blender
/
blender
129
397
Fork 571
Code Issues 6.2k Pull Requests 749 Projects 19 Wiki Activity

Cycles: Apple GPU threadgroup tuning 

This patch tunes maximum threads-per-threadgroup and threads-per-block for faster renders on Apple GPUs. Appropriate tuning is selected based on the GPU architecture (M1 or M2). We see a benchmark uplift of around 5-10% on M1 family chips. Similar uplift is expected on M2 with upcoming OS changes. (Ref T101931)

Reviewed By: brecht

Maniphest Tasks: T101931

Differential Revision: https://developer.blender.org/D16299
This commit is contained in:
Michael Jones (Apple) 2022-11-04 15:59:55 +00:00
parent 671c3e1fa4
commit 74140d41b1
Notes: blender-bot 2023-04-14 09:18:04 +02:00 
Referenced by issue #101931, Cycles Apple Silicon optimisations
1 changed files with 49 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

61
intern/cycles/device/metal/kernel.mm
View File

@ -45,6 +45,36 @@ bool kernel_has_intersection(DeviceKernel device_kernel)
struct ShaderCache {
ShaderCache(id<MTLDevice> _mtlDevice) : mtlDevice(_mtlDevice)
{
/* Initialize occupancy tuning LUT. */
if (MetalInfo::get_device_vendor(mtlDevice) == METAL_GPU_APPLE) {
switch (MetalInfo::get_apple_gpu_architecture(mtlDevice)) {
default:
case APPLE_M2:
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_COMPACT_SHADOW_STATES] = {32, 32};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_INIT_FROM_CAMERA] = {832, 32};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_INTERSECT_CLOSEST] = {64, 64};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW] = {64, 64};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_INTERSECT_SUBSURFACE] = {704, 32};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_QUEUED_PATHS_ARRAY] = {1024, 256};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_SHADE_BACKGROUND] = {64, 32};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_SHADE_SHADOW] = {256, 256};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE] = {448, 384};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_SORTED_PATHS_ARRAY] = {1024, 1024};
break;
case APPLE_M1:
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_COMPACT_SHADOW_STATES] = {256, 128};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_INIT_FROM_CAMERA] = {768, 32};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_INTERSECT_CLOSEST] = {512, 128};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW] = {384, 128};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_INTERSECT_SUBSURFACE] = {512, 64};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_QUEUED_PATHS_ARRAY] = {512, 256};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_SHADE_BACKGROUND] = {512, 128};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_SHADE_SHADOW] = {384, 32};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE] = {576, 384};
occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_SORTED_PATHS_ARRAY] = {832, 832};
break;
}
}
}
~ShaderCache();
@ -73,6 +103,11 @@ struct ShaderCache {
std::function<void(MetalKernelPipeline *)> completionHandler;
};
struct OccupancyTuningParameters {
int threads_per_threadgroup = 0;
int num_threads_per_block = 0;
} occupancy_tuning[DEVICE_KERNEL_NUM];
std::mutex cache_mutex;
PipelineCollection pipelines[DEVICE_KERNEL_NUM];
@ -230,6 +265,13 @@ void ShaderCache::load_kernel(DeviceKernel device_kernel,
request.pipeline->device_kernel = device_kernel;
request.pipeline->threads_per_threadgroup = device->max_threads_per_threadgroup;
if (occupancy_tuning[device_kernel].threads_per_threadgroup) {
request.pipeline->threads_per_threadgroup =
occupancy_tuning[device_kernel].threads_per_threadgroup;
request.pipeline->num_threads_per_block =
occupancy_tuning[device_kernel].num_threads_per_block;
}
/* metalrt options */
request.pipeline->use_metalrt = device->use_metalrt;
request.pipeline->metalrt_hair = device->use_metalrt &&
@ -374,13 +416,6 @@ void MetalKernelPipeline::compile()
const std::string function_name = std::string("cycles_metal_") +
device_kernel_as_string(device_kernel);
int threads_per_threadgroup = this->threads_per_threadgroup;
if (device_kernel > DEVICE_KERNEL_INTEGRATOR_MEGAKERNEL &&
device_kernel < DEVICE_KERNEL_INTEGRATOR_RESET) {
/* Always use 512 for the sorting kernels */
threads_per_threadgroup = 512;
}
NSString *entryPoint = [@(function_name.c_str()) copy];
NSError *error = NULL;
@ -644,12 +679,14 @@ void MetalKernelPipeline::compile()
return;
}
int num_threads_per_block = round_down(computePipelineState.maxTotalThreadsPerThreadgroup,
computePipelineState.threadExecutionWidth);
num_threads_per_block = std::max(num_threads_per_block,
(int)computePipelineState.threadExecutionWidth);
if (!num_threads_per_block) {
num_threads_per_block = round_down(computePipelineState.maxTotalThreadsPerThreadgroup,
computePipelineState.threadExecutionWidth);
num_threads_per_block = std::max(num_threads_per_block,
(int)computePipelineState.threadExecutionWidth);
}
this->pipeline = computePipelineState;
this->num_threads_per_block = num_threads_per_block;
if (@available(macOS 11.0, *)) {
if (creating_new_archive || recreate_archive) {