| R"( |
| |
| /* |
| * Copyright (c) 2019 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. |
| */ |
| /* |
| * Copyright (c) 2016-2020 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. |
| */ |
| #ifndef ARM_COMPUTE_HELPER_H |
| #define ARM_COMPUTE_HELPER_H |
| |
| #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16) |
| #pragma OPENCL EXTENSION cl_khr_fp16 : enable |
| #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16) |
| |
| #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
| #pragma OPENCL EXTENSION cl_arm_integer_dot_product_int8 : enable |
| #endif // 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) |
| #pragma OPENCL EXTENSION cl_arm_integer_dot_product_accumulate_int8 : enable |
| #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) |
| |
| #if defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf) |
| #pragma OPENCL EXTENSION cl_arm_printf : enable |
| #endif // defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf) |
| |
| #define GPU_ARCH_MIDGARD 0x100 |
| #define GPU_ARCH_BIFROST 0x200 |
| |
| /** Concatenate two inputs. |
| * |
| * @param[in] a The first input to be concatenated |
| * @param[in] b The second input to be concatenated |
| * |
| * @return The concatenated output |
| */ |
| #define CONCAT(a, b) a##b |
| |
| /** Expand the given vector |
| * |
| * @param[in] x The vector to be expanded |
| * |
| * @return The expanded output |
| */ |
| #define EXPAND(x) x |
| |
| /** Clamp the given value between an upper and lower bound. |
| * |
| * @param[in] x The value to be clamped |
| * @param[in] min_val The lower bound |
| * @param[in] max_val The upper bound |
| * |
| * @return The clamped value. |
| */ |
| #define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val) |
| |
| /** REVn reverses the given vector whose size is n. |
| * @name REVn |
| * |
| * @param[in] x The vector to be reversed |
| * |
| * @return The reversed vector |
| * @{ |
| */ |
| #define REV1(x) ((x)) |
| #define REV2(x) ((x).s10) |
| #define REV3(x) ((x).s210) |
| #define REV4(x) ((x).s3210) |
| #define REV8(x) ((x).s76543210) |
| #define REV16(x) ((x).sFEDCBA9876543210) |
| /** @} */ // end of group REVn |
| |
| /** Reverse the given vector. |
| * @name REVERSE |
| * |
| * @param[in] x The vector to be reversed |
| * @param[in] s The size of the vector |
| * |
| * @return The reversed vector |
| * @{ |
| */ |
| #define REVERSE_STR(x, s) REV##s((x)) |
| #define REVERSE(x, s) REVERSE_STR(x, s) |
| /** @} */ // end of group REVERSE |
| |
| /** Circular-right-shift (rotate-right) the vector of size s by the amount of n. |
| * @name ROTs_n |
| * |
| * @param[in] x The vector to be shifted |
| * |
| * @return The shifted vector |
| * @{ |
| */ |
| #define ROT1_0(x) ((x)) |
| |
| #define ROT2_0(x) ((x)) |
| #define ROT2_1(x) ((x).s10) |
| |
| #define ROT3_0(x) ((x)) |
| #define ROT3_1(x) ((x).s201) |
| #define ROT3_2(x) ((x).s120) |
| |
| #define ROT4_0(x) ((x)) |
| #define ROT4_1(x) ((x).s3012) |
| #define ROT4_2(x) ((x).s2301) |
| #define ROT4_3(x) ((x).s1230) |
| |
| #define ROT8_0(x) ((x)) |
| #define ROT8_1(x) ((x).s70123456) |
| #define ROT8_2(x) ((x).s67012345) |
| #define ROT8_3(x) ((x).s56701234) |
| #define ROT8_4(x) ((x).s45670123) |
| #define ROT8_5(x) ((x).s34567012) |
| #define ROT8_6(x) ((x).s23456701) |
| #define ROT8_7(x) ((x).s12345670) |
| |
| #define ROT16_0(x) ((x)) |
| #define ROT16_1(x) ((x).sF0123456789ABCDE) |
| #define ROT16_2(x) ((x).sEF0123456789ABCD) |
| #define ROT16_3(x) ((x).sDEF0123456789ABC) |
| #define ROT16_4(x) ((x).sCDEF0123456789AB) |
| #define ROT16_5(x) ((x).sBCDEF0123456789A) |
| #define ROT16_6(x) ((x).sABCDEF0123456789) |
| #define ROT16_7(x) ((x).s9ABCDEF012345678) |
| #define ROT16_8(x) ((x).s89ABCDEF01234567) |
| #define ROT16_9(x) ((x).s789ABCDEF0123456) |
| #define ROT16_10(x) ((x).s6789ABCDEF012345) |
| #define ROT16_11(x) ((x).s56789ABCDEF01234) |
| #define ROT16_12(x) ((x).s456789ABCDEF0123) |
| #define ROT16_13(x) ((x).s3456789ABCDEF012) |
| #define ROT16_14(x) ((x).s23456789ABCDEF01) |
| #define ROT16_15(x) ((x).s123456789ABCDEF0) |
| /** @} */ // end of group ROTs_n |
| |
| /** Circular-right-shift (rotate-right) the given vector by the given amount. |
| * @name ROTATE |
| * |
| * @param[in] x The vector to be shifted |
| * @param[in] s The size of the vector |
| * @param[in] n The amount to be shifted |
| * |
| * @return The shifted vector |
| * @{ |
| */ |
| #define ROTATE_STR(x, s, n) ROT##s##_##n(x) |
| #define ROTATE(x, s, n) ROTATE_STR(x, s, n) |
| /** @} */ // end of group ROTATE |
| |
| /** Creates a vector of size n filled with offset values corresponding to the location of each element. |
| * @name V_OFFSn |
| * |
| * @param[in] dt The data type of the output vector |
| * |
| * @return The vector filled with offset values |
| * @{ |
| */ |
| #define V_OFFS1(dt) (dt)(0) |
| #define V_OFFS2(dt) (dt)(0, 1) |
| #define V_OFFS3(dt) (dt)(0, 1, 3) |
| #define V_OFFS4(dt) (dt)(0, 1, 2, 3) |
| #define V_OFFS8(dt) (dt)(0, 1, 2, 3, 4, 5, 6, 7) |
| #define V_OFFS16(dt) (dt)(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15) |
| /** @} */ // end of group V_OFFSn |
| |
| /** Create a vector filled with offset values corresponding to the location of each element. |
| * @name VEC_OFFS |
| * |
| * @param[in] dt The data type of the output vector |
| * @param[in] s The size of the output vector |
| * |
| * @return The vector filled with offset values |
| * @{ |
| */ |
| #define VEC_OFFS_STR(dt, s) V_OFFS##s(dt) |
| #define VEC_OFFS(dt, s) VEC_OFFS_STR(dt, s) |
| /** @} */ // end of group VEC_OFFS |
| |
| #define VLOAD_STR(size) vload##size |
| #define VLOAD(size) VLOAD_STR(size) |
| |
| #define VSTORE_STR(size) vstore##size |
| #define VSTORE(size) VSTORE_STR(size) |
| |
| #define float1 float |
| #define half1 half |
| #define char1 char |
| #define uchar1 uchar |
| #define short1 short |
| #define ushort1 ushort |
| #define int1 int |
| #define uint1 uint |
| #define long1 long |
| #define ulong1 ulong |
| #define double1 double |
| |
| #define vload1(OFFSET, PTR) *(OFFSET + PTR) |
| #define vstore1(DATA, OFFSET, PTR) *(OFFSET + PTR) = DATA |
| |
| // Convert built-in functions with _sat modifier are not supported in floating point so we create defines |
| // without _sat to overcome this issue |
| #define convert_float_sat convert_float |
| #define convert_float1_sat convert_float |
| #define convert_float2_sat convert_float2 |
| #define convert_float3_sat convert_float3 |
| #define convert_float4_sat convert_float4 |
| #define convert_float8_sat convert_float8 |
| #define convert_float16_sat convert_float16 |
| #define convert_half_sat convert_float |
| #define convert_half1_sat convert_half |
| #define convert_half2_sat convert_half2 |
| #define convert_half3_sat convert_half3 |
| #define convert_half4_sat convert_half4 |
| #define convert_half8_sat convert_half8 |
| #define convert_half16_sat convert_half16 |
| |
| #define convert_float1 convert_float |
| #define convert_half1 convert_half |
| #define convert_char1 convert_char |
| #define convert_uchar1 convert_uchar |
| #define convert_short1 convert_short |
| #define convert_ushort1 convert_ushort |
| #define convert_int1 convert_int |
| #define convert_uint1 convert_uint |
| #define convert_long1 convert_long |
| #define convert_ulong1 convert_ulong |
| #define convert_double1 convert_double |
| |
| #define convert_char1_sat convert_char_sat |
| #define convert_uchar1_sat convert_uchar_sat |
| #define convert_short1_sat convert_short_sat |
| #define convert_ushort1_sat convert_ushort_sat |
| #define convert_int1_sat convert_int_sat |
| #define convert_uint1_sat convert_uint_sat |
| #define convert_long1_sat convert_long_sat |
| #define convert_ulong1_sat convert_ulong_sat |
| #define convert_double1_sat convert_double_sat |
| |
| #define VEC_DATA_TYPE_STR(type, size) type##size |
| #define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size) |
| |
| #define CL_VEC_DATA_TYPE_STR(type, size) type##size |
| #define CL_VEC_DATA_TYPE(type, size) CL_VEC_DATA_TYPE_STR(type, size) |
| |
| #define CONVERT_STR(x, type) (convert_##type((x))) |
| #define CONVERT(x, type) CONVERT_STR(x, type) |
| |
| #define CONVERT_SAT_STR(x, type) (convert_##type##_sat((x))) |
| #define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type) |
| |
| #define CONVERT_SAT_ROUND_STR(x, type, round) (convert_##type##_sat_##round((x))) |
| #define CONVERT_SAT_ROUND(x, type, round) CONVERT_SAT_ROUND_STR(x, type, round) |
| |
| #define VECTOR_DECLARATION(name) \ |
| __global uchar *name##_ptr, \ |
| uint name##_stride_x, \ |
| uint name##_step_x, \ |
| uint name##_offset_first_element_in_bytes |
| |
| #define IMAGE_DECLARATION(name) \ |
| __global uchar *name##_ptr, \ |
| uint name##_stride_x, \ |
| uint name##_step_x, \ |
| uint name##_stride_y, \ |
| uint name##_step_y, \ |
| uint name##_offset_first_element_in_bytes |
| |
| #define TENSOR3D_DECLARATION(name) \ |
| __global uchar *name##_ptr, \ |
| uint name##_stride_x, \ |
| uint name##_step_x, \ |
| uint name##_stride_y, \ |
| uint name##_step_y, \ |
| uint name##_stride_z, \ |
| uint name##_step_z, \ |
| uint name##_offset_first_element_in_bytes |
| |
| #define TENSOR4D_DECLARATION(name) \ |
| __global uchar *name##_ptr, \ |
| uint name##_stride_x, \ |
| uint name##_step_x, \ |
| uint name##_stride_y, \ |
| uint name##_step_y, \ |
| uint name##_stride_z, \ |
| uint name##_step_z, \ |
| uint name##_stride_w, \ |
| uint name##_step_w, \ |
| uint name##_offset_first_element_in_bytes |
| |
| #define CONVERT_TO_VECTOR_STRUCT(name) \ |
| update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x) |
| |
| #define CONVERT_TO_VECTOR_STRUCT_NO_STEP(name) \ |
| update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0) |
| |
| #define CONVERT_TO_IMAGE_STRUCT(name) \ |
| update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y) |
| |
| #define CONVERT_TO_IMAGE_STRUCT_NO_STEP(name) \ |
| update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0) |
| |
| #define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \ |
| update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z) |
| |
| #define CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(name) \ |
| update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, name##_step_z) |
| |
| #define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \ |
| update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z) |
| |
| #define CONVERT_TO_TENSOR3D_STRUCT(name) \ |
| update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \ |
| name##_stride_z, name##_step_z) |
| |
| #define CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(name) \ |
| update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0) |
| |
| #define CONVERT_TO_TENSOR4D_STRUCT(name, mod_size) \ |
| update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \ |
| name##_stride_z, name##_step_z, name##_stride_w, name##_step_w, mod_size) |
| |
| #define CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(name, mod_size) \ |
| update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0, name##_stride_w, 0, mod_size) |
| |
| /** Structure to hold Vector information */ |
| typedef struct Vector |
| { |
| __global uchar *ptr; /**< Pointer to the starting postion of the buffer */ |
| int offset_first_element_in_bytes; /**< The offset of the first element in the source image */ |
| int stride_x; /**< Stride of the image in X dimension (in bytes) */ |
| } Vector; |
| |
| /** Structure to hold Image information */ |
| typedef struct Image |
| { |
| __global uchar *ptr; /**< Pointer to the starting postion of the buffer */ |
| int offset_first_element_in_bytes; /**< The offset of the first element in the source image */ |
| int stride_x; /**< Stride of the image in X dimension (in bytes) */ |
| int stride_y; /**< Stride of the image in Y dimension (in bytes) */ |
| } Image; |
| |
| /** Structure to hold 3D tensor information */ |
| typedef struct Tensor3D |
| { |
| __global uchar *ptr; /**< Pointer to the starting postion of the buffer */ |
| int offset_first_element_in_bytes; /**< The offset of the first element in the source image */ |
| int stride_x; /**< Stride of the image in X dimension (in bytes) */ |
| int stride_y; /**< Stride of the image in Y dimension (in bytes) */ |
| int stride_z; /**< Stride of the image in Z dimension (in bytes) */ |
| } Tensor3D; |
| |
| /** Structure to hold 4D tensor information */ |
| typedef struct Tensor4D |
| { |
| __global uchar *ptr; /**< Pointer to the starting postion of the buffer */ |
| int offset_first_element_in_bytes; /**< The offset of the first element in the source image */ |
| int stride_x; /**< Stride of the image in X dimension (in bytes) */ |
| int stride_y; /**< Stride of the image in Y dimension (in bytes) */ |
| int stride_z; /**< Stride of the image in Z dimension (in bytes) */ |
| int stride_w; /**< Stride of the image in W dimension (in bytes) */ |
| } Tensor4D; |
| |
| /** Wrap vector information into an Vector structure, and make the pointer point at this workitem's data. |
| * |
| * @param[in] ptr Pointer to the starting postion of the buffer |
| * @param[in] offset_first_element_in_bytes The offset of the first element in the source vector |
| * @param[in] stride_x Stride of the vector in X dimension (in bytes) |
| * @param[in] step_x stride_x * number of elements along X processed per workitem(in bytes) |
| * |
| * @return An image object |
| */ |
| inline Vector update_vector_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x) |
| { |
| Vector vector = |
| { |
| .ptr = ptr, |
| .offset_first_element_in_bytes = offset_first_element_in_bytes, |
| .stride_x = stride_x, |
| }; |
| vector.ptr += vector.offset_first_element_in_bytes + get_global_id(0) * step_x; |
| return vector; |
| } |
| |
| /** Wrap image information into an Image structure, and make the pointer point at this workitem's data. |
| * |
| * @param[in] ptr Pointer to the starting postion of the buffer |
| * @param[in] offset_first_element_in_bytes The offset of the first element in the source image |
| * @param[in] stride_x Stride of the image in X dimension (in bytes) |
| * @param[in] step_x stride_x * number of elements along X processed per workitem(in bytes) |
| * @param[in] stride_y Stride of the image in Y dimension (in bytes) |
| * @param[in] step_y stride_y * number of elements along Y processed per workitem(in bytes) |
| * |
| * @return An image object |
| */ |
| inline Image update_image_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y) |
| { |
| Image img = |
| { |
| .ptr = ptr, |
| .offset_first_element_in_bytes = offset_first_element_in_bytes, |
| .stride_x = stride_x, |
| .stride_y = stride_y |
| }; |
| img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y; |
| return img; |
| } |
| |
| /** Wrap 3D tensor information into an image structure, and make the pointer point at this workitem's data. |
| * |
| * @param[in] ptr Pointer to the starting postion of the buffer |
| * @param[in] offset_first_element_in_bytes The offset of the first element in the source image |
| * @param[in] stride_x Stride of the image in X dimension (in bytes) |
| * @param[in] step_x stride_x * number of elements along X processed per workitem(in bytes) |
| * @param[in] stride_y Stride of the image in Y dimension (in bytes) |
| * @param[in] step_y stride_y * number of elements along Y processed per workitem(in bytes) |
| * @param[in] stride_z Stride of the image in Z dimension (in bytes) |
| * @param[in] step_z stride_z * number of elements along Z processed per workitem(in bytes) |
| * |
| * @return A 3D tensor object |
| */ |
| inline Image update_image_from_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z) |
| { |
| Image img = |
| { |
| .ptr = ptr, |
| .offset_first_element_in_bytes = offset_first_element_in_bytes, |
| .stride_x = stride_x, |
| .stride_y = stride_y |
| }; |
| img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z; |
| return img; |
| } |
| |
| /** Wrap 3D tensor information into an tensor structure, and make the pointer point at this workitem's data. |
| * |
| * @param[in] ptr Pointer to the starting postion of the buffer |
| * @param[in] offset_first_element_in_bytes The offset of the first element in the source image |
| * @param[in] stride_x Stride of the image in X dimension (in bytes) |
| * @param[in] step_x stride_x * number of elements along X processed per workitem(in bytes) |
| * @param[in] stride_y Stride of the image in Y dimension (in bytes) |
| * @param[in] step_y stride_y * number of elements along Y processed per workitem(in bytes) |
| * @param[in] stride_z Stride of the image in Z dimension (in bytes) |
| * @param[in] step_z stride_z * number of elements along Z processed per workitem(in bytes) |
| * |
| * @return A 3D tensor object |
| */ |
| inline Tensor3D update_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z) |
| { |
| Tensor3D tensor = |
| { |
| .ptr = ptr, |
| .offset_first_element_in_bytes = offset_first_element_in_bytes, |
| .stride_x = stride_x, |
| .stride_y = stride_y, |
| .stride_z = stride_z |
| }; |
| tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z; |
| return tensor; |
| } |
| |
| inline Tensor4D update_tensor4D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z, uint stride_w, |
| uint step_w, |
| uint mod_size) |
| { |
| Tensor4D tensor = |
| { |
| .ptr = ptr, |
| .offset_first_element_in_bytes = offset_first_element_in_bytes, |
| .stride_x = stride_x, |
| .stride_y = stride_y, |
| .stride_z = stride_z, |
| .stride_w = stride_w |
| }; |
| |
| tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + (get_global_id(2) % mod_size) * step_z + (get_global_id(2) / mod_size) * step_w; |
| return tensor; |
| } |
| |
| /** Get the pointer position of a Vector |
| * |
| * @param[in] vec Pointer to the starting position of the buffer |
| * @param[in] x Relative X position |
| */ |
| inline __global const uchar *vector_offset(const Vector *vec, int x) |
| { |
| return vec->ptr + x * vec->stride_x; |
| } |
| |
| /** Get the pointer position of a Image |
| * |
| * @param[in] img Pointer to the starting position of the buffer |
| * @param[in] x Relative X position |
| * @param[in] y Relative Y position |
| */ |
| inline __global uchar *offset(const Image *img, int x, int y) |
| { |
| return img->ptr + x * img->stride_x + y * img->stride_y; |
| } |
| |
| /** Get the pointer position of a Tensor3D |
| * |
| * @param[in] tensor Pointer to the starting position of the buffer |
| * @param[in] x Relative X position |
| * @param[in] y Relative Y position |
| * @param[in] z Relative Z position |
| */ |
| inline __global const uchar *tensor3D_offset(const Tensor3D *tensor, int x, int y, int z) |
| { |
| return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z; |
| } |
| |
| /** Get the pointer position of a Tensor4D |
| * |
| * @param[in] tensor Pointer to the starting position of the buffer |
| * @param[in] x Relative X position |
| * @param[in] y Relative Y position |
| * @param[in] z Relative Z position |
| * @param[in] w Relative W position |
| */ |
| inline __global const uchar *tensor4D_offset(const Tensor4D *tensor, int x, int y, int z, int w) |
| { |
| return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + w * tensor->stride_w; |
| } |
| |
| #endif // _HELPER_H |
| |
| #undef CONVERT_SAT |
| |
| #if defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) && defined(DATA_LAYOUT_NHWC) |
| |
| #define PTR_TO_VALUE(PTR, DATA_TYPE) *((__global DATA_TYPE *)(PTR)) |
| |
| #define CONVOLUTION1x9_STRIDE1_NHWC(acc, row_ptr, weights_ptr) \ |
| ({ \ |
| VEC_DATA_TYPE(DATA_TYPE, 8) \ |
| src0 = (VEC_DATA_TYPE(DATA_TYPE, 8))( \ |
| PTR_TO_VALUE(row_ptr + 0 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 1 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 2 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 3 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 4 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 5 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 6 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 7 * src_stride_y, DATA_TYPE)); \ |
| VEC_DATA_TYPE(DATA_TYPE, 8) \ |
| src1 = (VEC_DATA_TYPE(DATA_TYPE, 8))( \ |
| PTR_TO_VALUE(row_ptr + 8 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 9 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 10 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 11 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 12 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 13 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 14 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 15 * src_stride_y, DATA_TYPE)); \ |
| VEC_DATA_TYPE(DATA_TYPE, 8) \ |
| weights_values0 = (VEC_DATA_TYPE(DATA_TYPE, 8))( \ |
| PTR_TO_VALUE(weights_ptr + 0 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 1 * weights_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(weights_ptr + 2 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 3 * weights_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(weights_ptr + 4 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 5 * weights_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(weights_ptr + 6 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 7 * weights_stride_y, DATA_TYPE)); \ |
| DATA_TYPE weights_value1 = PTR_TO_VALUE(weights_ptr + 8 * weights_stride_y, DATA_TYPE); \ |
| acc += src0 * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1234, src0.s567, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s234, src0.s567, src1.s01) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s345, src0.s67, src1.s012) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s4567, src1.s0123) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s4; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s567, src1.s0123, src1.s4) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s5; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s67, src1.s012, src1.s345) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s6; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s7, src1.s0123, src1.s456) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s7; \ |
| acc += src1 * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1; \ |
| }) |
| |
| #define CONVOLUTION1x9_STRIDE2_NHWC(acc, row_ptr, weights_ptr) \ |
| ({ \ |
| VEC_DATA_TYPE(DATA_TYPE, 16) \ |
| src0 = (VEC_DATA_TYPE(DATA_TYPE, 16))( \ |
| PTR_TO_VALUE(row_ptr + 0 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 1 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 2 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 3 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 4 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 5 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 6 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 7 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 8 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 9 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 10 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 11 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 12 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 13 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 14 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 15 * src_stride_y, DATA_TYPE)); \ |
| VEC_DATA_TYPE(DATA_TYPE, 8) \ |
| src1 = (VEC_DATA_TYPE(DATA_TYPE, 8))( \ |
| PTR_TO_VALUE(row_ptr + 16 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 17 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 18 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 19 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 20 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 21 * src_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(row_ptr + 22 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 23 * src_stride_y, DATA_TYPE)); \ |
| VEC_DATA_TYPE(DATA_TYPE, 8) \ |
| weights_values0 = (VEC_DATA_TYPE(DATA_TYPE, 8))( \ |
| PTR_TO_VALUE(weights_ptr + 0 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 1 * weights_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(weights_ptr + 2 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 3 * weights_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(weights_ptr + 4 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 5 * weights_stride_y, DATA_TYPE), \ |
| PTR_TO_VALUE(weights_ptr + 6 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 7 * weights_stride_y, DATA_TYPE)); \ |
| DATA_TYPE weights_value1 = PTR_TO_VALUE(weights_ptr + 8 * weights_stride_y, DATA_TYPE); \ |
| acc += src0.s02468ACE * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1357, src0.s9BDF) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s3579, src0.sBDF, src1.s1) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s468A, src0.sCE, src1.s02) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s4; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s579, src0.sBDF, src1.s13) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s5; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s68A, src0.sCE, src1.s024) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s6; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s79B, src0.sDF, src1.s135) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s7; \ |
| acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s8AC, src0.sE, src1.s0246) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1; \ |
| }) |
| |
| #if defined(VEC_SIZE) |
| #define VFMA(acc, w, src0, src1, src2, src3, src4, src5, src6, src7) \ |
| ({ \ |
| acc##0 = fma(src0, w, acc##0); \ |
| acc##1 = fma(src1, w, acc##1); \ |
| acc##2 = fma(src2, w, acc##2); \ |
| acc##3 = fma(src3, w, acc##3); \ |
| acc##4 = fma(src4, w, acc##4); \ |
| acc##5 = fma(src5, w, acc##5); \ |
| acc##6 = fma(src6, w, acc##6); \ |
| acc##7 = fma(src7, w, acc##7); \ |
| }) |
| |
| #define CONVOLUTION1x9_STRIDE1_NHWC_BIFROST(acc, row_ptr, weights_ptr) \ |
| ({ \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)row_ptr); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 2 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 3 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 4 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 5 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 6 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 7 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 8 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src9 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 9 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src10 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 10 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src11 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 11 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src12 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 12 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src13 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 13 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src14 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 14 * src_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| src15 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 15 * src_stride_y)); \ |
| \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| w0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 0 * weights_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| w1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 1 * weights_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| w2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 2 * weights_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| w3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 3 * weights_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| w4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 4 * weights_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| w5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 5 * weights_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| w6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 6 * weights_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| w7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 7 * weights_stride_y)); \ |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ |
| w8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 8 * weights_stride_y)); \ |
| \ |
| VFMA(acc, w0, src0, src1, src2, src3, src4, src5, src6, src7); \ |
| VFMA(acc, w1, src1, src2, src3, src4, src5, src6, src7, src8); \ |
| VFMA(acc, w2, src2, src3, src4, src5, src6, src7, src8, src9); \ |
| VFMA(acc, w3, src3, src4, src5, src6, src7, src8, src9, src10); \ |
| VFMA(acc, w4, src4, src5, src6, src7, src8, src9, src10, src11); \ |
| VFMA(acc, w5, src5, src6, src7, src8, src9, src10, src11, src12); \ |
| VFMA(acc, w6, src6, src7, src8, src9, src10, src11, src12, src13); \ |
| VFMA(acc, w7, src7, src8, src9, src10, src11, src12, src13, src14); \ |
| VFMA(acc, w8, src8, src9, src10, src11, src12, src13, src14, src15); \ |
| }) |
| |
| #if VEC_SIZE == 4 |
| #define REDUCE(out, vec) \ |
| ({ \ |
| VEC_DATA_TYPE(DATA_TYPE, 2) \ |
| tmp1 = vec.s01 + vec.s23; \ |
| out = tmp1.s0 + tmp1.s1; \ |
| }) |
| #else // VEC_SIZE == 4 |
| #error("Not supported") |
| #endif // VEC_SIZE == 4 |
| |
| #if STRIDE_X == 1 |
| #define CONVOLUTION1x9_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x9_STRIDE1_NHWC_BIFROST(acc, row_ptr, weights_ptr) |
| #else // STRIDE_X == 1 |
| #error "Not supported" |
| #endif // STRIDE_X == 1 |
| |
| #else // defined(VEC_SIZE) |
| |
| #if STRIDE_X == 1 |
| #define CONVOLUTION1x9_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x9_STRIDE1_NHWC(acc, row_ptr, weights_ptr) |
| #elif STRIDE_X == 2 // STRIDE_X == 1 |
| #define CONVOLUTION1x9_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x9_STRIDE2_NHWC(acc, row_ptr, weights_ptr) |
| #else // STRIDE_X == 1 |
| #error "STRIDE_X larger than 2 is not supported" |
| #endif // STRIDE_X == 1 |
| |
| #endif // defined(VEC_SIZE) |
| |
| //#if defined(VEC_SIZE) |
| /** This kernel performs a direct convolution to convolve the low three dimensions in a tensor with the NHWC data layout |
| * |
| * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float |
| * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH |
| * @note If biases are used then -DHAS_BIAS has to be passed at compile time |
| * |
| * @param[in] src_ptr Pointer to the source tensor. Supported data types: 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) |
| * @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 types: same as @p src_ptr |
| * @param[in] dst_stride_x Stride of the destination tensor 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 tensor in Y dimension (in bytes) |
| * @param[in] dst_step_y dst_stride_y * number of elements along Z 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 tensor |
| * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr |
| * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) |
| * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) |
| * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| * @param[in] biases_ptr (Optional) Pointer to the biases tensor. 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[in] weights_stride_w (Optional) Stride of the weights tensor in the 4th dimension |
| */ |
| __kernel void direct_convolution9x9_nhwc( |
| TENSOR3D_DECLARATION(src), |
| TENSOR3D_DECLARATION(dst), |
| TENSOR3D_DECLARATION(weights), |
| #ifdef HAS_BIAS |
| VECTOR_DECLARATION(biases), |
| #endif /* defined(HAS_BIAS) */ |
| unsigned int weights_stride_w) |
| { |
| Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
| Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); |
| |
| VEC_DATA_TYPE(DATA_TYPE, 8) |
| values = 0; |
| |
| #if defined(VEC_SIZE) |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| values0 = 0; |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| values1 = 0; |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| values2 = 0; |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| values3 = 0; |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| values4 = 0; |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| values5 = 0; |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| values6 = 0; |
| VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| values7 = 0; |
| #define STEP_X (VEC_SIZE) |
| #else // defined(VEC_SIZE) |
| #define STEP_X (1) |
| #endif // defined(VEC_SIZE) |
| |
| const int id0 = get_global_id(0); |
| const int id1 = get_global_id(1); |
| const int id2 = get_global_id(2); |
| |
| __global uchar *weights_addr = (__global uchar *)tensor3D_offset(&weights, 0, 0, 0); |
| __global uchar *src_addr = (__global uchar *)offset(&src, 0, 0) - src_stride_x * id0 + ((id2 * STRIDE_Y) - PAD_TOP) * (int)src_stride_z; |
| |
| weights_addr += id0 * weights_stride_w; |
| |
| #if(PAD_TOP == 1) |
| const int coordy = id2 - PAD_TOP; |
| for(volatile int d = 0; d < WEIGHTS_DEPTH; d += STEP_X) |
| { |
| if(coordy < 0) // special case Z = -1 doesn't exists |
| { |
| //skip first row and load the two next ones |
| CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); |
| } |
| else if(coordy == (DST_HEIGHT - PAD_TOP - 1)) |
| { |
| // special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the |
| // Z axis has no padding at all. |
| CONVOLUTION1x9_NHWC(values, src_addr, weights_addr); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); |
| } |
| else |
| { |
| CONVOLUTION1x9_NHWC(values, src_addr, weights_addr); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); |
| } |
| src_addr += STEP_X * sizeof(DATA_TYPE); |
| weights_addr += STEP_X * sizeof(DATA_TYPE); |
| } |
| #elif(PAD_TOP == 2) // PAD_TOP == 1 |
| const int coordy = id2 * STRIDE_Y; |
| for(volatile int d = 0; d < WEIGHTS_DEPTH; d += STEP_X) |
| { |
| if(coordy == 0) // special case Z = -2 doesn't exists |
| { |
| //skip first row and load the two next ones |
| CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); |
| } |
| else if(coordy == 1) // special case Z = -1 doesn't exists |
| { |
| //skip first row and load the two next ones |
| CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); |
| } |
| else if(coordy == (SRC_HEIGHT - 5)) |
| { |
| // special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the |
| // Z axis has no padding at all. |
| CONVOLUTION1x9_NHWC(values, src_addr, weights_addr); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); |
| } |
| else if(coordy == (SRC_HEIGHT - 6)) |
| { |
| // special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the |
| // Z axis has no padding at all. |
| CONVOLUTION1x9_NHWC(values, src_addr, weights_addr); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); |
| } |
| else |
| { |
| CONVOLUTION1x9_NHWC(values, src_addr, weights_addr); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); |
| } |
| src_addr += STEP_X * sizeof(DATA_TYPE); |
| weights_addr += STEP_X * sizeof(DATA_TYPE); |
| } |
| |
| #else // PAD_TOP == 1 |
| for(volatile int d = 0; d < WEIGHTS_DEPTH; d += STEP_X) |
| { |
| CONVOLUTION1x9_NHWC(values, src_addr, weights_addr); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); |
| CONVOLUTION1x9_NHWC(values, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); |
| src_addr += STEP_X * sizeof(DATA_TYPE); |
| weights_addr += STEP_X * sizeof(DATA_TYPE); |
| } |
| #endif // PAD_TOP == 1 |
| |
| #if defined(VEC_SIZE) |
| REDUCE(values.s0, values0); |
| REDUCE(values.s1, values1); |
| REDUCE(values.s2, values2); |
| REDUCE(values.s3, values3); |
| REDUCE(values.s4, values4); |
| REDUCE(values.s5, values5); |
| REDUCE(values.s6, values6); |
| REDUCE(values.s7, values7); |
| #endif // defined(VEC_SIZE) |
| |
| #if defined(HAS_BIAS) |
| Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| values += (VEC_DATA_TYPE(DATA_TYPE, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, id0))); |
| #endif // defined(HAS_BIAS) |
| |
| *((__global DATA_TYPE *)(dst.ptr + 0 * dst_stride_y)) = values.s0; |
| *((__global DATA_TYPE *)(dst.ptr + 1 * dst_stride_y)) = values.s1; |
| *((__global DATA_TYPE *)(dst.ptr + 2 * dst_stride_y)) = values.s2; |
| *((__global DATA_TYPE *)(dst.ptr + 3 * dst_stride_y)) = values.s3; |
| *((__global DATA_TYPE *)(dst.ptr + 4 * dst_stride_y)) = values.s4; |
| *((__global DATA_TYPE *)(dst.ptr + 5 * dst_stride_y)) = values.s5; |
| *((__global DATA_TYPE *)(dst.ptr + 6 * dst_stride_y)) = values.s6; |
| *((__global DATA_TYPE *)(dst.ptr + 7 * dst_stride_y)) = values.s7; |
| #undef STEP_X |
| } |
| #endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) && defined(DATA_LAYOUT_NHWC) |
| |
| )" |