modeld: Move from SNPE to tinygrad (#25207)

* compiling, won't work yet

* running with inputs and outputs

* there's some magic chance this works

* no more dlc, include onnx

* yolo tests plz

* bump tinygrad

* files_common + delete dlc

* tinygrad_repo -> tinygrad

* pre commit config

* llops needed

* extra in files_common

* bump tinygrad

* fix indent

* tinygrad/nn/__init__

* tinygrad_repo

* bump tinygrad repo

* bump tinygrad

* bump with native_exp, match maybe

* native_explog is argument

* pyopencl no cache

* 5% chance this matches

* work in float32?

* bump tinygrad

* fix build

* no __init__

* fix recip

* dumb hack

* adding thneed PC support

* fix pc segfault

* pc thneed is working

* to_image

* prints stuff with debug=2

* it sort of works

* copy host ptr is simpler

* bug fix

* build on c3

* this correct?

* reenable float16

* fix private, fixup copy_inputs internal

* bump tinygrad and update ref commit

* fix OPTWG on PC

* maybe fix non determinism

* revert model replay ref commit

* comments, init zeroed out buffers

* upd ref commit

* bump tinygrad to fix initial image

* try this ref

Co-authored-by: Comma Device <[email protected]>

.gitignore

@@ -36,6 +36,7 @@ a.out
 config.json
 clcache
 compile_commands.json
+compare_runtime*.html
 
 persist
 board/obj/

.gitmodules

@@ -16,3 +16,6 @@
 [submodule "body"]
   path = body
   url = ../../commaai/body.git
+[submodule "tinygrad"]
+  path = tinygrad_repo
+  url = https://github.com/geohot/tinygrad.git

.pre-commit-config.yaml

@@ -28,7 +28,7 @@ repos:
     rev: v0.931
     hooks:
     -   id: mypy
-        exclude: '^(pyextra/)|(cereal/)|(rednose/)|(panda/)|(laika/)|(opendbc/)|(laika_repo/)|(rednose_repo/)/'
+        exclude: '^(pyextra/)|(cereal/)|(rednose/)|(panda/)|(laika/)|(opendbc/)|(laika_repo/)|(rednose_repo/)/|(tinygrad/)|(tinygrad_repo/)'
         additional_dependencies: ['types-PyYAML', 'lxml', 'numpy', 'types-atomicwrites', 'types-pycurl', 'types-requests', 'types-certifi']
         args:
         - --warn-redundant-casts
@@ -40,7 +40,7 @@ repos:
     rev: 4.0.1
     hooks:
     -   id: flake8
-        exclude: '^(pyextra/)|(cereal/)|(rednose/)|(panda/)|(laika/)|(opendbc/)|(laika_repo/)|(rednose_repo/)|(selfdrive/debug/)/'
+        exclude: '^(pyextra/)|(cereal/)|(rednose/)|(panda/)|(laika/)|(opendbc/)|(laika_repo/)|(rednose_repo/)|(tinygrad/)|(tinygrad_repo/)|(selfdrive/debug/)/'
         additional_dependencies: ['flake8-no-implicit-concat']
         args:
         - --indent-size=2
@@ -55,7 +55,7 @@ repos:
         entry: pylint
         language: system
         types: [python]
-        exclude: '^(pyextra/)|(cereal/)|(rednose/)|(panda/)|(laika/)|(laika_repo/)|(rednose_repo/)'
+        exclude: '^(pyextra/)|(cereal/)|(rednose/)|(panda/)|(laika/)|(laika_repo/)|(rednose_repo/)|(tinygrad/)|(tinygrad_repo/)'
         args:
         - -rn
         - -sn

SConstruct

@@ -49,6 +49,11 @@ AddOption('--no-thneed',
           dest='no_thneed',
           help='avoid using thneed')
 
+AddOption('--pc-thneed',
+          action='store_true',
+          dest='pc_thneed',
+          help='use thneed on pc')
+
 AddOption('--no-test',
           action='store_false',
           dest='test',

common/clutil.cc

@@ -78,7 +78,8 @@ cl_program cl_program_from_file(cl_context ctx, cl_device_id device_id, const ch
 }
 
 cl_program cl_program_from_source(cl_context ctx, cl_device_id device_id, const std::string& src, const char* args) {
-  cl_program prg = CL_CHECK_ERR(clCreateProgramWithSource(ctx, 1, (const char*[]){src.c_str()}, NULL, &err));
+  const char *csrc = src.c_str();
+  cl_program prg = CL_CHECK_ERR(clCreateProgramWithSource(ctx, 1, &csrc, NULL, &err));
   if (int err = clBuildProgram(prg, 1, &device_id, args, NULL, NULL); err != 0) {
     cl_print_build_errors(prg, device_id);
     assert(0);
@@ -87,7 +88,7 @@ cl_program cl_program_from_source(cl_context ctx, cl_device_id device_id, const
 }
 
 cl_program cl_program_from_binary(cl_context ctx, cl_device_id device_id, const uint8_t* binary, size_t length, const char* args) {
-  cl_program prg = CL_CHECK_ERR(clCreateProgramWithBinary(ctx, 1, &device_id, &length, (const uint8_t*[]){binary}, NULL, &err));
+  cl_program prg = CL_CHECK_ERR(clCreateProgramWithBinary(ctx, 1, &device_id, &length, &binary, NULL, &err));
   if (int err = clBuildProgram(prg, 1, &device_id, args, NULL, NULL); err != 0) {
     cl_print_build_errors(prg, device_id);
     assert(0);

release/build_release.sh

@@ -78,7 +78,7 @@ find . -name 'moc_*' -delete
 find . -name '__pycache__' -delete
 rm -rf panda/board panda/certs panda/crypto
 rm -rf .sconsign.dblite Jenkinsfile release/
-rm selfdrive/modeld/models/supercombo.dlc
+rm selfdrive/modeld/models/supercombo.onnx
 
 # Move back signed panda fw
 mkdir -p panda/board/obj

release/files_common

@@ -352,7 +352,7 @@ selfdrive/modeld/models/driving.cc
 selfdrive/modeld/models/driving.h
 selfdrive/modeld/models/dmonitoring.cc
 selfdrive/modeld/models/dmonitoring.h
-selfdrive/modeld/models/supercombo.dlc
+selfdrive/modeld/models/supercombo.onnx
 selfdrive/modeld/models/dmonitoring_model_q.dlc
 
 selfdrive/modeld/transforms/loadyuv.cc
@@ -561,3 +561,16 @@ opendbc/vw_mqb_2010.dbc
 opendbc/tesla_can.dbc
 opendbc/tesla_radar.dbc
 opendbc/tesla_powertrain.dbc
+
+tinygrad_repo/openpilot/compile.py
+tinygrad_repo/accel/opencl/*
+tinygrad_repo/extra/onnx.py
+tinygrad_repo/extra/utils.py
+tinygrad_repo/tinygrad/llops/ops_gpu.py
+tinygrad_repo/tinygrad/llops/ops_opencl.py
+tinygrad_repo/tinygrad/helpers.py
+tinygrad_repo/tinygrad/mlops.py
+tinygrad_repo/tinygrad/ops.py
+tinygrad_repo/tinygrad/shapetracker.py
+tinygrad_repo/tinygrad/tensor.py
+tinygrad_repo/tinygrad/nn/__init__.py

selfdrive/modeld/SConscript

@@ -62,25 +62,65 @@ else:
 
 common_model = lenv.Object(common_src)
 
-# build thneed model
-if use_thneed and arch == "larch64":
-  fn = File("models/supercombo").abspath
-  compiler = lenv.Program('thneed/compile', ["thneed/compile.cc"]+common_model, LIBS=libs)
-  cmd = f"cd {Dir('.').abspath} && {compiler[0].abspath} --in {fn}.dlc --out {fn}.thneed --binary --optimize"
-
-  lib_paths = ':'.join(Dir(p).abspath for p in lenv["LIBPATH"])
-  kernel_path = os.path.join(Dir('.').abspath, "thneed", "kernels")
-  cenv = Environment(ENV={'LD_LIBRARY_PATH': f"{lib_paths}:{lenv['ENV']['LD_LIBRARY_PATH']}", 'KERNEL_PATH': kernel_path})
-
-  kernels = [os.path.join(kernel_path, x) for x in os.listdir(kernel_path) if x.endswith(".cl")]
-  cenv.Command(fn + ".thneed", [fn + ".dlc", kernels, compiler], cmd)
-
 lenv.Program('_dmonitoringmodeld', [
     "dmonitoringmodeld.cc",
     "models/dmonitoring.cc",
   ]+common_model, LIBS=libs)
 
-lenv.Program('_modeld', [
+# build thneed model
+if use_thneed and arch == "larch64" or GetOption('pc_thneed'):
+  fn = File("models/supercombo").abspath
+
+  if GetOption('pc_thneed'):
+    cmd = f"cd {Dir('#').abspath}/tinygrad_repo && NATIVE_EXPLOG=1 OPTWG=1 UNSAFE_FLOAT4=1 DEBUGCL=1 python3 openpilot/compile.py {fn}.onnx {fn}.thneed"
+  else:
+    cmd = f"cd {Dir('#').abspath}/tinygrad_repo && FLOAT16=1 PYOPENCL_NO_CACHE=1 MATMUL=1 NATIVE_EXPLOG=1 OPTWG=1 UNSAFE_FLOAT4=1 DEBUGCL=1 python3 openpilot/compile.py {fn}.onnx {fn}.thneed"
+
+  # is there a better way then listing all of tinygrad?
+  lenv.Command(fn + ".thneed", [fn + ".onnx",
+    "#tinygrad_repo/openpilot/compile.py",
+    "#tinygrad_repo/accel/opencl/conv.cl",
+    "#tinygrad_repo/accel/opencl/matmul.cl",
+    "#tinygrad_repo/accel/opencl/ops_opencl.py",
+    "#tinygrad_repo/accel/opencl/preprocessing.py",
+    "#tinygrad_repo/extra/onnx.py",
+    "#tinygrad_repo/extra/utils.py",
+    "#tinygrad_repo/tinygrad/llops/ops_gpu.py",
+    "#tinygrad_repo/tinygrad/llops/ops_opencl.py",
+    "#tinygrad_repo/tinygrad/helpers.py",
+    "#tinygrad_repo/tinygrad/mlops.py",
+    "#tinygrad_repo/tinygrad/ops.py",
+    "#tinygrad_repo/tinygrad/shapetracker.py",
+    "#tinygrad_repo/tinygrad/tensor.py",
+    "#tinygrad_repo/tinygrad/nn/__init__.py"
+  ], cmd)
+
+  # old thneed compiler. TODO: remove this once tinygrad stuff is stable
+
+  #compiler = lenv.Program('thneed/compile', ["thneed/compile.cc"]+common_model, LIBS=libs)
+  #cmd = f"cd {Dir('.').abspath} && {compiler[0].abspath} --in {fn}.dlc --out {fn}.thneed --binary --optimize"
+
+  #lib_paths = ':'.join(Dir(p).abspath for p in lenv["LIBPATH"])
+  #kernel_path = os.path.join(Dir('.').abspath, "thneed", "kernels")
+  #cenv = Environment(ENV={'LD_LIBRARY_PATH': f"{lib_paths}:{lenv['ENV']['LD_LIBRARY_PATH']}", 'KERNEL_PATH': kernel_path})
+
+  #kernels = [os.path.join(kernel_path, x) for x in os.listdir(kernel_path) if x.endswith(".cl")]
+  #cenv.Command(fn + ".thneed", [fn + ".dlc", kernels, compiler], cmd)
+
+llenv = lenv.Clone()
+if GetOption('pc_thneed'):
+  pc_thneed_src = [
+    "thneed/thneed_common.cc",
+    "thneed/thneed_pc.cc",
+    "thneed/serialize.cc",
+    "runners/thneedmodel.cc",
+  ]
+  llenv['CFLAGS'].append("-DUSE_THNEED")
+  llenv['CXXFLAGS'].append("-DUSE_THNEED")
+  common_model += llenv.Object(pc_thneed_src)
+  libs += ['dl']
+
+llenv.Program('_modeld', [
     "modeld.cc",
     "models/driving.cc",
   ]+common_model, LIBS=libs + transformations)

selfdrive/modeld/thneed/serialize.cc

@@ -33,11 +33,14 @@ void Thneed::load(const char *filename) {
       assert(mobj["needs_load"].bool_value() == false);
     } else {
       if (mobj["needs_load"].bool_value()) {
-        //printf("loading %p %d @ 0x%X\n", clbuf, sz, ptr);
         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 {
-        clbuf = clCreateBuffer(context, CL_MEM_READ_WRITE, sz, NULL, NULL);
+        // 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);

selfdrive/modeld/thneed/thneed.h

@@ -122,7 +122,7 @@ class Thneed {
 
     // all CL kernels
     void find_inputs_outputs();
-    void copy_inputs(float **finputs);
+    void copy_inputs(float **finputs, bool internal=false);
     void copy_output(float *foutput);
     cl_int clexec();
     vector<shared_ptr<CLQueuedKernel> > kq;

selfdrive/modeld/thneed/thneed_common.cc

@@ -30,17 +30,16 @@ cl_int Thneed::clexec() {
   return clFinish(command_queue);
 }
 
-void Thneed::copy_inputs(float **finputs) {
-  //cl_int ret;
+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]);
 
-    // TODO: fix thneed caching
-    if (finputs[idx] != NULL) memcpy(inputs[idx], finputs[idx], input_sizes[idx]);
-    //if (finputs[idx] != NULL) CL_CHECK(clEnqueueWriteBuffer(command_queue, input_clmem[idx], CL_TRUE, 0, input_sizes[idx], finputs[idx], 0, NULL, NULL));
-
-    // HACK
-    //if (input_sizes[idx] == 16) memset((char*)inputs[idx] + 8, 0, 8);
+    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));
+    }
   }
 }
 
@@ -202,8 +201,8 @@ void CLQueuedKernel::debug_print(bool verbose) {
             assert(slice_pitch == 0);
 
             clGetImageInfo(val, CL_IMAGE_BUFFER, sizeof(buf), &buf, NULL);
-            size_t sz;
-            clGetMemObjectInfo(buf, CL_MEM_SIZE, sizeof(sz), &sz, 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;

selfdrive/modeld/thneed/thneed_pc.cc

@@ -0,0 +1,40 @@
+#include "selfdrive/modeld/thneed/thneed.h"
+
+#include <cassert>
+
+#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);
+  }
+}
+
+void Thneed::stop() {
+}
+
+void Thneed::find_inputs_outputs() {
+  // thneed on PC doesn't work on old style inputs/outputs
+}
+

selfdrive/modeld/thneed/thneed_qcom2.cc

@@ -269,7 +269,7 @@ void Thneed::execute(float **finputs, float *foutput, bool slow) {
   if (debug >= 1) tb = nanos_since_boot();
 
   // ****** copy inputs
-  copy_inputs(finputs);
+  copy_inputs(finputs, true);
 
   // ****** set power constraint
   int ret;

selfdrive/test/process_replay/model_replay_ref_commit

@@ -1 +1 @@
-ca90e11f8d59902af38d3785ddd91a27d0fbb411
+cffb4e720b0379bedd4ff802912d998ace775c37

tinygrad

@@ -0,0 +1 @@
+tinygrad_repo/tinygrad