fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-07 02:48:06 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

teflon: Add support for symmetric per-channel quantization

Browse source 
The old code would assert when a model has multiple scales but only
one zero_point. This is common for symmetric quantization where all
channels share the same zero_point (typically 0).

Handle this by replicating the single zero_point for all channels
instead of crashing.

Fixes MoveNet models using per-channel quantization.

Signed-off-by: Anders Roxell <anders.roxell@linaro.org>
Reviewed-by: Tomeu Vizoso <tomeu@tomeuvizoso.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39594>
This commit is contained in:
Anders Roxell 2026-01-15 14:17:11 +01:00 • committed by Marge Bot
parent 292fffac1b
commit 0887e6d89f
1 changed files with 23 additions and 8 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

31
src/gallium/frontends/teflon/tfl_device.c
View file

@ -334,14 +334,28 @@ fill_tensor(struct teflon_delegate *delegate, TfLiteContext *tf_context, struct
tensor->scale = quant->scale->data[0];
tensor->zero_point = quant->zero_point->data[0];
assert(quant->scale->size == quant->zero_point->size);
if (quant->scale->size > 1 &&
(!all_scales_equal(quant) || !all_zero_points_equal(quant))) {
/* Handle per-channel quantization */
if (quant->scale->size > 1 && !all_scales_equal(quant)) {
tensor->scales = calloc(quant->scale->size, sizeof(*tensor->scales));
memcpy(tensor->scales, quant->scale->data, quant->scale->size * sizeof(*tensor->scales));
tensor->zero_points = calloc(quant->zero_point->size, sizeof(*tensor->zero_points));
memcpy(tensor->zero_points, quant->zero_point->data, quant->zero_point->size * sizeof(*tensor->zero_points));
tensor->zero_points = calloc(quant->scale->size, sizeof(*tensor->zero_points));
if (quant->zero_point->size == quant->scale->size) {
/* Same number of zero_points as scales - copy directly */
memcpy(tensor->zero_points, quant->zero_point->data, quant->scale->size * sizeof(*tensor->zero_points));
} else if (quant->zero_point->size == 1) {
/* Single zero_point for all channels (common for symmetric quantization) - replicate it */
for (int i = 0; i < quant->scale->size; i++) {
tensor->zero_points[i] = quant->zero_point->data[0];
}
} else {
/* Unexpected case - use first zero_point for all */
fprintf(stderr, "teflon: WARNING: tensor %d has %d scales but %d zero_points, using first zero_point for all\n",
index, quant->scale->size, quant->zero_point->size);
for (int i = 0; i < quant->scale->size; i++) {
tensor->zero_points[i] = quant->zero_point->data[0];
}
}
}
}
@ -720,9 +734,10 @@ fused_relu6_supported(TfLiteTensor *tensor)
assert(tensor->quantization.type == kTfLiteAffineQuantization);
affine = (TfLiteAffineQuantization *)tensor->quantization.params;
assert(affine->scale->size == affine->zero_point->size);
for (int i = 0; i < affine->zero_point->size; i++) {
if ((quantized_max - affine->zero_point->data[i]) * affine->scale->data[i] > 6.0f)
/* Handle per-channel quantization where zero_point->size may be 1 */
for (int i = 0; i < affine->scale->size; i++) {
int zp_idx = (affine->zero_point->size == 1) ? 0 : i;
if ((quantized_max - affine->zero_point->data[zp_idx]) * affine->scale->data[i] > 6.0f)
return false;
}
return true;