115#ifndef EIGEN_ASSIGN_AOCL_H
116#define EIGEN_ASSIGN_AOCL_H
123template <
typename Xpr>
124struct aocl_is_contiguous
125 : bool_constant<(int(Xpr::Flags) & DirectAccessBit) && int(Xpr::InnerStrideAtCompileTime) == 1 &&
126 (int(evaluator<Xpr>::Flags) & LinearAccessBit)> {};
129template <
typename Dst,
typename Src>
130class aocl_assign_traits {
133 DstIsContiguous = aocl_is_contiguous<Dst>::value,
134 SrcIsContiguous = aocl_is_contiguous<Src>::value,
135 StorageOrdersAgree = (int(Dst::IsRowMajor) == int(Src::IsRowMajor)),
136 InnerSize = Dst::IsVectorAtCompileTime ?
int(Dst::SizeAtCompileTime)
137 : (Dst::Flags &
RowMajorBit) ?
int(Dst::ColsAtCompileTime)
138 : int(Dst::RowsAtCompileTime),
139 LargeEnough = (InnerSize == Dynamic) || (InnerSize >= EIGEN_AOCL_VML_THRESHOLD)
144 EnableAoclVML = DstIsContiguous && SrcIsContiguous && StorageOrdersAgree && LargeEnough,
145 Traversal = LinearTraversal
150template <
typename Dst,
typename Lhs,
typename Rhs>
151class aocl_assign_binary_traits {
154 DstIsContiguous = aocl_is_contiguous<Dst>::value,
155 LhsIsContiguous = aocl_is_contiguous<Lhs>::value,
156 RhsIsContiguous = aocl_is_contiguous<Rhs>::value,
158 (int(Dst::IsRowMajor) == int(Lhs::IsRowMajor)) && (
int(Dst::IsRowMajor) ==
int(Rhs::IsRowMajor)),
159 InnerSize = Dst::IsVectorAtCompileTime ? int(Dst::SizeAtCompileTime)
160 : (Dst::Flags &
RowMajorBit) ? int(Dst::ColsAtCompileTime)
161 : int(Dst::RowsAtCompileTime),
162 LargeEnough = (InnerSize == Dynamic) || (InnerSize >= EIGEN_AOCL_VML_THRESHOLD)
166 enum { EnableAoclVML = DstIsContiguous && LhsIsContiguous && RhsIsContiguous && StorageOrdersAgree && LargeEnough };
170#define EIGEN_AOCL_VML_UNARY_CALL_FLOAT(EIGENOP) \
171 template <typename DstXprType, typename SrcXprNested> \
172 struct Assignment<DstXprType, CwiseUnaryOp<scalar_##EIGENOP##_op<float>, SrcXprNested>, assign_op<float, float>, \
173 Dense2Dense, std::enable_if_t<aocl_assign_traits<DstXprType, SrcXprNested>::EnableAoclVML>> { \
174 using SrcXprType = CwiseUnaryOp<scalar_##EIGENOP##_op<float>, SrcXprNested>; \
175 static void run(DstXprType &dst, const SrcXprType &src, const assign_op<float, float> &func) { \
176 resize_if_allowed(dst, src, func); \
177 eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); \
178 Eigen::Index n = dst.size(); \
179 if (n <= 0) return; \
180 const float *input = reinterpret_cast<const float *>(src.nestedExpression().data()); \
181 float *output = reinterpret_cast<float *>(dst.data()); \
182 for (Eigen::Index i = 0; i < n; ++i) { \
183 output[i] = std::EIGENOP(input[i]); \
189#define EIGEN_AOCL_VML_UNARY_CALL_DOUBLE(EIGENOP, AOCLOP) \
190 template <typename DstXprType, typename SrcXprNested> \
191 struct Assignment<DstXprType, CwiseUnaryOp<scalar_##EIGENOP##_op<double>, SrcXprNested>, assign_op<double, double>, \
192 Dense2Dense, std::enable_if_t<aocl_assign_traits<DstXprType, SrcXprNested>::EnableAoclVML>> { \
193 using SrcXprType = CwiseUnaryOp<scalar_##EIGENOP##_op<double>, SrcXprNested>; \
194 static void run(DstXprType &dst, const SrcXprType &src, const assign_op<double, double> &func) { \
195 resize_if_allowed(dst, src, func); \
196 eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); \
197 Eigen::Index n = dst.size(); \
198 eigen_assert(n <= INT_MAX && "AOCL does not support arrays larger than INT_MAX"); \
199 if (n <= 0) return; \
200 const double *input = reinterpret_cast<const double *>(src.nestedExpression().data()); \
201 double *output = reinterpret_cast<double *>(dst.data()); \
202 int aocl_n = internal::convert_index<int>(n); \
203 AOCLOP(aocl_n, const_cast<double *>(input), output); \
211EIGEN_AOCL_VML_UNARY_CALL_DOUBLE(exp2, amd_vrda_exp2)
212EIGEN_AOCL_VML_UNARY_CALL_DOUBLE(exp, amd_vrda_exp)
213EIGEN_AOCL_VML_UNARY_CALL_DOUBLE(sin, amd_vrda_sin)
214EIGEN_AOCL_VML_UNARY_CALL_DOUBLE(cos, amd_vrda_cos)
215EIGEN_AOCL_VML_UNARY_CALL_DOUBLE(sqrt, amd_vrda_sqrt)
216EIGEN_AOCL_VML_UNARY_CALL_DOUBLE(cbrt, amd_vrda_cbrt)
217EIGEN_AOCL_VML_UNARY_CALL_DOUBLE(abs, amd_vrda_fabs)
218EIGEN_AOCL_VML_UNARY_CALL_DOUBLE(log, amd_vrda_log)
219EIGEN_AOCL_VML_UNARY_CALL_DOUBLE(log10, amd_vrda_log10)
220EIGEN_AOCL_VML_UNARY_CALL_DOUBLE(log2, amd_vrda_log2)
223#define EIGEN_AOCL_VML_BINARY_CALL_FLOAT(EIGENOP, STDFUNC) \
224 template <typename DstXprType, typename LhsXprNested, typename RhsXprNested> \
226 DstXprType, CwiseBinaryOp<scalar_##EIGENOP##_op<float, float>, LhsXprNested, RhsXprNested>, \
227 assign_op<float, float>, Dense2Dense, \
228 std::enable_if_t<aocl_assign_binary_traits<DstXprType, LhsXprNested, RhsXprNested>::EnableAoclVML>> { \
229 using SrcXprType = CwiseBinaryOp<scalar_##EIGENOP##_op<float, float>, LhsXprNested, RhsXprNested>; \
230 static void run(DstXprType &dst, const SrcXprType &src, const assign_op<float, float> &func) { \
231 resize_if_allowed(dst, src, func); \
232 eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); \
233 Eigen::Index n = dst.size(); \
234 if (n <= 0) return; \
235 const float *lhs = reinterpret_cast<const float *>(src.lhs().data()); \
236 const float *rhs = reinterpret_cast<const float *>(src.rhs().data()); \
237 float *output = reinterpret_cast<float *>(dst.data()); \
238 for (Eigen::Index i = 0; i < n; ++i) { \
239 output[i] = STDFUNC(lhs[i], rhs[i]); \
245#define EIGEN_AOCL_VML_BINARY_CALL_DOUBLE(EIGENOP, AOCLOP) \
246 template <typename DstXprType, typename LhsXprNested, typename RhsXprNested> \
248 DstXprType, CwiseBinaryOp<scalar_##EIGENOP##_op<double, double>, LhsXprNested, RhsXprNested>, \
249 assign_op<double, double>, Dense2Dense, \
250 std::enable_if_t<aocl_assign_binary_traits<DstXprType, LhsXprNested, RhsXprNested>::EnableAoclVML>> { \
251 using SrcXprType = CwiseBinaryOp<scalar_##EIGENOP##_op<double, double>, LhsXprNested, RhsXprNested>; \
252 static void run(DstXprType &dst, const SrcXprType &src, const assign_op<double, double> &func) { \
253 resize_if_allowed(dst, src, func); \
254 eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); \
255 Eigen::Index n = dst.size(); \
256 eigen_assert(n <= INT_MAX && "AOCL does not support arrays larger than INT_MAX"); \
257 if (n <= 0) return; \
258 const double *lhs = reinterpret_cast<const double *>(src.lhs().data()); \
259 const double *rhs = reinterpret_cast<const double *>(src.rhs().data()); \
260 double *output = reinterpret_cast<double *>(dst.data()); \
261 int aocl_n = internal::convert_index<int>(n); \
262 AOCLOP(aocl_n, const_cast<double *>(lhs), const_cast<double *>(rhs), output); \
271EIGEN_AOCL_VML_BINARY_CALL_DOUBLE(sum, amd_vrda_add)
272EIGEN_AOCL_VML_BINARY_CALL_DOUBLE(pow, amd_vrda_pow)
273EIGEN_AOCL_VML_BINARY_CALL_DOUBLE(max, amd_vrda_fmax)
274EIGEN_AOCL_VML_BINARY_CALL_DOUBLE(min, amd_vrda_fmin)
constexpr unsigned int RowMajorBit
Definition Constants.h:71