blender
/
blender
136
404
Fork 586
Code Issues 6.2k Pull Requests 753 Projects 19 Wiki Activity

Cycles: reduce kernel reserved local memory when not using shader raytracing 

Ref T87836
This commit is contained in:
Brecht Van Lommel 2021-10-20 14:21:01 +02:00
parent 2537b32392
commit 001f548227
Notes: blender-bot 2023-02-14 06:00:49 +01:00 
Referenced by issue #87836, Cycles: GPU Performance
2 changed files with 8 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
intern/cycles/device/cuda/device_impl.cpp
View File

@ -454,7 +454,7 @@ bool CUDADevice::load_kernels(const uint kernel_features)
return (result == CUDA_SUCCESS);
}
void CUDADevice::reserve_local_memory(const uint /* kernel_features */)
void CUDADevice::reserve_local_memory(const uint kernel_features)
{
/* Together with CU_CTX_LMEM_RESIZE_TO_MAX, this reserves local memory
* needed for kernel launches, so that we can reliably figure out when
@ -468,7 +468,9 @@ void CUDADevice::reserve_local_memory(const uint /* kernel_features */)
{
/* Use the biggest kernel for estimation. */
const DeviceKernel test_kernel = DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE;
const DeviceKernel test_kernel = (kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) ?
DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE :
DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE;
/* Launch kernel, using just 1 block appears sufficient to reserve memory for all
* multiprocessors. It would be good to do this in parallel for the multi GPU case

6
intern/cycles/device/hip/device_impl.cpp
View File

@ -430,7 +430,7 @@ bool HIPDevice::load_kernels(const uint kernel_features)
return (result == hipSuccess);
}
void HIPDevice::reserve_local_memory(const uint)
void HIPDevice::reserve_local_memory(const uint kernel_features)
{
/* Together with hipDeviceLmemResizeToMax, this reserves local memory
* needed for kernel launches, so that we can reliably figure out when
@ -444,7 +444,9 @@ void HIPDevice::reserve_local_memory(const uint)
{
/* Use the biggest kernel for estimation. */
const DeviceKernel test_kernel = DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE;
const DeviceKernel test_kernel = (kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) ?
DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE :
DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE;
/* Launch kernel, using just 1 block appears sufficient to reserve memory for all
* multiprocessors. It would be good to do this in parallel for the multi GPU case