Cycles Denoising: Speedup reconstruction by skipping near-zero weights

intern/cycles/kernel/filter/filter_nlm_cpu.h

@@ -101,7 +101,7 @@ ccl_device_inline void kernel_filter_nlm_calc_weight(const float *ccl_restrict d
 		for(int x = rect.x; x < rect.z; x++) {
 			const int low = max(rect.x, x-f);
 			const int high = min(rect.z, x+f+1);
-			out_image[y*w+x] = expf(-max(out_image[y*w+x] * (1.0f/(high - low)), 0.0f));
+			out_image[y*w+x] = fast_expf(-max(out_image[y*w+x] * (1.0f/(high - low)), 0.0f));
 		}
 	}
 }

intern/cycles/kernel/filter/filter_nlm_gpu.h

@@ -66,7 +66,7 @@ ccl_device_inline void kernel_filter_nlm_calc_weight(int x, int y,
 		sum += difference_image[y*w+x1];
 	}
 	sum *= 1.0f/(high-low);
-	out_image[y*w+x] = expf(-max(sum, 0.0f));
+	out_image[y*w+x] = fast_expf(-max(sum, 0.0f));
 }
 
 ccl_device_inline void kernel_filter_nlm_update_output(int x, int y,

intern/cycles/kernel/filter/filter_reconstruction.h

@@ -29,6 +29,10 @@ ccl_device_inline void kernel_filter_construct_gramian(int x, int y,
                                                        ccl_global float3 *XtWY,
                                                        int localIdx)
 {
+	if(weight < 1e-3f) {
+		return;
+	}
+
 	int p_offset =  y    *w +  x;
 	int q_offset = (y+dy)*w + (x+dx);