Cycles: Allow PTX targets for CUDA kernel build.

This is intended for developers on Windows primarily:
Now, CUDA architectures of type compute_xx are supported. This allows for quicker builds,
at the expense of the CUDA driver running ptxas the first time a kernel is loaded.

Differential Revision: https://developer.blender.org/D5953

intern/cycles/device/device_cuda.cpp

@@ -454,6 +454,12 @@ class CUDADevice : public Device {
         VLOG(1) << "Using precompiled kernel.";
         return cubin;
       }
+      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;
+      }
     }
 
     const string common_cflags = compile_kernel_get_common_cflags(

intern/cycles/kernel/CMakeLists.txt

@@ -389,11 +389,20 @@ if(WITH_CYCLES_CUDA_BINARIES)
   set(cuda_cubins)
 
   macro(CYCLES_CUDA_KERNEL_ADD arch prev_arch name flags sources experimental)
-    set(cuda_cubin ${name}_${arch}.cubin)
+    if(${arch} MATCHES "compute_.*")
+      set(format "ptx")
+    else()
+      set(format "cubin")
+    endif()
+    set(cuda_file ${name}_${arch}.${format})
 
     set(kernel_sources ${sources})
     if(NOT ${prev_arch} STREQUAL "none")
-      set(kernel_sources ${kernel_sources} ${name}_${prev_arch}.cubin)
+      if(${prev_arch} MATCHES "compute_.*")
+        set(kernel_sources ${kernel_sources} ${name}_${prev_arch}.ptx)
+      else()
+        set(kernel_sources ${kernel_sources} ${name}_${prev_arch}.cubin)
+      endif()
     endif()
 
     set(cuda_kernel_src "/kernels/cuda/${name}.cu")
@@ -406,7 +415,7 @@ if(WITH_CYCLES_CUDA_BINARIES)
       -I ${CMAKE_CURRENT_SOURCE_DIR}/..
       -I ${CMAKE_CURRENT_SOURCE_DIR}/kernels/cuda
       --use_fast_math
-      -o ${CMAKE_CURRENT_BINARY_DIR}/${cuda_cubin})
+      -o ${CMAKE_CURRENT_BINARY_DIR}/${cuda_file})
 
     if(${experimental})
       set(cuda_flags ${cuda_flags} -D __KERNEL_EXPERIMENTAL__)
@@ -440,20 +449,21 @@ if(WITH_CYCLES_CUDA_BINARIES)
             -v
             -cuda-toolkit-dir "${CUDA_TOOLKIT_ROOT_DIR}"
         DEPENDS ${kernel_sources} cycles_cubin_cc)
+      set(cuda_file ${cuda_cubin})
     else()
       add_custom_command(
-        OUTPUT ${cuda_cubin}
+        OUTPUT ${cuda_file}
         COMMAND ${CUDA_NVCC_EXECUTABLE}
             -arch=${arch}
             ${CUDA_NVCC_FLAGS}
-            --cubin
+            --${format}
             ${CMAKE_CURRENT_SOURCE_DIR}${cuda_kernel_src}
             --ptxas-options="-v"
             ${cuda_flags}
         DEPENDS ${kernel_sources})
     endif()
-    delayed_install("${CMAKE_CURRENT_BINARY_DIR}" "${cuda_cubin}" ${CYCLES_INSTALL_PATH}/lib)
-    list(APPEND cuda_cubins ${cuda_cubin})
+    delayed_install("${CMAKE_CURRENT_BINARY_DIR}" "${cuda_file}" ${CYCLES_INSTALL_PATH}/lib)
+    list(APPEND cuda_cubins ${cuda_file})
 
     unset(cuda_debug_flags)
   endmacro()