diff options
Diffstat (limited to 'src/core/CL/cl_kernels/gemm.cl')
| -rw-r--r-- | src/core/CL/cl_kernels/gemm.cl | 705 |
1 files changed, 4 insertions, 701 deletions
diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl index e969e847d7..f75161ca0a 100644 --- a/src/core/CL/cl_kernels/gemm.cl +++ b/src/core/CL/cl_kernels/gemm.cl @@ -23,10 +23,6 @@ */ #include "helpers.h" -#ifdef FIXED_POINT_POSITION -#include "fixed_point.h" -#endif // FIXED_POINT_POSITION - #if defined(TRANSPOSE_W) && defined(MULT_TRANSPOSE1XW_WIDTH) #if ELEMENT_SIZE == 1 @@ -44,7 +40,7 @@ * @note The transposition width must be passed at compile time using -DTRANSPOSE_W (i.e. -DTRANSPOSE_W) * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (i.e. -DMULT_TRANSPOSE1XW_WIDTH=2) * - * @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32 + * @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 * @param[in] src_stride_x Stride of the source matrix 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 matrix in Y dimension (in bytes) @@ -93,7 +89,7 @@ __kernel void gemm_transpose1xW(TENSOR3D_DECLARATION(src), * @note The data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=float) * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2) * - * @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32 + * @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 * @param[in] src_stride_x Stride of the source matrix 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 matrix in Y dimension (in bytes) @@ -1085,248 +1081,6 @@ __kernel void gemm_mm_interleaved_transposed_f16_bifrost(IMAGE_DECLARATION(src0) #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) -#if defined(FIXED_POINT_POSITION) -/** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1) in 8 bit fixed point precision - * Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_8bit and @ref gemm_transpose1x16 before running the matrix multiplication - * - * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA - * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (i.e. -DMULT_TRANSPOSE1XW_WIDTH=2) - * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2) - * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16) - * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16]) - * @note:ALPHA must be passed in 8 bit fixed point format - * - * @param[in] src0_ptr Pointer to the source matrix. Supported data types: QS8 - * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr - * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_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] src0_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - */ -__kernel void gemm_mm_interleaved_transposed_qs8(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst), - uint src0_stride_z, - uint src1_stride_z, - uint dst_stride_z) -{ - int x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH; - int y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT; - int z = get_global_id(2); - - // Offset - const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4; - const int offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 16; - - // src_addr_a = address of matrix A - // src_addr_b = address of matrix B - int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; - int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; - -#if defined(MATRIX_B_DEPTH) - // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 - src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; -#else // defined(MATRIX_B_DEPTH) - src1_addr_in_bytes += z * src1_stride_z; -#endif // defined(MATRIX_B_DEPTH) - - __global char *src_addr_a = (__global char *)(src0_ptr + src0_addr_in_bytes); - __global char *src_addr_b = (__global char *)(src1_ptr + src1_addr_in_bytes); - - // Compute end row address for matrix B - __global char *src_end_addr_b = src_addr_b + COLS_B; - - src_addr_a += offset_row_a; - src_addr_b += offset_row_b; - - // Reset accumulators - short8 c00 = 0.0f; - short8 c10 = 0.0f; - short8 c20 = 0.0f; - short8 c30 = 0.0f; - short8 c01 = 0.0f; - short8 c11 = 0.0f; - short8 c21 = 0.0f; - short8 c31 = 0.0f; - - // This for loop performs 1 accumulation for each iteration - for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 16 * MULT_TRANSPOSE1XW_WIDTH) - { - // Load values from matrix A (interleaved) and matrix B (transposed) - char4 a0 = vload4(0, src_addr_a); - char16 b0 = vload16(0, src_addr_b); - - c00 = mlal_sat_qs8x8(c00, (char8)a0.s0, b0.s01234567, FIXED_POINT_POSITION); - c10 = mlal_sat_qs8x8(c10, (char8)a0.s1, b0.s01234567, FIXED_POINT_POSITION); - c20 = mlal_sat_qs8x8(c20, (char8)a0.s2, b0.s01234567, FIXED_POINT_POSITION); - c30 = mlal_sat_qs8x8(c30, (char8)a0.s3, b0.s01234567, FIXED_POINT_POSITION); - - c01 = mlal_sat_qs8x8(c01, (char8)a0.s0, b0.s89ABCDEF, FIXED_POINT_POSITION); - c11 = mlal_sat_qs8x8(c11, (char8)a0.s1, b0.s89ABCDEF, FIXED_POINT_POSITION); - c21 = mlal_sat_qs8x8(c21, (char8)a0.s2, b0.s89ABCDEF, FIXED_POINT_POSITION); - c31 = mlal_sat_qs8x8(c31, (char8)a0.s3, b0.s89ABCDEF, FIXED_POINT_POSITION); - } - - // Compute destination address - Image dst = CONVERT_TO_IMAGE_STRUCT(dst); - - // Multiply by the weight of matrix product - char16 c00_qs8 = convert_char16_sat((short16)(c00, c01)); - char16 c10_qs8 = convert_char16_sat((short16)(c10, c11)); - char16 c20_qs8 = convert_char16_sat((short16)(c20, c21)); - char16 c30_qs8 = convert_char16_sat((short16)(c30, c31)); - -#if defined(ALPHA) - c00_qs8 = mul_sat_qs8x16(c00_qs8, (char16)ALPHA, FIXED_POINT_POSITION); - c10_qs8 = mul_sat_qs8x16(c10_qs8, (char16)ALPHA, FIXED_POINT_POSITION); - c20_qs8 = mul_sat_qs8x16(c20_qs8, (char16)ALPHA, FIXED_POINT_POSITION); - c30_qs8 = mul_sat_qs8x16(c30_qs8, (char16)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - - // Compute dst address - __global uchar *dst_addr = offset(&dst, 0, 0); - - // Add offset for batched GEMM - dst_addr += z * dst_stride_z; - - // Store 16x4 block - vstore16(c00_qs8, 0, (__global char *)(dst_addr + 0 * dst_stride_y)); - vstore16(c10_qs8, 0, (__global char *)(dst_addr + 1 * dst_stride_y)); - vstore16(c20_qs8, 0, (__global char *)(dst_addr + 2 * dst_stride_y)); - vstore16(c30_qs8, 0, (__global char *)(dst_addr + 3 * dst_stride_y)); -} - -/** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1) in 16 bit fixed point precision - * Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_16bit and @ref gemm_transpose1x8 before running the matrix multiplication - * - * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA - * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (i.e. -DMULT_TRANSPOSE1XW_WIDTH=2) - * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2) - * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16) - * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16]) - * @note:ALPHA must be passed in 16 bit fixed point format - * - * @param[in] src0_ptr Pointer to the source matrix. Supported data types: QS16 - * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr - * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_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] src0_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - */ -__kernel void gemm_mm_interleaved_transposed_qs16(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst), - uint src0_stride_z, - uint src1_stride_z, - uint dst_stride_z) -{ - int x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH; - int y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT; - int z = get_global_id(2); - - // Offset - const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4; - const int offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 8; - - // src_addr_a = address of matrix A - // src_addr_b = address of matrix B - int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; - int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; - -#if defined(MATRIX_B_DEPTH) - // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 - src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; -#else // defined(MATRIX_B_DEPTH) - src1_addr_in_bytes += z * src1_stride_z; -#endif // defined(MATRIX_B_DEPTH) - - __global short *src_addr_a = (__global short *)(src0_ptr + src0_addr_in_bytes); - __global short *src_addr_b = (__global short *)(src1_ptr + src1_addr_in_bytes); - - // Compute end row address for matrix B - __global short *src_end_addr_b = src_addr_b + COLS_B; - - src_addr_a += offset_row_a; - src_addr_b += offset_row_b; - - // Reset accumulators - int8 c00 = 0.0f; - int8 c10 = 0.0f; - int8 c20 = 0.0f; - int8 c30 = 0.0f; - - // This for loop performs 1 accumulation for each iteration - for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH) - { - /* Load values from matrix A (interleaved) and matrix B (transposed) */ - short4 a0 = vload4(0, src_addr_a); - short8 b0 = vload8(0, src_addr_b); - - c00 = mlal_sat_qs16x8(c00, (short8)a0.s0, b0, FIXED_POINT_POSITION); - c10 = mlal_sat_qs16x8(c10, (short8)a0.s1, b0, FIXED_POINT_POSITION); - c20 = mlal_sat_qs16x8(c20, (short8)a0.s2, b0, FIXED_POINT_POSITION); - c30 = mlal_sat_qs16x8(c30, (short8)a0.s3, b0, FIXED_POINT_POSITION); - } - - // Compute destination address - Image dst = CONVERT_TO_IMAGE_STRUCT(dst); - - // Multiply by the weight of matrix product - short8 c00_qs16 = convert_short8_sat(c00); - short8 c10_qs16 = convert_short8_sat(c10); - short8 c20_qs16 = convert_short8_sat(c20); - short8 c30_qs16 = convert_short8_sat(c30); - -#if defined(ALPHA) - c00_qs16 = mul_sat_qs16x8(c00_qs16, (short8)ALPHA, FIXED_POINT_POSITION); - c10_qs16 = mul_sat_qs16x8(c10_qs16, (short8)ALPHA, FIXED_POINT_POSITION); - c20_qs16 = mul_sat_qs16x8(c20_qs16, (short8)ALPHA, FIXED_POINT_POSITION); - c30_qs16 = mul_sat_qs16x8(c30_qs16, (short8)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - - // Compute dst address - __global uchar *dst_addr = offset(&dst, 0, 0); - - // Add offset for batched GEMM - dst_addr += z * dst_stride_z; - - // Store 8x4 block - vstore8(c00_qs16, 0, (__global short *)(dst_addr + 0 * dst_stride_y)); - vstore8(c10_qs16, 0, (__global short *)(dst_addr + 1 * dst_stride_y)); - vstore8(c20_qs16, 0, (__global short *)(dst_addr + 2 * dst_stride_y)); - vstore8(c30_qs16, 0, (__global short *)(dst_addr + 3 * dst_stride_y)); -} -#endif // defined(FIXED_POINT_POSITION) #endif // defined(COLS_B) && defined(MULT_TRANSPOSE1XW_WIDTH) && defined(MULT_INTERLEAVE4X4_HEIGHT) #if defined(COLS_A) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && (NUM_ELEMS_PROCESSED_PER_THREAD_Y) @@ -2543,365 +2297,6 @@ __kernel void gemm_mm_floating_point_f16_bifrost(IMAGE_DECLARATION(src0), } #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) -#if defined(FIXED_POINT_POSITION) -/** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped - * - * @note This OpenCL kernel works with fixed point data types QS8 - * @note The number of elements processed along the x and y directions must be passed at compile time using -DNUM_ELEMS_PROCESSED_PER_THREAD_X and -DNUM_ELEMS_PROCESSED_PER_THREAD_Y - * @note The number matrix A columns, the number of elements processed per thread along the Y direction and the alpha's value need to be passed at compile time using -DCOLS_A, -DNUM_ELEMS_PROCESSED_PER_THREAD_Y and -DALPHA - * @note The fixed point position need to be passed at compile time using -DFIXED_POINT_POSITION - * @note The optional alpha value must be passed in 8 bit fixed point format using -DALPHA - * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16) - * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16]) - * - * @param[in] src0_ptr Pointer to the source matrix. Supported data types: QS8/QS16 - * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr - * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - */ -__kernel void gemm_mm_qs8(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst), - uint src0_stride_z, - uint src1_stride_z, - uint dst_stride_z) -{ - int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; - - // Compute starting address for matrix A and Matrix B - int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); - - // Update address for the matrix A - src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; - - // Update address for the matrix B - src_addr.s1 += idx * sizeof(char); - - // Add offset for batched GEMM - src_addr.s0 += get_global_id(2) * src0_stride_z; - -#if defined(MATRIX_B_DEPTH) - // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 - src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; -#else // defined(MATRIX_B_DEPTH) - src_addr.s1 += get_global_id(2) * src1_stride_z; -#endif // defined(MATRIX_B_DEPTH) - - int end_row_vec_a = src_addr.s0 + (COLS_A * sizeof(char)); - - short8 acc00 = 0; - short8 acc01 = 0; -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - short8 acc10 = 0; - short8 acc11 = 0; -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - short8 acc20 = 0; - short8 acc21 = 0; -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - short8 acc30 = 0; - short8 acc31 = 0; -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - - // This for loop performs 4 accumulations per iteration - for(; src_addr.s0 <= (end_row_vec_a - 2); src_addr += (int2)(2, 2 * src1_stride_y)) - { - char2 a0 = vload2(0, (__global char *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - char2 a1 = vload2(0, (__global char *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - char2 a2 = vload2(0, (__global char *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - char2 a3 = vload2(0, (__global char *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - char16 b0 = vload16(0, (__global char *)(src1_ptr + src_addr.s1 + 0 * src1_stride_y)); - char16 b1 = vload16(0, (__global char *)(src1_ptr + src_addr.s1 + 1 * src1_stride_y)); - - acc00 = mlal_sat_qs8x8(acc00, (char8)a0.s0, b0.s01234567, FIXED_POINT_POSITION); - acc00 = mlal_sat_qs8x8(acc00, (char8)a0.s1, b1.s01234567, FIXED_POINT_POSITION); - acc01 = mlal_sat_qs8x8(acc01, (char8)a0.s0, b0.s89ABCDEF, FIXED_POINT_POSITION); - acc01 = mlal_sat_qs8x8(acc01, (char8)a0.s1, b1.s89ABCDEF, FIXED_POINT_POSITION); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - acc10 = mlal_sat_qs8x8(acc10, (char8)a1.s0, b0.s01234567, FIXED_POINT_POSITION); - acc10 = mlal_sat_qs8x8(acc10, (char8)a1.s1, b1.s01234567, FIXED_POINT_POSITION); - acc11 = mlal_sat_qs8x8(acc11, (char8)a1.s0, b0.s89ABCDEF, FIXED_POINT_POSITION); - acc11 = mlal_sat_qs8x8(acc11, (char8)a1.s1, b1.s89ABCDEF, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - acc20 = mlal_sat_qs8x8(acc20, (char8)a2.s0, b0.s01234567, FIXED_POINT_POSITION); - acc20 = mlal_sat_qs8x8(acc20, (char8)a2.s1, b1.s01234567, FIXED_POINT_POSITION); - acc21 = mlal_sat_qs8x8(acc21, (char8)a2.s0, b0.s89ABCDEF, FIXED_POINT_POSITION); - acc21 = mlal_sat_qs8x8(acc21, (char8)a2.s1, b1.s89ABCDEF, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - acc30 = mlal_sat_qs8x8(acc30, (char8)a3.s0, b0.s01234567, FIXED_POINT_POSITION); - acc30 = mlal_sat_qs8x8(acc30, (char8)a3.s1, b1.s01234567, FIXED_POINT_POSITION); - acc31 = mlal_sat_qs8x8(acc31, (char8)a3.s0, b0.s89ABCDEF, FIXED_POINT_POSITION); - acc31 = mlal_sat_qs8x8(acc31, (char8)a3.s1, b1.s89ABCDEF, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - } - - // Left-over accumulations - for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(1, src1_stride_y)) - { - char a0 = *((__global char *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - char a1 = *((__global char *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - char a2 = *((__global char *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - char a3 = *((__global char *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - char16 b0 = vload16(0, (__global char *)(src1_ptr + src_addr.s1)); - - acc00 = mlal_sat_qs8x8(acc00, (char8)a0, b0.s01234567, FIXED_POINT_POSITION); - acc01 = mlal_sat_qs8x8(acc01, (char8)a0, b0.s89ABCDEF, FIXED_POINT_POSITION); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - acc10 = mlal_sat_qs8x8(acc10, (char8)a1, b0.s01234567, FIXED_POINT_POSITION); - acc11 = mlal_sat_qs8x8(acc11, (char8)a1, b0.s89ABCDEF, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - acc20 = mlal_sat_qs8x8(acc20, (char8)a2, b0.s01234567, FIXED_POINT_POSITION); - acc21 = mlal_sat_qs8x8(acc21, (char8)a2, b0.s89ABCDEF, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - acc30 = mlal_sat_qs8x8(acc30, (char8)a3, b0.s01234567, FIXED_POINT_POSITION); - acc31 = mlal_sat_qs8x8(acc31, (char8)a3, b0.s89ABCDEF, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - } - - // Compute destination address - Image dst = CONVERT_TO_IMAGE_STRUCT(dst); - - // Compute dst address - __global uchar *dst_addr = offset(&dst, 0, 0); - - // Add offset for batched GEMM - dst_addr += get_global_id(2) * dst_stride_z; - - // Multiply by the weight of matrix product and store the result - char16 acc_qs8; - acc_qs8 = convert_char16_sat((short16)(acc00, acc01)); -#if defined(ALPHA) - acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore16(acc_qs8, 0, (__global char *)(dst_addr + 0 * dst_stride_y)); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - acc_qs8 = convert_char16_sat((short16)(acc10, acc11)); -#if defined(ALPHA) - acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore16(acc_qs8, 0, (__global char *)(dst_addr + 1 * dst_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - acc_qs8 = convert_char16_sat((short16)(acc20, acc21)); -#if defined(ALPHA) - acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore16(acc_qs8, 0, (__global char *)(dst_addr + 2 * dst_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - acc_qs8 = convert_char16_sat((short16)(acc30, acc31)); -#if defined(ALPHA) - acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore16(acc_qs8, 0, (__global char *)(dst_addr + 3 * dst_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 -} - -/** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped - * - * @note This OpenCL kernel works with fixed point data types QS16 - * @note The number of elements processed along the x and y directions must be passed at compile time using -DNUM_ELEMS_PROCESSED_PER_THREAD_X and -DNUM_ELEMS_PROCESSED_PER_THREAD_Y - * @note The number of matrix A columns, the number of elements processed per thread along the Y direction and the alpha's value need to be passed at compile time using -DCOLS_A, -DNUM_ELEMS_PROCESSED_PER_THREAD_Y and -DALPHA - * @note The fixed point position need to be passed at compile time using -DFIXED_POINT_POSITION - * @note The optional alpha value must be passed in 16 bit fixed point format using -DALPHA - * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16) - * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16]) - * - * @param[in] src0_ptr Pointer to the source matrix. Supported data types: QS8/QS16 - * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr - * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) - * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) - * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - */ -__kernel void gemm_mm_qs16(IMAGE_DECLARATION(src0), - IMAGE_DECLARATION(src1), - IMAGE_DECLARATION(dst), - uint src0_stride_z, - uint src1_stride_z, - uint dst_stride_z) -{ - int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; - - // Compute starting address for matrix A and Matrix B - int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); - - // Update address for the matrix A - src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; - - // Update address for the matrix B - src_addr.s1 += idx * sizeof(short); - - // Add offset for batched GEMM - src_addr.s0 += get_global_id(2) * src0_stride_z; - -#if defined(MATRIX_B_DEPTH) - // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 - src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; -#else // defined(MATRIX_B_DEPTH) - src_addr.s1 += get_global_id(2) * src1_stride_z; -#endif // defined(MATRIX_B_DEPTH) - - int end_row_vec_a = src_addr.s0 + (COLS_A * sizeof(short)); - - int8 acc0 = 0; -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - int8 acc1 = 0; -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - int8 acc2 = 0; -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - int8 acc3 = 0; -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - - // This for loop performs 4 accumulations per iteration - for(; src_addr.s0 <= (end_row_vec_a - 2 * (int)sizeof(short)); src_addr += (int2)(2 * sizeof(short), 2 * src1_stride_y)) - { - short2 a0 = vload2(0, (__global short *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - short2 a1 = vload2(0, (__global short *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - short2 a2 = vload2(0, (__global short *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - short2 a3 = vload2(0, (__global short *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - short8 b0 = vload8(0, (__global short *)(src1_ptr + src_addr.s1 + 0 * src1_stride_y)); - short8 b1 = vload8(0, (__global short *)(src1_ptr + src_addr.s1 + 1 * src1_stride_y)); - - acc0 = mlal_sat_qs16x8(acc0, (short8)a0.s0, b0, FIXED_POINT_POSITION); - acc0 = mlal_sat_qs16x8(acc0, (short8)a0.s1, b1, FIXED_POINT_POSITION); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - acc1 = mlal_sat_qs16x8(acc1, (short8)a1.s0, b0, FIXED_POINT_POSITION); - acc1 = mlal_sat_qs16x8(acc1, (short8)a1.s1, b1, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - acc2 = mlal_sat_qs16x8(acc2, (short8)a2.s0, b0, FIXED_POINT_POSITION); - acc2 = mlal_sat_qs16x8(acc2, (short8)a2.s1, b1, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - acc3 = mlal_sat_qs16x8(acc3, (short8)a3.s0, b0, FIXED_POINT_POSITION); - acc3 = mlal_sat_qs16x8(acc3, (short8)a3.s1, b1, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - } - - // Left-over accumulations - for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(short), src1_stride_y)) - { - short a0 = *((__global short *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - short a1 = *((__global short *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - short a2 = *((__global short *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - short a3 = *((__global short *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - short8 b0 = vload8(0, (__global short *)(src1_ptr + src_addr.s1)); - - acc0 = mlal_sat_qs16x8(acc0, (short8)a0, b0, FIXED_POINT_POSITION); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - acc1 = mlal_sat_qs16x8(acc1, (short8)a1, b0, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - acc2 = mlal_sat_qs16x8(acc2, (short8)a2, b0, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - acc3 = mlal_sat_qs16x8(acc3, (short8)a3, b0, FIXED_POINT_POSITION); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - } - - // Compute destination address - Image dst = CONVERT_TO_IMAGE_STRUCT(dst); - - // Compute dst address - __global uchar *dst_addr = offset(&dst, 0, 0); - - // Add offset for batched GEMM - dst_addr += get_global_id(2) * dst_stride_z; - - // Multiply by the weight of matrix product and store the result - short8 acc_qs16; - acc_qs16 = convert_short8_sat(acc0); -#if defined(ALPHA) - acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore8(acc_qs16, 0, (__global short *)(dst_addr + 0 * dst_stride_y)); -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 - acc_qs16 = convert_short8_sat(acc1); -#if defined(ALPHA) - acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore8(acc_qs16, 0, (__global short *)(dst_addr + 1 * dst_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 - acc_qs16 = convert_short8_sat(acc2); -#if defined(ALPHA) - acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore8(acc_qs16, 0, (__global short *)(dst_addr + 2 * dst_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 -#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 - acc_qs16 = convert_short8_sat(acc3); -#if defined(ALPHA) - acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION); -#endif // defined(ALPHA) - vstore8(acc_qs16, 0, (__global short *)(dst_addr + 3 * dst_stride_y)); -#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 -} -#endif // defined(FIXED_POINT_POSITION) #endif // defined(COLS_A) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && (NUM_ELEMS_PROCESSED_PER_THREAD_Y) #if defined(BETA) @@ -2988,94 +2383,6 @@ __kernel void gemm_ma_f16(TENSOR3D_DECLARATION(src), vstore8(out, 0, (__global half *)dst.ptr); } #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) - -#if defined(FIXED_POINT_POSITION) -/** This OpenCL kernel performs the in-place matrix addition between 2 matrices in 8 bit fixed point taking into account that the second matrix might be weighted by a scalar value beta: - * - * @note The beta's value and the fixed point position need to be passed at compile time using -DBETA and -DFIXED_POINT_POSITION - * - * @note: BETA must be passed in 8 bit fixed point format - * - * @param[in] src_ptr Pointer to the source matrix. Supported data types: QS8 - * @param[in] src_stride_x Stride of the source matrix 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 matrix 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 destination tensor in Z dimension (in bytes) - * @param[in] src_step_z dst_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 matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_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 matrix - */ -__kernel void gemm_ma_qs8(TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - // Compute source and destination addresses - Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - - // Load values from A x B - char16 alpha_ab = vload16(0, (__global char *)dst.ptr); - - // Load values from Matrix C - char16 c = vload16(0, (__global char *)src.ptr); - - // Computes alpha * axb + beta * c - char16 out = mla_sat_qs8x16(alpha_ab, (char16)BETA, c, FIXED_POINT_POSITION); - - // Store final result in axb matrix - vstore16(out, 0, (__global char *)dst.ptr); -} - -/** This OpenCL kernel performs the in-place matrix addition between 2 matrices in 16 bit fixed point taking into account that the second matrix might be weighted by a scalar value beta: - * - * @note The beta's value and the fixed point position need to be passed at compile time using -DBETA and -DFIXED_POINT_POSITION - * - * @note: BETA must be passed in 16 bit fixed point format - * - * @param[in] src_ptr Pointer to the source matrix. Supported data types: QS16 - * @param[in] src_stride_x Stride of the source matrix 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 matrix 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 destination tensor in Z dimension (in bytes) - * @param[in] src_step_z dst_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 matrix - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_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 matrix - */ -__kernel void gemm_ma_qs16(TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - // Compute source and destination addresses - Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - - // Load values from A x B - short8 alpha_ab = vload8(0, (__global short *)dst.ptr); - - // Load values from Matrix C - short8 c = vload8(0, (__global short *)src.ptr); - - // Computes alpha * axb + beta * c - short8 out = mla_sat_qs16x8(alpha_ab, (short8)BETA, c, FIXED_POINT_POSITION); - - // Store final result in axb matrix - vstore8(out, 0, (__global short *)dst.ptr); -} -#endif // defined(FIXED_POINT_POSITION) #endif // defined(BETA) #if defined(WIDTH_VECTOR_A) @@ -3151,7 +2458,7 @@ __kernel void gemm_lc_vm_f32(IMAGE_DECLARATION(src0), * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=short. * @note The vector size must be passed at compile time using -DVECTOR_SIZE e.g. -DVECTOR_SIZE=16. * - * @param[in, out] accum_ptr Pointer to the accumulate tensor. Supported data type: U8/S8/QS8/U16/S16/F16/U32/S32/F32 + * @param[in, out] accum_ptr Pointer to the accumulate tensor. Supported data type: U8/S8/U16/S16/F16/U32/S32/F32 * @param[in] accum_stride_x Stride of the accmulate tensor in X dimension (in bytes) * @param[in] accum_step_x accum_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] accum_stride_y Stride of the accumlulate tensor in Y dimension (in bytes) @@ -3175,11 +2482,7 @@ __kernel void gemm_accumulate_biases( accum_value = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)accum.ptr); VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE) biases_value = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)biases.ptr); -#ifdef FIXED_POINT_POSITION - accum_value = ADD_SAT_OP_EXPAND(biases_value, accum_value, DATA_TYPE, VECTOR_SIZE); -#else // FIXED_POINT_POSITION - accum_value = biases_value + accum_value; -#endif // FIXED_POINT_POSITION + accum_value = biases_value + accum_value; // Store result in the accumulate buffer VSTORE(VECTOR_SIZE) (accum_value, 0, (__global DATA_TYPE *)accum.ptr); |