Use right type in shmem calculation for kernels using reduce_output

src/cunumeric/index/advanced_indexing.cu

@@ -94,7 +94,7 @@ struct AdvancedIndexingImplBody<VariantKind::GPU, CODE, DIM, OUT_TYPE> {
 
     const size_t blocks = (volume + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
 
-    size_t shmem_size = THREADS_PER_BLOCK / 32 * sizeof(int64_t);
+    size_t shmem_size = THREADS_PER_BLOCK / 32 * sizeof(uint64_t);
 
     if (blocks >= MAX_REDUCTION_CTAS) {
       const size_t iters = (blocks + MAX_REDUCTION_CTAS - 1) / MAX_REDUCTION_CTAS;

src/cunumeric/index/repeat.cu

@@ -139,7 +139,7 @@ struct RepeatImplBody<VariantKind::GPU, CODE, DIM> {
 
     DeviceScalarReductionBuffer<SumReduction<uint64_t>> sum(stream);
     const size_t blocks_count = (extent + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
-    const size_t shmem_size   = THREADS_PER_BLOCK / 32 * sizeof(int64_t);
+    const size_t shmem_size   = THREADS_PER_BLOCK / 32 * sizeof(uint64_t);
 
     if (blocks_count > MAX_REDUCTION_CTAS) {
       const size_t iters = (blocks_count + MAX_REDUCTION_CTAS - 1) / MAX_REDUCTION_CTAS;

src/cunumeric/search/nonzero.cuh

@@ -63,7 +63,7 @@ int64_t compute_offsets(const AccessorRO<VAL, DIM>& in,
   DeviceScalarReductionBuffer<SumReduction<uint64_t>> size(stream);
 
   const size_t blocks = (volume + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
-  size_t shmem_size   = THREADS_PER_BLOCK / 32 * sizeof(int64_t);
+  size_t shmem_size   = THREADS_PER_BLOCK / 32 * sizeof(uint64_t);
 
   if (blocks >= MAX_REDUCTION_CTAS) {
     const size_t iters = (blocks + MAX_REDUCTION_CTAS - 1) / MAX_REDUCTION_CTAS;