)(dst.ptr + 4 * dst_stride_y + zout.s4));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
#else // defined(REINTERPRET_OUTPUT_AS_3D)
// Add offset for batched GEMM
dst.ptr += z * dst_stride_z;
// Store the result
VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X)
(CONVERT(acc0, VECTOR_INT), 0, (__global int *)(dst.ptr + 0 * dst_stride_y));
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X)
(CONVERT(acc1, VECTOR_INT), 0, (__global int *)(dst.ptr + 1 * dst_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X)
(CONVERT(acc2, VECTOR_INT), 0, (__global int *)(dst.ptr + 2 * dst_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X)
(CONVERT(acc3, VECTOR_INT), 0, (__global int *)(dst.ptr + 3 * dst_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X)
(CONVERT(acc4, VECTOR_INT), 0, (__global int *)(dst.ptr + 4 * dst_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
#endif // defined(REINTERPRET_OUTPUT_AS_3D)
}
/** OpenCL kernel optimized for Bifrost architectures that computes the matrix multiplication between matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped
*
* @attention The number of matrix A columns needs to be passed at compile time using -DCOLS_A
*
* @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
* -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
* -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
* -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
* -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
* (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
*
* @param[in] src0_ptr Pointer to the source matrix. Supported data type: QASYMM8
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported data type: same as @p src0_ptr
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data type: S32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_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 matrix
* @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D)
* @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements for the output tensor (only if defined REINTERPRET_OUTPUT_AS_3D)
*/
__kernel void gemmlowp_mm_bifrost(IMAGE_DECLARATION(src0),
IMAGE_DECLARATION(src1),
IMAGE_DECLARATION(dst),
uint src0_stride_z,
uint src1_stride_z,
uint dst_stride_z
#if defined(REINTERPRET_INPUT_AS_3D)
,
uint src_cross_plane_pad
#endif // REINTERPRET_INPUT_AS_3D
#if defined(REINTERPRET_OUTPUT_AS_3D)
,
uint dst_cross_plane_pad
#endif // REINTERPRET_OUTPUT_AS_3D
)
{
int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;
// Compute starting address for matrix A and Matrix B
int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
// Update address for the matrix A
src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;
// Update address for the matrix B
src_addr.s1 += idx;
#if defined(REINTERPRET_INPUT_AS_3D)
// Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension
// in order to take into account the presence of possible cross plane paddings
//
// | |
// | plane0 |
// | |
// |__________________|
// |******************|
// | cross_plane_pad |
// |******************|
// | |
// | plane1 |
// | |
// |__________________|
// The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D
uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;
zin = min(DEPTH_GEMM3D - 1, zin);
// Add offset due to the cross plane paddings
zin *= (src_cross_plane_pad * src0_stride_y);
// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
// multiply src0_stride_z by DEPTH_GEMM3D
src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;
#else // defined(REINTERPRET_INPUT_AS_3D)
// Add offset for batched GEMM
src_addr.s0 += get_global_id(2) * src0_stride_z;
#endif // defined(REINTERPRET_INPUT_AS_3D)
#if defined(MATRIX_B_DEPTH)
// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
#else // defined(MATRIX_B_DEPTH)
src_addr.s1 += get_global_id(2) * src1_stride_z;
#endif // defined(MATRIX_B_DEPTH)
int end_row_vec_a = src_addr.s0 + COLS_A;
uint acc00 = 0;
uint acc01 = 0;
uint acc02 = 0;
uint acc03 = 0;
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
uint acc10 = 0;
uint acc11 = 0;
uint acc12 = 0;
uint acc13 = 0;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
uint acc20 = 0;
uint acc21 = 0;
uint acc22 = 0;
uint acc23 = 0;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
uint acc30 = 0;
uint acc31 = 0;
uint acc32 = 0;
uint acc33 = 0;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
uint acc40 = 0;
uint acc41 = 0;
uint acc42 = 0;
uint acc43 = 0;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
for(; src_addr.s0 <= (end_row_vec_a - 4); src_addr += (int2)(4, 4 * src1_stride_y))
{
// Load values from matrix A
uchar4 a0 = vload4(0, src0_ptr + src_addr.s0 + 0 * src0_stride_y);
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
uchar4 a1 = vload4(0, src0_ptr + src_addr.s0 + 1 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
uchar4 a2 = vload4(0, src0_ptr + src_addr.s0 + 2 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
uchar4 a3 = vload4(0, src0_ptr + src_addr.s0 + 3 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
uchar4 a4 = vload4(0, src0_ptr + src_addr.s0 + 4 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
// Load values from matrix B
uchar4 b0 = vload4(0, src1_ptr + src_addr.s1 + 0 * src1_stride_y);
uchar4 b1 = vload4(0, src1_ptr + src_addr.s1 + 1 * src1_stride_y);
uchar4 b2 = vload4(0, src1_ptr + src_addr.s1 + 2 * src1_stride_y);
uchar4 b3 = vload4(0, src1_ptr + src_addr.s1 + 3 * src1_stride_y);
{
// Accumulate
ushort tmp0 = (ushort)b0.s0 * (ushort)a0.s0;
ushort tmp1 = (ushort)b0.s1 * (ushort)a0.s0;
ushort tmp2 = (ushort)b0.s2 * (ushort)a0.s0;
ushort tmp3 = (ushort)b0.s3 * (ushort)a0.s0;
ushort tmp4 = (ushort)b1.s0 * (ushort)a0.s1;
ushort tmp5 = (ushort)b1.s1 * (ushort)a0.s1;
ushort tmp6 = (ushort)b1.s2 * (ushort)a0.s1;
ushort tmp7 = (ushort)b1.s3 * (ushort)a0.s1;
ushort tmp8 = (ushort)b2.s0 * (ushort)a0.s2;
ushort tmp9 = (ushort)b2.s1 * (ushort)a0.s2;
ushort tmpA = (ushort)b2.s2 * (ushort)a0.s2;
ushort tmpB = (ushort)b2.s3 * (ushort)a0.s2;
ushort tmpC = (ushort)b3.s0 * (ushort)a0.s3;
ushort tmpD = (ushort)b3.s1 * (ushort)a0.s3;
ushort tmpE = (ushort)b3.s2 * (ushort)a0.s3;
ushort tmpF = (ushort)b3.s3 * (ushort)a0.s3;
acc00 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC);
acc01 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD);
acc02 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE);
acc03 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF);
}
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
{
// Accumulate
ushort tmp0 = (ushort)b0.s0 * (ushort)a1.s0;
ushort tmp1 = (ushort)b0.s1 * (ushort)a1.s0;
ushort tmp2 = (ushort)b0.s2 * (ushort)a1.s0;
ushort tmp3 = (ushort)b0.s3 * (ushort)a1.s0;
ushort tmp4 = (ushort)b1.s0 * (ushort)a1.s1;
ushort tmp5 = (ushort)b1.s1 * (ushort)a1.s1;
ushort tmp6 = (ushort)b1.s2 * (ushort)a1.s1;
ushort tmp7 = (ushort)b1.s3 * (ushort)a1.s1;
ushort tmp8 = (ushort)b2.s0 * (ushort)a1.s2;
ushort tmp9 = (ushort)b2.s1 * (ushort)a1.s2;
ushort tmpA = (ushort)b2.s2 * (ushort)a1.s2;
ushort tmpB = (ushort)b2.s3 * (ushort)a1.s2;
ushort tmpC = (ushort)b3.s0 * (ushort)a1.s3;
ushort tmpD = (ushort)b3.s1 * (ushort)a1.s3;
ushort tmpE = (ushort)b3.s2 * (ushort)a1.s3;
ushort tmpF = (ushort)b3.s3 * (ushort)a1.s3;
acc10 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC);
acc11 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD);
acc12 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE);
acc13 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF);
}
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
{
// Accumulate
ushort tmp0 = (ushort)b0.s0 * (ushort)a2.s0;
ushort tmp1 = (ushort)b0.s1 * (ushort)a2.s0;
ushort tmp2 = (ushort)b0.s2 * (ushort)a2.s0;
ushort tmp3 = (ushort)b0.s3 * (ushort)a2.s0;
ushort tmp4 = (ushort)b1.s0 * (ushort)a2.s1;
ushort tmp5 = (ushort)b1.s1 * (ushort)a2.s1;
ushort tmp6 = (ushort)b1.s2 * (ushort)a2.s1;
ushort tmp7 = (ushort)b1.s3 * (ushort)a2.s1;
ushort tmp8 = (ushort)b2.s0 * (ushort)a2.s2;
ushort tmp9 = (ushort)b2.s1 * (ushort)a2.s2;
ushort tmpA = (ushort)b2.s2 * (ushort)a2.s2;
ushort tmpB = (ushort)b2.s3 * (ushort)a2.s2;
ushort tmpC = (ushort)b3.s0 * (ushort)a2.s3;
ushort tmpD = (ushort)b3.s1 * (ushort)a2.s3;
ushort tmpE = (ushort)b3.s2 * (ushort)a2.s3;
ushort tmpF = (ushort)b3.s3 * (ushort)a2.s3;
acc20 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC);
acc21 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD);
acc22 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE);
acc23 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF);
}
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
{
// Accumulate
ushort tmp0 = (ushort)b0.s0 * (ushort)a3.s0;
ushort tmp1 = (ushort)b0.s1 * (ushort)a3.s0;
ushort tmp2 = (ushort)b0.s2 * (ushort)a3.s0;
ushort tmp3 = (ushort)b0.s3 * (ushort)a3.s0;
ushort tmp4 = (ushort)b1.s0 * (ushort)a3.s1;
ushort tmp5 = (ushort)b1.s1 * (ushort)a3.s1;
ushort tmp6 = (ushort)b1.s2 * (ushort)a3.s1;
ushort tmp7 = (ushort)b1.s3 * (ushort)a3.s1;
ushort tmp8 = (ushort)b2.s0 * (ushort)a3.s2;
ushort tmp9 = (ushort)b2.s1 * (ushort)a3.s2;
ushort tmpA = (ushort)b2.s2 * (ushort)a3.s2;
ushort tmpB = (ushort)b2.s3 * (ushort)a3.s2;
ushort tmpC = (ushort)b3.s0 * (ushort)a3.s3;
ushort tmpD = (ushort)b3.s1 * (ushort)a3.s3;
ushort tmpE = (ushort)b3.s2 * (ushort)a3.s3;
ushort tmpF = (ushort)b3.s3 * (ushort)a3.s3;
acc30 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC);
acc31 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD);
acc32 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE);
acc33 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF);
}
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
{
// Accumulate
ushort tmp0 = (ushort)b0.s0 * (ushort)a4.s0;
ushort tmp1 = (ushort)b0.s1 * (ushort)a4.s0;
ushort tmp2 = (ushort)b0.s2 * (ushort)a4.s0;
ushort tmp3 = (ushort)b0.s3 * (ushort)a4.s0;
ushort tmp4 = (ushort)b1.s0 * (ushort)a4.s1;
ushort tmp5 = (ushort)b1.s1 * (ushort)a4.s1;
ushort tmp6 = (ushort)b1.s2 * (ushort)a4.s1;
ushort tmp7 = (ushort)b1.s3 * (ushort)a4.s1;
ushort tmp8 = (ushort)b2.s0 * (ushort)a4.s2;
ushort tmp9 = (ushort)b2.s1 * (ushort)a4.s2;
ushort tmpA = (ushort)b2.s2 * (ushort)a4.s2;
ushort tmpB = (ushort)b2.s3 * (ushort)a4.s2;
ushort tmpC = (ushort)b3.s0 * (ushort)a4.s3;
ushort tmpD = (ushort)b3.s1 * (ushort)a4.s3;
ushort tmpE = (ushort)b3.s2 * (ushort)a4.s3;
ushort tmpF = (ushort)b3.s3 * (ushort)a4.s3;
acc40 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC);
acc41 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD);
acc42 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE);
acc43 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF);
}
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
}
for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(1, src1_stride_y))
{
// Load values from matrix A
uchar a0 = *(src0_ptr + src_addr.s0 + 0 * src0_stride_y);
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
uchar a1 = *(src0_ptr + src_addr.s0 + 1 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
uchar a2 = *(src0_ptr + src_addr.s0 + 2 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
uchar a3 = *(src0_ptr + src_addr.s0 + 3 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
uchar a4 = *(src0_ptr + src_addr.s0 + 4 * src0_stride_y);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
// Load values from matrix B
uchar4 b0 = vload4(0, src1_ptr + src_addr.s1);
// Accumulate
{
// Accumulate
ushort tmp0 = (ushort)b0.s0 * (ushort)a0;
ushort tmp1 = (ushort)b0.s1 * (ushort)a0;
ushort tmp2 = (ushort)b0.s2 * (ushort)a0;
ushort tmp3 = (ushort)b0.s3 * (ushort)a0;
acc00 += ((uint)tmp0);
acc01 += ((uint)tmp1);
acc02 += ((uint)tmp2);
acc03 += ((uint)tmp3);
}
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
{
// Accumulate
ushort tmp0 = (ushort)b0.s0 * (ushort)a1;
ushort tmp1 = (ushort)b0.s1 * (ushort)a1;
ushort tmp2 = (ushort)b0.s2 * (ushort)a1;
ushort tmp3 = (ushort)b0.s3 * (ushort)a1;
acc10 += ((uint)tmp0);
acc11 += ((uint)tmp1);
acc12 += ((uint)tmp2);
acc13 += ((uint)tmp3);
}
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
{
// Accumulate
ushort tmp0 = (ushort)b0.s0 * (ushort)a2;
ushort tmp1 = (ushort)b0.s1 * (ushort)a2;
ushort tmp2 = (ushort)b0.s2 * (ushort)a2;
ushort tmp3 = (ushort)b0.s3 * (ushort)a2;
acc20 += ((uint)tmp0);
acc21 += ((uint)tmp1);
acc22 += ((uint)tmp2);
acc23 += ((uint)tmp3);
}
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
{
// Accumulate
ushort tmp0 = (ushort)b0.s0 * (ushort)a3;
ushort tmp1 = (ushort)b0.s1 * (ushort)a3;
ushort tmp2 = (ushort)b0.s2 * (ushort)a3;
ushort tmp3 = (ushort)b0.s3 * (ushort)a3;
acc30 += ((uint)tmp0);
acc31 += ((uint)tmp1);
acc32 += ((uint)tmp2);
acc33 += ((uint)tmp3);
}
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
{
// Accumulate
ushort tmp0 = (ushort)b0.s0 * (ushort)a4;
ushort tmp1 = (ushort)b0.s1 * (ushort)a4;
ushort tmp2 = (ushort)b0.s2 * (ushort)a4;
ushort tmp3 = (ushort)b0.s3 * (ushort)a4;
acc40 += ((uint)tmp0);
acc41 += ((uint)tmp1);
acc42 += ((uint)tmp2);
acc43 += ((uint)tmp3);
}
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
}
const int z = get_global_id(2);
// Compute destination address
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
#if defined(REINTERPRET_OUTPUT_AS_3D)
// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
// in order to take into account the presence of possible cross plane paddings
//
// | |
// | plane0 |
// | |
// |__________________|
// |******************|
// | cross_plane_pad |
// |******************|
// | |
// | plane1 |
// | |
// |__________________|
// The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D
uint8 zout = ((uint8)(0, 1, 2, 3, 4, 5, 6, 7) + (uint8)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint8)HEIGHT_GEMM3D;
zout = min(DEPTH_GEMM3D - 1, zout);
// Add offset due to the cross plane paddings
zout *= (dst_cross_plane_pad * dst_stride_y);
// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
// multiply dst_stride_z by DEPTH_GEMM3D
dst.ptr += z * dst_stride_z * DEPTH_GEMM3D;
// Store the result
vstore4((int4)(acc00, acc01, acc02, acc03), 0, (__global int *)(dst.ptr + 0 * dst_stride_y + zout.s0));
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
vstore4((int4)(acc10, acc11, acc12, acc13), 0, (__global int *)(dst.ptr + 1 * dst_stride_y + zout.s1));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
vstore4((int4)(acc20, acc21, acc22, acc23), 0, (__global int *)(dst.ptr + 2 * dst_stride_y + zout.s2));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
vstore4((int4)(acc30, acc31, acc32, acc33), 0, (__global int *)(dst.ptr + 3 * dst_stride_y + zout.s3));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
vstore4((int4)(acc40, acc41, acc42, acc43), 0, (__global int *)(dst.ptr + 4 * dst_stride_y + zout.s4));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
#else // defined(REINTERPRET_OUTPUT_AS_3D)
// Add offset for batched GEMM
dst.ptr += z * dst_stride_z;
// Store the result
vstore4((int4)(acc00, acc01, acc02, acc03), 0, (__global int *)(dst.ptr + 0 * dst_stride_y));
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
vstore4((int4)(acc10, acc11, acc12, acc13), 0, (__global int *)(dst.ptr + 1 * dst_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
vstore4((int4)(acc20, acc21, acc22, acc23), 0, (__global int *)(dst.ptr + 2 * dst_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
vstore4((int4)(acc30, acc31, acc32, acc33), 0, (__global int *)(dst.ptr + 3 * dst_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
vstore4((int4)(acc40, acc41, acc42, acc43), 0, (__global int *)(dst.ptr + 4 * dst_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4
#endif // defined(REINTERPRET_OUTPUT_AS_3D)
}
#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
/** OpenCL kernel optimized to use dot product that computes the matrix multiplication between matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped
*
* @attention The number of matrix A columns needs to be passed at compile time using -DCOLS_A
*
* @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
* -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
* -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
* -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
* -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
* (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
*
* @param[in] src0_ptr Pointer to the source matrix. Supported data type: QASYMM8
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported data type: same as @p src0_ptr
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data type: S32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_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 matrix
* @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D)
* @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements for the output tensor (only if defined REINTERPRET_OUTPUT_AS_3D)
*/
__kernel void gemmlowp_mm_bifrost_dot8(IMAGE_DECLARATION(src0),
IMAGE_DECLARATION(src1),
IMAGE_DECLARATION(dst),
uint src0_stride_z,
uint src1_stride_z,
uint dst_stride_z
#if defined(REINTERPRET_INPUT_AS_3D)
,
uint src_cross_plane_pad
#endif // REINTERPRET_INPUT_AS_3D
#if defined(REINTERPRET_OUTPUT_AS_3D)
,
uint dst_cross_plane_pad
#endif // REINTERPRET_OUTPUT_AS_3D)
)
{
int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;
// Compute starting address for matrix A and Matrix B
int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
// Update address for the matrix A
src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;
// Update address for the matrix B
src_addr.s1 += idx;
#if defined(REINTERPRET_INPUT_AS_3D)
// Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension
// in order to take into account the presence of possible cross plane paddings
//
// | |
// | plane0 |
// | |
// |__________________|
// |******************|
// | cross_plane_pad |
// |******************|
// | |
// | plane1 |
// | |
// |__________________|
// The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D
uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;
zin = min(DEPTH_GEMM3D - 1, zin);
// Add offset due to the cross plane paddings
zin *= (src_cross_plane_pad * src0_stride_y);
zin += ((uint4)(0, 1, 2, 3)) * src0_stride_y;
// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
// multiply src0_stride_z by DEPTH_GEMM3D
src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;
#else // defined(REINTERPRET_INPUT_AS_3D)
// Add offset for batched GEMM
src_addr.s0 += get_global_id(2) * src0_stride_z;
#endif // defined(REINTERPRET_INPUT_AS_3D)
#if defined(MATRIX_B_DEPTH)
// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
#else // defined(MATRIX_B_DEPTH)
src_addr.s1 += get_global_id(2) * src1_stride_z;
#endif // defined(MATRIX_B_DEPTH)
uint acc00 = 0;
uint acc01 = 0;
uint acc02 = 0;
uint acc03 = 0;
uint acc04 = 0;
uint acc05 = 0;
uint acc06 = 0;
uint acc07 = 0;
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
uint acc10 = 0;
uint acc11 = 0;
uint acc12 = 0;
uint acc13 = 0;
uint acc14 = 0;
uint acc15 = 0;
uint acc16 = 0;
uint acc17 = 0;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
uint acc20 = 0;
uint acc21 = 0;
uint acc22 = 0;
uint acc23 = 0;
uint acc24 = 0;
uint acc25 = 0;
uint acc26 = 0;
uint acc27 = 0;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
uint acc30 = 0;
uint acc31 = 0;
uint acc32 = 0;
uint acc33 = 0;
uint acc34 = 0;
uint acc35 = 0;
uint acc36 = 0;
uint acc37 = 0;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
// A and B src indices get incremented at the same time.
int i = 0;
for(; i <= ((int)COLS_A - 8); i += 8)
{
#if defined(REINTERPRET_INPUT_AS_3D)
// Load values from matrix A and matrix B
uchar8 a0 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + zin.s0));
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
uchar8 a1 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + zin.s1));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
uchar8 a2 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + zin.s2));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
uchar8 a3 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + zin.s3));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#else // defined(REINTERPRET_INPUT_AS_3D)
// Load values from matrix A and matrix B
uchar8 a0 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
uchar8 a1 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
uchar8 a2 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
uchar8 a3 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#endif // defined(REINTERPRET_INPUT_AS_3D)
uchar8 b0 = vload8(0, src1_ptr + src_addr.s1 + 0 * src1_stride_y);
uchar8 b1 = vload8(0, src1_ptr + src_addr.s1 + 1 * src1_stride_y);
uchar8 b2 = vload8(0, src1_ptr + src_addr.s1 + 2 * src1_stride_y);
uchar8 b3 = vload8(0, src1_ptr + src_addr.s1 + 3 * src1_stride_y);
src_addr.s1 += 4 * src1_stride_y;
ARM_DOT(a0.s0123, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc00);
ARM_DOT(a0.s0123, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc01);
ARM_DOT(a0.s0123, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc02);
ARM_DOT(a0.s0123, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc03);
ARM_DOT(a0.s0123, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc04);
ARM_DOT(a0.s0123, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc05);
ARM_DOT(a0.s0123, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc06);
ARM_DOT(a0.s0123, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc07);
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
ARM_DOT(a1.s0123, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc10);
ARM_DOT(a1.s0123, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc11);
ARM_DOT(a1.s0123, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc12);
ARM_DOT(a1.s0123, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc13);
ARM_DOT(a1.s0123, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc14);
ARM_DOT(a1.s0123, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc15);
ARM_DOT(a1.s0123, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc16);
ARM_DOT(a1.s0123, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc17);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
ARM_DOT(a2.s0123, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc20);
ARM_DOT(a2.s0123, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc21);
ARM_DOT(a2.s0123, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc22);
ARM_DOT(a2.s0123, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc23);
ARM_DOT(a2.s0123, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc24);
ARM_DOT(a2.s0123, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc25);
ARM_DOT(a2.s0123, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc26);
ARM_DOT(a2.s0123, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc27);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
ARM_DOT(a3.s0123, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc30);
ARM_DOT(a3.s0123, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc31);
ARM_DOT(a3.s0123, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc32);
ARM_DOT(a3.s0123, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc33);
ARM_DOT(a3.s0123, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc34);
ARM_DOT(a3.s0123, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc35);
ARM_DOT(a3.s0123, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc36);
ARM_DOT(a3.s0123, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc37);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
b0 = vload8(0, src1_ptr + src_addr.s1 + 0 * src1_stride_y);
b1 = vload8(0, src1_ptr + src_addr.s1 + 1 * src1_stride_y);
b2 = vload8(0, src1_ptr + src_addr.s1 + 2 * src1_stride_y);
b3 = vload8(0, src1_ptr + src_addr.s1 + 3 * src1_stride_y);
src_addr.s1 += 4 * src1_stride_y;
ARM_DOT(a0.s4567, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc00);
ARM_DOT(a0.s4567, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc01);
ARM_DOT(a0.s4567, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc02);
ARM_DOT(a0.s4567, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc03);
ARM_DOT(a0.s4567, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc04);
ARM_DOT(a0.s4567, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc05);
ARM_DOT(a0.s4567, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc06);
ARM_DOT(a0.s4567, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc07);
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
ARM_DOT(a1.s4567, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc10);
ARM_DOT(a1.s4567, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc11);
ARM_DOT(a1.s4567, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc12);
ARM_DOT(a1.s4567, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc13);
ARM_DOT(a1.s4567, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc14);
ARM_DOT(a1.s4567, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc15);
ARM_DOT(a1.s4567, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc16);
ARM_DOT(a1.s4567, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc17);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
ARM_DOT(a2.s4567, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc20);
ARM_DOT(a2.s4567, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc21);
ARM_DOT(a2.s4567, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc22);
ARM_DOT(a2.s4567, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc23);
ARM_DOT(a2.s4567, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc24);
ARM_DOT(a2.s4567, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc25);
ARM_DOT(a2.s4567, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc26);
ARM_DOT(a2.s4567, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc27);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
ARM_DOT(a3.s4567, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc30);
ARM_DOT(a3.s4567, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc31);
ARM_DOT(a3.s4567, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc32);
ARM_DOT(a3.s4567, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc33);
ARM_DOT(a3.s4567, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc34);
ARM_DOT(a3.s4567, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc35);
ARM_DOT(a3.s4567, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc36);
ARM_DOT(a3.s4567, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc37);
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
src_addr.s0 += 8;
}
for(; i < (int)COLS_A; ++i)
{
#if defined(REINTERPRET_INPUT_AS_3D)
// Load values from matrix A
uchar a0 = *((__global uchar *)(src0_ptr + src_addr.s0 + zin.s0));
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
uchar a1 = *((__global uchar *)(src0_ptr + src_addr.s0 + zin.s1));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
uchar a2 = *((__global uchar *)(src0_ptr + src_addr.s0 + zin.s2));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
uchar a3 = *((__global uchar *)(src0_ptr + src_addr.s0 + zin.s3));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#else // defined(REINTERPRET_INPUT_AS_3D)
// Load values from matrix A
uchar a0 = *((__global uchar *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
uchar a1 = *((__global uchar *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
uchar a2 = *((__global uchar *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
uchar a3 = *((__global uchar *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#endif // defined(REINTERPRET_INPUT_AS_3D)
// Load values from matrix B
uchar8 b0 = vload8(0, src1_ptr + src_addr.s1);
src_addr.s1 += src1_stride_y;
acc00 += (uint)a0 * b0.s0;
acc01 += (uint)a0 * b0.s1;
acc02 += (uint)a0 * b0.s2;
acc03 += (uint)a0 * b0.s3;
acc04 += (uint)a0 * b0.s4;
acc05 += (uint)a0 * b0.s5;
acc06 += (uint)a0 * b0.s6;
acc07 += (uint)a0 * b0.s7;
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
acc10 += (uint)a1 * b0.s0;
acc11 += (uint)a1 * b0.s1;
acc12 += (uint)a1 * b0.s2;
acc13 += (uint)a1 * b0.s3;
acc14 += (uint)a1 * b0.s4;
acc15 += (uint)a1 * b0.s5;
acc16 += (uint)a1 * b0.s6;
acc17 += (uint)a1 * b0.s7;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
acc20 += (uint)a2 * b0.s0;
acc21 += (uint)a2 * b0.s1;
acc22 += (uint)a2 * b0.s2;
acc23 += (uint)a2 * b0.s3;
acc24 += (uint)a2 * b0.s4;
acc25 += (uint)a2 * b0.s5;
acc26 += (uint)a2 * b0.s6;
acc27 += (uint)a2 * b0.s7;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
acc30 += (uint)a3 * b0.s0;
acc31 += (uint)a3 * b0.s1;
acc32 += (uint)a3 * b0.s2;
acc33 += (uint)a3 * b0.s3;
acc34 += (uint)a3 * b0.s4;
acc35 += (uint)a3 * b0.s5;
acc36 += (uint)a3 * b0.s6;
acc37 += (uint)a3 * b0.s7;
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
src_addr.s0 += 1;
}
int z = get_global_id(2);
// Compute destination address
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
// Compute dst address
__global uchar *dst_addr = dst.ptr;
#if defined(REINTERPRET_OUTPUT_AS_3D)
// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
// in order to take into account the presence of possible cross plane paddings
//
// | |
// | plane0 |
// | |
// |__________________|
// |******************|
// | cross_plane_pad |
// |******************|
// | |
// | plane1 |
// | |
// |__________________|
// The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D
uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;
zout = min(DEPTH_GEMM3D - 1, zout);
// Add offset due to the cross plane paddings
zout *= (dst_cross_plane_pad * dst_stride_y);
// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
// multiply dst_stride_z by DEPTH_GEMM3D
dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
// Store the result
vstore4((int4)(acc00, acc01, acc02, acc03), 0, (__global int *)(dst_addr + 0 * dst_stride_y + zout.s0));
vstore4((int4)(acc04, acc05, acc06, acc07), 1, (__global int *)(dst_addr + 0 * dst_stride_y + zout.s0));
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
vstore4((int4)(acc10, acc11, acc12, acc13), 0, (__global int *)(dst_addr + 1 * dst_stride_y + zout.s1));
vstore4((int4)(acc14, acc15, acc16, acc17), 1, (__global int *)(dst_addr + 1 * dst_stride_y + zout.s1));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
vstore4((int4)(acc20, acc21, acc22, acc23), 0, (__global int *)(dst_addr + 2 * dst_stride_y + zout.s2));
vstore4((int4)(acc24, acc25, acc26, acc27), 1, (__global int *)(dst_addr + 2 * dst_stride_y + zout.s2));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
vstore4((int4)(acc30, acc31, acc32, acc33), 0, (__global int *)(dst_addr + 3 * dst_stride_y + zout.s3));
vstore4((int4)(acc34, acc35, acc36, acc37), 0, (__global int *)(dst_addr + 3 * dst_stride_y + zout.s3));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#else // defined(REINTERPRET_OUTPUT_AS_3D)
// Add offset for batched GEMM
dst_addr += z * dst_stride_z;
// Store the result
vstore4((int4)(acc00, acc01, acc02, acc03), 0, (__global int *)(dst_addr + 0 * dst_stride_y));
vstore4((int4)(acc04, acc05, acc06, acc07), 1, (__global int *)(dst_addr + 0 * dst_stride_y));
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
vstore4((int4)(acc10, acc11, acc12, acc13), 0, (__global int *)(dst_addr + 1 * dst_stride_y));
vstore4((int4)(acc14, acc15, acc16, acc17), 1, (__global int *)(dst_addr + 1 * dst_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
vstore4((int4)(acc20, acc21, acc22, acc23), 0, (__global int *)(dst_addr + 2 * dst_stride_y));
vstore4((int4)(acc24, acc25, acc26, acc27), 1, (__global int *)(dst_addr + 2 * dst_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
vstore4((int4)(acc30, acc31, acc32, acc33), 0, (__global int *)(dst_addr + 3 * dst_stride_y));
vstore4((int4)(acc34, acc35, acc36, acc37), 0, (__global int *)(dst_addr + 3 * dst_stride_y));
#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
#endif // defined(REINTERPRET_OUTPUT_AS_3D)
}
#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#endif // defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_Y) && defined(COLS_A)
#if defined(COLS_A)
/** OpenCL kernel used to compute the row-vectors of sums of all the entries in each row of Matrix A.
*
* @note This stage is needed to handle the offset of matrix product
* https://github.com/google/gemmlowp/blob/master/doc/low-precision.md
*
* @attention The number of matrix A columns needs to be passed at compile time using -DCOLS_A
*
* @param[in] src_ptr Pointer to the source tensor. Supported data type: QASYMM8
* @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 type: S32
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_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_gx_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 gemmlowp_matrix_a_reduction(TENSOR3D_DECLARATION(src),
IMAGE_DECLARATION(dst))
{
// Compute source and destination addresses
Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
uint4 sum_row_u32 = (uint4)0;
uint sum_row = 0;
__global const uchar *matrix_a = (__global const uchar *)(src.ptr + get_global_id(0) * src_stride_y + get_global_id(1) * src_stride_z);
int i = 0;
// This for loop performs 16 accumulations
for(; i <= ((int)COLS_A - 16); i += 16)
{
const uchar16 a0_u8 = vload16(0, matrix_a + i);
sum_row_u32 += convert_uint4(a0_u8.s0123) + convert_uint4(a0_u8.s4567) + convert_uint4(a0_u8.s89AB) + convert_uint4(a0_u8.sCDEF);
}
// This for loop performs the leftover accumulations
for(; i < COLS_A; ++i)
{
sum_row += matrix_a[i];
}
sum_row += sum_row_u32.s0 + sum_row_u32.s1 + sum_row_u32.s2 + sum_row_u32.s3;
*((__global int *)dst.ptr) = (int)sum_row;
}
#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
/** OpenCL kernel used to compute the row-vectors of sums of all the entries in each row of Matrix A using the arm dot product instruction
*
* @note This stage is needed to handle the offset of matrix product
* https://github.com/google/gemmlowp/blob/master/doc/low-precision.md
*
* @attention The number of matrix A columns needs to be passed at compile time using -DCOLS_A
*
* @param[in] src_ptr Pointer to the source tensor. Supported data type: QASYMM8
* @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 type: S32
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_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_gx_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 gemmlowp_matrix_a_reduction_dot8(TENSOR3D_DECLARATION(src),
IMAGE_DECLARATION(dst))
{
// Compute source and destination addresses
Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
uint sum_row = 0;
__global const uchar *matrix_a = (__global const uchar *)(src.ptr + get_global_id(0) * src_stride_y + get_global_id(1) * src_stride_z);
int i = 0;
// This for loop performs 16 accumulations
for(; i <= ((int)COLS_A - 32); i += 32)
{
uchar16 a0_u8 = vload16(0, matrix_a + i);
sum_row += arm_dot(a0_u8.s0123, (uchar4)(1));
sum_row += arm_dot(a0_u8.s4567, (uchar4)(1));
sum_row += arm_dot(a0_u8.s89AB, (uchar4)(1));
sum_row += arm_dot(a0_u8.sCDEF, (uchar4)(1));
a0_u8 = vload16(1, matrix_a + i);
sum_row += arm_dot(a0_u8.s0123, (uchar4)(1));
sum_row += arm_dot(a0_u8.s4567, (uchar4)(1));
sum_row += arm_dot(a0_u8.s89AB, (uchar4)(1));
sum_row += arm_dot(a0_u8.sCDEF, (uchar4)(1));
}
// This for loop performs the leftover accumulations
for(; i < COLS_A; ++i)
{
sum_row += matrix_a[i];
}
*((__global int *)dst.ptr) = (int)sum_row;
}
#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#endif // defined(COLS_A)
#if defined(COLS_B) && defined(ROWS_B)
/** OpenCL kernel used to compute the row-vectors of sums of all the entries in each column of Matrix B.
*
* @note This stage is needed to handle the offset of matrix product
* https://github.com/google/gemmlowp/blob/master/doc/low-precision.md
*
* @attention The number of matrix B columns and rows needs to be passed at compile time using -DCOLS_B and -DROWS_B
*
* @param[in] src_ptr Pointer to the source tensor. Supported data type: QASYMM8
* @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 type: S32
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_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_gx_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 gemmlowp_matrix_b_reduction(TENSOR3D_DECLARATION(src),
IMAGE_DECLARATION(dst))
{
// Compute source and destination addresses
Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
uint16 sum_col_u32 = (uint16)0;
__global const uchar *matrix_b = (__global const uchar *)(src.ptr + get_global_id(1) * src_stride_z);
int i = 0;
// This for loop performs 4 accumulations
for(; i <= ((int)ROWS_B - 4); i += 4)
{
const uchar16 b0_u8 = vload16(0, matrix_b + 0 * src_stride_y);
const uchar16 b1_u8 = vload16(0, matrix_b + 1 * src_stride_y);
const uchar16 b2_u8 = vload16(0, matrix_b + 2 * src_stride_y);
const uchar16 b3_u8 = vload16(0, matrix_b + 3 * src_stride_y);
sum_col_u32 += convert_uint16(b0_u8) + convert_uint16(b1_u8) + convert_uint16(b2_u8) + convert_uint16(b3_u8);
matrix_b += 4 * src_stride_y;
}
// This for loop perfoms the leftover accumulations
for(; i < (int)ROWS_B; ++i)
{
const uchar16 b0_u8 = vload16(0, matrix_b);
sum_col_u32 += convert_uint16(b0_u8);
matrix_b += src_stride_y;
}
vstore16(convert_int16(sum_col_u32), 0, (__global int *)dst.ptr);
}
#endif // defined(COLS_B) && defined(ROWS_B)
#if defined(K_OFFSET)
/* Helper function used to calculate the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel.
*
* This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel),
* and calculates the offset contribution of matrix A and matrix B.
*
* @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200)
* @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1)
* @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6)
* @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches
*
* @param[in] x get_global_id(0) * 4
* @param[in] y get_global_id(1)
* @param[in] z get_global_id(2)
* @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr
* @param[in] sum_col_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)
* @param[in] sum_col_step_x (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] sum_col_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)
* @param[in] sum_col_step_y (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] sum_col_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor
* @param[in] sum_row_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr
* @param[in] sum_row_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)
* @param[in] sum_row_step_x (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] sum_row_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)
* @param[in] sum_row_step_y (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] sum_row_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor
* @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr
* @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor
*/
inline int4 offset_contribution(
int x,
int y,
int z
#if defined(A_OFFSET)
,
IMAGE_DECLARATION(sum_col)
#endif // defined(A_OFFSET)
#if defined(B_OFFSET)
,
IMAGE_DECLARATION(sum_row)
#endif // defined(B_OFFSET)
#if defined(ADD_BIAS)
,
VECTOR_DECLARATION(biases)
#endif // defined(ADD_BIAS)
)
{
int4 a_offset_s32 = (int4)0;
int4 b_offset_s32 = (int4)0;
int batch_id = z;
#if defined(DEPTH_INPUT3D)
batch_id /= (int)DEPTH_INPUT3D;
#endif // defined(DEPTH_INPUT3D)
#if defined(A_OFFSET)
// Compute the offset contribution due to A_OFFSET
__global uchar *sum_col_addr = sum_col_ptr + sum_col_offset_first_element_in_bytes + x * sizeof(int);
// Compute the offset contribution due to A_OFFSET
#if defined(SUM_COL_HAS_BATCHES)
a_offset_s32 = vload4(0, (__global int *)(sum_col_addr + batch_id * sum_col_stride_y));
#else // defined(SUM_COL_HAS_BATCHES)
a_offset_s32 = vload4(0, (__global int *)sum_col_addr);
#endif // defined(SUM_COL_HAS_BATCHES)
a_offset_s32 *= (int4)A_OFFSET;
#endif // defined(A_OFFSET)
#if defined(B_OFFSET)
// Compute the offset contribution due to A_OFFSET
__global uchar *sum_row_addr = sum_row_ptr + sum_row_offset_first_element_in_bytes + y * sizeof(int);
// Compute the offset contribution due to B_OFFSET
#if defined(HEIGHT_INPUT3D) && defined(DEPTH_INPUT3D)
b_offset_s32 = (int4) * (((__global int *)(sum_row_addr + batch_id * sum_row_stride_y)) + (z % (int)DEPTH_INPUT3D) * (int)HEIGHT_INPUT3D);
#else // defined(HEIGHT_INPUT3D) && defined(DEPTH_INPUT3D)
b_offset_s32 = (int4) * (((__global int *)(sum_row_addr + batch_id * sum_row_stride_y)));
#endif // defined(HEIGHT_INPUT3D) && defined(DEPTH_INPUT3D)
b_offset_s32 *= (int4)B_OFFSET;
#endif // defined(B_OFFSET)
#if defined(ADD_BIAS)
// Add bias
__global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(int);
int4 biases_values = vload4(0, (__global int *)bias_addr);
b_offset_s32 += (int4)biases_values;
#endif // defined(ADD_BIAS)
return (int4)K_OFFSET + a_offset_s32 + b_offset_s32;
}
/* OpenCL kernel used to add the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel. The computation is performed in-place
*
* This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel),
* and adds to it the offset contribution of matrix A and matrix B in-place.
*
* @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200)
* @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1)
* @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6)
* @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches
*
* The final result is:
*
* mm_result[i][k] = mm_result[i][k] +
* (sum_col[k] * A_OFFSET) +
* (sum_row[i] * B_OFFSET) +
* (K_OFFSET)
*
* @param[in] mm_result_ptr Pointer to the source tensor. Supported data type: S32
* @param[in] mm_result_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] mm_result_step_x mm_result_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] mm_result_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] mm_result_step_y mm_result_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] mm_result_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] mm_result_step_z mm_result_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] mm_result_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr
* @param[in] sum_col_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)
* @param[in] sum_col_step_x (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] sum_col_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)
* @param[in] sum_col_step_y (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] sum_col_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor
* @param[in] sum_row_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr
* @param[in] sum_row_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)
* @param[in] sum_row_step_x (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] sum_row_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)
* @param[in] sum_row_step_y (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] sum_row_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor
* @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr
* @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor
*/
__kernel void gemmlowp_offset_contribution(TENSOR3D_DECLARATION(mm_result)
#if defined(A_OFFSET)
,
IMAGE_DECLARATION(sum_col)
#endif // defined(A_OFFSET)
#if defined(B_OFFSET)
,
IMAGE_DECLARATION(sum_row)
#endif // defined(B_OFFSET)
#if defined(ADD_BIAS)
,
VECTOR_DECLARATION(biases)
#endif // defined(ADD_BIAS))
)
{
const int x = get_global_id(0) * 4;
const int y = get_global_id(1);
const int z = get_global_id(2);
// Compute offset contribution
int4 offset_term_s32 = offset_contribution(
x, y, z
#if defined(A_OFFSET)
,
sum_col_ptr,
sum_col_stride_x,
sum_col_step_x,
sum_col_stride_y,
sum_col_step_y,
sum_col_offset_first_element_in_bytes
#endif // defined(A_OFFSET)
#if defined(B_OFFSET)
,
sum_row_ptr,
sum_row_stride_x,
sum_row_step_x,
sum_row_stride_y,
sum_row_step_y,
sum_row_offset_first_element_in_bytes
#endif // defined(B_OFFSET)
#if defined(ADD_BIAS)
,
biases_ptr,
biases_stride_x,
biases_step_x,
biases_offset_first_element_in_bytes
#endif // defined(ADD_BIAS)
);
__global uchar *mm_result_addr = mm_result_ptr + mm_result_offset_first_element_in_bytes + x * sizeof(int) + y * mm_result_stride_y + z * mm_result_stride_z;
int4 in_s32 = vload4(0, (__global int *)mm_result_addr);
// Add the offset terms to GEMM's result
in_s32 += offset_term_s32;
// Store the result with the offset contribution
vstore4(in_s32, 0, (__global int *)mm_result_addr);
}
#if defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT)
/* OpenCL kernel used to add the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel and it quantizes down to uint8.
*
* This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel), adds to it the offset contribution of matrix A and matrix B and quantizes to uint8 through the output stage.
*
*
* @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200)
* @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1)
* @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6)
* @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches
*
* The result before the output stage is:
*
* mm_result[i][k] = mm_result[i][k] +
* (sum_col[k] * A_OFFSET) +
* (sum_row[i] * B_OFFSET) +
* (K_OFFSET)
*
* This result is quantized down to uint8 using the output stage. The output stage computes the following operations:
*
* -# Add offset terms to final result
* -# Multiply each entry of result by result_mult_int
* -# Add bias to final result (if -DADD_BIAS is passed at compile time)
* -# Shift the int32 accumulator by result_shift
* -# Clamp the value between the specified min and max bounds (if -DMIN_BOUND and/or -DMAX_BOUND are passed at compile time)
* -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
*
* @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET, -RESULT_MULT_INT and -DRESULT_SHIFT
*
* @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time
* @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND.
* These values can be used to implement "rectified linear unit" activation functions
*
* @param[in] mm_result_ptr Pointer to the source tensor. Supported data type: S32
* @param[in] mm_result_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] mm_result_step_x mm_result_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] mm_result_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] mm_result_step_y mm_result_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] mm_result_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] mm_result_step_z mm_result_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] mm_result_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr
* @param[in] sum_col_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)
* @param[in] sum_col_step_x (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] sum_col_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)
* @param[in] sum_col_step_y (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] sum_col_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor
* @param[in] sum_row_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr
* @param[in] sum_row_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)
* @param[in] sum_row_step_x (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] sum_row_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)
* @param[in] sum_row_step_y (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] sum_row_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor
* @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr
* @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor
* @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_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_gx_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 src_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 gemmlowp_offset_contribution_quantize_down(TENSOR3D_DECLARATION(mm_result)
#if defined(A_OFFSET)
,
IMAGE_DECLARATION(sum_col)
#endif // defined(A_OFFSET)
#if defined(B_OFFSET)
,
IMAGE_DECLARATION(sum_row)
#endif // defined(B_OFFSET)
,
#if defined(ADD_BIAS)
VECTOR_DECLARATION(biases),
#endif // defined(ADD_BIAS)
TENSOR3D_DECLARATION(dst))
{
const int x = get_global_id(0) * 4;
const int y = get_global_id(1);
const int z = get_global_id(2);
__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z;
// Compute offset contribution
int4 offset_term_s32 = offset_contribution(
x, y, z
#if defined(A_OFFSET)
,
sum_col_ptr,
sum_col_stride_x,
sum_col_step_x,
sum_col_stride_y,
sum_col_step_y,
sum_col_offset_first_element_in_bytes
#endif // defined(A_OFFSET)
#if defined(B_OFFSET)
,
sum_row_ptr,
sum_row_stride_x,
sum_row_step_x,
sum_row_stride_y,
sum_row_step_y,
sum_row_offset_first_element_in_bytes
#endif // defined(B_OFFSET)
#if defined(ADD_BIAS)
,
biases_ptr,
biases_stride_x,
biases_step_x,
biases_offset_first_element_in_bytes
#endif // defined(ADD_BIAS)
);
__global uchar *mm_result_addr = mm_result_ptr + mm_result_offset_first_element_in_bytes + x * sizeof(int) + y * mm_result_stride_y + z * mm_result_stride_z;
int4 in_s32 = vload4(0, (__global int *)mm_result_addr);
// Add the offset terms to GEMM's result
in_s32 += offset_term_s32;
// -------------- OUTPUT STAGE
// Add the offset terms to GEMM's result
in_s32 += (int4)RESULT_OFFSET;
// Multiply by result_mult_int and shift
in_s32 *= RESULT_MULTIPLIER;
in_s32 >>= RESULT_SHIFT;
uchar4 res = convert_uchar4_sat(in_s32);
#if defined(MIN_BOUND)
res = max(res, (uchar4)MIN_BOUND);
#endif // defined(MIN_BOUND)
#if defined(MAX_BOUND)
res = min(res, (uchar4)MAX_BOUND);
#endif // defined(MAX_BOUND)
// Store the result
vstore4(res, 0, dst_addr);
}
/* OpenCL kernel used to add the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel and it quantizes down to uint8.
*
* This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel), adds to it the offset contribution of matrix A and matrix B and quantizes to uint8 through the output stage.
*
*
* @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200)
* @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1)
* @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6)
* @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches
*
* The result before the output stage is:
*
* mm_result[i][k] = mm_result[i][k] +
* (sum_col[k] * A_OFFSET) +
* (sum_row[i] * B_OFFSET) +
* (K_OFFSET)
*
* This result is quantized down to uint8 using the output stage. The output stage computes the following operations:
*
* -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
* -# Add bias to final result if bias tensor is not a nullptr
* -# Round to nearest division by a power-of-two using result_shift
* -# Add offset to each result
* -# Clamp the value between the specified min and max bounds
* -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
*
* @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET, -RESULT_MULT_INT and -DRESULT_SHIFT
*
* @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time
* @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND.
* These values can be used to implement "rectified linear unit" activation functions
*
* @param[in] mm_result_ptr Pointer to the source tensor. Supported data type: S32
* @param[in] mm_result_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] mm_result_step_x mm_result_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] mm_result_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] mm_result_step_y mm_result_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] mm_result_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] mm_result_step_z mm_result_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] mm_result_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr
* @param[in] sum_col_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)
* @param[in] sum_col_step_x (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] sum_col_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)
* @param[in] sum_col_step_y (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] sum_col_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor
* @param[in] sum_row_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr
* @param[in] sum_row_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)
* @param[in] sum_row_step_x (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] sum_row_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)
* @param[in] sum_row_step_y (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] sum_row_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor
* @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr
* @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor
* @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_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_gx_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 src_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 gemmlowp_offset_contribution_quantize_down_fixedpoint(TENSOR3D_DECLARATION(mm_result)
#if defined(A_OFFSET)
,
IMAGE_DECLARATION(sum_col)
#endif // defined(A_OFFSET)
#if defined(B_OFFSET)
,
IMAGE_DECLARATION(sum_row)
#endif // defined(B_OFFSET)
,
#if defined(ADD_BIAS)
VECTOR_DECLARATION(biases),
#endif // defined(ADD_BIAS)
TENSOR3D_DECLARATION(dst))
{
const int x = get_global_id(0) * 4;
const int y = get_global_id(1);
const int z = get_global_id(2);
// Compute offset contribution
int4 offset_term_s32 = offset_contribution(
x, y, z
#if defined(A_OFFSET)
,
sum_col_ptr,
sum_col_stride_x,
sum_col_step_x,
sum_col_stride_y,
sum_col_step_y,
sum_col_offset_first_element_in_bytes
#endif // defined(A_OFFSET)
#if defined(B_OFFSET)
,
sum_row_ptr,
sum_row_stride_x,
sum_row_step_x,
sum_row_stride_y,
sum_row_step_y,
sum_row_offset_first_element_in_bytes
#endif // defined(B_OFFSET)
#if defined(ADD_BIAS)
,
biases_ptr,
biases_stride_x,
biases_step_x,
biases_offset_first_element_in_bytes
#endif // defined(ADD_BIAS)
);
__global uchar *mm_result_addr = mm_result_ptr + mm_result_offset_first_element_in_bytes + x * sizeof(int) + y * mm_result_stride_y + z * mm_result_stride_z;
__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z;
int4 in_s32 = vload4(0, (__global int *)mm_result_addr);
// Add the offset terms to GEMM's result
in_s32 += offset_term_s32;
// -------------- OUTPUT STAGE
// Multiply by result_mult_int and shift
in_s32 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(in_s32, RESULT_MULTIPLIER, RESULT_SHIFT, 4);
// Add the offset terms to GEMM's result
in_s32 += (int4)RESULT_OFFSET;
uchar4 res = convert_uchar4_sat(in_s32);
#if defined(MIN_BOUND)
res = max(res, (uchar4)MIN_BOUND);
#endif // defined(MIN_BOUND)
#if defined(MAX_BOUND)
res = min(res, (uchar4)MAX_BOUND);
#endif // defined(MAX_BOUND)
// Store the result
vstore4(res, 0, dst_addr);
}
#endif // defined(K_OFFSET) && defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT)
#endif // defined(K_OFFSET)
#if defined(RESULT_OFFSET) && defined(RESULT_MULT_INT) && defined(RESULT_SHIFT)
/** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8
*
* This kernel takes a final int32 accumulator value and processes it to obtain the final QASYMM8 value.
* The following computations will be performed by the kernel:
*
* -# Add offset terms to final result
* -# Multiply each entry of result by result_mult_int
* -# Add bias to final result (if -DADD_BIAS is passed at compile time)
* -# Shift the int32 accumulator by result_shift
* -# Clamp the value between the specified min and max bounds (if -DMIN_BOUND and/or -DMAX_BOUND are passed at compile time)
* -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
*
* @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET, -RESULT_MULT_INT and -DRESULT_SHIFT
*
* @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time
* @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND.
* These values can be used to implement "rectified linear unit" activation functions
*
* @param[in] src_ptr Pointer to the source tensor. Supported data type: S32
* @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[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr
* @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor
* @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_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_gx_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 src_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 gemmlowp_output_stage_quantize_down(TENSOR3D_DECLARATION(src),
#if defined(ADD_BIAS)
VECTOR_DECLARATION(biases),
#endif // defined(ADD_BIAS)
TENSOR3D_DECLARATION(dst))
{
// Compute source and destination addresses
int x = get_global_id(0) * 4;
int y = get_global_id(1);
int z = get_global_id(2);
__global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(int) + y * src_stride_y + z * src_stride_z;
__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z;
int4 input_values = vload4(0, (__global int *)src_addr);
#if defined(ADD_BIAS)
// Add bias
__global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(int);
int4 biases_values = vload4(0, (__global int *)bias_addr);
input_values += (int4)biases_values;
#endif // defined(ADD_BIAS)
// Add the offset terms to GEMM's result
input_values += (int4)RESULT_OFFSET;
// Multiply by result_mult_int and shift
input_values *= RESULT_MULT_INT;
input_values >>= RESULT_SHIFT;
uchar4 res = convert_uchar4_sat(input_values);
#if defined(MIN_BOUND)
res = max(res, (uchar4)MIN_BOUND);
#endif // defined(MIN_BOUND)
#if defined(MAX_BOUND)
res = min(res, (uchar4)MAX_BOUND);
#endif // defined(MAX_BOUND)
// Store the result
vstore4(res, 0, dst_addr);
}
#endif // defined(RESULT_OFFSET) && defined(RESULT_MULT_INT) && defined(RESULT_SHIFT)
#if defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT)
/** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8
*
* This kernel takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value.
* The following computations will be performed by the kernel:
*
* -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
* -# Add bias to final result if bias tensor is not a nullptr
* -# Round to nearest division by a power-of-two using result_shift
* -# Add offset to each result
* -# Clamp the value between the specified min and max bounds
* -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
*
* @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET_AFTER_SHIFT, -DRESULT_FIXEDPOINT_MULTIPLIER and -DRESULT_SHIFT
*
* @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time
* @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND.
* These values can be used to implement "rectified linear unit" activation functions
*
* @param[in] src_ptr Pointer to the source tensor. Supported data type: S32
* @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[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr
* @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor
* @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_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_gx_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 src_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 gemmlowp_output_stage_quantize_down_fixedpoint(TENSOR3D_DECLARATION(src),
#if defined(ADD_BIAS)
VECTOR_DECLARATION(biases),
#endif // defined(ADD_BIAS)
TENSOR3D_DECLARATION(dst))
{
// Compute source and destination addresses
int x = get_global_id(0) * 4;
int y = get_global_id(1);
int z = get_global_id(2);
__global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(int) + y * src_stride_y + z * src_stride_z;
__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z;
int4 input_values = vload4(0, (__global int *)src_addr);
#if defined(ADD_BIAS)
// Add bias
__global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(int);
int4 biases_values = vload4(0, (__global int *)bias_addr);
input_values += (int4)biases_values;
#endif // defined(ADD_BIAS)
// Multiply by result_mult_int and shift
input_values = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(input_values, RESULT_FIXEDPOINT_MULTIPLIER, RESULT_SHIFT, 4);
// Add the offset terms to GEMM's result
input_values += (int4)RESULT_OFFSET_AFTER_SHIFT;
uchar4 res = convert_uchar4_sat(input_values);
#if defined(MIN_BOUND)
res = max(res, (uchar4)MIN_BOUND);
#endif // defined(MIN_BOUND)
#if defined(MAX_BOUND)
res = min(res, (uchar4)MAX_BOUND);
#endif // defined(MAX_BOUND)
// Store the result
vstore4(res, 0, dst_addr);
}
#endif // defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT)
#if defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET)
/** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8
*
* This kernel takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value.
* The following computations will be performed by the kernel:
*
* -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
* -# Add bias to final result if bias tensor is not a nullptr
* -# Requantize
* -# Add offset to each result
* -# Clamp the value between the specified min and max bounds
* -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
*
* @attention The offset and scalar scale factor must be passed at compile time using -DRESULT_OFFSET, -DREAL_MULTIPLIER
*
* @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time
* @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND.
* These values can be used to implement "rectified linear unit" activation functions
*
* @param[in] src_ptr Pointer to the source tensor. Supported data type: S32
* @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[in] biases_ptr Pointer to the biases tensor. Supported data type: 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[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_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_gx_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 src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] dst_stride_w Stride of the source tensor in W dimension (in bytes)
* @param[in] dst_step_w src_stride_w * number of elements along W 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 gemmlowp_output_stage_quantize_down_float(TENSOR3D_DECLARATION(src),
#if defined(ADD_BIAS)
VECTOR_DECLARATION(biases),
#endif // defined(ADD_BIAS)
#if defined(DST_HEIGHT)
TENSOR4D_DECLARATION(dst))
#else // defined(DST_HEIGHT)
TENSOR3D_DECLARATION(dst))
#endif // defined(DST_HEIGHT)
{
// Compute source and destination addresses
int x = get_global_id(0) * 4;
int y = get_global_id(1);
int z = get_global_id(2);
__global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(int) + y * src_stride_y + z * src_stride_z;
__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z;
int4 input_values = vload4(0, (__global int *)src_addr);
#if defined(ADD_BIAS)
// Add bias
__global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(int);
int4 biases_values = vload4(0, (__global int *)bias_addr);
input_values += (int4)biases_values;
#endif // defined(ADD_BIAS)
// Convert to float
float16 input_values_f = convert_float4(input_values);
input_values_f = round(input_values_f * (float)REAL_MULTIPLIER + (float)OUTPUT_OFFSET);
uchar4 res = convert_uchar4_sat(input_values_f);
#if defined(MIN_BOUND)
res = max(res, (uchar4)MIN_BOUND);
#endif // defined(MIN_BOUND)
#if defined(MAX_BOUND)
res = min(res, (uchar4)MAX_BOUND);
#endif // defined(MAX_BOUND)
// Store the result
vstore4(res, 0, dst_addr);
}
#endif // defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET)