Feature 对AMD的类CUDA支持?

[GuihongWang]

ROCm for windows

ROCm for linux

ZLUDA

---

Linux的ROCm支持更多显卡 但是还是兼容不了我这张假580（470）

所以我用了第三个ZLUDA方案 但是处理过程中出错了

C:\kits>C:\kits\zluda.exe -- cmd
Microsoft Windows [版本 10.0.19045.3570]
(c) Microsoft Corporation。保留所有权利。

C:\kits>E:

E:\>cd E:\workspace\MMD-OCR-v0.5.2

E:\workspace\MMD-OCR-v0.5.2>E:\workspace\MMD-OCR-v0.5.2\mmdocr.exe E:\workspace\MMD-OCR-v0.5.2\of.mp4
Device Number: 0
  Device name: AMD Radeon RX 580 2048SP [ZLUDA]
  Device VRAM (MB): 8388608
  Compute capability: 8.8
  Core Clock Rate (MHz): 1284
  Memory Clock Rate (MHz): 1750
  Memory Bus Width (bits): 0
  Peak Memory Bandwidth (GB/s): 0.000000

Reading video file E:\workspace\MMD-OCR-v0.5.2\of.mp4
Saving srt file to E:\workspace\MMD-OCR-v0.5.2\of.mp4.srt
Alphabet read, total 10444 chars
Loading models... done
Video width: 1920 height: 1080 frame count: 345
Video loaded
total time:  0.032813s
Exception: OpenCV(4.5.3-dev) g:\opencv-master\opencv\modules\core\src\cuda\gpu_mat.cu:116: error: (-217:Gpu API call) operation not supported in function '<unnamed>::DefaultAllocator::allocate'

CUDA-Z输出

CUDA-Z Report ============= Version: 0.10.251 64 bit http://cuda-z.sf.net/ OS Version: Windows x86 6.2.9200 Driver Version: Unknown Driver Dll Version: 12.20 Runtime Dll Version: 6.50

Core Information

Name: AMD Radeon RX 580 2048SP [ZLUDA]
Compute Capability: 8.8
Clock Rate: 1284 MHz
PCI Location: 0:1:0
Multiprocessors: 32
Threads Per Multiproc.: 2048
Warp Size: 64
Regs Per Block: 0
Threads Per Block: 1024
Threads Dimensions: 1024 x 1024 x 1024
Grid Dimensions: 2147483647 x 65536 x 65536
Watchdog Enabled: No
Integrated GPU: No
Concurrent Kernels: Yes
Compute Mode: Default
Stream Priorities: Yes

Memory Information

Total Global: 8192 MiB
Bus Width: 0 bits
Clock Rate: 1750 MHz
Error Correction: No
L2 Cache Size: 64 KiB
Shared Per Block: 64 KiB
Pitch: 2048 MiB
Total Constant: 2048 MiB
Texture Alignment: 256 B
Texture 1D Size: 16384
Texture 2D Size: 16384 x 16384
Texture 3D Size: 2048 x 2048 x 2048
GPU Overlap: Yes
Map Host Memory: Yes
Unified Addressing: Yes
Async Engine: Yes, Unidirectional

Performance Information

Memory Copy
Host Pinned to Device: --
Host Pageable to Device: --
Device to Host Pinned: --
Device to Host Pageable: --
Device to Device: --
GPU Core Performance
Single-precision Float: --
Double-precision Float: --
64-bit Integer: --
32-bit Integer: --
24-bit Integer: --

Generated: Thu Feb 15 14:22:30 2024