github/AMD-SHARK-Studio
1
1
Fork 0
mirror of https://github.com/nod-ai/AMD-SHARK-Studio.git synced 2026-04-03 03:00:17 -04:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'main' into ean-turbine-gen

Browse Source
This commit is contained in:
Ean Garvey
2023-11-30 17:44:38 -06:00
committed by GitHub
parent d5fd813db8 795fc33001
commit 678d851cfe
4 changed files with 18 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
apps/stable_diffusion/src/utils/resources/opt_flags.json
View File

@@ -59,24 +59,28 @@
"tuned": {
"fp16": {
"default_compilation_flags": [
"--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-preprocessing-pad-linalg-ops{pad-size=16}))"
"--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-preprocessing-pad-linalg-ops{pad-size=16}))",
"--iree-opt-data-tiling=False"
]
},
"fp32": {
"default_compilation_flags": [
"--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-preprocessing-pad-linalg-ops{pad-size=16}))"
"--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-preprocessing-pad-linalg-ops{pad-size=16}))",
"--iree-opt-data-tiling=False"
]
}
},
"untuned": {
"fp16": {
"default_compilation_flags": [
"--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-preprocessing-pad-linalg-ops{pad-size=16}))"
"--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-preprocessing-pad-linalg-ops{pad-size=16}))",
"--iree-opt-data-tiling=False"
]
},
"fp32": {
"default_compilation_flags": [
"--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-preprocessing-pad-linalg-ops{pad-size=16}))"
"--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-preprocessing-pad-linalg-ops{pad-size=16}))",
"--iree-opt-data-tiling=False"
]
}
}

7
apps/stable_diffusion/src/utils/stable_args.py
View File

@@ -587,6 +587,13 @@ p.add_argument(
help="Controls constant folding in iree-compile for all SD models.",
)
p.add_argument(
"--data_tiling",
default=False,
action=argparse.BooleanOptionalAction,
help="Controls data tiling in iree-compile for all SD models.",
)
##############################################################################
# Web UI flags
##############################################################################

2
apps/stable_diffusion/src/utils/utils.py
View File

@@ -541,6 +541,8 @@ def get_opt_flags(model, precision="fp16"):
iree_flags.append(
"--iree-codegen-linalg-max-constant-fold-elements=9223372036854775807"
)
if args.data_tiling == False:
iree_flags.append("--iree-opt-data-tiling=False")
if "default_compilation_flags" in opt_flags[model][is_tuned][precision]:
iree_flags += opt_flags[model][is_tuned][precision][

4
shark/iree_utils/compile_utils.py
View File

@@ -36,13 +36,11 @@ def get_iree_device_args(device, extra_args=[]):
if "cpu" in device:
from shark.iree_utils.cpu_utils import get_iree_cpu_args
data_tiling_flag = ["--iree-opt-data-tiling"]
u_kernel_flag = ["--iree-llvmcpu-enable-microkernels"]
u_kernel_flag = ["--iree-llvmcpu-enable-ukernels"]
stack_size_flag = ["--iree-llvmcpu-stack-allocation-limit=256000"]
return (
get_iree_cpu_args()
+ data_tiling_flag
+ u_kernel_flag
+ stack_size_flag
+ ["--iree-global-opt-enable-quantized-matmul-reassociation"]