Replies: 4 comments
-
图形API生成参考源，模拟我最常看的视频 ./ffmpeg -f lavfi -i testsrc2=s=1920x1080:r=24:d=120 -pix_fmt yuv420p10 -c:v hevc -b:v 6000k output_test.mp4基于 mpv-BenchMark 创建配置。 展开查看 pause = yes
osd-msg1 = "FPS: ${estimated-display-fps}"
vulkan-swap-mode = immediate
d3d11-sync-interval = 0
d3d11-flip = no
framedrop = no
hidpi-window-scale = no
video-unscaled = yes
sub = no
audio = no
keep-open = always
video-sync = display-desync
display-fps-override = 24
vo = gpu # <gpu|gpu-next>
hwdec = no
gpu-context = d3d11 # <d3d11|winvk>
dither-depth = 10
glsl-shaders = CfL_Prediction.glsl;Anime4K_Restore_CNN_L.glsl主观评价： 展开查看
|
Beta Was this translation helpful? Give feedback.
-
视频解码将切换到libplacebo，因此仅测试vo_gpu-next支持的各种解码模式 使用与上层相同的参考源。 基于 mpv-BenchMark 创建配置。 展开查看 pause = yes
osd-msg1 = "FPS: ${estimated-display-fps}"
vulkan-swap-mode = immediate
d3d11-sync-interval = 0
d3d11-flip = no
framedrop = no
hidpi-window-scale = no
video-unscaled = yes
sub = no
audio = no
keep-open = always
video-sync = display-desync
display-fps-override = 24
vo = gpu-next
hwdec = no # <no|d3d11va|d3d11va-copy|dxva2-copy|nvdec|nvdec-copy|vulkan|vulkan-copy>
gpu-context = d3d11 # <d3d11|winvk>略过部分解码方式（dxva2/qsv高概率出现色彩异常，暂无i卡跳过qsv-copy，cuda/cuda-copy几乎可被nvdec/nvdec-copy取代。） 主观评价： 展开查看
|
Beta Was this translation helpful? Give feedback.
-
着色器基于 “视频解码” 部分的结论，仅测试 --vo=gpu 下 --gpu-context=d3d11的成绩。 使用与上层相同的参考源。 基于 mpv-BenchMark 创建配置。 展开查看 pause = yes
osd-msg1 = "FPS: ${estimated-display-fps}"
opengl-swapinterval = 0
opengl-dwmflush = no
vulkan-swap-mode = immediate
d3d11-sync-interval = 0
d3d11-flip = no
framedrop = no
hidpi-window-scale = no
fullscreen = yes # 1080P → 1440P（本机）
sub = no
audio = no
keep-open = always
video-sync = display-desync
vo = gpu
hwdec = no
gpu-context = d3d11
#scale = ewa_lanczos # <bicubic|ewa_lanczos> 常规放大对比组
glsl-shaders = AMD_FSR.glsl # 着色器 仅测试主观上认定较“重”的放大类着色器。同类变体之间不存在明显效率差异（例如FSRCNNX_x2_8_0_4_1_LineArt） 结果： 展开查看
|
Beta Was this translation helpful? Give feedback.
-
1. RIFE 补帧模拟效率测试 将下方脚本保存为 ./vspipe -p BM_rife.vpy .跑分脚本内容 —— 展开查看import vapoursynth as vs
from vapoursynth import core
import k7sfunc as k7f
D4K = False
D10B = True
Model = 40
GPU = 0
clip = core.std.BlankClip(width=3840 if D4K else 1920, height=2160 if D4K else 1080, length=2000, format=vs.YUV420P10 if D10B else vs.YUV420P8, color=[128]*3)
clip = k7f.RIFE_NV(input=clip, lt_d2k=True, model=Model, sc_mode=0, fps_num=2, t_tta=False, ext_proc=True, gpu=GPU, gpu_t=2, st_eng=False, ws_size=0)
clip = core.std.Expr(clip, "x 1 +")
clip = core.std.Expr(clip, "x 1 -")
clip.set_output()跑分结果 展开查看
2. MVtools 补帧模拟效率测试 展开查看import vapoursynth as vs
from vapoursynth import core
import k7sfunc as k7f
D4K = True
D10B = True
Recal = True
Block = True
clip = core.std.BlankClip(width=3840 if D4K else 1920, height=2160 if D4K else 1080, length=2000, format=vs.YUV420P10 if D10B else vs.YUV420P8, color=[128]*3)
clip = k7f.MVT_LQ(input=clip, fps_in=24, fps_out=48, recal=Recal, block=Block)
clip = core.std.Expr(clip, "x 1 +")
clip = core.std.Expr(clip, "x 1 -")
clip.set_output()结果 —— 展开查看
3. ESRGAN 超分模拟效率测试 展开查看import vapoursynth as vs
from vapoursynth import core
import k7sfunc as k7f
Dfhd = True
D10B = True
Dhqmod = False
GPU = 0
clip = core.std.BlankClip(width=1920 if Dfhd else 1280, height=1080 if Dfhd else 720, length=2000, format=vs.YUV420P10 if D10B else vs.YUV420P8, color=[128]*3)
clip = k7f.UAI_NV_TRT(input=clip, model_pth="RealESRGANv2/animejanaiV2L3.onnx" if Dhqmod else "RealESRGANv2/animejanaiV2L1.onnx", fp16=True, gpu=GPU, res_opt=[1920, 1080] if Dfhd else [1280, 720], res_max=[1920, 1080] if Dfhd else [1280, 720])
clip = core.std.Expr(clip, "x 1 +")
clip = core.std.Expr(clip, "x 1 -")
clip.set_output()分别对 1080p10bit 720p8bit 目前两种能实用的条件进行测试，启用 Dhqmod 对应的模型是janaiV3，禁用对应janaiV1 展开查看
|
Beta Was this translation helpful? Give feedback.
-
测试设备
Windows 10 pro 22H2
7950X
DDR5 5600MHz 2x32GiB
RTX4080
楼层直达：
Beta Was this translation helpful? Give feedback.
All reactions