This patch adjusts OpenACC loop lowering in the AMD GCN target compiler
in such a way that the autovectorizer can vectorize the "vector" dimension
of those loops in more cases.
Rather than generating "SIMT" code that executes a scalar instruction
stream for each lane of a vector in lockstep, for GCN we model the GPU
like a typical CPU, with separate instructions to operate on scalar and
vector data. That means that unlike other offload targets, we rely on
the autovectorizer to handle the innermost OpenACC parallelism level,
which is "vector".
Because of this, the OpenACC builtin functions to return the current
vector lane and the vector width return 0 and 1 respectively, despite
the native vector width being 64 elements wide.
This allows generated code to work with our chosen compilation model,
but the way loops are lowered in omp-offload.c:oacc_xform_loop does not
understand the discrepancy between logical (OpenACC) and physical vector
sizes correctly. That means that if a loop is partitioned over e.g. the
worker AND vector dimensions, we actually lower with unit vector size --
meaning that if we then autovectorize, we end up trying to vectorize
over the "worker" dimension rather than the vector one! Then, because
the number of workers is not fixed at compile time, that means the
autovectorizer has a hard time analysing the loop and thus vectorization
often fails entirely.
We can fix this by deducing the true vector width in oacc_xform_loop,
and using that when we are on a "non-SIMT" offload target. We can then
rearrange how loops are lowered in that function so that the loop form
fed to the autovectorizer is more amenable to vectorization -- namely,
the innermost step is set to process each loop iteration sequentially.
For some benchmarks, allowing vectorization to succeed leads to quite
impressive performance improvements -- I've observed between 2.5x and
40x on one machine/GPU combination.
The low-level builtins available to user code (__builtin_goacc_parlevel_id
and __builtin_goacc_parlevel_size) continue to return 0/1 respectively
for the vector dimension for AMD GCN, even if their containing loop is
vectorized -- that's a quirk that we might possibly want to address at
some later date.
Only non-"chunking" loops are handled at present. "Chunking" loops are
still lowered as before.
2021-01-13 Julian Brown <julian@codesourcery.com>
gcc/
* omp-offload.c (oacc_thread_numbers): Add VF_BY_VECTORIZER parameter.
Add overloaded wrapper for previous arguments & behaviour.
(oacc_xform_loop): Lower vector loops to iterate a multiple of
omp_max_vf times over contiguous steps on non-SIMT targets.
projects / thirdparty / gcc.git / commit
