a3xx: make use of software clipping when hw can't handle it

The hw clipper only handles up to 6 UCPs. If there are more than 6 UCPs,
or a clip vertex, or clip distances are in use, then we must use the
fallback discard-based clipping from the frag shader.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: mesa-stable@lists.freedesktop.org

src/gallium/drivers/freedreno/a3xx/fd3_draw.c

@@ -156,6 +156,9 @@ fd3_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info)
 		.sprite_coord_mode = ctx->rasterizer->sprite_coord_mode,
 	};
 
+	if (fd3_needs_manual_clipping(ctx->prog.vp, ctx->rasterizer))
+		emit.key.ucp_enables = ctx->rasterizer->clip_plane_enable;
+
 	fixup_shader_state(ctx, &emit.key);
 
 	unsigned dirty = ctx->dirty;

src/gallium/drivers/freedreno/a3xx/fd3_emit.c

@@ -571,20 +571,24 @@ fd3_emit_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
 	if (dirty & (FD_DIRTY_RASTERIZER | FD_DIRTY_PROG)) {
 		uint32_t val = fd3_rasterizer_stateobj(ctx->rasterizer)
 				->gras_cl_clip_cntl;
+		uint8_t planes = ctx->rasterizer->clip_plane_enable;
 		val |= COND(fp->writes_pos, A3XX_GRAS_CL_CLIP_CNTL_ZCLIP_DISABLE);
 		val |= COND(fp->frag_coord, A3XX_GRAS_CL_CLIP_CNTL_ZCOORD |
 				A3XX_GRAS_CL_CLIP_CNTL_WCOORD);
-		/* TODO only use if prog doesn't use clipvertex/clipdist */
-		val |= A3XX_GRAS_CL_CLIP_CNTL_NUM_USER_CLIP_PLANES(
-				MIN2(util_bitcount(ctx->rasterizer->clip_plane_enable), 6));
+		if (!emit->key.ucp_enables)
+			val |= A3XX_GRAS_CL_CLIP_CNTL_NUM_USER_CLIP_PLANES(
+					MIN2(util_bitcount(planes), 6));
 		OUT_PKT0(ring, REG_A3XX_GRAS_CL_CLIP_CNTL, 1);
 		OUT_RING(ring, val);
 	}
 
-	if (dirty & (FD_DIRTY_RASTERIZER | FD_DIRTY_UCP)) {
+	if (dirty & (FD_DIRTY_RASTERIZER | FD_DIRTY_PROG | FD_DIRTY_UCP)) {
 		uint32_t planes = ctx->rasterizer->clip_plane_enable;
 		int count = 0;
 
+		if (emit->key.ucp_enables)
+			planes = 0;
+
 		while (planes && count < 6) {
 			int i = ffs(planes) - 1;
 

src/gallium/drivers/freedreno/a3xx/fd3_program.c

@@ -28,6 +28,7 @@
 
 #include "pipe/p_state.h"
 #include "util/u_string.h"
+#include "util/u_math.h"
 #include "util/u_memory.h"
 #include "util/u_inlines.h"
 #include "util/u_format.h"
@@ -85,6 +86,20 @@ fd3_vp_state_delete(struct pipe_context *pctx, void *hwcso)
 	delete_shader_stateobj(so);
 }
 
+bool
+fd3_needs_manual_clipping(const struct fd3_shader_stateobj *so,
+						  const struct pipe_rasterizer_state *rast)
+{
+	uint64_t outputs = ir3_shader_outputs(so->shader);
+
+	return (!rast->depth_clip ||
+			util_bitcount(rast->clip_plane_enable) > 6 ||
+			outputs & ((1ULL << VARYING_SLOT_CLIP_VERTEX) |
+					   (1ULL << VARYING_SLOT_CLIP_DIST0) |
+					   (1ULL << VARYING_SLOT_CLIP_DIST1)));
+}
+
+
 static void
 emit_shader(struct fd_ringbuffer *ring, const struct ir3_shader_variant *so)
 {

src/gallium/drivers/freedreno/a3xx/fd3_program.h

@@ -44,4 +44,7 @@ void fd3_program_emit(struct fd_ringbuffer *ring, struct fd3_emit *emit,
 
 void fd3_prog_init(struct pipe_context *pctx);
 
+bool fd3_needs_manual_clipping(const struct fd3_shader_stateobj *,
+							   const struct pipe_rasterizer_state *);
+
 #endif /* FD3_PROGRAM_H_ */

src/gallium/drivers/freedreno/ir3/ir3_shader.c

@@ -469,6 +469,12 @@ ir3_shader_disasm(struct ir3_shader_variant *so, uint32_t *bin)
 	debug_printf("\n");
 }
 
+uint64_t
+ir3_shader_outputs(const struct ir3_shader *so)
+{
+	return so->nir->info.outputs_written;
+}
+
 /* This has to reach into the fd_context a bit more than the rest of
  * ir3, but it needs to be aligned with the compiler, so both agree
  * on which const regs hold what.  And the logic is identical between

src/gallium/drivers/freedreno/ir3/ir3_shader.h

@@ -272,6 +272,7 @@ void ir3_shader_destroy(struct ir3_shader *shader);
 struct ir3_shader_variant * ir3_shader_variant(struct ir3_shader *shader,
 		struct ir3_shader_key key, struct pipe_debug_callback *debug);
 void ir3_shader_disasm(struct ir3_shader_variant *so, uint32_t *bin);
+uint64_t ir3_shader_outputs(const struct ir3_shader *so);
 
 struct fd_ringbuffer;
 struct fd_context;