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 bool
 105 lower_offset(nir_builder *b, nir_tex_instr *tex)
 106 {
 107    int offset_index = nir_tex_instr_src_index(tex, nir_tex_src_offset);
 108    if (offset_index < 0)
 109       return false;
 110
 111    int coord_index = nir_tex_instr_src_index(tex, nir_tex_src_coord);
 112    assert(coord_index >= 0);
 113
 114    assert(tex->src[offset_index].src.is_ssa);
 115    assert(tex->src[coord_index].src.is_ssa);
 116    nir_ssa_def *offset = tex->src[offset_index].src.ssa;
 117    nir_ssa_def *coord = tex->src[coord_index].src.ssa;
 118
 119    b->cursor = nir_before_instr(&tex->instr);
 120
 121    nir_ssa_def *offset_coord;
 122    if (nir_tex_instr_src_type(tex, coord_index) == nir_type_float) {
 123       assert(tex->sampler_dim == GLSL_SAMPLER_DIM_RECT);
 124       offset_coord = nir_fadd(b, coord, nir_i2f(b, offset));
 125    } else {
 126       offset_coord = nir_iadd(b, coord, offset);
 127    }
 128
 129    if (tex->is_array) {
 130       /* The offset is not applied to the array index */
 131       if (tex->coord_components == 2) {
 132          offset_coord = nir_vec2(b, nir_channel(b, offset_coord, 0),
 133                                     nir_channel(b, coord, 1));
 134       } else if (tex->coord_components == 3) {
 135          offset_coord = nir_vec3(b, nir_channel(b, offset_coord, 0),
 136                                     nir_channel(b, offset_coord, 1),
 137                                     nir_channel(b, coord, 2));
 138       } else {
 139          unreachable("Invalid number of components");
 140       }
 141    }
 142
 143    nir_instr_rewrite_src(&tex->instr, &tex->src[coord_index].src,
 144                          nir_src_for_ssa(offset_coord));
 145
 146    nir_tex_instr_remove_src(tex, offset_index);
 147
 148    return true;
 149 }
 150
 151
 152 static nir_ssa_def *
 153 get_texture_size(nir_builder *b, nir_tex_instr *tex)
 154 {
 155    b->cursor = nir_before_instr(&tex->instr);
 156
 157    nir_tex_instr *txs;
 158
 159    txs = nir_tex_instr_create(b->shader, 1);
 160    txs->op = nir_texop_txs;
 161    txs->sampler_dim = tex->sampler_dim;
 162    txs->is_array = tex->is_array;
 163    txs->is_shadow = tex->is_shadow;
 164    txs->is_new_style_shadow = tex->is_new_style_shadow;
 165    txs->texture_index = tex->texture_index;
 166    txs->texture = (nir_deref_var *)
 167       nir_copy_deref(txs, &tex->texture->deref);
 168    txs->sampler_index = tex->sampler_index;
 169    txs->sampler = (nir_deref_var *)
 170       nir_copy_deref(txs, &tex->sampler->deref);
 171    txs->dest_type = nir_type_int;
 172
 173    /* only single src, the lod: */
 174    txs->src[0].src = nir_src_for_ssa(nir_imm_int(b, 0));
 175    txs->src[0].src_type = nir_tex_src_lod;
 176
 177    nir_ssa_dest_init(&txs->instr, &txs->dest, tex->coord_components, 32, NULL);
 178    nir_builder_instr_insert(b, &txs->instr);
 179
 180    return nir_i2f(b, &txs->dest.ssa);
 181 }
 182
 183 static void
 184 lower_rect(nir_builder *b, nir_tex_instr *tex)
 185 {
 186    nir_ssa_def *txs = get_texture_size(b, tex);
 187    nir_ssa_def *scale = nir_frcp(b, txs);
 188
 189    /* Walk through the sources normalizing the requested arguments. */
 190    for (unsigned i = 0; i < tex->num_srcs; i++) {
 191       if (tex->src[i].src_type != nir_tex_src_coord)
 192          continue;
 193
 194       nir_ssa_def *coords =
 195          nir_ssa_for_src(b, tex->src[i].src, tex->coord_components);
 196       nir_instr_rewrite_src(&tex->instr,
 197                             &tex->src[i].src,
 198                             nir_src_for_ssa(nir_fmul(b, coords, scale)));
 199    }
 200
 201    tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
 202 }
 203
 204 static nir_ssa_def *
 205 sample_plane(nir_builder *b, nir_tex_instr *tex, int plane)
 206 {
 207    assert(tex->dest.is_ssa);
 208    assert(nir_tex_instr_dest_size(tex) == 4);
 209    assert(nir_alu_type_get_base_type(tex->dest_type) == nir_type_float);
 210    assert(tex->op == nir_texop_tex);
 211    assert(tex->coord_components == 2);
 212
 213    nir_tex_instr *plane_tex = nir_tex_instr_create(b->shader, 2);
 214    nir_src_copy(&plane_tex->src[0].src, &tex->src[0].src, plane_tex);
 215    plane_tex->src[0].src_type = nir_tex_src_coord;
 216    plane_tex->src[1].src = nir_src_for_ssa(nir_imm_int(b, plane));
 217    plane_tex->src[1].src_type = nir_tex_src_plane;
 218    plane_tex->op = nir_texop_tex;
 219    plane_tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
 220    plane_tex->dest_type = nir_type_float;
 221    plane_tex->coord_components = 2;
 222
 223    plane_tex->texture_index = tex->texture_index;
 224    plane_tex->texture = (nir_deref_var *)
 225       nir_copy_deref(plane_tex, &tex->texture->deref);
 226    plane_tex->sampler_index = tex->sampler_index;
 227    plane_tex->sampler = (nir_deref_var *)
 228       nir_copy_deref(plane_tex, &tex->sampler->deref);
 229
 230    nir_ssa_dest_init(&plane_tex->instr, &plane_tex->dest, 4, 32, NULL);
 231
 232    nir_builder_instr_insert(b, &plane_tex->instr);
 233
 234    return &plane_tex->dest.ssa;
 235 }
 236
 237 static void
 238 convert_yuv_to_rgb(nir_builder *b, nir_tex_instr *tex,
 239                    nir_ssa_def *y, nir_ssa_def *u, nir_ssa_def *v)
 240 {
 241    nir_const_value m[3] = {
 242       { .f32 = { 1.0f,  0.0f,         1.59602678f, 0.0f } },
 243       { .f32 = { 1.0f, -0.39176229f, -0.81296764f, 0.0f } },
 244       { .f32 = { 1.0f,  2.01723214f,  0.0f,        0.0f } }
 245    };
 246
 247    nir_ssa_def *yuv =
 248       nir_vec4(b,
 249                nir_fmul(b, nir_imm_float(b, 1.16438356f),
 250                         nir_fadd(b, y, nir_imm_float(b, -0.0625f))),
 251                nir_channel(b, nir_fadd(b, u, nir_imm_float(b, -0.5f)), 0),
 252                nir_channel(b, nir_fadd(b, v, nir_imm_float(b, -0.5f)), 0),
 253                nir_imm_float(b, 0.0));
 254
 255    nir_ssa_def *red = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, m[0]));
 256    nir_ssa_def *green = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, m[1]));
 257    nir_ssa_def *blue = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, m[2]));
 258
 259    nir_ssa_def *result = nir_vec4(b, red, green, blue, nir_imm_float(b, 1.0f));
 260
 261    nir_ssa_def_rewrite_uses(&tex->dest.ssa, nir_src_for_ssa(result));
 262 }
 263
 264 static void
 265 lower_y_uv_external(nir_builder *b, nir_tex_instr *tex)
 266 {
 267    b->cursor = nir_after_instr(&tex->instr);
 268
 269    nir_ssa_def *y = sample_plane(b, tex, 0);
 270    nir_ssa_def *uv = sample_plane(b, tex, 1);
 271
 272    convert_yuv_to_rgb(b, tex,
 273                       nir_channel(b, y, 0),
 274                       nir_channel(b, uv, 0),
 275                       nir_channel(b, uv, 1));
 276 }
 277
 278 static void
 279 lower_y_u_v_external(nir_builder *b, nir_tex_instr *tex)
 280 {
 281    b->cursor = nir_after_instr(&tex->instr);
 282
 283    nir_ssa_def *y = sample_plane(b, tex, 0);
 284    nir_ssa_def *u = sample_plane(b, tex, 1);
 285    nir_ssa_def *v = sample_plane(b, tex, 2);
 286
 287    convert_yuv_to_rgb(b, tex,
 288                       nir_channel(b, y, 0),
 289                       nir_channel(b, u, 0),
 290                       nir_channel(b, v, 0));
 291 }
 292
 293 static void
 294 lower_yx_xuxv_external(nir_builder *b, nir_tex_instr *tex)
 295 {
 296    b->cursor = nir_after_instr(&tex->instr);
 297
 298    nir_ssa_def *y = sample_plane(b, tex, 0);
 299    nir_ssa_def *xuxv = sample_plane(b, tex, 1);
 300
 301    convert_yuv_to_rgb(b, tex,
 302                       nir_channel(b, y, 0),
 303                       nir_channel(b, xuxv, 1),
 304                       nir_channel(b, xuxv, 3));
 305 }
 306
 307 /*
 308  * Emits a textureLod operation used to replace an existing
 309  * textureGrad instruction.
 310  */
 311 static void
 312 replace_gradient_with_lod(nir_builder *b, nir_ssa_def *lod, nir_tex_instr *tex)
 313 {
 314    /* We are going to emit a textureLod() with the same parameters except that
 315     * we replace ddx/ddy with lod.
 316     */
 317    int num_srcs = tex->num_srcs - 1;
 318    nir_tex_instr *txl = nir_tex_instr_create(b->shader, num_srcs);
 319
 320    txl->op = nir_texop_txl;
 321    txl->sampler_dim = tex->sampler_dim;
 322    txl->texture_index = tex->texture_index;
 323    txl->dest_type = tex->dest_type;
 324    txl->is_array = tex->is_array;
 325    txl->is_shadow = tex->is_shadow;
 326    txl->is_new_style_shadow = tex->is_new_style_shadow;
 327    txl->sampler_index = tex->sampler_index;
 328    txl->texture = (nir_deref_var *)
 329       nir_copy_deref(txl, &tex->texture->deref);
 330    txl->sampler = (nir_deref_var *)
 331       nir_copy_deref(txl, &tex->sampler->deref);
 332    txl->coord_components = tex->coord_components;
 333
 334    nir_ssa_dest_init(&txl->instr, &txl->dest, 4, 32, NULL);
 335
 336    int src_num = 0;
 337    for (int i = 0; i < tex->num_srcs; i++) {
 338       if (tex->src[i].src_type == nir_tex_src_ddx ||
 339           tex->src[i].src_type == nir_tex_src_ddy)
 340          continue;
 341       nir_src_copy(&txl->src[src_num].src, &tex->src[i].src, txl);
 342       txl->src[src_num].src_type = tex->src[i].src_type;
 343       src_num++;
 344    }
 345
 346    txl->src[src_num].src = nir_src_for_ssa(lod);
 347    txl->src[src_num].src_type = nir_tex_src_lod;
 348    src_num++;
 349
 350    assert(src_num == num_srcs);
 351
 352    nir_ssa_dest_init(&txl->instr, &txl->dest, 4, 32, NULL);
 353    nir_builder_instr_insert(b, &txl->instr);
 354
 355    nir_ssa_def_rewrite_uses(&tex->dest.ssa, nir_src_for_ssa(&txl->dest.ssa));
 356
 357    nir_instr_remove(&tex->instr);
 358 }
 359
 360 static void
 361 lower_gradient_cube_map(nir_builder *b, nir_tex_instr *tex)
 362 {
 363    assert(tex->sampler_dim == GLSL_SAMPLER_DIM_CUBE);
 364    assert(tex->op == nir_texop_txd);
 365    assert(tex->dest.is_ssa);
 366
 367    /* Use textureSize() to get the width and height of LOD 0 */
 368    nir_ssa_def *size = get_texture_size(b, tex);
 369
 370    /* Cubemap texture lookups first generate a texture coordinate normalized
 371     * to [-1, 1] on the appropiate face. The appropiate face is determined
 372     * by which component has largest magnitude and its sign. The texture
 373     * coordinate is the quotient of the remaining texture coordinates against
 374     * that absolute value of the component of largest magnitude. This
 375     * division requires that the computing of the derivative of the texel
 376     * coordinate must use the quotient rule. The high level GLSL code is as
 377     * follows:
 378     *
 379     * Step 1: selection
 380     *
 381     * vec3 abs_p, Q, dQdx, dQdy;
 382     * abs_p = abs(ir->coordinate);
 383     * if (abs_p.x >= max(abs_p.y, abs_p.z)) {
 384     *    Q = ir->coordinate.yzx;
 385     *    dQdx = ir->lod_info.grad.dPdx.yzx;
 386     *    dQdy = ir->lod_info.grad.dPdy.yzx;
 387     * }
 388     * if (abs_p.y >= max(abs_p.x, abs_p.z)) {
 389     *    Q = ir->coordinate.xzy;
 390     *    dQdx = ir->lod_info.grad.dPdx.xzy;
 391     *    dQdy = ir->lod_info.grad.dPdy.xzy;
 392     * }
 393     * if (abs_p.z >= max(abs_p.x, abs_p.y)) {
 394     *    Q = ir->coordinate;
 395     *    dQdx = ir->lod_info.grad.dPdx;
 396     *    dQdy = ir->lod_info.grad.dPdy;
 397     * }
 398     *
 399     * Step 2: use quotient rule to compute derivative. The normalized to
 400     * [-1, 1] texel coordinate is given by Q.xy / (sign(Q.z) * Q.z). We are
 401     * only concerned with the magnitudes of the derivatives whose values are
 402     * not affected by the sign. We drop the sign from the computation.
 403     *
 404     * vec2 dx, dy;
 405     * float recip;
 406     *
 407     * recip = 1.0 / Q.z;
 408     * dx = recip * ( dQdx.xy - Q.xy * (dQdx.z * recip) );
 409     * dy = recip * ( dQdy.xy - Q.xy * (dQdy.z * recip) );
 410     *
 411     * Step 3: compute LOD. At this point we have the derivatives of the
 412     * texture coordinates normalized to [-1,1]. We take the LOD to be
 413     *  result = log2(max(sqrt(dot(dx, dx)), sqrt(dy, dy)) * 0.5 * L)
 414     *         = -1.0 + log2(max(sqrt(dot(dx, dx)), sqrt(dy, dy)) * L)
 415     *         = -1.0 + log2(sqrt(max(dot(dx, dx), dot(dy,dy))) * L)
 416     *         = -1.0 + log2(sqrt(L * L * max(dot(dx, dx), dot(dy,dy))))
 417     *         = -1.0 + 0.5 * log2(L * L * max(dot(dx, dx), dot(dy,dy)))
 418     * where L is the dimension of the cubemap. The code is:
 419     *
 420     * float M, result;
 421     * M = max(dot(dx, dx), dot(dy, dy));
 422     * L = textureSize(sampler, 0).x;
 423     * result = -1.0 + 0.5 * log2(L * L * M);
 424     */
 425
 426    /* coordinate */
 427    nir_ssa_def *p =
 428       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_coord)].src.ssa;
 429
 430    /* unmodified dPdx, dPdy values */
 431    nir_ssa_def *dPdx =
 432       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_ddx)].src.ssa;
 433    nir_ssa_def *dPdy =
 434       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_ddy)].src.ssa;
 435
 436    nir_ssa_def *abs_p = nir_fabs(b, p);
 437    nir_ssa_def *abs_p_x = nir_channel(b, abs_p, 0);
 438    nir_ssa_def *abs_p_y = nir_channel(b, abs_p, 1);
 439    nir_ssa_def *abs_p_z = nir_channel(b, abs_p, 2);
 440
 441    /* 1. compute selector */
 442    nir_ssa_def *Q, *dQdx, *dQdy;
 443
 444    nir_ssa_def *cond_z = nir_fge(b, abs_p_z, nir_fmax(b, abs_p_x, abs_p_y));
 445    nir_ssa_def *cond_y = nir_fge(b, abs_p_y, nir_fmax(b, abs_p_x, abs_p_z));
 446
 447    unsigned yzx[4] = { 1, 2, 0, 0 };
 448    unsigned xzy[4] = { 0, 2, 1, 0 };
 449
 450    Q = nir_bcsel(b, cond_z,
 451                  p,
 452                  nir_bcsel(b, cond_y,
 453                            nir_swizzle(b, p, xzy, 3, false),
 454                            nir_swizzle(b, p, yzx, 3, false)));
 455
 456    dQdx = nir_bcsel(b, cond_z,
 457                     dPdx,
 458                     nir_bcsel(b, cond_y,
 459                               nir_swizzle(b, dPdx, xzy, 3, false),
 460                               nir_swizzle(b, dPdx, yzx, 3, false)));
 461
 462    dQdy = nir_bcsel(b, cond_z,
 463                     dPdy,
 464                     nir_bcsel(b, cond_y,
 465                               nir_swizzle(b, dPdy, xzy, 3, false),
 466                               nir_swizzle(b, dPdy, yzx, 3, false)));
 467
 468    /* 2. quotient rule */
 469
 470    /* tmp = Q.xy * recip;
 471     * dx = recip * ( dQdx.xy - (tmp * dQdx.z) );
 472     * dy = recip * ( dQdy.xy - (tmp * dQdy.z) );
 473     */
 474    nir_ssa_def *rcp_Q_z = nir_frcp(b, nir_channel(b, Q, 2));
 475
 476    unsigned xy[4] = { 0, 1, 0, 0 };
 477    nir_ssa_def *Q_xy = nir_swizzle(b, Q, xy, 2, false);
 478    nir_ssa_def *tmp = nir_fmul(b, Q_xy, rcp_Q_z);
 479
 480    nir_ssa_def *dQdx_xy = nir_swizzle(b, dQdx, xy, 2, false);
 481    nir_ssa_def *dQdx_z = nir_channel(b, dQdx, 2);
 482    nir_ssa_def *dx =
 483       nir_fmul(b, rcp_Q_z, nir_fsub(b, dQdx_xy, nir_fmul(b, tmp, dQdx_z)));
 484
 485    nir_ssa_def *dQdy_xy = nir_swizzle(b, dQdy, xy, 2, false);
 486    nir_ssa_def *dQdy_z = nir_channel(b, dQdy, 2);
 487    nir_ssa_def *dy =
 488       nir_fmul(b, rcp_Q_z, nir_fsub(b, dQdy_xy, nir_fmul(b, tmp, dQdy_z)));
 489
 490    /* M = max(dot(dx, dx), dot(dy, dy)); */
 491    nir_ssa_def *M = nir_fmax(b, nir_fdot(b, dx, dx), nir_fdot(b, dy, dy));
 492
 493    /* size has textureSize() of LOD 0 */
 494    nir_ssa_def *L = nir_channel(b, size, 0);
 495
 496    /* lod = -1.0 + 0.5 * log2(L * L * M); */
 497    nir_ssa_def *lod =
 498       nir_fadd(b,
 499                nir_imm_float(b, -1.0f),
 500                nir_fmul(b,
 501                         nir_imm_float(b, 0.5f),
 502                         nir_flog2(b, nir_fmul(b, L, nir_fmul(b, L, M)))));
 503
 504    /* 3. Replace the gradient instruction with an equivalent lod instruction */
 505    replace_gradient_with_lod(b, lod, tex);
 506 }
 507
 508 static void
 509 lower_gradient_shadow(nir_builder *b, nir_tex_instr *tex)
 510 {
 511    assert(tex->sampler_dim != GLSL_SAMPLER_DIM_CUBE);
 512    assert(tex->is_shadow);
 513    assert(tex->op == nir_texop_txd);
 514    assert(tex->dest.is_ssa);
 515
 516    /* Use textureSize() to get the width and height of LOD 0 */
 517    unsigned component_mask;
 518    switch (tex->sampler_dim) {
 519    case GLSL_SAMPLER_DIM_3D:
 520       component_mask = 7;
 521       break;
 522    case GLSL_SAMPLER_DIM_1D:
 523       component_mask = 1;
 524       break;
 525    default:
 526       component_mask = 3;
 527       break;
 528    }
 529
 530    nir_ssa_def *size =
 531       nir_channels(b, get_texture_size(b, tex), component_mask);
 532
 533    /* Scale the gradients by width and height.  Effectively, the incoming
 534     * gradients are s'(x,y), t'(x,y), and r'(x,y) from equation 3.19 in the
 535     * GL 3.0 spec; we want u'(x,y), which is w_t * s'(x,y).
 536     */
 537    nir_ssa_def *ddx =
 538       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_ddx)].src.ssa;
 539    nir_ssa_def *ddy =
 540       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_ddy)].src.ssa;
 541
 542    nir_ssa_def *dPdx = nir_fmul(b, ddx, size);
 543    nir_ssa_def *dPdy = nir_fmul(b, ddy, size);
 544
 545    nir_ssa_def *rho;
 546    if (dPdx->num_components == 1) {
 547       rho = nir_fmax(b, nir_fabs(b, dPdx), nir_fabs(b, dPdy));
 548    } else {
 549       rho = nir_fmax(b,
 550                      nir_fsqrt(b, nir_fdot(b, dPdx, dPdx)),
 551                      nir_fsqrt(b, nir_fdot(b, dPdy, dPdy)));
 552    }
 553
 554    /* lod = log2(rho).  We're ignoring GL state biases for now. */
 555    nir_ssa_def *lod = nir_flog2(b, rho);
 556
 557    /* Replace the gradient instruction with an equivalent lod instruction */
 558    replace_gradient_with_lod(b, lod, tex);
 559 }
 560
 561 static void
 562 saturate_src(nir_builder *b, nir_tex_instr *tex, unsigned sat_mask)
 563 {
 564    b->cursor = nir_before_instr(&tex->instr);
 565
 566    /* Walk through the sources saturating the requested arguments. */
 567    for (unsigned i = 0; i < tex->num_srcs; i++) {
 568       if (tex->src[i].src_type != nir_tex_src_coord)
 569          continue;
 570
 571       nir_ssa_def *src =
 572          nir_ssa_for_src(b, tex->src[i].src, tex->coord_components);
 573
 574       /* split src into components: */
 575       nir_ssa_def *comp[4];
 576
 577       assume(tex->coord_components >= 1);
 578
 579       for (unsigned j = 0; j < tex->coord_components; j++)
 580          comp[j] = nir_channel(b, src, j);
 581
 582       /* clamp requested components, array index does not get clamped: */
 583       unsigned ncomp = tex->coord_components;
 584       if (tex->is_array)
 585          ncomp--;
 586
 587       for (unsigned j = 0; j < ncomp; j++) {
 588          if ((1 << j) & sat_mask) {
 589             if (tex->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
 590                /* non-normalized texture coords, so clamp to texture
 591                 * size rather than [0.0, 1.0]
 592                 */
 593                nir_ssa_def *txs = get_texture_size(b, tex);
 594                comp[j] = nir_fmax(b, comp[j], nir_imm_float(b, 0.0));
 595                comp[j] = nir_fmin(b, comp[j], nir_channel(b, txs, j));
 596             } else {
 597                comp[j] = nir_fsat(b, comp[j]);
 598             }
 599          }
 600       }
 601
 602       /* and move the result back into a single vecN: */
 603       src = nir_vec(b, comp, tex->coord_components);
 604
 605       nir_instr_rewrite_src(&tex->instr,
 606                             &tex->src[i].src,
 607                             nir_src_for_ssa(src));
 608    }
 609 }
 610
 611 static nir_ssa_def *
 612 get_zero_or_one(nir_builder *b, nir_alu_type type, uint8_t swizzle_val)
 613 {
 614    nir_const_value v;
 615
 616    memset(&v, 0, sizeof(v));
 617
 618    if (swizzle_val == 4) {
 619       v.u32[0] = v.u32[1] = v.u32[2] = v.u32[3] = 0;
 620    } else {
 621       assert(swizzle_val == 5);
 622       if (type == nir_type_float)
 623          v.f32[0] = v.f32[1] = v.f32[2] = v.f32[3] = 1.0;
 624       else
 625          v.u32[0] = v.u32[1] = v.u32[2] = v.u32[3] = 1;
 626    }
 627
 628    return nir_build_imm(b, 4, 32, v);
 629 }
 630
 631 static void
 632 swizzle_result(nir_builder *b, nir_tex_instr *tex, const uint8_t swizzle[4])
 633 {
 634    assert(tex->dest.is_ssa);
 635
 636    b->cursor = nir_after_instr(&tex->instr);
 637
 638    nir_ssa_def *swizzled;
 639    if (tex->op == nir_texop_tg4) {
 640       if (swizzle[tex->component] < 4) {
 641          /* This one's easy */
 642          tex->component = swizzle[tex->component];
 643          return;
 644       } else {
 645          swizzled = get_zero_or_one(b, tex->dest_type, swizzle[tex->component]);
 646       }
 647    } else {
 648       assert(nir_tex_instr_dest_size(tex) == 4);
 649       if (swizzle[0] < 4 && swizzle[1] < 4 &&
 650           swizzle[2] < 4 && swizzle[3] < 4) {
 651          unsigned swiz[4] = { swizzle[0], swizzle[1], swizzle[2], swizzle[3] };
 652          /* We have no 0's or 1's, just emit a swizzling MOV */
 653          swizzled = nir_swizzle(b, &tex->dest.ssa, swiz, 4, false);
 654       } else {
 655          nir_ssa_def *srcs[4];
 656          for (unsigned i = 0; i < 4; i++) {
 657             if (swizzle[i] < 4) {
 658                srcs[i] = nir_channel(b, &tex->dest.ssa, swizzle[i]);
 659             } else {
 660                srcs[i] = get_zero_or_one(b, tex->dest_type, swizzle[i]);
 661             }
 662          }
 663          swizzled = nir_vec(b, srcs, 4);
 664       }
 665    }
 666
 667    nir_ssa_def_rewrite_uses_after(&tex->dest.ssa, nir_src_for_ssa(swizzled),
 668                                   swizzled->parent_instr);
 669 }
 670
 671 static void
 672 linearize_srgb_result(nir_builder *b, nir_tex_instr *tex)
 673 {
 674    assert(tex->dest.is_ssa);
 675    assert(nir_tex_instr_dest_size(tex) == 4);
 676    assert(nir_alu_type_get_base_type(tex->dest_type) == nir_type_float);
 677
 678    b->cursor = nir_after_instr(&tex->instr);
 679
 680    static const unsigned swiz[4] = {0, 1, 2, 0};
 681    nir_ssa_def *comp = nir_swizzle(b, &tex->dest.ssa, swiz, 3, true);
 682
 683    /* Formula is:
 684     *    (comp <= 0.04045) ?
 685     *          (comp / 12.92) :
 686     *          pow((comp + 0.055) / 1.055, 2.4)
 687     */
 688    nir_ssa_def *low  = nir_fmul(b, comp, nir_imm_float(b, 1.0 / 12.92));
 689    nir_ssa_def *high = nir_fpow(b,
 690                                 nir_fmul(b,
 691                                          nir_fadd(b,
 692                                                   comp,
 693                                                   nir_imm_float(b, 0.055)),
 694                                          nir_imm_float(b, 1.0 / 1.055)),
 695                                 nir_imm_float(b, 2.4));
 696    nir_ssa_def *cond = nir_fge(b, nir_imm_float(b, 0.04045), comp);
 697    nir_ssa_def *rgb  = nir_bcsel(b, cond, low, high);
 698
 699    /* alpha is untouched: */
 700    nir_ssa_def *result = nir_vec4(b,
 701                                   nir_channel(b, rgb, 0),
 702                                   nir_channel(b, rgb, 1),
 703                                   nir_channel(b, rgb, 2),
 704                                   nir_channel(b, &tex->dest.ssa, 3));
 705
 706    nir_ssa_def_rewrite_uses_after(&tex->dest.ssa, nir_src_for_ssa(result),
 707                                   result->parent_instr);
 708 }
 709
 710 static bool
 711 nir_lower_tex_block(nir_block *block, nir_builder *b,
 712                     const nir_lower_tex_options *options)
 713 {
 714    bool progress = false;
 715
 716    nir_foreach_instr_safe(instr, block) {
 717       if (instr->type != nir_instr_type_tex)
 718          continue;
 719
 720       nir_tex_instr *tex = nir_instr_as_tex(instr);
 721       bool lower_txp = !!(options->lower_txp & (1 << tex->sampler_dim));
 722
 723       /* mask of src coords to saturate (clamp): */
 724       unsigned sat_mask = 0;
 725
 726       if ((1 << tex->sampler_index) & options->saturate_r)
 727          sat_mask |= (1 << 2);    /* .z */
 728       if ((1 << tex->sampler_index) & options->saturate_t)
 729          sat_mask |= (1 << 1);    /* .y */
 730       if ((1 << tex->sampler_index) & options->saturate_s)
 731          sat_mask |= (1 << 0);    /* .x */
 732
 733       /* If we are clamping any coords, we must lower projector first
 734        * as clamping happens *after* projection:
 735        */
 736       if (lower_txp || sat_mask) {
 737          project_src(b, tex);
 738          progress = true;
 739       }
 740
 741       if ((tex->op == nir_texop_txf && options->lower_txf_offset) ||
 742           (tex->sampler_dim == GLSL_SAMPLER_DIM_RECT &&
 743            options->lower_rect_offset)) {
 744          progress = lower_offset(b, tex) || progress;
 745       }
 746
 747       if ((tex->sampler_dim == GLSL_SAMPLER_DIM_RECT) && options->lower_rect) {
 748          lower_rect(b, tex);
 749          progress = true;
 750       }
 751
 752       if ((1 << tex->texture_index) & options->lower_y_uv_external) {
 753          lower_y_uv_external(b, tex);
 754          progress = true;
 755       }
 756
 757       if ((1 << tex->texture_index) & options->lower_y_u_v_external) {
 758          lower_y_u_v_external(b, tex);
 759          progress = true;
 760       }
 761
 762       if ((1 << tex->texture_index) & options->lower_yx_xuxv_external) {
 763          lower_yx_xuxv_external(b, tex);
 764          progress = true;
 765       }
 766
 767
 768       if (sat_mask) {
 769          saturate_src(b, tex, sat_mask);
 770          progress = true;
 771       }
 772
 773       if (((1 << tex->texture_index) & options->swizzle_result) &&
 774           !nir_tex_instr_is_query(tex) &&
 775           !(tex->is_shadow && tex->is_new_style_shadow)) {
 776          swizzle_result(b, tex, options->swizzles[tex->texture_index]);
 777          progress = true;
 778       }
 779
 780       /* should be after swizzle so we know which channels are rgb: */
 781       if (((1 << tex->texture_index) & options->lower_srgb) &&
 782           !nir_tex_instr_is_query(tex) && !tex->is_shadow) {
 783          linearize_srgb_result(b, tex);
 784          progress = true;
 785       }
 786
 787       if (tex->op == nir_texop_txd && options->lower_txd_cube_map &&
 788           tex->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
 789          lower_gradient_cube_map(b, tex);
 790          progress = true;
 791          continue;
 792       }
 793
 794       if (tex->op == nir_texop_txd && options->lower_txd_shadow &&
 795           tex->is_shadow && tex->sampler_dim != GLSL_SAMPLER_DIM_CUBE) {
 796          lower_gradient_shadow(b, tex);
 797          progress = true;
 798          continue;
 799       }
 800    }
 801
 802    return progress;
 803 }
 804
 805 static bool
 806 nir_lower_tex_impl(nir_function_impl *impl,
 807                    const nir_lower_tex_options *options)
 808 {
 809    bool progress = false;
 810    nir_builder builder;
 811    nir_builder_init(&builder, impl);
 812
 813    nir_foreach_block(block, impl) {
 814       progress |= nir_lower_tex_block(block, &builder, options);
 815    }
 816
 817    nir_metadata_preserve(impl, nir_metadata_block_index |
 818                                nir_metadata_dominance);
 819    return progress;
 820 }
 821
 822 bool
 823 nir_lower_tex(nir_shader *shader, const nir_lower_tex_options *options)
 824 {
 825    bool progress = false;
 826
 827    nir_foreach_function(function, shader) {
 828       if (function->impl)
 829          progress |= nir_lower_tex_impl(function->impl, options);
 830    }
 831
 832    return progress;
 833 }