diff options
Diffstat (limited to 'src/core/CL/cl_kernels/common/gemmlowp.cl')
| -rw-r--r-- | src/core/CL/cl_kernels/common/gemmlowp.cl | 2162 |
1 files changed, 2162 insertions, 0 deletions
diff --git a/src/core/CL/cl_kernels/common/gemmlowp.cl b/src/core/CL/cl_kernels/common/gemmlowp.cl new file mode 100644 index 0000000000..62c4cd31f5 --- /dev/null +++ b/src/core/CL/cl_kernels/common/gemmlowp.cl @@ -0,0 +1,2162 @@ +/* + * Copyright (c) 2017-2022 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 "gemm_helpers.h" +#include "helpers_asymm.h" +#include "repeat.h" +#include "tile_helpers.h" + +#if defined(DATA_TYPE) && defined(ACC_DATA_TYPE) + +#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(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) + +#define ARM_DOT1(a, b, c) \ + ({ \ + ARM_DOT((VEC_DATA_TYPE(DATA_TYPE, 4))(a, (VEC_DATA_TYPE(DATA_TYPE, 3))0), (VEC_DATA_TYPE(DATA_TYPE, 4))(b, (VEC_DATA_TYPE(DATA_TYPE, 3))0), c); \ + }) +#define ARM_DOT2(a, b, c) \ + ({ \ + ARM_DOT((VEC_DATA_TYPE(DATA_TYPE, 4))(a, (VEC_DATA_TYPE(DATA_TYPE, 2))0), (VEC_DATA_TYPE(DATA_TYPE, 4))(b, (VEC_DATA_TYPE(DATA_TYPE, 2))0), c); \ + }) +#define ARM_DOT3(a, b, c) \ + ({ \ + ARM_DOT((VEC_DATA_TYPE(DATA_TYPE, 4))(a, (DATA_TYPE)0), (VEC_DATA_TYPE(DATA_TYPE, 4))(b, (DATA_TYPE)0), c); \ + }) +#define ARM_DOT4(a, b, c) \ + ({ \ + ARM_DOT(a, b, c); \ + }) +#define ARM_DOT8(a, b, c) \ + ({ \ + ARM_DOT4((a.lo), (b.lo), c); \ + ARM_DOT4((a.hi), (b.hi), c); \ + }) +#define ARM_DOT16(a, b, c) \ + ({ \ + ARM_DOT8((a.lo), (b.lo), c); \ + ARM_DOT8((a.hi), (b.hi), c); \ + }) + +#else // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) + +/** Specialized macros to perform the dot product instruction between two vectors of size K0 [1,16] without using the dot8 instruction. */ +#define ARM_DOT1(a, b, c) \ + ({ \ + c += (ACC_DATA_TYPE)a * b; \ + }) +#define ARM_DOT2(a, b, c) \ + ({ \ + c += (ACC_DATA_TYPE)a.s0 * b.s0; \ + c += (ACC_DATA_TYPE)a.s1 * b.s1; \ + }) +#define ARM_DOT3(a, b, c) \ + ({ \ + ARM_DOT2(a, b, c); \ + c += (ACC_DATA_TYPE)a.s2 * b.s2; \ + }) +#define ARM_DOT4(a, b, c) \ + ({ \ + ARM_DOT3(a, b, c); \ + c += (ACC_DATA_TYPE)a.s3 * b.s3; \ + }) +#define ARM_DOT8(a, b, c) \ + ({ \ + ARM_DOT4((a.lo), (b.lo), c); \ + ARM_DOT4((a.hi), (b.hi), c); \ + }) +#define ARM_DOT16(a, b, c) \ + ({ \ + ARM_DOT8((a.lo), (b.lo), c); \ + ARM_DOT8((a.hi), (b.hi), c); \ + }) +#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) + +/** Specialized macros to perform a broadcast dot product operation between one vector "a" and N0 vectors "b" of size K0 [1,16] */ +#define ARM_DOT_K0X1(k0, a, b, c) \ + ({ \ + ARM_DOT_K0(k0, (a), (b##0), (c)); \ + }) +#define ARM_DOT_K0X2(k0, a, b, c) \ + ({ \ + ARM_DOT_K0(k0, (a), (b##0), (c.s0)); \ + ARM_DOT_K0(k0, (a), (b##1), (c.s1)); \ + }) +#define ARM_DOT_K0X3(k0, a, b, c) \ + ({ \ + ARM_DOT_K0X2(k0, a, b, c); \ + ARM_DOT_K0(k0, (a), (b##2), (c.s2)); \ + }) +#define ARM_DOT_K0X4(k0, a, b, c) \ + ({ \ + ARM_DOT_K0X3(k0, a, b, c); \ + ARM_DOT_K0(k0, (a), (b##3), (c.s3)); \ + }) +#define ARM_DOT_K0X8(k0, a, b, c) \ + ({ \ + ARM_DOT_K0X4(k0, a, b, c); \ + ARM_DOT_K0(k0, (a), (b##4), (c.s4)); \ + ARM_DOT_K0(k0, (a), (b##5), (c.s5)); \ + ARM_DOT_K0(k0, (a), (b##6), (c.s6)); \ + ARM_DOT_K0(k0, (a), (b##7), (c.s7)); \ + }) +#define ARM_DOT_K0X16(k0, a, b, c) \ + ({ \ + ARM_DOT_K0X8(k0, a, b, c); \ + ARM_DOT_K0(k0, (a), (b##8), (c.s8)); \ + ARM_DOT_K0(k0, (a), (b##9), (c.s9)); \ + ARM_DOT_K0(k0, (a), (b##A), (c.sA)); \ + ARM_DOT_K0(k0, (a), (b##B), (c.sB)); \ + ARM_DOT_K0(k0, (a), (b##C), (c.sC)); \ + ARM_DOT_K0(k0, (a), (b##D), (c.sD)); \ + ARM_DOT_K0(k0, (a), (b##E), (c.sE)); \ + ARM_DOT_K0(k0, (a), (b##F), (c.sF)); \ + }) + +/** Specialized macros to perform a partial matrix multiplication with dimensions M0,N0,K0 */ +#define ARM_MM_K0XN0X1(n0, k0, a, b, c) \ + ({ \ + ARM_DOT_K0XN0(n0, k0, (a##0), b, (c##0)); \ + }) +#define ARM_MM_K0XN0X2(n0, k0, a, b, c) \ + ({ \ + ARM_MM_K0XN0X1(n0, k0, a, b, c); \ + ARM_DOT_K0XN0(n0, k0, (a##1), b, (c##1)); \ + }) +#define ARM_MM_K0XN0X3(n0, k0, a, b, c) \ + ({ \ + ARM_MM_K0XN0X2(n0, k0, a, b, c); \ + ARM_DOT_K0XN0(n0, k0, (a##2), b, (c##2)); \ + }) +#define ARM_MM_K0XN0X4(n0, k0, a, b, c) \ + ({ \ + ARM_MM_K0XN0X3(n0, k0, a, b, c); \ + ARM_DOT_K0XN0(n0, k0, (a##3), b, (c##3)); \ + }) +#define ARM_MM_K0XN0X5(n0, k0, a, b, c) \ + ({ \ + ARM_MM_K0XN0X4(n0, k0, a, b, c); \ + ARM_DOT_K0XN0(n0, k0, (a##4), b, (c##4)); \ + }) +#define ARM_MM_K0XN0X6(n0, k0, a, b, c) \ + ({ \ + ARM_MM_K0XN0X5(n0, k0, a, b, c); \ + ARM_DOT_K0XN0(n0, k0, (a##5), b, (c##5)); \ + }) +#define ARM_MM_K0XN0X7(n0, k0, a, b, c) \ + ({ \ + ARM_MM_K0XN0X6(n0, k0, a, b, c); \ + ARM_DOT_K0XN0(n0, k0, (a##6), b, (c##6)); \ + }) +#define ARM_MM_K0XN0X8(n0, k0, a, b, c) \ + ({ \ + ARM_MM_K0XN0X7(n0, k0, a, b, c); \ + ARM_DOT_K0XN0(n0, k0, (a##7), b, (c##7)); \ + }) + +#define ARM_DOT_K0(k0, a, b, c) \ + ({ \ + CONCAT(ARM_DOT, k0) \ + ((a), (b), (c)); \ + }) + +#define ARM_DOT_K0XN0(n0, k0, a, b, c) \ + ({ \ + CONCAT(ARM_DOT_K0X, n0) \ + (k0, (a), b, (c)); \ + }) + +#define ARM_MM_K0XN0XM0(m0, n0, k0, a, b, c) \ + ({ \ + CONCAT(ARM_MM_K0XN0X, m0) \ + (n0, k0, a, b, c); \ + }) + +/** Specialized macros to perform a broadcast dot product operation between one vector "a" and N0 vectors "b" of size K0 [1,16] */ +#define ARM_MUL_N0X1(VECTOR_ACC_TYPE, a, b, c) \ + ({ \ + c += CONVERT(b##0, VECTOR_ACC_TYPE) * a; \ + }) +#define ARM_MUL_N0X2(VECTOR_ACC_TYPE, a, b, c) \ + ({ \ + c += CONVERT(b##0, VECTOR_ACC_TYPE) * a.s##0; \ + c += CONVERT(b##1, VECTOR_ACC_TYPE) * a.s##1; \ + }) +#define ARM_MUL_N0X3(VECTOR_ACC_TYPE, a, b, c) \ + ({ \ + ARM_MUL_N0X2(VECTOR_ACC_TYPE, a, b, c); \ + c += CONVERT(b##2, VECTOR_ACC_TYPE) * a.s##2; \ + }) +#define ARM_MUL_N0X4(VECTOR_ACC_TYPE, a, b, c) \ + ({ \ + ARM_MUL_N0X3(VECTOR_ACC_TYPE, a, b, c); \ + c += CONVERT(b##3, VECTOR_ACC_TYPE) * a.s##3; \ + }) +#define ARM_MUL_N0X8(VECTOR_ACC_TYPE, a, b, c) \ + ({ \ + ARM_MUL_N0X4(VECTOR_ACC_TYPE, a, b, c); \ + c += CONVERT(b##4, VECTOR_ACC_TYPE) * a.s##4; \ + c += CONVERT(b##5, VECTOR_ACC_TYPE) * a.s##5; \ + c += CONVERT(b##6, VECTOR_ACC_TYPE) * a.s##6; \ + c += CONVERT(b##7, VECTOR_ACC_TYPE) * a.s##7; \ + }) +#define ARM_MUL_N0X16(VECTOR_ACC_TYPE, a, b, c) \ + ({ \ + ARM_MUL_N0X8(VECTOR_ACC_TYPE, a, b, c); \ + c += CONVERT(b##8, VECTOR_ACC_TYPE) * a.s##8; \ + c += CONVERT(b##9, VECTOR_ACC_TYPE) * a.s##9; \ + c += CONVERT(b##A, VECTOR_ACC_TYPE) * a.s##A; \ + c += CONVERT(b##B, VECTOR_ACC_TYPE) * a.s##B; \ + c += CONVERT(b##C, VECTOR_ACC_TYPE) * a.s##C; \ + c += CONVERT(b##D, VECTOR_ACC_TYPE) * a.s##D; \ + c += CONVERT(b##E, VECTOR_ACC_TYPE) * a.s##E; \ + c += CONVERT(b##F, VECTOR_ACC_TYPE) * a.s##F; \ + }) +/** Specialized macros to perform a a partial matrix multiplication with dimensions M0,N0,K0 */ +#define ARM_MM_NATIVE_N0XK0X1(VECTOR_ACC_TYPE, k0, a, b, c) \ + ({ \ + ARM_MUL_N0XK0(VECTOR_ACC_TYPE, k0, (a##0), b, (c##0)); \ + }) +#define ARM_MM_NATIVE_N0XK0X2(VECTOR_ACC_TYPE, k0, a, b, c) \ + ({ \ + ARM_MM_NATIVE_N0XK0X1(VECTOR_ACC_TYPE, k0, a, b, c); \ + ARM_MUL_N0XK0(VECTOR_ACC_TYPE, k0, (a##1), b, (c##1)); \ + }) +#define ARM_MM_NATIVE_N0XK0X3(VECTOR_ACC_TYPE, k0, a, b, c) \ + ({ \ + ARM_MM_NATIVE_N0XK0X2(VECTOR_ACC_TYPE, k0, a, b, c); \ + ARM_MUL_N0XK0(VECTOR_ACC_TYPE, k0, (a##2), b, (c##2)); \ + }) +#define ARM_MM_NATIVE_N0XK0X4(VECTOR_ACC_TYPE, k0, a, b, c) \ + ({ \ + ARM_MM_NATIVE_N0XK0X3(VECTOR_ACC_TYPE, k0, a, b, c); \ + ARM_MUL_N0XK0(VECTOR_ACC_TYPE, k0, (a##3), b, (c##3)); \ + }) +#define ARM_MM_NATIVE_N0XK0X5(VECTOR_ACC_TYPE, k0, a, b, c) \ + ({ \ + ARM_MM_NATIVE_N0XK0X4(VECTOR_ACC_TYPE, k0, a, b, c); \ + ARM_MUL_N0XK0(VECTOR_ACC_TYPE, k0, (a##4), b, (c##4)); \ + }) +#define ARM_MM_NATIVE_N0XK0X6(VECTOR_ACC_TYPE, k0, a, b, c) \ + ({ \ + ARM_MM_NATIVE_N0XK0X5(VECTOR_ACC_TYPE, k0, a, b, c); \ + ARM_MUL_N0XK0(VECTOR_ACC_TYPE, k0, (a##5), b, (c##5)); \ + }) +#define ARM_MM_NATIVE_N0XK0X7(VECTOR_ACC_TYPE, k0, a, b, c) \ + ({ \ + ARM_MM_NATIVE_N0XK0X6(VECTOR_ACC_TYPE, k0, a, b, c); \ + ARM_MUL_N0XK0(VECTOR_ACC_TYPE, k0, (a##6), b, (c##6)); \ + }) +#define ARM_MM_NATIVE_N0XK0X8(VECTOR_ACC_TYPE, k0, a, b, c) \ + ({ \ + ARM_MM_NATIVE_N0XK0X7(VECTOR_ACC_TYPE, k0, a, b, c); \ + ARM_MUL_N0XK0(VECTOR_ACC_TYPE, k0, (a##7), b, (c##7)); \ + }) +#define ARM_MUL_N0XK0(VECTOR_ACC_TYPE, k0, a, b, c) \ + ({ \ + CONCAT(ARM_MUL_N0X, k0) \ + (VECTOR_ACC_TYPE, (a), b, (c)); \ + }) +#define ARM_MM_NATIVE_N0XK0XM0(VECTOR_ACC_TYPE, m0, k0, a, b, c) \ + ({ \ + CONCAT(ARM_MM_NATIVE_N0XK0X, m0) \ + (VECTOR_ACC_TYPE, k0, a, b, c); \ + }) + +#if defined(GEMMLOWP_MM_RESHAPED_LHS_NT_RHS_T) +/** This OpenCL kernel computes the matrix multiplication between 2 matrices with QASYMM/QASYMM_SIGNED data type. + * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be NOT transposed + * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be transposed + * + * @note The input data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=uchar) + * @note The accumulator data type must be passed at compile time using -DACC_DATA_TYPE (i.e. -DACC_DATA_TYPE=uint) + * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. + * @note The GEMM's dimensions M and N must be passed at compile time using -DM and -DN (i.e. -DM=52 and -DN=90). + * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (i.e. -DM0=4, -DN0=8, -DK0=4). + * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (i.e. -DV0=2) + * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (i.e. -DH0=2) + * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time. + * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. + * @note Only the following configurations of M0, N0 and K0 are currently supported: + * - M0 = 2, 3, 4, 5, 6, 7, 8 + * - N0 = 2, 3, 4, 8, 16 + * - K0 = 2, 3, 4, 8, 16 + * - V0 >= 1 + * - H0 >= 1 + * + * @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 LHS matrix NOT reshaped + * + * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: QASYMM8/QASYMM_SIGNED + * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes) + * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes) + * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix + * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr + * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes) + * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes) + * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped 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_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_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] k Number of columns in LHS matrix and rows in RHS matrix not reshaped. + * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) + * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) + */ +__kernel void gemmlowp_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs), + IMAGE_DECLARATION(rhs), + IMAGE_DECLARATION(dst), + uint k, + uint lhs_stride_z, + uint rhs_stride_z, + uint dst_stride_z +#if defined(REINTERPRET_OUTPUT_AS_3D) + , + uint dst_cross_plane_pad +#endif // REINTERPRET_OUTPUT_AS_3D + ) +{ + // Block size +#define LHS_BLOCK_SIZE ((K0) * (M0)) + +#if defined(LHS_INTERLEAVE) +#define LHS_OFFSET_X (K0) +#define LHS_STEP_X ((K0) * (V0)) +#define LHS_STEP_LOOP (1) +#else // defined(INTERLEAVE) +#define LHS_OFFSET_X (LHS_BLOCK_SIZE) +#define LHS_STEP_X (K0) +#define LHS_STEP_LOOP (V0) +#endif // defined(INTERLEAVE) + + // Block size +#define RHS_BLOCK_SIZE ((K0) * (N0)) + + // RHS offset and step X +#if defined(RHS_INTERLEAVE) +#define RHS_OFFSET_X (K0) +#define RHS_STEP_X ((K0) * (H0)) +#define RHS_STEP_LOOP (1) +#else // defined(RHS_INTERLEAVE) +#define RHS_OFFSET_X (RHS_BLOCK_SIZE) +#define RHS_STEP_X (K0) +#define RHS_STEP_LOOP (H0) +#endif // defined(RHS_INTERLEAVE) + + uint x = get_global_id(0); + uint y = get_global_id(1); + uint z = get_global_id(2); + +#if defined(DUMMY_WORK_ITEMS) + if((x * N0 >= N) || (y * M0 >= M)) + { + return; + } +#endif // defined(DUMMY_WORK_ITEMS) + + // Compute LHS matrix address + __global DATA_TYPE *lhs_addr = (__global DATA_TYPE *)(lhs_ptr + lhs_offset_first_element_in_bytes + (y % V0) * (uint)LHS_OFFSET_X + (y / V0) * (uint)lhs_stride_y + (z * lhs_stride_z)); + + // Compute RHS matrix address + __global DATA_TYPE *rhs_addr = (__global DATA_TYPE *)(rhs_ptr + rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X + (x / (uint)H0) * rhs_stride_y); + +#if defined(MATRIX_B_DEPTH) + // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 + rhs_addr += (z % MATRIX_B_DEPTH) * rhs_stride_z; +#else // defined(MATRIX_B_DEPTH) + rhs_addr += z * rhs_stride_z; +#endif // defined(MATRIX_B_DEPTH) + + REPEAT_VAR_INIT_TO_CONST(8, uint, zlhs, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; + REPEAT_VAR_INIT_TO_CONST(16, uint, zrhs, 0); + + // Initialize the accumulators + REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(ACC_DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(ACC_DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(M0-1)=0; + + for(int i = 0; i < k; i += K0) + { + // Load values from LHS matrix + LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_addr, 0, LHS_STEP_X, zlhs); + + // Load values from RHS matrix + LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_addr, 0, RHS_STEP_X, zrhs); + + // Partial matrix multiplication M0,N0,K0 + ARM_MM_K0XN0XM0(M0, N0, K0, a, b, c); + + // Update address + lhs_addr += (M0 * LHS_STEP_X * LHS_STEP_LOOP); + rhs_addr += (N0 * RHS_STEP_X * RHS_STEP_LOOP); + } + + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(int)) + (y * (uint)M0 * dst_stride_y); + + REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; + +#if defined(REINTERPRET_OUTPUT_AS_3D) + // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, 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; + +#else // defined(REINTERPRET_OUTPUT_AS_3D) + + // Add offset for batched GEMM + dst_addr += z * dst_stride_z; + +#endif // defined(REINTERPRET_OUTPUT_AS_3D) + + // Convert and store output block + const bool cond_y = ((get_global_id(1) + 1) * M0 >= M); + const bool cond_x = ((get_global_id(0) + 1) * N0 >= N); + + // Store output block + REPEAT_VAR_INIT_CONVERT_SAT(M0, VEC_DATA_TYPE(int, N0), c, c_lp); + STORE_BLOCK_BOUNDARY_AWARE(M0, N0, int, c_lp, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + +#undef LHS_BLOCK_SIZE +#undef LHS_OFFSET_X +#undef LHS_STEP_X +#undef RHS_BLOCK_SIZE +#undef RHS_OFFSET_X +#undef RHS_STEP_X +} +#endif // defined(GEMMLOWP_MM_RESHAPED_LHS_NT_RHS_T) + +#if defined(GEMMLOWP_MM_RESHAPED_ONLY_RHS_T_FUSED_OUTPUT_STAGE_FIXEDPOINT) || defined(GEMMLOWP_MM_RESHAPED_ONLY_RHS_T) +#if defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT) +#define FUSED_OUTPUT_STAGE_FIXED_POINT +#endif // defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT) + +/** This OpenCL kernel computes the matrix multiplication between 2 matrices with fused output stage using fixed-point arithmetic. + * The LHS matrix is NOT reshaped + * The RHS matrix is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed + * + * @note The input data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=uchar) + * @note The accumulator data type must be passed at compile time using -DACC_DATA_TYPE (i.e. -DACC_DATA_TYPE=uint) + * @note The number of columns of LHS matrix must be passed at compile time using -DK (i.e. -DK=64) + * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (i.e. -DN0=8, -DK0=4). + * @note The number of M0 rows to process must be passed at compile time using -DM0 (i.e. -DM0=2) + * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (i.e. -DH0=2) + * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. + * @note Only the following configurations of M0, N0 and K0 are currently supported: + * - M0 = 1, 2, 3, 4, 5, 6, 7, 8 + * - N0 = 2, 3, 4, 8, 16 + * - K0 = 2, 3, 4, 8, 16 + * - H0 >= 1 + * + * @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 LHS matrix + * + * @note 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_MULTIPLIER and -DRESULT_SHIFT + * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time + * @note The output datatype should be passed at compile time using -DOUTPUT_DATA_TYPE + * @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 + * @note In case of per-channel quantization of matrix B, -DPER_CHANNEL_QUANTIZATION must be passed at compile time. + * + * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: QASYMM8/QASYMM8_SIGNED + * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes) + * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes) + * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix + * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr + * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes) + * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes) + * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix + * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr + * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix + * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) + * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) + * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) + * @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: S32 + * @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: S32 + * @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: S32 + * @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[in] result_multipliers_ptr (Optional) Pointer to the output multipliers vector for per-channel quantization. Supported data types: S32 + * @param[in] result_multipliers_stride_x (Optional) Stride of the output multipliers vector in X dimension (in bytes) + * @param[in] result_multipliers_step_x (Optional) output_multipliers_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] result_multipliers_offset_first_element_in_bytes (Optional) The offset of the first element in the output multipliers vector + * @param[in] result_shifts_ptr (Optional) Pointer to the output shifts vector for per-channel quantization. Supported data types: S32 + * @param[in] result_shifts_stride_x (Optional) Stride of the output shifts vector in X dimension (in bytes) + * @param[in] result_shifts_step_x (Optional) output_shifts_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] result_shifts_offset_first_element_in_bytes (Optional) The offset of the first element in the output shifts vector + */ +#if defined(GEMMLOWP_MM_RESHAPED_ONLY_RHS_T_FUSED_OUTPUT_STAGE_FIXEDPOINT) +__kernel void gemmlowp_mm_reshaped_only_rhs_t_fused_output_stage_fixedpoint +#elif defined(GEMMLOWP_MM_RESHAPED_ONLY_RHS_T) // defined(GEMMLOWP_MM_RESHAPED_ONLY_RHS_T_FUSED_OUTPUT_STAGE_FIXEDPOINT) +__kernel void gemmlowp_mm_reshaped_only_rhs_t +#endif // defined(GEMMLOWP_MM_RESHAPED_ONLY_RHS_T) +(IMAGE_DECLARATION(lhs), + IMAGE_DECLARATION(rhs), + IMAGE_DECLARATION(dst), + uint lhs_stride_z, + uint rhs_stride_z, + uint dst_stride_z +#if defined(REINTERPRET_INPUT_AS_3D) + , + uint lhs_cross_plane_pad +#endif // REINTERPRET_INPUT_AS_3D +#if defined(REINTERPRET_OUTPUT_AS_3D) + , + uint dst_cross_plane_pad +#endif // REINTERPRET_OUTPUT_AS_3D +#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) +#if defined(PER_CHANNEL_QUANTIZATION) + , + VECTOR_DECLARATION(result_multipliers), + VECTOR_DECLARATION(result_shifts) +#endif // defined(PER_CHANNEL_QUANTIZATION) +) +{ + // @note: replace with (DIMENSION + PAD) once we pass the relevant info at compile time +#define FULL_LHS_HEIGHT (lhs_stride_z / lhs_stride_y) +#define FULL_DST_HEIGHT (dst_stride_z / dst_stride_y) + + // RHS offset and step X +#if defined(RHS_INTERLEAVE) +#define RHS_OFFSET_X (K0) +#define RHS_STEP_X (K0 * H0) +#else // defined(RHS_INTERLEAVE) +#define RHS_OFFSET_X (K0 * N0) +#define RHS_STEP_X (K0) +#endif // defined(RHS_INTERLEAVE) +#define RHS_STEP_LOOP (N0 * K0 * H0) + + uint x = GET_SPATIAL_IDX(0, 1, 1); + uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0); + uint z = GET_SPATIAL_IDX(2, 1, 1); + int xo = (x * N0); + +#if defined(DUMMY_WORK_ITEMS) + if((xo >= N) || (y >= M)) + { + return; + } +#endif // defined(DUMMY_WORK_ITEMS) + + // Compute LHS matrix address + uint lhs_y = y + z * FULL_LHS_HEIGHT; + + // Compute RHS matrix address + uint rhs_offset_x = (x % H0) * RHS_OFFSET_X; + uint rhs_offset_y = (x / H0) * rhs_stride_y; + +#if defined(MATRIX_B_DEPTH) + // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 + rhs_offset_y += (z % MATRIX_B_DEPTH) * rhs_stride_z; +#else // defined(MATRIX_B_DEPTH) + rhs_offset_y += z * rhs_stride_z; +#endif // defined(MATRIX_B_DEPTH) + + // Initialize the accumulators + TILE(ACC_DATA_TYPE, M0, N0, c); + LOOP_UNROLLING(int, i, 0, 1, M0, + { + c[i].v = 0; + }) + + int i = 0; + for(; i <= (K - K0); i += K0) + { + TILE(DATA_TYPE, M0, K0, a); + TILE(DATA_TYPE, N0, K0, b); + + // Load values from LHS matrix + T_LOAD(DATA_TYPE, M0, K0, BUFFER, lhs, i, lhs_y, 1, lhs_stride_y, a); + + // // Load values from RHS matrix + LOOP_UNROLLING(int, _i, 0, 1, N0, + { + b[_i].v = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset_first_element_in_bytes + rhs_offset_x + rhs_offset_y + _i * RHS_STEP_X)); + }) + + // Partial matrix multiplication M0,N0,K0 + T_MMUL(DATA_TYPE, DATA_TYPE, ACC_DATA_TYPE, M0, N0, K0, NT, T, a, b, c); + + rhs_offset_x += RHS_STEP_LOOP; + } + +#if((K % K0) != 0) + + // Left-over accumulations + for(; i < K; ++i) + { + TILE(DATA_TYPE, M0, 1, a); + TILE(DATA_TYPE, N0, 1, b); + + // Load values from LHS matrix + T_LOAD(DATA_TYPE, M0, 1, BUFFER, lhs, i, lhs_y, 1, lhs_stride_y, a); + + LOOP_UNROLLING(int, _i, 0, 1, N0, + { + b[_i].v = *(__global DATA_TYPE *)(rhs_ptr + rhs_offset_first_element_in_bytes + rhs_offset_x + rhs_offset_y + _i * RHS_STEP_X); + }) + + T_MMUL(DATA_TYPE, DATA_TYPE, ACC_DATA_TYPE, M0, N0, 1, NT, T, a, b, c); + + rhs_offset_x += 1; + } +#endif // ((K % K0) != 0) + +#if defined(FUSED_OUTPUT_STAGE_FIXED_POINT) + + TILE(int, M0, N0, c_int); + TILE(int, M0, N0, offset_s32); + LOOP_UNROLLING(int, i, 0, 1, M0, + { + offset_s32[i].v = (VEC_DATA_TYPE(int, N0))K_OFFSET; + }) + + LOOP_UNROLLING(int, i, 0, 1, M0, + { + c_int[i].v = CONVERT_SAT(c[i].v, VEC_DATA_TYPE(int, N0)); + }) + +#if defined(A_OFFSET) + +#if defined(SUM_COL_HAS_BATCHES) + int sum_col_y = z; +#else // defined(SUM_COL_HAS_BATCHES) + int sum_col_y = 0; +#endif // defined(SUM_COL_HAS_BATCHES) + TILE(int, 1, N0, a_offset_s32); + + T_LOAD(int, 1, N0, BUFFER, sum_col, xo, sum_col_y, 1, sum_col_stride_y, a_offset_s32); + + a_offset_s32[0].v *= A_OFFSET; + + T_ELTWISE_BROADCAST_ADD_X(int, M0, N0, offset_s32, a_offset_s32, offset_s32); +#endif // defined(A_OFFSET) + +#if defined(B_OFFSET) + // Compute the offset contribution due to B_OFFSET + // Note: The sum_row tensor is generated through CLGEMMLowpMatrixAReductionKernel which + // does not introduce paddings. For this reason is safe to access the tensor in this manner + // without considering that the coordinate "y" could come from an input 3D tensor + TILE(int, M0, N0, b_offset_s32); + + T_LOAD(int, M0, 1, BUFFER, sum_row, y + z * (sum_row_stride_y / sizeof(int)), 0, 1, sum_row_stride_x, b_offset_s32); + + LOOP_UNROLLING(int, i, 0, 1, M0, + { + offset_s32[i].v += b_offset_s32[i].v *B_OFFSET; + }) + +#endif // defined(B_OFFSET) + +#if defined(ADD_BIAS) + + TILE(int, 1, N0, bias); + + T_LOAD(int, 1, N0, BUFFER, biases, xo, 0, 1, 0, bias); + + T_ELTWISE_BROADCAST_ADD_X(int, M0, N0, offset_s32, bias, offset_s32); +#endif // defined(ADD_BIAS) + + LOOP_UNROLLING(int, i, 0, 1, M0, + { + c_int[i].v += offset_s32[i].v; + }) + + TILE(DATA_TYPE, M0, N0, c_lp); + + // Multiply by result_mult_int and shift +#if defined(PER_CHANNEL_QUANTIZATION) + TILE(int, 1, N0, res_mul); + TILE(int, 1, N0, res_shift); + + T_LOAD(int, 1, N0, BUFFER, result_multipliers, xo, 0, 0, 0, res_mul); + T_LOAD(int, 1, N0, BUFFER, result_shifts, xo, 0, 0, 0, res_shift); + + T_QUANTIZE8(int, DATA_TYPE, PER_CHANNEL, M0, N0, RESULT_OFFSET, RESULT_SHIFT, RESULT_MULTIPLIER, c_int, res_mul, res_shift, c_lp); +#else // defined(PER_CHANNEL_QUANTIZATION) + T_QUANTIZE8(int, DATA_TYPE, PER_TENSOR, M0, N0, RESULT_OFFSET, RESULT_SHIFT, RESULT_MULTIPLIER, c_int, 0, 0, c_lp); +#endif // defined(PER_CHANNEL_QUANTIZATION) + +#if defined(MIN_BOUND) + LOOP_UNROLLING(int, i, 0, 1, M0, + { + c_lp[i].v = max(c_lp[i].v, (VEC_DATA_TYPE(DATA_TYPE, N0))MIN_BOUND); + }) +#endif // defined(MIN_BOUND) +#if defined(MAX_BOUND) + LOOP_UNROLLING(int, i, 0, 1, M0, + { + c_lp[i].v = min(c_lp[i].v, (VEC_DATA_TYPE(DATA_TYPE, N0))MAX_BOUND); + }) +#endif // defined(MAX_BOUND) + +#else // defined(FUSED_OUTPUT_STAGE_FIXED_POINT) + TILE(int, M0, N0, c_lp); + + LOOP_UNROLLING(int, i, 0, 1, M0, + { + c_lp[i].v = CONVERT_SAT(c[i].v, VEC_DATA_TYPE(int, N0)); + }) +#endif // defined(FUSED_OUTPUT_STAGE_FIXED_POINT) + + TILE(uint, M0, 1, dst_indirect_y); + + LOOP_UNROLLING(int, i, 0, 1, M0, + { +#if defined(REINTERPRET_OUTPUT_AS_3D) + dst_indirect_y[i].v = (uint)min((int)((y + i) % HEIGHT_GEMM3D), (int)HEIGHT_GEMM3D - 1); + dst_indirect_y[i].v += (uint)min((int)((y + i) / HEIGHT_GEMM3D), (int)DEPTH_GEMM3D - 1) * FULL_DST_HEIGHT; + dst_indirect_y[i].v += z *FULL_DST_HEIGHT *DEPTH_GEMM3D; +#else // (REINTERPRET_OUTPUT_AS_3D) + dst_indirect_y[i].v = (uint)min((int)y + i, (int)M - 1) + z *FULL_DST_HEIGHT; +#endif // defined(REINTERPRET_OUTPUT_AS_3D) + }) + + const bool cond_x = (xo > (N - N0)) & (PARTIAL_STORE_N0 != 0); + +#if defined(FUSED_OUTPUT_STAGE_FIXED_POINT) + T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, xo, dst_stride_y, cond_x, c_lp, dst_indirect_y); +#else // defined(FUSED_OUTPUT_STAGE_FIXED_POINT) + T_STORE_INDIRECT_WIDTH_SELECT(int, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, xo, dst_stride_y, cond_x, c_lp, dst_indirect_y); +#endif // defined(FUSED_OUTPUT_STAGE_FIXED_POINT) + +#undef RHS_OFFSET_X +#undef RHS_STEP_X +#undef RHS_STEP_LOOP +} +#endif // defined(GEMMLOWP_MM_RESHAPED_ONLY_RHS_T_FUSED_OUTPUT_STAGE_FIXEDPOINT) || defined(GEMMLOWP_MM_RESHAPED_ONLY_RHS_T) + +#if defined(GEMMLOWP_MM_NATIVE) + +/** This OpenCL kernel computes the matrix multiplication between 2 matrices. + * The LHS matrix is NOT reshaped + * The RHS matrix is NOT reshaped + * + * @note The input data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=uchar) + * @note The accumulator data type must be passed at compile time using -DACC_DATA_TYPE (i.e. -DACC_DATA_TYPE=uint) + * @note The number of columns of LHS matrix must be passed at compile time using -DK (i.e. -DK=64) + * @note The number of M0 rows to process must be passed at compile time using -DM0 (i.e. -DM0=2) + * @note The number of N0 columns to process must be passed at compile time using -DN0 (i.e. -DN0=2) + * @note The number of K0 partial accumulations must be passed at compile time using -DK0 (i.e., -DK0=2) + * @note Only the following configurations of M0, N0 and K0 are currently supported: + * - M0 = 1, 2, 3, 4, 5, 6, 7, 8 + * - N0 = 2, 3, 4, 8, 16 + * - K0 = 2, 3, 4, 8, 16 + * + * @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 LHS matrix + * + * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: QASYMM8 + * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes) + * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes) + * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix + * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr + * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes) + * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes) + * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped 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_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_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] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) + * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) + * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) + */ +__kernel void gemmlowp_mm_native(IMAGE_DECLARATION(lhs), + IMAGE_DECLARATION(rhs), + IMAGE_DECLARATION(dst), + uint lhs_stride_z, + uint rhs_stride_z, + uint dst_stride_z +#if defined(REINTERPRET_INPUT_AS_3D) + , + uint lhs_cross_plane_pad +#endif // REINTERPRET_INPUT_AS_3D +#if defined(REINTERPRET_OUTPUT_AS_3D) + , + uint dst_cross_plane_pad +#endif // REINTERPRET_OUTPUT_AS_3D + ) +{ + uint x = get_global_id(0); + uint y = get_global_id(1); + uint z = get_global_id(2); + +#if defined(DUMMY_WORK_ITEMS) + if((x * N0 >= N) || (y * M0 >= M)) + { + return; + } +#endif // defined(DUMMY_WORK_ITEMS) + + // Compute LHS matrix address + uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y; + + // Compute RHS matrix address + uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE); + +#if defined(MATRIX_B_DEPTH) + // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 + rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z; +#else // defined(MATRIX_B_DEPTH) + rhs_offset += z * rhs_stride_z; +#endif // defined(MATRIX_B_DEPTH) + + REPEAT_VAR_INIT_TO_CONST(8, uint, zlhs, 0); + REPEAT_VAR_INIT_TO_CONST(16, uint, zrhs, 0); + +#if defined(REINTERPRET_INPUT_AS_3D) + // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply lhs_stride_z by DEPTH_GEMM3D + lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_INPUT_AS_3D) + + // Add offset for batched GEMM + lhs_offset += z * lhs_stride_z; + +#endif // defined(REINTERPRET_INPUT_AS_3D) + + // Initialize the accumulators + REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(ACC_DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(ACC_DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(M0-1)=0; + + int i = 0; + + for(; i <= (K - K0); i += K0) + { + // Load values from LHS matrix + LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs); + + // Load values from RHS matrix + LOAD_BLOCK(K0, N0, DATA_TYPE, b, rhs_ptr, rhs_offset, rhs_stride_y, zrhs); + + // Partial matrix multiplication M0,N0,K0 +#if(GPU_ARCH == GPU_ARCH_MIDGARD) + ARM_MM_NATIVE_N0XK0XM0(VEC_DATA_TYPE(ACC_DATA_TYPE, N0), M0, K0, a, b, c); +#else // GPU_ARCH == GPU_ARCH_MIDGARD + // Transpose the values from RHS matrix + TRANSPOSE_K0XN0(K0, N0, b_t, b, DATA_TYPE); + + ARM_MM_K0XN0XM0(M0, N0, K0, a, b_t, c); +#endif // GPU_ARCH == GPU_ARCH_MIDGARD + + // Update the offset + lhs_offset += K0; + rhs_offset += K0 * rhs_stride_y; + } + + // Left-over for loop + for(; i < K; ++i) + { + // Load values from LHS matrix + LOAD_BLOCK(M0, 1, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs); + + // Load values from RHS matrix + LOAD_BLOCK(1, N0, DATA_TYPE, b, rhs_ptr, rhs_offset, rhs_stride_y, zrhs); + + // Partial matrix multiplication M0,N0,1 +#if(GPU_ARCH == GPU_ARCH_MIDGARD) + ARM_MM_NATIVE_N0XK0XM0(VEC_DATA_TYPE(ACC_DATA_TYPE, N0), M0, 1, a, b, c); +#else // GPU_ARCH == GPU_ARCH_MIDGARD + // Transpose the values from RHS matrix + TRANSPOSE_K0XN0(1, N0, b_t, b, DATA_TYPE); + + ARM_MM_K0XN0XM0(M0, N0, 1, a, b_t, c); +#endif // GPU_ARCH == GPU_ARCH_MIDGARD + + // Update the offset + lhs_offset += 1; + rhs_offset += rhs_stride_y; + } + + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(int)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); + + REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; + +#if defined(REINTERPRET_OUTPUT_AS_3D) + // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, 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; + +#else // defined(REINTERPRET_OUTPUT_AS_3D) + + // Add offset for batched GEMM + dst_addr += z * dst_stride_z; + +#endif // defined(REINTERPRET_OUTPUT_AS_3D) + const bool cond_y = y == 0; + const bool cond_x = ((x + 1) * N0 >= N); + + // Convert and store output block + REPEAT_VAR_INIT_CONVERT(M0, VEC_DATA_TYPE(int, N0), c, res); // resN = CONVERT(cN, VEC_DATA_TYPE(int, N0)); + STORE_BLOCK_BOUNDARY_AWARE(M0, N0, int, res, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); +} +#endif // defined(GEMMLOWP_MM_NATIVE) + +#if defined(GEMMLOWP_MATRIX_A_REDUCTION) +/** OpenCL kernel used to compute the row-vectors of sums of all the entries in each row of Matrix A. + * It is also possible to multiply each reduced row by a scalar value, if SCALAR is passed at compile time. + * + * @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 + * @note The input data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=uchar) + * @note The data type for the accumulation must be passed at compile time using -DACC_DATA_TYPE (i.e. -DACC_DATA_TYPE=uint) + * @note In case of scaling the scalar value must be passed at compile time using -DSCALAR (e.g. -DSCALAR=3) + * + * @param[in] src_ptr Pointer to the source tensor. Supported data type: QASYMM8/QASYMM8_SIGNED/QSYMM8 + * @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); + + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) + sum_row_32 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))0; + ACC_DATA_TYPE sum_row = 0; + + __global const DATA_TYPE *matrix_a = (__global const DATA_TYPE *)(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 VEC_DATA_TYPE(DATA_TYPE, 16) a0 = vload16(0, matrix_a + i); + + sum_row_32 += CONVERT(a0.s0123, VEC_DATA_TYPE(ACC_DATA_TYPE, 4)) + CONVERT(a0.s4567, VEC_DATA_TYPE(ACC_DATA_TYPE, 4)) + CONVERT(a0.s89AB, VEC_DATA_TYPE(ACC_DATA_TYPE, 4)) + CONVERT(a0.sCDEF, + VEC_DATA_TYPE(ACC_DATA_TYPE, 4)); + } + + // This for loop performs the leftover accumulations + for(; i < COLS_A; ++i) + { + sum_row += (ACC_DATA_TYPE)matrix_a[i]; + } + + sum_row += sum_row_32.s0 + sum_row_32.s1 + sum_row_32.s2 + sum_row_32.s3; + +#if defined(SCALAR) + sum_row *= (int)SCALAR; +#endif // defined(SCALAR) + *((__global int *)dst.ptr) = (int)sum_row; +} +#endif // defined(GEMMLOWP_MATRIX_A_REDUCTION) + +#if defined(GEMMLOWP_MATRIX_A_REDUCTION_DOT8) +/** 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. + * It is also possible to multiply each reduced row by a scalar value, if SCALAR is passed at compile time. + * + * @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 + * @note The input data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=uchar) + * @note The data type for the accumulation must be passed at compile time using -DACC_DATA_TYPE (i.e. -DACC_DATA_TYPE=uint) + * @note In case of scaling the scalar value must be passed at compile time using -DSCALAR (e.g. -DSCALAR=3) + * + * @param[in] src_ptr Pointer to the source tensor. Supported data type: QASYMM8/QASYMM8_SIGNED/QSYMM8 + * @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); + + ACC_DATA_TYPE sum_row = 0; + + __global const DATA_TYPE *matrix_a = (__global const DATA_TYPE *)(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) + { + VEC_DATA_TYPE(DATA_TYPE, 16) + a0 = vload16(0, matrix_a + i); + + DOT_PRODUCT4_INTEGER8(DATA_TYPE, DATA_TYPE, DATA_TYPE, a0.s0123, (VEC_DATA_TYPE(DATA_TYPE, 4))(1), sum_row); + DOT_PRODUCT4_INTEGER8(DATA_TYPE, DATA_TYPE, DATA_TYPE, a0.s4567, (VEC_DATA_TYPE(DATA_TYPE, 4))(1), sum_row); + DOT_PRODUCT4_INTEGER8(DATA_TYPE, DATA_TYPE, DATA_TYPE, a0.s89AB, (VEC_DATA_TYPE(DATA_TYPE, 4))(1), sum_row); + DOT_PRODUCT4_INTEGER8(DATA_TYPE, DATA_TYPE, DATA_TYPE, a0.sCDEF, (VEC_DATA_TYPE(DATA_TYPE, 4))(1), sum_row); + + a0 = vload16(1, matrix_a + i); + + DOT_PRODUCT4_INTEGER8(DATA_TYPE, DATA_TYPE, DATA_TYPE, a0.s0123, (VEC_DATA_TYPE(DATA_TYPE, 4))(1), sum_row); + DOT_PRODUCT4_INTEGER8(DATA_TYPE, DATA_TYPE, DATA_TYPE, a0.s4567, (VEC_DATA_TYPE(DATA_TYPE, 4))(1), sum_row); + DOT_PRODUCT4_INTEGER8(DATA_TYPE, DATA_TYPE, DATA_TYPE, a0.s89AB, (VEC_DATA_TYPE(DATA_TYPE, 4))(1), sum_row); + DOT_PRODUCT4_INTEGER8(DATA_TYPE, DATA_TYPE, DATA_TYPE, a0.sCDEF, (VEC_DATA_TYPE(DATA_TYPE, 4))(1), sum_row); + } + + // This for loop performs the leftover accumulations + for(; i < COLS_A; ++i) + { + sum_row += (ACC_DATA_TYPE)matrix_a[i]; + } + +#if defined(SCALAR) + sum_row *= (int)SCALAR; +#endif // defined(SCALAR) + *((__global int *)dst.ptr) = (int)sum_row; +} +#endif // defined(GEMMLOWP_MATRIX_A_REDUCTION_DOT8) + +#if defined(GEMMLOWP_MATRIX_B_REDUCTION) +/** OpenCL kernel used to compute the row-vectors of sums of all the entries in each column of Matrix B. + * It is also possible to multiply each reduced column by a scalar value, if SCALAR is passed at compile time. + * + * @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 + * @note The input data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=uchar) + * @note The data type for the accumulation must be passed at compile time using -DACC_DATA_TYPE (i.e. -DACC_DATA_TYPE=uint) + * @note In case of scaling the scalar value must be passed at compile time using -DSCALAR (i.e. -DSCALAR=3) + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE + * + * @param[in] src_ptr Pointer to the source tensor. Supported data type: QASYMM8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL + * @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 + const uint x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0); + const uint y = get_global_id(1); + + __global const DATA_TYPE *matrix_b = (__global const DATA_TYPE *)(src_ptr + src_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + y * src_step_y + y * src_stride_z); + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x_offs * sizeof(int) + y * dst_stride_y; + + VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE) + sum_col_32 = (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))0; + + int i = 0; + // This for loop performs 4 accumulations + for(; i <= ((int)ROWS_B - 4); i += 4) + { + const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + b0 = VLOAD(VEC_SIZE)(0, matrix_b + 0 * src_stride_y); + const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + b1 = VLOAD(VEC_SIZE)(0, matrix_b + 1 * src_stride_y); + const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + b2 = VLOAD(VEC_SIZE)(0, matrix_b + 2 * src_stride_y); + const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + b3 = VLOAD(VEC_SIZE)(0, matrix_b + 3 * src_stride_y); + + sum_col_32 += CONVERT(b0, VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)) + CONVERT(b1, VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)) + CONVERT(b2, VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)) + CONVERT(b3, + VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)); + + matrix_b += 4 * src_stride_y; + } + + // This for loop perfoms the leftover accumulations + for(; i < (int)ROWS_B; ++i) + { + const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + b0 = VLOAD(VEC_SIZE)(0, matrix_b); + + sum_col_32 += CONVERT(b0, VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)); + + matrix_b += src_stride_y; + } + +#if defined(SCALAR) + sum_col_32 *= (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))SCALAR; +#endif // defined(SCALAR) + VEC_DATA_TYPE(int, VEC_SIZE) + res0 = CONVERT(sum_col_32, VEC_DATA_TYPE(int, VEC_SIZE)); + + STORE_VECTOR_SELECT(res, int, dst_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) +} +#endif // defined(GEMMLOWP_MATRIX_B_REDUCTION) + +#endif // defined(DATA_TYPE) && defined(ACC_DATA_TYPE) + +#if defined(K_OFFSET) && defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) + +#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE) + +/* Helper function used to calculate the offset contribution after matrix multiplication. + * + * This kernel takes a final int32 accumulator value (the output of matrix multiplication), + * 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 + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE + * + * @param[in] x max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0) + * @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 VEC_INT 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) +) +{ + VEC_INT a_offset_s32 = (VEC_INT)0; + VEC_INT b_offset_s32 = (VEC_INT)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 = VLOAD(VEC_SIZE)(0, (__global int *)(sum_col_addr + batch_id * sum_col_stride_y)); +#else // defined(SUM_COL_HAS_BATCHES) + a_offset_s32 = VLOAD(VEC_SIZE)(0, (__global int *)sum_col_addr); +#endif // defined(SUM_COL_HAS_BATCHES) + + a_offset_s32 *= (VEC_INT)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 = (VEC_INT) * (((__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 = (VEC_INT) * (((__global int *)(sum_row_addr + batch_id * sum_row_stride_y))); +#endif // defined(HEIGHT_INPUT3D) && defined(DEPTH_INPUT3D) + b_offset_s32 *= (VEC_INT)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); + + VEC_INT biases_values = VLOAD(VEC_SIZE)(0, (__global int *)bias_addr); + b_offset_s32 += (VEC_INT)biases_values; +#endif // defined(ADD_BIAS) + + return (VEC_INT)K_OFFSET + a_offset_s32 + b_offset_s32; +} + +#if defined(GEMMLOWP_OFFSET_CONTRIBUTION) +/* OpenCL kernel used to add the offset contribution after matrix multiplication. The computation is performed in-place + * + * This kernel takes a final int32 accumulator value (the output of matrix multiplication), + * 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 + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE + * + * 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 = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0); + const int y = get_global_id(1); + const int z = get_global_id(2); + + // Compute offset contribution + VEC_INT 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; + + VEC_INT in_s32_0 = VLOAD(VEC_SIZE)(0, (__global int *)mm_result_addr); + + // Add the offset terms to GEMM's result + in_s32_0 += offset_term_s32; + + // Store the result with the offset contribution + STORE_VECTOR_SELECT(in_s32_, int, mm_result_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) +} +#endif // defined(GEMMLOWP_OFFSET_CONTRIBUTION) + +#if defined(GEMMLOWP_OFFSET_CONTRIBUTION_QUANTIZE_DOWN) +/* 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/int8 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. + * - to the [-128..127] range and cast to QASYMM8_SIGNED. + * + * @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 The output datatype should be passed at compile time using -DOUTPUT_DATA_TYPE + * @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 + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE + * + * @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/QASYMM8_SIGNED + * @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 + * @param[in] result_multipliers_ptr (Optional) Pointer to the output multipliers vector for per-channel quantization. Supported data types: S32 + * @param[in] result_multipliers_stride_x (Optional) Stride of the output multipliers vector in X dimension (in bytes) + * @param[in] result_multipliers_step_x (Optional) output_multipliers_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] result_multipliers_offset_first_element_in_bytes (Optional) The offset of the first element in the output multipliers vector + * @param[in] result_shifts_ptr (Optional) Pointer to the output shifts vector for per-channel quantization. Supported data types: S32 + * @param[in] result_shifts_stride_x (Optional) Stride of the output shifts vector in X dimension (in bytes) + * @param[in] result_shifts_step_x (Optional) output_shifts_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] result_shifts_offset_first_element_in_bytes (Optional) The offset of the first element in the output shifts vector + */ +__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) +#if defined(PER_CHANNEL_QUANTIZATION) + , + VECTOR_DECLARATION(result_multipliers), + VECTOR_DECLARATION(result_shifts) +#endif // defined(PER_CHANNEL_QUANTIZATION) + ) +{ + const int x = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0); + 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 + VEC_INT 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; + + VEC_INT in_s32 = VLOAD(VEC_SIZE)(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 += (VEC_INT)RESULT_OFFSET; + + // Multiply by result_mult_int and shift +#if defined(PER_CHANNEL_QUANTIZATION) + __global uchar *result_multipliers_addr = result_multipliers_ptr + result_multipliers_offset_first_element_in_bytes + x * sizeof(int); + __global uchar *result_shifts_addr = result_shifts_ptr + result_shifts_offset_first_element_in_bytes + x * sizeof(int); + VEC_INT result_multipliers_values = VLOAD(VEC_SIZE)(0, (__global int *)result_multipliers_addr); + VEC_INT result_shifts_values = VLOAD(VEC_SIZE)(0, (__global int *)result_shifts_addr); + + in_s32 *= result_multipliers_values; + in_s32 >>= result_shifts_values; +#else // defined(PER_CHANNEL_QUANTIZATION) + in_s32 *= RESULT_MULTIPLIER; + + in_s32 >>= RESULT_SHIFT; +#endif // defined(PER_CHANNEL_QUANTIZATION) + + VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE) + res0 = CONVERT_SAT(in_s32, VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE)); + +#if defined(MIN_BOUND) + res0 = max(res0, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE))MIN_BOUND); +#endif // defined(MIN_BOUND) +#if defined(MAX_BOUND) + res0 = min(res0, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE))MAX_BOUND); +#endif // defined(MAX_BOUND) + + // Store the result + STORE_VECTOR_SELECT(res, OUTPUT_DATA_TYPE, dst_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) +} +#endif // defined(GEMMLOWP_OFFSET_CONTRIBUTION_QUANTIZE_DOWN) + +#if defined(GEMMLOWP_OFFSET_CONTRIBUTION_QUANTIZE_DOWN_FIXEDPOINT) +/* OpenCL kernel used to add the offset contribution after matrix multiplication and it quantizes down to uint8. + * + * This kernel takes a final int32 accumulator value (the output of matrix multiplication), 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/int8 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. + * - to the [-128..127] range and cast to QASYMM8_SIGNED. + * + * @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 The output datatype should be passed at compile time using -DOUTPUT_DATA_TYPE + * @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 + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE + * + * @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/QASYMM8_SIGNED + * @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 + * @param[in] result_multipliers_ptr (Optional) Pointer to the output multipliers vector for per-channel quantization. Supported data types: S32 + * @param[in] result_multipliers_stride_x (Optional) Stride of the output multipliers vector in X dimension (in bytes) + * @param[in] result_multipliers_step_x (Optional) output_multipliers_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] result_multipliers_offset_first_element_in_bytes (Optional) The offset of the first element in the output multipliers vector + * @param[in] result_shifts_ptr (Optional) Pointer to the output shifts vector for per-channel quantization. Supported data types: S32 + * @param[in] result_shifts_stride_x (Optional) Stride of the output shifts vector in X dimension (in bytes) + * @param[in] result_shifts_step_x (Optional) output_shifts_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] result_shifts_offset_first_element_in_bytes (Optional) The offset of the first element in the output shifts vector + */ +__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) +#if defined(PER_CHANNEL_QUANTIZATION) + , + VECTOR_DECLARATION(result_multipliers), + VECTOR_DECLARATION(result_shifts) +#endif // defined(PER_CHANNEL_QUANTIZATION) + ) +{ + const int x = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0); + const int y = get_global_id(1); + const int z = get_global_id(2); + + // Compute offset contribution + VEC_INT 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; + + VEC_INT in_s32 = VLOAD(VEC_SIZE)(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 +#if defined(PER_CHANNEL_QUANTIZATION) + __global uchar *result_multipliers_addr = result_multipliers_ptr + result_multipliers_offset_first_element_in_bytes + x * sizeof(int); + __global uchar *result_shifts_addr = result_shifts_ptr + result_shifts_offset_first_element_in_bytes + x * sizeof(int); + VEC_INT result_multipliers_values = VLOAD(VEC_SIZE)(0, (__global int *)result_multipliers_addr); + VEC_INT result_shifts_values = VLOAD(VEC_SIZE)(0, (__global int *)result_shifts_addr); + + VEC_INT in_s32_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(in_s32, result_multipliers_values, result_shifts_values, VEC_SIZE); + VEC_INT in_s32_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(in_s32, result_multipliers_values, result_shifts_values, VEC_SIZE); + in_s32 = select(in_s32_shift_lt0, in_s32_shift_gt0, result_shifts_values >= 0); +#else // defined(PER_CHANNEL_QUANTIZATION) + +#if RESULT_SHIFT < 0 + in_s32 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(in_s32, RESULT_MULTIPLIER, RESULT_SHIFT, VEC_SIZE); +#else // RESULT_SHIFT >= 0 + in_s32 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(in_s32, RESULT_MULTIPLIER, RESULT_SHIFT, VEC_SIZE); +#endif // RESULT_SHIFT < 0 + +#endif // defined(PER_CHANNEL_QUANTIZATION) + + // Add the offset terms to GEMM's result + in_s32 += (VEC_INT)RESULT_OFFSET; + + VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE) + res0 = CONVERT_SAT(in_s32, VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE)); + +#if defined(MIN_BOUND) + res0 = max(res0, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE))MIN_BOUND); +#endif // defined(MIN_BOUND) +#if defined(MAX_BOUND) + res0 = min(res0, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE))MAX_BOUND); +#endif // defined(MAX_BOUND) + + // Store the result + STORE_VECTOR_SELECT(res, OUTPUT_DATA_TYPE, dst_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) +} +#endif // defined(GEMMLOWP_OFFSET_CONTRIBUTION_QUANTIZE_DOWN_FIXEDPOINT) + +#undef VEC_INT + +#endif // defined(K_OFFSET) && defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) + +#if defined(GEMMLOWP_OUTPUT_STAGE_QUANTIZE_DOWN) +/** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8/QASYMM8_SIGNED + * + * This kernel takes a final int32 accumulator value and processes it to obtain the final QASYMM8/QASYMM8_SIGNED 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. + * -# - to the [-128..127] range and cast to QASYMM8_SIGNED. + * + * @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 The output datatype should be passed at compile time using -DOUTPUT_DATA_TYPE + * @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 + * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE + * + * @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/QASYMM8_SIGNED + * @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 = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0); + 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; + + VEC_DATA_TYPE(int, VEC_SIZE) + input_values = VLOAD(VEC_SIZE)(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); + + VEC_DATA_TYPE(int, VEC_SIZE) + biases_values = VLOAD(VEC_SIZE)(0, (__global int *)bias_addr); + input_values += biases_values; +#endif // defined(ADD_BIAS) + + // Add the offset terms to GEMM's result + input_values += (VEC_DATA_TYPE(int, VEC_SIZE))RESULT_OFFSET; + + // Multiply by result_mult_int and shift + input_values *= RESULT_MULT_INT; + +#if RESULT_SHIFT < 0 + input_values >>= -RESULT_SHIFT; +#else // RESULT_SHIFT >= 0 + input_values >>= RESULT_SHIFT; +#endif // RESULT_SHIFT < 0 + + VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE) + res0 = CONVERT_SAT(input_values, VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE)); + +#if defined(MIN_BOUND) + res0 = max(res0, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE))MIN_BOUND); +#endif // defined(MIN_BOUND) +#if defined(MAX_BOUND) + res0 = min(res0, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE))MAX_BOUND); +#endif // defined(MAX_BOUND) + + // Store the result + STORE_VECTOR_SELECT(res, OUTPUT_DATA_TYPE, dst_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) +} +#endif // defined(GEMMLOWP_OUTPUT_STAGE_QUANTIZE_DOWN) + +#if defined(GEMMLOWP_OUTPUT_STAGE_QUANTIZE_DOWN_FIXEDPOINT) +/** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8/QASYMM8_SIGNED + * + * This kernel takes a final int32 accumulator value (the output of matrix multiplication), and processes it to obtain the final QASYMM8/QASYMM8_SIGNED 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. + * - to the [-128..127] range and cast to QASYMM8_SIGNED. + * + * @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 The output datatype should be passed at compile time using -DOUTPUT_DATA_TYPE + * @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 + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE + * + * @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/QASYMM8_SIGNED + * @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 = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0); + 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; + + VEC_DATA_TYPE(int, VEC_SIZE) + input_values = VLOAD(VEC_SIZE)(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); + + VEC_DATA_TYPE(int, VEC_SIZE) + biases_values = VLOAD(VEC_SIZE)(0, (__global int *)bias_addr); + input_values += biases_values; +#endif // defined(ADD_BIAS) + + // Multiply by result_mult_int and shift +#if RESULT_SHIFT < 0 + input_values = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(input_values, RESULT_FIXEDPOINT_MULTIPLIER, RESULT_SHIFT, VEC_SIZE); +#else // RESULT_SHIFT >= 0 + input_values = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(input_values, RESULT_FIXEDPOINT_MULTIPLIER, RESULT_SHIFT, VEC_SIZE); +#endif // RESULT_SHIFT < 0 + + // Add the offset terms to GEMM's result + input_values += (VEC_DATA_TYPE(int, VEC_SIZE))RESULT_OFFSET_AFTER_SHIFT; + + VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE) + res0 = CONVERT_SAT(input_values, VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE)); + +#if defined(MIN_BOUND) + res0 = max(res0, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE))MIN_BOUND); +#endif // defined(MIN_BOUND) +#if defined(MAX_BOUND) + res0 = min(res0, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE))MAX_BOUND); +#endif // defined(MAX_BOUND) + + // Store the result + STORE_VECTOR_SELECT(res, OUTPUT_DATA_TYPE, dst_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) +} +#endif // defined(GEMMLOWP_OUTPUT_STAGE_QUANTIZE_DOWN_FIXEDPOINT) + +#if defined(GEMMLOWP_OUTPUT_STAGE_QUANTIZE_DOWN_FIXEDPOINT_QSYMM16) +/** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QSYMM16 + * + * This kernel takes a final int32 accumulator value (the output of matrix multiplication), and processes it to obtain the final QSYMM16 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 [-32768..32767] range and cast to QSYMM16. + * + * @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -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 + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE + * + * @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: QSYMM16 + * @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_qsymm16(TENSOR3D_DECLARATION(src), +#if defined(ADD_BIAS) + VECTOR_DECLARATION(biases), +#endif // defined(ADD_BIAS) + TENSOR3D_DECLARATION(dst)) +{ + // Compute source and destination addresses + int x = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0); + 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 * sizeof(short) + y * dst_stride_y + z * dst_stride_z; + + VEC_DATA_TYPE(int, VEC_SIZE) + input_values = VLOAD(VEC_SIZE)(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); + + VEC_DATA_TYPE(int, VEC_SIZE) + biases_values = VLOAD(VEC_SIZE)(0, (__global int *)bias_addr); + input_values += biases_values; +#endif // defined(ADD_BIAS) + + // Multiply by result_mult_int and shift +#if RESULT_SHIFT < 0 + input_values = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(input_values, RESULT_FIXEDPOINT_MULTIPLIER, RESULT_SHIFT, VEC_SIZE); +#else // RESULT_SHIFT >= 0 + input_values = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(input_values, RESULT_FIXEDPOINT_MULTIPLIER, RESULT_SHIFT, VEC_SIZE); +#endif // RESULT_SHIFT < 0 + + VEC_DATA_TYPE(short, VEC_SIZE) + res0 = CONVERT_SAT(input_values, VEC_DATA_TYPE(short, VEC_SIZE)); + +#if defined(MIN_BOUND) + res0 = max(res0, (VEC_DATA_TYPE(short, VEC_SIZE))MIN_BOUND); +#endif // defined(MIN_BOUND) +#if defined(MAX_BOUND) + res0 = min(res0, (VEC_DATA_TYPE(short, VEC_SIZE))MAX_BOUND); +#endif // defined(MAX_BOUND) + + // Store the result + STORE_VECTOR_SELECT(res, short, dst_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) +} +#endif // defined(GEMMLOWP_OUTPUT_STAGE_QUANTIZE_DOWN_FIXEDPOINT_QSYMM16) + +#if defined(GEMMLOWP_OUTPUT_STAGE_QUANTIZE_DOWN_FLOAT) +/** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8/QASYMM8_SIGNED + * + * This kernel takes a final int32 accumulator value (the output of matrix multiplication), and processes it to obtain the final QASYMM8/QASYMM8_SIGNED 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. + * - to the [-128..127] range and cast to QASYMM8_SIGNED. + * + * @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 The output datatype should be passed at compile time using -DOUTPUT_DATA_TYPE + * @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 + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE + * + * @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 = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0); + 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; + + VEC_DATA_TYPE(int, VEC_SIZE) + input_values = VLOAD(VEC_SIZE)(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); + + VEC_DATA_TYPE(int, VEC_SIZE) + biases_values = VLOAD(VEC_SIZE)(0, (__global int *)bias_addr); + input_values += (VEC_DATA_TYPE(int, VEC_SIZE))biases_values; +#endif // defined(ADD_BIAS) + + // Convert to float + VEC_DATA_TYPE(float, VEC_SIZE) + input_values_f = CONVERT(input_values, VEC_DATA_TYPE(float, VEC_SIZE)); + input_values_f = round(input_values_f * (float)REAL_MULTIPLIER + (float)OUTPUT_OFFSET); + + VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE) + res0 = CONVERT_SAT(input_values_f, VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE)); + +#if defined(MIN_BOUND) + res0 = max(res0, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE))MIN_BOUND); +#endif // defined(MIN_BOUND) +#if defined(MAX_BOUND) + res0 = min(res0, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, VEC_SIZE))MAX_BOUND); +#endif // defined(MAX_BOUND) + + // Store the result + STORE_VECTOR_SELECT(res, OUTPUT_DATA_TYPE, dst_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) +} +#endif // defined(GEMMLOWP_OUTPUT_STAGE_QUANTIZE_DOWN_FLOAT) |