r600g: build packet header once

Build packet header once and allow to add fake register support so
we can handle things like indexed set of register (evergreen sampler
border registers for instance.

Signed-off-by: Jerome Glisse <jglisse@redhat.com>

4 files changed, 178 insertions, 85 deletions

diff --git a/src/gallium/drivers/r600/evergreend.h b/src/gallium/drivers/r600/evergreend.h
index 0a0a91eea2..486cb29005 100644
--- a/src/gallium/drivers/r600/evergreend.h
+++ b/src/gallium/drivers/r600/evergreend.h
@@ -1113,6 +1113,22 @@
 #define   S_03000C_DST_SEL_W(x)                        (((x) & 0x7) << 12)
 #define   G_03000C_DST_SEL_W(x)                        (((x) >> 12) & 0x7)
 
+#define R_00A400_TD_PS_SAMPLER0_BORDER_INDEX         0x00A400
+#define R_00A404_TD_PS_SAMPLER0_BORDER_RED           0x00A404
+#define R_00A408_TD_PS_SAMPLER0_BORDER_GREEN         0x00A408
+#define R_00A40C_TD_PS_SAMPLER0_BORDER_BLUE          0x00A40C
+#define R_00A410_TD_PS_SAMPLER0_BORDER_ALPHA         0x00A410
+#define R_00A414_TD_VS_SAMPLER0_BORDER_INDEX         0x00A414
+#define R_00A418_TD_VS_SAMPLER0_BORDER_RED           0x00A418
+#define R_00A41C_TD_VS_SAMPLER0_BORDER_GREEN         0x00A41C
+#define R_00A420_TD_VS_SAMPLER0_BORDER_BLUE          0x00A420
+#define R_00A424_TD_VS_SAMPLER0_BORDER_ALPHA         0x00A424
+#define R_00A428_TD_GS_SAMPLER0_BORDER_INDEX         0x00A428
+#define R_00A42C_TD_GS_SAMPLER0_BORDER_RED           0x00A42C
+#define R_00A430_TD_GS_SAMPLER0_BORDER_GREEN         0x00A430
+#define R_00A434_TD_GS_SAMPLER0_BORDER_BLUE          0x00A434
+#define R_00A438_TD_GS_SAMPLER0_BORDER_ALPHA         0x00A438
+
 #define R_03C000_SQ_TEX_SAMPLER_WORD0_0              0x03C000
 #define   S_03C000_CLAMP_X(x)                          (((x) & 0x7) << 0)
 #define   G_03C000_CLAMP_X(x)                          (((x) >> 0) & 0x7)
diff --git a/src/gallium/drivers/r600/r600.h b/src/gallium/drivers/r600/r600.h
index 44fae4bcef..b8c74675e6 100644
--- a/src/gallium/drivers/r600/r600.h
+++ b/src/gallium/drivers/r600/r600.h
@@ -137,6 +137,7 @@ enum evergreen_group_id {
 	EVERGREEN_GROUP_CTL_CONST,
 	EVERGREEN_GROUP_LOOP_CONST,
 	EVERGREEN_GROUP_BOOL_CONST,
+	EVERGREEN_GROUP_SAMPLER_BORDER,
 	EVERGREEN_NGROUPS
 };
 
@@ -183,6 +184,7 @@ struct r600_group_block {
 	unsigned		pm4_ndwords;
 	unsigned		nbo;
 	unsigned		nreg;
+	u32			*reg;
 	u32			pm4[R600_BLOCK_MAX_REG];
 	unsigned		pm4_bo_index[R600_BLOCK_MAX_REG];
 	struct r600_block_reloc	reloc[R600_BLOCK_MAX_BO];
diff --git a/src/gallium/winsys/r600/drm/evergreen_state.c b/src/gallium/winsys/r600/drm/evergreen_state.c
index b9d333060f..c2455e3a49 100644
--- a/src/gallium/winsys/r600/drm/evergreen_state.c
+++ b/src/gallium/winsys/r600/drm/evergreen_state.c
@@ -53,13 +53,13 @@ struct radeon_ws_bo {
 struct radeon_bo *radeon_bo_pb_get_bo(struct pb_buffer *_buf);
 
 struct radeon_bo *r600_context_reg_bo(struct r600_context *ctx, unsigned group_id, unsigned offset);
-void r600_context_group_emit_dirty(struct r600_context *ctx, struct r600_group *group, unsigned opcode);
+void r600_context_group_emit_dirty(struct r600_context *ctx, struct r600_group *group);
 void r600_context_bo_reloc(struct r600_context *ctx, u32 *pm4, struct radeon_bo *bo);
-int r600_context_add_block(struct r600_context *ctx, const struct r600_reg *reg, unsigned nreg);
+int r600_context_add_block(struct r600_context *ctx, const struct r600_reg *reg, unsigned nreg, unsigned opcode);
 int r600_group_init(struct r600_group *group, unsigned start_offset, unsigned end_offset);
 
 #define GROUP_FORCE_NEW_BLOCK	0
-static const struct r600_reg evergreen_reg_list[] = {
+static const struct r600_reg evergreen_config_reg_list[] = {
 	{0, 0, R_008958_VGT_PRIMITIVE_TYPE},
 	{0, 0, R_008A14_PA_CL_ENHANCE},
 	{0, 0, R_008C00_SQ_CONFIG},
@@ -74,6 +74,9 @@ static const struct r600_reg evergreen_reg_list[] = {
 	{0, 0, R_008D8C_SQ_DYN_GPR_CNTL_PS_FLUSH_REQ},
 	{0, 0, R_009100_SPI_CONFIG_CNTL},
 	{0, 0, R_00913C_SPI_CONFIG_CNTL_1},
+};
+
+static const struct r600_reg evergreen_context_reg_list[] = {
 	{0, 0, R_028000_DB_RENDER_CONTROL},
 	{0, 0, R_028004_DB_COUNT_CONTROL},
 	{0, 0, R_028008_DB_DEPTH_VIEW},
@@ -450,7 +453,7 @@ static int evergreen_state_resource_init(struct r600_context *ctx, u32 offset)
 	for (int i = 0; i < nreg; i++) {
 		r600_shader_resource[i].offset += offset;
 	}
-	return r600_context_add_block(ctx, r600_shader_resource, nreg);
+	return r600_context_add_block(ctx, r600_shader_resource, nreg, PKT3_SET_RESOURCE);
 }
 
 /* SHADER SAMPLER R600/R700 */
@@ -466,7 +469,39 @@ static int r600_state_sampler_init(struct r600_context *ctx, u32 offset)
 	for (int i = 0; i < nreg; i++) {
 		r600_shader_sampler[i].offset += offset;
 	}
-	return r600_context_add_block(ctx, r600_shader_sampler, nreg);
+	return r600_context_add_block(ctx, r600_shader_sampler, nreg, PKT3_SET_SAMPLER);
+}
+
+/* SHADER SAMPLER BORDER R600/R700 */
+static int evergreen_state_sampler_border_init(struct r600_context *ctx, u32 offset, unsigned id)
+{
+	struct r600_reg r600_shader_sampler_border[] = {
+		{0, 0, R_00A400_TD_PS_SAMPLER0_BORDER_INDEX},
+		{0, 0, R_00A404_TD_PS_SAMPLER0_BORDER_RED},
+		{0, 0, R_00A408_TD_PS_SAMPLER0_BORDER_GREEN},
+		{0, 0, R_00A40C_TD_PS_SAMPLER0_BORDER_BLUE},
+		{0, 0, R_00A410_TD_PS_SAMPLER0_BORDER_ALPHA},
+	};
+	unsigned nreg = sizeof(r600_shader_sampler_border)/sizeof(struct r600_reg);
+	unsigned fake_offset = (offset - R_00A400_TD_PS_SAMPLER0_BORDER_INDEX) * 0x10 + 0x40000 + id * 0x1C;
+	struct r600_group_block *block;
+	struct r600_group *group;
+	int r;
+
+	for (int i = 0; i < nreg; i++) {
+		r600_shader_sampler_border[i].offset -= R_00A400_TD_PS_SAMPLER0_BORDER_INDEX;
+		r600_shader_sampler_border[i].offset += fake_offset;
+	}
+	r = r600_context_add_block(ctx, r600_shader_sampler_border, nreg, PKT3_SET_CONFIG_REG);
+	if (r) {
+		return r;
+	}
+	/* set proper offset */
+	group = &ctx->groups[EVERGREEN_GROUP_SAMPLER_BORDER];
+	id = group->offset_block_id[((fake_offset - group->start_offset) >> 2)];
+	block = &group->blocks[id];
+	block->pm4[1] = (offset - EVERGREEN_CONFIG_REG_OFFSET) >> 2;
+	return 0;
 }
 
 int evergreen_context_init(struct r600_context *ctx, struct radeon *radeon)
@@ -502,10 +537,22 @@ int evergreen_context_init(struct r600_context *ctx, struct radeon *radeon)
 	if (r) {
 		goto out_err;
 	}
+	/* we use unassigned range of GPU reg to fake border color register */
+	r = r600_group_init(&ctx->groups[EVERGREEN_GROUP_SAMPLER_BORDER], 0x40000, 0x41000);
+	if (r) {
+		goto out_err;
+	}
 	ctx->ngroups = EVERGREEN_NGROUPS;
 
 	/* add blocks */
-	r = r600_context_add_block(ctx, evergreen_reg_list, sizeof(evergreen_reg_list)/sizeof(struct r600_reg));
+	r = r600_context_add_block(ctx, evergreen_config_reg_list,
+				 sizeof(evergreen_config_reg_list)/sizeof(struct r600_reg),
+					PKT3_SET_CONFIG_REG);
+	if (r)
+		goto out_err;
+	r = r600_context_add_block(ctx, evergreen_context_reg_list,
+				 sizeof(evergreen_context_reg_list)/sizeof(struct r600_reg),
+					PKT3_SET_CONTEXT_REG);
 	if (r)
 		goto out_err;
 
@@ -521,6 +568,18 @@ int evergreen_context_init(struct r600_context *ctx, struct radeon *radeon)
 		if (r)
 			goto out_err;
 	}
+	/* PS SAMPLER BORDER */
+	for (int j = 0; j < 18; j++) {
+		r = evergreen_state_sampler_border_init(ctx, R_00A400_TD_PS_SAMPLER0_BORDER_INDEX, j);
+		if (r)
+			goto out_err;
+	}
+	/* VS SAMPLER BORDER */
+	for (int j = 0; j < 18; j++) {
+		r = evergreen_state_sampler_border_init(ctx, R_00A414_TD_VS_SAMPLER0_BORDER_INDEX, j);
+		if (r)
+			goto out_err;
+	}
 	/* PS RESOURCE */
 	for (int j = 0, offset = 0; j < 176; j++, offset += 0x20) {
 		r = evergreen_state_resource_init(ctx, offset);
@@ -566,14 +625,14 @@ static inline void evergreen_context_pipe_state_set_resource(struct r600_context
 	offset -= ctx->groups[EVERGREEN_GROUP_RESOURCE].start_offset;
 	id = ctx->groups[EVERGREEN_GROUP_RESOURCE].offset_block_id[offset >> 2];
 	block = &ctx->groups[EVERGREEN_GROUP_RESOURCE].blocks[id];
-	block->pm4[0] = state->regs[0].value;
-	block->pm4[1] = state->regs[1].value;
-	block->pm4[2] = state->regs[2].value;
-	block->pm4[3] = state->regs[3].value;
-	block->pm4[4] = state->regs[4].value;
-	block->pm4[5] = state->regs[5].value;
-	block->pm4[6] = state->regs[6].value;
-	block->pm4[7] = state->regs[7].value;
+	block->reg[0] = state->regs[0].value;
+	block->reg[1] = state->regs[1].value;
+	block->reg[2] = state->regs[2].value;
+	block->reg[3] = state->regs[3].value;
+	block->reg[4] = state->regs[4].value;
+	block->reg[5] = state->regs[5].value;
+	block->reg[6] = state->regs[6].value;
+	block->reg[7] = state->regs[7].value;
 	radeon_ws_bo_reference(ctx->radeon, &block->reloc[1].bo, NULL);
 	radeon_ws_bo_reference(ctx->radeon , &block->reloc[2].bo, NULL);
 	if (state->regs[0].bo) {
@@ -589,7 +648,7 @@ static inline void evergreen_context_pipe_state_set_resource(struct r600_context
 	}
 	block->status |= R600_BLOCK_STATUS_ENABLED;
 	block->status |= R600_BLOCK_STATUS_DIRTY;
-	ctx->pm4_dirty_cdwords += 2 + block->pm4_ndwords;
+	ctx->pm4_dirty_cdwords += block->pm4_ndwords;
 }
 
 void evergreen_context_pipe_state_set_ps_resource(struct r600_context *ctx, struct r600_pipe_state *state, unsigned rid)
@@ -614,29 +673,29 @@ static inline void evergreen_context_pipe_state_set_sampler(struct r600_context
 	offset -= ctx->groups[EVERGREEN_GROUP_SAMPLER].start_offset;
 	id = ctx->groups[EVERGREEN_GROUP_SAMPLER].offset_block_id[offset >> 2];
 	block = &ctx->groups[EVERGREEN_GROUP_SAMPLER].blocks[id];
-	block->pm4[0] = state->regs[0].value;
-	block->pm4[1] = state->regs[1].value;
-	block->pm4[2] = state->regs[2].value;
+	block->reg[0] = state->regs[0].value;
+	block->reg[1] = state->regs[1].value;
+	block->reg[2] = state->regs[2].value;
 	block->status |= R600_BLOCK_STATUS_ENABLED;
 	block->status |= R600_BLOCK_STATUS_DIRTY;
-	ctx->pm4_dirty_cdwords += 2 + block->pm4_ndwords;
+	ctx->pm4_dirty_cdwords += block->pm4_ndwords;
 }
 
-static inline void evergreen_context_pipe_state_set_sampler_border(struct r600_context *ctx, struct r600_pipe_state *state, unsigned offset)
+static inline void evergreen_context_pipe_state_set_sampler_border(struct r600_context *ctx, struct r600_pipe_state *state, unsigned offset, unsigned id)
 {
 	struct r600_group_block *block;
-	unsigned id;
-
-	offset -= ctx->groups[EVERGREEN_GROUP_CONFIG].start_offset;
-	id = ctx->groups[EVERGREEN_GROUP_CONFIG].offset_block_id[offset >> 2];
-	block = &ctx->groups[EVERGREEN_GROUP_CONFIG].blocks[id];
-	block->pm4[0] = state->regs[3].value;
-	block->pm4[1] = state->regs[4].value;
-	block->pm4[2] = state->regs[5].value;
-	block->pm4[3] = state->regs[6].value;
+	unsigned fake_offset = (offset - R_00A400_TD_PS_SAMPLER0_BORDER_INDEX) * 0x10 + 0x40000 + id * 0x1C;
+
+	fake_offset -= ctx->groups[EVERGREEN_GROUP_SAMPLER_BORDER].start_offset;
+	id = ctx->groups[EVERGREEN_GROUP_SAMPLER_BORDER].offset_block_id[fake_offset >> 2];
+	block = &ctx->groups[EVERGREEN_GROUP_SAMPLER_BORDER].blocks[id];
+	block->reg[0] = id;
+	block->reg[1] = state->regs[3].value;
+	block->reg[2] = state->regs[4].value;
+	block->reg[3] = state->regs[5].value;
 	block->status |= R600_BLOCK_STATUS_ENABLED;
 	block->status |= R600_BLOCK_STATUS_DIRTY;
-	ctx->pm4_dirty_cdwords += 2 + block->pm4_ndwords;
+	ctx->pm4_dirty_cdwords += block->pm4_ndwords;
 }
 
 void evergreen_context_pipe_state_set_ps_sampler(struct r600_context *ctx, struct r600_pipe_state *state, unsigned id)
@@ -646,8 +705,7 @@ void evergreen_context_pipe_state_set_ps_sampler(struct r600_context *ctx, struc
 	offset = 0x0003C000 + id * 0xc;
 	evergreen_context_pipe_state_set_sampler(ctx, state, offset);
 	if (state->nregs > 3) {
-		offset = 0x0000A400 + id * 0x10;
-		//		evergreen_context_pipe_state_set_sampler_border(ctx, state, offset);
+		evergreen_context_pipe_state_set_sampler_border(ctx, state, R_00A400_TD_PS_SAMPLER0_BORDER_INDEX, id);
 	}
 }
 
@@ -658,8 +716,7 @@ void evergreen_context_pipe_state_set_vs_sampler(struct r600_context *ctx, struc
 	offset = 0x0003C0D8 + id * 0xc;
 	evergreen_context_pipe_state_set_sampler(ctx, state, offset);
 	if (state->nregs > 3) {
-		offset = 0x0000A600 + id * 0x10;
-		//		evergreen_context_pipe_state_set_sampler_border(ctx, state, offset);
+		evergreen_context_pipe_state_set_sampler_border(ctx, state, R_00A414_TD_VS_SAMPLER0_BORDER_INDEX, id);
 	}
 }
 
@@ -719,10 +776,11 @@ void evergreen_context_draw(struct r600_context *ctx, const struct r600_draw *dr
 	}
 
 	/* enough room to copy packet */
-	r600_context_group_emit_dirty(ctx, &ctx->groups[EVERGREEN_GROUP_CONFIG], PKT3_SET_CONFIG_REG);
-	r600_context_group_emit_dirty(ctx, &ctx->groups[EVERGREEN_GROUP_CONTEXT], PKT3_SET_CONTEXT_REG);
-	r600_context_group_emit_dirty(ctx, &ctx->groups[EVERGREEN_GROUP_RESOURCE], PKT3_SET_RESOURCE);
-	r600_context_group_emit_dirty(ctx, &ctx->groups[EVERGREEN_GROUP_SAMPLER], PKT3_SET_SAMPLER);
+	r600_context_group_emit_dirty(ctx, &ctx->groups[EVERGREEN_GROUP_CONFIG]);
+	r600_context_group_emit_dirty(ctx, &ctx->groups[EVERGREEN_GROUP_CONTEXT]);
+	r600_context_group_emit_dirty(ctx, &ctx->groups[EVERGREEN_GROUP_RESOURCE]);
+	r600_context_group_emit_dirty(ctx, &ctx->groups[EVERGREEN_GROUP_SAMPLER]);
+	r600_context_group_emit_dirty(ctx, &ctx->groups[EVERGREEN_GROUP_SAMPLER_BORDER]);
 
 	/* draw packet */
 	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_INDEX_TYPE, 0);
@@ -773,14 +831,14 @@ static inline void evergreen_resource_set(struct r600_context *ctx, struct r600_
 	offset -= ctx->groups[EVERGREEN_GROUP_RESOURCE].start_offset;
 	id = ctx->groups[EVERGREEN_GROUP_RESOURCE].offset_block_id[offset >> 2];
 	block = &ctx->groups[EVERGREEN_GROUP_RESOURCE].blocks[id];
-	block->pm4[0] = state->regs[0].value;
-	block->pm4[1] = state->regs[1].value;
-	block->pm4[2] = state->regs[2].value;
-	block->pm4[3] = state->regs[3].value;
-	block->pm4[4] = state->regs[4].value;
-	block->pm4[5] = state->regs[5].value;
-	block->pm4[6] = state->regs[6].value;
-	block->pm4[7] = state->regs[7].value;
+	block->reg[0] = state->regs[0].value;
+	block->reg[1] = state->regs[1].value;
+	block->reg[2] = state->regs[2].value;
+	block->reg[3] = state->regs[3].value;
+	block->reg[4] = state->regs[4].value;
+	block->reg[5] = state->regs[5].value;
+	block->reg[6] = state->regs[6].value;
+	block->reg[7] = state->regs[7].value;
 	radeon_ws_bo_reference(ctx->radeon, &block->reloc[1].bo, NULL);
 	radeon_ws_bo_reference(ctx->radeon , &block->reloc[2].bo, NULL);
 	if (state->regs[0].bo) {
@@ -796,7 +854,7 @@ static inline void evergreen_resource_set(struct r600_context *ctx, struct r600_
 	}
 	block->status |= R600_BLOCK_STATUS_ENABLED;
 	block->status |= R600_BLOCK_STATUS_DIRTY;
-	ctx->pm4_dirty_cdwords += 2 + block->pm4_ndwords;
+	ctx->pm4_dirty_cdwords += block->pm4_ndwords;
 }
 
 void evergreen_ps_resource_set(struct r600_context *ctx, struct r600_pipe_state *state, unsigned rid)
diff --git a/src/gallium/winsys/r600/drm/r600_state2.c b/src/gallium/winsys/r600/drm/r600_state2.c
index d60c37fc90..c4b45e565c 100644
--- a/src/gallium/winsys/r600/drm/r600_state2.c
+++ b/src/gallium/winsys/r600/drm/r600_state2.c
@@ -75,7 +75,7 @@ static int r600_group_id_register_offset(struct r600_context *ctx, unsigned offs
 	return -1;
 }
 
-int r600_context_add_block(struct r600_context *ctx, const struct r600_reg *reg, unsigned nreg)
+int r600_context_add_block(struct r600_context *ctx, const struct r600_reg *reg, unsigned nreg, unsigned opcode)
 {
 	struct r600_group_block *block, *tmp;
 	struct r600_group *group;
@@ -109,6 +109,12 @@ int r600_context_add_block(struct r600_context *ctx, const struct r600_reg *reg,
 		if (tmp == NULL) {
 			return -ENOMEM;
 		}
+		/* update reg pointer */
+		if (tmp != group->blocks) {
+			for (int j = 0; j < group->nblocks; j++) {
+				tmp[j].reg = &tmp[j].pm4[2];
+			}
+		}
 		group->blocks = tmp;
 		block = &group->blocks[group->nblocks++];
 		for (int j = 0; j < n; j++) {
@@ -118,7 +124,10 @@ int r600_context_add_block(struct r600_context *ctx, const struct r600_reg *reg,
 		/* initialize block */
 		memset(block, 0, sizeof(struct r600_group_block));
 		block->start_offset = reg[i].offset;
-		block->pm4_ndwords = n;
+		block->pm4[block->pm4_ndwords++] = PKT3(opcode, n);
+		block->pm4[block->pm4_ndwords++] = (block->start_offset - group->start_offset) >> 2;
+		block->reg = &block->pm4[block->pm4_ndwords];
+		block->pm4_ndwords += n;
 		block->nreg = n;
 		for (j = 0; j < n; j++) {
 			if (reg[i+j].need_bo) {
@@ -168,7 +177,8 @@ static void r600_group_fini(struct r600_group *group)
 }
 
 /* R600/R700 configuration */
-static const struct r600_reg r600_reg_list[] = {
+static const struct r600_reg r600_config_reg_list[] = {
+	{0, 0, R_008958_VGT_PRIMITIVE_TYPE},
 	{0, 0, R_008C00_SQ_CONFIG},
 	{0, 0, R_008C04_SQ_GPR_RESOURCE_MGMT_1},
 	{0, 0, R_008C08_SQ_GPR_RESOURCE_MGMT_2},
@@ -180,6 +190,9 @@ static const struct r600_reg r600_reg_list[] = {
 	{0, 0, R_009714_VC_ENHANCE},
 	{0, 0, R_009830_DB_DEBUG},
 	{0, 0, R_009838_DB_WATERMARKS},
+};
+
+static const struct r600_reg r600_context_reg_list[] = {
 	{0, 0, R_028350_SX_MISC},
 	{0, 0, R_0286C8_SPI_THREAD_GROUPING},
 	{0, 0, R_0288A8_SQ_ESGS_RING_ITEMSIZE},
@@ -509,7 +522,6 @@ static const struct r600_reg r600_reg_list[] = {
 	{0, 0, R_028850_SQ_PGM_RESOURCES_PS},
 	{0, 0, R_028854_SQ_PGM_EXPORTS_PS},
 	{0, 0, R_0288CC_SQ_PGM_CF_OFFSET_PS},
-	{0, 0, R_008958_VGT_PRIMITIVE_TYPE},
 	{0, 0, R_028400_VGT_MAX_VTX_INDX},
 	{0, 0, R_028404_VGT_MIN_VTX_INDX},
 	{0, 0, R_028408_VGT_INDX_OFFSET},
@@ -534,7 +546,7 @@ static int r600_state_constant_init(struct r600_context *ctx, u32 offset)
 	for (int i = 0; i < nreg; i++) {
 		r600_shader_constant[i].offset += offset;
 	}
-	return r600_context_add_block(ctx, r600_shader_constant, nreg);
+	return r600_context_add_block(ctx, r600_shader_constant, nreg, PKT3_SET_ALU_CONST);
 }
 
 /* SHADER RESOURCE R600/R700 */
@@ -554,7 +566,7 @@ static int r600_state_resource_init(struct r600_context *ctx, u32 offset)
 	for (int i = 0; i < nreg; i++) {
 		r600_shader_resource[i].offset += offset;
 	}
-	return r600_context_add_block(ctx, r600_shader_resource, nreg);
+	return r600_context_add_block(ctx, r600_shader_resource, nreg, PKT3_SET_RESOURCE);
 }
 
 /* SHADER SAMPLER R600/R700 */
@@ -570,7 +582,7 @@ static int r600_state_sampler_init(struct r600_context *ctx, u32 offset)
 	for (int i = 0; i < nreg; i++) {
 		r600_shader_sampler[i].offset += offset;
 	}
-	return r600_context_add_block(ctx, r600_shader_sampler, nreg);
+	return r600_context_add_block(ctx, r600_shader_sampler, nreg, PKT3_SET_SAMPLER);
 }
 
 /* SHADER SAMPLER BORDER R600/R700 */
@@ -587,7 +599,7 @@ static int r600_state_sampler_border_init(struct r600_context *ctx, u32 offset)
 	for (int i = 0; i < nreg; i++) {
 		r600_shader_sampler_border[i].offset += offset;
 	}
-	return r600_context_add_block(ctx, r600_shader_sampler_border, nreg);
+	return r600_context_add_block(ctx, r600_shader_sampler_border, nreg, PKT3_SET_CONFIG_REG);
 }
 
 /* initialize */
@@ -644,7 +656,14 @@ int r600_context_init(struct r600_context *ctx, struct radeon *radeon)
 	ctx->ngroups = R600_NGROUPS;
 
 	/* add blocks */
-	r = r600_context_add_block(ctx, r600_reg_list, sizeof(r600_reg_list)/sizeof(struct r600_reg));
+	r = r600_context_add_block(ctx, r600_config_reg_list,
+				sizeof(r600_config_reg_list)/sizeof(struct r600_reg),
+				PKT3_SET_CONFIG_REG);
+	if (r)
+		goto out_err;
+	r = r600_context_add_block(ctx, r600_context_reg_list,
+				sizeof(r600_context_reg_list)/sizeof(struct r600_reg),
+				PKT3_SET_CONTEXT_REG);
 	if (r)
 		goto out_err;
 
@@ -762,8 +781,8 @@ void r600_context_pipe_state_set(struct r600_context *ctx, struct r600_pipe_stat
 		id = group->offset_block_id[(state->regs[i].offset - group->start_offset) >> 2];
 		block = &group->blocks[id];
 		id = (state->regs[i].offset - block->start_offset) >> 2;
-		block->pm4[id] &= ~state->regs[i].mask;
-		block->pm4[id] |= state->regs[i].value;
+		block->reg[id] &= ~state->regs[i].mask;
+		block->reg[id] |= state->regs[i].value;
 		if (block->pm4_bo_index[id]) {
 			/* find relocation */
 			id = block->pm4_bo_index[id];
@@ -771,7 +790,7 @@ void r600_context_pipe_state_set(struct r600_context *ctx, struct r600_pipe_stat
 		}
 		block->status |= R600_BLOCK_STATUS_ENABLED;
 		block->status |= R600_BLOCK_STATUS_DIRTY;
-		ctx->pm4_dirty_cdwords += 2 + block->pm4_ndwords;
+		ctx->pm4_dirty_cdwords += block->pm4_ndwords;
 	}
 }
 
@@ -783,13 +802,13 @@ static inline void r600_context_pipe_state_set_resource(struct r600_context *ctx
 	offset -= ctx->groups[R600_GROUP_RESOURCE].start_offset;
 	id = ctx->groups[R600_GROUP_RESOURCE].offset_block_id[offset >> 2];
 	block = &ctx->groups[R600_GROUP_RESOURCE].blocks[id];
-	block->pm4[0] = state->regs[0].value;
-	block->pm4[1] = state->regs[1].value;
-	block->pm4[2] = state->regs[2].value;
-	block->pm4[3] = state->regs[3].value;
-	block->pm4[4] = state->regs[4].value;
-	block->pm4[5] = state->regs[5].value;
-	block->pm4[6] = state->regs[6].value;
+	block->reg[0] = state->regs[0].value;
+	block->reg[1] = state->regs[1].value;
+	block->reg[2] = state->regs[2].value;
+	block->reg[3] = state->regs[3].value;
+	block->reg[4] = state->regs[4].value;
+	block->reg[5] = state->regs[5].value;
+	block->reg[6] = state->regs[6].value;
 	radeon_ws_bo_reference(ctx->radeon, &block->reloc[1].bo, NULL);
 	radeon_ws_bo_reference(ctx->radeon , &block->reloc[2].bo, NULL);
 	if (state->regs[0].bo) {
@@ -805,7 +824,7 @@ static inline void r600_context_pipe_state_set_resource(struct r600_context *ctx
 	}
 	block->status |= R600_BLOCK_STATUS_ENABLED;
 	block->status |= R600_BLOCK_STATUS_DIRTY;
-	ctx->pm4_dirty_cdwords += 2 + block->pm4_ndwords;
+	ctx->pm4_dirty_cdwords += block->pm4_ndwords;
 }
 
 void r600_context_pipe_state_set_ps_resource(struct r600_context *ctx, struct r600_pipe_state *state, unsigned rid)
@@ -830,12 +849,12 @@ static inline void r600_context_pipe_state_set_sampler(struct r600_context *ctx,
 	offset -= ctx->groups[R600_GROUP_SAMPLER].start_offset;
 	id = ctx->groups[R600_GROUP_SAMPLER].offset_block_id[offset >> 2];
 	block = &ctx->groups[R600_GROUP_SAMPLER].blocks[id];
-	block->pm4[0] = state->regs[0].value;
-	block->pm4[1] = state->regs[1].value;
-	block->pm4[2] = state->regs[2].value;
+	block->reg[0] = state->regs[0].value;
+	block->reg[1] = state->regs[1].value;
+	block->reg[2] = state->regs[2].value;
 	block->status |= R600_BLOCK_STATUS_ENABLED;
 	block->status |= R600_BLOCK_STATUS_DIRTY;
-	ctx->pm4_dirty_cdwords += 2 + block->pm4_ndwords;
+	ctx->pm4_dirty_cdwords += block->pm4_ndwords;
 }
 
 static inline void r600_context_pipe_state_set_sampler_border(struct r600_context *ctx, struct r600_pipe_state *state, unsigned offset)
@@ -846,13 +865,13 @@ static inline void r600_context_pipe_state_set_sampler_border(struct r600_contex
 	offset -= ctx->groups[R600_GROUP_CONFIG].start_offset;
 	id = ctx->groups[R600_GROUP_CONFIG].offset_block_id[offset >> 2];
 	block = &ctx->groups[R600_GROUP_CONFIG].blocks[id];
-	block->pm4[0] = state->regs[3].value;
-	block->pm4[1] = state->regs[4].value;
-	block->pm4[2] = state->regs[5].value;
-	block->pm4[3] = state->regs[6].value;
+	block->reg[0] = state->regs[3].value;
+	block->reg[1] = state->regs[4].value;
+	block->reg[2] = state->regs[5].value;
+	block->reg[3] = state->regs[6].value;
 	block->status |= R600_BLOCK_STATUS_ENABLED;
 	block->status |= R600_BLOCK_STATUS_DIRTY;
-	ctx->pm4_dirty_cdwords += 2 + block->pm4_ndwords;
+	ctx->pm4_dirty_cdwords += block->pm4_ndwords;
 }
 
 void r600_context_pipe_state_set_ps_sampler(struct r600_context *ctx, struct r600_pipe_state *state, unsigned id)
@@ -879,7 +898,7 @@ void r600_context_pipe_state_set_vs_sampler(struct r600_context *ctx, struct r60
 	}
 }
 
-void r600_context_group_emit_dirty(struct r600_context *ctx, struct r600_group *group, unsigned opcode)
+void r600_context_group_emit_dirty(struct r600_context *ctx, struct r600_group *group)
 {
 	struct radeon_bo *bo;
 	int id;
@@ -898,8 +917,6 @@ void r600_context_group_emit_dirty(struct r600_context *ctx, struct r600_group *
 				}
 			}
 
-			ctx->pm4[ctx->pm4_cdwords++] = PKT3(opcode, block->nreg);
-			ctx->pm4[ctx->pm4_cdwords++] = (block->start_offset - group->start_offset) >> 2;
 			memcpy(&ctx->pm4[ctx->pm4_cdwords], block->pm4, block->pm4_ndwords * 4);
 			ctx->pm4_cdwords += block->pm4_ndwords;
 			block->status ^= R600_BLOCK_STATUS_DIRTY;
@@ -968,11 +985,11 @@ void r600_context_draw(struct r600_context *ctx, const struct r600_draw *draw)
 	}
 
 	/* enough room to copy packet */
-	r600_context_group_emit_dirty(ctx, &ctx->groups[R600_GROUP_CONFIG], PKT3_SET_CONFIG_REG);
-	r600_context_group_emit_dirty(ctx, &ctx->groups[R600_GROUP_CONTEXT], PKT3_SET_CONTEXT_REG);
-	r600_context_group_emit_dirty(ctx, &ctx->groups[R600_GROUP_ALU_CONST], PKT3_SET_ALU_CONST);
-	r600_context_group_emit_dirty(ctx, &ctx->groups[R600_GROUP_RESOURCE], PKT3_SET_RESOURCE);
-	r600_context_group_emit_dirty(ctx, &ctx->groups[R600_GROUP_SAMPLER], PKT3_SET_SAMPLER);
+	r600_context_group_emit_dirty(ctx, &ctx->groups[R600_GROUP_CONFIG]);
+	r600_context_group_emit_dirty(ctx, &ctx->groups[R600_GROUP_CONTEXT]);
+	r600_context_group_emit_dirty(ctx, &ctx->groups[R600_GROUP_ALU_CONST]);
+	r600_context_group_emit_dirty(ctx, &ctx->groups[R600_GROUP_RESOURCE]);
+	r600_context_group_emit_dirty(ctx, &ctx->groups[R600_GROUP_SAMPLER]);
 
 	/* draw packet */
 	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_INDEX_TYPE, 0);
@@ -1063,7 +1080,7 @@ void r600_context_flush(struct r600_context *ctx)
 			/* mark enabled block as dirty */
 			block = &ctx->groups[i].blocks[j];
 			if (block->status & R600_BLOCK_STATUS_ENABLED) {
-				ctx->pm4_dirty_cdwords += 2 + block->pm4_ndwords;
+				ctx->pm4_dirty_cdwords += block->pm4_ndwords;
 				block->status |= R600_BLOCK_STATUS_DIRTY;
 			}
 		}