freedreno/a4xx: MRT support

Signed-off-by: Rob Clark <[email protected]>

12 files changed, 212 insertions, 132 deletions

diff --git a/src/gallium/drivers/freedreno/a3xx/fd3_program.c b/src/gallium/drivers/freedreno/a3xx/fd3_program.c
index e98c6b5cff4..b5360797745 100644
--- a/src/gallium/drivers/freedreno/a3xx/fd3_program.c
+++ b/src/gallium/drivers/freedreno/a3xx/fd3_program.c
@@ -204,7 +204,7 @@ fd3_program_emit(struct fd_ringbuffer *ring, struct fd3_emit *emit,
 		color_regid[0] = color_regid[1] = color_regid[2] = color_regid[3] =
 			ir3_find_output_regid(fp, ir3_semantic_name(TGSI_SEMANTIC_COLOR, 0));
 	} else {
-		for (int i = 0; i < fp->outputs_count; i++) {
+		for (i = 0; i < fp->outputs_count; i++) {
 			ir3_semantic sem = fp->outputs[i].semantic;
 			unsigned idx = sem2idx(sem);
 			if (sem2name(sem) != TGSI_SEMANTIC_COLOR)
diff --git a/src/gallium/drivers/freedreno/a3xx/fd3_screen.c b/src/gallium/drivers/freedreno/a3xx/fd3_screen.c
index 094dcf376e5..722fe360202 100644
--- a/src/gallium/drivers/freedreno/a3xx/fd3_screen.c
+++ b/src/gallium/drivers/freedreno/a3xx/fd3_screen.c
@@ -105,7 +105,7 @@ void
 fd3_screen_init(struct pipe_screen *pscreen)
 {
 	struct fd_screen *screen = fd_screen(pscreen);
-	screen->max_rts = 4;
+	screen->max_rts = A3XX_MAX_RENDER_TARGETS;
 	screen->compiler = ir3_compiler_create(screen->gpu_id);
 	pscreen->context_create = fd3_context_create;
 	pscreen->is_format_supported = fd3_screen_is_format_supported;
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_blend.c b/src/gallium/drivers/freedreno/a4xx/fd4_blend.c
index 396caa532fc..9d5ae4242f9 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_blend.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_blend.c
@@ -61,7 +61,7 @@ fd4_blend_state_create(struct pipe_context *pctx,
 	struct fd4_blend_stateobj *so;
 //	enum a3xx_rop_code rop = ROP_COPY;
 	bool reads_dest = false;
-	int i;
+	unsigned i, mrt_blend = 0;
 
 	if (cso->logicop_enable) {
 //		rop = cso->logicop_func;  /* maps 1:1 */
@@ -115,7 +115,7 @@ fd4_blend_state_create(struct pipe_context *pctx,
 					A4XX_RB_MRT_CONTROL_READ_DEST_ENABLE |
 					A4XX_RB_MRT_CONTROL_BLEND |
 					A4XX_RB_MRT_CONTROL_BLEND2;
-			so->rb_fs_output |= A4XX_RB_FS_OUTPUT_ENABLE_BLEND(1);
+			mrt_blend |= (1 << i);
 		}
 
 		if (reads_dest)
@@ -125,5 +125,7 @@ fd4_blend_state_create(struct pipe_context *pctx,
 			so->rb_mrt[i].buf_info |= A4XX_RB_MRT_BUF_INFO_DITHER_MODE(DITHER_ALWAYS);
 	}
 
+	so->rb_fs_output = A4XX_RB_FS_OUTPUT_ENABLE_BLEND(mrt_blend);
+
 	return so;
 }
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_draw.c b/src/gallium/drivers/freedreno/a4xx/fd4_draw.c
index 154154190db..0927b0d7682 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_draw.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_draw.c
@@ -111,7 +111,6 @@ static void
 fd4_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info)
 {
 	struct fd4_context *fd4_ctx = fd4_context(ctx);
-	struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 	struct fd4_emit emit = {
 		.vtx  = &ctx->vtx,
 		.prog = &ctx->prog,
@@ -132,8 +131,6 @@ fd4_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info)
 			.fsaturate_t = fd4_ctx->fsaturate_t,
 			.fsaturate_r = fd4_ctx->fsaturate_r,
 		},
-		.format = fd4_emit_format(pfb->cbufs[0]),
-		.pformat = pipe_surface_format(pfb->cbufs[0]),
 	};
 	unsigned dirty;
 
@@ -173,20 +170,16 @@ fd4_clear(struct fd_context *ctx, unsigned buffers,
 	struct fd4_context *fd4_ctx = fd4_context(ctx);
 	struct fd_ringbuffer *ring = ctx->ring;
 	struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
+	unsigned char mrt_comp[A4XX_MAX_RENDER_TARGETS] = {0};
 	unsigned dirty = ctx->dirty;
-	unsigned ce, i;
+	unsigned i;
 	struct fd4_emit emit = {
 		.vtx  = &fd4_ctx->solid_vbuf_state,
 		.prog = &ctx->solid_prog,
 		.key = {
-			.half_precision = true,
+			.half_precision = fd_half_precision(pfb),
 		},
-		.format = fd4_emit_format(pfb->cbufs[0]),
 	};
-	uint32_t colr = 0;
-
-	if ((buffers & PIPE_CLEAR_COLOR) && pfb->nr_cbufs)
-		colr  = pack_rgba(pfb->cbufs[0]->format, color->f);
 
 	dirty &= FD_DIRTY_FRAMEBUFFER | FD_DIRTY_SCISSOR;
 	dirty |= FD_DIRTY_PROG;
@@ -260,16 +253,15 @@ fd4_clear(struct fd_context *ctx, unsigned buffers,
 	if (buffers & PIPE_CLEAR_COLOR) {
 		OUT_PKT0(ring, REG_A4XX_RB_ALPHA_CONTROL, 1);
 		OUT_RING(ring, A4XX_RB_ALPHA_CONTROL_ALPHA_TEST_FUNC(FUNC_NEVER));
-		ce = 0xf;
-	} else {
-		ce = 0x0;
 	}
 
 	for (i = 0; i < A4XX_MAX_RENDER_TARGETS; i++) {
+		mrt_comp[i] = (buffers & (PIPE_CLEAR_COLOR0 << i)) ? 0xf : 0x0;
+
 		OUT_PKT0(ring, REG_A4XX_RB_MRT_CONTROL(i), 1);
 		OUT_RING(ring, A4XX_RB_MRT_CONTROL_FASTCLEAR |
 				A4XX_RB_MRT_CONTROL_B11 |
-				A4XX_RB_MRT_CONTROL_COMPONENT_ENABLE(ce));
+				A4XX_RB_MRT_CONTROL_COMPONENT_ENABLE(0xf));
 
 		OUT_PKT0(ring, REG_A4XX_RB_MRT_BLEND_CONTROL(i), 1);
 		OUT_RING(ring, A4XX_RB_MRT_BLEND_CONTROL_RGB_SRC_FACTOR(FACTOR_ONE) |
@@ -280,6 +272,16 @@ fd4_clear(struct fd_context *ctx, unsigned buffers,
 				A4XX_RB_MRT_BLEND_CONTROL_ALPHA_DEST_FACTOR(FACTOR_ZERO));
 	}
 
+	OUT_PKT0(ring, REG_A4XX_RB_RENDER_COMPONENTS, 1);
+	OUT_RING(ring, A4XX_RB_RENDER_COMPONENTS_RT0(mrt_comp[0]) |
+			A4XX_RB_RENDER_COMPONENTS_RT1(mrt_comp[1]) |
+			A4XX_RB_RENDER_COMPONENTS_RT2(mrt_comp[2]) |
+			A4XX_RB_RENDER_COMPONENTS_RT3(mrt_comp[3]) |
+			A4XX_RB_RENDER_COMPONENTS_RT4(mrt_comp[4]) |
+			A4XX_RB_RENDER_COMPONENTS_RT5(mrt_comp[5]) |
+			A4XX_RB_RENDER_COMPONENTS_RT6(mrt_comp[6]) |
+			A4XX_RB_RENDER_COMPONENTS_RT7(mrt_comp[7]));
+
 	fd4_emit_vertex_bufs(ring, &emit);
 
 	OUT_PKT0(ring, REG_A4XX_GRAS_ALPHA_CONTROL, 1);
@@ -288,12 +290,6 @@ fd4_clear(struct fd_context *ctx, unsigned buffers,
 	OUT_PKT0(ring, REG_A4XX_GRAS_CLEAR_CNTL, 1);
 	OUT_RING(ring, 0x00000000);
 
-	OUT_PKT0(ring, REG_A4XX_RB_CLEAR_COLOR_DW0, 4);
-	OUT_RING(ring, colr);         /* RB_CLEAR_COLOR_DW0 */
-	OUT_RING(ring, colr);         /* RB_CLEAR_COLOR_DW1 */
-	OUT_RING(ring, colr);         /* RB_CLEAR_COLOR_DW2 */
-	OUT_RING(ring, colr);         /* RB_CLEAR_COLOR_DW3 */
-
 	/* until fastclear works: */
 	fd4_emit_const(ring, SHADER_FRAGMENT, 0, 0, 4, color->ui, NULL);
 
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_emit.c b/src/gallium/drivers/freedreno/a4xx/fd4_emit.c
index df96601c747..e7c210d3437 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_emit.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_emit.c
@@ -197,51 +197,90 @@ emit_textures(struct fd_context *ctx, struct fd_ringbuffer *ring,
  * special cases..
  */
 void
-fd4_emit_gmem_restore_tex(struct fd_ringbuffer *ring, struct pipe_surface *psurf)
+fd4_emit_gmem_restore_tex(struct fd_ringbuffer *ring, unsigned nr_bufs,
+		struct pipe_surface **bufs)
 {
-	struct fd_resource *rsc = fd_resource(psurf->texture);
-	unsigned lvl = psurf->u.tex.level;
-	struct fd_resource_slice *slice = fd_resource_slice(rsc, lvl);
-	uint32_t offset = fd_resource_offset(rsc, lvl, psurf->u.tex.first_layer);
-	enum pipe_format format = fd4_gmem_restore_format(psurf->format);
+	unsigned char mrt_comp[A4XX_MAX_RENDER_TARGETS] = {0};
+	int i;
 
-	debug_assert(psurf->u.tex.first_layer == psurf->u.tex.last_layer);
+	for (i = 0; i < A4XX_MAX_RENDER_TARGETS; i++) {
+		mrt_comp[i] = (i < nr_bufs) ? 0xf : 0;
+	}
+
+	OUT_PKT0(ring, REG_A4XX_RB_RENDER_COMPONENTS, 1);
+	OUT_RING(ring, A4XX_RB_RENDER_COMPONENTS_RT0(mrt_comp[0]) |
+			A4XX_RB_RENDER_COMPONENTS_RT1(mrt_comp[1]) |
+			A4XX_RB_RENDER_COMPONENTS_RT2(mrt_comp[2]) |
+			A4XX_RB_RENDER_COMPONENTS_RT3(mrt_comp[3]) |
+			A4XX_RB_RENDER_COMPONENTS_RT4(mrt_comp[4]) |
+			A4XX_RB_RENDER_COMPONENTS_RT5(mrt_comp[5]) |
+			A4XX_RB_RENDER_COMPONENTS_RT6(mrt_comp[6]) |
+			A4XX_RB_RENDER_COMPONENTS_RT7(mrt_comp[7]));
 
 	/* output sampler state: */
-	OUT_PKT3(ring, CP_LOAD_STATE, 4);
+	OUT_PKT3(ring, CP_LOAD_STATE, 2 + (2 * nr_bufs));
 	OUT_RING(ring, CP_LOAD_STATE_0_DST_OFF(0) |
 			CP_LOAD_STATE_0_STATE_SRC(SS_DIRECT) |
 			CP_LOAD_STATE_0_STATE_BLOCK(SB_FRAG_TEX) |
-			CP_LOAD_STATE_0_NUM_UNIT(1));
+			CP_LOAD_STATE_0_NUM_UNIT(nr_bufs));
 	OUT_RING(ring, CP_LOAD_STATE_1_STATE_TYPE(ST_SHADER) |
 			CP_LOAD_STATE_1_EXT_SRC_ADDR(0));
-	OUT_RING(ring, A4XX_TEX_SAMP_0_XY_MAG(A4XX_TEX_NEAREST) |
-			A4XX_TEX_SAMP_0_XY_MIN(A4XX_TEX_NEAREST) |
-			A4XX_TEX_SAMP_0_WRAP_S(A4XX_TEX_CLAMP_TO_EDGE) |
-			A4XX_TEX_SAMP_0_WRAP_T(A4XX_TEX_CLAMP_TO_EDGE) |
-			A4XX_TEX_SAMP_0_WRAP_R(A4XX_TEX_REPEAT));
-	OUT_RING(ring, 0x00000000);
+	for (i = 0; i < nr_bufs; i++) {
+		OUT_RING(ring, A4XX_TEX_SAMP_0_XY_MAG(A4XX_TEX_NEAREST) |
+				A4XX_TEX_SAMP_0_XY_MIN(A4XX_TEX_NEAREST) |
+				A4XX_TEX_SAMP_0_WRAP_S(A4XX_TEX_CLAMP_TO_EDGE) |
+				A4XX_TEX_SAMP_0_WRAP_T(A4XX_TEX_CLAMP_TO_EDGE) |
+				A4XX_TEX_SAMP_0_WRAP_R(A4XX_TEX_REPEAT));
+		OUT_RING(ring, 0x00000000);
+	}
 
 	/* emit texture state: */
-	OUT_PKT3(ring, CP_LOAD_STATE, 10);
+	OUT_PKT3(ring, CP_LOAD_STATE, 2 + (8 * nr_bufs));
 	OUT_RING(ring, CP_LOAD_STATE_0_DST_OFF(0) |
 			CP_LOAD_STATE_0_STATE_SRC(SS_DIRECT) |
 			CP_LOAD_STATE_0_STATE_BLOCK(SB_FRAG_TEX) |
-			CP_LOAD_STATE_0_NUM_UNIT(1));
+			CP_LOAD_STATE_0_NUM_UNIT(nr_bufs));
 	OUT_RING(ring, CP_LOAD_STATE_1_STATE_TYPE(ST_CONSTANTS) |
 			CP_LOAD_STATE_1_EXT_SRC_ADDR(0));
-	OUT_RING(ring, A4XX_TEX_CONST_0_FMT(fd4_pipe2tex(format)) |
-			A4XX_TEX_CONST_0_TYPE(A4XX_TEX_2D) |
-			fd4_tex_swiz(format,  PIPE_SWIZZLE_RED, PIPE_SWIZZLE_GREEN,
-					PIPE_SWIZZLE_BLUE, PIPE_SWIZZLE_ALPHA));
-	OUT_RING(ring, A4XX_TEX_CONST_1_WIDTH(psurf->width) |
-			A4XX_TEX_CONST_1_HEIGHT(psurf->height));
-	OUT_RING(ring, A4XX_TEX_CONST_2_PITCH(slice->pitch * rsc->cpp));
-	OUT_RING(ring, 0x00000000);
-	OUT_RELOC(ring, rsc->bo, offset, 0, 0);
-	OUT_RING(ring, 0x00000000);
-	OUT_RING(ring, 0x00000000);
-	OUT_RING(ring, 0x00000000);
+	for (i = 0; i < nr_bufs; i++) {
+		if (bufs[i]) {
+			struct fd_resource *rsc = fd_resource(bufs[i]->texture);
+			unsigned lvl = bufs[i]->u.tex.level;
+			struct fd_resource_slice *slice = fd_resource_slice(rsc, lvl);
+			uint32_t offset = fd_resource_offset(rsc, lvl, bufs[i]->u.tex.first_layer);
+			enum pipe_format format = fd4_gmem_restore_format(bufs[i]->format);
+
+			debug_assert(bufs[i]->u.tex.first_layer == bufs[i]->u.tex.last_layer);
+
+			OUT_RING(ring, A4XX_TEX_CONST_0_FMT(fd4_pipe2tex(format)) |
+					A4XX_TEX_CONST_0_TYPE(A4XX_TEX_2D) |
+					fd4_tex_swiz(format,  PIPE_SWIZZLE_RED, PIPE_SWIZZLE_GREEN,
+							PIPE_SWIZZLE_BLUE, PIPE_SWIZZLE_ALPHA));
+			OUT_RING(ring, A4XX_TEX_CONST_1_WIDTH(bufs[i]->width) |
+					A4XX_TEX_CONST_1_HEIGHT(bufs[i]->height));
+			OUT_RING(ring, A4XX_TEX_CONST_2_PITCH(slice->pitch * rsc->cpp));
+			OUT_RING(ring, 0x00000000);
+			OUT_RELOC(ring, rsc->bo, offset, 0, 0);
+			OUT_RING(ring, 0x00000000);
+			OUT_RING(ring, 0x00000000);
+			OUT_RING(ring, 0x00000000);
+		} else {
+			OUT_RING(ring, A4XX_TEX_CONST_0_FMT(0) |
+					A4XX_TEX_CONST_0_TYPE(A4XX_TEX_2D) |
+					A4XX_TEX_CONST_0_SWIZ_X(A4XX_TEX_ONE) |
+					A4XX_TEX_CONST_0_SWIZ_Y(A4XX_TEX_ONE) |
+					A4XX_TEX_CONST_0_SWIZ_Z(A4XX_TEX_ONE) |
+					A4XX_TEX_CONST_0_SWIZ_W(A4XX_TEX_ONE));
+			OUT_RING(ring, A4XX_TEX_CONST_1_WIDTH(0) |
+					A4XX_TEX_CONST_1_HEIGHT(0));
+			OUT_RING(ring, A4XX_TEX_CONST_2_PITCH(0));
+			OUT_RING(ring, 0x00000000);
+			OUT_RING(ring, 0x00000000);
+			OUT_RING(ring, 0x00000000);
+			OUT_RING(ring, 0x00000000);
+			OUT_RING(ring, 0x00000000);
+		}
+	}
 }
 
 void
@@ -348,6 +387,25 @@ fd4_emit_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
 
 	emit_marker(ring, 5);
 
+	if ((dirty & FD_DIRTY_FRAMEBUFFER) && !emit->key.binning_pass) {
+		struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
+		unsigned char mrt_comp[A4XX_MAX_RENDER_TARGETS] = {0};
+
+		for (unsigned i = 0; i < A4XX_MAX_RENDER_TARGETS; i++) {
+			mrt_comp[i] = ((i < pfb->nr_cbufs) && pfb->cbufs[i]) ? 0xf : 0;
+		}
+
+		OUT_PKT0(ring, REG_A4XX_RB_RENDER_COMPONENTS, 1);
+		OUT_RING(ring, A4XX_RB_RENDER_COMPONENTS_RT0(mrt_comp[0]) |
+				A4XX_RB_RENDER_COMPONENTS_RT1(mrt_comp[1]) |
+				A4XX_RB_RENDER_COMPONENTS_RT2(mrt_comp[2]) |
+				A4XX_RB_RENDER_COMPONENTS_RT3(mrt_comp[3]) |
+				A4XX_RB_RENDER_COMPONENTS_RT4(mrt_comp[4]) |
+				A4XX_RB_RENDER_COMPONENTS_RT5(mrt_comp[5]) |
+				A4XX_RB_RENDER_COMPONENTS_RT6(mrt_comp[6]) |
+				A4XX_RB_RENDER_COMPONENTS_RT7(mrt_comp[7]));
+	}
+
 	if ((dirty & (FD_DIRTY_ZSA | FD_DIRTY_PROG)) && !emit->key.binning_pass) {
 		uint32_t val = fd4_zsa_stateobj(ctx->zsa)->rb_render_control;
 
@@ -472,8 +530,10 @@ fd4_emit_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
 		OUT_RING(ring, A4XX_GRAS_CL_VPORT_ZSCALE_0(ctx->viewport.scale[2]));
 	}
 
-	if (dirty & FD_DIRTY_PROG)
-		fd4_program_emit(ring, emit);
+	if (dirty & FD_DIRTY_PROG) {
+		struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
+		fd4_program_emit(ring, emit, pfb->nr_cbufs, pfb->cbufs);
+	}
 
 	if (emit->prog == &ctx->prog) { /* evil hack to deal sanely with clear path */
 		ir3_emit_consts(vp, ring, emit->info, dirty);
@@ -690,9 +750,6 @@ fd4_emit_restore(struct fd_context *ctx)
 	OUT_PKT0(ring, REG_A4XX_RB_FS_OUTPUT, 1);
 	OUT_RING(ring, A4XX_RB_FS_OUTPUT_SAMPLE_MASK(0xffff));
 
-	OUT_PKT0(ring, REG_A4XX_RB_RENDER_COMPONENTS, 1);
-	OUT_RING(ring, A4XX_RB_RENDER_COMPONENTS_RT0(0xf));
-
 	OUT_PKT0(ring, REG_A4XX_GRAS_CLEAR_CNTL, 1);
 	OUT_RING(ring, A4XX_GRAS_CLEAR_CNTL_NOT_FASTCLEAR);
 
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_emit.h b/src/gallium/drivers/freedreno/a4xx/fd4_emit.h
index 7debee59471..99c7596fb82 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_emit.h
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_emit.h
@@ -43,7 +43,7 @@ void fd4_emit_const(struct fd_ringbuffer *ring, enum shader_t type,
 		const uint32_t *dwords, struct pipe_resource *prsc);
 
 void fd4_emit_gmem_restore_tex(struct fd_ringbuffer *ring,
-		struct pipe_surface *psurf);
+		unsigned nr_bufs, struct pipe_surface **bufs);
 
 /* grouped together emit-state for prog/vertex/state emit: */
 struct fd4_emit {
@@ -51,8 +51,6 @@ struct fd4_emit {
 	const struct fd_program_stateobj *prog;
 	const struct pipe_draw_info *info;
 	struct ir3_shader_key key;
-	enum a4xx_color_fmt format;
-	enum pipe_format pformat;
 	uint32_t dirty;
 
 	/* cached to avoid repeated lookups of same variants: */
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_gmem.c b/src/gallium/drivers/freedreno/a4xx/fd4_gmem.c
index 976255f5879..fce5d7a3930 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_gmem.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_gmem.c
@@ -44,12 +44,6 @@
 #include "fd4_format.h"
 #include "fd4_zsa.h"
 
-static const struct ir3_shader_key key = {
-		// XXX should set this based on render target format!  We don't
-		// want half_precision if float32 render target!!!
-		.half_precision = true,
-};
-
 static void
 emit_mrt(struct fd_ringbuffer *ring, unsigned nr_bufs,
 		struct pipe_surface **bufs, uint32_t *bases, uint32_t bin_w)
@@ -94,6 +88,8 @@ emit_mrt(struct fd_ringbuffer *ring, unsigned nr_bufs,
 			} else {
 				stride = slice->pitch * rsc->cpp;
 			}
+		} else if ((i < nr_bufs) && bases) {
+			base = bases[i];
 		}
 
 		OUT_PKT0(ring, REG_A4XX_RB_MRT_BUF_INFO(i), 3);
@@ -101,7 +97,7 @@ emit_mrt(struct fd_ringbuffer *ring, unsigned nr_bufs,
 				A4XX_RB_MRT_BUF_INFO_COLOR_TILE_MODE(tile_mode) |
 				A4XX_RB_MRT_BUF_INFO_COLOR_BUF_PITCH(stride) |
 				A4XX_RB_MRT_BUF_INFO_COLOR_SWAP(swap));
-		if (bin_w || (i >= nr_bufs)) {
+		if (bin_w || (i >= nr_bufs) || !bufs[i]) {
 			OUT_RING(ring, base);
 			OUT_RING(ring, A4XX_RB_MRT_CONTROL3_STRIDE(stride));
 		} else {
@@ -115,20 +111,6 @@ emit_mrt(struct fd_ringbuffer *ring, unsigned nr_bufs,
 	}
 }
 
-static uint32_t
-depth_base(struct fd_context *ctx)
-{
-	struct fd_gmem_stateobj *gmem = &ctx->gmem;
-	struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
-	uint32_t cpp = 4;
-	if (pfb->cbufs[0]) {
-		struct fd_resource *rsc =
-				fd_resource(pfb->cbufs[0]->texture);
-		cpp = rsc->cpp;
-	}
-	return align(gmem->bin_w * gmem->bin_h * cpp, 0x4000);
-}
-
 /* transfer from gmem to system memory (ie. normal RAM) */
 
 static void
@@ -163,13 +145,15 @@ static void
 fd4_emit_tile_gmem2mem(struct fd_context *ctx, struct fd_tile *tile)
 {
 	struct fd4_context *fd4_ctx = fd4_context(ctx);
+	struct fd_gmem_stateobj *gmem = &ctx->gmem;
 	struct fd_ringbuffer *ring = ctx->ring;
 	struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 	struct fd4_emit emit = {
 			.vtx = &fd4_ctx->solid_vbuf_state,
 			.prog = &ctx->solid_prog,
-			.key = key,
-			.format = fd4_emit_format(pfb->cbufs[0]),
+			.key = {
+				.half_precision = true,
+			},
 	};
 
 	OUT_PKT0(ring, REG_A4XX_RB_DEPTH_CONTROL, 1);
@@ -238,16 +222,22 @@ fd4_emit_tile_gmem2mem(struct fd_context *ctx, struct fd_tile *tile)
 	OUT_RING(ring, 0);            /* VFD_INDEX_OFFSET */
 	OUT_RING(ring, 0);            /* ??? UNKNOWN_2209 */
 
-	fd4_program_emit(ring, &emit);
+	fd4_program_emit(ring, &emit, 0, NULL);
 	fd4_emit_vertex_bufs(ring, &emit);
 
 	if (ctx->resolve & (FD_BUFFER_DEPTH | FD_BUFFER_STENCIL)) {
-		uint32_t base = depth_base(ctx);
-		emit_gmem2mem_surf(ctx, base, pfb->zsbuf);
+		emit_gmem2mem_surf(ctx, gmem->zsbuf_base[0], pfb->zsbuf);
 	}
 
 	if (ctx->resolve & FD_BUFFER_COLOR) {
-		emit_gmem2mem_surf(ctx, 0, pfb->cbufs[0]);
+		unsigned i;
+		for (i = 0; i < pfb->nr_cbufs; i++) {
+			if (!pfb->cbufs[i])
+				continue;
+			if (!(ctx->resolve & (PIPE_CLEAR_COLOR0 << i)))
+				continue;
+			emit_gmem2mem_surf(ctx, gmem->cbuf_base[i], pfb->cbufs[i]);
+		}
 	}
 
 	OUT_PKT0(ring, REG_A4XX_GRAS_SC_CONTROL, 1);
@@ -260,14 +250,14 @@ fd4_emit_tile_gmem2mem(struct fd_context *ctx, struct fd_tile *tile)
 /* transfer from system memory to gmem */
 
 static void
-emit_mem2gmem_surf(struct fd_context *ctx, uint32_t base,
-		struct pipe_surface *psurf, uint32_t bin_w)
+emit_mem2gmem_surf(struct fd_context *ctx, uint32_t *bases,
+		struct pipe_surface **bufs, uint32_t nr_bufs, uint32_t bin_w)
 {
 	struct fd_ringbuffer *ring = ctx->ring;
 
-	emit_mrt(ring, 1, &psurf, &base, bin_w);
+	emit_mrt(ring, nr_bufs, bufs, bases, bin_w);
 
-	fd4_emit_gmem_restore_tex(ring, psurf);
+	fd4_emit_gmem_restore_tex(ring, nr_bufs, bufs);
 
 	fd4_draw(ctx, ring, DI_PT_RECTLIST, IGNORE_VISIBILITY,
 			DI_SRC_SEL_AUTO_INDEX, 2, 1, INDEX_SIZE_IGN, 0, 0, NULL);
@@ -282,10 +272,13 @@ fd4_emit_tile_mem2gmem(struct fd_context *ctx, struct fd_tile *tile)
 	struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 	struct fd4_emit emit = {
 			.vtx = &fd4_ctx->blit_vbuf_state,
+			/* NOTE: They all use the same VP, this is for vtx bufs. */
 			.prog = &ctx->blit_prog[0],
-			.key = key,
-			.format = fd4_emit_format(pfb->cbufs[0]),
+			.key = {
+				.half_precision = fd_half_precision(pfb),
+			},
 	};
+	unsigned char mrt_comp[A4XX_MAX_RENDER_TARGETS] = {0};
 	float x0, y0, x1, y1;
 	unsigned bin_w = tile->bin_w;
 	unsigned bin_h = tile->bin_h;
@@ -304,7 +297,9 @@ fd4_emit_tile_mem2gmem(struct fd_context *ctx, struct fd_tile *tile)
 	OUT_RING(ring, fui(x1));
 	OUT_RING(ring, fui(y1));
 
-	for (i = 0; i < 8; i++) {
+	for (i = 0; i < A4XX_MAX_RENDER_TARGETS; i++) {
+		mrt_comp[i] = ((i < pfb->nr_cbufs) && pfb->cbufs[i]) ? 0xf : 0;
+
 		OUT_PKT0(ring, REG_A4XX_RB_MRT_CONTROL(i), 1);
 		OUT_RING(ring, A4XX_RB_MRT_CONTROL_FASTCLEAR |
 				A4XX_RB_MRT_CONTROL_B11 |
@@ -319,6 +314,16 @@ fd4_emit_tile_mem2gmem(struct fd_context *ctx, struct fd_tile *tile)
 				A4XX_RB_MRT_BLEND_CONTROL_ALPHA_DEST_FACTOR(FACTOR_ZERO));
 	}
 
+	OUT_PKT0(ring, REG_A4XX_RB_RENDER_COMPONENTS, 1);
+	OUT_RING(ring, A4XX_RB_RENDER_COMPONENTS_RT0(mrt_comp[0]) |
+			A4XX_RB_RENDER_COMPONENTS_RT1(mrt_comp[1]) |
+			A4XX_RB_RENDER_COMPONENTS_RT2(mrt_comp[2]) |
+			A4XX_RB_RENDER_COMPONENTS_RT3(mrt_comp[3]) |
+			A4XX_RB_RENDER_COMPONENTS_RT4(mrt_comp[4]) |
+			A4XX_RB_RENDER_COMPONENTS_RT5(mrt_comp[5]) |
+			A4XX_RB_RENDER_COMPONENTS_RT6(mrt_comp[6]) |
+			A4XX_RB_RENDER_COMPONENTS_RT7(mrt_comp[7]));
+
 	OUT_PKT0(ring, REG_A4XX_RB_RENDER_CONTROL, 1);
 	OUT_RING(ring, 0x8);          /* XXX RB_RENDER_CONTROL */
 
@@ -381,7 +386,6 @@ fd4_emit_tile_mem2gmem(struct fd_context *ctx, struct fd_tile *tile)
 	OUT_RING(ring, 0);            /* VFD_INDEX_OFFSET */
 	OUT_RING(ring, 0);            /* ??? UNKNOWN_2209 */
 
-	fd4_program_emit(ring, &emit);
 	fd4_emit_vertex_bufs(ring, &emit);
 
 	/* for gmem pitch/base calculations, we need to use the non-
@@ -390,11 +394,17 @@ fd4_emit_tile_mem2gmem(struct fd_context *ctx, struct fd_tile *tile)
 	bin_w = gmem->bin_w;
 	bin_h = gmem->bin_h;
 
-	if (fd_gmem_needs_restore(ctx, tile, FD_BUFFER_DEPTH | FD_BUFFER_STENCIL))
-		emit_mem2gmem_surf(ctx, depth_base(ctx), pfb->zsbuf, bin_w);
+	if (fd_gmem_needs_restore(ctx, tile, FD_BUFFER_DEPTH | FD_BUFFER_STENCIL)) {
+		emit.prog = &ctx->blit_prog[0];
+		fd4_program_emit(ring, &emit, 1, &pfb->zsbuf);
+		emit_mem2gmem_surf(ctx, gmem->zsbuf_base, &pfb->zsbuf, 1, bin_w);
+	}
 
-	if (fd_gmem_needs_restore(ctx, tile, FD_BUFFER_COLOR))
-		emit_mem2gmem_surf(ctx, 0, pfb->cbufs[0], bin_w);
+	if (fd_gmem_needs_restore(ctx, tile, FD_BUFFER_COLOR)) {
+		emit.prog = &ctx->blit_prog[pfb->nr_cbufs - 1];
+		fd4_program_emit(ring, &emit, pfb->nr_cbufs, pfb->cbufs);
+		emit_mem2gmem_surf(ctx, gmem->cbuf_base, pfb->cbufs, pfb->nr_cbufs, bin_w);
+	}
 
 	OUT_PKT0(ring, REG_A4XX_GRAS_SC_CONTROL, 1);
 	OUT_RING(ring, A4XX_GRAS_SC_CONTROL_RENDER_MODE(RB_RENDERING_PASS) |
@@ -537,7 +547,7 @@ fd4_emit_tile_prep(struct fd_context *ctx, struct fd_tile *tile)
 	uint32_t reg;
 
 	OUT_PKT0(ring, REG_A4XX_RB_DEPTH_INFO, 3);
-	reg = A4XX_RB_DEPTH_INFO_DEPTH_BASE(depth_base(ctx));
+	reg = A4XX_RB_DEPTH_INFO_DEPTH_BASE(gmem->zsbuf_base[0]);
 	if (pfb->zsbuf) {
 		reg |= A4XX_RB_DEPTH_INFO_DEPTH_FORMAT(fd4_pipe2depth(pfb->zsbuf->format));
 	}
@@ -586,7 +596,7 @@ fd4_emit_tile_renderprep(struct fd_context *ctx, struct fd_tile *tile)
 	OUT_RING(ring, CP_SET_BIN_1_X1(x1) | CP_SET_BIN_1_Y1(y1));
 	OUT_RING(ring, CP_SET_BIN_2_X2(x2) | CP_SET_BIN_2_Y2(y2));
 
-	emit_mrt(ring, pfb->nr_cbufs, pfb->cbufs, NULL, gmem->bin_w);
+	emit_mrt(ring, pfb->nr_cbufs, pfb->cbufs, gmem->cbuf_base, gmem->bin_w);
 
 	/* setup scissor/offset for current tile: */
 	OUT_PKT0(ring, REG_A4XX_RB_BIN_OFFSET, 1);
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_program.c b/src/gallium/drivers/freedreno/a4xx/fd4_program.c
index b3c06e3aae1..c7a6dffea7b 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_program.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_program.c
@@ -31,8 +31,6 @@
 #include "util/u_memory.h"
 #include "util/u_inlines.h"
 #include "util/u_format.h"
-#include "tgsi/tgsi_dump.h"
-#include "tgsi/tgsi_parse.h"
 
 #include "freedreno_program.h"
 
@@ -213,14 +211,17 @@ setup_stages(struct fd4_emit *emit, struct stage *s)
 }
 
 void
-fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit)
+fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit,
+		int nr, struct pipe_surface **bufs)
 {
 	struct stage s[MAX_STAGES];
-	uint32_t pos_regid, posz_regid, psize_regid, color_regid;
+	uint32_t pos_regid, posz_regid, psize_regid, color_regid[8];
 	uint32_t face_regid, coord_regid, zwcoord_regid;
 	int constmode;
 	int i, j, k;
 
+	debug_assert(nr <= ARRAY_SIZE(color_regid));
+
 	setup_stages(emit, s);
 
 	/* blob seems to always use constmode currently: */
@@ -232,11 +233,30 @@ fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit)
 		ir3_semantic_name(TGSI_SEMANTIC_POSITION, 0));
 	psize_regid = ir3_find_output_regid(s[VS].v,
 		ir3_semantic_name(TGSI_SEMANTIC_PSIZE, 0));
-	color_regid = ir3_find_output_regid(s[FS].v,
-		ir3_semantic_name(TGSI_SEMANTIC_COLOR, 0));
+	if (s[FS].v->color0_mrt) {
+		color_regid[0] = color_regid[1] = color_regid[2] = color_regid[3] =
+		color_regid[4] = color_regid[5] = color_regid[6] = color_regid[7] =
+			ir3_find_output_regid(s[FS].v, ir3_semantic_name(TGSI_SEMANTIC_COLOR, 0));
+	} else {
+		const struct ir3_shader_variant *fp = s[FS].v;
+		memset(color_regid, 0, sizeof(color_regid));
+		for (i = 0; i < fp->outputs_count; i++) {
+			ir3_semantic sem = fp->outputs[i].semantic;
+			unsigned idx = sem2idx(sem);
+			if (sem2name(sem) != TGSI_SEMANTIC_COLOR)
+				continue;
+			debug_assert(idx < ARRAY_SIZE(color_regid));
+			color_regid[idx] = fp->outputs[i].regid;
+		}
+	}
+
+	/* adjust regids for alpha output formats. there is no alpha render
+	 * format, so it's just treated like red
+	 */
+	for (i = 0; i < nr; i++)
+		if (util_format_is_alpha(pipe_surface_format(bufs[i])))
+			color_regid[i] += 3;
 
-	if (util_format_is_alpha(emit->pformat))
-		color_regid += 3;
 
 	/* TODO get these dynamically: */
 	face_regid = s[FS].v->frag_face ? regid(0,0) : regid(63,0);
@@ -419,29 +439,24 @@ fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit)
 					A4XX_RB_RENDER_CONTROL2_WCOORD));
 
 	OUT_PKT0(ring, REG_A4XX_RB_FS_OUTPUT_REG, 1);
-	OUT_RING(ring, A4XX_RB_FS_OUTPUT_REG_MRT(1) |
+	OUT_RING(ring, A4XX_RB_FS_OUTPUT_REG_MRT(MAX2(1, nr)) |
 			COND(s[FS].v->writes_pos, A4XX_RB_FS_OUTPUT_REG_FRAG_WRITES_Z));
 
 	OUT_PKT0(ring, REG_A4XX_SP_FS_OUTPUT_REG, 1);
-	if (s[FS].v->writes_pos) {
-		OUT_RING(ring, 0x00000001 |
-				A4XX_SP_FS_OUTPUT_REG_DEPTH_ENABLE |
-				A4XX_SP_FS_OUTPUT_REG_DEPTH_REGID(posz_regid));
-	} else {
-		OUT_RING(ring, 0x00000001);
-	}
+	OUT_RING(ring, A4XX_SP_FS_OUTPUT_REG_MRT(MAX2(1, nr)) |
+			COND(s[FS].v->writes_pos, A4XX_SP_FS_OUTPUT_REG_DEPTH_ENABLE) |
+			A4XX_SP_FS_OUTPUT_REG_DEPTH_REGID(posz_regid));
 
 	OUT_PKT0(ring, REG_A4XX_SP_FS_MRT_REG(0), 8);
-	OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(color_regid) |
-			A4XX_SP_FS_MRT_REG_MRTFORMAT(emit->format) |
-			COND(emit->key.half_precision, A4XX_SP_FS_MRT_REG_HALF_PRECISION));
-	OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
-	OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
-	OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
-	OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
-	OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
-	OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
-	OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
+	for (i = 0; i < 8; i++) {
+		enum a4xx_color_fmt format = 0;
+		if (i < nr)
+			format = fd4_emit_format(bufs[i]);
+		OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(color_regid[i]) |
+				A4XX_SP_FS_MRT_REG_MRTFORMAT(format) |
+				COND(emit->key.half_precision,
+					A4XX_SP_FS_MRT_REG_HALF_PRECISION));
+	}
 
 	if (emit->key.binning_pass) {
 		OUT_PKT0(ring, REG_A4XX_VPC_ATTR, 2);
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_program.h b/src/gallium/drivers/freedreno/a4xx/fd4_program.h
index 52306a4c60d..8dfccaf9d74 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_program.h
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_program.h
@@ -39,7 +39,8 @@ struct fd4_shader_stateobj {
 
 struct fd4_emit;
 
-void fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit);
+void fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit,
+		int nr, struct pipe_surface **bufs);
 
 void fd4_prog_init(struct pipe_context *pctx);
 
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_screen.c b/src/gallium/drivers/freedreno/a4xx/fd4_screen.c
index e8cbb2d201a..d8ea414f300 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_screen.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_screen.c
@@ -102,7 +102,7 @@ void
 fd4_screen_init(struct pipe_screen *pscreen)
 {
 	struct fd_screen *screen = fd_screen(pscreen);
-	screen->max_rts = 1;
+	screen->max_rts = A4XX_MAX_RENDER_TARGETS;
 	screen->compiler = ir3_compiler_create(screen->gpu_id);
 	pscreen->context_create = fd4_context_create;
 	pscreen->is_format_supported = fd4_screen_is_format_supported;
diff --git a/src/gallium/drivers/freedreno/freedreno_draw.c b/src/gallium/drivers/freedreno/freedreno_draw.c
index a04cc879630..6831a58749c 100644
--- a/src/gallium/drivers/freedreno/freedreno_draw.c
+++ b/src/gallium/drivers/freedreno/freedreno_draw.c
@@ -265,7 +265,8 @@ fd_clear(struct pipe_context *pctx, unsigned buffers,
 			FD_DIRTY_SAMPLE_MASK |
 			FD_DIRTY_PROG |
 			FD_DIRTY_CONSTBUF |
-			FD_DIRTY_BLEND;
+			FD_DIRTY_BLEND |
+			FD_DIRTY_FRAMEBUFFER;
 
 	if (fd_mesa_debug & FD_DBG_DCLEAR)
 		ctx->dirty = 0xffffffff;
diff --git a/src/gallium/drivers/freedreno/freedreno_util.h b/src/gallium/drivers/freedreno/freedreno_util.h
index 2880e890a1c..b5c5db91788 100644
--- a/src/gallium/drivers/freedreno/freedreno_util.h
+++ b/src/gallium/drivers/freedreno/freedreno_util.h
@@ -58,7 +58,7 @@ enum adreno_stencil_op fd_stencil_op(unsigned op);
 #define A3XX_MAX_RENDER_TARGETS 4
 #define A4XX_MAX_RENDER_TARGETS 8
 /* for now until a4xx MRT support: */
-#define MAX_RENDER_TARGETS A3XX_MAX_RENDER_TARGETS
+#define MAX_RENDER_TARGETS A4XX_MAX_RENDER_TARGETS
 
 #define FD_DBG_MSGS     0x0001
 #define FD_DBG_DISASM   0x0002