mawei/tinygrad
1
0
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-06-11 23:46:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

openpilot benchmark: add cast from numpy to benchmark (#7593)

Browse Source 
* openpilot benchmark: add cast from numpy to benchmark

* whitespace

* comment
This commit is contained in:
Harald Schäfer
2024-11-08 03:31:00 -08:00
committed by GitHub
parent d4e91b0de7
commit e7cbc29f48
2 changed files with 10 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
examples/openpilot/compile3.py
View File

@@ -58,9 +58,13 @@ def test(test_val=None):
Tensor.manual_seed(100)
new_inputs = {nm:Tensor.randn(*st.shape, dtype=dtype).mul(8).realize() for nm, (st, _, dtype, _) in
sorted(zip(run.captured.expected_names, run.captured.expected_st_vars_dtype_device))}
new_inputs_numpy = {k:v.numpy() for k,v in new_inputs.items()}
for _ in range(20):
st = time.perf_counter()
out = run(**new_inputs)
# Need to cast non-image inputs from numpy, this is only realistic way to run it
inputs = {**{k:v for k,v in new_inputs.items() if 'img' in k},
**{k:Tensor(v) for k,v in new_inputs_numpy.items() if 'img' not in k}}
out = run(**inputs)
mt = time.perf_counter()
val = out['outputs'].numpy()
et = time.perf_counter()

6
test/external/external_benchmark_openpilot.py vendored
View File

@@ -22,6 +22,7 @@ if __name__ == "__main__":
input_types = {inp.name: tensor_dtype_to_np_dtype(inp.type.tensor_type.elem_type) for inp in onnx_model.graph.input}
new_inputs = {k:Tensor.randn(*shp, dtype=_from_np_dtype(input_types[k])).mul(8).realize() for k,shp in input_shapes.items()}
new_inputs_junk = {k:Tensor.randn(*shp, dtype=_from_np_dtype(input_types[k])).mul(8).realize() for k,shp in input_shapes.items()}
new_inputs_junk_numpy = {k:v.numpy() for k,v in new_inputs_junk.items()}
# benchmark
for _ in range(5):
@@ -35,7 +36,10 @@ if __name__ == "__main__":
for _ in range(20):
GlobalCounters.reset()
st = time.perf_counter_ns()
ret = next(iter(run_onnx_jit(**new_inputs_junk).values())).cast(dtypes.float32).numpy()
# Need to cast non-image inputs from numpy, this is only realistic way to run model
inputs = {**{k:v for k,v in new_inputs_junk.items() if 'img' in k},
**{k:Tensor(v) for k,v in new_inputs_junk_numpy.items() if 'img' not in k}}
ret = next(iter(run_onnx_jit(**inputs).values())).cast(dtypes.float32).numpy()
print(f"jitted: {(time.perf_counter_ns() - st)*1e-6:7.4f} ms")
suffix = ""