diff --git a/src/mesa/drivers/dri/i965/gen8_ps_state.c b/src/mesa/drivers/dri/i965/gen8_ps_state.c
index c25b82757c1..c2810bb21b9 100644
--- a/src/mesa/drivers/dri/i965/gen8_ps_state.c
+++ b/src/mesa/drivers/dri/i965/gen8_ps_state.c
@@ -180,10 +180,6 @@ upload_ps_state(struct brw_context *brw)
    if (brw->wm.prog_data->base.nr_params > 0)
       dw6 |= GEN7_PS_PUSH_CONSTANT_ENABLE;
 
-   dw6 |= GEN7_PS_8_DISPATCH_ENABLE;
-   if (brw->wm.prog_data->prog_offset_16)
-      dw6 |= GEN7_PS_16_DISPATCH_ENABLE;
-
    /* From the documentation for this packet:
     * "If the PS kernel does not need the Position XY Offsets to
     *  compute a Position Value, then this field should be programmed
@@ -202,13 +198,40 @@ upload_ps_state(struct brw_context *brw)
    else
       dw6 |= GEN7_PS_POSOFFSET_NONE;
 
-   dw7 |=
-      brw->wm.prog_data->first_curbe_grf << GEN7_PS_DISPATCH_START_GRF_SHIFT_0 |
-      brw->wm.prog_data->first_curbe_grf_16<< GEN7_PS_DISPATCH_START_GRF_SHIFT_2;
+   /* _NEW_MULTISAMPLE
+    * In case of non 1x per sample shading, only one of SIMD8 and SIMD16
+    * should be enabled. We do 'SIMD16 only' dispatch if a SIMD16 shader
+    * is successfully compiled. In majority of the cases that bring us
+    * better performance than 'SIMD8 only' dispatch.
+    */
+   int min_invocations_per_fragment =
+      _mesa_get_min_invocations_per_fragment(ctx, brw->fragment_program, false);
+   assert(min_invocations_per_fragment >= 1);
+
+   if (brw->wm.prog_data->prog_offset_16) {
+      dw6 |= GEN7_PS_16_DISPATCH_ENABLE;
+      if (min_invocations_per_fragment == 1) {
+         dw6 |= GEN7_PS_8_DISPATCH_ENABLE;
+         dw7 |= (brw->wm.prog_data->first_curbe_grf <<
+                 GEN7_PS_DISPATCH_START_GRF_SHIFT_0);
+         dw7 |= (brw->wm.prog_data->first_curbe_grf_16 <<
+                 GEN7_PS_DISPATCH_START_GRF_SHIFT_2);
+      } else {
+         dw7 |= (brw->wm.prog_data->first_curbe_grf_16 <<
+                 GEN7_PS_DISPATCH_START_GRF_SHIFT_0);
+      }
+   } else {
+      dw6 |= GEN7_PS_8_DISPATCH_ENABLE;
+      dw7 |= (brw->wm.prog_data->first_curbe_grf <<
+              GEN7_PS_DISPATCH_START_GRF_SHIFT_0);
+   }
 
    BEGIN_BATCH(12);
    OUT_BATCH(_3DSTATE_PS << 16 | (12 - 2));
-   OUT_BATCH(brw->wm.base.prog_offset);
+   if (brw->wm.prog_data->prog_offset_16 && min_invocations_per_fragment > 1)
+      OUT_BATCH(brw->wm.base.prog_offset + brw->wm.prog_data->prog_offset_16);
+   else
+      OUT_BATCH(brw->wm.base.prog_offset);
    OUT_BATCH(0);
    OUT_BATCH(dw3);
    if (brw->wm.prog_data->total_scratch) {
@@ -230,7 +253,7 @@ upload_ps_state(struct brw_context *brw)
 
 const struct brw_tracked_state gen8_ps_state = {
    .dirty = {
-      .mesa  = _NEW_PROGRAM_CONSTANTS,
+      .mesa  = _NEW_PROGRAM_CONSTANTS | _NEW_MULTISAMPLE,
       .brw   = BRW_NEW_FRAGMENT_PROGRAM |
                BRW_NEW_PS_BINDING_TABLE |
                BRW_NEW_BATCH |