r600g: Implement GL_ARB_draw_indirect for EG/CM

Requires Evergreen/Cayman and radeon kernel module
2.41.0 or newer.

Expected piglit fails due to hardware limitations:
* arb_draw_indirect-draw-arrays-prim-restart
  Restarts not applied for DrawArrays commands
* arb_draw_indirect-vertexid
  Base vertex offset is not included in vertex id

Marek: bump vgt_state num_dw by 3 (= space needed for one register write)

Signed-off-by: Glenn Kennard <[email protected]>
Signed-off-by: Marek Olšák <[email protected]>

12 files changed, 133 insertions, 38 deletions

diff --git a/docs/GL3.txt b/docs/GL3.txt
index a5b81754564..43bbf854e9d 100644
--- a/docs/GL3.txt
+++ b/docs/GL3.txt
@@ -95,7 +95,7 @@ GL 3.3, GLSL 3.30 --- all DONE: i965, nv50, nvc0, r600, radeonsi, llvmpipe, soft
 GL 4.0, GLSL 4.00:
 
   GL_ARB_draw_buffers_blend                            DONE (i965, nv50, nvc0, r600, radeonsi, llvmpipe, softpipe)
-  GL_ARB_draw_indirect                                 DONE (i965, nvc0, radeonsi, llvmpipe, softpipe)
+  GL_ARB_draw_indirect                                 DONE (i965, nvc0, r600, radeonsi, llvmpipe, softpipe)
   GL_ARB_gpu_shader5                                   DONE (i965, nvc0)
   - 'precise' qualifier                                DONE
   - Dynamically uniform sampler array indices          DONE (r600)
@@ -159,7 +159,7 @@ GL 4.3, GLSL 4.30:
   GL_ARB_framebuffer_no_attachments                    not started
   GL_ARB_internalformat_query2                         not started
   GL_ARB_invalidate_subdata                            DONE (all drivers)
-  GL_ARB_multi_draw_indirect                           DONE (i965, nvc0, radeonsi, llvmpipe, softpipe)
+  GL_ARB_multi_draw_indirect                           DONE (i965, nvc0, r600, radeonsi, llvmpipe, softpipe)
   GL_ARB_program_interface_query                       not started
   GL_ARB_robust_buffer_access_behavior                 not started
   GL_ARB_shader_image_size                             not started
@@ -212,7 +212,7 @@ These are the extensions cherry-picked to make GLES 3.1
 GLES3.1, GLSL ES 3.1
   GL_ARB_arrays_of_arrays                              started (Timothy)
   GL_ARB_compute_shader                                in progress (jljusten)
-  GL_ARB_draw_indirect                                 DONE (i965, nvc0, radeonsi, llvmpipe, softpipe)
+  GL_ARB_draw_indirect                                 DONE (i965, nvc0, r600, radeonsi, llvmpipe, softpipe)
   GL_ARB_explicit_uniform_location                     DONE (all drivers that support GLSL)
   GL_ARB_framebuffer_no_attachments                    not started
   GL_ARB_program_interface_query                       not started
diff --git a/docs/relnotes/10.6.0.html b/docs/relnotes/10.6.0.html
index 056d3b00def..ceeca07ff38 100644
--- a/docs/relnotes/10.6.0.html
+++ b/docs/relnotes/10.6.0.html
@@ -49,6 +49,7 @@ Note: some of the new features are only available with certain drivers.
 <li>GL_ARB_gpu_shader_fp64 on nvc0, softpipe</li>
 <li>GL_ARB_instanced_arrays on freedreno</li>
 <li>GL_ARB_pipeline_statistics_query on i965, nv50, nvc0, r600, radeonsi, softpipe</li>
+<li>GL_ARB_draw_indirect, GL_ARB_multi_draw_indirect on r600</li>
 </ul>
 
 <h2>Bug fixes</h2>
diff --git a/src/gallium/drivers/r600/evergreen_state.c b/src/gallium/drivers/r600/evergreen_state.c
index ea58aeab0b4..bc634fe5e36 100644
--- a/src/gallium/drivers/r600/evergreen_state.c
+++ b/src/gallium/drivers/r600/evergreen_state.c
@@ -3438,7 +3438,7 @@ void evergreen_init_state_functions(struct r600_context *rctx)
 	r600_init_atom(rctx, &rctx->samplers[PIPE_SHADER_GEOMETRY].views.atom, id++, evergreen_emit_gs_sampler_views, 0);
 	r600_init_atom(rctx, &rctx->samplers[PIPE_SHADER_FRAGMENT].views.atom, id++, evergreen_emit_ps_sampler_views, 0);
 
-	r600_init_atom(rctx, &rctx->vgt_state.atom, id++, r600_emit_vgt_state, 7);
+	r600_init_atom(rctx, &rctx->vgt_state.atom, id++, r600_emit_vgt_state, 10);
 
 	if (rctx->b.chip_class == EVERGREEN) {
 		r600_init_atom(rctx, &rctx->sample_mask.atom, id++, evergreen_emit_sample_mask, 3);
diff --git a/src/gallium/drivers/r600/evergreend.h b/src/gallium/drivers/r600/evergreend.h
index 49899960367..cd4ff46b103 100644
--- a/src/gallium/drivers/r600/evergreend.h
+++ b/src/gallium/drivers/r600/evergreend.h
@@ -72,7 +72,6 @@
 #define PKT3_REG_RMW                           0x21
 #define PKT3_COND_EXEC                         0x22
 #define PKT3_PRED_EXEC                         0x23
-#define PKT3_START_3D_CMDBUF                   0x24
 #define PKT3_DRAW_INDEX_2                      0x27
 #define PKT3_CONTEXT_CONTROL                   0x28
 #define PKT3_DRAW_INDEX_IMMD_BE                0x29
diff --git a/src/gallium/drivers/r600/r600_asm.c b/src/gallium/drivers/r600/r600_asm.c
index 79e7f748cdf..762cc7fac44 100644
--- a/src/gallium/drivers/r600/r600_asm.c
+++ b/src/gallium/drivers/r600/r600_asm.c
@@ -2411,7 +2411,7 @@ void *r600_create_vertex_fetch_shader(struct pipe_context *ctx,
 
 		memset(&vtx, 0, sizeof(vtx));
 		vtx.buffer_id = elements[i].vertex_buffer_index + fetch_resource_start;
-		vtx.fetch_type = elements[i].instance_divisor ? 1 : 0;
+		vtx.fetch_type = elements[i].instance_divisor ? SQ_VTX_FETCH_INSTANCE_DATA : SQ_VTX_FETCH_VERTEX_DATA;
 		vtx.src_gpr = elements[i].instance_divisor > 1 ? i + 1 : 0;
 		vtx.src_sel_x = elements[i].instance_divisor ? 3 : 0;
 		vtx.mega_fetch_count = 0x1F;
diff --git a/src/gallium/drivers/r600/r600_pipe.c b/src/gallium/drivers/r600/r600_pipe.c
index 06dca25924a..c8a0e9cb624 100644
--- a/src/gallium/drivers/r600/r600_pipe.c
+++ b/src/gallium/drivers/r600/r600_pipe.c
@@ -317,6 +317,9 @@ static int r600_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
 		return family >= CHIP_CEDAR ? 1 : 0;
 	case PIPE_CAP_MAX_TEXTURE_GATHER_COMPONENTS:
 		return family >= CHIP_CEDAR ? 4 : 0;
+	case PIPE_CAP_DRAW_INDIRECT:
+		/* kernel command checker support is also required */
+		return family >= CHIP_CEDAR && rscreen->b.info.drm_minor >= 41;
 
 	/* Unsupported features. */
 	case PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT:
@@ -326,7 +329,6 @@ static int r600_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
 	case PIPE_CAP_VERTEX_COLOR_CLAMPED:
 	case PIPE_CAP_USER_VERTEX_BUFFERS:
 	case PIPE_CAP_TEXTURE_GATHER_OFFSETS:
-	case PIPE_CAP_DRAW_INDIRECT:
 	case PIPE_CAP_CONDITIONAL_RENDER_INVERTED:
 	case PIPE_CAP_SAMPLER_VIEW_TARGET:
 	case PIPE_CAP_VERTEXID_NOBASE:
diff --git a/src/gallium/drivers/r600/r600_pipe.h b/src/gallium/drivers/r600/r600_pipe.h
index e110efe7aea..72378548733 100644
--- a/src/gallium/drivers/r600/r600_pipe.h
+++ b/src/gallium/drivers/r600/r600_pipe.h
@@ -55,7 +55,7 @@
 
 /* the number of CS dwords for flushing and drawing */
 #define R600_MAX_FLUSH_CS_DWORDS	16
-#define R600_MAX_DRAW_CS_DWORDS		40
+#define R600_MAX_DRAW_CS_DWORDS		47
 #define R600_TRACE_CS_DWORDS		7
 
 #define R600_MAX_USER_CONST_BUFFERS 13
@@ -145,6 +145,7 @@ struct r600_vgt_state {
 	uint32_t vgt_multi_prim_ib_reset_en;
 	uint32_t vgt_multi_prim_ib_reset_indx;
 	uint32_t vgt_indx_offset;
+	bool last_draw_was_indirect;
 };
 
 struct r600_blend_color {
diff --git a/src/gallium/drivers/r600/r600_shader.c b/src/gallium/drivers/r600/r600_shader.c
index 77c9909f5d8..2ee59c8bac0 100644
--- a/src/gallium/drivers/r600/r600_shader.c
+++ b/src/gallium/drivers/r600/r600_shader.c
@@ -938,7 +938,7 @@ static int load_sample_position(struct r600_shader_ctx *ctx, struct r600_shader_
 	memset(&vtx, 0, sizeof(struct r600_bytecode_vtx));
 	vtx.op = FETCH_OP_VFETCH;
 	vtx.buffer_id = R600_SAMPLE_POSITIONS_CONST_BUFFER;
-	vtx.fetch_type = 2;	/* VTX_FETCH_NO_INDEX_OFFSET */
+	vtx.fetch_type = SQ_VTX_FETCH_NO_INDEX_OFFSET;
 	if (sample_id == NULL) {
 		vtx.src_gpr = ctx->fixed_pt_position_gpr; // SAMPLEID is in .w;
 		vtx.src_sel_x = 3;
@@ -1095,7 +1095,7 @@ static int tgsi_fetch_rel_const(struct r600_shader_ctx *ctx,
 
 	memset(&vtx, 0, sizeof(vtx));
 	vtx.buffer_id = cb_idx;
-	vtx.fetch_type = 2;		/* VTX_FETCH_NO_INDEX_OFFSET */
+	vtx.fetch_type = SQ_VTX_FETCH_NO_INDEX_OFFSET;
 	vtx.src_gpr = ar_reg;
 	vtx.src_sel_x = ar_chan;
 	vtx.mega_fetch_count = 16;
@@ -1173,7 +1173,7 @@ static int fetch_gs_input(struct r600_shader_ctx *ctx, struct tgsi_full_src_regi
 
 	memset(&vtx, 0, sizeof(vtx));
 	vtx.buffer_id = R600_GS_RING_CONST_BUFFER;
-	vtx.fetch_type = 2;		/* VTX_FETCH_NO_INDEX_OFFSET */
+	vtx.fetch_type = SQ_VTX_FETCH_NO_INDEX_OFFSET;
 	vtx.src_gpr = offset_reg;
 	vtx.src_sel_x = offset_chan;
 	vtx.offset = index * 16; /*bytes*/
@@ -1539,7 +1539,7 @@ static int generate_gs_copy_shader(struct r600_context *rctx,
 		memset(&vtx, 0, sizeof(vtx));
 		vtx.op = FETCH_OP_VFETCH;
 		vtx.buffer_id = R600_GS_RING_CONST_BUFFER;
-		vtx.fetch_type = 2;
+		vtx.fetch_type = SQ_VTX_FETCH_NO_INDEX_OFFSET;
 		vtx.offset = out->ring_offset;
 		vtx.dst_gpr = out->gpr;
 		vtx.dst_sel_x = 0;
@@ -5025,7 +5025,7 @@ static int do_vtx_fetch_inst(struct r600_shader_ctx *ctx, boolean src_requires_l
 	memset(&vtx, 0, sizeof(vtx));
 	vtx.op = FETCH_OP_VFETCH;
 	vtx.buffer_id = id + R600_MAX_CONST_BUFFERS;
-	vtx.fetch_type = 2;		/* VTX_FETCH_NO_INDEX_OFFSET */
+	vtx.fetch_type = SQ_VTX_FETCH_NO_INDEX_OFFSET;
 	vtx.src_gpr = src_gpr;
 	vtx.mega_fetch_count = 16;
 	vtx.dst_gpr = ctx->file_offset[inst->Dst[0].Register.File] + inst->Dst[0].Register.Index;
diff --git a/src/gallium/drivers/r600/r600_sq.h b/src/gallium/drivers/r600/r600_sq.h
index 587f88deb9e..1545cf1b521 100644
--- a/src/gallium/drivers/r600/r600_sq.h
+++ b/src/gallium/drivers/r600/r600_sq.h
@@ -481,4 +481,8 @@
 #define   INDEX_MODE_AR_W 3
 #define   INDEX_MODE_LOOP 4
 
+#define SQ_VTX_FETCH_VERTEX_DATA 0
+#define SQ_VTX_FETCH_INSTANCE_DATA 1
+#define SQ_VTX_FETCH_NO_INDEX_OFFSET 2
+
 #endif
diff --git a/src/gallium/drivers/r600/r600_state.c b/src/gallium/drivers/r600/r600_state.c
index 3c2fdfaed42..9d35c171708 100644
--- a/src/gallium/drivers/r600/r600_state.c
+++ b/src/gallium/drivers/r600/r600_state.c
@@ -3048,7 +3048,7 @@ void r600_init_state_functions(struct r600_context *rctx)
 	r600_init_atom(rctx, &rctx->samplers[PIPE_SHADER_FRAGMENT].views.atom, id++, r600_emit_ps_sampler_views, 0);
 	r600_init_atom(rctx, &rctx->vertex_buffer_state.atom, id++, r600_emit_vertex_buffers, 0);
 
-	r600_init_atom(rctx, &rctx->vgt_state.atom, id++, r600_emit_vgt_state, 7);
+	r600_init_atom(rctx, &rctx->vgt_state.atom, id++, r600_emit_vgt_state, 10);
 
 	r600_init_atom(rctx, &rctx->seamless_cube_map.atom, id++, r600_emit_seamless_cube_map, 3);
 	r600_init_atom(rctx, &rctx->sample_mask.atom, id++, r600_emit_sample_mask, 3);
diff --git a/src/gallium/drivers/r600/r600_state_common.c b/src/gallium/drivers/r600/r600_state_common.c
index b498d00b5b5..a08124bed16 100644
--- a/src/gallium/drivers/r600/r600_state_common.c
+++ b/src/gallium/drivers/r600/r600_state_common.c
@@ -196,6 +196,10 @@ void r600_emit_vgt_state(struct r600_context *rctx, struct r600_atom *atom)
 	r600_write_context_reg_seq(cs, R_028408_VGT_INDX_OFFSET, 2);
 	radeon_emit(cs, a->vgt_indx_offset); /* R_028408_VGT_INDX_OFFSET */
 	radeon_emit(cs, a->vgt_multi_prim_ib_reset_indx); /* R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX */
+	if (a->last_draw_was_indirect) {
+		a->last_draw_was_indirect = false;
+		r600_write_ctl_const(cs, R_03CFF0_SQ_VTX_BASE_VTX_LOC, 0);
+	}
 }
 
 static void r600_set_clip_state(struct pipe_context *ctx,
@@ -1353,7 +1357,7 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 	unsigned i;
 	struct radeon_winsys_cs *cs = rctx->b.rings.gfx.cs;
 
-	if (!info.count && (info.indexed || !info.count_from_stream_output)) {
+	if (!info.indirect && !info.count && (info.indexed || !info.count_from_stream_output)) {
 		return;
 	}
 
@@ -1379,19 +1383,44 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 		pipe_resource_reference(&ib.buffer, rctx->index_buffer.buffer);
 		ib.user_buffer = rctx->index_buffer.user_buffer;
 		ib.index_size = rctx->index_buffer.index_size;
-		ib.offset = rctx->index_buffer.offset + info.start * ib.index_size;
+		ib.offset = rctx->index_buffer.offset;
+		if (!info.indirect) {
+			ib.offset += info.start * ib.index_size;
+		}
 
 		/* Translate 8-bit indices to 16-bit. */
-		if (ib.index_size == 1) {
+		if (unlikely(ib.index_size == 1)) {
 			struct pipe_resource *out_buffer = NULL;
 			unsigned out_offset;
 			void *ptr;
+			unsigned start, count;
+
+			if (likely(!info.indirect)) {
+				start = 0;
+				count = info.count;
+			}
+			else {
+				/* Have to get start/count from indirect buffer, slow path ahead... */
+				struct r600_resource *indirect_resource = (struct r600_resource *)info.indirect;
+				unsigned *data = r600_buffer_map_sync_with_rings(&rctx->b, indirect_resource,
+					PIPE_TRANSFER_READ);
+				if (data) {
+					data += info.indirect_offset / sizeof(unsigned);
+					start = data[2] * ib.index_size;
+					count = data[0];
+					rctx->b.ws->buffer_unmap(indirect_resource->cs_buf);
+				}
+				else {
+					start = 0;
+					count = 0;
+				}
+			}
 
-			u_upload_alloc(rctx->b.uploader, 0, info.count * 2,
+			u_upload_alloc(rctx->b.uploader, start, count * 2,
 				       &out_offset, &out_buffer, &ptr);
 
 			util_shorten_ubyte_elts_to_userptr(
-						&rctx->b.b, &ib, 0, ib.offset, info.count, ptr);
+						&rctx->b.b, &ib, 0, ib.offset + start, count, ptr);
 
 			pipe_resource_reference(&ib.buffer, NULL);
 			ib.user_buffer = NULL;
@@ -1403,9 +1432,11 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 		/* Upload the index buffer.
 		 * The upload is skipped for small index counts on little-endian machines
 		 * and the indices are emitted via PKT3_DRAW_INDEX_IMMD.
+		 * Indirect draws never use immediate indices.
 		 * Note: Instanced rendering in combination with immediate indices hangs. */
-		if (ib.user_buffer && (R600_BIG_ENDIAN || info.instance_count > 1 ||
-				       info.count*ib.index_size > 20)) {
+		if (ib.user_buffer && (R600_BIG_ENDIAN || info.indirect ||
+						 info.instance_count > 1 ||
+						 info.count*ib.index_size > 20)) {
 			u_upload_data(rctx->b.uploader, 0, info.count * ib.index_size,
 				      ib.user_buffer, &ib.offset, &ib.buffer);
 			ib.user_buffer = NULL;
@@ -1417,7 +1448,8 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 	/* Set the index offset and primitive restart. */
 	if (rctx->vgt_state.vgt_multi_prim_ib_reset_en != info.primitive_restart ||
 	    rctx->vgt_state.vgt_multi_prim_ib_reset_indx != info.restart_index ||
-	    rctx->vgt_state.vgt_indx_offset != info.index_bias) {
+	    rctx->vgt_state.vgt_indx_offset != info.index_bias ||
+	    (rctx->vgt_state.last_draw_was_indirect && !info.indirect)) {
 		rctx->vgt_state.vgt_multi_prim_ib_reset_en = info.primitive_restart;
 		rctx->vgt_state.vgt_multi_prim_ib_reset_indx = info.restart_index;
 		rctx->vgt_state.vgt_indx_offset = info.index_bias;
@@ -1485,7 +1517,7 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 	}
 
 	/* Update start instance. */
-	if (rctx->last_start_instance != info.start_instance) {
+	if (!info.indirect && rctx->last_start_instance != info.start_instance) {
 		r600_write_ctl_const(cs, R_03CFF4_SQ_VTX_START_INST_LOC, info.start_instance);
 		rctx->last_start_instance = info.start_instance;
 	}
@@ -1510,8 +1542,30 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 	}
 
 	/* Draw packets. */
-	cs->buf[cs->cdw++] = PKT3(PKT3_NUM_INSTANCES, 0, rctx->b.predicate_drawing);
-	cs->buf[cs->cdw++] = info.instance_count;
+	if (!info.indirect) {
+		cs->buf[cs->cdw++] = PKT3(PKT3_NUM_INSTANCES, 0, rctx->b.predicate_drawing);
+		cs->buf[cs->cdw++] = info.instance_count;
+	}
+
+	if (unlikely(info.indirect)) {
+		uint64_t va = r600_resource(info.indirect)->gpu_address;
+		assert(rctx->b.chip_class >= EVERGREEN);
+
+		// Invalidate so non-indirect draw calls reset this state
+		rctx->vgt_state.last_draw_was_indirect = true;
+		rctx->last_start_instance = -1;
+
+		cs->buf[cs->cdw++] = PKT3(EG_PKT3_SET_BASE, 2, rctx->b.predicate_drawing);
+		cs->buf[cs->cdw++] = EG_DRAW_INDEX_INDIRECT_PATCH_TABLE_BASE;
+		cs->buf[cs->cdw++] = va;
+		cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
+
+		cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, rctx->b.predicate_drawing);
+		cs->buf[cs->cdw++] = r600_context_bo_reloc(&rctx->b, &rctx->b.rings.gfx,
+							   (struct r600_resource*)info.indirect,
+							   RADEON_USAGE_READ, RADEON_PRIO_MIN);
+	}
+
 	if (info.indexed) {
 		cs->buf[cs->cdw++] = PKT3(PKT3_INDEX_TYPE, 0, rctx->b.predicate_drawing);
 		cs->buf[cs->cdw++] = ib.index_size == 4 ?
@@ -1528,18 +1582,40 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 			cs->cdw += size_dw;
 		} else {
 			uint64_t va = r600_resource(ib.buffer)->gpu_address + ib.offset;
-			cs->buf[cs->cdw++] = PKT3(PKT3_DRAW_INDEX, 3, rctx->b.predicate_drawing);
-			cs->buf[cs->cdw++] = va;
-			cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
-			cs->buf[cs->cdw++] = info.count;
-			cs->buf[cs->cdw++] = V_0287F0_DI_SRC_SEL_DMA;
-			cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, rctx->b.predicate_drawing);
-			cs->buf[cs->cdw++] = r600_context_bo_reloc(&rctx->b, &rctx->b.rings.gfx,
-								   (struct r600_resource*)ib.buffer,
-								   RADEON_USAGE_READ, RADEON_PRIO_MIN);
+
+			if (likely(!info.indirect)) {
+				cs->buf[cs->cdw++] = PKT3(PKT3_DRAW_INDEX, 3, rctx->b.predicate_drawing);
+				cs->buf[cs->cdw++] = va;
+				cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
+				cs->buf[cs->cdw++] = info.count;
+				cs->buf[cs->cdw++] = V_0287F0_DI_SRC_SEL_DMA;
+				cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, rctx->b.predicate_drawing);
+				cs->buf[cs->cdw++] = r600_context_bo_reloc(&rctx->b, &rctx->b.rings.gfx,
+									   (struct r600_resource*)ib.buffer,
+									   RADEON_USAGE_READ, RADEON_PRIO_MIN);
+			}
+			else {
+				uint32_t max_size = (ib.buffer->width0 - ib.offset) / ib.index_size;
+
+				cs->buf[cs->cdw++] = PKT3(EG_PKT3_INDEX_BASE, 1, rctx->b.predicate_drawing);
+				cs->buf[cs->cdw++] = va;
+				cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
+
+				cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, rctx->b.predicate_drawing);
+				cs->buf[cs->cdw++] = r600_context_bo_reloc(&rctx->b, &rctx->b.rings.gfx,
+									   (struct r600_resource*)ib.buffer,
+									   RADEON_USAGE_READ, RADEON_PRIO_MIN);
+
+				cs->buf[cs->cdw++] = PKT3(EG_PKT3_INDEX_BUFFER_SIZE, 0, rctx->b.predicate_drawing);
+				cs->buf[cs->cdw++] = max_size;
+
+				cs->buf[cs->cdw++] = PKT3(EG_PKT3_DRAW_INDEX_INDIRECT, 1, rctx->b.predicate_drawing);
+				cs->buf[cs->cdw++] = info.indirect_offset;
+				cs->buf[cs->cdw++] = V_0287F0_DI_SRC_SEL_DMA;
+			}
 		}
 	} else {
-		if (info.count_from_stream_output) {
+		if (unlikely(info.count_from_stream_output)) {
 			struct r600_so_target *t = (struct r600_so_target*)info.count_from_stream_output;
 			uint64_t va = t->buf_filled_size->gpu_address + t->buf_filled_size_offset;
 
@@ -1558,8 +1634,14 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 								   RADEON_PRIO_MIN);
 		}
 
-		cs->buf[cs->cdw++] = PKT3(PKT3_DRAW_INDEX_AUTO, 1, rctx->b.predicate_drawing);
-		cs->buf[cs->cdw++] = info.count;
+		if (likely(!info.indirect)) {
+			cs->buf[cs->cdw++] = PKT3(PKT3_DRAW_INDEX_AUTO, 1, rctx->b.predicate_drawing);
+			cs->buf[cs->cdw++] = info.count;
+		}
+		else {
+			cs->buf[cs->cdw++] = PKT3(EG_PKT3_DRAW_INDIRECT, 1, rctx->b.predicate_drawing);
+			cs->buf[cs->cdw++] = info.indirect_offset;
+		}
 		cs->buf[cs->cdw++] = V_0287F0_DI_SRC_SEL_AUTO_INDEX |
 					(info.count_from_stream_output ? S_0287F0_USE_OPAQUE(1) : 0);
 	}
diff --git a/src/gallium/drivers/r600/r600d.h b/src/gallium/drivers/r600/r600d.h
index 6a5b9640ee0..bce8b4ea065 100644
--- a/src/gallium/drivers/r600/r600d.h
+++ b/src/gallium/drivers/r600/r600d.h
@@ -52,12 +52,18 @@
 
 
 #define PKT3_NOP                               0x10
+#define EG_PKT3_SET_BASE                       0x11 /* >= evergreen */
+#define     EG_DRAW_INDEX_INDIRECT_PATCH_TABLE_BASE 1 /* DX11 Draw_Index_Indirect Patch Table Base */
+#define EG_PKT3_INDEX_BUFFER_SIZE              0x13
 #define PKT3_INDIRECT_BUFFER_END               0x17
 #define PKT3_SET_PREDICATION                   0x20
 #define PKT3_REG_RMW                           0x21
 #define PKT3_COND_EXEC                         0x22
 #define PKT3_PRED_EXEC                         0x23
-#define PKT3_START_3D_CMDBUF                   0x24
+#define PKT3_START_3D_CMDBUF                   0x24 /* removed on evergreen */
+#define EG_PKT3_DRAW_INDIRECT                  0x24 /* >= evergreen */
+#define EG_PKT3_DRAW_INDEX_INDIRECT            0x25
+#define EG_PKT3_INDEX_BASE                     0x26
 #define PKT3_DRAW_INDEX_2                      0x27
 #define PKT3_CONTEXT_CONTROL                   0x28
 #define PKT3_DRAW_INDEX_IMMD_BE                0x29