radeonsi/gfx10: implement streamout

Acked-by: Bas Nieuwenhuizen <[email protected]>

4 files changed, 618 insertions, 33 deletions

diff --git a/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c b/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
index c97d9009164..2e7b42e056d 100644
--- a/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
+++ b/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
@@ -64,6 +64,14 @@ static LLVMValueRef ngg_get_prim_cnt(struct si_shader_context *ctx)
 			    false);
 }
 
+static LLVMValueRef ngg_get_ordered_id(struct si_shader_context *ctx)
+{
+	return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
+			    ctx->i32_0,
+			    LLVMConstInt(ctx->ac.i32, 11, false),
+			    false);
+}
+
 static LLVMValueRef ngg_get_query_buf(struct si_shader_context *ctx)
 {
 	LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
@@ -136,6 +144,371 @@ static void build_export_prim(struct si_shader_context *ctx,
 	ac_build_export(&ctx->ac, &args);
 }
 
+static void build_streamout_vertex(struct si_shader_context *ctx,
+				   LLVMValueRef *so_buffer, LLVMValueRef *wg_offset_dw,
+				   unsigned stream, LLVMValueRef offset_vtx,
+				   LLVMValueRef vertexptr)
+{
+	struct tgsi_shader_info *info = &ctx->shader->selector->info;
+	struct pipe_stream_output_info *so = &ctx->shader->selector->so;
+	LLVMBuilderRef builder = ctx->ac.builder;
+	LLVMValueRef offset[4] = {};
+	LLVMValueRef tmp;
+
+	for (unsigned buffer = 0; buffer < 4; ++buffer) {
+		if (!wg_offset_dw[buffer])
+			continue;
+
+		tmp = LLVMBuildMul(builder, offset_vtx,
+				   LLVMConstInt(ctx->i32, so->stride[buffer], false), "");
+		tmp = LLVMBuildAdd(builder, wg_offset_dw[buffer], tmp, "");
+		offset[buffer] = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->i32, 2, false), "");
+	}
+
+	for (unsigned i = 0; i < so->num_outputs; ++i) {
+		if (so->output[i].stream != stream)
+			continue;
+
+		unsigned reg = so->output[i].register_index;
+		struct si_shader_output_values out;
+		out.semantic_name = info->output_semantic_name[reg];
+		out.semantic_index = info->output_semantic_index[reg];
+
+		for (unsigned comp = 0; comp < 4; comp++) {
+			tmp = ac_build_gep0(&ctx->ac, vertexptr,
+					    LLVMConstInt(ctx->i32, 4 * reg + comp, false));
+			out.values[comp] = LLVMBuildLoad(builder, tmp, "");
+			out.vertex_stream[comp] =
+				(info->output_streams[reg] >> (2 * comp)) & 3;
+		}
+
+		si_emit_streamout_output(ctx, so_buffer, offset, &so->output[i], &out);
+	}
+}
+
+struct ngg_streamout {
+	LLVMValueRef num_vertices;
+
+	/* per-thread data */
+	LLVMValueRef prim_enable[4]; /* i1 per stream */
+	LLVMValueRef vertices[3]; /* [N x i32] addrspace(LDS)* */
+
+	/* Output */
+	LLVMValueRef emit[4]; /* per-stream emitted primitives (only valid for used streams) */
+};
+
+/**
+ * Build streamout logic.
+ *
+ * Implies a barrier.
+ *
+ * Writes number of emitted primitives to gs_ngg_scratch[4:8].
+ *
+ * Clobbers gs_ngg_scratch[8:].
+ */
+static void build_streamout(struct si_shader_context *ctx,
+			    struct ngg_streamout *nggso)
+{
+	struct tgsi_shader_info *info = &ctx->shader->selector->info;
+	struct pipe_stream_output_info *so = &ctx->shader->selector->so;
+	LLVMBuilderRef builder = ctx->ac.builder;
+	LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
+	LLVMValueRef tid = get_thread_id_in_tg(ctx);
+	LLVMValueRef tmp, tmp2;
+	LLVMValueRef i32_2 = LLVMConstInt(ctx->i32, 2, false);
+	LLVMValueRef i32_4 = LLVMConstInt(ctx->i32, 4, false);
+	LLVMValueRef i32_8 = LLVMConstInt(ctx->i32, 8, false);
+	LLVMValueRef so_buffer[4] = {};
+	unsigned max_num_vertices = 1 + (nggso->vertices[1] ? 1 : 0) +
+					(nggso->vertices[2] ? 1 : 0);
+	LLVMValueRef prim_stride_dw[4] = {};
+	LLVMValueRef prim_stride_dw_vgpr = LLVMGetUndef(ctx->i32);
+	int stream_for_buffer[4] = { -1, -1, -1, -1 };
+	unsigned bufmask_for_stream[4] = {};
+	bool isgs = ctx->type == PIPE_SHADER_GEOMETRY;
+	unsigned scratch_emit_base = isgs ? 4 : 0;
+	LLVMValueRef scratch_emit_basev = isgs ? i32_4 : ctx->i32_0;
+	unsigned scratch_offset_base = isgs ? 8 : 4;
+	LLVMValueRef scratch_offset_basev = isgs ? i32_8 : i32_4;
+
+	/* Determine the mapping of streamout buffers to vertex streams. */
+	for (unsigned i = 0; i < so->num_outputs; ++i) {
+		unsigned buf = so->output[i].output_buffer;
+		unsigned stream = so->output[i].stream;
+		assert(stream_for_buffer[buf] < 0 || stream_for_buffer[buf] == stream);
+		stream_for_buffer[buf] = stream;
+		bufmask_for_stream[stream] |= 1 << buf;
+	}
+
+	for (unsigned buffer = 0; buffer < 4; ++buffer) {
+		if (stream_for_buffer[buffer] == -1)
+			continue;
+
+		assert(so->stride[buffer]);
+
+		tmp = LLVMConstInt(ctx->i32, so->stride[buffer], false);
+		prim_stride_dw[buffer] = LLVMBuildMul(builder, tmp, nggso->num_vertices, "");
+		prim_stride_dw_vgpr = ac_build_writelane(
+			&ctx->ac, prim_stride_dw_vgpr, prim_stride_dw[buffer],
+			LLVMConstInt(ctx->i32, buffer, false));
+
+		so_buffer[buffer] = ac_build_load_to_sgpr(
+			&ctx->ac, buf_ptr,
+			LLVMConstInt(ctx->i32, SI_VS_STREAMOUT_BUF0 + buffer, false));
+	}
+
+	tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->i32_0, "");
+	ac_build_ifcc(&ctx->ac, tmp, 5200);
+	{
+		LLVMTypeRef gdsptr = LLVMPointerType(ctx->i32, AC_ADDR_SPACE_GDS);
+		LLVMValueRef gdsbase = LLVMBuildIntToPtr(builder, ctx->i32_0, gdsptr, "");
+
+		/* Advance the streamout offsets in GDS. */
+		LLVMValueRef offsets_vgpr = ac_build_alloca_undef(&ctx->ac, ctx->i32, "");
+		LLVMValueRef generated_by_stream_vgpr = ac_build_alloca_undef(&ctx->ac, ctx->i32, "");
+
+		tmp = LLVMBuildICmp(builder, LLVMIntULT, ac_get_thread_id(&ctx->ac), i32_4, "");
+		ac_build_ifcc(&ctx->ac, tmp, 5210);
+		{
+			if (isgs) {
+				tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tid);
+				tmp = LLVMBuildLoad(builder, tmp, "");
+			} else {
+				tmp = ac_build_writelane(&ctx->ac, ctx->i32_0,
+						ngg_get_prim_cnt(ctx), ctx->i32_0);
+			}
+			LLVMBuildStore(builder, tmp, generated_by_stream_vgpr);
+
+			unsigned swizzle[4];
+			int unused_stream = -1;
+			for (unsigned stream = 0; stream < 4; ++stream) {
+				if (!info->num_stream_output_components[stream]) {
+					unused_stream = stream;
+					break;
+				}
+			}
+			for (unsigned buffer = 0; buffer < 4; ++buffer) {
+				if (stream_for_buffer[buffer] >= 0) {
+					swizzle[buffer] = stream_for_buffer[buffer];
+				} else {
+					assert(unused_stream >= 0);
+					swizzle[buffer] = unused_stream;
+				}
+			}
+
+			tmp = ac_build_quad_swizzle(&ctx->ac, tmp,
+				swizzle[0], swizzle[1], swizzle[2], swizzle[3]);
+			tmp = LLVMBuildMul(builder, tmp, prim_stride_dw_vgpr, "");
+
+			LLVMValueRef args[] = {
+				LLVMBuildIntToPtr(builder, ngg_get_ordered_id(ctx), gdsptr, ""),
+				tmp,
+				ctx->i32_0, // ordering
+				ctx->i32_0, // scope
+				ctx->ac.i1false, // isVolatile
+				LLVMConstInt(ctx->i32, 4 << 24, false), // OA index
+				ctx->ac.i1true, // wave release
+				ctx->ac.i1true, // wave done
+			};
+			tmp = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ds.ordered.add",
+						 ctx->i32, args, ARRAY_SIZE(args), 0);
+
+			/* Keep offsets in a VGPR for quick retrieval via readlane by
+			 * the first wave for bounds checking, and also store in LDS
+			 * for retrieval by all waves later. */
+			LLVMBuildStore(builder, tmp, offsets_vgpr);
+
+			tmp2 = LLVMBuildAdd(builder, ac_get_thread_id(&ctx->ac),
+					    scratch_offset_basev, "");
+			tmp2 = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tmp2);
+			LLVMBuildStore(builder, tmp, tmp2);
+		}
+		ac_build_endif(&ctx->ac, 5210);
+
+		/* Determine the max emit per buffer. This is done via the SALU, in part
+		 * because LLVM can't generate divide-by-multiply if we try to do this
+		 * via VALU with one lane per buffer.
+		 */
+		LLVMValueRef max_emit[4] = {};
+		for (unsigned buffer = 0; buffer < 4; ++buffer) {
+			if (stream_for_buffer[buffer] == -1)
+				continue;
+
+			LLVMValueRef bufsize_dw =
+				LLVMBuildLShr(builder,
+					LLVMBuildExtractElement(builder, so_buffer[buffer], i32_2, ""),
+					i32_2, "");
+
+			tmp = LLVMBuildLoad(builder, offsets_vgpr, "");
+			LLVMValueRef offset_dw =
+				ac_build_readlane(&ctx->ac, tmp,
+						LLVMConstInt(ctx->i32, buffer, false));
+
+			tmp = LLVMBuildSub(builder, bufsize_dw, offset_dw, "");
+			tmp = LLVMBuildUDiv(builder, tmp, prim_stride_dw[buffer], "");
+
+			tmp2 = LLVMBuildICmp(builder, LLVMIntULT, bufsize_dw, offset_dw, "");
+			max_emit[buffer] = LLVMBuildSelect(builder, tmp2, ctx->i32_0, tmp, "");
+		}
+
+		/* Determine the number of emitted primitives per stream and fixup the
+		 * GDS counter if necessary.
+		 *
+		 * This is complicated by the fact that a single stream can emit to
+		 * multiple buffers (but luckily not vice versa).
+		 */
+		LLVMValueRef emit_vgpr = ctx->i32_0;
+
+		for (unsigned stream = 0; stream < 4; ++stream) {
+			if (!info->num_stream_output_components[stream])
+				continue;
+
+			tmp = LLVMBuildLoad(builder, generated_by_stream_vgpr, "");
+			LLVMValueRef generated =
+				ac_build_readlane(&ctx->ac, tmp,
+						  LLVMConstInt(ctx->i32, stream, false));
+
+			LLVMValueRef emit = generated;
+			for (unsigned buffer = 0; buffer < 4; ++buffer) {
+				if (stream_for_buffer[buffer] == stream)
+					emit = ac_build_umin(&ctx->ac, emit, max_emit[buffer]);
+			}
+
+			emit_vgpr = ac_build_writelane(&ctx->ac, emit_vgpr, emit,
+						       LLVMConstInt(ctx->i32, stream, false));
+
+			/* Fixup the offset using a plain GDS atomic if we overflowed. */
+			tmp = LLVMBuildICmp(builder, LLVMIntULT, emit, generated, "");
+			ac_build_ifcc(&ctx->ac, tmp, 5221); /* scalar branch */
+			tmp = LLVMBuildLShr(builder,
+					    LLVMConstInt(ctx->i32, bufmask_for_stream[stream], false),
+					    ac_get_thread_id(&ctx->ac), "");
+			tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
+			ac_build_ifcc(&ctx->ac, tmp, 5222);
+			{
+				tmp = LLVMBuildSub(builder, generated, emit, "");
+				tmp = LLVMBuildMul(builder, tmp, prim_stride_dw_vgpr, "");
+				tmp2 = LLVMBuildGEP(builder, gdsbase, &tid, 1, "");
+				LLVMBuildAtomicRMW(builder, LLVMAtomicRMWBinOpSub, tmp2, tmp,
+						   LLVMAtomicOrderingMonotonic, false);
+			}
+			ac_build_endif(&ctx->ac, 5222);
+			ac_build_endif(&ctx->ac, 5221);
+		}
+
+		tmp = LLVMBuildICmp(builder, LLVMIntULT, ac_get_thread_id(&ctx->ac), i32_4, "");
+		ac_build_ifcc(&ctx->ac, tmp, 5225);
+		{
+			tmp = LLVMBuildAdd(builder, ac_get_thread_id(&ctx->ac),
+					   scratch_emit_basev, "");
+			tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tmp);
+			LLVMBuildStore(builder, emit_vgpr, tmp);
+		}
+		ac_build_endif(&ctx->ac, 5225);
+	}
+	ac_build_endif(&ctx->ac, 5200);
+
+	/* Determine the workgroup-relative per-thread / primitive offset into
+	 * the streamout buffers */
+	struct ac_wg_scan primemit_scan[4] = {};
+
+	if (isgs) {
+		for (unsigned stream = 0; stream < 4; ++stream) {
+			if (!info->num_stream_output_components[stream])
+				continue;
+
+			primemit_scan[stream].enable_exclusive = true;
+			primemit_scan[stream].op = nir_op_iadd;
+			primemit_scan[stream].src = nggso->prim_enable[stream];
+			primemit_scan[stream].scratch =
+				ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch,
+					LLVMConstInt(ctx->i32, 12 + 8 * stream, false));
+			primemit_scan[stream].waveidx = get_wave_id_in_tg(ctx);
+			primemit_scan[stream].numwaves = get_tgsize(ctx);
+			primemit_scan[stream].maxwaves = 8;
+			ac_build_wg_scan_top(&ctx->ac, &primemit_scan[stream]);
+		}
+	}
+
+	ac_build_s_barrier(&ctx->ac);
+
+	/* Fetch the per-buffer offsets and per-stream emit counts in all waves. */
+	LLVMValueRef wgoffset_dw[4] = {};
+
+	{
+		LLVMValueRef scratch_vgpr;
+
+		tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, ac_get_thread_id(&ctx->ac));
+		scratch_vgpr = LLVMBuildLoad(builder, tmp, "");
+
+		for (unsigned buffer = 0; buffer < 4; ++buffer) {
+			if (stream_for_buffer[buffer] >= 0) {
+				wgoffset_dw[buffer] = ac_build_readlane(
+					&ctx->ac, scratch_vgpr,
+					LLVMConstInt(ctx->i32, scratch_offset_base + buffer, false));
+			}
+		}
+
+		for (unsigned stream = 0; stream < 4; ++stream) {
+			if (info->num_stream_output_components[stream]) {
+				nggso->emit[stream] = ac_build_readlane(
+					&ctx->ac, scratch_vgpr,
+					LLVMConstInt(ctx->i32, scratch_emit_base + stream, false));
+			}
+		}
+	}
+
+	/* Write out primitive data */
+	for (unsigned stream = 0; stream < 4; ++stream) {
+		if (!info->num_stream_output_components[stream])
+			continue;
+
+		if (isgs) {
+			ac_build_wg_scan_bottom(&ctx->ac, &primemit_scan[stream]);
+		} else {
+			primemit_scan[stream].result_exclusive = tid;
+		}
+
+		tmp = LLVMBuildICmp(builder, LLVMIntULT,
+				    primemit_scan[stream].result_exclusive,
+				    nggso->emit[stream], "");
+		tmp = LLVMBuildAnd(builder, tmp, nggso->prim_enable[stream], "");
+		ac_build_ifcc(&ctx->ac, tmp, 5240);
+		{
+			LLVMValueRef offset_vtx =
+				LLVMBuildMul(builder, primemit_scan[stream].result_exclusive,
+					     nggso->num_vertices, "");
+
+			for (unsigned i = 0; i < max_num_vertices; ++i) {
+				tmp = LLVMBuildICmp(builder, LLVMIntULT,
+						    LLVMConstInt(ctx->i32, i, false),
+						    nggso->num_vertices, "");
+				ac_build_ifcc(&ctx->ac, tmp, 5241);
+				build_streamout_vertex(ctx, so_buffer, wgoffset_dw,
+						       stream, offset_vtx, nggso->vertices[i]);
+				ac_build_endif(&ctx->ac, 5241);
+				offset_vtx = LLVMBuildAdd(builder, offset_vtx, ctx->i32_1, "");
+			}
+		}
+		ac_build_endif(&ctx->ac, 5240);
+	}
+}
+
+/**
+ * Returns an `[N x i32] addrspace(LDS)*` pointing at contiguous LDS storage
+ * for the vertex outputs.
+ */
+static LLVMValueRef ngg_nogs_vertex_ptr(struct si_shader_context *ctx,
+					LLVMValueRef vtxid)
+{
+	/* The extra dword is used to avoid LDS bank conflicts. */
+	unsigned vertex_size = 4 * ctx->shader->selector->info.num_outputs + 1;
+	LLVMTypeRef ai32 = LLVMArrayType(ctx->i32, vertex_size);
+	LLVMTypeRef pai32 = LLVMPointerType(ai32, AC_ADDR_SPACE_LDS);
+	LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, ctx->esgs_ring, pai32, "");
+	return LLVMBuildGEP(ctx->ac.builder, tmp, &vtxid, 1, "");
+}
+
 /**
  * Emit the epilogue of an API VS or TES shader compiled as ESGS shader.
  */
@@ -144,17 +517,23 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
 			     LLVMValueRef *addrs)
 {
 	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
-	struct tgsi_shader_info *info = &ctx->shader->selector->info;
+	struct si_shader_selector *sel = ctx->shader->selector;
+	struct tgsi_shader_info *info = &sel->info;
 	struct si_shader_output_values *outputs = NULL;
 	LLVMBuilderRef builder = ctx->ac.builder;
 	struct lp_build_if_state if_state;
-	LLVMValueRef tmp;
+	LLVMValueRef tmp, tmp2;
 
 	assert(!ctx->shader->is_gs_copy_shader);
 	assert(info->num_outputs <= max_outputs);
 
 	outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));
 
+	LLVMValueRef vertex_ptr = NULL;
+
+	if (sel->so.num_outputs)
+		vertex_ptr = ngg_nogs_vertex_ptr(ctx, get_thread_id_in_tg(ctx));
+
 	for (unsigned i = 0; i < info->num_outputs; i++) {
 		outputs[i].semantic_name = info->output_semantic_name[i];
 		outputs[i].semantic_index = info->output_semantic_index[i];
@@ -167,6 +546,13 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
 					      "");
 			outputs[i].vertex_stream[j] =
 				(info->output_streams[i] >> (2 * j)) & 3;
+
+			if (vertex_ptr) {
+				tmp = ac_build_gep0(&ctx->ac, vertex_ptr,
+					LLVMConstInt(ctx->i32, 4 * i + j, false));
+				tmp2 = ac_to_integer(&ctx->ac, outputs[i].values[j]);
+				LLVMBuildStore(builder, tmp2, tmp);
+			}
 		}
 	}
 
@@ -212,7 +598,21 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
 		num_vertices_val = LLVMConstInt(ctx->i32, num_vertices, false);
 	}
 
-	/* TODO: streamout */
+	/* Streamout */
+	LLVMValueRef emitted_prims = NULL;
+
+	if (sel->so.num_outputs) {
+		struct ngg_streamout nggso = {};
+
+		nggso.num_vertices = num_vertices_val;
+		nggso.prim_enable[0] = is_gs_thread;
+
+		for (unsigned i = 0; i < num_vertices; ++i)
+			nggso.vertices[i] = ngg_nogs_vertex_ptr(ctx, vtxindex[i]);
+
+		build_streamout(ctx, &nggso);
+		emitted_prims = nggso.emit[0];
+	}
 
 	/* TODO: primitive culling */
 
@@ -221,7 +621,8 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
 	/* Update query buffer */
 	tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
 	ac_build_ifcc(&ctx->ac, tmp, 5030);
-	tmp = LLVMBuildICmp(builder, LLVMIntULE, ac_get_thread_id(&ctx->ac), ctx->ac.i32_0, "");
+	tmp = LLVMBuildICmp(builder, LLVMIntULE, ac_get_thread_id(&ctx->ac),
+			    sel->so.num_outputs ? ctx->ac.i32_1 : ctx->ac.i32_0, "");
 	ac_build_ifcc(&ctx->ac, tmp, 5031);
 	{
 		LLVMValueRef args[] = {
@@ -232,6 +633,12 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
 			ctx->i32_0, /* cachepolicy */
 		};
 
+		if (sel->so.num_outputs) {
+			args[0] = ac_build_writelane(&ctx->ac, args[0], emitted_prims, ctx->i32_1);
+			args[2] = ac_build_writelane(&ctx->ac, args[2],
+					LLVMConstInt(ctx->i32, 24, false), ctx->i32_1);
+		}
+
 		/* TODO: should this be 64-bit atomics? */
 		ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32",
 				   ctx->i32, args, 5, 0);
@@ -563,9 +970,43 @@ void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx)
 	const LLVMValueRef tid = get_thread_id_in_tg(ctx);
 	LLVMValueRef num_emit_threads = ngg_get_prim_cnt(ctx);
 
-	/* TODO: streamout */
+	/* Streamout */
+	if (sel->so.num_outputs) {
+		struct ngg_streamout nggso = {};
 
-	tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, LLVMConstInt(ctx->i32, 4, false), "");
+		nggso.num_vertices = LLVMConstInt(ctx->i32, verts_per_prim, false);
+
+		LLVMValueRef vertexptr = ngg_gs_vertex_ptr(ctx, tid);
+		for (unsigned stream = 0; stream < 4; ++stream) {
+			if (!info->num_stream_output_components[stream])
+				continue;
+
+			LLVMValueRef gep_idx[3] = {
+				ctx->i32_0, /* implicit C-style array */
+				ctx->i32_1, /* second value of struct */
+				LLVMConstInt(ctx->i32, stream, false),
+			};
+			tmp = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
+			tmp = LLVMBuildLoad(builder, tmp, "");
+			tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
+			tmp2 = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
+			nggso.prim_enable[stream] = LLVMBuildAnd(builder, tmp, tmp2, "");
+		}
+
+		for (unsigned i = 0; i < verts_per_prim; ++i) {
+			tmp = LLVMBuildSub(builder, tid,
+					   LLVMConstInt(ctx->i32, verts_per_prim - i - 1, false), "");
+			tmp = ngg_gs_vertex_ptr(ctx, tmp);
+			nggso.vertices[i] = ac_build_gep0(&ctx->ac, tmp, ctx->i32_0);
+		}
+
+		build_streamout(ctx, &nggso);
+	}
+
+	/* Write shader query data. */
+	unsigned num_query_comps = sel->so.num_outputs ? 8 : 4;
+	tmp = LLVMBuildICmp(builder, LLVMIntULT, tid,
+			    LLVMConstInt(ctx->i32, num_query_comps, false), "");
 	ac_build_ifcc(&ctx->ac, tmp, 5110);
 	{
 		LLVMValueRef offset;
@@ -843,6 +1284,8 @@ void gfx10_ngg_calculate_subgroup_info(struct si_shader *shader)
 	} else {
 		/* TODO: This needs to be adjusted once LDS use for compaction
 		 * after culling is implemented. */
+		if (es_sel->so.num_outputs)
+			esvert_lds_size = 4 * es_sel->info.num_outputs + 1;
 	}
 
 	unsigned max_gsprims = max_gsprims_base;
diff --git a/src/gallium/drivers/radeonsi/si_gfx_cs.c b/src/gallium/drivers/radeonsi/si_gfx_cs.c
index bb34b07095a..4a4a7eecd63 100644
--- a/src/gallium/drivers/radeonsi/si_gfx_cs.c
+++ b/src/gallium/drivers/radeonsi/si_gfx_cs.c
@@ -376,7 +376,8 @@ void si_begin_new_gfx_cs(struct si_context *ctx)
 		si_mark_atom_dirty(ctx, &ctx->atoms.s.dpbb_state);
 	si_mark_atom_dirty(ctx, &ctx->atoms.s.stencil_ref);
 	si_mark_atom_dirty(ctx, &ctx->atoms.s.spi_map);
-	si_mark_atom_dirty(ctx, &ctx->atoms.s.streamout_enable);
+	if (ctx->chip_class < GFX10)
+		si_mark_atom_dirty(ctx, &ctx->atoms.s.streamout_enable);
 	si_mark_atom_dirty(ctx, &ctx->atoms.s.render_cond);
 	/* CLEAR_STATE disables all window rectangles. */
 	if (!has_clear_state || ctx->num_window_rectangles > 0)
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 68506b7a92c..6f11225c9af 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -5005,6 +5005,26 @@ static void create_function(struct si_shader_context *ctx)
 	}
 }
 
+/* Ensure that the esgs ring is declared.
+ *
+ * We declare it with 64KB alignment as a hint that the
+ * pointer value will always be 0.
+ */
+static void declare_esgs_ring(struct si_shader_context *ctx)
+{
+	if (ctx->esgs_ring)
+		return;
+
+	assert(!LLVMGetNamedGlobal(ctx->ac.module, "esgs_ring"));
+
+	ctx->esgs_ring = LLVMAddGlobalInAddressSpace(
+		ctx->ac.module, LLVMArrayType(ctx->i32, 0),
+		"esgs_ring",
+		AC_ADDR_SPACE_LDS);
+	LLVMSetLinkage(ctx->esgs_ring, LLVMExternalLinkage);
+	LLVMSetAlignment(ctx->esgs_ring, 64 * 1024);
+}
+
 /**
  * Load ESGS and GSVS ring buffer resource descriptors and save the variables
  * for later use.
@@ -5027,17 +5047,8 @@ static void preload_ring_buffers(struct si_shader_context *ctx)
 				ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
 		} else {
 			if (USE_LDS_SYMBOLS && HAVE_LLVM >= 0x0900) {
-				/* Declare the ESGS ring as an explicit LDS symbol.
-				 * For monolithic shaders, we declare the ring only once.
-				 *
-				 * We declare it with 64KB alignment as a hint that the
-				 * pointer value will always be 0.
-				 */
-				ctx->esgs_ring = LLVMAddGlobalInAddressSpace(
-					ctx->ac.module, LLVMArrayType(ctx->i32, 0),
-					"esgs_ring",
-					AC_ADDR_SPACE_LDS);
-				LLVMSetAlignment(ctx->esgs_ring, 64 * 1024);
+				/* Declare the ESGS ring as an explicit LDS symbol. */
+				declare_esgs_ring(ctx);
 			} else {
 				ac_declare_lds_as_pointer(&ctx->ac);
 				ctx->esgs_ring = ctx->ac.lds;
@@ -5192,15 +5203,26 @@ static bool si_shader_binary_open(struct si_screen *screen,
 
 	struct ac_rtld_symbol lds_symbols[2];
 	unsigned num_lds_symbols = 0;
+	unsigned esgs_ring_size = 0;
 
 	if (sel && screen->info.chip_class >= GFX9 &&
 	    sel->type == PIPE_SHADER_GEOMETRY && !shader->is_gs_copy_shader) {
+		esgs_ring_size = shader->gs_info.esgs_ring_size;;
+	}
+
+	if (sel && shader->key.as_ngg && sel->so.num_outputs) {
+		unsigned esgs_vertex_bytes = 4 * (4 * sel->info.num_outputs + 1);
+		esgs_ring_size = MAX2(esgs_ring_size,
+				      shader->ngg.max_out_verts * esgs_vertex_bytes);
+	}
+
+	if (esgs_ring_size) {
 		/* We add this symbol even on LLVM <= 8 to ensure that
 		 * shader->config.lds_size is set correctly below.
 		 */
 		struct ac_rtld_symbol *sym = &lds_symbols[num_lds_symbols++];
 		sym->name = "esgs_ring";
-		sym->size = shader->gs_info.esgs_ring_size;
+		sym->size = esgs_ring_size;
 		sym->align = 64 * 1024;
 	}
 
@@ -6086,10 +6108,14 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx)
 					lp_build_alloca(&ctx->gallivm, ctx->ac.i32, "");
 			}
 
-			LLVMTypeRef a8i32 = LLVMArrayType(ctx->i32, 8);
+			unsigned scratch_size = 8;
+			if (sel->so.num_outputs)
+				scratch_size = 44;
+
+			LLVMTypeRef ai32 = LLVMArrayType(ctx->i32, scratch_size);
 			ctx->gs_ngg_scratch = LLVMAddGlobalInAddressSpace(ctx->ac.module,
-				a8i32, "ngg_scratch", AC_ADDR_SPACE_LDS);
-			LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(a8i32));
+				ai32, "ngg_scratch", AC_ADDR_SPACE_LDS);
+			LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(ai32));
 			LLVMSetAlignment(ctx->gs_ngg_scratch, 4);
 
 			ctx->gs_ngg_emit = LLVMAddGlobalInAddressSpace(ctx->ac.module,
@@ -6099,6 +6125,26 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx)
 		}
 	}
 
+	if (shader->key.as_ngg && ctx->type != PIPE_SHADER_GEOMETRY) {
+		/* Unconditionally declare scratch space base for streamout and
+		 * vertex compaction. Whether space is actually allocated is
+		 * determined during linking / PM4 creation.
+		 *
+		 * Add an extra dword per vertex to ensure an odd stride, which
+		 * avoids bank conflicts for SoA accesses.
+		 */
+		declare_esgs_ring(ctx);
+
+		/* This is really only needed when streamout and / or vertex
+		 * compaction is enabled.
+		 */
+		LLVMTypeRef asi32 = LLVMArrayType(ctx->i32, 8);
+		ctx->gs_ngg_scratch = LLVMAddGlobalInAddressSpace(ctx->ac.module,
+			asi32, "ngg_scratch", AC_ADDR_SPACE_LDS);
+		LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(asi32));
+		LLVMSetAlignment(ctx->gs_ngg_scratch, 4);
+	}
+
 	/* For GFX9 merged shaders:
 	 * - Set EXEC for the first shader. If the prolog is present, set
 	 *   EXEC there instead.
diff --git a/src/gallium/drivers/radeonsi/si_state_streamout.c b/src/gallium/drivers/radeonsi/si_state_streamout.c
index e3c72ccdf49..6b727498ea1 100644
--- a/src/gallium/drivers/radeonsi/si_state_streamout.c
+++ b/src/gallium/drivers/radeonsi/si_state_streamout.c
@@ -50,7 +50,8 @@ si_create_so_target(struct pipe_context *ctx,
 		return NULL;
 	}
 
-	u_suballocator_alloc(sctx->allocator_zeroed_memory, 4, 4,
+	unsigned buf_filled_size_size = sctx->chip_class >= GFX10 ? 8 : 4;
+	u_suballocator_alloc(sctx->allocator_zeroed_memory, buf_filled_size_size, 4,
 			     &t->buf_filled_size_offset,
 			     (struct pipe_resource**)&t->buf_filled_size);
 	if (!t->buf_filled_size) {
@@ -122,8 +123,13 @@ static void si_set_streamout_targets(struct pipe_context *ctx,
 		 * used as an input immediately.
 		 */
 		sctx->flags |= SI_CONTEXT_INV_SCACHE |
-			       SI_CONTEXT_INV_VCACHE |
-			       SI_CONTEXT_VS_PARTIAL_FLUSH;
+			       SI_CONTEXT_INV_VCACHE;
+
+		/* The BUFFER_FILLED_SIZE is written using a PS_DONE event. */
+		if (sctx->chip_class >= GFX10)
+			sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH;
+		else
+			sctx->flags |= SI_CONTEXT_VS_PARTIAL_FLUSH;
 	}
 
 	/* All readers of the streamout targets need to be finished before we can
@@ -176,9 +182,16 @@ static void si_set_streamout_targets(struct pipe_context *ctx,
 		if (targets[i]) {
 			struct pipe_shader_buffer sbuf;
 			sbuf.buffer = targets[i]->buffer;
-			sbuf.buffer_offset = 0;
-			sbuf.buffer_size = targets[i]->buffer_offset +
-					   targets[i]->buffer_size;
+
+			if (sctx->chip_class >= GFX10) {
+				sbuf.buffer_offset = targets[i]->buffer_offset;
+				sbuf.buffer_size = targets[i]->buffer_size;
+			} else {
+				sbuf.buffer_offset = 0;
+				sbuf.buffer_size = targets[i]->buffer_offset +
+						   targets[i]->buffer_size;
+			}
+
 			si_set_rw_shader_buffer(sctx, SI_VS_STREAMOUT_BUF0 + i, &sbuf);
 			si_resource(targets[i]->buffer)->bind_history |= PIPE_BIND_STREAM_OUTPUT;
 		} else {
@@ -189,6 +202,74 @@ static void si_set_streamout_targets(struct pipe_context *ctx,
 		si_set_rw_shader_buffer(sctx, SI_VS_STREAMOUT_BUF0 + i, NULL);
 }
 
+static void gfx10_emit_streamout_begin(struct si_context *sctx)
+{
+	struct si_streamout_target **t = sctx->streamout.targets;
+	struct radeon_cmdbuf *cs = sctx->gfx_cs;
+	unsigned last_target = 0;
+
+	for (unsigned i = 0; i < sctx->streamout.num_targets; i++) {
+		if (t[i])
+			last_target = i;
+	}
+
+	for (unsigned i = 0; i < sctx->streamout.num_targets; i++) {
+		if (!t[i])
+			continue;
+
+		t[i]->stride_in_dw = sctx->streamout.stride_in_dw[i];
+
+		bool append = sctx->streamout.append_bitmask & (1 << i);
+		uint64_t va = 0;
+
+		if (append) {
+			radeon_add_to_buffer_list(sctx,  sctx->gfx_cs,
+						  t[i]->buf_filled_size,
+						  RADEON_USAGE_READ,
+						  RADEON_PRIO_SO_FILLED_SIZE);
+
+			va = t[i]->buf_filled_size->gpu_address +
+			     t[i]->buf_filled_size_offset;
+		}
+
+		radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, 0));
+		radeon_emit(cs, S_411_SRC_SEL(append ? V_411_SRC_ADDR_TC_L2 : V_411_DATA) |
+				S_411_DST_SEL(V_411_GDS) |
+				S_411_CP_SYNC(i == last_target));
+		radeon_emit(cs, va);
+		radeon_emit(cs, va >> 32);
+		radeon_emit(cs, 4 * i); /* destination in GDS */
+		radeon_emit(cs, 0);
+		radeon_emit(cs, S_414_BYTE_COUNT_GFX9(4) |
+				S_414_DISABLE_WR_CONFIRM_GFX9(i != last_target));
+	}
+
+	sctx->streamout.begin_emitted = true;
+}
+
+static void gfx10_emit_streamout_end(struct si_context *sctx)
+{
+	struct si_streamout_target **t = sctx->streamout.targets;
+
+	for (unsigned i = 0; i < sctx->streamout.num_targets; i++) {
+		if (!t[i])
+			continue;
+
+		uint64_t va = t[i]->buf_filled_size->gpu_address + t[i]->buf_filled_size_offset;
+
+		si_cp_release_mem(sctx, sctx->gfx_cs, V_028A90_PS_DONE, 0,
+				  EOP_DST_SEL_TC_L2,
+				  EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM,
+				  EOP_DATA_SEL_GDS,
+				  t[i]->buf_filled_size, va,
+				  EOP_DATA_GDS(i, 1), 0);
+
+		t[i]->buf_filled_size_valid = true;
+	}
+
+	sctx->streamout.begin_emitted = false;
+}
+
 static void si_flush_vgt_streamout(struct si_context *sctx)
 {
 	struct radeon_cmdbuf *cs = sctx->gfx_cs;
@@ -272,6 +353,11 @@ static void si_emit_streamout_begin(struct si_context *sctx)
 
 void si_emit_streamout_end(struct si_context *sctx)
 {
+	if (sctx->chip_class >= GFX10) {
+		gfx10_emit_streamout_end(sctx);
+		return;
+	}
+
 	struct radeon_cmdbuf *cs = sctx->gfx_cs;
 	struct si_streamout_target **t = sctx->streamout.targets;
 	unsigned i;
@@ -320,6 +406,8 @@ void si_emit_streamout_end(struct si_context *sctx)
 
 static void si_emit_streamout_enable(struct si_context *sctx)
 {
+	assert(sctx->chip_class < GFX10);
+
 	radeon_set_context_reg_seq(sctx->gfx_cs, R_028B94_VGT_STRMOUT_CONFIG, 2);
 	radeon_emit(sctx->gfx_cs,
 		    S_028B94_STREAMOUT_0_EN(si_get_strmout_en(sctx)) |
@@ -344,15 +432,17 @@ static void si_set_streamout_enable(struct si_context *sctx, bool enable)
 					  (sctx->streamout.enabled_mask << 8) |
 					  (sctx->streamout.enabled_mask << 12);
 
-	if ((old_strmout_en != si_get_strmout_en(sctx)) ||
-            (old_hw_enabled_mask != sctx->streamout.hw_enabled_mask))
+	if (sctx->chip_class < GFX10 &&
+	    ((old_strmout_en != si_get_strmout_en(sctx)) ||
+	     (old_hw_enabled_mask != sctx->streamout.hw_enabled_mask)))
 		si_mark_atom_dirty(sctx, &sctx->atoms.s.streamout_enable);
 }
 
 void si_update_prims_generated_query_state(struct si_context *sctx,
 					   unsigned type, int diff)
 {
-	if (type == PIPE_QUERY_PRIMITIVES_GENERATED) {
+	if (sctx->chip_class < GFX10 &&
+	    type == PIPE_QUERY_PRIMITIVES_GENERATED) {
 		bool old_strmout_en = si_get_strmout_en(sctx);
 
 		sctx->streamout.num_prims_gen_queries += diff;
@@ -371,6 +461,11 @@ void si_init_streamout_functions(struct si_context *sctx)
 	sctx->b.create_stream_output_target = si_create_so_target;
 	sctx->b.stream_output_target_destroy = si_so_target_destroy;
 	sctx->b.set_stream_output_targets = si_set_streamout_targets;
-	sctx->atoms.s.streamout_begin.emit = si_emit_streamout_begin;
-	sctx->atoms.s.streamout_enable.emit = si_emit_streamout_enable;
+
+	if (sctx->chip_class >= GFX10) {
+		sctx->atoms.s.streamout_begin.emit = gfx10_emit_streamout_begin;
+	} else {
+		sctx->atoms.s.streamout_begin.emit = si_emit_streamout_begin;
+		sctx->atoms.s.streamout_enable.emit = si_emit_streamout_enable;
+	}
 }