src/compiler/nir/nir_lower_tex.c

   1 /*
   2  * Copyright © 2015 Broadcom
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 /*
  25  * This lowering pass supports (as configured via nir_lower_tex_options)
  26  * various texture related conversions:
  27  *   + texture projector lowering: converts the coordinate division for
  28  *     texture projection to be done in ALU instructions instead of
  29  *     asking the texture operation to do so.
  30  *   + lowering RECT: converts the un-normalized RECT texture coordinates
  31  *     to normalized coordinates with txs plus ALU instructions
  32  *   + saturate s/t/r coords: to emulate certain texture clamp/wrap modes,
  33  *     inserts instructions to clamp specified coordinates to [0.0, 1.0].
  34  *     Note that this automatically triggers texture projector lowering if
  35  *     needed, since clamping must happen after projector lowering.
  36  */
  37
  38 #include "nir.h"
  39 #include "nir_builder.h"
  40
  41 static void
  42 project_src(nir_builder *b, nir_tex_instr *tex)
  43 {
  44    /* Find the projector in the srcs list, if present. */
  45    int proj_index = nir_tex_instr_src_index(tex, nir_tex_src_projector);
  46    if (proj_index < 0)
  47       return;
  48
  49    b->cursor = nir_before_instr(&tex->instr);
  50
  51    nir_ssa_def *inv_proj =
  52       nir_frcp(b, nir_ssa_for_src(b, tex->src[proj_index].src, 1));
  53
  54    /* Walk through the sources projecting the arguments. */
  55    for (unsigned i = 0; i < tex->num_srcs; i++) {
  56       switch (tex->src[i].src_type) {
  57       case nir_tex_src_coord:
  58       case nir_tex_src_comparator:
  59          break;
  60       default:
  61          continue;
  62       }
  63       nir_ssa_def *unprojected =
  64          nir_ssa_for_src(b, tex->src[i].src, nir_tex_instr_src_size(tex, i));
  65       nir_ssa_def *projected = nir_fmul(b, unprojected, inv_proj);
  66
  67       /* Array indices don't get projected, so make an new vector with the
  68        * coordinate's array index untouched.
  69        */
  70       if (tex->is_array && tex->src[i].src_type == nir_tex_src_coord) {
  71          switch (tex->coord_components) {
  72          case 4:
  73             projected = nir_vec4(b,
  74                                  nir_channel(b, projected, 0),
  75                                  nir_channel(b, projected, 1),
  76                                  nir_channel(b, projected, 2),
  77                                  nir_channel(b, unprojected, 3));
  78             break;
  79          case 3:
  80             projected = nir_vec3(b,
  81                                  nir_channel(b, projected, 0),
  82                                  nir_channel(b, projected, 1),
  83                                  nir_channel(b, unprojected, 2));
  84             break;
  85          case 2:
  86             projected = nir_vec2(b,
  87                                  nir_channel(b, projected, 0),
  88                                  nir_channel(b, unprojected, 1));
  89             break;
  90          default:
  91             unreachable("bad texture coord count for array");
  92             break;
  93          }
  94       }
  95
  96       nir_instr_rewrite_src(&tex->instr,
  97                             &tex->src[i].src,
  98                             nir_src_for_ssa(projected));
  99    }
 100
 101    nir_tex_instr_remove_src(tex, proj_index);
 102 }
 103
 104 static nir_ssa_def *
 105 get_texture_size(nir_builder *b, nir_tex_instr *tex)
 106 {
 107    b->cursor = nir_before_instr(&tex->instr);
 108
 109    nir_tex_instr *txs;
 110
 111    unsigned num_srcs = 1; /* One for the LOD */
 112    for (unsigned i = 0; i < tex->num_srcs; i++) {
 113       if (tex->src[i].src_type == nir_tex_src_texture_deref ||
 114           tex->src[i].src_type == nir_tex_src_sampler_deref ||
 115           tex->src[i].src_type == nir_tex_src_texture_offset ||
 116           tex->src[i].src_type == nir_tex_src_sampler_offset)
 117          num_srcs++;
 118    }
 119
 120    txs = nir_tex_instr_create(b->shader, num_srcs);
 121    txs->op = nir_texop_txs;
 122    txs->sampler_dim = tex->sampler_dim;
 123    txs->is_array = tex->is_array;
 124    txs->is_shadow = tex->is_shadow;
 125    txs->is_new_style_shadow = tex->is_new_style_shadow;
 126    txs->texture_index = tex->texture_index;
 127    txs->sampler_index = tex->sampler_index;
 128    txs->dest_type = nir_type_int;
 129
 130    unsigned idx = 0;
 131    for (unsigned i = 0; i < tex->num_srcs; i++) {
 132       if (tex->src[i].src_type == nir_tex_src_texture_deref ||
 133           tex->src[i].src_type == nir_tex_src_sampler_deref ||
 134           tex->src[i].src_type == nir_tex_src_texture_offset ||
 135           tex->src[i].src_type == nir_tex_src_sampler_offset) {
 136          nir_src_copy(&txs->src[idx].src, &tex->src[i].src, txs);
 137          txs->src[idx].src_type = tex->src[i].src_type;
 138          idx++;
 139       }
 140    }
 141    /* Add in an LOD because some back-ends require it */
 142    txs->src[idx].src = nir_src_for_ssa(nir_imm_int(b, 0));
 143    txs->src[idx].src_type = nir_tex_src_lod;
 144
 145    nir_ssa_dest_init(&txs->instr, &txs->dest,
 146                      nir_tex_instr_dest_size(txs), 32, NULL);
 147    nir_builder_instr_insert(b, &txs->instr);
 148
 149    return nir_i2f32(b, &txs->dest.ssa);
 150 }
 151
 152 static bool
 153 lower_offset(nir_builder *b, nir_tex_instr *tex)
 154 {
 155    int offset_index = nir_tex_instr_src_index(tex, nir_tex_src_offset);
 156    if (offset_index < 0)
 157       return false;
 158
 159    int coord_index = nir_tex_instr_src_index(tex, nir_tex_src_coord);
 160    assert(coord_index >= 0);
 161
 162    assert(tex->src[offset_index].src.is_ssa);
 163    assert(tex->src[coord_index].src.is_ssa);
 164    nir_ssa_def *offset = tex->src[offset_index].src.ssa;
 165    nir_ssa_def *coord = tex->src[coord_index].src.ssa;
 166
 167    b->cursor = nir_before_instr(&tex->instr);
 168
 169    nir_ssa_def *offset_coord;
 170    if (nir_tex_instr_src_type(tex, coord_index) == nir_type_float) {
 171       if (tex->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
 172          offset_coord = nir_fadd(b, coord, nir_i2f32(b, offset));
 173       } else {
 174          nir_ssa_def *txs = get_texture_size(b, tex);
 175          nir_ssa_def *scale = nir_frcp(b, txs);
 176
 177          offset_coord = nir_fadd(b, coord,
 178                                  nir_fmul(b,
 179                                           nir_i2f32(b, offset),
 180                                           scale));
 181       }
 182    } else {
 183       offset_coord = nir_iadd(b, coord, offset);
 184    }
 185
 186    if (tex->is_array) {
 187       /* The offset is not applied to the array index */
 188       if (tex->coord_components == 2) {
 189          offset_coord = nir_vec2(b, nir_channel(b, offset_coord, 0),
 190                                     nir_channel(b, coord, 1));
 191       } else if (tex->coord_components == 3) {
 192          offset_coord = nir_vec3(b, nir_channel(b, offset_coord, 0),
 193                                     nir_channel(b, offset_coord, 1),
 194                                     nir_channel(b, coord, 2));
 195       } else {
 196          unreachable("Invalid number of components");
 197       }
 198    }
 199
 200    nir_instr_rewrite_src(&tex->instr, &tex->src[coord_index].src,
 201                          nir_src_for_ssa(offset_coord));
 202
 203    nir_tex_instr_remove_src(tex, offset_index);
 204
 205    return true;
 206 }
 207
 208 static void
 209 lower_rect(nir_builder *b, nir_tex_instr *tex)
 210 {
 211    nir_ssa_def *txs = get_texture_size(b, tex);
 212    nir_ssa_def *scale = nir_frcp(b, txs);
 213
 214    /* Walk through the sources normalizing the requested arguments. */
 215    for (unsigned i = 0; i < tex->num_srcs; i++) {
 216       if (tex->src[i].src_type != nir_tex_src_coord)
 217          continue;
 218
 219       nir_ssa_def *coords =
 220          nir_ssa_for_src(b, tex->src[i].src, tex->coord_components);
 221       nir_instr_rewrite_src(&tex->instr,
 222                             &tex->src[i].src,
 223                             nir_src_for_ssa(nir_fmul(b, coords, scale)));
 224    }
 225
 226    tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
 227 }
 228
 229 static nir_ssa_def *
 230 sample_plane(nir_builder *b, nir_tex_instr *tex, int plane)
 231 {
 232    assert(tex->dest.is_ssa);
 233    assert(nir_tex_instr_dest_size(tex) == 4);
 234    assert(nir_alu_type_get_base_type(tex->dest_type) == nir_type_float);
 235    assert(tex->op == nir_texop_tex);
 236    assert(tex->coord_components == 2);
 237
 238    nir_tex_instr *plane_tex =
 239       nir_tex_instr_create(b->shader, tex->num_srcs + 1);
 240    for (unsigned i = 0; i < tex->num_srcs; i++) {
 241       nir_src_copy(&plane_tex->src[i].src, &tex->src[i].src, plane_tex);
 242       plane_tex->src[i].src_type = tex->src[i].src_type;
 243    }
 244    plane_tex->src[tex->num_srcs].src = nir_src_for_ssa(nir_imm_int(b, plane));
 245    plane_tex->src[tex->num_srcs].src_type = nir_tex_src_plane;
 246    plane_tex->op = nir_texop_tex;
 247    plane_tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
 248    plane_tex->dest_type = nir_type_float;
 249    plane_tex->coord_components = 2;
 250
 251    plane_tex->texture_index = tex->texture_index;
 252    plane_tex->sampler_index = tex->sampler_index;
 253
 254    nir_ssa_dest_init(&plane_tex->instr, &plane_tex->dest, 4, 32, NULL);
 255
 256    nir_builder_instr_insert(b, &plane_tex->instr);
 257
 258    return &plane_tex->dest.ssa;
 259 }
 260
 261 static void
 262 convert_yuv_to_rgb(nir_builder *b, nir_tex_instr *tex,
 263                    nir_ssa_def *y, nir_ssa_def *u, nir_ssa_def *v)
 264 {
 265    nir_const_value m[3] = {
 266       { .f32 = { 1.0f,  0.0f,         1.59602678f, 0.0f } },
 267       { .f32 = { 1.0f, -0.39176229f, -0.81296764f, 0.0f } },
 268       { .f32 = { 1.0f,  2.01723214f,  0.0f,        0.0f } }
 269    };
 270
 271    nir_ssa_def *yuv =
 272       nir_vec4(b,
 273                nir_fmul(b, nir_imm_float(b, 1.16438356f),
 274                         nir_fadd(b, y, nir_imm_float(b, -16.0f / 255.0f))),
 275                nir_channel(b, nir_fadd(b, u, nir_imm_float(b, -128.0f / 255.0f)), 0),
 276                nir_channel(b, nir_fadd(b, v, nir_imm_float(b, -128.0f / 255.0f)), 0),
 277                nir_imm_float(b, 0.0));
 278
 279    nir_ssa_def *red = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, m[0]));
 280    nir_ssa_def *green = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, m[1]));
 281    nir_ssa_def *blue = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, m[2]));
 282
 283    nir_ssa_def *result = nir_vec4(b, red, green, blue, nir_imm_float(b, 1.0f));
 284
 285    nir_ssa_def_rewrite_uses(&tex->dest.ssa, nir_src_for_ssa(result));
 286 }
 287
 288 static void
 289 lower_y_uv_external(nir_builder *b, nir_tex_instr *tex)
 290 {
 291    b->cursor = nir_after_instr(&tex->instr);
 292
 293    nir_ssa_def *y = sample_plane(b, tex, 0);
 294    nir_ssa_def *uv = sample_plane(b, tex, 1);
 295
 296    convert_yuv_to_rgb(b, tex,
 297                       nir_channel(b, y, 0),
 298                       nir_channel(b, uv, 0),
 299                       nir_channel(b, uv, 1));
 300 }
 301
 302 static void
 303 lower_y_u_v_external(nir_builder *b, nir_tex_instr *tex)
 304 {
 305    b->cursor = nir_after_instr(&tex->instr);
 306
 307    nir_ssa_def *y = sample_plane(b, tex, 0);
 308    nir_ssa_def *u = sample_plane(b, tex, 1);
 309    nir_ssa_def *v = sample_plane(b, tex, 2);
 310
 311    convert_yuv_to_rgb(b, tex,
 312                       nir_channel(b, y, 0),
 313                       nir_channel(b, u, 0),
 314                       nir_channel(b, v, 0));
 315 }
 316
 317 static void
 318 lower_yx_xuxv_external(nir_builder *b, nir_tex_instr *tex)
 319 {
 320    b->cursor = nir_after_instr(&tex->instr);
 321
 322    nir_ssa_def *y = sample_plane(b, tex, 0);
 323    nir_ssa_def *xuxv = sample_plane(b, tex, 1);
 324
 325    convert_yuv_to_rgb(b, tex,
 326                       nir_channel(b, y, 0),
 327                       nir_channel(b, xuxv, 1),
 328                       nir_channel(b, xuxv, 3));
 329 }
 330
 331 static void
 332 lower_xy_uxvx_external(nir_builder *b, nir_tex_instr *tex)
 333 {
 334   b->cursor = nir_after_instr(&tex->instr);
 335
 336   nir_ssa_def *y = sample_plane(b, tex, 0);
 337   nir_ssa_def *uxvx = sample_plane(b, tex, 1);
 338
 339   convert_yuv_to_rgb(b, tex,
 340                      nir_channel(b, y, 1),
 341                      nir_channel(b, uxvx, 0),
 342                      nir_channel(b, uxvx, 2));
 343 }
 344
 345 /*
 346  * Emits a textureLod operation used to replace an existing
 347  * textureGrad instruction.
 348  */
 349 static void
 350 replace_gradient_with_lod(nir_builder *b, nir_ssa_def *lod, nir_tex_instr *tex)
 351 {
 352    /* We are going to emit a textureLod() with the same parameters except that
 353     * we replace ddx/ddy with lod.
 354     */
 355    int num_srcs = tex->num_srcs - 1;
 356    nir_tex_instr *txl = nir_tex_instr_create(b->shader, num_srcs);
 357
 358    txl->op = nir_texop_txl;
 359    txl->sampler_dim = tex->sampler_dim;
 360    txl->texture_index = tex->texture_index;
 361    txl->dest_type = tex->dest_type;
 362    txl->is_array = tex->is_array;
 363    txl->is_shadow = tex->is_shadow;
 364    txl->is_new_style_shadow = tex->is_new_style_shadow;
 365    txl->sampler_index = tex->sampler_index;
 366    txl->coord_components = tex->coord_components;
 367
 368    nir_ssa_dest_init(&txl->instr, &txl->dest, 4, 32, NULL);
 369
 370    int src_num = 0;
 371    for (int i = 0; i < tex->num_srcs; i++) {
 372       if (tex->src[i].src_type == nir_tex_src_ddx ||
 373           tex->src[i].src_type == nir_tex_src_ddy)
 374          continue;
 375       nir_src_copy(&txl->src[src_num].src, &tex->src[i].src, txl);
 376       txl->src[src_num].src_type = tex->src[i].src_type;
 377       src_num++;
 378    }
 379
 380    txl->src[src_num].src = nir_src_for_ssa(lod);
 381    txl->src[src_num].src_type = nir_tex_src_lod;
 382    src_num++;
 383
 384    assert(src_num == num_srcs);
 385
 386    nir_ssa_dest_init(&txl->instr, &txl->dest,
 387                      tex->dest.ssa.num_components, 32, NULL);
 388    nir_builder_instr_insert(b, &txl->instr);
 389
 390    nir_ssa_def_rewrite_uses(&tex->dest.ssa, nir_src_for_ssa(&txl->dest.ssa));
 391
 392    nir_instr_remove(&tex->instr);
 393 }
 394
 395 static void
 396 lower_gradient_cube_map(nir_builder *b, nir_tex_instr *tex)
 397 {
 398    assert(tex->sampler_dim == GLSL_SAMPLER_DIM_CUBE);
 399    assert(tex->op == nir_texop_txd);
 400    assert(tex->dest.is_ssa);
 401
 402    /* Use textureSize() to get the width and height of LOD 0 */
 403    nir_ssa_def *size = get_texture_size(b, tex);
 404
 405    /* Cubemap texture lookups first generate a texture coordinate normalized
 406     * to [-1, 1] on the appropiate face. The appropiate face is determined
 407     * by which component has largest magnitude and its sign. The texture
 408     * coordinate is the quotient of the remaining texture coordinates against
 409     * that absolute value of the component of largest magnitude. This
 410     * division requires that the computing of the derivative of the texel
 411     * coordinate must use the quotient rule. The high level GLSL code is as
 412     * follows:
 413     *
 414     * Step 1: selection
 415     *
 416     * vec3 abs_p, Q, dQdx, dQdy;
 417     * abs_p = abs(ir->coordinate);
 418     * if (abs_p.x >= max(abs_p.y, abs_p.z)) {
 419     *    Q = ir->coordinate.yzx;
 420     *    dQdx = ir->lod_info.grad.dPdx.yzx;
 421     *    dQdy = ir->lod_info.grad.dPdy.yzx;
 422     * }
 423     * if (abs_p.y >= max(abs_p.x, abs_p.z)) {
 424     *    Q = ir->coordinate.xzy;
 425     *    dQdx = ir->lod_info.grad.dPdx.xzy;
 426     *    dQdy = ir->lod_info.grad.dPdy.xzy;
 427     * }
 428     * if (abs_p.z >= max(abs_p.x, abs_p.y)) {
 429     *    Q = ir->coordinate;
 430     *    dQdx = ir->lod_info.grad.dPdx;
 431     *    dQdy = ir->lod_info.grad.dPdy;
 432     * }
 433     *
 434     * Step 2: use quotient rule to compute derivative. The normalized to
 435     * [-1, 1] texel coordinate is given by Q.xy / (sign(Q.z) * Q.z). We are
 436     * only concerned with the magnitudes of the derivatives whose values are
 437     * not affected by the sign. We drop the sign from the computation.
 438     *
 439     * vec2 dx, dy;
 440     * float recip;
 441     *
 442     * recip = 1.0 / Q.z;
 443     * dx = recip * ( dQdx.xy - Q.xy * (dQdx.z * recip) );
 444     * dy = recip * ( dQdy.xy - Q.xy * (dQdy.z * recip) );
 445     *
 446     * Step 3: compute LOD. At this point we have the derivatives of the
 447     * texture coordinates normalized to [-1,1]. We take the LOD to be
 448     *  result = log2(max(sqrt(dot(dx, dx)), sqrt(dy, dy)) * 0.5 * L)
 449     *         = -1.0 + log2(max(sqrt(dot(dx, dx)), sqrt(dy, dy)) * L)
 450     *         = -1.0 + log2(sqrt(max(dot(dx, dx), dot(dy,dy))) * L)
 451     *         = -1.0 + log2(sqrt(L * L * max(dot(dx, dx), dot(dy,dy))))
 452     *         = -1.0 + 0.5 * log2(L * L * max(dot(dx, dx), dot(dy,dy)))
 453     * where L is the dimension of the cubemap. The code is:
 454     *
 455     * float M, result;
 456     * M = max(dot(dx, dx), dot(dy, dy));
 457     * L = textureSize(sampler, 0).x;
 458     * result = -1.0 + 0.5 * log2(L * L * M);
 459     */
 460
 461    /* coordinate */
 462    nir_ssa_def *p =
 463       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_coord)].src.ssa;
 464
 465    /* unmodified dPdx, dPdy values */
 466    nir_ssa_def *dPdx =
 467       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_ddx)].src.ssa;
 468    nir_ssa_def *dPdy =
 469       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_ddy)].src.ssa;
 470
 471    nir_ssa_def *abs_p = nir_fabs(b, p);
 472    nir_ssa_def *abs_p_x = nir_channel(b, abs_p, 0);
 473    nir_ssa_def *abs_p_y = nir_channel(b, abs_p, 1);
 474    nir_ssa_def *abs_p_z = nir_channel(b, abs_p, 2);
 475
 476    /* 1. compute selector */
 477    nir_ssa_def *Q, *dQdx, *dQdy;
 478
 479    nir_ssa_def *cond_z = nir_fge(b, abs_p_z, nir_fmax(b, abs_p_x, abs_p_y));
 480    nir_ssa_def *cond_y = nir_fge(b, abs_p_y, nir_fmax(b, abs_p_x, abs_p_z));
 481
 482    unsigned yzx[4] = { 1, 2, 0, 0 };
 483    unsigned xzy[4] = { 0, 2, 1, 0 };
 484
 485    Q = nir_bcsel(b, cond_z,
 486                  p,
 487                  nir_bcsel(b, cond_y,
 488                            nir_swizzle(b, p, xzy, 3, false),
 489                            nir_swizzle(b, p, yzx, 3, false)));
 490
 491    dQdx = nir_bcsel(b, cond_z,
 492                     dPdx,
 493                     nir_bcsel(b, cond_y,
 494                               nir_swizzle(b, dPdx, xzy, 3, false),
 495                               nir_swizzle(b, dPdx, yzx, 3, false)));
 496
 497    dQdy = nir_bcsel(b, cond_z,
 498                     dPdy,
 499                     nir_bcsel(b, cond_y,
 500                               nir_swizzle(b, dPdy, xzy, 3, false),
 501                               nir_swizzle(b, dPdy, yzx, 3, false)));
 502
 503    /* 2. quotient rule */
 504
 505    /* tmp = Q.xy * recip;
 506     * dx = recip * ( dQdx.xy - (tmp * dQdx.z) );
 507     * dy = recip * ( dQdy.xy - (tmp * dQdy.z) );
 508     */
 509    nir_ssa_def *rcp_Q_z = nir_frcp(b, nir_channel(b, Q, 2));
 510
 511    unsigned xy[4] = { 0, 1, 0, 0 };
 512    nir_ssa_def *Q_xy = nir_swizzle(b, Q, xy, 2, false);
 513    nir_ssa_def *tmp = nir_fmul(b, Q_xy, rcp_Q_z);
 514
 515    nir_ssa_def *dQdx_xy = nir_swizzle(b, dQdx, xy, 2, false);
 516    nir_ssa_def *dQdx_z = nir_channel(b, dQdx, 2);
 517    nir_ssa_def *dx =
 518       nir_fmul(b, rcp_Q_z, nir_fsub(b, dQdx_xy, nir_fmul(b, tmp, dQdx_z)));
 519
 520    nir_ssa_def *dQdy_xy = nir_swizzle(b, dQdy, xy, 2, false);
 521    nir_ssa_def *dQdy_z = nir_channel(b, dQdy, 2);
 522    nir_ssa_def *dy =
 523       nir_fmul(b, rcp_Q_z, nir_fsub(b, dQdy_xy, nir_fmul(b, tmp, dQdy_z)));
 524
 525    /* M = max(dot(dx, dx), dot(dy, dy)); */
 526    nir_ssa_def *M = nir_fmax(b, nir_fdot(b, dx, dx), nir_fdot(b, dy, dy));
 527
 528    /* size has textureSize() of LOD 0 */
 529    nir_ssa_def *L = nir_channel(b, size, 0);
 530
 531    /* lod = -1.0 + 0.5 * log2(L * L * M); */
 532    nir_ssa_def *lod =
 533       nir_fadd(b,
 534                nir_imm_float(b, -1.0f),
 535                nir_fmul(b,
 536                         nir_imm_float(b, 0.5f),
 537                         nir_flog2(b, nir_fmul(b, L, nir_fmul(b, L, M)))));
 538
 539    /* 3. Replace the gradient instruction with an equivalent lod instruction */
 540    replace_gradient_with_lod(b, lod, tex);
 541 }
 542
 543 static void
 544 lower_gradient(nir_builder *b, nir_tex_instr *tex)
 545 {
 546    assert(tex->sampler_dim != GLSL_SAMPLER_DIM_CUBE);
 547    assert(tex->op == nir_texop_txd);
 548    assert(tex->dest.is_ssa);
 549
 550    /* Use textureSize() to get the width and height of LOD 0 */
 551    unsigned component_mask;
 552    switch (tex->sampler_dim) {
 553    case GLSL_SAMPLER_DIM_3D:
 554       component_mask = 7;
 555       break;
 556    case GLSL_SAMPLER_DIM_1D:
 557       component_mask = 1;
 558       break;
 559    default:
 560       component_mask = 3;
 561       break;
 562    }
 563
 564    nir_ssa_def *size =
 565       nir_channels(b, get_texture_size(b, tex), component_mask);
 566
 567    /* Scale the gradients by width and height.  Effectively, the incoming
 568     * gradients are s'(x,y), t'(x,y), and r'(x,y) from equation 3.19 in the
 569     * GL 3.0 spec; we want u'(x,y), which is w_t * s'(x,y).
 570     */
 571    nir_ssa_def *ddx =
 572       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_ddx)].src.ssa;
 573    nir_ssa_def *ddy =
 574       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_ddy)].src.ssa;
 575
 576    nir_ssa_def *dPdx = nir_fmul(b, ddx, size);
 577    nir_ssa_def *dPdy = nir_fmul(b, ddy, size);
 578
 579    nir_ssa_def *rho;
 580    if (dPdx->num_components == 1) {
 581       rho = nir_fmax(b, nir_fabs(b, dPdx), nir_fabs(b, dPdy));
 582    } else {
 583       rho = nir_fmax(b,
 584                      nir_fsqrt(b, nir_fdot(b, dPdx, dPdx)),
 585                      nir_fsqrt(b, nir_fdot(b, dPdy, dPdy)));
 586    }
 587
 588    /* lod = log2(rho).  We're ignoring GL state biases for now. */
 589    nir_ssa_def *lod = nir_flog2(b, rho);
 590
 591    /* Replace the gradient instruction with an equivalent lod instruction */
 592    replace_gradient_with_lod(b, lod, tex);
 593 }
 594
 595 static void
 596 saturate_src(nir_builder *b, nir_tex_instr *tex, unsigned sat_mask)
 597 {
 598    b->cursor = nir_before_instr(&tex->instr);
 599
 600    /* Walk through the sources saturating the requested arguments. */
 601    for (unsigned i = 0; i < tex->num_srcs; i++) {
 602       if (tex->src[i].src_type != nir_tex_src_coord)
 603          continue;
 604
 605       nir_ssa_def *src =
 606          nir_ssa_for_src(b, tex->src[i].src, tex->coord_components);
 607
 608       /* split src into components: */
 609       nir_ssa_def *comp[4];
 610
 611       assume(tex->coord_components >= 1);
 612
 613       for (unsigned j = 0; j < tex->coord_components; j++)
 614          comp[j] = nir_channel(b, src, j);
 615
 616       /* clamp requested components, array index does not get clamped: */
 617       unsigned ncomp = tex->coord_components;
 618       if (tex->is_array)
 619          ncomp--;
 620
 621       for (unsigned j = 0; j < ncomp; j++) {
 622          if ((1 << j) & sat_mask) {
 623             if (tex->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
 624                /* non-normalized texture coords, so clamp to texture
 625                 * size rather than [0.0, 1.0]
 626                 */
 627                nir_ssa_def *txs = get_texture_size(b, tex);
 628                comp[j] = nir_fmax(b, comp[j], nir_imm_float(b, 0.0));
 629                comp[j] = nir_fmin(b, comp[j], nir_channel(b, txs, j));
 630             } else {
 631                comp[j] = nir_fsat(b, comp[j]);
 632             }
 633          }
 634       }
 635
 636       /* and move the result back into a single vecN: */
 637       src = nir_vec(b, comp, tex->coord_components);
 638
 639       nir_instr_rewrite_src(&tex->instr,
 640                             &tex->src[i].src,
 641                             nir_src_for_ssa(src));
 642    }
 643 }
 644
 645 static nir_ssa_def *
 646 get_zero_or_one(nir_builder *b, nir_alu_type type, uint8_t swizzle_val)
 647 {
 648    nir_const_value v;
 649
 650    memset(&v, 0, sizeof(v));
 651
 652    if (swizzle_val == 4) {
 653       v.u32[0] = v.u32[1] = v.u32[2] = v.u32[3] = 0;
 654    } else {
 655       assert(swizzle_val == 5);
 656       if (type == nir_type_float)
 657          v.f32[0] = v.f32[1] = v.f32[2] = v.f32[3] = 1.0;
 658       else
 659          v.u32[0] = v.u32[1] = v.u32[2] = v.u32[3] = 1;
 660    }
 661
 662    return nir_build_imm(b, 4, 32, v);
 663 }
 664
 665 static void
 666 swizzle_result(nir_builder *b, nir_tex_instr *tex, const uint8_t swizzle[4])
 667 {
 668    assert(tex->dest.is_ssa);
 669
 670    b->cursor = nir_after_instr(&tex->instr);
 671
 672    nir_ssa_def *swizzled;
 673    if (tex->op == nir_texop_tg4) {
 674       if (swizzle[tex->component] < 4) {
 675          /* This one's easy */
 676          tex->component = swizzle[tex->component];
 677          return;
 678       } else {
 679          swizzled = get_zero_or_one(b, tex->dest_type, swizzle[tex->component]);
 680       }
 681    } else {
 682       assert(nir_tex_instr_dest_size(tex) == 4);
 683       if (swizzle[0] < 4 && swizzle[1] < 4 &&
 684           swizzle[2] < 4 && swizzle[3] < 4) {
 685          unsigned swiz[4] = { swizzle[0], swizzle[1], swizzle[2], swizzle[3] };
 686          /* We have no 0s or 1s, just emit a swizzling MOV */
 687          swizzled = nir_swizzle(b, &tex->dest.ssa, swiz, 4, false);
 688       } else {
 689          nir_ssa_def *srcs[4];
 690          for (unsigned i = 0; i < 4; i++) {
 691             if (swizzle[i] < 4) {
 692                srcs[i] = nir_channel(b, &tex->dest.ssa, swizzle[i]);
 693             } else {
 694                srcs[i] = get_zero_or_one(b, tex->dest_type, swizzle[i]);
 695             }
 696          }
 697          swizzled = nir_vec(b, srcs, 4);
 698       }
 699    }
 700
 701    nir_ssa_def_rewrite_uses_after(&tex->dest.ssa, nir_src_for_ssa(swizzled),
 702                                   swizzled->parent_instr);
 703 }
 704
 705 static void
 706 linearize_srgb_result(nir_builder *b, nir_tex_instr *tex)
 707 {
 708    assert(tex->dest.is_ssa);
 709    assert(nir_tex_instr_dest_size(tex) == 4);
 710    assert(nir_alu_type_get_base_type(tex->dest_type) == nir_type_float);
 711
 712    b->cursor = nir_after_instr(&tex->instr);
 713
 714    static const unsigned swiz[4] = {0, 1, 2, 0};
 715    nir_ssa_def *comp = nir_swizzle(b, &tex->dest.ssa, swiz, 3, true);
 716
 717    /* Formula is:
 718     *    (comp <= 0.04045) ?
 719     *          (comp / 12.92) :
 720     *          pow((comp + 0.055) / 1.055, 2.4)
 721     */
 722    nir_ssa_def *low  = nir_fmul(b, comp, nir_imm_float(b, 1.0 / 12.92));
 723    nir_ssa_def *high = nir_fpow(b,
 724                                 nir_fmul(b,
 725                                          nir_fadd(b,
 726                                                   comp,
 727                                                   nir_imm_float(b, 0.055)),
 728                                          nir_imm_float(b, 1.0 / 1.055)),
 729                                 nir_imm_float(b, 2.4));
 730    nir_ssa_def *cond = nir_fge(b, nir_imm_float(b, 0.04045), comp);
 731    nir_ssa_def *rgb  = nir_bcsel(b, cond, low, high);
 732
 733    /* alpha is untouched: */
 734    nir_ssa_def *result = nir_vec4(b,
 735                                   nir_channel(b, rgb, 0),
 736                                   nir_channel(b, rgb, 1),
 737                                   nir_channel(b, rgb, 2),
 738                                   nir_channel(b, &tex->dest.ssa, 3));
 739
 740    nir_ssa_def_rewrite_uses_after(&tex->dest.ssa, nir_src_for_ssa(result),
 741                                   result->parent_instr);
 742 }
 743
 744 static bool
 745 nir_lower_tex_block(nir_block *block, nir_builder *b,
 746                     const nir_lower_tex_options *options)
 747 {
 748    bool progress = false;
 749
 750    nir_foreach_instr_safe(instr, block) {
 751       if (instr->type != nir_instr_type_tex)
 752          continue;
 753
 754       nir_tex_instr *tex = nir_instr_as_tex(instr);
 755       bool lower_txp = !!(options->lower_txp & (1 << tex->sampler_dim));
 756
 757       /* mask of src coords to saturate (clamp): */
 758       unsigned sat_mask = 0;
 759
 760       if ((1 << tex->sampler_index) & options->saturate_r)
 761          sat_mask |= (1 << 2);    /* .z */
 762       if ((1 << tex->sampler_index) & options->saturate_t)
 763          sat_mask |= (1 << 1);    /* .y */
 764       if ((1 << tex->sampler_index) & options->saturate_s)
 765          sat_mask |= (1 << 0);    /* .x */
 766
 767       /* If we are clamping any coords, we must lower projector first
 768        * as clamping happens *after* projection:
 769        */
 770       if (lower_txp || sat_mask) {
 771          project_src(b, tex);
 772          progress = true;
 773       }
 774
 775       if ((tex->op == nir_texop_txf && options->lower_txf_offset) ||
 776           (sat_mask && nir_tex_instr_src_index(tex, nir_tex_src_coord) >= 0) ||
 777           (tex->sampler_dim == GLSL_SAMPLER_DIM_RECT &&
 778            options->lower_rect_offset)) {
 779          progress = lower_offset(b, tex) || progress;
 780       }
 781
 782       if ((tex->sampler_dim == GLSL_SAMPLER_DIM_RECT) && options->lower_rect) {
 783          lower_rect(b, tex);
 784          progress = true;
 785       }
 786
 787       if ((1 << tex->texture_index) & options->lower_y_uv_external) {
 788          lower_y_uv_external(b, tex);
 789          progress = true;
 790       }
 791
 792       if ((1 << tex->texture_index) & options->lower_y_u_v_external) {
 793          lower_y_u_v_external(b, tex);
 794          progress = true;
 795       }
 796
 797       if ((1 << tex->texture_index) & options->lower_yx_xuxv_external) {
 798          lower_yx_xuxv_external(b, tex);
 799          progress = true;
 800       }
 801
 802       if ((1 << tex->texture_index) & options->lower_xy_uxvx_external) {
 803          lower_xy_uxvx_external(b, tex);
 804          progress = true;
 805       }
 806
 807       if (sat_mask) {
 808          saturate_src(b, tex, sat_mask);
 809          progress = true;
 810       }
 811
 812       if (((1 << tex->texture_index) & options->swizzle_result) &&
 813           !nir_tex_instr_is_query(tex) &&
 814           !(tex->is_shadow && tex->is_new_style_shadow)) {
 815          swizzle_result(b, tex, options->swizzles[tex->texture_index]);
 816          progress = true;
 817       }
 818
 819       /* should be after swizzle so we know which channels are rgb: */
 820       if (((1 << tex->texture_index) & options->lower_srgb) &&
 821           !nir_tex_instr_is_query(tex) && !tex->is_shadow) {
 822          linearize_srgb_result(b, tex);
 823          progress = true;
 824       }
 825
 826       if (tex->op == nir_texop_txd &&
 827           tex->sampler_dim == GLSL_SAMPLER_DIM_CUBE &&
 828           (options->lower_txd ||
 829            options->lower_txd_cube_map ||
 830            (tex->is_shadow && options->lower_txd_shadow))) {
 831          lower_gradient_cube_map(b, tex);
 832          progress = true;
 833          continue;
 834       }
 835
 836       if (tex->op == nir_texop_txd &&
 837           (options->lower_txd ||
 838            (options->lower_txd_shadow &&
 839             tex->is_shadow && tex->sampler_dim != GLSL_SAMPLER_DIM_CUBE))) {
 840          lower_gradient(b, tex);
 841          progress = true;
 842          continue;
 843       }
 844
 845       /* TXF, TXS and TXL require a LOD but not everything we implement using those
 846        * three opcodes provides one.  Provide a default LOD of 0.
 847        */
 848       if ((nir_tex_instr_src_index(tex, nir_tex_src_lod) == -1) &&
 849           (tex->op == nir_texop_txf || tex->op == nir_texop_txs ||
 850            tex->op == nir_texop_txl || tex->op == nir_texop_query_levels ||
 851            (tex->op == nir_texop_tex &&
 852             b->shader->info.stage != MESA_SHADER_FRAGMENT))) {
 853          b->cursor = nir_before_instr(&tex->instr);
 854          nir_tex_instr_add_src(tex, nir_tex_src_lod, nir_src_for_ssa(nir_imm_int(b, 0)));
 855          progress = true;
 856          continue;
 857       }
 858    }
 859
 860    return progress;
 861 }
 862
 863 static bool
 864 nir_lower_tex_impl(nir_function_impl *impl,
 865                    const nir_lower_tex_options *options)
 866 {
 867    bool progress = false;
 868    nir_builder builder;
 869    nir_builder_init(&builder, impl);
 870
 871    nir_foreach_block(block, impl) {
 872       progress |= nir_lower_tex_block(block, &builder, options);
 873    }
 874
 875    nir_metadata_preserve(impl, nir_metadata_block_index |
 876                                nir_metadata_dominance);
 877    return progress;
 878 }
 879
 880 bool
 881 nir_lower_tex(nir_shader *shader, const nir_lower_tex_options *options)
 882 {
 883    bool progress = false;
 884
 885    nir_assert_unlowered_derefs(shader, nir_lower_texture_derefs);
 886
 887    nir_foreach_function(function, shader) {
 888       if (function->impl)
 889          progress |= nir_lower_tex_impl(function->impl, options);
 890    }
 891
 892    return progress;
 893 }