Update defaults to IN_WG=128,IN_SIZEX=16,OUT_WG=128,OUT_SIZEX=16

src/cl/base.cl

@@ -4,8 +4,8 @@
 
 /* Tunable paramaters for -ctune :
 
-IN_WG, OUT_WG: 64, 128, 256. Default: 256.
-IN_SIZEX, OUT_SIZEX: 4, 8, 16, 32. Default: 32 on AMD, 4 on Nvidia.
+IN_WG, OUT_WG: 64, 128, 256. Default: 128.
+IN_SIZEX, OUT_SIZEX: 4, 8, 16, 32. Default: 16.
 UNROLL_W: 0, 1. Default: 0 on AMD, 1 on Nvidia.
 UNROLL_H: 0, 1. Default: 1.
 */

src/cl/middle.cl

@@ -1,37 +1,19 @@
 // Copyright (C) Mihai Preda and George Woltman
 
 #if !IN_WG
-#define IN_WG 256
+#define IN_WG 128
 #endif
 
 #if !OUT_WG
-#define OUT_WG 256
+#define OUT_WG 128
 #endif
 
 #if !IN_SIZEX
-#if AMDGPU
-#define IN_SIZEX 32
-#else // !AMDGPU
-#if G_W >= 64
-#define IN_SIZEX 4
-#else
-#define IN_SIZEX 32
-#endif
-#endif
+#define IN_SIZEX 16
 #endif
 
 #if !OUT_SIZEX
-#if AMDGPU
-// We realized that these (OUT_WG, OUT_SIZEX) combinations work well: (256, 32) and (64, 8)
-// so default OUT_SIZEX relative to OUT_WG
-#define OUT_SIZEX (OUT_WG / 8)
-#else
-#if G_W >= 64
-#define OUT_SIZEX 4
-#else
-#define OUT_SIZEX 32
-#endif
-#endif
+#define OUT_SIZEX 16
 #endif
 
 // Parameters we may want to let user tune.  WIDTH other than 512 and 1K is untested.  SMALL_HEIGHT other than 256 and 512 is untested.