radeonsi/gfx10: generate geometry shaders for NGG

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

4 files changed, 439 insertions, 4 deletions

diff --git a/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c b/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
index f5774b217ef..014fe1f96c9 100644
--- a/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
+++ b/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
@@ -27,12 +27,27 @@
 #include "sid.h"
 
 #include "util/u_memory.h"
+#include "util/u_prim.h"
 
 static LLVMValueRef get_wave_id_in_tg(struct si_shader_context *ctx)
 {
 	return si_unpack_param(ctx, ctx->param_merged_wave_info, 24, 4);
 }
 
+static LLVMValueRef get_tgsize(struct si_shader_context *ctx)
+{
+	return si_unpack_param(ctx, ctx->param_merged_wave_info, 28, 4);
+}
+
+static LLVMValueRef get_thread_id_in_tg(struct si_shader_context *ctx)
+{
+	LLVMBuilderRef builder = ctx->ac.builder;
+	LLVMValueRef tmp;
+	tmp = LLVMBuildMul(builder, get_wave_id_in_tg(ctx),
+			   LLVMConstInt(ctx->ac.i32, 64, false), "");
+	return LLVMBuildAdd(builder, tmp, ac_get_thread_id(&ctx->ac), "");
+}
+
 static LLVMValueRef ngg_get_vtx_cnt(struct si_shader_context *ctx)
 {
 	return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
@@ -263,3 +278,376 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
 
 	FREE(outputs);
 }
+
+static LLVMValueRef
+ngg_gs_get_vertex_storage(struct si_shader_context *ctx)
+{
+	const struct si_shader_selector *sel = ctx->shader->selector;
+	const struct tgsi_shader_info *info = &sel->info;
+
+	LLVMTypeRef elements[2] = {
+		LLVMArrayType(ctx->ac.i32, 4 * info->num_outputs),
+		LLVMArrayType(ctx->ac.i8, 4),
+	};
+	LLVMTypeRef type = LLVMStructTypeInContext(ctx->ac.context, elements, 2, false);
+	type = LLVMPointerType(LLVMArrayType(type, 0), AC_ADDR_SPACE_LDS);
+	return LLVMBuildBitCast(ctx->ac.builder, ctx->gs_ngg_emit, type, "");
+}
+
+/**
+ * Return a pointer to the LDS storage reserved for the N'th vertex, where N
+ * is in emit order; that is:
+ * - during the epilogue, N is the threadidx (relative to the entire threadgroup)
+ * - during vertex emit, i.e. while the API GS shader invocation is running,
+ *   N = threadidx * gs_max_out_vertices + emitidx
+ *
+ * Goals of the LDS memory layout:
+ * 1. Eliminate bank conflicts on write for geometry shaders that have all emits
+ *    in uniform control flow
+ * 2. Eliminate bank conflicts on read for export if, additionally, there is no
+ *    culling
+ * 3. Agnostic to the number of waves (since we don't know it before compiling)
+ * 4. Allow coalescing of LDS instructions (ds_write_b128 etc.)
+ * 5. Avoid wasting memory.
+ *
+ * We use an AoS layout due to point 4 (this also helps point 3). In an AoS
+ * layout, elimination of bank conflicts requires that each vertex occupy an
+ * odd number of dwords. We use the additional dword to store the output stream
+ * index as well as a flag to indicate whether this vertex ends a primitive
+ * for rasterization.
+ *
+ * Swizzling is required to satisfy points 1 and 2 simultaneously.
+ *
+ * Vertices are stored in export order (gsthread * gs_max_out_vertices + emitidx).
+ * Indices are swizzled in groups of 32, which ensures point 1 without
+ * disturbing point 2.
+ *
+ * \return an LDS pointer to type {[N x i32], [4 x i8]}
+ */
+static LLVMValueRef
+ngg_gs_vertex_ptr(struct si_shader_context *ctx, LLVMValueRef vertexidx)
+{
+	struct si_shader_selector *sel = ctx->shader->selector;
+	LLVMBuilderRef builder = ctx->ac.builder;
+	LLVMValueRef storage = ngg_gs_get_vertex_storage(ctx);
+
+	/* gs_max_out_vertices = 2^(write_stride_2exp) * some odd number */
+	unsigned write_stride_2exp = ffs(sel->gs_max_out_vertices) - 1;
+	if (write_stride_2exp) {
+		LLVMValueRef row =
+			LLVMBuildLShr(builder, vertexidx,
+				      LLVMConstInt(ctx->ac.i32, 5, false), "");
+		LLVMValueRef swizzle =
+			LLVMBuildAnd(builder, row,
+				     LLVMConstInt(ctx->ac.i32, (1u << write_stride_2exp) - 1,
+						  false), "");
+		vertexidx = LLVMBuildXor(builder, vertexidx, swizzle, "");
+	}
+
+	return ac_build_gep0(&ctx->ac, storage, vertexidx);
+}
+
+static LLVMValueRef
+ngg_gs_emit_vertex_ptr(struct si_shader_context *ctx, LLVMValueRef gsthread,
+		       LLVMValueRef emitidx)
+{
+	struct si_shader_selector *sel = ctx->shader->selector;
+	LLVMBuilderRef builder = ctx->ac.builder;
+	LLVMValueRef tmp;
+
+	tmp = LLVMConstInt(ctx->ac.i32, sel->gs_max_out_vertices, false);
+	tmp = LLVMBuildMul(builder, tmp, gsthread, "");
+	const LLVMValueRef vertexidx = LLVMBuildAdd(builder, tmp, emitidx, "");
+	return ngg_gs_vertex_ptr(ctx, vertexidx);
+}
+
+void gfx10_ngg_gs_emit_vertex(struct si_shader_context *ctx,
+			      unsigned stream,
+			      LLVMValueRef *addrs)
+{
+	const struct si_shader_selector *sel = ctx->shader->selector;
+	const struct tgsi_shader_info *info = &sel->info;
+	LLVMBuilderRef builder = ctx->ac.builder;
+	struct lp_build_if_state if_state;
+	LLVMValueRef tmp;
+	const LLVMValueRef vertexidx =
+		LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");
+
+	/* If this thread has already emitted the declared maximum number of
+	 * vertices, skip the write: excessive vertex emissions are not
+	 * supposed to have any effect.
+	 */
+	const LLVMValueRef can_emit =
+		LLVMBuildICmp(builder, LLVMIntULT, vertexidx,
+			      LLVMConstInt(ctx->i32, sel->gs_max_out_vertices, false), "");
+
+	tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
+	tmp = LLVMBuildSelect(builder, can_emit, tmp, vertexidx, "");
+	LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
+
+	lp_build_if(&if_state, &ctx->gallivm, can_emit);
+
+	const LLVMValueRef vertexptr =
+		ngg_gs_emit_vertex_ptr(ctx, get_thread_id_in_tg(ctx), vertexidx);
+	unsigned out_idx = 0;
+	for (unsigned i = 0; i < info->num_outputs; i++) {
+		for (unsigned chan = 0; chan < 4; chan++, out_idx++) {
+			if (!(info->output_usagemask[i] & (1 << chan)) ||
+			    ((info->output_streams[i] >> (2 * chan)) & 3) != stream)
+				continue;
+
+			LLVMValueRef out_val = LLVMBuildLoad(builder, addrs[4 * i + chan], "");
+			LLVMValueRef gep_idx[3] = {
+				ctx->ac.i32_0, /* implied C-style array */
+				ctx->ac.i32_0, /* first entry of struct */
+				LLVMConstInt(ctx->ac.i32, out_idx, false),
+			};
+			LLVMValueRef ptr = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
+
+			out_val = ac_to_integer(&ctx->ac, out_val);
+			LLVMBuildStore(builder, out_val, ptr);
+		}
+	}
+	assert(out_idx * 4 == sel->gsvs_vertex_size);
+
+	/* Determine and store whether this vertex completed a primitive. */
+	const LLVMValueRef curverts = LLVMBuildLoad(builder, ctx->gs_curprim_verts[stream], "");
+
+	tmp = LLVMConstInt(ctx->ac.i32, u_vertices_per_prim(sel->gs_output_prim) - 1, false);
+	const LLVMValueRef iscompleteprim =
+		LLVMBuildICmp(builder, LLVMIntUGE, curverts, tmp, "");
+
+	tmp = LLVMBuildAdd(builder, curverts, ctx->ac.i32_1, "");
+	LLVMBuildStore(builder, tmp, ctx->gs_curprim_verts[stream]);
+
+	LLVMValueRef gep_idx[3] = {
+		ctx->ac.i32_0, /* implied C-style array */
+		ctx->ac.i32_1, /* second struct entry */
+		LLVMConstInt(ctx->ac.i32, stream, false),
+	};
+	const LLVMValueRef primflagptr =
+		LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
+
+	tmp = LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i8, "");
+	LLVMBuildStore(builder, tmp, primflagptr);
+
+	lp_build_endif(&if_state);
+}
+
+void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx)
+{
+	const struct si_shader_selector *sel = ctx->shader->selector;
+	const struct tgsi_shader_info *info = &sel->info;
+	const unsigned verts_per_prim = u_vertices_per_prim(sel->gs_output_prim);
+	LLVMBuilderRef builder = ctx->ac.builder;
+	LLVMValueRef i8_0 = LLVMConstInt(ctx->ac.i8, 0, false);
+	LLVMValueRef tmp, tmp2;
+
+	/* Zero out remaining (non-emitted) primitive flags.
+	 *
+	 * Note: Alternatively, we could pass the relevant gs_next_vertex to
+	 *       the emit threads via LDS. This is likely worse in the expected
+	 *       typical case where each GS thread emits the full set of
+	 *       vertices.
+	 */
+	for (unsigned stream = 0; stream < 4; ++stream) {
+		if (!info->num_stream_output_components[stream])
+			continue;
+
+		const LLVMValueRef gsthread = get_thread_id_in_tg(ctx);
+
+		ac_build_bgnloop(&ctx->ac, 5100);
+
+		const LLVMValueRef vertexidx =
+			LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");
+		tmp = LLVMBuildICmp(builder, LLVMIntUGE, vertexidx,
+			LLVMConstInt(ctx->ac.i32, sel->gs_max_out_vertices, false), "");
+		ac_build_ifcc(&ctx->ac, tmp, 5101);
+		ac_build_break(&ctx->ac);
+		ac_build_endif(&ctx->ac, 5101);
+
+		tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
+		LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
+
+		tmp = ngg_gs_emit_vertex_ptr(ctx, gsthread, vertexidx);
+		LLVMValueRef gep_idx[3] = {
+			ctx->ac.i32_0, /* implied C-style array */
+			ctx->ac.i32_1, /* second entry of struct */
+			LLVMConstInt(ctx->ac.i32, stream, false),
+		};
+		tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
+		LLVMBuildStore(builder, i8_0, tmp);
+
+		ac_build_endloop(&ctx->ac, 5100);
+	}
+
+	lp_build_endif(&ctx->merged_wrap_if_state);
+
+	ac_build_s_barrier(&ctx->ac);
+
+	const LLVMValueRef tid = get_thread_id_in_tg(ctx);
+	LLVMValueRef num_emit_threads = ngg_get_prim_cnt(ctx);
+
+	/* TODO: streamout */
+
+	/* TODO: culling */
+
+	/* Determine vertex liveness. */
+	LLVMValueRef vertliveptr = lp_build_alloca(&ctx->gallivm, ctx->ac.i1, "vertexlive");
+
+	tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
+	ac_build_ifcc(&ctx->ac, tmp, 5120);
+	{
+		for (unsigned i = 0; i < verts_per_prim; ++i) {
+			const LLVMValueRef primidx =
+				LLVMBuildAdd(builder, tid,
+					     LLVMConstInt(ctx->ac.i32, i, false), "");
+
+			if (i > 0) {
+				tmp = LLVMBuildICmp(builder, LLVMIntULT, primidx, num_emit_threads, "");
+				ac_build_ifcc(&ctx->ac, tmp, 5121 + i);
+			}
+
+			/* Load primitive liveness */
+			tmp = ngg_gs_vertex_ptr(ctx, primidx);
+			LLVMValueRef gep_idx[3] = {
+				ctx->ac.i32_0, /* implicit C-style array */
+				ctx->ac.i32_1, /* second value of struct */
+				ctx->ac.i32_0, /* stream 0 */
+			};
+			tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
+			tmp = LLVMBuildLoad(builder, tmp, "");
+			const LLVMValueRef primlive =
+				LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
+
+			tmp = LLVMBuildLoad(builder, vertliveptr, "");
+			tmp = LLVMBuildOr(builder, tmp, primlive, ""),
+			LLVMBuildStore(builder, tmp, vertliveptr);
+
+			if (i > 0)
+				ac_build_endif(&ctx->ac, 5121 + i);
+		}
+	}
+	ac_build_endif(&ctx->ac, 5120);
+
+	/* Inclusive scan addition across the current wave. */
+	LLVMValueRef vertlive = LLVMBuildLoad(builder, vertliveptr, "");
+	struct ac_wg_scan vertlive_scan = {};
+	vertlive_scan.op = nir_op_iadd;
+	vertlive_scan.enable_reduce = true;
+	vertlive_scan.enable_exclusive = true;
+	vertlive_scan.src = vertlive;
+	vertlive_scan.scratch = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, ctx->i32_0);
+	vertlive_scan.waveidx = get_wave_id_in_tg(ctx);
+	vertlive_scan.numwaves = get_tgsize(ctx);
+	vertlive_scan.maxwaves = 8;
+
+	ac_build_wg_scan(&ctx->ac, &vertlive_scan);
+
+	/* Skip all exports (including index exports) when possible. At least on
+	 * early gfx10 revisions this is also to avoid hangs.
+	 */
+	LLVMValueRef have_exports =
+		LLVMBuildICmp(builder, LLVMIntNE, vertlive_scan.result_reduce, ctx->ac.i32_0, "");
+	num_emit_threads =
+		LLVMBuildSelect(builder, have_exports, num_emit_threads, ctx->ac.i32_0, "");
+
+	/* Allocate export space. Send this message as early as possible, to
+	 * hide the latency of the SQ <-> SPI roundtrip.
+	 *
+	 * Note: We could consider compacting primitives for export as well.
+	 *       PA processes 1 non-null prim / clock, but it fetches 4 DW of
+	 *       prim data per clock and skips null primitives at no additional
+	 *       cost. So compacting primitives can only be beneficial when
+	 *       there are 4 or more contiguous null primitives in the export
+	 *       (in the common case of single-dword prim exports).
+	 */
+	build_sendmsg_gs_alloc_req(ctx, vertlive_scan.result_reduce, num_emit_threads);
+
+	/* Setup the reverse vertex compaction permutation. We re-use stream 1
+	 * of the primitive liveness flags, relying on the fact that each
+	 * threadgroup can have at most 256 threads. */
+	ac_build_ifcc(&ctx->ac, vertlive, 5130);
+	{
+		tmp = ngg_gs_vertex_ptr(ctx, vertlive_scan.result_exclusive);
+		LLVMValueRef gep_idx[3] = {
+			ctx->ac.i32_0, /* implicit C-style array */
+			ctx->ac.i32_1, /* second value of struct */
+			ctx->ac.i32_1, /* stream 1 */
+		};
+		tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
+		tmp2 = LLVMBuildTrunc(builder, tid, ctx->ac.i8, "");
+		LLVMBuildStore(builder, tmp2, tmp);
+	}
+	ac_build_endif(&ctx->ac, 5130);
+
+	ac_build_s_barrier(&ctx->ac);
+
+	/* Export primitive data */
+	tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
+	ac_build_ifcc(&ctx->ac, tmp, 5140);
+	{
+		struct ngg_prim prim = {};
+		prim.num_vertices = verts_per_prim;
+
+		tmp = ngg_gs_vertex_ptr(ctx, tid);
+		LLVMValueRef gep_idx[3] = {
+			ctx->ac.i32_0, /* implicit C-style array */
+			ctx->ac.i32_1, /* second value of struct */
+			ctx->ac.i32_0, /* primflag */
+		};
+		tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
+		tmp = LLVMBuildLoad(builder, tmp, "");
+		prim.isnull = LLVMBuildICmp(builder, LLVMIntEQ, tmp,
+					    LLVMConstInt(ctx->ac.i8, 0, false), "");
+
+		for (unsigned i = 0; i < verts_per_prim; ++i) {
+			prim.index[i] = LLVMBuildSub(builder, vertlive_scan.result_exclusive,
+				LLVMConstInt(ctx->ac.i32, verts_per_prim - i - 1, false), "");
+			prim.edgeflag[i] = ctx->ac.i1false;
+		}
+
+		build_export_prim(ctx, &prim);
+	}
+	ac_build_endif(&ctx->ac, 5140);
+
+	/* Export position and parameter data */
+	tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, vertlive_scan.result_reduce, "");
+	ac_build_ifcc(&ctx->ac, tmp, 5145);
+	{
+		struct si_shader_output_values *outputs = NULL;
+		outputs = MALLOC(info->num_outputs * sizeof(outputs[0]));
+
+		tmp = ngg_gs_vertex_ptr(ctx, tid);
+		LLVMValueRef gep_idx[3] = {
+			ctx->ac.i32_0, /* implicit C-style array */
+			ctx->ac.i32_1, /* second value of struct */
+			ctx->ac.i32_1, /* stream 1: source data index */
+		};
+		tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
+		tmp = LLVMBuildLoad(builder, tmp, "");
+		tmp = LLVMBuildZExt(builder, tmp, ctx->ac.i32, "");
+		const LLVMValueRef vertexptr = ngg_gs_vertex_ptr(ctx, tmp);
+
+		unsigned out_idx = 0;
+		gep_idx[1] = ctx->ac.i32_0;
+		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];
+
+			for (unsigned j = 0; j < 4; j++, out_idx++) {
+				gep_idx[2] = LLVMConstInt(ctx->ac.i32, out_idx, false);
+				tmp = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
+				tmp = LLVMBuildLoad(builder, tmp, "");
+				outputs[i].values[j] = ac_to_float(&ctx->ac, tmp);
+				outputs[i].vertex_stream[j] =
+					(info->output_streams[i] >> (2 * j)) & 3;
+			}
+		}
+
+		si_llvm_export_vs(ctx, outputs, info->num_outputs);
+
+		FREE(outputs);
+	}
+	ac_build_endif(&ctx->ac, 5145);
+}
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 2ab1833579e..cc05b33ae1b 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -3401,11 +3401,15 @@ static void si_set_ls_return_value_for_tcs(struct si_shader_context *ctx)
 /* Pass GS inputs from ES to GS on GFX9. */
 static void si_set_es_return_value_for_gs(struct si_shader_context *ctx)
 {
+	LLVMBuilderRef builder = ctx->ac.builder;
 	LLVMValueRef ret = ctx->return_value;
 
 	ret = si_insert_input_ptr(ctx, ret, 0, 0);
 	ret = si_insert_input_ptr(ctx, ret, 1, 1);
-	ret = si_insert_input_ret(ctx, ret, ctx->param_gs2vs_offset, 2);
+	if (ctx->shader->key.as_ngg)
+		ret = LLVMBuildInsertValue(builder, ret, ctx->gs_tg_info, 2, "");
+	else
+		ret = si_insert_input_ret(ctx, ret, ctx->param_gs2vs_offset, 2);
 	ret = si_insert_input_ret(ctx, ret, ctx->param_merged_wave_info, 3);
 	ret = si_insert_input_ret(ctx, ret, ctx->param_merged_scratch_offset, 5);
 
@@ -3555,6 +3559,11 @@ static LLVMValueRef si_get_gs_wave_id(struct si_shader_context *ctx)
 
 static void emit_gs_epilogue(struct si_shader_context *ctx)
 {
+	if (ctx->shader->key.as_ngg) {
+		gfx10_ngg_gs_emit_epilogue(ctx);
+		return;
+	}
+
 	ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_NOP | AC_SENDMSG_GS_DONE,
 			 si_get_gs_wave_id(ctx));
 
@@ -4192,6 +4201,12 @@ static void si_llvm_emit_vertex(struct ac_shader_abi *abi,
 				LLVMValueRef *addrs)
 {
 	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
+
+	if (ctx->shader->key.as_ngg) {
+		gfx10_ngg_gs_emit_vertex(ctx, stream, addrs);
+		return;
+	}
+
 	struct tgsi_shader_info *info = &ctx->shader->selector->info;
 	struct si_shader *shader = ctx->shader;
 	struct lp_build_if_state if_state;
@@ -4284,6 +4299,11 @@ static void si_llvm_emit_primitive(struct ac_shader_abi *abi,
 {
 	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
 
+	if (ctx->shader->key.as_ngg) {
+		LLVMBuildStore(ctx->ac.builder, ctx->ac.i32_0, ctx->gs_curprim_verts[stream]);
+		return;
+	}
+
 	/* Signal primitive cut */
 	ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_CUT | AC_SENDMSG_GS | (stream << 8),
 			 si_get_gs_wave_id(ctx));
@@ -6087,11 +6107,27 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx)
 	}
 
 	if (ctx->type == PIPE_SHADER_GEOMETRY) {
-		int i;
-		for (i = 0; i < 4; i++) {
+		for (unsigned i = 0; i < 4; i++) {
 			ctx->gs_next_vertex[i] =
 				ac_build_alloca(&ctx->ac, ctx->i32, "");
 		}
+		if (shader->key.as_ngg) {
+			for (unsigned i = 0; i < 4; ++i) {
+				ctx->gs_curprim_verts[i] =
+					lp_build_alloca(&ctx->gallivm, ctx->ac.i32, "");
+			}
+
+			LLVMTypeRef a8i32 = LLVMArrayType(ctx->i32, 8);
+			ctx->gs_ngg_scratch = LLVMAddGlobalInAddressSpace(ctx->ac.module,
+				a8i32, "ngg_scratch", AC_ADDR_SPACE_LDS);
+			LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(a8i32));
+			LLVMSetAlignment(ctx->gs_ngg_scratch, 4);
+
+			ctx->gs_ngg_emit = LLVMAddGlobalInAddressSpace(ctx->ac.module,
+				LLVMArrayType(ctx->i32, 0), "ngg_emit", AC_ADDR_SPACE_LDS);
+			LLVMSetLinkage(ctx->gs_ngg_emit, LLVMExternalLinkage);
+			LLVMSetAlignment(ctx->gs_ngg_emit, 4);
+		}
 	}
 
 	if (sel->force_correct_derivs_after_kill) {
diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h b/src/gallium/drivers/radeonsi/si_shader_internal.h
index 5419a7312b1..55f32c66117 100644
--- a/src/gallium/drivers/radeonsi/si_shader_internal.h
+++ b/src/gallium/drivers/radeonsi/si_shader_internal.h
@@ -213,6 +213,9 @@ struct si_shader_context {
 
 	LLVMValueRef invoc0_tess_factors[6]; /* outer[4], inner[2] */
 	LLVMValueRef gs_next_vertex[4];
+	LLVMValueRef gs_curprim_verts[4];
+	LLVMValueRef gs_ngg_emit;
+	LLVMValueRef gs_ngg_scratch;
 	LLVMValueRef postponed_kill;
 	LLVMValueRef return_value;
 
@@ -382,5 +385,9 @@ LLVMValueRef si_unpack_param(struct si_shader_context *ctx,
 void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
 			     unsigned max_outputs,
 			     LLVMValueRef *addrs);
+void gfx10_ngg_gs_emit_vertex(struct si_shader_context *ctx,
+			      unsigned stream,
+			      LLVMValueRef *addrs);
+void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx);
 
 #endif
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index 2537dd90b5a..27835811cb7 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -2386,7 +2386,11 @@ 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) {