radeonsi: implement glDrawTransformFeedback functionality

Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

src/gallium/drivers/radeonsi/si_state.c

@@ -3241,6 +3241,7 @@ void si_init_config(struct r600_context *rctx)
 	si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE, 0x0);
 	si_pm4_set_reg(pm4, R_028A84_VGT_PRIMITIVEID_EN, 0x0);
 	si_pm4_set_reg(pm4, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
+	si_pm4_set_reg(pm4, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET, 0);
 	si_pm4_set_reg(pm4, R_028B94_VGT_STRMOUT_CONFIG, 0x0);
 	si_pm4_set_reg(pm4, R_028B98_VGT_STRMOUT_BUFFER_CONFIG, 0x0);
 	si_pm4_set_reg(pm4, R_028AA8_IA_MULTI_VGT_PARAM,

src/gallium/drivers/radeonsi/si_state_draw.c

@@ -557,6 +557,29 @@ static void si_state_draw(struct r600_context *rctx,
 			       S_02800C_NOOP_CULL_DISABLE(1));
 	}
 
+	if (info->count_from_stream_output) {
+		struct r600_so_target *t =
+			(struct r600_so_target*)info->count_from_stream_output;
+		uint64_t va = r600_resource_va(&rctx->screen->b.b,
+					       &t->buf_filled_size->b.b);
+		va += t->buf_filled_size_offset;
+
+		si_pm4_set_reg(pm4, R_028B30_VGT_STRMOUT_DRAW_OPAQUE_VERTEX_STRIDE,
+			       t->stride_in_dw);
+
+		si_pm4_cmd_begin(pm4, PKT3_COPY_DATA);
+		si_pm4_cmd_add(pm4,
+			       COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
+			       COPY_DATA_DST_SEL(COPY_DATA_REG) |
+			       COPY_DATA_WR_CONFIRM);
+		si_pm4_cmd_add(pm4, va);     /* src address lo */
+		si_pm4_cmd_add(pm4, va >> 32UL); /* src address hi */
+		si_pm4_cmd_add(pm4, R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE >> 2);
+		si_pm4_cmd_add(pm4, 0); /* unused */
+		si_pm4_add_bo(pm4, t->buf_filled_size, RADEON_USAGE_READ);
+		si_pm4_cmd_end(pm4, true);
+	}
+
 	/* draw packet */
 	si_pm4_cmd_begin(pm4, PKT3_INDEX_TYPE);
 	if (ib->index_size == 4) {

src/gallium/drivers/radeonsi/sid.h

@@ -103,6 +103,12 @@
 #define		WAIT_REG_MEM_EQUAL		3
 #define PKT3_MEM_WRITE                         0x3D /* not on CIK */
 #define PKT3_INDIRECT_BUFFER                   0x32
+#define PKT3_COPY_DATA			       0x40
+#define		COPY_DATA_SRC_SEL(x)		((x) & 0xf)
+#define			COPY_DATA_REG		0
+#define			COPY_DATA_MEM		1
+#define		COPY_DATA_DST_SEL(x)		(((x) & 0xf) << 8)
+#define		COPY_DATA_WR_CONFIRM		(1 << 20)
 #define PKT3_SURFACE_SYNC                      0x43 /* deprecated on CIK, use ACQUIRE_MEM */
 #define PKT3_ME_INITIALIZE                     0x44 /* not on CIK */
 #define PKT3_COND_WRITE                        0x45