CaptainDuke commented on a change in pull request #8479:
URL: https://github.com/apache/tvm/pull/8479#discussion_r677961390
##########
File path: python/tvm/topi/cuda/scatter.py
##########
@@ -789,42 +790,78 @@ def gen_ir(data_ptr, indices_ptr, updates_ptr, out_ptr):
# For now we avoid parallizing over dimensions indexed by `indices` as
# there may be repeated indices and hadling parallel accumulation can
- # be hard. So we parallelize over X_M .. X_{N-1} instead. This will
- # work well when these dimensions are large enough to saturate memory
- # bandwidth, but performance will be bad when these dimensions are
- # small.
- bx = te.thread_axis("blockIdx.x")
- tx = te.thread_axis("threadIdx.x")
+ # be hard. So we parallelize over X_M .. X_{N-1} instead.
+
+ # For better performance, we introduce blockIdx.y to implement
for-loops
+ # within one thread.
+ # The code is parallel over the scattered indices, so we use atomic_add
+ # to guarantee correctness when mode=="add"
+
max_threads =
int(tvm.target.Target.current(allow_none=False).max_num_threads)
tdim = min(max_threads, fused_updates_dimension)
- ib.scope_attr(tx, "thread_extent", tdim)
- bdim = ceil_div(fused_updates_dimension, tdim)
- ib.scope_attr(bx, "thread_extent", bdim)
-
- # Copy data into the output. This loop writes to the same portions of
- # memory as the following loop, so we do not need a memory sync.
- with ib.for_range(0, ceil_div(fused_shape, fused_updates_dimension),
name="i") as i:
- index = i * fused_updates_dimension + bx * tdim + tx
- with ib.if_scope(bx * tdim + tx < fused_updates_dimension):
+
+ with ib.new_scope():
+ bdim = ceil_div(fused_shape, tdim)
+ bx = te.thread_axis("blockIdx.x")
+ tx = te.thread_axis("threadIdx.x")
+ ib.scope_attr(bx, "thread_extent", bdim)
+ ib.scope_attr(tx, "thread_extent", tdim)
+
+ index = bx * tdim + tx
+ with ib.if_scope(index < fused_shape):
out[index] = data[index]
- with ib.for_range(0, fused_indices_dimension) as i:
- j = bx * tdim + tx
- with ib.if_scope(j < fused_updates_dimension):
- offset = fused_updates_dimension
- index = j # This is x_M, .. x_{N-1} part of the index into
out.
- # Build up the indices[0, y_0, .. y_{K-1}], .. indices[M-1,
y_0, .. y_{K-1}] part
- # of the index into out.
- for l in reversed(range(indices_ptr.shape[0].value)):
- # indices[i * l * fused_indices_dimension] = indices[l,
y_0, ... y_{k-1}]
- index += offset * indices[i + l * fused_indices_dimension]
- offset *= data_ptr.shape[l]
- if mode == "update":
- out[index] = updates[i * fused_updates_dimension + j]
- elif mode == "add":
- out[index] += updates[i * fused_updates_dimension + j]
- else:
- raise NotImplementedError("scatter_nd mode not in [update,
add]:", mode)
+ with ib.new_scope():
+ if updates.dtype == "int64" and mode == "add":
+ bdim_x = ceil_div(fused_updates_dimension, tdim)
+ bx = te.thread_axis("blockIdx.x")
+ tx = te.thread_axis("threadIdx.x")
+ ib.scope_attr(bx, "thread_extent", bdim_x)
+ ib.scope_attr(tx, "thread_extent", tdim)
+ with ib.for_range(0, fused_indices_dimension) as i:
+ j = bx * tdim + tx
+ with ib.if_scope(j < fused_updates_dimension):
+ offset = fused_updates_dimension
+ index = j # This is x_M, .. x_{N-1} part of the index
into out.
+ # Build up the
+ # indices[0, y_0, .. y_{K-1}], ... indices[M-1, y_0,
.. y_{K-1}]
+ # part of the index into out.
+ for l in reversed(range(indices_ptr.shape[0].value)):
+ # indices[i * l * fused_indices_dimension] =
indices[l, y_0,
+ #
... y_{k-1}]
+ index += offset * indices[i + l *
fused_indices_dimension]
+ offset *= data_ptr.shape[l]
+ out[index] += updates[i * fused_updates_dimension + j]
+ else:
+ bdim_x = ceil_div(fused_updates_dimension, tdim)
+ bdim_y = fused_indices_dimension
Review comment:
```
840 bx = te.thread_axis("blockIdx.y")
841 by = te.thread_axis("blockIdx.x")
```
According to [Maximum x-dimension of a grid of thread blocks = 2^31
-1](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#features-and-technical-specifications),
I exchange `blockIdx.x` and `blockIdx.y` to avoid out of bounds.
For the given input sizes, performance on `mode="add"` was
time: *784638840*
grid=(1,1,1), block=(1,1,1)
**_v.s._**
time: *105102068 + 2141897 = 107243965*.
grid=(34725600,1,1), block=(1,1,1)
grid=(551200,1,1), block=(1,1,1)
__7.3x faster__
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
