Files
tinygrad/openpilot/compile.py
2022-09-05 20:14:31 -07:00

340 lines
12 KiB
Python

#!/usr/bin/env python3
import pathlib, sys
sys.path.insert(0, str(pathlib.Path(__file__).parent.parent))
from collections import defaultdict
import pyopencl as cl
import os
import time
import io
os.environ['OPT'] = '99'
if os.getenv("GPU", None) is None:
os.environ['OPENCL'] = '1'
DEBUGCL = int(os.getenv("DEBUGCL", 0))
import onnx
import numpy as np
import tinygrad.ops as ops
from tinygrad.llops.ops_gpu import CL, CLProgram, CLBuffer
from extra.utils import fetch
from extra.onnx import get_run_onnx
from tinygrad.tensor import Tensor
from tinygrad.helpers import prod
OPENPILOT_MODEL = "https://github.com/commaai/openpilot/raw/ea449f1fe0bbff0eff5b12d64f0b5e75b7983998/selfdrive/modeld/models/supercombo.onnx"
np.random.seed(1337)
def get_random_input_tensors():
np_inputs = {
"input_imgs": np.random.randn(*(1, 12, 128, 256))*256,
"big_input_imgs": np.random.randn(*(1, 12, 128, 256))*256,
"desire": np.zeros((1, 8)),
"traffic_convention": np.array([[1., 0.]]),
"initial_state": np.random.randn(*(1, 512))
#"initial_state": np.zeros((1, 768))
}
#import pickle
#frames, big_frames, last_state, frame_inputs, policy_outs = pickle.load(open("openpilot/test/frame_0.pkl", "rb"))
#np_inputs["input_imgs"] = frames
#np_inputs["big_input_imgs"] = big_frames
#np_inputs["initial_state"] = last_state[0]
#for i,k in enumerate(np_inputs.keys()):
# dat = open("/home/batman/openpilot/xx/ml_tools/snpe/compile_test_data/dlc_input_%d" % i, "rb").read()
# np_inputs[k] = np.frombuffer(dat, np.float32).reshape(np_inputs[k].shape)
np_inputs = {k:v.astype(np.float32) for k,v in np_inputs.items()}
inputs = {k:Tensor(v.astype(np.float32), requires_grad=False) for k,v in np_inputs.items()}
for _,v in inputs.items(): v.realize()
return inputs, np_inputs
def compile(input, output_fn):
Tensor.no_grad = True
using_graph = ops.GRAPH
ops.GRAPH = False
inputs, _ = get_random_input_tensors()
if os.getenv("TEST_ENET", None) is not None:
from models.efficientnet import EfficientNet
Tensor.training = False
enet = EfficientNet(number=int(os.getenv("TEST_ENET", None)), has_se=False, input_channels=12, has_fc_output=False)
def run_onnx(x):
return {"outputs": enet.forward(x['input_imgs'])}
else:
onnx_model = onnx.load(io.BytesIO(dat))
run_onnx = get_run_onnx(onnx_model)
# initial run(s) to load weights
for _ in range(2):
st = time.monotonic()
tinygrad_out = run_onnx(inputs)['outputs']
mt = time.monotonic()
tinygrad_out.realize()
mt2 = time.monotonic()
tinygrad_out = tinygrad_out.numpy()
et = time.monotonic()
print(f"ran openpilot model in {(et-st)*1000.0:.2f} ms, waited {(mt2-mt)*1000.0:.2f} ms for realize, {(et-mt2)*1000.0:.2f} ms for GPU queue")
# realize all non GCed tensors (fix for batchnorm folding)
import gc
gc.collect()
for x in [x for x in gc.get_objects() if isinstance(x, Tensor)]:
x.realize()
# real run
inputs, np_inputs = get_random_input_tensors()
tinygrad_out = run_onnx(inputs)['outputs']
CL.CACHE = []
if using_graph: ops.GRAPH = True
CL.kernel_count = -1
tinygrad_out.realize()
ops.GRAPH = False
print("kernel count:", len(CL.CACHE))
# optimize local workgroups
OPTWG = int(os.getenv("OPTWG", 0))
if OPTWG:
MAX_WORKGROUP = CL.cl_ctx.devices[0].max_work_group_size
local_cl_cache = []
for i, (prg, args) in enumerate(CL.CACHE):
args = list(args)
if args[1] is None and len(args[0]) == 2:
args[1] = [min(MAX_WORKGROUP, args[0][0]), 1]
try:
e = prg.clprg(CL().cl_queue, *args)
except cl.LogicError:
# INVALID_WORK_GROUP_SIZE
args[1] = None
continue
if OPTWG == 2 and args[0][0] % args[1][0] != 0:
args[1] = None
if args[1] is None and len(args[0]) == 3:
"""
if args[0][1] == 1 and args[0][2] == 1:
args[1] = [min(1024, args[0][0]), 1, 1]
else:
args[1] = [1,min(16,args[0][1]),min(args[0][2], 4)]
args[1][0] = min(32, min(args[0][0], 1024 // (args[1][1] * args[1][2])))
"""
runtimes = []
for l2 in [16,args[0][1],MAX_WORKGROUP]:
for l3 in [4,16,args[0][2],MAX_WORKGROUP]:
for l1 in [max(1, MAX_WORKGROUP//(l2*l3)), args[0][0], 4, 16, MAX_WORKGROUP]:
if l1 > args[0][0] or l2 > args[0][1] or l3 > args[0][2]: continue
local_args = (l1, l2, l3)
if prod(local_args) > MAX_WORKGROUP: continue
args[1] = local_args
if OPTWG == 2:
bad = any(g%l != 0 for g,l in zip(args[0], args[1]))
if bad: continue
try:
e = prg.clprg(CL().cl_queue, *args)
except cl.LogicError:
# INVALID_WORK_GROUP_SIZE
continue
CL().cl_queue.finish()
runtime = e.profile.end - e.profile.start
#print(runtime, args[0], args[1])
runtimes.append((runtime, local_args))
#print(sorted(runtimes)[0:5])
if len(runtimes) > 0:
args[1] = sorted(runtimes)[0][1]
else:
args[1] = None
print("couldn't optimize", args[0])
local_cl_cache.append((prg, args))
else:
local_cl_cache = CL.CACHE[:]
CL.CACHE = None
# real CL ish
for j in range(1):
events = []
st = time.monotonic()
for i, (prg, args) in enumerate(local_cl_cache):
#print(args)
events.append(prg.clprg(CL().cl_queue, *args))
mt = time.monotonic()
CL().cl_queue.finish()
et = time.monotonic()
print(f"submit in {(mt-st)*1000.0:.2f} ms, total runtime is {(et-st)*1000.0:.2f} ms")
total_runtime = 0
runtimes = defaultdict(float)
print()
for i, ((prg, args), e) in enumerate(zip(local_cl_cache, events)):
# profile types https://www.khronos.org/registry/OpenCL/sdk/1.0/docs/man/xhtml/clGetEventProfilingInfo.html
runtime = (e.profile.end - e.profile.start)
if sys.platform == "darwin": runtime *= 45
runtimes[prg.name.rsplit("_", 1)[0]] += runtime
if DEBUGCL:
print(f"{i:3d} time {total_runtime/1e6:5.2f} ms running {prg.name:20s} with {str(args[0]):15s} {str(args[1]):15s} count {len(args)-2:2d} runtime {runtime/1e3:7.2f} us {prg.options}")
if DEBUGCL >=2 and prg.name == "elementwise_166": print(prg.prg)
#if prg.name == "matmul": print(f" {args[3].shape} {args[4].shape} -> {args[5].shape}")
total_runtime += runtime
for k,v in runtimes.items():
print(f"{k:20s} runtime: {v/1e6:.2f} ms")
print(f"total runtime: {total_runtime/1e6:.2f} ms")
tinygrad_out_np = tinygrad_out.numpy()
# float32 only
FLOAT16 = int(os.getenv("FLOAT16", 0))
if FLOAT16 == 0:
try:
from test.test_onnx import run_onnx_torch
torch_out = run_onnx_torch(onnx_model, np_inputs).numpy()
print(tinygrad_out_np, torch_out, "mse", np.sum((tinygrad_out_np-torch_out)**2), "max err", np.max(np.abs((tinygrad_out_np-torch_out))))
np.testing.assert_allclose(torch_out, tinygrad_out_np, atol=1e-4, rtol=1e-2)
except ModuleNotFoundError:
pass
# save local_cl_cache as thneed
import struct, json
jdat = {"binaries": [], "programs": {}, "kernels": [], "objects": []}
weights = []
binaries = []
saved_objs = set()
saved_binaries = set()
kernels_to_save = set()
kernels_to_not_save = set([x.lazydata.realized.cl for x in inputs.values()])
for self, args in local_cl_cache:
# output is always the first parameter
kernels_to_not_save.add(args[2])
for a in args[3:]:
kernels_to_save.add(a)
kernels_to_save -= kernels_to_not_save
gobj = 0
for self, args in local_cl_cache:
#if self.name not in jdat['programs']:
# jdat['programs'][self.name] = {"src": self.prg, "options": ' '.join(self.options)}
if self.name not in saved_binaries:
binary = self.clprogram.get_info(cl.program_info.BINARIES)
assert len(binary) == 1
jdat['binaries'].append({"name":self.name, "length":len(binary[0])})
binaries.append(binary[0])
saved_binaries.add(self.name)
targs, args_size = [], []
argdtypes = self.argdtypes if self.argdtypes is not None else [None]*(len(args)-2)
for a,d in zip(args[2:], argdtypes):
if d == np.int16:
targs.append(struct.pack("H", a).decode("latin_1"))
args_size.append(2)
elif d == np.int32:
targs.append(struct.pack("I", a).decode("latin_1"))
args_size.append(4)
elif isinstance(a, cl.LocalMemory):
targs.append("")
args_size.append(a.size)
elif d is None:
if getattr(a, "global_id", None) is None:
setattr(a, "global_id", gobj)
gobj += 1
ptr = struct.pack("Q", a.global_id).decode("latin_1")
if ptr not in saved_objs:
if isinstance(a, cl.Buffer):
needs_load = a in kernels_to_save
jdat['objects'].append({
"id": ptr, "arg_type": "float*", "needs_load": needs_load, "size": a.size,
})
if needs_load:
data = np.empty(a.size//4, dtype=np.float32)
CL.enqueue_copy(data, a, is_blocking=True)
weights.append(data.tobytes())
elif isinstance(a, cl.Image):
needs_load = a in kernels_to_save
row_pitch = (a.shape[0]*4*(2 if FLOAT16 else 4) + 63)//64 * 64
size = row_pitch * a.shape[1]
# this is *2 if float16 and *4 if float32
buf = CLBuffer(size * (2 if FLOAT16 else 1))
# zero out the buffer
zeros = np.zeros(size, dtype=np.uint8)
CL.enqueue_copy(buf.cl, zeros, is_blocking=True)
CLProgram("from_image_strided", """
__kernel void from_image_strided(read_only image2d_t in, __global float4 *out, int row_pitch) {
const sampler_t smp = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
int2 l;
l.y = get_global_id(1);
l.x = get_global_id(0);
out[l.y*row_pitch + l.x] = read_imagef(in, smp, l);
}
""", argdtypes=(None, None, np.int32))(a.shape, None, a, buf.cl, row_pitch//(4*(2 if FLOAT16 else 4)))
# multiple of 32 isn't enough
jdat['objects'].append({
"id": ptr, "needs_load": needs_load, "size": size, "arg_type": "image2d_t",
"width": a.shape[0], "height": a.shape[1], "row_pitch": row_pitch, "float32": not FLOAT16,
})
if needs_load:
data = np.empty(size//(2 if FLOAT16 else 4), dtype=np.float32)
CL.enqueue_copy(data, buf.cl, is_blocking=True)
if FLOAT16: data = data.astype(np.float16)
weights.append(data.tobytes())
else:
raise Exception("unknown object", a)
#print(jdat['objects'][-1])
saved_objs.add(ptr)
targs.append(ptr)
args_size.append(8)
else:
raise Exception("idk this type")
jdat['kernels'].append({
"name": self.name,
"work_dim": len(args[0]),
"global_work_size": args[0],
"local_work_size": [1 for x in args[0]] if args[1] is None else args[1],
"num_args": len(args)-2,
"args": targs,
"args_size": args_size
})
jdat['outputs'] = [{
"buffer_id": struct.pack("Q", tinygrad_out.lazydata.realized.cl.global_id).decode("latin_1"),
"size": tinygrad_out.lazydata.realized.cl.size,
}]
print(jdat['outputs'])
jdat['inputs'] = [{
"buffer_id": struct.pack("Q", v.lazydata.realized.cl.global_id).decode("latin_1"),
#"size": v.lazydata.realized.cl.size,
"size": prod(v.shape)*4,
"name": k
} for k,v in inputs.items()][::-1]
print(jdat['inputs'])
print(f"saving {len([x for x in jdat['objects'] if x['needs_load']])} objects")
print("saving thneed")
with open(output_fn, "wb") as f:
j = json.dumps(jdat, ensure_ascii=False).encode('latin_1')
f.write(struct.pack("I", len(j)))
f.write(j)
f.write(b''.join(weights))
f.write(b''.join(binaries))
# OPTWG=1 UNSAFE_FLOAT4=1 DEBUGCL=1 FLOAT16=1 MATMUL=1 python3 openpilot/compile.py
# 22.59 ms
if __name__ == "__main__":
if len(sys.argv) >= 3:
with open(sys.argv[1], "rb") as f:
dat = f.read()
compile(dat, sys.argv[2])
else:
dat = fetch(OPENPILOT_MODEL)
compile(dat, "/tmp/output.thneed")