From 07b67b894cfb71615658843b119c1d3e6357f803 Mon Sep 17 00:00:00 2001
From: Joongi Kim <joongi@an.kaist.ac.kr>
Date: Fri, 22 Jan 2016 17:36:35 +0900
Subject: [PATCH] refs #6: Refactor all occurrences of ShiftedInt to
 dev_offset_t.

 * This eases experiments to compare performance by different offset sizes.

 * Confirmed performance drop when dev_offset_t is ShiftedInt<uint32_t, 0>.
---
 include/nba/core/shiftedint.hh     | 4 ++++
 include/nba/engines/cuda/compat.hh | 4 ++--
 include/nba/framework/datablock.hh | 8 ++++----
 src/lib/offloadtask.cc             | 4 ++--
 4 files changed, 12 insertions(+), 8 deletions(-)

diff --git a/include/nba/core/shiftedint.hh b/include/nba/core/shiftedint.hh
index 73f9d9f..5c098a3 100644
--- a/include/nba/core/shiftedint.hh
+++ b/include/nba/core/shiftedint.hh
@@ -194,6 +194,10 @@ private:
     }
 };
 
+// Changing this to ShiftedInt<uint32_t, 0> or uint32_t exactly
+// reproduces the performance drop when we first changed offset types.
+typedef ShiftedInt<uint16_t, 2> dev_offset_t;
+
 } /* endns(nba) */
 
 #endif /* __NBA_CORE_SHIFTEDINT_HH__ */
diff --git a/include/nba/engines/cuda/compat.hh b/include/nba/engines/cuda/compat.hh
index a170917..5419c32 100644
--- a/include/nba/engines/cuda/compat.hh
+++ b/include/nba/engines/cuda/compat.hh
@@ -18,8 +18,8 @@ struct datablock_batch_info {
     uint32_t item_count_out;
     uint16_t *item_sizes_in;
     uint16_t *item_sizes_out;
-    nba::ShiftedInt<uint16_t, 2> *item_offsets_in;
-    nba::ShiftedInt<uint16_t, 2> *item_offsets_out;
+    nba::dev_offset_t *item_offsets_in;
+    nba::dev_offset_t *item_offsets_out;
 }; // __cuda_aligned
 
 struct datablock_kernel_arg {
diff --git a/include/nba/framework/datablock.hh b/include/nba/framework/datablock.hh
index b9bfaba..c36a474 100644
--- a/include/nba/framework/datablock.hh
+++ b/include/nba/framework/datablock.hh
@@ -80,7 +80,7 @@ struct item_size_info {
         uint16_t size;
         uint16_t sizes[NBA_MAX_COMP_BATCH_SIZE * 12];
     };
-    ShiftedInt<uint16_t, 2>[NBA_MAX_COMP_BATCH_SIZE * 12];
+    dev_offset_t offsets[NBA_MAX_COMP_BATCH_SIZE * 12];
 };
 #else
 struct item_size_info {
@@ -88,7 +88,7 @@ struct item_size_info {
         uint16_t size;
         uint16_t sizes[NBA_MAX_COMP_BATCH_SIZE * 96];
     };
-    ShiftedInt<uint16_t, 2> offsets[NBA_MAX_COMP_BATCH_SIZE * 96];
+    dev_offset_t offsets[NBA_MAX_COMP_BATCH_SIZE * 96];
 };
 #endif
 
@@ -126,8 +126,8 @@ struct datablock_batch_info {
     uint32_t item_count_out;
     uint16_t *item_sizes_in;
     uint16_t *item_sizes_out;
-    ShiftedInt<uint16_t, 2> *item_offsets_in;
-    ShiftedInt<uint16_t, 2> *item_offsets_out;
+    dev_offset_t *item_offsets_in;
+    dev_offset_t *item_offsets_out;
 }; // __cuda_aligned
 
 /**
diff --git a/src/lib/offloadtask.cc b/src/lib/offloadtask.cc
index 0a0634a..151e256 100644
--- a/src/lib/offloadtask.cc
+++ b/src/lib/offloadtask.cc
@@ -235,10 +235,10 @@ bool OffloadTask::copy_h2d()
                 dbarg_h->batches[b].item_sizes_out = (uint16_t *)
                         ((char *) t->aligned_item_sizes_d.ptr
                          + (uintptr_t) offsetof(struct item_size_info, sizes));
-                dbarg_h->batches[b].item_offsets_in = (ShiftedInt<uint16_t, 2> *)
+                dbarg_h->batches[b].item_offsets_in = (dev_offset_t *)
                         ((char *) t->aligned_item_sizes_d.ptr
                          + (uintptr_t) offsetof(struct item_size_info, offsets));
-                dbarg_h->batches[b].item_offsets_out = (ShiftedInt<uint16_t, 2> *)
+                dbarg_h->batches[b].item_offsets_out = (dev_offset_t *)
                         ((char *) t->aligned_item_sizes_d.ptr
                          + (uintptr_t) offsetof(struct item_size_info, offsets));
             } else {