src/mesa/drivers/dri/i965/brw_clear.c

   1 /*
   2  * Copyright 2003 VMware, Inc.
   3  * Copyright 2009, 2012 Intel Corporation.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sublicense, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  21  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  */
  26
  27 #include "main/mtypes.h"
  28 #include "main/condrender.h"
  29 #include "swrast/swrast.h"
  30 #include "drivers/common/meta.h"
  31
  32 #include "intel_batchbuffer.h"
  33 #include "intel_blit.h"
  34 #include "intel_fbo.h"
  35 #include "intel_mipmap_tree.h"
  36
  37 #include "brw_context.h"
  38 #include "brw_blorp.h"
  39 #include "brw_defines.h"
  40
  41 #define FILE_DEBUG_FLAG DEBUG_BLIT
  42
  43 static const char *buffer_names[] = {
  44    [BUFFER_FRONT_LEFT] = "front",
  45    [BUFFER_BACK_LEFT] = "back",
  46    [BUFFER_FRONT_RIGHT] = "front right",
  47    [BUFFER_BACK_RIGHT] = "back right",
  48    [BUFFER_DEPTH] = "depth",
  49    [BUFFER_STENCIL] = "stencil",
  50    [BUFFER_ACCUM] = "accum",
  51    [BUFFER_AUX0] = "aux0",
  52    [BUFFER_COLOR0] = "color0",
  53    [BUFFER_COLOR1] = "color1",
  54    [BUFFER_COLOR2] = "color2",
  55    [BUFFER_COLOR3] = "color3",
  56    [BUFFER_COLOR4] = "color4",
  57    [BUFFER_COLOR5] = "color5",
  58    [BUFFER_COLOR6] = "color6",
  59    [BUFFER_COLOR7] = "color7",
  60 };
  61
  62 static void
  63 debug_mask(const char *name, GLbitfield mask)
  64 {
  65    GLuint i;
  66
  67    if (unlikely(INTEL_DEBUG & DEBUG_BLIT)) {
  68       DBG("%s clear:", name);
  69       for (i = 0; i < BUFFER_COUNT; i++) {
  70          if (mask & (1 << i))
  71             DBG(" %s", buffer_names[i]);
  72       }
  73       DBG("\n");
  74    }
  75 }
  76
  77 /**
  78  * Returns true if the scissor is a noop (cuts out nothing).
  79  */
  80 static bool
  81 noop_scissor(struct gl_framebuffer *fb)
  82 {
  83    return fb->_Xmin <= 0 &&
  84           fb->_Ymin <= 0 &&
  85           fb->_Xmax >= fb->Width &&
  86           fb->_Ymax >= fb->Height;
  87 }
  88
  89 /**
  90  * Implements fast depth clears on gen6+.
  91  *
  92  * Fast clears basically work by setting a flag in each of the subspans
  93  * represented in the HiZ buffer that says "When you need the depth values for
  94  * this subspan, it's the hardware's current clear value."  Then later rendering
  95  * can just use the static clear value instead of referencing memory.
  96  *
  97  * The tricky part of the implementation is that you have to have the clear
  98  * value that was used on the depth buffer in place for all further rendering,
  99  * at least until a resolve to the real depth buffer happens.
 100  */
 101 static bool
 102 brw_fast_clear_depth(struct gl_context *ctx)
 103 {
 104    struct brw_context *brw = brw_context(ctx);
 105    struct gl_framebuffer *fb = ctx->DrawBuffer;
 106    struct intel_renderbuffer *depth_irb =
 107       intel_get_renderbuffer(fb, BUFFER_DEPTH);
 108    struct intel_mipmap_tree *mt = depth_irb->mt;
 109    struct gl_renderbuffer_attachment *depth_att = &fb->Attachment[BUFFER_DEPTH];
 110    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 111    bool same_clear_value = true;
 112
 113    if (devinfo->gen < 6)
 114       return false;
 115
 116    if (!intel_renderbuffer_has_hiz(depth_irb))
 117       return false;
 118
 119    /* We only handle full buffer clears -- otherwise you'd have to track whether
 120     * a previous clear had happened at a different clear value and resolve it
 121     * first.
 122     */
 123    if ((ctx->Scissor.EnableFlags & 1) && !noop_scissor(fb)) {
 124       perf_debug("Failed to fast clear %dx%d depth because of scissors.  "
 125                  "Possible 5%% performance win if avoided.\n",
 126                  mt->surf.logical_level0_px.width,
 127                  mt->surf.logical_level0_px.height);
 128       return false;
 129    }
 130
 131    switch (mt->format) {
 132    case MESA_FORMAT_Z32_FLOAT_S8X24_UINT:
 133    case MESA_FORMAT_Z24_UNORM_S8_UINT:
 134       /* From the Sandy Bridge PRM, volume 2 part 1, page 314:
 135        *
 136        *     "[DevSNB+]: Several cases exist where Depth Buffer Clear cannot be
 137        *      enabled (the legacy method of clearing must be performed):
 138        *
 139        *      - If the depth buffer format is D32_FLOAT_S8X24_UINT or
 140        *        D24_UNORM_S8_UINT.
 141        */
 142       return false;
 143
 144    case MESA_FORMAT_Z_UNORM16:
 145       /* From the Sandy Bridge PRM, volume 2 part 1, page 314:
 146        *
 147        *     "[DevSNB+]: Several cases exist where Depth Buffer Clear cannot be
 148        *      enabled (the legacy method of clearing must be performed):
 149        *
 150        *      - DevSNB{W/A}]: When depth buffer format is D16_UNORM and the
 151        *        width of the map (LOD0) is not multiple of 16, fast clear
 152        *        optimization must be disabled.
 153        */
 154       if (devinfo->gen == 6 &&
 155           (minify(mt->surf.phys_level0_sa.width,
 156                   depth_irb->mt_level - mt->first_level) % 16) != 0)
 157          return false;
 158       break;
 159
 160    default:
 161       break;
 162    }
 163
 164    /* Quantize the clear value to what can be stored in the actual depth
 165     * buffer.  This makes the following check more accurate because it now
 166     * checks if the actual depth bits will match.  It also prevents us from
 167     * getting a too-accurate depth value during depth testing or when sampling
 168     * with HiZ enabled.
 169     */
 170    float clear_value =
 171       mt->format == MESA_FORMAT_Z_FLOAT32 ? ctx->Depth.Clear :
 172       (unsigned)(ctx->Depth.Clear * fb->_DepthMax) / (float)fb->_DepthMax;
 173
 174    const uint32_t num_layers = depth_att->Layered ? depth_irb->layer_count : 1;
 175
 176    /* If we're clearing to a new clear value, then we need to resolve any clear
 177     * flags out of the HiZ buffer into the real depth buffer.
 178     */
 179    if (mt->fast_clear_color.f32[0] != clear_value) {
 180       for (uint32_t level = mt->first_level; level <= mt->last_level; level++) {
 181          if (!intel_miptree_level_has_hiz(mt, level))
 182             continue;
 183
 184          const unsigned level_layers = brw_get_num_logical_layers(mt, level);
 185
 186          for (uint32_t layer = 0; layer < level_layers; layer++) {
 187             if (level == depth_irb->mt_level &&
 188                 layer >= depth_irb->mt_layer &&
 189                 layer < depth_irb->mt_layer + num_layers) {
 190                /* We're going to clear this layer anyway.  Leave it alone. */
 191                continue;
 192             }
 193
 194             enum isl_aux_state aux_state =
 195                intel_miptree_get_aux_state(mt, level, layer);
 196
 197             if (aux_state != ISL_AUX_STATE_CLEAR &&
 198                 aux_state != ISL_AUX_STATE_COMPRESSED_CLEAR) {
 199                /* This slice doesn't have any fast-cleared bits. */
 200                continue;
 201             }
 202
 203             /* If we got here, then the level may have fast-clear bits that
 204              * use the old clear value.  We need to do a depth resolve to get
 205              * rid of their use of the clear value before we can change it.
 206              * Fortunately, few applications ever change their depth clear
 207              * value so this shouldn't happen often.
 208              */
 209             intel_hiz_exec(brw, mt, level, layer, 1,
 210                            ISL_AUX_OP_FULL_RESOLVE);
 211             intel_miptree_set_aux_state(brw, mt, level, layer, 1,
 212                                         ISL_AUX_STATE_RESOLVED);
 213          }
 214       }
 215
 216       intel_miptree_set_depth_clear_value(brw, mt, clear_value);
 217       same_clear_value = false;
 218    }
 219
 220    bool need_clear = false;
 221    for (unsigned a = 0; a < num_layers; a++) {
 222       enum isl_aux_state aux_state =
 223          intel_miptree_get_aux_state(mt, depth_irb->mt_level,
 224                                      depth_irb->mt_layer + a);
 225
 226       if (aux_state != ISL_AUX_STATE_CLEAR) {
 227          need_clear = true;
 228          break;
 229       }
 230    }
 231
 232    if (!need_clear) {
 233       /* If all of the layers we intend to clear are already in the clear
 234        * state then simply updating the miptree fast clear value is sufficient
 235        * to change their clear value.
 236        */
 237       if (devinfo->gen >= 10 && !same_clear_value) {
 238          /* Before gen10, it was enough to just update the clear value in the
 239           * miptree. But on gen10+, we let blorp update the clear value state
 240           * buffer when doing a fast clear. Since we are skipping the fast
 241           * clear here, we need to update the clear color ourselves.
 242           */
 243          uint32_t clear_offset = mt->aux_buf->clear_color_offset;
 244          union isl_color_value clear_color = { .f32 = { clear_value, } };
 245
 246          /* We can't update the clear color while the hardware is still using
 247           * the previous one for a resolve or sampling from it. So make sure
 248           * that there's no pending commands at this point.
 249           */
 250          brw_emit_pipe_control_flush(brw, PIPE_CONTROL_CS_STALL);
 251          for (int i = 0; i < 4; i++) {
 252             brw_store_data_imm32(brw, mt->aux_buf->clear_color_bo,
 253                                  clear_offset + i * 4, clear_color.u32[i]);
 254          }
 255          brw_emit_pipe_control_flush(brw, PIPE_CONTROL_STATE_CACHE_INVALIDATE);
 256       }
 257       return true;
 258    }
 259
 260    for (unsigned a = 0; a < num_layers; a++) {
 261       enum isl_aux_state aux_state =
 262          intel_miptree_get_aux_state(mt, depth_irb->mt_level,
 263                                      depth_irb->mt_layer + a);
 264
 265       if (aux_state != ISL_AUX_STATE_CLEAR) {
 266          intel_hiz_exec(brw, mt, depth_irb->mt_level,
 267                         depth_irb->mt_layer + a, 1,
 268                         ISL_AUX_OP_FAST_CLEAR);
 269       }
 270    }
 271
 272    /* Now, the HiZ buffer contains data that needs to be resolved to the depth
 273     * buffer.
 274     */
 275    intel_miptree_set_aux_state(brw, mt, depth_irb->mt_level,
 276                                depth_irb->mt_layer, num_layers,
 277                                ISL_AUX_STATE_CLEAR);
 278
 279    return true;
 280 }
 281
 282 /**
 283  * Called by ctx->Driver.Clear.
 284  */
 285 static void
 286 brw_clear(struct gl_context *ctx, GLbitfield mask)
 287 {
 288    struct brw_context *brw = brw_context(ctx);
 289    struct gl_framebuffer *fb = ctx->DrawBuffer;
 290    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 291    bool partial_clear = ctx->Scissor.EnableFlags && !noop_scissor(fb);
 292
 293    if (!_mesa_check_conditional_render(ctx))
 294       return;
 295
 296    if (mask & (BUFFER_BIT_FRONT_LEFT | BUFFER_BIT_FRONT_RIGHT)) {
 297       brw->front_buffer_dirty = true;
 298    }
 299
 300    intel_prepare_render(brw);
 301    brw_workaround_depthstencil_alignment(brw, partial_clear ? 0 : mask);
 302
 303    if (mask & BUFFER_BIT_DEPTH) {
 304       if (brw_fast_clear_depth(ctx)) {
 305          DBG("fast clear: depth\n");
 306          mask &= ~BUFFER_BIT_DEPTH;
 307       }
 308    }
 309
 310    if (mask & BUFFER_BIT_STENCIL) {
 311       struct intel_renderbuffer *stencil_irb =
 312          intel_get_renderbuffer(fb, BUFFER_STENCIL);
 313       struct intel_mipmap_tree *mt = stencil_irb->mt;
 314       if (mt && mt->stencil_mt)
 315          mt->stencil_mt->r8stencil_needs_update = true;
 316    }
 317
 318    if (mask & BUFFER_BITS_COLOR) {
 319       brw_blorp_clear_color(brw, fb, mask, partial_clear,
 320                             ctx->Color.sRGBEnabled);
 321       debug_mask("blorp color", mask & BUFFER_BITS_COLOR);
 322       mask &= ~BUFFER_BITS_COLOR;
 323    }
 324
 325    if (devinfo->gen >= 6 && (mask & BUFFER_BITS_DEPTH_STENCIL)) {
 326       brw_blorp_clear_depth_stencil(brw, fb, mask, partial_clear);
 327       debug_mask("blorp depth/stencil", mask & BUFFER_BITS_DEPTH_STENCIL);
 328       mask &= ~BUFFER_BITS_DEPTH_STENCIL;
 329    }
 330
 331    GLbitfield tri_mask = mask & (BUFFER_BIT_STENCIL |
 332                                  BUFFER_BIT_DEPTH);
 333
 334    if (tri_mask) {
 335       debug_mask("tri", tri_mask);
 336       mask &= ~tri_mask;
 337       _mesa_meta_glsl_Clear(&brw->ctx, tri_mask);
 338    }
 339
 340    /* Any strange buffers get passed off to swrast.  The only thing that
 341     * should be left at this point is the accumulation buffer.
 342     */
 343    assert((mask & ~BUFFER_BIT_ACCUM) == 0);
 344    if (mask) {
 345       debug_mask("swrast", mask);
 346       _swrast_Clear(ctx, mask);
 347    }
 348 }
 349
 350
 351 void
 352 intelInitClearFuncs(struct dd_function_table *functions)
 353 {
 354    functions->Clear = brw_clear;
 355 }