From e6ea6db38a036f2f11a89ce39494022d53a2eca8 Mon Sep 17 00:00:00 2001
From: Alexis Girault <agirault@nvidia.com>
Date: Mon, 31 Mar 2025 12:58:07 -0400
Subject: [PATCH 1/4] build: use sm list from nvcc

Signed-off-by: Alexis Girault <agirault@nvidia.com>
---
 scripts/get_cuda_gencode.sh | 94 ++++++++++++++++++++++---------------
 1 file changed, 55 insertions(+), 39 deletions(-)

diff --git a/scripts/get_cuda_gencode.sh b/scripts/get_cuda_gencode.sh
index 666d604..7c51526 100755
--- a/scripts/get_cuda_gencode.sh
+++ b/scripts/get_cuda_gencode.sh
@@ -17,50 +17,66 @@ if [ "$CUDA_VERSION_MAJOR" -lt 8 ]; then
     exit 1
 fi
 
-# Initialize with Pascal architecture
-COMPUTE_LIST="60 61 62"
-SM_LIST="60 61 62"
-
-# Add Volta (7.0) if CUDA >= 9.0
-if [ "$CUDA_VERSION_MAJOR" -ge 9 ]; then
-    COMPUTE_LIST="$COMPUTE_LIST 70 72"
-    SM_LIST="$SM_LIST 70 72"
-fi
+# Get the list of supported SM architectures (sm_XX) from nvcc
+# Filter to only include SM >= 60 (Pascal)
+SUPPORTED_SM_LIST=$("$NVCC" --list-gpu-code 2>/dev/null | sed 's/sm_//' | awk '$1 >= 60')
 
-# Add Turing (7.5) if CUDA >= 10.0
-if [ "$CUDA_VERSION_MAJOR" -ge 10 ]; then
-    COMPUTE_LIST="$COMPUTE_LIST 75"
-    SM_LIST="$SM_LIST 75"
-fi
+# Initialize empty lists
+COMPUTE_LIST=""
+SM_LIST=""
 
-# Add Ampere (8.0, 8.6, 8.7) if CUDA >= 11.1
-if [ "$CUDA_VERSION_MAJOR" -ge 11 ] && [ "$CUDA_VERSION_MINOR" -ge 1 ]; then
-    COMPUTE_LIST="$COMPUTE_LIST 80 86 87"
-    SM_LIST="$SM_LIST 80 86 87"
-fi
+if [ -n "$SUPPORTED_SM_LIST" ]; then
+    # Use the list of architectures supported by nvcc
+    SM_LIST=$SUPPORTED_SM_LIST
+    COMPUTE_LIST=$SUPPORTED_SM_LIST
+else
+    echo "Warning: Could not determine supported architectures from nvcc, falling back to version-based detection" >&2
 
-# Add Ada Lovelace (8.9) if CUDA >= 11.8
-if [ "$CUDA_VERSION_MAJOR" -ge 11 ] && [ "$CUDA_VERSION_MINOR" -ge 8 ]; then
-    COMPUTE_LIST="$COMPUTE_LIST 89"
-    SM_LIST="$SM_LIST 89"
-fi
+    # Initialize with Pascal architecture
+    COMPUTE_LIST="60 61 62"
+    SM_LIST="60 61 62"
 
-# Add Hopper (9.0) if CUDA >= 12.0
-if [ "$CUDA_VERSION_MAJOR" -ge 12 ]; then
-    COMPUTE_LIST="$COMPUTE_LIST 90"
-    SM_LIST="$SM_LIST 90"
-fi
+    # Add Volta (7.0) if CUDA >= 9.0
+    if [ "$CUDA_VERSION_MAJOR" -ge 9 ]; then
+        COMPUTE_LIST="$COMPUTE_LIST 70 72"
+        SM_LIST="$SM_LIST 70 72"
+    fi
 
-# Add Blackwell (10.0) if CUDA >= 12.6
-if [ "$CUDA_VERSION_MAJOR" -ge 12 ] && [ "$CUDA_VERSION_MINOR" -ge 6 ]; then
-    COMPUTE_LIST="$COMPUTE_LIST 100"
-    SM_LIST="$SM_LIST 100"
-fi
+    # Add Turing (7.5) if CUDA >= 10.0
+    if [ "$CUDA_VERSION_MAJOR" -ge 10 ]; then
+        COMPUTE_LIST="$COMPUTE_LIST 75"
+        SM_LIST="$SM_LIST 75"
+    fi
+
+    # Add Ampere (8.0, 8.6, 8.7) if CUDA >= 11.1
+    if [ "$CUDA_VERSION_MAJOR" -ge 11 ] && [ "$CUDA_VERSION_MINOR" -ge 1 ]; then
+        COMPUTE_LIST="$COMPUTE_LIST 80 86 87"
+        SM_LIST="$SM_LIST 80 86 87"
+    fi
+
+    # Add Ada Lovelace (8.9) if CUDA >= 11.8
+    if [ "$CUDA_VERSION_MAJOR" -ge 11 ] && [ "$CUDA_VERSION_MINOR" -ge 8 ]; then
+        COMPUTE_LIST="$COMPUTE_LIST 89"
+        SM_LIST="$SM_LIST 89"
+    fi
+
+    # Add Hopper (9.0) if CUDA >= 12.0
+    if [ "$CUDA_VERSION_MAJOR" -ge 12 ]; then
+        COMPUTE_LIST="$COMPUTE_LIST 90"
+        SM_LIST="$SM_LIST 90"
+    fi
+
+    # Add Blackwell (10.0) if CUDA >= 12.6
+    if [ "$CUDA_VERSION_MAJOR" -ge 12 ] && [ "$CUDA_VERSION_MINOR" -ge 6 ]; then
+        COMPUTE_LIST="$COMPUTE_LIST 100"
+        SM_LIST="$SM_LIST 100"
+    fi
 
-# Add Blackwell (12.0) if CUDA >= 12.8
-if [ "$CUDA_VERSION_MAJOR" -ge 12 ] && [ "$CUDA_VERSION_MINOR" -ge 8 ]; then
-    COMPUTE_LIST="$COMPUTE_LIST 120"
-    SM_LIST="$SM_LIST 120"
+    # Add Blackwell (12.0) if CUDA >= 12.8
+    if [ "$CUDA_VERSION_MAJOR" -ge 12 ] && [ "$CUDA_VERSION_MINOR" -ge 8 ]; then
+        COMPUTE_LIST="$COMPUTE_LIST 120"
+        SM_LIST="$SM_LIST 120"
+    fi
 fi
 
 # Generate NVCC flags
@@ -73,4 +89,4 @@ for sm in $SM_LIST; do
     GENCODE_FLAGS="$GENCODE_FLAGS -gencode arch=compute_$sm,code=sm_$sm"
 done
 
-echo "$GENCODE_FLAGS" 
\ No newline at end of file
+echo "$GENCODE_FLAGS"

From 72fa69c78d659d465d6b7a04f16f8ec373eacf0b Mon Sep 17 00:00:00 2001
From: Alexis Girault <agirault@nvidia.com>
Date: Mon, 31 Mar 2025 13:03:31 -0400
Subject: [PATCH 2/4] build: correct blackwell sm use

- 10.0 is from CTK 12.8+
- 10.1 was missing

Signed-off-by: Alexis Girault <agirault@nvidia.com>
---
 scripts/get_cuda_gencode.sh | 12 +++---------
 1 file changed, 3 insertions(+), 9 deletions(-)

diff --git a/scripts/get_cuda_gencode.sh b/scripts/get_cuda_gencode.sh
index 7c51526..ee411b6 100755
--- a/scripts/get_cuda_gencode.sh
+++ b/scripts/get_cuda_gencode.sh
@@ -66,16 +66,10 @@ else
         SM_LIST="$SM_LIST 90"
     fi
 
-    # Add Blackwell (10.0) if CUDA >= 12.6
+    # Add Blackwell (10.0, 10.1, 12.0) if CUDA >= 12.8
     if [ "$CUDA_VERSION_MAJOR" -ge 12 ] && [ "$CUDA_VERSION_MINOR" -ge 6 ]; then
-        COMPUTE_LIST="$COMPUTE_LIST 100"
-        SM_LIST="$SM_LIST 100"
-    fi
-
-    # Add Blackwell (12.0) if CUDA >= 12.8
-    if [ "$CUDA_VERSION_MAJOR" -ge 12 ] && [ "$CUDA_VERSION_MINOR" -ge 8 ]; then
-        COMPUTE_LIST="$COMPUTE_LIST 120"
-        SM_LIST="$SM_LIST 120"
+        COMPUTE_LIST="$COMPUTE_LIST 100 101 120"
+        SM_LIST="$SM_LIST 100 101 120"
     fi
 fi
 

From c521788ac86c057f79778bc48477b85df0e399b3 Mon Sep 17 00:00:00 2001
From: Alexis Girault <agirault@nvidia.com>
Date: Mon, 31 Mar 2025 13:09:24 -0400
Subject: [PATCH 3/4] build: consolidate arch lists for ptx and sass

The list was the same

Signed-off-by: Alexis Girault <agirault@nvidia.com>
---
 scripts/get_cuda_gencode.sh | 43 ++++++++++++-------------------------
 1 file changed, 14 insertions(+), 29 deletions(-)

diff --git a/scripts/get_cuda_gencode.sh b/scripts/get_cuda_gencode.sh
index ee411b6..c59969b 100755
--- a/scripts/get_cuda_gencode.sh
+++ b/scripts/get_cuda_gencode.sh
@@ -19,68 +19,53 @@ fi
 
 # Get the list of supported SM architectures (sm_XX) from nvcc
 # Filter to only include SM >= 60 (Pascal)
-SUPPORTED_SM_LIST=$("$NVCC" --list-gpu-code 2>/dev/null | sed 's/sm_//' | awk '$1 >= 60')
+ARCH_LIST=$("$NVCC" --list-gpu-code 2>/dev/null | sed 's/sm_//' | awk '$1 >= 60')
 
-# Initialize empty lists
-COMPUTE_LIST=""
-SM_LIST=""
-
-if [ -n "$SUPPORTED_SM_LIST" ]; then
-    # Use the list of architectures supported by nvcc
-    SM_LIST=$SUPPORTED_SM_LIST
-    COMPUTE_LIST=$SUPPORTED_SM_LIST
-else
+if [ -z "$ARCH_LIST" ]; then
     echo "Warning: Could not determine supported architectures from nvcc, falling back to version-based detection" >&2
 
     # Initialize with Pascal architecture
-    COMPUTE_LIST="60 61 62"
-    SM_LIST="60 61 62"
+    ARCH_LIST="60 61 62"
 
     # Add Volta (7.0) if CUDA >= 9.0
     if [ "$CUDA_VERSION_MAJOR" -ge 9 ]; then
-        COMPUTE_LIST="$COMPUTE_LIST 70 72"
-        SM_LIST="$SM_LIST 70 72"
+        ARCH_LIST="$ARCH_LIST 70 72"
     fi
 
     # Add Turing (7.5) if CUDA >= 10.0
     if [ "$CUDA_VERSION_MAJOR" -ge 10 ]; then
-        COMPUTE_LIST="$COMPUTE_LIST 75"
-        SM_LIST="$SM_LIST 75"
+        ARCH_LIST="$ARCH_LIST 75"
     fi
 
     # Add Ampere (8.0, 8.6, 8.7) if CUDA >= 11.1
     if [ "$CUDA_VERSION_MAJOR" -ge 11 ] && [ "$CUDA_VERSION_MINOR" -ge 1 ]; then
-        COMPUTE_LIST="$COMPUTE_LIST 80 86 87"
-        SM_LIST="$SM_LIST 80 86 87"
+        ARCH_LIST="$ARCH_LIST 80 86 87"
     fi
 
     # Add Ada Lovelace (8.9) if CUDA >= 11.8
     if [ "$CUDA_VERSION_MAJOR" -ge 11 ] && [ "$CUDA_VERSION_MINOR" -ge 8 ]; then
-        COMPUTE_LIST="$COMPUTE_LIST 89"
-        SM_LIST="$SM_LIST 89"
+        ARCH_LIST="$ARCH_LIST 89"
     fi
 
     # Add Hopper (9.0) if CUDA >= 12.0
     if [ "$CUDA_VERSION_MAJOR" -ge 12 ]; then
-        COMPUTE_LIST="$COMPUTE_LIST 90"
-        SM_LIST="$SM_LIST 90"
+        ARCH_LIST="$ARCH_LIST 90"
     fi
 
     # Add Blackwell (10.0, 10.1, 12.0) if CUDA >= 12.8
-    if [ "$CUDA_VERSION_MAJOR" -ge 12 ] && [ "$CUDA_VERSION_MINOR" -ge 6 ]; then
-        COMPUTE_LIST="$COMPUTE_LIST 100 101 120"
-        SM_LIST="$SM_LIST 100 101 120"
+    if [ "$CUDA_VERSION_MAJOR" -ge 12 ] && [ "$CUDA_VERSION_MINOR" -ge 8 ]; then
+        ARCH_LIST="$ARCH_LIST 100 101 120"
     fi
 fi
 
 # Generate NVCC flags
 GENCODE_FLAGS=""
-for compute in $COMPUTE_LIST; do
-    GENCODE_FLAGS="$GENCODE_FLAGS -gencode arch=compute_$compute,code=compute_$compute"
+for arch in $ARCH_LIST; do
+    GENCODE_FLAGS="$GENCODE_FLAGS -gencode arch=compute_$arch,code=compute_$arch"
 done
 
-for sm in $SM_LIST; do
-    GENCODE_FLAGS="$GENCODE_FLAGS -gencode arch=compute_$sm,code=sm_$sm"
+for arch in $ARCH_LIST; do
+    GENCODE_FLAGS="$GENCODE_FLAGS -gencode arch=compute_$arch,code=sm_$arch"
 done
 
 echo "$GENCODE_FLAGS"

From d1fde74a3f65ab59cf1a8958aa41e46505427175 Mon Sep 17 00:00:00 2001
From: Alexis Girault <agirault@nvidia.com>
Date: Mon, 31 Mar 2025 13:14:26 -0400
Subject: [PATCH 4/4] build: only build ptx for last arch

Signed-off-by: Alexis Girault <agirault@nvidia.com>
---
 scripts/get_cuda_gencode.sh | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/scripts/get_cuda_gencode.sh b/scripts/get_cuda_gencode.sh
index c59969b..946ea28 100755
--- a/scripts/get_cuda_gencode.sh
+++ b/scripts/get_cuda_gencode.sh
@@ -60,12 +60,14 @@ fi
 
 # Generate NVCC flags
 GENCODE_FLAGS=""
-for arch in $ARCH_LIST; do
-    GENCODE_FLAGS="$GENCODE_FLAGS -gencode arch=compute_$arch,code=compute_$arch"
-done
 
+# Generate SM-specific code for all architectures
 for arch in $ARCH_LIST; do
     GENCODE_FLAGS="$GENCODE_FLAGS -gencode arch=compute_$arch,code=sm_$arch"
 done
 
+# Generate PTX code only for the latest architecture
+LATEST_ARCH=$(echo "$ARCH_LIST" | tr ' ' '\n' | sort -n | tail -1)
+GENCODE_FLAGS="$GENCODE_FLAGS -gencode arch=compute_$LATEST_ARCH,code=compute_$LATEST_ARCH"
+
 echo "$GENCODE_FLAGS"