[Bf-blender-cvs] [555cdae6de8] cycles-x: Cleanup: add mechanism for devices to not have a megakernel

[Brecht Van Lommel]

Commit: 555cdae6de861edba6173d7ea9912131cb3e5ee8
Author: Brecht Van Lommel
Date:   Wed Apr 28 16:59:01 2021 +0200
Branches: cycles-x
https://developer.blender.org/rB555cdae6de861edba6173d7ea9912131cb3e5ee8

Cleanup: add mechanism for devices to not have a megakernel

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

M	intern/cycles/device/cuda/kernel.cpp
M	intern/cycles/device/cuda/kernel.h
M	intern/cycles/device/cuda/queue.cpp
M	intern/cycles/device/cuda/queue.h
M	intern/cycles/device/device_queue.h
M	intern/cycles/integrator/path_trace_work_gpu.cpp

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

diff --git a/intern/cycles/device/cuda/kernel.cpp b/intern/cycles/device/cuda/kernel.cpp
index 0dd776769d2..793c9efe72a 100644
--- a/intern/cycles/device/cuda/kernel.cpp
+++ b/intern/cycles/device/cuda/kernel.cpp
@@ -48,6 +48,11 @@ const CUDADeviceKernel &CUDADeviceKernels::get(DeviceKernel kernel) const
   return kernels_[(int)kernel];
 }
 
+bool CUDADeviceKernels::available(DeviceKernel kernel) const
+{
+  return kernels_[(int)kernel].function != nullptr;
+}
+
 CCL_NAMESPACE_END
 
 #endif /* WITH_CUDA*/
diff --git a/intern/cycles/device/cuda/kernel.h b/intern/cycles/device/cuda/kernel.h
index dba8702f691..b489547a350 100644
--- a/intern/cycles/device/cuda/kernel.h
+++ b/intern/cycles/device/cuda/kernel.h
@@ -33,10 +33,10 @@ class CUDADevice;
 /* CUDA kernel and associate occupancy information. */
 class CUDADeviceKernel {
  public:
-  CUfunction function;
+  CUfunction function = nullptr;
 
-  int num_threads_per_block;
-  int min_blocks;
+  int num_threads_per_block = 0;
+  int min_blocks = 0;
 };
 
 /* Cache of CUDA kernels for each DeviceKernel. */
@@ -44,6 +44,7 @@ class CUDADeviceKernels {
  public:
   void load(CUDADevice *device);
   const CUDADeviceKernel &get(DeviceKernel kernel) const;
+  bool available(DeviceKernel kernel) const;
 
  protected:
   CUDADeviceKernel kernels_[DEVICE_KERNEL_NUM];
diff --git a/intern/cycles/device/cuda/queue.cpp b/intern/cycles/device/cuda/queue.cpp
index 9ee0298aa2a..32af06d85df 100644
--- a/intern/cycles/device/cuda/queue.cpp
+++ b/intern/cycles/device/cuda/queue.cpp
@@ -47,6 +47,11 @@ void CUDADeviceQueue::init_execution()
   debug_init_execution();
 }
 
+bool CUDADeviceQueue::kernel_available(DeviceKernel kernel) const
+{
+  return cuda_device_->kernels.available(kernel);
+}
+
 bool CUDADeviceQueue::enqueue(DeviceKernel kernel, const int work_size, void *args[])
 {
   if (cuda_device_->have_error()) {
diff --git a/intern/cycles/device/cuda/queue.h b/intern/cycles/device/cuda/queue.h
index 59e10bcd711..acdcfd8a69e 100644
--- a/intern/cycles/device/cuda/queue.h
+++ b/intern/cycles/device/cuda/queue.h
@@ -37,6 +37,8 @@ class CUDADeviceQueue : public DeviceQueue {
 
   virtual void init_execution() override;
 
+  virtual bool kernel_available(DeviceKernel kernel) const override;
+
   virtual bool enqueue(DeviceKernel kernel, const int work_size, void *args[]) override;
 
   virtual bool synchronize() override;
diff --git a/intern/cycles/device/device_queue.h b/intern/cycles/device/device_queue.h
index 5d556ec44dc..dffa3ce26b7 100644
--- a/intern/cycles/device/device_queue.h
+++ b/intern/cycles/device/device_queue.h
@@ -43,6 +43,9 @@ class DeviceQueue {
    * Use this method after device synchronization has finished before enqueueing any kernels. */
   virtual void init_execution() = 0;
 
+  /* Test if an optional device kernel is available. */
+  virtual bool kernel_available(DeviceKernel kernel) const = 0;
+
   /* Enqueue kernel execution.
    *
    * Execute the kernel work_size times on the device.
diff --git a/intern/cycles/integrator/path_trace_work_gpu.cpp b/intern/cycles/integrator/path_trace_work_gpu.cpp
index 3590a788529..d477c92da13 100644
--- a/intern/cycles/integrator/path_trace_work_gpu.cpp
+++ b/intern/cycles/integrator/path_trace_work_gpu.cpp
@@ -204,7 +204,8 @@ bool PathTraceWorkGPU::enqueue_path_iteration()
 
   const int max_num_paths = get_max_num_paths();
   const float megakernel_threshold = 0.02f;
-  const bool use_megakernel = (num_paths < megakernel_threshold * max_num_paths);
+  const bool use_megakernel = queue_->kernel_available(DEVICE_KERNEL_INTEGRATOR_MEGAKERNEL) &&
+                              (num_paths < megakernel_threshold * max_num_paths_);
 
   if (use_megakernel) {
     enqueue_path_iteration(DEVICE_KERNEL_INTEGRATOR_MEGAKERNEL);