Cycles: Improve cache usage on Apple GPUs by chunking active indices

This patch partitions the active indices into chunks prior to sorting by material in order to tradeoff some material coherence for better locality. On Apple Silicon GPUs (particularly higher end M1-family GPUs), we observe overall render time speedups of up to 15%. The partitioning is implemented by repeating the range of `shader_sort_key` for each partition, and encoding a "locator" key which distributes the indices into sorted chunks.

Reviewed By: brecht

Differential Revision: https://developer.blender.org/D15331

1 files changed, 18 insertions, 0 deletions

diff --git a/intern/cycles/device/metal/util.mm b/intern/cycles/device/metal/util.mm
index a7a5b596b8f..c336dc310c8 100644
--- a/intern/cycles/device/metal/util.mm
+++ b/intern/cycles/device/metal/util.mm
@@ -72,6 +72,24 @@ MetalGPUVendor MetalInfo::get_device_vendor(id<MTLDevice> device)
   return METAL_GPU_UNKNOWN;
 }
 
+int MetalInfo::optimal_sort_partition_elements(id<MTLDevice> device)
+{
+  if (auto str = getenv("CYCLES_METAL_SORT_PARTITION_ELEMENTS")) {
+    return atoi(str);
+  }
+
+  /* On M1 and M2 GPUs, we see better cache utilization if we partition the active indices before
+   * sorting each partition by material. Partitioning into chunks of 65536 elements results in an
+   * overall render time speedup of up to 15%. */
+  if (get_device_vendor(device) == METAL_GPU_APPLE) {
+    AppleGPUArchitecture arch = get_apple_gpu_architecture(device);
+    if (arch == APPLE_M1 || arch == APPLE_M2) {
+      return 65536;
+    }
+  }
+  return 0;
+}
+
 vector<id<MTLDevice>> const &MetalInfo::get_usable_devices()
 {
   static vector<id<MTLDevice>> usable_devices;