GPU: more Metal bringup changes

GPU/GPUTracking/Base/metal/GPUReconstructionMetal.mm

@@ -15,6 +15,7 @@
 #include "GPUDefParametersLoad.inc"
 #include "GPUReconstructionMetal.h"
 
+#include <chrono>
 #include <cstddef>
 #include <map>
 
@@ -376,13 +377,27 @@ MetalMemRef metalResolve(const GPUReconstructionMetalInternals* internals, const
 
 void GPUReconstructionMetal::DrainStream(int32_t stream)
 {
-  if (mInternals->streamHasWork[stream]) {
-    CommitStream(stream);
+  // Hand the GPU what the other streams have encoded too, so it has work
+  // while the host waits and then works on what it waited for.
+  for (int32_t i = 0; i < mNStreams; i++) {
+    if (mInternals->streamHasWork[i]) {
+      CommitStream(i);
+    }
   }
   // Waiting on the buffer that is currently open would block forever: it is not
   // committed, so it never completes. Only what was submitted can be waited for.
+  // The last submitted buffer ends by signalling the stream's order event, once
+  // all the work before it is done. Watching the event returns sooner than
+  // waitUntilCompleted, which sleeps until the completion has been processed.
+  // Only short waits are worth a busy core, so give up after a while.
+  bool done = false;
+  const auto t0 = std::chrono::steady_clock::now();
+  while (!mInternals->submittedBuffers[stream].empty() && !(done = mInternals->streamOrder[stream].signaledValue >= mInternals->streamOrderValue[stream]) && std::chrono::steady_clock::now() - t0 < std::chrono::milliseconds(20)) {
+  }
   for (id<MTLCommandBuffer> cb : mInternals->submittedBuffers[stream]) {
-    [cb waitUntilCompleted];
+    if (!done) {
+      [cb waitUntilCompleted];
+    }
     if (cb.status == MTLCommandBufferStatusError) {
       GPUError("Metal command buffer on stream %d failed: %s", stream, [[cb.error description] UTF8String]);
     }

GPU/GPUTracking/Definitions/Parameters/GPUParameters.csv

@@ -28,7 +28,7 @@ GPUTPCCFDecodeZSDenseLink,"""GPUCA_WARP_SIZE""",,"[""GPUCA_WARP_SIZE"", 4]","[""
 GPUTPCCFGather,"[1024, 1]",,"[1024, 5]","[1024, 1]",,,,,,"[160, 11]","[1024, 1]",736,896,,"[928, 1]",,"[1024, 1]","[320, 2]",
 COMPRESSION_GATHER,1024,,1024,1024,,,,,,1024,1024,,1024,,,,,,
 GPUTPCGMMergerTrackFit,256,,"[192, 2]","[64, 7]",,,,,,"[32, 16]","[32, 8]","[32, 14]","[160, 2]",,"[32, 24]",,"[64, 8]","[64, 6]","[64, 4]"
-GPUTPCGMMergerFollowLoopers,256,,"[256, 5]","[256, 4]",,,,,,"[256, 4]","[128, 4]","[1024, 1]",640,,"[128, 16]",,"[224, 3]","[256, 7]",
+GPUTPCGMMergerFollowLoopers,256,,"[256, 5]","[256, 4]",,,,,,"[256, 4]","[128, 4]","[1024, 1]",640,,"[128, 16]",,"[224, 3]","[256, 7]","[1, 16]"
 GPUTPCGMMergerSectorRefit,256,,"[64, 4]","[256, 2]",,,,,,"[32, 8]","[64, 5]","[32, 7]","[32, 7]",,"[32, 20]",,"[32, 10]","[64, 4]",
 GPUTPCGMMergerUnpackResetIds,256,,256,256,,,,,,256,256,256,256,,256,,256,256,
 GPUTPCGMMergerUnpackGlobal,256,,256,256,,,,,,256,256,256,256,,256,,256,256,