Skip to content

Adding gfx908 support #137

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

Open
UD-mmcminn wants to merge 5 commits into
base: main
Choose a base branch
from
Open

Adding gfx908 support #137

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
4c5f934
Adding gfx908 support, needs testing
UD-mmcminn Nov 5, 2025
5772f27
Merge branch 'main' into main
UD-mmcminn Nov 7, 2025
e087808
Adding 908 to isArchExperimentallySupported, removed changes from v2.…
UD-mmcminn Nov 18, 2025
61ea89c
Merge branch 'main' of github.com:UD-mmcminn/aotriton
UD-mmcminn Nov 18, 2025
01fa9c8
Merge branch 'ROCm:main' into main
UD-mmcminn Nov 18, 2025
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
2 changes: 1 addition & 1 deletion CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -80,7 +80,7 @@ option(AOTRITON_INHERIT_SYSTEM_SITE_TRITON "Use system site packages and verify
set(AOTRITON_USE_LOCAL_TRITON_WHEEL "" CACHE STRING "Substitute install from third_party/triton with install from local pip wheel package.")
set(AOTRITON_GPU_BUILD_TIMEOUT "8.0" CACHE STRING "GPU kernel compiler times out after X minutes. 0 for indefinite. Highly recommended if AOTRITON_BUILD_FOR_TUNING=On.")
option(AOTRITON_TERMINATE_WHEN_GPU_BUILD_TIMEOUT "compile.py will return error when GPU built timeout. Defensive option for CI" OFF)
set(AOTRITON_TARGET_ARCH "gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1102;gfx1103;gfx1151;gfx1150;gfx1201;gfx1200" CACHE STRING "Target GPU Architecture. Select all GPUs within the given list")
set(AOTRITON_TARGET_ARCH "gfx908;gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1102;gfx1103;gfx1151;gfx1150;gfx1201;gfx1200" CACHE STRING "Target GPU Architecture. Select all GPUs within the given list")
set(TARGET_GPUS "OBSOLETE" CACHE STRING "OBSOLETE. To select only one GPU, use AOTRITON_TARGET_ARCH or AOTRITON_OVERRIDE_TARGET_GPUS.")
set(AOTRITON_OVERRIDE_TARGET_GPUS "" CACHE STRING "Override AOTRITON_TARGET_ARCH, and only build for GPUs within this list.")
set(AOTRITON_ALT_TRITON_WHEEL_CONFIG_FILE "" CACHE STRING "Specify how to use alternative Triton wheels for different architectures in a YAML config file.")
Expand Down
1 change: 1 addition & 0 deletions include/aotriton/util.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,6 +52,7 @@ enum AOTRITON_API GpuVendor : uint16_t {
// More bits for potential non-PCI architectures
enum AOTRITON_API Gpu : uint64_t {
GPU_ARCH_UNKNOWN = 0,
GPU_AMD_ARCH_GFX908_MOD0 = TRICAT(GpuVendor::kAMD, 0x908, 0),
GPU_AMD_ARCH_GFX90A_MOD0 = TRICAT(GpuVendor::kAMD, 0x90a, 0),
GPU_AMD_ARCH_GFX942_MOD0 = TRICAT(GpuVendor::kAMD, 0x942, 0),
GPU_AMD_ARCH_GFX942_MOD1 = TRICAT(GpuVendor::kAMD, 0x942, 1),
Expand Down
1 change: 1 addition & 0 deletions requirements-tuning.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,3 +5,4 @@ filelock
celery[pyamqp,sqlalchemy]
psycopg # the more mature psycopg2 needs extra binaries.
dacite
textual
2 changes: 2 additions & 0 deletions v2python/gpu_targets.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@
from collections import defaultdict

AOTRITON_SUPPORTED_GPUS = (
'gfx908_mod0',
'gfx90a_mod0',
'gfx942_mod0',
# 'gfx942_mod1',
Expand All @@ -21,6 +22,7 @@
}

AOTRITON_ARCH_WARPSIZE = {
'gfx908' : 64,
'gfx90a' : 64,
'gfx942' : 64,
'gfx950' : 64,
Expand Down
2 changes: 1 addition & 1 deletion v3python/compile.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,7 @@

from triton.backends.compiler import GPUTarget

KNOWN_TARGETS_64 = ['gfx90a', 'gfx942', 'gfx950']
KNOWN_TARGETS_64 = ['gfx908', 'gfx90a', 'gfx942', 'gfx950']
KNOWN_TARGETS_32 = ['gfx1100', 'gfx1101', 'gfx1102', 'gfx1103', 'gfx1150', 'gfx1151', 'gfx1200', 'gfx1201', 'gfx1250']

KNOWN_TARGETS = {
Expand Down
4 changes: 4 additions & 0 deletions v3python/gpu_targets.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@
from collections import defaultdict

AOTRITON_SUPPORTED_GPUS = (
'gfx908_mod0',
'gfx90a_mod0',
'gfx942_mod0',
# 'gfx942_mod1',
Expand Down Expand Up @@ -33,6 +34,7 @@
}

AOTRITON_ARCH_TO_PACK = {
'gfx908' : 'gfx908',
'gfx90a' : 'gfx90a',
'gfx942' : 'gfx942',
'gfx950' : 'gfx950',
Expand All @@ -52,6 +54,7 @@
}

AOTRITON_ARCH_WARPSIZE = {
'gfx908' : 64,
'gfx90a' : 64,
'gfx942' : 64,
'gfx950' : 64,
Expand All @@ -67,6 +70,7 @@
}

AOTRITON_ARCH_PRODUCTION_LINE = {
'gfx908' : 'CDNA',
'gfx90a' : 'CDNA',
'gfx942' : 'CDNA',
'gfx950' : 'CDNA',
Expand Down
4 changes: 3 additions & 1 deletion v3src/util.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,6 +46,7 @@ struct LazyGpu {
};

std::unordered_map<std::string, GpuClassifier> LazyGpu::string_to_classifier = {
{ "gfx908", DummyClassifier<GPU_AMD_ARCH_GFX908_MOD0 >() },
{ "gfx90a", DummyClassifier<GPU_AMD_ARCH_GFX90A_MOD0 >() },
{ "gfx942", DummyClassifier<GPU_AMD_ARCH_GFX942_MOD0 >() },
{"gfx1100", DummyClassifier<GPU_AMD_ARCH_GFX1100_MOD0>() },
Expand Down Expand Up @@ -74,7 +75,8 @@ getGpuFromStream(hipStream_t stream) {
bool isArchExperimentallySupported(hipStream_t stream) {
auto gpu = getGpuFromStream(stream);
uint32_t vendor_arch = Gpu2VendorArch(gpu);
return (vendor_arch == CAT32(GpuVendor::kAMD, 0x1150) ||
return (vendor_arch == CAT32(GpuVendor::kAMD, 0x908) ||
vendor_arch == CAT32(GpuVendor::kAMD, 0x1150) ||
vendor_arch == CAT32(GpuVendor::kAMD, 0x1151) ||
vendor_arch == CAT32(GpuVendor::kAMD, 0x1100) ||
vendor_arch == CAT32(GpuVendor::kAMD, 0x1101) ||
Expand Down