• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2017 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
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  * on the rights to use, copy, modify, merge, publish, distribute, sub
9  * license, and/or sell copies of the Software, and to permit persons to whom
10  * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
23  */
24 
25 #include "si_pipe.h"
26 #include "sid.h"
27 #include "util/format/u_format.h"
28 #include "util/u_pack_color.h"
29 #include "util/u_surface.h"
30 
31 enum
32 {
33    SI_CLEAR = SI_SAVE_FRAGMENT_STATE,
34    SI_CLEAR_SURFACE = SI_SAVE_FRAMEBUFFER | SI_SAVE_FRAGMENT_STATE,
35 };
36 
si_init_buffer_clear(struct si_clear_info * info,struct pipe_resource * resource,uint64_t offset,uint32_t size,uint32_t clear_value)37 void si_init_buffer_clear(struct si_clear_info *info,
38                           struct pipe_resource *resource, uint64_t offset,
39                           uint32_t size, uint32_t clear_value)
40 {
41    info->resource = resource;
42    info->offset = offset;
43    info->size = size;
44    info->clear_value = clear_value;
45    info->writemask = 0xffffffff;
46    info->is_dcc_msaa = false;
47 }
48 
si_init_buffer_clear_rmw(struct si_clear_info * info,struct pipe_resource * resource,uint64_t offset,uint32_t size,uint32_t clear_value,uint32_t writemask)49 static void si_init_buffer_clear_rmw(struct si_clear_info *info,
50                                      struct pipe_resource *resource, uint64_t offset,
51                                      uint32_t size, uint32_t clear_value, uint32_t writemask)
52 {
53    si_init_buffer_clear(info, resource, offset, size, clear_value);
54    info->writemask = writemask;
55 }
56 
si_execute_clears(struct si_context * sctx,struct si_clear_info * info,unsigned num_clears,unsigned types)57 void si_execute_clears(struct si_context *sctx, struct si_clear_info *info,
58                        unsigned num_clears, unsigned types)
59 {
60    if (!num_clears)
61       return;
62 
63    /* Flush caches and wait for idle. */
64    if (types & (SI_CLEAR_TYPE_CMASK | SI_CLEAR_TYPE_DCC))
65       sctx->flags |= si_get_flush_flags(sctx, SI_COHERENCY_CB_META, L2_LRU);
66 
67    if (types & SI_CLEAR_TYPE_HTILE)
68       sctx->flags |= si_get_flush_flags(sctx, SI_COHERENCY_DB_META, L2_LRU);
69 
70    /* Flush caches in case we use compute. */
71    sctx->flags |= SI_CONTEXT_INV_VCACHE;
72 
73    /* GFX6-8: CB and DB don't use L2. */
74    if (sctx->gfx_level <= GFX8)
75       sctx->flags |= SI_CONTEXT_INV_L2;
76 
77    /* Execute clears. */
78    for (unsigned i = 0; i < num_clears; i++) {
79       if (info[i].is_dcc_msaa) {
80          gfx9_clear_dcc_msaa(sctx, info[i].resource, info[i].clear_value,
81                              SI_OP_SKIP_CACHE_INV_BEFORE, SI_COHERENCY_CP);
82          continue;
83       }
84 
85       assert(info[i].size > 0);
86 
87       if (info[i].writemask != 0xffffffff) {
88          si_compute_clear_buffer_rmw(sctx, info[i].resource, info[i].offset, info[i].size,
89                                      info[i].clear_value, info[i].writemask,
90                                      SI_OP_SKIP_CACHE_INV_BEFORE, SI_COHERENCY_CP);
91       } else {
92          /* Compute shaders are much faster on both dGPUs and APUs. Don't use CP DMA. */
93          si_clear_buffer(sctx, info[i].resource, info[i].offset, info[i].size,
94                          &info[i].clear_value, 4, SI_OP_SKIP_CACHE_INV_BEFORE,
95                          SI_COHERENCY_CP, SI_COMPUTE_CLEAR_METHOD);
96       }
97    }
98 
99    /* Wait for idle. */
100    sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
101 
102    /* GFX6-8: CB and DB don't use L2. */
103    if (sctx->gfx_level <= GFX8)
104       sctx->flags |= SI_CONTEXT_WB_L2;
105 }
106 
si_alloc_separate_cmask(struct si_screen * sscreen,struct si_texture * tex)107 static bool si_alloc_separate_cmask(struct si_screen *sscreen, struct si_texture *tex)
108 {
109    assert(sscreen->info.gfx_level < GFX11);
110 
111    /* CMASK for MSAA is allocated in advance or always disabled
112     * by "nofmask" option.
113     */
114    if (tex->cmask_buffer)
115       return true;
116 
117    if (!tex->surface.cmask_size)
118       return false;
119 
120    tex->cmask_buffer =
121       si_aligned_buffer_create(&sscreen->b, PIPE_RESOURCE_FLAG_UNMAPPABLE, PIPE_USAGE_DEFAULT,
122                                tex->surface.cmask_size, 1 << tex->surface.cmask_alignment_log2);
123    if (tex->cmask_buffer == NULL)
124       return false;
125 
126    tex->cmask_base_address_reg = tex->cmask_buffer->gpu_address >> 8;
127    tex->cb_color_info |= S_028C70_FAST_CLEAR(1);
128 
129    p_atomic_inc(&sscreen->compressed_colortex_counter);
130    return true;
131 }
132 
si_set_clear_color(struct si_texture * tex,enum pipe_format surface_format,const union pipe_color_union * color)133 static bool si_set_clear_color(struct si_texture *tex, enum pipe_format surface_format,
134                                const union pipe_color_union *color)
135 {
136    union util_color uc;
137 
138    memset(&uc, 0, sizeof(uc));
139 
140    if (tex->surface.bpe == 16) {
141       /* DCC fast clear only:
142        *   CLEAR_WORD0 = R = G = B
143        *   CLEAR_WORD1 = A
144        */
145       assert(color->ui[0] == color->ui[1] && color->ui[0] == color->ui[2]);
146       uc.ui[0] = color->ui[0];
147       uc.ui[1] = color->ui[3];
148    } else {
149       if (tex->swap_rgb_to_bgr)
150          surface_format = util_format_rgb_to_bgr(surface_format);
151 
152       util_pack_color_union(surface_format, &uc, color);
153    }
154 
155    if (memcmp(tex->color_clear_value, &uc, 2 * sizeof(uint32_t)) == 0)
156       return false;
157 
158    memcpy(tex->color_clear_value, &uc, 2 * sizeof(uint32_t));
159    return true;
160 }
161 
162 /** Linearize and convert luminance/intensity to red. */
si_simplify_cb_format(enum pipe_format format)163 enum pipe_format si_simplify_cb_format(enum pipe_format format)
164 {
165    format = util_format_linear(format);
166    format = util_format_luminance_to_red(format);
167    return util_format_intensity_to_red(format);
168 }
169 
vi_alpha_is_on_msb(struct si_screen * sscreen,enum pipe_format format)170 bool vi_alpha_is_on_msb(struct si_screen *sscreen, enum pipe_format format)
171 {
172    format = si_simplify_cb_format(format);
173    const struct util_format_description *desc = util_format_description(format);
174    unsigned comp_swap = si_translate_colorswap(sscreen->info.gfx_level, format, false);
175 
176    /* The following code matches the hw behavior. */
177    if (desc->nr_channels == 1) {
178       return (comp_swap == V_028C70_SWAP_ALT_REV) != (sscreen->info.family == CHIP_RAVEN2 ||
179                                                       sscreen->info.family == CHIP_RENOIR);
180    }
181 
182    return comp_swap != V_028C70_SWAP_STD_REV && comp_swap != V_028C70_SWAP_ALT_REV;
183 }
184 
gfx8_get_dcc_clear_parameters(struct si_screen * sscreen,enum pipe_format base_format,enum pipe_format surface_format,const union pipe_color_union * color,uint32_t * clear_value,bool * eliminate_needed)185 static bool gfx8_get_dcc_clear_parameters(struct si_screen *sscreen, enum pipe_format base_format,
186                                           enum pipe_format surface_format,
187                                           const union pipe_color_union *color, uint32_t *clear_value,
188                                           bool *eliminate_needed)
189 {
190    /* If we want to clear without needing a fast clear eliminate step, we
191     * can set color and alpha independently to 0 or 1 (or 0/max for integer
192     * formats).
193     */
194    bool values[4] = {};      /* whether to clear to 0 or 1 */
195    bool color_value = false; /* clear color to 0 or 1 */
196    bool alpha_value = false; /* clear alpha to 0 or 1 */
197    int alpha_channel;        /* index of the alpha component */
198    bool has_color = false;
199    bool has_alpha = false;
200 
201    const struct util_format_description *desc =
202       util_format_description(si_simplify_cb_format(surface_format));
203 
204    /* 128-bit fast clear with different R,G,B values is unsupported. */
205    if (desc->block.bits == 128 && (color->ui[0] != color->ui[1] || color->ui[0] != color->ui[2]))
206       return false;
207 
208    *eliminate_needed = true;
209    *clear_value = GFX8_DCC_CLEAR_REG;
210 
211    if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
212       return true; /* need ELIMINATE_FAST_CLEAR */
213 
214    bool base_alpha_is_on_msb = vi_alpha_is_on_msb(sscreen, base_format);
215    bool surf_alpha_is_on_msb = vi_alpha_is_on_msb(sscreen, surface_format);
216 
217    /* Formats with 3 channels can't have alpha. */
218    if (desc->nr_channels == 3)
219       alpha_channel = -1;
220    else if (surf_alpha_is_on_msb)
221       alpha_channel = desc->nr_channels - 1;
222    else
223       alpha_channel = 0;
224 
225    for (int i = 0; i < 4; ++i) {
226       if (desc->swizzle[i] >= PIPE_SWIZZLE_0)
227          continue;
228 
229       if (desc->channel[i].pure_integer && desc->channel[i].type == UTIL_FORMAT_TYPE_SIGNED) {
230          /* Use the maximum value for clamping the clear color. */
231          int max = u_bit_consecutive(0, desc->channel[i].size - 1);
232 
233          values[i] = color->i[i] != 0;
234          if (color->i[i] != 0 && MIN2(color->i[i], max) != max)
235             return true; /* need ELIMINATE_FAST_CLEAR */
236       } else if (desc->channel[i].pure_integer &&
237                  desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
238          /* Use the maximum value for clamping the clear color. */
239          unsigned max = u_bit_consecutive(0, desc->channel[i].size);
240 
241          values[i] = color->ui[i] != 0U;
242          if (color->ui[i] != 0U && MIN2(color->ui[i], max) != max)
243             return true; /* need ELIMINATE_FAST_CLEAR */
244       } else {
245          values[i] = color->f[i] != 0.0F;
246          if (color->f[i] != 0.0F && color->f[i] != 1.0F)
247             return true; /* need ELIMINATE_FAST_CLEAR */
248       }
249 
250       if (desc->swizzle[i] == alpha_channel) {
251          alpha_value = values[i];
252          has_alpha = true;
253       } else {
254          color_value = values[i];
255          has_color = true;
256       }
257    }
258 
259    /* If alpha isn't present, make it the same as color, and vice versa. */
260    if (!has_alpha)
261       alpha_value = color_value;
262    else if (!has_color)
263       color_value = alpha_value;
264 
265    if (color_value != alpha_value && base_alpha_is_on_msb != surf_alpha_is_on_msb)
266       return true; /* require ELIMINATE_FAST_CLEAR */
267 
268    /* Check if all color values are equal if they are present. */
269    for (int i = 0; i < 4; ++i) {
270       if (desc->swizzle[i] <= PIPE_SWIZZLE_W && desc->swizzle[i] != alpha_channel &&
271           values[i] != color_value)
272          return true; /* require ELIMINATE_FAST_CLEAR */
273    }
274 
275    /* This doesn't need ELIMINATE_FAST_CLEAR.
276     * On chips predating Raven2, the DCC clear codes and the CB clear
277     * color registers must match.
278     */
279    *eliminate_needed = false;
280 
281    if (color_value) {
282       if (alpha_value)
283          *clear_value = GFX8_DCC_CLEAR_1111;
284       else
285          *clear_value = GFX8_DCC_CLEAR_1110;
286    } else {
287       if (alpha_value)
288          *clear_value = GFX8_DCC_CLEAR_0001;
289       else
290          *clear_value = GFX8_DCC_CLEAR_0000;
291    }
292    return true;
293 }
294 
gfx11_get_dcc_clear_parameters(struct si_screen * sscreen,enum pipe_format surface_format,const union pipe_color_union * color,uint32_t * clear_value)295 static bool gfx11_get_dcc_clear_parameters(struct si_screen *sscreen, enum pipe_format surface_format,
296                                            const union pipe_color_union *color, uint32_t *clear_value)
297 {
298    const struct util_format_description *desc =
299       util_format_description(si_simplify_cb_format(surface_format));
300    unsigned start_bit = UINT_MAX;
301    unsigned end_bit = 0;
302 
303    /* TODO: 8bpp and 16bpp fast DCC clears don't work. */
304    if (desc->block.bits <= 16)
305       return false;
306 
307    /* Find the used bit range. */
308    for (unsigned i = 0; i < 4; i++) {
309       unsigned swizzle = desc->swizzle[i];
310 
311       if (swizzle >= PIPE_SWIZZLE_0)
312          continue;
313 
314       start_bit = MIN2(start_bit, desc->channel[swizzle].shift);
315       end_bit = MAX2(end_bit, desc->channel[swizzle].shift + desc->channel[swizzle].size);
316    }
317 
318    union {
319       uint8_t ub[16];
320       uint16_t us[8];
321       uint32_t ui[4];
322    } value = {};
323    util_pack_color_union(surface_format, (union util_color*)&value, color);
324 
325    /* Check the cases where all components or bits are either all 0 or all 1. */
326    bool all_bits_are_0 = true;
327    bool all_bits_are_1 = true;
328    bool all_words_are_fp16_1 = false;
329    bool all_words_are_fp32_1 = false;
330 
331    for (unsigned i = start_bit; i < end_bit; i++) {
332       bool bit = value.ub[i / 8] & BITFIELD_BIT(i % 8);
333 
334       all_bits_are_0 &= !bit;
335       all_bits_are_1 &= bit;
336    }
337 
338    if (start_bit % 16 == 0 && end_bit % 16 == 0) {
339       all_words_are_fp16_1 = true;
340       for (unsigned i = start_bit / 16; i < end_bit / 16; i++)
341          all_words_are_fp16_1 &= value.us[i] == 0x3c00;
342    }
343 
344    if (start_bit % 32 == 0 && end_bit % 32 == 0) {
345       all_words_are_fp32_1 = true;
346       for (unsigned i = start_bit / 32; i < end_bit / 32; i++)
347          all_words_are_fp32_1 &= value.ui[i] == 0x3f800000;
348    }
349 
350 #if 0 /* debug code */
351    int i = util_format_get_first_non_void_channel(surface_format);
352    if (desc->channel[i].type == UTIL_FORMAT_TYPE_SIGNED && desc->channel[i].pure_integer) {
353       printf("%i %i %i %i\n", color->i[0], color->i[1], color->i[2], color->i[3]);
354    } else if (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED && desc->channel[i].pure_integer) {
355       printf("%u %u %u %u\n", color->ui[0], color->ui[1], color->ui[2], color->ui[3]);
356    } else {
357       printf("%f %f %f %f\n", color->f[0], color->f[1], color->f[2], color->f[3]);
358    }
359    for (unsigned i = 0; i < end_bit / 8; i++)
360       printf("%02x", value.ub[i]);
361    printf("\n");
362    printf("bits=[%u..%u)%s%s%s%s\n", start_bit, end_bit,
363           all_bits_are_0 ? ", all 0" : "",
364           all_bits_are_1 ? ", all 1" : "",
365           all_words_are_fp16_1 ? ", all fp16 1" : "",
366           all_words_are_fp32_1 ? ", all fp32 1" : "");
367 #endif
368 
369    *clear_value = 0;
370 
371    if (all_bits_are_0 || all_bits_are_1 || all_words_are_fp16_1 || all_words_are_fp32_1) {
372       if (all_bits_are_0)
373          *clear_value = GFX11_DCC_CLEAR_0000;
374       else if (all_bits_are_1)
375          *clear_value = GFX11_DCC_CLEAR_1111_UNORM;
376       else if (all_words_are_fp16_1)
377          *clear_value = GFX11_DCC_CLEAR_1111_FP16;
378       else if (all_words_are_fp32_1)
379          *clear_value = GFX11_DCC_CLEAR_1111_FP32;
380 
381       return true;
382    }
383 
384    /* Check 0001 and 1110 cases. */
385    if (vi_alpha_is_on_msb(sscreen, surface_format)) {
386       if (desc->nr_channels == 2 && desc->channel[0].size == 8) {
387          if (value.ub[0] == 0x00 && value.ub[1] == 0xff) {
388             *clear_value = GFX11_DCC_CLEAR_0001_UNORM;
389             return true;
390          } else if (value.ub[0] == 0xff && value.ub[1] == 0x00) {
391             *clear_value = GFX11_DCC_CLEAR_1110_UNORM;
392             return true;
393          }
394       } else if (desc->nr_channels == 4 && desc->channel[0].size == 8) {
395          if (value.ub[0] == 0x00 && value.ub[1] == 0x00 &&
396              value.ub[2] == 0x00 && value.ub[3] == 0xff) {
397             *clear_value = GFX11_DCC_CLEAR_0001_UNORM;
398             return true;
399          } else if (value.ub[0] == 0xff && value.ub[1] == 0xff &&
400                     value.ub[2] == 0xff && value.ub[3] == 0x00) {
401             *clear_value = GFX11_DCC_CLEAR_1110_UNORM;
402             return true;
403          }
404       } else if (desc->nr_channels == 4 && desc->channel[0].size == 16) {
405          if (value.us[0] == 0x0000 && value.us[1] == 0x0000 &&
406              value.us[2] == 0x0000 && value.us[3] == 0xffff) {
407             *clear_value = GFX11_DCC_CLEAR_0001_UNORM;
408             return true;
409          } else if (value.us[0] == 0xffff && value.us[1] == 0xffff &&
410                     value.us[2] == 0xffff && value.us[3] == 0x0000) {
411             *clear_value = GFX11_DCC_CLEAR_1110_UNORM;
412             return true;
413          }
414       }
415    }
416 
417    return false;
418 }
419 
vi_dcc_get_clear_info(struct si_context * sctx,struct si_texture * tex,unsigned level,unsigned clear_value,struct si_clear_info * out)420 bool vi_dcc_get_clear_info(struct si_context *sctx, struct si_texture *tex, unsigned level,
421                            unsigned clear_value, struct si_clear_info *out)
422 {
423    struct pipe_resource *dcc_buffer = &tex->buffer.b.b;
424    uint64_t dcc_offset = tex->surface.meta_offset;
425    uint32_t clear_size;
426 
427    assert(vi_dcc_enabled(tex, level));
428 
429    if (sctx->gfx_level >= GFX10) {
430       /* 4x and 8x MSAA needs a sophisticated compute shader for
431        * the clear. GFX11 doesn't need that.
432        */
433       if (sctx->gfx_level < GFX11 && tex->buffer.b.b.nr_storage_samples >= 4)
434          return false;
435 
436       unsigned num_layers = util_num_layers(&tex->buffer.b.b, level);
437 
438       if (num_layers == 1) {
439          /* Clear a specific level. */
440          dcc_offset += tex->surface.u.gfx9.meta_levels[level].offset;
441          clear_size = tex->surface.u.gfx9.meta_levels[level].size;
442       } else if (tex->buffer.b.b.last_level == 0) {
443          /* Clear all layers having only 1 level. */
444          clear_size = tex->surface.meta_size;
445       } else {
446          /* Clearing DCC with both multiple levels and multiple layers is not
447           * implemented.
448           */
449          return false;
450       }
451    } else if (sctx->gfx_level == GFX9) {
452       /* TODO: Implement DCC fast clear for level 0 of mipmapped textures. Mipmapped
453        * DCC has to clear a rectangular area of DCC for level 0 (because the whole miptree
454        * is organized in a 2D plane).
455        */
456       if (tex->buffer.b.b.last_level > 0)
457          return false;
458 
459       /* 4x and 8x MSAA need to clear only sample 0 and 1 in a compute shader and leave other
460        * samples untouched. (only the first 2 samples are compressed) */
461       if (tex->buffer.b.b.nr_storage_samples >= 4) {
462          si_init_buffer_clear(out, dcc_buffer, 0, 0, clear_value);
463          out->is_dcc_msaa = true;
464          return true;
465       }
466 
467       clear_size = tex->surface.meta_size;
468    } else {
469       unsigned num_layers = util_num_layers(&tex->buffer.b.b, level);
470 
471       /* If this is 0, fast clear isn't possible. (can occur with MSAA) */
472       if (!tex->surface.u.legacy.color.dcc_level[level].dcc_fast_clear_size)
473          return false;
474 
475       /* Layered 4x and 8x MSAA DCC fast clears need to clear
476        * dcc_fast_clear_size bytes for each layer. A compute shader
477        * would be more efficient than separate per-layer clear operations.
478        */
479       if (tex->buffer.b.b.nr_storage_samples >= 4 && num_layers > 1)
480          return false;
481 
482       dcc_offset += tex->surface.u.legacy.color.dcc_level[level].dcc_offset;
483       clear_size = tex->surface.u.legacy.color.dcc_level[level].dcc_fast_clear_size;
484    }
485 
486    si_init_buffer_clear(out, dcc_buffer, dcc_offset, clear_size, clear_value);
487    return true;
488 }
489 
490 /* Set the same micro tile mode as the destination of the last MSAA resolve.
491  * This allows hitting the MSAA resolve fast path, which requires that both
492  * src and dst micro tile modes match.
493  */
si_set_optimal_micro_tile_mode(struct si_screen * sscreen,struct si_texture * tex)494 static void si_set_optimal_micro_tile_mode(struct si_screen *sscreen, struct si_texture *tex)
495 {
496    if (sscreen->info.gfx_level >= GFX10 || tex->buffer.b.is_shared ||
497        tex->buffer.b.b.nr_samples <= 1 ||
498        tex->surface.micro_tile_mode == tex->last_msaa_resolve_target_micro_mode)
499       return;
500 
501    assert(sscreen->info.gfx_level >= GFX9 ||
502           tex->surface.u.legacy.level[0].mode == RADEON_SURF_MODE_2D);
503    assert(tex->buffer.b.b.last_level == 0);
504 
505    if (sscreen->info.gfx_level >= GFX9) {
506       /* 4K or larger tiles only. 0 is linear. 1-3 are 256B tiles. */
507       assert(tex->surface.u.gfx9.swizzle_mode >= 4);
508 
509       /* If you do swizzle_mode % 4, you'll get:
510        *   0 = Depth
511        *   1 = Standard,
512        *   2 = Displayable
513        *   3 = Rotated
514        *
515        * Depth-sample order isn't allowed:
516        */
517       assert(tex->surface.u.gfx9.swizzle_mode % 4 != 0);
518 
519       switch (tex->last_msaa_resolve_target_micro_mode) {
520       case RADEON_MICRO_MODE_DISPLAY:
521          tex->surface.u.gfx9.swizzle_mode &= ~0x3;
522          tex->surface.u.gfx9.swizzle_mode += 2; /* D */
523          break;
524       case RADEON_MICRO_MODE_STANDARD:
525          tex->surface.u.gfx9.swizzle_mode &= ~0x3;
526          tex->surface.u.gfx9.swizzle_mode += 1; /* S */
527          break;
528       case RADEON_MICRO_MODE_RENDER:
529          tex->surface.u.gfx9.swizzle_mode &= ~0x3;
530          tex->surface.u.gfx9.swizzle_mode += 3; /* R */
531          break;
532       default: /* depth */
533          assert(!"unexpected micro mode");
534          return;
535       }
536    } else if (sscreen->info.gfx_level >= GFX7) {
537       /* These magic numbers were copied from addrlib. It doesn't use
538        * any definitions for them either. They are all 2D_TILED_THIN1
539        * modes with different bpp and micro tile mode.
540        */
541       switch (tex->last_msaa_resolve_target_micro_mode) {
542       case RADEON_MICRO_MODE_DISPLAY:
543          tex->surface.u.legacy.tiling_index[0] = 10;
544          break;
545       case RADEON_MICRO_MODE_STANDARD:
546          tex->surface.u.legacy.tiling_index[0] = 14;
547          break;
548       case RADEON_MICRO_MODE_RENDER:
549          tex->surface.u.legacy.tiling_index[0] = 28;
550          break;
551       default: /* depth, thick */
552          assert(!"unexpected micro mode");
553          return;
554       }
555    } else { /* GFX6 */
556       switch (tex->last_msaa_resolve_target_micro_mode) {
557       case RADEON_MICRO_MODE_DISPLAY:
558          switch (tex->surface.bpe) {
559          case 1:
560             tex->surface.u.legacy.tiling_index[0] = 10;
561             break;
562          case 2:
563             tex->surface.u.legacy.tiling_index[0] = 11;
564             break;
565          default: /* 4, 8 */
566             tex->surface.u.legacy.tiling_index[0] = 12;
567             break;
568          }
569          break;
570       case RADEON_MICRO_MODE_STANDARD:
571          switch (tex->surface.bpe) {
572          case 1:
573             tex->surface.u.legacy.tiling_index[0] = 14;
574             break;
575          case 2:
576             tex->surface.u.legacy.tiling_index[0] = 15;
577             break;
578          case 4:
579             tex->surface.u.legacy.tiling_index[0] = 16;
580             break;
581          default: /* 8, 16 */
582             tex->surface.u.legacy.tiling_index[0] = 17;
583             break;
584          }
585          break;
586       default: /* depth, thick */
587          assert(!"unexpected micro mode");
588          return;
589       }
590    }
591 
592    tex->surface.micro_tile_mode = tex->last_msaa_resolve_target_micro_mode;
593 
594    p_atomic_inc(&sscreen->dirty_tex_counter);
595 }
596 
si_get_htile_clear_value(struct si_texture * tex,float depth)597 static uint32_t si_get_htile_clear_value(struct si_texture *tex, float depth)
598 {
599    /* Maximum 14-bit UINT value. */
600    const uint32_t max_z_value = 0x3FFF;
601 
602    /* For clears, Zmask and Smem will always be set to zero. */
603    const uint32_t zmask = 0;
604    const uint32_t smem  = 0;
605 
606    /* Convert depthValue to 14-bit zmin/zmax uint values. */
607    const uint32_t zmin = lroundf(depth * max_z_value);
608    const uint32_t zmax = zmin;
609 
610    if (tex->htile_stencil_disabled) {
611       /* Z-only HTILE is laid out as follows:
612        * |31     18|17      4|3     0|
613        * +---------+---------+-------+
614        * |  Max Z  |  Min Z  | ZMask |
615        */
616       return ((zmax & 0x3FFF) << 18) |
617              ((zmin & 0x3FFF) << 4) |
618              ((zmask & 0xF) << 0);
619    } else {
620       /* Z+S HTILE is laid out as-follows:
621        * |31       12|11 10|9    8|7   6|5   4|3     0|
622        * +-----------+-----+------+-----+-----+-------+
623        * |  Z Range  |     | SMem | SR1 | SR0 | ZMask |
624        *
625        * The base value for zRange is either zMax or zMin, depending on ZRANGE_PRECISION.
626        * For a fast clear, zMin == zMax == clearValue. This means that the base will
627        * always be the clear value (converted to 14-bit UINT).
628        *
629        * When abs(zMax-zMin) < 16, the delta is equal to the difference. In the case of
630        * fast clears, where zMax == zMin, the delta is always zero.
631        */
632       const uint32_t delta = 0;
633       const uint32_t zrange = (zmax << 6) | delta;
634 
635       /* SResults 0 & 1 are set based on the stencil compare state.
636        * For fast-clear, the default value of sr0 and sr1 are both 0x3.
637        */
638       const uint32_t sresults = 0xf;
639 
640       return ((zrange & 0xFFFFF) << 12) |
641              ((smem & 0x3) <<  8) |
642              ((sresults & 0xF) <<  4) |
643              ((zmask & 0xF) <<  0);
644    }
645 }
646 
si_can_fast_clear_depth(struct si_texture * zstex,unsigned level,float depth,unsigned buffers)647 static bool si_can_fast_clear_depth(struct si_texture *zstex, unsigned level, float depth,
648                                     unsigned buffers)
649 {
650    /* TC-compatible HTILE only supports depth clears to 0 or 1. */
651    return buffers & PIPE_CLEAR_DEPTH &&
652           si_htile_enabled(zstex, level, PIPE_MASK_Z) &&
653           (!zstex->tc_compatible_htile || depth == 0 || depth == 1);
654 }
655 
si_can_fast_clear_stencil(struct si_texture * zstex,unsigned level,uint8_t stencil,unsigned buffers)656 static bool si_can_fast_clear_stencil(struct si_texture *zstex, unsigned level, uint8_t stencil,
657                                       unsigned buffers)
658 {
659    /* TC-compatible HTILE only supports stencil clears to 0. */
660    return buffers & PIPE_CLEAR_STENCIL &&
661           si_htile_enabled(zstex, level, PIPE_MASK_S) &&
662           (!zstex->tc_compatible_htile || stencil == 0);
663 }
664 
si_fast_clear(struct si_context * sctx,unsigned * buffers,const union pipe_color_union * color,float depth,uint8_t stencil)665 static void si_fast_clear(struct si_context *sctx, unsigned *buffers,
666                           const union pipe_color_union *color, float depth, uint8_t stencil)
667 {
668    struct pipe_framebuffer_state *fb = &sctx->framebuffer.state;
669    struct si_clear_info info[8 * 2 + 1]; /* MRTs * (CMASK + DCC) + ZS */
670    unsigned num_clears = 0;
671    unsigned clear_types = 0;
672    unsigned num_pixels = fb->width * fb->height;
673 
674    /* This function is broken in BE, so just disable this path for now */
675 #if UTIL_ARCH_BIG_ENDIAN
676    return;
677 #endif
678 
679    if (sctx->render_cond)
680       return;
681 
682    /* Gather information about what to clear. */
683    unsigned color_buffer_mask = (*buffers & PIPE_CLEAR_COLOR) >> util_logbase2(PIPE_CLEAR_COLOR0);
684    while (color_buffer_mask) {
685       unsigned i = u_bit_scan(&color_buffer_mask);
686 
687       struct si_texture *tex = (struct si_texture *)fb->cbufs[i]->texture;
688       unsigned level = fb->cbufs[i]->u.tex.level;
689       unsigned num_layers = util_num_layers(&tex->buffer.b.b, level);
690 
691       /* the clear is allowed if all layers are bound */
692       if (fb->cbufs[i]->u.tex.first_layer != 0 ||
693           fb->cbufs[i]->u.tex.last_layer != num_layers - 1) {
694          continue;
695       }
696 
697       /* We can change the micro tile mode before a full clear. */
698       /* This is only used for MSAA textures when clearing all layers. */
699       si_set_optimal_micro_tile_mode(sctx->screen, tex);
700 
701       if (tex->swap_rgb_to_bgr_on_next_clear) {
702          assert(!tex->swap_rgb_to_bgr);
703          assert(tex->buffer.b.b.nr_samples >= 2);
704          tex->swap_rgb_to_bgr = true;
705          tex->swap_rgb_to_bgr_on_next_clear = false;
706 
707          /* Update all sampler views and images. */
708          p_atomic_inc(&sctx->screen->dirty_tex_counter);
709       }
710 
711       /* only supported on tiled surfaces */
712       if (tex->surface.is_linear) {
713          continue;
714       }
715 
716       /* Use a slow clear for small surfaces where the cost of
717        * the eliminate pass can be higher than the benefit of fast
718        * clear. The closed driver does this, but the numbers may differ.
719        *
720        * This helps on both dGPUs and APUs, even small APUs like Mullins.
721        */
722       bool fb_too_small = num_pixels * num_layers <= 512 * 512;
723       bool too_small = tex->buffer.b.b.nr_samples <= 1 && fb_too_small;
724       bool eliminate_needed = false;
725       bool fmask_decompress_needed = false;
726 
727       /* Try to clear DCC first, otherwise try CMASK. */
728       if (vi_dcc_enabled(tex, level)) {
729          uint32_t reset_value;
730 
731          if (sctx->screen->debug_flags & DBG(NO_DCC_CLEAR))
732             continue;
733 
734          if (sctx->gfx_level >= GFX11) {
735             if (!gfx11_get_dcc_clear_parameters(sctx->screen, fb->cbufs[i]->format, color,
736                                                 &reset_value))
737                continue;
738          } else {
739             if (!gfx8_get_dcc_clear_parameters(sctx->screen, tex->buffer.b.b.format,
740                                                fb->cbufs[i]->format, color, &reset_value,
741                                                &eliminate_needed))
742                continue;
743          }
744 
745          /* Shared textures can't use fast clear without an explicit flush
746           * because the clear color is not exported.
747           *
748           * Chips without DCC constant encoding must set the clear color registers
749           * correctly even if the fast clear eliminate pass is not needed.
750           */
751          if ((eliminate_needed || !sctx->screen->info.has_dcc_constant_encode) &&
752              tex->buffer.b.is_shared &&
753              !(tex->buffer.external_usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH))
754             continue;
755 
756          if (eliminate_needed && too_small)
757             continue;
758 
759          /* We can clear any level, but we only set up the clear value registers for the first
760           * level. Therefore, all other levels can be cleared only if the clear value registers
761           * are not used, which is only the case with DCC constant encoding and 0/1 clear values.
762           */
763          if (level > 0 && (eliminate_needed || !sctx->screen->info.has_dcc_constant_encode))
764             continue;
765 
766          if (tex->buffer.b.b.nr_samples >= 2 && eliminate_needed &&
767              !sctx->screen->allow_dcc_msaa_clear_to_reg_for_bpp[util_logbase2(tex->surface.bpe)])
768             continue;
769 
770          assert(num_clears < ARRAY_SIZE(info));
771 
772          if (!vi_dcc_get_clear_info(sctx, tex, level, reset_value, &info[num_clears]))
773             continue;
774 
775          num_clears++;
776          clear_types |= SI_CLEAR_TYPE_DCC;
777 
778          si_mark_display_dcc_dirty(sctx, tex);
779 
780          /* DCC fast clear with MSAA should clear CMASK to 0xC. */
781          if (tex->buffer.b.b.nr_samples >= 2 && tex->cmask_buffer) {
782             assert(sctx->gfx_level < GFX11); /* no FMASK/CMASK on GFX11 */
783             assert(num_clears < ARRAY_SIZE(info));
784             si_init_buffer_clear(&info[num_clears++], &tex->cmask_buffer->b.b,
785                                  tex->surface.cmask_offset, tex->surface.cmask_size, 0xCCCCCCCC);
786             clear_types |= SI_CLEAR_TYPE_CMASK;
787             fmask_decompress_needed = true;
788          }
789       } else {
790          /* No CMASK on GFX11. */
791          if (sctx->gfx_level >= GFX11)
792             continue;
793 
794          if (level > 0)
795             continue;
796 
797          /* Shared textures can't use fast clear without an explicit flush
798           * because the clear color is not exported.
799           */
800          if (tex->buffer.b.is_shared &&
801              !(tex->buffer.external_usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH))
802             continue;
803 
804          if (too_small)
805             continue;
806 
807          /* 128-bit formats are unsupported */
808          if (tex->surface.bpe > 8) {
809             continue;
810          }
811 
812          /* RB+ doesn't work with CMASK fast clear on Stoney. */
813          if (sctx->family == CHIP_STONEY)
814             continue;
815 
816          /* Disable fast clear if tex is encrypted */
817          if (tex->buffer.flags & RADEON_FLAG_ENCRYPTED)
818             continue;
819 
820          uint64_t cmask_offset = 0;
821          unsigned clear_size = 0;
822 
823          if (sctx->gfx_level >= GFX10) {
824             assert(level == 0);
825 
826             /* Clearing CMASK with both multiple levels and multiple layers is not
827              * implemented.
828              */
829             if (num_layers > 1 && tex->buffer.b.b.last_level > 0)
830                continue;
831 
832             if (!si_alloc_separate_cmask(sctx->screen, tex))
833                continue;
834 
835             if (num_layers == 1) {
836                /* Clear level 0. */
837                cmask_offset = tex->surface.cmask_offset + tex->surface.u.gfx9.color.cmask_level0.offset;
838                clear_size = tex->surface.u.gfx9.color.cmask_level0.size;
839             } else if (tex->buffer.b.b.last_level == 0) {
840                /* Clear all layers having only 1 level. */
841                cmask_offset = tex->surface.cmask_offset;
842                clear_size = tex->surface.cmask_size;
843             } else {
844                assert(0); /* this is prevented above */
845             }
846          } else if (sctx->gfx_level == GFX9) {
847             /* TODO: Implement CMASK fast clear for level 0 of mipmapped textures. Mipmapped
848              * CMASK has to clear a rectangular area of CMASK for level 0 (because the whole
849              * miptree is organized in a 2D plane).
850              */
851             if (tex->buffer.b.b.last_level > 0)
852                continue;
853 
854             if (!si_alloc_separate_cmask(sctx->screen, tex))
855                continue;
856 
857             cmask_offset = tex->surface.cmask_offset;
858             clear_size = tex->surface.cmask_size;
859          } else {
860             if (!si_alloc_separate_cmask(sctx->screen, tex))
861                continue;
862 
863             /* GFX6-8: This only covers mipmap level 0. */
864             cmask_offset = tex->surface.cmask_offset;
865             clear_size = tex->surface.cmask_size;
866          }
867 
868          /* Do the fast clear. */
869          assert(num_clears < ARRAY_SIZE(info));
870          si_init_buffer_clear(&info[num_clears++], &tex->cmask_buffer->b.b,
871                               cmask_offset, clear_size, 0);
872          clear_types |= SI_CLEAR_TYPE_CMASK;
873          eliminate_needed = true;
874       }
875 
876       if ((eliminate_needed || fmask_decompress_needed) &&
877           !(tex->dirty_level_mask & (1 << level))) {
878          assert(sctx->gfx_level < GFX11); /* no decompression needed on GFX11 */
879          tex->dirty_level_mask |= 1 << level;
880          si_set_sampler_depth_decompress_mask(sctx, tex);
881          p_atomic_inc(&sctx->screen->compressed_colortex_counter);
882       }
883 
884       *buffers &= ~(PIPE_CLEAR_COLOR0 << i);
885 
886       /* Chips with DCC constant encoding don't need to set the clear
887        * color registers for DCC clear values 0 and 1.
888        */
889       if (sctx->screen->info.has_dcc_constant_encode && !eliminate_needed)
890          continue;
891 
892       /* There are no clear color registers on GFX11. */
893       assert(sctx->gfx_level < GFX11);
894 
895       if (si_set_clear_color(tex, fb->cbufs[i]->format, color)) {
896          sctx->framebuffer.dirty_cbufs |= 1 << i;
897          si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
898       }
899    }
900 
901    /* Depth/stencil clears. */
902    struct pipe_surface *zsbuf = fb->zsbuf;
903    struct si_texture *zstex = zsbuf ? (struct si_texture *)zsbuf->texture : NULL;
904    unsigned zs_num_layers = zstex ? util_num_layers(&zstex->buffer.b.b, zsbuf->u.tex.level) : 0;
905 
906    if (zstex && zsbuf->u.tex.first_layer == 0 &&
907        zsbuf->u.tex.last_layer == zs_num_layers - 1 &&
908        si_htile_enabled(zstex, zsbuf->u.tex.level, PIPE_MASK_ZS)) {
909       unsigned level = zsbuf->u.tex.level;
910       bool update_db_depth_clear = false;
911       bool update_db_stencil_clear = false;
912       bool fb_too_small = num_pixels * zs_num_layers <= 512 * 512;
913 
914       /* Transition from TC-incompatible to TC-compatible HTILE if requested. */
915       if (zstex->enable_tc_compatible_htile_next_clear) {
916           /* If both depth and stencil are present, they must be cleared together. */
917          if ((*buffers & PIPE_CLEAR_DEPTHSTENCIL) == PIPE_CLEAR_DEPTHSTENCIL ||
918              (*buffers & PIPE_CLEAR_DEPTH && (!zstex->surface.has_stencil ||
919                                               zstex->htile_stencil_disabled))) {
920             /* The conversion from TC-incompatible to TC-compatible can only be done in one clear. */
921             assert(zstex->buffer.b.b.last_level == 0);
922             assert(!zstex->tc_compatible_htile);
923 
924             /* Enable TC-compatible HTILE. */
925             zstex->enable_tc_compatible_htile_next_clear = false;
926             zstex->tc_compatible_htile = true;
927 
928             /* Update the framebuffer state to reflect the change. */
929             sctx->framebuffer.DB_has_shader_readable_metadata = true;
930             sctx->framebuffer.dirty_zsbuf = true;
931             si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
932 
933             /* Update all sampler views and shader images in all contexts. */
934             p_atomic_inc(&sctx->screen->dirty_tex_counter);
935 
936             /* Perform the clear here if possible, else clear to uncompressed. */
937             uint32_t clear_value;
938 
939             if (zstex->htile_stencil_disabled || !zstex->surface.has_stencil) {
940                if (si_can_fast_clear_depth(zstex, level, depth, *buffers)) {
941                   /* Z-only clear. */
942                   clear_value = si_get_htile_clear_value(zstex, depth);
943                   *buffers &= ~PIPE_CLEAR_DEPTH;
944                   zstex->depth_cleared_level_mask_once |= BITFIELD_BIT(level);
945                   zstex->depth_cleared_level_mask |= BITFIELD_BIT(level);
946                   update_db_depth_clear = true;
947                }
948             } else if ((*buffers & PIPE_CLEAR_DEPTHSTENCIL) == PIPE_CLEAR_DEPTHSTENCIL) {
949                if (si_can_fast_clear_depth(zstex, level, depth, *buffers) &&
950                    si_can_fast_clear_stencil(zstex, level, stencil, *buffers)) {
951                   /* Combined Z+S clear. */
952                   clear_value = si_get_htile_clear_value(zstex, depth);
953                   *buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
954                   zstex->depth_cleared_level_mask_once |= BITFIELD_BIT(level);
955                   zstex->depth_cleared_level_mask |= BITFIELD_BIT(level);
956                   zstex->stencil_cleared_level_mask_once |= BITFIELD_BIT(level);
957                   update_db_depth_clear = true;
958                   update_db_stencil_clear = true;
959                }
960             }
961 
962             if (!update_db_depth_clear) {
963                /* Clear to uncompressed, so that it doesn't contain values incompatible
964                 * with the new TC-compatible HTILE setting.
965                 *
966                 * 0xfffff30f = uncompressed Z + S
967                 * 0xfffc000f = uncompressed Z only
968                 */
969                clear_value = !zstex->htile_stencil_disabled ? 0xfffff30f : 0xfffc000f;
970             }
971 
972             zstex->need_flush_after_depth_decompression = sctx->gfx_level == GFX10_3;
973 
974             assert(num_clears < ARRAY_SIZE(info));
975             si_init_buffer_clear(&info[num_clears++], &zstex->buffer.b.b,
976                                  zstex->surface.meta_offset, zstex->surface.meta_size, clear_value);
977             clear_types |= SI_CLEAR_TYPE_HTILE;
978          }
979       } else if (num_clears || !fb_too_small) {
980          /* This is where the HTILE buffer clear is done.
981           *
982           * If there is no clear scheduled and the framebuffer size is too small, we should use
983           * the draw-based clear that is without waits. If there is some other clear scheduled,
984           * we will have to wait anyway, so add the HTILE buffer clear to the batch here.
985           * If the framebuffer size is large enough, use this codepath too.
986           */
987          uint64_t htile_offset = zstex->surface.meta_offset;
988          unsigned htile_size = 0;
989 
990          /* Determine the HTILE subset to clear. */
991          if (sctx->gfx_level >= GFX10) {
992             /* This can only clear a layered texture with 1 level or a mipmap texture
993              * with 1 layer. Other cases are unimplemented.
994              */
995             if (zs_num_layers == 1) {
996                /* Clear a specific level. */
997                htile_offset += zstex->surface.u.gfx9.meta_levels[level].offset;
998                htile_size = zstex->surface.u.gfx9.meta_levels[level].size;
999             } else if (zstex->buffer.b.b.last_level == 0) {
1000                /* Clear all layers having only 1 level. */
1001                htile_size = zstex->surface.meta_size;
1002             }
1003          } else {
1004             /* This can only clear a layered texture with 1 level. Other cases are
1005              * unimplemented.
1006              */
1007             if (zstex->buffer.b.b.last_level == 0)
1008                htile_size = zstex->surface.meta_size;
1009          }
1010 
1011          /* Perform the clear if it's possible. */
1012          if (zstex->htile_stencil_disabled || !zstex->surface.has_stencil) {
1013             if (htile_size &&
1014                 si_can_fast_clear_depth(zstex, level, depth, *buffers)) {
1015                /* Z-only clear. */
1016                assert(num_clears < ARRAY_SIZE(info));
1017                si_init_buffer_clear(&info[num_clears++], &zstex->buffer.b.b, htile_offset,
1018                                     htile_size, si_get_htile_clear_value(zstex, depth));
1019                clear_types |= SI_CLEAR_TYPE_HTILE;
1020                *buffers &= ~PIPE_CLEAR_DEPTH;
1021                zstex->depth_cleared_level_mask_once |= BITFIELD_BIT(level);
1022                zstex->depth_cleared_level_mask |= BITFIELD_BIT(level);
1023                update_db_depth_clear = true;
1024             }
1025          } else if ((*buffers & PIPE_CLEAR_DEPTHSTENCIL) == PIPE_CLEAR_DEPTHSTENCIL) {
1026             if (htile_size &&
1027                 si_can_fast_clear_depth(zstex, level, depth, *buffers) &&
1028                 si_can_fast_clear_stencil(zstex, level, stencil, *buffers)) {
1029                /* Combined Z+S clear. */
1030                assert(num_clears < ARRAY_SIZE(info));
1031                si_init_buffer_clear(&info[num_clears++], &zstex->buffer.b.b, htile_offset,
1032                                     htile_size, si_get_htile_clear_value(zstex, depth));
1033                clear_types |= SI_CLEAR_TYPE_HTILE;
1034                *buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
1035                zstex->depth_cleared_level_mask_once |= BITFIELD_BIT(level);
1036                zstex->depth_cleared_level_mask |= BITFIELD_BIT(level);
1037                zstex->stencil_cleared_level_mask_once |= BITFIELD_BIT(level);
1038                update_db_depth_clear = true;
1039                update_db_stencil_clear = true;
1040             }
1041          } else {
1042             /* Z-only or S-only clear when both Z/S are present using a read-modify-write
1043              * compute shader.
1044              *
1045              * If we get both clears but only one of them can be fast-cleared, we use
1046              * the draw-based fast clear to do both at the same time.
1047              */
1048             const uint32_t htile_depth_writemask = 0xfffffc0f;
1049             const uint32_t htile_stencil_writemask = 0x000003f0;
1050 
1051             if (htile_size &&
1052                 !(*buffers & PIPE_CLEAR_STENCIL) &&
1053                 si_can_fast_clear_depth(zstex, level, depth, *buffers)) {
1054                /* Z-only clear with stencil left intact. */
1055                assert(num_clears < ARRAY_SIZE(info));
1056                si_init_buffer_clear_rmw(&info[num_clears++], &zstex->buffer.b.b, htile_offset,
1057                                         htile_size, si_get_htile_clear_value(zstex, depth),
1058                                         htile_depth_writemask);
1059                clear_types |= SI_CLEAR_TYPE_HTILE;
1060                *buffers &= ~PIPE_CLEAR_DEPTH;
1061                zstex->depth_cleared_level_mask_once |= BITFIELD_BIT(level);
1062                zstex->depth_cleared_level_mask |= BITFIELD_BIT(level);
1063                update_db_depth_clear = true;
1064             } else if (htile_size &&
1065                        !(*buffers & PIPE_CLEAR_DEPTH) &&
1066                        si_can_fast_clear_stencil(zstex, level, stencil, *buffers)) {
1067                /* Stencil-only clear with depth left intact. */
1068                assert(num_clears < ARRAY_SIZE(info));
1069                si_init_buffer_clear_rmw(&info[num_clears++], &zstex->buffer.b.b, htile_offset,
1070                                         htile_size, si_get_htile_clear_value(zstex, depth),
1071                                         htile_stencil_writemask);
1072                clear_types |= SI_CLEAR_TYPE_HTILE;
1073                *buffers &= ~PIPE_CLEAR_STENCIL;
1074                zstex->stencil_cleared_level_mask_once |= BITFIELD_BIT(level);
1075                update_db_stencil_clear = true;
1076             }
1077          }
1078 
1079          zstex->need_flush_after_depth_decompression = update_db_depth_clear && sctx->gfx_level == GFX10_3;
1080 
1081          /* Update DB_DEPTH_CLEAR. */
1082          if (update_db_depth_clear &&
1083              zstex->depth_clear_value[level] != (float)depth) {
1084             zstex->depth_clear_value[level] = depth;
1085             sctx->framebuffer.dirty_zsbuf = true;
1086             si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
1087          }
1088 
1089          /* Update DB_STENCIL_CLEAR. */
1090          if (update_db_stencil_clear &&
1091              zstex->stencil_clear_value[level] != stencil) {
1092             zstex->stencil_clear_value[level] = stencil;
1093             sctx->framebuffer.dirty_zsbuf = true;
1094             si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
1095          }
1096       }
1097    }
1098 
1099    si_execute_clears(sctx, info, num_clears, clear_types);
1100 }
1101 
si_clear(struct pipe_context * ctx,unsigned buffers,const struct pipe_scissor_state * scissor_state,const union pipe_color_union * color,double depth,unsigned stencil)1102 static void si_clear(struct pipe_context *ctx, unsigned buffers,
1103                      const struct pipe_scissor_state *scissor_state,
1104                      const union pipe_color_union *color, double depth, unsigned stencil)
1105 {
1106    struct si_context *sctx = (struct si_context *)ctx;
1107    struct pipe_framebuffer_state *fb = &sctx->framebuffer.state;
1108    struct pipe_surface *zsbuf = fb->zsbuf;
1109    struct si_texture *zstex = zsbuf ? (struct si_texture *)zsbuf->texture : NULL;
1110    bool needs_db_flush = false;
1111 
1112    /* Unset clear flags for non-existent buffers. */
1113    for (unsigned i = 0; i < 8; i++) {
1114       if (i >= fb->nr_cbufs || !fb->cbufs[i])
1115          buffers &= ~(PIPE_CLEAR_COLOR0 << i);
1116    }
1117    if (!zsbuf)
1118       buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
1119    else if (!util_format_has_stencil(util_format_description(zsbuf->format)))
1120       buffers &= ~PIPE_CLEAR_STENCIL;
1121 
1122    si_fast_clear(sctx, &buffers, color, depth, stencil);
1123    if (!buffers)
1124       return; /* all buffers have been cleared */
1125 
1126    if (buffers & PIPE_CLEAR_COLOR) {
1127       /* These buffers cannot use fast clear, make sure to disable expansion. */
1128       unsigned color_buffer_mask = (buffers & PIPE_CLEAR_COLOR) >> util_logbase2(PIPE_CLEAR_COLOR0);
1129       while (color_buffer_mask) {
1130          unsigned i = u_bit_scan(&color_buffer_mask);
1131          struct si_texture *tex = (struct si_texture *)fb->cbufs[i]->texture;
1132          if (tex->surface.fmask_size == 0)
1133             tex->dirty_level_mask &= ~(1 << fb->cbufs[i]->u.tex.level);
1134       }
1135    }
1136 
1137    if (zstex && zsbuf->u.tex.first_layer == 0 &&
1138        zsbuf->u.tex.last_layer == util_max_layer(&zstex->buffer.b.b, 0)) {
1139       unsigned level = zsbuf->u.tex.level;
1140 
1141       if (si_can_fast_clear_depth(zstex, level, depth, buffers)) {
1142          /* Need to disable EXPCLEAR temporarily if clearing
1143           * to a new value. */
1144          if (!(zstex->depth_cleared_level_mask_once & BITFIELD_BIT(level)) ||
1145              zstex->depth_clear_value[level] != depth) {
1146             sctx->db_depth_disable_expclear = true;
1147          }
1148 
1149          if (zstex->depth_clear_value[level] != (float)depth) {
1150             if ((zstex->depth_clear_value[level] != 0) != (depth != 0)) {
1151                /* ZRANGE_PRECISION register of a bound surface will change so we
1152                 * must flush the DB caches. */
1153                needs_db_flush = true;
1154             }
1155             /* Update DB_DEPTH_CLEAR. */
1156             zstex->depth_clear_value[level] = depth;
1157             sctx->framebuffer.dirty_zsbuf = true;
1158             si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
1159          }
1160          sctx->db_depth_clear = true;
1161          si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1162       }
1163 
1164       if (si_can_fast_clear_stencil(zstex, level, stencil, buffers)) {
1165          stencil &= 0xff;
1166 
1167          /* Need to disable EXPCLEAR temporarily if clearing
1168           * to a new value. */
1169          if (!(zstex->stencil_cleared_level_mask_once & BITFIELD_BIT(level)) ||
1170              zstex->stencil_clear_value[level] != stencil) {
1171             sctx->db_stencil_disable_expclear = true;
1172          }
1173 
1174          if (zstex->stencil_clear_value[level] != (uint8_t)stencil) {
1175             /* Update DB_STENCIL_CLEAR. */
1176             zstex->stencil_clear_value[level] = stencil;
1177             sctx->framebuffer.dirty_zsbuf = true;
1178             si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
1179          }
1180          sctx->db_stencil_clear = true;
1181          si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1182       }
1183 
1184       if (needs_db_flush)
1185          sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_DB;
1186    }
1187 
1188    if (unlikely(sctx->thread_trace_enabled)) {
1189       if (buffers & PIPE_CLEAR_COLOR)
1190          sctx->sqtt_next_event = EventCmdClearColorImage;
1191       else if (buffers & PIPE_CLEAR_DEPTHSTENCIL)
1192          sctx->sqtt_next_event = EventCmdClearDepthStencilImage;
1193    }
1194 
1195    si_blitter_begin(sctx, SI_CLEAR);
1196    util_blitter_clear(sctx->blitter, fb->width, fb->height, util_framebuffer_get_num_layers(fb),
1197                       buffers, color, depth, stencil, sctx->framebuffer.nr_samples > 1);
1198    si_blitter_end(sctx);
1199 
1200    if (sctx->db_depth_clear) {
1201       sctx->db_depth_clear = false;
1202       sctx->db_depth_disable_expclear = false;
1203       zstex->depth_cleared_level_mask_once |= BITFIELD_BIT(zsbuf->u.tex.level);
1204       zstex->depth_cleared_level_mask |= BITFIELD_BIT(zsbuf->u.tex.level);
1205       si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1206    }
1207 
1208    if (sctx->db_stencil_clear) {
1209       sctx->db_stencil_clear = false;
1210       sctx->db_stencil_disable_expclear = false;
1211       zstex->stencil_cleared_level_mask_once |= BITFIELD_BIT(zsbuf->u.tex.level);
1212       si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1213    }
1214 }
1215 
si_try_normal_clear(struct si_context * sctx,struct pipe_surface * dst,unsigned dstx,unsigned dsty,unsigned width,unsigned height,bool render_condition_enabled,unsigned buffers,const union pipe_color_union * color,float depth,unsigned stencil)1216 static bool si_try_normal_clear(struct si_context *sctx, struct pipe_surface *dst,
1217                                 unsigned dstx, unsigned dsty, unsigned width, unsigned height,
1218                                 bool render_condition_enabled, unsigned buffers,
1219                                 const union pipe_color_union *color,
1220                                 float depth, unsigned stencil)
1221 {
1222    /* This is worth it only if it's a whole image clear, so that we just clear DCC/HTILE. */
1223    if (dstx == 0 && dsty == 0 &&
1224        width == dst->width &&
1225        height == dst->height &&
1226        dst->u.tex.first_layer == 0 &&
1227        dst->u.tex.last_layer == util_max_layer(dst->texture, dst->u.tex.level) &&
1228        /* pipe->clear honors render_condition, so only use it if it's unset or if it's set and enabled. */
1229        (!sctx->render_cond || render_condition_enabled) &&
1230        sctx->has_graphics) {
1231       struct pipe_context *ctx = &sctx->b;
1232       struct pipe_framebuffer_state saved_fb = {}, fb = {};
1233 
1234       util_copy_framebuffer_state(&saved_fb, &sctx->framebuffer.state);
1235 
1236       if (buffers & PIPE_CLEAR_COLOR) {
1237          fb.cbufs[0] = dst;
1238          fb.nr_cbufs = 1;
1239       } else {
1240          fb.zsbuf = dst;
1241       }
1242 
1243       fb.width = dst->width;
1244       fb.height = dst->height;
1245 
1246       ctx->set_framebuffer_state(ctx, &fb);
1247       ctx->clear(ctx, buffers, NULL, color, depth, stencil);
1248       ctx->set_framebuffer_state(ctx, &saved_fb);
1249 
1250       util_copy_framebuffer_state(&saved_fb, NULL);
1251 
1252       return true;
1253    }
1254 
1255    return false;
1256 }
1257 
si_clear_render_target(struct pipe_context * ctx,struct pipe_surface * dst,const union pipe_color_union * color,unsigned dstx,unsigned dsty,unsigned width,unsigned height,bool render_condition_enabled)1258 static void si_clear_render_target(struct pipe_context *ctx, struct pipe_surface *dst,
1259                                    const union pipe_color_union *color, unsigned dstx,
1260                                    unsigned dsty, unsigned width, unsigned height,
1261                                    bool render_condition_enabled)
1262 {
1263    struct si_context *sctx = (struct si_context *)ctx;
1264    struct si_texture *sdst = (struct si_texture *)dst->texture;
1265 
1266    /* Fast path that just clears DCC. */
1267    if (si_try_normal_clear(sctx, dst, dstx, dsty, width, height, render_condition_enabled,
1268                            PIPE_CLEAR_COLOR0, color, 0, 0))
1269       return;
1270 
1271    if (dst->texture->nr_samples <= 1 &&
1272        (sctx->gfx_level >= GFX10 || !vi_dcc_enabled(sdst, dst->u.tex.level))) {
1273       si_compute_clear_render_target(ctx, dst, color, dstx, dsty, width, height,
1274                                      render_condition_enabled);
1275       return;
1276    }
1277 
1278    si_blitter_begin(sctx,
1279                     SI_CLEAR_SURFACE | (render_condition_enabled ? 0 : SI_DISABLE_RENDER_COND));
1280    util_blitter_clear_render_target(sctx->blitter, dst, color, dstx, dsty, width, height);
1281    si_blitter_end(sctx);
1282 }
1283 
si_clear_depth_stencil(struct pipe_context * ctx,struct pipe_surface * dst,unsigned clear_flags,double depth,unsigned stencil,unsigned dstx,unsigned dsty,unsigned width,unsigned height,bool render_condition_enabled)1284 static void si_clear_depth_stencil(struct pipe_context *ctx, struct pipe_surface *dst,
1285                                    unsigned clear_flags, double depth, unsigned stencil,
1286                                    unsigned dstx, unsigned dsty, unsigned width, unsigned height,
1287                                    bool render_condition_enabled)
1288 {
1289    struct si_context *sctx = (struct si_context *)ctx;
1290    union pipe_color_union unused = {};
1291 
1292    /* Fast path that just clears HTILE. */
1293    if (si_try_normal_clear(sctx, dst, dstx, dsty, width, height, render_condition_enabled,
1294                            clear_flags, &unused, depth, stencil))
1295       return;
1296 
1297    si_blitter_begin(sctx,
1298                     SI_CLEAR_SURFACE | (render_condition_enabled ? 0 : SI_DISABLE_RENDER_COND));
1299    util_blitter_clear_depth_stencil(sctx->blitter, dst, clear_flags, depth, stencil, dstx, dsty,
1300                                     width, height);
1301    si_blitter_end(sctx);
1302 }
1303 
si_clear_texture(struct pipe_context * pipe,struct pipe_resource * tex,unsigned level,const struct pipe_box * box,const void * data)1304 static void si_clear_texture(struct pipe_context *pipe, struct pipe_resource *tex, unsigned level,
1305                              const struct pipe_box *box, const void *data)
1306 {
1307    struct pipe_screen *screen = pipe->screen;
1308    struct si_texture *stex = (struct si_texture *)tex;
1309    struct pipe_surface tmpl = {{0}};
1310    struct pipe_surface *sf;
1311 
1312    tmpl.format = tex->format;
1313    tmpl.u.tex.first_layer = box->z;
1314    tmpl.u.tex.last_layer = box->z + box->depth - 1;
1315    tmpl.u.tex.level = level;
1316    sf = pipe->create_surface(pipe, tex, &tmpl);
1317    if (!sf)
1318       return;
1319 
1320    if (stex->is_depth) {
1321       unsigned clear;
1322       float depth;
1323       uint8_t stencil = 0;
1324 
1325       /* Depth is always present. */
1326       clear = PIPE_CLEAR_DEPTH;
1327       util_format_unpack_z_float(tex->format, &depth, data, 1);
1328 
1329       if (stex->surface.has_stencil) {
1330          clear |= PIPE_CLEAR_STENCIL;
1331          util_format_unpack_s_8uint(tex->format, &stencil, data, 1);
1332       }
1333 
1334       si_clear_depth_stencil(pipe, sf, clear, depth, stencil, box->x, box->y, box->width,
1335                              box->height, false);
1336    } else {
1337       union pipe_color_union color;
1338 
1339       util_format_unpack_rgba(tex->format, color.ui, data, 1);
1340 
1341       if (screen->is_format_supported(screen, tex->format, tex->target, 0, 0,
1342                                       PIPE_BIND_RENDER_TARGET)) {
1343          si_clear_render_target(pipe, sf, &color, box->x, box->y, box->width, box->height, false);
1344       } else {
1345          /* Software fallback - just for R9G9B9E5_FLOAT */
1346          util_clear_render_target(pipe, sf, &color, box->x, box->y, box->width, box->height);
1347       }
1348    }
1349    pipe_surface_reference(&sf, NULL);
1350 }
1351 
si_init_clear_functions(struct si_context * sctx)1352 void si_init_clear_functions(struct si_context *sctx)
1353 {
1354    sctx->b.clear_render_target = si_clear_render_target;
1355    sctx->b.clear_texture = si_clear_texture;
1356 
1357    if (sctx->has_graphics) {
1358       sctx->b.clear = si_clear;
1359       sctx->b.clear_depth_stencil = si_clear_depth_stencil;
1360    }
1361 }
1362