nir: Don't use ffma in nir_lower_wpos_ytransform().
authorKenneth Graunke <kenneth@whitecape.org>
Wed, 18 May 2016 19:14:02 +0000 (12:14 -0700)
committerKenneth Graunke <kenneth@whitecape.org>
Fri, 20 May 2016 21:29:04 +0000 (14:29 -0700)
ffma is an explicitly fused multiply add with higher precision.
The optimizer will take care of promoting mul/add to fma when
it's beneficial to do so.

This fixes failures on Gen4-5 when using this pass, as those platforms
don't actually implement fma().

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
src/compiler/nir/nir_lower_wpos_ytransform.c

index 36e25b933ae24659e8991f2ea9fda9b48716e987..ccf0fd351a72958d23d0828548ffe7992e029685 100644 (file)
@@ -123,19 +123,15 @@ emit_wpos_adjustment(lower_wpos_ytransform_state *state,
     * inversion/identity, or the other way around if we're drawing to an FBO.
     */
    if (invert) {
-      /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
-       */
-      wpos_temp_y = nir_ffma(b,
-                             nir_channel(b, wpos_temp, 1),
-                             nir_channel(b, wpostrans, 0),
-                             nir_channel(b, wpostrans, 1));
+      /* wpos_temp.y = wpos_input * wpostrans.xxxx + wpostrans.yyyy */
+      wpos_temp_y = nir_fadd(b, nir_fmul(b, nir_channel(b, wpos_temp, 1),
+                                            nir_channel(b, wpostrans, 0)),
+                                nir_channel(b, wpostrans, 1));
    } else {
-      /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
-       */
-      wpos_temp_y = nir_ffma(b,
-                             nir_channel(b, wpos_temp, 1),
-                             nir_channel(b, wpostrans, 2),
-                             nir_channel(b, wpostrans, 3));
+      /* wpos_temp.y = wpos_input * wpostrans.zzzz + wpostrans.wwww */
+      wpos_temp_y = nir_fadd(b, nir_fmul(b, nir_channel(b, wpos_temp, 1),
+                                            nir_channel(b, wpostrans, 2)),
+                                nir_channel(b, wpostrans, 3));
    }
 
    wpos_temp = nir_vec4(b,