Triton Blocked Matrix Multiplication | A MyTorch Sidequest! скачать в хорошем качестве

Triton Blocked Matrix Multiplication | A MyTorch Sidequest!

Code: https://github.com/priyammaz/TritonKe... Previously we implemented a very slow Matrix Multiplication kernel in Triton. The main reason was that each PID was responsible for one of the outputs in the result of our matmul. Instead, we will now make every PID responsible for a block of outputs in the matmul, reducing the overall scheduling overhead and increasing throughput! Timestamps: 00:00:00 - What is Blocked Matmul? 00:03:00 - Lets write the Pseudocode! 00:10:00 - Start Triton Blocked MatMul 00:13:00 - How Offsets are computed 00:17:20 - Get our pointers to block of A and B 00:21:00 - Loop over inner dimension K in blocks 00:22:08 - Perform the Multiply Accumulate Operation 00:28:30 - Store results in final output array 00:30:13 - Triton Autotuner 00:33:48 - Dynamic grid calculation 00:36:05 - Quick bug fix 00:36:40 - Tensorfloat32 Precision Issues 00:39:00 - Performance comparisons Socials! X   / data_adventurer   Instagram   / nixielights   Linkedin   / priyammaz   Discord   / discord   🚀 Github: https://github.com/priyammaz 🌐 Website: https://www.priyammazumdar.com/