[SRC] Enable unroll=1 for gfx950 (#1602)

* [SRC] Enable unroll=1 for gfx950

* Fix typo from rebase in generate.py

* Support for unroll=1 and gfx90a when building for all GPU targets

---------

Signed-off-by: nileshnegi <Nilesh.Negi@amd.com>
This commit is contained in:
Nilesh M Negi
2025-03-27 18:21:35 -05:00
committed by GitHub
parent 9dc23d9265
commit 307bc10781
8 changed files with 116 additions and 40 deletions
+9 -6
View File
@@ -30,17 +30,20 @@ struct ncclKernelMatch {
};
#ifdef ENABLE_COLLTRACE
#define ncclGetKernelIndex(p_comm) ((p_comm)->unroll + ((p_comm)->collTraceEnabled ? 2 : 0))
static ncclKernelMatch const ncclKerns[4] = {
{(void *)ncclDevKernel_Generic, true},
#define ncclGetKernelIndex(p_comm) ((p_comm)->unroll + ((p_comm)->collTraceEnabled ? 3 : 0))
static ncclKernelMatch const ncclKerns[6] = {
{(void *)ncclDevKernel_Generic_1, true},
{(void *)ncclDevKernel_Generic_2, true},
{(void *)ncclDevKernel_Generic_4, true},
{(void *)ncclDevKernelDebug_Generic, true},
{(void *)ncclDevKernelDebug_Generic_1, true},
{(void *)ncclDevKernelDebug_Generic_2, true},
{(void *)ncclDevKernelDebug_Generic_4, true}
};
#else
#define ncclGetKernelIndex(p_comm) ((p_comm)->unroll)
static ncclKernelMatch const ncclKerns[2] = {
{(void*)ncclDevKernel_Generic, true},
static ncclKernelMatch const ncclKerns[3] = {
{(void*)ncclDevKernel_Generic_1, true},
{(void*)ncclDevKernel_Generic_2, true},
{(void*)ncclDevKernel_Generic_4, true}
};
#endif