brw_nir_lower_image_load_store.c 31.8 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
/*
 * Copyright © 2018 Intel Corporation
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 */

#include "isl/isl.h"

#include "brw_nir.h"
#include "compiler/nir/nir_builder.h"
#include "compiler/nir/nir_format_convert.h"

/* The higher compiler layers use the GL enums for image formats even if
 * they come in from SPIR-V or Vulkan.  We need to turn them into an ISL
 * enum before we can use them.
 */
static enum isl_format
isl_format_for_gl_format(uint32_t gl_format)
{
   switch (gl_format) {
   case GL_R8:             return ISL_FORMAT_R8_UNORM;
   case GL_R8_SNORM:       return ISL_FORMAT_R8_SNORM;
   case GL_R8UI:           return ISL_FORMAT_R8_UINT;
   case GL_R8I:            return ISL_FORMAT_R8_SINT;
   case GL_RG8:            return ISL_FORMAT_R8G8_UNORM;
   case GL_RG8_SNORM:      return ISL_FORMAT_R8G8_SNORM;
   case GL_RG8UI:          return ISL_FORMAT_R8G8_UINT;
   case GL_RG8I:           return ISL_FORMAT_R8G8_SINT;
   case GL_RGBA8:          return ISL_FORMAT_R8G8B8A8_UNORM;
   case GL_RGBA8_SNORM:    return ISL_FORMAT_R8G8B8A8_SNORM;
   case GL_RGBA8UI:        return ISL_FORMAT_R8G8B8A8_UINT;
   case GL_RGBA8I:         return ISL_FORMAT_R8G8B8A8_SINT;
   case GL_R11F_G11F_B10F: return ISL_FORMAT_R11G11B10_FLOAT;
   case GL_RGB10_A2:       return ISL_FORMAT_R10G10B10A2_UNORM;
   case GL_RGB10_A2UI:     return ISL_FORMAT_R10G10B10A2_UINT;
   case GL_R16:            return ISL_FORMAT_R16_UNORM;
   case GL_R16_SNORM:      return ISL_FORMAT_R16_SNORM;
   case GL_R16F:           return ISL_FORMAT_R16_FLOAT;
   case GL_R16UI:          return ISL_FORMAT_R16_UINT;
   case GL_R16I:           return ISL_FORMAT_R16_SINT;
   case GL_RG16:           return ISL_FORMAT_R16G16_UNORM;
   case GL_RG16_SNORM:     return ISL_FORMAT_R16G16_SNORM;
   case GL_RG16F:          return ISL_FORMAT_R16G16_FLOAT;
   case GL_RG16UI:         return ISL_FORMAT_R16G16_UINT;
   case GL_RG16I:          return ISL_FORMAT_R16G16_SINT;
   case GL_RGBA16:         return ISL_FORMAT_R16G16B16A16_UNORM;
   case GL_RGBA16_SNORM:   return ISL_FORMAT_R16G16B16A16_SNORM;
   case GL_RGBA16F:        return ISL_FORMAT_R16G16B16A16_FLOAT;
   case GL_RGBA16UI:       return ISL_FORMAT_R16G16B16A16_UINT;
   case GL_RGBA16I:        return ISL_FORMAT_R16G16B16A16_SINT;
   case GL_R32F:           return ISL_FORMAT_R32_FLOAT;
   case GL_R32UI:          return ISL_FORMAT_R32_UINT;
   case GL_R32I:           return ISL_FORMAT_R32_SINT;
   case GL_RG32F:          return ISL_FORMAT_R32G32_FLOAT;
   case GL_RG32UI:         return ISL_FORMAT_R32G32_UINT;
   case GL_RG32I:          return ISL_FORMAT_R32G32_SINT;
   case GL_RGBA32F:        return ISL_FORMAT_R32G32B32A32_FLOAT;
   case GL_RGBA32UI:       return ISL_FORMAT_R32G32B32A32_UINT;
   case GL_RGBA32I:        return ISL_FORMAT_R32G32B32A32_SINT;
   case GL_NONE:           return ISL_FORMAT_UNSUPPORTED;
   default:
      assert(!"Invalid image format");
      return ISL_FORMAT_UNSUPPORTED;
   }
}

static nir_ssa_def *
_load_image_param(nir_builder *b, nir_deref_instr *deref, unsigned offset)
{
   nir_intrinsic_instr *load =
      nir_intrinsic_instr_create(b->shader,
                                 nir_intrinsic_image_deref_load_param_intel);
   load->src[0] = nir_src_for_ssa(&deref->dest.ssa);
   nir_intrinsic_set_base(load, offset / 4);

   switch (offset) {
   case BRW_IMAGE_PARAM_OFFSET_OFFSET:
   case BRW_IMAGE_PARAM_SWIZZLING_OFFSET:
      load->num_components = 2;
      break;
   case BRW_IMAGE_PARAM_TILING_OFFSET:
   case BRW_IMAGE_PARAM_SIZE_OFFSET:
      load->num_components = 3;
      break;
   case BRW_IMAGE_PARAM_STRIDE_OFFSET:
      load->num_components = 4;
      break;
   default:
      unreachable("Invalid param offset");
   }
   nir_ssa_dest_init(&load->instr, &load->dest,
                     load->num_components, 32, NULL);

   nir_builder_instr_insert(b, &load->instr);
   return &load->dest.ssa;
}

#define load_image_param(b, d, o) \
   _load_image_param(b, d, BRW_IMAGE_PARAM_##o##_OFFSET)

static nir_ssa_def *
image_coord_is_in_bounds(nir_builder *b, nir_deref_instr *deref,
                         nir_ssa_def *coord)
{
   nir_ssa_def *size = load_image_param(b, deref, SIZE);
   nir_ssa_def *cmp = nir_ilt(b, coord, size);
124
125

   unsigned coord_comps = glsl_get_sampler_coordinate_components(deref->type);
126
   nir_ssa_def *in_bounds = nir_imm_true(b);
127
   for (unsigned i = 0; i < coord_comps; i++)
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
      in_bounds = nir_iand(b, in_bounds, nir_channel(b, cmp, i));

   return in_bounds;
}

/** Calculate the offset in memory of the texel given by \p coord.
 *
 * This is meant to be used with untyped surface messages to access a tiled
 * surface, what involves taking into account the tiling and swizzling modes
 * of the surface manually so it will hopefully not happen very often.
 *
 * The tiling algorithm implemented here matches either the X or Y tiling
 * layouts supported by the hardware depending on the tiling coefficients
 * passed to the program as uniforms.  See Volume 1 Part 2 Section 4.5
 * "Address Tiling Function" of the IVB PRM for an in-depth explanation of
 * the hardware tiling format.
 */
static nir_ssa_def *
image_address(nir_builder *b, const struct gen_device_info *devinfo,
              nir_deref_instr *deref, nir_ssa_def *coord)
{
149
150
151
152
153
154
155
156
157
158
   if (glsl_get_sampler_dim(deref->type) == GLSL_SAMPLER_DIM_1D &&
       glsl_sampler_type_is_array(deref->type)) {
      /* It's easier if 1D arrays are treated like 2D arrays */
      coord = nir_vec3(b, nir_channel(b, coord, 0),
                          nir_imm_int(b, 0),
                          nir_channel(b, coord, 1));
   } else {
      unsigned dims = glsl_get_sampler_coordinate_components(deref->type);
      coord = nir_channels(b, coord, (1 << dims) - 1);
   }
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557

   nir_ssa_def *offset = load_image_param(b, deref, OFFSET);
   nir_ssa_def *tiling = load_image_param(b, deref, TILING);
   nir_ssa_def *stride = load_image_param(b, deref, STRIDE);

   /* Shift the coordinates by the fixed surface offset.  It may be non-zero
    * if the image is a single slice of a higher-dimensional surface, or if a
    * non-zero mipmap level of the surface is bound to the pipeline.  The
    * offset needs to be applied here rather than at surface state set-up time
    * because the desired slice-level may start mid-tile, so simply shifting
    * the surface base address wouldn't give a well-formed tiled surface in
    * the general case.
    */
   nir_ssa_def *xypos = (coord->num_components == 1) ?
                        nir_vec2(b, coord, nir_imm_int(b, 0)) :
                        nir_channels(b, coord, 0x3);
   xypos = nir_iadd(b, xypos, offset);

   /* The layout of 3-D textures in memory is sort-of like a tiling
    * format.  At each miplevel, the slices are arranged in rows of
    * 2^level slices per row.  The slice row is stored in tmp.y and
    * the slice within the row is stored in tmp.x.
    *
    * The layout of 2-D array textures and cubemaps is much simpler:
    * Depending on whether the ARYSPC_LOD0 layout is in use it will be
    * stored in memory as an array of slices, each one being a 2-D
    * arrangement of miplevels, or as a 2D arrangement of miplevels,
    * each one being an array of slices.  In either case the separation
    * between slices of the same LOD is equal to the qpitch value
    * provided as stride.w.
    *
    * This code can be made to handle either 2D arrays and 3D textures
    * by passing in the miplevel as tile.z for 3-D textures and 0 in
    * tile.z for 2-D array textures.
    *
    * See Volume 1 Part 1 of the Gen7 PRM, sections 6.18.4.7 "Surface
    * Arrays" and 6.18.6 "3D Surfaces" for a more extensive discussion
    * of the hardware 3D texture and 2D array layouts.
    */
   if (coord->num_components > 2) {
      /* Decompose z into a major (tmp.y) and a minor (tmp.x)
       * index.
       */
      nir_ssa_def *z = nir_channel(b, coord, 2);
      nir_ssa_def *z_x = nir_ubfe(b, z, nir_imm_int(b, 0),
                                  nir_channel(b, tiling, 2));
      nir_ssa_def *z_y = nir_ushr(b, z, nir_channel(b, tiling, 2));

      /* Take into account the horizontal (tmp.x) and vertical (tmp.y)
       * slice offset.
       */
      xypos = nir_iadd(b, xypos, nir_imul(b, nir_vec2(b, z_x, z_y),
                                             nir_channels(b, stride, 0xc)));
   }

   nir_ssa_def *addr;
   if (coord->num_components > 1) {
      /* Calculate the major/minor x and y indices.  In order to
       * accommodate both X and Y tiling, the Y-major tiling format is
       * treated as being a bunch of narrow X-tiles placed next to each
       * other.  This means that the tile width for Y-tiling is actually
       * the width of one sub-column of the Y-major tile where each 4K
       * tile has 8 512B sub-columns.
       *
       * The major Y value is the row of tiles in which the pixel lives.
       * The major X value is the tile sub-column in which the pixel
       * lives; for X tiling, this is the same as the tile column, for Y
       * tiling, each tile has 8 sub-columns.  The minor X and Y indices
       * are the position within the sub-column.
       */

      /* Calculate the minor x and y indices. */
      nir_ssa_def *minor = nir_ubfe(b, xypos, nir_imm_int(b, 0),
                                       nir_channels(b, tiling, 0x3));
      nir_ssa_def *major = nir_ushr(b, xypos, nir_channels(b, tiling, 0x3));

      /* Calculate the texel index from the start of the tile row and the
       * vertical coordinate of the row.
       * Equivalent to:
       *   tmp.x = (major.x << tile.y << tile.x) +
       *           (minor.y << tile.x) + minor.x
       *   tmp.y = major.y << tile.y
       */
      nir_ssa_def *idx_x, *idx_y;
      idx_x = nir_ishl(b, nir_channel(b, major, 0), nir_channel(b, tiling, 1));
      idx_x = nir_iadd(b, idx_x, nir_channel(b, minor, 1));
      idx_x = nir_ishl(b, idx_x, nir_channel(b, tiling, 0));
      idx_x = nir_iadd(b, idx_x, nir_channel(b, minor, 0));
      idx_y = nir_ishl(b, nir_channel(b, major, 1), nir_channel(b, tiling, 1));

      /* Add it to the start of the tile row. */
      nir_ssa_def *idx;
      idx = nir_imul(b, idx_y, nir_channel(b, stride, 1));
      idx = nir_iadd(b, idx, idx_x);

      /* Multiply by the Bpp value. */
      addr = nir_imul(b, idx, nir_channel(b, stride, 0));

      if (devinfo->gen < 8 && !devinfo->is_baytrail) {
         /* Take into account the two dynamically specified shifts.  Both are
          * used to implement swizzling of X-tiled surfaces.  For Y-tiled
          * surfaces only one bit needs to be XOR-ed with bit 6 of the memory
          * address, so a swz value of 0xff (actually interpreted as 31 by the
          * hardware) will be provided to cause the relevant bit of tmp.y to
          * be zero and turn the first XOR into the identity.  For linear
          * surfaces or platforms lacking address swizzling both shifts will
          * be 0xff causing the relevant bits of both tmp.x and .y to be zero,
          * what effectively disables swizzling.
          */
         nir_ssa_def *swizzle = load_image_param(b, deref, SWIZZLING);
         nir_ssa_def *shift0 = nir_ushr(b, addr, nir_channel(b, swizzle, 0));
         nir_ssa_def *shift1 = nir_ushr(b, addr, nir_channel(b, swizzle, 1));

         /* XOR tmp.x and tmp.y with bit 6 of the memory address. */
         nir_ssa_def *bit = nir_iand(b, nir_ixor(b, shift0, shift1),
                                        nir_imm_int(b, 1 << 6));
         addr = nir_ixor(b, addr, bit);
      }
   } else {
      /* Multiply by the Bpp/stride value.  Note that the addr.y may be
       * non-zero even if the image is one-dimensional because a vertical
       * offset may have been applied above to select a non-zero slice or
       * level of a higher-dimensional texture.
       */
      nir_ssa_def *idx;
      idx = nir_imul(b, nir_channel(b, xypos, 1), nir_channel(b, stride, 1));
      idx = nir_iadd(b, nir_channel(b, xypos, 0), idx);
      addr = nir_imul(b, idx, nir_channel(b, stride, 0));
   }

   return addr;
}

struct format_info {
   const struct isl_format_layout *fmtl;
   unsigned chans;
   unsigned bits[4];
};

static struct format_info
get_format_info(enum isl_format fmt)
{
   const struct isl_format_layout *fmtl = isl_format_get_layout(fmt);

   return (struct format_info) {
      .fmtl = fmtl,
      .chans = isl_format_get_num_channels(fmt),
      .bits = {
         fmtl->channels.r.bits,
         fmtl->channels.g.bits,
         fmtl->channels.b.bits,
         fmtl->channels.a.bits
      },
   };
}

static nir_ssa_def *
nir_zero_vec(nir_builder *b, unsigned num_components)
{
   nir_const_value v;
   memset(&v, 0, sizeof(v));

   return nir_build_imm(b, num_components, 32, v);
}

static nir_ssa_def *
convert_color_for_load(nir_builder *b, const struct gen_device_info *devinfo,
                       nir_ssa_def *color,
                       enum isl_format image_fmt, enum isl_format lower_fmt,
                       unsigned dest_components)
{
   if (image_fmt == lower_fmt)
      goto expand_vec;

   if (image_fmt == ISL_FORMAT_R11G11B10_FLOAT) {
      assert(lower_fmt == ISL_FORMAT_R32_UINT);
      color = nir_format_unpack_11f11f10f(b, color);
      goto expand_vec;
   }

   struct format_info image = get_format_info(image_fmt);
   struct format_info lower = get_format_info(lower_fmt);

   const bool needs_sign_extension =
      isl_format_has_snorm_channel(image_fmt) ||
      isl_format_has_sint_channel(image_fmt);

   /* We only check the red channel to detect if we need to pack/unpack */
   assert(image.bits[0] != lower.bits[0] ||
          memcmp(image.bits, lower.bits, sizeof(image.bits)) == 0);

   if (image.bits[0] != lower.bits[0] && lower_fmt == ISL_FORMAT_R32_UINT) {
      if (needs_sign_extension)
         color = nir_format_unpack_sint(b, color, image.bits, image.chans);
      else
         color = nir_format_unpack_uint(b, color, image.bits, image.chans);
   } else {
      /* All these formats are homogeneous */
      for (unsigned i = 1; i < image.chans; i++)
         assert(image.bits[i] == image.bits[0]);

      /* On IVB, we rely on the undocumented behavior that typed reads from
       * surfaces of the unsupported R8 and R16 formats return useful data in
       * their least significant bits.  However, the data in the high bits is
       * garbage so we have to discard it.
       */
      if (devinfo->gen == 7 && !devinfo->is_haswell &&
          (lower_fmt == ISL_FORMAT_R16_UINT ||
           lower_fmt == ISL_FORMAT_R8_UINT))
         color = nir_format_mask_uvec(b, color, lower.bits);

      if (image.bits[0] != lower.bits[0]) {
         color = nir_format_bitcast_uvec_unmasked(b, color, lower.bits[0],
                                                  image.bits[0]);
      }

      if (needs_sign_extension)
         color = nir_format_sign_extend_ivec(b, color, image.bits);
   }

   switch (image.fmtl->channels.r.type) {
   case ISL_UNORM:
      assert(isl_format_has_uint_channel(lower_fmt));
      color = nir_format_unorm_to_float(b, color, image.bits);
      break;

   case ISL_SNORM:
      assert(isl_format_has_uint_channel(lower_fmt));
      color = nir_format_snorm_to_float(b, color, image.bits);
      break;

   case ISL_SFLOAT:
      if (image.bits[0] == 16)
         color = nir_unpack_half_2x16_split_x(b, color);
      break;

   case ISL_UINT:
   case ISL_SINT:
      break;

   default:
      unreachable("Invalid image channel type");
   }

expand_vec:
   assert(dest_components == 1 || dest_components == 4);
   assert(color->num_components <= dest_components);
   if (color->num_components == dest_components)
      return color;

   nir_ssa_def *comps[4];
   for (unsigned i = 0; i < color->num_components; i++)
      comps[i] = nir_channel(b, color, i);

   for (unsigned i = color->num_components; i < 3; i++)
      comps[i] = nir_imm_int(b, 0);

   if (color->num_components < 4) {
      if (isl_format_has_int_channel(image_fmt))
         comps[3] = nir_imm_int(b, 1);
      else
         comps[3] = nir_imm_float(b, 1);
   }

   return nir_vec(b, comps, dest_components);
}

static bool
lower_image_load_instr(nir_builder *b,
                       const struct gen_device_info *devinfo,
                       nir_intrinsic_instr *intrin)
{
   nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
   nir_variable *var = nir_deref_instr_get_variable(deref);
   const enum isl_format image_fmt =
      isl_format_for_gl_format(var->data.image.format);

   if (isl_has_matching_typed_storage_image_format(devinfo, image_fmt)) {
      const enum isl_format lower_fmt =
         isl_lower_storage_image_format(devinfo, image_fmt);
      const unsigned dest_components = intrin->num_components;

      /* Use an undef to hold the uses of the load while we do the color
       * conversion.
       */
      nir_ssa_def *placeholder = nir_ssa_undef(b, 4, 32);
      nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(placeholder));

      intrin->num_components = isl_format_get_num_channels(lower_fmt);
      intrin->dest.ssa.num_components = intrin->num_components;

      b->cursor = nir_after_instr(&intrin->instr);

      nir_ssa_def *color = convert_color_for_load(b, devinfo,
                                                  &intrin->dest.ssa,
                                                  image_fmt, lower_fmt,
                                                  dest_components);

      nir_ssa_def_rewrite_uses(placeholder, nir_src_for_ssa(color));
      nir_instr_remove(placeholder->parent_instr);
   } else {
      const struct isl_format_layout *image_fmtl =
         isl_format_get_layout(image_fmt);
      /* We have a matching typed format for everything 32b and below */
      assert(image_fmtl->bpb == 64 || image_fmtl->bpb == 128);
      enum isl_format raw_fmt = (image_fmtl->bpb == 64) ?
                                ISL_FORMAT_R32G32_UINT :
                                ISL_FORMAT_R32G32B32A32_UINT;
      const unsigned dest_components = intrin->num_components;

      b->cursor = nir_instr_remove(&intrin->instr);

      nir_ssa_def *coord = intrin->src[1].ssa;

      nir_ssa_def *do_load = image_coord_is_in_bounds(b, deref, coord);
      if (devinfo->gen == 7 && !devinfo->is_haswell) {
         /* Check whether the first stride component (i.e. the Bpp value)
          * is greater than four, what on Gen7 indicates that a surface of
          * type RAW has been bound for untyped access.  Reading or writing
          * to a surface of type other than RAW using untyped surface
          * messages causes a hang on IVB and VLV.
          */
         nir_ssa_def *stride = load_image_param(b, deref, STRIDE);
         nir_ssa_def *is_raw =
            nir_ilt(b, nir_imm_int(b, 4), nir_channel(b, stride, 0));
         do_load = nir_iand(b, do_load, is_raw);
      }
      nir_push_if(b, do_load);

      nir_ssa_def *addr = image_address(b, devinfo, deref, coord);
      nir_intrinsic_instr *load =
         nir_intrinsic_instr_create(b->shader,
                                    nir_intrinsic_image_deref_load_raw_intel);
      load->src[0] = nir_src_for_ssa(&deref->dest.ssa);
      load->src[1] = nir_src_for_ssa(addr);
      load->num_components = image_fmtl->bpb / 32;
      nir_ssa_dest_init(&load->instr, &load->dest,
                        load->num_components, 32, NULL);
      nir_builder_instr_insert(b, &load->instr);

      nir_push_else(b, NULL);

      nir_ssa_def *zero = nir_zero_vec(b, load->num_components);

      nir_pop_if(b, NULL);

      nir_ssa_def *value = nir_if_phi(b, &load->dest.ssa, zero);

      nir_ssa_def *color = convert_color_for_load(b, devinfo, value,
                                                  image_fmt, raw_fmt,
                                                  dest_components);

      nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(color));
   }

   return true;
}

static nir_ssa_def *
convert_color_for_store(nir_builder *b, const struct gen_device_info *devinfo,
                        nir_ssa_def *color,
                        enum isl_format image_fmt, enum isl_format lower_fmt)
{
   struct format_info image = get_format_info(image_fmt);
   struct format_info lower = get_format_info(lower_fmt);

   color = nir_channels(b, color, (1 << image.chans) - 1);

   if (image_fmt == lower_fmt)
      return color;

   if (image_fmt == ISL_FORMAT_R11G11B10_FLOAT) {
      assert(lower_fmt == ISL_FORMAT_R32_UINT);
      return nir_format_pack_11f11f10f(b, color);
   }

   switch (image.fmtl->channels.r.type) {
   case ISL_UNORM:
      assert(isl_format_has_uint_channel(lower_fmt));
      color = nir_format_float_to_unorm(b, color, image.bits);
      break;

   case ISL_SNORM:
      assert(isl_format_has_uint_channel(lower_fmt));
      color = nir_format_float_to_snorm(b, color, image.bits);
      break;

   case ISL_SFLOAT:
      if (image.bits[0] == 16) {
         nir_ssa_def *f16comps[4];
         for (unsigned i = 0; i < image.chans; i++) {
            f16comps[i] = nir_pack_half_2x16_split(b, nir_channel(b, color, i),
                                                      nir_imm_float(b, 0));
         }
         color = nir_vec(b, f16comps, image.chans);
      }
      break;

   case ISL_UINT:
558
      color = nir_format_clamp_uint(b, color, image.bits);
559
560
561
      break;

   case ISL_SINT:
562
      color = nir_format_clamp_sint(b, color, image.bits);
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
      break;

   default:
      unreachable("Invalid image channel type");
   }

   if (image.bits[0] < 32 &&
       (isl_format_has_snorm_channel(image_fmt) ||
        isl_format_has_sint_channel(image_fmt)))
      color = nir_format_mask_uvec(b, color, image.bits);

   if (image.bits[0] != lower.bits[0] && lower_fmt == ISL_FORMAT_R32_UINT) {
      color = nir_format_pack_uint(b, color, image.bits, image.chans);
   } else {
      /* All these formats are homogeneous */
      for (unsigned i = 1; i < image.chans; i++)
         assert(image.bits[i] == image.bits[0]);

      if (image.bits[0] != lower.bits[0]) {
         color = nir_format_bitcast_uvec_unmasked(b, color, image.bits[0],
                                                  lower.bits[0]);
      }
   }

   return color;
}

static bool
lower_image_store_instr(nir_builder *b,
                        const struct gen_device_info *devinfo,
                        nir_intrinsic_instr *intrin)
{
   nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
   nir_variable *var = nir_deref_instr_get_variable(deref);

   /* For write-only surfaces, we trust that the hardware can just do the
    * conversion for us.
    */
601
   if (var->data.image.access & ACCESS_NON_READABLE)
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
      return false;

   const enum isl_format image_fmt =
      isl_format_for_gl_format(var->data.image.format);

   if (isl_has_matching_typed_storage_image_format(devinfo, image_fmt)) {
      const enum isl_format lower_fmt =
         isl_lower_storage_image_format(devinfo, image_fmt);

      /* Color conversion goes before the store */
      b->cursor = nir_before_instr(&intrin->instr);

      nir_ssa_def *color = convert_color_for_store(b, devinfo,
                                                   intrin->src[3].ssa,
                                                   image_fmt, lower_fmt);
      intrin->num_components = isl_format_get_num_channels(lower_fmt);
      nir_instr_rewrite_src(&intrin->instr, &intrin->src[3],
                            nir_src_for_ssa(color));
   } else {
      const struct isl_format_layout *image_fmtl =
         isl_format_get_layout(image_fmt);
      /* We have a matching typed format for everything 32b and below */
      assert(image_fmtl->bpb == 64 || image_fmtl->bpb == 128);
      enum isl_format raw_fmt = (image_fmtl->bpb == 64) ?
                                ISL_FORMAT_R32G32_UINT :
                                ISL_FORMAT_R32G32B32A32_UINT;

      b->cursor = nir_instr_remove(&intrin->instr);

      nir_ssa_def *coord = intrin->src[1].ssa;

      nir_ssa_def *do_store = image_coord_is_in_bounds(b, deref, coord);
      if (devinfo->gen == 7 && !devinfo->is_haswell) {
         /* Check whether the first stride component (i.e. the Bpp value)
          * is greater than four, what on Gen7 indicates that a surface of
          * type RAW has been bound for untyped access.  Reading or writing
          * to a surface of type other than RAW using untyped surface
          * messages causes a hang on IVB and VLV.
          */
         nir_ssa_def *stride = load_image_param(b, deref, STRIDE);
         nir_ssa_def *is_raw =
            nir_ilt(b, nir_imm_int(b, 4), nir_channel(b, stride, 0));
         do_store = nir_iand(b, do_store, is_raw);
      }
      nir_push_if(b, do_store);

      nir_ssa_def *addr = image_address(b, devinfo, deref, coord);
      nir_ssa_def *color = convert_color_for_store(b, devinfo,
                                                   intrin->src[3].ssa,
                                                   image_fmt, raw_fmt);

      nir_intrinsic_instr *store =
         nir_intrinsic_instr_create(b->shader,
                                    nir_intrinsic_image_deref_store_raw_intel);
      store->src[0] = nir_src_for_ssa(&deref->dest.ssa);
      store->src[1] = nir_src_for_ssa(addr);
      store->src[2] = nir_src_for_ssa(color);
      store->num_components = image_fmtl->bpb / 32;
      nir_builder_instr_insert(b, &store->instr);

      nir_pop_if(b, NULL);
   }

   return true;
}

static bool
lower_image_atomic_instr(nir_builder *b,
                         const struct gen_device_info *devinfo,
                         nir_intrinsic_instr *intrin)
{
   if (devinfo->is_haswell || devinfo->gen >= 8)
      return false;

   nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);

   b->cursor = nir_instr_remove(&intrin->instr);

   /* Use an undef to hold the uses of the load conversion. */
   nir_ssa_def *placeholder = nir_ssa_undef(b, 4, 32);
   nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(placeholder));

   /* Check the first component of the size field to find out if the
    * image is bound.  Necessary on IVB for typed atomics because
    * they don't seem to respect null surfaces and will happily
    * corrupt or read random memory when no image is bound.
    */
   nir_ssa_def *size = load_image_param(b, deref, SIZE);
   nir_ssa_def *zero = nir_imm_int(b, 0);
   nir_push_if(b, nir_ine(b, nir_channel(b, size, 0), zero));

   nir_builder_instr_insert(b, &intrin->instr);

   nir_pop_if(b, NULL);

   nir_ssa_def *result = nir_if_phi(b, &intrin->dest.ssa, zero);
   nir_ssa_def_rewrite_uses(placeholder, nir_src_for_ssa(result));

   return true;
}

static bool
lower_image_size_instr(nir_builder *b,
                       const struct gen_device_info *devinfo,
                       nir_intrinsic_instr *intrin)
{
   nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
   nir_variable *var = nir_deref_instr_get_variable(deref);

   /* For write-only images, we have an actual image surface so we fall back
    * and let the back-end emit a TXS for this.
    */
   if (var->data.image.access & ACCESS_NON_READABLE)
      return false;

   /* If we have a matching typed format, then we have an actual image surface
    * so we fall back and let the back-end emit a TXS for this.
    */
   const enum isl_format image_fmt =
      isl_format_for_gl_format(var->data.image.format);
   if (isl_has_matching_typed_storage_image_format(devinfo, image_fmt))
      return false;
724
725
726
727
728
729
730
731
732
733

   b->cursor = nir_instr_remove(&intrin->instr);

   nir_ssa_def *size = load_image_param(b, deref, SIZE);

   nir_ssa_def *comps[4] = { NULL, NULL, NULL, NULL };

   enum glsl_sampler_dim dim = glsl_get_sampler_dim(deref->type);
   unsigned coord_comps = glsl_get_sampler_coordinate_components(deref->type);
   for (unsigned c = 0; c < coord_comps; c++) {
734
      if (c == 2 && dim == GLSL_SAMPLER_DIM_CUBE) {
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
         comps[2] = nir_idiv(b, nir_channel(b, size, 2), nir_imm_int(b, 6));
      } else {
         comps[c] = nir_channel(b, size, c);
      }
   }

   for (unsigned c = coord_comps; c < intrin->dest.ssa.num_components; ++c)
      comps[c] = nir_imm_int(b, 1);

   nir_ssa_def *vec = nir_vec(b, comps, intrin->dest.ssa.num_components);
   nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(vec));

   return true;
}

bool
brw_nir_lower_image_load_store(nir_shader *shader,
                               const struct gen_device_info *devinfo)
{
   bool progress = false;

   nir_foreach_function(function, shader) {
      if (function->impl == NULL)
         continue;

      nir_foreach_block_safe(block, function->impl) {
         nir_builder b;
         nir_builder_init(&b, function->impl);

         nir_foreach_instr_safe(instr, block) {
            if (instr->type != nir_instr_type_intrinsic)
               continue;

            nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
            switch (intrin->intrinsic) {
            case nir_intrinsic_image_deref_load:
               if (lower_image_load_instr(&b, devinfo, intrin))
                  progress = true;
               break;

            case nir_intrinsic_image_deref_store:
               if (lower_image_store_instr(&b, devinfo, intrin))
                  progress = true;
               break;

            case nir_intrinsic_image_deref_atomic_add:
            case nir_intrinsic_image_deref_atomic_min:
            case nir_intrinsic_image_deref_atomic_max:
            case nir_intrinsic_image_deref_atomic_and:
            case nir_intrinsic_image_deref_atomic_or:
            case nir_intrinsic_image_deref_atomic_xor:
            case nir_intrinsic_image_deref_atomic_exchange:
            case nir_intrinsic_image_deref_atomic_comp_swap:
               if (lower_image_atomic_instr(&b, devinfo, intrin))
                  progress = true;
               break;

            case nir_intrinsic_image_deref_size:
               if (lower_image_size_instr(&b, devinfo, intrin))
                  progress = true;
               break;

            default:
               /* Nothing to do */
               break;
            }
         }
      }

804
805
      if (progress)
         nir_metadata_preserve(function->impl, nir_metadata_none);
806
807
808
809
   }

   return progress;
}
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850

void
brw_nir_rewrite_image_intrinsic(nir_intrinsic_instr *intrin,
                                nir_ssa_def *index)
{
   nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
   nir_variable *var = nir_deref_instr_get_variable(deref);

   switch (intrin->intrinsic) {
#define CASE(op) \
   case nir_intrinsic_image_deref_##op: \
      intrin->intrinsic = nir_intrinsic_image_##op; \
      break;
   CASE(load)
   CASE(store)
   CASE(atomic_add)
   CASE(atomic_min)
   CASE(atomic_max)
   CASE(atomic_and)
   CASE(atomic_or)
   CASE(atomic_xor)
   CASE(atomic_exchange)
   CASE(atomic_comp_swap)
   CASE(atomic_fadd)
   CASE(size)
   CASE(samples)
   CASE(load_raw_intel)
   CASE(store_raw_intel)
#undef CASE
   default:
      unreachable("Unhanded image intrinsic");
   }

   nir_intrinsic_set_image_dim(intrin, glsl_get_sampler_dim(deref->type));
   nir_intrinsic_set_image_array(intrin, glsl_sampler_type_is_array(deref->type));
   nir_intrinsic_set_access(intrin, var->data.image.access);
   nir_intrinsic_set_format(intrin, var->data.image.format);

   nir_instr_rewrite_src(&intrin->instr, &intrin->src[0],
                         nir_src_for_ssa(index));
}