diff --git a/makefile b/makefile
index af91e67f8..a7757f9e1 100644
--- a/makefile
+++ b/makefile
@@ -172,8 +172,9 @@ update-examples/%: examples/%.dx build
 	$(dex) script --allow-errors $< > $<.tmp
 	mv $<.tmp $<
 
-run-gpu-tests: export DEX_ALLOC_CONTRACTIONS=0
-run-gpu-tests: tests/gpu-tests.dx build
+gpu-tests: run-gpu-tests/gpu-tests
+
+run-gpu-tests/%: tests/%.dx build
 	misc/check-quine $< $(dex) --backend llvm-cuda script --allow-errors
 
 update-gpu-tests: export DEX_ALLOW_CONTRACTIONS=0
diff --git a/src/lib/dexrt.cpp b/src/lib/dexrt.cpp
index e12d21b76..02ce53c1d 100644
--- a/src/lib/dexrt.cpp
+++ b/src/lib/dexrt.cpp
@@ -234,17 +234,23 @@ void dex_cuMemcpyHtoD(int64_t bytes, char* device_ptr, char* host_ptr) {
   CHECK(cuMemcpyHtoD, reinterpret_cast<CUdeviceptr>(device_ptr), host_ptr, bytes);
 }
 
-void dex_queryParallelismCUDA(const char* kernel_func, int64_t iters,
+void dex_queryParallelismCUDA(char* kernel_func, int64_t iters,
                               int32_t* numWorkgroups, int32_t* workgroupSize) {
   if (iters == 0) {
     *numWorkgroups = 0;
     *workgroupSize = 0;
     return;
   }
-  // TODO: Use the occupancy calculator, or at least use a fixed number of blocks?
-  const int64_t fixedWgSize = 1024;
-  *workgroupSize = fixedWgSize;
-  *numWorkgroups = std::min((iters + fixedWgSize - 1) / fixedWgSize, fixedWgSize);
+  int min_grid_size_for_max_occupancy;
+       int block_size_32;
+  CUfunction kernel = reinterpret_cast<CUfunction>(kernel_func);
+  CHECK(cuOccupancyMaxPotentialBlockSize,
+        &min_grid_size_for_max_occupancy,
+        &block_size_32,
+        kernel, nullptr, 0, 0);
+  int64_t block_size = block_size_32;
+  *workgroupSize = block_size;
+  *numWorkgroups = (iters + block_size - 1) / block_size;
 }
 
 void dex_loadKernelCUDA(const char* kernel_text, char** module_storage, char** kernel_storage) {