Compute du/dv/dw in no-mask SF prog

1 files changed, 70 insertions, 8 deletions

diff --git a/src/exa_sf.g4a b/src/exa_sf.g4a
index 5a023992..4da5ebac 100644
--- a/src/exa_sf.g4a
+++ b/src/exa_sf.g4a
@@ -26,14 +26,76 @@
  *
  */
 
-send (1) 0 g6<1>F g1.12<0,1,0>F math inv scalar mlen 1 rlen 1 { align1 };
-send (1) 0 g6.4<1>F g1.20<0,1,0>F math inv scalar mlen 1 rlen 1 { align1 };
-add (8) g7<1>F g4<8,8,1>F -g3<8,8,1>F { align1 };
-mul (1) g7<1>F g7<0,1,0>F g6<0,1,0>F { align1 };
-mul (1) g7.4<1>F g7.4<0,1,0>F g6.4<0,1,0>F { align1 };
-mov (8) m1<1>F g7<0,1,0>F { align1 };
-mov (8) m2<1>F g7.4<0,1,0>F { align1 };
-mov (8) m3<1>F g3<8,8,1>F { align1 };
+/*
+ * Inputs (note all sub-register addresses are bytes, not float indices)
+ *
+ * Note that the vertices will have been reordered:
+ *
+ * V0 is topmost (leftmost among topmost) (upper left)
+ * V1 is next clockwise (lower right)
+ * V2 is remaining (lower left)
+ *
+ *  V0 ...................... XX
+ *  |                          .
+ *  |                          .
+ *  |                          .
+ *  V2------------------------V1
+ *
+ *  G0	    thread state -- just pass along
+ *
+ *  G1 and G2 are fixed by SF spec
+ *
+ *  G1.0    reserved
+ *  G1.4    Provoking vertex
+ *  G1.8    Determinant
+ *  G1.12   X1 - X0
+ *  G1.16   X2 - X0
+ *  G1.20   Y1 - Y0
+ *  G1.24   Y2 - Y0
+ *  G1.30   reserved
+ *
+ *  G2.0    Z0
+ *  G2.4    1/W0
+ *  G2.8    Z1
+ *  G2.12   1/W1
+ *  G2.16   Z2
+ *  G2.20   1/W2
+ *  G2.24   reserved
+ *  G2.30   reserved
+ *
+ *  G3 is V0 Vertex Attribute Data from URB (upper left)
+ *
+ *  G3.0    u0
+ *  G3.4    v0
+ *
+ *  G4 is V1 Vertex Attribute Data from URB (lower right)
+ *
+ *  G4.0    u1
+ *  G4.4    v1
+ *
+ *  G5 is V2 Vertex Attribute Data from URB (lower left)
+ *
+ */
+
+/* Compute inverses of the input deltas */
+send (4) 0 g6<1>F g1.12<4,4,1>F math inv mlen 1 rlen 1 { align1 };
+
+/* texture value at V0 */
+mov (4) m3<1>F g3<4,4,1>F { align1 };
+
+/* compute V1 - V2 (motion in X) for texture coordinates */
+add (4) g7<1>F g4<4,4,1>F -g5<4,4,1>F { align1 };
+
+/* multiply by 1/dx */
+mul (4) m1<1>F g7<4,4,1>F g6.0<0,1,0>F { align1 };
+
+/* Compute V2 - V0 (motion in Y) for texture coordinates */
+add (4) g7<1>F g5<4,4,1>F -g3<4,4,1>F { align1 };
+
+/* multiply by 1/dy */
+mul (4) m2<1>F g7<4,4,1>F g6.8<0,1,0>F {align1 };
+
+/* and we're done */
 send (8) 0 null g0<8,8,1>F urb 0 transpose used complete mlen 4 rlen 0 { align1 EOT };
 nop;
 nop;