diff options
Diffstat (limited to 'src/routines')
38 files changed, 194 insertions, 85 deletions
diff --git a/src/routines/level1/xaxpy.cc b/src/routines/level1/xaxpy.cc index c5acaf49..37d23543 100644 --- a/src/routines/level1/xaxpy.cc +++ b/src/routines/level1/xaxpy.cc @@ -29,7 +29,7 @@ template <> const Precision Xaxpy<double2>::precision_ = Precision::kComplexDoub // Constructor: forwards to base class constructor template <typename T> -Xaxpy<T>::Xaxpy(Queue &queue, Event &event, const std::string &name): +Xaxpy<T>::Xaxpy(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Xaxpy"}, precision_) { source_string_ = #include "../../kernels/level1/level1.opencl" @@ -89,13 +89,13 @@ StatusCode Xaxpy<T>::DoAxpy(const size_t n, const T alpha, if (use_fast_kernel) { auto global = std::vector<size_t>{CeilDiv(n, db_["WPT"]*db_["VW"])}; auto local = std::vector<size_t>{db_["WGS"]}; - status = RunKernel(kernel, global, local); + status = RunKernel(kernel, global, local, event_); } else { auto n_ceiled = Ceil(n, db_["WGS"]*db_["WPT"]); auto global = std::vector<size_t>{n_ceiled/db_["WPT"]}; auto local = std::vector<size_t>{db_["WGS"]}; - status = RunKernel(kernel, global, local); + status = RunKernel(kernel, global, local, event_); } if (ErrorIn(status)) { return status; } diff --git a/src/routines/level1/xcopy.cc b/src/routines/level1/xcopy.cc index 8c7f8671..04508383 100644 --- a/src/routines/level1/xcopy.cc +++ b/src/routines/level1/xcopy.cc @@ -29,7 +29,7 @@ template <> const Precision Xcopy<double2>::precision_ = Precision::kComplexDoub // Constructor: forwards to base class constructor template <typename T> -Xcopy<T>::Xcopy(Queue &queue, Event &event, const std::string &name): +Xcopy<T>::Xcopy(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Xaxpy"}, precision_) { source_string_ = #include "../../kernels/level1/level1.opencl" @@ -87,13 +87,13 @@ StatusCode Xcopy<T>::DoCopy(const size_t n, if (use_fast_kernel) { auto global = std::vector<size_t>{CeilDiv(n, db_["WPT"]*db_["VW"])}; auto local = std::vector<size_t>{db_["WGS"]}; - status = RunKernel(kernel, global, local); + status = RunKernel(kernel, global, local, event_); } else { auto n_ceiled = Ceil(n, db_["WGS"]*db_["WPT"]); auto global = std::vector<size_t>{n_ceiled/db_["WPT"]}; auto local = std::vector<size_t>{db_["WGS"]}; - status = RunKernel(kernel, global, local); + status = RunKernel(kernel, global, local, event_); } if (ErrorIn(status)) { return status; } diff --git a/src/routines/level1/xdot.cc b/src/routines/level1/xdot.cc index e22b0f8b..4813a004 100644 --- a/src/routines/level1/xdot.cc +++ b/src/routines/level1/xdot.cc @@ -29,7 +29,7 @@ template <> const Precision Xdot<double2>::precision_ = Precision::kComplexDoubl // Constructor: forwards to base class constructor template <typename T> -Xdot<T>::Xdot(Queue &queue, Event &event, const std::string &name): +Xdot<T>::Xdot(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Xdot"}, precision_) { source_string_ = #include "../../kernels/level1/xdot.opencl" @@ -78,11 +78,16 @@ StatusCode Xdot<T>::DoDot(const size_t n, kernel1.SetArgument(7, temp_buffer()); kernel1.SetArgument(8, static_cast<int>(do_conjugate)); + // Event waiting list + auto eventWaitList = std::vector<Event>(); + // Launches the main kernel auto global1 = std::vector<size_t>{db_["WGS1"]*temp_size}; auto local1 = std::vector<size_t>{db_["WGS1"]}; - status = RunKernel(kernel1, global1, local1); + auto kernelEvent = Event(); + status = RunKernel(kernel1, global1, local1, kernelEvent.pointer()); if (ErrorIn(status)) { return status; } + eventWaitList.push_back(kernelEvent); // Sets the arguments for the epilogue kernel kernel2.SetArgument(0, temp_buffer()); @@ -92,7 +97,7 @@ StatusCode Xdot<T>::DoDot(const size_t n, // Launches the epilogue kernel auto global2 = std::vector<size_t>{db_["WGS2"]}; auto local2 = std::vector<size_t>{db_["WGS2"]}; - status = RunKernel(kernel2, global2, local2); + status = RunKernel(kernel2, global2, local2, event_, eventWaitList); if (ErrorIn(status)) { return status; } // Succesfully finished the computation diff --git a/src/routines/level1/xdotc.cc b/src/routines/level1/xdotc.cc index f414f556..b3a01079 100644 --- a/src/routines/level1/xdotc.cc +++ b/src/routines/level1/xdotc.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xdotc<T>::Xdotc(Queue &queue, Event &event, const std::string &name): +Xdotc<T>::Xdotc(Queue &queue, EventPointer event, const std::string &name): Xdot<T>(queue, event, name) { } diff --git a/src/routines/level1/xdotu.cc b/src/routines/level1/xdotu.cc index 28d9b730..8dded6e0 100644 --- a/src/routines/level1/xdotu.cc +++ b/src/routines/level1/xdotu.cc @@ -20,7 +20,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xdotu<T>::Xdotu(Queue &queue, Event &event, const std::string &name): +Xdotu<T>::Xdotu(Queue &queue, EventPointer event, const std::string &name): Xdot<T>(queue, event, name) { } diff --git a/src/routines/level1/xnrm2.cc b/src/routines/level1/xnrm2.cc index 685eb29f..04e4137c 100644 --- a/src/routines/level1/xnrm2.cc +++ b/src/routines/level1/xnrm2.cc @@ -29,7 +29,7 @@ template <> const Precision Xnrm2<double2>::precision_ = Precision::kComplexDoub // Constructor: forwards to base class constructor template <typename T> -Xnrm2<T>::Xnrm2(Queue &queue, Event &event, const std::string &name): +Xnrm2<T>::Xnrm2(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Xdot"}, precision_) { source_string_ = #include "../../kernels/level1/xnrm2.opencl" @@ -69,12 +69,16 @@ StatusCode Xnrm2<T>::DoNrm2(const size_t n, kernel1.SetArgument(2, static_cast<int>(x_offset)); kernel1.SetArgument(3, static_cast<int>(x_inc)); kernel1.SetArgument(4, temp_buffer()); + // Event waiting list + auto eventWaitList = std::vector<Event>(); // Launches the main kernel auto global1 = std::vector<size_t>{db_["WGS1"]*temp_size}; auto local1 = std::vector<size_t>{db_["WGS1"]}; - status = RunKernel(kernel1, global1, local1); + auto kernelEvent = Event(); + status = RunKernel(kernel1, global1, local1, kernelEvent.pointer()); if (ErrorIn(status)) { return status; } + eventWaitList.push_back(kernelEvent); // Sets the arguments for the epilogue kernel kernel2.SetArgument(0, temp_buffer()); @@ -84,7 +88,7 @@ StatusCode Xnrm2<T>::DoNrm2(const size_t n, // Launches the epilogue kernel auto global2 = std::vector<size_t>{db_["WGS2"]}; auto local2 = std::vector<size_t>{db_["WGS2"]}; - status = RunKernel(kernel2, global2, local2); + status = RunKernel(kernel2, global2, local2, event_, eventWaitList); if (ErrorIn(status)) { return status; } // Succesfully finished the computation diff --git a/src/routines/level1/xscal.cc b/src/routines/level1/xscal.cc index 57bbe9e8..e83e73fd 100644 --- a/src/routines/level1/xscal.cc +++ b/src/routines/level1/xscal.cc @@ -29,7 +29,7 @@ template <> const Precision Xscal<double2>::precision_ = Precision::kComplexDoub // Constructor: forwards to base class constructor template <typename T> -Xscal<T>::Xscal(Queue &queue, Event &event, const std::string &name): +Xscal<T>::Xscal(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Xaxpy"}, precision_) { source_string_ = #include "../../kernels/level1/level1.opencl" @@ -81,13 +81,13 @@ StatusCode Xscal<T>::DoScal(const size_t n, const T alpha, if (use_fast_kernel) { auto global = std::vector<size_t>{CeilDiv(n, db_["WPT"]*db_["VW"])}; auto local = std::vector<size_t>{db_["WGS"]}; - status = RunKernel(kernel, global, local); + status = RunKernel(kernel, global, local, event_); } else { auto n_ceiled = Ceil(n, db_["WGS"]*db_["WPT"]); auto global = std::vector<size_t>{n_ceiled/db_["WPT"]}; auto local = std::vector<size_t>{db_["WGS"]}; - status = RunKernel(kernel, global, local); + status = RunKernel(kernel, global, local, event_); } if (ErrorIn(status)) { return status; } diff --git a/src/routines/level1/xswap.cc b/src/routines/level1/xswap.cc index c986b3fb..bc425f40 100644 --- a/src/routines/level1/xswap.cc +++ b/src/routines/level1/xswap.cc @@ -29,7 +29,7 @@ template <> const Precision Xswap<double2>::precision_ = Precision::kComplexDoub // Constructor: forwards to base class constructor template <typename T> -Xswap<T>::Xswap(Queue &queue, Event &event, const std::string &name): +Xswap<T>::Xswap(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Xaxpy"}, precision_) { source_string_ = #include "../../kernels/level1/level1.opencl" @@ -87,13 +87,13 @@ StatusCode Xswap<T>::DoSwap(const size_t n, if (use_fast_kernel) { auto global = std::vector<size_t>{CeilDiv(n, db_["WPT"]*db_["VW"])}; auto local = std::vector<size_t>{db_["WGS"]}; - status = RunKernel(kernel, global, local); + status = RunKernel(kernel, global, local, event_); } else { auto n_ceiled = Ceil(n, db_["WGS"]*db_["WPT"]); auto global = std::vector<size_t>{n_ceiled/db_["WPT"]}; auto local = std::vector<size_t>{db_["WGS"]}; - status = RunKernel(kernel, global, local); + status = RunKernel(kernel, global, local, event_); } if (ErrorIn(status)) { return status; } diff --git a/src/routines/level2/xgbmv.cc b/src/routines/level2/xgbmv.cc index 14d391ca..f90e26b2 100644 --- a/src/routines/level2/xgbmv.cc +++ b/src/routines/level2/xgbmv.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xgbmv<T>::Xgbmv(Queue &queue, Event &event, const std::string &name): +Xgbmv<T>::Xgbmv(Queue &queue, EventPointer event, const std::string &name): Xgemv<T>(queue, event, name) { } diff --git a/src/routines/level2/xgemv.cc b/src/routines/level2/xgemv.cc index bf7ae6fa..24e87db0 100644 --- a/src/routines/level2/xgemv.cc +++ b/src/routines/level2/xgemv.cc @@ -29,7 +29,7 @@ template <> const Precision Xgemv<double2>::precision_ = Precision::kComplexDoub // Constructor: forwards to base class constructor template <typename T> -Xgemv<T>::Xgemv(Queue &queue, Event &event, const std::string &name): +Xgemv<T>::Xgemv(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Pad", "Xgemv"}, precision_) { source_string_ = #include "../../kernels/level2/xgemv.opencl" @@ -162,7 +162,7 @@ StatusCode Xgemv<T>::MatVec(const Layout layout, const Transpose a_transpose, // Launches the kernel auto global = std::vector<size_t>{global_size}; auto local = std::vector<size_t>{local_size}; - status = RunKernel(kernel, global, local); + status = RunKernel(kernel, global, local, event_); if (ErrorIn(status)) { return status; } // Succesfully finished the computation diff --git a/src/routines/level2/xger.cc b/src/routines/level2/xger.cc index 9ab21bfb..dda78232 100644 --- a/src/routines/level2/xger.cc +++ b/src/routines/level2/xger.cc @@ -29,7 +29,7 @@ template <> const Precision Xger<double2>::precision_ = Precision::kComplexDoubl // Constructor: forwards to base class constructor template <typename T> -Xger<T>::Xger(Queue &queue, Event &event, const std::string &name): +Xger<T>::Xger(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Xger"}, precision_) { source_string_ = #include "../../kernels/level2/level2.opencl" @@ -89,7 +89,7 @@ StatusCode Xger<T>::DoGer(const Layout layout, auto a_two_ceiled = Ceil(CeilDiv(a_two, db_["WPT"]), db_["WGS2"]); auto global = std::vector<size_t>{a_one_ceiled, a_two_ceiled}; auto local = std::vector<size_t>{db_["WGS1"], db_["WGS2"]}; - status = RunKernel(kernel, global, local); + status = RunKernel(kernel, global, local, event_); if (ErrorIn(status)) { return status; } // Succesfully finished the computation diff --git a/src/routines/level2/xgerc.cc b/src/routines/level2/xgerc.cc index 09408898..73284b52 100644 --- a/src/routines/level2/xgerc.cc +++ b/src/routines/level2/xgerc.cc @@ -20,7 +20,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xgerc<T>::Xgerc(Queue &queue, Event &event, const std::string &name): +Xgerc<T>::Xgerc(Queue &queue, EventPointer event, const std::string &name): Xger<T>(queue, event, name) { } diff --git a/src/routines/level2/xgeru.cc b/src/routines/level2/xgeru.cc index 36fd9d0a..7730d6a5 100644 --- a/src/routines/level2/xgeru.cc +++ b/src/routines/level2/xgeru.cc @@ -20,7 +20,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xgeru<T>::Xgeru(Queue &queue, Event &event, const std::string &name): +Xgeru<T>::Xgeru(Queue &queue, EventPointer event, const std::string &name): Xger<T>(queue, event, name) { } diff --git a/src/routines/level2/xhbmv.cc b/src/routines/level2/xhbmv.cc index f59a7cb3..58591b50 100644 --- a/src/routines/level2/xhbmv.cc +++ b/src/routines/level2/xhbmv.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xhbmv<T>::Xhbmv(Queue &queue, Event &event, const std::string &name): +Xhbmv<T>::Xhbmv(Queue &queue, EventPointer event, const std::string &name): Xgemv<T>(queue, event, name) { } diff --git a/src/routines/level2/xhemv.cc b/src/routines/level2/xhemv.cc index 5a58b28b..b4ef0fa4 100644 --- a/src/routines/level2/xhemv.cc +++ b/src/routines/level2/xhemv.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xhemv<T>::Xhemv(Queue &queue, Event &event, const std::string &name): +Xhemv<T>::Xhemv(Queue &queue, EventPointer event, const std::string &name): Xgemv<T>(queue, event, name) { } diff --git a/src/routines/level2/xher.cc b/src/routines/level2/xher.cc index 1aefa240..aba665b0 100644 --- a/src/routines/level2/xher.cc +++ b/src/routines/level2/xher.cc @@ -28,7 +28,7 @@ template <> const Precision Xher<double2, double>::precision_ = Precision::kComp // Constructor: forwards to base class constructor template <typename T, typename U> -Xher<T,U>::Xher(Queue &queue, Event &event, const std::string &name): +Xher<T,U>::Xher(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Xger"}, precision_) { source_string_ = #include "../../kernels/level2/level2.opencl" @@ -99,7 +99,7 @@ StatusCode Xher<T,U>::DoHer(const Layout layout, const Triangle triangle, auto global_two = Ceil(CeilDiv(n, db_["WPT"]), db_["WGS2"]); auto global = std::vector<size_t>{global_one, global_two}; auto local = std::vector<size_t>{db_["WGS1"], db_["WGS2"]}; - status = RunKernel(kernel, global, local); + status = RunKernel(kernel, global, local, event_); if (ErrorIn(status)) { return status; } // Succesfully finished the computation diff --git a/src/routines/level2/xher2.cc b/src/routines/level2/xher2.cc index 364add12..bcd6488f 100644 --- a/src/routines/level2/xher2.cc +++ b/src/routines/level2/xher2.cc @@ -28,7 +28,7 @@ template <> const Precision Xher2<double2>::precision_ = Precision::kComplexDoub // Constructor: forwards to base class constructor template <typename T> -Xher2<T>::Xher2(Queue &queue, Event &event, const std::string &name): +Xher2<T>::Xher2(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Xger"}, precision_) { source_string_ = #include "../../kernels/level2/level2.opencl" @@ -91,7 +91,7 @@ StatusCode Xher2<T>::DoHer2(const Layout layout, const Triangle triangle, auto global_two = Ceil(CeilDiv(n, db_["WPT"]), db_["WGS2"]); auto global = std::vector<size_t>{global_one, global_two}; auto local = std::vector<size_t>{db_["WGS1"], db_["WGS2"]}; - status = RunKernel(kernel, global, local); + status = RunKernel(kernel, global, local, event_); if (ErrorIn(status)) { return status; } // Succesfully finished the computation diff --git a/src/routines/level2/xhpmv.cc b/src/routines/level2/xhpmv.cc index 2269255d..92686dbe 100644 --- a/src/routines/level2/xhpmv.cc +++ b/src/routines/level2/xhpmv.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xhpmv<T>::Xhpmv(Queue &queue, Event &event, const std::string &name): +Xhpmv<T>::Xhpmv(Queue &queue, EventPointer event, const std::string &name): Xgemv<T>(queue, event, name) { } diff --git a/src/routines/level2/xhpr.cc b/src/routines/level2/xhpr.cc index b0cea72f..4b31ad09 100644 --- a/src/routines/level2/xhpr.cc +++ b/src/routines/level2/xhpr.cc @@ -20,7 +20,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T, typename U> -Xhpr<T,U>::Xhpr(Queue &queue, Event &event, const std::string &name): +Xhpr<T,U>::Xhpr(Queue &queue, EventPointer event, const std::string &name): Xher<T,U>(queue, event, name) { } diff --git a/src/routines/level2/xhpr2.cc b/src/routines/level2/xhpr2.cc index ded35e53..9be24f43 100644 --- a/src/routines/level2/xhpr2.cc +++ b/src/routines/level2/xhpr2.cc @@ -20,7 +20,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xhpr2<T>::Xhpr2(Queue &queue, Event &event, const std::string &name): +Xhpr2<T>::Xhpr2(Queue &queue, EventPointer event, const std::string &name): Xher2<T>(queue, event, name) { } diff --git a/src/routines/level2/xsbmv.cc b/src/routines/level2/xsbmv.cc index 457bd762..bc82c88d 100644 --- a/src/routines/level2/xsbmv.cc +++ b/src/routines/level2/xsbmv.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xsbmv<T>::Xsbmv(Queue &queue, Event &event, const std::string &name): +Xsbmv<T>::Xsbmv(Queue &queue, EventPointer event, const std::string &name): Xgemv<T>(queue, event, name) { } diff --git a/src/routines/level2/xspmv.cc b/src/routines/level2/xspmv.cc index 4f1a9c61..6e00dcfa 100644 --- a/src/routines/level2/xspmv.cc +++ b/src/routines/level2/xspmv.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xspmv<T>::Xspmv(Queue &queue, Event &event, const std::string &name): +Xspmv<T>::Xspmv(Queue &queue, EventPointer event, const std::string &name): Xgemv<T>(queue, event, name) { } diff --git a/src/routines/level2/xspr.cc b/src/routines/level2/xspr.cc index 2d998e0b..55af2f29 100644 --- a/src/routines/level2/xspr.cc +++ b/src/routines/level2/xspr.cc @@ -20,7 +20,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xspr<T>::Xspr(Queue &queue, Event &event, const std::string &name): +Xspr<T>::Xspr(Queue &queue, EventPointer event, const std::string &name): Xher<T,T>(queue, event, name) { } diff --git a/src/routines/level2/xspr2.cc b/src/routines/level2/xspr2.cc index fd5232da..9a3f97ce 100644 --- a/src/routines/level2/xspr2.cc +++ b/src/routines/level2/xspr2.cc @@ -20,7 +20,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xspr2<T>::Xspr2(Queue &queue, Event &event, const std::string &name): +Xspr2<T>::Xspr2(Queue &queue, EventPointer event, const std::string &name): Xher2<T>(queue, event, name) { } diff --git a/src/routines/level2/xsymv.cc b/src/routines/level2/xsymv.cc index ec12324b..a9eb284f 100644 --- a/src/routines/level2/xsymv.cc +++ b/src/routines/level2/xsymv.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xsymv<T>::Xsymv(Queue &queue, Event &event, const std::string &name): +Xsymv<T>::Xsymv(Queue &queue, EventPointer event, const std::string &name): Xgemv<T>(queue, event, name) { } diff --git a/src/routines/level2/xsyr.cc b/src/routines/level2/xsyr.cc index c01fa2d3..4b3928e5 100644 --- a/src/routines/level2/xsyr.cc +++ b/src/routines/level2/xsyr.cc @@ -20,7 +20,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xsyr<T>::Xsyr(Queue &queue, Event &event, const std::string &name): +Xsyr<T>::Xsyr(Queue &queue, EventPointer event, const std::string &name): Xher<T,T>(queue, event, name) { } diff --git a/src/routines/level2/xsyr2.cc b/src/routines/level2/xsyr2.cc index 6db55085..3ae389e0 100644 --- a/src/routines/level2/xsyr2.cc +++ b/src/routines/level2/xsyr2.cc @@ -20,7 +20,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xsyr2<T>::Xsyr2(Queue &queue, Event &event, const std::string &name): +Xsyr2<T>::Xsyr2(Queue &queue, EventPointer event, const std::string &name): Xher2<T>(queue, event, name) { } diff --git a/src/routines/level2/xtbmv.cc b/src/routines/level2/xtbmv.cc index 2e1aebff..47371c87 100644 --- a/src/routines/level2/xtbmv.cc +++ b/src/routines/level2/xtbmv.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xtbmv<T>::Xtbmv(Queue &queue, Event &event, const std::string &name): +Xtbmv<T>::Xtbmv(Queue &queue, EventPointer event, const std::string &name): Xgemv<T>(queue, event, name) { } diff --git a/src/routines/level2/xtpmv.cc b/src/routines/level2/xtpmv.cc index aa0e099b..c63cb9b2 100644 --- a/src/routines/level2/xtpmv.cc +++ b/src/routines/level2/xtpmv.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xtpmv<T>::Xtpmv(Queue &queue, Event &event, const std::string &name): +Xtpmv<T>::Xtpmv(Queue &queue, EventPointer event, const std::string &name): Xgemv<T>(queue, event, name) { } diff --git a/src/routines/level2/xtrmv.cc b/src/routines/level2/xtrmv.cc index 94424743..9111d41d 100644 --- a/src/routines/level2/xtrmv.cc +++ b/src/routines/level2/xtrmv.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xtrmv<T>::Xtrmv(Queue &queue, Event &event, const std::string &name): +Xtrmv<T>::Xtrmv(Queue &queue, EventPointer event, const std::string &name): Xgemv<T>(queue, event, name) { } diff --git a/src/routines/level3/xgemm.cc b/src/routines/level3/xgemm.cc index 5dc2ad7f..7557dcc3 100644 --- a/src/routines/level3/xgemm.cc +++ b/src/routines/level3/xgemm.cc @@ -29,7 +29,7 @@ template <> const Precision Xgemm<double2>::precision_ = Precision::kComplexDoub // Constructor: forwards to base class constructor template <typename T> -Xgemm<T>::Xgemm(Queue &queue, Event &event, const std::string &name): +Xgemm<T>::Xgemm(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, precision_) { source_string_ = #include "../../kernels/level3/copy.opencl" @@ -122,30 +122,43 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout, auto b_temp = (b_no_temp) ? b_buffer : Buffer<T>(context_, k_ceiled*n_ceiled); auto c_temp = (c_no_temp) ? c_buffer : Buffer<T>(context_, m_ceiled*n_ceiled); + // 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. if (!a_no_temp) { - status = PadCopyTransposeMatrix(a_one, a_two, a_ld, a_offset, a_buffer, + auto eventProcessA = Event(); + status = PadCopyTransposeMatrix(eventProcessA.pointer(), emptyEventList, + a_one, a_two, a_ld, a_offset, a_buffer, m_ceiled, k_ceiled, m_ceiled, 0, a_temp, program, true, a_do_transpose, a_conjugate); if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessA); } // As above, but now for matrix B if (!b_no_temp) { - status = PadCopyTransposeMatrix(b_one, b_two, b_ld, b_offset, b_buffer, + auto eventProcessB = Event(); + status = PadCopyTransposeMatrix(eventProcessB.pointer(), emptyEventList, + b_one, b_two, b_ld, b_offset, b_buffer, n_ceiled, k_ceiled, n_ceiled, 0, b_temp, program, true, b_do_transpose, b_conjugate); if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessB); } // 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)) { - status = PadCopyTransposeMatrix(c_one, c_two, c_ld, c_offset, c_buffer, + auto eventProcessC = Event(); + status = PadCopyTransposeMatrix(eventProcessC.pointer(), emptyEventList, + c_one, c_two, c_ld, c_offset, c_buffer, m_ceiled, n_ceiled, m_ceiled, 0, c_temp, program, true, c_do_transpose, false); if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessC); } // Retrieves the Xgemm kernel from the compiled binary @@ -170,12 +183,15 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout, auto local = std::vector<size_t>{db_["MDIMC"], db_["NDIMC"]}; // Launches the kernel - status = RunKernel(kernel, global, local); + auto eventKernel = Event(); + status = RunKernel(kernel, global, local, eventKernel.pointer(), eventWaitList); if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventKernel); // Runs the post-processing kernel if needed if (!c_no_temp) { - status = PadCopyTransposeMatrix(m_ceiled, n_ceiled, m_ceiled, 0, c_temp, + status = PadCopyTransposeMatrix(event_, eventWaitList, + m_ceiled, n_ceiled, m_ceiled, 0, c_temp, c_one, c_two, c_ld, c_offset, c_buffer, program, false, c_do_transpose, false); if (ErrorIn(status)) { return status; } diff --git a/src/routines/level3/xhemm.cc b/src/routines/level3/xhemm.cc index bcc60dee..c0a4306a 100644 --- a/src/routines/level3/xhemm.cc +++ b/src/routines/level3/xhemm.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xhemm<T>::Xhemm(Queue &queue, Event &event, const std::string &name): +Xhemm<T>::Xhemm(Queue &queue, EventPointer event, const std::string &name): Xgemm<T>(queue, event, name) { } @@ -79,9 +79,13 @@ StatusCode Xhemm<T>::DoHemm(const Layout layout, const Side side, const Triangle auto global = std::vector<size_t>{Ceil(CeilDiv(k, db_["PAD_WPTX"]), db_["PAD_DIMX"]), Ceil(CeilDiv(k, db_["PAD_WPTY"]), db_["PAD_DIMY"])}; auto local = std::vector<size_t>{db_["PAD_DIMX"], db_["PAD_DIMY"]}; - status = RunKernel(kernel, global, local); + auto kernelEvent = Event(); + status = RunKernel(kernel, global, local, kernelEvent.pointer()); if (ErrorIn(status)) { return status; } + // Synchronize now: 'DoGemm' does not accept a list of events to wait for + kernelEvent.WaitForCompletion(); + // Runs the regular Xgemm code with either "C := AB+C" or ... if (side == Side::kLeft) { status = DoGemm(layout, Transpose::kNo, Transpose::kNo, diff --git a/src/routines/level3/xher2k.cc b/src/routines/level3/xher2k.cc index 1711905d..4d5a4d35 100644 --- a/src/routines/level3/xher2k.cc +++ b/src/routines/level3/xher2k.cc @@ -27,7 +27,7 @@ template <> const Precision Xher2k<double2,double>::precision_ = Precision::kCom // Constructor: forwards to base class constructor template <typename T, typename U> -Xher2k<T,U>::Xher2k(Queue &queue, Event &event, const std::string &name): +Xher2k<T,U>::Xher2k(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, precision_) { source_string_ = #include "../../kernels/level3/copy.opencl" @@ -112,39 +112,58 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co auto b2_temp = (b2_no_temp) ? b_buffer : Buffer<T>(context_, k_ceiled*n_ceiled); auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled); + // Events of all kernels (including pre/post processing kernels) + auto eventWaitList = std::vector<Event>(); + auto emptyEventList = std::vector<Event>(); + // 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 (!a1_no_temp) { - status = PadCopyTransposeMatrix(ab_one, ab_two, a_ld, a_offset, a_buffer, + auto eventProcessA1 = Event(); + status = PadCopyTransposeMatrix(eventProcessA1.pointer(), emptyEventList, + ab_one, ab_two, a_ld, a_offset, a_buffer, n_ceiled, k_ceiled, n_ceiled, 0, a1_temp, program, true, ab_rotated, ab_conjugate); + eventWaitList.push_back(eventProcessA1); if (ErrorIn(status)) { return status; } } if (!a2_no_temp) { - status = PadCopyTransposeMatrix(ab_one, ab_two, a_ld, a_offset, a_buffer, + auto eventProcessA2 = Event(); + status = PadCopyTransposeMatrix(eventProcessA2.pointer(), emptyEventList, + ab_one, ab_two, a_ld, a_offset, a_buffer, n_ceiled, k_ceiled, n_ceiled, 0, a2_temp, program, true, ab_rotated, !ab_conjugate); + eventWaitList.push_back(eventProcessA2); if (ErrorIn(status)) { return status; } } if (!b1_no_temp) { - status = PadCopyTransposeMatrix(ab_one, ab_two, b_ld, b_offset, b_buffer, + auto eventProcessB1 = Event(); + status = PadCopyTransposeMatrix(eventProcessB1.pointer(), emptyEventList, + ab_one, ab_two, b_ld, b_offset, b_buffer, n_ceiled, k_ceiled, n_ceiled, 0, b1_temp, program, true, ab_rotated, ab_conjugate); + eventWaitList.push_back(eventProcessB1); if (ErrorIn(status)) { return status; } } if (!b2_no_temp) { - status = PadCopyTransposeMatrix(ab_one, ab_two, b_ld, b_offset, b_buffer, + auto eventProcessB2 = Event(); + status = PadCopyTransposeMatrix(eventProcessB2.pointer(), emptyEventList, + ab_one, ab_two, b_ld, b_offset, b_buffer, n_ceiled, k_ceiled, n_ceiled, 0, b2_temp, program, true, ab_rotated, !ab_conjugate); + eventWaitList.push_back(eventProcessB2); 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. - status = PadCopyTransposeMatrix(n, n, c_ld, c_offset, c_buffer, + auto eventProcessC = Event(); + status = PadCopyTransposeMatrix(eventProcessC.pointer(), emptyEventList, + n, n, c_ld, c_offset, c_buffer, n_ceiled, n_ceiled, n_ceiled, 0, c_temp, program, true, c_rotated, false); + eventWaitList.push_back(eventProcessC); if (ErrorIn(status)) { return status; } // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary @@ -169,8 +188,10 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co auto local = std::vector<size_t>{db_["MDIMC"], db_["NDIMC"]}; // Launches the kernel - status = RunKernel(kernel, global, local); + auto eventKernel1 = Event(); + status = RunKernel(kernel, global, local, eventKernel1.pointer(), eventWaitList); if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventKernel1); // Swaps the arguments for matrices A and B, sets 'beta' to 1, and conjugate alpha auto conjugate_alpha = T{alpha.real(), -alpha.imag()}; @@ -181,13 +202,16 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co kernel.SetArgument(5, a2_temp()); // Runs the kernel again - status = RunKernel(kernel, global, local); + auto eventKernel2 = Event(); + status = RunKernel(kernel, 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(n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + status = PadCopyTransposeMatrix(event_, eventWaitList, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, n, n, c_ld, c_offset, c_buffer, program, false, c_rotated, false, upper, lower, true); if (ErrorIn(status)) { return status; } diff --git a/src/routines/level3/xherk.cc b/src/routines/level3/xherk.cc index cbd0a188..574debe4 100644 --- a/src/routines/level3/xherk.cc +++ b/src/routines/level3/xherk.cc @@ -27,7 +27,7 @@ template <> const Precision Xherk<double2,double>::precision_ = Precision::kComp // Constructor: forwards to base class constructor template <typename T, typename U> -Xherk<T,U>::Xherk(Queue &queue, Event &event, const std::string &name): +Xherk<T,U>::Xherk(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, precision_) { source_string_ = #include "../../kernels/level3/copy.opencl" @@ -103,27 +103,40 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons 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); + // 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) { - status = PadCopyTransposeMatrix(a_one, a_two, a_ld, a_offset, a_buffer, + auto eventProcessA = Event(); + status = PadCopyTransposeMatrix(eventProcessA.pointer(), emptyEventList, + a_one, a_two, a_ld, a_offset, a_buffer, n_ceiled, k_ceiled, n_ceiled, 0, a_temp, program, true, a_rotated, a_conjugate); + eventWaitList.push_back(eventProcessA); if (ErrorIn(status)) { return status; } } if (!b_no_temp) { - status = PadCopyTransposeMatrix(a_one, a_two, a_ld, a_offset, a_buffer, + auto eventProcessB = Event(); + status = PadCopyTransposeMatrix(eventProcessB.pointer(), emptyEventList, + a_one, a_two, a_ld, a_offset, a_buffer, n_ceiled, k_ceiled, n_ceiled, 0, b_temp, 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. - status = PadCopyTransposeMatrix(n, n, c_ld, c_offset, c_buffer, + auto eventProcessC = Event(); + status = PadCopyTransposeMatrix(eventProcessC.pointer(), emptyEventList, + n, n, c_ld, c_offset, c_buffer, n_ceiled, n_ceiled, n_ceiled, 0, c_temp, program, true, c_rotated, false); + eventWaitList.push_back(eventProcessC); if (ErrorIn(status)) { return status; } // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary @@ -149,13 +162,16 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons auto local = std::vector<size_t>{db_["MDIMC"], db_["NDIMC"]}; // Launches the kernel - status = RunKernel(kernel, global, local); + auto eventKernel = Event(); + status = RunKernel(kernel, 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(n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + status = PadCopyTransposeMatrix(event_, eventWaitList, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, n, n, c_ld, c_offset, c_buffer, program, false, c_rotated, false, upper, lower, true); if (ErrorIn(status)) { return status; } diff --git a/src/routines/level3/xsymm.cc b/src/routines/level3/xsymm.cc index 583d5c7d..914a326a 100644 --- a/src/routines/level3/xsymm.cc +++ b/src/routines/level3/xsymm.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xsymm<T>::Xsymm(Queue &queue, Event &event, const std::string &name): +Xsymm<T>::Xsymm(Queue &queue, EventPointer event, const std::string &name): Xgemm<T>(queue, event, name) { } @@ -79,9 +79,13 @@ StatusCode Xsymm<T>::DoSymm(const Layout layout, const Side side, const Triangle auto global = std::vector<size_t>{Ceil(CeilDiv(k, db_["PAD_WPTX"]), db_["PAD_DIMX"]), Ceil(CeilDiv(k, db_["PAD_WPTY"]), db_["PAD_DIMY"])}; auto local = std::vector<size_t>{db_["PAD_DIMX"], db_["PAD_DIMY"]}; - status = RunKernel(kernel, global, local); + auto kernelEvent = Event(); + status = RunKernel(kernel, global, local, kernelEvent.pointer()); if (ErrorIn(status)) { return status; } + // Synchronize now: 'DoGemm' does not accept a list of events to wait for + kernelEvent.WaitForCompletion(); + // Runs the regular Xgemm code with either "C := AB+C" or ... if (side == Side::kLeft) { status = DoGemm(layout, Transpose::kNo, Transpose::kNo, diff --git a/src/routines/level3/xsyr2k.cc b/src/routines/level3/xsyr2k.cc index 79090871..44d0024e 100644 --- a/src/routines/level3/xsyr2k.cc +++ b/src/routines/level3/xsyr2k.cc @@ -29,7 +29,7 @@ template <> const Precision Xsyr2k<double2>::precision_ = Precision::kComplexDou // Constructor: forwards to base class constructor template <typename T> -Xsyr2k<T>::Xsyr2k(Queue &queue, Event &event, const std::string &name): +Xsyr2k<T>::Xsyr2k(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, precision_) { source_string_ = #include "../../kernels/level3/copy.opencl" @@ -104,28 +104,41 @@ StatusCode Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, cons auto b_temp = (b_no_temp) ? b_buffer : Buffer<T>(context_, k_ceiled*n_ceiled); auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled); + // Events of all kernels (including pre/post processing kernels) + auto eventWaitList = std::vector<Event>(); + auto emptyEventList = std::vector<Event>(); + // 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) { - status = PadCopyTransposeMatrix(ab_one, ab_two, a_ld, a_offset, a_buffer, + auto eventProcessA = Event(); + status = PadCopyTransposeMatrix(eventProcessA.pointer(), emptyEventList, + ab_one, ab_two, a_ld, a_offset, a_buffer, n_ceiled, k_ceiled, n_ceiled, 0, a_temp, program, true, ab_rotated, false); if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessA); } if (!b_no_temp) { - status = PadCopyTransposeMatrix(ab_one, ab_two, b_ld, b_offset, b_buffer, + auto eventProcessB = Event(); + status = PadCopyTransposeMatrix(eventProcessB.pointer(), emptyEventList, + ab_one, ab_two, b_ld, b_offset, b_buffer, n_ceiled, k_ceiled, n_ceiled, 0, b_temp, 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. - status = PadCopyTransposeMatrix(n, n, c_ld, c_offset, c_buffer, + auto eventProcessC = Event(); + status = PadCopyTransposeMatrix(eventProcessC.pointer(), emptyEventList, + n, n, c_ld, c_offset, c_buffer, n_ceiled, n_ceiled, n_ceiled, 0, c_temp, 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 { @@ -148,8 +161,10 @@ StatusCode Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, cons auto local = std::vector<size_t>{db_["MDIMC"], db_["NDIMC"]}; // Launches the kernel - status = RunKernel(kernel, global, local); + auto eventKernel1 = Event(); + status = RunKernel(kernel, 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<T>(1); @@ -158,13 +173,16 @@ StatusCode Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, cons kernel.SetArgument(5, a_temp()); // Runs the kernel again - status = RunKernel(kernel, global, local); + auto eventKernel2 = Event(); + status = RunKernel(kernel, 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(n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + status = PadCopyTransposeMatrix(event_, eventWaitList, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, n, n, c_ld, c_offset, c_buffer, program, false, c_rotated, false, upper, lower, false); if (ErrorIn(status)) { return status; } diff --git a/src/routines/level3/xsyrk.cc b/src/routines/level3/xsyrk.cc index ca429bd7..44ed8d35 100644 --- a/src/routines/level3/xsyrk.cc +++ b/src/routines/level3/xsyrk.cc @@ -29,7 +29,7 @@ template <> const Precision Xsyrk<double2>::precision_ = Precision::kComplexDoub // Constructor: forwards to base class constructor template <typename T> -Xsyrk<T>::Xsyrk(Queue &queue, Event &event, const std::string &name): +Xsyrk<T>::Xsyrk(Queue &queue, EventPointer event, const std::string &name): Routine<T>(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, precision_) { source_string_ = #include "../../kernels/level3/copy.opencl" @@ -97,22 +97,32 @@ StatusCode Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const auto a_temp = (a_no_temp) ? a_buffer : Buffer<T>(context_, k_ceiled*n_ceiled); auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled); + // 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. if (!a_no_temp) { - status = PadCopyTransposeMatrix(a_one, a_two, a_ld, a_offset, a_buffer, + auto eventProcessA = Event(); + status = PadCopyTransposeMatrix(eventProcessA.pointer(), emptyEventList, + a_one, a_two, a_ld, a_offset, a_buffer, n_ceiled, k_ceiled, n_ceiled, 0, a_temp, program, true, a_rotated, false); if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessA); } // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to // modify the other triangle. - status = PadCopyTransposeMatrix(n, n, c_ld, c_offset, c_buffer, + auto eventProcessC = Event(); + status = PadCopyTransposeMatrix(eventProcessC.pointer(), emptyEventList, + n, n, c_ld, c_offset, c_buffer, n_ceiled, n_ceiled, n_ceiled, 0, c_temp, 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 { @@ -135,17 +145,21 @@ StatusCode Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const auto local = std::vector<size_t>{db_["MDIMC"], db_["NDIMC"]}; // Launches the kernel - status = RunKernel(kernel, global, local); + auto eventKernel = Event(); + status = RunKernel(kernel, 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(n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + status = PadCopyTransposeMatrix(event_, eventWaitList, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, n, n, c_ld, c_offset, c_buffer, program, false, c_rotated, false, upper, lower, false); if (ErrorIn(status)) { return status; } + // Successfully finished the computation return StatusCode::kSuccess; } catch (...) { return StatusCode::kInvalidKernel; } diff --git a/src/routines/level3/xtrmm.cc b/src/routines/level3/xtrmm.cc index 1180c026..484cf040 100644 --- a/src/routines/level3/xtrmm.cc +++ b/src/routines/level3/xtrmm.cc @@ -21,7 +21,7 @@ namespace clblast { // Constructor: forwards to base class constructor template <typename T> -Xtrmm<T>::Xtrmm(Queue &queue, Event &event, const std::string &name): +Xtrmm<T>::Xtrmm(Queue &queue, EventPointer event, const std::string &name): Xgemm<T>(queue, event, name) { } @@ -82,9 +82,13 @@ StatusCode Xtrmm<T>::DoTrmm(const Layout layout, const Side side, const Triangle auto global = std::vector<size_t>{Ceil(CeilDiv(k, db_["PAD_WPTX"]), db_["PAD_DIMX"]), Ceil(CeilDiv(k, db_["PAD_WPTY"]), db_["PAD_DIMY"])}; auto local = std::vector<size_t>{db_["PAD_DIMX"], db_["PAD_DIMY"]}; - status = RunKernel(kernel, global, local); + auto kernelEvent = Event(); + status = RunKernel(kernel, global, local, kernelEvent.pointer()); if (ErrorIn(status)) { return status; } + // Synchronize now: 'DoGemm' does not accept a list of events to wait for + kernelEvent.WaitForCompletion(); + // Runs the regular Xgemm code with either "B := alpha*A*B" or ... if (side == Side::kLeft) { status = DoGemm(layout, a_transpose, Transpose::kNo, |