From 7f06cf219ce05de1b463e53fbdc6c14db4b97d61 Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Thu, 8 Nov 2012 23:57:28 -0600
Subject: [PATCH] Introduce work group size floor in scan.
---
pyopencl/scan.py | 15 +++++++++++++--
1 file changed, 13 insertions(+), 2 deletions(-)
diff --git a/pyopencl/scan.py b/pyopencl/scan.py
index 3a82172a..af428afe 100644
--- a/pyopencl/scan.py
+++ b/pyopencl/scan.py
@@ -975,7 +975,6 @@ class GenericScanKernel(_GenericScanKernelBase):
trip_count = 0
-
if self.devices[0].type == cl.device_type.CPU:
# (about the widest vector a CPU can support, also taking
# into account that CPUs don't hide latency by large work groups
@@ -1002,6 +1001,19 @@ class GenericScanKernel(_GenericScanKernelBase):
wg_size, k_group_size) + 256 <= avail_local_mem):
solutions.append((wg_size*k_group_size, k_group_size, wg_size))
+ if self.devices[0].type == cl.device_type.GPU:
+ from pytools import any
+ for wg_size_floor in [256, 192, 128]:
+ have_sol_above_floor = any(wg_size >= wg_size_floor
+ for _, _, wg_size in solutions)
+
+ if have_sol_above_floor:
+ # delete all the others
+ solutions = [(total, k_group_size, wg_size)
+ for total, k_group_size, wg_size in solutions
+ if wg_size >= wg_size_floor]
+ break
+
_, k_group_size, max_scan_wg_size = max(solutions)
while True:
@@ -1140,7 +1152,6 @@ class GenericScanKernel(_GenericScanKernelBase):
wg_size = _round_down_to_power_of_2(
min(max_wg_size, 128))
-
scan_tpl = _make_template(SCAN_INTERVALS_SOURCE)
scan_src = str(scan_tpl.render(
wg_size=wg_size,
--
GitLab