diff options
| author | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-10-14 14:05:41 -0400 |
|---|---|---|
| committer | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-10-14 14:05:41 -0400 |
| commit | fce5ab8d025799d71fbd260c4ef91a715497b54b (patch) | |
| tree | b832bd80bc53748eaf0574532793dc958b67aced /kernel/gsp_sema_service.c | |
| parent | 6b090f70816f60a1a98fcf5889a3275177e8c132 (diff) | |
| download | open-iov-fce5ab8d025799d71fbd260c4ef91a715497b54b.tar.gz open-iov-fce5ab8d025799d71fbd260c4ef91a715497b54b.zip | |
vgsp-gr: enlarge the data bounce window to 208 pages (256 KB operands)
victory_matmul (256x256 float: 256 KB per operand) failed with all-wrong
results while 128 KB workloads passed: the relocation bounce window was
112 pages with a 100-page data cap and 33 pages per pointer, so each 64-page
operand was truncated and the engine ran on partial/mismatched windows.
Grow the planted bounce window to 208 pages (main-PT slots 0x110..0x1df),
move the mail page to the last data-free slot (206, marker 207), and raise
the per-pointer cap to 68 pages and the data cap to 205 (three 256 KB
operands = 3x68 = 204 <= 205). Operands larger than 256 KB still hit the cap
and need a chunked data path (documented at the define).
Diffstat (limited to 'kernel/gsp_sema_service.c')
| -rw-r--r-- | kernel/gsp_sema_service.c | 16 |
1 files changed, 12 insertions, 4 deletions
diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c index 10992c3..c07915f 100644 --- a/kernel/gsp_sema_service.c +++ b/kernel/gsp_sema_service.c @@ -12366,10 +12366,18 @@ static bool svc_gr_page_inline_covers(u64 page) * accesses the tail: with 32 pages the tail read the NEXT candidate's * relocated page (wrong values; vadd tails). 3x33=99 <= the 100-page cap. */ -#define SVC_GR_DATA_PAGES_PER_PTR 33u -/* m114: total pages the pre-commit bounce window can carry (slot 100 is - * the mail message). */ -#define SVC_GR_DATA_BOUNCE_MAX 100u +/* + * m150 (large buffers): the bounce window was grown to 208 pages (mail page + * at slot 206). 68 pages per candidate/pointer covers a 256 KB cudaMalloc'd + * operand whose pointer sits 0x400 into its page (64 full + offset + tail), + * and up to three such operands (3x68 = 204 <= 205) fit below the mail page. + * Larger-than-256 KB operands still hit the cap; they need a chunked data + * path (future work). + */ +#define SVC_GR_DATA_PAGES_PER_PTR 68u +/* Total pages the pre-commit bounce window can carry below the mail page + * (GSP_PHYS_RMCALL_GR_MAIL_SLOT = 206). */ +#define SVC_GR_DATA_BOUNCE_MAX 205u static u32 svc_gr_data_log; /* m142: bounded per-page attribution log */ static u32 svc_gr_src_log; /* m150: bounded per-candidate source log */ |