#5 Added original 1d gauss for comparison

src/lib/cpp/gpu/diffusion.cc

@@ -79,6 +79,42 @@ namespace gpu {
             }
         }
     }
+
+    void diffusion_core_og(const float *__restrict__ input, const float *__restrict__ kernel, float *__restrict__ output, const shape_t &N, const int32_t dim, const int32_t radius) {
+        // Note: the use of 32-bit is intentional to reduce register pressure on GPU. Each of the 32-bit values shouldn't exceed 2^32, but the indices (address) to the arrays can be.
+
+        const int32_t
+            strides[3] = {(N.y)*(N.x), N.x, 1},
+            stride = strides[dim],
+            Ns[3] = {N.z, N.y, N.x};
+
+        #pragma acc parallel loop collapse(3) present(input, kernel, output)
+        for (int32_t i = 0; i < N.z; i++) {
+            for (int32_t j = 0; j < N.y; j++) {
+                for (int32_t k = 0; k < N.x; k++) {
+                    const int32_t
+                        X[3] = {i, j, k},
+                        ranges[2] = {
+                            -std::min(radius, X[dim]), std::min(radius, Ns[dim]-X[dim]-1)
+                        };
+                    const int64_t output_index = (uint64_t)i*(uint64_t)strides[0] + (uint64_t)j*(uint64_t)strides[1] + (uint64_t)k*(uint64_t)strides[2];
+
+                    float sum = 0.0f;
+
+                    for (int32_t r = -radius; r <= radius; r++) {
+                        const int64_t input_index = output_index + (int64_t)r*(int64_t)stride;
+                        bool cond = r >= ranges[0] && r <= ranges[1];
+
+                        // Original:
+                        float val = cond ? input[input_index] : 0.0f;
+                        sum += val * kernel[radius+r];
+                    }
+
+                    output[output_index] = sum;
+                }
+            }
+        }
+    }
     void diffusion_core_x(const float *__restrict__ input, const float *__restrict__ kernel, float *__restrict__ output, const shape_t &N, const int32_t radius) {
         // Note: the use of 32-bit is intentional to reduce register pressure on GPU. Each of the 32-bit values shouldn't exceed 2^32, but the indices (address) to the arrays can be.
         constexpr int veclen = 128;