2ffbfda47d
test: print SMEM verify data
2026-05-28 11:51:08 +00:00
4fd41365de
test: add SMEM verify for HD=64 K-tile offsets
2026-05-28 11:49:44 +00:00
4483539f01
test: HD=64 random data, 4 K-tiles, accumulate
2026-05-28 11:47:56 +00:00
73bd21ce01
test: force 1 K-tile for HD=64 debug
2026-05-28 11:46:12 +00:00
abe1870429
test: HD=64 all-ones, expected S[0,j]=64 (unscaled) or 8.0 scaled
2026-05-28 11:44:31 +00:00
73f9ff98c9
test: UMMA QK HD=64 (4 K-tiles, accumulate) — multi-K-tile test
2026-05-28 11:42:29 +00:00
df34cae9c6
UMMA QK GEMM WORKING! Update docs — 4x was scale factor, not bug
...
Major milestone: UMMA QK GEMM produces correct attention scores at HD=16!
- MMA computes raw dot product; apply 1/sqrt(HD) scaling manually
- tcgen05.fence::after_thread_sync for MMA→TMEM fence
- 32x32b.x8 TMEM reads for Layout D output
- 4 warps (128 threads) required for M=128
- Next: HD=64 multi-K-tile, PV GEMM, full FMHA pipeline
2026-05-28 11:41:19 +00:00
1874a70a6d
test: fix var ref
2026-05-28 11:39:15 +00:00
8426d13285
test: fix comparison — row 0 is S[0,c], rows 1-127 should be zero
2026-05-28 11:38:22 +00:00
6f40fafa91
test: verify ALL 128 rows × 8 cols match scalar reference
2026-05-28 11:36:46 +00:00
3c7d9d9303
test: apply 1/sqrt(HD) scale to MMA output — 4x was the scale factor, not a bug!
2026-05-28 11:34:45 +00:00
013f370046
test: all-ones data, expected S[0,j]=16.0 for every j
2026-05-28 11:32:56 +00:00
f5a0966afc
test: 4 warp leaders (lane==0) call MMA simultaneously
2026-05-28 11:30:19 +00:00
c01d6fddf4
test: gau-nernst pattern — fence::after_thread_sync, 4 warps, 128 threads, 32x32b.x8 loop
2026-05-28 11:28:47 +00:00
a048b56886
test: single-thread MMA + 0.25 scaling for 4× factor
2026-05-28 10:23:06 +00:00
57d67e6b51
test: revert to 64-bit descriptors, 4 warp leaders, 32x32b read
2026-05-28 10:21:06 +00:00
32f7fa7bce
Update CURRENT_ISSUE.md and MEMORY.md with UMMA 4× bug details
...
- MMA produces exactly 4× scalar reference for all output values
- SMEM data verified correct, descriptor values correct
- 4× persists across different N, warp counts, padding
- TMEM multi-store bug documented (16x256b.x1 crashes on 2nd store)
- Layout D read with 32x32b.x8 works
- Next: study CUTLASS FMHA TMEM output layout to fix 4× factor
2026-05-28 10:15:14 +00:00
3f95f1c5d4
test: try LBO with block_mn=32 (1/4 of M=128)
2026-05-28 10:11:38 +00:00
d03e353972
test: 4 warp leaders call MMA (Layout D requires 4 warps)
2026-05-28 10:10:07 +00:00
8059ed15ad
test: explicitly zero padding between Q and K
2026-05-28 10:08:35 +00:00
9e98c067ab
test: Layout D TMEM read using 32x32b.x8 format, 4 warps
2026-05-28 10:07:15 +00:00
68d1a7920c
test: M=64 in both desc and idesc
2026-05-28 10:04:17 +00:00
0f51fda0da
test: try N=8 in idesc
2026-05-28 10:02:52 +00:00
4f7c9649fd
test: clean UMMA QK test, debug 4x factor, 8KB padding, 128 TMEM cols
2026-05-28 10:01:39 +00:00
ac65ece33b
test: TMEM 2-store with fence outside wid guard, 64 threads
2026-05-28 09:59:43 +00:00
2c89eea6be
test: fence+sync between 2 tmem_stores
2026-05-28 09:58:51 +00:00
24c5afe1dc
test: 64 threads, 2 stores to col 0
2026-05-28 09:57:53 +00:00
987f2c8917
test: 2 tmem_stores to SAME column 0
2026-05-28 09:57:07 +00:00
494149f034
test: 32 threads (1 warp), no guards, all participate
2026-05-28 09:56:17 +00:00
f0cb71da5c
test: TMEM 2-col with fence+sync between stores, separate wid==0 blocks
2026-05-28 09:54:19 +00:00
b69a538ab1
test: add fence+sync between 2 tmem_stores
2026-05-28 09:53:10 +00:00
7a21fa4bd8
test: add 2nd tmem_store to column 1
2026-05-28 09:52:05 +00:00
4b129c146e
test: add 1 tmem_load back
2026-05-28 09:51:21 +00:00
61f19ce891
test: skip tmem_load, only store+dealloc
2026-05-28 09:50:48 +00:00
2513e1a692
test: use 64 threads, fence outside warp guard, 1 store
2026-05-28 09:50:09 +00:00
abfe9dbaa1
test: only 1 tmem_store to verify single column works
2026-05-28 09:49:21 +00:00
5795589abc
test: TMEM 4 columns, individual store calls + loop load
2026-05-28 09:48:27 +00:00
8a428f6127
test: TMEM column addressing test (128 cols, store+load)
2026-05-28 09:46:49 +00:00
ee3fe6d6b2
test: tmem_load column 1 only
2026-05-28 09:45:34 +00:00
6c38c6e442
test: read 8 TMEM columns individually (no loop)
2026-05-28 09:44:30 +00:00
bcc6ed114d
test: add 8KB padding after sQ to prevent MMA read overrun
2026-05-28 09:43:17 +00:00
764ed01d6f
test: try M=64 in descriptor + idesc to debug 4x factor
2026-05-28 09:41:50 +00:00
4cb656e583
test: try idesc=0 (same as gau-nernst)
2026-05-28 09:40:19 +00:00
cfba8484da
test: try idesc with N=128 (full extent) + 128 TMEM cols
2026-05-28 09:39:19 +00:00
30f0056b11
test: clean rewrite with SMEM Q/K verification and dot product check
2026-05-28 09:38:26 +00:00
7eb85a71fc
test: add Q SMEM verification output + bf16_to_f32_host
2026-05-28 09:37:07 +00:00
8f23c2aaf6
test: verify SMEM Q layout by reading back canonical data
2026-05-28 09:35:58 +00:00
004046a6a8
test: read only 1 TMEM column after MMA
2026-05-28 09:35:02 +00:00
41128122e3
test: clean rewrite, 32 TMEM cols, MMA N=32, tmem_load loop
2026-05-28 09:33:45 +00:00
58be79957d
test: 32 TMEM cols, add MMA call with N=32, read S from TMEM
2026-05-28 09:32:33 +00:00
