Support cublasLt

Makefile.am

@@ -128,6 +128,7 @@ versinclude_HEADERS = 				\
 	include/starpu_disk.h			\
 	include/starpu_cublas.h			\
 	include/starpu_cublas_v2.h		\
+	include/starpu_cublasLt.h		\
 	include/starpu_cusolver.h		\
 	include/starpu_cusparse.h		\
 	include/starpu_hipblas.h		\

configure.ac

@@ -1844,6 +1844,10 @@ if test x$enable_cuda = xyes; then
 	SAVED_LIBS="${LIBS}"
 	AC_CHECK_HEADERS([cuda_gl_interop.h])
 
+    AC_CHECK_LIB([cublasLt], [cublasLtCreate],
+             [AC_DEFINE([STARPU_HAVE_LIBCUBLASLT], [1], [Define to 1 if you have the cublasLt library])
+              STARPU_CUDA_LDFLAGS="$STARPU_CUDA_LDFLAGS -lcublasLt"])
+
 	AC_CHECK_LIB([cusparse], [cusparseCreate],
 		     [AC_DEFINE([STARPU_HAVE_LIBCUSPARSE], [1], [Define to 1 if you have the cusparse library])
 		      STARPU_CUDA_LDFLAGS="$STARPU_CUDA_LDFLAGS -lcusparse"])

doc/doxygen/Makefile.am

@@ -173,6 +173,7 @@ dox_inputs = $(DOX_CONFIG) 				\
 	$(top_srcdir)/include/starpu_bound.h		\
 	$(top_srcdir)/include/starpu_cublas.h		\
 	$(top_srcdir)/include/starpu_cublas_v2.h	\
+	$(top_srcdir)/include/starpu_cublasLt.h		\
 	$(top_srcdir)/include/starpu_cusparse.h		\
 	$(top_srcdir)/include/starpu_cuda.h		\
 	$(top_srcdir)/include/starpu_cusolver.h		\

doc/doxygen/chapters/files.doxy

@@ -24,6 +24,7 @@
 \file starpu_cublas.h
 \file starpu_cublas_v2.h
 \file starpu_cuda.h
+\file starpu_cublasLt.h
 \file starpu_cusolver.h
 \file starpu_cusparse.h
 \file starpu_data_filters.h

doc/doxygen/chapters/starpu_faq/check_list_performance.doxy

@@ -151,6 +151,13 @@ the beginning of the codelet to make sure that CUBLAS is really using the proper
 stream. When using CUBLAS v2, starpu_cublas_get_local_handle() can be called to queue CUBLAS
 kernels with the proper configuration.
 
+Similary, calling starpu_cublasLt_init() makes StarPU create CUBLASLT handles
+on each CUDA device, starpu_cublasLt_get_local_handle() can then be used to
+queue CUSPARSE kernels with the proper configuration. starpu_cublasLt_shutdown() will synchronously deinitialize the CUBLASLT library on every CUDA device.
+Since CUBLASLT handles are not along with CUDA streams, users may need to call
+starpu_cuda_get_local_stream() to get a CUDA stream before calling a
+CUBLASLT API.
+
 Similarly, calling starpu_cusparse_init() makes StarPU create CUSPARSE handles
 on each CUDA device, starpu_cusparse_get_local_handle() can then be used to
 queue CUSPARSE kernels with the proper configuration. starpu_cusparse_shutdown() will synchronously deinitialize the CUSPARSE library on every CUDA device.

doc/doxygen/doxygen-config-include.cfg.in

@@ -25,6 +25,7 @@ INPUT                  +=  @top_builddir@/doc/doxygen/starpu_config.h \
 			 @top_srcdir@/include/starpu_task_list.h \
 			 @top_srcdir@/include/starpu_task_util.h \
 			 @top_srcdir@/include/starpu_cuda.h \
+			 @top_srcdir@/include/starpu_cublasLt.h \
 			 @top_srcdir@/include/starpu_cusparse.h \
 			 @top_srcdir@/include/starpu_cublas.h \
 			 @top_srcdir@/include/starpu_cublas_v2.h \

include/starpu.h

@@ -73,6 +73,7 @@ typedef INT_PTR intptr_t;
 #include <starpu_hip.h>
 #include <starpu_hipblas.h>
 #include <starpu_cublas.h>
+#include <starpu_cublasLt.h>
 #include <starpu_cusparse.h>
 #include <starpu_bound.h>
 #include <starpu_hash.h>

include/starpu_cublasLt.h

@@ -0,0 +1,61 @@
+/* StarPU --- Runtime system for heterogeneous multicore architectures.
+*
+ * Copyright (C) 2010-2024  University of Bordeaux, CNRS (LaBRI UMR 5800), Inria
+ *
+ * StarPU is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU Lesser General Public License as published by
+ * the Free Software Foundation; either version 2.1 of the License, or (at
+ * your option) any later version.
+ *
+ * StarPU is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * See the GNU Lesser General Public License in COPYING.LGPL for more details.
+ */
+
+#ifndef __STARPU_CUBLASLT_H__
+#define __STARPU_CUBLASLT_H__
+
+#ifdef STARPU_USE_CUDA
+#include <cublasLt.h>
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+    /**
+       @ingroup API_CUDA_Extensions
+       @{
+    */
+
+    /**
+       Initialize CUBLASLT on every CUDA device
+       controlled by StarPU. This call blocks until CUBLASLT has been properly
+       initialized on every device. See \ref CUDA-specificOptimizations for more details.
+    */
+    void starpu_cublasLt_init(void);
+
+    /**
+       Synchronously deinitialize the CUBLASLT library on
+       every CUDA device. See \ref CUDA-specificOptimizations for more details.
+    */
+    void starpu_cublasLt_shutdown(void);
+
+#ifdef STARPU_USE_CUDA
+    /**
+       Return the CUBLASLT handle to be used to queue CUBLASLT
+       kernels. It is properly initialized and configured for multistream by
+       starpu_cublasLt_init(). See \ref CUDA-specificOptimizations for more details.
+    */
+    cublasLtHandle_t starpu_cublasLt_get_local_handle(void);
+#endif
+
+    /** @} */
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* __STARPU_CUBLASLT_H__ */

src/Makefile.am

@@ -375,6 +375,7 @@ endif
 
 libstarpu_@STARPU_EFFECTIVE_VERSION@_la_SOURCES += drivers/cuda/starpu_cublas.c
 libstarpu_@STARPU_EFFECTIVE_VERSION@_la_SOURCES += drivers/cuda/starpu_cublas_v2.c
+libstarpu_@STARPU_EFFECTIVE_VERSION@_la_SOURCES += drivers/cuda/starpu_cublasLt.c
 libstarpu_@STARPU_EFFECTIVE_VERSION@_la_SOURCES += drivers/cuda/starpu_cusparse.c
 libstarpu_@STARPU_EFFECTIVE_VERSION@_la_SOURCES += drivers/cuda/starpu_cusolver.c
 

src/drivers/cuda/starpu_cublasLt.c

@@ -0,0 +1,76 @@
+/* StarPU --- Runtime system for heterogeneous multicore architectures.
+ *
+ * Copyright (C) 2009-2024  University of Bordeaux, CNRS (LaBRI UMR 5800), Inria
+ *
+ * StarPU is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU Lesser General Public License as published by
+ * the Free Software Foundation; either version 2.1 of the License, or (at
+ * your option) any later version.
+ *
+ * StarPU is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * See the GNU Lesser General Public License in COPYING.LGPL for more details.
+ */
+
+#include <common/config.h>
+
+#include <starpu.h>
+#include <starpu_cuda.h>
+#include <core/workers.h>
+
+#ifdef STARPU_HAVE_LIBCUBLASLT
+#include <cublasLt.h>
+
+static cublasLtHandle_t cublasLt_handles[STARPU_NMAXWORKERS];
+static cublasLtHandle_t main_handle;
+
+static void init_cublasLt_func(void *args STARPU_ATTRIBUTE_UNUSED)
+{
+	cublasLtCreate(&cublasLt_handles[starpu_worker_get_id_check()]);
+	// No need for setting streams, because the cublasLt handles are not along with streams
+}
+
+static void shutdown_cublasLt_func(void *args STARPU_ATTRIBUTE_UNUSED)
+{
+	cublasLtDestroy(cublasLt_handles[starpu_worker_get_id_check()]);
+}
+#endif
+
+void starpu_cublasLt_init(void)
+{
+#ifdef STARPU_HAVE_LIBCUBLASLT
+	if (!starpu_cuda_worker_get_count())
+		return;
+	starpu_execute_on_each_worker_ex(init_cublasLt_func, NULL, STARPU_CUDA, "init_cublasLt");
+
+	if (cublasLtCreate(&main_handle) != CUBLAS_STATUS_SUCCESS)
+		main_handle = NULL;
+#endif
+}
+
+void starpu_cublasLt_shutdown(void)
+{
+#ifdef STARPU_HAVE_LIBCUBLASLT
+	if (!starpu_cuda_worker_get_count())
+		return;
+	starpu_execute_on_each_worker_ex(shutdown_cublasLt_func, NULL, STARPU_CUDA, "shutdown_cublasLt");
+
+	if (main_handle)
+		cublasLtDestroy(main_handle);
+#endif
+}
+
+#ifdef STARPU_HAVE_LIBCUBLASLT
+cublasLtHandle_t starpu_cublasLt_get_local_handle(void)
+{
+	if (!starpu_cuda_worker_get_count())
+		return NULL;
+	int workerid = starpu_worker_get_id();
+	if (workerid >= 0)
+		return cublasLt_handles[workerid];
+	else
+		return main_handle;
+}
+#endif

tests/helper/cublasLt_init.c

@@ -0,0 +1,71 @@
+/* StarPU --- Runtime system for heterogeneous multicore architectures.
+*
+ * Copyright (C) 2009-2024  University of Bordeaux, CNRS (LaBRI UMR 5800), Inria
+ *
+ * StarPU is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU Lesser General Public License as published by
+ * the Free Software Foundation; either version 2.1 of the License, or (at
+ * your option) any later version.
+ *
+ * StarPU is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * See the GNU Lesser General Public License in COPYING.LGPL for more details.
+ */
+
+#include <stdio.h>
+#include <unistd.h>
+#include <errno.h>
+#include <starpu.h>
+#include <stdlib.h>
+#include "../helper.h"
+
+/*
+ * Test initializing cublasLt, and how much time that takes
+ */
+
+static double start;
+static double end;
+
+//static float *data = NULL;
+
+int main(int argc, char **argv)
+{
+    int ret;
+
+    ret = starpu_initialize(NULL, &argc, &argv);
+    if (ret == -ENODEV) return STARPU_TEST_SKIPPED;
+    STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
+
+    unsigned ngpus = starpu_cuda_worker_get_count();
+
+    double init_timing;
+    double shutdown_timing;
+
+    start = starpu_timing_now();
+    starpu_cublasLt_init();
+    end = starpu_timing_now();
+    init_timing = end - start;
+
+    start = starpu_timing_now();
+    starpu_cublasLt_shutdown();
+    end = starpu_timing_now();
+    shutdown_timing = end - start;
+
+    FPRINTF(stderr, "Total:\n");
+    FPRINTF(stderr, "\tinit: %2.2f ms\n", init_timing/(1000));
+    FPRINTF(stderr, "\tshutdown: %2.2f ms\n", shutdown_timing/(1000));
+
+    if (ngpus != 0)
+    {
+        FPRINTF(stderr, "per-GPU (#gpu = %u):\n", ngpus);
+
+        FPRINTF(stderr, "\tinit: %2.2f ms\n", init_timing/(1000*ngpus));
+        FPRINTF(stderr, "\tshutdown: %2.2f ms\n", shutdown_timing/(1000*ngpus));
+    }
+
+    starpu_shutdown();
+
+    return EXIT_SUCCESS;
+}