GPU Metal: size the automatic grids by the number of GPU cores

The multiprocessor count, of which the automatically sized grids launch
minBlocks blocks each, was never set and stayed 1, so those kernels ran
one or a few threadgroups on a 32-core M1 Max. Metal does not report the
core count; read gpu-core-count from the GPU's IORegistry entry, as
CUDA uses multiProcessorCount, and link IOKit for that.

Author: ktf

GPU/GPUTracking/Base/metal/CMakeLists.txt

@@ -46,7 +46,7 @@ if(ALIGPU_BUILD_TYPE STREQUAL "O2")
                  PUBLIC_LINK_LIBRARIES O2::GPUTracking
                  TARGETVARNAME targetName)
 
-  target_link_libraries(${targetName} PUBLIC ${METAL_FRAMEWORKS})
+  target_link_libraries(${targetName} PUBLIC ${METAL_FRAMEWORKS} "-framework IOKit")
 
   target_compile_definitions(${targetName} PRIVATE $<TARGET_PROPERTY:O2::GPUTracking,COMPILE_DEFINITIONS>)
   # the compile_defitions are not propagated automatically on purpose (they are

GPU/GPUTracking/Base/metal/GPUReconstructionMetal.mm

@@ -18,6 +18,7 @@
 #include <cstddef>
 #include <map>
 
+#include <IOKit/IOKitLib.h>
 #include <mach-o/dyld.h>
 #include <mach-o/getsect.h>
 #include <mach-o/ldsyms.h>  // _mh_execute_header
@@ -28,6 +29,28 @@
     return (1);             \
   }
 
+namespace
+{
+// Metal does not report the number of GPU cores, but the IORegistry entry of
+// an Apple silicon GPU does. 0 if it is not there.
+uint32_t metalGPUCoreCount()
+{
+  uint32_t n = 0;
+  io_service_t service = IOServiceGetMatchingService(kIOMainPortDefault, IOServiceMatching("AGXAccelerator"));
+  if (service) {
+    CFTypeRef p = IORegistryEntrySearchCFProperty(service, kIOServicePlane, CFSTR("gpu-core-count"), kCFAllocatorDefault, kIORegistryIterateRecursively);
+    if (p) {
+      if (CFGetTypeID(p) == CFNumberGetTypeID()) {
+        CFNumberGetValue((CFNumberRef)p, kCFNumberSInt32Type, &n);
+      }
+      CFRelease(p);
+    }
+    IOObjectRelease(service);
+  }
+  return n;
+}
+}  // namespace
+
 #include "utils/qGetLdBinarySymbols.h"
 QGET_LD_BINARY_SYMBOLS(GPUReconstructionMetalCode_src);
 
@@ -74,6 +97,14 @@
     mDeviceName = device_name.c_str();
     // Basically a random number for now.
     mMaxBackendThreads = 1000;
+    // The automatic grids are minBlocks per core; with the default of 1 they
+    // left all but one core idle.
+    mMultiprocessorCount = metalGPUCoreCount();
+    if (mMultiprocessorCount == 0) {
+      GPUWarning("Could not read the number of GPU cores, assuming 1");
+      mMultiprocessorCount = 1;
+    }
+    mMaxBackendThreads = std::max<int32_t>(mMaxBackendThreads, deviceMaxWorkGroup.width * mMultiprocessorCount);
 
     if (GetMetalPrograms()) {
       return 1;