[wakamex]

On master, calculate_motion_score_kernel_16bpc (libvmaf/src/feature/cuda/integer_motion/motion_score.cu:88) sets the rounding offset with

unsigned add_before_shift_y = pow(2, (src.bpc - 1));

in every thread, so every pixel of a 10-bit frame pays for a double-precision pow(). Consumer GPUs run FP64 slowly (1/64 of the FP32 rate on the RTX 3090's GA102), and in the 10-bit run below the motion kernel is the most expensive kernel, 53% of all GPU kernel time. The 8-bit kernel doesn't have this.

This PR replaces it with 1u << (src.bpc - 1), which is the same value for every supported bit depth.

Kernel times from Nsight Systems (nsys profile -t cuda, then the cuda_gpu_kern_sum report), RTX 3090, driver 615.71.09, libvmaf built with CUDA 13.1. The input is 96 frames of 1080p 10-bit 4:2:0: FFmpeg's testsrc2 as the reference and a noised and blurred copy as the distorted clip, scored with vmaf -b 10 -m path=model/vmaf_v0.6.1.json --gpumask 0. Lower is better.

GPU kernel time	master 229f247	this branch
calculate_motion_score_kernel_16bpc, mean per frame	0.689 ms	0.089 ms
all kernels, 96 frames	125.0 ms	66.1 ms

Every per-frame feature and VMAF score in the CSV output is the same as master on this clip and on an 8-bit version of it.

This is the only pow() in the CUDA device code. The other pow() calls under libvmaf/src/feature/cuda/ are in integer_adm_cuda.c and run on the host.

Validation

libvmaf built with -Denable_cuda=true -Denable_float=true -Dbuilt_in_models=true (CUDA 13.1, GCC 13), master 229f247 against this branch. The unit tests give the same result on both: 30 of 31 pass, and test_cuda_pic_preallocation segfaults in test_cuda_picture_preallocation_method_host_pinned on master too.