From 5a2c1af2387c35dd5dd6ce39ce3d339ef2504efb Mon Sep 17 00:00:00 2001
From: "Rebecca N. Palmer" <rebecca_palmer@zoho.com>
Date: Sat, 11 Aug 2018 17:16:11 +0100
Subject: [PATCH] Disable use_bank_conflict_avoidance only on beignet, improve
comment
---
pyopencl/scan.py | 11 ++++++++---
1 file changed, 8 insertions(+), 3 deletions(-)
diff --git a/pyopencl/scan.py b/pyopencl/scan.py
index c17acb2c..ab8aee30 100644
--- a/pyopencl/scan.py
+++ b/pyopencl/scan.py
@@ -1227,12 +1227,17 @@ class GenericScanKernel(_GenericScanKernelBase):
max_scan_wg_size = min(dev.max_work_group_size for dev in self.devices)
wg_size_multiples = 64
- # Intel beignet asserts or gives wrong results with packed structs
+ # Intel beignet fails "Out of shared local memory" in test_scan int64
+ # and asserts in test_sort with this enabled:
+ # https://github.com/inducer/pyopencl/pull/238
+ # A beignet bug report (outside of pyopencl) suggests packed structs
+ # (which this is) can even give wrong results:
# https://bugs.freedesktop.org/show_bug.cgi?id=98717
- # TODO: is this all Intel ICDs or only beignet?
+ # TODO: does this also affect Intel Compute Runtime?
use_bank_conflict_avoidance = (
self.dtype.itemsize > 4 and self.dtype.itemsize % 8 == 0
- and is_gpu and "Intel" not in self.devices[0].platform.name)
+ and is_gpu
+ and "beignet" not in self.devices[0].platform.version.lower())
# k_group_size should be a power of two because of in-kernel
# division by that number.
--
GitLab