compat/cuda: switch from powf to __powf intrinsic

The powf builtin causes crashes on older clang, so manually implement
the (faster) intrinsic.
The code it spawns is identical to that of nvcc.

1 files changed, 4 insertions, 4 deletions

diff --git a/libavfilter/vf_bilateral_cuda.cu b/libavfilter/vf_bilateral_cuda.cu
index 8aba3a079f..bbcfc81db5 100644
--- a/libavfilter/vf_bilateral_cuda.cu
+++ b/libavfilter/vf_bilateral_cuda.cu
@@ -34,9 +34,9 @@ extern "C"
 __device__ static inline float norm_squared(float4 first_yuv, float4 second_yuv)
 {
     float ans = 0;
-    ans += powf(first_yuv.x - second_yuv.x, 2);
-    ans += powf(first_yuv.y - second_yuv.y, 2);
-    ans += powf(first_yuv.z - second_yuv.z, 2);
+    ans += __powf(first_yuv.x - second_yuv.x, 2);
+    ans += __powf(first_yuv.y - second_yuv.y, 2);
+    ans += __powf(first_yuv.z - second_yuv.z, 2);
     return ans;
 }
 
@@ -52,7 +52,7 @@ __device__ static inline float calculate_w(int x, int y, int r, int c,
                                            float sigma_space, float sigma_color)
 {
     float first_term, second_term;
-    first_term = (powf(x - r, 2) + powf(y - c, 2)) / (2 * sigma_space * sigma_space);
+    first_term = (__powf(x - r, 2) + __powf(y - c, 2)) / (2 * sigma_space * sigma_space);
     second_term = norm_squared(pixel_value, neighbor_value) / (2 * sigma_color * sigma_color);
     return __expf(-first_term - second_term);
 }