From 05bca95f0275578fd7d8eba2ae7f7f432bad3480 Mon Sep 17 00:00:00 2001
From: akolliasAMD <99202231+akolliasAMD@users.noreply.github.com>
Date: Wed, 19 Jun 2024 09:27:29 -0600
Subject: [PATCH] BUILD: add support for specific gpu arch with ROCM (#987)

---
 config/m4/rocm.m4                         | 35 ++++++++++++++++++++---
 configure.ac                              |  3 ++
 cuda_lt.sh                                |  4 +--
 src/components/ec/rocm/kernel/Makefile.am |  4 +--
 4 files changed, 38 insertions(+), 8 deletions(-)

diff --git a/config/m4/rocm.m4 b/config/m4/rocm.m4
index c7c4ca141d..8a2bdc2fb9 100644
--- a/config/m4/rocm.m4
+++ b/config/m4/rocm.m4
@@ -4,6 +4,17 @@
 # See file LICENSE for terms.
 #
 
+ROCM_ARCH_NATIVE="--offload-arch=native"
+ROCM_ARCH908="--offload-arch=gfx908"
+ROCM_ARCH90A="--offload-arch=gfx90a"
+ROCM_ARCH94="--offload-arch=gfx940 \
+--offload-arch=gfx941 \
+--offload-arch=gfx942"
+ROCM_ARCH10="--offload-arch=gfx1030"
+ROCM_ARCH11="--offload-arch=gfx1100 \
+--offload-arch=gfx1101 \
+--offload-arch=gfx1102"
+
 # ROCM_PARSE_FLAGS(ARG, VAR_LIBS, VAR_LDFLAGS, VAR_CPPFLAGS)
 # ----------------------------------------------------------
 # Parse whitespace-separated ARG into appropriate LIBS, LDFLAGS, and
@@ -73,7 +84,12 @@ AC_ARG_WITH([rocm],
         [Enable the use of ROCm (default is autodetect).])],
     [],
     [with_rocm=guess])
-
+AC_ARG_WITH([rocm-arch],
+    [AS_HELP_STRING([--with-rocm-arch=arch-code],
+        [Defines target GPU architecture,
+            see rocm documentation for valid --offload-arch options for details
+            'all-arch-no-native' for all default architectures but not native])],
+        [], [with_rocm_arch=all])
 rocm_happy=no
 hip_happy=no
 AS_IF([test "x$with_rocm" != "xno"],
@@ -115,15 +131,26 @@ AS_IF([test "x$with_rocm" != "xno"],
            AC_SUBST([ROCM_ROOT])],
           [AC_MSG_WARN([ROCm not found])])
 
+
+    # Check whether we run on ROCm 6.0 or higher
+    CHECK_ROCM_VERSION(6, ROCM_VERSION_60_OR_GREATER)
+    AC_MSG_CHECKING([if ROCm version is 6.0 or above])
+
+    AS_IF([test "x$rocm_happy" = "xyes"],
+        [AS_IF([test "x$with_rocm_arch" = "xall"],
+          [ROCM_ARCH="${ROCM_ARCH908} ${ROCM_ARCH90A} ${ROCM_ARCH94} ${ROCM_ARCH10} ${ROCM_ARCH11} ${ROCM_ARCH_NATIVE}"],
+        [AS_IF([test "x$with_rocm_arch" = "xall-arch-no-native"],
+          [ROCM_ARCH="${ROCM_ARCH908} ${ROCM_ARCH90A} ${ROCM_ARCH94} ${ROCM_ARCH10} ${ROCM_ARCH11}"],
+        [ROCM_ARCH="$with_rocm_arch"])])
+        AS_IF([test "$ROCM_VERSION_60_OR_GREATER" = "1"],
+          AC_SUBST([ROCM_ARCH], ["$ROCM_ARCH"]),
+          AC_SUBST([ROCM_ARCH], [""]))])
     CPPFLAGS="$SAVE_CPPFLAGS"
     LDFLAGS="$SAVE_LDFLAGS"
     LIBS="$SAVE_LIBS"
 
     HIP_BUILD_FLAGS([$with_rocm], [HIP_LIBS], [HIP_LDFLAGS], [HIP_CPPFLAGS])
 
-    # Check whether we run on ROCm 6.0 or higher
-    CHECK_ROCM_VERSION(6, ROCM_VERSION_60_OR_GREATER)
-    AC_MSG_CHECKING([if ROCm version is 6.0 or above])
     if test "$ROCM_VERSION_60_OR_GREATER" = "1" ; then
         AC_MSG_RESULT([yes])
     else
diff --git a/configure.ac b/configure.ac
index 246e297c90..0022b147fb 100644
--- a/configure.ac
+++ b/configure.ac
@@ -265,6 +265,9 @@ AC_MSG_NOTICE([      C++ compiler:   ${CXX} ${CXXFLAGS} ${BASE_CXXFLAGS}])
 AS_IF([test "x$cuda_happy" = "xyes"],[
 AC_MSG_NOTICE([     NVCC gencodes:   ${NVCC_ARCH}])
 ])
+AS_IF([test "x$rocm_happy" = xyes],[
+AC_MSG_NOTICE([ROCM architectures:   ${ROCM_ARCH}])
+])
 AC_MSG_NOTICE([          Perftest:   ${mpi_enable}])
 AC_MSG_NOTICE([             Gtest:   ${gtest_enable}])
 AC_MSG_NOTICE([        MC modules:   <$(echo ${mc_modules}|tr ':' ' ') >])
diff --git a/cuda_lt.sh b/cuda_lt.sh
index 6601e3edcb..9b7b505d6f 100755
--- a/cuda_lt.sh
+++ b/cuda_lt.sh
@@ -28,7 +28,7 @@ mkdir -p $pic_dir
 
 tmpcmd="${@:3}"
 if [[ "$tmpcmd" == *"amdclang"* ]]; then
-  cmd="${@:3:2} -x hip -target x86_64-unknown-linux-gnu --offload-arch=gfx908 --offload-arch=gfx90a --offload-arch=gfx940 --offload-arch=gfx941 --offload-arch=gfx942 --offload-arch=gfx1030 --offload-arch=gfx1100 --offload-arch=gfx1101 --offload-arch=gfx1102 --offload-arch=native ${@:5} -fPIC -O3 -o ${pic_filepath}"
+  cmd="${@:3:2} -x hip -target x86_64-unknown-linux-gnu ${@:5} -fPIC -O3 -o ${pic_filepath}"
 elif [[ "$tmpcmd" == *"hipcc"* ]]; then
   cmd="${@:3} -fPIC -o ${pic_filepath}"
 else
@@ -38,7 +38,7 @@ echo $cmd
 $cmd
 
 if [[ "$tmpcmd" == *"amdclang"* ]]; then
-  cmd="${@:3:2} -x hip -target x86_64-unknown-linux-gnu --offload-arch=gfx908 --offload-arch=gfx90a --offload-arch=gfx940 --offload-arch=gfx941 --offload-arch=gfx942 --offload-arch=gfx1030 --offload-arch=gfx1100 --offload-arch=gfx1101 --offload-arch=gfx1102 --offload-arch=native ${@:5} -O3 -o ${npic_filepath}"
+  cmd="${@:3:2} -x hip -target x86_64-unknown-linux-gnu ${@:5} -O3 -o ${npic_filepath}"
 else
   cmd="${@:3} -o ${npic_filepath}"
 fi
diff --git a/src/components/ec/rocm/kernel/Makefile.am b/src/components/ec/rocm/kernel/Makefile.am
index 6f95d2b33b..17c00bf95d 100644
--- a/src/components/ec/rocm/kernel/Makefile.am
+++ b/src/components/ec/rocm/kernel/Makefile.am
@@ -17,10 +17,10 @@ HIPCCFLAGS =                                     \
 LINK = $(LIBTOOL) --mode=link $(CC) -o $@
 
 .cu.o:
-	$(HIPCC) -c $< -o $@  $(HIPCCFLAGS)
+	$(HIPCC) -c $< -o $@  $(ROCM_ARCH) $(HIPCCFLAGS)
 
 .cu.lo:
-	/bin/bash $(top_srcdir)/cuda_lt.sh "$(LIBTOOL)" $@ $(HIPCC) -c  $< $(HIPCCFLAGS)
+	/bin/bash $(top_srcdir)/cuda_lt.sh "$(LIBTOOL)" $@ $(HIPCC) -c  $< $(ROCM_ARCH) $(HIPCCFLAGS)
 
 comp_noinst = libucc_ec_rocm_kernels.la