freedreno/a5xx: transform-feedback support

We'll need to revisit when adding hw binning pass support, whether we
can still do this in main draw step, as we do w/ a3xx/a4xx, or if we
needed to move it to the binning stage.

Still some failing piglits but most tests pass and the common cases seem
to work.

Signed-off-by: Rob Clark <robdclark@gmail.com>

src/gallium/drivers/freedreno/a5xx/fd5_draw.c

@@ -173,6 +173,17 @@ fd5_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info)
 //	emit.fp = NULL;
 //	draw_impl(ctx, ctx->batch->binning, &emit);
 
+	if (emit.streamout_mask) {
+		struct fd_ringbuffer *ring = ctx->batch->draw;
+
+		for (unsigned i = 0; i < PIPE_MAX_SO_BUFFERS; i++) {
+			if (emit.streamout_mask & (1 << i)) {
+				OUT_PKT7(ring, CP_EVENT_WRITE, 1);
+				OUT_RING(ring, FLUSH_SO_0 + i);
+			}
+		}
+	}
+
 	return true;
 }
 

src/gallium/drivers/freedreno/a5xx/fd5_emit.c

@@ -526,19 +526,63 @@ fd5_emit_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
 		OUT_RING(ring, A5XX_GRAS_CL_VPORT_ZSCALE_0(ctx->viewport.scale[2]));
 	}
 
-	if (dirty & (FD_DIRTY_PROG | FD_DIRTY_FRAMEBUFFER)) {
+	if (dirty & FD_DIRTY_PROG)
+		fd5_program_emit(ring, emit);
+
+	if (dirty & (FD_DIRTY_FRAMEBUFFER | FD_DIRTY_RASTERIZER)) {
 		struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
-		unsigned n = pfb->nr_cbufs;
-		/* if we have depth/stencil, we need at least on MRT: */
-		if (pfb->zsbuf)
-			n = MAX2(1, n);
-		fd5_program_emit(ring, emit, n, pfb->cbufs);
+		uint32_t posz_regid = ir3_find_output_regid(fp, FRAG_RESULT_DEPTH);
+		unsigned nr = pfb->nr_cbufs;
+
+		if (emit->key.binning_pass)
+			nr = 0;
+		else if (ctx->rasterizer->rasterizer_discard)
+			nr = 0;
+
+		OUT_PKT4(ring, REG_A5XX_RB_FS_OUTPUT_CNTL, 1);
+		OUT_RING(ring, A5XX_RB_FS_OUTPUT_CNTL_MRT(nr) |
+				COND(fp->writes_pos, A5XX_RB_FS_OUTPUT_CNTL_FRAG_WRITES_Z));
+
+		OUT_PKT4(ring, REG_A5XX_SP_FS_OUTPUT_CNTL, 1);
+		OUT_RING(ring, A5XX_SP_FS_OUTPUT_CNTL_MRT(nr) |
+				A5XX_SP_FS_OUTPUT_CNTL_DEPTH_REGID(posz_regid) |
+				A5XX_SP_FS_OUTPUT_CNTL_SAMPLEMASK_REGID(regid(63, 0)));
 	}
 
 	if (emit->prog == &ctx->prog) { /* evil hack to deal sanely with clear path */
 		ir3_emit_consts(vp, ring, ctx, emit->info, dirty);
 		if (!emit->key.binning_pass)
 			ir3_emit_consts(fp, ring, ctx, emit->info, dirty);
+
+		struct pipe_stream_output_info *info = &vp->shader->stream_output;
+		if (info->num_outputs) {
+			struct fd_streamout_stateobj *so = &ctx->streamout;
+
+			for (unsigned i = 0; i < so->num_targets; i++) {
+				struct pipe_stream_output_target *target = so->targets[i];
+
+				if (!target)
+					continue;
+
+				unsigned offset = (so->offsets[i] * info->stride[i] * 4) +
+						target->buffer_offset;
+
+				OUT_PKT4(ring, REG_A5XX_VPC_SO_BUFFER_BASE_LO(i), 3);
+				/* VPC_SO[i].BUFFER_BASE_LO: */
+				OUT_RELOCW(ring, fd_resource(target->buffer)->bo, 0, 0, 0);
+				OUT_RING(ring, target->buffer_size + offset);
+
+				OUT_PKT4(ring, REG_A5XX_VPC_SO_BUFFER_OFFSET(i), 3);
+				OUT_RING(ring, offset);
+				/* VPC_SO[i].FLUSH_BASE_LO/HI: */
+				// TODO just give hw a dummy addr for now.. we should
+				// be using this an then CP_MEM_TO_REG to set the
+				// VPC_SO[i].BUFFER_OFFSET for the next draw..
+				OUT_RELOCW(ring, fd5_context(ctx)->blit_mem, 0x100, 0, 0);
+
+				emit->streamout_mask |= (1 << i);
+			}
+		}
 	}
 
 	if ((dirty & FD_DIRTY_BLEND)) {

src/gallium/drivers/freedreno/a5xx/fd5_emit.h

@@ -54,6 +54,8 @@ struct fd5_emit {
 	/* cached to avoid repeated lookups of same variants: */
 	const struct ir3_shader_variant *vp, *fp;
 	/* TODO: other shader stages.. */
+
+	unsigned streamout_mask;
 };
 
 static inline enum a5xx_color_fmt fd5_emit_format(struct pipe_surface *surf)

src/gallium/drivers/freedreno/a5xx/fd5_program.c

@@ -132,6 +132,122 @@ emit_shader(struct fd_ringbuffer *ring, const struct ir3_shader_variant *so)
 	}
 }
 
+/* Add any missing varyings needed for stream-out.  Otherwise varyings not
+ * used by fragment shader will be stripped out.
+ */
+static void
+link_stream_out(struct ir3_shader_linkage *l, const struct ir3_shader_variant *v)
+{
+	const struct pipe_stream_output_info *strmout = &v->shader->stream_output;
+
+	/*
+	 * First, any stream-out varyings not already in linkage map (ie. also
+	 * consumed by frag shader) need to be added:
+	 */
+	for (unsigned i = 0; i < strmout->num_outputs; i++) {
+		const struct pipe_stream_output *out = &strmout->output[i];
+		unsigned k = out->register_index;
+		unsigned compmask =
+			(1 << (out->num_components + out->start_component)) - 1;
+		unsigned idx, nextloc = 0;
+
+		/* psize/pos need to be the last entries in linkage map, and will
+		 * get added link_stream_out, so skip over them:
+		 */
+		if ((v->outputs[k].slot == VARYING_SLOT_PSIZ) ||
+				(v->outputs[k].slot == VARYING_SLOT_POS))
+			continue;
+
+		for (idx = 0; idx < l->cnt; idx++) {
+			if (l->var[idx].regid == v->outputs[k].regid)
+				break;
+			nextloc = MAX2(nextloc, l->var[idx].loc + 4);
+		}
+
+		/* add if not already in linkage map: */
+		if (idx == l->cnt)
+			ir3_link_add(l, v->outputs[k].regid, compmask, nextloc);
+
+		/* expand component-mask if needed, ie streaming out all components
+		 * but frag shader doesn't consume all components:
+		 */
+		if (compmask & ~l->var[idx].compmask) {
+			l->var[idx].compmask |= compmask;
+			l->max_loc = MAX2(l->max_loc,
+				l->var[idx].loc + util_last_bit(l->var[idx].compmask));
+		}
+	}
+}
+
+/* TODO maybe some of this we could pre-compute once rather than having
+ * so much draw-time logic?
+ */
+static void
+emit_stream_out(struct fd_ringbuffer *ring, const struct ir3_shader_variant *v,
+		struct ir3_shader_linkage *l)
+{
+	const struct pipe_stream_output_info *strmout = &v->shader->stream_output;
+	unsigned ncomp[PIPE_MAX_SO_BUFFERS] = {0};
+	unsigned prog[align(l->max_loc, 2) / 2];
+
+	memset(prog, 0, sizeof(prog));
+
+	for (unsigned i = 0; i < strmout->num_outputs; i++) {
+		const struct pipe_stream_output *out = &strmout->output[i];
+		unsigned k = out->register_index;
+		unsigned idx;
+
+		ncomp[out->output_buffer] += out->num_components;
+
+		/* linkage map sorted by order frag shader wants things, so
+		 * a bit less ideal here..
+		 */
+		for (idx = 0; idx < l->cnt; idx++)
+			if (l->var[idx].regid == v->outputs[k].regid)
+				break;
+
+		debug_assert(idx < l->cnt);
+
+		for (unsigned j = 0; j < out->num_components; j++) {
+			unsigned c   = j + out->start_component;
+			unsigned loc = l->var[idx].loc + c;
+			unsigned off = j + out->dst_offset;  /* in dwords */
+
+			if (loc & 1) {
+				prog[loc/2] |= A5XX_VPC_SO_PROG_B_EN |
+						A5XX_VPC_SO_PROG_B_BUF(out->output_buffer) |
+						A5XX_VPC_SO_PROG_B_OFF(off * 4);
+			} else {
+				prog[loc/2] |= A5XX_VPC_SO_PROG_A_EN |
+						A5XX_VPC_SO_PROG_A_BUF(out->output_buffer) |
+						A5XX_VPC_SO_PROG_A_OFF(off * 4);
+			}
+		}
+	}
+
+	OUT_PKT7(ring, CP_CONTEXT_REG_BUNCH, 12 + (2 * ARRAY_SIZE(prog)));
+	OUT_RING(ring, REG_A5XX_VPC_SO_BUF_CNTL);
+	OUT_RING(ring, A5XX_VPC_SO_BUF_CNTL_ENABLE |
+			COND(ncomp[0] > 0, A5XX_VPC_SO_BUF_CNTL_BUF0) |
+			COND(ncomp[1] > 0, A5XX_VPC_SO_BUF_CNTL_BUF1) |
+			COND(ncomp[2] > 0, A5XX_VPC_SO_BUF_CNTL_BUF2) |
+			COND(ncomp[3] > 0, A5XX_VPC_SO_BUF_CNTL_BUF3));
+	OUT_RING(ring, REG_A5XX_VPC_SO_NCOMP(0));
+	OUT_RING(ring, ncomp[0]);
+	OUT_RING(ring, REG_A5XX_VPC_SO_NCOMP(1));
+	OUT_RING(ring, ncomp[1]);
+	OUT_RING(ring, REG_A5XX_VPC_SO_NCOMP(2));
+	OUT_RING(ring, ncomp[2]);
+	OUT_RING(ring, REG_A5XX_VPC_SO_NCOMP(3));
+	OUT_RING(ring, ncomp[3]);
+	OUT_RING(ring, REG_A5XX_VPC_SO_CNTL);
+	OUT_RING(ring, A5XX_VPC_SO_CNTL_ENABLE);
+	for (unsigned i = 0; i < ARRAY_SIZE(prog); i++) {
+		OUT_RING(ring, REG_A5XX_VPC_SO_PROG);
+		OUT_RING(ring, prog[i]);
+	}
+}
+
 struct stage {
 	const struct ir3_shader_variant *v;
 	const struct ir3_info *i;
@@ -214,24 +330,17 @@ setup_stages(struct fd5_emit *emit, struct stage *s)
 }
 
 void
-fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
-		int nr, struct pipe_surface **bufs)
+fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit)
 {
 	struct stage s[MAX_STAGES];
-	uint32_t pos_regid, posz_regid, psize_regid, color_regid[8];
+	uint32_t pos_regid, psize_regid, color_regid[8];
 	uint32_t face_regid, coord_regid, zwcoord_regid;
 	uint32_t vcoord_regid, vertex_regid, instance_regid;
 	int i, j;
 
-	debug_assert(nr <= ARRAY_SIZE(color_regid));
-
-	if (emit->key.binning_pass)
-		nr = 0;
-
 	setup_stages(emit, s);
 
 	pos_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_POS);
-	posz_regid = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DEPTH);
 	psize_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_PSIZ);
 	vertex_regid = ir3_find_output_regid(s[VS].v, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE);
 	instance_regid = ir3_find_output_regid(s[VS].v, SYSTEM_VALUE_INSTANCE_ID);
@@ -342,6 +451,10 @@ fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
 	struct ir3_shader_linkage l = {0};
 	ir3_link_shaders(&l, s[VS].v, s[FS].v);
 
+	if ((s[VS].v->shader->stream_output.num_outputs > 0) &&
+			!emit->key.binning_pass)
+		link_stream_out(&l, s[VS].v);
+
 	BITSET_DECLARE(varbs, 128) = {0};
 	uint32_t *varmask = (uint32_t *)varbs;
 
@@ -362,6 +475,17 @@ fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
 	if (psize_regid != regid(63,0))
 		ir3_link_add(&l, psize_regid, 0x1, l.max_loc);
 
+	if ((s[VS].v->shader->stream_output.num_outputs > 0) &&
+			!emit->key.binning_pass) {
+		emit_stream_out(ring, s[VS].v, &l);
+
+		OUT_PKT4(ring, REG_A5XX_VPC_SO_OVERRIDE, 1);
+		OUT_RING(ring, 0x00000000);
+	} else {
+		OUT_PKT4(ring, REG_A5XX_VPC_SO_OVERRIDE, 1);
+		OUT_RING(ring, A5XX_VPC_SO_OVERRIDE_SO_DISABLE);
+	}
+
 	for (i = 0, j = 0; (i < 16) && (j < l.cnt); i++) {
 		uint32_t reg = 0;
 
@@ -406,11 +530,6 @@ fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
 		OUT_RING(ring, 0x00000000);    /* SP_FS_OBJ_START_LO */
 		OUT_RING(ring, 0x00000000);    /* SP_FS_OBJ_START_HI */
 	} else {
-		uint32_t stride_in_vpc = align(s[FS].v->total_in, 4) + 4;
-
-		if (s[VS].v->writes_psize)
-			stride_in_vpc++;
-
 		// TODO if some of these other bits depend on something other than
 		// program state we should probably move these next three regs:
 
@@ -418,12 +537,12 @@ fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
 		OUT_RING(ring, A5XX_SP_PRIMITIVE_CNTL_VSOUT(l.cnt));
 
 		OUT_PKT4(ring, REG_A5XX_VPC_CNTL_0, 1);
-		OUT_RING(ring, A5XX_VPC_CNTL_0_STRIDE_IN_VPC(stride_in_vpc) |
+		OUT_RING(ring, A5XX_VPC_CNTL_0_STRIDE_IN_VPC(l.max_loc) |
 				COND(s[FS].v->total_in > 0, A5XX_VPC_CNTL_0_VARYING) |
 				0x10000);    // XXX
 
 		OUT_PKT4(ring, REG_A5XX_PC_PRIMITIVE_CNTL, 1);
-		OUT_RING(ring, A5XX_PC_PRIMITIVE_CNTL_STRIDE_IN_VPC(stride_in_vpc) |
+		OUT_RING(ring, A5XX_PC_PRIMITIVE_CNTL_STRIDE_IN_VPC(l.max_loc) |
 				0x400);      // XXX
 
 		OUT_PKT4(ring, REG_A5XX_SP_FS_OBJ_START_LO, 2);
@@ -467,25 +586,17 @@ fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
 					A5XX_GRAS_CNTL_UNK3) |
 			COND(s[FS].v->frag_face, A5XX_GRAS_CNTL_UNK3));
 
-	OUT_PKT4(ring, REG_A5XX_RB_RENDER_CONTROL0, 3);
-	OUT_RING(ring,
-			COND(s[FS].v->total_in > 0, A5XX_RB_RENDER_CONTROL0_VARYING) |
+	OUT_PKT4(ring, REG_A5XX_RB_RENDER_CONTROL0, 2);
+	OUT_RING(ring, COND(s[FS].v->total_in > 0, A5XX_RB_RENDER_CONTROL0_VARYING) |
 			COND(s[FS].v->frag_coord, A5XX_RB_RENDER_CONTROL0_XCOORD |
 					A5XX_RB_RENDER_CONTROL0_YCOORD |
 					A5XX_RB_RENDER_CONTROL0_ZCOORD |
 					A5XX_RB_RENDER_CONTROL0_WCOORD |
 					A5XX_RB_RENDER_CONTROL0_UNK3) |
 			COND(s[FS].v->frag_face, A5XX_RB_RENDER_CONTROL0_UNK3));
+	OUT_RING(ring, COND(s[FS].v->frag_face, A5XX_RB_RENDER_CONTROL1_FACENESS));
 
-	OUT_RING(ring,
-			COND(s[FS].v->frag_face, A5XX_RB_RENDER_CONTROL1_FACENESS));
-	OUT_RING(ring, A5XX_RB_FS_OUTPUT_CNTL_MRT(nr) |
-			COND(s[FS].v->writes_pos, A5XX_RB_FS_OUTPUT_CNTL_FRAG_WRITES_Z));
-
-	OUT_PKT4(ring, REG_A5XX_SP_FS_OUTPUT_CNTL, 9);
-	OUT_RING(ring, A5XX_SP_FS_OUTPUT_CNTL_MRT(nr) |
-			A5XX_SP_FS_OUTPUT_CNTL_DEPTH_REGID(posz_regid) |
-			A5XX_SP_FS_OUTPUT_CNTL_SAMPLEMASK_REGID(regid(63, 0)));
+	OUT_PKT4(ring, REG_A5XX_SP_FS_OUTPUT_REG(0), 8);
 	for (i = 0; i < 8; i++) {
 		OUT_RING(ring, A5XX_SP_FS_OUTPUT_REG_REGID(color_regid[i]) |
 				COND(emit->key.half_precision,

src/gallium/drivers/freedreno/a5xx/fd5_program.h

@@ -37,8 +37,7 @@ struct fd5_shader_stateobj {
 
 struct fd5_emit;
 
-void fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
-		int nr, struct pipe_surface **bufs);
+void fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit);
 
 void fd5_prog_init(struct pipe_context *pctx);
 

src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c

@@ -218,7 +218,8 @@ compile_init(struct ir3_compiler *compiler,
 		so->constbase.driver_param = constoff;
 		constoff += align(IR3_DP_COUNT, 4) / 4;
 
-		if (so->shader->stream_output.num_outputs > 0) {
+		if ((compiler->gpu_id < 500) &&
+				so->shader->stream_output.num_outputs > 0) {
 			so->constbase.tfbo = constoff;
 			constoff += align(PIPE_MAX_SO_BUFFERS * ptrsz, 4) / 4;
 		}
@@ -2001,7 +2002,8 @@ emit_function(struct ir3_compile *ctx, nir_function_impl *impl)
 	 * out, we guarantee that all exit paths flow into the stream-
 	 * out instructions.
 	 */
-	if ((ctx->so->shader->stream_output.num_outputs > 0) &&
+	if ((ctx->compiler->gpu_id < 500) &&
+			(ctx->so->shader->stream_output.num_outputs > 0) &&
 			!ctx->so->key.binning_pass) {
 		debug_assert(ctx->so->type == SHADER_VERTEX);
 		emit_stream_out(ctx);

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

@@ -614,6 +614,8 @@ max_tf_vtx(struct fd_context *ctx, const struct ir3_shader_variant *v)
 	struct pipe_stream_output_info *info = &v->shader->stream_output;
 	uint32_t maxvtxcnt = 0x7fffffff;
 
+	if (ctx->screen->gpu_id >= 500)
+		return 0;
 	if (v->key.binning_pass)
 		return 0;
 	if (v->shader->stream_output.num_outputs == 0)