[XNNPACK EP] Read Gemm M from the input tensor at Compute time (#31189)
Fixes #31153
`Gemm::Gemm` cached the row count from the static `NodeArg` shape, and
`dim_value()` returns 0
for a symbolic dim, so `M_ = dim(0).dim_value() > 1 ? dim(0).dim_value()
: 1` turned a dynamic
M into 1. `Compute()` then used that cached `M_` for both the output
shape and the
`xnn_reshape_fully_connected_nc_*` row count, so the kernel produced a
`[1, N]` output and
multiplied only the first row.
`Compute()` now reads M from `A->Shape()`, which is what the XNNPACK
reshape API wants anyway.
`M_` is removed since nothing else used it. The `transA = 1` path is
unchanged: M still comes
from `A->Shape()[1]` and `K_` is still passed as the row count.
Verified on x86_64 Linux (Ubuntu 20.04, gcc 13, Release,
`--use_xnnpack`):
- The reporter's repro script fails on `Reshape` with `Input
shape:{1,16}` before the change
and runs clean after.
- New `XnnpackEP.TestGemm_DynamicM` fails before and passes after. It
feeds `[10, 3]` to a
symbolic-M Gemm and compares against the CPU EP.
- `onnx_test_runner -e xnnpack` on a symbolic-M Gemm with four data sets
(M = 10, 1, 7, 32)
against numpy-computed outputs: 3 of 4 fail before, 4 of 4 pass after.
This covers
re-reshaping the cached operator across runs, not just one M.
- Full `onnxruntime_provider_test`: 5587 ran, 5461 passed, rest skipped,
exit 0.
Not verified: Android ARM64 (the reporter's platform, no device here),
and the fp16 branch,
which has no test coverage before or after. Both take the same one line
of row-count plumbing.
Two adjacent things I left alone because they are out of scope for this
issue, but reviewers
may want to know: `IsOnnxNodeSupported` still accepts `transA = 1`,
which looks broken
independently of M (for a `[K, M]` A the kernel passes `K_` as the row
count against a
K-input-channel operator, which only lines up when `M == K`), and it
still accepts rank-1 A
while the constructor indexes `shapeA->dim(1)`.
Thanks to @whousemyname for the report, which pinpointed the exact line
and both candidate
fixes.
Co-authored-by: guptaishaan <guptaishaan@users.noreply.github.com>
onnxruntime
4c4d4923 - [XNNPACK EP] Read Gemm M from the input tensor at Compute time (#31189)
