[Beignet] [PATCH] GBE: Fix bitcast between long and other type.

[Ruiling Song]

As we store long low/high 32bits separately, when we do bitcast
like int64 --> int16, the horizontal stride of the int64's low/high
half should be set as 2 instead of 4.

This fix an regression of opencv test:
Imgproc/Threshold.Mat/40, where GetParam() = (16SC1, 0, 0, false)

Signed-off-by: Ruiling Song <ruiling.song at intel.com>
---
 backend/src/backend/gen_insn_selection.cpp |   11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/backend/src/backend/gen_insn_selection.cpp b/backend/src/backend/gen_insn_selection.cpp
index 74f7cfe..6e6e6eb 100644
--- a/backend/src/backend/gen_insn_selection.cpp
+++ b/backend/src/backend/gen_insn_selection.cpp
@@ -3081,7 +3081,7 @@ namespace gbe
           wideReg = sel.selReg(insn.getDst(index/multiple), narrowType);
           narrowReg = sel.selReg(insn.getSrc(i), narrowType);  //retype to narrow type
         }
-        if(wideReg.hstride != GEN_VERTICAL_STRIDE_0) {
+        if(wideReg.hstride != GEN_HORIZONTAL_STRIDE_0) {
           if(multiple == 2) {
             wideReg = sel.unpacked_uw(wideReg.reg());
             wideReg = GenRegister::retype(wideReg, getGenType(narrowType));
@@ -3098,6 +3098,15 @@ namespace gbe
           wideReg.subphysical = 1;
         }
         if(isInt64) {
+          if(wideReg.hstride != GEN_HORIZONTAL_STRIDE_0) {
+            // as we store long by bottom & high part separately, we have to divide hstride by 2
+            if (wideReg.hstride == GEN_HORIZONTAL_STRIDE_2)
+              wideReg.hstride = GEN_HORIZONTAL_STRIDE_1;
+            else if (wideReg.hstride == GEN_HORIZONTAL_STRIDE_4)
+              wideReg.hstride = GEN_HORIZONTAL_STRIDE_2;
+            else
+              GBE_ASSERT(0);
+          }
           // offset to next half
           wideReg.subphysical = 1;
           if(i >= multiple/2)
-- 
1.7.10.4