[PR #48] [CLOSED] Add TMA optimization hint for large FFMA segments #220

New issue

Closed

opened 2026-10-26 21:15:00 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 21:15:00 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/48
Author: @A-transformer
Created: 3/7/2025
Status: ❌ Closed

Base: main ← Head: add_TMA_optimization_hint

📝 Commits (1)

82d4b99 Add TMA optimization hint for large FFMA segments

📊 Changes

1 file changed (+5 additions, -0 deletions)

View changed files

📝 deep_gemm/jit/interleave_ffma.py (+5 -0)

📄 Description

This check is added to identify segments with a high number of FFMA (Fused Multiply-Add) instructions, which may benefit from Tensor Memory Accelerator (TMA) optimizations on NVIDIA Hopper architecture. TMA enables asynchronous data transfers from global memory to shared memory, reducing register pressure and improving data access efficiency. Segments with a large number of FFMA instructions often involve significant data movement, making them prime candidates for TMA to accelerate memory operations and enhance overall GEMM performance.

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/deepseek-ai/DeepGEMM/pull/48 **Author:** [@A-transformer](https://github.com/A-transformer) **Created:** 3/7/2025 **Status:** ❌ Closed **Base:** `main` ← **Head:** `add_TMA_optimization_hint` --- ### 📝 Commits (1) - [`82d4b99`](https://github.com/deepseek-ai/DeepGEMM/commit/82d4b996f7898124c22831dd1926f97749af0bc3) Add TMA optimization hint for large FFMA segments ### 📊 Changes **1 file changed** (+5 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `deep_gemm/jit/interleave_ffma.py` (+5 -0) </details> ### 📄 Description This check is added to identify segments with a high number of FFMA (Fused Multiply-Add) instructions, which may benefit from Tensor Memory Accelerator (TMA) optimizations on NVIDIA Hopper architecture. TMA enables asynchronous data transfers from global memory to shared memory, reducing register pressure and improving data access efficiency. Segments with a large number of FFMA instructions often involve significant data movement, making them prime candidates for TMA to accelerate memory operations and enhance overall GEMM performance. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 21:15:00 +00:00