diff options
Diffstat (limited to 'src/core/CL/cl_kernels')
| -rw-r--r-- | src/core/CL/cl_kernels/convolution_layer.cl | 19 | ||||
| -rw-r--r-- | src/core/CL/cl_kernels/gemmlowp.cl | 8 |
2 files changed, 15 insertions, 12 deletions
diff --git a/src/core/CL/cl_kernels/convolution_layer.cl b/src/core/CL/cl_kernels/convolution_layer.cl index ce0849bf7a..77b9b64945 100644 --- a/src/core/CL/cl_kernels/convolution_layer.cl +++ b/src/core/CL/cl_kernels/convolution_layer.cl @@ -97,13 +97,14 @@ __kernel void reshape_to_columns( } } -#if defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(PAD_RIGHT) && defined(PAD_BOTTOM) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(KERNEL_DEPTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) +#if defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(PAD_RIGHT) && defined(PAD_BOTTOM) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(KERNEL_DEPTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(PAD_VALUE) /** This kernel performs a reshaping of the input tensor to a tensor used to perform convolution using GEMM. * * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float + * @note The value to use for the paddings must be passed at compile time using -DPAD_VALUE: e.g. -DPAD_VALUE=0 * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32 * @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) @@ -149,14 +150,10 @@ __kernel void im2col_generic( { #if PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0 *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y)); -#else // PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0 +#else // PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0 if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT) { -#if defined(OFFSET) - *output_ptr = OFFSET; -#else /* OFFSET */ - *output_ptr = 0; -#endif /* OFFSET */ + *output_ptr = PAD_VALUE; } else { @@ -183,7 +180,7 @@ __kernel void im2col_generic( * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32 * @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) @@ -252,7 +249,7 @@ __kernel void im2col_kernel3x3_padx0_pady0( * * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32 * @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) @@ -291,7 +288,7 @@ __kernel void col2im( * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float * @note In case biases will be added in late stage, -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/F16/F32 + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/F16/F32 * @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) diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl index a8e8e600fe..a92881320e 100644 --- a/src/core/CL/cl_kernels/gemmlowp.cl +++ b/src/core/CL/cl_kernels/gemmlowp.cl @@ -380,6 +380,7 @@ __kernel void gemmlowp_matrix_b_reduction(TENSOR3D_DECLARATION(src), * @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200) * @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1) * @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6) + * @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches * * The final result is: * @@ -429,7 +430,12 @@ __kernel void gemmlowp_offset_contribution(TENSOR3D_DECLARATION(mm_result) Image sum_col = CONVERT_TO_IMAGE_STRUCT(sum_col); // Compute the offset contribution due to A_OFFSET +#if defined(SUM_COL_HAS_BATCHES) + a_offset_s32 = vload16(0, (__global int *)(sum_col.ptr + get_global_id(2) * sum_col_stride_y)); +#else // defined(MATRIX_B_HAS_BATCHES) a_offset_s32 = vload16(0, (__global int *)(sum_col.ptr)); +#endif // defined(MATRIX_B_HAS_BATCHES) + a_offset_s32 *= (int16)A_OFFSET; #endif // defined(A_OFFSET) @@ -615,4 +621,4 @@ __kernel void gemmlowp_output_stage_quantize_down_fixedpoint(TENSOR3D_DECLARATIO // Store the result vstore16(res, 0, dst.ptr); } -#endif // defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT)
\ No newline at end of file +#endif // defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT) |