Set thread work size to 4 for elementwise kernels similar to cuda

Author: pruthvistony

aten/src/ATen/native/cuda/DistributionTemplates.h

@@ -34,11 +34,7 @@ const uint32_t block_size_bound = 256;
 const uint32_t grid_size_bound = 4;
 // number of randoms given by distributions like curand_uniform4, curand_uniform2_double
 // used in calculating philox offset.
-#if defined(USE_ROCM)
-const uint32_t curand4_engine_calls = 8;
-#else
 const uint32_t curand4_engine_calls = 4;
-#endif
 
 // utility function that calculates proper philox_offset
 // for distributions utilizing TensorIterator. For distributions using

aten/src/ATen/native/cuda/thread_constants.h

@@ -18,9 +18,5 @@ constexpr uint32_t num_threads() {
 }
 #endif
 
-#if defined(USE_ROCM)
-constexpr int thread_work_size() { return 8; }
-#else
 constexpr int thread_work_size() { return 4; }
-#endif
 constexpr int block_work_size() { return thread_work_size() * num_threads(); }