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_i2f32(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_clone(tex->texture, txs);
 167    txs->sampler_index = tex->sampler_index;
 168    txs->sampler = nir_deref_var_clone(tex->sampler, txs);
 169    txs->dest_type = nir_type_int;
 170
 171    /* only single src, the lod: */
 172    txs->src[0].src = nir_src_for_ssa(nir_imm_int(b, 0));
 173    txs->src[0].src_type = nir_tex_src_lod;
 174
 175    nir_ssa_dest_init(&txs->instr, &txs->dest,
 176                      nir_tex_instr_dest_size(txs), 32, NULL);
 177    nir_builder_instr_insert(b, &txs->instr);
 178
 179    return nir_i2f32(b, &txs->dest.ssa);
 180 }
 181
 182 static void
 183 lower_rect(nir_builder *b, nir_tex_instr *tex)
 184 {
 185    nir_ssa_def *txs = get_texture_size(b, tex);
 186    nir_ssa_def *scale = nir_frcp(b, txs);
 187
 188    /* Walk through the sources normalizing the requested arguments. */
 189    for (unsigned i = 0; i < tex->num_srcs; i++) {
 190       if (tex->src[i].src_type != nir_tex_src_coord)
 191          continue;
 192
 193       nir_ssa_def *coords =
 194          nir_ssa_for_src(b, tex->src[i].src, tex->coord_components);
 195       nir_instr_rewrite_src(&tex->instr,
 196                             &tex->src[i].src,
 197                             nir_src_for_ssa(nir_fmul(b, coords, scale)));
 198    }
 199
 200    tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
 201 }
 202
 203 static nir_ssa_def *
 204 sample_plane(nir_builder *b, nir_tex_instr *tex, int plane)
 205 {
 206    assert(tex->dest.is_ssa);
 207    assert(nir_tex_instr_dest_size(tex) == 4);
 208    assert(nir_alu_type_get_base_type(tex->dest_type) == nir_type_float);
 209    assert(tex->op == nir_texop_tex);
 210    assert(tex->coord_components == 2);
 211
 212    nir_tex_instr *plane_tex = nir_tex_instr_create(b->shader, 2);
 213    nir_src_copy(&plane_tex->src[0].src, &tex->src[0].src, plane_tex);
 214    plane_tex->src[0].src_type = nir_tex_src_coord;
 215    plane_tex->src[1].src = nir_src_for_ssa(nir_imm_int(b, plane));
 216    plane_tex->src[1].src_type = nir_tex_src_plane;
 217    plane_tex->op = nir_texop_tex;
 218    plane_tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
 219    plane_tex->dest_type = nir_type_float;
 220    plane_tex->coord_components = 2;
 221
 222    plane_tex->texture_index = tex->texture_index;
 223    plane_tex->texture = nir_deref_var_clone(tex->texture, plane_tex);
 224    plane_tex->sampler_index = tex->sampler_index;
 225    plane_tex->sampler = nir_deref_var_clone(tex->sampler, plane_tex);
 226
 227    nir_ssa_dest_init(&plane_tex->instr, &plane_tex->dest, 4, 32, NULL);
 228
 229    nir_builder_instr_insert(b, &plane_tex->instr);
 230
 231    return &plane_tex->dest.ssa;
 232 }
 233
 234 static void
 235 convert_yuv_to_rgb(nir_builder *b, nir_tex_instr *tex,
 236                    nir_ssa_def *y, nir_ssa_def *u, nir_ssa_def *v)
 237 {
 238    nir_const_value m[3] = {
 239       { .f32 = { 1.0f,  0.0f,         1.59602678f, 0.0f } },
 240       { .f32 = { 1.0f, -0.39176229f, -0.81296764f, 0.0f } },
 241       { .f32 = { 1.0f,  2.01723214f,  0.0f,        0.0f } }
 242    };
 243
 244    nir_ssa_def *yuv =
 245       nir_vec4(b,
 246                nir_fmul(b, nir_imm_float(b, 1.16438356f),
 247                         nir_fadd(b, y, nir_imm_float(b, -16.0f / 255.0f))),
 248                nir_channel(b, nir_fadd(b, u, nir_imm_float(b, -128.0f / 255.0f)), 0),
 249                nir_channel(b, nir_fadd(b, v, nir_imm_float(b, -128.0f / 255.0f)), 0),
 250                nir_imm_float(b, 0.0));
 251
 252    nir_ssa_def *red = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, m[0]));
 253    nir_ssa_def *green = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, m[1]));
 254    nir_ssa_def *blue = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, m[2]));
 255
 256    nir_ssa_def *result = nir_vec4(b, red, green, blue, nir_imm_float(b, 1.0f));
 257
 258    nir_ssa_def_rewrite_uses(&tex->dest.ssa, nir_src_for_ssa(result));
 259 }
 260
 261 static void
 262 lower_y_uv_external(nir_builder *b, nir_tex_instr *tex)
 263 {
 264    b->cursor = nir_after_instr(&tex->instr);
 265
 266    nir_ssa_def *y = sample_plane(b, tex, 0);
 267    nir_ssa_def *uv = sample_plane(b, tex, 1);
 268
 269    convert_yuv_to_rgb(b, tex,
 270                       nir_channel(b, y, 0),
 271                       nir_channel(b, uv, 0),
 272                       nir_channel(b, uv, 1));
 273 }
 274
 275 static void
 276 lower_y_u_v_external(nir_builder *b, nir_tex_instr *tex)
 277 {
 278    b->cursor = nir_after_instr(&tex->instr);
 279
 280    nir_ssa_def *y = sample_plane(b, tex, 0);
 281    nir_ssa_def *u = sample_plane(b, tex, 1);
 282    nir_ssa_def *v = sample_plane(b, tex, 2);
 283
 284    convert_yuv_to_rgb(b, tex,
 285                       nir_channel(b, y, 0),
 286                       nir_channel(b, u, 0),
 287                       nir_channel(b, v, 0));
 288 }
 289
 290 static void
 291 lower_yx_xuxv_external(nir_builder *b, nir_tex_instr *tex)
 292 {
 293    b->cursor = nir_after_instr(&tex->instr);
 294
 295    nir_ssa_def *y = sample_plane(b, tex, 0);
 296    nir_ssa_def *xuxv = sample_plane(b, tex, 1);
 297
 298    convert_yuv_to_rgb(b, tex,
 299                       nir_channel(b, y, 0),
 300                       nir_channel(b, xuxv, 1),
 301                       nir_channel(b, xuxv, 3));
 302 }
 303
 304 static void
 305 lower_xy_uxvx_external(nir_builder *b, nir_tex_instr *tex)
 306 {
 307   b->cursor = nir_after_instr(&tex->instr);
 308
 309   nir_ssa_def *y = sample_plane(b, tex, 0);
 310   nir_ssa_def *uxvx = sample_plane(b, tex, 1);
 311
 312   convert_yuv_to_rgb(b, tex,
 313                      nir_channel(b, y, 1),
 314                      nir_channel(b, uxvx, 0),
 315                      nir_channel(b, uxvx, 2));
 316 }
 317
 318 /*
 319  * Emits a textureLod operation used to replace an existing
 320  * textureGrad instruction.
 321  */
 322 static void
 323 replace_gradient_with_lod(nir_builder *b, nir_ssa_def *lod, nir_tex_instr *tex)
 324 {
 325    /* We are going to emit a textureLod() with the same parameters except that
 326     * we replace ddx/ddy with lod.
 327     */
 328    int num_srcs = tex->num_srcs - 1;
 329    nir_tex_instr *txl = nir_tex_instr_create(b->shader, num_srcs);
 330
 331    txl->op = nir_texop_txl;
 332    txl->sampler_dim = tex->sampler_dim;
 333    txl->texture_index = tex->texture_index;
 334    txl->dest_type = tex->dest_type;
 335    txl->is_array = tex->is_array;
 336    txl->is_shadow = tex->is_shadow;
 337    txl->is_new_style_shadow = tex->is_new_style_shadow;
 338    txl->sampler_index = tex->sampler_index;
 339    txl->texture = nir_deref_var_clone(tex->texture, txl);
 340    txl->sampler = nir_deref_var_clone(tex->sampler, txl);
 341    txl->coord_components = tex->coord_components;
 342
 343    nir_ssa_dest_init(&txl->instr, &txl->dest, 4, 32, NULL);
 344
 345    int src_num = 0;
 346    for (int i = 0; i < tex->num_srcs; i++) {
 347       if (tex->src[i].src_type == nir_tex_src_ddx ||
 348           tex->src[i].src_type == nir_tex_src_ddy)
 349          continue;
 350       nir_src_copy(&txl->src[src_num].src, &tex->src[i].src, txl);
 351       txl->src[src_num].src_type = tex->src[i].src_type;
 352       src_num++;
 353    }
 354
 355    txl->src[src_num].src = nir_src_for_ssa(lod);
 356    txl->src[src_num].src_type = nir_tex_src_lod;
 357    src_num++;
 358
 359    assert(src_num == num_srcs);
 360
 361    nir_ssa_dest_init(&txl->instr, &txl->dest,
 362                      tex->dest.ssa.num_components, 32, NULL);
 363    nir_builder_instr_insert(b, &txl->instr);
 364
 365    nir_ssa_def_rewrite_uses(&tex->dest.ssa, nir_src_for_ssa(&txl->dest.ssa));
 366
 367    nir_instr_remove(&tex->instr);
 368 }
 369
 370 static void
 371 lower_gradient_cube_map(nir_builder *b, nir_tex_instr *tex)
 372 {
 373    assert(tex->sampler_dim == GLSL_SAMPLER_DIM_CUBE);
 374    assert(tex->op == nir_texop_txd);
 375    assert(tex->dest.is_ssa);
 376
 377    /* Use textureSize() to get the width and height of LOD 0 */
 378    nir_ssa_def *size = get_texture_size(b, tex);
 379
 380    /* Cubemap texture lookups first generate a texture coordinate normalized
 381     * to [-1, 1] on the appropiate face. The appropiate face is determined
 382     * by which component has largest magnitude and its sign. The texture
 383     * coordinate is the quotient of the remaining texture coordinates against
 384     * that absolute value of the component of largest magnitude. This
 385     * division requires that the computing of the derivative of the texel
 386     * coordinate must use the quotient rule. The high level GLSL code is as
 387     * follows:
 388     *
 389     * Step 1: selection
 390     *
 391     * vec3 abs_p, Q, dQdx, dQdy;
 392     * abs_p = abs(ir->coordinate);
 393     * if (abs_p.x >= max(abs_p.y, abs_p.z)) {
 394     *    Q = ir->coordinate.yzx;
 395     *    dQdx = ir->lod_info.grad.dPdx.yzx;
 396     *    dQdy = ir->lod_info.grad.dPdy.yzx;
 397     * }
 398     * if (abs_p.y >= max(abs_p.x, abs_p.z)) {
 399     *    Q = ir->coordinate.xzy;
 400     *    dQdx = ir->lod_info.grad.dPdx.xzy;
 401     *    dQdy = ir->lod_info.grad.dPdy.xzy;
 402     * }
 403     * if (abs_p.z >= max(abs_p.x, abs_p.y)) {
 404     *    Q = ir->coordinate;
 405     *    dQdx = ir->lod_info.grad.dPdx;
 406     *    dQdy = ir->lod_info.grad.dPdy;
 407     * }
 408     *
 409     * Step 2: use quotient rule to compute derivative. The normalized to
 410     * [-1, 1] texel coordinate is given by Q.xy / (sign(Q.z) * Q.z). We are
 411     * only concerned with the magnitudes of the derivatives whose values are
 412     * not affected by the sign. We drop the sign from the computation.
 413     *
 414     * vec2 dx, dy;
 415     * float recip;
 416     *
 417     * recip = 1.0 / Q.z;
 418     * dx = recip * ( dQdx.xy - Q.xy * (dQdx.z * recip) );
 419     * dy = recip * ( dQdy.xy - Q.xy * (dQdy.z * recip) );
 420     *
 421     * Step 3: compute LOD. At this point we have the derivatives of the
 422     * texture coordinates normalized to [-1,1]. We take the LOD to be
 423     *  result = log2(max(sqrt(dot(dx, dx)), sqrt(dy, dy)) * 0.5 * L)
 424     *         = -1.0 + log2(max(sqrt(dot(dx, dx)), sqrt(dy, dy)) * L)
 425     *         = -1.0 + log2(sqrt(max(dot(dx, dx), dot(dy,dy))) * L)
 426     *         = -1.0 + log2(sqrt(L * L * max(dot(dx, dx), dot(dy,dy))))
 427     *         = -1.0 + 0.5 * log2(L * L * max(dot(dx, dx), dot(dy,dy)))
 428     * where L is the dimension of the cubemap. The code is:
 429     *
 430     * float M, result;
 431     * M = max(dot(dx, dx), dot(dy, dy));
 432     * L = textureSize(sampler, 0).x;
 433     * result = -1.0 + 0.5 * log2(L * L * M);
 434     */
 435
 436    /* coordinate */
 437    nir_ssa_def *p =
 438       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_coord)].src.ssa;
 439
 440    /* unmodified dPdx, dPdy values */
 441    nir_ssa_def *dPdx =
 442       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_ddx)].src.ssa;
 443    nir_ssa_def *dPdy =
 444       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_ddy)].src.ssa;
 445
 446    nir_ssa_def *abs_p = nir_fabs(b, p);
 447    nir_ssa_def *abs_p_x = nir_channel(b, abs_p, 0);
 448    nir_ssa_def *abs_p_y = nir_channel(b, abs_p, 1);
 449    nir_ssa_def *abs_p_z = nir_channel(b, abs_p, 2);
 450
 451    /* 1. compute selector */
 452    nir_ssa_def *Q, *dQdx, *dQdy;
 453
 454    nir_ssa_def *cond_z = nir_fge(b, abs_p_z, nir_fmax(b, abs_p_x, abs_p_y));
 455    nir_ssa_def *cond_y = nir_fge(b, abs_p_y, nir_fmax(b, abs_p_x, abs_p_z));
 456
 457    unsigned yzx[4] = { 1, 2, 0, 0 };
 458    unsigned xzy[4] = { 0, 2, 1, 0 };
 459
 460    Q = nir_bcsel(b, cond_z,
 461                  p,
 462                  nir_bcsel(b, cond_y,
 463                            nir_swizzle(b, p, xzy, 3, false),
 464                            nir_swizzle(b, p, yzx, 3, false)));
 465
 466    dQdx = nir_bcsel(b, cond_z,
 467                     dPdx,
 468                     nir_bcsel(b, cond_y,
 469                               nir_swizzle(b, dPdx, xzy, 3, false),
 470                               nir_swizzle(b, dPdx, yzx, 3, false)));
 471
 472    dQdy = nir_bcsel(b, cond_z,
 473                     dPdy,
 474                     nir_bcsel(b, cond_y,
 475                               nir_swizzle(b, dPdy, xzy, 3, false),
 476                               nir_swizzle(b, dPdy, yzx, 3, false)));
 477
 478    /* 2. quotient rule */
 479
 480    /* tmp = Q.xy * recip;
 481     * dx = recip * ( dQdx.xy - (tmp * dQdx.z) );
 482     * dy = recip * ( dQdy.xy - (tmp * dQdy.z) );
 483     */
 484    nir_ssa_def *rcp_Q_z = nir_frcp(b, nir_channel(b, Q, 2));
 485
 486    unsigned xy[4] = { 0, 1, 0, 0 };
 487    nir_ssa_def *Q_xy = nir_swizzle(b, Q, xy, 2, false);
 488    nir_ssa_def *tmp = nir_fmul(b, Q_xy, rcp_Q_z);
 489
 490    nir_ssa_def *dQdx_xy = nir_swizzle(b, dQdx, xy, 2, false);
 491    nir_ssa_def *dQdx_z = nir_channel(b, dQdx, 2);
 492    nir_ssa_def *dx =
 493       nir_fmul(b, rcp_Q_z, nir_fsub(b, dQdx_xy, nir_fmul(b, tmp, dQdx_z)));
 494
 495    nir_ssa_def *dQdy_xy = nir_swizzle(b, dQdy, xy, 2, false);
 496    nir_ssa_def *dQdy_z = nir_channel(b, dQdy, 2);
 497    nir_ssa_def *dy =
 498       nir_fmul(b, rcp_Q_z, nir_fsub(b, dQdy_xy, nir_fmul(b, tmp, dQdy_z)));
 499
 500    /* M = max(dot(dx, dx), dot(dy, dy)); */
 501    nir_ssa_def *M = nir_fmax(b, nir_fdot(b, dx, dx), nir_fdot(b, dy, dy));
 502
 503    /* size has textureSize() of LOD 0 */
 504    nir_ssa_def *L = nir_channel(b, size, 0);
 505
 506    /* lod = -1.0 + 0.5 * log2(L * L * M); */
 507    nir_ssa_def *lod =
 508       nir_fadd(b,
 509                nir_imm_float(b, -1.0f),
 510                nir_fmul(b,
 511                         nir_imm_float(b, 0.5f),
 512                         nir_flog2(b, nir_fmul(b, L, nir_fmul(b, L, M)))));
 513
 514    /* 3. Replace the gradient instruction with an equivalent lod instruction */
 515    replace_gradient_with_lod(b, lod, tex);
 516 }
 517
 518 static void
 519 lower_gradient_shadow(nir_builder *b, nir_tex_instr *tex)
 520 {
 521    assert(tex->sampler_dim != GLSL_SAMPLER_DIM_CUBE);
 522    assert(tex->is_shadow);
 523    assert(tex->op == nir_texop_txd);
 524    assert(tex->dest.is_ssa);
 525
 526    /* Use textureSize() to get the width and height of LOD 0 */
 527    unsigned component_mask;
 528    switch (tex->sampler_dim) {
 529    case GLSL_SAMPLER_DIM_3D:
 530       component_mask = 7;
 531       break;
 532    case GLSL_SAMPLER_DIM_1D:
 533       component_mask = 1;
 534       break;
 535    default:
 536       component_mask = 3;
 537       break;
 538    }
 539
 540    nir_ssa_def *size =
 541       nir_channels(b, get_texture_size(b, tex), component_mask);
 542
 543    /* Scale the gradients by width and height.  Effectively, the incoming
 544     * gradients are s'(x,y), t'(x,y), and r'(x,y) from equation 3.19 in the
 545     * GL 3.0 spec; we want u'(x,y), which is w_t * s'(x,y).
 546     */
 547    nir_ssa_def *ddx =
 548       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_ddx)].src.ssa;
 549    nir_ssa_def *ddy =
 550       tex->src[nir_tex_instr_src_index(tex, nir_tex_src_ddy)].src.ssa;
 551
 552    nir_ssa_def *dPdx = nir_fmul(b, ddx, size);
 553    nir_ssa_def *dPdy = nir_fmul(b, ddy, size);
 554
 555    nir_ssa_def *rho;
 556    if (dPdx->num_components == 1) {
 557       rho = nir_fmax(b, nir_fabs(b, dPdx), nir_fabs(b, dPdy));
 558    } else {
 559       rho = nir_fmax(b,
 560                      nir_fsqrt(b, nir_fdot(b, dPdx, dPdx)),
 561                      nir_fsqrt(b, nir_fdot(b, dPdy, dPdy)));
 562    }
 563
 564    /* lod = log2(rho).  We're ignoring GL state biases for now. */
 565    nir_ssa_def *lod = nir_flog2(b, rho);
 566
 567    /* Replace the gradient instruction with an equivalent lod instruction */
 568    replace_gradient_with_lod(b, lod, tex);
 569 }
 570
 571 static void
 572 saturate_src(nir_builder *b, nir_tex_instr *tex, unsigned sat_mask)
 573 {
 574    b->cursor = nir_before_instr(&tex->instr);
 575
 576    /* Walk through the sources saturating the requested arguments. */
 577    for (unsigned i = 0; i < tex->num_srcs; i++) {
 578       if (tex->src[i].src_type != nir_tex_src_coord)
 579          continue;
 580
 581       nir_ssa_def *src =
 582          nir_ssa_for_src(b, tex->src[i].src, tex->coord_components);
 583
 584       /* split src into components: */
 585       nir_ssa_def *comp[4];
 586
 587       assume(tex->coord_components >= 1);
 588
 589       for (unsigned j = 0; j < tex->coord_components; j++)
 590          comp[j] = nir_channel(b, src, j);
 591
 592       /* clamp requested components, array index does not get clamped: */
 593       unsigned ncomp = tex->coord_components;
 594       if (tex->is_array)
 595          ncomp--;
 596
 597       for (unsigned j = 0; j < ncomp; j++) {
 598          if ((1 << j) & sat_mask) {
 599             if (tex->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
 600                /* non-normalized texture coords, so clamp to texture
 601                 * size rather than [0.0, 1.0]
 602                 */
 603                nir_ssa_def *txs = get_texture_size(b, tex);
 604                comp[j] = nir_fmax(b, comp[j], nir_imm_float(b, 0.0));
 605                comp[j] = nir_fmin(b, comp[j], nir_channel(b, txs, j));
 606             } else {
 607                comp[j] = nir_fsat(b, comp[j]);
 608             }
 609          }
 610       }
 611
 612       /* and move the result back into a single vecN: */
 613       src = nir_vec(b, comp, tex->coord_components);
 614
 615       nir_instr_rewrite_src(&tex->instr,
 616                             &tex->src[i].src,
 617                             nir_src_for_ssa(src));
 618    }
 619 }
 620
 621 static nir_ssa_def *
 622 get_zero_or_one(nir_builder *b, nir_alu_type type, uint8_t swizzle_val)
 623 {
 624    nir_const_value v;
 625
 626    memset(&v, 0, sizeof(v));
 627
 628    if (swizzle_val == 4) {
 629       v.u32[0] = v.u32[1] = v.u32[2] = v.u32[3] = 0;
 630    } else {
 631       assert(swizzle_val == 5);
 632       if (type == nir_type_float)
 633          v.f32[0] = v.f32[1] = v.f32[2] = v.f32[3] = 1.0;
 634       else
 635          v.u32[0] = v.u32[1] = v.u32[2] = v.u32[3] = 1;
 636    }
 637
 638    return nir_build_imm(b, 4, 32, v);
 639 }
 640
 641 static void
 642 swizzle_result(nir_builder *b, nir_tex_instr *tex, const uint8_t swizzle[4])
 643 {
 644    assert(tex->dest.is_ssa);
 645
 646    b->cursor = nir_after_instr(&tex->instr);
 647
 648    nir_ssa_def *swizzled;
 649    if (tex->op == nir_texop_tg4) {
 650       if (swizzle[tex->component] < 4) {
 651          /* This one's easy */
 652          tex->component = swizzle[tex->component];
 653          return;
 654       } else {
 655          swizzled = get_zero_or_one(b, tex->dest_type, swizzle[tex->component]);
 656       }
 657    } else {
 658       assert(nir_tex_instr_dest_size(tex) == 4);
 659       if (swizzle[0] < 4 && swizzle[1] < 4 &&
 660           swizzle[2] < 4 && swizzle[3] < 4) {
 661          unsigned swiz[4] = { swizzle[0], swizzle[1], swizzle[2], swizzle[3] };
 662          /* We have no 0s or 1s, just emit a swizzling MOV */
 663          swizzled = nir_swizzle(b, &tex->dest.ssa, swiz, 4, false);
 664       } else {
 665          nir_ssa_def *srcs[4];
 666          for (unsigned i = 0; i < 4; i++) {
 667             if (swizzle[i] < 4) {
 668                srcs[i] = nir_channel(b, &tex->dest.ssa, swizzle[i]);
 669             } else {
 670                srcs[i] = get_zero_or_one(b, tex->dest_type, swizzle[i]);
 671             }
 672          }
 673          swizzled = nir_vec(b, srcs, 4);
 674       }
 675    }
 676
 677    nir_ssa_def_rewrite_uses_after(&tex->dest.ssa, nir_src_for_ssa(swizzled),
 678                                   swizzled->parent_instr);
 679 }
 680
 681 static void
 682 linearize_srgb_result(nir_builder *b, nir_tex_instr *tex)
 683 {
 684    assert(tex->dest.is_ssa);
 685    assert(nir_tex_instr_dest_size(tex) == 4);
 686    assert(nir_alu_type_get_base_type(tex->dest_type) == nir_type_float);
 687
 688    b->cursor = nir_after_instr(&tex->instr);
 689
 690    static const unsigned swiz[4] = {0, 1, 2, 0};
 691    nir_ssa_def *comp = nir_swizzle(b, &tex->dest.ssa, swiz, 3, true);
 692
 693    /* Formula is:
 694     *    (comp <= 0.04045) ?
 695     *          (comp / 12.92) :
 696     *          pow((comp + 0.055) / 1.055, 2.4)
 697     */
 698    nir_ssa_def *low  = nir_fmul(b, comp, nir_imm_float(b, 1.0 / 12.92));
 699    nir_ssa_def *high = nir_fpow(b,
 700                                 nir_fmul(b,
 701                                          nir_fadd(b,
 702                                                   comp,
 703                                                   nir_imm_float(b, 0.055)),
 704                                          nir_imm_float(b, 1.0 / 1.055)),
 705                                 nir_imm_float(b, 2.4));
 706    nir_ssa_def *cond = nir_fge(b, nir_imm_float(b, 0.04045), comp);
 707    nir_ssa_def *rgb  = nir_bcsel(b, cond, low, high);
 708
 709    /* alpha is untouched: */
 710    nir_ssa_def *result = nir_vec4(b,
 711                                   nir_channel(b, rgb, 0),
 712                                   nir_channel(b, rgb, 1),
 713                                   nir_channel(b, rgb, 2),
 714                                   nir_channel(b, &tex->dest.ssa, 3));
 715
 716    nir_ssa_def_rewrite_uses_after(&tex->dest.ssa, nir_src_for_ssa(result),
 717                                   result->parent_instr);
 718 }
 719
 720 static bool
 721 nir_lower_tex_block(nir_block *block, nir_builder *b,
 722                     const nir_lower_tex_options *options)
 723 {
 724    bool progress = false;
 725
 726    nir_foreach_instr_safe(instr, block) {
 727       if (instr->type != nir_instr_type_tex)
 728          continue;
 729
 730       nir_tex_instr *tex = nir_instr_as_tex(instr);
 731       bool lower_txp = !!(options->lower_txp & (1 << tex->sampler_dim));
 732
 733       /* mask of src coords to saturate (clamp): */
 734       unsigned sat_mask = 0;
 735
 736       if ((1 << tex->sampler_index) & options->saturate_r)
 737          sat_mask |= (1 << 2);    /* .z */
 738       if ((1 << tex->sampler_index) & options->saturate_t)
 739          sat_mask |= (1 << 1);    /* .y */
 740       if ((1 << tex->sampler_index) & options->saturate_s)
 741          sat_mask |= (1 << 0);    /* .x */
 742
 743       /* If we are clamping any coords, we must lower projector first
 744        * as clamping happens *after* projection:
 745        */
 746       if (lower_txp || sat_mask) {
 747          project_src(b, tex);
 748          progress = true;
 749       }
 750
 751       if ((tex->op == nir_texop_txf && options->lower_txf_offset) ||
 752           (tex->sampler_dim == GLSL_SAMPLER_DIM_RECT &&
 753            options->lower_rect_offset)) {
 754          progress = lower_offset(b, tex) || progress;
 755       }
 756
 757       if ((tex->sampler_dim == GLSL_SAMPLER_DIM_RECT) && options->lower_rect) {
 758          lower_rect(b, tex);
 759          progress = true;
 760       }
 761
 762       if ((1 << tex->texture_index) & options->lower_y_uv_external) {
 763          lower_y_uv_external(b, tex);
 764          progress = true;
 765       }
 766
 767       if ((1 << tex->texture_index) & options->lower_y_u_v_external) {
 768          lower_y_u_v_external(b, tex);
 769          progress = true;
 770       }
 771
 772       if ((1 << tex->texture_index) & options->lower_yx_xuxv_external) {
 773          lower_yx_xuxv_external(b, tex);
 774          progress = true;
 775       }
 776
 777       if ((1 << tex->texture_index) & options->lower_xy_uxvx_external) {
 778          lower_xy_uxvx_external(b, tex);
 779          progress = true;
 780       }
 781
 782       if (sat_mask) {
 783          saturate_src(b, tex, sat_mask);
 784          progress = true;
 785       }
 786
 787       if (((1 << tex->texture_index) & options->swizzle_result) &&
 788           !nir_tex_instr_is_query(tex) &&
 789           !(tex->is_shadow && tex->is_new_style_shadow)) {
 790          swizzle_result(b, tex, options->swizzles[tex->texture_index]);
 791          progress = true;
 792       }
 793
 794       /* should be after swizzle so we know which channels are rgb: */
 795       if (((1 << tex->texture_index) & options->lower_srgb) &&
 796           !nir_tex_instr_is_query(tex) && !tex->is_shadow) {
 797          linearize_srgb_result(b, tex);
 798          progress = true;
 799       }
 800
 801       if (tex->op == nir_texop_txd &&
 802           tex->sampler_dim == GLSL_SAMPLER_DIM_CUBE &&
 803           (options->lower_txd_cube_map ||
 804            (tex->is_shadow && options->lower_txd_shadow))) {
 805          lower_gradient_cube_map(b, tex);
 806          progress = true;
 807          continue;
 808       }
 809
 810       if (tex->op == nir_texop_txd && options->lower_txd_shadow &&
 811           tex->is_shadow && tex->sampler_dim != GLSL_SAMPLER_DIM_CUBE) {
 812          lower_gradient_shadow(b, tex);
 813          progress = true;
 814          continue;
 815       }
 816
 817       /* TXF, TXS and TXL require a LOD but not everything we implement using those
 818        * three opcodes provides one.  Provide a default LOD of 0.
 819        */
 820       if ((nir_tex_instr_src_index(tex, nir_tex_src_lod) == -1) &&
 821           (tex->op == nir_texop_txf || tex->op == nir_texop_txs ||
 822            tex->op == nir_texop_txl || tex->op == nir_texop_query_levels ||
 823            (tex->op == nir_texop_tex &&
 824             b->shader->info.stage != MESA_SHADER_FRAGMENT))) {
 825          b->cursor = nir_before_instr(&tex->instr);
 826          nir_tex_instr_add_src(tex, nir_tex_src_lod, nir_src_for_ssa(nir_imm_int(b, 0)));
 827          progress = true;
 828          continue;
 829       }
 830    }
 831
 832    return progress;
 833 }
 834
 835 static bool
 836 nir_lower_tex_impl(nir_function_impl *impl,
 837                    const nir_lower_tex_options *options)
 838 {
 839    bool progress = false;
 840    nir_builder builder;
 841    nir_builder_init(&builder, impl);
 842
 843    nir_foreach_block(block, impl) {
 844       progress |= nir_lower_tex_block(block, &builder, options);
 845    }
 846
 847    nir_metadata_preserve(impl, nir_metadata_block_index |
 848                                nir_metadata_dominance);
 849    return progress;
 850 }
 851
 852 bool
 853 nir_lower_tex(nir_shader *shader, const nir_lower_tex_options *options)
 854 {
 855    bool progress = false;
 856
 857    nir_foreach_function(function, shader) {
 858       if (function->impl)
 859          progress |= nir_lower_tex_impl(function->impl, options);
 860    }
 861
 862    return progress;
 863 }