r600g: simplify states

Directly build PM4 packet, avoid using malloc (no states are
bigger than 128 dwords), remove unecessary informations,
remove pm4 building in favor of prebuild pm4 packet.

Signed-off-by: Jerome Glisse <[email protected]>

8 files changed, 398 insertions, 496 deletions

diff --git a/src/gallium/drivers/r600/r600_blit.c b/src/gallium/drivers/r600/r600_blit.c
index 72175fbbd5e..d3b722c82f7 100644
--- a/src/gallium/drivers/r600/r600_blit.c
+++ b/src/gallium/drivers/r600/r600_blit.c
@@ -132,7 +132,7 @@ static void r600_resource_copy_region(struct pipe_context *ctx,
 				      unsigned srcx, unsigned srcy, unsigned srcz,
 				      unsigned width, unsigned height)
 {
-	util_resource_copy_region(pipe, dst, subdst, dstx, dsty, dstz,
+	util_resource_copy_region(ctx, dst, subdst, dstx, dsty, dstz,
 				  src, subsrc, srcx, srcy, srcz, width, height);
 }
 
@@ -190,7 +190,7 @@ static int r600_blit_state_vs_resources(struct r600_screen *rscreen, struct r600
 	memcpy(bo->data, vbo, 128);
 	radeon_bo_unmap(rscreen->rw, bo);
 
-	rstate = radeon_state(rscreen->rw, R600_VS_RESOURCE_TYPE, R600_VS_RESOURCE + 0);
+	rstate = radeon_state(rscreen->rw, R600_VS_RESOURCE0 + 0);
 	if (rstate == NULL) {
 		radeon_bo_decref(rscreen->rw, bo);
 		return -ENOMEM;
@@ -199,33 +199,35 @@ static int r600_blit_state_vs_resources(struct r600_screen *rscreen, struct r600
 	/* set states (most default value are 0 and struct already
 	 * initialized to 0, thus avoid resetting them)
 	 */
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD0] = 0x00000000;
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD1] = 0x00000080;
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD2] = 0x02302000;
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD3] = 0x00000000;
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD4] = 0x00000000;
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD5] = 0x00000000;
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD6] = 0xC0000000;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD0] = 0x00000000;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD1] = 0x00000080;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD2] = 0x02302000;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD3] = 0x00000000;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD4] = 0x00000000;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD5] = 0x00000000;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD6] = 0xC0000000;
 	rstate->bo[0] = bo;
 	rstate->nbo = 1;
 	rstate->placement[0] = RADEON_GEM_DOMAIN_GTT;
+	rstate->reloc_pm4_id[0] = R600_RESOURCE__RESOURCE_BO0_ID;
+	rstate->reloc_pm4_id[1] = R600_RESOURCE__RESOURCE_BO1_ID;
 	if (radeon_state_pm4(rstate)) {
 		radeon_state_decref(rstate);
 		return -ENOMEM;
 	}
 	bstates->vs_resource0 = rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_VS_RESOURCE_TYPE, R600_VS_RESOURCE + 1);
+	rstate = radeon_state(rscreen->rw, R600_VS_RESOURCE0 + 1);
 	if (rstate == NULL) {
 		return -ENOMEM;
 	}
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD0] = 0x00000010;
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD1] = 0x00000070;
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD2] = 0x02302000;
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD3] = 0x00000000;
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD4] = 0x00000000;
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD5] = 0x00000000;
-	rstate->states[R600_VS_RESOURCE__RESOURCE160_WORD6] = 0xC0000000;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD0] = 0x00000010;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD1] = 0x00000070;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD2] = 0x02302000;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD3] = 0x00000000;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD4] = 0x00000000;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD5] = 0x00000000;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD6] = 0xC0000000;
 	rstate->bo[0] = radeon_bo_incref(rscreen->rw, bo);
 	rstate->nbo = 1;
 	rstate->placement[0] = RADEON_GEM_DOMAIN_GTT;
@@ -303,7 +305,7 @@ static struct radeon_state *r600_blit_state_vs_shader(struct r600_screen *rscree
 	}
 	radeon_bo_unmap(rscreen->rw, bo);
 
-	rstate = radeon_state(rscreen->rw, R600_VS_SHADER_TYPE, R600_VS_SHADER);
+	rstate = radeon_state(rscreen->rw, R600_VS_SHADER);
 	if (rstate == NULL) {
 		radeon_bo_decref(rscreen->rw, bo);
 		return NULL;
@@ -321,6 +323,8 @@ static struct radeon_state *r600_blit_state_vs_shader(struct r600_screen *rscree
 	rstate->nbo = 2;
 	rstate->placement[0] = RADEON_GEM_DOMAIN_GTT;
 	rstate->placement[2] = RADEON_GEM_DOMAIN_GTT;
+	rstate->reloc_pm4_id[0] = R600_VS_SHADER__SQ_PGM_START_VS_BO_ID;
+	rstate->reloc_pm4_id[1] = R600_VS_SHADER__SQ_PGM_START_FS_BO_ID;
 
 	if (radeon_state_pm4(rstate)) {
 		radeon_state_decref(rstate);
@@ -374,7 +378,7 @@ static struct radeon_state *r600_blit_state_ps_shader(struct r600_screen *rscree
 	}
 	radeon_bo_unmap(rscreen->rw, bo);
 
-	rstate = radeon_state(rscreen->rw, R600_PS_SHADER_TYPE, R600_PS_SHADER);
+	rstate = radeon_state(rscreen->rw, R600_PS_SHADER);
 	if (rstate == NULL) {
 		radeon_bo_decref(rscreen->rw, bo);
 		return NULL;
@@ -391,6 +395,7 @@ static struct radeon_state *r600_blit_state_ps_shader(struct r600_screen *rscree
 	rstate->bo[0] = bo;
 	rstate->nbo = 1;
 	rstate->placement[0] = RADEON_GEM_DOMAIN_GTT;
+	rstate->reloc_pm4_id[0] = R600_PS_SHADER__SQ_PGM_START_PS_BO_ID;
 
 	if (radeon_state_pm4(rstate)) {
 		radeon_state_decref(rstate);
@@ -403,7 +408,7 @@ static struct radeon_state *r600_blit_state_vgt(struct r600_screen *rscreen)
 {
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_VGT_TYPE, R600_VGT);
+	rstate = radeon_state(rscreen->rw, R600_VGT);
 	if (rstate == NULL)
 		return NULL;
 
@@ -425,7 +430,7 @@ static struct radeon_state *r600_blit_state_draw(struct r600_screen *rscreen)
 {
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_DRAW_TYPE, R600_DRAW);
+	rstate = radeon_state(rscreen->rw, R600_DRAW);
 	if (rstate == NULL)
 		return NULL;
 
@@ -448,7 +453,7 @@ static struct radeon_state *r600_blit_state_vs_constant(struct r600_screen *rscr
 {
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_VS_CONSTANT_TYPE, R600_VS_CONSTANT + id);
+	rstate = radeon_state(rscreen->rw, R600_VS_CONSTANT0 + id);
 	if (rstate == NULL)
 		return NULL;
 
@@ -471,7 +476,7 @@ static struct radeon_state *r600_blit_state_rasterizer(struct r600_screen *rscre
 {
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_RASTERIZER_TYPE, R600_RASTERIZER);
+	rstate = radeon_state(rscreen->rw, R600_RASTERIZER);
 	if (rstate == NULL)
 		return NULL;
 
@@ -500,7 +505,7 @@ static struct radeon_state *r600_blit_state_dsa(struct r600_screen *rscreen)
 {
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_DSA_TYPE, R600_DSA);
+	rstate = radeon_state(rscreen->rw, R600_DSA);
 	if (rstate == NULL)
 		return NULL;
 
@@ -524,7 +529,7 @@ static struct radeon_state *r600_blit_state_blend(struct r600_screen *rscreen)
 {
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_BLEND_TYPE, R600_BLEND);
+	rstate = radeon_state(rscreen->rw, R600_BLEND);
 	if (rstate == NULL)
 		return NULL;
 
@@ -543,7 +548,7 @@ static struct radeon_state *r600_blit_state_cb_cntl(struct r600_screen *rscreen)
 {
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_CB_CNTL_TYPE, R600_CB_CNTL);
+	rstate = radeon_state(rscreen->rw, R600_CB_CNTL);
 	if (rstate == NULL)
 		return NULL;
 
@@ -786,10 +791,10 @@ int r600_blit_uncompress_depth(struct pipe_context *ctx, struct r600_resource_te
 	r600_queries_suspend(ctx);
 
 	/* schedule draw*/
-	r = radeon_ctx_set_draw_new(rctx->ctx, draw);
+	r = radeon_ctx_set_draw(rctx->ctx, draw);
 	if (r == -EBUSY) {
 		r600_flush(ctx, 0, NULL);
-		r = radeon_ctx_set_draw_new(rctx->ctx, draw);
+		r = radeon_ctx_set_draw(rctx->ctx, draw);
 	}
 	if (r) {
 		goto out;
diff --git a/src/gallium/drivers/r600/r600_context.c b/src/gallium/drivers/r600/r600_context.c
index 9af28356c5c..790a85110bf 100644
--- a/src/gallium/drivers/r600/r600_context.c
+++ b/src/gallium/drivers/r600/r600_context.c
@@ -53,12 +53,10 @@ void r600_flush(struct pipe_context *ctx, unsigned flags,
 
 	/* suspend queries */
 	r600_queries_suspend(ctx);
-	if (radeon_ctx_pm4(rctx->ctx))
-		goto out;
 	/* FIXME dumping should be removed once shader support instructions
 	 * without throwing bad code
 	 */
-	if (!rctx->ctx->cpm4)
+	if (!rctx->ctx->id)
 		goto out;
 	sprintf(dname, "gallium-%08d.bof", dc);
 	if (dc < 2) {
@@ -73,8 +71,7 @@ void r600_flush(struct pipe_context *ctx, unsigned flags,
 	}
 	dc++;
 out:
-	rctx->ctx = radeon_ctx_decref(rctx->ctx);
-	rctx->ctx = radeon_ctx(rscreen->rw);
+	radeon_ctx_clear(rctx->ctx);
 	/* resume queries */
 	r600_queries_resume(ctx);
 }
@@ -218,7 +215,7 @@ static void r600_init_config(struct r600_context *rctx)
 		num_es_stack_entries = 0;
 		break;
 	}
-	rctx->hw_states.config = radeon_state(rctx->rw, R600_CONFIG_TYPE, R600_CONFIG);
+	rctx->hw_states.config = radeon_state(rctx->rw, R600_CONFIG);
 
 	rctx->hw_states.config->states[R600_CONFIG__SQ_CONFIG] = 0x00000000;
 	switch (family) {
diff --git a/src/gallium/drivers/r600/r600_draw.c b/src/gallium/drivers/r600/r600_draw.c
index 1eb868c4c77..a1a392ad2b5 100644
--- a/src/gallium/drivers/r600/r600_draw.c
+++ b/src/gallium/drivers/r600/r600_draw.c
@@ -101,19 +101,21 @@ static int r600_draw_common(struct r600_draw *draw)
 		rbuffer = (struct r600_resource*)vertex_buffer->buffer;
 		offset = rctx->vertex_elements->elements[i].src_offset + vertex_buffer->buffer_offset;
 		format = r600_translate_colorformat(rctx->vertex_elements->elements[i].src_format);
-		vs_resource = radeon_state(rscreen->rw, R600_VS_RESOURCE_TYPE, R600_VS_RESOURCE + i);
+		vs_resource = radeon_state(rscreen->rw, R600_VS_RESOURCE0 + i);
 		if (vs_resource == NULL)
 			return -ENOMEM;
 		vs_resource->bo[0] = radeon_bo_incref(rscreen->rw, rbuffer->bo);
 		vs_resource->nbo = 1;
-		vs_resource->states[R600_PS_RESOURCE__RESOURCE0_WORD0] = offset;
-		vs_resource->states[R600_PS_RESOURCE__RESOURCE0_WORD1] = rbuffer->bo->size - offset;
-		vs_resource->states[R600_PS_RESOURCE__RESOURCE0_WORD2] = S_038008_STRIDE(vertex_buffer->stride) |
+		vs_resource->reloc_pm4_id[0] = R600_RESOURCE__RESOURCE_BO0_ID;
+		vs_resource->reloc_pm4_id[1] = R600_RESOURCE__RESOURCE_BO1_ID;
+		vs_resource->states[R600_RESOURCE__RESOURCE_WORD0] = offset;
+		vs_resource->states[R600_RESOURCE__RESOURCE_WORD1] = rbuffer->bo->size - offset;
+		vs_resource->states[R600_RESOURCE__RESOURCE_WORD2] = S_038008_STRIDE(vertex_buffer->stride) |
 								S_038008_DATA_FORMAT(format);
-		vs_resource->states[R600_PS_RESOURCE__RESOURCE0_WORD3] = 0x00000000;
-		vs_resource->states[R600_PS_RESOURCE__RESOURCE0_WORD4] = 0x00000000;
-		vs_resource->states[R600_PS_RESOURCE__RESOURCE0_WORD5] = 0x00000000;
-		vs_resource->states[R600_PS_RESOURCE__RESOURCE0_WORD6] = 0xC0000000;
+		vs_resource->states[R600_RESOURCE__RESOURCE_WORD3] = 0x00000000;
+		vs_resource->states[R600_RESOURCE__RESOURCE_WORD4] = 0x00000000;
+		vs_resource->states[R600_RESOURCE__RESOURCE_WORD5] = 0x00000000;
+		vs_resource->states[R600_RESOURCE__RESOURCE_WORD6] = 0xC0000000;
 		vs_resource->placement[0] = RADEON_GEM_DOMAIN_GTT;
 		vs_resource->placement[1] = RADEON_GEM_DOMAIN_GTT;
 		r = radeon_draw_set_new(rctx->draw, vs_resource);
@@ -121,22 +123,29 @@ static int r600_draw_common(struct r600_draw *draw)
 			return r;
 	}
 	/* FIXME start need to change winsys */
-	draw->draw = radeon_state(rscreen->rw, R600_DRAW_TYPE, R600_DRAW);
-	if (draw->draw == NULL)
-		return -ENOMEM;
-	draw->draw->states[R600_DRAW__VGT_NUM_INDICES] = draw->count;
-	draw->draw->states[R600_DRAW__VGT_DRAW_INITIATOR] = vgt_draw_initiator;
 	if (draw->index_buffer) {
+		draw->draw = radeon_state(rscreen->rw, R600_DRAW);
+		if (draw->draw == NULL)
+			return -ENOMEM;
+		draw->draw->states[R600_DRAW__VGT_NUM_INDICES] = draw->count;
+		draw->draw->states[R600_DRAW__VGT_DRAW_INITIATOR] = vgt_draw_initiator;
 		rbuffer = (struct r600_resource*)draw->index_buffer;
 		draw->draw->bo[0] = radeon_bo_incref(rscreen->rw, rbuffer->bo);
 		draw->draw->placement[0] = RADEON_GEM_DOMAIN_GTT;
 		draw->draw->placement[1] = RADEON_GEM_DOMAIN_GTT;
 		draw->draw->nbo = 1;
+		draw->draw->reloc_pm4_id[0] = R600_DRAW__INDICES_BO_ID;
+	} else {
+		draw->draw = radeon_state(rscreen->rw, R600_DRAW_AUTO);
+		if (draw->draw == NULL)
+			return -ENOMEM;
+		draw->draw->states[R600_DRAW_AUTO__VGT_NUM_INDICES] = draw->count;
+		draw->draw->states[R600_DRAW_AUTO__VGT_DRAW_INITIATOR] = vgt_draw_initiator;
 	}
 	r = radeon_draw_set_new(rctx->draw, draw->draw);
 	if (r)
 		return r;
-	draw->vgt = radeon_state(rscreen->rw, R600_VGT_TYPE, R600_VGT);
+	draw->vgt = radeon_state(rscreen->rw, R600_VGT);
 	if (draw->vgt == NULL)
 		return -ENOMEM;
 	draw->vgt->states[R600_VGT__VGT_PRIMITIVE_TYPE] = prim;
@@ -145,23 +154,18 @@ static int r600_draw_common(struct r600_draw *draw)
 	draw->vgt->states[R600_VGT__VGT_INDX_OFFSET] = draw->start;
 	draw->vgt->states[R600_VGT__VGT_MULTI_PRIM_IB_RESET_INDX] = 0x00000000;
 	draw->vgt->states[R600_VGT__VGT_DMA_INDEX_TYPE] = vgt_dma_index_type;
-	draw->vgt->states[R600_VGT__VGT_PRIMITIVEID_EN] = 0x00000000;
 	draw->vgt->states[R600_VGT__VGT_DMA_NUM_INSTANCES] = 0x00000001;
-	draw->vgt->states[R600_VGT__VGT_MULTI_PRIM_IB_RESET_EN] = 0x00000000;
-	draw->vgt->states[R600_VGT__VGT_INSTANCE_STEP_RATE_0] = 0x00000000;
-	draw->vgt->states[R600_VGT__VGT_INSTANCE_STEP_RATE_1] = 0x00000000;
 	r = radeon_draw_set_new(rctx->draw, draw->vgt);
 	if (r)
 		return r;
 	/* FIXME */
-	r = radeon_ctx_set_draw_new(rctx->ctx, rctx->draw);
+	r = radeon_ctx_set_draw(rctx->ctx, rctx->draw);
 	if (r == -EBUSY) {
 		r600_flush(draw->ctx, 0, NULL);
-		r = radeon_ctx_set_draw_new(rctx->ctx, rctx->draw);
+		r = radeon_ctx_set_draw(rctx->ctx, rctx->draw);
 	}
 	if (r)
 		return r;
-	rctx->draw = radeon_draw_duplicate(rctx->draw);
 	return 0;
 }
 
diff --git a/src/gallium/drivers/r600/r600_query.c b/src/gallium/drivers/r600/r600_query.c
index 5929606cd28..8b4fe8999f3 100644
--- a/src/gallium/drivers/r600/r600_query.c
+++ b/src/gallium/drivers/r600/r600_query.c
@@ -36,10 +36,11 @@ static struct radeon_state *r600_query_begin(struct r600_context *rctx, struct r
 	struct r600_screen *rscreen = rctx->screen;
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_QUERY_BEGIN_TYPE, R600_QUERY_BEGIN);
+	rstate = radeon_state(rscreen->rw, R600_QUERY_BEGIN);
 	if (rstate == NULL)
 		return NULL;
 	rstate->states[R600_QUERY__OFFSET] = rquery->num_results;
+	rstate->reloc_pm4_id[0] = R600_QUERY__BO_ID;
 	rstate->bo[0] = radeon_bo_incref(rscreen->rw, rquery->buffer);
 	rstate->nbo = 1;
 	rstate->placement[0] = RADEON_GEM_DOMAIN_GTT;
@@ -55,10 +56,11 @@ static struct radeon_state *r600_query_end(struct r600_context *rctx, struct r60
 	struct r600_screen *rscreen = rctx->screen;
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_QUERY_END_TYPE, R600_QUERY_END);
+	rstate = radeon_state(rscreen->rw, R600_QUERY_END);
 	if (rstate == NULL)
 		return NULL;
 	rstate->states[R600_QUERY__OFFSET] = rquery->num_results + 8;
+	rstate->reloc_pm4_id[0] = R600_QUERY__BO_ID;
 	rstate->bo[0] = radeon_bo_incref(rscreen->rw, rquery->buffer);
 	rstate->nbo = 1;
 	rstate->placement[0] = RADEON_GEM_DOMAIN_GTT;
diff --git a/src/gallium/drivers/r600/r600_shader.c b/src/gallium/drivers/r600/r600_shader.c
index b2d1a1bf016..f0b7df5a6f3 100644
--- a/src/gallium/drivers/r600/r600_shader.c
+++ b/src/gallium/drivers/r600/r600_shader.c
@@ -132,7 +132,7 @@ static int r600_pipe_shader_vs(struct pipe_context *ctx, struct r600_context_sta
 	unsigned i, tmp;
 
 	rpshader->rstate = radeon_state_decref(rpshader->rstate);
-	state = radeon_state(rscreen->rw, R600_VS_SHADER_TYPE, R600_VS_SHADER);
+	state = radeon_state(rscreen->rw, R600_VS_SHADER);
 	if (state == NULL)
 		return -ENOMEM;
 	for (i = 0; i < 10; i++) {
@@ -151,6 +151,8 @@ static int r600_pipe_shader_vs(struct pipe_context *ctx, struct r600_context_sta
 	rpshader->rstate->nbo = 2;
 	rpshader->rstate->placement[0] = RADEON_GEM_DOMAIN_GTT;
 	rpshader->rstate->placement[2] = RADEON_GEM_DOMAIN_GTT;
+	state->reloc_pm4_id[0] = R600_VS_SHADER__SQ_PGM_START_VS_BO_ID;
+	state->reloc_pm4_id[1] = R600_VS_SHADER__SQ_PGM_START_FS_BO_ID;
 	return radeon_state_pm4(state);
 }
 
@@ -165,7 +167,7 @@ static int r600_pipe_shader_ps(struct pipe_context *ctx, struct r600_context_sta
 
 	rasterizer = &rctx->rasterizer->state.rasterizer;
 	rpshader->rstate = radeon_state_decref(rpshader->rstate);
-	state = radeon_state(rscreen->rw, R600_PS_SHADER_TYPE, R600_PS_SHADER);
+	state = radeon_state(rscreen->rw, R600_PS_SHADER);
 	if (state == NULL)
 		return -ENOMEM;
 	for (i = 0; i < rshader->ninput; i++) {
@@ -204,6 +206,7 @@ static int r600_pipe_shader_ps(struct pipe_context *ctx, struct r600_context_sta
 	rpshader->rstate->bo[0] = radeon_bo_incref(rscreen->rw, rpshader->bo);
 	rpshader->rstate->nbo = 1;
 	rpshader->rstate->placement[0] = RADEON_GEM_DOMAIN_GTT;
+	state->reloc_pm4_id[0] = R600_PS_SHADER__SQ_PGM_START_PS_BO_ID;
 	return radeon_state_pm4(state);
 }
 
diff --git a/src/gallium/drivers/r600/r600_state.c b/src/gallium/drivers/r600/r600_state.c
index b5e5346163c..e75575da792 100644
--- a/src/gallium/drivers/r600/r600_state.c
+++ b/src/gallium/drivers/r600/r600_state.c
@@ -283,19 +283,17 @@ static void r600_set_constant_buffer(struct pipe_context *ctx,
 {
 	struct r600_screen *rscreen = r600_screen(ctx->screen);
 	struct r600_context *rctx = r600_context(ctx);
-	unsigned nconstant = 0, i, type, id;
+	unsigned nconstant = 0, i, id;
 	struct radeon_state *rstate;
 	struct pipe_transfer *transfer;
 	u32 *ptr;
 
 	switch (shader) {
 	case PIPE_SHADER_VERTEX:
-		id = R600_VS_CONSTANT;
-		type = R600_VS_CONSTANT_TYPE;
+		id = R600_VS_CONSTANT0;
 		break;
 	case PIPE_SHADER_FRAGMENT:
-		id = R600_PS_CONSTANT;
-		type = R600_PS_CONSTANT_TYPE;
+		id = R600_PS_CONSTANT0;
 		break;
 	default:
 		R600_ERR("unsupported %d\n", shader);
@@ -307,7 +305,7 @@ static void r600_set_constant_buffer(struct pipe_context *ctx,
 		if (ptr == NULL)
 			return;
 		for (i = 0; i < nconstant; i++) {
-			rstate = radeon_state(rscreen->rw, type, id + i);
+			rstate = radeon_state(rscreen->rw, id + i);
 			if (rstate == NULL)
 				return;
 			rstate->states[R600_PS_CONSTANT__SQ_ALU_CONSTANT0_0] = ptr[i * 4 + 0];
@@ -622,7 +620,7 @@ static struct radeon_state *r600_blend(struct r600_context *rctx)
 	const struct pipe_blend_state *state = &rctx->blend->state.blend;
 	int i;
 
-	rstate = radeon_state(rscreen->rw, R600_BLEND_TYPE, R600_BLEND);
+	rstate = radeon_state(rscreen->rw, R600_BLEND);
 	if (rstate == NULL)
 		return NULL;
 	rstate->states[R600_BLEND__CB_BLEND_RED] = fui(rctx->blend_color.color[0]);
@@ -681,14 +679,14 @@ static struct radeon_state *r600_ucp(struct r600_context *rctx, int clip)
 	struct radeon_state *rstate;
 	const struct pipe_clip_state *state = &rctx->clip->state.clip;
 
-	rstate = radeon_state(rscreen->rw, R600_CLIP_TYPE, R600_CLIP + clip);
+	rstate = radeon_state(rscreen->rw, R600_UCP0 + clip);
 	if (rstate == NULL)
 		return NULL;
 
-	rstate->states[R600_CLIP__PA_CL_UCP_X_0] = fui(state->ucp[clip][0]);
-	rstate->states[R600_CLIP__PA_CL_UCP_Y_0] = fui(state->ucp[clip][1]);
-	rstate->states[R600_CLIP__PA_CL_UCP_Z_0] = fui(state->ucp[clip][2]);
-	rstate->states[R600_CLIP__PA_CL_UCP_W_0] = fui(state->ucp[clip][3]);
+	rstate->states[R600_UCP__PA_CL_UCP_X_0] = fui(state->ucp[clip][0]);
+	rstate->states[R600_UCP__PA_CL_UCP_Y_0] = fui(state->ucp[clip][1]);
+	rstate->states[R600_UCP__PA_CL_UCP_Z_0] = fui(state->ucp[clip][2]);
+	rstate->states[R600_UCP__PA_CL_UCP_W_0] = fui(state->ucp[clip][3]);
 
 	if (radeon_state_pm4(rstate)) {
 		radeon_state_decref(rstate);
@@ -711,7 +709,7 @@ static struct radeon_state *r600_cb(struct r600_context *rctx, int cb)
 	unsigned format, swap, ntype;
 	const struct util_format_description *desc;
 
-	rstate = radeon_state(rscreen->rw, R600_CB0_TYPE + cb, R600_CB0 + cb);
+	rstate = radeon_state(rscreen->rw, R600_CB0 + cb);
 	if (rstate == NULL)
 		return NULL;
 	rtex = (struct r600_resource_texture*)state->cbufs[cb]->texture;
@@ -722,6 +720,9 @@ static struct radeon_state *r600_cb(struct r600_context *rctx, int cb)
 	rstate->placement[0] = RADEON_GEM_DOMAIN_GTT;
 	rstate->placement[2] = RADEON_GEM_DOMAIN_GTT;
 	rstate->placement[4] = RADEON_GEM_DOMAIN_GTT;
+	rstate->reloc_pm4_id[0] = R600_CB__CB_COLOR0_BASE_BO_ID;
+	rstate->reloc_pm4_id[1] = R600_CB__CB_COLOR0_FRAG_BO_ID;
+	rstate->reloc_pm4_id[2] = R600_CB__CB_COLOR0_TILE_BO_ID;
 	rstate->nbo = 3;
 	pitch = (rtex->pitch[level] / rtex->bpt) / 8 - 1;
 	slice = (rtex->pitch[level] / rtex->bpt) * state->cbufs[cb]->height / 64 - 1;
@@ -740,14 +741,14 @@ static struct radeon_state *r600_cb(struct r600_context *rctx, int cb)
 		S_0280A0_SOURCE_FORMAT(1) |
 		S_0280A0_NUMBER_TYPE(ntype);
 
-	rstate->states[R600_CB0__CB_COLOR0_BASE] = rtex->offset[level] >> 8;
-	rstate->states[R600_CB0__CB_COLOR0_INFO] = color_info;
-	rstate->states[R600_CB0__CB_COLOR0_SIZE] = S_028060_PITCH_TILE_MAX(pitch) |
+	rstate->states[R600_CB__CB_COLOR0_BASE] = rtex->offset[level] >> 8;
+	rstate->states[R600_CB__CB_COLOR0_INFO] = color_info;
+	rstate->states[R600_CB__CB_COLOR0_SIZE] = S_028060_PITCH_TILE_MAX(pitch) |
 						S_028060_SLICE_TILE_MAX(slice);
-	rstate->states[R600_CB0__CB_COLOR0_VIEW] = 0x00000000;
-	rstate->states[R600_CB0__CB_COLOR0_FRAG] = 0x00000000;
-	rstate->states[R600_CB0__CB_COLOR0_TILE] = 0x00000000;
-	rstate->states[R600_CB0__CB_COLOR0_MASK] = 0x00000000;
+	rstate->states[R600_CB__CB_COLOR0_VIEW] = 0x00000000;
+	rstate->states[R600_CB__CB_COLOR0_FRAG] = 0x00000000;
+	rstate->states[R600_CB__CB_COLOR0_TILE] = 0x00000000;
+	rstate->states[R600_CB__CB_COLOR0_MASK] = 0x00000000;
 	if (radeon_state_pm4(rstate)) {
 		radeon_state_decref(rstate);
 		return NULL;
@@ -768,7 +769,7 @@ static struct radeon_state *r600_db(struct r600_context *rctx)
 	if (state->zsbuf == NULL)
 		return NULL;
 
-	rstate = radeon_state(rscreen->rw, R600_DB_TYPE, R600_DB);
+	rstate = radeon_state(rscreen->rw, R600_DB);
 	if (rstate == NULL)
 		return NULL;
 
@@ -782,6 +783,7 @@ static struct radeon_state *r600_db(struct r600_context *rctx)
 	rstate->bo[0] = radeon_bo_incref(rscreen->rw, rbuffer->bo);
 	rstate->nbo = 1;
 	rstate->placement[0] = RADEON_GEM_DOMAIN_VRAM;
+	rstate->reloc_pm4_id[0] = R600_DB__DB_DEPTH_BASE_BO_ID;
 	level = state->zsbuf->level;
 	pitch = (rtex->pitch[level] / rtex->bpt) / 8 - 1;
 	slice = (rtex->pitch[level] / rtex->bpt) * state->zsbuf->height / 64 - 1;
@@ -844,7 +846,7 @@ static struct radeon_state *r600_rasterizer(struct r600_context *rctx)
 		prov_vtx = 0;
 
 	rctx->flat_shade = state->flatshade;
-	rstate = radeon_state(rscreen->rw, R600_RASTERIZER_TYPE, R600_RASTERIZER);
+	rstate = radeon_state(rscreen->rw, R600_RASTERIZER);
 	if (rstate == NULL)
 		return NULL;
 	rstate->states[R600_RASTERIZER__SPI_INTERP_CONTROL_0] = 0x00000001;
@@ -925,7 +927,7 @@ static struct radeon_state *r600_scissor(struct r600_context *rctx)
 	}
 	tl = S_028240_TL_X(minx) | S_028240_TL_Y(miny) | S_028240_WINDOW_OFFSET_DISABLE(1);
 	br = S_028244_BR_X(maxx) | S_028244_BR_Y(maxy);
-	rstate = radeon_state(rscreen->rw, R600_SCISSOR_TYPE, R600_SCISSOR);
+	rstate = radeon_state(rscreen->rw, R600_SCISSOR);
 	if (rstate == NULL)
 		return NULL;
 	rstate->states[R600_SCISSOR__PA_SC_SCREEN_SCISSOR_TL] = tl;
@@ -960,7 +962,7 @@ static struct radeon_state *r600_viewport(struct r600_context *rctx)
 	struct r600_screen *rscreen = rctx->screen;
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_VIEWPORT_TYPE, R600_VIEWPORT);
+	rstate = radeon_state(rscreen->rw, R600_VIEWPORT);
 	if (rstate == NULL)
 		return NULL;
 	rstate->states[R600_VIEWPORT__PA_SC_VPORT_ZMIN_0] = 0x00000000;
@@ -993,7 +995,7 @@ static struct radeon_state *r600_dsa(struct r600_context *rctx)
 	if (rctx->ps_shader == NULL) {
 		return NULL;
 	}
-	rstate = radeon_state(rscreen->rw, R600_DSA_TYPE, R600_DSA);
+	rstate = radeon_state(rscreen->rw, R600_DSA);
 	if (rstate == NULL)
 		return NULL;
 
@@ -1145,7 +1147,7 @@ static struct radeon_state *r600_sampler(struct r600_context *rctx,
 	struct r600_screen *rscreen = rctx->screen;
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_PS_SAMPLER_TYPE, id);
+	rstate = radeon_state(rscreen->rw, id);
 	if (rstate == NULL)
 		return NULL;
 	rstate->states[R600_PS_SAMPLER__SQ_TEX_SAMPLER_WORD0_0] =
@@ -1246,7 +1248,7 @@ static struct radeon_state *r600_resource(struct pipe_context *ctx,
 		R600_ERR("unknow format %d\n", view->texture->format);
 		return NULL;
 	}
-	rstate = radeon_state(rscreen->rw, R600_PS_RESOURCE_TYPE, id);
+	rstate = radeon_state(rscreen->rw, id);
 	if (rstate == NULL) {
 		return NULL;
 	}
@@ -1268,34 +1270,36 @@ static struct radeon_state *r600_resource(struct pipe_context *ctx,
 	rstate->placement[1] = RADEON_GEM_DOMAIN_GTT;
 	rstate->placement[2] = RADEON_GEM_DOMAIN_GTT;
 	rstate->placement[3] = RADEON_GEM_DOMAIN_GTT;
+	rstate->reloc_pm4_id[0] = R600_RESOURCE__RESOURCE_BO0_ID;
+	rstate->reloc_pm4_id[1] = R600_RESOURCE__RESOURCE_BO1_ID;
 
 	pitch = (tmp->pitch[0] / tmp->bpt);
 	pitch = (pitch + 0x7) & ~0x7;
 	
 	/* FIXME properly handle first level != 0 */
-	rstate->states[R600_PS_RESOURCE__RESOURCE0_WORD0] =
+	rstate->states[R600_RESOURCE__RESOURCE_WORD0] =
 			S_038000_DIM(r600_tex_dim(view->texture->target)) |
 			S_038000_TILE_MODE(array_mode) |
 			S_038000_TILE_TYPE(tile_type) |
 			S_038000_PITCH((pitch / 8) - 1) |
 			S_038000_TEX_WIDTH(view->texture->width0 - 1);
-	rstate->states[R600_PS_RESOURCE__RESOURCE0_WORD1] =
+	rstate->states[R600_RESOURCE__RESOURCE_WORD1] =
 			S_038004_TEX_HEIGHT(view->texture->height0 - 1) |
 			S_038004_TEX_DEPTH(view->texture->depth0 - 1) |
 			S_038004_DATA_FORMAT(format);
-	rstate->states[R600_PS_RESOURCE__RESOURCE0_WORD2] = tmp->offset[0] >> 8;
-	rstate->states[R600_PS_RESOURCE__RESOURCE0_WORD3] = tmp->offset[1] >> 8;
-	rstate->states[R600_PS_RESOURCE__RESOURCE0_WORD4] =
+	rstate->states[R600_RESOURCE__RESOURCE_WORD2] = tmp->offset[0] >> 8;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD3] = tmp->offset[1] >> 8;
+	rstate->states[R600_RESOURCE__RESOURCE_WORD4] =
 		        word4 | 
 			S_038010_NUM_FORMAT_ALL(V_038010_SQ_NUM_FORMAT_NORM) |
 			S_038010_SRF_MODE_ALL(V_038010_SFR_MODE_NO_ZERO) |
 			S_038010_REQUEST_SIZE(1) |
 			S_038010_BASE_LEVEL(view->first_level);
-	rstate->states[R600_PS_RESOURCE__RESOURCE0_WORD5] =
+	rstate->states[R600_RESOURCE__RESOURCE_WORD5] =
 			S_038014_LAST_LEVEL(view->last_level) |
 			S_038014_BASE_ARRAY(0) |
 			S_038014_LAST_ARRAY(0);
-	rstate->states[R600_PS_RESOURCE__RESOURCE0_WORD6] =
+	rstate->states[R600_RESOURCE__RESOURCE_WORD6] =
 			S_038018_TYPE(V_038010_SQ_TEX_VTX_VALID_TEXTURE);
 	if (radeon_state_pm4(rstate)) {
 		radeon_state_decref(rstate);
@@ -1342,7 +1346,7 @@ static struct radeon_state *r600_cb_cntl(struct r600_context *rctx)
 			target_mask |= (pbs->rt[0].colormask << (4 * i));
 		}
 	}
-	rstate = radeon_state(rscreen->rw, R600_CB_CNTL_TYPE, R600_CB_CNTL);
+	rstate = radeon_state(rscreen->rw, R600_CB_CNTL);
 	rstate->states[R600_CB_CNTL__CB_SHADER_MASK] = shader_mask;
 	rstate->states[R600_CB_CNTL__CB_TARGET_MASK] = target_mask;
 	rstate->states[R600_CB_CNTL__CB_COLOR_CONTROL] = color_control;
@@ -1419,7 +1423,7 @@ int r600_context_hw_states(struct pipe_context *ctx)
 		if (rctx->ps_sampler[i]) {
 			rctx->hw_states.ps_sampler[i] = r600_sampler(rctx,
 							&rctx->ps_sampler[i]->state.sampler,
-							R600_PS_SAMPLER + i);
+							R600_PS_SAMPLER0 + i);
 		}
 	}
 	rctx->hw_states.ps_nsampler = rctx->ps_nsampler;
@@ -1427,7 +1431,7 @@ int r600_context_hw_states(struct pipe_context *ctx)
 		if (rctx->ps_sampler_view[i]) {
 			rctx->hw_states.ps_resource[i] = r600_resource(ctx,
 							&rctx->ps_sampler_view[i]->state.sampler_view,
-							R600_PS_RESOURCE + i);
+							R600_PS_RESOURCE0 + i);
 		}
 	}
 	rctx->hw_states.ps_nresource = rctx->ps_nsampler_view;
diff --git a/src/gallium/drivers/r600/r600_texture.c b/src/gallium/drivers/r600/r600_texture.c
index fb84ed9cfea..9dc0208eb13 100644
--- a/src/gallium/drivers/r600/r600_texture.c
+++ b/src/gallium/drivers/r600/r600_texture.c
@@ -663,7 +663,7 @@ static struct radeon_state *r600_texture_state_scissor(struct r600_screen *rscre
 {
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_SCISSOR_TYPE, R600_SCISSOR);
+	rstate = radeon_state(rscreen->rw, R600_SCISSOR);
 	if (rstate == NULL)
 		return NULL;
 
@@ -707,7 +707,7 @@ static struct radeon_state *r600_texture_state_cb0(struct r600_screen *rscreen,
 	unsigned format, swap, ntype;
 	const struct util_format_description *desc;
 
-	rstate = radeon_state(rscreen->rw, R600_CB0_TYPE, R600_CB0);
+	rstate = radeon_state(rscreen->rw, R600_CB0);
 	if (rstate == NULL)
 		return NULL;
 	rbuffer = &rtexture->resource;
@@ -742,13 +742,16 @@ static struct radeon_state *r600_texture_state_cb0(struct r600_screen *rscreen,
 		rstate->nbo = 3;
 		color_info = S_0280A0_SOURCE_FORMAT(1);
 	}
+	rstate->reloc_pm4_id[0] = R600_CB__CB_COLOR0_BASE_BO_ID;
+	rstate->reloc_pm4_id[1] = R600_CB__CB_COLOR0_FRAG_BO_ID;
+	rstate->reloc_pm4_id[2] = R600_CB__CB_COLOR0_TILE_BO_ID;
 	color_info |= S_0280A0_FORMAT(format) |
 		S_0280A0_COMP_SWAP(swap) |
 		S_0280A0_BLEND_CLAMP(1) |
 		S_0280A0_NUMBER_TYPE(ntype);
-	rstate->states[R600_CB0__CB_COLOR0_BASE] = rtexture->offset[level] >> 8;
-	rstate->states[R600_CB0__CB_COLOR0_INFO] = color_info;
-	rstate->states[R600_CB0__CB_COLOR0_SIZE] = S_028060_PITCH_TILE_MAX(pitch) |
+	rstate->states[R600_CB__CB_COLOR0_BASE] = rtexture->offset[level] >> 8;
+	rstate->states[R600_CB__CB_COLOR0_INFO] = color_info;
+	rstate->states[R600_CB__CB_COLOR0_SIZE] = S_028060_PITCH_TILE_MAX(pitch) |
 						S_028060_SLICE_TILE_MAX(slice);
 
 	if (radeon_state_pm4(rstate)) {
@@ -766,7 +769,7 @@ static struct radeon_state *r600_texture_state_db(struct r600_screen *rscreen,
 	struct r600_resource *rbuffer;
 	unsigned pitch, slice, format;
 
-	rstate = radeon_state(rscreen->rw, R600_DB_TYPE, R600_DB);
+	rstate = radeon_state(rscreen->rw, R600_DB);
 	if (rstate == NULL)
 		return NULL;
 	rbuffer = &rtexture->resource;
@@ -784,6 +787,7 @@ static struct radeon_state *r600_texture_state_db(struct r600_screen *rscreen,
 	rstate->states[R600_DB__DB_PREFETCH_LIMIT] = (rtexture->height[level] / 8) -1;
 	rstate->states[R600_DB__DB_DEPTH_SIZE] = S_028000_PITCH_TILE_MAX(pitch) |
 						S_028000_SLICE_TILE_MAX(slice);
+	rstate->reloc_pm4_id[0] = R600_DB__DB_DEPTH_BASE_BO_ID;
 	rstate->bo[0] = radeon_bo_incref(rscreen->rw, rbuffer->bo);
 	rstate->placement[0] = RADEON_GEM_DOMAIN_GTT;
 	rstate->nbo = 1;
@@ -815,7 +819,7 @@ static struct radeon_state *r600_texture_state_viewport(struct r600_screen *rscr
 {
 	struct radeon_state *rstate;
 
-	rstate = radeon_state(rscreen->rw, R600_VIEWPORT_TYPE, R600_VIEWPORT);
+	rstate = radeon_state(rscreen->rw, R600_VIEWPORT);
 	if (rstate == NULL)
 		return NULL;
 
diff --git a/src/gallium/drivers/r600/radeon.h b/src/gallium/drivers/r600/radeon.h
index b2cc74f6967..777fe3e7922 100644
--- a/src/gallium/drivers/r600/radeon.h
+++ b/src/gallium/drivers/r600/radeon.h
@@ -104,26 +104,18 @@ int radeon_bo_wait(struct radeon *radeon, struct radeon_bo *bo);
 struct radeon_state {
 	struct radeon			*radeon;
 	unsigned			refcount;
-	unsigned			type;
 	unsigned			id;
-	unsigned			nstates;
-	u32				*states;
-	unsigned			npm4;
 	unsigned			cpm4;
+	u32				states[128];
 	u32				pm4_crc;
-	u32				*pm4;
-	u32				nimmd;
-	u32				*immd;
 	unsigned			nbo;
 	struct radeon_bo		*bo[4];
-	unsigned			nreloc;
-	unsigned			reloc_pm4_id[8];
-	unsigned			reloc_bo_id[8];
+	unsigned			reloc_pm4_id[4];
 	u32				placement[8];
 	unsigned			bo_dirty[4];
 };
 
-struct radeon_state *radeon_state(struct radeon *radeon, u32 type, u32 id);
+struct radeon_state *radeon_state(struct radeon *radeon, u32 id);
 struct radeon_state *radeon_state_incref(struct radeon_state *state);
 struct radeon_state *radeon_state_decref(struct radeon_state *state);
 int radeon_state_pm4(struct radeon_state *state);
@@ -147,16 +139,6 @@ int radeon_draw_set(struct radeon_draw *draw, struct radeon_state *state);
 int radeon_draw_set_new(struct radeon_draw *draw, struct radeon_state *state);
 int radeon_draw_check(struct radeon_draw *draw);
 
-struct radeon_ctx *radeon_ctx(struct radeon *radeon);
-struct radeon_ctx *radeon_ctx_decref(struct radeon_ctx *ctx);
-struct radeon_ctx *radeon_ctx_incref(struct radeon_ctx *ctx);
-int radeon_ctx_set_draw(struct radeon_ctx *ctx, struct radeon_draw *draw);
-int radeon_ctx_set_query_state(struct radeon_ctx *ctx, struct radeon_state *state);
-int radeon_ctx_set_draw_new(struct radeon_ctx *ctx, struct radeon_draw *draw);
-int radeon_ctx_pm4(struct radeon_ctx *ctx);
-int radeon_ctx_submit(struct radeon_ctx *ctx);
-void radeon_ctx_dump_bof(struct radeon_ctx *ctx, const char *file);
-
 /*
  * radeon context functions
  */
@@ -169,261 +151,216 @@ struct radeon_cs_reloc {
 };
 #pragma pack()
 
+struct radeon_ctx_bo {
+	struct radeon_bo		*bo;
+	u32				bo_flushed;
+	unsigned			state_id;
+};
+
 struct radeon_ctx {
 	int				refcount;
 	struct radeon			*radeon;
 	u32				*pm4;
-	u32				cpm4;
-	u32				draw_cpm4;
+	int				npm4;
 	unsigned			id;
-	unsigned			next_id;
 	unsigned			nreloc;
+	unsigned			max_reloc;
 	struct radeon_cs_reloc		*reloc;
 	unsigned			nbo;
-	struct radeon_bo		**bo;
-	unsigned			ndraw;
-	struct radeon_draw		*cdraw;
-	struct radeon_draw		**draw;
-	unsigned			nstate;
-	struct radeon_state		**state;
+	struct radeon_ctx_bo		*bo;
+	unsigned			max_bo;
+	u32				*state_crc32;
 };
 
+struct radeon_ctx *radeon_ctx(struct radeon *radeon);
+struct radeon_ctx *radeon_ctx_decref(struct radeon_ctx *ctx);
+struct radeon_ctx *radeon_ctx_incref(struct radeon_ctx *ctx);
+void radeon_ctx_clear(struct radeon_ctx *ctx);
+int radeon_ctx_set_draw(struct radeon_ctx *ctx, struct radeon_draw *draw);
+int radeon_ctx_set_query_state(struct radeon_ctx *ctx, struct radeon_state *state);
+int radeon_ctx_pm4(struct radeon_ctx *ctx);
+int radeon_ctx_submit(struct radeon_ctx *ctx);
+void radeon_ctx_dump_bof(struct radeon_ctx *ctx, const char *file);
+
+
 /*
  * R600/R700
  */
-
-#define R600_NSTATE				1288
-#define R600_NTYPE				35
-
-#define R600_CONFIG				0
-#define R600_CONFIG_TYPE				0
-#define R600_CB_CNTL				1
-#define R600_CB_CNTL_TYPE				1
-#define R600_RASTERIZER				2
-#define R600_RASTERIZER_TYPE				2
-#define R600_VIEWPORT				3
-#define R600_VIEWPORT_TYPE				3
-#define R600_SCISSOR				4
-#define R600_SCISSOR_TYPE				4
-#define R600_BLEND				5
-#define R600_BLEND_TYPE				5
-#define R600_DSA				6
-#define R600_DSA_TYPE				6
-#define R600_VS_SHADER				7
-#define R600_VS_SHADER_TYPE				7
-#define R600_PS_SHADER				8
-#define R600_PS_SHADER_TYPE				8
-#define R600_PS_CONSTANT				9
-#define R600_PS_CONSTANT_TYPE				9
-#define R600_VS_CONSTANT				265
-#define R600_VS_CONSTANT_TYPE				10
-#define R600_PS_RESOURCE				521
-#define R600_PS_RESOURCE_TYPE				11
-#define R600_VS_RESOURCE				681
-#define R600_VS_RESOURCE_TYPE				12
-#define R600_FS_RESOURCE				841
-#define R600_FS_RESOURCE_TYPE				13
-#define R600_GS_RESOURCE				1001
-#define R600_GS_RESOURCE_TYPE				14
-#define R600_PS_SAMPLER				1161
-#define R600_PS_SAMPLER_TYPE				15
-#define R600_VS_SAMPLER				1179
-#define R600_VS_SAMPLER_TYPE				16
-#define R600_GS_SAMPLER				1197
-#define R600_GS_SAMPLER_TYPE				17
-#define R600_PS_SAMPLER_BORDER				1215
-#define R600_PS_SAMPLER_BORDER_TYPE				18
-#define R600_VS_SAMPLER_BORDER				1233
-#define R600_VS_SAMPLER_BORDER_TYPE				19
-#define R600_GS_SAMPLER_BORDER				1251
-#define R600_GS_SAMPLER_BORDER_TYPE				20
-#define R600_CB0				1269
-#define R600_CB0_TYPE				21
-#define R600_CB1				1270
-#define R600_CB1_TYPE				22
-#define R600_CB2				1271
-#define R600_CB2_TYPE				23
-#define R600_CB3				1272
-#define R600_CB3_TYPE				24
-#define R600_CB4				1273
-#define R600_CB4_TYPE				25
-#define R600_CB5				1274
-#define R600_CB5_TYPE				26
-#define R600_CB6				1275
-#define R600_CB6_TYPE				27
-#define R600_CB7				1276
-#define R600_CB7_TYPE				28
-#define R600_QUERY_BEGIN			1277
-#define R600_QUERY_BEGIN_TYPE			29
-#define R600_QUERY_END				1278
-#define R600_QUERY_END_TYPE			30
-#define R600_DB					1279
-#define R600_DB_TYPE				31
-#define R600_CLIP				1280
-#define R600_CLIP_TYPE				32
-#define R600_VGT				1286
-#define R600_VGT_TYPE				33
-#define R600_DRAW				1287
-#define R600_DRAW_TYPE				34
+#define R600_CONFIG		0
+#define R600_CB_CNTL		1
+#define R600_RASTERIZER		2
+#define R600_VIEWPORT		3
+#define R600_SCISSOR		4
+#define R600_BLEND		5
+#define R600_DSA		6
+#define R600_VGT		7
+#define R600_QUERY_BEGIN	8
+#define R600_QUERY_END		9
+#define R600_VS_SHADER		10
+#define R600_PS_SHADER		11
+#define R600_DB			12
+#define R600_CB0		13
+#define R600_UCP0		21
+#define R600_PS_RESOURCE0	27
+#define R600_VS_RESOURCE0	187
+#define R600_FS_RESOURCE0	347
+#define R600_GS_RESOURCE0	363
+#define R600_PS_CONSTANT0	523
+#define R600_VS_CONSTANT0	779
+#define R600_PS_SAMPLER0	1035
+#define R600_VS_SAMPLER0	1053
+#define R600_GS_SAMPLER0	1071
+#define R600_PS_SAMPLER_BORDER0	1089
+#define R600_VS_SAMPLER_BORDER0	1107
+#define R600_GS_SAMPLER_BORDER0	1125
+#define R600_DRAW_AUTO		1143
+#define R600_DRAW		1144
+#define R600_NSTATE		1145
 
 /* R600_CONFIG */
-#define R600_CONFIG__SQ_CONFIG			0
+#define R600_CONFIG__SQ_CONFIG					0
 #define R600_CONFIG__SQ_GPR_RESOURCE_MGMT_1			1
 #define R600_CONFIG__SQ_GPR_RESOURCE_MGMT_2			2
 #define R600_CONFIG__SQ_THREAD_RESOURCE_MGMT			3
 #define R600_CONFIG__SQ_STACK_RESOURCE_MGMT_1			4
 #define R600_CONFIG__SQ_STACK_RESOURCE_MGMT_2			5
-#define R600_CONFIG__SQ_DYN_GPR_CNTL_PS_FLUSH_REQ			6
-#define R600_CONFIG__TA_CNTL_AUX			7
-#define R600_CONFIG__VC_ENHANCE			8
-#define R600_CONFIG__DB_DEBUG			9
-#define R600_CONFIG__DB_WATERMARKS			10
-#define R600_CONFIG__SX_MISC			11
-#define R600_CONFIG__SPI_THREAD_GROUPING			12
-#define R600_CONFIG__CB_SHADER_CONTROL			13
-#define R600_CONFIG__SQ_ESGS_RING_ITEMSIZE			14
-#define R600_CONFIG__SQ_GSVS_RING_ITEMSIZE			15
-#define R600_CONFIG__SQ_ESTMP_RING_ITEMSIZE			16
-#define R600_CONFIG__SQ_GSTMP_RING_ITEMSIZE			17
-#define R600_CONFIG__SQ_VSTMP_RING_ITEMSIZE			18
-#define R600_CONFIG__SQ_PSTMP_RING_ITEMSIZE			19
-#define R600_CONFIG__SQ_FBUF_RING_ITEMSIZE			20
-#define R600_CONFIG__SQ_REDUC_RING_ITEMSIZE			21
-#define R600_CONFIG__SQ_GS_VERT_ITEMSIZE			22
-#define R600_CONFIG__VGT_OUTPUT_PATH_CNTL			23
-#define R600_CONFIG__VGT_HOS_CNTL			24
-#define R600_CONFIG__VGT_HOS_MAX_TESS_LEVEL			25
-#define R600_CONFIG__VGT_HOS_MIN_TESS_LEVEL			26
-#define R600_CONFIG__VGT_HOS_REUSE_DEPTH			27
-#define R600_CONFIG__VGT_GROUP_PRIM_TYPE			28
-#define R600_CONFIG__VGT_GROUP_FIRST_DECR			29
-#define R600_CONFIG__VGT_GROUP_DECR			30
-#define R600_CONFIG__VGT_GROUP_VECT_0_CNTL			31
-#define R600_CONFIG__VGT_GROUP_VECT_1_CNTL			32
-#define R600_CONFIG__VGT_GROUP_VECT_0_FMT_CNTL			33
-#define R600_CONFIG__VGT_GROUP_VECT_1_FMT_CNTL			34
-#define R600_CONFIG__VGT_GS_MODE			35
-#define R600_CONFIG__PA_SC_MODE_CNTL			36
-#define R600_CONFIG__VGT_STRMOUT_EN			37
-#define R600_CONFIG__VGT_REUSE_OFF			38
-#define R600_CONFIG__VGT_VTX_CNT_EN			39
-#define R600_CONFIG__VGT_STRMOUT_BUFFER_EN			40
-#define R600_CONFIG_SIZE				41
-#define R600_CONFIG_PM4				128
+#define R600_CONFIG__SQ_DYN_GPR_CNTL_PS_FLUSH_REQ		8
+#define R600_CONFIG__TA_CNTL_AUX				11
+#define R600_CONFIG__VC_ENHANCE					14
+#define R600_CONFIG__DB_DEBUG					17
+#define R600_CONFIG__DB_WATERMARKS				20
+#define R600_CONFIG__SX_MISC					23
+#define R600_CONFIG__SPI_THREAD_GROUPING			26
+#define R600_CONFIG__CB_SHADER_CONTROL				29
+#define R600_CONFIG__SQ_ESGS_RING_ITEMSIZE			32
+#define R600_CONFIG__SQ_GSVS_RING_ITEMSIZE			33
+#define R600_CONFIG__SQ_ESTMP_RING_ITEMSIZE			34
+#define R600_CONFIG__SQ_GSTMP_RING_ITEMSIZE			35
+#define R600_CONFIG__SQ_VSTMP_RING_ITEMSIZE			36
+#define R600_CONFIG__SQ_PSTMP_RING_ITEMSIZE			37
+#define R600_CONFIG__SQ_FBUF_RING_ITEMSIZE			38
+#define R600_CONFIG__SQ_REDUC_RING_ITEMSIZE			39
+#define R600_CONFIG__SQ_GS_VERT_ITEMSIZE			40
+#define R600_CONFIG__VGT_OUTPUT_PATH_CNTL			43
+#define R600_CONFIG__VGT_HOS_CNTL				44
+#define R600_CONFIG__VGT_HOS_MAX_TESS_LEVEL			45
+#define R600_CONFIG__VGT_HOS_MIN_TESS_LEVEL			46
+#define R600_CONFIG__VGT_HOS_REUSE_DEPTH			47
+#define R600_CONFIG__VGT_GROUP_PRIM_TYPE			48
+#define R600_CONFIG__VGT_GROUP_FIRST_DECR			49
+#define R600_CONFIG__VGT_GROUP_DECR				50
+#define R600_CONFIG__VGT_GROUP_VECT_0_CNTL			51
+#define R600_CONFIG__VGT_GROUP_VECT_1_CNTL			52
+#define R600_CONFIG__VGT_GROUP_VECT_0_FMT_CNTL			53
+#define R600_CONFIG__VGT_GROUP_VECT_1_FMT_CNTL			54
+#define R600_CONFIG__VGT_GS_MODE				55
+#define R600_CONFIG__PA_SC_MODE_CNTL				58
+#define R600_CONFIG__VGT_STRMOUT_EN				61
+#define R600_CONFIG__VGT_REUSE_OFF				62
+#define R600_CONFIG__VGT_VTX_CNT_EN				63
+#define R600_CONFIG__VGT_STRMOUT_BUFFER_EN			66
 /* R600_CB_CNTL */
-#define R600_CB_CNTL__CB_CLEAR_RED			0
-#define R600_CB_CNTL__CB_CLEAR_GREEN			1
-#define R600_CB_CNTL__CB_CLEAR_BLUE			2
-#define R600_CB_CNTL__CB_CLEAR_ALPHA			3
-#define R600_CB_CNTL__CB_SHADER_MASK			4
-#define R600_CB_CNTL__CB_TARGET_MASK			5
-#define R600_CB_CNTL__CB_FOG_RED			6
-#define R600_CB_CNTL__CB_FOG_GREEN			7
-#define R600_CB_CNTL__CB_FOG_BLUE			8
-#define R600_CB_CNTL__CB_COLOR_CONTROL			9
-#define R600_CB_CNTL__PA_SC_AA_CONFIG			10
-#define R600_CB_CNTL__PA_SC_AA_SAMPLE_LOCS_MCTX			11
-#define R600_CB_CNTL__PA_SC_AA_SAMPLE_LOCS_8S_WD1_MCTX			12
-#define R600_CB_CNTL__CB_CLRCMP_CONTROL			13
-#define R600_CB_CNTL__CB_CLRCMP_SRC			14
-#define R600_CB_CNTL__CB_CLRCMP_DST			15
-#define R600_CB_CNTL__CB_CLRCMP_MSK			16
-#define R600_CB_CNTL__PA_SC_AA_MASK			17
-#define R600_CB_CNTL_SIZE				18
-#define R600_CB_CNTL_PM4				128
+#define R600_CB_CNTL__CB_CLEAR_RED				0
+#define R600_CB_CNTL__CB_CLEAR_GREEN				1
+#define R600_CB_CNTL__CB_CLEAR_BLUE				2
+#define R600_CB_CNTL__CB_CLEAR_ALPHA				3
+#define R600_CB_CNTL__CB_SHADER_MASK				6
+#define R600_CB_CNTL__CB_TARGET_MASK				7
+#define R600_CB_CNTL__CB_FOG_RED				10
+#define R600_CB_CNTL__CB_FOG_GREEN				11
+#define R600_CB_CNTL__CB_FOG_BLUE				12
+#define R600_CB_CNTL__CB_COLOR_CONTROL				15
+#define R600_CB_CNTL__PA_SC_AA_CONFIG				18
+#define R600_CB_CNTL__PA_SC_AA_SAMPLE_LOCS_MCTX			21
+#define R600_CB_CNTL__PA_SC_AA_SAMPLE_LOCS_8S_WD1_MCTX		22
+#define R600_CB_CNTL__CB_CLRCMP_CONTROL				25
+#define R600_CB_CNTL__CB_CLRCMP_SRC				26
+#define R600_CB_CNTL__CB_CLRCMP_DST				27
+#define R600_CB_CNTL__CB_CLRCMP_MSK				28
+#define R600_CB_CNTL__PA_SC_AA_MASK				31
 /* R600_RASTERIZER */
 #define R600_RASTERIZER__SPI_INTERP_CONTROL_0			0
-#define R600_RASTERIZER__PA_CL_CLIP_CNTL			1
-#define R600_RASTERIZER__PA_SU_SC_MODE_CNTL			2
-#define R600_RASTERIZER__PA_CL_VS_OUT_CNTL			3
-#define R600_RASTERIZER__PA_CL_NANINF_CNTL			4
-#define R600_RASTERIZER__PA_SU_POINT_SIZE			5
-#define R600_RASTERIZER__PA_SU_POINT_MINMAX			6
-#define R600_RASTERIZER__PA_SU_LINE_CNTL			7
-#define R600_RASTERIZER__PA_SC_LINE_STIPPLE			8
-#define R600_RASTERIZER__PA_SC_MPASS_PS_CNTL			9
-#define R600_RASTERIZER__PA_SC_LINE_CNTL			10
-#define R600_RASTERIZER__PA_CL_GB_VERT_CLIP_ADJ			11
-#define R600_RASTERIZER__PA_CL_GB_VERT_DISC_ADJ			12
-#define R600_RASTERIZER__PA_CL_GB_HORZ_CLIP_ADJ			13
-#define R600_RASTERIZER__PA_CL_GB_HORZ_DISC_ADJ			14
-#define R600_RASTERIZER__PA_SU_POLY_OFFSET_DB_FMT_CNTL			15
-#define R600_RASTERIZER__PA_SU_POLY_OFFSET_CLAMP			16
-#define R600_RASTERIZER__PA_SU_POLY_OFFSET_FRONT_SCALE			17
-#define R600_RASTERIZER__PA_SU_POLY_OFFSET_FRONT_OFFSET			18
-#define R600_RASTERIZER__PA_SU_POLY_OFFSET_BACK_SCALE			19
-#define R600_RASTERIZER__PA_SU_POLY_OFFSET_BACK_OFFSET			20
-#define R600_RASTERIZER_SIZE				21
-#define R600_RASTERIZER_PM4				128
+#define R600_RASTERIZER__PA_CL_CLIP_CNTL			3
+#define R600_RASTERIZER__PA_SU_SC_MODE_CNTL			4
+#define R600_RASTERIZER__PA_CL_VS_OUT_CNTL			7
+#define R600_RASTERIZER__PA_CL_NANINF_CNTL			8
+#define R600_RASTERIZER__PA_SU_POINT_SIZE			11
+#define R600_RASTERIZER__PA_SU_POINT_MINMAX			12
+#define R600_RASTERIZER__PA_SU_LINE_CNTL			13
+#define R600_RASTERIZER__PA_SC_LINE_STIPPLE			14
+#define R600_RASTERIZER__PA_SC_MPASS_PS_CNTL			17
+#define R600_RASTERIZER__PA_SC_LINE_CNTL			20
+#define R600_RASTERIZER__PA_CL_GB_VERT_CLIP_ADJ			23
+#define R600_RASTERIZER__PA_CL_GB_VERT_DISC_ADJ			24
+#define R600_RASTERIZER__PA_CL_GB_HORZ_CLIP_ADJ			25
+#define R600_RASTERIZER__PA_CL_GB_HORZ_DISC_ADJ			26
+#define R600_RASTERIZER__PA_SU_POLY_OFFSET_DB_FMT_CNTL		29
+#define R600_RASTERIZER__PA_SU_POLY_OFFSET_CLAMP		30
+#define R600_RASTERIZER__PA_SU_POLY_OFFSET_FRONT_SCALE		31
+#define R600_RASTERIZER__PA_SU_POLY_OFFSET_FRONT_OFFSET		32
+#define R600_RASTERIZER__PA_SU_POLY_OFFSET_BACK_SCALE		33
+#define R600_RASTERIZER__PA_SU_POLY_OFFSET_BACK_OFFSET		34
 /* R600_VIEWPORT */
 #define R600_VIEWPORT__PA_SC_VPORT_ZMIN_0			0
 #define R600_VIEWPORT__PA_SC_VPORT_ZMAX_0			1
-#define R600_VIEWPORT__PA_CL_VPORT_XSCALE_0			2
-#define R600_VIEWPORT__PA_CL_VPORT_YSCALE_0			3
-#define R600_VIEWPORT__PA_CL_VPORT_ZSCALE_0			4
-#define R600_VIEWPORT__PA_CL_VPORT_XOFFSET_0			5
-#define R600_VIEWPORT__PA_CL_VPORT_YOFFSET_0			6
-#define R600_VIEWPORT__PA_CL_VPORT_ZOFFSET_0			7
-#define R600_VIEWPORT__PA_CL_VTE_CNTL			8
-#define R600_VIEWPORT_SIZE				9
-#define R600_VIEWPORT_PM4				128
+#define R600_VIEWPORT__PA_CL_VPORT_XSCALE_0			4
+#define R600_VIEWPORT__PA_CL_VPORT_YSCALE_0			7
+#define R600_VIEWPORT__PA_CL_VPORT_ZSCALE_0			10
+#define R600_VIEWPORT__PA_CL_VPORT_XOFFSET_0			13
+#define R600_VIEWPORT__PA_CL_VPORT_YOFFSET_0			16
+#define R600_VIEWPORT__PA_CL_VPORT_ZOFFSET_0			19
+#define R600_VIEWPORT__PA_CL_VTE_CNTL				22
 /* R600_SCISSOR */
 #define R600_SCISSOR__PA_SC_SCREEN_SCISSOR_TL			0
 #define R600_SCISSOR__PA_SC_SCREEN_SCISSOR_BR			1
-#define R600_SCISSOR__PA_SC_WINDOW_OFFSET			2
-#define R600_SCISSOR__PA_SC_WINDOW_SCISSOR_TL			3
-#define R600_SCISSOR__PA_SC_WINDOW_SCISSOR_BR			4
-#define R600_SCISSOR__PA_SC_CLIPRECT_RULE			5
-#define R600_SCISSOR__PA_SC_CLIPRECT_0_TL			6
-#define R600_SCISSOR__PA_SC_CLIPRECT_0_BR			7
-#define R600_SCISSOR__PA_SC_CLIPRECT_1_TL			8
-#define R600_SCISSOR__PA_SC_CLIPRECT_1_BR			9
-#define R600_SCISSOR__PA_SC_CLIPRECT_2_TL			10
-#define R600_SCISSOR__PA_SC_CLIPRECT_2_BR			11
-#define R600_SCISSOR__PA_SC_CLIPRECT_3_TL			12
-#define R600_SCISSOR__PA_SC_CLIPRECT_3_BR			13
-#define R600_SCISSOR__PA_SC_EDGERULE			14
-#define R600_SCISSOR__PA_SC_GENERIC_SCISSOR_TL			15
-#define R600_SCISSOR__PA_SC_GENERIC_SCISSOR_BR			16
-#define R600_SCISSOR__PA_SC_VPORT_SCISSOR_0_TL			17
-#define R600_SCISSOR__PA_SC_VPORT_SCISSOR_0_BR			18
-#define R600_SCISSOR_SIZE				19
-#define R600_SCISSOR_PM4				128
+#define R600_SCISSOR__PA_SC_WINDOW_OFFSET			4
+#define R600_SCISSOR__PA_SC_WINDOW_SCISSOR_TL			5
+#define R600_SCISSOR__PA_SC_WINDOW_SCISSOR_BR			6
+#define R600_SCISSOR__PA_SC_CLIPRECT_RULE			7
+#define R600_SCISSOR__PA_SC_CLIPRECT_0_TL			8
+#define R600_SCISSOR__PA_SC_CLIPRECT_0_BR			9
+#define R600_SCISSOR__PA_SC_CLIPRECT_1_TL			10
+#define R600_SCISSOR__PA_SC_CLIPRECT_1_BR			11
+#define R600_SCISSOR__PA_SC_CLIPRECT_2_TL			12
+#define R600_SCISSOR__PA_SC_CLIPRECT_2_BR			13
+#define R600_SCISSOR__PA_SC_CLIPRECT_3_TL			14
+#define R600_SCISSOR__PA_SC_CLIPRECT_3_BR			15
+#define R600_SCISSOR__PA_SC_EDGERULE				16
+#define R600_SCISSOR__PA_SC_GENERIC_SCISSOR_TL			19
+#define R600_SCISSOR__PA_SC_GENERIC_SCISSOR_BR			20
+#define R600_SCISSOR__PA_SC_VPORT_SCISSOR_0_TL			23
+#define R600_SCISSOR__PA_SC_VPORT_SCISSOR_0_BR			24
 /* R600_BLEND */
-#define R600_BLEND__CB_BLEND_RED			0
-#define R600_BLEND__CB_BLEND_GREEN			1
-#define R600_BLEND__CB_BLEND_BLUE			2
-#define R600_BLEND__CB_BLEND_ALPHA			3
-#define R600_BLEND__CB_BLEND0_CONTROL			4
-#define R600_BLEND__CB_BLEND1_CONTROL			5
-#define R600_BLEND__CB_BLEND2_CONTROL			6
-#define R600_BLEND__CB_BLEND3_CONTROL			7
-#define R600_BLEND__CB_BLEND4_CONTROL			8
-#define R600_BLEND__CB_BLEND5_CONTROL			9
-#define R600_BLEND__CB_BLEND6_CONTROL			10
-#define R600_BLEND__CB_BLEND7_CONTROL			11
-#define R600_BLEND__CB_BLEND_CONTROL			12
-#define R600_BLEND_SIZE				13
-#define R600_BLEND_PM4				128
+#define R600_BLEND__CB_BLEND_RED				0
+#define R600_BLEND__CB_BLEND_GREEN				1
+#define R600_BLEND__CB_BLEND_BLUE				2
+#define R600_BLEND__CB_BLEND_ALPHA				3
+#define R600_BLEND__CB_BLEND0_CONTROL				6
+#define R600_BLEND__CB_BLEND1_CONTROL				7
+#define R600_BLEND__CB_BLEND2_CONTROL				8
+#define R600_BLEND__CB_BLEND3_CONTROL				9
+#define R600_BLEND__CB_BLEND4_CONTROL				10
+#define R600_BLEND__CB_BLEND5_CONTROL				11
+#define R600_BLEND__CB_BLEND6_CONTROL				12
+#define R600_BLEND__CB_BLEND7_CONTROL				13
+#define R600_BLEND__CB_BLEND_CONTROL				16
 /* R600_DSA */
-#define R600_DSA__DB_STENCIL_CLEAR			0
-#define R600_DSA__DB_DEPTH_CLEAR			1
-#define R600_DSA__SX_ALPHA_TEST_CONTROL			2
-#define R600_DSA__DB_STENCILREFMASK			3
-#define R600_DSA__DB_STENCILREFMASK_BF			4
-#define R600_DSA__SX_ALPHA_REF			5
-#define R600_DSA__SPI_FOG_FUNC_SCALE			6
-#define R600_DSA__SPI_FOG_FUNC_BIAS			7
-#define R600_DSA__SPI_FOG_CNTL			8
-#define R600_DSA__DB_DEPTH_CONTROL			9
-#define R600_DSA__DB_SHADER_CONTROL			10
-#define R600_DSA__DB_RENDER_CONTROL			11
-#define R600_DSA__DB_RENDER_OVERRIDE			12
-#define R600_DSA__DB_SRESULTS_COMPARE_STATE1			13
-#define R600_DSA__DB_PRELOAD_CONTROL			14
-#define R600_DSA__DB_ALPHA_TO_MASK			15
-#define R600_DSA_SIZE				16
-#define R600_DSA_PM4				128
+#define R600_DSA__DB_STENCIL_CLEAR				0
+#define R600_DSA__DB_DEPTH_CLEAR				1
+#define R600_DSA__SX_ALPHA_TEST_CONTROL				4
+#define R600_DSA__DB_STENCILREFMASK				7
+#define R600_DSA__DB_STENCILREFMASK_BF				8
+#define R600_DSA__SX_ALPHA_REF					9
+#define R600_DSA__SPI_FOG_FUNC_SCALE				12
+#define R600_DSA__SPI_FOG_FUNC_BIAS				13
+#define R600_DSA__SPI_FOG_CNTL					16
+#define R600_DSA__DB_DEPTH_CONTROL				19
+#define R600_DSA__DB_SHADER_CONTROL				22
+#define R600_DSA__DB_RENDER_CONTROL				25
+#define R600_DSA__DB_RENDER_OVERRIDE				26
+#define R600_DSA__DB_SRESULTS_COMPARE_STATE1			29
+#define R600_DSA__DB_PRELOAD_CONTROL				30
+#define R600_DSA__DB_ALPHA_TO_MASK				33
 /* R600_VS_SHADER */
 #define R600_VS_SHADER__SQ_VTX_SEMANTIC_0			0
 #define R600_VS_SHADER__SQ_VTX_SEMANTIC_1			1
@@ -457,25 +394,25 @@ struct radeon_ctx {
 #define R600_VS_SHADER__SQ_VTX_SEMANTIC_29			29
 #define R600_VS_SHADER__SQ_VTX_SEMANTIC_30			30
 #define R600_VS_SHADER__SQ_VTX_SEMANTIC_31			31
-#define R600_VS_SHADER__SPI_VS_OUT_ID_0			32
-#define R600_VS_SHADER__SPI_VS_OUT_ID_1			33
-#define R600_VS_SHADER__SPI_VS_OUT_ID_2			34
-#define R600_VS_SHADER__SPI_VS_OUT_ID_3			35
-#define R600_VS_SHADER__SPI_VS_OUT_ID_4			36
-#define R600_VS_SHADER__SPI_VS_OUT_ID_5			37
-#define R600_VS_SHADER__SPI_VS_OUT_ID_6			38
-#define R600_VS_SHADER__SPI_VS_OUT_ID_7			39
-#define R600_VS_SHADER__SPI_VS_OUT_ID_8			40
-#define R600_VS_SHADER__SPI_VS_OUT_ID_9			41
-#define R600_VS_SHADER__SPI_VS_OUT_CONFIG			42
-#define R600_VS_SHADER__SQ_PGM_START_VS			43
-#define R600_VS_SHADER__SQ_PGM_RESOURCES_VS			44
-#define R600_VS_SHADER__SQ_PGM_START_FS			45
-#define R600_VS_SHADER__SQ_PGM_RESOURCES_FS			46
-#define R600_VS_SHADER__SQ_PGM_CF_OFFSET_VS			47
-#define R600_VS_SHADER__SQ_PGM_CF_OFFSET_FS			48
-#define R600_VS_SHADER_SIZE				49
-#define R600_VS_SHADER_PM4				128
+#define R600_VS_SHADER__SPI_VS_OUT_ID_0				34
+#define R600_VS_SHADER__SPI_VS_OUT_ID_1				35
+#define R600_VS_SHADER__SPI_VS_OUT_ID_2				36
+#define R600_VS_SHADER__SPI_VS_OUT_ID_3				37
+#define R600_VS_SHADER__SPI_VS_OUT_ID_4				38
+#define R600_VS_SHADER__SPI_VS_OUT_ID_5				39
+#define R600_VS_SHADER__SPI_VS_OUT_ID_6				40
+#define R600_VS_SHADER__SPI_VS_OUT_ID_7				41
+#define R600_VS_SHADER__SPI_VS_OUT_ID_8				42
+#define R600_VS_SHADER__SPI_VS_OUT_ID_9				43
+#define R600_VS_SHADER__SPI_VS_OUT_CONFIG			46
+#define R600_VS_SHADER__SQ_PGM_START_VS				49
+#define R600_VS_SHADER__SQ_PGM_START_VS_BO_ID			51
+#define R600_VS_SHADER__SQ_PGM_RESOURCES_VS			54
+#define R600_VS_SHADER__SQ_PGM_START_FS				57
+#define R600_VS_SHADER__SQ_PGM_START_FS_BO_ID			59
+#define R600_VS_SHADER__SQ_PGM_RESOURCES_FS			62
+#define R600_VS_SHADER__SQ_PGM_CF_OFFSET_VS			65
+#define R600_VS_SHADER__SQ_PGM_CF_OFFSET_FS			68
 /* R600_PS_SHADER */
 #define R600_PS_SHADER__SPI_PS_INPUT_CNTL_0			0
 #define R600_PS_SHADER__SPI_PS_INPUT_CNTL_1			1
@@ -509,158 +446,104 @@ struct radeon_ctx {
 #define R600_PS_SHADER__SPI_PS_INPUT_CNTL_29			29
 #define R600_PS_SHADER__SPI_PS_INPUT_CNTL_30			30
 #define R600_PS_SHADER__SPI_PS_INPUT_CNTL_31			31
-#define R600_PS_SHADER__SPI_PS_IN_CONTROL_0			32
-#define R600_PS_SHADER__SPI_PS_IN_CONTROL_1			33
-#define R600_PS_SHADER__SPI_INPUT_Z			34
-#define R600_PS_SHADER__SQ_PGM_START_PS			35
-#define R600_PS_SHADER__SQ_PGM_RESOURCES_PS			36
-#define R600_PS_SHADER__SQ_PGM_EXPORTS_PS			37
-#define R600_PS_SHADER__SQ_PGM_CF_OFFSET_PS			38
-#define R600_PS_SHADER_SIZE				39
-#define R600_PS_SHADER_PM4				128
+#define R600_PS_SHADER__SPI_PS_IN_CONTROL_0			34
+#define R600_PS_SHADER__SPI_PS_IN_CONTROL_1			35
+#define R600_PS_SHADER__SPI_INPUT_Z				38
+#define R600_PS_SHADER__SQ_PGM_START_PS				41
+#define R600_PS_SHADER__SQ_PGM_START_PS_BO_ID			43
+#define R600_PS_SHADER__SQ_PGM_RESOURCES_PS			46
+#define R600_PS_SHADER__SQ_PGM_EXPORTS_PS			47
+#define R600_PS_SHADER__SQ_PGM_CF_OFFSET_PS			50
 /* R600_PS_CONSTANT */
 #define R600_PS_CONSTANT__SQ_ALU_CONSTANT0_0			0
 #define R600_PS_CONSTANT__SQ_ALU_CONSTANT1_0			1
 #define R600_PS_CONSTANT__SQ_ALU_CONSTANT2_0			2
 #define R600_PS_CONSTANT__SQ_ALU_CONSTANT3_0			3
-#define R600_PS_CONSTANT_SIZE				4
-#define R600_PS_CONSTANT_PM4				128
 /* R600_VS_CONSTANT */
 #define R600_VS_CONSTANT__SQ_ALU_CONSTANT0_256			0
 #define R600_VS_CONSTANT__SQ_ALU_CONSTANT1_256			1
 #define R600_VS_CONSTANT__SQ_ALU_CONSTANT2_256			2
 #define R600_VS_CONSTANT__SQ_ALU_CONSTANT3_256			3
-#define R600_VS_CONSTANT_SIZE				4
-#define R600_VS_CONSTANT_PM4				128
 /* R600_PS_RESOURCE */
-#define R600_PS_RESOURCE__RESOURCE0_WORD0			0
-#define R600_PS_RESOURCE__RESOURCE0_WORD1			1
-#define R600_PS_RESOURCE__RESOURCE0_WORD2			2
-#define R600_PS_RESOURCE__RESOURCE0_WORD3			3
-#define R600_PS_RESOURCE__RESOURCE0_WORD4			4
-#define R600_PS_RESOURCE__RESOURCE0_WORD5			5
-#define R600_PS_RESOURCE__RESOURCE0_WORD6			6
-#define R600_PS_RESOURCE_SIZE				7
-#define R600_PS_RESOURCE_PM4				128
-/* R600_VS_RESOURCE */
-#define R600_VS_RESOURCE__RESOURCE160_WORD0			0
-#define R600_VS_RESOURCE__RESOURCE160_WORD1			1
-#define R600_VS_RESOURCE__RESOURCE160_WORD2			2
-#define R600_VS_RESOURCE__RESOURCE160_WORD3			3
-#define R600_VS_RESOURCE__RESOURCE160_WORD4			4
-#define R600_VS_RESOURCE__RESOURCE160_WORD5			5
-#define R600_VS_RESOURCE__RESOURCE160_WORD6			6
-#define R600_VS_RESOURCE_SIZE				7
-#define R600_VS_RESOURCE_PM4				128
-/* R600_FS_RESOURCE */
-#define R600_FS_RESOURCE__RESOURCE320_WORD0			0
-#define R600_FS_RESOURCE__RESOURCE320_WORD1			1
-#define R600_FS_RESOURCE__RESOURCE320_WORD2			2
-#define R600_FS_RESOURCE__RESOURCE320_WORD3			3
-#define R600_FS_RESOURCE__RESOURCE320_WORD4			4
-#define R600_FS_RESOURCE__RESOURCE320_WORD5			5
-#define R600_FS_RESOURCE__RESOURCE320_WORD6			6
-#define R600_FS_RESOURCE_SIZE				7
-#define R600_FS_RESOURCE_PM4				128
-/* R600_GS_RESOURCE */
-#define R600_GS_RESOURCE__RESOURCE336_WORD0			0
-#define R600_GS_RESOURCE__RESOURCE336_WORD1			1
-#define R600_GS_RESOURCE__RESOURCE336_WORD2			2
-#define R600_GS_RESOURCE__RESOURCE336_WORD3			3
-#define R600_GS_RESOURCE__RESOURCE336_WORD4			4
-#define R600_GS_RESOURCE__RESOURCE336_WORD5			5
-#define R600_GS_RESOURCE__RESOURCE336_WORD6			6
-#define R600_GS_RESOURCE_SIZE				7
-#define R600_GS_RESOURCE_PM4				128
+#define R600_RESOURCE__RESOURCE_WORD0				0
+#define R600_RESOURCE__RESOURCE_WORD1				1
+#define R600_RESOURCE__RESOURCE_WORD2				2
+#define R600_RESOURCE__RESOURCE_WORD3				3
+#define R600_RESOURCE__RESOURCE_WORD4				4
+#define R600_RESOURCE__RESOURCE_WORD5				5
+#define R600_RESOURCE__RESOURCE_WORD6				6
+#define R600_RESOURCE__RESOURCE_BO0_ID				8
+#define R600_RESOURCE__RESOURCE_BO1_ID				10
 /* R600_PS_SAMPLER */
 #define R600_PS_SAMPLER__SQ_TEX_SAMPLER_WORD0_0			0
 #define R600_PS_SAMPLER__SQ_TEX_SAMPLER_WORD1_0			1
 #define R600_PS_SAMPLER__SQ_TEX_SAMPLER_WORD2_0			2
-#define R600_PS_SAMPLER_SIZE				3
-#define R600_PS_SAMPLER_PM4				128
 /* R600_VS_SAMPLER */
-#define R600_VS_SAMPLER__SQ_TEX_SAMPLER_WORD0_18			0
-#define R600_VS_SAMPLER__SQ_TEX_SAMPLER_WORD1_18			1
-#define R600_VS_SAMPLER__SQ_TEX_SAMPLER_WORD2_18			2
-#define R600_VS_SAMPLER_SIZE				3
-#define R600_VS_SAMPLER_PM4				128
+#define R600_VS_SAMPLER__SQ_TEX_SAMPLER_WORD0_18		0
+#define R600_VS_SAMPLER__SQ_TEX_SAMPLER_WORD1_18		1
+#define R600_VS_SAMPLER__SQ_TEX_SAMPLER_WORD2_18		2
 /* R600_GS_SAMPLER */
-#define R600_GS_SAMPLER__SQ_TEX_SAMPLER_WORD0_36			0
-#define R600_GS_SAMPLER__SQ_TEX_SAMPLER_WORD1_36			1
-#define R600_GS_SAMPLER__SQ_TEX_SAMPLER_WORD2_36			2
-#define R600_GS_SAMPLER_SIZE				3
-#define R600_GS_SAMPLER_PM4				128
+#define R600_GS_SAMPLER__SQ_TEX_SAMPLER_WORD0_36		0
+#define R600_GS_SAMPLER__SQ_TEX_SAMPLER_WORD1_36		1
+#define R600_GS_SAMPLER__SQ_TEX_SAMPLER_WORD2_36		2
 /* R600_PS_SAMPLER_BORDER */
-#define R600_PS_SAMPLER_BORDER__TD_PS_SAMPLER0_BORDER_RED			0
-#define R600_PS_SAMPLER_BORDER__TD_PS_SAMPLER0_BORDER_GREEN			1
-#define R600_PS_SAMPLER_BORDER__TD_PS_SAMPLER0_BORDER_BLUE			2
-#define R600_PS_SAMPLER_BORDER__TD_PS_SAMPLER0_BORDER_ALPHA			3
-#define R600_PS_SAMPLER_BORDER_SIZE				4
-#define R600_PS_SAMPLER_BORDER_PM4				128
+#define R600_PS_SAMPLER_BORDER__TD_PS_SAMPLER0_BORDER_RED	0
+#define R600_PS_SAMPLER_BORDER__TD_PS_SAMPLER0_BORDER_GREEN	1
+#define R600_PS_SAMPLER_BORDER__TD_PS_SAMPLER0_BORDER_BLUE	2
+#define R600_PS_SAMPLER_BORDER__TD_PS_SAMPLER0_BORDER_ALPHA	3
 /* R600_VS_SAMPLER_BORDER */
-#define R600_VS_SAMPLER_BORDER__TD_VS_SAMPLER0_BORDER_RED			0
-#define R600_VS_SAMPLER_BORDER__TD_VS_SAMPLER0_BORDER_GREEN			1
-#define R600_VS_SAMPLER_BORDER__TD_VS_SAMPLER0_BORDER_BLUE			2
-#define R600_VS_SAMPLER_BORDER__TD_VS_SAMPLER0_BORDER_ALPHA			3
-#define R600_VS_SAMPLER_BORDER_SIZE				4
-#define R600_VS_SAMPLER_BORDER_PM4				128
+#define R600_VS_SAMPLER_BORDER__TD_VS_SAMPLER0_BORDER_RED	0
+#define R600_VS_SAMPLER_BORDER__TD_VS_SAMPLER0_BORDER_GREEN	1
+#define R600_VS_SAMPLER_BORDER__TD_VS_SAMPLER0_BORDER_BLUE	2
+#define R600_VS_SAMPLER_BORDER__TD_VS_SAMPLER0_BORDER_ALPHA	3
 /* R600_GS_SAMPLER_BORDER */
-#define R600_GS_SAMPLER_BORDER__TD_GS_SAMPLER0_BORDER_RED			0
-#define R600_GS_SAMPLER_BORDER__TD_GS_SAMPLER0_BORDER_GREEN			1
-#define R600_GS_SAMPLER_BORDER__TD_GS_SAMPLER0_BORDER_BLUE			2
-#define R600_GS_SAMPLER_BORDER__TD_GS_SAMPLER0_BORDER_ALPHA			3
-#define R600_GS_SAMPLER_BORDER_SIZE				4
-#define R600_GS_SAMPLER_BORDER_PM4				128
-/* R600_CB0 */
-#define R600_CB0__CB_COLOR0_BASE			0
-#define R600_CB0__CB_COLOR0_INFO			1
-#define R600_CB0__CB_COLOR0_SIZE			2
-#define R600_CB0__CB_COLOR0_VIEW			3
-#define R600_CB0__CB_COLOR0_FRAG			4
-#define R600_CB0__CB_COLOR0_TILE			5
-#define R600_CB0__CB_COLOR0_MASK			6
-#define R600_CB0_SIZE				7
-#define R600_CB0_PM4				128
+#define R600_GS_SAMPLER_BORDER__TD_GS_SAMPLER0_BORDER_RED	0
+#define R600_GS_SAMPLER_BORDER__TD_GS_SAMPLER0_BORDER_GREEN	1
+#define R600_GS_SAMPLER_BORDER__TD_GS_SAMPLER0_BORDER_BLUE	2
+#define R600_GS_SAMPLER_BORDER__TD_GS_SAMPLER0_BORDER_ALPHA	3
+/* R600_CB */
+#define R600_CB__CB_COLOR0_BASE					0
+#define R600_CB__CB_COLOR0_BASE_BO_ID				2
+#define R600_CB__CB_COLOR0_INFO					5
+#define R600_CB__CB_COLOR0_SIZE					8
+#define R600_CB__CB_COLOR0_VIEW					11
+#define R600_CB__CB_COLOR0_FRAG					14
+#define R600_CB__CB_COLOR0_FRAG_BO_ID				16
+#define R600_CB__CB_COLOR0_TILE					19
+#define R600_CB__CB_COLOR0_TILE_BO_ID				21
+#define R600_CB__CB_COLOR0_MASK					24
 /* R600_DB */
-#define R600_DB__DB_DEPTH_BASE			0
-#define R600_DB__DB_DEPTH_SIZE			1
-#define R600_DB__DB_DEPTH_VIEW			2
-#define R600_DB__DB_DEPTH_INFO			3
-#define R600_DB__DB_HTILE_SURFACE			4
-#define R600_DB__DB_PREFETCH_LIMIT			5
-#define R600_DB_SIZE				6
-#define R600_DB_PM4				128
+#define R600_DB__DB_DEPTH_BASE					0
+#define R600_DB__DB_DEPTH_BASE_BO_ID				2
+#define R600_DB__DB_DEPTH_SIZE					5
+#define R600_DB__DB_DEPTH_VIEW					6
+#define R600_DB__DB_DEPTH_INFO					9
+#define R600_DB__DB_HTILE_SURFACE				12
+#define R600_DB__DB_PREFETCH_LIMIT				15
 /* R600_VGT */
-#define R600_VGT__VGT_PRIMITIVE_TYPE			0
-#define R600_VGT__VGT_MAX_VTX_INDX			1
-#define R600_VGT__VGT_MIN_VTX_INDX			2
-#define R600_VGT__VGT_INDX_OFFSET			3
-#define R600_VGT__VGT_MULTI_PRIM_IB_RESET_INDX			4
-#define R600_VGT__VGT_DMA_INDEX_TYPE			5
-#define R600_VGT__VGT_PRIMITIVEID_EN			6
-#define R600_VGT__VGT_DMA_NUM_INSTANCES			7
-#define R600_VGT__VGT_MULTI_PRIM_IB_RESET_EN			8
-#define R600_VGT__VGT_INSTANCE_STEP_RATE_0			9
-#define R600_VGT__VGT_INSTANCE_STEP_RATE_1			10
-#define R600_VGT_SIZE				11
-#define R600_VGT_PM4				128
+#define R600_VGT__VGT_MAX_VTX_INDX				0
+#define R600_VGT__VGT_MIN_VTX_INDX				1
+#define R600_VGT__VGT_INDX_OFFSET				2
+#define R600_VGT__VGT_MULTI_PRIM_IB_RESET_INDX			3
+#define R600_VGT__VGT_PRIMITIVE_TYPE				6
+#define R600_VGT__VGT_DMA_INDEX_TYPE				8
+#define R600_VGT__VGT_DMA_NUM_INSTANCES				10
+/* R600_DRAW_AUTO */
+#define R600_DRAW_AUTO__VGT_NUM_INDICES				0
+#define R600_DRAW_AUTO__VGT_DRAW_INITIATOR			1
 /* R600_DRAW */
-#define R600_DRAW__VGT_NUM_INDICES			0
-#define R600_DRAW__VGT_DMA_BASE_HI			1
-#define R600_DRAW__VGT_DMA_BASE			2
-#define R600_DRAW__VGT_DRAW_INITIATOR			3
-#define R600_DRAW_SIZE				4
-#define R600_DRAW_PM4				128
-/* R600_CLIP */
-#define R600_CLIP__PA_CL_UCP_X_0  0
-#define R600_CLIP__PA_CL_UCP_Y_0  1
-#define R600_CLIP__PA_CL_UCP_Z_0  2
-#define R600_CLIP__PA_CL_UCP_W_0  3
-#define R600_CLIP_SIZE  4
-#define R600_CLIP_PM4 128
+#define R600_DRAW__VGT_DMA_BASE					0
+#define R600_DRAW__VGT_DMA_BASE_HI				1
+#define R600_DRAW__VGT_NUM_INDICES				2
+#define R600_DRAW__VGT_DRAW_INITIATOR				3
+#define R600_DRAW__INDICES_BO_ID				5
+/* R600_UCP */
+#define R600_UCP__PA_CL_UCP_X_0					0
+#define R600_UCP__PA_CL_UCP_Y_0					1
+#define R600_UCP__PA_CL_UCP_Z_0					2
+#define R600_UCP__PA_CL_UCP_W_0					3
 /* R600 QUERY BEGIN/END */
-#define R600_QUERY__OFFSET			0
-#define R600_QUERY_SIZE				1
-#define R600_QUERY_PM4				128
+#define R600_QUERY__OFFSET					0
+#define R600_QUERY__BO_ID					3
 
 #endif