src/gallium/drivers/freedreno/a6xx/fd6_resource.c

   1 /*
   2  * Copyright (C) 2018 Rob Clark <robclark@freedesktop.org>
   3  * Copyright © 2018 Google, Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  *
  24  * Authors:
  25  *    Rob Clark <robclark@freedesktop.org>
  26  */
  27
  28 #include "fd6_resource.h"
  29 #include "fd6_format.h"
  30
  31 #include "a6xx.xml.h"
  32
  33 /* indexed by cpp, including msaa 2x and 4x: */
  34 static const struct {
  35         unsigned pitchalign;
  36         unsigned heightalign;
  37 } tile_alignment[] = {
  38         [1]  = { 128, 32 },
  39         [2]  = {  64, 32 },
  40         [3]  = {  64, 32 },
  41         [4]  = {  64, 16 },
  42         [6]  = {  64, 16 },
  43         [8]  = {  64, 16 },
  44         [12] = {  64, 16 },
  45         [16] = {  64, 16 },
  46         [24] = {  64, 16 },
  47         [32] = {  64, 16 },
  48         [48] = {  64, 16 },
  49         [64] = {  64, 16 },
  50
  51         /* special cases for r16: */
  52         [0]  = { 128, 16 },
  53 };
  54
  55 /* NOTE: good way to test this is:  (for example)
  56  *  piglit/bin/texelFetch fs sampler3D 100x100x8
  57  */
  58 static uint32_t
  59 setup_slices(struct fd_resource *rsc, uint32_t alignment, enum pipe_format format)
  60 {
  61         struct pipe_resource *prsc = &rsc->base;
  62         struct fd_screen *screen = fd_screen(prsc->screen);
  63         enum util_format_layout layout = util_format_description(format)->layout;
  64         uint32_t pitchalign = screen->gmem_alignw;
  65         uint32_t level, size = 0;
  66         uint32_t depth = prsc->depth0;
  67         /* linear dimensions: */
  68         uint32_t lwidth = prsc->width0;
  69         uint32_t lheight = prsc->height0;
  70         /* tile_mode dimensions: */
  71         uint32_t twidth = util_next_power_of_two(lwidth);
  72         uint32_t theight = util_next_power_of_two(lheight);
  73         /* in layer_first layout, the level (slice) contains just one
  74          * layer (since in fact the layer contains the slices)
  75          */
  76         uint32_t layers_in_level = rsc->layout.layer_first ? 1 : prsc->array_size;
  77         int ta = rsc->layout.cpp;
  78
  79         /* The z16/r16 formats seem to not play by the normal tiling rules: */
  80         if ((rsc->layout.cpp == 2) && (util_format_get_nr_components(format) == 1))
  81                 ta = 0;
  82
  83         debug_assert(ta < ARRAY_SIZE(tile_alignment));
  84         debug_assert(tile_alignment[ta].pitchalign);
  85
  86         for (level = 0; level <= prsc->last_level; level++) {
  87                 struct fdl_slice *slice = fd_resource_slice(rsc, level);
  88                 uint32_t tile_mode = fd_resource_tile_mode(prsc, level);
  89                 uint32_t width, height;
  90
  91                 /* tiled levels of 3D textures are rounded up to PoT dimensions: */
  92                 if ((prsc->target == PIPE_TEXTURE_3D) && tile_mode) {
  93                         width = twidth;
  94                         height = theight;
  95                 } else {
  96                         width = lwidth;
  97                         height = lheight;
  98                 }
  99                 uint32_t aligned_height = height;
 100                 uint32_t blocks;
 101
 102                 if (tile_mode) {
 103                         pitchalign = tile_alignment[ta].pitchalign;
 104                         aligned_height = align(aligned_height,
 105                                         tile_alignment[ta].heightalign);
 106                 } else {
 107                         pitchalign = 64;
 108                 }
 109
 110                 /* The blits used for mem<->gmem work at a granularity of
 111                  * 32x32, which can cause faults due to over-fetch on the
 112                  * last level.  The simple solution is to over-allocate a
 113                  * bit the last level to ensure any over-fetch is harmless.
 114                  * The pitch is already sufficiently aligned, but height
 115                  * may not be:
 116                  */
 117                 if ((level == prsc->last_level) && (prsc->target != PIPE_BUFFER))
 118                         aligned_height = align(aligned_height, 32);
 119
 120                 if (layout == UTIL_FORMAT_LAYOUT_ASTC)
 121                         slice->pitch =
 122                                 util_align_npot(width, pitchalign * util_format_get_blockwidth(format));
 123                 else
 124                         slice->pitch = align(width, pitchalign);
 125
 126                 slice->offset = size;
 127                 blocks = util_format_get_nblocks(format, slice->pitch, aligned_height);
 128
 129                 /* 1d array and 2d array textures must all have the same layer size
 130                  * for each miplevel on a6xx. 3d textures can have different layer
 131                  * sizes for high levels, but the hw auto-sizer is buggy (or at least
 132                  * different than what this code does), so as soon as the layer size
 133                  * range gets into range, we stop reducing it.
 134                  */
 135                 if (prsc->target == PIPE_TEXTURE_3D) {
 136                         if (level < 1 || fd_resource_slice(rsc, level - 1)->size0 > 0xf000) {
 137                                 slice->size0 = align(blocks * rsc->layout.cpp, alignment);
 138                         } else {
 139                                 slice->size0 = fd_resource_slice(rsc, level - 1)->size0;
 140                         }
 141                 } else {
 142                         slice->size0 = align(blocks * rsc->layout.cpp, alignment);
 143                 }
 144
 145                 size += slice->size0 * depth * layers_in_level;
 146
 147 #if 0
 148                 fprintf(stderr, "%s: %ux%ux%u@%u:\t%2u: stride=%4u, size=%6u,%7u, aligned_height=%3u, blocks=%u, offset=0x%x tiling=%d\n",
 149                                 util_format_name(prsc->format),
 150                                 width, height, depth, rsc->layout.cpp,
 151                                 level, slice->pitch * rsc->layout.cpp,
 152                                 slice->size0, size, aligned_height, blocks,
 153                                 slice->offset, fd_resource_tile_mode(prsc, level));
 154 #endif
 155
 156                 depth = u_minify(depth, 1);
 157                 lwidth = u_minify(lwidth, 1);
 158                 lheight = u_minify(lheight, 1);
 159                 twidth = u_minify(twidth, 1);
 160                 theight = u_minify(theight, 1);
 161         }
 162
 163         return size;
 164 }
 165
 166 /* A subset of the valid tiled formats can be compressed.  We do
 167  * already require tiled in order to be compressed, but just because
 168  * it can be tiled doesn't mean it can be compressed.
 169  */
 170 static bool
 171 ok_ubwc_format(enum pipe_format pfmt)
 172 {
 173         /* NOTE: both x24s8 and z24s8 map to RB6_X8Z24_UNORM, but UBWC
 174          * does not seem to work properly when sampling x24s8.. possibly
 175          * because we sample it as TFMT6_8_8_8_8_UINT.
 176          *
 177          * This could possibly be a hw limitation, or maybe something
 178          * else wrong somewhere (although z24s8 blits and sampling with
 179          * UBWC seem fine).  Recheck on a later revision of a6xx
 180          */
 181         if (pfmt == PIPE_FORMAT_X24S8_UINT)
 182                 return false;
 183
 184         switch (fd6_pipe2color(pfmt)) {
 185         case RB6_R10G10B10A2_UINT:
 186         case RB6_R10G10B10A2_UNORM:
 187         case RB6_R11G11B10_FLOAT:
 188         case RB6_R16_FLOAT:
 189         case RB6_R16G16B16A16_FLOAT:
 190         case RB6_R16G16B16A16_SINT:
 191         case RB6_R16G16B16A16_UINT:
 192         case RB6_R16G16_FLOAT:
 193         case RB6_R16G16_SINT:
 194         case RB6_R16G16_UINT:
 195         case RB6_R16_SINT:
 196         case RB6_R16_UINT:
 197         case RB6_R32G32B32A32_SINT:
 198         case RB6_R32G32B32A32_UINT:
 199         case RB6_R32G32_SINT:
 200         case RB6_R32G32_UINT:
 201         case RB6_R5G6B5_UNORM:
 202         case RB6_R8G8B8A8_SINT:
 203         case RB6_R8G8B8A8_UINT:
 204         case RB6_R8G8B8A8_UNORM:
 205         case RB6_R8G8B8_UNORM:
 206         case RB6_R8G8_SINT:
 207         case RB6_R8G8_UINT:
 208         case RB6_R8G8_UNORM:
 209         case RB6_Z24_UNORM_S8_UINT:
 210         case RB6_Z24_UNORM_S8_UINT_AS_R8G8B8A8:
 211                 return true;
 212         default:
 213                 return false;
 214         }
 215 }
 216
 217 uint32_t
 218 fd6_fill_ubwc_buffer_sizes(struct fd_resource *rsc)
 219 {
 220 #define RBG_TILE_WIDTH_ALIGNMENT 64
 221 #define RGB_TILE_HEIGHT_ALIGNMENT 16
 222 #define UBWC_PLANE_SIZE_ALIGNMENT 4096
 223
 224         struct pipe_resource *prsc = &rsc->base;
 225         uint32_t width = prsc->width0;
 226         uint32_t height = prsc->height0;
 227
 228         if (!ok_ubwc_format(prsc->format))
 229                 return 0;
 230
 231         /* limit things to simple single level 2d for now: */
 232         if ((prsc->depth0 != 1) || (prsc->array_size != 1) || (prsc->last_level != 0))
 233                 return 0;
 234
 235         uint32_t block_width, block_height;
 236         switch (rsc->layout.cpp) {
 237         case 2:
 238         case 4:
 239                 block_width = 16;
 240                 block_height = 4;
 241                 break;
 242         case 8:
 243                 block_width = 8;
 244                 block_height = 4;
 245                 break;
 246         case 16:
 247                 block_width = 4;
 248                 block_height = 4;
 249                 break;
 250         default:
 251                 return 0;
 252         }
 253
 254         uint32_t meta_stride =
 255                 ALIGN_POT(DIV_ROUND_UP(width, block_width), RBG_TILE_WIDTH_ALIGNMENT);
 256         uint32_t meta_height =
 257                 ALIGN_POT(DIV_ROUND_UP(height, block_height), RGB_TILE_HEIGHT_ALIGNMENT);
 258         uint32_t meta_size =
 259                 ALIGN_POT(meta_stride * meta_height, UBWC_PLANE_SIZE_ALIGNMENT);
 260
 261         /* UBWC goes first, then color data.. this constraint is mainly only
 262          * because it is what the kernel expects for scanout.  For non-2D we
 263          * could just use a separate UBWC buffer..
 264          */
 265         for (int level = 0; level <= prsc->last_level; level++) {
 266                 struct fdl_slice *slice = fd_resource_slice(rsc, level);
 267                 slice->offset += meta_size;
 268         }
 269
 270         rsc->layout.ubwc_slices[0].offset = 0;
 271         rsc->layout.ubwc_slices[0].pitch = meta_stride;
 272         rsc->layout.ubwc_size = meta_size >> 2;   /* in dwords??? */
 273         rsc->layout.tile_mode = TILE6_3;
 274
 275         return meta_size;
 276 }
 277
 278 /**
 279  * Ensure the rsc is in an ok state to be used with the specified format.
 280  * This handles the case of UBWC buffers used with non-UBWC compatible
 281  * formats, by triggering an uncompress.
 282  */
 283 void
 284 fd6_validate_format(struct fd_context *ctx, struct fd_resource *rsc,
 285                 enum pipe_format format)
 286 {
 287         if (!rsc->layout.ubwc_size)
 288                 return;
 289
 290         if (ok_ubwc_format(format))
 291                 return;
 292
 293         fd_resource_uncompress(ctx, rsc);
 294 }
 295
 296 static void
 297 setup_lrz(struct fd_resource *rsc)
 298 {
 299         struct fd_screen *screen = fd_screen(rsc->base.screen);
 300         const uint32_t flags = DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
 301                         DRM_FREEDRENO_GEM_TYPE_KMEM; /* TODO */
 302         unsigned width0 = rsc->base.width0;
 303         unsigned height0 = rsc->base.height0;
 304
 305         /* LRZ buffer is super-sampled: */
 306         switch (rsc->base.nr_samples) {
 307         case 4:
 308                 width0 *= 2;
 309                 /* fallthru */
 310         case 2:
 311                 height0 *= 2;
 312         }
 313
 314         unsigned lrz_pitch  = align(DIV_ROUND_UP(width0, 8), 32);
 315         unsigned lrz_height = align(DIV_ROUND_UP(height0, 8), 16);
 316
 317         unsigned size = lrz_pitch * lrz_height * 2;
 318
 319         rsc->lrz_height = lrz_height;
 320         rsc->lrz_width = lrz_pitch;
 321         rsc->lrz_pitch = lrz_pitch;
 322         rsc->lrz = fd_bo_new(screen->dev, size, flags, "lrz");
 323 }
 324
 325 uint32_t
 326 fd6_setup_slices(struct fd_resource *rsc)
 327 {
 328         uint32_t alignment;
 329
 330         if (!(fd_mesa_debug & FD_DBG_NOLRZ) && has_depth(rsc->base.format))
 331                 setup_lrz(rsc);
 332
 333         switch (rsc->base.target) {
 334         case PIPE_TEXTURE_3D:
 335                 rsc->layout.layer_first = false;
 336                 alignment = 4096;
 337                 break;
 338         default:
 339                 rsc->layout.layer_first = true;
 340                 alignment = 1;
 341                 break;
 342         }
 343
 344         return setup_slices(rsc, alignment, rsc->base.format);
 345 }