From 39f259f89332e936122f7e02d6088a1dae3fb628 Mon Sep 17 00:00:00 2001
From: Gabriele Bozzola <gbozzola@amazon.com>
Date: Tue, 4 Aug 2026 02:38:58 -0700
Subject: [PATCH] Add support for multi-arch CUDA binaries

This PR follows MAGMA's example and adds `CUDA_TARGETS`. `CUDA_TARGETS`
can be used to bake multiple `CUDA_ARCHs` in a single binary.

In particular, like in MAGMA, I build real targets (SASS) for all the
architectures as well as a virtual target (PTX) for the last one. This
allows architectures that were not in CUDA_ARCH to run by compiling the
PTX directly.
---
 Makefile  | 16 ++++++++++++++--
 README.md |  6 ++++++
 2 files changed, 20 insertions(+), 2 deletions(-)

diff --git a/Makefile b/Makefile
index d4a44790d5..803b371052 100644
--- a/Makefile
+++ b/Makefile
@@ -45,6 +45,7 @@ XSMM_DIR ?= ../libxsmm
 # Often /opt/cuda or /usr/local/cuda, but sometimes present on machines that don't support CUDA
 CUDA_DIR  ?=
 CUDA_ARCH ?=
+CUDA_TARGETS ?=
 
 # Often /opt/rocm, but sometimes present on machines that don't support HIP
 ROCM_DIR ?=
@@ -188,7 +189,18 @@ CXXFLAGS ?= $(OPT) $(CXXFLAGS.$(CC_VENDOR)) $(if $(PEDANTIC),$(PEDANTICFLAGS))
 FFLAGS ?= $(OPT) $(FFLAGS.$(FC_VENDOR))
 LIBCXX ?= -lstdc++
 NVCCFLAGS ?= -ccbin $(CXX) -Xcompiler '$(OPT)' -Xcompiler -fPIC
-ifneq ($(CUDA_ARCH),)
+CUDA_TARGETS_UNKNOWN := $(filter-out sm_%,$(CUDA_TARGETS))
+CUDA_SMS := $(patsubst sm_%,%,$(filter sm_%,$(CUDA_TARGETS)))
+CUDA_SMS := $(shell printf "%s\n" $(CUDA_SMS) | sort -n)
+cuda_gencode_sm = -gencode arch=compute_$(1),code=sm_$(1)
+cuda_gencode_compute = -gencode arch=compute_$(1),code=compute_$(1)
+ifneq ($(strip $(CUDA_TARGETS)),)
+  ifneq ($(strip $(CUDA_TARGETS_UNKNOWN)),)
+    $(error Unknown CUDA target(s): $(CUDA_TARGETS_UNKNOWN); expected targets such as sm_80)
+  endif
+  NVCCFLAGS += $(foreach sm,$(CUDA_SMS),$(call cuda_gencode_sm,$(sm))) \
+    $(call cuda_gencode_compute,$(lastword $(CUDA_SMS)))
+else ifneq ($(strip $(CUDA_ARCH)),)
   NVCCFLAGS += -arch=$(CUDA_ARCH)
 endif
 HIPCCFLAGS ?= $(filter-out $(OMP_SIMD_FLAG),$(OPT)) -fPIC -munsafe-fp-atomics
@@ -1049,7 +1061,7 @@ print-% :
 CONFIG_VARS = CC CXX FC NVCC NVCC_CXX HIPCC \
   OPT CFLAGS CPPFLAGS CXXFLAGS FFLAGS NVCCFLAGS HIPCCFLAGS SYCLFLAGS \
   AR ARFLAGS LDFLAGS LDLIBS LIBCXX SED \
-  MAGMA_DIR XSMM_DIR CUDA_DIR CUDA_ARCH MFEM_DIR PETSC_DIR NEK5K_DIR ROCM_DIR HIP_ARCH SYCL_DIR
+  MAGMA_DIR XSMM_DIR CUDA_DIR CUDA_ARCH CUDA_TARGETS MFEM_DIR PETSC_DIR NEK5K_DIR ROCM_DIR HIP_ARCH SYCL_DIR
 
 # $(call needs_save,CFLAGS) returns true (a nonempty string) if CFLAGS
 # was set on the command line or in config.mk (where it will appear as
diff --git a/README.md b/README.md
index c8aa8f3e26..e669d01b53 100644
--- a/README.md
+++ b/README.md
@@ -62,6 +62,12 @@ $ make AVX=0
 if your compiler does not support gcc-style options, if you are cross compiling, etc.
 
 To enable CUDA support, add `CUDA_DIR=/opt/cuda` or an appropriate directory to your `make` invocation.
+To target one or more CUDA architectures, pass a space-separated `CUDA_TARGETS` list:
+
+```console
+$ make CUDA_DIR=/opt/cuda CUDA_TARGETS='sm_80 sm_90'
+```
+
 To enable HIP support, add `ROCM_DIR=/opt/rocm` or an appropriate directory.
 To enable SYCL support, add `SYCL_DIR=/opt/sycl` or an appropriate directory.
 Note that SYCL backends require building with oneAPI compilers as well: