From d9d57e5d6f8c040172a771f8be356d1f0586fd25 Mon Sep 17 00:00:00 2001 From: Adeeb Shihadeh Date: Wed, 13 Nov 2024 19:27:11 -0800 Subject: [PATCH] Revert ISP image processing + tinygrad bump (#34020) * Revert "Replace ThneedModel with TinygradModel (#33532)" This reverts commit da952e9b6418b07df27a51de6629d8f4127cb03c. * Revert "camerad: move E + D cams image pipelines to the IFE (#33959)" This reverts commit f2a1cce42b47306a9551cbfa65c5edbbaec037ae. --- release/release_files.py | 2 +- selfdrive/modeld/SConscript | 37 ++- selfdrive/modeld/dmonitoringmodeld | 6 + selfdrive/modeld/dmonitoringmodeld.py | 47 ++-- selfdrive/modeld/modeld.py | 68 ++--- selfdrive/modeld/models/commonmodel.cc | 27 +- selfdrive/modeld/models/commonmodel.h | 7 +- selfdrive/modeld/models/commonmodel.pxd | 3 +- selfdrive/modeld/models/commonmodel_pyx.pyx | 26 +- selfdrive/modeld/runners/__init__.py | 10 +- selfdrive/modeld/runners/runmodel_pyx.pyx | 12 +- selfdrive/modeld/runners/thneedmodel.cc | 58 ++++ selfdrive/modeld/runners/thneedmodel.h | 17 ++ selfdrive/modeld/runners/thneedmodel.pxd | 9 + selfdrive/modeld/runners/thneedmodel_pyx.pyx | 14 + selfdrive/modeld/runners/tinygrad_helpers.py | 8 - selfdrive/modeld/thneed/README | 8 + selfdrive/modeld/thneed/__init__.py | 0 selfdrive/modeld/thneed/serialize.cc | 154 +++++++++++ selfdrive/modeld/thneed/thneed.h | 133 +++++++++ selfdrive/modeld/thneed/thneed_common.cc | 216 +++++++++++++++ selfdrive/modeld/thneed/thneed_pc.cc | 32 +++ selfdrive/modeld/thneed/thneed_qcom2.cc | 258 ++++++++++++++++++ selfdrive/test/test_onroad.py | 12 +- system/camerad/cameras/camera_qcom2.cc | 2 +- system/hardware/tici/tests/test_power_draw.py | 2 +- tinygrad_repo | 2 +- 27 files changed, 1020 insertions(+), 150 deletions(-) create mode 100644 selfdrive/modeld/runners/thneedmodel.cc create mode 100644 selfdrive/modeld/runners/thneedmodel.h create mode 100644 selfdrive/modeld/runners/thneedmodel.pxd create mode 100644 selfdrive/modeld/runners/thneedmodel_pyx.pyx delete mode 100644 selfdrive/modeld/runners/tinygrad_helpers.py create mode 100644 selfdrive/modeld/thneed/README create mode 100644 selfdrive/modeld/thneed/__init__.py create mode 100644 selfdrive/modeld/thneed/serialize.cc create mode 100644 selfdrive/modeld/thneed/thneed.h create mode 100644 selfdrive/modeld/thneed/thneed_common.cc create mode 100644 selfdrive/modeld/thneed/thneed_pc.cc create mode 100644 selfdrive/modeld/thneed/thneed_qcom2.cc diff --git a/release/release_files.py b/release/release_files.py index 2802f10a2d..afd0d468b6 100755 --- a/release/release_files.py +++ b/release/release_files.py @@ -55,7 +55,7 @@ whitelist = [ "tools/joystick/", "tools/longitudinal_maneuvers/", - "tinygrad_repo/examples/openpilot/compile3.py", + "tinygrad_repo/openpilot/compile2.py", "tinygrad_repo/extra/onnx.py", "tinygrad_repo/extra/onnx_ops.py", "tinygrad_repo/extra/thneed.py", diff --git a/selfdrive/modeld/SConscript b/selfdrive/modeld/SConscript index 54eeb4aa02..d472998416 100644 --- a/selfdrive/modeld/SConscript +++ b/selfdrive/modeld/SConscript @@ -13,6 +13,15 @@ common_src = [ "transforms/transform.cc", ] +thneed_src_common = [ + "thneed/thneed_common.cc", + "thneed/serialize.cc", +] + +thneed_src_qcom = thneed_src_common + ["thneed/thneed_qcom2.cc"] +thneed_src_pc = thneed_src_common + ["thneed/thneed_pc.cc"] +thneed_src = thneed_src_qcom if arch == "larch64" else thneed_src_pc + # SNPE except on Mac and ARM Linux snpe_lib = [] if arch != "Darwin" and arch != "aarch64": @@ -50,18 +59,20 @@ fn = File("models/supercombo").abspath cmd = f'python3 {Dir("#selfdrive/modeld").abspath}/get_model_metadata.py {fn}.onnx' lenv.Command(fn + "_metadata.pkl", [fn + ".onnx"] + tinygrad_files, cmd) -# Compile tinygrad model -# TODO this is all super hacky -pythonpath_string = 'PYTHONPATH="${PYTHONPATH}:' + env.Dir("#tinygrad_repo").abspath + '"' -if arch == 'larch64': - device_string = 'QCOM=1' -elif arch == 'Darwin' or arch == 'aarch64': - device_string = 'CLANG=1 IMAGE=0' -else: - device_string = 'GPU=1' +# Build thneed model +if arch == "larch64" or GetOption('pc_thneed'): + tinygrad_opts = [] + if not GetOption('pc_thneed'): + # use FLOAT16 on device for speed + don't cache the CL kernels for space + tinygrad_opts += ["FLOAT16=1", "PYOPENCL_NO_CACHE=1"] + cmd = f"cd {Dir('#').abspath}/tinygrad_repo && " + ' '.join(tinygrad_opts) + f" python3 openpilot/compile2.py {fn}.onnx {fn}.thneed" -for model_name in ['supercombo', 'dmonitoring_model']: - fn = File(f"models/{model_name}").abspath - cmd = f'{pythonpath_string} {device_string} python3 {Dir("#tinygrad_repo").abspath}/examples/openpilot/compile3.py {fn}.onnx {fn}_tinygrad.pkl' - lenv.Command(fn + "_tinygrad.pkl", [fn + ".onnx"] + tinygrad_files, cmd) + lenv.Command(fn + ".thneed", [fn + ".onnx"] + tinygrad_files, cmd) + fn_dm = File("models/dmonitoring_model").abspath + cmd = f"cd {Dir('#').abspath}/tinygrad_repo && " + ' '.join(tinygrad_opts) + f" python3 openpilot/compile2.py {fn_dm}.onnx {fn_dm}.thneed" + lenv.Command(fn_dm + ".thneed", [fn_dm + ".onnx"] + tinygrad_files, cmd) + + thneed_lib = env.SharedLibrary('thneed', thneed_src, LIBS=[gpucommon, common, 'OpenCL', 'dl']) + thneedmodel_lib = env.Library('thneedmodel', ['runners/thneedmodel.cc']) + lenvCython.Program('runners/thneedmodel_pyx.so', 'runners/thneedmodel_pyx.pyx', LIBS=envCython["LIBS"]+[thneedmodel_lib, thneed_lib, gpucommon, common, 'dl', 'OpenCL']) diff --git a/selfdrive/modeld/dmonitoringmodeld b/selfdrive/modeld/dmonitoringmodeld index 90b43800fe..80157e1751 100755 --- a/selfdrive/modeld/dmonitoringmodeld +++ b/selfdrive/modeld/dmonitoringmodeld @@ -1,4 +1,10 @@ #!/usr/bin/env bash DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" >/dev/null && pwd)" +cd "$DIR/../../" + +if [ -f "$DIR/libthneed.so" ]; then + export LD_PRELOAD="$DIR/libthneed.so" +fi + exec "$DIR/dmonitoringmodeld.py" "$@" diff --git a/selfdrive/modeld/dmonitoringmodeld.py b/selfdrive/modeld/dmonitoringmodeld.py index 7e69d3afde..31440c1295 100755 --- a/selfdrive/modeld/dmonitoringmodeld.py +++ b/selfdrive/modeld/dmonitoringmodeld.py @@ -1,16 +1,8 @@ #!/usr/bin/env python3 import os -from openpilot.system.hardware import TICI -## TODO this is hack -if TICI: - GPU_BACKEND = 'QCOM' -else: - GPU_BACKEND = 'GPU' -os.environ[GPU_BACKEND] = '1' import gc import math import time -import pickle import ctypes import numpy as np from pathlib import Path @@ -22,11 +14,9 @@ from msgq.visionipc import VisionIpcClient, VisionStreamType, VisionBuf from openpilot.common.swaglog import cloudlog from openpilot.common.params import Params from openpilot.common.realtime import set_realtime_priority -from openpilot.selfdrive.modeld.models.commonmodel_pyx import CLContext #, cl_from_visionbuf +from openpilot.selfdrive.modeld.runners import ModelRunner, Runtime +from openpilot.selfdrive.modeld.models.commonmodel_pyx import CLContext from openpilot.selfdrive.modeld.parse_model_outputs import sigmoid -#from openpilot.selfdrive.modeld.runners.tinygrad_helpers import qcom_tensor_from_opencl_address -from tinygrad.tensor import Tensor -#from tinygrad.dtype import dtypes CALIB_LEN = 3 MODEL_WIDTH = 1440 @@ -36,7 +26,9 @@ OUTPUT_SIZE = 84 + FEATURE_LEN PROCESS_NAME = "selfdrive.modeld.dmonitoringmodeld" SEND_RAW_PRED = os.getenv('SEND_RAW_PRED') -MODEL_PKL_PATH = Path(__file__).parent / 'models/dmonitoring_model_tinygrad.pkl' +MODEL_PATHS = { + ModelRunner.THNEED: Path(__file__).parent / 'models/dmonitoring_model.thneed', + ModelRunner.ONNX: Path(__file__).parent / 'models/dmonitoring_model.onnx'} class DriverStateResult(ctypes.Structure): _fields_ = [ @@ -67,32 +59,33 @@ class DMonitoringModelResult(ctypes.Structure): class ModelState: inputs: dict[str, np.ndarray] output: np.ndarray + model: ModelRunner def __init__(self, cl_ctx): assert ctypes.sizeof(DMonitoringModelResult) == OUTPUT_SIZE * ctypes.sizeof(ctypes.c_float) - self.numpy_inputs = {'calib': np.zeros((1, CALIB_LEN), dtype=np.float32), - 'input_img': np.zeros((1,MODEL_HEIGHT * MODEL_WIDTH), dtype=np.uint8)} - self.img = None + self.output = np.zeros(OUTPUT_SIZE, dtype=np.float32) + self.inputs = { + 'input_img': np.zeros(MODEL_HEIGHT * MODEL_WIDTH, dtype=np.uint8), + 'calib': np.zeros(CALIB_LEN, dtype=np.float32)} - - with open(MODEL_PKL_PATH, "rb") as f: - self.model_run = pickle.load(f) + self.model = ModelRunner(MODEL_PATHS, self.output, Runtime.GPU, False, cl_ctx) + self.model.addInput("input_img", None) + self.model.addInput("calib", self.inputs['calib']) def run(self, buf:VisionBuf, calib:np.ndarray) -> tuple[np.ndarray, float]: - self.numpy_inputs['calib'][0,:] = calib + self.inputs['calib'][:] = calib - t1 = time.perf_counter() - # TODO use opencl buffer directly to make tensor v_offset = buf.height - MODEL_HEIGHT h_offset = (buf.width - MODEL_WIDTH) // 2 buf_data = buf.data.reshape(-1, buf.stride) - self.numpy_inputs['input_img'][:] = buf_data[v_offset:v_offset+MODEL_HEIGHT, h_offset:h_offset+MODEL_WIDTH].reshape((1, -1)) - - tensor_inputs = {k: Tensor(v) for k,v in self.numpy_inputs.items()} - output = self.model_run(**tensor_inputs)['outputs'].numpy().flatten() + input_data = self.inputs['input_img'].reshape(MODEL_HEIGHT, MODEL_WIDTH) + input_data[:] = buf_data[v_offset:v_offset+MODEL_HEIGHT, h_offset:h_offset+MODEL_WIDTH] + self.model.setInputBuffer("input_img", self.inputs['input_img'].view(np.float32)) + t1 = time.perf_counter() + self.model.execute() t2 = time.perf_counter() - return output, t2 - t1 + return self.output, t2 - t1 def fill_driver_state(msg, ds_result: DriverStateResult): diff --git a/selfdrive/modeld/modeld.py b/selfdrive/modeld/modeld.py index e4690c6fe2..4e91d32400 100755 --- a/selfdrive/modeld/modeld.py +++ b/selfdrive/modeld/modeld.py @@ -1,12 +1,5 @@ #!/usr/bin/env python3 import os -from openpilot.system.hardware import TICI -## TODO this is hack -if TICI: - GPU_BACKEND = 'QCOM' -else: - GPU_BACKEND = 'GPU' -os.environ[GPU_BACKEND] = '1' import time import pickle import numpy as np @@ -25,24 +18,21 @@ from openpilot.common.transformations.camera import DEVICE_CAMERAS from openpilot.common.transformations.model import get_warp_matrix from openpilot.system import sentry from openpilot.selfdrive.controls.lib.desire_helper import DesireHelper +from openpilot.selfdrive.modeld.runners import ModelRunner, Runtime from openpilot.selfdrive.modeld.parse_model_outputs import Parser from openpilot.selfdrive.modeld.fill_model_msg import fill_model_msg, fill_pose_msg, PublishState from openpilot.selfdrive.modeld.constants import ModelConstants from openpilot.selfdrive.modeld.models.commonmodel_pyx import ModelFrame, CLContext -from openpilot.selfdrive.modeld.runners.tinygrad_helpers import qcom_tensor_from_opencl_address - -from tinygrad.tensor import Tensor -from tinygrad.dtype import dtypes PROCESS_NAME = "selfdrive.modeld.modeld" SEND_RAW_PRED = os.getenv('SEND_RAW_PRED') -MODEL_PATH = Path(__file__).parent / 'models/supercombo.onnx' -MODEL_PKL_PATH = Path(__file__).parent / 'models/supercombo_tinygrad.pkl' +MODEL_PATHS = { + ModelRunner.THNEED: Path(__file__).parent / 'models/supercombo.thneed', + ModelRunner.ONNX: Path(__file__).parent / 'models/supercombo.onnx'} + METADATA_PATH = Path(__file__).parent / 'models/supercombo_metadata.pkl' -# TODO: should not hardcoded -IMG_INPUT_SHAPE = (1, 12, 128, 256) class FrameMeta: frame_id: int = 0 @@ -59,6 +49,7 @@ class ModelState: inputs: dict[str, np.ndarray] output: np.ndarray prev_desire: np.ndarray # for tracking the rising edge of the pulse + model: ModelRunner def __init__(self, context: CLContext): self.frame = ModelFrame(context) @@ -69,14 +60,13 @@ class ModelState: self.prev_desired_curv_20hz = np.zeros((ModelConstants.FULL_HISTORY_BUFFER_LEN + 1, ModelConstants.PREV_DESIRED_CURV_LEN), dtype=np.float32) # img buffers are managed in openCL transform code - self.numpy_inputs = { - 'desire': np.zeros((1, (ModelConstants.HISTORY_BUFFER_LEN+1), ModelConstants.DESIRE_LEN), dtype=np.float32), - 'traffic_convention': np.zeros((1, ModelConstants.TRAFFIC_CONVENTION_LEN), dtype=np.float32), - 'lateral_control_params': np.zeros((1, ModelConstants.LATERAL_CONTROL_PARAMS_LEN), dtype=np.float32), - 'prev_desired_curv': np.zeros((1,(ModelConstants.HISTORY_BUFFER_LEN+1), ModelConstants.PREV_DESIRED_CURV_LEN), dtype=np.float32), - 'features_buffer': np.zeros((1, ModelConstants.HISTORY_BUFFER_LEN, ModelConstants.FEATURE_LEN), dtype=np.float32), + self.inputs = { + 'desire': np.zeros(ModelConstants.DESIRE_LEN * (ModelConstants.HISTORY_BUFFER_LEN+1), dtype=np.float32), + 'traffic_convention': np.zeros(ModelConstants.TRAFFIC_CONVENTION_LEN, dtype=np.float32), + 'lateral_control_params': np.zeros(ModelConstants.LATERAL_CONTROL_PARAMS_LEN, dtype=np.float32), + 'prev_desired_curv': np.zeros(ModelConstants.PREV_DESIRED_CURV_LEN * (ModelConstants.HISTORY_BUFFER_LEN+1), dtype=np.float32), + 'features_buffer': np.zeros(ModelConstants.HISTORY_BUFFER_LEN * ModelConstants.FEATURE_LEN, dtype=np.float32), } - self.img_inputs = {} # type: ignore with open(METADATA_PATH, 'rb') as f: model_metadata = pickle.load(f) @@ -86,8 +76,11 @@ class ModelState: self.output = np.zeros(net_output_size, dtype=np.float32) self.parser = Parser() - with open(MODEL_PKL_PATH, "rb") as f: - self.model_run = pickle.load(f) + self.model = ModelRunner(MODEL_PATHS, self.output, Runtime.GPU, False, context) + self.model.addInput("input_imgs", None) + self.model.addInput("big_input_imgs", None) + for k,v in self.inputs.items(): + self.model.addInput(k, v) def slice_outputs(self, model_outputs: np.ndarray) -> dict[str, np.ndarray]: parsed_model_outputs = {k: model_outputs[np.newaxis, v] for k,v in self.output_slices.items()} @@ -104,27 +97,18 @@ class ModelState: self.desire_20Hz[:-1] = self.desire_20Hz[1:] self.desire_20Hz[-1] = new_desire - self.numpy_inputs['desire'][:] = self.desire_20Hz.reshape((1,25,4,-1)).max(axis=2) + self.inputs['desire'][:] = self.desire_20Hz.reshape((25,4,-1)).max(axis=1).flatten() - self.numpy_inputs['traffic_convention'][:] = inputs['traffic_convention'] - self.numpy_inputs['lateral_control_params'][:] = inputs['lateral_control_params'] - input_imgs_cl = self.frame.prepare(buf, transform.flatten()) - big_input_imgs_cl = self.wide_frame.prepare(wbuf, transform_wide.flatten()) + self.inputs['traffic_convention'][:] = inputs['traffic_convention'] + self.inputs['lateral_control_params'][:] = inputs['lateral_control_params'] - if TICI: - # The imgs tensors are backed by opencl memory, only need init once - if 'input_imgs' not in self.img_inputs: - self.img_inputs['input_imgs'] = qcom_tensor_from_opencl_address(input_imgs_cl.mem_address, IMG_INPUT_SHAPE, dtype=dtypes.uint8) - self.img_inputs['big_input_imgs'] = qcom_tensor_from_opencl_address(big_input_imgs_cl.mem_address, IMG_INPUT_SHAPE, dtype=dtypes.uint8) - else: - self.img_inputs['input_imgs'] = Tensor(self.frame.buffer_from_cl(input_imgs_cl)).reshape(IMG_INPUT_SHAPE) - self.img_inputs['big_input_imgs'] = Tensor(self.wide_frame.buffer_from_cl(big_input_imgs_cl)).reshape(IMG_INPUT_SHAPE) + self.model.setInputBuffer("input_imgs", self.frame.prepare(buf, transform.flatten(), self.model.getCLBuffer("input_imgs"))) + self.model.setInputBuffer("big_input_imgs", self.wide_frame.prepare(wbuf, transform_wide.flatten(), self.model.getCLBuffer("big_input_imgs"))) - tensor_inputs = {**self.img_inputs, **{k: Tensor(v) for k,v in self.numpy_inputs.items()}} if prepare_only: return None - self.output = self.model_run(**tensor_inputs)['outputs'].numpy().flatten() + self.model.execute() outputs = self.parser.parse_outputs(self.slice_outputs(self.output)) self.full_features_20Hz[:-1] = self.full_features_20Hz[1:] @@ -134,9 +118,9 @@ class ModelState: self.prev_desired_curv_20hz[-1] = outputs['desired_curvature'][0, :] idxs = np.arange(-4,-100,-4)[::-1] - self.numpy_inputs['features_buffer'][:] = self.full_features_20Hz[idxs] + self.inputs['features_buffer'][:] = self.full_features_20Hz[idxs].flatten() # TODO model only uses last value now, once that changes we need to input strided action history buffer - self.numpy_inputs['prev_desired_curv'][-ModelConstants.PREV_DESIRED_CURV_LEN:] = 0. * self.prev_desired_curv_20hz[-4, :] + self.inputs['prev_desired_curv'][-ModelConstants.PREV_DESIRED_CURV_LEN:] = 0. * self.prev_desired_curv_20hz[-4, :] return outputs @@ -205,7 +189,7 @@ def main(demo=False): cloudlog.info("modeld got CarParams: %s", CP.carName) # TODO this needs more thought, use .2s extra for now to estimate other delays - steer_delay = .2 + steer_delay = CP.steerActuatorDelay + .2 DH = DesireHelper() diff --git a/selfdrive/modeld/models/commonmodel.cc b/selfdrive/modeld/models/commonmodel.cc index ef730e01aa..e8a5a7ed52 100644 --- a/selfdrive/modeld/models/commonmodel.cc +++ b/selfdrive/modeld/models/commonmodel.cc @@ -8,7 +8,6 @@ ModelFrame::ModelFrame(cl_device_id device_id, cl_context context) { input_frames = std::make_unique(buf_size); - input_frames_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, buf_size, NULL, &err)); q = CL_CHECK_ERR(clCreateCommandQueue(context, device_id, 0, &err)); y_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, MODEL_WIDTH * MODEL_HEIGHT, NULL, &err)); @@ -23,7 +22,7 @@ ModelFrame::ModelFrame(cl_device_id device_id, cl_context context) { loadyuv_init(&loadyuv, context, device_id, MODEL_WIDTH, MODEL_HEIGHT); } -cl_mem* ModelFrame::prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3 &projection) { +uint8_t* ModelFrame::prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3 &projection, cl_mem *output) { transform_queue(&this->transform, q, yuv_cl, frame_width, frame_height, frame_stride, frame_uv_offset, y_cl, u_cl, v_cl, MODEL_WIDTH, MODEL_HEIGHT, projection); @@ -32,19 +31,19 @@ cl_mem* ModelFrame::prepare(cl_mem yuv_cl, int frame_width, int frame_height, in CL_CHECK(clEnqueueCopyBuffer(q, img_buffer_20hz_cl, img_buffer_20hz_cl, (i+1)*frame_size_bytes, i*frame_size_bytes, frame_size_bytes, 0, nullptr, nullptr)); } loadyuv_queue(&loadyuv, q, y_cl, u_cl, v_cl, last_img_cl); + if (output == NULL) { + CL_CHECK(clEnqueueReadBuffer(q, img_buffer_20hz_cl, CL_TRUE, 0, frame_size_bytes, &input_frames[0], 0, nullptr, nullptr)); + CL_CHECK(clEnqueueReadBuffer(q, last_img_cl, CL_TRUE, 0, frame_size_bytes, &input_frames[MODEL_FRAME_SIZE], 0, nullptr, nullptr)); + clFinish(q); + return &input_frames[0]; + } else { + copy_queue(&loadyuv, q, img_buffer_20hz_cl, *output, 0, 0, frame_size_bytes); + copy_queue(&loadyuv, q, last_img_cl, *output, 0, frame_size_bytes, frame_size_bytes); - copy_queue(&loadyuv, q, img_buffer_20hz_cl, input_frames_cl, 0, 0, frame_size_bytes); - copy_queue(&loadyuv, q, last_img_cl, input_frames_cl, 0, frame_size_bytes, frame_size_bytes); - - // NOTE: Since thneed is using a different command queue, this clFinish is needed to ensure the image is ready. - clFinish(q); - return &input_frames_cl; -} - -uint8_t* ModelFrame::buffer_from_cl(cl_mem *in_frames) { - CL_CHECK(clEnqueueReadBuffer(q, *in_frames, CL_TRUE, 0, MODEL_FRAME_SIZE * 2 * sizeof(uint8_t), &input_frames[0], 0, nullptr, nullptr)); - clFinish(q); - return &input_frames[0]; + // NOTE: Since thneed is using a different command queue, this clFinish is needed to ensure the image is ready. + clFinish(q); + return NULL; + } } ModelFrame::~ModelFrame() { diff --git a/selfdrive/modeld/models/commonmodel.h b/selfdrive/modeld/models/commonmodel.h index 91cbbcddd3..1c7360f159 100644 --- a/selfdrive/modeld/models/commonmodel.h +++ b/selfdrive/modeld/models/commonmodel.h @@ -20,8 +20,7 @@ class ModelFrame { public: ModelFrame(cl_device_id device_id, cl_context context); ~ModelFrame(); - cl_mem* prepare(cl_mem yuv_cl, int width, int height, int frame_stride, int frame_uv_offset, const mat3& transform); - uint8_t* buffer_from_cl(cl_mem *in_frames); + uint8_t* prepare(cl_mem yuv_cl, int width, int height, int frame_stride, int frame_uv_offset, const mat3& transform, cl_mem *output); const int MODEL_WIDTH = 512; const int MODEL_HEIGHT = 256; @@ -33,7 +32,7 @@ private: Transform transform; LoadYUVState loadyuv; cl_command_queue q; - cl_mem y_cl, u_cl, v_cl, img_buffer_20hz_cl, last_img_cl, input_frames_cl; + cl_mem y_cl, u_cl, v_cl, img_buffer_20hz_cl, last_img_cl; cl_buffer_region region; std::unique_ptr input_frames; -}; +}; \ No newline at end of file diff --git a/selfdrive/modeld/models/commonmodel.pxd b/selfdrive/modeld/models/commonmodel.pxd index 676defe15c..3348af3f17 100644 --- a/selfdrive/modeld/models/commonmodel.pxd +++ b/selfdrive/modeld/models/commonmodel.pxd @@ -15,5 +15,4 @@ cdef extern from "selfdrive/modeld/models/commonmodel.h": cppclass ModelFrame: int buf_size ModelFrame(cl_device_id, cl_context) - cl_mem * prepare(cl_mem, int, int, int, int, mat3) - unsigned char * buffer_from_cl(cl_mem*); + unsigned char * prepare(cl_mem, int, int, int, int, mat3, cl_mem*) diff --git a/selfdrive/modeld/models/commonmodel_pyx.pyx b/selfdrive/modeld/models/commonmodel_pyx.pyx index e6ae349e00..99f9c5dc17 100644 --- a/selfdrive/modeld/models/commonmodel_pyx.pyx +++ b/selfdrive/modeld/models/commonmodel_pyx.pyx @@ -4,7 +4,6 @@ import numpy as np cimport numpy as cnp from libc.string cimport memcpy -from libc.stdint cimport uintptr_t from msgq.visionipc.visionipc cimport cl_mem from msgq.visionipc.visionipc_pyx cimport VisionBuf, CLContext as BaseCLContext @@ -24,13 +23,6 @@ cdef class CLMem: mem.mem = cmem return mem - @property - def mem_address(self): - return (self.mem) - -def cl_from_visionbuf(VisionBuf buf): - return CLMem.create(&buf.buf.buf_cl) - cdef class ModelFrame: cdef cppModelFrame * frame @@ -40,14 +32,14 @@ cdef class ModelFrame: def __dealloc__(self): del self.frame - def prepare(self, VisionBuf buf, float[:] projection): + def prepare(self, VisionBuf buf, float[:] projection, CLMem output): cdef mat3 cprojection memcpy(cprojection.v, &projection[0], 9*sizeof(float)) - cdef cl_mem * data - data = self.frame.prepare(buf.buf.buf_cl, buf.width, buf.height, buf.stride, buf.uv_offset, cprojection) - return CLMem.create(data) - - def buffer_from_cl(self, CLMem in_frames): - cdef unsigned char * data2 - data2 = self.frame.buffer_from_cl(in_frames.mem) - return np.asarray( data2) + cdef unsigned char * data + if output is None: + data = self.frame.prepare(buf.buf.buf_cl, buf.width, buf.height, buf.stride, buf.uv_offset, cprojection, NULL) + else: + data = self.frame.prepare(buf.buf.buf_cl, buf.width, buf.height, buf.stride, buf.uv_offset, cprojection, output.mem) + if not data: + return None + return np.asarray( data) diff --git a/selfdrive/modeld/runners/__init__.py b/selfdrive/modeld/runners/__init__.py index afcb97df52..4c29bf3f1c 100644 --- a/selfdrive/modeld/runners/__init__.py +++ b/selfdrive/modeld/runners/__init__.py @@ -3,18 +3,18 @@ from openpilot.system.hardware import TICI from openpilot.selfdrive.modeld.runners.runmodel_pyx import RunModel, Runtime assert Runtime -USE_TINYGRAD = int(os.getenv('USE_TINYGRAD', str(int(TICI)))) +USE_THNEED = int(os.getenv('USE_THNEED', str(int(TICI)))) USE_SNPE = int(os.getenv('USE_SNPE', str(int(TICI)))) class ModelRunner(RunModel): - TINYGRAD = 'TINYGRAD' + THNEED = 'THNEED' SNPE = 'SNPE' ONNX = 'ONNX' def __new__(cls, paths, *args, **kwargs): - if ModelRunner.TINYGRAD in paths and USE_TINYGRAD: - from openpilot.selfdrive.modeld.runners.tinygradmodel import TinygradModel as Runner - runner_type = ModelRunner.TINYGRAD + if ModelRunner.THNEED in paths and USE_THNEED: + from openpilot.selfdrive.modeld.runners.thneedmodel_pyx import ThneedModel as Runner + runner_type = ModelRunner.THNEED elif ModelRunner.SNPE in paths and USE_SNPE: from openpilot.selfdrive.modeld.runners.snpemodel_pyx import SNPEModel as Runner runner_type = ModelRunner.SNPE diff --git a/selfdrive/modeld/runners/runmodel_pyx.pyx b/selfdrive/modeld/runners/runmodel_pyx.pyx index 8ef41ea17f..12b8ec10ff 100644 --- a/selfdrive/modeld/runners/runmodel_pyx.pyx +++ b/selfdrive/modeld/runners/runmodel_pyx.pyx @@ -5,7 +5,6 @@ from libcpp.string cimport string from .runmodel cimport USE_CPU_RUNTIME, USE_GPU_RUNTIME, USE_DSP_RUNTIME from selfdrive.modeld.models.commonmodel_pyx cimport CLMem -import numpy as np class Runtime: CPU = USE_CPU_RUNTIME @@ -22,12 +21,11 @@ cdef class RunModel: else: self.model.addInput(name, NULL, 0) - def setInputBuffer(self, string name, unsigned char[:] input_buffer): - cdef int num_floats = len(input_buffer) // sizeof(float) - cdef float* float_ptr = &input_buffer[0] - cdef float[:] float_buffer_view = float_ptr - if float_buffer_view is not None: - self.model.setInputBuffer(name, &float_buffer_view[0], num_floats) + def setInputBuffer(self, string name, float[:] buffer): + if buffer is not None: + self.model.setInputBuffer(name, &buffer[0], len(buffer)) + else: + self.model.setInputBuffer(name, NULL, 0) def getCLBuffer(self, string name): cdef void * cl_buf = self.model.getCLBuffer(name) diff --git a/selfdrive/modeld/runners/thneedmodel.cc b/selfdrive/modeld/runners/thneedmodel.cc new file mode 100644 index 0000000000..a16d8b42aa --- /dev/null +++ b/selfdrive/modeld/runners/thneedmodel.cc @@ -0,0 +1,58 @@ +#include "selfdrive/modeld/runners/thneedmodel.h" + +#include + +#include "common/swaglog.h" + +ThneedModel::ThneedModel(const std::string path, float *_output, size_t _output_size, int runtime, bool luse_tf8, cl_context context) { + thneed = new Thneed(true, context); + thneed->load(path.c_str()); + thneed->clexec(); + + recorded = false; + output = _output; +} + +void* ThneedModel::getCLBuffer(const std::string name) { + int index = -1; + for (int i = 0; i < inputs.size(); i++) { + if (name == inputs[i]->name) { + index = i; + break; + } + } + + if (index == -1) { + LOGE("Tried to get CL buffer for input `%s` but no input with this name exists", name.c_str()); + assert(false); + } + + if (thneed->input_clmem.size() >= inputs.size()) { + return &thneed->input_clmem[inputs.size() - index - 1]; + } else { + return nullptr; + } +} + +void ThneedModel::execute() { + if (!recorded) { + thneed->record = true; + float *input_buffers[inputs.size()]; + for (int i = 0; i < inputs.size(); i++) { + input_buffers[inputs.size() - i - 1] = inputs[i]->buffer; + } + + thneed->copy_inputs(input_buffers); + thneed->clexec(); + thneed->copy_output(output); + thneed->stop(); + + recorded = true; + } else { + float *input_buffers[inputs.size()]; + for (int i = 0; i < inputs.size(); i++) { + input_buffers[inputs.size() - i - 1] = inputs[i]->buffer; + } + thneed->execute(input_buffers, output); + } +} diff --git a/selfdrive/modeld/runners/thneedmodel.h b/selfdrive/modeld/runners/thneedmodel.h new file mode 100644 index 0000000000..6ed479c081 --- /dev/null +++ b/selfdrive/modeld/runners/thneedmodel.h @@ -0,0 +1,17 @@ +#pragma once + +#include + +#include "selfdrive/modeld/runners/runmodel.h" +#include "selfdrive/modeld/thneed/thneed.h" + +class ThneedModel : public RunModel { +public: + ThneedModel(const std::string path, float *_output, size_t _output_size, int runtime, bool use_tf8 = false, cl_context context = NULL); + void *getCLBuffer(const std::string name); + void execute(); +private: + Thneed *thneed = NULL; + bool recorded; + float *output; +}; diff --git a/selfdrive/modeld/runners/thneedmodel.pxd b/selfdrive/modeld/runners/thneedmodel.pxd new file mode 100644 index 0000000000..79e24dbdd6 --- /dev/null +++ b/selfdrive/modeld/runners/thneedmodel.pxd @@ -0,0 +1,9 @@ +# distutils: language = c++ + +from libcpp.string cimport string + +from msgq.visionipc.visionipc cimport cl_context + +cdef extern from "selfdrive/modeld/runners/thneedmodel.h": + cdef cppclass ThneedModel: + ThneedModel(string, float*, size_t, int, bool, cl_context) diff --git a/selfdrive/modeld/runners/thneedmodel_pyx.pyx b/selfdrive/modeld/runners/thneedmodel_pyx.pyx new file mode 100644 index 0000000000..6f8fdd255f --- /dev/null +++ b/selfdrive/modeld/runners/thneedmodel_pyx.pyx @@ -0,0 +1,14 @@ +# distutils: language = c++ +# cython: c_string_encoding=ascii, language_level=3 + +from libcpp cimport bool +from libcpp.string cimport string + +from .thneedmodel cimport ThneedModel as cppThneedModel +from selfdrive.modeld.models.commonmodel_pyx cimport CLContext +from selfdrive.modeld.runners.runmodel_pyx cimport RunModel +from selfdrive.modeld.runners.runmodel cimport RunModel as cppRunModel + +cdef class ThneedModel(RunModel): + def __cinit__(self, string path, float[:] output, int runtime, bool use_tf8, CLContext context): + self.model = new cppThneedModel(path, &output[0], len(output), runtime, use_tf8, context.context) diff --git a/selfdrive/modeld/runners/tinygrad_helpers.py b/selfdrive/modeld/runners/tinygrad_helpers.py deleted file mode 100644 index 776381341c..0000000000 --- a/selfdrive/modeld/runners/tinygrad_helpers.py +++ /dev/null @@ -1,8 +0,0 @@ - -from tinygrad.tensor import Tensor -from tinygrad.helpers import to_mv - -def qcom_tensor_from_opencl_address(opencl_address, shape, dtype): - cl_buf_desc_ptr = to_mv(opencl_address, 8).cast('Q')[0] - rawbuf_ptr = to_mv(cl_buf_desc_ptr, 0x100).cast('Q')[20] # offset 0xA0 is a raw gpu pointer. - return Tensor.from_blob(rawbuf_ptr, shape, dtype=dtype, device='QCOM') diff --git a/selfdrive/modeld/thneed/README b/selfdrive/modeld/thneed/README new file mode 100644 index 0000000000..f3bc66d8fc --- /dev/null +++ b/selfdrive/modeld/thneed/README @@ -0,0 +1,8 @@ +thneed is an SNPE accelerator. I know SNPE is already an accelerator, but sometimes things need to go even faster.. + +It runs on the local device, and caches a single model run. Then it replays it, but fast. + +thneed slices through abstraction layers like a fish. + +You need a thneed. + diff --git a/selfdrive/modeld/thneed/__init__.py b/selfdrive/modeld/thneed/__init__.py new file mode 100644 index 0000000000..e69de29bb2 diff --git a/selfdrive/modeld/thneed/serialize.cc b/selfdrive/modeld/thneed/serialize.cc new file mode 100644 index 0000000000..3dc2bef414 --- /dev/null +++ b/selfdrive/modeld/thneed/serialize.cc @@ -0,0 +1,154 @@ +#include +#include + +#include "third_party/json11/json11.hpp" +#include "common/util.h" +#include "common/clutil.h" +#include "common/swaglog.h" +#include "selfdrive/modeld/thneed/thneed.h" +using namespace json11; + +extern map g_program_source; + +void Thneed::load(const char *filename) { + LOGD("Thneed::load: loading from %s\n", filename); + + string buf = util::read_file(filename); + int jsz = *(int *)buf.data(); + string jsonerr; + string jj(buf.data() + sizeof(int), jsz); + Json jdat = Json::parse(jj, jsonerr); + + map real_mem; + real_mem[NULL] = NULL; + + int ptr = sizeof(int)+jsz; + for (auto &obj : jdat["objects"].array_items()) { + auto mobj = obj.object_items(); + int sz = mobj["size"].int_value(); + cl_mem clbuf = NULL; + + if (mobj["buffer_id"].string_value().size() > 0) { + // image buffer must already be allocated + clbuf = real_mem[*(cl_mem*)(mobj["buffer_id"].string_value().data())]; + assert(mobj["needs_load"].bool_value() == false); + } else { + if (mobj["needs_load"].bool_value()) { + clbuf = clCreateBuffer(context, CL_MEM_COPY_HOST_PTR | CL_MEM_READ_WRITE, sz, &buf[ptr], NULL); + if (debug >= 1) printf("loading %p %d @ 0x%X\n", clbuf, sz, ptr); + ptr += sz; + } else { + // TODO: is there a faster way to init zeroed out buffers? + void *host_zeros = calloc(sz, 1); + clbuf = clCreateBuffer(context, CL_MEM_COPY_HOST_PTR | CL_MEM_READ_WRITE, sz, host_zeros, NULL); + free(host_zeros); + } + } + assert(clbuf != NULL); + + if (mobj["arg_type"] == "image2d_t" || mobj["arg_type"] == "image1d_t") { + cl_image_desc desc = {0}; + desc.image_type = (mobj["arg_type"] == "image2d_t") ? CL_MEM_OBJECT_IMAGE2D : CL_MEM_OBJECT_IMAGE1D_BUFFER; + desc.image_width = mobj["width"].int_value(); + desc.image_height = mobj["height"].int_value(); + desc.image_row_pitch = mobj["row_pitch"].int_value(); + assert(sz == desc.image_height*desc.image_row_pitch); +#ifdef QCOM2 + desc.buffer = clbuf; +#else + // TODO: we are creating unused buffers on PC + clReleaseMemObject(clbuf); +#endif + cl_image_format format = {0}; + format.image_channel_order = CL_RGBA; + format.image_channel_data_type = mobj["float32"].bool_value() ? CL_FLOAT : CL_HALF_FLOAT; + + cl_int errcode; + +#ifndef QCOM2 + if (mobj["needs_load"].bool_value()) { + clbuf = clCreateImage(context, CL_MEM_COPY_HOST_PTR | CL_MEM_READ_WRITE, &format, &desc, &buf[ptr-sz], &errcode); + } else { + clbuf = clCreateImage(context, CL_MEM_READ_WRITE, &format, &desc, NULL, &errcode); + } +#else + clbuf = clCreateImage(context, CL_MEM_READ_WRITE, &format, &desc, NULL, &errcode); +#endif + if (clbuf == NULL) { + LOGE("clError: %s create image %zux%zu rp %zu with buffer %p\n", cl_get_error_string(errcode), + desc.image_width, desc.image_height, desc.image_row_pitch, desc.buffer); + } + assert(clbuf != NULL); + } + + real_mem[*(cl_mem*)(mobj["id"].string_value().data())] = clbuf; + } + + map g_programs; + for (const auto &[name, source] : jdat["programs"].object_items()) { + if (debug >= 1) printf("building %s with size %zu\n", name.c_str(), source.string_value().size()); + g_programs[name] = cl_program_from_source(context, device_id, source.string_value()); + } + + for (auto &obj : jdat["inputs"].array_items()) { + auto mobj = obj.object_items(); + int sz = mobj["size"].int_value(); + cl_mem aa = real_mem[*(cl_mem*)(mobj["buffer_id"].string_value().data())]; + input_clmem.push_back(aa); + input_sizes.push_back(sz); + LOGD("Thneed::load: adding input %s with size %d\n", mobj["name"].string_value().data(), sz); + + cl_int cl_err; + void *ret = clEnqueueMapBuffer(command_queue, aa, CL_TRUE, CL_MAP_WRITE, 0, sz, 0, NULL, NULL, &cl_err); + if (cl_err != CL_SUCCESS) LOGE("clError: %s map %p %d\n", cl_get_error_string(cl_err), aa, sz); + assert(cl_err == CL_SUCCESS); + inputs.push_back(ret); + } + + for (auto &obj : jdat["outputs"].array_items()) { + auto mobj = obj.object_items(); + int sz = mobj["size"].int_value(); + LOGD("Thneed::save: adding output with size %d\n", sz); + // TODO: support multiple outputs + output = real_mem[*(cl_mem*)(mobj["buffer_id"].string_value().data())]; + assert(output != NULL); + } + + for (auto &obj : jdat["binaries"].array_items()) { + string name = obj["name"].string_value(); + size_t length = obj["length"].int_value(); + if (debug >= 1) printf("binary %s with size %zu\n", name.c_str(), length); + g_programs[name] = cl_program_from_binary(context, device_id, (const uint8_t*)&buf[ptr], length); + ptr += length; + } + + for (auto &obj : jdat["kernels"].array_items()) { + auto gws = obj["global_work_size"]; + auto lws = obj["local_work_size"]; + auto kk = shared_ptr(new CLQueuedKernel(this)); + + kk->name = obj["name"].string_value(); + kk->program = g_programs[kk->name]; + kk->work_dim = obj["work_dim"].int_value(); + for (int i = 0; i < kk->work_dim; i++) { + kk->global_work_size[i] = gws[i].int_value(); + kk->local_work_size[i] = lws[i].int_value(); + } + kk->num_args = obj["num_args"].int_value(); + for (int i = 0; i < kk->num_args; i++) { + string arg = obj["args"].array_items()[i].string_value(); + int arg_size = obj["args_size"].array_items()[i].int_value(); + kk->args_size.push_back(arg_size); + if (arg_size == 8) { + cl_mem val = *(cl_mem*)(arg.data()); + val = real_mem[val]; + kk->args.push_back(string((char*)&val, sizeof(val))); + } else { + kk->args.push_back(arg); + } + } + kq.push_back(kk); + } + + clFinish(command_queue); +} diff --git a/selfdrive/modeld/thneed/thneed.h b/selfdrive/modeld/thneed/thneed.h new file mode 100644 index 0000000000..47e18e0be3 --- /dev/null +++ b/selfdrive/modeld/thneed/thneed.h @@ -0,0 +1,133 @@ +#pragma once + +#ifndef __user +#define __user __attribute__(()) +#endif + +#include +#include +#include +#include +#include + +#include + +#include "third_party/linux/include/msm_kgsl.h" + +using namespace std; + +cl_int thneed_clSetKernelArg(cl_kernel kernel, cl_uint arg_index, size_t arg_size, const void *arg_value); + +namespace json11 { + class Json; +} +class Thneed; + +class GPUMalloc { + public: + GPUMalloc(int size, int fd); + ~GPUMalloc(); + void *alloc(int size); + private: + uint64_t base; + int remaining; +}; + +class CLQueuedKernel { + public: + CLQueuedKernel(Thneed *lthneed) { thneed = lthneed; } + CLQueuedKernel(Thneed *lthneed, + cl_kernel _kernel, + cl_uint _work_dim, + const size_t *_global_work_size, + const size_t *_local_work_size); + cl_int exec(); + void debug_print(bool verbose); + int get_arg_num(const char *search_arg_name); + cl_program program; + string name; + cl_uint num_args; + vector arg_names; + vector arg_types; + vector args; + vector args_size; + cl_kernel kernel = NULL; + json11::Json to_json() const; + + cl_uint work_dim; + size_t global_work_size[3] = {0}; + size_t local_work_size[3] = {0}; + private: + Thneed *thneed; +}; + +class CachedIoctl { + public: + virtual void exec() {} +}; + +class CachedSync: public CachedIoctl { + public: + CachedSync(Thneed *lthneed, string ldata) { thneed = lthneed; data = ldata; } + void exec(); + private: + Thneed *thneed; + string data; +}; + +class CachedCommand: public CachedIoctl { + public: + CachedCommand(Thneed *lthneed, struct kgsl_gpu_command *cmd); + void exec(); + private: + void disassemble(int cmd_index); + struct kgsl_gpu_command cache; + unique_ptr cmds; + unique_ptr objs; + Thneed *thneed; + vector > kq; +}; + +class Thneed { + public: + Thneed(bool do_clinit=false, cl_context _context = NULL); + void stop(); + void execute(float **finputs, float *foutput, bool slow=false); + void wait(); + + vector input_clmem; + vector inputs; + vector input_sizes; + cl_mem output = NULL; + + cl_context context = NULL; + cl_command_queue command_queue; + cl_device_id device_id; + int context_id; + + // protected? + bool record = false; + int debug; + int timestamp; + +#ifdef QCOM2 + unique_ptr ram; + vector > cmds; + int fd; +#endif + + // all CL kernels + void copy_inputs(float **finputs, bool internal=false); + void copy_output(float *foutput); + cl_int clexec(); + vector > kq; + + // pending CL kernels + vector > ckq; + + // loading + void load(const char *filename); + private: + void clinit(); +}; + diff --git a/selfdrive/modeld/thneed/thneed_common.cc b/selfdrive/modeld/thneed/thneed_common.cc new file mode 100644 index 0000000000..ecdf1237e3 --- /dev/null +++ b/selfdrive/modeld/thneed/thneed_common.cc @@ -0,0 +1,216 @@ +#include "selfdrive/modeld/thneed/thneed.h" + +#include +#include +#include + +#include "common/clutil.h" +#include "common/timing.h" + +map, string> g_args; +map, int> g_args_size; +map g_program_source; + +void Thneed::stop() { + //printf("Thneed::stop: recorded %lu commands\n", cmds.size()); + record = false; +} + +void Thneed::clinit() { + device_id = cl_get_device_id(CL_DEVICE_TYPE_DEFAULT); + if (context == NULL) context = CL_CHECK_ERR(clCreateContext(NULL, 1, &device_id, NULL, NULL, &err)); + //cl_command_queue_properties props[3] = {CL_QUEUE_PROPERTIES, CL_QUEUE_PROFILING_ENABLE, 0}; + cl_command_queue_properties props[3] = {CL_QUEUE_PROPERTIES, 0, 0}; + command_queue = CL_CHECK_ERR(clCreateCommandQueueWithProperties(context, device_id, props, &err)); + printf("Thneed::clinit done\n"); +} + +cl_int Thneed::clexec() { + if (debug >= 1) printf("Thneed::clexec: running %lu queued kernels\n", kq.size()); + for (auto &k : kq) { + if (record) ckq.push_back(k); + cl_int ret = k->exec(); + assert(ret == CL_SUCCESS); + } + return clFinish(command_queue); +} + +void Thneed::copy_inputs(float **finputs, bool internal) { + for (int idx = 0; idx < inputs.size(); ++idx) { + if (debug >= 1) printf("copying %lu -- %p -> %p (cl %p)\n", input_sizes[idx], finputs[idx], inputs[idx], input_clmem[idx]); + + if (internal) { + // if it's internal, using memcpy is fine since the buffer sync is cached in the ioctl layer + if (finputs[idx] != NULL) memcpy(inputs[idx], finputs[idx], input_sizes[idx]); + } else { + if (finputs[idx] != NULL) CL_CHECK(clEnqueueWriteBuffer(command_queue, input_clmem[idx], CL_TRUE, 0, input_sizes[idx], finputs[idx], 0, NULL, NULL)); + } + } +} + +void Thneed::copy_output(float *foutput) { + if (output != NULL) { + size_t sz; + clGetMemObjectInfo(output, CL_MEM_SIZE, sizeof(sz), &sz, NULL); + if (debug >= 1) printf("copying %lu for output %p -> %p\n", sz, output, foutput); + CL_CHECK(clEnqueueReadBuffer(command_queue, output, CL_TRUE, 0, sz, foutput, 0, NULL, NULL)); + } else { + printf("CAUTION: model output is NULL, does it have no outputs?\n"); + } +} + +// *********** CLQueuedKernel *********** + +CLQueuedKernel::CLQueuedKernel(Thneed *lthneed, + cl_kernel _kernel, + cl_uint _work_dim, + const size_t *_global_work_size, + const size_t *_local_work_size) { + thneed = lthneed; + kernel = _kernel; + work_dim = _work_dim; + assert(work_dim <= 3); + for (int i = 0; i < work_dim; i++) { + global_work_size[i] = _global_work_size[i]; + local_work_size[i] = _local_work_size[i]; + } + + char _name[0x100]; + clGetKernelInfo(kernel, CL_KERNEL_FUNCTION_NAME, sizeof(_name), _name, NULL); + name = string(_name); + clGetKernelInfo(kernel, CL_KERNEL_NUM_ARGS, sizeof(num_args), &num_args, NULL); + + // get args + for (int i = 0; i < num_args; i++) { + char arg_name[0x100] = {0}; + clGetKernelArgInfo(kernel, i, CL_KERNEL_ARG_NAME, sizeof(arg_name), arg_name, NULL); + arg_names.push_back(string(arg_name)); + clGetKernelArgInfo(kernel, i, CL_KERNEL_ARG_TYPE_NAME, sizeof(arg_name), arg_name, NULL); + arg_types.push_back(string(arg_name)); + + args.push_back(g_args[make_pair(kernel, i)]); + args_size.push_back(g_args_size[make_pair(kernel, i)]); + } + + // get program + clGetKernelInfo(kernel, CL_KERNEL_PROGRAM, sizeof(program), &program, NULL); +} + +int CLQueuedKernel::get_arg_num(const char *search_arg_name) { + for (int i = 0; i < num_args; i++) { + if (arg_names[i] == search_arg_name) return i; + } + printf("failed to find %s in %s\n", search_arg_name, name.c_str()); + assert(false); +} + +cl_int CLQueuedKernel::exec() { + if (kernel == NULL) { + kernel = clCreateKernel(program, name.c_str(), NULL); + arg_names.clear(); + arg_types.clear(); + + for (int j = 0; j < num_args; j++) { + char arg_name[0x100] = {0}; + clGetKernelArgInfo(kernel, j, CL_KERNEL_ARG_NAME, sizeof(arg_name), arg_name, NULL); + arg_names.push_back(string(arg_name)); + clGetKernelArgInfo(kernel, j, CL_KERNEL_ARG_TYPE_NAME, sizeof(arg_name), arg_name, NULL); + arg_types.push_back(string(arg_name)); + + cl_int ret; + if (args[j].size() != 0) { + assert(args[j].size() == args_size[j]); + ret = thneed_clSetKernelArg(kernel, j, args[j].size(), args[j].data()); + } else { + ret = thneed_clSetKernelArg(kernel, j, args_size[j], NULL); + } + assert(ret == CL_SUCCESS); + } + } + + if (thneed->debug >= 1) { + debug_print(thneed->debug >= 2); + } + + return clEnqueueNDRangeKernel(thneed->command_queue, + kernel, work_dim, NULL, global_work_size, local_work_size, 0, NULL, NULL); +} + +void CLQueuedKernel::debug_print(bool verbose) { + printf("%p %56s -- ", kernel, name.c_str()); + for (int i = 0; i < work_dim; i++) { + printf("%4zu ", global_work_size[i]); + } + printf(" -- "); + for (int i = 0; i < work_dim; i++) { + printf("%4zu ", local_work_size[i]); + } + printf("\n"); + + if (verbose) { + for (int i = 0; i < num_args; i++) { + string arg = args[i]; + printf(" %s %s", arg_types[i].c_str(), arg_names[i].c_str()); + void *arg_value = (void*)arg.data(); + int arg_size = arg.size(); + if (arg_size == 0) { + printf(" (size) %d", args_size[i]); + } else if (arg_size == 1) { + printf(" = %d", *((char*)arg_value)); + } else if (arg_size == 2) { + printf(" = %d", *((short*)arg_value)); + } else if (arg_size == 4) { + if (arg_types[i] == "float") { + printf(" = %f", *((float*)arg_value)); + } else { + printf(" = %d", *((int*)arg_value)); + } + } else if (arg_size == 8) { + cl_mem val = (cl_mem)(*((uintptr_t*)arg_value)); + printf(" = %p", val); + if (val != NULL) { + cl_mem_object_type obj_type; + clGetMemObjectInfo(val, CL_MEM_TYPE, sizeof(obj_type), &obj_type, NULL); + if (arg_types[i] == "image2d_t" || arg_types[i] == "image1d_t" || obj_type == CL_MEM_OBJECT_IMAGE2D) { + cl_image_format format; + size_t width, height, depth, array_size, row_pitch, slice_pitch; + cl_mem buf; + clGetImageInfo(val, CL_IMAGE_FORMAT, sizeof(format), &format, NULL); + assert(format.image_channel_order == CL_RGBA); + assert(format.image_channel_data_type == CL_HALF_FLOAT || format.image_channel_data_type == CL_FLOAT); + clGetImageInfo(val, CL_IMAGE_WIDTH, sizeof(width), &width, NULL); + clGetImageInfo(val, CL_IMAGE_HEIGHT, sizeof(height), &height, NULL); + clGetImageInfo(val, CL_IMAGE_ROW_PITCH, sizeof(row_pitch), &row_pitch, NULL); + clGetImageInfo(val, CL_IMAGE_DEPTH, sizeof(depth), &depth, NULL); + clGetImageInfo(val, CL_IMAGE_ARRAY_SIZE, sizeof(array_size), &array_size, NULL); + clGetImageInfo(val, CL_IMAGE_SLICE_PITCH, sizeof(slice_pitch), &slice_pitch, NULL); + assert(depth == 0); + assert(array_size == 0); + assert(slice_pitch == 0); + + clGetImageInfo(val, CL_IMAGE_BUFFER, sizeof(buf), &buf, NULL); + size_t sz = 0; + if (buf != NULL) clGetMemObjectInfo(buf, CL_MEM_SIZE, sizeof(sz), &sz, NULL); + printf(" image %zu x %zu rp %zu @ %p buffer %zu", width, height, row_pitch, buf, sz); + } else { + size_t sz; + clGetMemObjectInfo(val, CL_MEM_SIZE, sizeof(sz), &sz, NULL); + printf(" buffer %zu", sz); + } + } + } + printf("\n"); + } + } +} + +cl_int thneed_clSetKernelArg(cl_kernel kernel, cl_uint arg_index, size_t arg_size, const void *arg_value) { + g_args_size[make_pair(kernel, arg_index)] = arg_size; + if (arg_value != NULL) { + g_args[make_pair(kernel, arg_index)] = string((char*)arg_value, arg_size); + } else { + g_args[make_pair(kernel, arg_index)] = string(""); + } + cl_int ret = clSetKernelArg(kernel, arg_index, arg_size, arg_value); + return ret; +} diff --git a/selfdrive/modeld/thneed/thneed_pc.cc b/selfdrive/modeld/thneed/thneed_pc.cc new file mode 100644 index 0000000000..8d0037628e --- /dev/null +++ b/selfdrive/modeld/thneed/thneed_pc.cc @@ -0,0 +1,32 @@ +#include "selfdrive/modeld/thneed/thneed.h" + +#include + +#include "common/clutil.h" +#include "common/timing.h" + +Thneed::Thneed(bool do_clinit, cl_context _context) { + context = _context; + if (do_clinit) clinit(); + char *thneed_debug_env = getenv("THNEED_DEBUG"); + debug = (thneed_debug_env != NULL) ? atoi(thneed_debug_env) : 0; +} + +void Thneed::execute(float **finputs, float *foutput, bool slow) { + uint64_t tb, te; + if (debug >= 1) tb = nanos_since_boot(); + + // ****** copy inputs + copy_inputs(finputs); + + // ****** run commands + clexec(); + + // ****** copy outputs + copy_output(foutput); + + if (debug >= 1) { + te = nanos_since_boot(); + printf("model exec in %lu us\n", (te-tb)/1000); + } +} diff --git a/selfdrive/modeld/thneed/thneed_qcom2.cc b/selfdrive/modeld/thneed/thneed_qcom2.cc new file mode 100644 index 0000000000..21de15d17c --- /dev/null +++ b/selfdrive/modeld/thneed/thneed_qcom2.cc @@ -0,0 +1,258 @@ +#include "selfdrive/modeld/thneed/thneed.h" + +#include +#include + +#include +#include +#include +#include +#include + +#include "common/clutil.h" +#include "common/timing.h" + +Thneed *g_thneed = NULL; +int g_fd = -1; + +void hexdump(uint8_t *d, int len) { + assert((len%4) == 0); + printf(" dumping %p len 0x%x\n", d, len); + for (int i = 0; i < len/4; i++) { + if (i != 0 && (i%0x10) == 0) printf("\n"); + printf("%8x ", d[i]); + } + printf("\n"); +} + +// *********** ioctl interceptor *********** + +extern "C" { + +int (*my_ioctl)(int filedes, unsigned long request, void *argp) = NULL; +#undef ioctl +int ioctl(int filedes, unsigned long request, void *argp) { + request &= 0xFFFFFFFF; // needed on QCOM2 + if (my_ioctl == NULL) my_ioctl = reinterpret_cast(dlsym(RTLD_NEXT, "ioctl")); + Thneed *thneed = g_thneed; + + // save the fd + if (request == IOCTL_KGSL_GPUOBJ_ALLOC) g_fd = filedes; + + // note that this runs always, even without a thneed object + if (request == IOCTL_KGSL_DRAWCTXT_CREATE) { + struct kgsl_drawctxt_create *create = (struct kgsl_drawctxt_create *)argp; + create->flags &= ~KGSL_CONTEXT_PRIORITY_MASK; + create->flags |= 6 << KGSL_CONTEXT_PRIORITY_SHIFT; // priority from 1-15, 1 is max priority + printf("IOCTL_KGSL_DRAWCTXT_CREATE: creating context with flags 0x%x\n", create->flags); + } + + if (thneed != NULL) { + if (request == IOCTL_KGSL_GPU_COMMAND) { + struct kgsl_gpu_command *cmd = (struct kgsl_gpu_command *)argp; + if (thneed->record) { + thneed->timestamp = cmd->timestamp; + thneed->context_id = cmd->context_id; + thneed->cmds.push_back(unique_ptr(new CachedCommand(thneed, cmd))); + } + if (thneed->debug >= 1) { + printf("IOCTL_KGSL_GPU_COMMAND(%2zu): flags: 0x%lx context_id: %u timestamp: %u numcmds: %d numobjs: %d\n", + thneed->cmds.size(), + cmd->flags, + cmd->context_id, cmd->timestamp, cmd->numcmds, cmd->numobjs); + } + } else if (request == IOCTL_KGSL_GPUOBJ_SYNC) { + struct kgsl_gpuobj_sync *cmd = (struct kgsl_gpuobj_sync *)argp; + struct kgsl_gpuobj_sync_obj *objs = (struct kgsl_gpuobj_sync_obj *)(cmd->objs); + + if (thneed->debug >= 2) { + printf("IOCTL_KGSL_GPUOBJ_SYNC count:%d ", cmd->count); + for (int i = 0; i < cmd->count; i++) { + printf(" -- offset:0x%lx len:0x%lx id:%d op:%d ", objs[i].offset, objs[i].length, objs[i].id, objs[i].op); + } + printf("\n"); + } + + if (thneed->record) { + thneed->cmds.push_back(unique_ptr(new + CachedSync(thneed, string((char *)objs, sizeof(struct kgsl_gpuobj_sync_obj)*cmd->count)))); + } + } else if (request == IOCTL_KGSL_DEVICE_WAITTIMESTAMP_CTXTID) { + struct kgsl_device_waittimestamp_ctxtid *cmd = (struct kgsl_device_waittimestamp_ctxtid *)argp; + if (thneed->debug >= 1) { + printf("IOCTL_KGSL_DEVICE_WAITTIMESTAMP_CTXTID: context_id: %d timestamp: %d timeout: %d\n", + cmd->context_id, cmd->timestamp, cmd->timeout); + } + } else if (request == IOCTL_KGSL_SETPROPERTY) { + if (thneed->debug >= 1) { + struct kgsl_device_getproperty *prop = (struct kgsl_device_getproperty *)argp; + printf("IOCTL_KGSL_SETPROPERTY: 0x%x sizebytes:%zu\n", prop->type, prop->sizebytes); + if (thneed->debug >= 2) { + hexdump((uint8_t *)prop->value, prop->sizebytes); + if (prop->type == KGSL_PROP_PWR_CONSTRAINT) { + struct kgsl_device_constraint *constraint = (struct kgsl_device_constraint *)prop->value; + hexdump((uint8_t *)constraint->data, constraint->size); + } + } + } + } else if (request == IOCTL_KGSL_DRAWCTXT_CREATE || request == IOCTL_KGSL_DRAWCTXT_DESTROY) { + // this happens + } else if (request == IOCTL_KGSL_GPUOBJ_ALLOC || request == IOCTL_KGSL_GPUOBJ_FREE) { + // this happens + } else { + if (thneed->debug >= 1) { + printf("other ioctl %lx\n", request); + } + } + } + + int ret = my_ioctl(filedes, request, argp); + // NOTE: This error message goes into stdout and messes up pyenv + // if (ret != 0) printf("ioctl returned %d with errno %d\n", ret, errno); + return ret; +} + +} + +// *********** GPUMalloc *********** + +GPUMalloc::GPUMalloc(int size, int fd) { + struct kgsl_gpuobj_alloc alloc; + memset(&alloc, 0, sizeof(alloc)); + alloc.size = size; + alloc.flags = 0x10000a00; + ioctl(fd, IOCTL_KGSL_GPUOBJ_ALLOC, &alloc); + void *addr = mmap64(NULL, alloc.mmapsize, 0x3, 0x1, fd, alloc.id*0x1000); + assert(addr != MAP_FAILED); + + base = (uint64_t)addr; + remaining = size; +} + +GPUMalloc::~GPUMalloc() { + // TODO: free the GPU malloced area +} + +void *GPUMalloc::alloc(int size) { + void *ret = (void*)base; + size = (size+0xff) & (~0xFF); + assert(size <= remaining); + remaining -= size; + base += size; + return ret; +} + +// *********** CachedSync, at the ioctl layer *********** + +void CachedSync::exec() { + struct kgsl_gpuobj_sync cmd; + + cmd.objs = (uint64_t)data.data(); + cmd.obj_len = data.length(); + cmd.count = data.length() / sizeof(struct kgsl_gpuobj_sync_obj); + + int ret = ioctl(thneed->fd, IOCTL_KGSL_GPUOBJ_SYNC, &cmd); + assert(ret == 0); +} + +// *********** CachedCommand, at the ioctl layer *********** + +CachedCommand::CachedCommand(Thneed *lthneed, struct kgsl_gpu_command *cmd) { + thneed = lthneed; + assert(cmd->numsyncs == 0); + + memcpy(&cache, cmd, sizeof(cache)); + + if (cmd->numcmds > 0) { + cmds = make_unique(cmd->numcmds); + memcpy(cmds.get(), (void *)cmd->cmdlist, sizeof(struct kgsl_command_object)*cmd->numcmds); + cache.cmdlist = (uint64_t)cmds.get(); + for (int i = 0; i < cmd->numcmds; i++) { + void *nn = thneed->ram->alloc(cmds[i].size); + memcpy(nn, (void*)cmds[i].gpuaddr, cmds[i].size); + cmds[i].gpuaddr = (uint64_t)nn; + } + } + + if (cmd->numobjs > 0) { + objs = make_unique(cmd->numobjs); + memcpy(objs.get(), (void *)cmd->objlist, sizeof(struct kgsl_command_object)*cmd->numobjs); + cache.objlist = (uint64_t)objs.get(); + for (int i = 0; i < cmd->numobjs; i++) { + void *nn = thneed->ram->alloc(objs[i].size); + memset(nn, 0, objs[i].size); + objs[i].gpuaddr = (uint64_t)nn; + } + } + + kq = thneed->ckq; + thneed->ckq.clear(); +} + +void CachedCommand::exec() { + cache.timestamp = ++thneed->timestamp; + int ret = ioctl(thneed->fd, IOCTL_KGSL_GPU_COMMAND, &cache); + + if (thneed->debug >= 1) printf("CachedCommand::exec got %d\n", ret); + + if (thneed->debug >= 2) { + for (auto &it : kq) { + it->debug_print(false); + } + } + + assert(ret == 0); +} + +// *********** Thneed *********** + +Thneed::Thneed(bool do_clinit, cl_context _context) { + // TODO: QCOM2 actually requires a different context + //context = _context; + if (do_clinit) clinit(); + assert(g_fd != -1); + fd = g_fd; + ram = make_unique(0x80000, fd); + timestamp = -1; + g_thneed = this; + char *thneed_debug_env = getenv("THNEED_DEBUG"); + debug = (thneed_debug_env != NULL) ? atoi(thneed_debug_env) : 0; +} + +void Thneed::wait() { + struct kgsl_device_waittimestamp_ctxtid wait; + wait.context_id = context_id; + wait.timestamp = timestamp; + wait.timeout = -1; + + uint64_t tb = nanos_since_boot(); + int wret = ioctl(fd, IOCTL_KGSL_DEVICE_WAITTIMESTAMP_CTXTID, &wait); + uint64_t te = nanos_since_boot(); + + if (debug >= 1) printf("wait %d after %lu us\n", wret, (te-tb)/1000); +} + +void Thneed::execute(float **finputs, float *foutput, bool slow) { + uint64_t tb, te; + if (debug >= 1) tb = nanos_since_boot(); + + // ****** copy inputs + copy_inputs(finputs, true); + + // ****** run commands + int i = 0; + for (auto &it : cmds) { + ++i; + if (debug >= 1) printf("run %2d @ %7lu us: ", i, (nanos_since_boot()-tb)/1000); + it->exec(); + if ((i == cmds.size()) || slow) wait(); + } + + // ****** copy outputs + copy_output(foutput); + + if (debug >= 1) { + te = nanos_since_boot(); + printf("model exec in %lu us\n", (te-tb)/1000); + } +} diff --git a/selfdrive/test/test_onroad.py b/selfdrive/test/test_onroad.py index 2e67791675..9b131a639a 100644 --- a/selfdrive/test/test_onroad.py +++ b/selfdrive/test/test_onroad.py @@ -36,7 +36,7 @@ CPU usage budget TEST_DURATION = 25 LOG_OFFSET = 8 -MAX_TOTAL_CPU = 275. # total for all 8 cores +MAX_TOTAL_CPU = 265. # total for all 8 cores PROCS = { # Baseline CPU usage by process "selfdrive.controls.controlsd": 16.0, @@ -50,8 +50,8 @@ PROCS = { "selfdrive.locationd.paramsd": 9.0, "./sensord": 7.0, "selfdrive.controls.radard": 2.0, - "selfdrive.modeld.modeld": 22.0, - "selfdrive.modeld.dmonitoringmodeld": 21.0, + "selfdrive.modeld.modeld": 17.0, + "selfdrive.modeld.dmonitoringmodeld": 11.0, "system.hardware.hardwared": 4.0, "selfdrive.locationd.calibrationd": 2.0, "selfdrive.locationd.torqued": 5.0, @@ -361,15 +361,13 @@ class TestOnroad: result += "------------------------------------------------\n" result += "----------------- Model Timing -----------------\n" result += "------------------------------------------------\n" - # TODO: Decrease again when tinygrad speeds ups + # TODO: this went up when plannerd cpu usage increased, why? cfgs = [ - ("modelV2", 0.050, 0.040), + ("modelV2", 0.050, 0.036), ("driverStateV2", 0.050, 0.026), ] for (s, instant_max, avg_max) in cfgs: ts = [getattr(m, s).modelExecutionTime for m in self.msgs[s]] - # TODO some tinygrad init happens in first iteration - ts = ts[1:] assert max(ts) < instant_max, f"high '{s}' execution time: {max(ts)}" assert np.mean(ts) < avg_max, f"high avg '{s}' execution time: {np.mean(ts)}" result += f"'{s}' execution time: min {min(ts):.5f}s\n" diff --git a/system/camerad/cameras/camera_qcom2.cc b/system/camerad/cameras/camera_qcom2.cc index 0c102051a5..fb325ac772 100644 --- a/system/camerad/cameras/camera_qcom2.cc +++ b/system/camerad/cameras/camera_qcom2.cc @@ -55,7 +55,7 @@ public: float fl_pix = 0; - CameraState(SpectraMaster *master, const CameraConfig &config) : camera(master, config, config.stream_type == VISION_STREAM_ROAD) {}; + CameraState(SpectraMaster *master, const CameraConfig &config) : camera(master, config, true /*config.stream_type == VISION_STREAM_ROAD*/) {}; ~CameraState(); void init(VisionIpcServer *v, cl_device_id device_id, cl_context ctx); void update_exposure_score(float desired_ev, int exp_t, int exp_g_idx, float exp_gain); diff --git a/system/hardware/tici/tests/test_power_draw.py b/system/hardware/tici/tests/test_power_draw.py index 8598b2faa2..0ef34549b5 100644 --- a/system/hardware/tici/tests/test_power_draw.py +++ b/system/hardware/tici/tests/test_power_draw.py @@ -31,7 +31,7 @@ class Proc: PROCS = [ - Proc(['camerad'], 1.75, msgs=['roadCameraState', 'wideRoadCameraState', 'driverCameraState']), + Proc(['camerad'], 2.1, msgs=['roadCameraState', 'wideRoadCameraState', 'driverCameraState']), Proc(['modeld'], 1.12, atol=0.2, msgs=['modelV2']), Proc(['dmonitoringmodeld'], 0.5, msgs=['driverStateV2']), Proc(['encoderd'], 0.23, msgs=[]), diff --git a/tinygrad_repo b/tinygrad_repo index ad119af6a5..9dda6d260d 160000 --- a/tinygrad_repo +++ b/tinygrad_repo @@ -1 +1 @@ -Subproject commit ad119af6a511373e1c016a6525ab733f14a60c51 +Subproject commit 9dda6d260db0255750bacff61e3cee1e580567e1