nve0: fix and improve NV12 shader

Only the 1st set of coordinates is valid.
NVC0 was fixed in 9a3e579f637267b32efc46cfce5d1a36a41323b1.

1 files changed, 23 insertions, 43 deletions

diff --git a/src/nve0_shader.h b/src/nve0_shader.h
index b727d68..4705075 100644
--- a/src/nve0_shader.h
+++ b/src/nve0_shader.h
@@ -399,7 +399,7 @@ NVE0FP_NV12[] = {
 	0x00000000,
 	0x00000000,
 	0x80000000,
-	0x00000a0a,
+	0x0000000a,
 	0x00000000,
 	0x00000000,
 	0x00000000,
@@ -413,48 +413,28 @@ NVE0FP_NV12[] = {
 	0x00000000,
 	0x0000000f,
 	0x00000000,
-	0xfff09c00,
-	0xc07e007c,
-	0x10209c00,
-	0xc8000000,
-	0x0bf01c40,
-	0xc07e0080,
-	0x0bf05c40,
-	0xc07e0084,
-	0xfc001e86,
-	0x80120000,
-	0x00001de6,
-	0xf0000000, /* texbar */
-	0x00015c40,
-	0x58004000,
-	0x1050dc20,
-	0x50004000,
-	0x20511c20,
-	0x50004000,
-	0x30515c20,
-	0x50004000,
-	0x0bf01c40,
-	0xc07e0090,
-	0x0bf05c40,
-	0xc07e0094,
-	0xfc001e86,
-	0x80130001,
-	0x00001de6,
-	0xf0000000, /* texbar */
-	0x4000dc40,
-	0x30064000,
-	0x50011c40,
-	0x30084000,
-	0x60015c40,
-	0x300a4000,
-	0x70101c40,
-	0x30064000,
-	0x90109c40,
-	0x300a4000,
-	0x80105c40,
-	0x30084000,
-	0x00001de7,
-	0x80000000,
+	0xe23282e7, 0x2c220202, /* sched 0x2e 0x28 0x23 0x2e 0x20 0x20 0xc2 */
+	0xfff09c00, 0xc07e007c, /* interp pass f32 $r2 a[0x7c] */
+	0x10209c00, 0xc8000000, /* rcp f32 $r2 $r2 */
+	0x0bf01c40, 0xc07e0080, /* interp mul f32 $r0 a[0x80] $r2 0 */
+	0x0bf05c40, 0xc07e0084, /* interp mul f32 $r1 a[0x84] $r2 0 */
+	0xfc011e86, 0x80120000, /* tex t { # # # $r4 } t2d c[0x0] { $r0 $r1 } */
+	0xfc001f06, 0x80130001, /* tex p { # # $r0 $r1 } t2d c[0x4] { $r0 $r1 } */
+	0x04001de6, 0xf0000000, /* texbar 1 */
+	0x02004287, 0x22004a22, /* sched 0x28 0x4 0x20 0x20 0xa2 0x4 0x20 */
+	0x00415c40, 0x58004000, /* mul ftz rn f32 $r5 $r4 c0[0] */
+	0x1050dc20, 0x50004000, /* add ftz rn f32 $r3 $r5 c0[0x4] */
+	0x20511c20, 0x50004000, /* add ftz rn f32 $r4 $r5 c0[0x8] */
+	0x30515c20, 0x50004000, /* add ftz rn f32 $r5 $r5 c0[0xc] */
+	0x00001de6, 0xf0000000, /* texbar 0 */
+	0x4000dc40, 0x30064000, /* fma ftz rn f32 $r3 $r0 c0[0x10] $r3 */
+	0x50011c40, 0x30084000, /* fma ftz rn f32 $r4 $r0 c0[0x14] $r4 */
+	0x42004287, 0x200002e0, /* sched 0x28 0x4 0x20 0x4 0x2e 0 0 */
+	0x60015c40, 0x300a4000, /* fma ftz rn f32 $r5 $r0 c0[0x18] $r5 */
+	0x70101c40, 0x30064000, /* fma ftz rn f32 $r0 $r1 c0[0x1c] $r3 */
+	0x90109c40, 0x300a4000, /* fma ftz rn f32 $r2 $r1 c0[0x24] $r5 */
+	0x80105c40, 0x30084000, /* fma ftz rn f32 $r1 $r1 c0[0x20] $r4 */
+	0x00001de7, 0x80000000, /* exit */
 };
 
 #endif