Some faster/better kernels. Tensors with points are now kept in (b, c, ...) format as this is easier for pytorch

utils/cinclude/cuda_utils.h

@@ -3,10 +3,21 @@
 
 #include <cmath>
 
+#define TOTAL_THREADS 512
+
 inline int opt_n_threads(int work_size) {
     const int pow_2 = std::log(static_cast<double>(work_size)) / std::log(2.0);
 
-    return max(min(1 << pow_2, 512), 32);
+    return max(min(1 << pow_2, TOTAL_THREADS), 1);
+}
+
+inline dim3 opt_block_config(int x, int y) {
+    const int x_threads = opt_n_threads(x);
+    const int y_threads =
+        max(min(opt_n_threads(y), TOTAL_THREADS / x_threads), 1);
+    dim3 block_config(x_threads, y_threads, 1);
+
+    return block_config;
 }
 
 #endif