Scatter on Cuda (apache#6533)

* working cuda scatter

fix lint

fix pylint again

* cuda scatter with threading

* add dynamic shape tests

* remove unused variable

python/tvm/relay/op/_transform.py

@@ -104,7 +104,7 @@ def compute_scatter(attrs, inputs, output_type):
     return [topi.scatter(inputs[0], inputs[1], inputs[2], attrs.axis)]
 
 
-_reg.register_schedule("scatter", strategy.schedule_scatter)
+_reg.register_strategy("scatter", strategy.scatter_strategy)
 
 # scatter_add
 @_reg.register_compute("scatter_add")

python/tvm/relay/op/strategy/cuda.py

@@ -651,6 +651,19 @@ def sparse_dense_padded_strategy_cuda(attrs, inputs, out_type, target):
     return strategy
 
 
+@scatter_strategy.register(["cuda", "gpu"])
+def scatter_cuda(attrs, inputs, out_type, target):
+    """sparse dense cuda strategy"""
+    strategy = _op.OpStrategy()
+    strategy.add_implementation(
+        wrap_compute_scatter(topi.cuda.scatter),
+        wrap_topi_schedule(topi.generic.schedule_extern),
+        name="scatter.cuda",
+        plevel=10,
+    )
+    return strategy
+
+
 @argsort_strategy.register(["cuda", "gpu"])
 def argsort_strategy_cuda(attrs, inputs, out_type, target):
     """argsort cuda strategy"""

python/tvm/relay/op/strategy/generic.py

@@ -1032,11 +1032,24 @@ def schedule_argwhere(attrs, outs, target):
 
 
 # scatter
-@generic_func
-def schedule_scatter(attrs, outs, target):
-    """schedule scatter"""
-    with target:
-        return topi.generic.schedule_scatter(outs)
+@override_native_generic_func("scatter_strategy")
+def scatter_strategy(attrs, outs, out_type, target):
+    strategy = _op.OpStrategy()
+    strategy.add_implementation(
+        wrap_compute_scatter(topi.scatter),
+        wrap_topi_schedule(topi.generic.schedule_scatter),
+        name="scatter.generic",
+    )
+    return strategy
+
+
+def wrap_compute_scatter(topi_compute):
+    """Wrap scatter topi compute"""
+
+    def _compute_scatter(attrs, inputs, _):
+        return [topi_compute(inputs[0], inputs[1], inputs[2], axis=attrs.axis)]
+
+    return _compute_scatter
 
 
 # scatter_add

python/tvm/topi/cuda/__init__.py

@@ -46,6 +46,7 @@
 from .ssd import *
 from .nms import get_valid_counts, non_max_suppression
 from .rcnn import *
+from .scatter import *
 from .sort import *
 from .conv2d_nhwc_tensorcore import *
 from .conv3d_ndhwc_tensorcore import *