| R"( |
| |
| /* |
| * Copyright (c) 2018-2020 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. |
| */ |
| /* |
| * Copyright (c) 2017-2020 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_HELPERS_ASYMM_H |
| #define ARM_COMPUTE_HELPERS_ASYMM_H |
| |
| /* |
| * 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 |
| |
| /** Convert the given vector with round to nearest even rounding mode |
| * |
| * @param[in] x The target to be converted |
| * @param[in] type The target type |
| * |
| * @return The converted vector |
| */ |
| #define CONVERT_DOWN_RTE_STR(x, type) (convert_##type##_rte((x))) |
| #define CONVERT_DOWN_RTE(x, type) CONVERT_DOWN_RTE_STR(x, type) |
| |
| /** Quantize a floating-point scalar value to 8-bit asymmetric |
| * |
| * @param[in] input Input value to quantize |
| * @param[in] offset Quantization offset |
| * @param[in] scale Quantization scale |
| * |
| * @return quantized value |
| */ |
| inline uchar quantize_qasymm8(float input, float offset, float scale) |
| { |
| float out_f32 = input / scale + offset; |
| uchar res_u8 = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, int), uchar); |
| return res_u8; |
| } |
| |
| /** Dequantize a scalar value from 8-bit asymmetric to floating-point |
| * |
| * @param[in] input Input value to quantize |
| * @param[in] offset Quantization offset |
| * @param[in] scale Quantization scale |
| * |
| * @return quantized value |
| */ |
| inline float dequantize_qasymm8(uchar input, float offset, float scale) |
| { |
| return ((float)input - offset) * scale; |
| } |
| |
| /** Dequantize a scalar value from signed 8-bit asymmetric to floating-point |
| * |
| * @param[in] input Input value to quantize |
| * @param[in] offset Quantization offset |
| * @param[in] scale Quantization scale |
| * |
| * @return quantized value |
| */ |
| inline float dequantize_qasymm8_signed(char input, float offset, float scale) |
| { |
| return ((float)input - offset) * scale; |
| } |
| |
| /** Quantize a vector of values from floating-point |
| * |
| * @param[in] type Output data type. |
| * @param[in] size Size of vector. |
| * |
| * @return quantized values |
| */ |
| #define QUANTIZE_IMPL(type, size) \ |
| inline VEC_DATA_TYPE(type, size) quantize_##type##size(VEC_DATA_TYPE(float, size) input, float offset, float scale) \ |
| { \ |
| VEC_DATA_TYPE(float, size) \ |
| out_f32 = input / (VEC_DATA_TYPE(float, size))(scale) + (VEC_DATA_TYPE(float, size))(offset); \ |
| VEC_DATA_TYPE(type, size) \ |
| res = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, VEC_DATA_TYPE(int, size)), VEC_DATA_TYPE(type, size)); \ |
| return res; \ |
| } |
| |
| /** Dequantize a vector of values to floating-point |
| * |
| * @param[in] type Input data type. |
| * @param[in] size Size of vector. |
| * |
| * @return dequantized values in floating point |
| */ |
| #define DEQUANTIZE_IMPL(type, size) \ |
| inline VEC_DATA_TYPE(float, size) dequantize_##type##size(VEC_DATA_TYPE(type, size) input, float offset, float scale) \ |
| { \ |
| return (CONVERT(input, VEC_DATA_TYPE(float, size)) - offset) * scale; \ |
| } |
| |
| /** Correctly-rounded-to-nearest division by a power-of-two. |
| * |
| * @param[in] size Size of vector. |
| * |
| * @return Correctly-rounded-to-nearest division by a power-of-two. |
| */ |
| #define ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(size) \ |
| inline VEC_DATA_TYPE(int, size) asymm_rounding_divide_by_POW2_##size(VEC_DATA_TYPE(int, size) x, VEC_DATA_TYPE(int, size) exponent) \ |
| { \ |
| const VEC_DATA_TYPE(int, size) \ |
| zero = (VEC_DATA_TYPE(int, size))0; \ |
| const VEC_DATA_TYPE(int, size) \ |
| one = (VEC_DATA_TYPE(int, size))1; \ |
| VEC_DATA_TYPE(int, size) \ |
| mask = (one << exponent) - one; \ |
| VEC_DATA_TYPE(int, size) \ |
| threshold = (mask >> 1) + select(zero, one, x < 0); \ |
| return (x >> exponent) + select(zero, one, (x & mask) > threshold); \ |
| } |
| |
| /** Product of two numbers, interpreting them as fixed-point values in the interval [-1, 1), |
| * rounding to the nearest value, and saturating -1 * -1 to the maximum value. |
| * |
| * @param[in] size Size of vector. |
| * |
| * @return Product of two fixed-point numbers. |
| */ |
| #define ASYMM_MULT_IMPL(size) \ |
| inline VEC_DATA_TYPE(int, size) asymm_mult##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \ |
| { \ |
| VEC_DATA_TYPE(int, size) \ |
| overflow = a == b && a == INT_MIN; \ |
| VEC_DATA_TYPE(long, size) \ |
| a_64 = convert_long##size(a); \ |
| VEC_DATA_TYPE(long, size) \ |
| b_64 = convert_long##size(b); \ |
| VEC_DATA_TYPE(long, size) \ |
| ab_64 = a_64 * b_64; \ |
| /* COMPMID-907 */ \ |
| VEC_DATA_TYPE(int, size) \ |
| ab_x2_high32 = convert_int##size(((ab_64 + (1 << 30)) >> 31)); \ |
| return select(ab_x2_high32, INT_MAX, overflow); \ |
| } |
| |
| /** Calculates \f$ exp(x) \f$ for x in [-1/4, 0). |
| * |
| * @param[in] size Size of vector. |
| * |
| * @return Result in fixed-point format Q0. |
| */ |
| #define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(size) \ |
| inline VEC_DATA_TYPE(int, size) asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(VEC_DATA_TYPE(int, size) a) \ |
| { \ |
| const VEC_DATA_TYPE(int, size) constant_term = 1895147668; \ |
| const VEC_DATA_TYPE(int, size) constant_1_over_3 = 715827883; \ |
| const int k_fractional_bits = 31; \ |
| VEC_DATA_TYPE(int, size) \ |
| x = a + (1 << (k_fractional_bits - 3)); \ |
| VEC_DATA_TYPE(int, size) \ |
| x2 = ASYMM_MULT(x, x, size); \ |
| VEC_DATA_TYPE(int, size) \ |
| x3 = ASYMM_MULT(x2, x, size); \ |
| VEC_DATA_TYPE(int, size) \ |
| x4 = ASYMM_MULT(x2, x2, size); \ |
| VEC_DATA_TYPE(int, size) \ |
| x4_over_4 = ASYMM_ROUNDING_DIVIDE_BY_POW2(x4, 2, size); \ |
| VEC_DATA_TYPE(int, size) \ |
| x4_over_24_plus_x3_over_6_plus_x2 = ASYMM_MULT((x4_over_4 + x3), constant_1_over_3, size) + x2; \ |
| VEC_DATA_TYPE(int, size) \ |
| x4_over_24_plus_x3_over_6_plus_x2_over_2 = ASYMM_ROUNDING_DIVIDE_BY_POW2(x4_over_24_plus_x3_over_6_plus_x2, 1, size); \ |
| return constant_term + ASYMM_MULT(constant_term, x + x4_over_24_plus_x3_over_6_plus_x2_over_2, size); \ |
| } |
| |
| /** Each bit of the result is set to the corresponding bit of either then_val or |
| * else_val depending on whether the corresponding bit of if_mask is set. |
| * Equivalent to the VBSL instruction in ARM NEON. |
| * |
| * @param[in] size Size of vector. |
| * |
| * @returns Result contaning bits from @p then_val or from @p else_val depending on corresponding bit in @p if_mask is set or not. |
| */ |
| #define ASYMM_SELECT_USING_MASK_IMPL(size) \ |
| inline VEC_DATA_TYPE(int, size) asymm_select_using_mask##size(VEC_DATA_TYPE(int, size) if_mask, VEC_DATA_TYPE(int, size) then_val, VEC_DATA_TYPE(int, size) else_val) \ |
| { \ |
| return (if_mask & then_val) ^ (~if_mask & else_val); \ |
| } |
| |
| /** For each element of input vector, the corresponding bits of the result item are set |
| * if the input item is zero. |
| * |
| * @param[in] size Size of vector. |
| * |
| * @returns Output vector with bits set when corresponding bit in @p a is zero. |
| */ |
| #define ASYMM_MASK_IF_ZERO_IMPL(size) \ |
| inline VEC_DATA_TYPE(int, size) asymm_mask_if_zero##size(VEC_DATA_TYPE(int, size) a) \ |
| { \ |
| const VEC_DATA_TYPE(int, size) all_zeros = 0; \ |
| const VEC_DATA_TYPE(int, size) all_ones = ~0; \ |
| return select(all_zeros, all_ones, a == 0); \ |
| } |
| |
| /** For each element of input vector, the corresponding bits of the result item are set |
| * if the input item is non-zero. |
| * |
| * @param[in] size Size of vector. |
| * |
| * @returns Output vector with bits set when corresponding bit in @p a is non zero. |
| */ |
| #define ASYMM_MASK_IF_NON_ZERO_IMPL(size) \ |
| inline VEC_DATA_TYPE(int, size) asymm_mask_if_non_zero##size(VEC_DATA_TYPE(int, size) a) \ |
| { \ |
| const VEC_DATA_TYPE(int, size) all_zeros = 0; \ |
| const VEC_DATA_TYPE(int, size) all_ones = ~0; \ |
| return select(all_zeros, all_ones, a != 0); \ |
| } |
| |
| #define EXP_BARREL_SHIFTER_IMPL(size) \ |
| inline VEC_DATA_TYPE(int, size) exp_barrel_shifter##size(VEC_DATA_TYPE(int, size) result, int exponent, int fp_multiplier, int k_integer_bits, int k_fractional_bits, VEC_DATA_TYPE(int, size) remainder) \ |
| { \ |
| if(k_integer_bits > exponent) \ |
| { \ |
| const int k_shift_amount = k_integer_bits > exponent ? k_fractional_bits + exponent : 0; \ |
| return ASYMM_SELECT_USING_MASK( \ |
| ASYMM_MASK_IF_NON_ZERO(remainder & (1 << k_shift_amount), size), \ |
| ASYMM_MULT(result, fp_multiplier, size), result, size); \ |
| } \ |
| \ |
| return result; \ |
| } |
| |
| /** Calculates \f$ exp(x) \f$ for x < 0. |
| * |
| * @param[in] size Size of vector. |
| * |
| * @return Result in fixed-point format Q0. |
| */ |
| #define ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(size) \ |
| inline VEC_DATA_TYPE(int, size) asymm_exp_on_negative_values##size(VEC_DATA_TYPE(int, size) a, int k_integer_bits) \ |
| { \ |
| const int k_fractional_bits = 31 - k_integer_bits; \ |
| VEC_DATA_TYPE(int, size) \ |
| k_one_quarter = 1 << (k_fractional_bits - 2); \ |
| VEC_DATA_TYPE(int, size) \ |
| mask = k_one_quarter - 1; \ |
| VEC_DATA_TYPE(int, size) \ |
| a_mod_quarter_minus_one_quarter = (a & mask) - k_one_quarter; \ |
| VEC_DATA_TYPE(int, size) \ |
| a_mod_quarter_minus_one_quarter_scaled = a_mod_quarter_minus_one_quarter << k_integer_bits; \ |
| VEC_DATA_TYPE(int, size) \ |
| result = ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(a_mod_quarter_minus_one_quarter_scaled, size); \ |
| VEC_DATA_TYPE(int, size) \ |
| remainder = a_mod_quarter_minus_one_quarter - a; \ |
| \ |
| result = EXP_BARREL_SHIFTER(result, -2, 1672461947, k_integer_bits, k_fractional_bits, remainder, size); \ |
| result = EXP_BARREL_SHIFTER(result, -1, 1302514674, k_integer_bits, k_fractional_bits, remainder, size); \ |
| result = EXP_BARREL_SHIFTER(result, +0, 790015084, k_integer_bits, k_fractional_bits, remainder, size); \ |
| result = EXP_BARREL_SHIFTER(result, +1, 290630308, k_integer_bits, k_fractional_bits, remainder, size); \ |
| result = EXP_BARREL_SHIFTER(result, +2, 39332535, k_integer_bits, k_fractional_bits, remainder, size); \ |
| result = EXP_BARREL_SHIFTER(result, +3, 720401, k_integer_bits, k_fractional_bits, remainder, size); \ |
| result = EXP_BARREL_SHIFTER(result, +4, 242, k_integer_bits, k_fractional_bits, remainder, size); \ |
| \ |
| if(k_integer_bits > 5) \ |
| { \ |
| const VEC_DATA_TYPE(int, size) clamp = -(1 << (k_fractional_bits + 5)); \ |
| result = ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_NON_ZERO(a < clamp, size), 0, result, size); \ |
| } \ |
| \ |
| const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX; \ |
| return ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_ZERO(a, size), Q0_one, result, size); \ |
| } |
| |
| /** Calculates the product of a integer value by a power of two, with either a positive exponent |
| * (equivalent to an arithmetic left shift, saturating) or a negative exponent |
| * (equivalent to an arithmetic right shift, rounding to nearest). |
| * |
| * @param[in] size Size of vector. |
| * |
| * @return Arithmetic left or right shift. |
| */ |
| #define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(size) \ |
| inline VEC_DATA_TYPE(int, size) asymm_saturating_rounding_mult_by_pow2##size(VEC_DATA_TYPE(int, size) x, int exponent) \ |
| { \ |
| if(exponent < 0) \ |
| { \ |
| return ASYMM_ROUNDING_DIVIDE_BY_POW2(x, -exponent, size); \ |
| } \ |
| \ |
| const VEC_DATA_TYPE(int, size) min = INT_MIN; \ |
| const VEC_DATA_TYPE(int, size) max = INT_MAX; \ |
| int threshold = ((1 << (31 - exponent)) - 1); \ |
| VEC_DATA_TYPE(int, size) \ |
| positive_mask = ASYMM_MASK_IF_NON_ZERO(x > threshold, size); \ |
| VEC_DATA_TYPE(int, size) \ |
| negative_mask = ASYMM_MASK_IF_NON_ZERO(x < -threshold, size); \ |
| VEC_DATA_TYPE(int, size) \ |
| result = x << exponent; \ |
| result = ASYMM_SELECT_USING_MASK(positive_mask, max, result, size); \ |
| result = ASYMM_SELECT_USING_MASK(negative_mask, min, result, size); \ |
| return result; \ |
| } |
| |
| /** Calculates (a+b)/2, rounded to the nearest integer. |
| * Equivalent to VRHADD in the ARM NEON instruction set. |
| * |
| * @param[in] size Size of vector. |
| * |
| * @return (a+b)/2, rounded to the nearest integer. |
| */ |
| #define ASYMM_ROUNDING_HALF_SUM_IMPL(size) \ |
| inline VEC_DATA_TYPE(int, size) asymm_rounding_half_sum##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \ |
| { \ |
| VEC_DATA_TYPE(long, size) \ |
| a64 = convert_long##size(a); \ |
| VEC_DATA_TYPE(long, size) \ |
| b64 = convert_long##size(b); \ |
| VEC_DATA_TYPE(long, size) \ |
| sum = a64 + b64; \ |
| const VEC_DATA_TYPE(long, size) one = 1; \ |
| const VEC_DATA_TYPE(long, size) minus_one = -1; \ |
| VEC_DATA_TYPE(long, size) \ |
| sign = select(minus_one, one, sum >= 0); \ |
| return convert_int##size((sum + sign) / 2); \ |
| } |
| |
| /** Calculates \f$ 1 / (1 + x) \f$ for x in (0, 1). |
| * |
| * @param[in] size Size of vector. |
| * |
| * @return Result in fixed-point format Q0. |
| */ |
| #define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(size) \ |
| inline VEC_DATA_TYPE(int, size) asymm_one_over_one_plus_x_for_x_in_0_1##size(VEC_DATA_TYPE(int, size) a) \ |
| { \ |
| const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX; \ |
| const VEC_DATA_TYPE(int, size) Q2_one = 1 << (31 - 2); \ |
| VEC_DATA_TYPE(int, size) \ |
| half_denominator = ASYMM_ROUNDING_HALF_SUM(a, Q0_one, size); \ |
| const VEC_DATA_TYPE(int, size) Q2_48_over_17 = 1515870810; \ |
| const VEC_DATA_TYPE(int, size) Q2_neg_32_over_17 = -1010580540; \ |
| VEC_DATA_TYPE(int, size) \ |
| x = Q2_48_over_17 + ASYMM_MULT(half_denominator, Q2_neg_32_over_17, size); \ |
| for(int i = 0; i < 3; i++) \ |
| { \ |
| VEC_DATA_TYPE(int, size) \ |
| half_denominator_times_x = ASYMM_MULT(half_denominator, x, size); \ |
| VEC_DATA_TYPE(int, size) \ |
| one_minus_half_denominator_times_x = Q2_one - half_denominator_times_x; \ |
| VEC_DATA_TYPE(int, size) \ |
| tmp = ASYMM_MULT(x, one_minus_half_denominator_times_x, size); \ |
| x = x + ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(tmp, 2, size); \ |
| } \ |
| return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, 1, size); \ |
| } |
| |
| /** Considering the integer value as fixed-point, change the number of integer bits and update value accordingly. |
| * |
| * @param[in] size Size of vector. |
| * |
| * @return Rescaled value. |
| */ |
| #define ASYMM_RESCALE_IMPL(size) \ |
| inline VEC_DATA_TYPE(int, size) asymm_rescale##size(VEC_DATA_TYPE(int, size) value, int src_integer_bits, int dst_integer_bits) \ |
| { \ |
| int exponent = src_integer_bits - dst_integer_bits; \ |
| return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(value, exponent, size); \ |
| } |
| |
| #define QUANTIZE_STR(input, offset, scale, type, size) quantize_##type##size(input, offset, scale) |
| #define QUANTIZE(input, offset, scale, type, size) QUANTIZE_STR(input, offset, scale, type, size) |
| #define DEQUANTIZE_STR(input, offset, scale, type, size) dequantize_##type##size(input, offset, scale) |
| #define DEQUANTIZE(input, offset, scale, type, size) DEQUANTIZE_STR(input, offset, scale, type, size) |
| |
| #define ASYMM_ROUNDING_DIVIDE_BY_POW2(x, exponent, size) asymm_rounding_divide_by_POW2_##size(x, exponent) |
| #define ASYMM_MULT(a, b, size) asymm_mult##size(a, b) |
| #define ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(x, quantized_multiplier, left_shift, size) \ |
| ASYMM_MULT(x *((VEC_DATA_TYPE(int, size))(1) << (-left_shift)), quantized_multiplier, size) |
| #define ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(x, quantized_multiplier, right_shift, size) \ |
| ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(x, quantized_multiplier, size), right_shift, size) |
| #define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(a, size) asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(a) |
| #define ASYMM_SELECT_USING_MASK(if_mask, then_val, else_val, size) asymm_select_using_mask##size(if_mask, then_val, else_val) |
| #define ASYMM_MASK_IF_ZERO(a, size) asymm_mask_if_zero##size(a) |
| #define ASYMM_MASK_IF_NON_ZERO(a, size) asymm_mask_if_non_zero##size(a) |
| #define EXP_BARREL_SHIFTER(result, exponent, fp_multiplier, k_integer_bits, k_fractional_bits, remainder, size) exp_barrel_shifter##size(result, exponent, fp_multiplier, k_integer_bits, k_fractional_bits, remainder) |
| #define ASYMM_EXP_ON_NEGATIVE_VALUES(a, k_integer_bits, size) asymm_exp_on_negative_values##size(a, k_integer_bits) |
| #define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1(a, size) asymm_one_over_one_plus_x_for_x_in_0_1##size(a) |
| #define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, exponent, size) asymm_saturating_rounding_mult_by_pow2##size(x, exponent) |
| #define ASYMM_ROUNDING_HALF_SUM(a, b, size) asymm_rounding_half_sum##size(a, b) |
| #define ASYMM_RESCALE(value, src_integer_bits, dst_integer_bits, size) asymm_rescale##size(value, src_integer_bits, dst_integer_bits) |
| |
| QUANTIZE_IMPL(uchar, 1) |
| QUANTIZE_IMPL(char, 1) |
| QUANTIZE_IMPL(uint, 1) |
| QUANTIZE_IMPL(int, 1) |
| QUANTIZE_IMPL(uchar, 4) |
| QUANTIZE_IMPL(ushort, 4) |
| QUANTIZE_IMPL(short, 4) |
| QUANTIZE_IMPL(uchar, 16) |
| QUANTIZE_IMPL(char, 16) |
| QUANTIZE_IMPL(ushort, 16) |
| QUANTIZE_IMPL(short, 16) |
| QUANTIZE_IMPL(uint, 16) |
| QUANTIZE_IMPL(int, 16) |
| |
| DEQUANTIZE_IMPL(uchar, 1) |
| DEQUANTIZE_IMPL(char, 1) |
| DEQUANTIZE_IMPL(uint, 1) |
| DEQUANTIZE_IMPL(int, 1) |
| DEQUANTIZE_IMPL(uchar, 4) |
| DEQUANTIZE_IMPL(ushort, 4) |
| DEQUANTIZE_IMPL(short, 4) |
| DEQUANTIZE_IMPL(uchar, 16) |
| DEQUANTIZE_IMPL(char, 16) |
| DEQUANTIZE_IMPL(ushort, 16) |
| DEQUANTIZE_IMPL(short, 16) |
| DEQUANTIZE_IMPL(uint, 16) |
| DEQUANTIZE_IMPL(int, 16) |
| |
| ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(1) |
| ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(2) |
| ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(4) |
| ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(8) |
| ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(16) |
| |
| ASYMM_MULT_IMPL(1) |
| ASYMM_MULT_IMPL(2) |
| ASYMM_MULT_IMPL(4) |
| ASYMM_MULT_IMPL(8) |
| ASYMM_MULT_IMPL(16) |
| |
| ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(2) |
| ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(4) |
| ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(8) |
| ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(16) |
| |
| ASYMM_SELECT_USING_MASK_IMPL(2) |
| ASYMM_SELECT_USING_MASK_IMPL(4) |
| ASYMM_SELECT_USING_MASK_IMPL(8) |
| ASYMM_SELECT_USING_MASK_IMPL(16) |
| |
| ASYMM_MASK_IF_ZERO_IMPL(2) |
| ASYMM_MASK_IF_ZERO_IMPL(4) |
| ASYMM_MASK_IF_ZERO_IMPL(8) |
| ASYMM_MASK_IF_ZERO_IMPL(16) |
| |
| ASYMM_MASK_IF_NON_ZERO_IMPL(2) |
| ASYMM_MASK_IF_NON_ZERO_IMPL(4) |
| ASYMM_MASK_IF_NON_ZERO_IMPL(8) |
| ASYMM_MASK_IF_NON_ZERO_IMPL(16) |
| |
| EXP_BARREL_SHIFTER_IMPL(2) |
| EXP_BARREL_SHIFTER_IMPL(4) |
| EXP_BARREL_SHIFTER_IMPL(8) |
| EXP_BARREL_SHIFTER_IMPL(16) |
| |
| ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(2) |
| ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(4) |
| ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(8) |
| ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(16) |
| |
| ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(2) |
| ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(4) |
| ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(8) |
| ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(16) |
| |
| ASYMM_ROUNDING_HALF_SUM_IMPL(2) |
| ASYMM_ROUNDING_HALF_SUM_IMPL(4) |
| ASYMM_ROUNDING_HALF_SUM_IMPL(8) |
| ASYMM_ROUNDING_HALF_SUM_IMPL(16) |
| |
| ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(2) |
| ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(4) |
| ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(8) |
| ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(16) |
| |
| ASYMM_RESCALE_IMPL(2) |
| ASYMM_RESCALE_IMPL(4) |
| ASYMM_RESCALE_IMPL(8) |
| ASYMM_RESCALE_IMPL(16) |
| |
| #endif // ARM_COMPUTE_HELPERS_ASYMM_H |
| |
| /** Clamps the given coordinates to the borders according to the border size. |
| * |
| * @param[in] coords Vector of 2D coordinates to clamp. Even positions are X coords, odd positions are Y coords. |
| * @param[in] width Width of the image |
| * @param[in] height Height of the image |
| * @param[in] border_size Border size of the image |
| * |
| */ |
| inline const float8 clamp_to_border_with_size_quantized(float8 coords, const float width, const float height, const float border_size) |
| { |
| const float4 clamped_x = clamp(coords.even, 0.0f - border_size, width - 1 + border_size); |
| const float4 clamped_y = clamp(coords.odd, 0.0f - border_size, height - 1 + border_size); |
| return (float8)(clamped_x.s0, clamped_y.s0, clamped_x.s1, clamped_y.s1, clamped_x.s2, clamped_y.s2, clamped_x.s3, clamped_y.s3); |
| } |
| |
| /* FIXME(COMPMID-682): Clamp border properly in UNDEFINED border mode in Warp, Scale, Remap */ |
| /** Clamps the given coordinates to the borders. |
| * |
| * @param[in] coords Vector of 2D coordinates to clamp. Even positions are X coords, odd positions are Y coords. |
| * @param[in] width Width of the image |
| * @param[in] height Height of the image |
| * |
| */ |
| inline const float8 clamp_to_border_quantized(float8 coords, const float width, const float height) |
| { |
| return clamp_to_border_with_size_quantized(coords, width, height, 1); |
| } |
| |
| /** Given a texel coordinates this function will return the following array of coordinates: |
| * [ P, right neighbour, below neighbour, below right neighbour ] |
| * |
| * @note No checks to see if the coordinates are out of the image are done here. |
| * |
| * @param[in] coord Input coordinates |
| * |
| * @return vector of 8 floats with the coordinates, even positions are x and odd y. |
| */ |
| inline const float8 get_neighbour_coords_quantized(const float2 coord) |
| { |
| return (float8)(/*tl*/ coord.s0, coord.s1, /*tr*/ coord.s0 + 1, coord.s1, /*bl*/ coord.s0, coord.s1 + 1, /*br*/ coord.s0 + 1, coord.s1 + 1); |
| } |
| |
| /** Returns the current thread coordinates. */ |
| inline const float2 get_current_coords_quantized() |
| { |
| return (float2)(get_global_id(0) * 4, get_global_id(1)); |
| } |
| |
| /** Computes the bilinear interpolation for each set of coordinates in the vector coords and returns the values |
| * |
| * @param[in] in Pointer to the source image. |
| * @param[in] coords Vector of four 2D coordinates. Even pos is x and odd y. |
| * @param[in] width Width of the image |
| * @param[in] height Height of the image |
| * @param[in] border_size Border size |
| * @param[in] scale Scale value |
| * @param[in] offset_qasymm Offset value |
| */ |
| inline const VEC_DATA_TYPE(DATA_TYPE, 4) bilinear_interpolate_with_border_quantized(const Image *in, const float8 coords, const float width, const float height, const float border_size, |
| const float scale, const int offset_qasymm) |
| { |
| // If any of the 4 texels is out of the image's boundaries we use the border value (REPLICATE or CONSTANT) for any texel out of the image. |
| |
| // Sets the 4x4 coordinates for each of the four input texels |
| const float8 fc = floor(coords); |
| const float16 c1 = (float16)( |
| clamp_to_border_with_size_quantized(get_neighbour_coords_quantized((float2)(fc.s0, fc.s1)), width, height, border_size), |
| clamp_to_border_with_size_quantized(get_neighbour_coords_quantized((float2)(fc.s2, fc.s3)), width, height, border_size)); |
| const float16 c2 = (float16)( |
| clamp_to_border_with_size_quantized(get_neighbour_coords_quantized((float2)(fc.s4, fc.s5)), width, height, border_size), |
| clamp_to_border_with_size_quantized(get_neighbour_coords_quantized((float2)(fc.s6, fc.s7)), width, height, border_size)); |
| |
| // Loads the values from the input image |
| const int16 t = (int16)( |
| /* tl, tr, bl, br */ |
| * ((__global DATA_TYPE *)offset(in, c1.s0, c1.s1)), *((__global DATA_TYPE *)offset(in, c1.s2, c1.s3)), |
| *((__global DATA_TYPE *)offset(in, c1.s4, c1.s5)), *((__global DATA_TYPE *)offset(in, c1.s6, c1.s7)), |
| *((__global DATA_TYPE *)offset(in, c1.s8, c1.s9)), *((__global DATA_TYPE *)offset(in, c1.sa, c1.sb)), |
| *((__global DATA_TYPE *)offset(in, c1.sc, c1.sd)), *((__global DATA_TYPE *)offset(in, c1.se, c1.sf)), |
| *((__global DATA_TYPE *)offset(in, c2.s0, c2.s1)), *((__global DATA_TYPE *)offset(in, c2.s2, c2.s3)), |
| *((__global DATA_TYPE *)offset(in, c2.s4, c2.s5)), *((__global DATA_TYPE *)offset(in, c2.s6, c2.s7)), |
| *((__global DATA_TYPE *)offset(in, c2.s8, c2.s9)), *((__global DATA_TYPE *)offset(in, c2.sa, c2.sb)), |
| *((__global DATA_TYPE *)offset(in, c2.sc, c2.sd)), *((__global DATA_TYPE *)offset(in, c2.se, c2.sf))); |
| |
| const float16 inf32 = convert_float16(t - (int16)offset_qasymm) * (float16)scale; |
| |
| const float8 a = coords - fc; |
| const float8 b = ((float8)(1.f)) - a; |
| const float4 fr = (float4)( |
| ((inf32.s0 * b.s0 * b.s1) + (inf32.s1 * a.s0 * b.s1) + (inf32.s2 * b.s0 * a.s1) + (inf32.s3 * a.s0 * a.s1)), |
| ((inf32.s4 * b.s2 * b.s3) + (inf32.s5 * a.s2 * b.s3) + (inf32.s6 * b.s2 * a.s3) + (inf32.s7 * a.s2 * a.s3)), |
| ((inf32.s8 * b.s4 * b.s5) + (inf32.s9 * a.s4 * b.s5) + (inf32.sa * b.s4 * a.s5) + (inf32.sb * a.s4 * a.s5)), |
| ((inf32.sc * b.s6 * b.s7) + (inf32.sd * a.s6 * b.s7) + (inf32.se * b.s6 * a.s7) + (inf32.sf * a.s6 * a.s7))); |
| |
| const VEC_DATA_TYPE(DATA_TYPE, 4) res = CONVERT_SAT(convert_int4_sat_rtp(fr / scale) + offset_qasymm, VEC_DATA_TYPE(DATA_TYPE, 4)); |
| |
| return res; |
| } |
| |
| /* FIXME(COMPMID-682): Clamp border properly in UNDEFINED border mode in Warp, Scale, Remap */ |
| /** Computes the bilinear interpolation for each set of coordinates in the vector coords and returns the values |
| * |
| * @param[in] in Pointer to the source image. |
| * @param[in] coords Vector of four 2D coordinates. Even pos is x and odd y. |
| * @param[in] width Width of the image |
| * @param[in] height Height of the image |
| * @param[in] scale Scale value |
| * @param[in] offset_qasymm Offset value |
| */ |
| inline const VEC_DATA_TYPE(DATA_TYPE, 4) bilinear_interpolate_quantized(const Image *in, const float8 coords, const float width, const float height, const float scale, const int offset_qasymm) |
| { |
| return bilinear_interpolate_with_border_quantized(in, coords, width, height, 1, scale, offset_qasymm); |
| } |
| |
| )" |
