[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:
+9
-6
@@ -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
|
||||
|
||||
Reference in New Issue
Block a user