Search results

I just completed a 2GB capture (3+ minutes of footage) at 82Mb/s (10MB/s). This is using a prototype of the rate control I'm working on. Here is my new console output. Every two frames the Qp is lowered by one. If the bitrate is stressing buffers, the Qp is increased by 2 or 4, depending on...

@ThePhage Thank you for the nice feedback. I have never done anything like this before, any firmware reverse engineering. I've never worked with MIPS code, or Novatek SoC. My experience is as a C developer in camera development and video compression, so normal "forward" engineering. So I do...

I'm using a SanDisc Extreme Pro, rated to 100MB/s, the scanning is only using ~0.5MB/s. It is an internal limitation to the way the buffering is done (I'm guessing.) This is old dashcam tech, not designed for quality. Reels was originally putting out only 8Mb/s (1MB/s at playback speed, encode...

Here is another post to document all my firmware changes to date. Address map of changes to Type-C Reels firmware (should help in supporting other types) 0x13a2f0 - I'm hooking a function call to 0x13b4b0 that set up the input buffer address (either the read or write location) Hook code is...

Attached in V5.0 firmware. Locked quantization: QP25 Locked white balance: RGB weights 1.75, 1.0, 1.0 Pipeline resolution: 1600x1200 (to use all of this you need to replace the lens) Encode resolution: 1600x1200 I moving the seconds input buffer any to just after the encoding working space...

I went through a lot of lenses. It has to be a macro lens, and they are rare (or at least poorly described.) Also with any lens you will need a spacer to move the lens further away from the sensor to focus. Optically a 12mm lens would be 12mm from the sensor to focus at infinity, here we need to...

3D printed the lens offset. Attached is more model for the 3D print. Now my frames in focus and not skewed. There is even space between the lens, and the open film loading door. V4.2 (Type C) still occasional saw a buffer collision between the compression data, and the second input frame...

V4.1 (Type C) I was incorrectly running QP23, which failed on some scenes, so this is running QP28.

Now with 12mm macro. Better focus, and much lower chromatic aberration. Yes is it askew. I need to 3D print a new camera mount, as the 12mm needs to be 5mm higher.

Resolution hack update V4. Type C firmware attached I wasn't able to limit the size of the H264 compression buffer, but by dropping the processing resolution to 1600x1200, on was able to fit one of the two buffers lower in memory, and now have no conflicts. The bit-rate control was struggling...

Resolution hacking updates. I was finally able to move the input buffers so that will not collide with each other. This took ages to work out, as there are multiple areas that the read and write these buffer addresses, some seem to be within interrupts. So I can now capture 1920x1440, process...

Looking for the AE is hard, and there doesn't seem to be any active error messages in that code. It might be near AWB, but I haven't looked hard for it yet.

Slow slow progress. I believe the function FUN_002f5094() set up the addresses for the two capture buffers, as it has these telling error logs: ERR:%s() Input buffer0 Y channel address - 0x%08x ERR:%s() Input buffer1 Y channel address - 0x%08x This function is called at the video frame rate...

128MB has been confirmed, as the memory repeats at the 128MB boundaries. 1Gb = 128MB

Today's progress. I did a series of captures, erasing the memory between captures so I could more easily see how the memory was being used (from the shell: mem fill 82000000 04000000 ff). Each capture use a different source resolution, all encoding to 1920x1440. The encoding resolution is not...

Thank you. "Built-in 1Gb DDR3 SDRAM" , now triply confirmed. This feature sheet is good to see.

Something doesn't quite add up, which where the video corruption could be coming from. Here we see Luma data straddling 0x06000000. Yet on the console there is this info: H264 is initializing something at 0x060A55C0, which in the middle not the top a the frame buffer. This address doesn't...

I also found the NT96655, I hoped to find what the differences are to the 96658, as the 55 doesn't have integrated RAM. I think I'm over my confusion able the tiny RAM size. There seem to be plenty of unused memory. Here is an crudely annotated memory dump during 1920x1440 capture (1920x1440...

Thank you @Mac84 and @wnt2fly. I going to proceed assuming the memory integrated and only 128MB. Crazy how there are no leaked datasheets for this obsolete part.

I'm diving into where the frames are stored in memory, and have run into a new question. When the system boots it reports: I always read that 1024 as 1GByte. Yet that Mb <- seems to be Mbits. Only 128MBytes, that is tiny. Does anyone have the parts list, or images on the underside on the...