X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fradeonsi%2Fsi_state_shaders.c;h=6e77ca9fc59538d7ca35059ce3587c4845c0dcd3;hb=3be4ed2fe1ee8629a486156a764b451513d34aac;hp=de33d2503018c7c16829a7dea9dabfc810aab907;hpb=86b52d42368ac496fe24bc6674e754c323381635;p=mesa.git

diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index de33d250301..6e77ca9fc59 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -23,11 +23,10 @@
  */
 
 #include "si_build_pm4.h"
-#include "gfx9d.h"
+#include "sid.h"
 
 #include "compiler/nir/nir_serialize.h"
 #include "tgsi/tgsi_parse.h"
-#include "tgsi/tgsi_ureg.h"
 #include "util/hash_table.h"
 #include "util/crc32.h"
 #include "util/u_async_debug.h"
@@ -128,21 +127,21 @@ static uint32_t *read_chunk(uint32_t *ptr, void **data, unsigned *size)
 static void *si_get_shader_binary(struct si_shader *shader)
 {
 	/* There is always a size of data followed by the data itself. */
-	unsigned relocs_size = shader->binary.reloc_count *
-			       sizeof(shader->binary.relocs[0]);
-	unsigned disasm_size = shader->binary.disasm_string ?
-			       strlen(shader->binary.disasm_string) + 1 : 0;
 	unsigned llvm_ir_size = shader->binary.llvm_ir_string ?
 				strlen(shader->binary.llvm_ir_string) + 1 : 0;
+
+	/* Refuse to allocate overly large buffers and guard against integer
+	 * overflow. */
+	if (shader->binary.elf_size > UINT_MAX / 4 ||
+	    llvm_ir_size > UINT_MAX / 4)
+		return NULL;
+
 	unsigned size =
 		4 + /* total size */
 		4 + /* CRC32 of the data below */
 		align(sizeof(shader->config), 4) +
 		align(sizeof(shader->info), 4) +
-		4 + align(shader->binary.code_size, 4) +
-		4 + align(shader->binary.rodata_size, 4) +
-		4 + align(relocs_size, 4) +
-		4 + align(disasm_size, 4) +
+		4 + align(shader->binary.elf_size, 4) +
 		4 + align(llvm_ir_size, 4);
 	void *buffer = CALLOC(1, size);
 	uint32_t *ptr = (uint32_t*)buffer;
@@ -155,10 +154,7 @@ static void *si_get_shader_binary(struct si_shader *shader)
 
 	ptr = write_data(ptr, &shader->config, sizeof(shader->config));
 	ptr = write_data(ptr, &shader->info, sizeof(shader->info));
-	ptr = write_chunk(ptr, shader->binary.code, shader->binary.code_size);
-	ptr = write_chunk(ptr, shader->binary.rodata, shader->binary.rodata_size);
-	ptr = write_chunk(ptr, shader->binary.relocs, relocs_size);
-	ptr = write_chunk(ptr, shader->binary.disasm_string, disasm_size);
+	ptr = write_chunk(ptr, shader->binary.elf_buffer, shader->binary.elf_size);
 	ptr = write_chunk(ptr, shader->binary.llvm_ir_string, llvm_ir_size);
 	assert((char *)ptr - (char *)buffer == size);
 
@@ -176,6 +172,7 @@ static bool si_load_shader_binary(struct si_shader *shader, void *binary)
 	uint32_t size = *ptr++;
 	uint32_t crc32 = *ptr++;
 	unsigned chunk_size;
+	unsigned elf_size;
 
 	if (util_hash_crc32(ptr, size - 8) != crc32) {
 		fprintf(stderr, "radeonsi: binary shader has invalid CRC32\n");
@@ -184,13 +181,9 @@ static bool si_load_shader_binary(struct si_shader *shader, void *binary)
 
 	ptr = read_data(ptr, &shader->config, sizeof(shader->config));
 	ptr = read_data(ptr, &shader->info, sizeof(shader->info));
-	ptr = read_chunk(ptr, (void**)&shader->binary.code,
-			 &shader->binary.code_size);
-	ptr = read_chunk(ptr, (void**)&shader->binary.rodata,
-			 &shader->binary.rodata_size);
-	ptr = read_chunk(ptr, (void**)&shader->binary.relocs, &chunk_size);
-	shader->binary.reloc_count = chunk_size / sizeof(shader->binary.relocs[0]);
-	ptr = read_chunk(ptr, (void**)&shader->binary.disasm_string, &chunk_size);
+	ptr = read_chunk(ptr, (void**)&shader->binary.elf_buffer,
+			 &elf_size);
+	shader->binary.elf_size = elf_size;
 	ptr = read_chunk(ptr, (void**)&shader->binary.llvm_ir_string, &chunk_size);
 
 	return true;
@@ -338,10 +331,10 @@ void si_destroy_shader_cache(struct si_screen *sscreen)
 /* SHADER STATES */
 
 static void si_set_tesseval_regs(struct si_screen *sscreen,
-				 struct si_shader_selector *tes,
+				 const struct si_shader_selector *tes,
 				 struct si_pm4_state *pm4)
 {
-	struct tgsi_shader_info *info = &tes->info;
+	const struct tgsi_shader_info *info = &tes->info;
 	unsigned tes_prim_mode = info->properties[TGSI_PROPERTY_TES_PRIM_MODE];
 	unsigned tes_spacing = info->properties[TGSI_PROPERTY_TES_SPACING];
 	bool tes_vertex_order_cw = info->properties[TGSI_PROPERTY_TES_VERTEX_ORDER_CW];
@@ -397,11 +390,11 @@ static void si_set_tesseval_regs(struct si_screen *sscreen,
 	} else
 		distribution_mode = V_028B6C_DISTRIBUTION_MODE_NO_DIST;
 
-	si_pm4_set_reg(pm4, R_028B6C_VGT_TF_PARAM,
-		       S_028B6C_TYPE(type) |
-		       S_028B6C_PARTITIONING(partitioning) |
-		       S_028B6C_TOPOLOGY(topology) |
-		       S_028B6C_DISTRIBUTION_MODE(distribution_mode));
+	assert(pm4->shader);
+	pm4->shader->vgt_tf_param = S_028B6C_TYPE(type) |
+				    S_028B6C_PARTITIONING(partitioning) |
+				    S_028B6C_TOPOLOGY(topology) |
+				    S_028B6C_DISTRIBUTION_MODE(distribution_mode);
 }
 
 /* Polaris needs different VTX_REUSE_DEPTH settings depending on
@@ -425,7 +418,8 @@ static void polaris_set_vgt_vertex_reuse(struct si_screen *sscreen,
 {
 	unsigned type = sel->type;
 
-	if (sscreen->info.family < CHIP_POLARIS10)
+	if (sscreen->info.family < CHIP_POLARIS10 ||
+	    sscreen->info.chip_class >= GFX10)
 		return;
 
 	/* VS as VS, or VS as ES: */
@@ -441,8 +435,8 @@ static void polaris_set_vgt_vertex_reuse(struct si_screen *sscreen,
 		    PIPE_TESS_SPACING_FRACTIONAL_ODD)
 			vtx_reuse_depth = 14;
 
-		si_pm4_set_reg(pm4, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
-			       vtx_reuse_depth);
+		assert(pm4->shader);
+		pm4->shader->vgt_vertex_reuse_block_cntl = vtx_reuse_depth;
 	}
 }
 
@@ -453,18 +447,19 @@ static struct si_pm4_state *si_get_shader_pm4_state(struct si_shader *shader)
 	else
 		shader->pm4 = CALLOC_STRUCT(si_pm4_state);
 
-	return shader->pm4;
+	if (shader->pm4) {
+		shader->pm4->shader = shader;
+		return shader->pm4;
+	} else {
+		fprintf(stderr, "radeonsi: Failed to create pm4 state.\n");
+		return NULL;
+	}
 }
 
 static unsigned si_get_num_vs_user_sgprs(unsigned num_always_on_user_sgprs)
 {
 	/* Add the pointer to VBO descriptors. */
-	if (HAVE_32BIT_POINTERS) {
-		return num_always_on_user_sgprs + 1;
-	} else {
-		assert(num_always_on_user_sgprs % 2 == 0);
-		return num_always_on_user_sgprs + 2;
-	}
+	return num_always_on_user_sgprs + 1;
 }
 
 static void si_shader_ls(struct si_screen *sscreen, struct si_shader *shader)
@@ -473,7 +468,7 @@ static void si_shader_ls(struct si_screen *sscreen, struct si_shader *shader)
 	unsigned vgpr_comp_cnt;
 	uint64_t va;
 
-	assert(sscreen->info.chip_class <= VI);
+	assert(sscreen->info.chip_class <= GFX8);
 
 	pm4 = si_get_shader_pm4_state(shader);
 	if (!pm4)
@@ -514,22 +509,39 @@ static void si_shader_hs(struct si_screen *sscreen, struct si_shader *shader)
 	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
 
 	if (sscreen->info.chip_class >= GFX9) {
-		si_pm4_set_reg(pm4, R_00B410_SPI_SHADER_PGM_LO_LS, va >> 8);
-		si_pm4_set_reg(pm4, R_00B414_SPI_SHADER_PGM_HI_LS, S_00B414_MEM_BASE(va >> 40));
+		if (sscreen->info.chip_class >= GFX10) {
+			si_pm4_set_reg(pm4, R_00B520_SPI_SHADER_PGM_LO_LS, va >> 8);
+			si_pm4_set_reg(pm4, R_00B524_SPI_SHADER_PGM_HI_LS, S_00B524_MEM_BASE(va >> 40));
+		} else {
+			si_pm4_set_reg(pm4, R_00B410_SPI_SHADER_PGM_LO_LS, va >> 8);
+			si_pm4_set_reg(pm4, R_00B414_SPI_SHADER_PGM_HI_LS, S_00B414_MEM_BASE(va >> 40));
+		}
 
 		/* We need at least 2 components for LS.
-		 * VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
-		 * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
+		 * GFX9  VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
+		 * GFX10 VGPR0-3: (VertexID, RelAutoindex, UserVGPR1, InstanceID).
+		 * On gfx9, StepRate0 is set to 1 so that VGPR3 doesn't have to
+		 * be loaded.
 		 */
-		ls_vgpr_comp_cnt = shader->info.uses_instanceid ? 2 : 1;
+		ls_vgpr_comp_cnt = 1;
+		if (shader->info.uses_instanceid) {
+			if (sscreen->info.chip_class >= GFX10)
+				ls_vgpr_comp_cnt = 3;
+			else
+				ls_vgpr_comp_cnt = 2;
+		}
 
 		unsigned num_user_sgprs =
 			si_get_num_vs_user_sgprs(GFX9_TCS_NUM_USER_SGPR);
 
 		shader->config.rsrc2 =
 			S_00B42C_USER_SGPR(num_user_sgprs) |
-			S_00B42C_USER_SGPR_MSB(num_user_sgprs >> 5) |
 			S_00B42C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
+
+		if (sscreen->info.chip_class >= GFX10)
+			shader->config.rsrc2 |= S_00B42C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5);
+		else
+			shader->config.rsrc2 |= S_00B42C_USER_SGPR_MSB_GFX9(num_user_sgprs >> 5);
 	} else {
 		si_pm4_set_reg(pm4, R_00B420_SPI_SHADER_PGM_LO_HS, va >> 8);
 		si_pm4_set_reg(pm4, R_00B424_SPI_SHADER_PGM_HI_HS, S_00B424_MEM_BASE(va >> 40));
@@ -542,17 +554,45 @@ static void si_shader_hs(struct si_screen *sscreen, struct si_shader *shader)
 
 	si_pm4_set_reg(pm4, R_00B428_SPI_SHADER_PGM_RSRC1_HS,
 		       S_00B428_VGPRS((shader->config.num_vgprs - 1) / 4) |
-		       S_00B428_SGPRS((shader->config.num_sgprs - 1) / 8) |
+		       (sscreen->info.chip_class <= GFX9 ?
+				S_00B428_SGPRS((shader->config.num_sgprs - 1) / 8) : 0) |
 		       S_00B428_DX10_CLAMP(1) |
+		       S_00B428_MEM_ORDERED(sscreen->info.chip_class >= GFX10) |
 		       S_00B428_FLOAT_MODE(shader->config.float_mode) |
 		       S_00B428_LS_VGPR_COMP_CNT(ls_vgpr_comp_cnt));
 
-	if (sscreen->info.chip_class <= VI) {
+	if (sscreen->info.chip_class <= GFX8) {
 		si_pm4_set_reg(pm4, R_00B42C_SPI_SHADER_PGM_RSRC2_HS,
 			       shader->config.rsrc2);
 	}
 }
 
+static void si_emit_shader_es(struct si_context *sctx)
+{
+	struct si_shader *shader = sctx->queued.named.es->shader;
+	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+	if (!shader)
+		return;
+
+	radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
+				   SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
+				   shader->selector->esgs_itemsize / 4);
+
+	if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
+		radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
+					   SI_TRACKED_VGT_TF_PARAM,
+					   shader->vgt_tf_param);
+
+	if (shader->vgt_vertex_reuse_block_cntl)
+		radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
+					   SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
+					   shader->vgt_vertex_reuse_block_cntl);
+
+	if (initial_cdw != sctx->gfx_cs->current.cdw)
+		sctx->context_roll = true;
+}
+
 static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
 {
 	struct si_pm4_state *pm4;
@@ -561,12 +601,13 @@ static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
 	uint64_t va;
 	unsigned oc_lds_en;
 
-	assert(sscreen->info.chip_class <= VI);
+	assert(sscreen->info.chip_class <= GFX8);
 
 	pm4 = si_get_shader_pm4_state(shader);
 	if (!pm4)
 		return;
 
+	pm4->atom.emit = si_emit_shader_es;
 	va = shader->bo->gpu_address;
 	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
 
@@ -582,8 +623,6 @@ static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
 
 	oc_lds_en = shader->selector->type == PIPE_SHADER_TESS_EVAL ? 1 : 0;
 
-	si_pm4_set_reg(pm4, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
-		       shader->selector->esgs_itemsize / 4);
 	si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
 	si_pm4_set_reg(pm4, R_00B324_SPI_SHADER_PGM_HI_ES, S_00B324_MEM_BASE(va >> 40));
 	si_pm4_set_reg(pm4, R_00B328_SPI_SHADER_PGM_RSRC1_ES,
@@ -603,41 +642,9 @@ static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
 	polaris_set_vgt_vertex_reuse(sscreen, shader->selector, shader, pm4);
 }
 
-static unsigned si_conv_prim_to_gs_out(unsigned mode)
-{
-	static const int prim_conv[] = {
-		[PIPE_PRIM_POINTS]			= V_028A6C_OUTPRIM_TYPE_POINTLIST,
-		[PIPE_PRIM_LINES]			= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
-		[PIPE_PRIM_LINE_LOOP]			= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
-		[PIPE_PRIM_LINE_STRIP]			= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
-		[PIPE_PRIM_TRIANGLES]			= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
-		[PIPE_PRIM_TRIANGLE_STRIP]		= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
-		[PIPE_PRIM_TRIANGLE_FAN]		= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
-		[PIPE_PRIM_QUADS]			= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
-		[PIPE_PRIM_QUAD_STRIP]			= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
-		[PIPE_PRIM_POLYGON]			= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
-		[PIPE_PRIM_LINES_ADJACENCY]		= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
-		[PIPE_PRIM_LINE_STRIP_ADJACENCY]	= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
-		[PIPE_PRIM_TRIANGLES_ADJACENCY]		= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
-		[PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY]	= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
-		[PIPE_PRIM_PATCHES]			= V_028A6C_OUTPRIM_TYPE_POINTLIST,
-	};
-	assert(mode < ARRAY_SIZE(prim_conv));
-
-	return prim_conv[mode];
-}
-
-struct gfx9_gs_info {
-	unsigned es_verts_per_subgroup;
-	unsigned gs_prims_per_subgroup;
-	unsigned gs_inst_prims_in_subgroup;
-	unsigned max_prims_per_subgroup;
-	unsigned lds_size;
-};
-
-static void gfx9_get_gs_info(struct si_shader_selector *es,
-				   struct si_shader_selector *gs,
-				   struct gfx9_gs_info *out)
+void gfx9_get_gs_info(struct si_shader_selector *es,
+		      struct si_shader_selector *gs,
+		      struct gfx9_gs_info *out)
 {
 	unsigned gs_num_invocations = MAX2(gs->gs_num_invocations, 1);
 	unsigned input_prim = gs->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
@@ -658,8 +665,6 @@ static void gfx9_get_gs_info(struct si_shader_selector *es,
 	unsigned max_gs_prims, gs_prims;
 	unsigned min_es_verts, es_verts, worst_case_es_verts;
 
-	assert(gs_num_invocations <= 32); /* GL maximum */
-
 	if (uses_adjacency || gs_num_invocations > 1)
 		max_gs_prims = 127 / gs_num_invocations;
 	else
@@ -730,11 +735,79 @@ static void gfx9_get_gs_info(struct si_shader_selector *es,
 	out->gs_inst_prims_in_subgroup = gs_prims * gs_num_invocations;
 	out->max_prims_per_subgroup = out->gs_inst_prims_in_subgroup *
 				      gs->gs_max_out_vertices;
-	out->lds_size = align(esgs_lds_size, 128) / 128;
+	out->esgs_ring_size = 4 * esgs_lds_size;
 
 	assert(out->max_prims_per_subgroup <= max_out_prims);
 }
 
+static void si_emit_shader_gs(struct si_context *sctx)
+{
+	struct si_shader *shader = sctx->queued.named.gs->shader;
+	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+	if (!shader)
+		return;
+
+	/* R_028A60_VGT_GSVS_RING_OFFSET_1, R_028A64_VGT_GSVS_RING_OFFSET_2
+	 * R_028A68_VGT_GSVS_RING_OFFSET_3 */
+	radeon_opt_set_context_reg3(sctx, R_028A60_VGT_GSVS_RING_OFFSET_1,
+				    SI_TRACKED_VGT_GSVS_RING_OFFSET_1,
+				    shader->ctx_reg.gs.vgt_gsvs_ring_offset_1,
+				    shader->ctx_reg.gs.vgt_gsvs_ring_offset_2,
+				    shader->ctx_reg.gs.vgt_gsvs_ring_offset_3);
+
+	/* R_028AB0_VGT_GSVS_RING_ITEMSIZE */
+	radeon_opt_set_context_reg(sctx, R_028AB0_VGT_GSVS_RING_ITEMSIZE,
+				   SI_TRACKED_VGT_GSVS_RING_ITEMSIZE,
+				   shader->ctx_reg.gs.vgt_gsvs_ring_itemsize);
+
+	/* R_028B38_VGT_GS_MAX_VERT_OUT */
+	radeon_opt_set_context_reg(sctx, R_028B38_VGT_GS_MAX_VERT_OUT,
+				   SI_TRACKED_VGT_GS_MAX_VERT_OUT,
+				   shader->ctx_reg.gs.vgt_gs_max_vert_out);
+
+	/* R_028B5C_VGT_GS_VERT_ITEMSIZE, R_028B60_VGT_GS_VERT_ITEMSIZE_1
+	 * R_028B64_VGT_GS_VERT_ITEMSIZE_2, R_028B68_VGT_GS_VERT_ITEMSIZE_3 */
+	radeon_opt_set_context_reg4(sctx, R_028B5C_VGT_GS_VERT_ITEMSIZE,
+				    SI_TRACKED_VGT_GS_VERT_ITEMSIZE,
+				    shader->ctx_reg.gs.vgt_gs_vert_itemsize,
+				    shader->ctx_reg.gs.vgt_gs_vert_itemsize_1,
+				    shader->ctx_reg.gs.vgt_gs_vert_itemsize_2,
+				    shader->ctx_reg.gs.vgt_gs_vert_itemsize_3);
+
+	/* R_028B90_VGT_GS_INSTANCE_CNT */
+	radeon_opt_set_context_reg(sctx, R_028B90_VGT_GS_INSTANCE_CNT,
+				   SI_TRACKED_VGT_GS_INSTANCE_CNT,
+				   shader->ctx_reg.gs.vgt_gs_instance_cnt);
+
+	if (sctx->chip_class >= GFX9) {
+		/* R_028A44_VGT_GS_ONCHIP_CNTL */
+		radeon_opt_set_context_reg(sctx, R_028A44_VGT_GS_ONCHIP_CNTL,
+					   SI_TRACKED_VGT_GS_ONCHIP_CNTL,
+					   shader->ctx_reg.gs.vgt_gs_onchip_cntl);
+		/* R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP */
+		radeon_opt_set_context_reg(sctx, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP,
+					   SI_TRACKED_VGT_GS_MAX_PRIMS_PER_SUBGROUP,
+					   shader->ctx_reg.gs.vgt_gs_max_prims_per_subgroup);
+		/* R_028AAC_VGT_ESGS_RING_ITEMSIZE */
+		radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
+					   SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
+					   shader->ctx_reg.gs.vgt_esgs_ring_itemsize);
+
+		if (shader->key.part.gs.es->type == PIPE_SHADER_TESS_EVAL)
+			radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
+						   SI_TRACKED_VGT_TF_PARAM,
+						   shader->vgt_tf_param);
+		if (shader->vgt_vertex_reuse_block_cntl)
+			radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
+						   SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
+						   shader->vgt_vertex_reuse_block_cntl);
+	}
+
+	if (initial_cdw != sctx->gfx_cs->current.cdw)
+		sctx->context_roll = true;
+}
+
 static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
 {
 	struct si_shader_selector *sel = shader->selector;
@@ -749,33 +822,35 @@ static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
 	if (!pm4)
 		return;
 
+	pm4->atom.emit = si_emit_shader_gs;
+
 	offset = num_components[0] * sel->gs_max_out_vertices;
-	si_pm4_set_reg(pm4, R_028A60_VGT_GSVS_RING_OFFSET_1, offset);
+	shader->ctx_reg.gs.vgt_gsvs_ring_offset_1 = offset;
+
 	if (max_stream >= 1)
 		offset += num_components[1] * sel->gs_max_out_vertices;
-	si_pm4_set_reg(pm4, R_028A64_VGT_GSVS_RING_OFFSET_2, offset);
+	shader->ctx_reg.gs.vgt_gsvs_ring_offset_2 = offset;
+
 	if (max_stream >= 2)
 		offset += num_components[2] * sel->gs_max_out_vertices;
-	si_pm4_set_reg(pm4, R_028A68_VGT_GSVS_RING_OFFSET_3, offset);
-	si_pm4_set_reg(pm4, R_028A6C_VGT_GS_OUT_PRIM_TYPE,
-		       si_conv_prim_to_gs_out(sel->gs_output_prim));
+	shader->ctx_reg.gs.vgt_gsvs_ring_offset_3 = offset;
+
 	if (max_stream >= 3)
 		offset += num_components[3] * sel->gs_max_out_vertices;
-	si_pm4_set_reg(pm4, R_028AB0_VGT_GSVS_RING_ITEMSIZE, offset);
+	shader->ctx_reg.gs.vgt_gsvs_ring_itemsize = offset;
 
 	/* The GSVS_RING_ITEMSIZE register takes 15 bits */
 	assert(offset < (1 << 15));
 
-	si_pm4_set_reg(pm4, R_028B38_VGT_GS_MAX_VERT_OUT, sel->gs_max_out_vertices);
+	shader->ctx_reg.gs.vgt_gs_max_vert_out = sel->gs_max_out_vertices;
 
-	si_pm4_set_reg(pm4, R_028B5C_VGT_GS_VERT_ITEMSIZE, num_components[0]);
-	si_pm4_set_reg(pm4, R_028B60_VGT_GS_VERT_ITEMSIZE_1, (max_stream >= 1) ? num_components[1] : 0);
-	si_pm4_set_reg(pm4, R_028B64_VGT_GS_VERT_ITEMSIZE_2, (max_stream >= 2) ? num_components[2] : 0);
-	si_pm4_set_reg(pm4, R_028B68_VGT_GS_VERT_ITEMSIZE_3, (max_stream >= 3) ? num_components[3] : 0);
+	shader->ctx_reg.gs.vgt_gs_vert_itemsize = num_components[0];
+	shader->ctx_reg.gs.vgt_gs_vert_itemsize_1 = (max_stream >= 1) ? num_components[1] : 0;
+	shader->ctx_reg.gs.vgt_gs_vert_itemsize_2 = (max_stream >= 2) ? num_components[2] : 0;
+	shader->ctx_reg.gs.vgt_gs_vert_itemsize_3 = (max_stream >= 3) ? num_components[3] : 0;
 
-	si_pm4_set_reg(pm4, R_028B90_VGT_GS_INSTANCE_CNT,
-		       S_028B90_CNT(MIN2(gs_num_invocations, 127)) |
-		       S_028B90_ENABLE(gs_num_invocations > 0));
+	shader->ctx_reg.gs.vgt_gs_instance_cnt = S_028B90_CNT(MIN2(gs_num_invocations, 127)) |
+						 S_028B90_ENABLE(gs_num_invocations > 0);
 
 	va = shader->bo->gpu_address;
 	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
@@ -784,7 +859,6 @@ static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
 		unsigned input_prim = sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
 		unsigned es_type = shader->key.part.gs.es->type;
 		unsigned es_vgpr_comp_cnt, gs_vgpr_comp_cnt;
-		struct gfx9_gs_info gs_info;
 
 		if (es_type == PIPE_SHADER_VERTEX)
 			/* VGPR0-3: (VertexID, InstanceID / StepRate0, ...) */
@@ -812,33 +886,45 @@ static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
 		else
 			num_user_sgprs = GFX9_TESGS_NUM_USER_SGPR;
 
-		gfx9_get_gs_info(shader->key.part.gs.es, sel, &gs_info);
+		if (sscreen->info.chip_class >= GFX10) {
+			si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
+			si_pm4_set_reg(pm4, R_00B324_SPI_SHADER_PGM_HI_ES, S_00B324_MEM_BASE(va >> 40));
+		} else {
+			si_pm4_set_reg(pm4, R_00B210_SPI_SHADER_PGM_LO_ES, va >> 8);
+			si_pm4_set_reg(pm4, R_00B214_SPI_SHADER_PGM_HI_ES, S_00B214_MEM_BASE(va >> 40));
+		}
 
-		si_pm4_set_reg(pm4, R_00B210_SPI_SHADER_PGM_LO_ES, va >> 8);
-		si_pm4_set_reg(pm4, R_00B214_SPI_SHADER_PGM_HI_ES, S_00B214_MEM_BASE(va >> 40));
+		uint32_t rsrc1 =
+			S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
+			S_00B228_DX10_CLAMP(1) |
+			S_00B228_MEM_ORDERED(sscreen->info.chip_class >= GFX10) |
+			S_00B228_FLOAT_MODE(shader->config.float_mode) |
+			S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt);
+		uint32_t rsrc2 =
+			S_00B22C_USER_SGPR(num_user_sgprs) |
+			S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
+			S_00B22C_OC_LDS_EN(es_type == PIPE_SHADER_TESS_EVAL) |
+			S_00B22C_LDS_SIZE(shader->config.lds_size) |
+			S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
+
+		if (sscreen->info.chip_class >= GFX10) {
+			rsrc2 |= S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5);
+		} else {
+			rsrc1 |= S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8);
+			rsrc2 |= S_00B22C_USER_SGPR_MSB_GFX9(num_user_sgprs >> 5);
+		}
 
-		si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
-			       S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
-			       S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8) |
-			       S_00B228_DX10_CLAMP(1) |
-			       S_00B228_FLOAT_MODE(shader->config.float_mode) |
-			       S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt));
-		si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
-			       S_00B22C_USER_SGPR(num_user_sgprs) |
-			       S_00B22C_USER_SGPR_MSB(num_user_sgprs >> 5) |
-			       S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
-			       S_00B22C_OC_LDS_EN(es_type == PIPE_SHADER_TESS_EVAL) |
-			       S_00B22C_LDS_SIZE(gs_info.lds_size) |
-			       S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
+		si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS, rsrc1);
+		si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS, rsrc2);
 
-		si_pm4_set_reg(pm4, R_028A44_VGT_GS_ONCHIP_CNTL,
-			       S_028A44_ES_VERTS_PER_SUBGRP(gs_info.es_verts_per_subgroup) |
-			       S_028A44_GS_PRIMS_PER_SUBGRP(gs_info.gs_prims_per_subgroup) |
-			       S_028A44_GS_INST_PRIMS_IN_SUBGRP(gs_info.gs_inst_prims_in_subgroup));
-		si_pm4_set_reg(pm4, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP,
-			       S_028A94_MAX_PRIMS_PER_SUBGROUP(gs_info.max_prims_per_subgroup));
-		si_pm4_set_reg(pm4, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
-			       shader->key.part.gs.es->esgs_itemsize / 4);
+		shader->ctx_reg.gs.vgt_gs_onchip_cntl =
+			S_028A44_ES_VERTS_PER_SUBGRP(shader->gs_info.es_verts_per_subgroup) |
+			S_028A44_GS_PRIMS_PER_SUBGRP(shader->gs_info.gs_prims_per_subgroup) |
+			S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader->gs_info.gs_inst_prims_in_subgroup);
+		shader->ctx_reg.gs.vgt_gs_max_prims_per_subgroup =
+			S_028A94_MAX_PRIMS_PER_SUBGROUP(shader->gs_info.max_prims_per_subgroup);
+		shader->ctx_reg.gs.vgt_esgs_ring_itemsize =
+			shader->key.part.gs.es->esgs_itemsize / 4;
 
 		if (es_type == PIPE_SHADER_TESS_EVAL)
 			si_set_tesseval_regs(sscreen, shader->key.part.gs.es, pm4);
@@ -860,6 +946,348 @@ static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
 	}
 }
 
+/* Common tail code for NGG primitive shaders. */
+static void gfx10_emit_shader_ngg_tail(struct si_context *sctx,
+				       struct si_shader *shader,
+				       unsigned initial_cdw)
+{
+	radeon_opt_set_context_reg(sctx, R_0287FC_GE_MAX_OUTPUT_PER_SUBGROUP,
+				   SI_TRACKED_GE_MAX_OUTPUT_PER_SUBGROUP,
+				   shader->ctx_reg.ngg.ge_max_output_per_subgroup);
+	radeon_opt_set_context_reg(sctx, R_028B4C_GE_NGG_SUBGRP_CNTL,
+				   SI_TRACKED_GE_NGG_SUBGRP_CNTL,
+				   shader->ctx_reg.ngg.ge_ngg_subgrp_cntl);
+	radeon_opt_set_context_reg(sctx, R_028A84_VGT_PRIMITIVEID_EN,
+				   SI_TRACKED_VGT_PRIMITIVEID_EN,
+				   shader->ctx_reg.ngg.vgt_primitiveid_en);
+	radeon_opt_set_context_reg(sctx, R_028A44_VGT_GS_ONCHIP_CNTL,
+				   SI_TRACKED_VGT_GS_ONCHIP_CNTL,
+				   shader->ctx_reg.ngg.vgt_gs_onchip_cntl);
+	radeon_opt_set_context_reg(sctx, R_028B90_VGT_GS_INSTANCE_CNT,
+				   SI_TRACKED_VGT_GS_INSTANCE_CNT,
+				   shader->ctx_reg.ngg.vgt_gs_instance_cnt);
+	radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
+				   SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
+				   shader->ctx_reg.ngg.vgt_esgs_ring_itemsize);
+	radeon_opt_set_context_reg(sctx, R_028AB4_VGT_REUSE_OFF,
+				   SI_TRACKED_VGT_REUSE_OFF,
+				   shader->ctx_reg.ngg.vgt_reuse_off);
+	radeon_opt_set_context_reg(sctx, R_0286C4_SPI_VS_OUT_CONFIG,
+				   SI_TRACKED_SPI_VS_OUT_CONFIG,
+				   shader->ctx_reg.ngg.spi_vs_out_config);
+	radeon_opt_set_context_reg2(sctx, R_028708_SPI_SHADER_IDX_FORMAT,
+				   SI_TRACKED_SPI_SHADER_IDX_FORMAT,
+				   shader->ctx_reg.ngg.spi_shader_idx_format,
+				   shader->ctx_reg.ngg.spi_shader_pos_format);
+	radeon_opt_set_context_reg(sctx, R_028818_PA_CL_VTE_CNTL,
+				   SI_TRACKED_PA_CL_VTE_CNTL,
+				   shader->ctx_reg.ngg.pa_cl_vte_cntl);
+	radeon_opt_set_context_reg(sctx, R_028838_PA_CL_NGG_CNTL,
+				   SI_TRACKED_PA_CL_NGG_CNTL,
+				   shader->ctx_reg.ngg.pa_cl_ngg_cntl);
+
+	if (initial_cdw != sctx->gfx_cs->current.cdw)
+		sctx->context_roll = true;
+
+	if (shader->ge_cntl != sctx->last_multi_vgt_param) {
+		radeon_set_uconfig_reg(sctx->gfx_cs, R_03096C_GE_CNTL, shader->ge_cntl);
+		sctx->last_multi_vgt_param = shader->ge_cntl;
+	}
+}
+
+static void gfx10_emit_shader_ngg_notess_nogs(struct si_context *sctx)
+{
+	struct si_shader *shader = sctx->queued.named.gs->shader;
+	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+	if (!shader)
+		return;
+
+	gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
+}
+
+static void gfx10_emit_shader_ngg_tess_nogs(struct si_context *sctx)
+{
+	struct si_shader *shader = sctx->queued.named.gs->shader;
+	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+	if (!shader)
+		return;
+
+	radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
+				   SI_TRACKED_VGT_TF_PARAM,
+				   shader->vgt_tf_param);
+
+	gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
+}
+
+static void gfx10_emit_shader_ngg_notess_gs(struct si_context *sctx)
+{
+	struct si_shader *shader = sctx->queued.named.gs->shader;
+	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+	if (!shader)
+		return;
+
+	radeon_opt_set_context_reg(sctx, R_028B38_VGT_GS_MAX_VERT_OUT,
+				   SI_TRACKED_VGT_GS_MAX_VERT_OUT,
+				   shader->ctx_reg.ngg.vgt_gs_max_vert_out);
+
+	gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
+}
+
+static void gfx10_emit_shader_ngg_tess_gs(struct si_context *sctx)
+{
+	struct si_shader *shader = sctx->queued.named.gs->shader;
+	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+	if (!shader)
+		return;
+
+	radeon_opt_set_context_reg(sctx, R_028B38_VGT_GS_MAX_VERT_OUT,
+				   SI_TRACKED_VGT_GS_MAX_VERT_OUT,
+				   shader->ctx_reg.ngg.vgt_gs_max_vert_out);
+	radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
+				   SI_TRACKED_VGT_TF_PARAM,
+				   shader->vgt_tf_param);
+
+	gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
+}
+
+unsigned si_get_input_prim(const struct si_shader_selector *gs)
+{
+	if (gs->type == PIPE_SHADER_GEOMETRY)
+		return gs->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
+
+	if (gs->type == PIPE_SHADER_TESS_EVAL) {
+		if (gs->info.properties[TGSI_PROPERTY_TES_POINT_MODE])
+			return PIPE_PRIM_POINTS;
+		if (gs->info.properties[TGSI_PROPERTY_TES_PRIM_MODE] == PIPE_PRIM_LINES)
+			return PIPE_PRIM_LINES;
+		return PIPE_PRIM_TRIANGLES;
+	}
+
+	/* TODO: Set this correctly if the primitive type is set in the shader key. */
+	return PIPE_PRIM_TRIANGLES;
+}
+
+/**
+ * Prepare the PM4 image for \p shader, which will run as a merged ESGS shader
+ * in NGG mode.
+ */
+static void gfx10_shader_ngg(struct si_screen *sscreen, struct si_shader *shader)
+{
+	const struct si_shader_selector *gs_sel = shader->selector;
+	const struct tgsi_shader_info *gs_info = &gs_sel->info;
+	enum pipe_shader_type gs_type = shader->selector->type;
+	const struct si_shader_selector *es_sel =
+		shader->previous_stage_sel ? shader->previous_stage_sel : shader->selector;
+	const struct tgsi_shader_info *es_info = &es_sel->info;
+	enum pipe_shader_type es_type = es_sel->type;
+	unsigned num_user_sgprs;
+	unsigned nparams, es_vgpr_comp_cnt, gs_vgpr_comp_cnt;
+	uint64_t va;
+	unsigned window_space =
+		gs_info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
+	bool es_enable_prim_id = shader->key.mono.u.vs_export_prim_id || es_info->uses_primid;
+	unsigned gs_num_invocations = MAX2(gs_sel->gs_num_invocations, 1);
+	unsigned input_prim = si_get_input_prim(gs_sel);
+	bool break_wave_at_eoi = false;
+	struct si_pm4_state *pm4 = si_get_shader_pm4_state(shader);
+	if (!pm4)
+		return;
+
+	if (es_type == PIPE_SHADER_TESS_EVAL) {
+		pm4->atom.emit = gs_type == PIPE_SHADER_GEOMETRY ? gfx10_emit_shader_ngg_tess_gs
+								 : gfx10_emit_shader_ngg_tess_nogs;
+	} else {
+		pm4->atom.emit = gs_type == PIPE_SHADER_GEOMETRY ? gfx10_emit_shader_ngg_notess_gs
+								 : gfx10_emit_shader_ngg_notess_nogs;
+	}
+
+	va = shader->bo->gpu_address;
+	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
+
+	if (es_type == PIPE_SHADER_VERTEX) {
+		/* VGPR5-8: (VertexID, UserVGPR0, UserVGPR1, UserVGPR2 / InstanceID) */
+		es_vgpr_comp_cnt = shader->info.uses_instanceid ? 3 : 0;
+
+		if (es_info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS]) {
+			num_user_sgprs = SI_SGPR_VS_BLIT_DATA +
+					 es_info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS];
+		} else {
+			num_user_sgprs = si_get_num_vs_user_sgprs(GFX9_VSGS_NUM_USER_SGPR);
+		}
+	} else {
+		assert(es_type == PIPE_SHADER_TESS_EVAL);
+		es_vgpr_comp_cnt = es_enable_prim_id ? 3 : 2;
+		num_user_sgprs = GFX9_TESGS_NUM_USER_SGPR;
+
+		if (es_enable_prim_id || gs_info->uses_primid)
+			break_wave_at_eoi = true;
+	}
+
+	/* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
+	 * VGPR[0:4] are always loaded.
+	 *
+	 * Vertex shaders always need to load VGPR3, because they need to
+	 * pass edge flags for decomposed primitives (such as quads) to the PA
+	 * for the GL_LINE polygon mode to skip rendering lines on inner edges.
+	 */
+	if (gs_info->uses_invocationid || gs_type == PIPE_SHADER_VERTEX)
+		gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID, edge flags. */
+	else if (gs_info->uses_primid)
+		gs_vgpr_comp_cnt = 2; /* VGPR2 contains PrimitiveID. */
+	else if (input_prim >= PIPE_PRIM_TRIANGLES)
+		gs_vgpr_comp_cnt = 1; /* VGPR1 contains offsets 2, 3 */
+	else
+		gs_vgpr_comp_cnt = 0; /* VGPR0 contains offsets 0, 1 */
+
+	si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
+	si_pm4_set_reg(pm4, R_00B324_SPI_SHADER_PGM_HI_ES, va >> 40);
+	si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
+		       S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
+		       S_00B228_FLOAT_MODE(shader->config.float_mode) |
+		       S_00B228_DX10_CLAMP(1) |
+		       S_00B228_MEM_ORDERED(1) |
+		       S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt));
+	si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
+		       S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0) |
+		       S_00B22C_USER_SGPR(num_user_sgprs) |
+		       S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
+		       S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5) |
+		       S_00B22C_OC_LDS_EN(es_type == PIPE_SHADER_TESS_EVAL) |
+		       S_00B22C_LDS_SIZE(shader->config.lds_size));
+
+	/* TODO: Use NO_PC_EXPORT when applicable. */
+	nparams = MAX2(shader->info.nr_param_exports, 1);
+	shader->ctx_reg.ngg.spi_vs_out_config =
+		S_0286C4_VS_EXPORT_COUNT(nparams - 1);
+
+	shader->ctx_reg.ngg.spi_shader_idx_format =
+		S_028708_IDX0_EXPORT_FORMAT(V_028708_SPI_SHADER_1COMP);
+	shader->ctx_reg.ngg.spi_shader_pos_format =
+		S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
+		S_02870C_POS1_EXPORT_FORMAT(shader->info.nr_pos_exports > 1 ?
+					    V_02870C_SPI_SHADER_4COMP :
+					    V_02870C_SPI_SHADER_NONE) |
+		S_02870C_POS2_EXPORT_FORMAT(shader->info.nr_pos_exports > 2 ?
+					    V_02870C_SPI_SHADER_4COMP :
+					    V_02870C_SPI_SHADER_NONE) |
+		S_02870C_POS3_EXPORT_FORMAT(shader->info.nr_pos_exports > 3 ?
+					    V_02870C_SPI_SHADER_4COMP :
+					    V_02870C_SPI_SHADER_NONE);
+
+	shader->ctx_reg.ngg.vgt_primitiveid_en =
+		S_028A84_PRIMITIVEID_EN(es_enable_prim_id) |
+		S_028A84_NGG_DISABLE_PROVOK_REUSE(es_enable_prim_id);
+
+	if (gs_type == PIPE_SHADER_GEOMETRY) {
+		shader->ctx_reg.ngg.vgt_esgs_ring_itemsize = es_sel->esgs_itemsize / 4;
+		shader->ctx_reg.ngg.vgt_gs_max_vert_out = gs_sel->gs_max_out_vertices;
+	} else {
+		shader->ctx_reg.ngg.vgt_esgs_ring_itemsize = 1;
+	}
+
+	if (es_type == PIPE_SHADER_TESS_EVAL)
+		si_set_tesseval_regs(sscreen, es_sel, pm4);
+
+	shader->ctx_reg.ngg.vgt_gs_onchip_cntl =
+		S_028A44_ES_VERTS_PER_SUBGRP(shader->ngg.hw_max_esverts) |
+		S_028A44_GS_PRIMS_PER_SUBGRP(shader->ngg.max_gsprims) |
+		S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader->ngg.max_gsprims * gs_num_invocations);
+	shader->ctx_reg.ngg.ge_max_output_per_subgroup =
+		S_0287FC_MAX_VERTS_PER_SUBGROUP(shader->ngg.max_out_verts);
+	shader->ctx_reg.ngg.ge_ngg_subgrp_cntl =
+		S_028B4C_PRIM_AMP_FACTOR(shader->ngg.prim_amp_factor) |
+		S_028B4C_THDS_PER_SUBGRP(0); /* for fast launch */
+	shader->ctx_reg.ngg.vgt_gs_instance_cnt =
+		S_028B90_CNT(gs_num_invocations) |
+		S_028B90_ENABLE(gs_num_invocations > 1) |
+		S_028B90_EN_MAX_VERT_OUT_PER_GS_INSTANCE(
+			shader->ngg.max_vert_out_per_gs_instance);
+
+	/* User edge flags are set by the pos exports. If user edge flags are
+	 * not used, we must use hw-generated edge flags and pass them via
+	 * the prim export to prevent drawing lines on internal edges of
+	 * decomposed primitives (such as quads) with polygon mode = lines.
+	 *
+	 * TODO: We should combine hw-generated edge flags with user edge
+	 *       flags in the shader.
+	 */
+	shader->ctx_reg.ngg.pa_cl_ngg_cntl =
+		S_028838_INDEX_BUF_EDGE_FLAG_ENA(gs_type == PIPE_SHADER_VERTEX &&
+						 !gs_info->writes_edgeflag);
+
+	shader->ge_cntl =
+		S_03096C_PRIM_GRP_SIZE(shader->ngg.max_gsprims) |
+		S_03096C_VERT_GRP_SIZE(shader->ngg.hw_max_esverts) |
+		S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi);
+
+	if (window_space) {
+		shader->ctx_reg.ngg.pa_cl_vte_cntl =
+			S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
+	} else {
+		shader->ctx_reg.ngg.pa_cl_vte_cntl =
+			S_028818_VTX_W0_FMT(1) |
+			S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
+			S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
+			S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
+	}
+
+	shader->ctx_reg.ngg.vgt_reuse_off =
+		S_028AB4_REUSE_OFF(sscreen->info.family == CHIP_NAVI10 &&
+				   sscreen->info.chip_external_rev == 0x1 &&
+				   es_type == PIPE_SHADER_TESS_EVAL);
+}
+
+static void si_emit_shader_vs(struct si_context *sctx)
+{
+	struct si_shader *shader = sctx->queued.named.vs->shader;
+	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+	if (!shader)
+		return;
+
+	radeon_opt_set_context_reg(sctx, R_028A40_VGT_GS_MODE,
+				   SI_TRACKED_VGT_GS_MODE,
+				   shader->ctx_reg.vs.vgt_gs_mode);
+	radeon_opt_set_context_reg(sctx, R_028A84_VGT_PRIMITIVEID_EN,
+				   SI_TRACKED_VGT_PRIMITIVEID_EN,
+				   shader->ctx_reg.vs.vgt_primitiveid_en);
+
+	if (sctx->chip_class <= GFX8) {
+		radeon_opt_set_context_reg(sctx, R_028AB4_VGT_REUSE_OFF,
+					   SI_TRACKED_VGT_REUSE_OFF,
+					   shader->ctx_reg.vs.vgt_reuse_off);
+	}
+
+	radeon_opt_set_context_reg(sctx, R_0286C4_SPI_VS_OUT_CONFIG,
+				   SI_TRACKED_SPI_VS_OUT_CONFIG,
+				   shader->ctx_reg.vs.spi_vs_out_config);
+
+	radeon_opt_set_context_reg(sctx, R_02870C_SPI_SHADER_POS_FORMAT,
+				   SI_TRACKED_SPI_SHADER_POS_FORMAT,
+				   shader->ctx_reg.vs.spi_shader_pos_format);
+
+	radeon_opt_set_context_reg(sctx, R_028818_PA_CL_VTE_CNTL,
+				   SI_TRACKED_PA_CL_VTE_CNTL,
+				   shader->ctx_reg.vs.pa_cl_vte_cntl);
+
+	if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
+		radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
+					   SI_TRACKED_VGT_TF_PARAM,
+					   shader->vgt_tf_param);
+
+	if (shader->vgt_vertex_reuse_block_cntl)
+		radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
+					   SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
+					   shader->vgt_vertex_reuse_block_cntl);
+
+	if (initial_cdw != sctx->gfx_cs->current.cdw)
+		sctx->context_roll = true;
+}
+
 /**
  * Compute the state for \p shader, which will run as a vertex shader on the
  * hardware.
@@ -872,18 +1300,19 @@ static void si_shader_vs(struct si_screen *sscreen, struct si_shader *shader,
 {
 	const struct tgsi_shader_info *info = &shader->selector->info;
 	struct si_pm4_state *pm4;
-	unsigned num_user_sgprs;
-	unsigned nparams, vgpr_comp_cnt;
+	unsigned num_user_sgprs, vgpr_comp_cnt;
 	uint64_t va;
-	unsigned oc_lds_en;
+	unsigned nparams, oc_lds_en;
 	unsigned window_space =
-	   info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
+		info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
 	bool enable_prim_id = shader->key.mono.u.vs_export_prim_id || info->uses_primid;
 
 	pm4 = si_get_shader_pm4_state(shader);
 	if (!pm4)
 		return;
 
+	pm4->atom.emit = si_emit_shader_vs;
+
 	/* We always write VGT_GS_MODE in the VS state, because every switch
 	 * between different shader pipelines involving a different GS or no
 	 * GS at all involves a switch of the VS (different GS use different
@@ -898,19 +1327,18 @@ static void si_shader_vs(struct si_screen *sscreen, struct si_shader *shader,
 		if (enable_prim_id)
 			mode = V_028A40_GS_SCENARIO_A;
 
-		si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE, S_028A40_MODE(mode));
-		si_pm4_set_reg(pm4, R_028A84_VGT_PRIMITIVEID_EN, enable_prim_id);
+		shader->ctx_reg.vs.vgt_gs_mode = S_028A40_MODE(mode);
+		shader->ctx_reg.vs.vgt_primitiveid_en = enable_prim_id;
 	} else {
-		si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE,
-			       ac_vgt_gs_mode(gs->gs_max_out_vertices,
-					      sscreen->info.chip_class));
-		si_pm4_set_reg(pm4, R_028A84_VGT_PRIMITIVEID_EN, 0);
+		shader->ctx_reg.vs.vgt_gs_mode = ac_vgt_gs_mode(gs->gs_max_out_vertices,
+								sscreen->info.chip_class);
+		shader->ctx_reg.vs.vgt_primitiveid_en = 0;
 	}
 
-	if (sscreen->info.chip_class <= VI) {
+	if (sscreen->info.chip_class <= GFX8) {
 		/* Reuse needs to be set off if we write oViewport. */
-		si_pm4_set_reg(pm4, R_028AB4_VGT_REUSE_OFF,
-			       S_028AB4_REUSE_OFF(info->writes_viewport_index));
+		shader->ctx_reg.vs.vgt_reuse_off =
+				S_028AB4_REUSE_OFF(info->writes_viewport_index);
 	}
 
 	va = shader->bo->gpu_address;
@@ -940,49 +1368,55 @@ static void si_shader_vs(struct si_screen *sscreen, struct si_shader *shader,
 
 	/* VS is required to export at least one param. */
 	nparams = MAX2(shader->info.nr_param_exports, 1);
-	si_pm4_set_reg(pm4, R_0286C4_SPI_VS_OUT_CONFIG,
-		       S_0286C4_VS_EXPORT_COUNT(nparams - 1));
-
-	si_pm4_set_reg(pm4, R_02870C_SPI_SHADER_POS_FORMAT,
-		       S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
-		       S_02870C_POS1_EXPORT_FORMAT(shader->info.nr_pos_exports > 1 ?
-						   V_02870C_SPI_SHADER_4COMP :
-						   V_02870C_SPI_SHADER_NONE) |
-		       S_02870C_POS2_EXPORT_FORMAT(shader->info.nr_pos_exports > 2 ?
-						   V_02870C_SPI_SHADER_4COMP :
-						   V_02870C_SPI_SHADER_NONE) |
-		       S_02870C_POS3_EXPORT_FORMAT(shader->info.nr_pos_exports > 3 ?
-						   V_02870C_SPI_SHADER_4COMP :
-						   V_02870C_SPI_SHADER_NONE));
+	shader->ctx_reg.vs.spi_vs_out_config = S_0286C4_VS_EXPORT_COUNT(nparams - 1);
+
+	shader->ctx_reg.vs.spi_shader_pos_format =
+			S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
+			S_02870C_POS1_EXPORT_FORMAT(shader->info.nr_pos_exports > 1 ?
+						    V_02870C_SPI_SHADER_4COMP :
+						    V_02870C_SPI_SHADER_NONE) |
+			S_02870C_POS2_EXPORT_FORMAT(shader->info.nr_pos_exports > 2 ?
+						    V_02870C_SPI_SHADER_4COMP :
+						    V_02870C_SPI_SHADER_NONE) |
+			S_02870C_POS3_EXPORT_FORMAT(shader->info.nr_pos_exports > 3 ?
+						    V_02870C_SPI_SHADER_4COMP :
+						    V_02870C_SPI_SHADER_NONE);
 
 	oc_lds_en = shader->selector->type == PIPE_SHADER_TESS_EVAL ? 1 : 0;
 
 	si_pm4_set_reg(pm4, R_00B120_SPI_SHADER_PGM_LO_VS, va >> 8);
 	si_pm4_set_reg(pm4, R_00B124_SPI_SHADER_PGM_HI_VS, S_00B124_MEM_BASE(va >> 40));
-	si_pm4_set_reg(pm4, R_00B128_SPI_SHADER_PGM_RSRC1_VS,
-		       S_00B128_VGPRS((shader->config.num_vgprs - 1) / 4) |
-		       S_00B128_SGPRS((shader->config.num_sgprs - 1) / 8) |
-		       S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt) |
-		       S_00B128_DX10_CLAMP(1) |
-		       S_00B128_FLOAT_MODE(shader->config.float_mode));
-	si_pm4_set_reg(pm4, R_00B12C_SPI_SHADER_PGM_RSRC2_VS,
-		       S_00B12C_USER_SGPR(num_user_sgprs) |
-		       S_00B12C_OC_LDS_EN(oc_lds_en) |
-		       S_00B12C_SO_BASE0_EN(!!shader->selector->so.stride[0]) |
-		       S_00B12C_SO_BASE1_EN(!!shader->selector->so.stride[1]) |
-		       S_00B12C_SO_BASE2_EN(!!shader->selector->so.stride[2]) |
-		       S_00B12C_SO_BASE3_EN(!!shader->selector->so.stride[3]) |
-		       S_00B12C_SO_EN(!!shader->selector->so.num_outputs) |
-		       S_00B12C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
+
+	uint32_t rsrc1 = S_00B128_VGPRS((shader->config.num_vgprs - 1) / 4) |
+			 S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt) |
+			 S_00B128_DX10_CLAMP(1) |
+			 S_00B128_MEM_ORDERED(sscreen->info.chip_class >= GFX10) |
+			 S_00B128_FLOAT_MODE(shader->config.float_mode);
+	uint32_t rsrc2 = S_00B12C_USER_SGPR(num_user_sgprs) |
+			 S_00B12C_OC_LDS_EN(oc_lds_en) |
+			 S_00B12C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
+
+	if (sscreen->info.chip_class <= GFX9) {
+		rsrc1 |= S_00B128_SGPRS((shader->config.num_sgprs - 1) / 8);
+		rsrc2 |= S_00B12C_SO_BASE0_EN(!!shader->selector->so.stride[0]) |
+			 S_00B12C_SO_BASE1_EN(!!shader->selector->so.stride[1]) |
+			 S_00B12C_SO_BASE2_EN(!!shader->selector->so.stride[2]) |
+			 S_00B12C_SO_BASE3_EN(!!shader->selector->so.stride[3]) |
+			 S_00B12C_SO_EN(!!shader->selector->so.num_outputs);
+	}
+
+	si_pm4_set_reg(pm4, R_00B128_SPI_SHADER_PGM_RSRC1_VS, rsrc1);
+	si_pm4_set_reg(pm4, R_00B12C_SPI_SHADER_PGM_RSRC2_VS, rsrc2);
+
 	if (window_space)
-		si_pm4_set_reg(pm4, R_028818_PA_CL_VTE_CNTL,
-			       S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1));
+		shader->ctx_reg.vs.pa_cl_vte_cntl =
+				S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
 	else
-		si_pm4_set_reg(pm4, R_028818_PA_CL_VTE_CNTL,
-			       S_028818_VTX_W0_FMT(1) |
-			       S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
-			       S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
-			       S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
+		shader->ctx_reg.vs.pa_cl_vte_cntl =
+				S_028818_VTX_W0_FMT(1) |
+				S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
+				S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
+				S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
 
 	if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
 		si_set_tesseval_regs(sscreen, shader->selector, pm4);
@@ -1017,7 +1451,42 @@ static unsigned si_get_spi_shader_col_format(struct si_shader *shader)
 	return value;
 }
 
-static void si_shader_ps(struct si_shader *shader)
+static void si_emit_shader_ps(struct si_context *sctx)
+{
+	struct si_shader *shader = sctx->queued.named.ps->shader;
+	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+	if (!shader)
+		return;
+
+	/* R_0286CC_SPI_PS_INPUT_ENA, R_0286D0_SPI_PS_INPUT_ADDR*/
+	radeon_opt_set_context_reg2(sctx, R_0286CC_SPI_PS_INPUT_ENA,
+				    SI_TRACKED_SPI_PS_INPUT_ENA,
+				    shader->ctx_reg.ps.spi_ps_input_ena,
+				    shader->ctx_reg.ps.spi_ps_input_addr);
+
+	radeon_opt_set_context_reg(sctx, R_0286E0_SPI_BARYC_CNTL,
+				   SI_TRACKED_SPI_BARYC_CNTL,
+				   shader->ctx_reg.ps.spi_baryc_cntl);
+	radeon_opt_set_context_reg(sctx, R_0286D8_SPI_PS_IN_CONTROL,
+				   SI_TRACKED_SPI_PS_IN_CONTROL,
+				   shader->ctx_reg.ps.spi_ps_in_control);
+
+	/* R_028710_SPI_SHADER_Z_FORMAT, R_028714_SPI_SHADER_COL_FORMAT */
+	radeon_opt_set_context_reg2(sctx, R_028710_SPI_SHADER_Z_FORMAT,
+				    SI_TRACKED_SPI_SHADER_Z_FORMAT,
+				    shader->ctx_reg.ps.spi_shader_z_format,
+				    shader->ctx_reg.ps.spi_shader_col_format);
+
+	radeon_opt_set_context_reg(sctx, R_02823C_CB_SHADER_MASK,
+				   SI_TRACKED_CB_SHADER_MASK,
+				   shader->ctx_reg.ps.cb_shader_mask);
+
+	if (initial_cdw != sctx->gfx_cs->current.cdw)
+		sctx->context_roll = true;
+}
+
+static void si_shader_ps(struct si_screen *sscreen, struct si_shader *shader)
 {
 	struct tgsi_shader_info *info = &shader->selector->info;
 	struct si_pm4_state *pm4;
@@ -1074,6 +1543,8 @@ static void si_shader_ps(struct si_shader *shader)
 	if (!pm4)
 		return;
 
+	pm4->atom.emit = si_emit_shader_ps;
+
 	/* SPI_BARYC_CNTL.POS_FLOAT_LOCATION
 	 * Possible vaules:
 	 * 0 -> Position = pixel center
@@ -1116,35 +1587,37 @@ static void si_shader_ps(struct si_shader *shader)
 	    !info->writes_z && !info->writes_stencil && !info->writes_samplemask)
 		spi_shader_col_format = V_028714_SPI_SHADER_32_R;
 
-	si_pm4_set_reg(pm4, R_0286CC_SPI_PS_INPUT_ENA, input_ena);
-	si_pm4_set_reg(pm4, R_0286D0_SPI_PS_INPUT_ADDR,
-		       shader->config.spi_ps_input_addr);
+	shader->ctx_reg.ps.spi_ps_input_ena = input_ena;
+	shader->ctx_reg.ps.spi_ps_input_addr = shader->config.spi_ps_input_addr;
 
 	/* Set interpolation controls. */
 	spi_ps_in_control = S_0286D8_NUM_INTERP(si_get_ps_num_interp(shader));
 
-	/* Set registers. */
-	si_pm4_set_reg(pm4, R_0286E0_SPI_BARYC_CNTL, spi_baryc_cntl);
-	si_pm4_set_reg(pm4, R_0286D8_SPI_PS_IN_CONTROL, spi_ps_in_control);
-
-	si_pm4_set_reg(pm4, R_028710_SPI_SHADER_Z_FORMAT,
-		       ac_get_spi_shader_z_format(info->writes_z,
-						  info->writes_stencil,
-						  info->writes_samplemask));
-
-	si_pm4_set_reg(pm4, R_028714_SPI_SHADER_COL_FORMAT, spi_shader_col_format);
-	si_pm4_set_reg(pm4, R_02823C_CB_SHADER_MASK, cb_shader_mask);
+	shader->ctx_reg.ps.spi_baryc_cntl = spi_baryc_cntl;
+	shader->ctx_reg.ps.spi_ps_in_control = spi_ps_in_control;
+	shader->ctx_reg.ps.spi_shader_z_format =
+			ac_get_spi_shader_z_format(info->writes_z,
+						   info->writes_stencil,
+						   info->writes_samplemask);
+	shader->ctx_reg.ps.spi_shader_col_format = spi_shader_col_format;
+	shader->ctx_reg.ps.cb_shader_mask = cb_shader_mask;
 
 	va = shader->bo->gpu_address;
 	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
 	si_pm4_set_reg(pm4, R_00B020_SPI_SHADER_PGM_LO_PS, va >> 8);
 	si_pm4_set_reg(pm4, R_00B024_SPI_SHADER_PGM_HI_PS, S_00B024_MEM_BASE(va >> 40));
 
-	si_pm4_set_reg(pm4, R_00B028_SPI_SHADER_PGM_RSRC1_PS,
-		       S_00B028_VGPRS((shader->config.num_vgprs - 1) / 4) |
-		       S_00B028_SGPRS((shader->config.num_sgprs - 1) / 8) |
-		       S_00B028_DX10_CLAMP(1) |
-		       S_00B028_FLOAT_MODE(shader->config.float_mode));
+	uint32_t rsrc1 =
+		S_00B028_VGPRS((shader->config.num_vgprs - 1) / 4) |
+		S_00B028_DX10_CLAMP(1) |
+		S_00B028_MEM_ORDERED(sscreen->info.chip_class >= GFX10) |
+		S_00B028_FLOAT_MODE(shader->config.float_mode);
+
+	if (sscreen->info.chip_class < GFX10) {
+		rsrc1 |= S_00B028_SGPRS((shader->config.num_sgprs - 1) / 8);
+	}
+
+	si_pm4_set_reg(pm4, R_00B028_SPI_SHADER_PGM_RSRC1_PS, rsrc1);
 	si_pm4_set_reg(pm4, R_00B02C_SPI_SHADER_PGM_RSRC2_PS,
 		       S_00B02C_EXTRA_LDS_SIZE(shader->config.lds_size) |
 		       S_00B02C_USER_SGPR(SI_PS_NUM_USER_SGPR) |
@@ -1160,6 +1633,8 @@ static void si_shader_init_pm4_state(struct si_screen *sscreen,
 			si_shader_ls(sscreen, shader);
 		else if (shader->key.as_es)
 			si_shader_es(sscreen, shader);
+		else if (shader->key.as_ngg)
+			gfx10_shader_ngg(sscreen, shader);
 		else
 			si_shader_vs(sscreen, shader, NULL);
 		break;
@@ -1169,14 +1644,19 @@ static void si_shader_init_pm4_state(struct si_screen *sscreen,
 	case PIPE_SHADER_TESS_EVAL:
 		if (shader->key.as_es)
 			si_shader_es(sscreen, shader);
+		else if (shader->key.as_ngg)
+			gfx10_shader_ngg(sscreen, shader);
 		else
 			si_shader_vs(sscreen, shader, NULL);
 		break;
 	case PIPE_SHADER_GEOMETRY:
-		si_shader_gs(sscreen, shader);
+		if (shader->key.as_ngg)
+			gfx10_shader_ngg(sscreen, shader);
+		else
+			si_shader_gs(sscreen, shader);
 		break;
 	case PIPE_SHADER_FRAGMENT:
-		si_shader_ps(shader);
+		si_shader_ps(sscreen, shader);
 		break;
 	default:
 		assert(0);
@@ -1192,27 +1672,53 @@ static unsigned si_get_alpha_test_func(struct si_context *sctx)
 	return PIPE_FUNC_ALWAYS;
 }
 
-static void si_shader_selector_key_vs(struct si_context *sctx,
-				      struct si_shader_selector *vs,
-				      struct si_shader_key *key,
-				      struct si_vs_prolog_bits *prolog_key)
+void si_shader_selector_key_vs(struct si_context *sctx,
+			       struct si_shader_selector *vs,
+			       struct si_shader_key *key,
+			       struct si_vs_prolog_bits *prolog_key)
 {
-	if (!sctx->vertex_elements)
+	if (!sctx->vertex_elements ||
+	    vs->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS])
 		return;
 
-	prolog_key->instance_divisor_is_one =
-		sctx->vertex_elements->instance_divisor_is_one;
-	prolog_key->instance_divisor_is_fetched =
-		sctx->vertex_elements->instance_divisor_is_fetched;
+	struct si_vertex_elements *elts = sctx->vertex_elements;
+
+	prolog_key->instance_divisor_is_one = elts->instance_divisor_is_one;
+	prolog_key->instance_divisor_is_fetched = elts->instance_divisor_is_fetched;
+	prolog_key->unpack_instance_id_from_vertex_id =
+		sctx->prim_discard_cs_instancing;
 
 	/* Prefer a monolithic shader to allow scheduling divisions around
 	 * VBO loads. */
 	if (prolog_key->instance_divisor_is_fetched)
 		key->opt.prefer_mono = 1;
 
-	unsigned count = MIN2(vs->info.num_inputs,
-			      sctx->vertex_elements->count);
-	memcpy(key->mono.vs_fix_fetch, sctx->vertex_elements->fix_fetch, count);
+	unsigned count = MIN2(vs->info.num_inputs, elts->count);
+	unsigned count_mask = (1 << count) - 1;
+	unsigned fix = elts->fix_fetch_always & count_mask;
+	unsigned opencode = elts->fix_fetch_opencode & count_mask;
+
+	if (sctx->vertex_buffer_unaligned & elts->vb_alignment_check_mask) {
+		uint32_t mask = elts->fix_fetch_unaligned & count_mask;
+		while (mask) {
+			unsigned i = u_bit_scan(&mask);
+			unsigned log_hw_load_size = 1 + ((elts->hw_load_is_dword >> i) & 1);
+			unsigned vbidx = elts->vertex_buffer_index[i];
+			struct pipe_vertex_buffer *vb = &sctx->vertex_buffer[vbidx];
+			unsigned align_mask = (1 << log_hw_load_size) - 1;
+			if (vb->buffer_offset & align_mask ||
+			    vb->stride & align_mask) {
+				fix |= 1 << i;
+				opencode |= 1 << i;
+			}
+		}
+	}
+
+	while (fix) {
+		unsigned i = u_bit_scan(&fix);
+		key->mono.vs_fix_fetch[i].bits = elts->fix_fetch[i];
+	}
+	key->mono.vs_fetch_opencode = opencode;
 }
 
 static void si_shader_selector_key_hw_vs(struct si_context *sctx,
@@ -1267,6 +1773,7 @@ static void si_shader_selector_key_hw_vs(struct si_context *sctx,
 /* Compute the key for the hw shader variant */
 static inline void si_shader_selector_key(struct pipe_context *ctx,
 					  struct si_shader_selector *sel,
+					  union si_vgt_stages_key stages_key,
 					  struct si_shader_key *key)
 {
 	struct si_context *sctx = (struct si_context *)ctx;
@@ -1282,6 +1789,7 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
 		else if (sctx->gs_shader.cso)
 			key->as_es = 1;
 		else {
+			key->as_ngg = stages_key.u.ngg;
 			si_shader_selector_key_hw_vs(sctx, sel, key);
 
 			if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
@@ -1323,6 +1831,7 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
 		if (sctx->gs_shader.cso)
 			key->as_es = 1;
 		else {
+			key->as_ngg = stages_key.u.ngg;
 			si_shader_selector_key_hw_vs(sctx, sel, key);
 
 			if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
@@ -1340,6 +1849,8 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
 				key->part.gs.prolog.gfx9_prev_is_vs = 1;
 			}
 
+			key->as_ngg = stages_key.u.ngg;
+
 			/* Merged ES-GS can have unbalanced wave usage.
 			 *
 			 * ES threads are per-vertex, while GS threads are
@@ -1400,11 +1911,11 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
 		    blend && blend->alpha_to_coverage)
 			key->part.ps.epilog.spi_shader_col_format |= V_028710_SPI_SHADER_32_AR;
 
-		/* On SI and CIK except Hawaii, the CB doesn't clamp outputs
+		/* On GFX6 and GFX7 except Hawaii, the CB doesn't clamp outputs
 		 * to the range supported by the type if a channel has less
 		 * than 16 bits and the export format is 16_ABGR.
 		 */
-		if (sctx->chip_class <= CIK && sctx->family != CHIP_HAWAII) {
+		if (sctx->chip_class <= GFX7 && sctx->family != CHIP_HAWAII) {
 			key->part.ps.epilog.color_is_int8 = sctx->framebuffer.color_is_int8;
 			key->part.ps.epilog.color_is_int10 = sctx->framebuffer.color_is_int10;
 		}
@@ -1478,7 +1989,7 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
 		key->part.ps.epilog.alpha_func = si_get_alpha_test_func(sctx);
 
 		/* ps_uses_fbfetch is true only if the color buffer is bound. */
-		if (sctx->ps_uses_fbfetch) {
+		if (sctx->ps_uses_fbfetch && !sctx->blitter->running) {
 			struct pipe_surface *cb0 = sctx->framebuffer.state.cbufs[0];
 			struct pipe_resource *tex = cb0->texture;
 
@@ -1511,7 +2022,6 @@ static void si_build_shader_variant(struct si_shader *shader,
 	struct si_screen *sscreen = sel->screen;
 	struct ac_llvm_compiler *compiler;
 	struct pipe_debug_callback *debug = &shader->compiler_ctx_state.debug;
-	int r;
 
 	if (thread_index >= 0) {
 		if (low_priority) {
@@ -1528,10 +2038,9 @@ static void si_build_shader_variant(struct si_shader *shader,
 		compiler = shader->compiler_ctx_state.compiler;
 	}
 
-	r = si_shader_create(sscreen, compiler, shader, debug);
-	if (unlikely(r)) {
-		PRINT_ERR("Failed to build shader variant (type=%u) %d\n",
-			 sel->type, r);
+	if (unlikely(!si_shader_create(sscreen, compiler, shader, debug))) {
+		PRINT_ERR("Failed to build shader variant (type=%u)\n",
+			  sel->type);
 		shader->compilation_failed = true;
 		return;
 	}
@@ -1540,7 +2049,7 @@ static void si_build_shader_variant(struct si_shader *shader,
 		FILE *f = open_memstream(&shader->shader_log,
 					 &shader->shader_log_size);
 		if (f) {
-			si_shader_dump(sscreen, shader, NULL, sel->type, f, false);
+			si_shader_dump(sscreen, shader, NULL, f, false);
 			fclose(f);
 		}
 	}
@@ -1580,6 +2089,7 @@ static bool si_check_missing_main_part(struct si_screen *sscreen,
 		main_part->selector = sel;
 		main_part->key.as_es = key->as_es;
 		main_part->key.as_ls = key->as_ls;
+		main_part->key.as_ngg = key->as_ngg;
 		main_part->is_monolithic = false;
 
 		if (si_compile_tgsi_shader(sscreen, compiler_state->compiler,
@@ -1592,12 +2102,19 @@ static bool si_check_missing_main_part(struct si_screen *sscreen,
 	return true;
 }
 
-/* Select the hw shader variant depending on the current state. */
-static int si_shader_select_with_key(struct si_screen *sscreen,
-				     struct si_shader_ctx_state *state,
-				     struct si_compiler_ctx_state *compiler_state,
-				     struct si_shader_key *key,
-				     int thread_index)
+/**
+ * Select a shader variant according to the shader key.
+ *
+ * \param optimized_or_none  If the key describes an optimized shader variant and
+ *                           the compilation isn't finished, don't select any
+ *                           shader and return an error.
+ */
+int si_shader_select_with_key(struct si_screen *sscreen,
+			      struct si_shader_ctx_state *state,
+			      struct si_compiler_ctx_state *compiler_state,
+			      struct si_shader_key *key,
+			      int thread_index,
+			      bool optimized_or_none)
 {
 	struct si_shader_selector *sel = state->cso;
 	struct si_shader_selector *previous_stage_sel = NULL;
@@ -1613,6 +2130,9 @@ again:
 		   memcmp(&current->key, key, sizeof(*key)) == 0)) {
 		if (unlikely(!util_queue_fence_is_signalled(&current->ready))) {
 			if (current->is_optimized) {
+				if (optimized_or_none)
+					return -1;
+
 				memset(&key->opt, 0, sizeof(key->opt));
 				goto current_not_ready;
 			}
@@ -1649,6 +2169,8 @@ current_not_ready:
 				 * shader so as not to cause a stall due to compilation.
 				 */
 				if (iter->is_optimized) {
+					if (optimized_or_none)
+						return -1;
 					memset(&key->opt, 0, sizeof(key->opt));
 					goto again;
 				}
@@ -1690,19 +2212,26 @@ current_not_ready:
 			util_queue_fence_wait(&previous_stage_sel->ready);
 	}
 
-	/* Compile the main shader part if it doesn't exist. This can happen
-	 * if the initial guess was wrong. */
 	bool is_pure_monolithic =
 		sscreen->use_monolithic_shaders ||
 		memcmp(&key->mono, &zeroed.mono, sizeof(key->mono)) != 0;
 
-	if (!is_pure_monolithic) {
-		bool ok;
+	/* Compile the main shader part if it doesn't exist. This can happen
+	 * if the initial guess was wrong.
+	 *
+	 * The prim discard CS doesn't need the main shader part.
+	 */
+	if (!is_pure_monolithic &&
+	    !key->opt.vs_as_prim_discard_cs) {
+		bool ok = true;
 
 		/* Make sure the main shader part is present. This is needed
 		 * for shaders that can be compiled as VS, LS, or ES, and only
 		 * one of them is compiled at creation.
 		 *
+		 * It is also needed for GS, which can be compiled as non-NGG
+		 * and NGG.
+		 *
 		 * For merged shaders, check that the starting shader's main
 		 * part is present.
 		 */
@@ -1721,10 +2250,13 @@ current_not_ready:
 							previous_stage_sel,
 							compiler_state, &shader1_key);
 			mtx_unlock(&previous_stage_sel->mutex);
-		} else {
+		}
+
+		if (ok) {
 			ok = si_check_missing_main_part(sscreen, sel,
 							compiler_state, key);
 		}
+
 		if (!ok) {
 			FREE(shader);
 			mtx_unlock(&sel->mutex);
@@ -1747,14 +2279,13 @@ current_not_ready:
 		is_pure_monolithic ||
 		memcmp(&key->opt, &zeroed.opt, sizeof(key->opt)) != 0;
 
+	/* The prim discard CS is always optimized. */
 	shader->is_optimized =
-		!is_pure_monolithic &&
-		memcmp(&key->opt, &zeroed.opt, sizeof(key->opt)) != 0;
+		(!is_pure_monolithic || key->opt.vs_as_prim_discard_cs) &&
+		 memcmp(&key->opt, &zeroed.opt, sizeof(key->opt)) != 0;
 
 	/* If it's an optimized shader, compile it asynchronously. */
-	if (shader->is_optimized &&
-	    !is_pure_monolithic &&
-	    thread_index < 0) {
+	if (shader->is_optimized && thread_index < 0) {
 		/* Compile it asynchronously. */
 		util_queue_add_job(&sscreen->shader_compiler_queue_low_priority,
 				   shader, &shader->ready,
@@ -1773,6 +2304,12 @@ current_not_ready:
 		/* Use the default (unoptimized) shader for now. */
 		memset(&key->opt, 0, sizeof(key->opt));
 		mtx_unlock(&sel->mutex);
+
+		if (sscreen->options.sync_compile)
+			util_queue_fence_wait(&shader->ready);
+
+		if (optimized_or_none)
+			return -1;
 		goto again;
 	}
 
@@ -1802,14 +2339,15 @@ current_not_ready:
 
 static int si_shader_select(struct pipe_context *ctx,
 			    struct si_shader_ctx_state *state,
+			    union si_vgt_stages_key stages_key,
 			    struct si_compiler_ctx_state *compiler_state)
 {
 	struct si_context *sctx = (struct si_context *)ctx;
 	struct si_shader_key key;
 
-	si_shader_selector_key(ctx, state->cso, &key);
+	si_shader_selector_key(ctx, state->cso, stages_key, &key);
 	return si_shader_select_with_key(sctx->screen, state, compiler_state,
-					 &key, -1);
+					 &key, -1, false);
 }
 
 static void si_parse_next_shader_property(const struct tgsi_shader_info *info,
@@ -1864,6 +2402,9 @@ static void si_init_shader_selector_async(void *job, int thread_index)
 	assert(thread_index < ARRAY_SIZE(sscreen->compiler));
 	compiler = &sscreen->compiler[thread_index];
 
+	if (sel->nir)
+		si_lower_nir(sel);
+
 	/* Compile the main shader part for use with a prolog and/or epilog.
 	 * If this fails, the driver will try to compile a monolithic shader
 	 * on demand.
@@ -1886,6 +2427,13 @@ static void si_init_shader_selector_async(void *job, int thread_index)
 		si_parse_next_shader_property(&sel->info,
 					      sel->so.num_outputs != 0,
 					      &shader->key);
+		if (sscreen->info.chip_class >= GFX10 &&
+		    !sscreen->options.disable_ngg &&
+		    (((sel->type == PIPE_SHADER_VERTEX ||
+		       sel->type == PIPE_SHADER_TESS_EVAL) &&
+		      !shader->key.as_ls && !shader->key.as_es) ||
+		     sel->type == PIPE_SHADER_GEOMETRY))
+			shader->key.as_ngg = 1;
 
 		if (sel->tokens || sel->nir)
 			ir_binary = si_get_ir_binary(sel);
@@ -1896,7 +2444,7 @@ static void si_init_shader_selector_async(void *job, int thread_index)
 		if (ir_binary &&
 		    si_shader_cache_load_shader(sscreen, ir_binary, shader)) {
 			mtx_unlock(&sscreen->shader_cache_mutex);
-			si_shader_dump_stats_for_shader_db(shader, debug);
+			si_shader_dump_stats_for_shader_db(sscreen, shader, debug);
 		} else {
 			mtx_unlock(&sscreen->shader_cache_mutex);
 
@@ -1962,7 +2510,11 @@ static void si_init_shader_selector_async(void *job, int thread_index)
 		}
 	}
 
-	/* The GS copy shader is always pre-compiled. */
+	/* The GS copy shader is always pre-compiled.
+	 *
+	 * TODO-GFX10: We could compile the GS copy shader on demand, since it
+	 * is only used in the (rare) non-NGG case.
+	 */
 	if (sel->type == PIPE_SHADER_GEOMETRY) {
 		sel->gs_copy_shader = si_generate_gs_copy_shader(sscreen, compiler, sel, debug);
 		if (!sel->gs_copy_shader) {
@@ -1982,12 +2534,12 @@ void si_schedule_initial_compile(struct si_context *sctx, unsigned processor,
 	util_queue_fence_init(ready_fence);
 
 	struct util_async_debug_callback async_debug;
-	bool wait =
+	bool debug =
 		(sctx->debug.debug_message && !sctx->debug.async) ||
 		sctx->is_debug ||
 		si_can_dump_shader(sctx->screen, processor);
 
-	if (wait) {
+	if (debug) {
 		u_async_debug_init(&async_debug);
 		compiler_ctx_state->debug = async_debug.base;
 	}
@@ -1995,11 +2547,14 @@ void si_schedule_initial_compile(struct si_context *sctx, unsigned processor,
 	util_queue_add_job(&sctx->screen->shader_compiler_queue, job,
 			   ready_fence, execute, NULL);
 
-	if (wait) {
+	if (debug) {
 		util_queue_fence_wait(ready_fence);
 		u_async_debug_drain(&async_debug, &sctx->debug);
 		u_async_debug_cleanup(&async_debug);
 	}
+
+	if (sctx->screen->options.sync_compile)
+		util_queue_fence_wait(ready_fence);
 }
 
 /* Return descriptor slot usage masks from the given shader info. */
@@ -2058,10 +2613,9 @@ static void *si_create_shader_selector(struct pipe_context *ctx,
 
 		sel->nir = state->ir.nir;
 
+		si_nir_opts(sel->nir);
 		si_nir_scan_shader(sel->nir, &sel->info);
-		si_nir_scan_tess_ctrl(sel->nir, &sel->info, &sel->tcs_info);
-
-		si_lower_nir(sel);
+		si_nir_scan_tess_ctrl(sel->nir, &sel->tcs_info);
 	}
 
 	sel->type = sel->info.processor;
@@ -2088,6 +2642,15 @@ static void *si_create_shader_selector(struct pipe_context *ctx,
 		sel->info.uses_kill &&
 		sctx->screen->debug_flags & DBG(FS_CORRECT_DERIVS_AFTER_KILL);
 
+	sel->prim_discard_cs_allowed =
+		sel->type == PIPE_SHADER_VERTEX &&
+		!sel->info.uses_bindless_images &&
+		!sel->info.uses_bindless_samplers &&
+		!sel->info.writes_memory &&
+		!sel->info.writes_viewport_index &&
+		!sel->info.properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION] &&
+		!sel->so.num_outputs;
+
 	/* Set which opcode uses which (i,j) pair. */
 	if (sel->info.uses_persp_opcode_interp_centroid)
 		sel->info.uses_persp_centroid = true;
@@ -2107,6 +2670,12 @@ static void *si_create_shader_selector(struct pipe_context *ctx,
 	case PIPE_SHADER_GEOMETRY:
 		sel->gs_output_prim =
 			sel->info.properties[TGSI_PROPERTY_GS_OUTPUT_PRIM];
+
+		/* Only possibilities: POINTS, LINE_STRIP, TRIANGLES */
+		sel->rast_prim = sel->gs_output_prim;
+		if (util_rast_prim_is_triangles(sel->rast_prim))
+			sel->rast_prim = PIPE_PRIM_TRIANGLES;
+
 		sel->gs_max_out_vertices =
 			sel->info.properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
 		sel->gs_num_invocations =
@@ -2175,6 +2744,14 @@ static void *si_create_shader_selector(struct pipe_context *ctx,
 			sel->esgs_itemsize += 4;
 
 		assert(((sel->esgs_itemsize / 4) & C_028AAC_ITEMSIZE) == 0);
+
+		/* Only for TES: */
+		if (sel->info.properties[TGSI_PROPERTY_TES_POINT_MODE])
+			sel->rast_prim = PIPE_PRIM_POINTS;
+		else if (sel->info.properties[TGSI_PROPERTY_TES_PRIM_MODE] == PIPE_PRIM_LINES)
+			sel->rast_prim = PIPE_PRIM_LINE_STRIP;
+		else
+			sel->rast_prim = PIPE_PRIM_TRIANGLES;
 		break;
 
 	case PIPE_SHADER_FRAGMENT:
@@ -2208,6 +2785,7 @@ static void *si_create_shader_selector(struct pipe_context *ctx,
 			}
 		}
 		break;
+	default:;
 	}
 
 	/* PA_CL_VS_OUT_CNTL */
@@ -2368,6 +2946,27 @@ static void si_update_tess_uses_prim_id(struct si_context *sctx)
 		 sctx->ps_shader.cso->info.uses_primid);
 }
 
+static bool si_update_ngg(struct si_context *sctx)
+{
+	if (sctx->chip_class <= GFX9 ||
+	    sctx->screen->options.disable_ngg)
+		return false;
+
+	bool new_ngg = true;
+
+	/* EN_MAX_VERT_OUT_PER_GS_INSTANCE does not work with tesselation. */
+	if (sctx->gs_shader.cso && sctx->tes_shader.cso &&
+	    sctx->gs_shader.cso->gs_num_invocations * sctx->gs_shader.cso->gs_max_out_vertices > 256)
+		new_ngg = false;
+
+	if (new_ngg != sctx->ngg) {
+		sctx->ngg = new_ngg;
+		sctx->last_rast_prim = -1; /* reset this so that it gets updated */
+		return true;
+	}
+	return false;
+}
+
 static void si_bind_gs_shader(struct pipe_context *ctx, void *state)
 {
 	struct si_context *sctx = (struct si_context *)ctx;
@@ -2375,6 +2974,7 @@ static void si_bind_gs_shader(struct pipe_context *ctx, void *state)
 	struct si_shader *old_hw_vs_variant = si_get_vs_state(sctx);
 	struct si_shader_selector *sel = state;
 	bool enable_changed = !!sctx->gs_shader.cso != !!sel;
+	bool ngg_changed;
 
 	if (sctx->gs_shader.cso == sel)
 		return;
@@ -2386,8 +2986,10 @@ static void si_bind_gs_shader(struct pipe_context *ctx, void *state)
 	si_update_common_shader_state(sctx);
 	sctx->last_rast_prim = -1; /* reset this so that it gets updated */
 
-	if (enable_changed) {
+	ngg_changed = si_update_ngg(sctx);
+	if (ngg_changed || enable_changed)
 		si_shader_change_notify(sctx);
+	if (enable_changed) {
 		if (sctx->ia_multi_vgt_param_key.u.uses_tess)
 			si_update_tess_uses_prim_id(sctx);
 	}
@@ -2439,6 +3041,7 @@ static void si_bind_tes_shader(struct pipe_context *ctx, void *state)
 	sctx->last_rast_prim = -1; /* reset this so that it gets updated */
 
 	if (enable_changed) {
+		si_update_ngg(sctx);
 		si_shader_change_notify(sctx);
 		sctx->last_tes_sh_base = -1; /* invalidate derived tess state */
 	}
@@ -2492,14 +3095,21 @@ static void si_delete_shader(struct si_context *sctx, struct si_shader *shader)
 	util_queue_fence_destroy(&shader->ready);
 
 	if (shader->pm4) {
+		/* If destroyed shaders were not unbound, the next compiled
+		 * shader variant could get the same pointer address and so
+		 * binding it to the same shader stage would be considered
+		 * a no-op, causing random behavior.
+		 */
 		switch (shader->selector->type) {
 		case PIPE_SHADER_VERTEX:
 			if (shader->key.as_ls) {
-				assert(sctx->chip_class <= VI);
+				assert(sctx->chip_class <= GFX8);
 				si_pm4_delete_state(sctx, ls, shader->pm4);
 			} else if (shader->key.as_es) {
-				assert(sctx->chip_class <= VI);
+				assert(sctx->chip_class <= GFX8);
 				si_pm4_delete_state(sctx, es, shader->pm4);
+			} else if (shader->key.as_ngg) {
+				si_pm4_delete_state(sctx, gs, shader->pm4);
 			} else {
 				si_pm4_delete_state(sctx, vs, shader->pm4);
 			}
@@ -2509,8 +3119,10 @@ static void si_delete_shader(struct si_context *sctx, struct si_shader *shader)
 			break;
 		case PIPE_SHADER_TESS_EVAL:
 			if (shader->key.as_es) {
-				assert(sctx->chip_class <= VI);
+				assert(sctx->chip_class <= GFX8);
 				si_pm4_delete_state(sctx, es, shader->pm4);
+			} else if (shader->key.as_ngg) {
+				si_pm4_delete_state(sctx, gs, shader->pm4);
 			} else {
 				si_pm4_delete_state(sctx, vs, shader->pm4);
 			}
@@ -2524,6 +3136,7 @@ static void si_delete_shader(struct si_context *sctx, struct si_shader *shader)
 		case PIPE_SHADER_FRAGMENT:
 			si_pm4_delete_state(sctx, ps, shader->pm4);
 			break;
+		default:;
 		}
 	}
 
@@ -2563,6 +3176,8 @@ void si_destroy_shader_selector(struct si_context *sctx,
 		si_delete_shader(sctx, sel->main_shader_part_ls);
 	if (sel->main_shader_part_es)
 		si_delete_shader(sctx, sel->main_shader_part_es);
+	if (sel->main_shader_part_ngg)
+		si_delete_shader(sctx, sel->main_shader_part_ngg);
 	if (sel->gs_copy_shader)
 		si_delete_shader(sctx, sel->gs_copy_shader);
 
@@ -2589,7 +3204,8 @@ static unsigned si_get_ps_input_cntl(struct si_context *sctx,
 	unsigned j, offset, ps_input_cntl = 0;
 
 	if (interpolate == TGSI_INTERPOLATE_CONSTANT ||
-	    (interpolate == TGSI_INTERPOLATE_COLOR && sctx->flatshade))
+	    (interpolate == TGSI_INTERPOLATE_COLOR && sctx->flatshade) ||
+	    name == TGSI_SEMANTIC_PRIMID)
 		ps_input_cntl |= S_028644_FLAT_SHADE(1);
 
 	if (name == TGSI_SEMANTIC_PCOORD ||
@@ -2624,8 +3240,8 @@ static unsigned si_get_ps_input_cntl(struct si_context *sctx,
 		}
 	}
 
-	if (name == TGSI_SEMANTIC_PRIMID)
-		/* PrimID is written after the last output. */
+	if (j == vsinfo->num_outputs && name == TGSI_SEMANTIC_PRIMID)
+		/* PrimID is written after the last output when HW VS is used. */
 		ps_input_cntl |= S_028644_OFFSET(vs->info.vs_output_param_offset[vsinfo->num_outputs]);
 	else if (j == vsinfo->num_outputs && !G_028644_PT_SPRITE_TEX(ps_input_cntl)) {
 		/* No corresponding output found, load defaults into input.
@@ -2684,9 +3300,13 @@ static void si_emit_spi_map(struct si_context *sctx)
 	/* R_028644_SPI_PS_INPUT_CNTL_0 */
 	/* Dota 2: Only ~16% of SPI map updates set different values. */
 	/* Talos: Only ~9% of SPI map updates set different values. */
+	unsigned initial_cdw = sctx->gfx_cs->current.cdw;
 	radeon_opt_set_context_regn(sctx, R_028644_SPI_PS_INPUT_CNTL_0,
 				    spi_ps_input_cntl,
 				    sctx->tracked_regs.spi_ps_input_cntl, num_interp);
+
+	if (initial_cdw != sctx->gfx_cs->current.cdw)
+		sctx->context_roll = true;
 }
 
 /**
@@ -2722,10 +3342,10 @@ static bool si_update_gs_ring_buffers(struct si_context *sctx)
 	unsigned num_se = sctx->screen->info.max_se;
 	unsigned wave_size = 64;
 	unsigned max_gs_waves = 32 * num_se; /* max 32 per SE on GCN */
-	/* On SI-CI, the value comes from VGT_GS_VERTEX_REUSE = 16.
-	 * On VI+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
+	/* On GFX6-GFX7, the value comes from VGT_GS_VERTEX_REUSE = 16.
+	 * On GFX8+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
 	 */
-	unsigned gs_vertex_reuse = (sctx->chip_class >= VI ? 32 : 16) * num_se;
+	unsigned gs_vertex_reuse = (sctx->chip_class >= GFX8 ? 32 : 16) * num_se;
 	unsigned alignment = 256 * num_se;
 	/* The maximum size is 63.999 MB per SE. */
 	unsigned max_size = ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se;
@@ -2752,7 +3372,7 @@ static bool si_update_gs_ring_buffers(struct si_context *sctx)
 	 *
 	 * GFX9 doesn't have the ESGS ring.
 	 */
-	bool update_esgs = sctx->chip_class <= VI &&
+	bool update_esgs = sctx->chip_class <= GFX8 &&
 			   esgs_ring_size &&
 			   (!sctx->esgs_ring ||
 			    sctx->esgs_ring->width0 < esgs_ring_size);
@@ -2790,9 +3410,9 @@ static bool si_update_gs_ring_buffers(struct si_context *sctx)
 	if (!pm4)
 		return false;
 
-	if (sctx->chip_class >= CIK) {
+	if (sctx->chip_class >= GFX7) {
 		if (sctx->esgs_ring) {
-			assert(sctx->chip_class <= VI);
+			assert(sctx->chip_class <= GFX8);
 			si_pm4_set_reg(pm4, R_030900_VGT_ESGS_RING_SIZE,
 				       sctx->esgs_ring->width0 / 256);
 		}
@@ -2824,7 +3444,7 @@ static bool si_update_gs_ring_buffers(struct si_context *sctx)
 
 	/* Set ring bindings. */
 	if (sctx->esgs_ring) {
-		assert(sctx->chip_class <= VI);
+		assert(sctx->chip_class <= GFX8);
 		si_set_ring_buffer(sctx, SI_ES_RING_ESGS,
 				   sctx->esgs_ring, 0, sctx->esgs_ring->width0,
 				   true, true, 4, 64, 0);
@@ -2866,7 +3486,6 @@ static int si_update_scratch_buffer(struct si_context *sctx,
 				    struct si_shader *shader)
 {
 	uint64_t scratch_va = sctx->scratch_buffer->gpu_address;
-	int r;
 
 	if (!shader)
 		return 0;
@@ -2891,22 +3510,16 @@ static int si_update_scratch_buffer(struct si_context *sctx,
 
 	assert(sctx->scratch_buffer);
 
-	if (shader->previous_stage)
-		si_shader_apply_scratch_relocs(shader->previous_stage, scratch_va);
-
-	si_shader_apply_scratch_relocs(shader, scratch_va);
-
 	/* Replace the shader bo with a new bo that has the relocs applied. */
-	r = si_shader_binary_upload(sctx->screen, shader);
-	if (r) {
+	if (!si_shader_binary_upload(sctx->screen, shader, scratch_va)) {
 		si_shader_unlock(shader);
-		return r;
+		return -1;
 	}
 
 	/* Update the shader state to use the new shader bo. */
 	si_shader_init_pm4_state(sctx->screen, shader);
 
-	r600_resource_reference(&shader->scratch_bo, sctx->scratch_buffer);
+	si_resource_reference(&shader->scratch_bo, sctx->scratch_buffer);
 
 	si_shader_unlock(shader);
 	return 1;
@@ -2981,10 +3594,12 @@ static bool si_update_scratch_relocs(struct si_context *sctx)
 	if (r < 0)
 		return false;
 	if (r == 1) {
-		if (sctx->tes_shader.current)
+		if (sctx->vs_shader.current->key.as_ls)
 			si_pm4_bind_state(sctx, ls, sctx->vs_shader.current->pm4);
-		else if (sctx->gs_shader.current)
+		else if (sctx->vs_shader.current->key.as_es)
 			si_pm4_bind_state(sctx, es, sctx->vs_shader.current->pm4);
+		else if (sctx->vs_shader.current->key.as_ngg)
+			si_pm4_bind_state(sctx, gs, sctx->vs_shader.current->pm4);
 		else
 			si_pm4_bind_state(sctx, vs, sctx->vs_shader.current->pm4);
 	}
@@ -2994,8 +3609,10 @@ static bool si_update_scratch_relocs(struct si_context *sctx)
 	if (r < 0)
 		return false;
 	if (r == 1) {
-		if (sctx->gs_shader.current)
+		if (sctx->tes_shader.current->key.as_es)
 			si_pm4_bind_state(sctx, es, sctx->tes_shader.current->pm4);
+		else if (sctx->tes_shader.current->key.as_ngg)
+			si_pm4_bind_state(sctx, gs, sctx->tes_shader.current->pm4);
 		else
 			si_pm4_bind_state(sctx, vs, sctx->tes_shader.current->pm4);
 	}
@@ -3016,7 +3633,7 @@ static bool si_update_spi_tmpring_size(struct si_context *sctx)
 	if (scratch_needed_size > 0) {
 		if (scratch_needed_size > current_scratch_buffer_size) {
 			/* Create a bigger scratch buffer */
-			r600_resource_reference(&sctx->scratch_buffer, NULL);
+			si_resource_reference(&sctx->scratch_buffer, NULL);
 
 			sctx->scratch_buffer =
 				si_aligned_buffer_create(&sctx->screen->b,
@@ -3066,19 +3683,22 @@ static void si_init_tess_factor_ring(struct si_context *sctx)
 
 	si_init_config_add_vgt_flush(sctx);
 
-	si_pm4_add_bo(sctx->init_config, r600_resource(sctx->tess_rings),
+	si_pm4_add_bo(sctx->init_config, si_resource(sctx->tess_rings),
 		      RADEON_USAGE_READWRITE, RADEON_PRIO_SHADER_RINGS);
 
-	uint64_t factor_va = r600_resource(sctx->tess_rings)->gpu_address +
+	uint64_t factor_va = si_resource(sctx->tess_rings)->gpu_address +
 			     sctx->screen->tess_offchip_ring_size;
 
 	/* Append these registers to the init config state. */
-	if (sctx->chip_class >= CIK) {
+	if (sctx->chip_class >= GFX7) {
 		si_pm4_set_reg(sctx->init_config, R_030938_VGT_TF_RING_SIZE,
 			       S_030938_SIZE(sctx->screen->tess_factor_ring_size / 4));
 		si_pm4_set_reg(sctx->init_config, R_030940_VGT_TF_MEMORY_BASE,
 			       factor_va >> 8);
-		if (sctx->chip_class >= GFX9)
+		if (sctx->chip_class >= GFX10)
+			si_pm4_set_reg(sctx->init_config, R_030984_VGT_TF_MEMORY_BASE_HI_UMD,
+				       S_030984_BASE_HI(factor_va >> 40));
+		else if (sctx->chip_class == GFX9)
 			si_pm4_set_reg(sctx->init_config, R_030944_VGT_TF_MEMORY_BASE_HI,
 				       S_030944_BASE_HI(factor_va >> 40));
 		si_pm4_set_reg(sctx->init_config, R_03093C_VGT_HS_OFFCHIP_PARAM,
@@ -3100,71 +3720,51 @@ static void si_init_tess_factor_ring(struct si_context *sctx)
 	si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
 }
 
-/**
- * This is used when TCS is NULL in the VS->TCS->TES chain. In this case,
- * VS passes its outputs to TES directly, so the fixed-function shader only
- * has to write TESSOUTER and TESSINNER.
- */
-static void si_generate_fixed_func_tcs(struct si_context *sctx)
+static struct si_pm4_state *si_build_vgt_shader_config(struct si_screen *screen,
+						       union si_vgt_stages_key key)
 {
-	struct ureg_src outer, inner;
-	struct ureg_dst tessouter, tessinner;
-	struct ureg_program *ureg = ureg_create(PIPE_SHADER_TESS_CTRL);
+	struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
+	uint32_t stages = 0;
 
-	if (!ureg)
-		return; /* if we get here, we're screwed */
+	if (key.u.tess) {
+		stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) |
+		          S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
 
-	assert(!sctx->fixed_func_tcs_shader.cso);
-
-	outer = ureg_DECL_system_value(ureg,
-				       TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI, 0);
-	inner = ureg_DECL_system_value(ureg,
-				       TGSI_SEMANTIC_DEFAULT_TESSINNER_SI, 0);
+		if (key.u.gs)
+			stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) |
+				  S_028B54_GS_EN(1);
+		else if (key.u.ngg)
+			stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS);
+		else
+			stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
+	} else if (key.u.gs) {
+		stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) |
+			  S_028B54_GS_EN(1);
+	} else if (key.u.ngg) {
+		stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL);
+	}
 
-	tessouter = ureg_DECL_output(ureg, TGSI_SEMANTIC_TESSOUTER, 0);
-	tessinner = ureg_DECL_output(ureg, TGSI_SEMANTIC_TESSINNER, 0);
+	if (key.u.ngg) {
+		stages |= S_028B54_PRIMGEN_EN(1);
+		if (key.u.streamout)
+			stages |= S_028B54_NGG_WAVE_ID_EN(1);
+	} else if (key.u.gs)
+		stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
 
-	ureg_MOV(ureg, tessouter, outer);
-	ureg_MOV(ureg, tessinner, inner);
-	ureg_END(ureg);
+	if (screen->info.chip_class >= GFX9)
+		stages |= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
 
-	sctx->fixed_func_tcs_shader.cso =
-		ureg_create_shader_and_destroy(ureg, &sctx->b);
+	si_pm4_set_reg(pm4, R_028B54_VGT_SHADER_STAGES_EN, stages);
+	return pm4;
 }
 
-static void si_update_vgt_shader_config(struct si_context *sctx)
+static void si_update_vgt_shader_config(struct si_context *sctx,
+					union si_vgt_stages_key key)
 {
-	/* Calculate the index of the config.
-	 * 0 = VS, 1 = VS+GS, 2 = VS+Tess, 3 = VS+Tess+GS */
-	unsigned index = 2*!!sctx->tes_shader.cso + !!sctx->gs_shader.cso;
-	struct si_pm4_state **pm4 = &sctx->vgt_shader_config[index];
-
-	if (!*pm4) {
-		uint32_t stages = 0;
+	struct si_pm4_state **pm4 = &sctx->vgt_shader_config[key.index];
 
-		*pm4 = CALLOC_STRUCT(si_pm4_state);
-
-		if (sctx->tes_shader.cso) {
-			stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) |
-				  S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
-
-			if (sctx->gs_shader.cso)
-				stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) |
-					  S_028B54_GS_EN(1) |
-				          S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
-			else
-				stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
-		} else if (sctx->gs_shader.cso) {
-			stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) |
-				  S_028B54_GS_EN(1) |
-			          S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
-		}
-
-		if (sctx->chip_class >= GFX9)
-			stages |= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
-
-		si_pm4_set_reg(*pm4, R_028B54_VGT_SHADER_STAGES_EN, stages);
-	}
+	if (unlikely(!*pm4))
+		*pm4 = si_build_vgt_shader_config(sctx->screen, key);
 	si_pm4_bind_state(sctx, vgt_shader_config, *pm4);
 }
 
@@ -3176,6 +3776,7 @@ bool si_update_shaders(struct si_context *sctx)
 	struct si_shader *old_vs = si_get_vs_state(sctx);
 	bool old_clip_disable = old_vs ? old_vs->key.opt.clip_disable : false;
 	struct si_shader *old_ps = sctx->ps_shader.current;
+	union si_vgt_stages_key key;
 	unsigned old_spi_shader_col_format =
 		old_ps ? old_ps->key.part.ps.epilog.spi_shader_col_format : 0;
 	int r;
@@ -3184,7 +3785,25 @@ bool si_update_shaders(struct si_context *sctx)
 	compiler_state.debug = sctx->debug;
 	compiler_state.is_debug_context = sctx->is_debug;
 
-	/* Update stages before GS. */
+	key.index = 0;
+
+	if (sctx->tes_shader.cso)
+		key.u.tess = 1;
+	if (sctx->gs_shader.cso)
+		key.u.gs = 1;
+
+	if (sctx->chip_class >= GFX10) {
+		key.u.ngg = sctx->ngg;
+
+		if (sctx->gs_shader.cso)
+			key.u.streamout = !!sctx->gs_shader.cso->so.num_outputs;
+		else if (sctx->tes_shader.cso)
+			key.u.streamout = !!sctx->tes_shader.cso->so.num_outputs;
+		else
+			key.u.streamout = !!sctx->vs_shader.cso->so.num_outputs;
+	}
+
+	/* Update TCS and TES. */
 	if (sctx->tes_shader.cso) {
 		if (!sctx->tess_rings) {
 			si_init_tess_factor_ring(sctx);
@@ -3192,92 +3811,93 @@ bool si_update_shaders(struct si_context *sctx)
 				return false;
 		}
 
-		/* VS as LS */
-		if (sctx->chip_class <= VI) {
-			r = si_shader_select(ctx, &sctx->vs_shader,
-					     &compiler_state);
-			if (r)
-				return false;
-			si_pm4_bind_state(sctx, ls, sctx->vs_shader.current->pm4);
-		}
-
 		if (sctx->tcs_shader.cso) {
-			r = si_shader_select(ctx, &sctx->tcs_shader,
+			r = si_shader_select(ctx, &sctx->tcs_shader, key,
 					     &compiler_state);
 			if (r)
 				return false;
 			si_pm4_bind_state(sctx, hs, sctx->tcs_shader.current->pm4);
 		} else {
 			if (!sctx->fixed_func_tcs_shader.cso) {
-				si_generate_fixed_func_tcs(sctx);
+				sctx->fixed_func_tcs_shader.cso =
+					si_create_fixed_func_tcs(sctx);
 				if (!sctx->fixed_func_tcs_shader.cso)
 					return false;
 			}
 
 			r = si_shader_select(ctx, &sctx->fixed_func_tcs_shader,
-					     &compiler_state);
+					     key, &compiler_state);
 			if (r)
 				return false;
 			si_pm4_bind_state(sctx, hs,
 					  sctx->fixed_func_tcs_shader.current->pm4);
 		}
 
-		if (sctx->gs_shader.cso) {
-			/* TES as ES */
-			if (sctx->chip_class <= VI) {
-				r = si_shader_select(ctx, &sctx->tes_shader,
-						     &compiler_state);
-				if (r)
-					return false;
-				si_pm4_bind_state(sctx, es, sctx->tes_shader.current->pm4);
-			}
-		} else {
-			/* TES as VS */
-			r = si_shader_select(ctx, &sctx->tes_shader,
-					     &compiler_state);
+		if (!sctx->gs_shader.cso || sctx->chip_class <= GFX8) {
+			r = si_shader_select(ctx, &sctx->tes_shader, key, &compiler_state);
 			if (r)
 				return false;
-			si_pm4_bind_state(sctx, vs, sctx->tes_shader.current->pm4);
-		}
-	} else if (sctx->gs_shader.cso) {
-		if (sctx->chip_class <= VI) {
-			/* VS as ES */
-			r = si_shader_select(ctx, &sctx->vs_shader,
-					     &compiler_state);
-			if (r)
-				return false;
-			si_pm4_bind_state(sctx, es, sctx->vs_shader.current->pm4);
 
-			si_pm4_bind_state(sctx, ls, NULL);
-			si_pm4_bind_state(sctx, hs, NULL);
+			if (sctx->gs_shader.cso) {
+				/* TES as ES */
+				assert(sctx->chip_class <= GFX8);
+				si_pm4_bind_state(sctx, es, sctx->tes_shader.current->pm4);
+			} else if (key.u.ngg) {
+				si_pm4_bind_state(sctx, gs, sctx->tes_shader.current->pm4);
+			} else {
+				si_pm4_bind_state(sctx, vs, sctx->tes_shader.current->pm4);
+			}
 		}
 	} else {
-		/* VS as VS */
-		r = si_shader_select(ctx, &sctx->vs_shader, &compiler_state);
-		if (r)
-			return false;
-		si_pm4_bind_state(sctx, vs, sctx->vs_shader.current->pm4);
-		si_pm4_bind_state(sctx, ls, NULL);
+		if (sctx->chip_class <= GFX8)
+			si_pm4_bind_state(sctx, ls, NULL);
 		si_pm4_bind_state(sctx, hs, NULL);
 	}
 
 	/* Update GS. */
 	if (sctx->gs_shader.cso) {
-		r = si_shader_select(ctx, &sctx->gs_shader, &compiler_state);
+		r = si_shader_select(ctx, &sctx->gs_shader, key, &compiler_state);
 		if (r)
 			return false;
 		si_pm4_bind_state(sctx, gs, sctx->gs_shader.current->pm4);
-		si_pm4_bind_state(sctx, vs, sctx->gs_shader.cso->gs_copy_shader->pm4);
+		if (!key.u.ngg) {
+			si_pm4_bind_state(sctx, vs, sctx->gs_shader.cso->gs_copy_shader->pm4);
 
-		if (!si_update_gs_ring_buffers(sctx))
-			return false;
+			if (!si_update_gs_ring_buffers(sctx))
+				return false;
+		} else {
+			si_pm4_bind_state(sctx, vs, NULL);
+		}
 	} else {
-		si_pm4_bind_state(sctx, gs, NULL);
-		if (sctx->chip_class <= VI)
-			si_pm4_bind_state(sctx, es, NULL);
+		if (!key.u.ngg) {
+			si_pm4_bind_state(sctx, gs, NULL);
+			if (sctx->chip_class <= GFX8)
+				si_pm4_bind_state(sctx, es, NULL);
+		}
 	}
 
-	si_update_vgt_shader_config(sctx);
+	/* Update VS. */
+	if ((!key.u.tess && !key.u.gs) || sctx->chip_class <= GFX8) {
+		r = si_shader_select(ctx, &sctx->vs_shader, key, &compiler_state);
+		if (r)
+			return false;
+
+		if (!key.u.tess && !key.u.gs) {
+			if (key.u.ngg) {
+				si_pm4_bind_state(sctx, gs, sctx->vs_shader.current->pm4);
+				si_pm4_bind_state(sctx, vs, NULL);
+			} else {
+				si_pm4_bind_state(sctx, vs, sctx->vs_shader.current->pm4);
+			}
+		} else if (sctx->tes_shader.cso) {
+			si_pm4_bind_state(sctx, ls, sctx->vs_shader.current->pm4);
+		} else {
+			assert(sctx->gs_shader.cso);
+			si_pm4_bind_state(sctx, es, sctx->vs_shader.current->pm4);
+		}
+	}
+
+	si_update_vgt_shader_config(sctx, key);
 
 	if (old_clip_disable != si_get_vs_state(sctx)->key.opt.clip_disable)
 		si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_regs);
@@ -3285,7 +3905,7 @@ bool si_update_shaders(struct si_context *sctx)
 	if (sctx->ps_shader.cso) {
 		unsigned db_shader_control;
 
-		r = si_shader_select(ctx, &sctx->ps_shader, &compiler_state);
+		r = si_shader_select(ctx, &sctx->ps_shader, key, &compiler_state);
 		if (r)
 			return false;
 		si_pm4_bind_state(sctx, ps, sctx->ps_shader.current->pm4);
@@ -3294,7 +3914,9 @@ bool si_update_shaders(struct si_context *sctx)
 			sctx->ps_shader.cso->db_shader_control |
 			S_02880C_KILL_ENABLE(si_get_alpha_test_func(sctx) != PIPE_FUNC_ALWAYS);
 
-		if (si_pm4_state_changed(sctx, ps) || si_pm4_state_changed(sctx, vs) ||
+		if (si_pm4_state_changed(sctx, ps) ||
+		    si_pm4_state_changed(sctx, vs) ||
+		    (key.u.ngg && si_pm4_state_changed(sctx, gs)) ||
 		    sctx->sprite_coord_enable != rs->sprite_coord_enable ||
 		    sctx->flatshade != rs->flatshade) {
 			sctx->sprite_coord_enable = rs->sprite_coord_enable;
@@ -3320,7 +3942,7 @@ bool si_update_shaders(struct si_context *sctx)
 			sctx->smoothing_enabled = sctx->ps_shader.current->key.part.ps.epilog.poly_line_smoothing;
 			si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
 
-			if (sctx->chip_class == SI)
+			if (sctx->chip_class == GFX6)
 				si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
 
 			if (sctx->framebuffer.nr_samples <= 1)
@@ -3338,7 +3960,7 @@ bool si_update_shaders(struct si_context *sctx)
 			return false;
 	}
 
-	if (sctx->chip_class >= CIK) {
+	if (sctx->chip_class >= GFX7) {
 		if (si_pm4_state_enabled_and_changed(sctx, ls))
 			sctx->prefetch_L2_mask |= SI_PREFETCH_LS;
 		else if (!sctx->queued.named.ls)
@@ -3388,70 +4010,6 @@ static void si_emit_scratch_state(struct si_context *sctx)
 	}
 }
 
-void *si_get_blit_vs(struct si_context *sctx, enum blitter_attrib_type type,
-		     unsigned num_layers)
-{
-	unsigned vs_blit_property;
-	void **vs;
-
-	switch (type) {
-	case UTIL_BLITTER_ATTRIB_NONE:
-		vs = num_layers > 1 ? &sctx->vs_blit_pos_layered :
-				      &sctx->vs_blit_pos;
-		vs_blit_property = SI_VS_BLIT_SGPRS_POS;
-		break;
-	case UTIL_BLITTER_ATTRIB_COLOR:
-		vs = num_layers > 1 ? &sctx->vs_blit_color_layered :
-				      &sctx->vs_blit_color;
-		vs_blit_property = SI_VS_BLIT_SGPRS_POS_COLOR;
-		break;
-	case UTIL_BLITTER_ATTRIB_TEXCOORD_XY:
-	case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW:
-		assert(num_layers == 1);
-		vs = &sctx->vs_blit_texcoord;
-		vs_blit_property = SI_VS_BLIT_SGPRS_POS_TEXCOORD;
-		break;
-	default:
-		assert(0);
-		return NULL;
-	}
-	if (*vs)
-		return *vs;
-
-	struct ureg_program *ureg = ureg_create(PIPE_SHADER_VERTEX);
-	if (!ureg)
-		return NULL;
-
-	/* Tell the shader to load VS inputs from SGPRs: */
-	ureg_property(ureg, TGSI_PROPERTY_VS_BLIT_SGPRS, vs_blit_property);
-	ureg_property(ureg, TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION, true);
-
-	/* This is just a pass-through shader with 1-3 MOV instructions. */
-	ureg_MOV(ureg,
-		 ureg_DECL_output(ureg, TGSI_SEMANTIC_POSITION, 0),
-		 ureg_DECL_vs_input(ureg, 0));
-
-	if (type != UTIL_BLITTER_ATTRIB_NONE) {
-		ureg_MOV(ureg,
-			 ureg_DECL_output(ureg, TGSI_SEMANTIC_GENERIC, 0),
-			 ureg_DECL_vs_input(ureg, 1));
-	}
-
-	if (num_layers > 1) {
-		struct ureg_src instance_id =
-			ureg_DECL_system_value(ureg, TGSI_SEMANTIC_INSTANCEID, 0);
-		struct ureg_dst layer =
-			ureg_DECL_output(ureg, TGSI_SEMANTIC_LAYER, 0);
-
-		ureg_MOV(ureg, ureg_writemask(layer, TGSI_WRITEMASK_X),
-			 ureg_scalar(instance_id, TGSI_SWIZZLE_X));
-	}
-	ureg_END(ureg);
-
-	*vs = ureg_create_shader_and_destroy(ureg, &sctx->b);
-	return *vs;
-}
-
 void si_init_shader_functions(struct si_context *sctx)
 {
 	sctx->atoms.s.spi_map.emit = si_emit_spi_map;