Revert "[BE] Cleanup stale comments/copy from gemm (#162001)"

This reverts commit b40d9432be.

Reverted https://github.com/pytorch/pytorch/pull/162001 on behalf of https://github.com/jeanschmidt due to break a few internal tests ([comment](https://github.com/pytorch/pytorch/pull/161999#issuecomment-3255381925))
This commit is contained in:
PyTorch MergeBot 2025-09-04 19:56:48 +00:00
parent 9e5247f51d
commit afa6e5604d

View File

@ -457,9 +457,24 @@ void gemm(
return; return;
} }
#endif #endif
// for the fallback path, first compute gemm with beta = 0,
// and then add c in full precision.
int64_t c_size = n * m;
std::vector<float> float_c(c_size, 0.f);
gemm_no_downcast_stub( gemm_no_downcast_stub(
at::kCPU, at::kBFloat16, at::kCPU, at::kBFloat16,
transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc); transa, transb, m, n, k, alpha, a, lda, b, ldb, 0.f, float_c.data(), m);
for (const auto j : c10::irange(n)) {
for (const auto i : c10::irange(m)) {
auto offset = j * ldc + i;
// beta == 0 won't propagate NaN from C
if (beta == 0.f) {
c[offset] = float_c[j * m + i];
} else {
c[offset] = beta * c[offset] + float_c[j * m + i];
}
}
}
} }
void gemm( void gemm(
@ -478,9 +493,24 @@ void gemm(
return; return;
} }
#endif #endif
// for the fallback path, first compute gemm with beta = 0,
// and then add c in full precision.
int64_t c_size = n * m;
std::vector<float> float_c(c_size, 0.f);
gemm_no_downcast_stub( gemm_no_downcast_stub(
at::kCPU, at::kHalf, at::kCPU, at::kHalf,
transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc); transa, transb, m, n, k, alpha, a, lda, b, ldb, 0.f, float_c.data(), m);
for (const auto j : c10::irange(n)) {
for (const auto i : c10::irange(m)) {
auto offset = j * ldc + i;
// beta == 0 won't propagate NaN from C
if (beta == 0.f) {
c[offset] = float_c[j * m + i];
} else {
c[offset] = beta * c[offset] + float_c[j * m + i];
}
}
}
} }
void gemm( void gemm(