+static LLVMValueRef get_wave_id_in_tg(struct radv_shader_context *ctx)
+{
+ return ac_unpack_param(&ctx->ac, ctx->merged_wave_info, 24, 4);
+}
+
+static LLVMValueRef get_tgsize(struct radv_shader_context *ctx)
+{
+ return ac_unpack_param(&ctx->ac, ctx->merged_wave_info, 28, 4);
+}
+
+static LLVMValueRef get_thread_id_in_tg(struct radv_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 radv_shader_context *ctx)
+{
+ return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
+ LLVMConstInt(ctx->ac.i32, 12, false),
+ LLVMConstInt(ctx->ac.i32, 9, false),
+ false);
+}
+
+static LLVMValueRef ngg_get_prim_cnt(struct radv_shader_context *ctx)
+{
+ return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
+ LLVMConstInt(ctx->ac.i32, 22, false),
+ LLVMConstInt(ctx->ac.i32, 9, false),
+ false);
+}
+
+static LLVMValueRef
+ngg_gs_get_vertex_storage(struct radv_shader_context *ctx)
+{
+ unsigned num_outputs = util_bitcount64(ctx->output_mask);
+
+ LLVMTypeRef elements[2] = {
+ LLVMArrayType(ctx->ac.i32, 4 * 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 radv_shader_context *ctx, LLVMValueRef vertexidx)
+{
+ 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(ctx->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 radv_shader_context *ctx, LLVMValueRef gsthread,
+ LLVMValueRef emitidx)
+{
+ LLVMBuilderRef builder = ctx->ac.builder;
+ LLVMValueRef tmp;
+
+ tmp = LLVMConstInt(ctx->ac.i32, ctx->gs_max_out_vertices, false);
+ tmp = LLVMBuildMul(builder, tmp, gsthread, "");
+ const LLVMValueRef vertexidx = LLVMBuildAdd(builder, tmp, emitidx, "");
+ return ngg_gs_vertex_ptr(ctx, vertexidx);
+}
+
+/* Send GS Alloc Req message from the first wave of the group to SPI.
+ * Message payload is:
+ * - bits 0..10: vertices in group
+ * - bits 12..22: primitives in group
+ */
+static void build_sendmsg_gs_alloc_req(struct radv_shader_context *ctx,
+ LLVMValueRef vtx_cnt,
+ LLVMValueRef prim_cnt)
+{
+ LLVMBuilderRef builder = ctx->ac.builder;
+ LLVMValueRef tmp;
+
+ tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
+ ac_build_ifcc(&ctx->ac, tmp, 5020);
+
+ tmp = LLVMBuildShl(builder, prim_cnt, LLVMConstInt(ctx->ac.i32, 12, false),"");
+ tmp = LLVMBuildOr(builder, tmp, vtx_cnt, "");
+ ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_ALLOC_REQ, tmp);
+
+ ac_build_endif(&ctx->ac, 5020);
+}
+
+struct ngg_prim {
+ unsigned num_vertices;
+ LLVMValueRef isnull;
+ LLVMValueRef index[3];
+ LLVMValueRef edgeflag[3];
+};
+
+static void build_export_prim(struct radv_shader_context *ctx,
+ const struct ngg_prim *prim)
+{
+ LLVMBuilderRef builder = ctx->ac.builder;
+ struct ac_export_args args;
+ LLVMValueRef tmp;
+
+ tmp = LLVMBuildZExt(builder, prim->isnull, ctx->ac.i32, "");
+ args.out[0] = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->ac.i32, 31, false), "");
+
+ for (unsigned i = 0; i < prim->num_vertices; ++i) {
+ tmp = LLVMBuildShl(builder, prim->index[i],
+ LLVMConstInt(ctx->ac.i32, 10 * i, false), "");
+ args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
+ tmp = LLVMBuildZExt(builder, prim->edgeflag[i], ctx->ac.i32, "");
+ tmp = LLVMBuildShl(builder, tmp,
+ LLVMConstInt(ctx->ac.i32, 10 * i + 9, false), "");
+ args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
+ }
+
+ args.out[0] = LLVMBuildBitCast(builder, args.out[0], ctx->ac.f32, "");
+ args.out[1] = LLVMGetUndef(ctx->ac.f32);
+ args.out[2] = LLVMGetUndef(ctx->ac.f32);
+ args.out[3] = LLVMGetUndef(ctx->ac.f32);
+
+ args.target = V_008DFC_SQ_EXP_PRIM;
+ args.enabled_channels = 1;
+ args.done = true;
+ args.valid_mask = false;
+ args.compr = false;
+
+ ac_build_export(&ctx->ac, &args);
+}
+
+static void
+handle_ngg_outputs_post(struct radv_shader_context *ctx)
+{
+ LLVMBuilderRef builder = ctx->ac.builder;
+ struct ac_build_if_state if_state;
+ unsigned num_vertices = 3;
+ LLVMValueRef tmp;
+
+ assert((ctx->stage == MESA_SHADER_VERTEX ||
+ ctx->stage == MESA_SHADER_TESS_EVAL) && !ctx->is_gs_copy_shader);
+
+ LLVMValueRef prims_in_wave = ac_unpack_param(&ctx->ac, ctx->merged_wave_info, 8, 8);
+ LLVMValueRef vtx_in_wave = ac_unpack_param(&ctx->ac, ctx->merged_wave_info, 0, 8);
+ LLVMValueRef is_gs_thread = LLVMBuildICmp(builder, LLVMIntULT,
+ ac_get_thread_id(&ctx->ac), prims_in_wave, "");
+ LLVMValueRef is_es_thread = LLVMBuildICmp(builder, LLVMIntULT,
+ ac_get_thread_id(&ctx->ac), vtx_in_wave, "");
+ LLVMValueRef vtxindex[] = {
+ ac_unpack_param(&ctx->ac, ctx->gs_vtx_offset[0], 0, 16),
+ ac_unpack_param(&ctx->ac, ctx->gs_vtx_offset[0], 16, 16),
+ ac_unpack_param(&ctx->ac, ctx->gs_vtx_offset[2], 0, 16),
+ };
+
+ /* TODO: streamout */
+
+ /* Copy Primitive IDs from GS threads to the LDS address corresponding
+ * to the ES thread of the provoking vertex.
+ */
+ if (ctx->stage == MESA_SHADER_VERTEX &&
+ ctx->options->key.vs_common_out.export_prim_id) {
+ /* TODO: streamout */
+
+ ac_build_ifcc(&ctx->ac, is_gs_thread, 5400);
+ /* Extract the PROVOKING_VTX_INDEX field. */
+ LLVMValueRef provoking_vtx_in_prim =
+ LLVMConstInt(ctx->ac.i32, 0, false);
+
+ /* provoking_vtx_index = vtxindex[provoking_vtx_in_prim]; */
+ LLVMValueRef indices = ac_build_gather_values(&ctx->ac, vtxindex, 3);
+ LLVMValueRef provoking_vtx_index =
+ LLVMBuildExtractElement(builder, indices, provoking_vtx_in_prim, "");
+
+ LLVMBuildStore(builder, ctx->abi.gs_prim_id,
+ ac_build_gep0(&ctx->ac, ctx->esgs_ring, provoking_vtx_index));
+ ac_build_endif(&ctx->ac, 5400);
+ }
+
+ /* TODO: primitive culling */
+
+ build_sendmsg_gs_alloc_req(ctx, ngg_get_vtx_cnt(ctx), ngg_get_prim_cnt(ctx));
+
+ /* TODO: streamout queries */
+ /* Export primitive data to the index buffer. Format is:
+ * - bits 0..8: index 0
+ * - bit 9: edge flag 0
+ * - bits 10..18: index 1
+ * - bit 19: edge flag 1
+ * - bits 20..28: index 2
+ * - bit 29: edge flag 2
+ * - bit 31: null primitive (skip)
+ *
+ * For the first version, we will always build up all three indices
+ * independent of the primitive type. The additional garbage data
+ * shouldn't hurt.
+ *
+ * TODO: culling depends on the primitive type, so can have some
+ * interaction here.
+ */
+ ac_nir_build_if(&if_state, ctx, is_gs_thread);
+ {
+ struct ngg_prim prim = {};
+
+ prim.num_vertices = num_vertices;
+ prim.isnull = ctx->ac.i1false;
+ memcpy(prim.index, vtxindex, sizeof(vtxindex[0]) * 3);
+
+ for (unsigned i = 0; i < num_vertices; ++i) {
+ tmp = LLVMBuildLShr(builder, ctx->abi.gs_invocation_id,
+ LLVMConstInt(ctx->ac.i32, 8 + i, false), "");
+ prim.edgeflag[i] = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
+ }
+
+ build_export_prim(ctx, &prim);
+ }
+ ac_nir_build_endif(&if_state);
+
+ /* Export per-vertex data (positions and parameters). */
+ ac_nir_build_if(&if_state, ctx, is_es_thread);
+ {
+ struct radv_vs_output_info *outinfo =
+ ctx->stage == MESA_SHADER_TESS_EVAL ? &ctx->shader_info->tes.outinfo : &ctx->shader_info->vs.outinfo;
+
+ /* Exporting the primitive ID is handled below. */
+ /* TODO: use the new VS export path */
+ handle_vs_outputs_post(ctx, false,
+ ctx->options->key.vs_common_out.export_clip_dists,
+ outinfo);
+
+ if (ctx->options->key.vs_common_out.export_prim_id) {
+ unsigned param_count = outinfo->param_exports;
+ LLVMValueRef values[4];
+
+ if (ctx->stage == MESA_SHADER_VERTEX) {
+ /* Wait for GS stores to finish. */
+ ac_build_s_barrier(&ctx->ac);
+
+ tmp = ac_build_gep0(&ctx->ac, ctx->esgs_ring,
+ get_thread_id_in_tg(ctx));
+ values[0] = LLVMBuildLoad(builder, tmp, "");
+ } else {
+ assert(ctx->stage == MESA_SHADER_TESS_EVAL);
+ values[0] = ctx->abi.tes_patch_id;
+ }
+
+ values[0] = ac_to_float(&ctx->ac, values[0]);
+ for (unsigned j = 1; j < 4; j++)
+ values[j] = ctx->ac.f32_0;
+
+ radv_export_param(ctx, param_count, values, 0x1);
+
+ outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID] = param_count++;
+ outinfo->export_prim_id = true;
+ outinfo->param_exports = param_count;
+ }
+ }
+ ac_nir_build_endif(&if_state);
+}
+
+static void gfx10_ngg_gs_emit_prologue(struct radv_shader_context *ctx)
+{
+ /* Zero out the part of LDS scratch that is used to accumulate the
+ * per-stream generated primitive count.
+ */
+ LLVMBuilderRef builder = ctx->ac.builder;
+ LLVMValueRef scratchptr = ctx->gs_ngg_scratch;
+ LLVMValueRef tid = get_thread_id_in_tg(ctx);
+ LLVMBasicBlockRef merge_block;
+ LLVMValueRef cond;
+
+ LLVMValueRef fn = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->ac.builder));
+ LLVMBasicBlockRef then_block = LLVMAppendBasicBlockInContext(ctx->ac.context, fn, "");
+ merge_block = LLVMAppendBasicBlockInContext(ctx->ac.context, fn, "");
+
+ cond = LLVMBuildICmp(builder, LLVMIntULT, tid, LLVMConstInt(ctx->ac.i32, 4, false), "");
+ LLVMBuildCondBr(ctx->ac.builder, cond, then_block, merge_block);
+ LLVMPositionBuilderAtEnd(ctx->ac.builder, then_block);
+
+ LLVMValueRef ptr = ac_build_gep0(&ctx->ac, scratchptr, tid);
+ LLVMBuildStore(builder, ctx->ac.i32_0, ptr);
+
+ LLVMBuildBr(ctx->ac.builder, merge_block);
+ LLVMPositionBuilderAtEnd(ctx->ac.builder, merge_block);
+
+ ac_build_s_barrier(&ctx->ac);
+}
+
+static void gfx10_ngg_gs_emit_epilogue_1(struct radv_shader_context *ctx)
+{
+ LLVMBuilderRef builder = ctx->ac.builder;
+ LLVMValueRef i8_0 = LLVMConstInt(ctx->ac.i8, 0, false);
+ LLVMValueRef tmp;
+
+ /* 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) {
+ unsigned num_components;
+
+ num_components =
+ ctx->shader_info->info.gs.num_stream_output_components[stream];
+ if (!num_components)
+ 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, ctx->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);
+ }
+}
+
+static void gfx10_ngg_gs_emit_epilogue_2(struct radv_shader_context *ctx)
+{
+ const unsigned verts_per_prim = si_conv_gl_prim_to_vertices(ctx->gs_output_prim);
+ LLVMBuilderRef builder = ctx->ac.builder;
+ LLVMValueRef tmp, tmp2;
+
+ 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 = ac_build_alloca(&ctx->ac, 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->ac.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 radv_vs_output_info *outinfo = &ctx->shader_info->vs.outinfo;
+ bool export_view_index = ctx->options->key.has_multiview_view_index;
+ struct radv_shader_output_values *outputs;
+ unsigned noutput = 0;
+
+ /* Allocate a temporary array for the output values. */
+ unsigned num_outputs = util_bitcount64(ctx->output_mask) + export_view_index;
+ outputs = calloc(num_outputs, sizeof(outputs[0]));
+
+ memset(outinfo->vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
+ sizeof(outinfo->vs_output_param_offset));
+ outinfo->pos_exports = 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);
+
+ if (ctx->output_mask & (1ull << VARYING_SLOT_PSIZ)) {
+ outinfo->writes_pointsize = true;
+ }
+
+ if (ctx->output_mask & (1ull << VARYING_SLOT_LAYER)) {
+ outinfo->writes_layer = true;
+ }
+
+ if (ctx->output_mask & (1ull << VARYING_SLOT_VIEWPORT)) {
+ outinfo->writes_viewport_index = true;
+ }
+
+ unsigned out_idx = 0;
+ gep_idx[1] = ctx->ac.i32_0;
+ for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
+ if (!(ctx->output_mask & (1ull << i)))
+ continue;
+
+ outputs[noutput].slot_name = i;
+ outputs[noutput].slot_index = i == VARYING_SLOT_CLIP_DIST1;
+
+ outputs[noutput].usage_mask = ctx->shader_info->info.gs.output_usage_mask[i];
+ int length = util_last_bit(outputs[noutput].usage_mask);
+
+ for (unsigned j = 0; j < length; j++, out_idx++) {
+ gep_idx[2] = LLVMConstInt(ctx->ac.i32, out_idx, false);
+ tmp = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
+ tmp = LLVMBuildLoad(builder, tmp, "");
+
+ LLVMTypeRef type = LLVMGetAllocatedType(ctx->abi.outputs[ac_llvm_reg_index_soa(i, j)]);
+ if (ac_get_type_size(type) == 2) {
+ tmp = ac_to_integer(&ctx->ac, tmp);
+ tmp = LLVMBuildTrunc(ctx->ac.builder, tmp, ctx->ac.i16, "");
+ }
+
+ outputs[noutput].values[j] = ac_to_float(&ctx->ac, tmp);
+ }
+
+ for (unsigned j = length; j < 4; j++)
+ outputs[noutput].values[j] = LLVMGetUndef(ctx->ac.f32);
+
+ noutput++;
+ }
+
+ /* Export ViewIndex. */
+ if (export_view_index) {
+ outinfo->writes_layer = true;
+
+ outputs[noutput].slot_name = VARYING_SLOT_LAYER;
+ outputs[noutput].slot_index = 0;
+ outputs[noutput].usage_mask = 0x1;
+ outputs[noutput].values[0] = ac_to_float(&ctx->ac, ctx->abi.view_index);
+ for (unsigned j = 1; j < 4; j++)
+ outputs[noutput].values[j] = ctx->ac.f32_0;
+ noutput++;
+ }
+
+ radv_llvm_export_vs(ctx, outputs, noutput, outinfo,
+ ctx->options->key.vs_common_out.export_clip_dists);
+ FREE(outputs);
+ }
+ ac_build_endif(&ctx->ac, 5145);
+}
+
+static void gfx10_ngg_gs_emit_vertex(struct radv_shader_context *ctx,
+ unsigned stream,
+ LLVMValueRef *addrs)
+{
+ LLVMBuilderRef builder = ctx->ac.builder;
+ 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->ac.i32, ctx->gs_max_out_vertices, false), "");
+ ac_build_kill_if_false(&ctx->ac, can_emit);
+
+ tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
+ tmp = LLVMBuildSelect(builder, can_emit, tmp, vertexidx, "");
+ LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
+
+ 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 < AC_LLVM_MAX_OUTPUTS; ++i) {
+ unsigned output_usage_mask =
+ ctx->shader_info->info.gs.output_usage_mask[i];
+ uint8_t output_stream =
+ ctx->shader_info->info.gs.output_streams[i];
+ LLVMValueRef *out_ptr = &addrs[i * 4];
+ int length = util_last_bit(output_usage_mask);
+
+ if (!(ctx->output_mask & (1ull << i)) ||
+ output_stream != stream)
+ continue;
+
+ for (unsigned j = 0; j < length; j++, out_idx++) {
+ if (!(output_usage_mask & (1 << j)))
+ continue;
+
+ LLVMValueRef out_val = LLVMBuildLoad(ctx->ac.builder,
+ out_ptr[j], "");
+ 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);
+ out_val = LLVMBuildZExtOrBitCast(ctx->ac.builder, out_val, ctx->ac.i32, "");
+
+ LLVMBuildStore(builder, out_val, ptr);
+ }
+ }
+ assert(out_idx * 4 <= ctx->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, si_conv_gl_prim_to_vertices(ctx->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);
+
+ tmp = LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
+ tmp = LLVMBuildAdd(builder, tmp, LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i32, ""), "");
+ LLVMBuildStore(builder, tmp, ctx->gs_generated_prims[stream]);
+}
+