fix nonstatic weights

accel/opencl/preprocessing.py

@@ -62,7 +62,12 @@ def preprocessing_op(x,w,C):
   w = w.contiguous_op()
 
   # early realize on the weights
-  w.realize().image
+  bw = w
+  while getattr(bw, 'op', None) and len(bw.op.src) == 1:
+    bw = bw.op.src[0]
+  if bw.realized:
+    # weights are static
+    w.realize().image
   return x,w,C
 
 def postprocessing_op(ret, C, C_initial):

extra/onnx.py

@@ -87,10 +87,7 @@ def get_run_onnx(onnx_model):
       elif n.op_type == "Sigmoid": ret = inp[0].sigmoid()
       elif n.op_type == "Tanh": ret = inp[0].tanh()
       elif n.op_type == "Softmax": ret = inp[0].softmax()
-      elif n.op_type == "MatMul":
-        if inp[1].lazydata.realized is not None:
-          print("WARNING: matmul weights are not static")
-        ret = inp[0].matmul(inp[1])
+      elif n.op_type == "MatMul": ret = inp[0].matmul(inp[1])
       # one liners
       elif n.op_type == "Elu": ret = inp[0].elu(alpha=opt['alpha'])
       elif n.op_type == "Clip": ret = inp[0].clip(*(inp[1:] if len(inp) > 1 else (opt.get('min', -3.4e38), opt.get('max', 3.4e38))))

extra/thneed.py

@@ -281,7 +281,8 @@ class Thneed:
     local_cl_cache = []
     for prg, args in self.cl_cache:
       args = list(args)
-      if args[1] is None and len(args[0]) == 2:
+      # TODO: WTF is wrong with to_image?
+      if args[1] is None and len(args[0]) == 2 and 'to_image' not in prg.name:
         args[1] = [min(MAX_WORKGROUP, args[0][0]), 1]
         try:
           e = prg.clprg(CL().cl_queue, *args)

openpilot/compile.py

@@ -93,6 +93,7 @@ def compile(dat, output_fn):
 
   # real run
   inputs, np_inputs = get_random_input_tensors(input_shapes)
+  print("***** REAL RUN *****")
   tinygrad_out = run_onnx(inputs)['outputs']
 
   # note, since CL.CACHE is enabled, it doesn't actually run the kernels