Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Disable GPU accelerated row-column transpose for Pascal GPUs: #5122

Merged
merged 3 commits into from
Apr 5, 2022
Merged

Disable GPU accelerated row-column transpose for Pascal GPUs: #5122

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
764e959
Disable GPU accelerated row-column transpose for Pascal GPUs:
mythrocks Apr 1, 2022
393ad54
Review feedback:
mythrocks Apr 1, 2022
dc6a561
Review feedback:
mythrocks Apr 4, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 20 additions & 2 deletions sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuColumnarToRowExec.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,7 +19,7 @@ package com.nvidia.spark.rapids
import scala.annotation.tailrec
import scala.collection.mutable.Queue

import ai.rapids.cudf.{HostColumnVector, NvtxColor, Table}
import ai.rapids.cudf.{Cuda, HostColumnVector, NvtxColor, Table}
import com.nvidia.spark.rapids.GpuColumnarToRowExecParent.makeIteratorFunc
import com.nvidia.spark.rapids.shims.ShimUnaryExecNode

Expand Down Expand Up @@ -343,6 +343,19 @@ object GpuColumnarToRowExecParent {
Option(Tuple2(arg.child, arg.exportColumnarRdd))
}

/**
* Helper to check if GPU accelerated row-column transpose is supported.
* This is a workaround for [[https://github.com/rapidsai/cudf/issues/10569]],
* where CUDF JNI column->row transposition works incorrectly on certain
* GPU architectures.
*/
private lazy val isAcceleratedTransposeSupported: Boolean = {
// Check if the current CUDA device architecture exceeds Pascal.
// i.e. CUDA compute capability > 6.x.
// Reference: https://developer.nvidia.com/cuda-gpus
Cuda.getComputeCapabilityMajor > 6
}

def makeIteratorFunc(
output: Seq[Attribute],
numOutputRows: GpuMetric,
Expand All @@ -356,7 +369,12 @@ object GpuColumnarToRowExecParent {
// This number includes the 1-bit validity per column, but doesn't include padding.
// We are being conservative by only allowing 100M columns until we feel the need to
// increase this number
output.length <= 100000000) {
output.length <= 100000000 &&
// Work around {@link https://github.com/rapidsai/cudf/issues/10569}, where CUDF JNI
// acceleration of column->row transposition produces incorrect results on certain
// GPU architectures.
// Check that the accelerated transpose works correctly on the current CUDA device.
isAcceleratedTransposeSupported) {
(batches: Iterator[ColumnarBatch]) => {
// UnsafeProjection is not serializable so do it on the executor side
val toUnsafe = UnsafeProjection.create(output, output)
Expand Down