#5 Updated the automatic choosing of whether to use the in-memory or …

…out-of-core diffusion implementation.
jamesavery · Sep 11, 2024 · 5d5741a · 5d5741a
1 parent a7d8e21
commit 5d5741a
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/src/pybind/diffusion-pybind.cc b/src/pybind/diffusion-pybind.cc
@@ -17,9 +17,9 @@ namespace python_api {
 
 #ifdef _OPENACC
         const int64_t total_size = N.z * N.y * N.x;
-        //if (total_size * (sizeof(uint8_t) + (2 * sizeof(float)) + sizeof(uint16_t)) > (1 * 1e9)) { // TODO make automatic
+        if (total_size * (sizeof(uint8_t) + (2 * sizeof(float)) + sizeof(uint16_t)) > (9 * 1e9)) { // TODO make automatic - set to ~90% of the 3080's 10 GB
         if (true) {
-            const shape_t global_shape = {64, 64, 64};
+            const shape_t global_shape = { 128, 128, 128 }; // TODO balancing act. Larger results in less wasted compute, while smaller results in more concurrency.
             NS::diffusion_out_of_core(voxels, N, global_shape, kernel, kernel_size, repititions, output);
         } else {
 #endif