[GitHub] [incubator-tvm] tkonolige commented on a change in pull request #6533: Scatter on Cuda

[GitBox <gi...@apache.org>]

tkonolige commented on a change in pull request #6533:
URL: https://github.com/apache/incubator-tvm/pull/6533#discussion_r502532481



##########
File path: python/tvm/topi/cuda/scatter.py
##########
@@ -0,0 +1,444 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=invalid-name, no-member, too-many-locals, too-many-arguments, too-many-statements, singleton-comparison, unused-argument
+"""Scatter operator """
+import tvm
+from tvm import te
+
+
+def ceil_div(a, b):
+    return (a + b - 1) // b
+
+
+def gen_ir_1d(data, indices, updates, axis, out):
+    """Generate scatter ir for 1d inputs
+
+    Parameters
+    ----------
+    data : tir.Tensor
+        The input data to the operator.
+
+    indices : tir.Tensor
+        The index locations to update.
+
+    updates : tir.Tensor
+        The values to update.
+
+    axis : int
+        The axis to scatter on
+
+    out : tir.Tensor
+        The output tensor.
+
+    Returns
+    -------
+    ret : tir
+        The computational ir.
+    """
+    assert axis == 0
+    n = data.shape[0]
+
+    ib = tvm.tir.ir_builder.create()
+
+    out_ptr = ib.buffer_ptr(out)
+    data_ptr = ib.buffer_ptr(data)
+
+    with ib.new_scope():
+        bx = te.thread_axis("blockIdx.x")
+        ib.scope_attr(bx, "thread_extent", n)
+        out_ptr[bx] = data_ptr[bx]
+
+    indices_ptr = ib.buffer_ptr(indices)
+    updates_ptr = ib.buffer_ptr(updates)
+
+    ni = indices.shape[0]
+
+    with ib.new_scope():
+        bx = te.thread_axis("blockIdx.x")
+        ib.scope_attr(bx, "thread_extent", 1)
+        with ib.for_range(0, ni, name="i") as i:
+            index = indices_ptr[i]
+            with ib.if_scope(index < 0):
+                out_ptr[index + n] = updates_ptr[i]
+            with ib.else_scope():
+                out_ptr[index] = updates_ptr[i]
+
+    return ib.get()
+
+
+def gen_ir_2d(data, indices, updates, axis, out):
+    """Generate scatter ir for 2d inputs
+
+    Parameters
+    ----------
+    data : tir.Tensor
+        The input data to the operator.
+
+    indices : tir.Tensor
+        The index locations to update.
+
+    updates : tir.Tensor
+        The values to update.
+
+    axis : int
+        The axis to scatter on
+
+    out : tir.Tensor
+        The output tensor.
+
+    Returns
+    -------
+    ret : tir
+        The computational ir.
+    """
+    warp_size = tvm.target.Target.current(False).thread_warp_size
+
+    n = data.shape[0]
+    c = data.shape[1]
+
+    ib = tvm.tir.ir_builder.create()
+
+    out_ptr = ib.buffer_ptr(out)
+    data_ptr = ib.buffer_ptr(data)
+
+    with ib.new_scope():
+        bx = te.thread_axis("blockIdx.x")
+        ib.scope_attr(bx, "thread_extent", n)
+        tx = te.thread_axis("threadIdx.x")
+        ib.scope_attr(tx, "thread_extent", warp_size)
+        i = bx
+        with ib.for_range(0, ceil_div(c, warp_size), name="j") as j_:

Review comment:
       Yeah, that would work too. I'm not sure what would be faster. My guess is they are equivalent unless n is really small and c is large (in which case the current code would be worse).




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org