nvenc SDK v8

[totaam]

Issue migrated from trac ticket # 1552

component: encodings | priority: major | resolution: fixed

2017-06-18 11:14:56: antoine created the issue

---

Version 8 is available for download.

The existing "nvenc7" codec compiles against it just fine.
TODO:

• rename the codec to just "nvenc" and use versioned checks with pkg-config.
• check for new useful API extensions we can use
• try to fix nvenc v7 support #1260#comment:16 (especially now that the problem also occurs on other cards: NVENC7 active but not used? #1550#comment:4)
• update the buildbot

[totaam]

2017-06-18 11:47:28: antoine changed status from new to assigned

[totaam]

2017-06-18 11:47:28: antoine commented

---

Interesting things found in the SDK 8 headers:

• encode from an opengl texture (NV_ENC_INPUT_RESOURCE_TYPE_OPENGL_TEX / NV_ENC_DEVICE_TYPE_OPENGL): could save us a lot of processing if we can stay on the GPU (don't copy pixmap data to ram: avoid the round-trips and stay on the GPU if we can #365)
• NV_ENC_MEMORY_HEAP should not be used any more (maybe the alternative will fix the lockups?)
• better "Long Term Reference" frames: could be used to send the whole window (including widgets, etc) encoded with lossless quality?
• sps / pps changes: meh
• "Motion Estimation" structures: meh
• NV_ENC_PIC_PARAMS_HEVC can now be used

So not a lot there.
r16083 renames the codec to "nvenc" and uses versionned pkgconfig checks.

[totaam]

2017-06-18 14:51:52: antoine changed status from assigned to closed

[totaam]

2017-06-18 14:51:52: antoine set resolution to fixed

[totaam]

2017-06-18 14:51:52: antoine commented

---

• r16084 API additions (commented out)

This will do for now, the API fix will be tracked in #1550.

[totaam]

2017-07-21 14:23:29: antoine commented

---

r16444 fixes RGB upload mode, so r16446 enables it by default again.

[totaam]

2017-07-22 14:10:52: antoine commented

---

• with NVFBC (nvidia capture sdk support #1317), the pixel buffers were bigger than the pixel contents, so r16456 sliced them to the right size for the device upload - but r16457 makes this redundant: the right place to fix this was in the CUDA image wrapper since the buffers are used elsewhere and we can't just go around and fix them all - this change is kept in place because the slicing should be (almost) free with numpy arrays
• r16455: disable RGB mode on win32, no idea why it doesn't work "an API argument is invalid" error...