Differential D14923

Enable inlining on Apple Silicon. Use new process-wide ShaderCache in order to safely re-enable binary archives
ClosedPublic
Actions

Authored by Michael Jones (michael_jones) on May 11 2022, 3:12 PM.

Tags

None

Subscribers

Aled Brown (Gelert)

Tokens

"Love" token, awarded by DaveDeer."Love" token, awarded by skyscapeparadise.

Details

Reviewers

Brecht Van Lommel (brecht)

Commits

rB007184bcf212: Enable inlining on Apple Silicon. Use new process-wide ShaderCache in order to…

Summary

This patch is the same as D14763, but with a fix for unit test failures caused by ShaderCache fetch logic not working in the non-MetalRT case:

diff --git a/intern/cycles/device/metal/kernel.mm b/intern/cycles/device/metal/kernel.mm
index ad268ae7057..6aa1a56056e 100644
--- a/intern/cycles/device/metal/kernel.mm
+++ b/intern/cycles/device/metal/kernel.mm
@@ -203,9 +203,12 @@ bool kernel_has_intersection(DeviceKernel device_kernel)
 
   /* metalrt options */
   request.pipeline->use_metalrt = device->use_metalrt;
-  request.pipeline->metalrt_hair = device->kernel_features & KERNEL_FEATURE_HAIR;
-  request.pipeline->metalrt_hair_thick = device->kernel_features & KERNEL_FEATURE_HAIR_THICK;
-  request.pipeline->metalrt_pointcloud = device->kernel_features & KERNEL_FEATURE_POINTCLOUD;
+  request.pipeline->metalrt_hair = device->use_metalrt &&
+                                   (device->kernel_features & KERNEL_FEATURE_HAIR);
+  request.pipeline->metalrt_hair_thick = device->use_metalrt &&
+                                         (device->kernel_features & KERNEL_FEATURE_HAIR_THICK);
+  request.pipeline->metalrt_pointcloud = device->use_metalrt &&
+                                         (device->kernel_features & KERNEL_FEATURE_POINTCLOUD);
 
   {
     thread_scoped_lock lock(cache_mutex);
@@ -225,9 +228,9 @@ bool kernel_has_intersection(DeviceKernel device_kernel)
 
   /* metalrt options */
   bool use_metalrt = device->use_metalrt;
-  bool metalrt_hair = device->kernel_features & KERNEL_FEATURE_HAIR;
-  bool metalrt_hair_thick = device->kernel_features & KERNEL_FEATURE_HAIR_THICK;
-  bool metalrt_pointcloud = device->kernel_features & KERNEL_FEATURE_POINTCLOUD;
+  bool metalrt_hair = use_metalrt && (device->kernel_features & KERNEL_FEATURE_HAIR);
+  bool metalrt_hair_thick = use_metalrt && (device->kernel_features & KERNEL_FEATURE_HAIR_THICK);
+  bool metalrt_pointcloud = use_metalrt && (device->kernel_features & KERNEL_FEATURE_POINTCLOUD);
 
   MetalKernelPipeline *best_pipeline = nullptr;
   for (auto &pipeline : collection) {

Diff Detail

Repository

Branch

arcpatch-D14923 (branched from master)

Build Status

Buildable 22070
Build 22070: arc lint + arc unit

Event Timeline

Michael Jones (michael_jones) requested review of this revision.May 11 2022, 3:12 PM

Michael Jones (michael_jones) created this revision.

Harbormaster completed remote builds in B22068: Diff 51384.May 11 2022, 3:12 PM

Michael Jones (michael_jones) edited the summary of this revision. (Show Details)May 11 2022, 3:18 PM

Michael Jones (michael_jones) added a reviewer: Brecht Van Lommel (brecht).

Remove diagnostic printf

Harbormaster completed remote builds in B22069: Diff 51385.May 11 2022, 3:29 PM

Remove unnecessary functions & add PSO NULL-check during MetalDeviceQueue::enqueue

Harbormaster completed remote builds in B22070: Diff 51387.May 11 2022, 3:50 PM

Remove undefined method

Harbormaster completed remote builds in B22071: Diff 51388.May 11 2022, 3:53 PM

Metal tests are passing on the buildbot now.
https://builder.blender.org/admin/#/builders/19/builds/386

This can be merged into master.

This revision is now accepted and ready to land.May 11 2022, 3:54 PM

Brecht Van Lommel (brecht) mentioned this in D14763: Cycles: Enable inlining on Apple Silicon for 1.1x speedup.May 11 2022, 3:56 PM

Aled Brown (Gelert) added a subscriber: Aled Brown (Gelert).May 11 2022, 4:13 PM

Steve Hanff (Tiwaz) added a subscriber: Steve Hanff (Tiwaz).May 11 2022, 5:09 PM

Closed by commit rB007184bcf212: Enable inlining on Apple Silicon. Use new process-wide ShaderCache in order to… (authored by Michael Jones (michael_jones)). · Explain WhyMay 11 2022, 5:21 PM

This revision was automatically updated to reflect the committed changes.

Michael Jones (michael_jones) added a commit: rB007184bcf212: Enable inlining on Apple Silicon. Use new process-wide ShaderCache in order to….

Steve Hanff (Tiwaz) removed a subscriber: Steve Hanff (Tiwaz).May 11 2022, 5:26 PM

Kady Richardson (skyscapeparadise) awarded a token.May 11 2022, 10:33 PM

Davide Caprioli (DaveDeer) awarded a token.May 16 2022, 12:39 PM

Michael Jones (michael_jones) mentioned this in D14645: Cycles: Apple Silicon optimizations (~20% uplift on M1 Max).Jul 11 2022, 4:18 PM

Michael Jones (michael_jones) mentioned this in D15456: Cycles: refactor to move part of KernelData definition to template header.Jul 14 2022, 6:47 PM

Brecht Van Lommel (brecht) mentioned this in rBe11c899e715b: Cycles: disable Metal inlining optimization on Apple GPUs.Aug 18 2022, 8:17 PM

Revision Contents
Changeset List

Path

Size

intern/

cycles/

device/

metal/

8 lines

96 lines

108 lines

856 lines

31 lines

kernel/

device/

metal/

24 lines

Diff 51387

intern/cycles/device/metal/device_impl.h

Loading...

intern/cycles/device/metal/device_impl.mm

Loading...

intern/cycles/device/metal/kernel.h

Loading...

intern/cycles/device/metal/kernel.mm

Loading...

intern/cycles/device/metal/queue.mm

Loading...

intern/cycles/kernel/device/metal/compat.h

Loading...