deepseek-ai/DeepGEMM

Fork 0

mirror of https://github.com/deepseek-ai/DeepGEMM.git synced 2026-11-05 00:47:26 +00:00

[GH-ISSUE #145] Question about restriction on BLK_N on SM100 when type is 1d2d #82

New issue

Closed

opened 2026-10-26 21:12:06 +00:00 by bsakaguchi · 3 comments

bsakaguchi commented

2026-10-26 21:12:06 +00:00

Owner

Originally created by @pengwubj on GitHub (Jul 30, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/145

        // 1D2D kernels' maximum block N is 128
        // 1D2D kernels require more friendly block Ns
        if (kernel_type == KernelType::Kernel1D2D and (block_n > 128 or 128 % block_n != 0))
            return false;

Based on the above functions, the restriction on block_n is more strict on sm100 when type is 1d2d, where BLK_N cannot span across two 128x128 blocks, but there's no such restriction on sm90. Since sm90 has already been optimized to allow the compiler to statically deduce predicates, reducing register usage and corresponding sel instructions, why is this restriction added on sm100? Thanks!
2. Another question is why BLK_N does not use 192 in the sm90 kernel. This shape also meets the requirement of 'At most 2 scales in the N direction'.

Originally created by @pengwubj on GitHub (Jul 30, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/145 1. ```c++ // 1D2D kernels' maximum block N is 128 // 1D2D kernels require more friendly block Ns if (kernel_type == KernelType::Kernel1D2D and (block_n > 128 or 128 % block_n != 0)) return false; ``` Based on the above functions, the restriction on block_n is more strict on sm100 when type is 1d2d, where BLK_N cannot span across two 128x128 blocks, but there's no such restriction on sm90. Since sm90 has already been optimized to allow the compiler to statically deduce predicates, reducing register usage and corresponding sel instructions, why is this restriction added on sm100? Thanks! 2. Another question is why BLK_N does not use 192 in the sm90 kernel. This shape also meets the requirement of 'At most 2 scales in the N direction'.

bsakaguchi closed this issue

2026-10-26 21:12:07 +00:00

bsakaguchi commented

2026-10-26 21:12:08 +00:00

Author

Owner

@LyricZhao commented on GitHub (Jul 30, 2025):

cc @RayWang96
192 may lead to insufficient registers (maybe sufficient but requires BLOCK_M <= 128). We have already implemented another way of larger block sizes (supporting BLOCK_M == 256).

@LyricZhao commented on GitHub (Jul 30, 2025): 1. cc @RayWang96 2. 192 may lead to insufficient registers (maybe sufficient but requires `BLOCK_M <= 128`). We have already implemented another way of larger block sizes (supporting `BLOCK_M == 256`).

bsakaguchi commented

2026-10-26 21:12:09 +00:00

Author

Owner

@RayWang96 commented on GitHub (Jul 30, 2025):

First of all, the Blackwell 1D2D kernel was implemented as a workaround to maintain compatibility with the old recipe, so it hasn’t been deeply optimized. These constraints were chosen because our tests showed the best performance under the above limits. If you remove the assert, the code should still run.

@RayWang96 commented on GitHub (Jul 30, 2025): First of all, the Blackwell 1D2D kernel was implemented as a workaround to maintain compatibility with the old recipe, so it hasn’t been deeply optimized. These constraints were chosen because our tests showed the best performance under the above limits. If you remove the assert, the code should still run.

bsakaguchi commented

2026-10-26 21:12:09 +00:00

Author

Owner

@pengwubj commented on GitHub (Jul 31, 2025):

Thanks for quick reply.

@pengwubj commented on GitHub (Jul 31, 2025): Thanks for quick reply.

No labels

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

deepseek-ai/DeepGEMM#82

No description provided.

Rows
Columns