Cycles: Use pre-compiled PTX kernel for older generation when no matching one is found

This patch changes the discovery of pre-compiled kernels, to look for any PTX, even if
it does not match the current architecture version exactly. It works because the driver can
JIT-compile PTX generated for architectures less than or equal to the current one.
This e.g. makes it possible to render on a new GPU architecture even if no pre-compiled
binary kernel was distributed for it as part of the Blender installation.

Reviewed By: brecht

Differential Revision: https://developer.blender.org/D8332

2 files changed, 20 insertions, 7 deletions

diff --git a/intern/cycles/device/cuda/device_cuda_impl.cpp b/intern/cycles/device/cuda/device_cuda_impl.cpp
index 0be2c322dfa..3a2eb8df95b 100644
--- a/intern/cycles/device/cuda/device_cuda_impl.cpp
+++ b/intern/cycles/device/cuda/device_cuda_impl.cpp
@@ -383,11 +383,24 @@ string CUDADevice::compile_kernel(const DeviceRequestedFeatures &requested_featu
       }
     }
 
-    const string ptx = path_get(string_printf("lib/%s_compute_%d%d.ptx", name, major, minor));
-    VLOG(1) << "Testing for pre-compiled kernel " << ptx << ".";
-    if (path_exists(ptx)) {
-      VLOG(1) << "Using precompiled kernel.";
-      return ptx;
+    /* The driver can JIT-compile PTX generated for older generations, so find the closest one. */
+    int ptx_major = major, ptx_minor = minor;
+    while (ptx_major >= 3) {
+      const string ptx = path_get(
+          string_printf("lib/%s_compute_%d%d.ptx", name, ptx_major, ptx_minor));
+      VLOG(1) << "Testing for pre-compiled kernel " << ptx << ".";
+      if (path_exists(ptx)) {
+        VLOG(1) << "Using precompiled kernel.";
+        return ptx;
+      }
+
+      if (ptx_minor > 0) {
+        ptx_minor--;
+      }
+      else {
+        ptx_major--;
+        ptx_minor = 9;
+      }
     }
   }
 
diff --git a/intern/cycles/kernel/CMakeLists.txt b/intern/cycles/kernel/CMakeLists.txt
index 7cc0d32d521..5533eeb006d 100644
--- a/intern/cycles/kernel/CMakeLists.txt
+++ b/intern/cycles/kernel/CMakeLists.txt
@@ -539,7 +539,7 @@ if(WITH_CYCLES_DEVICE_OPTIX AND WITH_CYCLES_CUDA_BINARIES)
           ${SRC_UTIL_HEADERS}
         COMMAND ${CUBIN_CC_ENV}
             "$<TARGET_FILE:cycles_cubin_cc>"
-            -target 30
+            -target 52
             -ptx
             -i ${CMAKE_CURRENT_SOURCE_DIR}/${input}
             ${cuda_flags}
@@ -563,7 +563,7 @@ if(WITH_CYCLES_DEVICE_OPTIX AND WITH_CYCLES_CUDA_BINARIES)
         COMMAND
           ${CUDA_NVCC_EXECUTABLE}
           --ptx
-          -arch=sm_30
+          -arch=sm_52
           ${cuda_flags}
           ${input}
         WORKING_DIRECTORY