[IREE dispatch profiler] Separate iree build folder and generated folder (#13536)
diff --git a/experimental/dispatch_profiler/README.md b/experimental/dispatch_profiler/README.md
index 5be486c..c09152d 100644
--- a/experimental/dispatch_profiler/README.md
+++ b/experimental/dispatch_profiler/README.md
@@ -18,7 +18,7 @@
IREE dispatch profiler provides [`generator.py`](generator.py) that can be used to generate dispatches. Please find a sample run below:
```bash
-build-debug $ python3 ../iree/experimental/dispatch_profiler/generator.py
+$ python3 dispatch_profiler/generator.py --generated-dir </path/to/create/`generated`/dir>
[Generating]: ./generated/linalg/matmul/matmul_128x128x256_f16t_f16t_f16t/matmul_128x128x256_f16t_f16t_f16t.mlir
Emitting tuning configuration : tile_config_128x128_64x4_tensorcore_mmasync
Emitting tuning configuration : tile_config_128x128_32x5_tensorcore_mmasync
@@ -35,7 +35,7 @@
#### Generating user-specified matmul shape `768x512x1024`
```bash
-python3 ../iree/experimental/dispatch_profiler/generator.py --problem-m=768 --problem-n=512 --problem-k=1024
+python3 ../iree/experimental/dispatch_profiler/generator.py --generated-dir </path/to/create/`generated`/dir> --problem-m=768 --problem-n=512 --problem-k=1024
...
[Generating]: ./generated/linalg/matmul/matmul_768x512x1024_f16t_f16t_f16t/matmul_768x512x1024_f16t_f16t_f16t.mlir
[Generating]: ./generated/linalg/matmul/matmul_768x512x1024_f32t_f32t_f32t/matmul_768x512x1024_f32t_f32t_f32t.mlir
@@ -47,7 +47,7 @@
Generate matmuls where M ranges from 64 to 1024 in increments of 128, N varies from 64 to 1024 in steps of 128, and K is fixed at 4096.
```bash
-$ python3 ../iree/experimental/dispatch_profiler/generator.py --problem-m=64:1024:128 --problem-n=64:1024:128 --problem-k=4096
+$ python3 ../iree/experimental/dispatch_profiler/generator.py --generated-dir </path/to/create/`generated`/dir> --problem-m=64:1024:128 --problem-n=64:1024:128 --problem-k=4096
...
```
@@ -56,7 +56,7 @@
IREE dispatch profiler provies `compile.py` that trigges `iree-compile` with appropiate compilation flags. The output of `iree-compile` vmfb files are placed in `mlir_dialect/operation_path/operation_name.mlir`. The `compiler.py` uses all the possible cpus on your machine to compile all different generated mlir source files.
```bash
-python3 ../iree/experimental/dispatch_profiler/compile.py
+python3 ../iree/experimental/dispatch_profiler/compile.py --build-dir </path/to/iree/build/dir> --generated-dir </path/to/create/`generated`/dir>
```
Compiles all the generated source mlir dispatches. One can check the generated dispatched folder to find the vmfb files.
@@ -73,7 +73,7 @@
### Functional verification and performance profiling of a _single_ dispatch
```
-$ python3 ../iree/experimental/dispatch_profiler/profiler.py --dispatches=matmul_3456x1024x2048_f16t_f16t_f16t_tile_config_128x128_32x5_tensorcore_mmasync --verification-enabled=true --profiling-enabled=true
+$ python3 profiler.py --build-dir </path/to/iree/build/dir> --generated-dir </path/to/create/`generated`/dir> --dispatches=matmul_3456x1024x2048_f16t_f16t_f16t_tile_config_128x128_32x5_tensorcore_mmasync --verification-enabled=true --profiling-enabled=true
----------------------------------------------------------------
Dispatch : matmul_3456x1024x2048_f16t_f16t_f16t_tile_config_128x128_32x5_tensorcore_mmasync
Provider : IREE Codegen
@@ -92,7 +92,7 @@
Verification, particularly for large matrix multiplications, can be time-consuming when using a CPU-based numpy reference. To prioritize profiling speed and when functional correctness is assured, disable verification using `--verification-enabled=false`.
```bash
- python3 ../iree/experimental/dispatch_profiler/profiler.py --dispatches=matmul_3456x1024x2048_f16t_f16t_f16t_tile_config_128x128_32x5_tensorcore_mmasync --verification-enabled=false --profiling-enabled=true
+python3 profiler.py --build-dir </path/to/iree/build/dir> --generated-dir </path/to/create/`generated`/dir> --dispatches=matmul_3456x1024x2048_f16t_f16t_f16t_tile_config_128x128_32x5_tensorcore_mmasync --verification-enabled=false --profiling-enabled=true
```
### Performance profile _single_ operation and _sweep_ tunning configurations
@@ -100,7 +100,7 @@
The `--dispatch` option accepts a comma-separated list of regex patterns to profile all tuning configurations generated for a operation. The command-line argument is formatted as `--dispatch=<regex>,<regex>`. Additionally, you can export the profiled output to a CSV file for further analysis using `--output=<filepath>`.
```bash
-$ python3 ../iree/experimental/dispatch_profiler/profiler.py --dispatches=matmul_3456x1024x2048_f16t_f16t_f16t_*_tensorcore_mmasync --verification-enabled=false --profiling-enabled=true --output=data.csv
+$ python3 profiler.py --build-dir </path/to/iree/build/dir> --generated-dir </path/to/create/`generated`/dir> --dispatches=matmul_3456x1024x2048_f16t_f16t_f16t_*_tensorcore_mmasync --verification-enabled=false --profiling-enabled=true --output=data.csv
----------------------------------------------------------------
Dispatch : matmul_3456x1024x2048_f16t_f16t_f16t_tile_config_128x256_32x3_tensorcore_mmasync
Provider : IREE Codegen
@@ -146,7 +146,7 @@
Another example showcasing the use of `--dispatch` to profile a matmul_3456x1024x2048 targetting F16 and F32 NVIDIA A100 Tensor Cores.
```bash
-$ python3 ../iree/experimental/dispatch_profiler/profiler.py --dispatches=matmul_3456x1024x2048_f16t_f16t_f16t_tile_config_128x128_32x5_tensorcore_mmasync,matmul_3456x1024x2048_f32t_f32t_f32t_tile_config_128x128_16x5_tensorcore_mmasync
+$ python3 profiler.py --build-dir </path/to/iree/build/dir> --generated-dir </path/to/create/`generated`/dir> --dispatches=matmul_3456x1024x2048_f16t_f16t_f16t_tile_config_128x128_32x5_tensorcore_mmasync,matmul_3456x1024x2048_f32t_f32t_f32t_tile_config_128x128_16x5_tensorcore_mmasync
----------------------------------------------------------------
Dispatch : matmul_3456x1024x2048_f16t_f16t_f16t_tile_config_128x128_32x5_tensorcore_mmasync
Provider : IREE Codegen
diff --git a/experimental/dispatch_profiler/launchers.py b/experimental/dispatch_profiler/launchers.py
index 968f991..0ea8a30 100644
--- a/experimental/dispatch_profiler/launchers.py
+++ b/experimental/dispatch_profiler/launchers.py
@@ -17,7 +17,8 @@
def __init__(self, args, operation):
self.operation = operation
- self.generated_path = Path(args.build_dir, 'generated', args.mlir_dialect)
+ self.generated_path = Path(args.generated_dir, 'generated',
+ args.mlir_dialect)
self.args = args
self.benchmark_dispatch_repeat_count = args.batch_size
@@ -31,7 +32,7 @@
operation.name()).with_suffix(".mlir")
# path to cached numpy refernece input and expected output files.
- self.op_reference_cache_path = Path(args.build_dir, 'generated',
+ self.op_reference_cache_path = Path(args.generated_dir, 'generated',
'reference_cache', operation.name())
if not self.op_reference_cache_path.exists():
diff --git a/experimental/dispatch_profiler/manifest.py b/experimental/dispatch_profiler/manifest.py
index 6f29e69..8fa1263 100644
--- a/experimental/dispatch_profiler/manifest.py
+++ b/experimental/dispatch_profiler/manifest.py
@@ -118,7 +118,7 @@
self.dispatch_names = [x for x in args.dispatches.split(',') if x != '']
# Paths to the generated directory (e.g. `./generated/linalg`).
- self.generated_path = Path(self.args.build_dir, 'generated',
+ self.generated_path = Path(self.args.generated_dir, 'generated',
self.args.mlir_dialect)
# Create the directories in self.generated_path, if it does not exist.
diff --git a/experimental/dispatch_profiler/options.py b/experimental/dispatch_profiler/options.py
index 5f6039a..521888b 100644
--- a/experimental/dispatch_profiler/options.py
+++ b/experimental/dispatch_profiler/options.py
@@ -22,9 +22,14 @@
def add_typical_arguments(parser):
"""Adds typical command line arguments to the parser."""
parser.add_argument("--build-dir", default=".", \
- help="IREE top-level build directory is used to generate "\
- "operations and npy files.This should be same that used "\
- "to call generated.py")
+ help="IREE top-level build directory. Looks for "\
+ "`iree/build/tools` dir in the build directory. "\
+ "The dispatch profiler scripts uses iree-compile, "\
+ "iree-benchmark-module from `build_dir` directory.")
+ parser.add_argument("--generated-dir", default=".", \
+ help="The dispatch profiler scripts generate "\
+ "mlir dispatches, compiled vmfbs, and reference_chache "\
+ "containing golden npy files in the generated-dir")
parser.add_argument("--operation-kind","--op-kind", \
dest="operation_kind", default="all", \
help="Specifies the operation kinds to generate.", \
