Skip to content

[ROCm] Add mixed precision support to the HIP optimized warp per row TBE backward kernel #163

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
avbokovoy wants to merge 1 commit into
base: abokovoi/upstream
Choose a base branch
from
Open

[ROCm] Add mixed precision support to the HIP optimized warp per row TBE backward kernel #163

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 2 additions & 4 deletions fbgemm_gpu/codegen/training/backward/embedding_backward_split_template.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1257,17 +1257,15 @@ Tensor {{ embedding_cuda_op }}(
const static auto use_hip_kernel = fbgemm_gpu::config::is_feature_enabled(fbgemm_gpu::config::FeatureGateName::TBE_ROCM_HIP_BACKWARD_KERNEL);

constexpr bool supported_weights_type = std::is_same_v<emb_t, float> || std::is_same_v<emb_t, at::Half>;
constexpr bool supported_grad_type = std::is_same_v<grad_t, float> || std::is_same_v<grad_t, at::Half>;
constexpr bool supported_grad_type = std::is_same_v<grad_t, float> || std::is_same_v<grad_t, at::Half> || std::is_same_v<grad_t, at::BFloat16>;
const bool cached = uvm_weights.numel() > 0 || lxu_cache_weights.numel() > 0;

constexpr bool same_precision = std::is_same_v<emb_t, grad_t>;

// The optimized HIP backward kernel uses warpSize-64-only
// intrinsics; it must only be dispatched on warpSize 64
// devices. warpSize 32 devices fall through to the generic
// shuffle-based backward kernel.

if (use_hip_kernel && at::cuda::warp_size() == 64 && !mixed_D && !cached && supported_weights_type && supported_grad_type && same_precision && rocm::is_supported_cdna())
if (use_hip_kernel && at::cuda::warp_size() == 64 && !mixed_D && !cached && supported_weights_type && supported_grad_type && rocm::is_supported_cdna())
{
constexpr int segments_per_workgroup = 4;
{%- for kDimSize in [64, 128, 160, 192, 256, 320] %}
Expand Down
Loading