cluda_opencl: Use round to nearest even to match cpu version

jvesely · jvesely · commit ac58a7907147 · 2018-04-04T21:48:08.000-04:00
Fixes test_elemwise_f16 add/iadd on amdgcn and possibly other OpenCL implementations. Fixes Theano#462 Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
diff --git a/src/cluda_opencl.h b/src/cluda_opencl.h
@@ -50,7 +50,7 @@ typedef struct _ga_half {
 #define ga_half2float(p) vload_half(0, &((p).data))
 static inline ga_half ga_float2half(ga_float f) {
   ga_half r;
-  vstore_half_rtn(f, 0, &r.data);
+  vstore_half_rte(f, 0, &r.data);
   return r;
 }
 
diff --git a/src/cluda_opencl.h.c b/src/cluda_opencl.h.c
@@ -122,7 +122,7 @@ static const char cluda_opencl_h[] = {
 0x6f, 0x61, 0x74, 0x20, 0x66, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
 0x67, 0x61, 0x5f, 0x68, 0x61, 0x6c, 0x66, 0x20, 0x72, 0x3b, 0x0a,
 0x20, 0x20, 0x76, 0x73, 0x74, 0x6f, 0x72, 0x65, 0x5f, 0x68, 0x61,
-0x6c, 0x66, 0x5f, 0x72, 0x74, 0x6e, 0x28, 0x66, 0x2c, 0x20, 0x30,
+0x6c, 0x66, 0x5f, 0x72, 0x74, 0x65, 0x28, 0x66, 0x2c, 0x20, 0x30,
 0x2c, 0x20, 0x26, 0x72, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b,
 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x72,
 0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x23, 0x70, 0x72, 0x61, 0x67, 0x6d,

Original file line number	Diff line number	Diff line change
`@@ -50,7 +50,7 @@ typedef struct _ga_half {`
`50`	`50`	`#define ga_half2float(p) vload_half(0, &((p).data))`
`51`	`51`	`static inline ga_half ga_float2half(ga_float f) {`
`52`	`52`	`ga_half r;`
`53`		`- vstore_half_rtn(f, 0, &r.data);`
	`53`	`+ vstore_half_rte(f, 0, &r.data);`
`54`	`54`	`return r;`
`55`	`55`	`}`
`56`	`56`