mesa/state_tracker/st_draw_hw_select.c

/*
 * Copyright 2022 Advanced Micro Devices, Inc.
 * All Rights Reserved.
 *
 * 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 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 "main/enums.h"
#include "main/context.h"

#include "st_context.h"
#include "st_nir.h"
#include "st_draw.h"

#include "nir.h"
#include "nir_builtin_builder.h"

#include "u_memory.h"

union state_key {
   struct {
      unsigned num_user_clip_planes:4;
      unsigned face_culling_enabled:1;
      unsigned result_offset_from_attribute:1;
      unsigned primitive:4;
   };
   uint32_t u32;
};

enum primitive_state {
   HW_SELECT_PRIM_NONE,
   HW_SELECT_PRIM_POINTS,
   HW_SELECT_PRIM_LINES,
   HW_SELECT_PRIM_TRIANGLES,
   HW_SELECT_PRIM_QUADS,
};

struct geometry_constant {
   float depth_scale;
   float depth_transport;
   uint32_t culling_config;
   uint32_t result_offset;
   float clip_planes[MAX_CLIP_PLANES][4];
};

#define set_uniform_location(var, field, packed)                 \
   do {                                                          \
      unsigned offset = Offset(struct geometry_constant, field); \
      var->data.driver_location = offset >> (packed ? 2 : 4);    \
      var->data.location_frac = (offset >> 2) & 0x3;             \
   } while (0)

static nir_ssa_def *
has_nan_or_inf(nir_builder *b, nir_ssa_def *v)
{
   nir_ssa_def *nan = nir_bany_fnequal4(b, v, v);

   nir_ssa_def *imm = nir_imm_float(b, INFINITY);
   nir_ssa_def *inf = nir_bany(b, nir_feq(b, nir_fabs(b, v), imm));

   return nir_ior(b, nan, inf);
}

static void
return_if_true(nir_builder *b, nir_ssa_def *cond)
{
   nir_if *if_cond = nir_push_if(b, cond);
   nir_jump(b, nir_jump_return);
   nir_pop_if(b, if_cond);
}

static void
get_input_vertices(nir_builder *b, nir_ssa_def **v)
{
   const int num_in_vert = b->shader->info.gs.vertices_in;

   nir_variable *in_pos = nir_variable_create(
      b->shader, nir_var_shader_in, glsl_array_type(glsl_vec4_type(), num_in_vert, 0),
      "gl_Position");
   in_pos->data.location = VARYING_SLOT_POS;

   nir_ssa_def *is_nan_or_inf = NULL;
   for (int i = 0; i < num_in_vert; i++) {
      v[i] = nir_load_array_var_imm(b, in_pos, i);
      nir_ssa_def *r = has_nan_or_inf(b, v[i]);
      is_nan_or_inf = i ? nir_ior(b, is_nan_or_inf, r) : r;
   }
   return_if_true(b, is_nan_or_inf);
}

static void
face_culling(nir_builder *b, nir_ssa_def **v, bool packed)
{
   /* use the z value of the face normal to determine if the face points to us:
    *   Nz = (x1 - x0) * (y2 - y0) - (y1 - y0) * (x2 - x0)
    *
    * it should be in NDC (Normalized Device Coordinate), but now we are in clip
    * space (Vd = Vc / Vc.w), so multiply Nz with w0*w1*w2 to get the clip space
    * value:
    *   det = x0 * (y1 * w2 - y2 * w1) +
    *         x1 * (y2 * w0 - y0 * w2) +
    *         x2 * (y0 * w1 - y1 * w0)
    *
    * we only care about the sign of the det, but also need to count the sign of
    * w0/w1/w2 as a negtive w would change the direction of Nz < 0
    */
   nir_ssa_def *y1w2 = nir_fmul(b, nir_channel(b, v[1], 1), nir_channel(b, v[2], 3));
   nir_ssa_def *y2w1 = nir_fmul(b, nir_channel(b, v[2], 1), nir_channel(b, v[1], 3));
   nir_ssa_def *y2w0 = nir_fmul(b, nir_channel(b, v[2], 1), nir_channel(b, v[0], 3));
   nir_ssa_def *y0w2 = nir_fmul(b, nir_channel(b, v[0], 1), nir_channel(b, v[2], 3));
   nir_ssa_def *y0w1 = nir_fmul(b, nir_channel(b, v[0], 1), nir_channel(b, v[1], 3));
   nir_ssa_def *y1w0 = nir_fmul(b, nir_channel(b, v[1], 1), nir_channel(b, v[0], 3));
   nir_ssa_def *t0 = nir_fmul(b, nir_channel(b, v[0], 0), nir_fsub(b, y1w2, y2w1));
   nir_ssa_def *t1 = nir_fmul(b, nir_channel(b, v[1], 0), nir_fsub(b, y2w0, y0w2));
   nir_ssa_def *t2 = nir_fmul(b, nir_channel(b, v[2], 0), nir_fsub(b, y0w1, y1w0));
   nir_ssa_def *det = nir_fadd(b, nir_fadd(b, t0, t1), t2);

   /* invert det sign once any vertex w < 0 */
   nir_ssa_def *n0 = nir_flt(b, nir_channel(b, v[0], 3), nir_imm_float(b, 0));
   nir_ssa_def *n1 = nir_flt(b, nir_channel(b, v[1], 3), nir_imm_float(b, 0));
   nir_ssa_def *n2 = nir_flt(b, nir_channel(b, v[2], 3), nir_imm_float(b, 0));
   nir_ssa_def *cond = nir_ixor(b, nir_ixor(b, n0, n1), n2);
   det = nir_bcsel(b, cond, nir_fneg(b, det), det);

   nir_variable *culling_config = nir_variable_create(
      b->shader, nir_var_uniform, glsl_uint_type(), "culling_config");
   set_uniform_location(culling_config, culling_config, packed);
   nir_ssa_def *config = nir_i2b(b, nir_load_var(b, culling_config));

   /* det < 0 then z points to camera */
   nir_ssa_def *zero = nir_imm_zero(b, 1, det->bit_size);
   nir_ssa_def *is_zero = nir_feq(b, det, zero);
   nir_ssa_def *is_neg = nir_flt(b, det, zero);
   nir_ssa_def *cull = nir_ixor(b, is_neg, config);
   return_if_true(b, nir_ior(b, is_zero, cull));
}

static void
fast_frustum_culling(nir_builder *b, nir_ssa_def **v)
{
   nir_ssa_def *cull = NULL;

   /* there are six culling planes for the visible volume:
    *   1.  x + w = 0
    *   2. -x + w = 0
    *   3.  y + w = 0
    *   4. -y + w = 0
    *   5.  z + w = 0
    *   6. -z + w = 0
    *
    * if all vertices of the primitive are outside (plane equation <0) of
    * any plane, the primitive must be invisible.
    */
   for (int i = 0; i < 6; i++) {
      nir_ssa_def *outside = NULL;

      for (int j = 0; j < b->shader->info.gs.vertices_in; j++) {
         nir_ssa_def *c = nir_channel(b, v[j], i >> 1);
         if (i & 1)
            c = nir_fneg(b, c);

         nir_ssa_def *r = nir_flt(b, nir_channel(b, v[j], 3), c);
         outside = j ? nir_iand(b, outside, r) : r;
      }

      cull = i ? nir_ior(b, cull, outside) : outside;
   }

   return_if_true(b, cull);
}

static nir_ssa_def *
get_intersection(nir_builder *b, nir_ssa_def *v1, nir_ssa_def *v2,
                 nir_ssa_def *d1, nir_ssa_def *d2)
{
   nir_ssa_def *factor = nir_fdiv(b, d1, nir_fsub(b, d1, d2));
   return nir_fmad(b, nir_fsub(b, v2, v1), factor, v1);
}

#define begin_for_loop(name, max)                                       \
   nir_variable *name##_index =                                         \
      nir_local_variable_create(b->impl, glsl_int_type(), #name "_i");  \
   nir_store_var(b, name##_index, nir_imm_int(b, 0), 1);                \
                                                                        \
   nir_loop *name = nir_push_loop(b);                                   \
   {                                                                    \
      nir_ssa_def *idx = nir_load_var(b, name##_index);                 \
      nir_if *if_in_loop = nir_push_if(b, nir_ilt(b, idx, max));

#define end_for_loop(name)                                              \
         nir_store_var(b, name##_index, nir_iadd_imm(b, idx, 1), 1);    \
      nir_push_else(b, if_in_loop);                                     \
         nir_jump(b, nir_jump_break);                                   \
      nir_pop_if(b, if_in_loop);                                        \
   }                                                                    \
   nir_pop_loop(b, name);

static void
clip_with_plane(nir_builder *b, nir_variable *vert, nir_variable *num_vert,
                int max_vert, nir_ssa_def *plane)
{
   nir_variable *all_clipped = nir_local_variable_create(
      b->impl, glsl_bool_type(), "all_clipped");
   nir_store_var(b, all_clipped, nir_imm_true(b), 1);

   nir_variable *dist = nir_local_variable_create(
      b->impl, glsl_array_type(glsl_float_type(), max_vert, 0), "dist");

   nir_ssa_def *num = nir_load_var(b, num_vert);
   begin_for_loop(dist_loop, num)
   {
      nir_ssa_def *v = nir_load_array_var(b, vert, idx);
      nir_ssa_def *d = nir_fdot(b, v, plane);
      nir_store_array_var(b, dist, idx, d, 1);

      nir_ssa_def *clipped = nir_flt(b, d, nir_imm_float(b, 0));
      nir_store_var(b, all_clipped,
                    nir_iand(b, nir_load_var(b, all_clipped), clipped), 1);
   }
   end_for_loop(dist_loop)

   return_if_true(b, nir_load_var(b, all_clipped));

   /* Use +/0/- to denote the dist[i] sign, which means:
    * +: inside plane
    * -: outside plane
    * 0: just on the plane
    *
    * Some example:
    * ++++: all vertex not clipped
    * ----: all vertex clipped
    * +-++: one vertex clipped, need to insert two vertex at '-', array grow
    * +--+: two vertex clipped, need to insert two vertex at '--', array same
    * +---: three vertex clipped, need to insert two vertex at '---', array trim
    * +-0+: one vertex clipped, need to insert one vertex at '-', array same
    *
    * Plane clip only produce convex polygon, so '-' must be contigous, there's
    * no '+-+-', so one clip plane can only grow array by 1.
    */

   /* when array grow or '-' has been replaced with inserted vertex, save the
    * original vert to be used by following calculation.
    */
   nir_variable *saved =
      nir_local_variable_create(b->impl, glsl_vec4_type(), "saved");

   nir_variable *vert_index =
      nir_local_variable_create(b->impl, glsl_int_type(), "vert_index");
   nir_store_var(b, vert_index, nir_imm_int(b, 0), 1);

   begin_for_loop(vert_loop, num)
   {
      nir_ssa_def *di = nir_load_array_var(b, dist, idx);
      nir_if *if_clipped = nir_push_if(b, nir_flt(b, di, nir_imm_float(b, 0)));
      {
         /* - case, we need to take care of sign change and insert vertex */

         nir_ssa_def *prev = nir_bcsel(b, nir_ieq_imm(b, idx, 0),
                                       nir_iadd_imm(b, num, -1),
                                       nir_iadd_imm(b, idx, -1));
         nir_ssa_def *dp = nir_load_array_var(b, dist, prev);
         nir_if *prev_if = nir_push_if(b, nir_flt(b, nir_imm_float(b, 0), dp));
         {
            /* +- case, replace - with inserted vertex
             * assert(vert_index <= idx), array is sure to not grow here
             * but need to save vert[idx] when vert_index==idx
             */

            nir_ssa_def *vi = nir_load_array_var(b, vert, idx);
            nir_store_var(b, saved, vi, 0xf);

            nir_ssa_def *vp = nir_load_array_var(b, vert, prev);
            nir_ssa_def *iv = get_intersection(b, vp, vi, dp, di);
            nir_ssa_def *index = nir_load_var(b, vert_index);
            nir_store_array_var(b, vert, index, iv, 0xf);

            nir_store_var(b, vert_index, nir_iadd_imm(b, index, 1), 1);
         }
         nir_pop_if(b, prev_if);

         nir_ssa_def *next = nir_bcsel(b, nir_ieq(b, idx, nir_iadd_imm(b, num, -1)),
                                       nir_imm_int(b, 0), nir_iadd_imm(b, idx, 1));
         nir_ssa_def *dn = nir_load_array_var(b, dist, next);
         nir_if *next_if = nir_push_if(b, nir_flt(b, nir_imm_float(b, 0), dn));
         {
            /* -+ case, may grow array:
             *   vert_index > idx: +-+ case, grow array, current vertex in 'saved',
             *     save next + to 'saved', will replace it with inserted vertex.
             *   vert_index <= idx: --+ case, will replace last - with inserted vertex,
             *     no need to save last -, because + case won't use - value.
             */

            nir_ssa_def *index = nir_load_var(b, vert_index);
            nir_ssa_def *vi = nir_bcsel(b, nir_flt(b, idx, index),
                                        nir_load_var(b, saved),
                                        nir_load_array_var(b, vert, idx));
            nir_ssa_def *vn = nir_load_array_var(b, vert, next);
            nir_ssa_def *iv = get_intersection(b, vn, vi, dn, di);

            nir_store_var(b, saved, nir_load_array_var(b, vert, index), 0xf);
            nir_store_array_var(b, vert, index, iv, 0xf);

            nir_store_var(b, vert_index, nir_iadd_imm(b, index, 1), 1);
         }
         nir_pop_if(b, next_if);
      }
      nir_push_else(b, if_clipped);
      {
         /* +/0 case, just keep the vert
          *   vert_index > idx: array grew case, vert[idx] is inserted vertex or prev
          *     +/0 vertex, current vertex is in 'saved', need to save next vertex
          *   vert_index < idx: array trim case
          */

         nir_ssa_def *index = nir_load_var(b, vert_index);
         nir_ssa_def *vi = nir_bcsel(b, nir_flt(b, idx, index),
                                     nir_load_var(b, saved),
                                     nir_load_array_var(b, vert, idx));

         nir_store_var(b, saved, nir_load_array_var(b, vert, index), 0xf);
         nir_store_array_var(b, vert, index, vi, 0xf);

         nir_store_var(b, vert_index, nir_iadd_imm(b, index, 1), 1);
      }
      nir_pop_if(b, if_clipped);
   }
   end_for_loop(vert_loop);

   nir_copy_var(b, num_vert, vert_index);
}

static nir_ssa_def *
get_user_clip_plane(nir_builder *b, int index, bool packed)
{
   char name[16];
   snprintf(name, sizeof(name), "gl_ClipPlane%d", index);
   nir_variable *plane = nir_variable_create(
      b->shader, nir_var_uniform, glsl_vec4_type(), name);

   set_uniform_location(plane, clip_planes[index][0], packed);

   return nir_load_var(b, plane);
}

static void
get_depth_range_transform(nir_builder *b, bool packed, nir_ssa_def **trans)
{
   nir_variable *depth_scale = nir_variable_create(
      b->shader, nir_var_uniform, glsl_float_type(), "depth_scale");
   set_uniform_location(depth_scale, depth_scale, packed);

   nir_variable *depth_transport = nir_variable_create(
      b->shader, nir_var_uniform, glsl_float_type(), "depth_transport");
   set_uniform_location(depth_transport, depth_transport, packed);

   trans[0] = nir_load_var(b, depth_scale);
   trans[1] = nir_load_var(b, depth_transport);
}

static nir_ssa_def *
get_window_space_depth(nir_builder *b, nir_ssa_def *v, nir_ssa_def **trans)
{
   nir_ssa_def *z = nir_channel(b, v, 2);
   nir_ssa_def *w = nir_channel(b, v, 3);

   /* do perspective division, if w==0, xyz must be 0 too (otherwise can't pass
    * the clip test), 0/0=NaN, but we want it to be the nearest point.
    */
   nir_ssa_def *c = nir_feq(b, w, nir_imm_float(b, 0));
   nir_ssa_def *d = nir_bcsel(b, c, nir_imm_float(b, -1), nir_fdiv(b, z, w));

   /* map [-1, 1] to [near, far] set by glDepthRange(near, far) */
   return nir_fmad(b, trans[0], d, trans[1]);
}

static void
update_result_buffer(nir_builder *b, nir_ssa_def *dmin, nir_ssa_def *dmax,
                     bool offset_from_attribute, bool packed)
{
   nir_ssa_def *offset;
   if (offset_from_attribute) {
      nir_variable *in_offset = nir_variable_create(
         b->shader, nir_var_shader_in,
         glsl_array_type(glsl_uint_type(), b->shader->info.gs.vertices_in, 0),
         "result_offset");
      in_offset->data.location = VARYING_SLOT_VAR0;
      offset = nir_load_array_var_imm(b, in_offset, 0);
   } else {
      nir_variable *uni_offset = nir_variable_create(
         b->shader, nir_var_uniform, glsl_uint_type(), "result_offset");
      set_uniform_location(uni_offset, result_offset, packed);
      offset = nir_load_var(b, uni_offset);
   }

   nir_variable_create(b->shader, nir_var_mem_ssbo,
                       glsl_array_type(glsl_uint_type(), 0, 0), "result");
   /* driver_location = 0 (slot 0) */

   nir_ssa_def *ssbo = nir_imm_int(b, 0);
   nir_ssbo_atomic_exchange(b, 32, ssbo, offset, nir_imm_int(b, 1));
   nir_ssbo_atomic_umin(b, 32, ssbo, nir_iadd_imm(b, offset, 4), dmin);
   nir_ssbo_atomic_umax(b, 32, ssbo, nir_iadd_imm(b, offset, 8), dmax);
}

static void
build_point_nir_shader(nir_builder *b, union state_key state, bool packed)
{
   assert(b->shader->info.gs.vertices_in == 1);

   nir_ssa_def *v;
   get_input_vertices(b, &v);

   fast_frustum_culling(b, &v);

   nir_ssa_def *outside = NULL;
   for (int i = 0; i < state.num_user_clip_planes; i++) {
      nir_ssa_def *p = get_user_clip_plane(b, i, packed);
      nir_ssa_def *d = nir_fdot(b, v, p);
      nir_ssa_def *r = nir_flt(b, d, nir_imm_float(b, 0));
      outside = i ? nir_ior(b, outside, r) : r;
   }
   if (outside)
      return_if_true(b, outside);

   nir_ssa_def *trans[2];
   get_depth_range_transform(b, packed, trans);

   nir_ssa_def *depth = get_window_space_depth(b, v, trans);
   nir_ssa_def *fdepth = nir_fmul_imm(b, depth, 4294967295.0);
   nir_ssa_def *idepth = nir_f2uN(b, fdepth, 32);

   update_result_buffer(b, idepth, idepth, state.result_offset_from_attribute, packed);
}

static nir_variable *
create_clip_planes(nir_builder *b, int num_clip_planes, bool packed)
{
   nir_variable *clip_planes = nir_local_variable_create(
      b->impl, glsl_array_type(glsl_vec4_type(), num_clip_planes, 0), "clip_planes");

   nir_ssa_def *unit_clip_planes[6] = {
      nir_imm_vec4(b,  1,  0,  0,  1),
      nir_imm_vec4(b, -1,  0,  0,  1),
      nir_imm_vec4(b,  0,  1,  0,  1),
      nir_imm_vec4(b,  0, -1,  0,  1),
      nir_imm_vec4(b,  0,  0,  1,  1),
      nir_imm_vec4(b,  0,  0, -1,  1),
   };
   for (int i = 0; i < 6; i++)
      nir_store_array_var_imm(b, clip_planes, i, unit_clip_planes[i], 0xf);

   for (int i = 6; i < num_clip_planes; i++) {
      nir_ssa_def *p = get_user_clip_plane(b, i - 6, packed);
      nir_store_array_var_imm(b, clip_planes, i, p, 0xf);
   }

   return clip_planes;
}

static void
build_line_nir_shader(nir_builder *b, union state_key state, bool packed)
{
   assert(b->shader->info.gs.vertices_in == 2);

   nir_ssa_def *v[2];
   get_input_vertices(b, v);

   fast_frustum_culling(b, v);

   nir_variable *vert0 = nir_local_variable_create(b->impl, glsl_vec4_type(), "vert0");
   nir_store_var(b, vert0, v[0], 0xf);

   nir_variable *vert1 = nir_local_variable_create(b->impl, glsl_vec4_type(), "vert1");
   nir_store_var(b, vert1, v[1], 0xf);

   const int num_clip_planes = 6 + state.num_user_clip_planes;
   nir_variable *clip_planes = create_clip_planes(b, num_clip_planes, packed);

   begin_for_loop(clip_loop, nir_imm_int(b, num_clip_planes))
   {
      nir_ssa_def *plane = nir_load_array_var(b, clip_planes, idx);
      nir_ssa_def *v0 = nir_load_var(b, vert0);
      nir_ssa_def *v1 = nir_load_var(b, vert1);
      nir_ssa_def *d0 = nir_fdot(b, v0, plane);
      nir_ssa_def *d1 = nir_fdot(b, v1, plane);
      nir_ssa_def *n0 = nir_flt(b, d0, nir_imm_float(b, 0));
      nir_ssa_def *n1 = nir_flt(b, d1, nir_imm_float(b, 0));

      return_if_true(b, nir_iand(b, n0, n1));

      nir_if *clip_if = nir_push_if(b, nir_ior(b, n0, n1));
      {
         nir_ssa_def *iv = get_intersection(b, v0, v1, d0, d1);
         nir_store_var(b, vert0, nir_bcsel(b, n0, iv, v0), 0xf);
         nir_store_var(b, vert1, nir_bcsel(b, n1, iv, v1), 0xf);
      }
      nir_pop_if(b, clip_if);
   }
   end_for_loop(clip_loop)

   nir_ssa_def *trans[2];
   get_depth_range_transform(b, packed, trans);

   nir_ssa_def *d0 = get_window_space_depth(b, nir_load_var(b, vert0), trans);
   nir_ssa_def *d1 = get_window_space_depth(b, nir_load_var(b, vert1), trans);

   nir_ssa_def *dmin = nir_fmin(b, d0, d1);
   nir_ssa_def *dmax = nir_fmax(b, d0, d1);

   nir_ssa_def *fdmin = nir_fmul_imm(b, dmin, 4294967295.0);
   nir_ssa_def *idmin = nir_f2uN(b, fdmin, 32);

   nir_ssa_def *fdmax = nir_fmul_imm(b, dmax, 4294967295.0);
   nir_ssa_def *idmax = nir_f2uN(b, fdmax, 32);

   update_result_buffer(b, idmin, idmax, state.result_offset_from_attribute, packed);
}

static void
build_planar_primitive_nir_shader(nir_builder *b, union state_key state, bool packed)
{
   const int num_in_vert = b->shader->info.gs.vertices_in;
   assert(num_in_vert == 3 || num_in_vert == 4);

   nir_ssa_def *v[4];
   get_input_vertices(b, v);

   if (state.face_culling_enabled)
      face_culling(b, v, packed);

   /* fast frustum culling, this should filter out most primitives */
   fast_frustum_culling(b, v);

   const int num_clip_planes = 6 + state.num_user_clip_planes;
   const int max_vert = num_in_vert + num_clip_planes;

   /* TODO: could use shared memory (ie. AMD GPU LDS) for this array
    * to reduce register usage.
    */
   nir_variable *vert = nir_local_variable_create(
      b->impl, glsl_array_type(glsl_vec4_type(), max_vert, 0), "vert");
   for (int i = 0; i < num_in_vert; i++)
      nir_store_array_var_imm(b, vert, i, v[i], 0xf);

   nir_variable *num_vert =
      nir_local_variable_create(b->impl, glsl_int_type(), "num_vert");
   nir_store_var(b, num_vert, nir_imm_int(b, num_in_vert), 1);

   nir_variable *clip_planes = create_clip_planes(b, num_clip_planes, packed);

   /* accurate clipping with all clip planes */
   begin_for_loop(clip_loop, nir_imm_int(b, num_clip_planes))
   {
      nir_ssa_def *plane = nir_load_array_var(b, clip_planes, idx);
      clip_with_plane(b, vert, num_vert, max_vert, plane);
   }
   end_for_loop(clip_loop)

   nir_ssa_def *trans[2];
   get_depth_range_transform(b, packed, trans);

   nir_variable *dmin =
      nir_local_variable_create(b->impl, glsl_float_type(), "dmin");
   nir_store_var(b, dmin, nir_imm_float(b, 1), 1);

   nir_variable *dmax =
      nir_local_variable_create(b->impl, glsl_float_type(), "dmax");
   nir_store_var(b, dmax, nir_imm_float(b, 0), 1);

   begin_for_loop(depth_loop, nir_load_var(b, num_vert))
   {
      nir_ssa_def *vtx = nir_load_array_var(b, vert, idx);
      nir_ssa_def *depth = get_window_space_depth(b, vtx, trans);
      nir_store_var(b, dmin, nir_fmin(b, nir_load_var(b, dmin), depth), 1);
      nir_store_var(b, dmax, nir_fmax(b, nir_load_var(b, dmax), depth), 1);
   }
   end_for_loop(depth_loop)

   nir_ssa_def *fdmin = nir_fmul_imm(b, nir_load_var(b, dmin), 4294967295.0);
   nir_ssa_def *idmin = nir_f2uN(b, fdmin, 32);

   nir_ssa_def *fdmax = nir_fmul_imm(b, nir_load_var(b, dmax), 4294967295.0);
   nir_ssa_def *idmax = nir_f2uN(b, fdmax, 32);

   update_result_buffer(b, idmin, idmax, state.result_offset_from_attribute, packed);
}

static void *
hw_select_create_gs(struct st_context *st, union state_key state)
{
   const nir_shader_compiler_options *options =
      st_get_nir_compiler_options(st, MESA_SHADER_GEOMETRY);

   nir_builder b = nir_builder_init_simple_shader(MESA_SHADER_GEOMETRY, options,
                                                  "hw select GS");

   nir_shader *nir = b.shader;
   nir->info.inputs_read = VARYING_BIT_POS;
   nir->num_uniforms = DIV_ROUND_UP(sizeof(struct geometry_constant), (4 * sizeof(float)));
   nir->info.num_ssbos = 1;
   nir->info.gs.output_primitive = SHADER_PRIM_POINTS;
   nir->info.gs.vertices_out = 1;
   nir->info.gs.invocations = 1;
   nir->info.gs.active_stream_mask = 1;

   if (state.result_offset_from_attribute)
      nir->info.inputs_read |= VARYING_BIT_VAR(0);

   bool packed = st->ctx->Const.PackedDriverUniformStorage;

   switch (state.primitive) {
   case HW_SELECT_PRIM_POINTS:
      nir->info.gs.input_primitive = SHADER_PRIM_POINTS;
      nir->info.gs.vertices_in = 1;
      build_point_nir_shader(&b, state, packed);
      break;
   case HW_SELECT_PRIM_LINES:
      nir->info.gs.input_primitive = SHADER_PRIM_LINES;
      nir->info.gs.vertices_in = 2;
      build_line_nir_shader(&b, state, packed);
      break;
   case HW_SELECT_PRIM_TRIANGLES:
      nir->info.gs.input_primitive = SHADER_PRIM_TRIANGLES;
      nir->info.gs.vertices_in = 3;
      build_planar_primitive_nir_shader(&b, state, packed);
      break;
   case HW_SELECT_PRIM_QUADS:
      /* geometry shader has no quad primitive, use lines_adjacency instead */
      nir->info.gs.input_primitive = SHADER_PRIM_LINES_ADJACENCY;
      nir->info.gs.vertices_in = 4;
      build_planar_primitive_nir_shader(&b, state, packed);
      break;
   default:
      unreachable("unexpected primitive");
   }

   nir_lower_returns(nir);

   return st_nir_finish_builtin_shader(st, nir);
}

bool
st_draw_hw_select_prepare_common(struct gl_context *ctx)
{
   struct st_context *st = st_context(ctx);
   if (st->gp || st->tcp || st->tep) {
      fprintf(stderr, "HW GL_SELECT does not support user geometry/tessellation shader\n");
      return false;
   }

   struct geometry_constant consts;

   float n = ctx->ViewportArray[0].Near;
   float f = ctx->ViewportArray[0].Far;
   consts.depth_scale = (f - n) / 2;
   consts.depth_transport = (f + n) / 2;

   /* this field is not used when face culling disabled */
   consts.culling_config =
      (ctx->Polygon.CullFaceMode == GL_BACK) ^
      (ctx->Polygon.FrontFace == GL_CCW);

   /* this field is not used when passing result offset by attribute */
   consts.result_offset = st->ctx->Select.ResultOffset;

   int num_planes = 0;
   u_foreach_bit(i, ctx->Transform.ClipPlanesEnabled) {
      COPY_4V(consts.clip_planes[num_planes], ctx->Transform._ClipUserPlane[i]);
      num_planes++;
   }

   struct pipe_constant_buffer cb;
   cb.buffer = NULL;
   cb.user_buffer = &consts;
   cb.buffer_offset = 0;
   cb.buffer_size = sizeof(consts) - (MAX_CLIP_PLANES - num_planes) * 4 * sizeof(float);

   struct pipe_context *pipe = st->pipe;
   pipe->set_constant_buffer(pipe, PIPE_SHADER_GEOMETRY, 0, false, &cb);

   struct pipe_shader_buffer buffer;
   memset(&buffer, 0, sizeof(buffer));
   buffer.buffer = ctx->Select.Result->buffer;
   buffer.buffer_size = MAX_NAME_STACK_RESULT_NUM * 3 * sizeof(int);

   pipe->set_shader_buffers(pipe, PIPE_SHADER_GEOMETRY, 0, 1, &buffer, 0x1);

   return true;
}

static union state_key
make_state_key(struct gl_context *ctx, int mode)
{
   union state_key state = {0};

   switch (mode) {
   case GL_POINTS:
      state.primitive = HW_SELECT_PRIM_POINTS;
      break;
   case GL_LINES:
   case GL_LINE_STRIP:
   case GL_LINE_LOOP:
      state.primitive = HW_SELECT_PRIM_LINES;
      break;
   case GL_QUADS:
      state.primitive = HW_SELECT_PRIM_QUADS;
      break;
   case GL_TRIANGLES:
   case GL_TRIANGLE_STRIP:
   case GL_TRIANGLE_FAN:
      /* These will be broken into triangles. */
   case GL_QUAD_STRIP:
   case GL_POLYGON:
      state.primitive = HW_SELECT_PRIM_TRIANGLES;
      break;
   default:
      fprintf(stderr, "HW GL_SELECT does not support draw mode %s\n",
              _mesa_enum_to_string(mode));
      return (union state_key){0};
   }

   /* TODO: support gl_ClipDistance/gl_CullDistance, but it costs more regs */
   struct gl_program *vp = ctx->st->vp;
   if (vp->info.clip_distance_array_size || vp->info.cull_distance_array_size) {
      fprintf(stderr, "HW GL_SELECT does not support gl_ClipDistance/gl_CullDistance\n");
      return (union state_key){0};
   }

   state.num_user_clip_planes = util_bitcount(ctx->Transform.ClipPlanesEnabled);

   /* face culling only apply to 2D primitives */
   if (state.primitive == HW_SELECT_PRIM_QUADS ||
       state.primitive == HW_SELECT_PRIM_TRIANGLES)
      state.face_culling_enabled = ctx->Polygon.CullFlag;

   state.result_offset_from_attribute =
      ctx->VertexProgram._VPMode == VP_MODE_FF &&
      (ctx->VertexProgram._VaryingInputs & VERT_BIT_SELECT_RESULT_OFFSET);

   return state;
}

bool
st_draw_hw_select_prepare_mode(struct gl_context *ctx, struct pipe_draw_info *info)
{
   union state_key key = make_state_key(ctx, info->mode);
   if (!key.u32)
      return false;

   struct st_context *st = st_context(ctx);
   if (!st->hw_select_shaders)
      st->hw_select_shaders = _mesa_hash_table_create_u32_keys(NULL);

   struct hash_entry *he = _mesa_hash_table_search(st->hw_select_shaders,
                                                   (void*)(uintptr_t)key.u32);
   void *gs;
   if (!he) {
      gs = hw_select_create_gs(st, key);
      if (!gs)
         return false;

      _mesa_hash_table_insert(st->hw_select_shaders, (void*)(uintptr_t)key.u32, gs);
   } else
      gs = he->data;

   struct cso_context *cso = st->cso_context;
   cso_set_geometry_shader_handle(cso, gs);

   /* Replace draw mode with equivalent one which geometry shader support.
    *
    * New mode consume same vertex buffer structure and produce primitive with
    * same vertices (no need to be same type of primitive, because geometry shader
    * operate on vertives and emit nothing).
    *
    * We can break QUAD and POLYGON to triangles with same shape. But we can't futher
    * break them into single line or point because new primitive need to contain >=3
    * vertices so that it's still handled in 2D (planar) way instead of 1D (line) or
    * 0D (point) way which have different algorithm.
    */
   switch (info->mode) {
   case GL_QUADS:
      info->mode = GL_LINES_ADJACENCY;
      break;
   case GL_QUAD_STRIP:
      info->mode = GL_TRIANGLE_STRIP;
      break;
   case GL_POLYGON:
      info->mode = GL_TRIANGLE_FAN;
      break;
   default:
      break;
   }

   /* Only normal glBegin/End draws pass result offset by attribute to avoid flush
    * vertices when change name stack, so multiple glBegin/End sections before/after
    * name stack calls can be merged to a single draw call. To achieve this We mark
    * name stack result buffer used in glEnd instead of the last draw call.
    *
    * Other case like glDrawArrays and display list replay won't merge draws cross
    * name stack calls, so we just mark name stack result buffer used here.
    */
   if (!key.result_offset_from_attribute)
      ctx->Select.ResultUsed = GL_TRUE;

   return true;
}