diff options
author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2019-03-11 16:07:12 +0000 |
---|---|---|
committer | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2019-04-01 11:28:12 +0000 |
commit | 62251f71792c06dbe4c9d1985816ba15bcad14e4 (patch) | |
tree | d9d3778a2431486bffbcc80bd2cc51f3000ef116 /src/core/CL/cl_kernels/gemmlowp.cl | |
parent | b4a44ff3aa98d2b51f1621a7525db3f81108a1bd (diff) | |
download | ComputeLibrary-62251f71792c06dbe4c9d1985816ba15bcad14e4.tar.gz |
COMPMID-2002: Implement CLGEMMLowpMatrixMultiplyReshapedOnlyRHS - Transposed
Change-Id: I3907d151107766dc34749fe5710d7219e810b39f
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/875
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Diffstat (limited to 'src/core/CL/cl_kernels/gemmlowp.cl')
-rw-r--r-- | src/core/CL/cl_kernels/gemmlowp.cl | 549 |
1 files changed, 547 insertions, 2 deletions
diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl index 8dc22d7d56..52ce0f1ed0 100644 --- a/src/core/CL/cl_kernels/gemmlowp.cl +++ b/src/core/CL/cl_kernels/gemmlowp.cl @@ -2099,7 +2099,7 @@ __kernel void gemmlowp_mm_bifrost_dot8(IMAGE_DECLARATION(src0), #error "N0 value not supported" #endif // N0 conditions -/** This OpenCL kernel computes the matrix multiplication between 2 matrices. +/** This OpenCL kernel computes the matrix multiplication between 2 matrices with QASYMM 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 * @@ -2114,6 +2114,8 @@ __kernel void gemmlowp_mm_bifrost_dot8(IMAGE_DECLARATION(src0), * - 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 @@ -2432,7 +2434,7 @@ __kernel void gemmlowp_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs), } #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) -/** This OpenCL kernel computes the matrix multiplication between 2 matrices unsing the dot8 instruction. +/** This OpenCL kernel computes the matrix multiplication between 2 matrices with QASYMM8 data type using the dot8 instruction. * 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 * @@ -2521,6 +2523,549 @@ __kernel void gemmlowp_mm_reshaped_lhs_nt_rhs_t_dot8(IMAGE_DECLARATION(lhs), #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) #endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(K) +#if defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(K) + +#define CONCAT(a, b) a##b + +#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) + +#define ARM_DOT1(a, b, c) \ + ({ \ + ARM_DOT((uchar4)(a, (uchar3)0), (uchar4)(b, (uchar3)0), c); \ + }) +#define ARM_DOT2(a, b, c) \ + ({ \ + ARM_DOT((uchar4)(a, (uchar2)0), (uchar4)(b, (uchar2)0), c); \ + }) +#define ARM_DOT3(a, b, c) \ + ({ \ + ARM_DOT((uchar4)(a, (uchar)0), (uchar4)(b, (uchar)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) + +#define ARM_DOT1(a, b, c) \ + ({ \ + c += (uint)a.s0 * b.s0; \ + }) +#define ARM_DOT2(a, b, c) \ + ({ \ + ARM_DOT1(a, b, c); \ + c += (uint)a.s1 * b.s1; \ + }) +#define ARM_DOT3(a, b, c) \ + ({ \ + ARM_DOT2(a, b, c); \ + c += (uint)a.s2 * b.s2; \ + }) +#define ARM_DOT4(a, b, c) \ + ({ \ + ARM_DOT3(a, b, c); \ + c += (uint)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) + +#if N0 == 2 +#define ARM_DOT_K0XN0(k0, a, b, c) \ + ({ \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##0), (c.s0)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##1), (c.s1)); \ + }) +#elif N0 == 3 // N0 == 3 +#define ARM_DOT_K0XN0(k0, a, b, c) \ + ({ \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##0), (c.s0)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##1), (c.s1)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##2), (c.s2)); \ + }) +#elif N0 == 4 // N0 == 4 +#define ARM_DOT_K0XN0(k0, a, b, c) \ + ({ \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##0), (c.s0)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##1), (c.s1)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##2), (c.s2)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##3), (c.s3)); \ + }) +#elif N0 == 8 // N0 == 8 +#define ARM_DOT_K0XN0(k0, a, b, c) \ + ({ \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##0), (c.s0)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##1), (c.s1)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##2), (c.s2)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##3), (c.s3)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##4), (c.s4)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##5), (c.s5)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##6), (c.s6)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##7), (c.s7)); \ + }) +#elif N0 == 16 // N0 == 16 +#define ARM_DOT_K0XN0(k0, a, b, c) \ + ({ \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##0), (c.s0)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##1), (c.s1)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##2), (c.s2)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##3), (c.s3)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##4), (c.s4)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##5), (c.s5)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##6), (c.s6)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##7), (c.s7)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##8), (c.s8)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##9), (c.s9)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##A), (c.sA)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##B), (c.sB)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##C), (c.sC)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##D), (c.sD)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##E), (c.sE)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##F), (c.sF)); \ + }) +#else // N0 not supported +#error "N0 value not supported" +#endif // N0 conditions + +/** This OpenCL kernel computes the matrix multiplication between 2 matrices. + * The LHS matrix is NOT reshaped + * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed + * + * @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 + * + * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F16/F32 + * @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) + */ +__kernel void 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 + ) +{ + // 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); + + // Compute LHS matrix address + uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; + + // Compute RHS matrix address + uint rhs_offset = 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_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, zin, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; + +#if defined(REINTERPRET_INPUT_AS_3D) + // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension + // in order to take into account the presence of possible cross plane paddings + // + // | | + // | plane0 | + // | | + // |__________________| + // |******************| + // | cross_plane_pad | + // |******************| + // | | + // | plane1 | + // | | + // |__________________| + + // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + zin0 = (0 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zin0 = min((uint)(DEPTH_GEMM3D - 1), zin0); + zin0 *= (lhs_cross_plane_pad * lhs_stride_y); +#if M0 > 1 + zin1 = (1 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zin1 = min((uint)(DEPTH_GEMM3D - 1), zin1); + zin1 *= (lhs_cross_plane_pad * lhs_stride_y); +#endif // M0 > 1 +#if M0 > 2 + zin2 = (2 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zin2 = min((uint)(DEPTH_GEMM3D - 1), zin2); + zin2 *= (lhs_cross_plane_pad * lhs_stride_y); +#endif // M0 > 2 +#if M0 > 3 + zin3 = (3 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zin3 = min((uint)(DEPTH_GEMM3D - 1), zin3); + zin3 *= (lhs_cross_plane_pad * lhs_stride_y); +#endif // M0 > 3 +#if M0 > 4 + zin4 = (4 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zin4 = min((uint)(DEPTH_GEMM3D - 1), zin4); + zin4 *= (lhs_cross_plane_pad * lhs_stride_y); +#endif // M0 > 4 +#if M0 > 5 + zin5 = (5 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zin5 = min((uint)(DEPTH_GEMM3D - 1), zin5); + zin5 *= (lhs_cross_plane_pad * lhs_stride_y); +#endif // M0 > 5 +#if M0 > 6 + zin6 = (6 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zin6 = min((uint)(DEPTH_GEMM3D - 1), zin6); + zin6 *= (lhs_cross_plane_pad * lhs_stride_y); +#endif // M0 > 6 +#if M0 > 7 + zin7 = (7 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zin7 = min((uint)(DEPTH_GEMM3D - 1), zout7); + zin7 *= (lhs_cross_plane_pad * lhs_stride_y); +#endif // M0 > 7 + + // 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(uint, N0), c, 0); //VEC_DATA_TYPE(uint, N0) c0=0,c1=0,c2=0,... c(N0-1)=0; + + for(int i = 0; i < K; i += K0) + { + // Supported cases (M0, K0): + // 1,2 - 1,3 - 1,4 - 1,8 - 1,16 + // 2,2 - 2,3 - 2,4 - 2,8 - 2,16 + // 3,2 - 3,3 - 3,4 - 3,8 - 3,16 + // 4,2 - 4,3 - 4,4 - 4,8 - 4,16 + // 5,2 - 5,3 - 5,4 - 5,8 - 5,16 + // 6,2 - 6,3 - 6,4 - 6,8 - 6,16 + // 7,2 - 7,3 - 7,4 - 7,8 - 7,16 + // 8,2 - 8,3 - 8,4 - 8,8 - 8,16 + // Load values from LHS matrix + VEC_DATA_TYPE(uchar, K0) + a0 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 0 * lhs_stride_y + zin0); +#if M0 > 1 + VEC_DATA_TYPE(uchar, K0) + a1 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1); +#endif // M0 > 1 +#if M0 > 2 + VEC_DATA_TYPE(uchar, K0) + a2 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2); +#endif // M0 > 2 +#if M0 > 3 + VEC_DATA_TYPE(uchar, K0) + a3 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3); +#endif // M0 > 3 +#if M0 > 4 + VEC_DATA_TYPE(uchar, K0) + a4 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4); +#endif // M0 > 4 +#if M0 > 5 + VEC_DATA_TYPE(uchar, K0) + a5 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5); +#endif // M0 > 5 +#if M0 > 6 + VEC_DATA_TYPE(uchar, K0) + a6 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6); +#endif // M0 > 6 +#if M0 > 7 + VEC_DATA_TYPE(uchar, K0) + a7 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7); +#endif // M0 > 7 + + // Load values from RHS matrix + VEC_DATA_TYPE(uchar, K0) + b0 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 0 * RHS_STEP_X); + VEC_DATA_TYPE(uchar, K0) + b1 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 1 * RHS_STEP_X); +#if N0 > 2 + VEC_DATA_TYPE(uchar, K0) + b2 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 2 * RHS_STEP_X); +#endif // N0 > 2 +#if N0 > 3 + VEC_DATA_TYPE(uchar, K0) + b3 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 3 * RHS_STEP_X); +#endif // N0 > 3 +#if N0 > 4 + VEC_DATA_TYPE(uchar, K0) + b4 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 4 * RHS_STEP_X); + VEC_DATA_TYPE(uchar, K0) + b5 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 5 * RHS_STEP_X); + VEC_DATA_TYPE(uchar, K0) + b6 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 6 * RHS_STEP_X); + VEC_DATA_TYPE(uchar, K0) + b7 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 7 * RHS_STEP_X); +#endif // N0 > 4 +#if N0 > 8 + VEC_DATA_TYPE(uchar, K0) + b8 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 8 * RHS_STEP_X); + VEC_DATA_TYPE(uchar, K0) + b9 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 9 * RHS_STEP_X); + VEC_DATA_TYPE(uchar, K0) + bA = VLOAD(K0)(0, rhs_ptr + rhs_offset + 10 * RHS_STEP_X); + VEC_DATA_TYPE(uchar, K0) + bB = VLOAD(K0)(0, rhs_ptr + rhs_offset + 11 * RHS_STEP_X); + VEC_DATA_TYPE(uchar, K0) + bC = VLOAD(K0)(0, rhs_ptr + rhs_offset + 12 * RHS_STEP_X); + VEC_DATA_TYPE(uchar, K0) + bD = VLOAD(K0)(0, rhs_ptr + rhs_offset + 13 * RHS_STEP_X); + VEC_DATA_TYPE(uchar, K0) + bE = VLOAD(K0)(0, rhs_ptr + rhs_offset + 14 * RHS_STEP_X); + VEC_DATA_TYPE(uchar, K0) + bF = VLOAD(K0)(0, rhs_ptr + rhs_offset + 15 * RHS_STEP_X); +#endif // N0 > 8 + + // Accumulate + ARM_DOT_K0XN0(K0, a0, b, c0); +#if M0 > 1 + ARM_DOT_K0XN0(K0, a1, b, c1); +#endif // M0 > 1 +#if M0 > 2 + ARM_DOT_K0XN0(K0, a2, b, c2); +#endif // M0 > 2 +#if M0 > 3 + ARM_DOT_K0XN0(K0, a3, b, c3); +#endif // M0 > 3 +#if M0 > 4 + ARM_DOT_K0XN0(K0, a4, b, c4); +#endif // M0 > 4 +#if M0 > 5 + ARM_DOT_K0XN0(K0, a5, b, c5); +#endif // M0 > 5 +#if M0 > 6 + ARM_DOT_K0XN0(K0, a6, b, c6); +#endif // M0 > 6 +#if M0 > 7 + ARM_DOT_K0XN0(K0, a7, b, c7); +#endif // M0 > 7 + + lhs_offset += K0; + rhs_offset += 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) + // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension + // in order to take into account the presence of possible cross plane paddings + // + // | | + // | plane0 | + // | | + // |__________________| + // |******************| + // | cross_plane_pad | + // |******************| + // | | + // | plane1 | + // | | + // |__________________| + + // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + zout0 = (0 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zout0 = min((uint)(DEPTH_GEMM3D - 1), zout0); + zout0 *= (dst_cross_plane_pad * dst_stride_y); +#if M0 > 1 + zout1 = (1 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zout1 = min((uint)(DEPTH_GEMM3D - 1), zout1); + zout1 *= (dst_cross_plane_pad * dst_stride_y); +#endif // M0 > 1 +#if M0 > 2 + zout2 = (2 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zout2 = min((uint)(DEPTH_GEMM3D - 1), zout2); + zout2 *= (dst_cross_plane_pad * dst_stride_y); +#endif // M0 > 2 +#if M0 > 3 + zout3 = (3 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zout3 = min((uint)(DEPTH_GEMM3D - 1), zout3); + zout3 *= (dst_cross_plane_pad * dst_stride_y); +#endif // M0 > 3 +#if M0 > 4 + zout4 = (4 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zout4 = min((uint)(DEPTH_GEMM3D - 1), zout4); + zout4 *= (dst_cross_plane_pad * dst_stride_y); +#endif // M0 > 4 +#if M0 > 5 + zout5 = (5 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zout5 = min((uint)(DEPTH_GEMM3D - 1), zout5); + zout5 *= (dst_cross_plane_pad * dst_stride_y); +#endif // M0 > 5 +#if M0 > 6 + zout6 = (6 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zout6 = min((uint)(DEPTH_GEMM3D - 1), zout6); + zout6 *= (dst_cross_plane_pad * dst_stride_y); +#endif // M0 > 6 +#if M0 > 7 + zout7 = (7 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; + zout7 = min((uint)(DEPTH_GEMM3D - 1), zout7); + zout7 *= (dst_cross_plane_pad * dst_stride_y); +#endif // M0 > 7 + + // 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) + + // Store output block + VSTORE(N0) + (CONVERT_SAT(c0, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 0 * dst_stride_y + zout0)); +#if M0 > 1 + VSTORE(N0) + (CONVERT_SAT(c1, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 1 * dst_stride_y + zout1)); +#endif // M0 > 1 +#if M0 > 2 + VSTORE(N0) + (CONVERT_SAT(c2, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 2 * dst_stride_y + zout2)); +#endif // M0 > 2 +#if M0 > 3 + VSTORE(N0) + (CONVERT_SAT(c3, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 3 * dst_stride_y + zout3)); +#endif // M0 > 3 +#if M0 > 4 + VSTORE(N0) + (CONVERT_SAT(c4, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 4 * dst_stride_y + zout4)); +#endif // M0 > 4 +#if M0 > 5 + VSTORE(N0) + (CONVERT_SAT(c5, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 5 * dst_stride_y + zout5)); +#endif // M0 > 5 +#if M0 > 6 + VSTORE(N0) + (CONVERT_SAT(c6, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 6 * dst_stride_y + zout6)); +#endif // M0 > 6 +#if M0 > 7 + VSTORE(N0) + (CONVERT_SAT(c7, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 7 * dst_stride_y + zout7)); +#endif // M0 > 7 + +#undef RHS_BLOCK_SIZE +#undef RHS_OFFSET_X +#undef RHS_STEP_X +} +#endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(K) + #if defined(COLS_A) /** OpenCL kernel used to compute the row-vectors of sums of all the entries in each row of Matrix A. * |