nvc0: add support for clip distance shader outputs

6 files changed, 88 insertions, 16 deletions

diff --git a/src/gallium/drivers/nvc0/nvc0_3d.xml.h b/src/gallium/drivers/nvc0/nvc0_3d.xml.h
index 9264b72dadd..a8d91082740 100644
--- a/src/gallium/drivers/nvc0/nvc0_3d.xml.h
+++ b/src/gallium/drivers/nvc0/nvc0_3d.xml.h
@@ -697,15 +697,15 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #define NVC0_3D_CLIPID_FILL_RECT_VERT_HIGH__MASK		0xffff0000
 #define NVC0_3D_CLIPID_FILL_RECT_VERT_HIGH__SHIFT		16
 
-#define NVC0_3D_VP_CLIP_DISTANCE_ENABLE				0x00001510
-#define NVC0_3D_VP_CLIP_DISTANCE_ENABLE_0			0x00000001
-#define NVC0_3D_VP_CLIP_DISTANCE_ENABLE_1			0x00000002
-#define NVC0_3D_VP_CLIP_DISTANCE_ENABLE_2			0x00000004
-#define NVC0_3D_VP_CLIP_DISTANCE_ENABLE_3			0x00000008
-#define NVC0_3D_VP_CLIP_DISTANCE_ENABLE_4			0x00000010
-#define NVC0_3D_VP_CLIP_DISTANCE_ENABLE_5			0x00000020
-#define NVC0_3D_VP_CLIP_DISTANCE_ENABLE_6			0x00000040
-#define NVC0_3D_VP_CLIP_DISTANCE_ENABLE_7			0x00000080
+#define NVC0_3D_CLIP_DISTANCE_ENABLE				0x00001510
+#define NVC0_3D_CLIP_DISTANCE_ENABLE_0				0x00000001
+#define NVC0_3D_CLIP_DISTANCE_ENABLE_1				0x00000002
+#define NVC0_3D_CLIP_DISTANCE_ENABLE_2				0x00000004
+#define NVC0_3D_CLIP_DISTANCE_ENABLE_3				0x00000008
+#define NVC0_3D_CLIP_DISTANCE_ENABLE_4				0x00000010
+#define NVC0_3D_CLIP_DISTANCE_ENABLE_5				0x00000020
+#define NVC0_3D_CLIP_DISTANCE_ENABLE_6				0x00000040
+#define NVC0_3D_CLIP_DISTANCE_ENABLE_7				0x00000080
 
 #define NVC0_3D_SAMPLECNT_ENABLE				0x00001514
 
@@ -1022,6 +1022,40 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #define NVC0_3D_VIEW_VOLUME_CLIP_CTRL_UNK12_UNK1		0x00001000
 #define NVC0_3D_VIEW_VOLUME_CLIP_CTRL_UNK12_UNK2		0x00002000
 
+#define NVC0_3D_CLIP_DISTANCE_MODE				0x00001940
+#define NVC0_3D_CLIP_DISTANCE_MODE_0__MASK			0x00000001
+#define NVC0_3D_CLIP_DISTANCE_MODE_0__SHIFT			0
+#define NVC0_3D_CLIP_DISTANCE_MODE_0_CLIP			0x00000000
+#define NVC0_3D_CLIP_DISTANCE_MODE_0_CULL			0x00000001
+#define NVC0_3D_CLIP_DISTANCE_MODE_1__MASK			0x00000010
+#define NVC0_3D_CLIP_DISTANCE_MODE_1__SHIFT			4
+#define NVC0_3D_CLIP_DISTANCE_MODE_1_CLIP			0x00000000
+#define NVC0_3D_CLIP_DISTANCE_MODE_1_CULL			0x00000010
+#define NVC0_3D_CLIP_DISTANCE_MODE_2__MASK			0x00000100
+#define NVC0_3D_CLIP_DISTANCE_MODE_2__SHIFT			8
+#define NVC0_3D_CLIP_DISTANCE_MODE_2_CLIP			0x00000000
+#define NVC0_3D_CLIP_DISTANCE_MODE_2_CULL			0x00000100
+#define NVC0_3D_CLIP_DISTANCE_MODE_3__MASK			0x00001000
+#define NVC0_3D_CLIP_DISTANCE_MODE_3__SHIFT			12
+#define NVC0_3D_CLIP_DISTANCE_MODE_3_CLIP			0x00000000
+#define NVC0_3D_CLIP_DISTANCE_MODE_3_CULL			0x00001000
+#define NVC0_3D_CLIP_DISTANCE_MODE_4__MASK			0x00010000
+#define NVC0_3D_CLIP_DISTANCE_MODE_4__SHIFT			16
+#define NVC0_3D_CLIP_DISTANCE_MODE_4_CLIP			0x00000000
+#define NVC0_3D_CLIP_DISTANCE_MODE_4_CULL			0x00010000
+#define NVC0_3D_CLIP_DISTANCE_MODE_5__MASK			0x00100000
+#define NVC0_3D_CLIP_DISTANCE_MODE_5__SHIFT			20
+#define NVC0_3D_CLIP_DISTANCE_MODE_5_CLIP			0x00000000
+#define NVC0_3D_CLIP_DISTANCE_MODE_5_CULL			0x00100000
+#define NVC0_3D_CLIP_DISTANCE_MODE_6__MASK			0x01000000
+#define NVC0_3D_CLIP_DISTANCE_MODE_6__SHIFT			24
+#define NVC0_3D_CLIP_DISTANCE_MODE_6_CLIP			0x00000000
+#define NVC0_3D_CLIP_DISTANCE_MODE_6_CULL			0x01000000
+#define NVC0_3D_CLIP_DISTANCE_MODE_7__MASK			0x10000000
+#define NVC0_3D_CLIP_DISTANCE_MODE_7__SHIFT			28
+#define NVC0_3D_CLIP_DISTANCE_MODE_7_CLIP			0x00000000
+#define NVC0_3D_CLIP_DISTANCE_MODE_7_CULL			0x10000000
+
 #define NVC0_3D_CLIP_RECTS_EN					0x0000194c
 
 #define NVC0_3D_CLIP_RECTS_MODE					0x00001950
diff --git a/src/gallium/drivers/nvc0/nvc0_context.h b/src/gallium/drivers/nvc0/nvc0_context.h
index c739b1b68e2..c46e7430f85 100644
--- a/src/gallium/drivers/nvc0/nvc0_context.h
+++ b/src/gallium/drivers/nvc0/nvc0_context.h
@@ -74,13 +74,15 @@ struct nvc0_context {
       int32_t index_bias;
       boolean prim_restart;
       boolean early_z;
+      uint16_t scissor;
       uint8_t num_vtxbufs;
       uint8_t num_vtxelts;
       uint8_t num_textures[5];
       uint8_t num_samplers[5];
       uint8_t tls_required; /* bitmask of shader types using l[] */
       uint8_t c14_bound; /* whether immediate array constbuf is bound */
-      uint16_t scissor;
+      uint8_t clip_enable;
+      uint32_t clip_mode;
       uint32_t uniform_buffer_bound[5];
    } state;
 
diff --git a/src/gallium/drivers/nvc0/nvc0_program.c b/src/gallium/drivers/nvc0/nvc0_program.c
index c422c9b1087..f3185b488e8 100644
--- a/src/gallium/drivers/nvc0/nvc0_program.c
+++ b/src/gallium/drivers/nvc0/nvc0_program.c
@@ -268,6 +268,11 @@ nvc0_vtgp_gen_header(struct nvc0_program *vp, struct nv50_ir_prog_info *info)
       }
    }
 
+   vp->vp.clip_enable = (1 << info->io.clipDistanceCount) - 1;
+   for (i = 0; i < 8; ++i)
+      if (info->io.cullDistanceMask & (1 << i))
+         vp->vp.clip_mode |= 1 << (i * 4);
+
    return 0;
 }
 
diff --git a/src/gallium/drivers/nvc0/nvc0_program.h b/src/gallium/drivers/nvc0/nvc0_program.h
index 239890bd89a..b107850ea92 100644
--- a/src/gallium/drivers/nvc0/nvc0_program.h
+++ b/src/gallium/drivers/nvc0/nvc0_program.h
@@ -27,6 +27,8 @@ struct nvc0_program {
    uint32_t flags[2];
 
    struct {
+      uint32_t clip_mode; /* clip/cull selection */
+      uint8_t clip_enable; /* only applies if num_ucps == 0 */
       uint8_t edgeflag;
       uint8_t num_ucps;
       uint8_t out_pos[PIPE_MAX_SHADER_OUTPUTS];
diff --git a/src/gallium/drivers/nvc0/nvc0_shader_state.c b/src/gallium/drivers/nvc0/nvc0_shader_state.c
index 472bdb11aba..f4a12fbee11 100644
--- a/src/gallium/drivers/nvc0/nvc0_shader_state.c
+++ b/src/gallium/drivers/nvc0/nvc0_shader_state.c
@@ -59,6 +59,25 @@ nvc0_program_update_context_state(struct nvc0_context *nvc0,
    }
 }
 
+static void
+nvc0_program_validate_clip(struct nvc0_context *nvc0, struct nvc0_program *vp)
+{
+   struct nouveau_channel *chan = nvc0->screen->base.channel;
+
+   if (nvc0->vertprog->vp.num_ucps)
+      return;
+
+   if (nvc0->state.clip_enable != vp->vp.clip_enable) {
+      nvc0->state.clip_enable = vp->vp.clip_enable;
+      IMMED_RING(chan, RING_3D(CLIP_DISTANCE_ENABLE), vp->vp.clip_enable);
+   }
+   if (nvc0->state.clip_mode != vp->vp.clip_mode) {
+      nvc0->state.clip_mode = vp->vp.clip_mode;
+      BEGIN_RING(chan, RING_3D(CLIP_DISTANCE_MODE), 1);
+      OUT_RING  (chan, vp->vp.clip_mode);
+   }
+}
+
 static INLINE boolean
 nvc0_program_validate(struct nvc0_context *nvc0, struct nvc0_program *prog)
 {
@@ -98,6 +117,9 @@ nvc0_vertprog_validate(struct nvc0_context *nvc0)
    BEGIN_RING(chan, RING_3D(SP_GPR_ALLOC(1)), 1);
    OUT_RING  (chan, vp->max_gpr);
 
+   if (!nvc0->gmtyprog && !nvc0->tevlprog)
+      nvc0_program_validate_clip(nvc0, vp);
+
    // BEGIN_RING(chan, RING_3D_(0x163c), 1);
    // OUT_RING  (chan, 0);
 }
@@ -178,7 +200,10 @@ nvc0_tevlprog_validate(struct nvc0_context *nvc0)
    BEGIN_RING(chan, RING_3D(SP_START_ID(3)), 1);
    OUT_RING  (chan, tp->code_base);
    BEGIN_RING(chan, RING_3D(SP_GPR_ALLOC(3)), 1);
-   OUT_RING  (chan, tp->max_gpr);   
+   OUT_RING  (chan, tp->max_gpr);
+
+   if (!nvc0->gmtyprog)
+      nvc0_program_validate_clip(nvc0, tp);
 }
 
 void
@@ -205,6 +230,8 @@ nvc0_gmtyprog_validate(struct nvc0_context *nvc0)
    OUT_RING  (chan, gp->max_gpr);
    BEGIN_RING(chan, RING_3D(LAYER), 1);
    OUT_RING  (chan, (gp->hdr[13] & (1 << 9)) ? NVC0_3D_LAYER_USE_GP : 0);
+
+   nvc0_program_validate_clip(nvc0, gp);
 }
 
 /* It's *is* kind of shader related. We need to inspect the program
diff --git a/src/gallium/drivers/nvc0/nvc0_state_validate.c b/src/gallium/drivers/nvc0/nvc0_state_validate.c
index 2fca2919ce6..5cf68c7b6f2 100644
--- a/src/gallium/drivers/nvc0/nvc0_state_validate.c
+++ b/src/gallium/drivers/nvc0/nvc0_state_validate.c
@@ -267,11 +267,13 @@ nvc0_validate_clip(struct nvc0_context *nvc0)
       BEGIN_RING_1I(chan, RING_3D(CB_POS), nvc0->clip.nr * 4 + 1);
       OUT_RING  (chan, 0);
       OUT_RINGp (chan, &nvc0->clip.ucp[0][0], nvc0->clip.nr * 4);
+   }
 
-      BEGIN_RING(chan, RING_3D(VP_CLIP_DISTANCE_ENABLE), 1);
-      OUT_RING  (chan, (1 << nvc0->clip.nr) - 1);
-   } else {
-      IMMED_RING(chan, RING_3D(VP_CLIP_DISTANCE_ENABLE), 0);
+   if (nvc0->vertprog->vp.num_ucps) {
+      nvc0->state.clip_mode = 0;
+      nvc0->state.clip_enable = (1 << nvc0->clip.nr) - 1;
+      IMMED_RING(chan, RING_3D(CLIP_DISTANCE_ENABLE), nvc0->state.clip_enable);
+      IMMED_RING(chan, RING_3D(CLIP_DISTANCE_MODE), 0);
    }
 }
 
@@ -474,13 +476,13 @@ static struct state_validate {
     { nvc0_validate_stipple,       NVC0_NEW_STIPPLE },
     { nvc0_validate_scissor,       NVC0_NEW_SCISSOR | NVC0_NEW_RASTERIZER },
     { nvc0_validate_viewport,      NVC0_NEW_VIEWPORT },
-    { nvc0_validate_clip,          NVC0_NEW_CLIP },
     { nvc0_vertprog_validate,      NVC0_NEW_VERTPROG },
     { nvc0_tctlprog_validate,      NVC0_NEW_TCTLPROG },
     { nvc0_tevlprog_validate,      NVC0_NEW_TEVLPROG },
     { nvc0_gmtyprog_validate,      NVC0_NEW_GMTYPROG },
     { nvc0_fragprog_validate,      NVC0_NEW_FRAGPROG },
     { nvc0_validate_derived_1,     NVC0_NEW_FRAGPROG | NVC0_NEW_ZSA },
+    { nvc0_validate_clip,          NVC0_NEW_CLIP },
     { nvc0_constbufs_validate,     NVC0_NEW_CONSTBUF },
     { nvc0_validate_textures,      NVC0_NEW_TEXTURES },
     { nvc0_validate_samplers,      NVC0_NEW_SAMPLERS },