[Bf-blender-cvs] [2d8c59ccb9a] master: Fix T77095: fix Cycles performance regression with AMD RX cards

[Brecht Van Lommel]

Commit: 2d8c59ccb9ad50de9d56d0d6b6d14815348d123f
Author: Brecht Van Lommel
Date:   Fri Jun 26 17:43:32 2020 +0200
Branches: master
https://developer.blender.org/rB2d8c59ccb9ad50de9d56d0d6b6d14815348d123f

Fix T77095: fix Cycles performance regression with AMD RX cards

Apply the workaround only for known problematic drivers. The latest pro driver
appears to work correctly, hopefully the regular driver will as well once it
is updated to the same OpenCL driver version (3075.13).

===================================================================

M	intern/cycles/device/opencl/device_opencl_impl.cpp

===================================================================

diff --git a/intern/cycles/device/opencl/device_opencl_impl.cpp b/intern/cycles/device/opencl/device_opencl_impl.cpp
index 431aaf2ee45..8c94815b193 100644
--- a/intern/cycles/device/opencl/device_opencl_impl.cpp
+++ b/intern/cycles/device/opencl/device_opencl_impl.cpp
@@ -1942,10 +1942,8 @@ void OpenCLDevice::bake(DeviceTask &task, RenderTile &rtile)
   clFinish(cqCommandQueue);
 }
 
-string OpenCLDevice::kernel_build_options(const string *debug_src)
+static bool kernel_build_opencl_2(cl_device_id cdDevice)
 {
-  string build_options = "-cl-no-signed-zeros -cl-mad-enable ";
-
   /* Build with OpenCL 2.0 if available, this improves performance
    * with AMD OpenCL drivers on Windows and Linux (legacy drivers).
    * Note that OpenCL selects the highest 1.x version by default,
@@ -1953,18 +1951,36 @@ string OpenCLDevice::kernel_build_options(const string *debug_src)
   int version_major, version_minor;
   if (OpenCLInfo::get_device_version(cdDevice, &version_major, &version_minor)) {
     if (version_major >= 2) {
-      /* This appears to trigger a driver bug in Radeon RX cards, so we
-       * don't use OpenCL 2.0 for those. */
+      /* This appears to trigger a driver bug in Radeon RX cards with certain
+       * driver version, so don't use OpenCL 2.0 for those. */
       string device_name = OpenCLInfo::get_readable_device_name(cdDevice);
-      if (!(string_startswith(device_name, "Radeon RX 4") ||
-            string_startswith(device_name, "Radeon (TM) RX 4") ||
-            string_startswith(device_name, "Radeon RX 5") ||
-            string_startswith(device_name, "Radeon (TM) RX 5"))) {
-        build_options += "-cl-std=CL2.0 ";
+      if (string_startswith(device_name, "Radeon RX 4") ||
+          string_startswith(device_name, "Radeon (TM) RX 4") ||
+          string_startswith(device_name, "Radeon RX 5") ||
+          string_startswith(device_name, "Radeon (TM) RX 5")) {
+        char version[256] = "";
+        int driver_major, driver_minor;
+        clGetDeviceInfo(cdDevice, CL_DEVICE_VERSION, sizeof(version), &version, NULL);
+        if (sscanf(version, "OpenCL 2.0 AMD-APP (%d.%d)", &driver_major, &driver_minor) == 2) {
+          return !(driver_major == 3075 && driver_minor <= 12);
+        }
       }
+
+      return true;
     }
   }
 
+  return false;
+}
+
+string OpenCLDevice::kernel_build_options(const string *debug_src)
+{
+  string build_options = "-cl-no-signed-zeros -cl-mad-enable ";
+
+  if (kernel_build_opencl_2(cdDevice)) {
+    build_options += "-cl-std=CL2.0 ";
+  }
+
   if (platform_name == "NVIDIA CUDA") {
     build_options +=
         "-D__KERNEL_OPENCL_NVIDIA__ "