freedreno: a2xx: fd2_draw update

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Signed-off-by: Rob Clark <robdclark@gmail.com>

6 files changed, 114 insertions, 20 deletions

diff --git a/src/gallium/drivers/freedreno/a2xx/fd2_draw.c b/src/gallium/drivers/freedreno/a2xx/fd2_draw.c
index f00bec6efcd..6dac8ca6a9d 100644
--- a/src/gallium/drivers/freedreno/a2xx/fd2_draw.c
+++ b/src/gallium/drivers/freedreno/a2xx/fd2_draw.c
@@ -77,29 +77,44 @@ emit_vertexbufs(struct fd_context *ctx)
 	fd2_emit_vertex_bufs(ctx->batch->draw, 0x78, bufs, vtx->num_elements);
 }
 
-static bool
-fd2_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info,
-             unsigned index_offset)
+static void
+draw_impl(struct fd_context *ctx, const struct pipe_draw_info *info,
+		   struct fd_ringbuffer *ring, unsigned index_offset)
 {
-	struct fd_ringbuffer *ring = ctx->batch->draw;
-
-	if (ctx->dirty & FD_DIRTY_VTXBUF)
-		emit_vertexbufs(ctx);
-
-	fd2_emit_state(ctx, ctx->dirty);
-
 	OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 	OUT_RING(ring, CP_REG(REG_A2XX_VGT_INDX_OFFSET));
-	OUT_RING(ring, info->start);
+	OUT_RING(ring, info->index_size ? 0 : info->start);
 
 	OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 	OUT_RING(ring, CP_REG(REG_A2XX_VGT_VERTEX_REUSE_BLOCK_CNTL));
-	OUT_RING(ring, 0x0000003b);
+	OUT_RING(ring, is_a20x(ctx->screen) ? 0x00000002 : 0x0000003b);
 
 	OUT_PKT0(ring, REG_A2XX_TC_CNTL_STATUS, 1);
 	OUT_RING(ring, A2XX_TC_CNTL_STATUS_L2_INVALIDATE);
 
-	if (!is_a20x(ctx->screen)) {
+	if (is_a20x(ctx->screen)) {
+		/* wait for DMA to finish and
+		 * dummy draw one triangle with indexes 0,0,0.
+		 * with PRE_FETCH_CULL_ENABLE | GRP_CULL_ENABLE.
+		 *
+		 * this workaround is for a HW bug related to DMA alignment:
+		 * it is necessary for indexed draws and possibly also
+		 * draws that read binning data
+		 */
+		OUT_PKT3(ring, CP_WAIT_REG_EQ, 4);
+		OUT_RING(ring, 0x000005d0); /* RBBM_STATUS */
+		OUT_RING(ring, 0x00000000);
+		OUT_RING(ring, 0x00001000); /* bit: 12: VGT_BUSY_NO_DMA */
+		OUT_RING(ring, 0x00000001);
+
+		OUT_PKT3(ring, CP_DRAW_INDX_BIN, 6);
+		OUT_RING(ring, 0x00000000);
+		OUT_RING(ring, 0x0003c004);
+		OUT_RING(ring, 0x00000000);
+		OUT_RING(ring, 0x00000003);
+		OUT_RELOC(ring, fd_resource(fd2_context(ctx)->solid_vertexbuf)->bo, 0x80, 0, 0);
+		OUT_RING(ring, 0x00000006);
+	} else {
 		OUT_WFI (ring);
 
 		OUT_PKT3(ring, CP_SET_CONSTANT, 3);
@@ -111,11 +126,62 @@ fd2_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info,
 	fd_draw_emit(ctx->batch, ring, ctx->primtypes[info->mode],
 				 IGNORE_VISIBILITY, info, index_offset);
 
-	OUT_PKT3(ring, CP_SET_CONSTANT, 2);
-	OUT_RING(ring, CP_REG(REG_A2XX_UNKNOWN_2010));
-	OUT_RING(ring, 0x00000000);
+	if (is_a20x(ctx->screen)) {
+		/* not sure why this is required, but it fixes some hangs */
+		OUT_WFI(ring);
+	} else {
+		OUT_PKT3(ring, CP_SET_CONSTANT, 2);
+		OUT_RING(ring, CP_REG(REG_A2XX_UNKNOWN_2010));
+		OUT_RING(ring, 0x00000000);
+	}
 
 	emit_cacheflush(ring);
+}
+
+
+static bool
+fd2_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *pinfo,
+			 unsigned index_offset)
+{
+	if (!ctx->prog.fp || !ctx->prog.vp)
+		return false;
+
+	if (ctx->dirty & FD_DIRTY_VTXBUF)
+		emit_vertexbufs(ctx);
+
+	fd2_emit_state(ctx, ctx->dirty);
+
+	/* a2xx can draw only 65535 vertices at once
+	 * on a22x the field in the draw command is 32bits but seems limited too
+	 * using a limit of 32k because it fixes an unexplained hang
+	 * 32766 works for all primitives (multiple of 2 and 3)
+	 */
+	if (pinfo->count > 32766) {
+		static const uint16_t step_tbl[PIPE_PRIM_MAX] = {
+			[0 ... PIPE_PRIM_MAX - 1]  = 32766,
+			[PIPE_PRIM_LINE_STRIP]     = 32765,
+			[PIPE_PRIM_TRIANGLE_STRIP] = 32764,
+
+			/* needs more work */
+			[PIPE_PRIM_TRIANGLE_FAN]   = 0,
+			[PIPE_PRIM_LINE_LOOP]      = 0,
+		};
+
+		struct pipe_draw_info info = *pinfo;
+		unsigned count = info.count;
+		unsigned step = step_tbl[info.mode];
+
+		if (!step)
+			return false;
+
+		for (; count + step > 32766; count -= step) {
+			info.count = MIN2(count, 32766);
+			draw_impl(ctx, &info, ctx->batch->draw, index_offset);
+			info.start += step;
+		}
+	} else {
+		draw_impl(ctx, pinfo, ctx->batch->draw, index_offset);
+	}
 
 	fd_context_all_clean(ctx);
 
diff --git a/src/gallium/drivers/freedreno/freedreno_batch.c b/src/gallium/drivers/freedreno/freedreno_batch.c
index 28b4942f9f8..eae2f68ce11 100644
--- a/src/gallium/drivers/freedreno/freedreno_batch.c
+++ b/src/gallium/drivers/freedreno/freedreno_batch.c
@@ -83,6 +83,7 @@ batch_init(struct fd_batch *batch)
 	batch->flushed = false;
 	batch->gmem_reason = 0;
 	batch->num_draws = 0;
+	batch->num_vertices = 0;
 	batch->stage = FD_STAGE_NULL;
 
 	fd_reset_wfi(batch);
diff --git a/src/gallium/drivers/freedreno/freedreno_batch.h b/src/gallium/drivers/freedreno/freedreno_batch.h
index d4feadd5590..a40d36094cd 100644
--- a/src/gallium/drivers/freedreno/freedreno_batch.h
+++ b/src/gallium/drivers/freedreno/freedreno_batch.h
@@ -124,6 +124,7 @@ struct fd_batch {
 		FD_GMEM_LOGICOP_ENABLED      = 0x20,
 	} gmem_reason;
 	unsigned num_draws;   /* number of draws in current batch */
+	unsigned num_vertices;   /* number of vertices in current batch */
 
 	/* Track the maximal bounds of the scissor of all the draws within a
 	 * batch.  Used at the tile rendering step (fd_gmem_render_tiles(),
diff --git a/src/gallium/drivers/freedreno/freedreno_draw.c b/src/gallium/drivers/freedreno/freedreno_draw.c
index 90d0e440361..f17cb563063 100644
--- a/src/gallium/drivers/freedreno/freedreno_draw.c
+++ b/src/gallium/drivers/freedreno/freedreno_draw.c
@@ -291,6 +291,8 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 	if (ctx->draw_vbo(ctx, info, index_offset))
 		batch->needs_flush = true;
 
+	batch->num_vertices += info->count * info->instance_count;
+
 	for (i = 0; i < ctx->streamout.num_targets; i++)
 		ctx->streamout.offsets[i] += info->count;
 
diff --git a/src/gallium/drivers/freedreno/freedreno_draw.h b/src/gallium/drivers/freedreno/freedreno_draw.h
index 8ae13411cef..c2197f21a9f 100644
--- a/src/gallium/drivers/freedreno/freedreno_draw.h
+++ b/src/gallium/drivers/freedreno/freedreno_draw.h
@@ -73,9 +73,29 @@ fd_draw(struct fd_batch *batch, struct fd_ringbuffer *ring,
 	}
 
 	if (is_a20x(batch->ctx->screen)) {
-		OUT_PKT3(ring, CP_DRAW_INDX, idx_buffer ? 4 : 2);
+		/* a20x has a different draw command for drawing with binning data
+		 * note: if we do patching we will have to insert a NOP
+		 *
+		 * binning data is is 1 byte/vertex (8x8x4 bin position of vertex)
+		 * base ptr set by the CP_SET_DRAW_INIT_FLAGS command
+		 *
+		 * TODO: investigate the faceness_cull_select parameter to see how
+		 * it is used with hw binning to use "faceness" bits
+		 */
+		uint32_t size = 2;
+		if (vismode)
+			size += 2;
+		if (idx_buffer)
+			size += 2;
+
+		OUT_PKT3(ring, vismode ? CP_DRAW_INDX_BIN : CP_DRAW_INDX, size);
 		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, DRAW_A20X(primtype, src_sel, idx_type, vismode, count));
+		OUT_RING(ring, DRAW_A20X(primtype, DI_FACE_CULL_NONE, src_sel,
+								 idx_type, vismode, vismode, count));
+		if (vismode == USE_VISIBILITY) {
+			OUT_RING(ring, batch->num_vertices);
+			OUT_RING(ring, count);
+		}
 	} else {
 		OUT_PKT3(ring, CP_DRAW_INDX, idx_buffer ? 5 : 3);
 		OUT_RING(ring, 0x00000000);        /* viz query info. */
diff --git a/src/gallium/drivers/freedreno/freedreno_util.h b/src/gallium/drivers/freedreno/freedreno_util.h
index 125ad83523c..b0ed3fa8e4f 100644
--- a/src/gallium/drivers/freedreno/freedreno_util.h
+++ b/src/gallium/drivers/freedreno/freedreno_util.h
@@ -114,15 +114,19 @@ static inline uint32_t DRAW(enum pc_di_primtype prim_type,
 }
 
 static inline uint32_t DRAW_A20X(enum pc_di_primtype prim_type,
+		enum pc_di_face_cull_sel faceness_cull_select,
 		enum pc_di_src_sel source_select, enum pc_di_index_size index_size,
-		enum pc_di_vis_cull_mode vis_cull_mode,
+		bool pre_fetch_cull_enable,
+		bool grp_cull_enable,
 		uint16_t count)
 {
 	return (prim_type         << 0) |
 			(source_select     << 6) |
+			(faceness_cull_select << 8) |
 			((index_size & 1)  << 11) |
 			((index_size >> 1) << 13) |
-			(vis_cull_mode     << 9) |
+			(pre_fetch_cull_enable << 14) |
+			(grp_cull_enable << 15) |
 			(count         << 16);
 }