From b98af44fcf89b9946e1de438b1f5527e6bf28905 Mon Sep 17 00:00:00 2001 From: Ivan Shapovalov Date: Sat, 22 Oct 2016 05:14:19 +0300 Subject: treewide: use C++ exceptions properly Since the codebase is designed around proper C++ idioms such as RAII, it makes sense to only use C++ exceptions internally instead of mixing exceptions and error codes. The exceptions are now caught at top level to preserve compatibility with the existing error code-based API. Note that we deliberately do not catch C++ runtime errors (such as `std::bad_alloc`) nor logic errors (aka failed assertions) because no actual handling can ever happen for such errors. However, in the C interface we do catch _all_ exceptions (...) and convert them into a wild-card error code. --- src/routines/level3/xsyr2k.cpp | 214 +++++++++++++++++++---------------------- 1 file changed, 98 insertions(+), 116 deletions(-) (limited to 'src/routines/level3/xsyr2k.cpp') diff --git a/src/routines/level3/xsyr2k.cpp b/src/routines/level3/xsyr2k.cpp index 158cd9e5..b10ee586 100644 --- a/src/routines/level3/xsyr2k.cpp +++ b/src/routines/level3/xsyr2k.cpp @@ -39,7 +39,7 @@ Xsyr2k::Xsyr2k(Queue &queue, EventPointer event, const std::string &name): // The main routine template -StatusCode Xsyr2k::DoSyr2k(const Layout layout, const Triangle triangle, const Transpose ab_transpose, +void Xsyr2k::DoSyr2k(const Layout layout, const Triangle triangle, const Transpose ab_transpose, const size_t n, const size_t k, const T alpha, const Buffer &a_buffer, const size_t a_offset, const size_t a_ld, @@ -48,7 +48,7 @@ StatusCode Xsyr2k::DoSyr2k(const Layout layout, const Triangle triangle, cons const Buffer &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); } // Computes whether or not the matrices are transposed in memory. This is based on their layout // (row or column-major) and whether or not they are requested to be pre-transposed. @@ -67,12 +67,9 @@ StatusCode Xsyr2k::DoSyr2k(const Layout layout, const Triangle triangle, cons // matrix A cannot be less than N when rotated, or less than K when not-rotated // matrix B cannot be less than N when rotated, or less than K when not-rotated // matrix C cannot be less than N - auto status = TestMatrixA(ab_one, ab_two, a_buffer, a_offset, a_ld); - if (ErrorIn(status)) { return status; } - status = TestMatrixB(ab_one, ab_two, b_buffer, b_offset, b_ld); - if (ErrorIn(status)) { return status; } - status = TestMatrixC(n, n, c_buffer, c_offset, c_ld); - if (ErrorIn(status)) { return status; } + TestMatrixA(ab_one, ab_two, a_buffer, a_offset, a_ld); + TestMatrixB(ab_one, ab_two, b_buffer, b_offset, b_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"]); @@ -81,114 +78,99 @@ StatusCode Xsyr2k::DoSyr2k(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(), routine_name_); - - // Determines whether or not temporary matrices are needed - auto a_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && - ab_rotated == false; - auto b_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && - ab_rotated == false; - - // Creates the temporary matrices - auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto b_temp = (b_no_temp) ? b_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto c_temp = Buffer(context_, n_ceiled*n_ceiled); - - // Events of all kernels (including pre/post processing kernels) - auto eventWaitList = std::vector(); - auto emptyEventList = std::vector(); - - // Runs the pre-processing kernels. This transposes the matrices A and B, but also pads zeros to - // 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. - if (!a_no_temp) { - auto eventProcessA = Event(); - status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList, - ab_one, ab_two, a_ld, a_offset, a_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, a_temp, - ConstantOne(), program, - true, ab_rotated, false); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventProcessA); - } - if (!b_no_temp) { - auto eventProcessB = Event(); - status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList, - ab_one, ab_two, b_ld, b_offset, b_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, b_temp, - ConstantOne(), program, - true, ab_rotated, false); - if (ErrorIn(status)) { return status; } - 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(); - 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(), program, - true, c_rotated, false); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventProcessC); - - // 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(n_ceiled)); - kernel.SetArgument(1, static_cast(k_ceiled)); - kernel.SetArgument(2, GetRealArg(alpha)); - kernel.SetArgument(3, GetRealArg(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{ - (n_ceiled * db_["MDIMC"]) / db_["MWG"], - (n_ceiled * db_["NDIMC"]) / db_["NWG"] - }; - auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; - - // Launches the kernel - auto eventKernel1 = Event(); - status = RunKernel(kernel, queue_, device_, global, local, eventKernel1.pointer(), eventWaitList); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventKernel1); - - // Swaps the arguments for matrices A and B, and sets 'beta' to 1 - auto one = static_cast(1); - kernel.SetArgument(3, GetRealArg(one)); - kernel.SetArgument(4, b_temp()); - kernel.SetArgument(5, a_temp()); - - // Runs the kernel again - auto eventKernel2 = Event(); - status = RunKernel(kernel, queue_, device_, global, local, eventKernel2.pointer(), eventWaitList); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventKernel2); - - // 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(), program, - false, c_rotated, false, upper, lower, false); - 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(), routine_name_); + + // Determines whether or not temporary matrices are needed + auto a_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && + ab_rotated == false; + auto b_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && + ab_rotated == false; + + // Creates the temporary matrices + auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); + auto b_temp = (b_no_temp) ? b_buffer : Buffer(context_, k_ceiled*n_ceiled); + auto c_temp = Buffer(context_, n_ceiled*n_ceiled); + + // Events of all kernels (including pre/post processing kernels) + auto eventWaitList = std::vector(); + auto emptyEventList = std::vector(); + + // Runs the pre-processing kernels. This transposes the matrices A and B, but also pads zeros to + // 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. + if (!a_no_temp) { + auto eventProcessA = Event(); + PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList, + ab_one, ab_two, a_ld, a_offset, a_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, a_temp, + ConstantOne(), program, + true, ab_rotated, false); + eventWaitList.push_back(eventProcessA); + } + if (!b_no_temp) { + auto eventProcessB = Event(); + PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList, + ab_one, ab_two, b_ld, b_offset, b_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, b_temp, + ConstantOne(), program, + true, ab_rotated, false); + 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(), 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(n_ceiled)); + kernel.SetArgument(1, static_cast(k_ceiled)); + kernel.SetArgument(2, GetRealArg(alpha)); + kernel.SetArgument(3, GetRealArg(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{ + (n_ceiled * db_["MDIMC"]) / db_["MWG"], + (n_ceiled * db_["NDIMC"]) / db_["NWG"] + }; + auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; + + // Launches the kernel + auto eventKernel1 = Event(); + RunKernel(kernel, queue_, device_, global, local, eventKernel1.pointer(), eventWaitList); + eventWaitList.push_back(eventKernel1); + + // Swaps the arguments for matrices A and B, and sets 'beta' to 1 + auto one = static_cast(1); + kernel.SetArgument(3, GetRealArg(one)); + kernel.SetArgument(4, b_temp()); + kernel.SetArgument(5, a_temp()); + + // Runs the kernel again + auto eventKernel2 = Event(); + RunKernel(kernel, queue_, device_, global, local, eventKernel2.pointer(), eventWaitList); + eventWaitList.push_back(eventKernel2); + + // 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(), program, + false, c_rotated, false, upper, lower, false); } // ================================================================================================= -- cgit v1.2.3