[TDA4M] MSC output is not NV12 but NV21

Dongwon Choi

Intellectual 845 points

Hi.

I used MSC for scaling such as below.

The result is not NV12 but NV21.

Can you check this issue?

I'm attaching coeff , scaling ratio, the source file and result file.

[scaling info]

[coeff]

void scale_set_coeff(tivx_vpac_msc_coefficients_t* coeff, uint32_t interpolation)
{
  uint32_t i;
  uint32_t idx;
  uint32_t weight;

  idx = 0;
  coeff->single_phase[0][idx ++] = 0;
  coeff->single_phase[0][idx ++] = 0;
  coeff->single_phase[0][idx ++] = 256;
  coeff->single_phase[0][idx ++] = 0;
  coeff->single_phase[0][idx ++] = 0;
  idx = 0;
  coeff->single_phase[1][idx ++] = 0;
  coeff->single_phase[1][idx ++] = 0;
  coeff->single_phase[1][idx ++] = 256;
  coeff->single_phase[1][idx ++] = 0;
  coeff->single_phase[1][idx ++] = 0;

  if (VX_INTERPOLATION_BILINEAR == interpolation) {
    idx = 0;
    for (i = 0; i < 32; i++) {
      weight = i << 2;
      coeff->multi_phase[0][idx ++] = 0;
      coeff->multi_phase[0][idx ++] = 0;
      coeff->multi_phase[0][idx ++] = 256 - weight;
      coeff->multi_phase[0][idx ++] = weight;
      coeff->multi_phase[0][idx ++] = 0;
    }
    idx = 0;
    for (i = 0; i < 32; i++) {
      weight = (i + 32) << 2;
      coeff->multi_phase[1][idx ++] = 0;
      coeff->multi_phase[1][idx ++] = 0;
      coeff->multi_phase[1][idx ++] = 256 - weight;
      coeff->multi_phase[1][idx ++] = weight;
      coeff->multi_phase[1][idx ++] = 0;
    }
    idx = 0;
    for (i = 0; i < 32; i++) {
      weight = i << 2;
      coeff->multi_phase[2][idx ++] = 0;
      coeff->multi_phase[2][idx ++] = 0;
      coeff->multi_phase[2][idx ++] = 256 - weight;
      coeff->multi_phase[2][idx ++] = weight;
      coeff->multi_phase[2][idx ++] = 0;
    }
    idx = 0;
    for (i = 0; i < 32; i++) {
      weight = (i + 32) << 2;
      coeff->multi_phase[3][idx ++] = 0;
      coeff->multi_phase[3][idx ++] = 0;
      coeff->multi_phase[3][idx ++] = 256 - weight;
      coeff->multi_phase[3][idx ++] = weight;
      coeff->multi_phase[3][idx ++] = 0;
    }
  } else { /* STR_VX_INTERPOLATION_NEAREST_NEIGHBOR */
    idx = 0;
    for (i = 0; i < 32; i++) {
      coeff->multi_phase[0][idx ++] = 0;
      coeff->multi_phase[0][idx ++] = 0;
      coeff->multi_phase[0][idx ++] = 256;
      coeff->multi_phase[0][idx ++] = 0;
      coeff->multi_phase[0][idx ++] = 0;
    }
    idx = 0;
    for (i = 0; i < 32; i++) {
      coeff->multi_phase[1][idx ++] = 0;
      coeff->multi_phase[1][idx ++] = 0;
      coeff->multi_phase[1][idx ++] = 0;
      coeff->multi_phase[1][idx ++] = 256;
      coeff->multi_phase[1][idx ++] = 0;
    }
    idx = 0;
    for (i = 0; i < 32; i++) {
      coeff->multi_phase[2][idx ++] = 0;
      coeff->multi_phase[2][idx ++] = 0;
      coeff->multi_phase[2][idx ++] = 256;
      coeff->multi_phase[2][idx ++] = 0;
      coeff->multi_phase[2][idx ++] = 0;
    }
    idx = 0;
    for (i = 0; i < 32; i++) {
      coeff->multi_phase[3][idx ++] = 0;
      coeff->multi_phase[3][idx ++] = 0;
      coeff->multi_phase[3][idx ++] = 0;
      coeff->multi_phase[3][idx ++] = 256;
      coeff->multi_phase[3][idx ++] = 0;
    }
  }
}

[input file (2048x1280)]

0000000001_2048x1280_input.zip

[output file (640x256)]

0000000001_640x256_lfr.zip

over 5 years ago

0 Brijesh Jadav over 5 years ago

TI__Guru**** 495135 points

Hi Choi,

Could you please tell me which scalar node are you using? Also which SDK release are you using?

Rgds,

Brijesh

0 Brijesh Jadav over 5 years ago in reply to Brijesh Jadav

TI__Guru**** 495135 points

Hi Choi,

Can you try commenting out call to tivxVpacMscScaleUpdateStartPhases aPI in the file psdk_rtos_auto_j7_07_00_00_11\tiovx\kernels_j7\hwa\vpac_msc\vx_vpac_msc_multi_scale_output_target.c, at around linke 490?

Rgds,

Brijesh

0 Brijesh Jadav over 5 years ago in reply to Brijesh Jadav

TI__Guru**** 495135 points

Hi Choi,

Did this change work fine? Could you please let us know?

Rgds,

Brijesh

0 Dongwon Choi over 5 years ago in reply to Brijesh Jadav

Intellectual 845 points

Hi Brijesh.

I'm sorry for late response.

I'm using psdk 6.2.

Is this the known issue of psdk6.2?

I commented tivxVpacMscScaleUpdateStartPhases() such as below, the result was same as before(Output is NV21)

tiovx/kernels_j7/hwa/vpac_msc/vx_vpac_msc_multi_scale_output_target.c

   487                 else
   488                 {
~  489 //                    tivxVpacMscScaleUpdateStartPhases(msc_obj, sc_cfg, cnt);
   490                 }

Thanks.

0 Brijesh Jadav over 5 years ago in reply to Dongwon Choi

TI__Guru**** 495135 points

Hi Choi,

This is an issue in SDK6.2 release. Can you try commenting out call to API tivxVpacMscScaleUpdateStartPhases in the entire file psdk_rtos_auto_j7_06_02_00_21\tiovx\kernels_j7\hwa\vpac_msc\vx_vpac_msc_multi_scale_output_target.c?

It seems chroma offset is going wrong somewhere.

Could you please confirm that you are using multi-scale node? This change is only for multi-scale node.

Rgds,

Brijesh

0 Dongwon Choi over 5 years ago in reply to Brijesh Jadav

Intellectual 845 points

Hi Brijesh

The output of MSC became NV12 when I commented all tivxVpacMscScaleUpdateStartPhases() in vx_vpac_msc_multi_scale_output_target.c

Is this issue fixed at psdk 7.0?

I'm using msc node such as below

scalerObj->node = tivxVpacMscScaleNode(graph, input, output_fsd, output_ld, output_od, output_lbf, output_tsr);

Regards

Dongwon Choi

0 Brijesh Jadav over 5 years ago in reply to Dongwon Choi

TI__Guru**** 495135 points

Hi Choi,

This is not yet fixed in SDK7.0. I will add it in next release.

Rgds,

Brijesh

0 Dongwon Choi over 5 years ago in reply to Brijesh Jadav

Intellectual 845 points

Hi Brijesh.

I have additional issue of MSC.

I need to use the below crop and scaling with MSC

However result was abnormal

Can you check this output at your side?

[ROI information]

crop.crop_width = 2048;

crop.crop_height = 953;

crop.crop_start_x = 0;

crop.crop_start_y = 7;

input file : 0000000001_2048x960_input.yuv

output file : 0000000001_1376x640_od.yuv

msc_out.zip

Regards

Dongwon Choi

0 Brijesh Jadav over 5 years ago in reply to Dongwon Choi

TI__Guru**** 495135 points

Hi Choi,

How are you saving the output file? Do you use pitch parameter of image buffer to save image?

The reason i am asking, OpenVX allocates pitch to be 64byte aligned and so for 1376 output width, pitch would be 1408. So could you please take care while saving frame?

Regards,

Brijesh

0 Dongwon Choi over 5 years ago in reply to Brijesh Jadav

Intellectual 845 points

Hi Brijesh.

I used the below function which is same function in vision_apps/apps/dl_demos/app_tidl_avp3/avp_scaler_module.c

vx_status writeScalerOutput(char* file_name, vx_object_array img_arr)

Regards

Dongwon Choi

0 Karthik Ramanan over 5 years ago in reply to Dongwon Choi

TI__Guru** 113810 points

Hi Choi,

Can you please update the latest on this thread? Can you please call out what is required as the next step?

Regards

Karthik

Processors

Processors forum

[TDA4M] MSC output is not NV12 but NV21