Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Navi card subgroup shuffle support for gemm #512

Open
wants to merge 7 commits into
base: master
Choose a base branch
from
Open

Navi card subgroup shuffle support for gemm #512

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
fde7943
Some draft
fancyIX Oct 24, 2023
4e35c89
modifiy draft with sub_group func
fancyIX Oct 24, 2023
a871605
Add Navi subgroup shuffle support
fancyIX Oct 28, 2023
b5fe348
try to fix build error
fancyIX Oct 29, 2023
3ef0f88
try to fix build error
fancyIX Oct 29, 2023
27648f5
try to fix build error of too long string
fancyIX Oct 29, 2023
ed32af0
only apply asm for shuffle for single precision
fancyIX Nov 1, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 19 additions & 0 deletions src/kernels/level3/xgemm_part1.opencl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -138,6 +138,25 @@ R"(
#endif
#endif

#if USE_SUBGROUP_SHUFFLING == 1 && SUBGROUP_SHUFFLING_GCN == 1
#define SUBGROUP_SIZE 32 // Assumes subgroup size is always 4 on AMD GCN GPUs
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

On the left you write 32, on the right you write 4, one of them probably is incorrect?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes. It should be 32 (for Navi cards). Will change the comment.

#define NAVI_SHFL(s0, l) \
{ \
__asm ( \
"ds_bpermute_b32 %[dos0], %[ol0], %[os0]\n" \
"s_waitcnt lgkmcnt(0)\n" \
: [dos0] "=&v" (s0) \
: [ol0] "v" (l), \
[os0] "0" (s0)); \
}
#define NAVI_LID() \
if (get_work_dim() == 2) { \
return (get_local_size(0) * get_local_id(1) + get_local_id(0)) % SUBGROUP_SIZE; \
} else { \
return (get_local_id(0)) % SUBGROUP_SIZE; \
}
#endif

#if NWI != SUBGROUP_SIZE || MDIMC < SUBGROUP_SIZE
#undef USE_SUBGROUP_SHUFFLING
#define USE_SUBGROUP_SHUFFLING 0 // Disables subgroups in case the assumptions don't hold
Expand Down
6 changes: 6 additions & 0 deletions src/kernels/level3/xgemm_part3.opencl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,8 @@ INLINE_FUNC int clblast_get_sub_group_local_id() {
int ret;
asm volatile("mov.u32 %0, %%laneid;" : "=r"(ret) );
return ret;
#elif SUBGROUP_SHUFFLING_GCN == 1
NAVI_LID()
#endif
}

Expand All @@ -49,6 +51,10 @@ INLINE_FUNC realN clblast_sub_group_shuffle(realN reg, int src) {
asm volatile("shfl.idx.b32 %0, %1, %2, 0x1f;" : "=f"(ret): "f"(reg), "r"(src));
#endif
return ret;
#elif SUBGROUP_SHUFFLING_GCN == 1
realN ret = reg;
NAVI_SHFL(ret, ((src) << 2))
return ret;
#endif
}
#endif
Expand Down
5 changes: 5 additions & 0 deletions src/utilities/compile.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,6 +78,11 @@ std::shared_ptr<Program> CompileFromSource(
}
}

if (device.IsGPU() && device.IsAMD() && device.Name().find("gfx1") != std::string::npos) {
header_string += "#define USE_SUBGROUP_SHUFFLING 1\n";
header_string += "#define SUBGROUP_SHUFFLING_GCN 1\n";
}
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Implementing it like this forces it on for all these AMD GPUs. Did you verify that using subgroup shuffling is always better compared to not using subgroup shuffling? The easiest way to test is to run the XGEMM kernel tuner with and without this modification and compare execution times of a good portion of the first chunk of kernels.

The alternative to what you implemented here is to make it a AMD-specific tuning parameter, and then at tuning time it will decide whether subgroup shuffling was a good idea or not.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It only is supposed to turn on with device name beginning with "gfx1", which I assume only applies on Navi cards, like gfx1010, gfx 1030, etc

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I have run xgemm tuner and it seems a little bit better on rx 6900 xt card. I can test on rx 5700 xt too later.


// For Qualcomm devices, specifying the OpenCL kernel attribute reqd_work_group_size reduces performance.
// This option compiles without the workgroup size requirement and does not affect correctness.
if (device.IsQualcomm()) {
Expand Down