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