[macOS GPU Support] Tune dispatching of persistent threads for Apple silicon GPUs (#3978)

* Use 768 instead of 384 threads in generic kernels

* Use 1536 instead of 1024 threads in force kernels.

platforms/opencl/src/OpenCLContext.cpp

@@ -218,6 +218,9 @@ OpenCLContext::OpenCLContext(const System& system, int platformIndex, int device
       
         if (vendor.size() >= 5 && vendor.substr(0, 5) == "Apple") {
             simdWidth = 32;
+
+            // 768 threads per GPU core.
+            numThreadBlocksPerComputeUnit = 12;
         }
         else if (vendor.size() >= 6 && vendor.substr(0, 6) == "NVIDIA") {
             compilationDefines["WARPS_ARE_ATOMIC"] = "";

platforms/opencl/src/OpenCLNonbondedUtilities.cpp

@@ -65,8 +65,14 @@ OpenCLNonbondedUtilities::OpenCLNonbondedUtilities(OpenCLContext& context) : con
         forceThreadBlockSize = 1;
     }
     else if (context.getSIMDWidth() == 32) {
-            numForceThreadBlocks = 4*context.getDevice().getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>();
-            forceThreadBlockSize = 256;
+        int blocksPerComputeUnit = 4;
+        std::string vendor = context.getDevice().getInfo<CL_DEVICE_VENDOR>();
+        if (vendor.size() >= 5 && vendor.substr(0, 5) == "Apple") {
+            // 1536 threads per GPU core.
+            blocksPerComputeUnit = 6;
+        }
+        numForceThreadBlocks = blocksPerComputeUnit*context.getDevice().getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>();
+        forceThreadBlockSize = 256;
     }
     else {
         numForceThreadBlocks = context.getNumThreadBlocks();