From eaa4feee433e05ce69c5e3926b0e8d367b4ea545 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Simon=20B=C3=BCrger?= <simon.buerger@rwth-aachen.de>
Date: Mon, 24 Apr 2023 14:58:53 +0100
Subject: [PATCH] benchmark Dslash(...) instead of full M(...)

---
 Quda/Benchmark_Quda.cpp | 18 +++++++++++++-----
 1 file changed, 13 insertions(+), 5 deletions(-)

diff --git a/Quda/Benchmark_Quda.cpp b/Quda/Benchmark_Quda.cpp
index 3a0b795..37b3dea 100644
--- a/Quda/Benchmark_Quda.cpp
+++ b/Quda/Benchmark_Quda.cpp
@@ -106,21 +106,29 @@ cudaGaugeField make_gauge_field(int L)
 // create a random source vector (L = local size)
 ColorSpinorField make_source(int L)
 {
+  // NOTE: `param.x` directly determines the size of the (local, per rank) memory
+  // allocation. Thus for checkerboarding, we have to specifly x=(L/2,L,L,L) to get a
+  // physical local volume of L^4, thus implicity choosing a dimension for the
+  // checkerboarding (shouldnt really matter of course which one).
   ColorSpinorParam param;
   param.nColor = 3;
   param.nSpin = 4;
   param.nVec = 1; // only a single vector
   param.pad = 0;
-  param.siteSubset = QUDA_FULL_SITE_SUBSET;
+  param.siteSubset = QUDA_PARITY_SITE_SUBSET;
   param.nDim = 4;
-  param.x[0] = L;
+  param.x[0] = L / 2;
   param.x[1] = L;
   param.x[2] = L;
   param.x[3] = L;
   param.x[4] = 1; // no fifth dimension
   param.pc_type = QUDA_4D_PC;
   param.siteOrder = QUDA_EVEN_ODD_SITE_ORDER;
-  param.gammaBasis = QUDA_DEGRAND_ROSSI_GAMMA_BASIS;
+
+  // somewhat surprisingly, the DiracWilson::Dslash(...) function only works with the
+  // UKQCD_GAMMA_BASIS
+  param.gammaBasis = QUDA_UKQCD_GAMMA_BASIS;
+
   param.create = QUDA_NULL_FIELD_CREATE; // do not (zero-) initilize the field
   param.setPrecision(QUDA_DOUBLE_PRECISION);
   param.location = QUDA_CUDA_FIELD_LOCATION;
@@ -168,14 +176,14 @@ void benchmark_wilson()
 
     // couple iterations without timing to warm up
     for (int iter = 0; iter < niter_warmup; ++iter)
-      dirac.M(tmp, src);
+      dirac.Dslash(tmp, src, QUDA_EVEN_PARITY);
 
     // actual benchmark with timings
     dirac.Flops(); // reset flops counter
     device_timer_t device_timer;
     device_timer.start();
     for (int iter = 0; iter < niter; ++iter)
-      dirac.M(tmp, src);
+      dirac.Dslash(tmp, src, QUDA_EVEN_PARITY);
     device_timer.stop();
 
     double secs = device_timer.last() / niter;