4 appearances on the backlist front page in the last 30 days.
Do u know that Triton can specialize on pointer alignment (16B or not) and non-constexpr int (1, multiple of 2/4/8/16.., or not). There are good reasons for it, but can be unexpected if unaware.
A kernel-tuning note shows how high-level fp32 max operations lower into predicate/select PTX and then fuse into a new Blackwell SASS instruction
Quickly whipped up a few CuteDSL kernels for sm120 (TMA+mma.sync). There is quite a gap wrt CuBLAS @ 600W, but surprisingly, my kernels come up slightly ahead @ 400W. I have a feeling TMA is more energy-efficient than cp.async? (CuBLAS disp
Finally got some time to port my tcgen05 kernels to CuteDSL. For PTX enjoyers, this should feel natural (except TMA ). A BF16 MMA mainloop is shown below. I also worked up an example for MXFP8 and NVFP4.