cudaPackages.cuda_nvcc: patch nvcc.profile

pkgs/development/cuda-modules/cuda/overrides.nix

@@ -61,7 +61,59 @@ attrsets.filterAttrs (attr: _: (builtins.hasAttr attr prev)) {
 
   cuda_nvcc = prev.cuda_nvcc.overrideAttrs (
     oldAttrs: {
-      propagatedBuildInputs = [final.setupCudaHook];
+
+      outputs = oldAttrs.outputs ++ [ "lib" ];
+
+      # Patch the nvcc.profile.
+      # Syntax:
+      # - `=` for assignment,
+      # - `?=` for conditional assignment,
+      # - `+=` to "prepend",
+      # - `=+` to "append".
+
+      # Cf. https://web.archive.org/web/20230308044351/https://arcb.csc.ncsu.edu/~mueller/cluster/nvidia/2.0/nvcc_2.0.pdf
+
+      # We set all variables with the lowest priority (=+), but we do force
+      # nvcc to use the fixed backend toolchain. Cf. comments in
+      # backend-stdenv.nix
+
+      postPatch =
+        (oldAttrs.postPatch or "")
+        + ''
+          substituteInPlace bin/nvcc.profile \
+            --replace \
+              '$(TOP)/lib' \
+              "''${!outputLib}/lib" \
+            --replace \
+              '$(TOP)/$(_NVVM_BRANCH_)' \
+              "''${!outputBin}/nvvm" \
+            --replace \
+              '$(TOP)/$(_TARGET_DIR_)/include' \
+              "''${!outputDev}/include"
+
+          cat << EOF >> bin/nvcc.profile
+
+          # Fix a compatible backend compiler
+          PATH += ${lib.getBin final.backendStdenv.cc}/bin:
+          LIBRARIES += "-L${lib.getLib final.backendStdenv.nixpkgsCompatibleLibstdcxx}/lib"
+
+          # Expose the split-out nvvm
+          LIBRARIES =+ -L''${!outputBin}/nvvm/lib
+          INCLUDES =+ -I''${!outputBin}/nvvm/include
+
+          # Expose cudart and the libcuda stubs
+          LIBRARIES =+ -L$static/lib" "-L${final.cuda_cudart.lib}/lib -L${final.cuda_cudart.lib}/lib/stubs
+          INCLUDES =+ -I${final.cuda_cudart.dev}/include
+          EOF
+        '';
+
+      propagatedBuildInputs = [ final.setupCudaHook ];
+
+      postInstall =
+        (oldAttrs.postInstall or "")
+        + ''
+          moveToOutput "nvvm" "''${!outputBin}"
+        '';
 
       meta = (oldAttrs.meta or {}) // {
         mainProgram = "nvcc";

pkgs/development/cuda-modules/saxpy/default.nix

@@ -14,6 +14,7 @@ let
     cudaVersion
     flags
     libcublas
+    setupCudaHook
     ;
 in
 backendStdenv.mkDerivation {

pkgs/development/cuda-modules/setup-hooks/setup-cuda-hook.sh

@@ -93,26 +93,6 @@ setupCUDAToolkitCompilers() {
     if [[ -z "${dontCompressFatbin-}" ]]; then
         export NVCC_PREPEND_FLAGS+=" -Xfatbin=-compress-all"
     fi
-
-    # CMake's enable_language(CUDA) runs a compiler test and it doesn't account for
-    # CUDAToolkit_ROOT. We have to help it locate libcudart
-    if [[ -z "${nvccDontPrependCudartFlags-}" ]] ; then
-        if [[ ! -v cudaOutputToPath["cuda_cudart-out"] ]] ; then
-            echo "setupCUDAToolkitCompilers: missing cudaPackages.cuda_cudart. This may become an an error in the future" >&2
-            # exit 1
-        fi
-        for pkg in "${!cudaOutputToPath[@]}" ; do
-            [[ ! "$pkg" = cuda_cudart* ]] && continue
-
-            local path="${cudaOutputToPath[$pkg]}"
-            if [[ -d "$path/include" ]] ; then
-                export NVCC_PREPEND_FLAGS+=" -I$path/include"
-            fi
-            if [[ -d "$path/lib" ]] ; then
-                export NVCC_PREPEND_FLAGS+=" -L$path/lib"
-            fi
-        done
-    fi
 }
 preConfigureHooks+=(setupCUDAToolkitCompilers)