Something impacting the logic for blockDimensions.C == 48 on iOS 26.1

Author: liuliu

lib/nnc/mfa/v2/NAAttentionDescriptor.cpp

@@ -52,20 +52,7 @@ NAAttentionKernelDescriptor NAAttentionDescriptor::kernelDescriptor(MTL::Device
     } else if (matrixDimensions[1] % 48 == 0) {
       return simd::ushort3 { 16, 48, revisedHead };
     }
-    // Prefer no trailing involved, so the compute is more evenly distributed.
-    if (matrixDimensions[1] % 128 > 64 && matrixDimensions[1] % 96 < 48) {
-      return simd::ushort3 { 16, 64, revisedHead };
-    } else if (matrixDimensions[1] % 128 < 64 && matrixDimensions[1] % 96 > 48) {
-      return simd::ushort3 { 16, 48, revisedHead };
-    }
-    // If we have to use matrix multiplication, calculate how much wasted compute we are going to be with.
-    const unsigned short remainder64 = matrixDimensions[1] % 64;
-    const unsigned short remainder48 = matrixDimensions[1] % 48;
-    if (remainder64 * 48 < remainder48 * 64) {
-      return simd::ushort3 { 16, 48, revisedHead };
-    } else {
-      return simd::ushort3 { 16, 64, revisedHead };
-    }
+    return simd::ushort3 { 16, 64, revisedHead };
   };
   auto createExecutionSIMDGroups = 
   [=]() -> uint16_t {