diff options
Diffstat (limited to 'src/routines/level3/xherk.cpp')
-rw-r--r-- | src/routines/level3/xherk.cpp | 196 |
1 files changed, 90 insertions, 106 deletions
diff --git a/src/routines/level3/xherk.cpp b/src/routines/level3/xherk.cpp index 3063f3bc..6e36714e 100644 --- a/src/routines/level3/xherk.cpp +++ b/src/routines/level3/xherk.cpp @@ -39,7 +39,7 @@ Xherk<T,U>::Xherk(Queue &queue, EventPointer event, const std::string &name): // The main routine template <typename T, typename U> -StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Transpose a_transpose, +void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Transpose a_transpose, const size_t n, const size_t k, const U alpha, const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, @@ -47,7 +47,7 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld) { // Makes sure all dimensions are larger than zero - if ((n == 0) || (k == 0) ) { return StatusCode::kInvalidDimension; } + if ((n == 0) || (k == 0) ) { throw BLASError(StatusCode::kInvalidDimension); } // Determines whether to apply the conjugate transpose to matrix B (argument: no transpose) or // to matrix A (argument: conjugate transpose) @@ -70,10 +70,8 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons // space. Also tests that the leading dimensions of: // matrix A cannot be less than N when rotated, or less than K when not-rotated // matrix C cannot be less than N - auto status = TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld); - if (ErrorIn(status)) { return status; } - status = TestMatrixC(n, n, c_buffer, c_offset, c_ld); - if (ErrorIn(status)) { return status; } + TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld); + TestMatrixC(n, n, c_buffer, c_offset, c_ld); // Calculates the ceiled versions of n and k auto n_ceiled = Ceil(n, db_["NWG"]); @@ -82,106 +80,92 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons // Decides which kernel to run: the upper-triangular or lower-triangular version auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower"; - // The padded/transposed input/output matrices: if memory allocation fails, throw an exception - try { - - // Loads the program from the database - const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_); - - // Determines whether or not temporary matrices are needed - auto a_no_temp = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && - a_rotated == false && a_conjugate == false; - auto b_no_temp = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && - a_rotated == false && b_conjugate == false; - - // Creates the temporary matrices - auto a_temp = (a_no_temp) ? a_buffer : Buffer<T>(context_, k_ceiled*n_ceiled); - auto b_temp = (b_no_temp) ? a_buffer : Buffer<T>(context_, k_ceiled*n_ceiled); - auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled); - - // Convert the arguments to complex versions - auto complex_alpha = T{alpha, static_cast<U>(0.0)}; - auto complex_beta = T{beta, static_cast<U>(0.0)}; - - // Events of all kernels (including pre/post processing kernels) - auto eventWaitList = std::vector<Event>(); - auto emptyEventList = std::vector<Event>(); - - // Runs the pre-processing kernel for matrix A. This transposes the matrix, but also pads zeros - // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In - // case nothing has to be done, these kernels can be skipped. Two copies are created. - if (!a_no_temp) { - auto eventProcessA = Event(); - status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList, - a_one, a_two, a_ld, a_offset, a_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, a_temp, - ConstantOne<T>(), program, - true, a_rotated, a_conjugate); - eventWaitList.push_back(eventProcessA); - if (ErrorIn(status)) { return status; } - } - if (!b_no_temp) { - auto eventProcessB = Event(); - status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList, - a_one, a_two, a_ld, a_offset, a_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, b_temp, - ConstantOne<T>(), program, - true, a_rotated, b_conjugate); - eventWaitList.push_back(eventProcessB); - if (ErrorIn(status)) { return status; } - } - - // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to - // modify the other triangle. - auto eventProcessC = Event(); - status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList, - n, n, c_ld, c_offset, c_buffer, - n_ceiled, n_ceiled, n_ceiled, 0, c_temp, - ConstantOne<T>(), program, - true, c_rotated, false); - eventWaitList.push_back(eventProcessC); - if (ErrorIn(status)) { return status; } - - // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary - try { - auto kernel = Kernel(program, kernel_name); - - // Sets the kernel arguments - kernel.SetArgument(0, static_cast<int>(n_ceiled)); - kernel.SetArgument(1, static_cast<int>(k_ceiled)); - kernel.SetArgument(2, GetRealArg(complex_alpha)); - kernel.SetArgument(3, GetRealArg(complex_beta)); - kernel.SetArgument(4, a_temp()); - kernel.SetArgument(5, b_temp()); - kernel.SetArgument(6, c_temp()); - - // Computes the global and local thread sizes - auto global = std::vector<size_t>{ - (n_ceiled * db_["MDIMC"]) / db_["MWG"], - (n_ceiled * db_["NDIMC"]) / db_["NWG"] - }; - auto local = std::vector<size_t>{db_["MDIMC"], db_["NDIMC"]}; - - // Launches the kernel - auto eventKernel = Event(); - status = RunKernel(kernel, queue_, device_, global, local, eventKernel.pointer(), eventWaitList); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventKernel); - - // Runs the post-processing kernel - auto upper = (triangle == Triangle::kUpper); - auto lower = (triangle == Triangle::kLower); - status = PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList, - n_ceiled, n_ceiled, n_ceiled, 0, c_temp, - n, n, c_ld, c_offset, c_buffer, - ConstantOne<T>(), program, - false, c_rotated, false, upper, lower, true); - if (ErrorIn(status)) { return status; } - - // Successfully finished the computation - return StatusCode::kSuccess; - } catch (...) { return StatusCode::kInvalidKernel; } - } catch (...) { return StatusCode::kTempBufferAllocFailure; } + // Loads the program from the database + const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_); + + // Determines whether or not temporary matrices are needed + auto a_no_temp = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && + a_rotated == false && a_conjugate == false; + auto b_no_temp = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && + a_rotated == false && b_conjugate == false; + + // Creates the temporary matrices + auto a_temp = (a_no_temp) ? a_buffer : Buffer<T>(context_, k_ceiled*n_ceiled); + auto b_temp = (b_no_temp) ? a_buffer : Buffer<T>(context_, k_ceiled*n_ceiled); + auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled); + + // Convert the arguments to complex versions + auto complex_alpha = T{alpha, static_cast<U>(0.0)}; + auto complex_beta = T{beta, static_cast<U>(0.0)}; + + // Events of all kernels (including pre/post processing kernels) + auto eventWaitList = std::vector<Event>(); + auto emptyEventList = std::vector<Event>(); + + // Runs the pre-processing kernel for matrix A. This transposes the matrix, but also pads zeros + // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In + // case nothing has to be done, these kernels can be skipped. Two copies are created. + if (!a_no_temp) { + auto eventProcessA = Event(); + PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList, + a_one, a_two, a_ld, a_offset, a_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, a_temp, + ConstantOne<T>(), program, + true, a_rotated, a_conjugate); + eventWaitList.push_back(eventProcessA); + } + if (!b_no_temp) { + auto eventProcessB = Event(); + PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList, + a_one, a_two, a_ld, a_offset, a_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, b_temp, + ConstantOne<T>(), program, + true, a_rotated, b_conjugate); + eventWaitList.push_back(eventProcessB); + } + + // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to + // modify the other triangle. + auto eventProcessC = Event(); + PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList, + n, n, c_ld, c_offset, c_buffer, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + ConstantOne<T>(), program, + true, c_rotated, false); + eventWaitList.push_back(eventProcessC); + + // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary + auto kernel = Kernel(program, kernel_name); + + // Sets the kernel arguments + kernel.SetArgument(0, static_cast<int>(n_ceiled)); + kernel.SetArgument(1, static_cast<int>(k_ceiled)); + kernel.SetArgument(2, GetRealArg(complex_alpha)); + kernel.SetArgument(3, GetRealArg(complex_beta)); + kernel.SetArgument(4, a_temp()); + kernel.SetArgument(5, b_temp()); + kernel.SetArgument(6, c_temp()); + + // Computes the global and local thread sizes + auto global = std::vector<size_t>{ + (n_ceiled * db_["MDIMC"]) / db_["MWG"], + (n_ceiled * db_["NDIMC"]) / db_["NWG"] + }; + auto local = std::vector<size_t>{db_["MDIMC"], db_["NDIMC"]}; + + // Launches the kernel + auto eventKernel = Event(); + RunKernel(kernel, queue_, device_, global, local, eventKernel.pointer(), eventWaitList); + eventWaitList.push_back(eventKernel); + + // Runs the post-processing kernel + auto upper = (triangle == Triangle::kUpper); + auto lower = (triangle == Triangle::kLower); + PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + n, n, c_ld, c_offset, c_buffer, + ConstantOne<T>(), program, + false, c_rotated, false, upper, lower, true); } // ================================================================================================= |