src/mesa/drivers/dri/i965/brw_blorp.cpp

   1 /*
   2  * Copyright © 2012 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 <errno.h>
  25 #include "intel_batchbuffer.h"
  26 #include "intel_fbo.h"
  27
  28 #include "brw_blorp.h"
  29 #include "brw_defines.h"
  30 #include "brw_state.h"
  31 #include "gen6_blorp.h"
  32 #include "gen7_blorp.h"
  33
  34 #define FILE_DEBUG_FLAG DEBUG_BLORP
  35
  36 brw_blorp_mip_info::brw_blorp_mip_info()
  37    : mt(NULL),
  38      level(0),
  39      layer(0),
  40      width(0),
  41      height(0),
  42      x_offset(0),
  43      y_offset(0)
  44 {
  45 }
  46
  47 brw_blorp_surface_info::brw_blorp_surface_info()
  48    : map_stencil_as_y_tiled(false),
  49      num_samples(0)
  50 {
  51 }
  52
  53 void
  54 brw_blorp_mip_info::set(struct intel_mipmap_tree *mt,
  55                         unsigned int level, unsigned int layer)
  56 {
  57    /* Layer is a physical layer, so if this is a 2D multisample array texture
  58     * using INTEL_MSAA_LAYOUT_UMS or INTEL_MSAA_LAYOUT_CMS, then it had better
  59     * be a multiple of num_samples.
  60     */
  61    if (mt->msaa_layout == INTEL_MSAA_LAYOUT_UMS ||
  62        mt->msaa_layout == INTEL_MSAA_LAYOUT_CMS) {
  63       assert(layer % mt->num_samples == 0);
  64    }
  65
  66    intel_miptree_check_level_layer(mt, level, layer);
  67
  68    this->mt = mt;
  69    this->level = level;
  70    this->layer = layer;
  71    this->width = minify(mt->physical_width0, level - mt->first_level);
  72    this->height = minify(mt->physical_height0, level - mt->first_level);
  73
  74    intel_miptree_get_image_offset(mt, level, layer, &x_offset, &y_offset);
  75 }
  76
  77 void
  78 brw_blorp_surface_info::set(struct brw_context *brw,
  79                             struct intel_mipmap_tree *mt,
  80                             unsigned int level, unsigned int layer,
  81                             mesa_format format, bool is_render_target)
  82 {
  83    brw_blorp_mip_info::set(mt, level, layer);
  84    this->num_samples = mt->num_samples;
  85    this->array_layout = mt->array_layout;
  86    this->map_stencil_as_y_tiled = false;
  87    this->msaa_layout = mt->msaa_layout;
  88
  89    if (format == MESA_FORMAT_NONE)
  90       format = mt->format;
  91
  92    switch (format) {
  93    case MESA_FORMAT_S_UINT8:
  94       /* The miptree is a W-tiled stencil buffer.  Surface states can't be set
  95        * up for W tiling, so we'll need to use Y tiling and have the WM
  96        * program swizzle the coordinates.
  97        */
  98       this->map_stencil_as_y_tiled = true;
  99       this->brw_surfaceformat = BRW_SURFACEFORMAT_R8_UNORM;
 100       break;
 101    case MESA_FORMAT_Z24_UNORM_X8_UINT:
 102       /* It would make sense to use BRW_SURFACEFORMAT_R24_UNORM_X8_TYPELESS
 103        * here, but unfortunately it isn't supported as a render target, which
 104        * would prevent us from blitting to 24-bit depth.
 105        *
 106        * The miptree consists of 32 bits per pixel, arranged as 24-bit depth
 107        * values interleaved with 8 "don't care" bits.  Since depth values don't
 108        * require any blending, it doesn't matter how we interpret the bit
 109        * pattern as long as we copy the right amount of data, so just map it
 110        * as 8-bit BGRA.
 111        */
 112       this->brw_surfaceformat = BRW_SURFACEFORMAT_B8G8R8A8_UNORM;
 113       break;
 114    case MESA_FORMAT_Z_FLOAT32:
 115       this->brw_surfaceformat = BRW_SURFACEFORMAT_R32_FLOAT;
 116       break;
 117    case MESA_FORMAT_Z_UNORM16:
 118       this->brw_surfaceformat = BRW_SURFACEFORMAT_R16_UNORM;
 119       break;
 120    default: {
 121       mesa_format linear_format = _mesa_get_srgb_format_linear(format);
 122       if (is_render_target) {
 123          assert(brw->format_supported_as_render_target[linear_format]);
 124          this->brw_surfaceformat = brw->render_target_format[linear_format];
 125       } else {
 126          this->brw_surfaceformat = brw_format_for_mesa_format(linear_format);
 127       }
 128       break;
 129    }
 130    }
 131 }
 132
 133
 134 /**
 135  * Split x_offset and y_offset into a base offset (in bytes) and a remaining
 136  * x/y offset (in pixels).  Note: we can't do this by calling
 137  * intel_renderbuffer_tile_offsets(), because the offsets may have been
 138  * adjusted to account for Y vs. W tiling differences.  So we compute it
 139  * directly from the adjusted offsets.
 140  */
 141 uint32_t
 142 brw_blorp_surface_info::compute_tile_offsets(uint32_t *tile_x,
 143                                              uint32_t *tile_y) const
 144 {
 145    uint32_t mask_x, mask_y;
 146
 147    intel_get_tile_masks(mt->tiling, mt->tr_mode, mt->cpp,
 148                         map_stencil_as_y_tiled,
 149                         &mask_x, &mask_y);
 150
 151    *tile_x = x_offset & mask_x;
 152    *tile_y = y_offset & mask_y;
 153
 154    return intel_miptree_get_aligned_offset(mt, x_offset & ~mask_x,
 155                                            y_offset & ~mask_y,
 156                                            map_stencil_as_y_tiled);
 157 }
 158
 159
 160 brw_blorp_params::brw_blorp_params(unsigned num_varyings,
 161                                    unsigned num_draw_buffers,
 162                                    unsigned num_layers)
 163    : x0(0),
 164      y0(0),
 165      x1(0),
 166      y1(0),
 167      depth_format(0),
 168      hiz_op(GEN6_HIZ_OP_NONE),
 169      use_wm_prog(false),
 170      num_varyings(num_varyings),
 171      num_draw_buffers(num_draw_buffers),
 172      num_layers(num_layers)
 173 {
 174 }
 175
 176 extern "C" {
 177 void
 178 intel_hiz_exec(struct brw_context *brw, struct intel_mipmap_tree *mt,
 179                unsigned int level, unsigned int layer, gen6_hiz_op op)
 180 {
 181    const char *opname = NULL;
 182
 183    switch (op) {
 184    case GEN6_HIZ_OP_DEPTH_RESOLVE:
 185       opname = "depth resolve";
 186       break;
 187    case GEN6_HIZ_OP_HIZ_RESOLVE:
 188       opname = "hiz ambiguate";
 189       break;
 190    case GEN6_HIZ_OP_DEPTH_CLEAR:
 191       opname = "depth clear";
 192       break;
 193    case GEN6_HIZ_OP_NONE:
 194       opname = "noop?";
 195       break;
 196    }
 197
 198    DBG("%s %s to mt %p level %d layer %d\n",
 199        __func__, opname, mt, level, layer);
 200
 201    if (brw->gen >= 8) {
 202       gen8_hiz_exec(brw, mt, level, layer, op);
 203    } else {
 204       brw_hiz_op_params params(mt, level, layer, op);
 205       brw_blorp_exec(brw, &params);
 206    }
 207 }
 208
 209 } /* extern "C" */
 210
 211 void
 212 brw_blorp_exec(struct brw_context *brw, const brw_blorp_params *params)
 213 {
 214    struct gl_context *ctx = &brw->ctx;
 215    uint32_t estimated_max_batch_usage = 1500;
 216    bool check_aperture_failed_once = false;
 217
 218    /* Flush the sampler and render caches.  We definitely need to flush the
 219     * sampler cache so that we get updated contents from the render cache for
 220     * the glBlitFramebuffer() source.  Also, we are sometimes warned in the
 221     * docs to flush the cache between reinterpretations of the same surface
 222     * data with different formats, which blorp does for stencil and depth
 223     * data.
 224     */
 225    brw_emit_mi_flush(brw);
 226
 227 retry:
 228    intel_batchbuffer_require_space(brw, estimated_max_batch_usage, RENDER_RING);
 229    intel_batchbuffer_save_state(brw);
 230    drm_intel_bo *saved_bo = brw->batch.bo;
 231    uint32_t saved_used = USED_BATCH(brw->batch);
 232    uint32_t saved_state_batch_offset = brw->batch.state_batch_offset;
 233
 234    switch (brw->gen) {
 235    case 6:
 236       gen6_blorp_exec(brw, params);
 237       break;
 238    case 7:
 239       gen7_blorp_exec(brw, params);
 240       break;
 241    default:
 242       /* BLORP is not supported before Gen6. */
 243       unreachable("not reached");
 244    }
 245
 246    /* Make sure we didn't wrap the batch unintentionally, and make sure we
 247     * reserved enough space that a wrap will never happen.
 248     */
 249    assert(brw->batch.bo == saved_bo);
 250    assert((USED_BATCH(brw->batch) - saved_used) * 4 +
 251           (saved_state_batch_offset - brw->batch.state_batch_offset) <
 252           estimated_max_batch_usage);
 253    /* Shut up compiler warnings on release build */
 254    (void)saved_bo;
 255    (void)saved_used;
 256    (void)saved_state_batch_offset;
 257
 258    /* Check if the blorp op we just did would make our batch likely to fail to
 259     * map all the BOs into the GPU at batch exec time later.  If so, flush the
 260     * batch and try again with nothing else in the batch.
 261     */
 262    if (dri_bufmgr_check_aperture_space(&brw->batch.bo, 1)) {
 263       if (!check_aperture_failed_once) {
 264          check_aperture_failed_once = true;
 265          intel_batchbuffer_reset_to_saved(brw);
 266          intel_batchbuffer_flush(brw);
 267          goto retry;
 268       } else {
 269          int ret = intel_batchbuffer_flush(brw);
 270          WARN_ONCE(ret == -ENOSPC,
 271                    "i965: blorp emit exceeded available aperture space\n");
 272       }
 273    }
 274
 275    if (unlikely(brw->always_flush_batch))
 276       intel_batchbuffer_flush(brw);
 277
 278    /* We've smashed all state compared to what the normal 3D pipeline
 279     * rendering tracks for GL.
 280     */
 281    brw->ctx.NewDriverState = ~0ull;
 282    brw->no_depth_or_stencil = false;
 283    brw->ib.type = -1;
 284
 285    /* Flush the sampler cache so any texturing from the destination is
 286     * coherent.
 287     */
 288    brw_emit_mi_flush(brw);
 289 }
 290
 291 brw_hiz_op_params::brw_hiz_op_params(struct intel_mipmap_tree *mt,
 292                                      unsigned int level,
 293                                      unsigned int layer,
 294                                      gen6_hiz_op op)
 295 {
 296    this->hiz_op = op;
 297
 298    depth.set(mt, level, layer);
 299
 300    /* Align the rectangle primitive to 8x4 pixels.
 301     *
 302     * During fast depth clears, the emitted rectangle primitive  must be
 303     * aligned to 8x4 pixels.  From the Ivybridge PRM, Vol 2 Part 1 Section
 304     * 11.5.3.1 Depth Buffer Clear (and the matching section in the Sandybridge
 305     * PRM):
 306     *     If Number of Multisamples is NUMSAMPLES_1, the rectangle must be
 307     *     aligned to an 8x4 pixel block relative to the upper left corner
 308     *     of the depth buffer [...]
 309     *
 310     * For hiz resolves, the rectangle must also be 8x4 aligned. Item
 311     * WaHizAmbiguate8x4Aligned from the Haswell workarounds page and the
 312     * Ivybridge simulator require the alignment.
 313     *
 314     * To be safe, let's just align the rect for all hiz operations and all
 315     * hardware generations.
 316     *
 317     * However, for some miptree slices of a Z24 texture, emitting an 8x4
 318     * aligned rectangle that covers the slice may clobber adjacent slices if
 319     * we strictly adhered to the texture alignments specified in the PRM.  The
 320     * Ivybridge PRM, Section "Alignment Unit Size", states that
 321     * SURFACE_STATE.Surface_Horizontal_Alignment should be 4 for Z24 surfaces,
 322     * not 8. But commit 1f112cc increased the alignment from 4 to 8, which
 323     * prevents the clobbering.
 324     */
 325    depth.width = ALIGN(depth.width, 8);
 326    depth.height = ALIGN(depth.height, 4);
 327
 328    x1 = depth.width;
 329    y1 = depth.height;
 330
 331    assert(intel_miptree_level_has_hiz(mt, level));
 332
 333    switch (mt->format) {
 334    case MESA_FORMAT_Z_UNORM16:       depth_format = BRW_DEPTHFORMAT_D16_UNORM; break;
 335    case MESA_FORMAT_Z_FLOAT32: depth_format = BRW_DEPTHFORMAT_D32_FLOAT; break;
 336    case MESA_FORMAT_Z24_UNORM_X8_UINT:    depth_format = BRW_DEPTHFORMAT_D24_UNORM_X8_UINT; break;
 337    default:                    unreachable("not reached");
 338    }
 339 }
 340
 341 uint32_t
 342 brw_hiz_op_params::get_wm_prog(struct brw_context *brw,
 343                                brw_blorp_prog_data **prog_data) const
 344 {
 345    return 0;
 346 }