Google Git
Sign in
android / platform / external / ComputeLibrary / eb0170eb99a8d03384e87ff20a38f76ce16de252 / . / build / android-arm64v8a / src / core / CL / cl_kernels / gemm_helpers.hembed
blob: 9b7438587b07f061aabe3e835530aa010934013f [file] [log] [blame]
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) 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
#if GPU_ARCH == GPU_ARCH_BIFROST
#define MLA(a, b, c) (fma(c, b, a))
#else // GPU_ARCH == GPU_ARCH_BIFROST
#define MLA(a, b, c) ((b) * (c) + (a))
#endif // GPU_ARCH == GPU_ARCH_BIFROST
// Logistic Activation
#define logistic_op(DATA_TYPE, x, A_VAL, B_VAL) ((DATA_TYPE)1.0 / ((DATA_TYPE)1.0 + exp(-x)))
// Hyperbolic Tangent Activation
#define tanh_op(DATA_TYPE, x, A_VAL, B_VAL) ((DATA_TYPE)A_VAL * tanh((DATA_TYPE)B_VAL * x))
// RELU Tangent Activation
#define relu_op(DATA_TYPE, x, A_VAL, B_VAL) (max((DATA_TYPE)0.0, x))
// Bounded RELU Activation
#define brelu_op(DATA_TYPE, x, A_VAL, B_VAL) (min((DATA_TYPE)A_VAL, max((DATA_TYPE)0.0, x)))
// Lower Upper Bounded RELU Activation
#define lu_brelu_op(DATA_TYPE, x, A_VAL, B_VAL) (min(max(x, (DATA_TYPE)B_VAL), (DATA_TYPE)A_VAL))
// Leaky RELU Activation
#define lrelu_op(DATA_TYPE, x, A_VAL, B_VAL) ((min(x, (DATA_TYPE)0.0) * (DATA_TYPE)A_VAL) + max(x, (DATA_TYPE)0.0))
// Soft RELU Activation
#define srelu_op(DATA_TYPE, x, A_VAL, B_VAL) (log((DATA_TYPE)1.0 + exp(x)))
// ELU Activation
#define elu_op(DATA_TYPE, x, A_VAL, B_VAL) (select(((DATA_TYPE)A_VAL * (exp(x) - (DATA_TYPE)1.0)), x, isgreaterequal(x, (DATA_TYPE)0.0)))
// Absolute Activation
#define abs_op(DATA_TYPE, x, A_VAL, B_VAL) (fabs(x))
// Square Activation
#define square_op(DATA_TYPE, x, A_VAL, B_VAL) (x * x)
// Square-root Activation
#define sqrt_op(DATA_TYPE, x, A_VAL, B_VAL) (sqrt(x))
// Linear Activation
#define linear_op(DATA_TYPE, x, A_VAL, B_VAL) (MLA((DATA_TYPE)B_VAL, (DATA_TYPE)A_VAL, x))
// Identity Activation
#define identity_op(DATA_TYPE, x, A_VAL, B_VAL) (x)
#define OP(op, DATA_TYPE, x, A_VAL, B_VAL) op##_op(DATA_TYPE, x, A_VAL, B_VAL)
#define ACTIVATION(op, DATA_TYPE, x, A_VAL, B_VAL) OP(op, DATA_TYPE, x, A_VAL, B_VAL)
/*
* 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
/** Loads the rows from 0 to n-1 in the given variables (BASENAME0 to BASENAMEn-1).
* @name LOAD_ROW_n
*
* @param[in] N0 The number of rows to load
* @param[in] DATA_TYPE The data type of variables
* @param[in] BASENAME The basename of the destination variables for the loaded rows
* @param[in] PTR The base pointer
* @param[in] OFFSET The offset within a row
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The z-axis offset vector
* @{
*/
#define LOAD_ROW_1(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##0 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 0 * STRIDE_Y + Z##0));
#define LOAD_ROW_2(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_1(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##1 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 1 * STRIDE_Y + Z##1));
#define LOAD_ROW_3(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_2(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##2 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 2 * STRIDE_Y + Z##2));
#define LOAD_ROW_4(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_3(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##3 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 3 * STRIDE_Y + Z##3));
#define LOAD_ROW_5(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_4(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##4 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 4 * STRIDE_Y + Z##4));
#define LOAD_ROW_6(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_5(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##5 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 5 * STRIDE_Y + Z##5));
#define LOAD_ROW_7(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_6(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##6 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 6 * STRIDE_Y + Z##6));
#define LOAD_ROW_8(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_7(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##7 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 7 * STRIDE_Y + Z##7));
#define LOAD_ROW_9(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_8(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##8 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 8 * STRIDE_Y + Z##8));
#define LOAD_ROW_10(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_9(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##9 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 9 * STRIDE_Y + Z##9));
#define LOAD_ROW_11(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_10(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##A = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 10 * STRIDE_Y + Z##A));
#define LOAD_ROW_12(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_11(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##B = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 11 * STRIDE_Y + Z##B));
#define LOAD_ROW_13(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_12(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##C = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 12 * STRIDE_Y + Z##C));
#define LOAD_ROW_14(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_13(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##D = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 13 * STRIDE_Y + Z##D));
#define LOAD_ROW_15(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_14(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##E = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 14 * STRIDE_Y + Z##E));
#define LOAD_ROW_16(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
LOAD_ROW_15(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
VEC_DATA_TYPE(DATA_TYPE, N0) \
BASENAME##F = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 15 * STRIDE_Y + Z##F));
/** @}*/ // end of group LOAD_ROW_n
/** Load Blocks (consecutive rows and columns) with Z offset.
* @name LOAD_BLOCK
*
* Supported cases are M0=1,2,3,...,16 and N0=1,2,3,4,8,16
* The data to load is expected to have consecutive names for each row.
* E.g., for M0=3, and BASENAME=c, the expected data is c0, c1 and c2.
* The Z offset is expected to have consecutive names.
* E.g., for M0=3, and Z=zin, the expected Z offsets are zin0, zin1 and zin2.
*
* @param[in] M0 The number of consecutive rows
* @param[in] N0 The number of consecutive columns
* @param[in] DATA_TYPE The data type of the target
* @param[in] BASENAME The basename of the result variables
* @param[in] PTR The base pointer for the data
* @param[in] OFFSET The offset within a row
* @param[in] STRIDE_Y The stride in y-axis direction
* @param[in] Z The z-axis offset vector
* @{
*/
#define LOAD_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) LOAD_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)
#define LOAD_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) LOAD_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)
/** @} */ // end of group LOAD_BLOCK
/** Basic macros to calculate Z offset values from Z0 to Zn-1
* @name CALCULATE_Z_OFFSET_n
*
* @param[in] M0 The number of offset values to calculate
* @param[in] DATA_TYPE The data type of the results
* @param[in] Z The basename of the result variables
* @param[in] Y The work-itme ID of y-axis
* @param[in] HEIGHT_GEMM3D The height of GEMM3D
* @param[in] DEPTH_GEMM3D The depth of GEMM3D
* @param[in] CROSS_PLANE_PAD The padding required for plane changes accross the z-dimension
* @param[in] STRIDE_Y The stride value in y-axis direction
*
* @{
*/
#define CALCULATE_Z_OFFSET_1(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
Z##0 = (0 + (DATA_TYPE)(Y * (DATA_TYPE)M0)) / (DATA_TYPE)HEIGHT_GEMM3D; \
Z##0 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##0); \
Z##0 *= (CROSS_PLANE_PAD * STRIDE_Y);
#define CALCULATE_Z_OFFSET_2(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
CALCULATE_Z_OFFSET_1(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
Z##1 = (1 + (DATA_TYPE)(Y * (DATA_TYPE)M0)) / (DATA_TYPE)HEIGHT_GEMM3D; \
Z##1 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##1); \
Z##1 *= (CROSS_PLANE_PAD * STRIDE_Y);
#define CALCULATE_Z_OFFSET_3(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
CALCULATE_Z_OFFSET_2(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
Z##2 = (2 + (DATA_TYPE)(Y * (DATA_TYPE)M0)) / (DATA_TYPE)HEIGHT_GEMM3D; \
Z##2 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##2); \
Z##2 *= (CROSS_PLANE_PAD * STRIDE_Y);
#define CALCULATE_Z_OFFSET_4(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
CALCULATE_Z_OFFSET_3(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
Z##3 = (3 + (DATA_TYPE)(Y * (DATA_TYPE)M0)) / (DATA_TYPE)HEIGHT_GEMM3D; \
Z##3 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##3); \
Z##3 *= (CROSS_PLANE_PAD * STRIDE_Y);
#define CALCULATE_Z_OFFSET_5(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
CALCULATE_Z_OFFSET_4(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
Z##4 = (4 + (DATA_TYPE)(Y * (DATA_TYPE)M0)) / (DATA_TYPE)HEIGHT_GEMM3D; \
Z##4 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##4); \
Z##4 *= (CROSS_PLANE_PAD * STRIDE_Y);
#define CALCULATE_Z_OFFSET_6(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
CALCULATE_Z_OFFSET_5(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
Z##5 = (5 + (DATA_TYPE)(Y * (DATA_TYPE)M0)) / (DATA_TYPE)HEIGHT_GEMM3D; \
Z##5 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##5); \
Z##5 *= (CROSS_PLANE_PAD * STRIDE_Y);
#define CALCULATE_Z_OFFSET_7(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
CALCULATE_Z_OFFSET_6(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
Z##6 = (6 + (DATA_TYPE)(Y * (DATA_TYPE)M0)) / (DATA_TYPE)HEIGHT_GEMM3D; \
Z##6 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##6); \
Z##6 *= (CROSS_PLANE_PAD * STRIDE_Y);
#define CALCULATE_Z_OFFSET_8(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
CALCULATE_Z_OFFSET_7(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
Z##7 = (7 + (DATA_TYPE)(Y * (DATA_TYPE)M0)) / (DATA_TYPE)HEIGHT_GEMM3D; \
Z##7 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##7); \
Z##7 *= (CROSS_PLANE_PAD * STRIDE_Y);
/** @} */ // end of group CALCULATE_Z_OFFSET_n
/** Calculate Z offset values from Z0 to Zn-1
* @name CALCULATE_Z_OFFSET
*
* The Z offsets are expected to have consecutive names.
* E.g., for M0=3 and Z=zin, the expected names of Z offsets are zin1, zin2, zin3.
* Note that, CROSS_PLANE_PAD (cross plain padding) is required to take into account
* the possible cross plane paddings in case of the plance changes across the z-dimension.
*
* <!--
* | |
* | plane0 |
* | |
* |__________________|
* |******************|
* | cross_plane_pad |
* |******************|
* | |
* | plane1 |
* | |
* |__________________|
* -->
*
* @param[in] M0 The number of offset values to calculate
* @param[in] DATA_TYPE The data type of the results
* @param[in] Z The basename of the result variables
* @param[in] Y The work-itme ID of y-axis
* @param[in] HEIGHT_GEMM3D The height of GEMM3D
* @param[in] DEPTH_GEMM3D The depth of GEMM3D
* @param[in] CROSS_PLANE_PAD The padding required for plane changes accross the z-dimension
* @param[in] STRIDE_Y The stride value in y-axis direction
* @{
*/
#define CALCULATE_Z_OFFSET_STR(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) CALCULATE_Z_OFFSET_##M0(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y)
#define CALCULATE_Z_OFFSET(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) CALCULATE_Z_OFFSET_STR(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y)
/** @} */ // end of group CALCULATE_Z_OFFSET
/** Store the 0 to (n-1)th rows of the given variables
* @name STORE_ROW_n
*
* @param[in] N0 The size of the vectors
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @{
*/
#define STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));
#define STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));
#define STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));
#define STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));
#define STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));
#define STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));
#define STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));
#define STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));
#define STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));
#define STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));
#define STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));
#define STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));
#define STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));
#define STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));
#define STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));
#define STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
/** @} */ // end of groupd STORE_ROW_n
/** Convert and store the 0th to (n-1)th rows of the given variables
* @name CONVERT_STORE_ROW_n
*
* @param[in] N0 The size of the vectors
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @{
*/
#define CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##0), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));
#define CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##1), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));
#define CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##2), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));
#define CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##3), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));
#define CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##4), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));
#define CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##5), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));
#define CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##6), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));
#define CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##7), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));
#define CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##8), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));
#define CONVERT_STORE_ROW_10(N0, DATA, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##9), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));
#define CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##A), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));
#define CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##B), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));
#define CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##C), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));
#define CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##D), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));
#define CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##E), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));
#define CONVERT_STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##F), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
/** @} */ // end of groupd CONVERT_STORE_ROW_n
/** Store a block of the given size M0xN0
* @name STORE_BLOCK
*
* Supported cases are M0=1,2,3,...,16 and N0=2,3,4,8,16.
* The data to store is expected to have consecutive names for each row.
* E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
* The Z offset is expected to have consecutive names.
* E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
*
* @param[in] M0 The number of rows to store
* @param[in] N0 The size of each vector
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @{
*/
#define STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
#define STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
/** @} */ // end of group STORE_BLOCK
/** Convert and store a block of the given size M0xN0
* @name CONVERT_STORE_BLOCK
*
* Supported cases are M0=1,2,3,...,16 and N0=2,3,4,8,16.
* The data to store is expected to have consecutive names for each row.
* E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
* The Z offset is expected to have consecutive names.
* E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
*
* @param[in] M0 The number of rows to store
* @param[in] N0 The size of each vector
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @{
*/
#define CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
#define CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
/** @} */ // end of group CONVERT_STORE_BLOCK
/** Scale the rows in the given variables (BASENAME0 to BASENAMEn-1)
* @name SCALE_ROW_n
*
* @param[in] DATA_TYPE The data type of the variables
* @param[in] BASENAME The basename of the variables
* @param[in] SCALE The scale factor
* @{
*/
#define SCALE_ROW_1(DATA_TYPE, BASENAME, SCALE) \
BASENAME##0 *= (DATA_TYPE)SCALE;
#define SCALE_ROW_2(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_1(DATA_TYPE, BASENAME, SCALE) \
BASENAME##1 *= (DATA_TYPE)SCALE;
#define SCALE_ROW_3(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_2(DATA_TYPE, BASENAME, SCALE) \
BASENAME##2 *= (DATA_TYPE)SCALE;
#define SCALE_ROW_4(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_3(DATA_TYPE, BASENAME, SCALE) \
BASENAME##3 *= (DATA_TYPE)SCALE;
#define SCALE_ROW_5(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_4(DATA_TYPE, BASENAME, SCALE) \
BASENAME##4 *= (DATA_TYPE)SCALE;
#define SCALE_ROW_6(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_5(DATA_TYPE, BASENAME, SCALE) \
BASENAME##5 *= (DATA_TYPE)SCALE;
#define SCALE_ROW_7(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_6(DATA_TYPE, BASENAME, SCALE) \
BASENAME##6 *= (DATA_TYPE)SCALE;
#define SCALE_ROW_8(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_7(DATA_TYPE, BASENAME, SCALE) \
BASENAME##7 *= (DATA_TYPE)SCALE;
#define SCALE_ROW_9(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_8(DATA_TYPE, BASENAME, SCALE) \
BASENAME##8 *= (DATA_TYPE)SCALE;
#define SCALE_ROW_10(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_9(DATA_TYPE, BASENAME, SCALE) \
BASENAME##9 *= (DATA_TYPE)SCALE;
#define SCALE_ROW_11(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_10(DATA_TYPE, BASENAME, SCALE) \
BASENAME##A *= (DATA_TYPE)SCALE;
#define SCALE_ROW_12(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_11(DATA_TYPE, BASENAME, SCALE) \
BASENAME##B *= (DATA_TYPE)SCALE;
#define SCALE_ROW_13(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_12(DATA_TYPE, BASENAME, SCALE) \
BASENAME##C *= (DATA_TYPE)SCALE;
#define SCALE_ROW_14(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_13(DATA_TYPE, BASENAME, SCALE) \
BASENAME##D *= (DATA_TYPE)SCALE;
#define SCALE_ROW_15(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_14(DATA_TYPE, BASENAME, SCALE) \
BASENAME##E *= (DATA_TYPE)SCALE;
#define SCALE_ROW_16(DATA_TYPE, BASENAME, SCALE) \
SCALE_ROW_15(DATA_TYPE, BASENAME, SCALE) \
BASENAME##F *= (DATA_TYPE)SCALE;
/** @} */ // end of group SCALE_ROW_n
/** Scale elements stored in a block (BASENAME)
* @name SCALE_BLOCK
*
* Supported cases are N=1,2,3,...,16
*
* @param[in] N The number of rows in the block
* @param[in] DATA_TYPE The data type of the block
* @param[in] BASENAME The basename of the block
* @param[in] SCALE The scale factor
* @{
*/
#define SCALE_BLOCK_STR(N, DATA_TYPE, BASENAME, SCALE) SCALE_ROW_##N(DATA_TYPE, BASENAME, SCALE)
#define SCALE_BLOCK(N, DATA_TYPE, BASENAME, SCALE) SCALE_BLOCK_STR(N, DATA_TYPE, BASENAME, SCALE)
/** @} */ // end of group SCALE_BLOCK
/** Create a new vector containing the values at the given index for a set of given vectors
* @name COLUMN_VECTORn
*
* @param[in] IDX_COL The index value
* @param[in] BASENAME The basename of the destination vectors
* @param[in] X The basename of the source vectors
* @param[in] TYPE The data type of the destination vectors
* @{
*/
#define COLUMN_VECTOR1(IDX_COL, BASENAME, X, TYPE) \
TYPE BASENAME##IDX_COL = (TYPE)((X##0).s##IDX_COL);
#define COLUMN_VECTOR2(IDX_COL, BASENAME, X, TYPE) \
VEC_DATA_TYPE(TYPE, 2) \
BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 2))((X##0).s##IDX_COL, (X##1).s##IDX_COL);
#define COLUMN_VECTOR3(IDX_COL, BASENAME, X, TYPE) \
VEC_DATA_TYPE(TYPE, 3) \
BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 3))((X##0).s##IDX_COL, (X##1).s##IDX_COL, (X##2).s##IDX_COL);
#define COLUMN_VECTOR4(IDX_COL, BASENAME, X, TYPE) \
VEC_DATA_TYPE(TYPE, 4) \
BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 4))((X##0).s##IDX_COL, (X##1).s##IDX_COL, (X##2).s##IDX_COL, (X##3).s##IDX_COL);
#define COLUMN_VECTOR8(IDX_COL, BASENAME, X, TYPE) \
VEC_DATA_TYPE(TYPE, 8) \
BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 8))((X##0).s##IDX_COL, (X##1).s##IDX_COL, (X##2).s##IDX_COL, (X##3).s##IDX_COL, (X##4).s##IDX_COL, (X##5).s##IDX_COL, (X##6).s##IDX_COL, (X##7).s##IDX_COL);
#define COLUMN_VECTOR16(IDX_COL, BASENAME, X, TYPE) \
VEC_DATA_TYPE(TYPE, 16) \
BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 16))((X##0).s##IDX_COL, (X##1).s##IDX_COL, (X##2).s##IDX_COL, (X##3).s##IDX_COL, (X##4).s##IDX_COL, (X##5).s##IDX_COL, (X##6).s##IDX_COL, (X##7).s##IDX_COL, (X##8).s##IDX_COL, (X##9).s##IDX_COL, (X##A).s##IDX_COL, (X##B).s##IDX_COL, (X##C).s##IDX_COL, (X##D).s##IDX_COL, (X##E).s##IDX_COL, (X##F).s##IDX_COL);
/** @} */ // end of group COLUMN_VECTORn
/** Create transposed vectors of the given vectors
* @name TRANSPOSE_K0Xn
*
* @param[in] K0 The size of the source vectors
* @param[in] BASENAME The basename of transposed vectors
* @param[in] B The basename of source vectors for transposition
* @param[in] TYPE The data type of the transposed vectors
* @{
*/
#define TRANSPOSE_K0X1(K0, BASENAME, B, TYPE) \
COLUMN_VECTOR(K0, 0, BASENAME, B, TYPE);
#define TRANSPOSE_K0X2(K0, BASENAME, B, TYPE) \
TRANSPOSE_K0X1(K0, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, 1, BASENAME, B, TYPE);
#define TRANSPOSE_K0X3(K0, BASENAME, B, TYPE) \
TRANSPOSE_K0X2(K0, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, 2, BASENAME, B, TYPE);
#define TRANSPOSE_K0X4(K0, BASENAME, B, TYPE) \
TRANSPOSE_K0X3(K0, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, 3, BASENAME, B, TYPE);
#define TRANSPOSE_K0X8(K0, BASENAME, B, TYPE) \
TRANSPOSE_K0X4(K0, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, 4, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, 5, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, 6, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, 7, BASENAME, B, TYPE);
#define TRANSPOSE_K0X16(K0, BASENAME, B, TYPE) \
TRANSPOSE_K0X8(K0, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, 8, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, 9, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, A, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, B, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, C, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, D, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, E, BASENAME, B, TYPE); \
COLUMN_VECTOR(K0, F, BASENAME, B, TYPE);
/** @} */ // end of group TRANSPOSE_K0Xn
/** Create column vectors to contain the values at the given index for a set of given vectors
*
* @param[in] K0 The number of source vectors
* @param[in] IDX_COL The index value
* @param[in] BASENAME The basename of the destination vectors
* @param[in] B The basename of the source vectors
* @param[in] TYPE The data type of the destination vectors
*/
#define COLUMN_VECTOR(K0, IDX_COL, BASENAME, B, TYPE) \
CONCAT(COLUMN_VECTOR, K0) \
(IDX_COL, BASENAME, B, TYPE);
/** Create transposed vectors form the given source vectors
*
* @param[in] K0 The size of source vectors
* @param[in] N0 The number of source vectors
* @param[in] BASENAME The basename of transposed vectors
* @param[in] B The basename of source vectors for transposition
* @param[in] TYPE The data type of the transposed vectors
*
*/
#define TRANSPOSE_K0XN0(K0, N0, BASENAME, B, TYPE) \
CONCAT(TRANSPOSE_K0X, N0) \
(K0, BASENAME, B, TYPE);
/** Add the variables (BIAS0 to BIASn-1) to the others (BASENAME0 to BASENAMEn-1)
* @name ADD_ROW_n
*
* @param[in] BASENAME The basename of the destination variables
* @param[in] BIAS The basename of the added variables
* @{
*/
#define ADD_ROW_1(BASENAME, BIAS) \
BASENAME##0 += BIAS##0;
#define ADD_ROW_2(BASENAME, BIAS) \
ADD_ROW_1(BASENAME, BIAS) \
BASENAME##1 += BIAS##1;
#define ADD_ROW_3(BASENAME, BIAS) \
ADD_ROW_2(BASENAME, BIAS) \
BASENAME##2 += BIAS##2;
#define ADD_ROW_4(BASENAME, BIAS) \
ADD_ROW_3(BASENAME, BIAS) \
BASENAME##3 += BIAS##3;
#define ADD_ROW_5(BASENAME, BIAS) \
ADD_ROW_4(BASENAME, BIAS) \
BASENAME##4 += BIAS##4;
#define ADD_ROW_6(BASENAME, BIAS) \
ADD_ROW_5(BASENAME, BIAS) \
BASENAME##5 += BIAS##5;
#define ADD_ROW_7(BASENAME, BIAS) \
ADD_ROW_6(BASENAME, BIAS) \
BASENAME##6 += BIAS##6;
#define ADD_ROW_8(BASENAME, BIAS) \
ADD_ROW_7(BASENAME, BIAS) \
BASENAME##7 += BIAS##7;
#define ADD_ROW_9(BASENAME, BIAS) \
ADD_ROW_8(BASENAME, BIAS) \
BASENAME##8 += BIAS##8;
#define ADD_ROW_10(BASENAME, BIAS) \
ADD_ROW_9(BASENAME, BIAS) \
BASENAME##9 += BIAS##9;
#define ADD_ROW_11(BASENAME, BIAS) \
ADD_ROW_10(BASENAME, BIAS) \
BASENAME##A += BIAS##A;
#define ADD_ROW_12(BASENAME, BIAS) \
ADD_ROW_11(BASENAME, BIAS) \
BASENAME##B += BIAS##B;
#define ADD_ROW_13(BASENAME, BIAS) \
ADD_ROW_12(BASENAME, BIAS) \
BASENAME##C += BIAS##C;
#define ADD_ROW_14(BASENAME, BIAS) \
ADD_ROW_13(BASENAME, BIAS) \
BASENAME##D += BIAS##D;
#define ADD_ROW_15(BASENAME, BIAS) \
ADD_ROW_14(BASENAME, BIAS) \
BASENAME##E += BIAS##E;
#define ADD_ROW_16(BASENAME, BIAS) \
ADD_ROW_15(BASENAME, BIAS) \
BASENAME##F += BIAS##F;
/** @} */ // end of group ADD_ROW_n
/** Add the block (BIAS) to another block (BASENAME)
* @name ADD_BLOCK
*
* Supported cases are N=1,2,3,...,16
*
* @param[in] N The number of vectors in the block
* @param[in] BASENAME The basename of the destination variables
* @param[in] BIAS The basename of the added variables
* @{
*/
#define ADD_BLOCK_STR(N, BASENAME, BIAS) ADD_ROW_##N(BASENAME, BIAS)
#define ADD_BLOCK(N, BASENAME, BIAS) ADD_BLOCK_STR(N, BASENAME, BIAS)
/** @} */ // end of group ADD_BLOCK
/** Broadcast (add single value) to the each element of the destination variables
* @name ADD_ROW_BROADCAST_n
*
* @param[in] BASENAME The basename of the destination variables
* @param[in] BIAS The variable containing the value to add
* @{
*/
#define ADD_ROW_BROADCAST_1(BASENAME, BIAS) \
BASENAME##0 += BIAS;
#define ADD_ROW_BROADCAST_2(BASENAME, BIAS) \
ADD_ROW_BROADCAST_1(BASENAME, BIAS) \
BASENAME##1 += BIAS;
#define ADD_ROW_BROADCAST_3(BASENAME, BIAS) \
ADD_ROW_BROADCAST_2(BASENAME, BIAS) \
BASENAME##2 += BIAS;
#define ADD_ROW_BROADCAST_4(BASENAME, BIAS) \
ADD_ROW_BROADCAST_3(BASENAME, BIAS) \
BASENAME##3 += BIAS;
#define ADD_ROW_BROADCAST_5(BASENAME, BIAS) \
ADD_ROW_BROADCAST_4(BASENAME, BIAS) \
BASENAME##4 += BIAS;
#define ADD_ROW_BROADCAST_6(BASENAME, BIAS) \
ADD_ROW_BROADCAST_5(BASENAME, BIAS) \
BASENAME##5 += BIAS;
#define ADD_ROW_BROADCAST_7(BASENAME, BIAS) \
ADD_ROW_BROADCAST_6(BASENAME, BIAS) \
BASENAME##6 += BIAS;
#define ADD_ROW_BROADCAST_8(BASENAME, BIAS) \
ADD_ROW_BROADCAST_7(BASENAME, BIAS) \
BASENAME##7 += BIAS;
#define ADD_ROW_BROADCAST_9(BASENAME, BIAS) \
ADD_ROW_BROADCAST_8(BASENAME, BIAS) \
BASENAME##8 += BIAS;
#define ADD_ROW_BROADCAST_10(BASENAME, BIAS) \
ADD_ROW_BROADCAST_9(BASENAME, BIAS) \
BASENAME##9 += BIAS;
#define ADD_ROW_BROADCAST_11(BASENAME, BIAS) \
ADD_ROW_BROADCAST_10(BASENAME, BIAS) \
BASENAME##A += BIAS;
#define ADD_ROW_BROADCAST_12(BASENAME, BIAS) \
ADD_ROW_BROADCAST_11(BASENAME, BIAS) \
BASENAME##B += BIAS;
#define ADD_ROW_BROADCAST_13(BASENAME, BIAS) \
ADD_ROW_BROADCAST_12(BASENAME, BIAS) \
BASENAME##C += BIAS;
#define ADD_ROW_BROADCAST_14(BASENAME, BIAS) \
ADD_ROW_BROADCAST_13(BASENAME, BIAS) \
BASENAME##D += BIAS;
#define ADD_ROW_BROADCAST_15(BASENAME, BIAS) \
ADD_ROW_BROADCAST_14(BASENAME, BIAS) \
BASENAME##E += BIAS;
#define ADD_ROW_BROADCAST_16(BASENAME, BIAS) \
ADD_ROW_BROADCAST_15(BASENAME, BIAS) \
BASENAME##F += BIAS;
/** Broadcast (add a value) to the each element of the destination block (BASENAME)
* @name ADD_BLOCK_BROADCAST
*
* Supported cases are N=1,2,3,...,16.
*
* @param[in] N The number of vectors in the block
* @param[in] BASENAME The basename of the destination variables
* @param[in] BIAS The variable containing the value to add
* @{
*/
#define ADD_BLOCK_BROADCAST_STR(N, BASENAME, BIAS) ADD_ROW_BROADCAST_##N(BASENAME, BIAS)
#define ADD_BLOCK_BROADCAST(N, BASENAME, BIAS) ADD_BLOCK_BROADCAST_STR(N, BASENAME, BIAS)
/** @} */ // end of group ADD_BLOCK_BROADCAST
/** Apply activation to the given variables
* @name ACTIVATION_ROW_n
*
* @param[in] ACTIVATION_TYPE The type of the activation
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] A_VAL Additional value required by the activation
* @param[in] B_VAL Additional value required by the activation
* @{
*/
#define ACTIVATION_ROW_1(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##0 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##0, A_VAL, B_VAL);
#define ACTIVATION_ROW_2(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_1(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##1 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##1, A_VAL, B_VAL);
#define ACTIVATION_ROW_3(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_2(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##2 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##2, A_VAL, B_VAL);
#define ACTIVATION_ROW_4(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_3(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##3 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##3, A_VAL, B_VAL);
#define ACTIVATION_ROW_5(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_4(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##4 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##4, A_VAL, B_VAL);
#define ACTIVATION_ROW_6(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_5(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##5 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##5, A_VAL, B_VAL);
#define ACTIVATION_ROW_7(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_6(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##6 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##6, A_VAL, B_VAL);
#define ACTIVATION_ROW_8(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_7(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##7 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##7, A_VAL, B_VAL);
#define ACTIVATION_ROW_9(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_8(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##8 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##8, A_VAL, B_VAL);
#define ACTIVATION_ROW_10(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_9(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##9 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##9, A_VAL, B_VAL);
#define ACTIVATION_ROW_11(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_10(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##A = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##A, A_VAL, B_VAL);
#define ACTIVATION_ROW_12(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_11(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##B = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##B, A_VAL, B_VAL);
#define ACTIVATION_ROW_13(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_12(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##C = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##C, A_VAL, B_VAL);
#define ACTIVATION_ROW_14(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_13(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##D = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##D, A_VAL, B_VAL);
#define ACTIVATION_ROW_15(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_14(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##E = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##E, A_VAL, B_VAL);
#define ACTIVATION_ROW_16(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
ACTIVATION_ROW_15(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) \
BASENAME##F = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, BASENAME##F, A_VAL, B_VAL);
/** @} */ // end of group ACTIVATION_ROW_n
/** Apply activation to a block (BASENAME)
* @name ACTIVATION_BLOCK
*
* Supported cases are N=1,2,3,...,16.
*
* @param[in] N The number of vectors in the block
* @param[in] ACTIVATION_TYPE The type of the activation
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] A_VAL Additional value required by the activation
* @param[in] B_VAL Additional value required by the activation
* @{
*/
#define ACTIVATION_BLOCK_STR(N, ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) ACTIVATION_ROW_##N(ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL)
#define ACTIVATION_BLOCK(N, ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL) ACTIVATION_BLOCK_STR(N, ACTIVATION_TYPE, DATA_TYPE, BASENAME, A_VAL, B_VAL)
/** @} */ // end of group ACTIVATION_BLOCK
/** Apply convert_<data_type> to the given variables
* @name CONVERT_ROW_n
*
* @param[in] N The size of the vectors
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME_SRC The basename of the source variables
* @param[in] BASENAME_DST The basename of the destination variables
*/
#define CONVERT_ROW_1(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##0 = CONVERT(BASENAME_SRC##0, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_2(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_1(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##1 = CONVERT(BASENAME_SRC##1, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_3(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_2(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##2 = CONVERT(BASENAME_SRC##2, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_4(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_3(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##3 = CONVERT(BASENAME_SRC##3, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_5(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_4(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##4 = CONVERT(BASENAME_SRC##4, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_6(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_5(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##5 = CONVERT(BASENAME_SRC##5, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_7(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_6(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##6 = CONVERT(BASENAME_SRC##6, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_8(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_7(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##7 = CONVERT(BASENAME_SRC##7, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_9(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_8(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##8 = CONVERT(BASENAME_SRC##8, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_10(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_9(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##9 = CONVERT(BASENAME_SRC##9, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_11(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_10(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##A = CONVERT(BASENAME_SRC##A, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_12(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_11(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##B = CONVERT(BASENAME_SRC##B, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_13(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_12(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##C = CONVERT(BASENAME_SRC##C, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_14(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_13(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##D = CONVERT(BASENAME_SRC##D, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_15(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_14(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##E = CONVERT(BASENAME_SRC##E, VEC_DATA_TYPE(DATA_TYPE, N));
#define CONVERT_ROW_16(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
CONVERT_ROW_15(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
VEC_DATA_TYPE(DATA_TYPE, N) \
BASENAME_DST##F = CONVERT(BASENAME_SRC##F, VEC_DATA_TYPE(DATA_TYPE, N));
/** @} */ // end of group CONVERT_ROW_n
/** Apply convert_<data_type> to a block (BASENAME_SRC) and save to another block (BASENAME_DST)
* @name CONVERT_BLOCK
*
* Supported cases N=1,2,3,...,16.
*
* @param[in] M The number of vectors to convert
* @param[in] N The size of the vectors
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME_SRC The basename of the source variables
* @param[in] BASENAME_DST The basename of the destination variables
*/
#define CONVERT_BLOCK_STR(M, N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) CONVERT_ROW_##M(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)
#define CONVERT_BLOCK(M, N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) CONVERT_BLOCK_STR(M, N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)
/** @} */ // end of group CONVERT_BLOCK
)"