ac_nir_lower_ngg.c

/*
 * Copyright © 2021 Valve Corporation
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 *
 */

#include "ac_nir.h"
#include "nir_builder.h"
#include "u_math.h"
#include "u_vector.h"

enum {
   nggc_passflag_used_by_pos = 1,
   nggc_passflag_used_by_other = 2,
   nggc_passflag_used_by_both = nggc_passflag_used_by_pos | nggc_passflag_used_by_other,
};

typedef struct
{
   nir_ssa_def *ssa;
   nir_variable *var;
} saved_uniform;

typedef struct
{
   nir_variable *position_value_var;
   nir_variable *prim_exp_arg_var;
   nir_variable *es_accepted_var;
   nir_variable *gs_accepted_var;

   struct u_vector saved_uniforms;

   bool passthrough;
   bool export_prim_id;
   bool early_prim_export;
   unsigned wave_size;
   unsigned max_num_waves;
   unsigned num_vertices_per_primitives;
   unsigned provoking_vtx_idx;
   unsigned max_es_num_vertices;
   unsigned total_lds_bytes;

   uint64_t inputs_needed_by_pos;
   uint64_t inputs_needed_by_others;
} lower_ngg_nogs_state;

typedef struct
{
   /* bitsize of this component (max 32), or 0 if it's never written at all */
   uint8_t bit_size : 6;
   /* output stream index  */
   uint8_t stream : 2;
} gs_output_component_info;

typedef struct
{
   nir_variable *output_vars[VARYING_SLOT_MAX][4];
   nir_variable *current_clear_primflag_idx_var;
   int const_out_vtxcnt[4];
   int const_out_prmcnt[4];
   unsigned wave_size;
   unsigned max_num_waves;
   unsigned num_vertices_per_primitive;
   unsigned lds_addr_gs_out_vtx;
   unsigned lds_addr_gs_scratch;
   unsigned lds_bytes_per_gs_out_vertex;
   unsigned lds_offs_primflags;
   bool found_out_vtxcnt[4];
   bool output_compile_time_known;
   bool provoking_vertex_last;
   gs_output_component_info output_component_info[VARYING_SLOT_MAX][4];
} lower_ngg_gs_state;

typedef struct {
   nir_variable *pre_cull_position_value_var;
} remove_culling_shader_outputs_state;

typedef struct {
   nir_variable *pos_value_replacement;
} remove_extra_position_output_state;

typedef struct {
   nir_ssa_def *reduction_result;
   nir_ssa_def *excl_scan_result;
} wg_scan_result;

/* Per-vertex LDS layout of culling shaders */
enum {
   /* Position of the ES vertex (at the beginning for alignment reasons) */
   lds_es_pos_x = 0,
   lds_es_pos_y = 4,
   lds_es_pos_z = 8,
   lds_es_pos_w = 12,

   /* 1 when the vertex is accepted, 0 if it should be culled */
   lds_es_vertex_accepted = 16,
   /* ID of the thread which will export the current thread's vertex */
   lds_es_exporter_tid = 17,

   /* Repacked arguments - also listed separately for VS and TES */
   lds_es_arg_0 = 20,

   /* VS arguments which need to be repacked */
   lds_es_vs_vertex_id = 20,
   lds_es_vs_instance_id = 24,

   /* TES arguments which need to be repacked */
   lds_es_tes_u = 20,
   lds_es_tes_v = 24,
   lds_es_tes_rel_patch_id = 28,
   lds_es_tes_patch_id = 32,
};

typedef struct {
   nir_ssa_def *num_repacked_invocations;
   nir_ssa_def *repacked_invocation_index;
} wg_repack_result;

/**
 * Repacks invocations in the current workgroup to eliminate gaps between them.
 *
 * Uses 1 dword of LDS per 4 waves (1 byte of LDS per wave).
 * Assumes that all invocations in the workgroup are active (exec = -1).
 */
static wg_repack_result
repack_invocations_in_workgroup(nir_builder *b, nir_ssa_def *input_bool,
                                unsigned lds_addr_base, unsigned max_num_waves,
                                unsigned wave_size)
{
   /* Input boolean: 1 if the current invocation should survive the repack. */
   assert(input_bool->bit_size == 1);

   /* STEP 1. Count surviving invocations in the current wave.
    *
    * Implemented by a scalar instruction that simply counts the number of bits set in a 32/64-bit mask.
    */

   nir_ssa_def *input_mask = nir_build_ballot(b, 1, wave_size, input_bool);
   nir_ssa_def *surviving_invocations_in_current_wave = nir_bit_count(b, input_mask);

   /* If we know at compile time that the workgroup has only 1 wave, no further steps are necessary. */
   if (max_num_waves == 1) {
      wg_repack_result r = {
         .num_repacked_invocations = surviving_invocations_in_current_wave,
         .repacked_invocation_index = nir_build_mbcnt_amd(b, input_mask, nir_imm_int(b, 0)),
      };
      return r;
   }

   /* STEP 2. Waves tell each other their number of surviving invocations.
    *
    * Each wave activates only its first lane (exec = 1), which stores the number of surviving
    * invocations in that wave into the LDS, then reads the numbers from every wave.
    *
    * The workgroup size of NGG shaders is at most 256, which means
    * the maximum number of waves is 4 in Wave64 mode and 8 in Wave32 mode.
    * Each wave writes 1 byte, so it's up to 8 bytes, so at most 2 dwords are necessary.
    */

   const unsigned num_lds_dwords = DIV_ROUND_UP(max_num_waves, 4);
   assert(num_lds_dwords <= 2);

   nir_ssa_def *wave_id = nir_build_load_subgroup_id(b);
   nir_ssa_def *dont_care = nir_ssa_undef(b, 1, num_lds_dwords * 32);
   nir_if *if_first_lane = nir_push_if(b, nir_build_elect(b, 1));

   nir_build_store_shared(b, nir_u2u8(b, surviving_invocations_in_current_wave), wave_id, .base = lds_addr_base, .align_mul = 1u, .write_mask = 0x1u);

   nir_scoped_barrier(b, .execution_scope=NIR_SCOPE_WORKGROUP, .memory_scope=NIR_SCOPE_WORKGROUP,
                         .memory_semantics=NIR_MEMORY_ACQ_REL, .memory_modes=nir_var_mem_shared);

   nir_ssa_def *packed_counts = nir_build_load_shared(b, 1, num_lds_dwords * 32, nir_imm_int(b, 0), .base = lds_addr_base, .align_mul = 8u);

   nir_pop_if(b, if_first_lane);

   packed_counts = nir_if_phi(b, packed_counts, dont_care);

   /* STEP 3. Compute the repacked invocation index and the total number of surviving invocations.
    *
    * By now, every wave knows the number of surviving invocations in all waves.
    * Each number is 1 byte, and they are packed into up to 2 dwords.
    *
    * Each lane N will sum the number of surviving invocations from waves 0 to N-1.
    * If the workgroup has M waves, then each wave will use only its first M+1 lanes for this.
    * (Other lanes are not deactivated but their calculation is not used.)
    *
    * - We read the sum from the lane whose id is the current wave's id.
    *   Add the masked bitcount to this, and we get the repacked invocation index.
    * - We read the sum from the lane whose id is the number of waves in the workgroup.
    *   This is the total number of surviving invocations in the workgroup.
    */

   nir_ssa_def *num_waves = nir_build_load_num_subgroups(b);

   /* sel = 0x01010101 * lane_id + 0x03020100 */
   nir_ssa_def *lane_id = nir_load_subgroup_invocation(b);
   nir_ssa_def *packed_id = nir_build_byte_permute_amd(b, nir_imm_int(b, 0), lane_id, nir_imm_int(b, 0));
   nir_ssa_def *sel = nir_iadd_imm_nuw(b, packed_id, 0x03020100);
   nir_ssa_def *sum = NULL;

   if (num_lds_dwords == 1) {
      /* Broadcast the packed data we read from LDS (to the first 16 lanes, but we only care up to num_waves). */
      nir_ssa_def *packed_dw = nir_build_lane_permute_16_amd(b, packed_counts, nir_imm_int(b, 0), nir_imm_int(b, 0));

      /* Use byte-permute to filter out the bytes not needed by the current lane. */
      nir_ssa_def *filtered_packed = nir_build_byte_permute_amd(b, packed_dw, nir_imm_int(b, 0), sel);

      /* Horizontally add the packed bytes. */
      sum = nir_sad_u8x4(b, filtered_packed, nir_imm_int(b, 0), nir_imm_int(b, 0));
   } else if (num_lds_dwords == 2) {
      /* Create selectors for the byte-permutes below. */
      nir_ssa_def *dw0_selector = nir_build_lane_permute_16_amd(b, sel, nir_imm_int(b, 0x44443210), nir_imm_int(b, 0x4));
      nir_ssa_def *dw1_selector = nir_build_lane_permute_16_amd(b, sel, nir_imm_int(b, 0x32100000), nir_imm_int(b, 0x4));

      /* Broadcast the packed data we read from LDS (to the first 16 lanes, but we only care up to num_waves). */
      nir_ssa_def *packed_dw0 = nir_build_lane_permute_16_amd(b, nir_unpack_64_2x32_split_x(b, packed_counts), nir_imm_int(b, 0), nir_imm_int(b, 0));
      nir_ssa_def *packed_dw1 = nir_build_lane_permute_16_amd(b, nir_unpack_64_2x32_split_y(b, packed_counts), nir_imm_int(b, 0), nir_imm_int(b, 0));

      /* Use byte-permute to filter out the bytes not needed by the current lane. */
      nir_ssa_def *filtered_packed_dw0 = nir_build_byte_permute_amd(b, packed_dw0, nir_imm_int(b, 0), dw0_selector);
      nir_ssa_def *filtered_packed_dw1 = nir_build_byte_permute_amd(b, packed_dw1, nir_imm_int(b, 0), dw1_selector);

      /* Horizontally add the packed bytes. */
      sum = nir_sad_u8x4(b, filtered_packed_dw0, nir_imm_int(b, 0), nir_imm_int(b, 0));
      sum = nir_sad_u8x4(b, filtered_packed_dw1, nir_imm_int(b, 0), sum);
   } else {
      unreachable("Unimplemented NGG wave count");
   }

   nir_ssa_def *wg_repacked_index_base = nir_build_read_invocation(b, sum, wave_id);
   nir_ssa_def *wg_num_repacked_invocations = nir_build_read_invocation(b, sum, num_waves);
   nir_ssa_def *wg_repacked_index = nir_build_mbcnt_amd(b, input_mask, wg_repacked_index_base);

   wg_repack_result r = {
      .num_repacked_invocations = wg_num_repacked_invocations,
      .repacked_invocation_index = wg_repacked_index,
   };

   return r;
}

static nir_ssa_def *
pervertex_lds_addr(nir_builder *b, nir_ssa_def *vertex_idx, unsigned per_vtx_bytes)
{
   return nir_imul_imm(b, vertex_idx, per_vtx_bytes);
}

static nir_ssa_def *
emit_pack_ngg_prim_exp_arg(nir_builder *b, unsigned num_vertices_per_primitives,
                           nir_ssa_def *vertex_indices[3], nir_ssa_def *is_null_prim)
{
   nir_ssa_def *arg = vertex_indices[0];

   for (unsigned i = 0; i < num_vertices_per_primitives; ++i) {
      assert(vertex_indices[i]);

      if (i)
         arg = nir_ior(b, arg, nir_ishl(b, vertex_indices[i], nir_imm_int(b, 10u * i)));

      if (b->shader->info.stage == MESA_SHADER_VERTEX) {
         nir_ssa_def *edgeflag = nir_build_load_initial_edgeflag_amd(b, 32, nir_imm_int(b, i));
         arg = nir_ior(b, arg, nir_ishl(b, edgeflag, nir_imm_int(b, 10u * i + 9u)));
      }
   }

   if (is_null_prim) {
      if (is_null_prim->bit_size == 1)
         is_null_prim = nir_b2i32(b, is_null_prim);
      assert(is_null_prim->bit_size == 32);
      arg = nir_ior(b, arg, nir_ishl(b, is_null_prim, nir_imm_int(b, 31u)));
   }

   return arg;
}

static nir_ssa_def *
ngg_input_primitive_vertex_index(nir_builder *b, unsigned vertex)
{
   /* TODO: This is RADV specific. We'll need to refactor RADV and/or RadeonSI to match. */
   return nir_ubfe(b, nir_build_load_gs_vertex_offset_amd(b, .base = vertex / 2u * 2u),
                      nir_imm_int(b, (vertex % 2u) * 16u), nir_imm_int(b, 16u));
}

static nir_ssa_def *
emit_ngg_nogs_prim_exp_arg(nir_builder *b, lower_ngg_nogs_state *st)
{
   if (st->passthrough) {
      assert(!st->export_prim_id || b->shader->info.stage != MESA_SHADER_VERTEX);
      return nir_build_load_packed_passthrough_primitive_amd(b);
   } else {
      nir_ssa_def *vtx_idx[3] = {0};

      vtx_idx[0] = ngg_input_primitive_vertex_index(b, 0);
      vtx_idx[1] = st->num_vertices_per_primitives >= 2
               ? ngg_input_primitive_vertex_index(b, 1)
               : nir_imm_zero(b, 1, 32);
      vtx_idx[2] = st->num_vertices_per_primitives >= 3
               ? ngg_input_primitive_vertex_index(b, 2)
               : nir_imm_zero(b, 1, 32);

      return emit_pack_ngg_prim_exp_arg(b, st->num_vertices_per_primitives, vtx_idx, NULL);
   }
}

static void
emit_ngg_nogs_prim_export(nir_builder *b, lower_ngg_nogs_state *st, nir_ssa_def *arg)
{
   nir_if *if_gs_thread = nir_push_if(b, nir_build_has_input_primitive_amd(b));
   {
      if (!arg)
         arg = emit_ngg_nogs_prim_exp_arg(b, st);

      if (st->export_prim_id && b->shader->info.stage == MESA_SHADER_VERTEX) {
         /* Copy Primitive IDs from GS threads to the LDS address corresponding to the ES thread of the provoking vertex. */
         nir_ssa_def *prim_id = nir_build_load_primitive_id(b);
         nir_ssa_def *provoking_vtx_idx = ngg_input_primitive_vertex_index(b, st->provoking_vtx_idx);
         nir_ssa_def *addr = pervertex_lds_addr(b, provoking_vtx_idx, 4u);

         nir_build_store_shared(b,  prim_id, addr, .write_mask = 1u, .align_mul = 4u);
      }

      nir_build_export_primitive_amd(b, arg);
   }
   nir_pop_if(b, if_gs_thread);
}

static void
emit_store_ngg_nogs_es_primitive_id(nir_builder *b)
{
   nir_ssa_def *prim_id = NULL;

   if (b->shader->info.stage == MESA_SHADER_VERTEX) {
      /* Workgroup barrier - wait for GS threads to store primitive ID in LDS. */
      nir_scoped_barrier(b, .execution_scope = NIR_SCOPE_WORKGROUP, .memory_scope = NIR_SCOPE_WORKGROUP,
                            .memory_semantics = NIR_MEMORY_ACQ_REL, .memory_modes = nir_var_mem_shared);

      /* LDS address where the primitive ID is stored */
      nir_ssa_def *thread_id_in_threadgroup = nir_build_load_local_invocation_index(b);
      nir_ssa_def *addr =  pervertex_lds_addr(b, thread_id_in_threadgroup, 4u);

      /* Load primitive ID from LDS */
      prim_id = nir_build_load_shared(b, 1, 32, addr, .align_mul = 4u);
   } else if (b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
      /* Just use tess eval primitive ID, which is the same as the patch ID. */
      prim_id = nir_build_load_primitive_id(b);
   }

   nir_io_semantics io_sem = {
      .location = VARYING_SLOT_PRIMITIVE_ID,
      .num_slots = 1,
   };

   nir_build_store_output(b, prim_id, nir_imm_zero(b, 1, 32),
                          .base = io_sem.location,
                          .write_mask = 1u, .src_type = nir_type_uint32, .io_semantics = io_sem);
}

static bool
remove_culling_shader_output(nir_builder *b, nir_instr *instr, void *state)
{
   remove_culling_shader_outputs_state *s = (remove_culling_shader_outputs_state *) state;

   if (instr->type != nir_instr_type_intrinsic)
      return false;

   nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);

   /* These are not allowed in VS / TES */
   assert(intrin->intrinsic != nir_intrinsic_store_per_vertex_output &&
          intrin->intrinsic != nir_intrinsic_load_per_vertex_input);

   /* We are only interested in output stores now */
   if (intrin->intrinsic != nir_intrinsic_store_output)
      return false;

   b->cursor = nir_before_instr(instr);

   /* Position output - store the value to a variable, remove output store */
   nir_io_semantics io_sem = nir_intrinsic_io_semantics(intrin);
   if (io_sem.location == VARYING_SLOT_POS) {
      /* TODO: check if it's indirect, etc? */
      unsigned writemask = nir_intrinsic_write_mask(intrin);
      nir_ssa_def *store_val = intrin->src[0].ssa;
      nir_store_var(b, s->pre_cull_position_value_var, store_val, writemask);
   }

   /* Remove all output stores */
   nir_instr_remove(instr);
   return true;
}

static void
remove_culling_shader_outputs(nir_shader *culling_shader, lower_ngg_nogs_state *nogs_state, nir_variable *pre_cull_position_value_var)
{
   remove_culling_shader_outputs_state s = {
      .pre_cull_position_value_var = pre_cull_position_value_var,
   };

   nir_shader_instructions_pass(culling_shader, remove_culling_shader_output,
                                nir_metadata_block_index | nir_metadata_dominance, &s);

   /* Remove dead code resulting from the deleted outputs. */
   bool progress;
   do {
      progress = false;
      NIR_PASS(progress, culling_shader, nir_opt_dead_write_vars);
      NIR_PASS(progress, culling_shader, nir_opt_dce);
      NIR_PASS(progress, culling_shader, nir_opt_dead_cf);
   } while (progress);
}

static void
rewrite_uses_to_var(nir_builder *b, nir_ssa_def *old_def, nir_variable *replacement_var, unsigned replacement_var_channel)
{
   if (old_def->parent_instr->type == nir_instr_type_load_const)
      return;

   b->cursor = nir_after_instr(old_def->parent_instr);
   if (b->cursor.instr->type == nir_instr_type_phi)
      b->cursor = nir_after_phis(old_def->parent_instr->block);

   nir_ssa_def *pos_val_rep = nir_load_var(b, replacement_var);
   nir_ssa_def *replacement = nir_channel(b, pos_val_rep, replacement_var_channel);

   if (old_def->num_components > 1) {
      /* old_def uses a swizzled vector component.
       * There is no way to replace the uses of just a single vector component,
       * so instead create a new vector and replace all uses of the old vector.
       */
      nir_ssa_def *old_def_elements[NIR_MAX_VEC_COMPONENTS] = {0};
      for (unsigned j = 0; j < old_def->num_components; ++j)
         old_def_elements[j] = nir_channel(b, old_def, j);
      replacement = nir_vec(b, old_def_elements, old_def->num_components);
   }

   nir_ssa_def_rewrite_uses_after(old_def, replacement, replacement->parent_instr);
}

static bool
remove_extra_pos_output(nir_builder *b, nir_instr *instr, void *state)
{
   remove_extra_position_output_state *s = (remove_extra_position_output_state *) state;

   if (instr->type != nir_instr_type_intrinsic)
      return false;

   nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);

   /* These are not allowed in VS / TES */
   assert(intrin->intrinsic != nir_intrinsic_store_per_vertex_output &&
          intrin->intrinsic != nir_intrinsic_load_per_vertex_input);

   /* We are only interested in output stores now */
   if (intrin->intrinsic != nir_intrinsic_store_output)
      return false;

   nir_io_semantics io_sem = nir_intrinsic_io_semantics(intrin);
   if (io_sem.location != VARYING_SLOT_POS)
      return false;

   b->cursor = nir_before_instr(instr);

   /* In case other outputs use what we calculated for pos,
    * try to avoid calculating it again by rewriting the usages
    * of the store components here.
    */
   nir_ssa_def *store_val = intrin->src[0].ssa;
   unsigned store_pos_component = nir_intrinsic_component(intrin);

   nir_instr_remove(instr);

   if (store_val->parent_instr->type == nir_instr_type_alu) {
      nir_alu_instr *alu = nir_instr_as_alu(store_val->parent_instr);
      if (nir_op_is_vec(alu->op)) {
         /* Output store uses a vector, we can easily rewrite uses of each vector element. */

         unsigned num_vec_src = 0;
         if (alu->op == nir_op_mov)
            num_vec_src = 1;
         else if (alu->op == nir_op_vec2)
            num_vec_src = 2;
         else if (alu->op == nir_op_vec3)
            num_vec_src = 3;
         else if (alu->op == nir_op_vec4)
            num_vec_src = 4;
         assert(num_vec_src);

         /* Remember the current components whose uses we wish to replace.
          * This is needed because rewriting one source can affect the others too.
          */
         nir_ssa_def *vec_comps[NIR_MAX_VEC_COMPONENTS] = {0};
         for (unsigned i = 0; i < num_vec_src; i++)
            vec_comps[i] = alu->src[i].src.ssa;

         for (unsigned i = 0; i < num_vec_src; i++)
            rewrite_uses_to_var(b, vec_comps[i], s->pos_value_replacement, store_pos_component + i);
      } else {
         rewrite_uses_to_var(b, store_val, s->pos_value_replacement, store_pos_component);
      }
   } else {
      rewrite_uses_to_var(b, store_val, s->pos_value_replacement, store_pos_component);
   }

   return true;
}

static void
remove_extra_pos_outputs(nir_shader *shader, lower_ngg_nogs_state *nogs_state)
{
   remove_extra_position_output_state s = {
      .pos_value_replacement = nogs_state->position_value_var,
   };

   nir_shader_instructions_pass(shader, remove_extra_pos_output,
                                nir_metadata_block_index | nir_metadata_dominance, &s);
}

/**
 * Perform vertex compaction after culling.
 *
 * 1. Repack surviving ES invocations (this determines which lane will export which vertex)
 * 2. Surviving ES vertex invocations store their data to LDS
 * 3. Emit GS_ALLOC_REQ
 * 4. Repacked invocations load the vertex data from LDS
 * 5. GS threads update their vertex indices
 */
static void
compact_vertices_after_culling(nir_builder *b,
                               lower_ngg_nogs_state *nogs_state,
                               nir_variable *vertices_in_wave_var,
                               nir_variable *primitives_in_wave_var,
                               nir_variable **repacked_arg_vars,
                               nir_variable **gs_vtxaddr_vars,
                               nir_ssa_def *invocation_index,
                               nir_ssa_def *es_vertex_lds_addr,
                               unsigned ngg_scratch_lds_base_addr,
                               unsigned pervertex_lds_bytes,
                               unsigned max_exported_args)
{
   nir_variable *es_accepted_var = nogs_state->es_accepted_var;
   nir_variable *gs_accepted_var = nogs_state->gs_accepted_var;
   nir_variable *position_value_var = nogs_state->position_value_var;
   nir_variable *prim_exp_arg_var = nogs_state->prim_exp_arg_var;

   nir_ssa_def *es_accepted = nir_load_var(b, es_accepted_var);

   /* Repack the vertices that survived the culling. */
   wg_repack_result rep = repack_invocations_in_workgroup(b, es_accepted, ngg_scratch_lds_base_addr,
                                                          nogs_state->max_num_waves, nogs_state->wave_size);
   nir_ssa_def *num_live_vertices_in_workgroup = rep.num_repacked_invocations;
   nir_ssa_def *es_exporter_tid = rep.repacked_invocation_index;

   nir_if *if_es_accepted = nir_push_if(b, es_accepted);
   {
      nir_ssa_def *exporter_addr = pervertex_lds_addr(b, es_exporter_tid, pervertex_lds_bytes);

      /* Store the exporter thread's index to the LDS space of the current thread so GS threads can load it */
      nir_build_store_shared(b, nir_u2u8(b, es_exporter_tid), es_vertex_lds_addr, .base = lds_es_exporter_tid, .align_mul = 1u, .write_mask = 0x1u);

      /* Store the current thread's position output to the exporter thread's LDS space */
      nir_ssa_def *pos = nir_load_var(b, position_value_var);
      nir_build_store_shared(b, pos, exporter_addr, .base = lds_es_pos_x, .align_mul = 4u, .write_mask = 0xfu);

      /* Store the current thread's repackable arguments to the exporter thread's LDS space */
      for (unsigned i = 0; i < max_exported_args; ++i) {
         nir_ssa_def *arg_val = nir_load_var(b, repacked_arg_vars[i]);
         nir_build_store_shared(b, arg_val, exporter_addr, .base = lds_es_arg_0 + 4u * i, .align_mul = 4u, .write_mask = 0x1u);
      }
   }
   nir_pop_if(b, if_es_accepted);

   /* If all vertices are culled, set primitive count to 0 as well. */
   nir_ssa_def *num_exported_prims = nir_build_load_workgroup_num_input_primitives_amd(b);
   num_exported_prims = nir_bcsel(b, nir_ieq_imm(b, num_live_vertices_in_workgroup, 0u), nir_imm_int(b, 0u), num_exported_prims);

   nir_if *if_wave_0 = nir_push_if(b, nir_ieq(b, nir_build_load_subgroup_id(b), nir_imm_int(b, 0)));
   {
      /* Tell the final vertex and primitive count to the HW.
       * We do this here to mask some of the latency of the LDS.
       */
      nir_build_alloc_vertices_and_primitives_amd(b, num_live_vertices_in_workgroup, num_exported_prims);
   }
   nir_pop_if(b, if_wave_0);

   /* Calculate the number of vertices and primitives left in the current wave */
   nir_ssa_def *has_vtx_after_culling = nir_ilt(b, invocation_index, num_live_vertices_in_workgroup);
   nir_ssa_def *has_prm_after_culling = nir_ilt(b, invocation_index, num_exported_prims);
   nir_ssa_def *vtx_cnt = nir_bit_count(b, nir_build_ballot(b, 1, nogs_state->wave_size, has_vtx_after_culling));
   nir_ssa_def *prm_cnt = nir_bit_count(b, nir_build_ballot(b, 1, nogs_state->wave_size, has_prm_after_culling));
   nir_store_var(b, vertices_in_wave_var, vtx_cnt, 0x1u);
   nir_store_var(b, primitives_in_wave_var, prm_cnt, 0x1u);

   /* TODO: Consider adding a shortcut exit.
    * Waves that have no vertices and primitives left can s_endpgm right here.
    */

   nir_scoped_barrier(b, .execution_scope=NIR_SCOPE_WORKGROUP, .memory_scope=NIR_SCOPE_WORKGROUP,
                         .memory_semantics=NIR_MEMORY_ACQ_REL, .memory_modes=nir_var_mem_shared);

   nir_if *if_packed_es_thread = nir_push_if(b, nir_ilt(b, invocation_index, num_live_vertices_in_workgroup));
   {
      /* Read position from the current ES thread's LDS space (written by the exported vertex's ES thread) */
      nir_ssa_def *exported_pos = nir_build_load_shared(b, 4, 32, es_vertex_lds_addr, .base = lds_es_pos_x, .align_mul = 4u);
      nir_store_var(b, position_value_var, exported_pos, 0xfu);

      /* Read the repacked arguments */
      for (unsigned i = 0; i < max_exported_args; ++i) {
         nir_ssa_def *arg_val = nir_build_load_shared(b, 1, 32, es_vertex_lds_addr, .base = lds_es_arg_0 + 4u * i, .align_mul = 4u);
         nir_store_var(b, repacked_arg_vars[i], arg_val, 0x1u);
      }
   }
   nir_pop_if(b, if_packed_es_thread);

   nir_if *if_gs_accepted = nir_push_if(b, nir_load_var(b, gs_accepted_var));
   {
      nir_ssa_def *exporter_vtx_indices[3] = {0};

      /* Load the index of the ES threads that will export the current GS thread's vertices */
      for (unsigned v = 0; v < 3; ++v) {
         nir_ssa_def *vtx_addr = nir_load_var(b, gs_vtxaddr_vars[v]);
         nir_ssa_def *exporter_vtx_idx = nir_build_load_shared(b, 1, 8, vtx_addr, .base = lds_es_exporter_tid, .align_mul = 1u);
         exporter_vtx_indices[v] = nir_u2u32(b, exporter_vtx_idx);
      }

      nir_ssa_def *prim_exp_arg = emit_pack_ngg_prim_exp_arg(b, 3, exporter_vtx_indices, NULL);
      nir_store_var(b, prim_exp_arg_var, prim_exp_arg, 0x1u);
   }
   nir_pop_if(b, if_gs_accepted);
}

static void
analyze_shader_before_culling_walk(nir_ssa_def *ssa,
                                   uint8_t flag,
                                   lower_ngg_nogs_state *nogs_state)
{
   nir_instr *instr = ssa->parent_instr;
   uint8_t old_pass_flags = instr->pass_flags;
   instr->pass_flags |= flag;

   if (instr->pass_flags == old_pass_flags)
      return; /* Already visited. */

   switch (instr->type) {
   case nir_instr_type_intrinsic: {
      nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);

      /* VS input loads and SSBO loads are actually VRAM reads on AMD HW. */
      switch (intrin->intrinsic) {
      case nir_intrinsic_load_input: {
         nir_io_semantics in_io_sem = nir_intrinsic_io_semantics(intrin);
         uint64_t in_mask = UINT64_C(1) << (uint64_t) in_io_sem.location;
         if (instr->pass_flags & nggc_passflag_used_by_pos)
            nogs_state->inputs_needed_by_pos |= in_mask;
         else if (instr->pass_flags & nggc_passflag_used_by_other)
            nogs_state->inputs_needed_by_others |= in_mask;
         break;
      }
      default:
         break;
      }

      break;
   }
   case nir_instr_type_alu: {
      nir_alu_instr *alu = nir_instr_as_alu(instr);
      unsigned num_srcs = nir_op_infos[alu->op].num_inputs;

      for (unsigned i = 0; i < num_srcs; ++i) {
         analyze_shader_before_culling_walk(alu->src[i].src.ssa, flag, nogs_state);
      }

      break;
   }
   case nir_instr_type_phi: {
      nir_phi_instr *phi = nir_instr_as_phi(instr);
      nir_foreach_phi_src_safe(phi_src, phi) {
         analyze_shader_before_culling_walk(phi_src->src.ssa, flag, nogs_state);
      }

      break;
   }
   default:
      break;
   }
}

static void
analyze_shader_before_culling(nir_shader *shader, lower_ngg_nogs_state *nogs_state)
{
   nir_foreach_function(func, shader) {
      nir_foreach_block(block, func->impl) {
         nir_foreach_instr(instr, block) {
            instr->pass_flags = 0;

            if (instr->type != nir_instr_type_intrinsic)
               continue;

            nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
            if (intrin->intrinsic != nir_intrinsic_store_output)
               continue;

            nir_io_semantics io_sem = nir_intrinsic_io_semantics(intrin);
            nir_ssa_def *store_val = intrin->src[0].ssa;
            uint8_t flag = io_sem.location == VARYING_SLOT_POS ? nggc_passflag_used_by_pos : nggc_passflag_used_by_other;
            analyze_shader_before_culling_walk(store_val, flag, nogs_state);
         }
      }
   }
}

/**
 * Save the reusable SSA definitions to variables so that the
 * bottom shader part can reuse them from the top part.
 *
 * 1. We create a new function temporary variable for reusables,
 *    and insert a store+load.
 * 2. The shader is cloned (the top part is created), then the
 *    control flow is reinserted (for the bottom part.)
 * 3. For reusables, we delete the variable stores from the
 *    bottom part. This will make them use the variables from
 *    the top part and DCE the redundant instructions.
 */
static void
save_reusable_variables(nir_builder *b, lower_ngg_nogs_state *nogs_state)
{
   ASSERTED int vec_ok = u_vector_init(&nogs_state->saved_uniforms, sizeof(saved_uniform), 4 * sizeof(saved_uniform));
   assert(vec_ok);

   unsigned loop_depth = 0;

   nir_foreach_block_safe(block, b->impl) {
      /* Check whether we're in a loop. */
      nir_cf_node *next_cf_node = nir_cf_node_next(&block->cf_node);
      nir_cf_node *prev_cf_node = nir_cf_node_prev(&block->cf_node);
      if (next_cf_node && next_cf_node->type == nir_cf_node_loop)
         loop_depth++;
      if (prev_cf_node && prev_cf_node->type == nir_cf_node_loop)
         loop_depth--;

      /* The following code doesn't make sense in loops, so just skip it then. */
      if (loop_depth)
         continue;

      nir_foreach_instr_safe(instr, block) {
         /* Find instructions whose SSA definitions are used by both
          * the top and bottom parts of the shader. In this case, it
          * makes sense to try to reuse these from the top part.
          */
         if ((instr->pass_flags & nggc_passflag_used_by_both) != nggc_passflag_used_by_both)
            continue;

         nir_ssa_def *ssa = NULL;

         switch (instr->type) {
         case nir_instr_type_alu: {
            nir_alu_instr *alu = nir_instr_as_alu(instr);
            if (alu->dest.dest.ssa.divergent)
               continue;
            /* Ignore uniform floats because they regress VGPR usage too much */
            if (nir_op_infos[alu->op].output_type & nir_type_float)
               continue;
            ssa = &alu->dest.dest.ssa;
            break;
         }
         case nir_instr_type_intrinsic: {
            nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
            if (!nir_intrinsic_can_reorder(intrin) ||
                !nir_intrinsic_infos[intrin->intrinsic].has_dest ||
                intrin->dest.ssa.divergent)
               continue;
            ssa = &intrin->dest.ssa;
            break;
         }
         case nir_instr_type_phi: {
            nir_phi_instr *phi = nir_instr_as_phi(instr);
            if (phi->dest.ssa.divergent)
               continue;
            ssa = &phi->dest.ssa;
            break;
         }
         default:
            continue;
         }

         assert(ssa);

         enum glsl_base_type base_type = GLSL_TYPE_UINT;
         switch (ssa->bit_size) {
         case 8: base_type = GLSL_TYPE_UINT8; break;
         case 16: base_type = GLSL_TYPE_UINT16; break;
         case 32: base_type = GLSL_TYPE_UINT; break;
         case 64: base_type = GLSL_TYPE_UINT64; break;
         default: continue;
         }

         const struct glsl_type *t = ssa->num_components == 1
                                     ? glsl_scalar_type(base_type)
                                     : glsl_vector_type(base_type, ssa->num_components);

         saved_uniform *saved = (saved_uniform *) u_vector_add(&nogs_state->saved_uniforms);
         assert(saved);

         saved->var = nir_local_variable_create(b->impl, t, NULL);
         saved->ssa = ssa;

         b->cursor = instr->type == nir_instr_type_phi
                     ? nir_after_instr_and_phis(instr)
                     : nir_after_instr(instr);
         nir_store_var(b, saved->var, saved->ssa, BITFIELD_MASK(ssa->num_components));
         nir_ssa_def *reloaded = nir_load_var(b, saved->var);
         nir_ssa_def_rewrite_uses_after(ssa, reloaded, reloaded->parent_instr);
      }
   }
}

/**
 * Reuses suitable variables from the top part of the shader,
 * by deleting their stores from the bottom part.
 */
static void
apply_reusable_variables(nir_builder *b, lower_ngg_nogs_state *nogs_state)
{
   if (!u_vector_length(&nogs_state->saved_uniforms)) {
      u_vector_finish(&nogs_state->saved_uniforms);
      return;
   }

   nir_foreach_block_reverse_safe(block, b->impl) {
      nir_foreach_instr_reverse_safe(instr, block) {
         if (instr->type != nir_instr_type_intrinsic)
            continue;
         nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);

         /* When we found any of these intrinsics, it means
          * we reached the top part and we must stop.
          */
         if (intrin->intrinsic == nir_intrinsic_overwrite_subgroup_num_vertices_and_primitives_amd ||
             intrin->intrinsic == nir_intrinsic_alloc_vertices_and_primitives_amd ||
             intrin->intrinsic == nir_intrinsic_export_primitive_amd)
            goto done;

         if (intrin->intrinsic != nir_intrinsic_store_deref)
            continue;
         nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
         if (deref->deref_type != nir_deref_type_var)
            continue;

         saved_uniform *saved;
         u_vector_foreach(saved, &nogs_state->saved_uniforms) {
            if (saved->var == deref->var) {
               nir_instr_remove(instr);
            }
         }
      }
   }

   done:
   u_vector_finish(&nogs_state->saved_uniforms);
}

static void
add_deferred_attribute_culling(nir_builder *b, nir_cf_list *original_extracted_cf, lower_ngg_nogs_state *nogs_state)
{
   assert(b->shader->info.outputs_written & (1 << VARYING_SLOT_POS));

   bool uses_instance_id = BITSET_TEST(b->shader->info.system_values_read, SYSTEM_VALUE_INSTANCE_ID);
   bool uses_tess_primitive_id = BITSET_TEST(b->shader->info.system_values_read, SYSTEM_VALUE_PRIMITIVE_ID);

   unsigned max_exported_args = b->shader->info.stage == MESA_SHADER_VERTEX ? 2 : 4;
   if (b->shader->info.stage == MESA_SHADER_VERTEX && !uses_instance_id)
      max_exported_args--;
   else if (b->shader->info.stage == MESA_SHADER_TESS_EVAL && !uses_tess_primitive_id)
      max_exported_args--;

   unsigned pervertex_lds_bytes = lds_es_arg_0 + max_exported_args * 4u;
   unsigned total_es_lds_bytes = pervertex_lds_bytes * nogs_state->max_es_num_vertices;
   unsigned max_num_waves = nogs_state->max_num_waves;
   unsigned ngg_scratch_lds_base_addr = ALIGN(total_es_lds_bytes, 8u);
   unsigned ngg_scratch_lds_bytes = DIV_ROUND_UP(max_num_waves, 4u);
   nogs_state->total_lds_bytes = ngg_scratch_lds_base_addr + ngg_scratch_lds_bytes;

   nir_function_impl *impl = nir_shader_get_entrypoint(b->shader);

   /* Create some helper variables. */
   nir_variable *position_value_var = nogs_state->position_value_var;
   nir_variable *prim_exp_arg_var = nogs_state->prim_exp_arg_var;
   nir_variable *gs_accepted_var = nogs_state->gs_accepted_var;
   nir_variable *es_accepted_var = nogs_state->es_accepted_var;
   nir_variable *vertices_in_wave_var = nir_local_variable_create(impl, glsl_uint_type(), "vertices_in_wave");
   nir_variable *primitives_in_wave_var = nir_local_variable_create(impl, glsl_uint_type(), "primitives_in_wave");
   nir_variable *gs_vtxaddr_vars[3] = {
      nir_local_variable_create(impl, glsl_uint_type(), "gs_vtx0_addr"),
      nir_local_variable_create(impl, glsl_uint_type(), "gs_vtx1_addr"),
      nir_local_variable_create(impl, glsl_uint_type(), "gs_vtx2_addr"),
   };
   nir_variable *repacked_arg_vars[4] = {
      nir_local_variable_create(impl, glsl_uint_type(), "repacked_arg_0"),
      nir_local_variable_create(impl, glsl_uint_type(), "repacked_arg_1"),
      nir_local_variable_create(impl, glsl_uint_type(), "repacked_arg_2"),
      nir_local_variable_create(impl, glsl_uint_type(), "repacked_arg_3"),
   };

   /* Top part of the culling shader (aka. position shader part)
    *
    * We clone the full ES shader and emit it here, but we only really care
    * about its position output, so we delete every other output from this part.
    * The position output is stored into a temporary variable, and reloaded later.
    */

   b->cursor = nir_before_cf_list(&impl->body);

   nir_if *if_es_thread = nir_push_if(b, nir_build_has_input_vertex_amd(b));
   {
      /* Initialize the position output variable to zeroes, in case not all VS/TES invocations store the output.
       * The spec doesn't require it, but we use (0, 0, 0, 1) because some games rely on that.
       */
      nir_store_var(b, position_value_var, nir_imm_vec4(b, 0.0f, 0.0f, 0.0f, 1.0f), 0xfu);

      /* Now reinsert a clone of the shader code */
      struct hash_table *remap_table = _mesa_pointer_hash_table_create(NULL);
      nir_cf_list_clone_and_reinsert(original_extracted_cf, &if_es_thread->cf_node, b->cursor, remap_table);
      _mesa_hash_table_destroy(remap_table, NULL);
      b->cursor = nir_after_cf_list(&if_es_thread->then_list);

      /* Remember the current thread's shader arguments */
      if (b->shader->info.stage == MESA_SHADER_VERTEX) {
         nir_store_var(b, repacked_arg_vars[0], nir_build_load_vertex_id_zero_base(b), 0x1u);
         if (uses_instance_id)
            nir_store_var(b, repacked_arg_vars[1], nir_build_load_instance_id(b), 0x1u);
      } else if (b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
         nir_ssa_def *tess_coord = nir_build_load_tess_coord(b);
         nir_store_var(b, repacked_arg_vars[0], nir_channel(b, tess_coord, 0), 0x1u);
         nir_store_var(b, repacked_arg_vars[1], nir_channel(b, tess_coord, 1), 0x1u);
         nir_store_var(b, repacked_arg_vars[2], nir_build_load_tess_rel_patch_id_amd(b), 0x1u);
         if (uses_tess_primitive_id)
            nir_store_var(b, repacked_arg_vars[3], nir_build_load_primitive_id(b), 0x1u);
      } else {
         unreachable("Should be VS or TES.");
      }
   }
   nir_pop_if(b, if_es_thread);

   /* Remove all non-position outputs, and put the position output into the variable. */
   nir_metadata_preserve(impl, nir_metadata_none);
   remove_culling_shader_outputs(b->shader, nogs_state, position_value_var);
   b->cursor = nir_after_cf_list(&impl->body);

   /* Run culling algorithms if culling is enabled.
    *
    * NGG culling can be enabled or disabled in runtime.
    * This is determined by a SGPR shader argument which is acccessed
    * by the following NIR intrinsic.
    */

   nir_if *if_cull_en = nir_push_if(b, nir_build_load_cull_any_enabled_amd(b));
   {
      nir_ssa_def *invocation_index = nir_build_load_local_invocation_index(b);
      nir_ssa_def *es_vertex_lds_addr = pervertex_lds_addr(b, invocation_index, pervertex_lds_bytes);

      /* ES invocations store their vertex data to LDS for GS threads to read. */
      if_es_thread = nir_push_if(b, nir_build_has_input_vertex_amd(b));
      {
         /* Store position components that are relevant to culling in LDS */
         nir_ssa_def *pre_cull_pos = nir_load_var(b, position_value_var);
         nir_ssa_def *pre_cull_w = nir_channel(b, pre_cull_pos, 3);
         nir_build_store_shared(b, pre_cull_w, es_vertex_lds_addr, .write_mask = 0x1u, .align_mul = 4, .base = lds_es_pos_w);
         nir_ssa_def *pre_cull_x_div_w = nir_fdiv(b, nir_channel(b, pre_cull_pos, 0), pre_cull_w);
         nir_ssa_def *pre_cull_y_div_w = nir_fdiv(b, nir_channel(b, pre_cull_pos, 1), pre_cull_w);
         nir_build_store_shared(b, nir_vec2(b, pre_cull_x_div_w, pre_cull_y_div_w), es_vertex_lds_addr, .write_mask = 0x3u, .align_mul = 4, .base = lds_es_pos_x);

         /* Clear out the ES accepted flag in LDS */
         nir_build_store_shared(b, nir_imm_zero(b, 1, 8), es_vertex_lds_addr, .write_mask = 0x1u, .align_mul = 4, .base = lds_es_vertex_accepted);
      }
      nir_pop_if(b, if_es_thread);

      nir_scoped_barrier(b, .execution_scope=NIR_SCOPE_WORKGROUP, .memory_scope=NIR_SCOPE_WORKGROUP,
                            .memory_semantics=NIR_MEMORY_ACQ_REL, .memory_modes=nir_var_mem_shared);

      nir_store_var(b, gs_accepted_var, nir_imm_bool(b, false), 0x1u);
      nir_store_var(b, prim_exp_arg_var, nir_imm_int(b, 1 << 31), 0x1u);

      /* GS invocations load the vertex data and perform the culling. */
      nir_if *if_gs_thread = nir_push_if(b, nir_build_has_input_primitive_amd(b));
      {
         /* Load vertex indices from input VGPRs */
         nir_ssa_def *vtx_idx[3] = {0};
         for (unsigned vertex = 0; vertex < 3; ++vertex)
            vtx_idx[vertex] = ngg_input_primitive_vertex_index(b, vertex);

         nir_ssa_def *vtx_addr[3] = {0};
         nir_ssa_def *pos[3][4] = {0};

         /* Load W positions of vertices first because the culling code will use these first */
         for (unsigned vtx = 0; vtx < 3; ++vtx) {
            vtx_addr[vtx] = pervertex_lds_addr(b, vtx_idx[vtx], pervertex_lds_bytes);
            pos[vtx][3] = nir_build_load_shared(b, 1, 32, vtx_addr[vtx], .align_mul = 4u, .base = lds_es_pos_w);
            nir_store_var(b, gs_vtxaddr_vars[vtx], vtx_addr[vtx], 0x1u);
         }

         /* Load the X/W, Y/W positions of vertices */
         for (unsigned vtx = 0; vtx < 3; ++vtx) {
            nir_ssa_def *xy = nir_build_load_shared(b, 2, 32, vtx_addr[vtx], .align_mul = 4u, .base = lds_es_pos_x);
            pos[vtx][0] = nir_channel(b, xy, 0);
            pos[vtx][1] = nir_channel(b, xy, 1);
         }

         /* See if the current primitive is accepted */
         nir_ssa_def *accepted = ac_nir_cull_triangle(b, nir_imm_bool(b, true), pos);
         nir_store_var(b, gs_accepted_var, accepted, 0x1u);

         nir_if *if_gs_accepted = nir_push_if(b, accepted);
         {
            /* Store the accepted state to LDS for ES threads */
            for (unsigned vtx = 0; vtx < 3; ++vtx)
               nir_build_store_shared(b, nir_imm_intN_t(b, 0xff, 8), vtx_addr[vtx], .base = lds_es_vertex_accepted, .align_mul = 4u, .write_mask = 0x1u);
         }
         nir_pop_if(b, if_gs_accepted);
      }
      nir_pop_if(b, if_gs_thread);

      nir_scoped_barrier(b, .execution_scope=NIR_SCOPE_WORKGROUP, .memory_scope=NIR_SCOPE_WORKGROUP,
                            .memory_semantics=NIR_MEMORY_ACQ_REL, .memory_modes=nir_var_mem_shared);

      nir_store_var(b, es_accepted_var, nir_imm_bool(b, false), 0x1u);

      /* ES invocations load their accepted flag from LDS. */
      if_es_thread = nir_push_if(b, nir_build_has_input_vertex_amd(b));
      {
         nir_ssa_def *accepted = nir_build_load_shared(b, 1, 8u, es_vertex_lds_addr, .base = lds_es_vertex_accepted, .align_mul = 4u);
         nir_ssa_def *accepted_bool = nir_ine(b, accepted, nir_imm_intN_t(b, 0, 8));
         nir_store_var(b, es_accepted_var, accepted_bool, 0x1u);
      }
      nir_pop_if(b, if_es_thread);

      /* Vertex compaction. */
      compact_vertices_after_culling(b, nogs_state,
                                     vertices_in_wave_var, primitives_in_wave_var,
                                     repacked_arg_vars, gs_vtxaddr_vars,
                                     invocation_index, es_vertex_lds_addr,
                                     ngg_scratch_lds_base_addr, pervertex_lds_bytes, max_exported_args);
   }
   nir_push_else(b, if_cull_en);
   {
      /* When culling is disabled, we do the same as we would without culling. */
      nir_if *if_wave_0 = nir_push_if(b, nir_ieq(b, nir_build_load_subgroup_id(b), nir_imm_int(b, 0)));
      {
         nir_ssa_def *vtx_cnt = nir_build_load_workgroup_num_input_vertices_amd(b);
         nir_ssa_def *prim_cnt = nir_build_load_workgroup_num_input_primitives_amd(b);
         nir_build_alloc_vertices_and_primitives_amd(b, vtx_cnt, prim_cnt);
      }
      nir_pop_if(b, if_wave_0);
      nir_store_var(b, prim_exp_arg_var, emit_ngg_nogs_prim_exp_arg(b, nogs_state), 0x1u);

      nir_ssa_def *vtx_cnt = nir_bit_count(b, nir_build_ballot(b, 1, nogs_state->wave_size, nir_build_has_input_vertex_amd(b)));
      nir_ssa_def *prm_cnt = nir_bit_count(b, nir_build_ballot(b, 1, nogs_state->wave_size, nir_build_has_input_primitive_amd(b)));
      nir_store_var(b, vertices_in_wave_var, vtx_cnt, 0x1u);
      nir_store_var(b, primitives_in_wave_var, prm_cnt, 0x1u);
   }
   nir_pop_if(b, if_cull_en);

   /* Update shader arguments.
    *
    * The registers which hold information about the subgroup's
    * vertices and primitives are updated here, so the rest of the shader
    * doesn't need to worry about the culling.
    *
    * These "overwrite" intrinsics must be at top level control flow,
    * otherwise they can mess up the backend (eg. ACO's SSA).
    *
    * TODO:
    * A cleaner solution would be to simply replace all usages of these args
    * with the load of the variables.
    * However, this wouldn't work right now because the backend uses the arguments
    * for purposes not expressed in NIR, eg. VS input loads, etc.
    * This can change if VS input loads and other stuff are lowered to eg. load_buffer_amd.
    */

   if (b->shader->info.stage == MESA_SHADER_VERTEX)
      nir_build_overwrite_vs_arguments_amd(b,
         nir_load_var(b, repacked_arg_vars[0]), nir_load_var(b, repacked_arg_vars[1]));
   else if (b->shader->info.stage == MESA_SHADER_TESS_EVAL)
      nir_build_overwrite_tes_arguments_amd(b,
         nir_load_var(b, repacked_arg_vars[0]), nir_load_var(b, repacked_arg_vars[1]),
         nir_load_var(b, repacked_arg_vars[2]), nir_load_var(b, repacked_arg_vars[3]));
   else
      unreachable("Should be VS or TES.");

   nir_ssa_def *vertices_in_wave = nir_load_var(b, vertices_in_wave_var);
   nir_ssa_def *primitives_in_wave = nir_load_var(b, primitives_in_wave_var);
   nir_build_overwrite_subgroup_num_vertices_and_primitives_amd(b, vertices_in_wave, primitives_in_wave);
}

static bool
can_use_deferred_attribute_culling(nir_shader *shader)
{
   /* When the shader writes memory, it is difficult to guarantee correctness.
    * Future work:
    * - if only write-only SSBOs are used
    * - if we can prove that non-position outputs don't rely on memory stores
    * then may be okay to keep the memory stores in the 1st shader part, and delete them from the 2nd.
    */
   if (shader->info.writes_memory)
      return false;

   /* When the shader relies on the subgroup invocation ID, we'd break it, because the ID changes after the culling.
    * Future work: try to save this to LDS and reload, but it can still be broken in subtle ways.
    */
   if (BITSET_TEST(shader->info.system_values_read, SYSTEM_VALUE_SUBGROUP_INVOCATION))
      return false;

   return true;
}

ac_nir_ngg_config
ac_nir_lower_ngg_nogs(nir_shader *shader,
                      unsigned max_num_es_vertices,
                      unsigned num_vertices_per_primitives,
                      unsigned max_workgroup_size,
                      unsigned wave_size,
                      bool consider_culling,
                      bool consider_passthrough,
                      bool export_prim_id,
                      bool provoking_vtx_last)
{
   nir_function_impl *impl = nir_shader_get_entrypoint(shader);
   assert(impl);
   assert(max_num_es_vertices && max_workgroup_size && wave_size);

   bool can_cull = consider_culling && (num_vertices_per_primitives == 3) &&
                   can_use_deferred_attribute_culling(shader);
   bool passthrough = consider_passthrough && !can_cull &&
                      !(shader->info.stage == MESA_SHADER_VERTEX && export_prim_id);

   nir_variable *position_value_var = nir_local_variable_create(impl, glsl_vec4_type(), "position_value");
   nir_variable *prim_exp_arg_var = nir_local_variable_create(impl, glsl_uint_type(), "prim_exp_arg");
   nir_variable *es_accepted_var = can_cull ? nir_local_variable_create(impl, glsl_bool_type(), "es_accepted") : NULL;
   nir_variable *gs_accepted_var = can_cull ? nir_local_variable_create(impl, glsl_bool_type(), "gs_accepted") : NULL;

   lower_ngg_nogs_state state = {
      .passthrough = passthrough,
      .export_prim_id = export_prim_id,
      .early_prim_export = exec_list_is_singular(&impl->body),
      .num_vertices_per_primitives = num_vertices_per_primitives,
      .provoking_vtx_idx = provoking_vtx_last ? (num_vertices_per_primitives - 1) : 0,
      .position_value_var = position_value_var,
      .prim_exp_arg_var = prim_exp_arg_var,
      .es_accepted_var = es_accepted_var,
      .gs_accepted_var = gs_accepted_var,
      .max_num_waves = DIV_ROUND_UP(max_workgroup_size, wave_size),
      .max_es_num_vertices = max_num_es_vertices,
      .wave_size = wave_size,
   };

   /* We need LDS space when VS needs to export the primitive ID. */
   if (shader->info.stage == MESA_SHADER_VERTEX && export_prim_id)
      state.total_lds_bytes = max_num_es_vertices * 4u;

   /* The shader only needs this much LDS when culling is turned off. */
   unsigned lds_bytes_if_culling_off = state.total_lds_bytes;

   nir_builder builder;
   nir_builder *b = &builder; /* This is to avoid the & */
   nir_builder_init(b, impl);

   if (can_cull) {
      /* We need divergence info for culling shaders. */
      nir_divergence_analysis(shader);
      analyze_shader_before_culling(shader, &state);
      save_reusable_variables(b, &state);
   }

   nir_cf_list extracted;
   nir_cf_extract(&extracted, nir_before_cf_list(&impl->body), nir_after_cf_list(&impl->body));
   b->cursor = nir_before_cf_list(&impl->body);

   if (!can_cull) {
      /* Allocate export space on wave 0 - confirm to the HW that we want to use all possible space */
      nir_if *if_wave_0 = nir_push_if(b, nir_ieq(b, nir_build_load_subgroup_id(b), nir_imm_int(b, 0)));
      {
         nir_ssa_def *vtx_cnt = nir_build_load_workgroup_num_input_vertices_amd(b);
         nir_ssa_def *prim_cnt = nir_build_load_workgroup_num_input_primitives_amd(b);
         nir_build_alloc_vertices_and_primitives_amd(b, vtx_cnt, prim_cnt);
      }
      nir_pop_if(b, if_wave_0);

      /* Take care of early primitive export, otherwise just pack the primitive export argument */
      if (state.early_prim_export)
         emit_ngg_nogs_prim_export(b, &state, NULL);
      else
         nir_store_var(b, prim_exp_arg_var, emit_ngg_nogs_prim_exp_arg(b, &state), 0x1u);
   } else {
      add_deferred_attribute_culling(b, &extracted, &state);
      b->cursor = nir_after_cf_list(&impl->body);

      if (state.early_prim_export)
         emit_ngg_nogs_prim_export(b, &state, nir_load_var(b, state.prim_exp_arg_var));
   }

   nir_intrinsic_instr *export_vertex_instr;

   nir_if *if_es_thread = nir_push_if(b, nir_build_has_input_vertex_amd(b));
   {
      /* Run the actual shader */
      nir_cf_reinsert(&extracted, b->cursor);
      b->cursor = nir_after_cf_list(&if_es_thread->then_list);

      /* Export all vertex attributes (except primitive ID) */
      export_vertex_instr = nir_build_export_vertex_amd(b);

      /* Export primitive ID (in case of early primitive export or TES) */
      if (state.export_prim_id && (state.early_prim_export || shader->info.stage != MESA_SHADER_VERTEX))
         emit_store_ngg_nogs_es_primitive_id(b);
   }
   nir_pop_if(b, if_es_thread);

   /* Take care of late primitive export */
   if (!state.early_prim_export) {
      emit_ngg_nogs_prim_export(b, &state, nir_load_var(b, prim_exp_arg_var));
      if (state.export_prim_id && shader->info.stage == MESA_SHADER_VERTEX) {
         if_es_thread = nir_push_if(b, nir_build_has_input_vertex_amd(b));
         emit_store_ngg_nogs_es_primitive_id(b);
         nir_pop_if(b, if_es_thread);
      }
   }

   if (can_cull) {
      /* Replace uniforms. */
      apply_reusable_variables(b, &state);

      /* Remove the redundant position output. */
      remove_extra_pos_outputs(shader, &state);

      /* After looking at the performance in apps eg. Doom Eternal, and The Witcher 3,
       * it seems that it's best to put the position export always at the end, and
       * then let ACO schedule it up (slightly) only when early prim export is used.
       */
      b->cursor = nir_before_instr(&export_vertex_instr->instr);

      nir_ssa_def *pos_val = nir_load_var(b, state.position_value_var);
      nir_io_semantics io_sem = { .location = VARYING_SLOT_POS, .num_slots = 1 };
      nir_build_store_output(b, pos_val, nir_imm_int(b, 0), .base = VARYING_SLOT_POS, .component = 0, .io_semantics = io_sem, .write_mask = 0xfu);
   }

   nir_metadata_preserve(impl, nir_metadata_none);
   nir_validate_shader(shader, "after emitting NGG VS/TES");

   /* Cleanup */
   nir_opt_dead_write_vars(shader);
   nir_lower_vars_to_ssa(shader);
   nir_remove_dead_variables(shader, nir_var_function_temp, NULL);
   nir_lower_alu_to_scalar(shader, NULL, NULL);
   nir_lower_phis_to_scalar(shader, true);

   if (can_cull) {
      /* It's beneficial to redo these opts after splitting the shader. */
      nir_opt_sink(shader, nir_move_load_input | nir_move_const_undef | nir_move_copies);
      nir_opt_move(shader, nir_move_load_input | nir_move_copies | nir_move_const_undef);
   }

   bool progress;
   do {
      progress = false;
      NIR_PASS(progress, shader, nir_opt_undef);
      NIR_PASS(progress, shader, nir_opt_cse);
      NIR_PASS(progress, shader, nir_opt_dce);
      NIR_PASS(progress, shader, nir_opt_dead_cf);
   } while (progress);

   shader->info.shared_size = state.total_lds_bytes;

   ac_nir_ngg_config ret = {
      .lds_bytes_if_culling_off = lds_bytes_if_culling_off,
      .can_cull = can_cull,
      .passthrough = passthrough,
      .early_prim_export = state.early_prim_export,
      .nggc_inputs_read_by_pos = state.inputs_needed_by_pos,
      .nggc_inputs_read_by_others = state.inputs_needed_by_others,
   };

   return ret;
}

static nir_ssa_def *
ngg_gs_out_vertex_addr(nir_builder *b, nir_ssa_def *out_vtx_idx, lower_ngg_gs_state *s)
{
   unsigned write_stride_2exp = ffs(MAX2(b->shader->info.gs.vertices_out, 1)) - 1;

   /* gs_max_out_vertices = 2^(write_stride_2exp) * some odd number */
   if (write_stride_2exp) {
      nir_ssa_def *row = nir_ushr_imm(b, out_vtx_idx, 5);
      nir_ssa_def *swizzle = nir_iand_imm(b, row, (1u << write_stride_2exp) - 1u);
      out_vtx_idx = nir_ixor(b, out_vtx_idx, swizzle);
   }

   nir_ssa_def *out_vtx_offs = nir_imul_imm(b, out_vtx_idx, s->lds_bytes_per_gs_out_vertex);
   return nir_iadd_imm_nuw(b, out_vtx_offs, s->lds_addr_gs_out_vtx);
}

static nir_ssa_def *
ngg_gs_emit_vertex_addr(nir_builder *b, nir_ssa_def *gs_vtx_idx, lower_ngg_gs_state *s)
{
   nir_ssa_def *tid_in_tg = nir_build_load_local_invocation_index(b);
   nir_ssa_def *gs_out_vtx_base = nir_imul_imm(b, tid_in_tg, b->shader->info.gs.vertices_out);
   nir_ssa_def *out_vtx_idx = nir_iadd_nuw(b, gs_out_vtx_base, gs_vtx_idx);

   return ngg_gs_out_vertex_addr(b, out_vtx_idx, s);
}

static void
ngg_gs_clear_primflags(nir_builder *b, nir_ssa_def *num_vertices, unsigned stream, lower_ngg_gs_state *s)
{
   nir_ssa_def *zero_u8 = nir_imm_zero(b, 1, 8);
   nir_store_var(b, s->current_clear_primflag_idx_var, num_vertices, 0x1u);

   nir_loop *loop = nir_push_loop(b);
   {
      nir_ssa_def *current_clear_primflag_idx = nir_load_var(b, s->current_clear_primflag_idx_var);
      nir_if *if_break = nir_push_if(b, nir_uge(b, current_clear_primflag_idx, nir_imm_int(b, b->shader->info.gs.vertices_out)));
      {
         nir_jump(b, nir_jump_break);
      }
      nir_push_else(b, if_break);
      {
         nir_ssa_def *emit_vtx_addr = ngg_gs_emit_vertex_addr(b, current_clear_primflag_idx, s);
         nir_build_store_shared(b, zero_u8, emit_vtx_addr, .base = s->lds_offs_primflags + stream, .align_mul = 1, .write_mask = 0x1u);
         nir_store_var(b, s->current_clear_primflag_idx_var, nir_iadd_imm_nuw(b, current_clear_primflag_idx, 1), 0x1u);
      }
      nir_pop_if(b, if_break);
   }
   nir_pop_loop(b, loop);
}

static void
ngg_gs_shader_query(nir_builder *b, nir_intrinsic_instr *intrin, lower_ngg_gs_state *s)
{
   nir_if *if_shader_query = nir_push_if(b, nir_build_load_shader_query_enabled_amd(b));
   nir_ssa_def *num_prims_in_wave = NULL;

   /* Calculate the "real" number of emitted primitives from the emitted GS vertices and primitives.
    * GS emits points, line strips or triangle strips.
    * Real primitives are points, lines or triangles.
    */
   if (nir_src_is_const(intrin->src[0]) && nir_src_is_const(intrin->src[1])) {
      unsigned gs_vtx_cnt = nir_src_as_uint(intrin->src[0]);
      unsigned gs_prm_cnt = nir_src_as_uint(intrin->src[1]);
      unsigned total_prm_cnt = gs_vtx_cnt - gs_prm_cnt * (s->num_vertices_per_primitive - 1u);
      nir_ssa_def *num_threads = nir_bit_count(b, nir_build_ballot(b, 1, s->wave_size, nir_imm_bool(b, true)));
      num_prims_in_wave = nir_imul_imm(b, num_threads, total_prm_cnt);
   } else {
      nir_ssa_def *gs_vtx_cnt = intrin->src[0].ssa;
      nir_ssa_def *prm_cnt = intrin->src[1].ssa;
      if (s->num_vertices_per_primitive > 1)
         prm_cnt = nir_iadd_nuw(b, nir_imul_imm(b, prm_cnt, -1u * (s->num_vertices_per_primitive - 1)), gs_vtx_cnt);
      num_prims_in_wave = nir_build_reduce(b, prm_cnt, .reduction_op = nir_op_iadd);
   }

   /* Store the query result to GDS using an atomic add. */
   nir_if *if_first_lane = nir_push_if(b, nir_build_elect(b, 1));
   nir_build_gds_atomic_add_amd(b, 32, num_prims_in_wave, nir_imm_int(b, 0), nir_imm_int(b, 0x100));
   nir_pop_if(b, if_first_lane);

   nir_pop_if(b, if_shader_query);
}

static bool
lower_ngg_gs_store_output(nir_builder *b, nir_intrinsic_instr *intrin, lower_ngg_gs_state *s)
{
   assert(nir_src_is_const(intrin->src[1]));
   b->cursor = nir_before_instr(&intrin->instr);

   unsigned writemask = nir_intrinsic_write_mask(intrin);
   unsigned base = nir_intrinsic_base(intrin);
   unsigned component_offset = nir_intrinsic_component(intrin);
   unsigned base_offset = nir_src_as_uint(intrin->src[1]);
   nir_io_semantics io_sem = nir_intrinsic_io_semantics(intrin);

   assert((base + base_offset) < VARYING_SLOT_MAX);

   nir_ssa_def *store_val = intrin->src[0].ssa;

   for (unsigned comp = 0; comp < 4; ++comp) {
      if (!(writemask & (1 << comp)))
         continue;
      unsigned stream = (io_sem.gs_streams >> (comp * 2)) & 0x3;
      if (!(b->shader->info.gs.active_stream_mask & (1 << stream)))
         continue;

      /* Small bitsize components consume the same amount of space as 32-bit components,
       * but 64-bit ones consume twice as many. (Vulkan spec 15.1.5)
       */
      unsigned num_consumed_components = MIN2(1, DIV_ROUND_UP(store_val->bit_size, 32));
      nir_ssa_def *element = nir_channel(b, store_val, comp);
      if (num_consumed_components > 1)
         element = nir_extract_bits(b, &element, 1, 0, num_consumed_components, 32);

      for (unsigned c = 0; c < num_consumed_components; ++c) {
         unsigned component_index =  (comp * num_consumed_components) + c + component_offset;
         unsigned base_index = base + base_offset + component_index / 4;
         component_index %= 4;

         /* Save output usage info */
         gs_output_component_info *info = &s->output_component_info[base_index][component_index];
         info->bit_size = MAX2(info->bit_size, MIN2(store_val->bit_size, 32));
         info->stream = stream;

         /* Store the current component element */
         nir_ssa_def *component_element = element;
         if (num_consumed_components > 1)
            component_element = nir_channel(b, component_element, c);
         if (component_element->bit_size != 32)
            component_element = nir_u2u32(b, component_element);

         nir_store_var(b, s->output_vars[base_index][component_index], component_element, 0x1u);
      }
   }

   nir_instr_remove(&intrin->instr);
   return true;
}

static bool
lower_ngg_gs_emit_vertex_with_counter(nir_builder *b, nir_intrinsic_instr *intrin, lower_ngg_gs_state *s)
{
   b->cursor = nir_before_instr(&intrin->instr);

   unsigned stream = nir_intrinsic_stream_id(intrin);
   if (!(b->shader->info.gs.active_stream_mask & (1 << stream))) {
      nir_instr_remove(&intrin->instr);
      return true;
   }

   nir_ssa_def *gs_emit_vtx_idx = intrin->src[0].ssa;
   nir_ssa_def *current_vtx_per_prim = intrin->src[1].ssa;
   nir_ssa_def *gs_emit_vtx_addr = ngg_gs_emit_vertex_addr(b, gs_emit_vtx_idx, s);

   for (unsigned slot = 0; slot < VARYING_SLOT_MAX; ++slot) {
      unsigned packed_location = util_bitcount64((b->shader->info.outputs_written & BITFIELD64_MASK(slot)));

      for (unsigned comp = 0; comp < 4; ++comp) {
         gs_output_component_info *info = &s->output_component_info[slot][comp];
         if (info->stream != stream || !info->bit_size)
            continue;

         /* Store the output to LDS */
         nir_ssa_def *out_val = nir_load_var(b, s->output_vars[slot][comp]);
         if (info->bit_size != 32)
            out_val = nir_u2u(b, out_val, info->bit_size);

         nir_build_store_shared(b, out_val, gs_emit_vtx_addr, .base = packed_location * 16 + comp * 4, .align_mul = 4, .write_mask = 0x1u);

         /* Clear the variable that holds the output */
         nir_store_var(b, s->output_vars[slot][comp], nir_ssa_undef(b, 1, 32), 0x1u);
      }
   }

   /* Calculate and store per-vertex primitive flags based on vertex counts:
    * - bit 0: whether this vertex finishes a primitive (a real primitive, not the strip)
    * - bit 1: whether the primitive index is odd (if we are emitting triangle strips, otherwise always 0)
    * - bit 2: always 1 (so that we can use it for determining vertex liveness)
    */

   nir_ssa_def *completes_prim = nir_ige(b, current_vtx_per_prim, nir_imm_int(b, s->num_vertices_per_primitive - 1));
   nir_ssa_def *prim_flag = nir_bcsel(b, completes_prim, nir_imm_int(b, 0b101u), nir_imm_int(b, 0b100u));

   if (s->num_vertices_per_primitive == 3) {
      nir_ssa_def *odd = nir_iand_imm(b, current_vtx_per_prim, 1);
      prim_flag = nir_iadd_nuw(b, prim_flag, nir_ishl(b, odd, nir_imm_int(b, 1)));
   }

   nir_build_store_shared(b, nir_u2u8(b, prim_flag), gs_emit_vtx_addr, .base = s->lds_offs_primflags + stream, .align_mul = 4u, .write_mask = 0x1u);
   nir_instr_remove(&intrin->instr);
   return true;
}

static bool
lower_ngg_gs_end_primitive_with_counter(nir_builder *b, nir_intrinsic_instr *intrin, UNUSED lower_ngg_gs_state *s)
{
   b->cursor = nir_before_instr(&intrin->instr);

   /* These are not needed, we can simply remove them */
   nir_instr_remove(&intrin->instr);
   return true;
}

static bool
lower_ngg_gs_set_vertex_and_primitive_count(nir_builder *b, nir_intrinsic_instr *intrin, lower_ngg_gs_state *s)
{
   b->cursor = nir_before_instr(&intrin->instr);

   unsigned stream = nir_intrinsic_stream_id(intrin);
   if (stream > 0 && !(b->shader->info.gs.active_stream_mask & (1 << stream))) {
      nir_instr_remove(&intrin->instr);
      return true;
   }

   s->found_out_vtxcnt[stream] = true;

   /* Clear the primitive flags of non-emitted vertices */
   if (!nir_src_is_const(intrin->src[0]) || nir_src_as_uint(intrin->src[0]) < b->shader->info.gs.vertices_out)
      ngg_gs_clear_primflags(b, intrin->src[0].ssa, stream, s);

   ngg_gs_shader_query(b, intrin, s);
   nir_instr_remove(&intrin->instr);
   return true;
}

static bool
lower_ngg_gs_intrinsic(nir_builder *b, nir_instr *instr, void *state)
{
   lower_ngg_gs_state *s = (lower_ngg_gs_state *) state;

   if (instr->type != nir_instr_type_intrinsic)
      return false;

   nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);

   if (intrin->intrinsic == nir_intrinsic_store_output)
      return lower_ngg_gs_store_output(b, intrin, s);
   else if (intrin->intrinsic == nir_intrinsic_emit_vertex_with_counter)
      return lower_ngg_gs_emit_vertex_with_counter(b, intrin, s);
   else if (intrin->intrinsic == nir_intrinsic_end_primitive_with_counter)
      return lower_ngg_gs_end_primitive_with_counter(b, intrin, s);
   else if (intrin->intrinsic == nir_intrinsic_set_vertex_and_primitive_count)
      return lower_ngg_gs_set_vertex_and_primitive_count(b, intrin, s);

   return false;
}

static void
lower_ngg_gs_intrinsics(nir_shader *shader, lower_ngg_gs_state *s)
{
   nir_shader_instructions_pass(shader, lower_ngg_gs_intrinsic, nir_metadata_none, s);
}

static void
ngg_gs_export_primitives(nir_builder *b, nir_ssa_def *max_num_out_prims, nir_ssa_def *tid_in_tg,
                         nir_ssa_def *exporter_tid_in_tg, nir_ssa_def *primflag_0,
                         lower_ngg_gs_state *s)
{
   nir_if *if_prim_export_thread = nir_push_if(b, nir_ilt(b, tid_in_tg, max_num_out_prims));

   /* Only bit 0 matters here - set it to 1 when the primitive should be null */
   nir_ssa_def *is_null_prim = nir_ixor(b, primflag_0, nir_imm_int(b, -1u));

   nir_ssa_def *vtx_indices[3] = {0};
   vtx_indices[s->num_vertices_per_primitive - 1] = exporter_tid_in_tg;
   if (s->num_vertices_per_primitive >= 2)
      vtx_indices[s->num_vertices_per_primitive - 2] = nir_isub(b, exporter_tid_in_tg, nir_imm_int(b, 1));
   if (s->num_vertices_per_primitive == 3)
      vtx_indices[s->num_vertices_per_primitive - 3] = nir_isub(b, exporter_tid_in_tg, nir_imm_int(b, 2));

   if (s->num_vertices_per_primitive == 3) {
      /* API GS outputs triangle strips, but NGG HW understands triangles.
       * We already know the triangles due to how we set the primitive flags, but we need to
       * make sure the vertex order is so that the front/back is correct, and the provoking vertex is kept.
       */

      nir_ssa_def *is_odd = nir_ubfe(b, primflag_0, nir_imm_int(b, 1), nir_imm_int(b, 1));
      if (!s->provoking_vertex_last) {
         vtx_indices[1] = nir_iadd(b, vtx_indices[1], is_odd);
         vtx_indices[2] = nir_isub(b, vtx_indices[2], is_odd);
      } else {
         vtx_indices[0] = nir_iadd(b, vtx_indices[0], is_odd);
         vtx_indices[1] = nir_isub(b, vtx_indices[1], is_odd);
      }
   }

   nir_ssa_def *arg = emit_pack_ngg_prim_exp_arg(b, s->num_vertices_per_primitive, vtx_indices, is_null_prim);
   nir_build_export_primitive_amd(b, arg);
   nir_pop_if(b, if_prim_export_thread);
}

static void
ngg_gs_export_vertices(nir_builder *b, nir_ssa_def *max_num_out_vtx, nir_ssa_def *tid_in_tg,
                       nir_ssa_def *out_vtx_lds_addr, lower_ngg_gs_state *s)
{
   nir_if *if_vtx_export_thread = nir_push_if(b, nir_ilt(b, tid_in_tg, max_num_out_vtx));
   nir_ssa_def *exported_out_vtx_lds_addr = out_vtx_lds_addr;

   if (!s->output_compile_time_known) {
      /* Vertex compaction.
       * The current thread will export a vertex that was live in another invocation.
       * Load the index of the vertex that the current thread will have to export.
       */
      nir_ssa_def *exported_vtx_idx = nir_build_load_shared(b, 1, 8, out_vtx_lds_addr, .base = s->lds_offs_primflags + 1, .align_mul = 1u);
      exported_out_vtx_lds_addr = ngg_gs_out_vertex_addr(b, nir_u2u32(b, exported_vtx_idx), s);
   }

   for (unsigned slot = 0; slot < VARYING_SLOT_MAX; ++slot) {
      if (!(b->shader->info.outputs_written & BITFIELD64_BIT(slot)))
         continue;

      unsigned packed_location = util_bitcount64((b->shader->info.outputs_written & BITFIELD64_MASK(slot)));
      nir_io_semantics io_sem = { .location = slot, .num_slots = 1 };

      for (unsigned comp = 0; comp < 4; ++comp) {
         gs_output_component_info *info = &s->output_component_info[slot][comp];
         if (info->stream != 0 || info->bit_size == 0)
            continue;

         nir_ssa_def *load = nir_build_load_shared(b, 1, info->bit_size, exported_out_vtx_lds_addr, .base = packed_location * 16u + comp * 4u, .align_mul = 4u);
         nir_build_store_output(b, load, nir_imm_int(b, 0), .write_mask = 0x1u, .base = slot, .component = comp, .io_semantics = io_sem);
      }
   }

   nir_build_export_vertex_amd(b);
   nir_pop_if(b, if_vtx_export_thread);
}

static void
ngg_gs_setup_vertex_compaction(nir_builder *b, nir_ssa_def *vertex_live, nir_ssa_def *tid_in_tg,
                               nir_ssa_def *exporter_tid_in_tg, lower_ngg_gs_state *s)
{
   assert(vertex_live->bit_size == 1);
   nir_if *if_vertex_live = nir_push_if(b, vertex_live);
   {
      /* Setup the vertex compaction.
       * Save the current thread's id for the thread which will export the current vertex.
       * We reuse stream 1 of the primitive flag of the other thread's vertex for storing this.
       */

      nir_ssa_def *exporter_lds_addr = ngg_gs_out_vertex_addr(b, exporter_tid_in_tg, s);
      nir_ssa_def *tid_in_tg_u8 = nir_u2u8(b, tid_in_tg);
      nir_build_store_shared(b, tid_in_tg_u8, exporter_lds_addr, .base = s->lds_offs_primflags + 1, .align_mul = 1u, .write_mask = 0x1u);
   }
   nir_pop_if(b, if_vertex_live);
}

static nir_ssa_def *
ngg_gs_load_out_vtx_primflag_0(nir_builder *b, nir_ssa_def *tid_in_tg, nir_ssa_def *vtx_lds_addr,
                               nir_ssa_def *max_num_out_vtx, lower_ngg_gs_state *s)
{
   nir_ssa_def *zero = nir_imm_int(b, 0);

   nir_if *if_outvtx_thread = nir_push_if(b, nir_ilt(b, tid_in_tg, max_num_out_vtx));
   nir_ssa_def *primflag_0 = nir_build_load_shared(b, 1, 8, vtx_lds_addr, .base = s->lds_offs_primflags, .align_mul = 4u);
   primflag_0 = nir_u2u32(b, primflag_0);
   nir_pop_if(b, if_outvtx_thread);

   return nir_if_phi(b, primflag_0, zero);
}

static void
ngg_gs_finale(nir_builder *b, lower_ngg_gs_state *s)
{
   nir_ssa_def *tid_in_tg = nir_build_load_local_invocation_index(b);
   nir_ssa_def *max_vtxcnt = nir_build_load_workgroup_num_input_vertices_amd(b);
   nir_ssa_def *max_prmcnt = max_vtxcnt; /* They are currently practically the same; both RADV and RadeonSI do this. */
   nir_ssa_def *out_vtx_lds_addr = ngg_gs_out_vertex_addr(b, tid_in_tg, s);

   if (s->output_compile_time_known) {
      /* When the output is compile-time known, the GS writes all possible vertices and primitives it can.
       * The gs_alloc_req needs to happen on one wave only, otherwise the HW hangs.
       */
      nir_if *if_wave_0 = nir_push_if(b, nir_ieq(b, nir_build_load_subgroup_id(b), nir_imm_zero(b, 1, 32)));
      nir_build_alloc_vertices_and_primitives_amd(b, max_vtxcnt, max_prmcnt);
      nir_pop_if(b, if_wave_0);
   }

   /* Workgroup barrier: wait for all GS threads to finish */
   nir_scoped_barrier(b, .execution_scope=NIR_SCOPE_WORKGROUP, .memory_scope=NIR_SCOPE_WORKGROUP,
                         .memory_semantics=NIR_MEMORY_ACQ_REL, .memory_modes=nir_var_mem_shared);

   nir_ssa_def *out_vtx_primflag_0 = ngg_gs_load_out_vtx_primflag_0(b, tid_in_tg, out_vtx_lds_addr, max_vtxcnt, s);

   if (s->output_compile_time_known) {
      ngg_gs_export_primitives(b, max_vtxcnt, tid_in_tg, tid_in_tg, out_vtx_primflag_0, s);
      ngg_gs_export_vertices(b, max_vtxcnt, tid_in_tg, out_vtx_lds_addr, s);
      return;
   }

   /* When the output vertex count is not known at compile time:
    * There may be gaps between invocations that have live vertices, but NGG hardware
    * requires that the invocations that export vertices are packed (ie. compact).
    * To ensure this, we need to repack invocations that have a live vertex.
    */
   nir_ssa_def *vertex_live = nir_ine(b, out_vtx_primflag_0, nir_imm_zero(b, 1, out_vtx_primflag_0->bit_size));
   wg_repack_result rep = repack_invocations_in_workgroup(b, vertex_live, s->lds_addr_gs_scratch, s->max_num_waves, s->wave_size);

   nir_ssa_def *workgroup_num_vertices = rep.num_repacked_invocations;
   nir_ssa_def *exporter_tid_in_tg = rep.repacked_invocation_index;

   /* When the workgroup emits 0 total vertices, we also must export 0 primitives (otherwise the HW can hang). */
   nir_ssa_def *any_output = nir_ine(b, workgroup_num_vertices, nir_imm_int(b, 0));
   max_prmcnt = nir_bcsel(b, any_output, max_prmcnt, nir_imm_int(b, 0));

   /* Allocate export space. We currently don't compact primitives, just use the maximum number. */
   nir_if *if_wave_0 = nir_push_if(b, nir_ieq(b, nir_build_load_subgroup_id(b), nir_imm_zero(b, 1, 32)));
   nir_build_alloc_vertices_and_primitives_amd(b, workgroup_num_vertices, max_prmcnt);
   nir_pop_if(b, if_wave_0);

   /* Vertex compaction. This makes sure there are no gaps between threads that export vertices. */
   ngg_gs_setup_vertex_compaction(b, vertex_live, tid_in_tg, exporter_tid_in_tg, s);

   /* Workgroup barrier: wait for all LDS stores to finish. */
   nir_scoped_barrier(b, .execution_scope=NIR_SCOPE_WORKGROUP, .memory_scope=NIR_SCOPE_WORKGROUP,
                        .memory_semantics=NIR_MEMORY_ACQ_REL, .memory_modes=nir_var_mem_shared);

   ngg_gs_export_primitives(b, max_prmcnt, tid_in_tg, exporter_tid_in_tg, out_vtx_primflag_0, s);
   ngg_gs_export_vertices(b, workgroup_num_vertices, tid_in_tg, out_vtx_lds_addr, s);
}

void
ac_nir_lower_ngg_gs(nir_shader *shader,
                    unsigned wave_size,
                    unsigned max_workgroup_size,
                    unsigned esgs_ring_lds_bytes,
                    unsigned gs_out_vtx_bytes,
                    unsigned gs_total_out_vtx_bytes,
                    bool provoking_vertex_last)
{
   nir_function_impl *impl = nir_shader_get_entrypoint(shader);
   assert(impl);

   lower_ngg_gs_state state = {
      .max_num_waves = DIV_ROUND_UP(max_workgroup_size, wave_size),
      .wave_size = wave_size,
      .lds_addr_gs_out_vtx = esgs_ring_lds_bytes,
      .lds_addr_gs_scratch = ALIGN(esgs_ring_lds_bytes + gs_total_out_vtx_bytes, 8u /* for the repacking code */),
      .lds_offs_primflags = gs_out_vtx_bytes,
      .lds_bytes_per_gs_out_vertex = gs_out_vtx_bytes + 4u,
      .provoking_vertex_last = provoking_vertex_last,
   };

   unsigned lds_scratch_bytes = DIV_ROUND_UP(state.max_num_waves, 4u) * 4u;
   unsigned total_lds_bytes = state.lds_addr_gs_scratch + lds_scratch_bytes;
   shader->info.shared_size = total_lds_bytes;

   nir_gs_count_vertices_and_primitives(shader, state.const_out_vtxcnt, state.const_out_prmcnt, 4u);
   state.output_compile_time_known = state.const_out_vtxcnt[0] == shader->info.gs.vertices_out &&
                                     state.const_out_prmcnt[0] != -1;

   if (!state.output_compile_time_known)
      state.current_clear_primflag_idx_var = nir_local_variable_create(impl, glsl_uint_type(), "current_clear_primflag_idx");

   if (shader->info.gs.output_primitive == GL_POINTS)
      state.num_vertices_per_primitive = 1;
   else if (shader->info.gs.output_primitive == GL_LINE_STRIP)
      state.num_vertices_per_primitive = 2;
   else if (shader->info.gs.output_primitive == GL_TRIANGLE_STRIP)
      state.num_vertices_per_primitive = 3;
   else
      unreachable("Invalid GS output primitive.");

   /* Extract the full control flow. It is going to be wrapped in an if statement. */
   nir_cf_list extracted;
   nir_cf_extract(&extracted, nir_before_cf_list(&impl->body), nir_after_cf_list(&impl->body));

   nir_builder builder;
   nir_builder *b = &builder; /* This is to avoid the & */
   nir_builder_init(b, impl);
   b->cursor = nir_before_cf_list(&impl->body);

   /* Workgroup barrier: wait for ES threads */
   nir_scoped_barrier(b, .execution_scope=NIR_SCOPE_WORKGROUP, .memory_scope=NIR_SCOPE_WORKGROUP,
                         .memory_semantics=NIR_MEMORY_ACQ_REL, .memory_modes=nir_var_mem_shared);

   /* Wrap the GS control flow. */
   nir_if *if_gs_thread = nir_push_if(b, nir_build_has_input_primitive_amd(b));

   /* Create and initialize output variables */
   for (unsigned slot = 0; slot < VARYING_SLOT_MAX; ++slot) {
      for (unsigned comp = 0; comp < 4; ++comp) {
         state.output_vars[slot][comp] = nir_local_variable_create(impl, glsl_uint_type(), "output");
      }
   }

   nir_cf_reinsert(&extracted, b->cursor);
   b->cursor = nir_after_cf_list(&if_gs_thread->then_list);
   nir_pop_if(b, if_gs_thread);

   /* Lower the GS intrinsics */
   lower_ngg_gs_intrinsics(shader, &state);
   b->cursor = nir_after_cf_list(&impl->body);

   if (!state.found_out_vtxcnt[0]) {
      fprintf(stderr, "Could not find set_vertex_and_primitive_count for stream 0. This would hang your GPU.");
      abort();
   }

   /* Emit the finale sequence */
   ngg_gs_finale(b, &state);
   nir_validate_shader(shader, "after emitting NGG GS");

   /* Cleanup */
   nir_lower_vars_to_ssa(shader);
   nir_remove_dead_variables(shader, nir_var_function_temp, NULL);
   nir_metadata_preserve(impl, nir_metadata_none);
}