From 93836c0e3c9111eb57d539ad0ae9e160117742a8 Mon Sep 17 00:00:00 2001
From: mclark <mclark@nvidia.com>
Date: Thu, 16 Jul 2015 22:25:01 +0900
Subject: [PATCH] Fixed memory bandwidth computations for 4-d preconditioned
 5-d dslash operators (closes #315).

---
 lib/dslash_domain_wall_4d.cu | 22 ++++++++++++++++++++++
 lib/dslash_mobius.cu         | 23 +++++++++++++++++++++++
 2 files changed, 45 insertions(+)

diff --git a/lib/dslash_domain_wall_4d.cu b/lib/dslash_domain_wall_4d.cu
index 15730330be..5051392c36 100644
--- a/lib/dslash_domain_wall_4d.cu
+++ b/lib/dslash_domain_wall_4d.cu
@@ -215,6 +215,28 @@ namespace quda {
       }
       return flops;
     }
+
+    long long bytes() const {
+      bool isHalf = in->Precision() == sizeof(short) ? true : false;
+      int spinor_bytes = 2 * in->Ncolor() * in->Nspin() * in->Precision() + (isHalf ? sizeof(float) : 0);
+      long long Ls = in->X(4);
+      long long bytes;
+
+      switch(DS_type){
+      case 0:
+	bytes = DslashCuda::bytes();
+	break;
+      case 1:
+	bytes = (x ? 5ll : 4ll ) * spinor_bytes * in->VolumeCB();
+	break;
+      case 2:
+	bytes = (x ? Ls + 2 : Ls + 1) * spinor_bytes * in->VolumeCB();
+	break;
+      default:
+	errorQuda("invalid Dslash type");
+      }
+      return bytes;
+    }
   };
 #endif // GPU_DOMAIN_WALL_DIRAC
 
diff --git a/lib/dslash_mobius.cu b/lib/dslash_mobius.cu
index 144f79b7b8..ed5dadd68f 100644
--- a/lib/dslash_mobius.cu
+++ b/lib/dslash_mobius.cu
@@ -228,6 +228,29 @@ namespace quda {
       }
       return flops;
     }
+
+    long long bytes() const {
+      bool isHalf = in->Precision() == sizeof(short) ? true : false;
+      int spinor_bytes = 2 * in->Ncolor() * in->Nspin() * in->Precision() + (isHalf ? sizeof(float) : 0);
+      long long Ls = in->X(4);
+      long long bytes;
+
+      switch(DS_type){
+      case 0:
+	bytes = DslashCuda::bytes();
+	break;
+      case 1:
+      case 2:
+	bytes = (x ? 5ll : 4ll) * spinor_bytes * in->VolumeCB();
+	break;
+      case 3:
+	bytes = (x ? Ls + 2 : Ls + 1) * spinor_bytes * in->VolumeCB();
+	break;
+      default:
+	errorQuda("invalid Dslash type");
+      }
+      return bytes;
+    }
   };
 #endif // GPU_DOMAIN_WALL_DIRAC