src/intel/isl/isl_gen7.c

   1 /*
   2  * Copyright 2015 Intel Corporation
   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 #include "isl_gen7.h"
  25 #include "isl_priv.h"
  26
  27 static bool
  28 gen7_format_needs_valign2(const struct isl_device *dev,
  29                           enum isl_format format)
  30 {
  31    assert(ISL_DEV_GEN(dev) == 7);
  32
  33    /* From the Ivybridge PRM (2012-05-31), Volume 4, Part 1, Section 2.12.1,
  34     * RENDER_SURFACE_STATE Surface Vertical Alignment:
  35     *
  36     *    - Value of 1 [VALIGN_4] is not supported for format YCRCB_NORMAL
  37     *      (0x182), YCRCB_SWAPUVY (0x183), YCRCB_SWAPUV (0x18f), YCRCB_SWAPY
  38     *      (0x190)
  39     *
  40     *    - VALIGN_4 is not supported for surface format R32G32B32_FLOAT.
  41     *
  42     * The R32G32B32_FLOAT restriction is dropped on Haswell.
  43     */
  44    return isl_format_is_yuv(format) ||
  45           (format == ISL_FORMAT_R32G32B32_FLOAT && !ISL_DEV_IS_HASWELL(dev));
  46 }
  47
  48 bool
  49 isl_gen7_choose_msaa_layout(const struct isl_device *dev,
  50                             const struct isl_surf_init_info *info,
  51                             enum isl_tiling tiling,
  52                             enum isl_msaa_layout *msaa_layout)
  53 {
  54    bool require_array = false;
  55    bool require_interleaved = false;
  56
  57    assert(ISL_DEV_GEN(dev) == 7);
  58    assert(info->samples >= 1);
  59
  60    if (info->samples == 1) {
  61       *msaa_layout = ISL_MSAA_LAYOUT_NONE;
  62       return true;
  63    }
  64
  65    if (!isl_format_supports_multisampling(dev->info, info->format))
  66       return false;
  67
  68    /* From the Ivybridge PRM, Volume 4 Part 1 p73, SURFACE_STATE, Number of
  69     * Multisamples:
  70     *
  71     *    - If this field is any value other than MULTISAMPLECOUNT_1, the
  72     *      Surface Type must be SURFTYPE_2D.
  73     *
  74     *    - If this field is any value other than MULTISAMPLECOUNT_1, Surface
  75     *      Min LOD, Mip Count / LOD, and Resource Min LOD must be set to zero
  76     */
  77    if (info->dim != ISL_SURF_DIM_2D)
  78       return false;
  79    if (info->levels > 1)
  80       return false;
  81
  82    /* The Ivyrbridge PRM insists twice that signed integer formats cannot be
  83     * multisampled.
  84     *
  85     * From the Ivybridge PRM, Volume 4 Part 1 p73, SURFACE_STATE, Number of
  86     * Multisamples:
  87     *
  88     *    - This field must be set to MULTISAMPLECOUNT_1 for SINT MSRTs when
  89     *      all RT channels are not written.
  90     *
  91     * And errata from the Ivybridge PRM, Volume 4 Part 1 p77,
  92     * RENDER_SURFACE_STATE, MCS Enable:
  93     *
  94     *   This field must be set to 0 [MULTISAMPLECOUNT_1] for all SINT MSRTs
  95     *   when all RT channels are not written.
  96     *
  97     * Note that the above SINT restrictions apply only to *MSRTs* (that is,
  98     * *multisampled* render targets). The restrictions seem to permit an MCS
  99     * if the render target is singlesampled.
 100     *
 101     * Moreover, empirically it looks that hardware can render multisampled
 102     * surfaces with RGBA8I, RGBA16I and RGBA32I.
 103     */
 104
 105    /* Multisampling requires vertical alignment of four. */
 106    if (info->samples > 1 && gen7_format_needs_valign2(dev, info->format))
 107       return false;
 108
 109    /* More obvious restrictions */
 110    if (isl_surf_usage_is_display(info->usage))
 111       return false;
 112    if (tiling == ISL_TILING_LINEAR)
 113       return false;
 114
 115    /* From the Ivybridge PRM, Volume 4 Part 1 p72, SURFACE_STATE, Multisampled
 116     * Suface Storage Format:
 117     *
 118     *    +---------------------+----------------------------------------------------------------+
 119     *    | MSFMT_MSS           | Multsampled surface was/is rendered as a render target         |
 120     *    | MSFMT_DEPTH_STENCIL | Multisampled surface was rendered as a depth or stencil buffer |
 121     *    +---------------------+----------------------------------------------------------------+
 122     *
 123     * In the table above, MSFMT_MSS refers to ISL_MSAA_LAYOUT_ARRAY, and
 124     * MSFMT_DEPTH_STENCIL refers to ISL_MSAA_LAYOUT_INTERLEAVED.
 125     */
 126    if (isl_surf_usage_is_depth_or_stencil(info->usage) ||
 127        (info->usage & ISL_SURF_USAGE_HIZ_BIT))
 128       require_interleaved = true;
 129
 130    /* From the Ivybridge PRM, Volume 4 Part 1 p72, SURFACE_STATE, Multisampled
 131     * Suface Storage Format:
 132     *
 133     *    If the surface’s Number of Multisamples is MULTISAMPLECOUNT_8, Width
 134     *    is >= 8192 (meaning the actual surface width is >= 8193 pixels), this
 135     *    field must be set to MSFMT_MSS.
 136     */
 137    if (info->samples == 8 && info->width > 8192)
 138       require_array = true;
 139
 140    /* From the Ivybridge PRM, Volume 4 Part 1 p72, SURFACE_STATE, Multisampled
 141     * Suface Storage Format:
 142     *
 143     *    If the surface’s Number of Multisamples is MULTISAMPLECOUNT_8,
 144     *    ((Depth+1) * (Height+1)) is > 4,194,304, OR if the surface’s Number
 145     *    of Multisamples is MULTISAMPLECOUNT_4, ((Depth+1) * (Height+1)) is
 146     *    > 8,388,608, this field must be set to MSFMT_DEPTH_STENCIL.
 147     */
 148    if ((info->samples == 8 && info->height > 4194304u) ||
 149        (info->samples == 4 && info->height > 8388608u))
 150       require_interleaved = true;
 151
 152    /* From the Ivybridge PRM, Volume 4 Part 1 p72, SURFACE_STATE, Multisampled
 153     * Suface Storage Format:
 154     *
 155     *    This field must be set to MSFMT_DEPTH_STENCIL if Surface Format is
 156     *    one of the following: I24X8_UNORM, L24X8_UNORM, A24X8_UNORM, or
 157     *    R24_UNORM_X8_TYPELESS.
 158     */
 159    if (info->format == ISL_FORMAT_I24X8_UNORM ||
 160        info->format == ISL_FORMAT_L24X8_UNORM ||
 161        info->format == ISL_FORMAT_A24X8_UNORM ||
 162        info->format == ISL_FORMAT_R24_UNORM_X8_TYPELESS)
 163       require_interleaved = true;
 164
 165    if (require_array && require_interleaved)
 166       return false;
 167
 168    if (require_interleaved) {
 169       *msaa_layout = ISL_MSAA_LAYOUT_INTERLEAVED;
 170       return true;
 171    }
 172
 173    /* Default to the array layout because it permits multisample
 174     * compression.
 175     */
 176    *msaa_layout = ISL_MSAA_LAYOUT_ARRAY;
 177    return true;
 178 }
 179
 180 /**
 181  * @brief Filter out tiling flags that are incompatible with the surface.
 182  *
 183  * The resultant outgoing @a flags is a subset of the incoming @a flags. The
 184  * outgoing flags may be empty (0x0) if the incoming flags were too
 185  * restrictive.
 186  *
 187  * For example, if the surface will be used for a display
 188  * (ISL_SURF_USAGE_DISPLAY_BIT), then this function filters out all tiling
 189  * flags except ISL_TILING_X_BIT and ISL_TILING_LINEAR_BIT.
 190  */
 191 void
 192 isl_gen6_filter_tiling(const struct isl_device *dev,
 193                        const struct isl_surf_init_info *restrict info,
 194                        isl_tiling_flags_t *flags)
 195 {
 196    /* IVB+ requires separate stencil */
 197    assert(ISL_DEV_USE_SEPARATE_STENCIL(dev));
 198
 199    /* Clear flags unsupported on this hardware */
 200    if (ISL_DEV_GEN(dev) < 9) {
 201       *flags &= ~ISL_TILING_Yf_BIT;
 202       *flags &= ~ISL_TILING_Ys_BIT;
 203    }
 204
 205    /* And... clear the Yf and Ys bits anyway because Anvil doesn't support
 206     * them yet.
 207     */
 208    *flags &= ~ISL_TILING_Yf_BIT; /* FINISHME[SKL]: Support Yf */
 209    *flags &= ~ISL_TILING_Ys_BIT; /* FINISHME[SKL]: Support Ys */
 210
 211    if (isl_surf_usage_is_depth(info->usage)) {
 212       /* Depth requires Y. */
 213       *flags &= ISL_TILING_ANY_Y_MASK;
 214    }
 215
 216    /* Separate stencil requires W tiling, and W tiling requires separate
 217     * stencil.
 218     */
 219    if (isl_surf_usage_is_stencil(info->usage)) {
 220       *flags &= ISL_TILING_W_BIT;
 221    } else {
 222       *flags &= ~ISL_TILING_W_BIT;
 223    }
 224
 225    /* From the SKL+ PRMs, RENDER_SURFACE_STATE:TileMode,
 226     *    If Surface Format is ASTC*, this field must be TILEMODE_YMAJOR.
 227     */
 228    if (isl_format_get_layout(info->format)->txc == ISL_TXC_ASTC)
 229       *flags &= ISL_TILING_Y0_BIT;
 230
 231    /* MCS buffers are always Y-tiled */
 232    if (isl_format_get_layout(info->format)->txc == ISL_TXC_MCS)
 233       *flags &= ISL_TILING_Y0_BIT;
 234
 235    if (info->usage & (ISL_SURF_USAGE_DISPLAY_ROTATE_90_BIT |
 236                       ISL_SURF_USAGE_DISPLAY_ROTATE_180_BIT |
 237                       ISL_SURF_USAGE_DISPLAY_ROTATE_270_BIT)) {
 238       assert(*flags & ISL_SURF_USAGE_DISPLAY_BIT);
 239       isl_finishme("%s:%s: handle rotated display surfaces",
 240                    __FILE__, __func__);
 241    }
 242
 243    if (info->usage & (ISL_SURF_USAGE_DISPLAY_FLIP_X_BIT |
 244                       ISL_SURF_USAGE_DISPLAY_FLIP_Y_BIT)) {
 245       assert(*flags & ISL_SURF_USAGE_DISPLAY_BIT);
 246       isl_finishme("%s:%s: handle flipped display surfaces",
 247                    __FILE__, __func__);
 248    }
 249
 250    if (info->usage & ISL_SURF_USAGE_DISPLAY_BIT) {
 251       /* Before Skylake, the display engine does not accept Y */
 252       /* FINISHME[SKL]: Y tiling for display surfaces */
 253       *flags &= (ISL_TILING_LINEAR_BIT | ISL_TILING_X_BIT);
 254    }
 255
 256    if (info->samples > 1) {
 257       /* From the Sandybridge PRM, Volume 4 Part 1, SURFACE_STATE Tiled
 258        * Surface:
 259        *
 260        *   For multisample render targets, this field must be 1 (true). MSRTs
 261        *   can only be tiled.
 262        *
 263        * From the Broadwell PRM >> Volume2d: Command Structures >>
 264        * RENDER_SURFACE_STATE Tile Mode:
 265        *
 266        *   If Number of Multisamples is not MULTISAMPLECOUNT_1, this field
 267        *   must be YMAJOR.
 268        *
 269        * As usual, though, stencil is special and requires W-tiling.
 270        */
 271       *flags &= (ISL_TILING_ANY_Y_MASK | ISL_TILING_W_BIT);
 272    }
 273
 274    /* workaround */
 275    if (ISL_DEV_GEN(dev) == 7 &&
 276        gen7_format_needs_valign2(dev, info->format) &&
 277        (info->usage & ISL_SURF_USAGE_RENDER_TARGET_BIT) &&
 278        info->samples == 1) {
 279       /* Y tiling is illegal. From the Ivybridge PRM, Vol4 Part1 2.12.2.1,
 280        * SURFACE_STATE Surface Vertical Alignment:
 281        *
 282        *     This field must be set to VALIGN_4 for all tiled Y Render Target
 283        *     surfaces.
 284        */
 285       *flags &= ~ISL_TILING_Y0_BIT;
 286    }
 287
 288    /* From the Sandybridge PRM, Volume 1, Part 2, page 32:
 289     *
 290     *    "NOTE: 128BPE Format Color Buffer ( render target ) MUST be either
 291     *     TileX or Linear."
 292     *
 293     * This is necessary all the way back to 965, but is permitted on Gen7+.
 294     */
 295    if (ISL_DEV_GEN(dev) < 7 && isl_format_get_layout(info->format)->bpb >= 128)
 296       *flags &= ~ISL_TILING_Y0_BIT;
 297
 298    /* From the BDW and SKL PRMs, Volume 2d,
 299     * RENDER_SURFACE_STATE::Width - Programming Notes:
 300     *
 301     *   A known issue exists if a primitive is rendered to the first 2 rows and
 302     *   last 2 columns of a 16K width surface. If any geometry is drawn inside
 303     *   this square it will be copied to column X=2 and X=3 (arrangement on Y
 304     *   position will stay the same). If any geometry exceeds the boundaries of
 305     *   this 2x2 region it will be drawn normally. The issue also only occurs
 306     *   if the surface has TileMode != Linear.
 307     *
 308     * [Internal documentation notes that this issue isn't present on SKL GT4.]
 309     * To prevent this rendering corruption, only allow linear tiling for
 310     * surfaces with widths greater than 16K-2 pixels.
 311     *
 312     * TODO: Is this an issue for multisampled surfaces as well?
 313     */
 314    if (info->width > 16382 && info->samples == 1 &&
 315        info->usage & ISL_SURF_USAGE_RENDER_TARGET_BIT &&
 316        (ISL_DEV_GEN(dev) == 8 ||
 317         (dev->info->is_skylake && dev->info->gt != 4))) {
 318           *flags &= ISL_TILING_LINEAR_BIT;
 319    }
 320 }
 321
 322 void
 323 isl_gen7_choose_image_alignment_el(const struct isl_device *dev,
 324                                    const struct isl_surf_init_info *restrict info,
 325                                    enum isl_tiling tiling,
 326                                    enum isl_dim_layout dim_layout,
 327                                    enum isl_msaa_layout msaa_layout,
 328                                    struct isl_extent3d *image_align_el)
 329 {
 330    assert(ISL_DEV_GEN(dev) == 7);
 331
 332    /* Handled by isl_choose_image_alignment_el */
 333    assert(info->format != ISL_FORMAT_HIZ);
 334
 335    /* IVB+ does not support combined depthstencil. */
 336    assert(!isl_surf_usage_is_depth_and_stencil(info->usage));
 337
 338    /* From the Ivy Bridge PRM, Vol. 2, Part 2, Section 6.18.4.4,
 339     * "Alignment unit size", the alignment parameters are summarized in the
 340     * following table:
 341     *
 342     *     Surface Defined By | Surface Format  | Align Width | Align Height
 343     *    --------------------+-----------------+-------------+--------------
 344     *       DEPTH_BUFFER     |   D16_UNORM     |      8      |      4
 345     *                        |     other       |      4      |      4
 346     *    --------------------+-----------------+-------------+--------------
 347     *       STENCIL_BUFFER   |      N/A        |      8      |      8
 348     *    --------------------+-----------------+-------------+--------------
 349     *       SURFACE_STATE    | BC*, ETC*, EAC* |      4      |      4
 350     *                        |      FXT1       |      8      |      4
 351     *                        |   all others    |   HALIGN    |   VALIGN
 352     *    -------------------------------------------------------------------
 353     */
 354    if (isl_surf_usage_is_depth(info->usage)) {
 355       *image_align_el = info->format == ISL_FORMAT_R16_UNORM ?
 356                         isl_extent3d(8, 4, 1) : isl_extent3d(4, 4, 1);
 357       return;
 358    } else if (isl_surf_usage_is_stencil(info->usage)) {
 359       *image_align_el = isl_extent3d(8, 8, 1);
 360       return;
 361    } else if (isl_format_is_compressed(info->format)) {
 362       /* Compressed formats all have alignment equal to block size. */
 363       *image_align_el = isl_extent3d(1, 1, 1);
 364       return;
 365    }
 366
 367    /* Everything after this point is in the "set by Surface Horizontal or
 368     * Vertical Alignment" case.  Now it's just a matter of applying
 369     * restrictions.
 370     */
 371
 372    /* There are no restrictions on halign beyond what's given in the table
 373     * above.  We set it to the minimum value of 4 because that uses the least
 374     * memory.
 375     */
 376    const uint32_t halign = 4;
 377
 378    bool require_valign4 = false;
 379
 380    /* From the Ivybridge PRM, Volume 4, Part 1, Section 2.12.1:
 381     * RENDER_SURFACE_STATE Surface Vertical Alignment:
 382     *
 383     *    * This field is intended to be set to VALIGN_4 if the surface was
 384     *      rendered as a depth buffer,
 385     *
 386     *    * for a multisampled (4x) render target, or for a multisampled (8x)
 387     *      render target, since these surfaces support only alignment of 4.
 388     *
 389     *    * This field must be set to VALIGN_4 for all tiled Y Render Target
 390     *      surfaces
 391     *
 392     *    * Value of 1 is not supported for format YCRCB_NORMAL (0x182),
 393     *      YCRCB_SWAPUVY (0x183), YCRCB_SWAPUV (0x18f), YCRCB_SWAPY (0x190)
 394     *
 395     *    * If Number of Multisamples is not MULTISAMPLECOUNT_1, this field
 396     *      must be set to VALIGN_4."
 397     *
 398     * The first restriction is already handled by the table above and the
 399     * second restriction is redundant with the fifth.
 400     */
 401    if (info->samples > 1)
 402       require_valign4 = true;
 403
 404    if (tiling == ISL_TILING_Y0 &&
 405        (info->usage & ISL_SURF_USAGE_RENDER_TARGET_BIT))
 406       require_valign4 = true;
 407
 408    assert(!(require_valign4 && gen7_format_needs_valign2(dev, info->format)));
 409
 410    /* We default to VALIGN_2 because it uses the least memory. */
 411    const uint32_t valign = require_valign4 ? 4 : 2;
 412
 413    *image_align_el = isl_extent3d(halign, valign, 1);
 414 }