[TOSA] does TOSA only support symmetry quantization?

For below TOSA MLIR, we could see tosa.conv’s quant info has only one weight zero point, but its next tosa.rescale op have multiplers and scales per channel. So dose this mean TOSA only support symmetry quantization, which means weight zero point is 0 for all channels? If not, where is other channels’ weight zero point saved in the TOSA IR?

  %191 = "tosa.conv2d"(%190, %153, %152) {dilation = [1, 1], pad = [0, 1, 0, 1], quantization_info = {input_zp = -1 : i32, weight_zp = 0 : i32}, stride = [2, 2]} : (tensor<1x192x192x3xi8>, tensor<32x3x3x3xi8>, tensor<32xi32>) -> tensor<1x96x96x32xi32>
  %192 = "tosa.rescale"(%191) {double_round = true, input_zp = 0 : i32, multiplier = [1210412592 : i32, 1804062962 : i32, 1593447828 : i32, 1501199931 : i32, 1508738108 : i32, 1643332332 : i32, 1366333750 : i32, 1501199931 : i32, 1112171479 : i32, 2118326527 : i32, 1398876642 : i32, 1501199931 : i32, 1501199931 : i32, 1501199931 : i32, 1995278927 : i32, 1089739904 : i32, 1552628968 : i32, 1080650795 : i32, 1855542004 : i32, 1501199931 : i32, 2072892157 : i32, 2053420967 : i32, 1347224629 : i32, 1584419561 : i32, 1629273664 : i32, 1128014807 : i32, 1407624077 : i32, 2064658172 : i32, 1501199931 : i32, 1864139338 : i32, 1530191250 : i32, 1458142436 : i32], output_zp = -128 : i32, per_channel = true, scale32 = true, shift = [38 : i32, 37 : i32, 39 : i32, 52 : i32, 38 : i32, 40 : i32, 36 : i32, 52 : i32, 39 : i32, 40 : i32, 37 : i32, 52 : i32, 52 : i32, 52 : i32, 38 : i32, 39 : i32, 43 : i32, 39 : i32, 37 : i32, 52 : i32, 38 : i32, 38 : i32, 38 : i32, 40 : i32, 40 : i32, 39 : i32, 43 : i32, 37 : i32, 52 : i32, 40 : i32, 37 : i32, 40 : i32]} : (tensor<1x96x96x32xi32>) -> tensor<1x96x96x32xi8>

That is correct. TOSA convolution only supports symmetric per channel quantization. As you point out, there is only one zero point for the weights available. (which would be expected to be 0 for per-channel quantized weights)

This has been the common quantized convolution that we’ve seen in use. Do you have cases where asymmetric per-channel quantization is providing a significant benefit?

Thanks for your reply.