diff options
Diffstat (limited to 'src/routines/level3/xgemm.cpp')
| -rw-r--r-- | src/routines/level3/xgemm.cpp | 61 |
1 files changed, 34 insertions, 27 deletions
diff --git a/src/routines/level3/xgemm.cpp b/src/routines/level3/xgemm.cpp index 9ea5559c..0b8e768f 100644 --- a/src/routines/level3/xgemm.cpp +++ b/src/routines/level3/xgemm.cpp @@ -34,6 +34,7 @@ Xgemm<T>::Xgemm(Queue &queue, EventPointer event, const std::string &name): #include "../../kernels/level3/convert_hermitian.opencl" #include "../../kernels/level3/xgemm_part1.opencl" #include "../../kernels/level3/xgemm_part2.opencl" + #include "../../kernels/level3/xgemm_part3.opencl" ; } @@ -63,9 +64,12 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout, const auto b_rotated = (layout == Layout::kColMajor && b_transpose != Transpose::kNo) || (layout == Layout::kRowMajor && b_transpose == Transpose::kNo); const auto c_rotated = (layout == Layout::kRowMajor); - const auto a_do_transpose = a_rotated; - const auto b_do_transpose = !b_rotated; - const auto c_do_transpose = c_rotated; + static const auto a_want_rotated = false; + static const auto b_want_rotated = true; + static const auto c_want_rotated = false; + const auto a_do_transpose = a_rotated != a_want_rotated; + const auto b_do_transpose = b_rotated != b_want_rotated; + const auto c_do_transpose = c_rotated != c_want_rotated; // In case of complex data-types, the transpose can also become a conjugate transpose const auto a_conjugate = (a_transpose == Transpose::kConjugate); @@ -99,6 +103,15 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout, const auto n_ceiled = Ceil(n, db_["NWG"]); const auto k_ceiled = Ceil(k, db_["KWG"]); + // Computes the first and second "internal" (ceiled) dimensions of the 3 matrices taking into account + // whether the matrices need to be rotated or not for the kernel. + const auto a_one_i = (a_want_rotated) ? k_ceiled : m_ceiled; + const auto a_two_i = (a_want_rotated) ? m_ceiled : k_ceiled; + const auto b_one_i = (b_want_rotated) ? n_ceiled : k_ceiled; + const auto b_two_i = (b_want_rotated) ? k_ceiled : n_ceiled; + const auto c_one_i = (c_want_rotated) ? n_ceiled : m_ceiled; + const auto c_two_i = (c_want_rotated) ? m_ceiled : n_ceiled; + // The padded/transposed input/output matrices: if memory allocation fails, throw an exception try { @@ -106,23 +119,17 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout, const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_); // Determines whether or not temporary matrices are needed - auto a_no_temp = a_one == m_ceiled && a_two == k_ceiled && a_ld == m_ceiled && a_offset == 0 && + auto a_no_temp = a_one == a_one_i && a_two == a_two_i && a_ld == a_one && a_offset == 0 && a_do_transpose == false && a_conjugate == false; - auto b_no_temp = b_one == n_ceiled && b_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && + auto b_no_temp = b_one == b_one_i && b_two == b_two_i && b_ld == b_one && b_offset == 0 && b_do_transpose == false && b_conjugate == false; - auto c_no_temp = c_one == m_ceiled && c_two == n_ceiled && c_ld == m_ceiled && c_offset == 0 && + auto c_no_temp = c_one == c_one_i && c_two == c_two_i && c_ld == c_one && c_offset == 0 && c_do_transpose == false; // Creates the temporary matrices - const auto a_temp = (a_no_temp) ? a_buffer : Buffer<T>(context_, k_ceiled*m_ceiled); - const auto b_temp = (b_no_temp) ? b_buffer : Buffer<T>(context_, k_ceiled*n_ceiled); - const auto c_temp = (c_no_temp) ? c_buffer : Buffer<T>(context_, m_ceiled*n_ceiled); - - // Upload the scalar arguments as constant buffers to the device (needed for half-precision) - auto alpha_buffer = Buffer<T>(context_, 1); - auto beta_buffer = Buffer<T>(context_, 1); - alpha_buffer.Write(queue_, 1, &alpha); - beta_buffer.Write(queue_, 1, &beta); + const auto a_temp = (a_no_temp) ? a_buffer : Buffer<T>(context_, a_one_i*a_two_i); + const auto b_temp = (b_no_temp) ? b_buffer : Buffer<T>(context_, b_one_i*b_two_i); + const auto c_temp = (c_no_temp) ? c_buffer : Buffer<T>(context_, c_one_i*c_two_i); // Events of all kernels (including pre/post processing kernels) auto eventWaitList = std::vector<Event>(); @@ -133,9 +140,9 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout, // case nothing has to be done, these kernels can be skipped. if (!a_no_temp) { auto eventProcessA = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList, + status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList, a_one, a_two, a_ld, a_offset, a_buffer, - m_ceiled, k_ceiled, m_ceiled, 0, a_temp, + a_one_i, a_two_i, a_one_i, 0, a_temp, ConstantOne<T>(), program, true, a_do_transpose, a_conjugate); if (ErrorIn(status)) { return status; } @@ -145,9 +152,9 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout, // As above, but now for matrix B if (!b_no_temp) { auto eventProcessB = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB.pointer(), emptyEventList, + status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList, b_one, b_two, b_ld, b_offset, b_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, b_temp, + b_one_i, b_two_i, b_one_i, 0, b_temp, ConstantOne<T>(), program, true, b_do_transpose, b_conjugate); if (ErrorIn(status)) { return status; } @@ -157,9 +164,9 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout, // As above, but now for matrix C. This is only necessary if C is used both as input and output. if (!c_no_temp && beta != static_cast<T>(0)) { auto eventProcessC = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, + status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList, c_one, c_two, c_ld, c_offset, c_buffer, - m_ceiled, n_ceiled, m_ceiled, 0, c_temp, + c_one_i, c_two_i, c_one_i, 0, c_temp, ConstantOne<T>(), program, true, c_do_transpose, false); if (ErrorIn(status)) { return status; } @@ -174,16 +181,16 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout, kernel.SetArgument(0, static_cast<int>(m_ceiled)); kernel.SetArgument(1, static_cast<int>(n_ceiled)); kernel.SetArgument(2, static_cast<int>(k_ceiled)); - kernel.SetArgument(3, alpha_buffer()); - kernel.SetArgument(4, beta_buffer()); + kernel.SetArgument(3, GetRealArg(alpha)); + kernel.SetArgument(4, GetRealArg(beta)); kernel.SetArgument(5, a_temp()); kernel.SetArgument(6, b_temp()); kernel.SetArgument(7, c_temp()); // Computes the global and local thread sizes const auto global = std::vector<size_t>{ - (m_ceiled * db_["MDIMC"]) / db_["MWG"], - (n_ceiled * db_["NDIMC"]) / db_["NWG"] + (c_one_i * db_["MDIMC"]) / db_["MWG"], + (c_two_i * db_["NDIMC"]) / db_["NWG"] }; const auto local = std::vector<size_t>{db_["MDIMC"], db_["NDIMC"]}; @@ -196,8 +203,8 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout, // Runs the post-processing kernel if needed if (!c_no_temp) { eventWaitList.push_back(eventKernel); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, - m_ceiled, n_ceiled, m_ceiled, 0, c_temp, + status = PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList, + c_one_i, c_two_i, c_one_i, 0, c_temp, c_one, c_two, c_ld, c_offset, c_buffer, ConstantOne<T>(), program, false, c_do_transpose, false); |