Optional offloading to AMD GPUs

build/bli_config.h.in

@@ -183,5 +183,10 @@
 #define BLIS_DISABLE_COMPLEX_RETURN_INTEL
 #endif
 
+#if @enable_amd_offload@
+#define BLIS_ENABLE_AMD_OFFLOAD
+#else
+#define BLIS_DISABLE_AMD_OFFLOAD
+#endif
 
 #endif

build/config.mk.in

@@ -208,5 +208,8 @@ LIBPTHREAD        := @libpthread@
 # Whether we should use AMD-customized versions of certain framework files.
 ENABLE_AMD_FRAME_TWEAKS := @enable_amd_frame_tweaks@
 
+# Whether offloading to AMD accelerators should be attempted
+ENABLE_AMD_OFFLOAD := @enable_amd_offload@
+
 # end of ifndef CONFIG_MK_INCLUDED conditional block
 endif

common.mk

@@ -552,6 +552,11 @@ ifeq ($(DEBUG_TYPE),sde)
 LDFLAGS    := $(filter-out $(LIBMEMKIND),$(LDFLAGS))
 endif
 
+ifeq ($(ENABLE_AMD_OFFLOAD),yes)
+LDFLAGS    += -ldl
+LDFLAGS    += -L/opt/rocm/lib -lamdhip64 -lrocblas
+endif
+
 # Specify the shared library's 'soname' field.
 # NOTE: The flag for creating shared objects is different for Linux and OS X.
 ifeq ($(OS_NAME),Darwin)
@@ -1134,6 +1139,11 @@ ifeq ($(MK_ENABLE_CBLAS),yes)
 CINCFLAGS       += -I$(CBLAS_H_DIRPATH)
 endif
 
+# If AMD offloading is enabled, we also add the ROCm include directory
+ifeq ($(ENABLE_AMD_OFFLOAD),yes)
+CINCFLAGS	+= -I/opt/rocm/include -D__HIP_PLATFORM_AMD__=1
+endif
+
 # Obtain a list of header paths in the configured addons. Then add -I to each
 # header path.
 CADDONINCFLAGS  := $(strip $(patsubst %, -I%, $(ADDON_HDR_DIRPATHS)))

configure

@@ -285,6 +285,15 @@ print_usage()
 	echo "                 which are determined by the BLIS subconfiguration used at"
 	echo "                 runtime.) By default, these customized files are disabled."
 	echo " "
+	echo "   --enable-amd-offload, --disable-amd-offload"
+	echo " "
+	echo "                 Enable conditional offloading of some Level-3 BLAS calls"
+	echo "                 to AMD accelerators such as MI100, MI200."
+	echo "                 Enabling this option requires ROCm to be installed and"
+	echo "                 uses rocBLAS as a backend."
+	echo "                 Introduces rocblas-dev and hip-dev as dependencies."
+	echo "                 By default, the offloading path are disabled."
+	echo " "
 	echo "   -a NAME --enable-addon=NAME"
 	echo " "
 	echo "                 Enable the code provided by an addon. An addon consists"
@@ -2469,6 +2478,7 @@ main()
 	enable_mixed_dt_extra_mem='yes'
 	enable_sup_handling='yes'
 	enable_amd_frame_tweaks='no'
+	enable_amd_offload='no'
 	enable_memkind='' # The default memkind value is determined later on.
 	enable_trsm_preinversion='yes'
 	force_version='no'
@@ -2687,6 +2697,12 @@ main()
 						disable-amd-frame-tweaks)
 							enable_amd_frame_tweaks='no'
 							;;
+						enable-amd-offload)
+							enable_amd_offload='yes'
+							;;
+						disable-amd-offload)
+							enable_amd_offload='no'
+							;;
 						with-memkind)
 							enable_memkind='yes'
 							;;
@@ -3616,6 +3632,29 @@ main()
 		echo "${script_name}: AMD-specific framework files will not be considered."
 	fi
 
+	# Check whether anything should be offloaded to AMD accelerators
+	enable_amd_offload_01=0
+	if [ "x${enable_amd_offload}" = "xyes" ]; then
+		echo "${script_name}: Offloading to AMD accelerators will be considered."
+		echo "${script_name}:   checking for ROCm installation and availability."
+
+		# Make sure there's a ROCm installation present
+		# use rocm_agent_enumerator to see if there's a gfx != gfx000
+		gfxs=`rocm_agent_enumerator`
+		if [ -z "$gfxs" ]; then
+			echo "${script_name}:   rocm_agent_enumerator returns no agents."
+			enable_amd_offload='no'
+		elif [[ "$gfxs" =~ "gfx9" ]] || [[ "$gfxs" =~ "gfx10" ]]; then
+			echo "${script_name}:   found AMD accelerator(s)."
+			enable_amd_offload_01=1
+		else
+			echo "${script_name}:   Illegal rocm_agent_enumerator output. $gfsx"
+			enable_amd_offload='no'
+		fi
+	else
+		echo "${script_name}: Offloading to AMD accelerators will not be considered."
+	fi
+
 	# Check if addons were given.
 	if [ -n "${addon_flag}" ]; then
 
@@ -3871,6 +3910,7 @@ main()
 		| sed -e "s/@enable_blas@/${enable_blas}/g" \
 		| sed -e "s/@enable_cblas@/${enable_cblas}/g" \
 		| sed -e "s/@enable_amd_frame_tweaks@/${enable_amd_frame_tweaks}/g" \
+		| sed -e "s/@enable_amd_offload@/${enable_amd_offload}/g" \
 		| sed -e "s/@enable_memkind@/${enable_memkind}/g" \
 		| sed -e "s/@pragma_omp_simd@/${pragma_omp_simd}/g" \
 		| sed -e "s/@addon_list@/${addon_list}/g" \
@@ -3910,6 +3950,7 @@ main()
 		| sed   -e "s/@enable_sandbox@/${enable_sandbox_01}/g" \
 		| sed   -e "s/@enable_shared@/${enable_shared_01}/g" \
 		| sed   -e "s/@complex_return_intel@/${complex_return_intel01}/g" \
+		| sed   -e "s/@enable_amd_offload@/${enable_amd_offload_01}/g" \
 		> "${bli_config_h_out_path}"
 
 	# -- Instantiate bli_addon.h file from template ----------------------------

frame/3/bli_l3_oapi_ex.c

@@ -33,6 +33,9 @@
 */
 
 #include "blis.h"
+#ifdef BLIS_ENABLE_AMD_OFFLOAD
+#include "../base/bli_offloader.h"
+#endif
 
 //
 // Define object-based interfaces (expert).
@@ -73,6 +76,21 @@ void PASTEMAC(gemm,BLIS_OAPI_EX_SUF)
 			return;
 		}
 	}
+#ifdef BLIS_ENABLE_AMD_OFFLOAD
+	// check if we should offload - since attempting to offload and fail
+	// incurrs a non-trivial cost, we only want to fail and fall through
+	// in rare cases
+	const bool do_offload = bli_do_offload_gemmex( alpha, a, b, beta, c);
+	if ( do_offload )
+	{
+		// attempts to offload
+		const err_t result = bli_offload_gemmex( alpha, a, b, beta, c);
+		if ( result == BLIS_SUCCESS )
+		{
+			return;
+		}
+	}
+#endif
 
 	// Initialize a local runtime with global settings if necessary. Note
 	// that in the case that a runtime is passed in, we make a local copy.

frame/base/bli_init.c

@@ -34,6 +34,9 @@
 */
 
 #include "blis.h"
+#ifdef BLIS_ENABLE_AMD_OFFLOAD
+#include "bli_offloader.h"
+#endif
 
 // -----------------------------------------------------------------------------
 
@@ -87,7 +90,11 @@ int bli_init_apis( void )
 	bli_pack_init();
 	bli_memsys_init();
 
-	return 0;
+#ifdef BLIS_ENABLE_AMD_OFFLOAD
+        bli_offloader_init();
+#endif
+
+        return 0;
 }
 
 int bli_finalize_apis( void )
@@ -99,6 +106,10 @@ int bli_finalize_apis( void )
 	bli_ind_finalize();
 	bli_gks_finalize();
 
+#ifdef BLIS_ENABLE_AMD_OFFLOAD
+        bli_offloader_finalize();
+#endif
+
 	return 0;
 }