From 07ca1ee6df51ee463fdf321c2b4b2e3c7673ddf3 Mon Sep 17 00:00:00 2001
From: Isuru Fernando <isuruf@gmail.com>
Date: Fri, 6 Jan 2023 06:38:33 +0530
Subject: [PATCH] Use a vec array for local_isrc* for coalesced array access
(#151)
* Merge local_isrc and local_isrc_strength and tag as vec for coalescced access
* Use new name
* Add an explanation about the optimization
* Back to loopy main with renamed transform
Co-authored-by: Andreas Kloeckner <inform@tiker.net>
---
sumpy/p2p.py | 12 ++++++++++++
1 file changed, 12 insertions(+)
diff --git a/sumpy/p2p.py b/sumpy/p2p.py
index c24fa8d9..385e90b0 100644
--- a/sumpy/p2p.py
+++ b/sumpy/p2p.py
@@ -656,6 +656,18 @@ class P2PFromCSR(P2PBase):
knl = lp.tag_inames(knl, {"itgt_box": "g.0", "inner": "l.0"})
knl = lp.set_temporary_address_space(knl,
["local_isrc", "local_isrc_strength"], lp.AddressSpace.LOCAL)
+
+ # By having a concatenated memory layout of the temporaries
+ # and marking the first axis as vec, we are transposing the
+ # the arrays and also making the access of the source
+ # co-ordinates and the strength for each source a coalesced
+ # access of 256 bits (assuming double precision) which is
+ # optimized for NVIDIA GPUs. On an NVIDIA Titan V, this
+ # optimization led to a 8% speedup in the performance.
+ knl = lp.concatenate_arrays(knl,
+ ["local_isrc", "local_isrc_strength"], "local_isrc")
+ knl = lp.tag_array_axes(knl, "local_isrc", "vec,C")
+
knl = lp.add_inames_for_unused_hw_axes(knl)
# knl = lp.set_options(knl, write_code=True)
--
GitLab