dsg/mesa
2
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ethosu: Fix ublock selection for 8-bit depthwise/pooling on U85-256

Browse source 
For U85-256 with 8-bit IFM, Vela's _uBlockToOpTable restricts which
microblocks are valid per operation type:

  {2,2,8}  and {4,1,8}:  conv, matmul, vectorprod, reducesum, eltwise, resize
  {2,1,16}:              depthwise, pool, eltwise, reduceminmax, argmax, resize

Mesa's find_ublock() was not enforcing these constraints, allowing
{4,1,8} or {2,2,8} to be selected for depthwise/pooling based on
minimum waste. For depthwise ops with OFM shapes that aligned better
to {4,1,8}, the wrong ublock was chosen, causing incorrect weight
encoding and NPU hangs.

Fix by skipping {4,1,8} and {2,2,8} for depthwise/pooling operations,
matching Vela's operation-validity table.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39611>
This commit is contained in:
Tomeu Vizoso 2026-03-22 16:12:21 +01:00 • committed by Marge Bot
parent dc36a32214
commit f9cd399eb0
1 changed files with 18 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

28
src/gallium/drivers/ethosu/ethosu_sched_u85.c
View file

@ -656,23 +656,31 @@ find_ublock(struct ethosu_operation *operation, bool is_part_kernel)
continue; /* Skip 2x1x16 for part-kernel 1x1 */
}
/* The 2x1x16 microblock is only valid for 16-bit IFM convolutions.
* For 8-bit IFM, it can only be used for depthwise/pooling operations.
* This matches the bitsToOperations table in Vela's ethos_u85.cpp.
/* U85-256 ublock-to-operation validity for 8-bit IFM
* (from Vela's _uBlockToOpTable in ethos_u85.cpp):
*
* {2,2,8} / Shape(2,2,8): conv, matmul, vectorprod, reducesum, eltwise, resize
* {4,1,8} / Shape(1,4,8): conv, matmul, vectorprod, reducesum, eltwise, resize
* {2,1,16} / Shape(1,2,16): depthwise, pool, eltwise, reduceminmax, argmax, resize
*
* So for 8-bit IFM:
* - depthwise/pooling can ONLY use {2,1,16}
* - convolution (non-depthwise) CANNOT use {2,1,16}
*/
/* Skip {2,1,16} for 8-bit non-depthwise convolutions */
if (ublk.width == 2 && ublk.height == 1 && ublk.depth == 16 &&
operation->type == ETHOSU_OPERATION_TYPE_CONVOLUTION &&
!is_depthwise) {
continue; /* Skip 2x1x16 for 8-bit IFM non-depthwise convolutions */
continue;
}
/* For non-1x1 regular convolutions, skip 2x1x16 unless it's depthwise */
if (!is_pointwise && !is_depthwise &&
operation->type == ETHOSU_OPERATION_TYPE_CONVOLUTION &&
ublk.width == 2 &&
ublk.height == 1 &&
ublk.depth == 16)
/* Skip {4,1,8} and {2,2,8} for 8-bit depthwise/pooling —
* only {2,1,16} is valid for these operations at 8-bit */
if (!(ublk.width == 2 && ublk.height == 1 && ublk.depth == 16) &&
(is_depthwise || is_pooling)) {
continue;
}
/* Minimum waste is better than aspect correct */
struct ethosu_block tmp = block_round_away(operation->ofm.shape, ublk);