1/*
2 * Copyright © 2013 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 "util/ralloc.h"
25
26#include "main/macros.h" /* Needed for MAX3 and MAX2 for format_rgb9e5 */
27#include "util/format_rgb9e5.h"
28#include "util/format_srgb.h"
29
30#include "blorp_priv.h"
31#include "compiler/brw_eu_defines.h"
32
33#include "blorp_nir_builder.h"
34
35#define FILE_DEBUG_FLAG DEBUG_BLORP
36
37struct brw_blorp_const_color_prog_key
38{
39   enum blorp_shader_type shader_type; /* Must be BLORP_SHADER_TYPE_CLEAR */
40   bool use_simd16_replicated_data;
41   bool clear_rgb_as_red;
42   bool pad[3];
43};
44
45static bool
46blorp_params_get_clear_kernel(struct blorp_batch *batch,
47                              struct blorp_params *params,
48                              bool use_replicated_data,
49                              bool clear_rgb_as_red)
50{
51   struct blorp_context *blorp = batch->blorp;
52
53   const struct brw_blorp_const_color_prog_key blorp_key = {
54      .shader_type = BLORP_SHADER_TYPE_CLEAR,
55      .use_simd16_replicated_data = use_replicated_data,
56      .clear_rgb_as_red = clear_rgb_as_red,
57   };
58
59   if (blorp->lookup_shader(batch, &blorp_key, sizeof(blorp_key),
60                            &params->wm_prog_kernel, &params->wm_prog_data))
61      return true;
62
63   void *mem_ctx = ralloc_context(NULL);
64
65   nir_builder b;
66   blorp_nir_init_shader(&b, mem_ctx, MESA_SHADER_FRAGMENT, "BLORP-clear");
67
68   nir_variable *v_color =
69      BLORP_CREATE_NIR_INPUT(b.shader, clear_color, glsl_vec4_type());
70   nir_ssa_def *color = nir_load_var(&b, v_color);
71
72   if (clear_rgb_as_red) {
73      nir_variable *frag_coord =
74         nir_variable_create(b.shader, nir_var_shader_in,
75                             glsl_vec4_type(), "gl_FragCoord");
76      frag_coord->data.location = VARYING_SLOT_POS;
77
78      nir_ssa_def *pos = nir_f2i32(&b, nir_load_var(&b, frag_coord));
79      nir_ssa_def *comp = nir_umod(&b, nir_channel(&b, pos, 0),
80                                       nir_imm_int(&b, 3));
81      nir_ssa_def *color_component =
82         nir_bcsel(&b, nir_ieq(&b, comp, nir_imm_int(&b, 0)),
83                       nir_channel(&b, color, 0),
84                       nir_bcsel(&b, nir_ieq(&b, comp, nir_imm_int(&b, 1)),
85                                     nir_channel(&b, color, 1),
86                                     nir_channel(&b, color, 2)));
87
88      nir_ssa_def *u = nir_ssa_undef(&b, 1, 32);
89      color = nir_vec4(&b, color_component, u, u, u);
90   }
91
92   nir_variable *frag_color = nir_variable_create(b.shader, nir_var_shader_out,
93                                                  glsl_vec4_type(),
94                                                  "gl_FragColor");
95   frag_color->data.location = FRAG_RESULT_COLOR;
96   nir_store_var(&b, frag_color, color, 0xf);
97
98   struct brw_wm_prog_key wm_key;
99   brw_blorp_init_wm_prog_key(&wm_key);
100
101   struct brw_wm_prog_data prog_data;
102   const unsigned *program =
103      blorp_compile_fs(blorp, mem_ctx, b.shader, &wm_key, use_replicated_data,
104                       &prog_data);
105
106   bool result =
107      blorp->upload_shader(batch, &blorp_key, sizeof(blorp_key),
108                           program, prog_data.base.program_size,
109                           &prog_data.base, sizeof(prog_data),
110                           &params->wm_prog_kernel, &params->wm_prog_data);
111
112   ralloc_free(mem_ctx);
113   return result;
114}
115
116struct layer_offset_vs_key {
117   enum blorp_shader_type shader_type;
118   unsigned num_inputs;
119};
120
121/* In the case of doing attachment clears, we are using a surface state that
122 * is handed to us so we can't set (and don't even know) the base array layer.
123 * In order to do a layered clear in this scenario, we need some way of adding
124 * the base array layer to the instance id.  Unfortunately, our hardware has
125 * no real concept of "base instance", so we have to do it manually in a
126 * vertex shader.
127 */
128static bool
129blorp_params_get_layer_offset_vs(struct blorp_batch *batch,
130                                 struct blorp_params *params)
131{
132   struct blorp_context *blorp = batch->blorp;
133   struct layer_offset_vs_key blorp_key = {
134      .shader_type = BLORP_SHADER_TYPE_LAYER_OFFSET_VS,
135   };
136
137   if (params->wm_prog_data)
138      blorp_key.num_inputs = params->wm_prog_data->num_varying_inputs;
139
140   if (blorp->lookup_shader(batch, &blorp_key, sizeof(blorp_key),
141                            &params->vs_prog_kernel, &params->vs_prog_data))
142      return true;
143
144   void *mem_ctx = ralloc_context(NULL);
145
146   nir_builder b;
147   blorp_nir_init_shader(&b, mem_ctx, MESA_SHADER_VERTEX, "BLORP-layer-offset-vs");
148
149   const struct glsl_type *uvec4_type = glsl_vector_type(GLSL_TYPE_UINT, 4);
150
151   /* First we deal with the header which has instance and base instance */
152   nir_variable *a_header = nir_variable_create(b.shader, nir_var_shader_in,
153                                                uvec4_type, "header");
154   a_header->data.location = VERT_ATTRIB_GENERIC0;
155
156   nir_variable *v_layer = nir_variable_create(b.shader, nir_var_shader_out,
157                                               glsl_int_type(), "layer_id");
158   v_layer->data.location = VARYING_SLOT_LAYER;
159
160   /* Compute the layer id */
161   nir_ssa_def *header = nir_load_var(&b, a_header);
162   nir_ssa_def *base_layer = nir_channel(&b, header, 0);
163   nir_ssa_def *instance = nir_channel(&b, header, 1);
164   nir_store_var(&b, v_layer, nir_iadd(&b, instance, base_layer), 0x1);
165
166   /* Then we copy the vertex from the next slot to VARYING_SLOT_POS */
167   nir_variable *a_vertex = nir_variable_create(b.shader, nir_var_shader_in,
168                                                glsl_vec4_type(), "a_vertex");
169   a_vertex->data.location = VERT_ATTRIB_GENERIC1;
170
171   nir_variable *v_pos = nir_variable_create(b.shader, nir_var_shader_out,
172                                             glsl_vec4_type(), "v_pos");
173   v_pos->data.location = VARYING_SLOT_POS;
174
175   nir_copy_var(&b, v_pos, a_vertex);
176
177   /* Then we copy everything else */
178   for (unsigned i = 0; i < blorp_key.num_inputs; i++) {
179      nir_variable *a_in = nir_variable_create(b.shader, nir_var_shader_in,
180                                               uvec4_type, "input");
181      a_in->data.location = VERT_ATTRIB_GENERIC2 + i;
182
183      nir_variable *v_out = nir_variable_create(b.shader, nir_var_shader_out,
184                                                uvec4_type, "output");
185      v_out->data.location = VARYING_SLOT_VAR0 + i;
186
187      nir_copy_var(&b, v_out, a_in);
188   }
189
190   struct brw_vs_prog_data vs_prog_data;
191   memset(&vs_prog_data, 0, sizeof(vs_prog_data));
192
193   const unsigned *program =
194      blorp_compile_vs(blorp, mem_ctx, b.shader, &vs_prog_data);
195
196   bool result =
197      blorp->upload_shader(batch, &blorp_key, sizeof(blorp_key),
198                           program, vs_prog_data.base.base.program_size,
199                           &vs_prog_data.base.base, sizeof(vs_prog_data),
200                           &params->vs_prog_kernel, &params->vs_prog_data);
201
202   ralloc_free(mem_ctx);
203   return result;
204}
205
206/* The x0, y0, x1, and y1 parameters must already be populated with the render
207 * area of the framebuffer to be cleared.
208 */
209static void
210get_fast_clear_rect(const struct isl_device *dev,
211                    const struct isl_surf *aux_surf,
212                    unsigned *x0, unsigned *y0,
213                    unsigned *x1, unsigned *y1)
214{
215   unsigned int x_align, y_align;
216   unsigned int x_scaledown, y_scaledown;
217
218   /* Only single sampled surfaces need to (and actually can) be resolved. */
219   if (aux_surf->usage == ISL_SURF_USAGE_CCS_BIT) {
220      /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
221       * Target(s)", beneath the "Fast Color Clear" bullet (p327):
222       *
223       *     Clear pass must have a clear rectangle that must follow
224       *     alignment rules in terms of pixels and lines as shown in the
225       *     table below. Further, the clear-rectangle height and width
226       *     must be multiple of the following dimensions. If the height
227       *     and width of the render target being cleared do not meet these
228       *     requirements, an MCS buffer can be created such that it
229       *     follows the requirement and covers the RT.
230       *
231       * The alignment size in the table that follows is related to the
232       * alignment size that is baked into the CCS surface format but with X
233       * alignment multiplied by 16 and Y alignment multiplied by 32.
234       */
235      x_align = isl_format_get_layout(aux_surf->format)->bw;
236      y_align = isl_format_get_layout(aux_surf->format)->bh;
237
238      x_align *= 16;
239
240      /* SKL+ line alignment requirement for Y-tiled are half those of the prior
241       * generations.
242       */
243      if (dev->info->gen >= 9)
244         y_align *= 16;
245      else
246         y_align *= 32;
247
248      /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
249       * Target(s)", beneath the "Fast Color Clear" bullet (p327):
250       *
251       *     In order to optimize the performance MCS buffer (when bound to
252       *     1X RT) clear similarly to MCS buffer clear for MSRT case,
253       *     clear rect is required to be scaled by the following factors
254       *     in the horizontal and vertical directions:
255       *
256       * The X and Y scale down factors in the table that follows are each
257       * equal to half the alignment value computed above.
258       */
259      x_scaledown = x_align / 2;
260      y_scaledown = y_align / 2;
261
262      /* From BSpec: 3D-Media-GPGPU Engine > 3D Pipeline > Pixel > Pixel
263       * Backend > MCS Buffer for Render Target(s) [DevIVB+] > Table "Color
264       * Clear of Non-MultiSampled Render Target Restrictions":
265       *
266       *   Clear rectangle must be aligned to two times the number of
267       *   pixels in the table shown below due to 16x16 hashing across the
268       *   slice.
269       */
270      x_align *= 2;
271      y_align *= 2;
272   } else {
273      assert(aux_surf->usage == ISL_SURF_USAGE_MCS_BIT);
274
275      /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
276       * Target(s)", beneath the "MSAA Compression" bullet (p326):
277       *
278       *     Clear pass for this case requires that scaled down primitive
279       *     is sent down with upper left co-ordinate to coincide with
280       *     actual rectangle being cleared. For MSAA, clear rectangle’s
281       *     height and width need to as show in the following table in
282       *     terms of (width,height) of the RT.
283       *
284       *     MSAA  Width of Clear Rect  Height of Clear Rect
285       *      2X     Ceil(1/8*width)      Ceil(1/2*height)
286       *      4X     Ceil(1/8*width)      Ceil(1/2*height)
287       *      8X     Ceil(1/2*width)      Ceil(1/2*height)
288       *     16X         width            Ceil(1/2*height)
289       *
290       * The text "with upper left co-ordinate to coincide with actual
291       * rectangle being cleared" is a little confusing--it seems to imply
292       * that to clear a rectangle from (x,y) to (x+w,y+h), one needs to
293       * feed the pipeline using the rectangle (x,y) to
294       * (x+Ceil(w/N),y+Ceil(h/2)), where N is either 2 or 8 depending on
295       * the number of samples.  Experiments indicate that this is not
296       * quite correct; actually, what the hardware appears to do is to
297       * align whatever rectangle is sent down the pipeline to the nearest
298       * multiple of 2x2 blocks, and then scale it up by a factor of N
299       * horizontally and 2 vertically.  So the resulting alignment is 4
300       * vertically and either 4 or 16 horizontally, and the scaledown
301       * factor is 2 vertically and either 2 or 8 horizontally.
302       */
303      switch (aux_surf->format) {
304      case ISL_FORMAT_MCS_2X:
305      case ISL_FORMAT_MCS_4X:
306         x_scaledown = 8;
307         break;
308      case ISL_FORMAT_MCS_8X:
309         x_scaledown = 2;
310         break;
311      case ISL_FORMAT_MCS_16X:
312         x_scaledown = 1;
313         break;
314      default:
315         unreachable("Unexpected MCS format for fast clear");
316      }
317      y_scaledown = 2;
318      x_align = x_scaledown * 2;
319      y_align = y_scaledown * 2;
320   }
321
322   *x0 = ROUND_DOWN_TO(*x0,  x_align) / x_scaledown;
323   *y0 = ROUND_DOWN_TO(*y0, y_align) / y_scaledown;
324   *x1 = ALIGN(*x1, x_align) / x_scaledown;
325   *y1 = ALIGN(*y1, y_align) / y_scaledown;
326}
327
328void
329blorp_fast_clear(struct blorp_batch *batch,
330                 const struct blorp_surf *surf, enum isl_format format,
331                 uint32_t level, uint32_t start_layer, uint32_t num_layers,
332                 uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1)
333{
334   /* Ensure that all layers undergoing the clear have an auxiliary buffer. */
335   assert(start_layer + num_layers <=
336          MAX2(surf->aux_surf->logical_level0_px.depth >> level,
337               surf->aux_surf->logical_level0_px.array_len));
338
339   struct blorp_params params;
340   blorp_params_init(&params);
341   params.num_layers = num_layers;
342
343   params.x0 = x0;
344   params.y0 = y0;
345   params.x1 = x1;
346   params.y1 = y1;
347
348   memset(&params.wm_inputs.clear_color, 0xff, 4*sizeof(float));
349   params.fast_clear_op = ISL_AUX_OP_FAST_CLEAR;
350
351   get_fast_clear_rect(batch->blorp->isl_dev, surf->aux_surf,
352                       &params.x0, &params.y0, &params.x1, &params.y1);
353
354   if (!blorp_params_get_clear_kernel(batch, &params, true, false))
355      return;
356
357   brw_blorp_surface_info_init(batch->blorp, &params.dst, surf, level,
358                               start_layer, format, true);
359   params.num_samples = params.dst.surf.samples;
360
361   batch->blorp->exec(batch, &params);
362}
363
364union isl_color_value
365swizzle_color_value(union isl_color_value src, struct isl_swizzle swizzle)
366{
367   union isl_color_value dst = { .u32 = { 0, } };
368
369   /* We assign colors in ABGR order so that the first one will be taken in
370    * RGBA precedence order.  According to the PRM docs for shader channel
371    * select, this matches Haswell hardware behavior.
372    */
373   if ((unsigned)(swizzle.a - ISL_CHANNEL_SELECT_RED) < 4)
374      dst.u32[swizzle.a - ISL_CHANNEL_SELECT_RED] = src.u32[3];
375   if ((unsigned)(swizzle.b - ISL_CHANNEL_SELECT_RED) < 4)
376      dst.u32[swizzle.b - ISL_CHANNEL_SELECT_RED] = src.u32[2];
377   if ((unsigned)(swizzle.g - ISL_CHANNEL_SELECT_RED) < 4)
378      dst.u32[swizzle.g - ISL_CHANNEL_SELECT_RED] = src.u32[1];
379   if ((unsigned)(swizzle.r - ISL_CHANNEL_SELECT_RED) < 4)
380      dst.u32[swizzle.r - ISL_CHANNEL_SELECT_RED] = src.u32[0];
381
382   return dst;
383}
384
385void
386blorp_clear(struct blorp_batch *batch,
387            const struct blorp_surf *surf,
388            enum isl_format format, struct isl_swizzle swizzle,
389            uint32_t level, uint32_t start_layer, uint32_t num_layers,
390            uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1,
391            union isl_color_value clear_color,
392            const bool color_write_disable[4])
393{
394   struct blorp_params params;
395   blorp_params_init(&params);
396
397   /* Manually apply the clear destination swizzle.  This way swizzled clears
398    * will work for swizzles which we can't normally use for rendering and it
399    * also ensures that they work on pre-Haswell hardware which can't swizlle
400    * at all.
401    */
402   clear_color = swizzle_color_value(clear_color, swizzle);
403   swizzle = ISL_SWIZZLE_IDENTITY;
404
405   bool clear_rgb_as_red = false;
406   if (format == ISL_FORMAT_R9G9B9E5_SHAREDEXP) {
407      clear_color.u32[0] = float3_to_rgb9e5(clear_color.f32);
408      format = ISL_FORMAT_R32_UINT;
409   } else if (format == ISL_FORMAT_L8_UNORM_SRGB) {
410      clear_color.f32[0] = util_format_linear_to_srgb_float(clear_color.f32[0]);
411      format = ISL_FORMAT_R8_UNORM;
412   } else if (format == ISL_FORMAT_A4B4G4R4_UNORM) {
413      /* Broadwell and earlier cannot render to this format so we need to work
414       * around it by swapping the colors around and using B4G4R4A4 instead.
415       */
416      const struct isl_swizzle ARGB = ISL_SWIZZLE(ALPHA, RED, GREEN, BLUE);
417      clear_color = swizzle_color_value(clear_color, ARGB);
418      format = ISL_FORMAT_B4G4R4A4_UNORM;
419   } else if (isl_format_get_layout(format)->bpb % 3 == 0) {
420      clear_rgb_as_red = true;
421      if (format == ISL_FORMAT_R8G8B8_UNORM_SRGB) {
422         clear_color.f32[0] = util_format_linear_to_srgb_float(clear_color.f32[0]);
423         clear_color.f32[1] = util_format_linear_to_srgb_float(clear_color.f32[1]);
424         clear_color.f32[2] = util_format_linear_to_srgb_float(clear_color.f32[2]);
425      }
426   }
427
428   memcpy(&params.wm_inputs.clear_color, clear_color.f32, sizeof(float) * 4);
429
430   bool use_simd16_replicated_data = true;
431
432   /* From the SNB PRM (Vol4_Part1):
433    *
434    *     "Replicated data (Message Type = 111) is only supported when
435    *      accessing tiled memory.  Using this Message Type to access linear
436    *      (untiled) memory is UNDEFINED."
437    */
438   if (surf->surf->tiling == ISL_TILING_LINEAR)
439      use_simd16_replicated_data = false;
440
441   /* Replicated clears don't work yet before gen6 */
442   if (batch->blorp->isl_dev->info->gen < 6)
443      use_simd16_replicated_data = false;
444
445   /* Constant color writes ignore everyting in blend and color calculator
446    * state.  This is not documented.
447    */
448   if (color_write_disable) {
449      for (unsigned i = 0; i < 4; i++) {
450         params.color_write_disable[i] = color_write_disable[i];
451         if (color_write_disable[i])
452            use_simd16_replicated_data = false;
453      }
454   }
455
456   if (!blorp_params_get_clear_kernel(batch, &params,
457                                      use_simd16_replicated_data,
458                                      clear_rgb_as_red))
459      return;
460
461   if (!blorp_ensure_sf_program(batch, &params))
462      return;
463
464   while (num_layers > 0) {
465      brw_blorp_surface_info_init(batch->blorp, &params.dst, surf, level,
466                                  start_layer, format, true);
467      params.dst.view.swizzle = swizzle;
468
469      params.x0 = x0;
470      params.y0 = y0;
471      params.x1 = x1;
472      params.y1 = y1;
473
474      if (params.dst.tile_x_sa || params.dst.tile_y_sa) {
475         assert(params.dst.surf.samples == 1);
476         assert(num_layers == 1);
477         params.x0 += params.dst.tile_x_sa;
478         params.y0 += params.dst.tile_y_sa;
479         params.x1 += params.dst.tile_x_sa;
480         params.y1 += params.dst.tile_y_sa;
481      }
482
483      /* The MinLOD and MinimumArrayElement don't work properly for cube maps.
484       * Convert them to a single slice on gen4.
485       */
486      if (batch->blorp->isl_dev->info->gen == 4 &&
487          (params.dst.surf.usage & ISL_SURF_USAGE_CUBE_BIT)) {
488         blorp_surf_convert_to_single_slice(batch->blorp->isl_dev, &params.dst);
489      }
490
491      if (clear_rgb_as_red) {
492         surf_fake_rgb_with_red(batch->blorp->isl_dev, &params.dst);
493         params.x0 *= 3;
494         params.x1 *= 3;
495      }
496
497      if (isl_format_is_compressed(params.dst.surf.format)) {
498         blorp_surf_convert_to_uncompressed(batch->blorp->isl_dev, &params.dst,
499                                            NULL, NULL, NULL, NULL);
500                                            //&dst_x, &dst_y, &dst_w, &dst_h);
501      }
502
503      if (params.dst.tile_x_sa || params.dst.tile_y_sa) {
504         /* Either we're on gen4 where there is no multisampling or the
505          * surface is compressed which also implies no multisampling.
506          * Therefore, sa == px and we don't need to do a conversion.
507          */
508         assert(params.dst.surf.samples == 1);
509         params.x0 += params.dst.tile_x_sa;
510         params.y0 += params.dst.tile_y_sa;
511         params.x1 += params.dst.tile_x_sa;
512         params.y1 += params.dst.tile_y_sa;
513      }
514
515      params.num_samples = params.dst.surf.samples;
516
517      /* We may be restricted on the number of layers we can bind at any one
518       * time.  In particular, Sandy Bridge has a maximum number of layers of
519       * 512 but a maximum 3D texture size is much larger.
520       */
521      params.num_layers = MIN2(params.dst.view.array_len, num_layers);
522
523      const unsigned max_image_width = 16 * 1024;
524      if (params.dst.surf.logical_level0_px.width > max_image_width) {
525         /* Clearing an RGB image as red multiplies the surface width by 3
526          * so it may now be too wide for the hardware surface limits.  We
527          * have to break the clear up into pieces in order to clear wide
528          * images.
529          */
530         assert(clear_rgb_as_red);
531         assert(params.dst.surf.dim == ISL_SURF_DIM_2D);
532         assert(params.dst.surf.tiling == ISL_TILING_LINEAR);
533         assert(params.dst.surf.logical_level0_px.depth == 1);
534         assert(params.dst.surf.logical_level0_px.array_len == 1);
535         assert(params.dst.surf.levels == 1);
536         assert(params.dst.surf.samples == 1);
537         assert(params.dst.tile_x_sa == 0 || params.dst.tile_y_sa == 0);
538         assert(params.dst.aux_usage == ISL_AUX_USAGE_NONE);
539
540         /* max_image_width rounded down to a multiple of 3 */
541         const unsigned max_fake_rgb_width = (max_image_width / 3) * 3;
542         const unsigned cpp =
543            isl_format_get_layout(params.dst.surf.format)->bpb / 8;
544
545         params.dst.surf.logical_level0_px.width = max_fake_rgb_width;
546         params.dst.surf.phys_level0_sa.width = max_fake_rgb_width;
547
548         uint32_t orig_x0 = params.x0, orig_x1 = params.x1;
549         uint64_t orig_offset = params.dst.addr.offset;
550         for (uint32_t x = orig_x0; x < orig_x1; x += max_fake_rgb_width) {
551            /* Offset to the surface.  It's easy because we're linear */
552            params.dst.addr.offset = orig_offset + x * cpp;
553
554            params.x0 = 0;
555            params.x1 = MIN2(orig_x1 - x, max_image_width);
556
557            batch->blorp->exec(batch, &params);
558         }
559      } else {
560         batch->blorp->exec(batch, &params);
561      }
562
563      start_layer += params.num_layers;
564      num_layers -= params.num_layers;
565   }
566}
567
568void
569blorp_clear_depth_stencil(struct blorp_batch *batch,
570                          const struct blorp_surf *depth,
571                          const struct blorp_surf *stencil,
572                          uint32_t level, uint32_t start_layer,
573                          uint32_t num_layers,
574                          uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1,
575                          bool clear_depth, float depth_value,
576                          uint8_t stencil_mask, uint8_t stencil_value)
577{
578   struct blorp_params params;
579   blorp_params_init(&params);
580
581   params.x0 = x0;
582   params.y0 = y0;
583   params.x1 = x1;
584   params.y1 = y1;
585
586   if (ISL_DEV_GEN(batch->blorp->isl_dev) == 6) {
587      /* For some reason, Sandy Bridge gets occlusion queries wrong if we
588       * don't have a shader.  In particular, it records samples even though
589       * we disable statistics in 3DSTATE_WM.  Give it the usual clear shader
590       * to work around the issue.
591       */
592      if (!blorp_params_get_clear_kernel(batch, &params, false, false))
593         return;
594   }
595
596   while (num_layers > 0) {
597      params.num_layers = num_layers;
598
599      if (stencil_mask) {
600         brw_blorp_surface_info_init(batch->blorp, &params.stencil, stencil,
601                                     level, start_layer,
602                                     ISL_FORMAT_UNSUPPORTED, true);
603         params.stencil_mask = stencil_mask;
604         params.stencil_ref = stencil_value;
605
606         params.dst.surf.samples = params.stencil.surf.samples;
607         params.dst.surf.logical_level0_px =
608            params.stencil.surf.logical_level0_px;
609         params.dst.view = params.depth.view;
610
611         params.num_samples = params.stencil.surf.samples;
612
613         /* We may be restricted on the number of layers we can bind at any
614          * one time.  In particular, Sandy Bridge has a maximum number of
615          * layers of 512 but a maximum 3D texture size is much larger.
616          */
617         if (params.stencil.view.array_len < params.num_layers)
618            params.num_layers = params.stencil.view.array_len;
619      }
620
621      if (clear_depth) {
622         brw_blorp_surface_info_init(batch->blorp, &params.depth, depth,
623                                     level, start_layer,
624                                     ISL_FORMAT_UNSUPPORTED, true);
625         params.z = depth_value;
626         params.depth_format =
627            isl_format_get_depth_format(depth->surf->format, false);
628
629         params.dst.surf.samples = params.depth.surf.samples;
630         params.dst.surf.logical_level0_px =
631            params.depth.surf.logical_level0_px;
632         params.dst.view = params.depth.view;
633
634         params.num_samples = params.depth.surf.samples;
635
636         /* We may be restricted on the number of layers we can bind at any
637          * one time.  In particular, Sandy Bridge has a maximum number of
638          * layers of 512 but a maximum 3D texture size is much larger.
639          */
640         if (params.depth.view.array_len < params.num_layers)
641            params.num_layers = params.depth.view.array_len;
642      }
643
644      batch->blorp->exec(batch, &params);
645
646      start_layer += params.num_layers;
647      num_layers -= params.num_layers;
648   }
649}
650
651bool
652blorp_can_hiz_clear_depth(uint8_t gen, enum isl_format format,
653                          uint32_t num_samples,
654                          uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1)
655{
656   /* This function currently doesn't support any gen prior to gen8 */
657   assert(gen >= 8);
658
659   if (gen == 8 && format == ISL_FORMAT_R16_UNORM) {
660      /* Apply the D16 alignment restrictions. On BDW, HiZ has an 8x4 sample
661       * block with the following property: as the number of samples increases,
662       * the number of pixels representable by this block decreases by a factor
663       * of the sample dimensions. Sample dimensions scale following the MSAA
664       * interleaved pattern.
665       *
666       * Sample|Sample|Pixel
667       * Count |Dim   |Dim
668       * ===================
669       *    1  | 1x1  | 8x4
670       *    2  | 2x1  | 4x4
671       *    4  | 2x2  | 4x2
672       *    8  | 4x2  | 2x2
673       *   16  | 4x4  | 2x1
674       *
675       * Table: Pixel Dimensions in a HiZ Sample Block Pre-SKL
676       */
677      const struct isl_extent2d sa_block_dim =
678         isl_get_interleaved_msaa_px_size_sa(num_samples);
679      const uint8_t align_px_w = 8 / sa_block_dim.w;
680      const uint8_t align_px_h = 4 / sa_block_dim.h;
681
682      /* Fast depth clears clear an entire sample block at a time. As a result,
683       * the rectangle must be aligned to the dimensions of the encompassing
684       * pixel block for a successful operation.
685       *
686       * Fast clears can still work if the upper-left corner is aligned and the
687       * bottom-rigtht corner touches the edge of a depth buffer whose extent
688       * is unaligned. This is because each miplevel in the depth buffer is
689       * padded by the Pixel Dim (similar to a standard compressed texture).
690       * In this case, the clear rectangle could be padded by to match the full
691       * depth buffer extent but to support multiple clearing techniques, we
692       * chose to be unaware of the depth buffer's extent and thus don't handle
693       * this case.
694       */
695      if (x0 % align_px_w || y0 % align_px_h ||
696          x1 % align_px_w || y1 % align_px_h)
697         return false;
698   }
699   return true;
700}
701
702void
703blorp_hiz_clear_depth_stencil(struct blorp_batch *batch,
704                              const struct blorp_surf *depth,
705                              const struct blorp_surf *stencil,
706                              uint32_t level,
707                              uint32_t start_layer, uint32_t num_layers,
708                              uint32_t x0, uint32_t y0,
709                              uint32_t x1, uint32_t y1,
710                              bool clear_depth, float depth_value,
711                              bool clear_stencil, uint8_t stencil_value)
712{
713   struct blorp_params params;
714   blorp_params_init(&params);
715
716   /* This requires WM_HZ_OP which only exists on gen8+ */
717   assert(ISL_DEV_GEN(batch->blorp->isl_dev) >= 8);
718
719   params.hiz_op = ISL_AUX_OP_FAST_CLEAR;
720   params.num_layers = 1;
721
722   params.x0 = x0;
723   params.y0 = y0;
724   params.x1 = x1;
725   params.y1 = y1;
726
727   for (uint32_t l = 0; l < num_layers; l++) {
728      const uint32_t layer = start_layer + l;
729      if (clear_stencil) {
730         brw_blorp_surface_info_init(batch->blorp, &params.stencil, stencil,
731                                     level, layer,
732                                     ISL_FORMAT_UNSUPPORTED, true);
733         params.stencil_mask = 0xff;
734         params.stencil_ref = stencil_value;
735         params.num_samples = params.stencil.surf.samples;
736      }
737
738      if (clear_depth) {
739         /* If we're clearing depth, we must have HiZ */
740         assert(depth && depth->aux_usage == ISL_AUX_USAGE_HIZ);
741
742         brw_blorp_surface_info_init(batch->blorp, &params.depth, depth,
743                                     level, layer,
744                                     ISL_FORMAT_UNSUPPORTED, true);
745         params.depth.clear_color.f32[0] = depth_value;
746         params.depth_format =
747            isl_format_get_depth_format(depth->surf->format, false);
748         params.num_samples = params.depth.surf.samples;
749      }
750
751      batch->blorp->exec(batch, &params);
752   }
753}
754
755/* Given a depth stencil attachment, this function performs a fast depth clear
756 * on a depth portion and a regular clear on the stencil portion. When
757 * performing a fast depth clear on the depth portion, the HiZ buffer is simply
758 * tagged as cleared so the depth clear value is not actually needed.
759 */
760void
761blorp_gen8_hiz_clear_attachments(struct blorp_batch *batch,
762                                 uint32_t num_samples,
763                                 uint32_t x0, uint32_t y0,
764                                 uint32_t x1, uint32_t y1,
765                                 bool clear_depth, bool clear_stencil,
766                                 uint8_t stencil_value)
767{
768   assert(batch->flags & BLORP_BATCH_NO_EMIT_DEPTH_STENCIL);
769
770   struct blorp_params params;
771   blorp_params_init(&params);
772   params.num_layers = 1;
773   params.hiz_op = ISL_AUX_OP_FAST_CLEAR;
774   params.x0 = x0;
775   params.y0 = y0;
776   params.x1 = x1;
777   params.y1 = y1;
778   params.num_samples = num_samples;
779   params.depth.enabled = clear_depth;
780   params.stencil.enabled = clear_stencil;
781   params.stencil_ref = stencil_value;
782   batch->blorp->exec(batch, &params);
783}
784
785/** Clear active color/depth/stencili attachments
786 *
787 * This function performs a clear operation on the currently bound
788 * color/depth/stencil attachments.  It is assumed that any information passed
789 * in here is valid, consistent, and in-bounds relative to the currently
790 * attached depth/stencil.  The binding_table_offset parameter is the 32-bit
791 * offset relative to surface state base address where pre-baked binding table
792 * that we are to use lives.  If clear_color is false, binding_table_offset
793 * must point to a binding table with one entry which is a valid null surface
794 * that matches the currently bound depth and stencil.
795 */
796void
797blorp_clear_attachments(struct blorp_batch *batch,
798                        uint32_t binding_table_offset,
799                        enum isl_format depth_format,
800                        uint32_t num_samples,
801                        uint32_t start_layer, uint32_t num_layers,
802                        uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1,
803                        bool clear_color, union isl_color_value color_value,
804                        bool clear_depth, float depth_value,
805                        uint8_t stencil_mask, uint8_t stencil_value)
806{
807   struct blorp_params params;
808   blorp_params_init(&params);
809
810   assert(batch->flags & BLORP_BATCH_NO_EMIT_DEPTH_STENCIL);
811
812   params.x0 = x0;
813   params.y0 = y0;
814   params.x1 = x1;
815   params.y1 = y1;
816
817   params.use_pre_baked_binding_table = true;
818   params.pre_baked_binding_table_offset = binding_table_offset;
819
820   params.num_layers = num_layers;
821   params.num_samples = num_samples;
822
823   if (clear_color) {
824      params.dst.enabled = true;
825
826      memcpy(&params.wm_inputs.clear_color, color_value.f32, sizeof(float) * 4);
827
828      /* Unfortunately, without knowing whether or not our destination surface
829       * is tiled or not, we have to assume it may be linear.  This means no
830       * SIMD16_REPDATA for us. :-(
831       */
832      if (!blorp_params_get_clear_kernel(batch, &params, false, false))
833         return;
834   }
835
836   if (clear_depth) {
837      params.depth.enabled = true;
838
839      params.z = depth_value;
840      params.depth_format = isl_format_get_depth_format(depth_format, false);
841   }
842
843   if (stencil_mask) {
844      params.stencil.enabled = true;
845
846      params.stencil_mask = stencil_mask;
847      params.stencil_ref = stencil_value;
848   }
849
850   if (!blorp_params_get_layer_offset_vs(batch, &params))
851      return;
852
853   params.vs_inputs.base_layer = start_layer;
854
855   batch->blorp->exec(batch, &params);
856}
857
858void
859blorp_ccs_resolve(struct blorp_batch *batch,
860                  struct blorp_surf *surf, uint32_t level,
861                  uint32_t start_layer, uint32_t num_layers,
862                  enum isl_format format,
863                  enum isl_aux_op resolve_op)
864{
865   struct blorp_params params;
866
867   blorp_params_init(&params);
868   brw_blorp_surface_info_init(batch->blorp, &params.dst, surf,
869                               level, start_layer, format, true);
870
871   /* From the Ivy Bridge PRM, Vol2 Part1 11.9 "Render Target Resolve":
872    *
873    *     A rectangle primitive must be scaled down by the following factors
874    *     with respect to render target being resolved.
875    *
876    * The scaledown factors in the table that follows are related to the block
877    * size of the CCS format.  For IVB and HSW, we divide by two, for BDW we
878    * multiply by 8 and 16. On Sky Lake, we multiply by 8.
879    */
880   const struct isl_format_layout *aux_fmtl =
881      isl_format_get_layout(params.dst.aux_surf.format);
882   assert(aux_fmtl->txc == ISL_TXC_CCS);
883
884   unsigned x_scaledown, y_scaledown;
885   if (ISL_DEV_GEN(batch->blorp->isl_dev) >= 9) {
886      x_scaledown = aux_fmtl->bw * 8;
887      y_scaledown = aux_fmtl->bh * 8;
888   } else if (ISL_DEV_GEN(batch->blorp->isl_dev) >= 8) {
889      x_scaledown = aux_fmtl->bw * 8;
890      y_scaledown = aux_fmtl->bh * 16;
891   } else {
892      x_scaledown = aux_fmtl->bw / 2;
893      y_scaledown = aux_fmtl->bh / 2;
894   }
895   params.x0 = params.y0 = 0;
896   params.x1 = minify(params.dst.aux_surf.logical_level0_px.width, level);
897   params.y1 = minify(params.dst.aux_surf.logical_level0_px.height, level);
898   params.x1 = ALIGN(params.x1, x_scaledown) / x_scaledown;
899   params.y1 = ALIGN(params.y1, y_scaledown) / y_scaledown;
900
901   if (batch->blorp->isl_dev->info->gen >= 9) {
902      assert(resolve_op == ISL_AUX_OP_FULL_RESOLVE ||
903             resolve_op == ISL_AUX_OP_PARTIAL_RESOLVE);
904   } else {
905      /* Broadwell and earlier do not have a partial resolve */
906      assert(resolve_op == ISL_AUX_OP_FULL_RESOLVE);
907   }
908   params.fast_clear_op = resolve_op;
909   params.num_layers = num_layers;
910
911   /* Note: there is no need to initialize push constants because it doesn't
912    * matter what data gets dispatched to the render target.  However, we must
913    * ensure that the fragment shader delivers the data using the "replicated
914    * color" message.
915    */
916
917   if (!blorp_params_get_clear_kernel(batch, &params, true, false))
918      return;
919
920   batch->blorp->exec(batch, &params);
921}
922
923static nir_ssa_def *
924blorp_nir_bit(nir_builder *b, nir_ssa_def *src, unsigned bit)
925{
926   return nir_iand(b, nir_ushr(b, src, nir_imm_int(b, bit)),
927                      nir_imm_int(b, 1));
928}
929
930struct blorp_mcs_partial_resolve_key
931{
932   enum blorp_shader_type shader_type;
933   bool indirect_clear_color;
934   bool int_format;
935   uint32_t num_samples;
936};
937
938static bool
939blorp_params_get_mcs_partial_resolve_kernel(struct blorp_batch *batch,
940                                            struct blorp_params *params)
941{
942   struct blorp_context *blorp = batch->blorp;
943   const struct blorp_mcs_partial_resolve_key blorp_key = {
944      .shader_type = BLORP_SHADER_TYPE_MCS_PARTIAL_RESOLVE,
945      .indirect_clear_color = params->dst.clear_color_addr.buffer != NULL,
946      .int_format = isl_format_has_int_channel(params->dst.view.format),
947      .num_samples = params->num_samples,
948   };
949
950   if (blorp->lookup_shader(batch, &blorp_key, sizeof(blorp_key),
951                            &params->wm_prog_kernel, &params->wm_prog_data))
952      return true;
953
954   void *mem_ctx = ralloc_context(NULL);
955
956   nir_builder b;
957   blorp_nir_init_shader(&b, mem_ctx, MESA_SHADER_FRAGMENT,
958                         "BLORP-mcs-partial-resolve");
959
960   nir_variable *v_color =
961      BLORP_CREATE_NIR_INPUT(b.shader, clear_color, glsl_vec4_type());
962
963   nir_variable *frag_color =
964      nir_variable_create(b.shader, nir_var_shader_out,
965                          glsl_vec4_type(), "gl_FragColor");
966   frag_color->data.location = FRAG_RESULT_COLOR;
967
968   /* Do an MCS fetch and check if it is equal to the magic clear value */
969   nir_ssa_def *mcs =
970      blorp_nir_txf_ms_mcs(&b, nir_f2i32(&b, blorp_nir_frag_coord(&b)),
971                               nir_load_layer_id(&b));
972   nir_ssa_def *is_clear =
973      blorp_nir_mcs_is_clear_color(&b, mcs, blorp_key.num_samples);
974
975   /* If we aren't the clear value, discard. */
976   nir_intrinsic_instr *discard =
977      nir_intrinsic_instr_create(b.shader, nir_intrinsic_discard_if);
978   discard->src[0] = nir_src_for_ssa(nir_inot(&b, is_clear));
979   nir_builder_instr_insert(&b, &discard->instr);
980
981   nir_ssa_def *clear_color = nir_load_var(&b, v_color);
982   if (blorp_key.indirect_clear_color && blorp->isl_dev->info->gen <= 8) {
983      /* Gen7-8 clear colors are stored as single 0/1 bits */
984      clear_color = nir_vec4(&b, blorp_nir_bit(&b, clear_color, 31),
985                                 blorp_nir_bit(&b, clear_color, 30),
986                                 blorp_nir_bit(&b, clear_color, 29),
987                                 blorp_nir_bit(&b, clear_color, 28));
988
989      if (!blorp_key.int_format)
990         clear_color = nir_i2f32(&b, clear_color);
991   }
992   nir_store_var(&b, frag_color, clear_color, 0xf);
993
994   struct brw_wm_prog_key wm_key;
995   brw_blorp_init_wm_prog_key(&wm_key);
996   wm_key.tex.compressed_multisample_layout_mask = 1;
997   wm_key.tex.msaa_16 = blorp_key.num_samples == 16;
998   wm_key.multisample_fbo = true;
999
1000   struct brw_wm_prog_data prog_data;
1001   const unsigned *program =
1002      blorp_compile_fs(blorp, mem_ctx, b.shader, &wm_key, false,
1003                       &prog_data);
1004
1005   bool result =
1006      blorp->upload_shader(batch, &blorp_key, sizeof(blorp_key),
1007                           program, prog_data.base.program_size,
1008                           &prog_data.base, sizeof(prog_data),
1009                           &params->wm_prog_kernel, &params->wm_prog_data);
1010
1011   ralloc_free(mem_ctx);
1012   return result;
1013}
1014
1015void
1016blorp_mcs_partial_resolve(struct blorp_batch *batch,
1017                          struct blorp_surf *surf,
1018                          enum isl_format format,
1019                          uint32_t start_layer, uint32_t num_layers)
1020{
1021   struct blorp_params params;
1022   blorp_params_init(&params);
1023
1024   assert(batch->blorp->isl_dev->info->gen >= 7);
1025
1026   params.x0 = 0;
1027   params.y0 = 0;
1028   params.x1 = surf->surf->logical_level0_px.width;
1029   params.y1 = surf->surf->logical_level0_px.height;
1030
1031   brw_blorp_surface_info_init(batch->blorp, &params.src, surf, 0,
1032                               start_layer, format, false);
1033   brw_blorp_surface_info_init(batch->blorp, &params.dst, surf, 0,
1034                               start_layer, format, true);
1035
1036   params.num_samples = params.dst.surf.samples;
1037   params.num_layers = num_layers;
1038   params.dst_clear_color_as_input = surf->clear_color_addr.buffer != NULL;
1039
1040   memcpy(&params.wm_inputs.clear_color,
1041          surf->clear_color.f32, sizeof(float) * 4);
1042
1043   if (!blorp_params_get_mcs_partial_resolve_kernel(batch, &params))
1044      return;
1045
1046   batch->blorp->exec(batch, &params);
1047}
1048
1049/** Clear a CCS to the "uncompressed" state
1050 *
1051 * This pass is the CCS equivalent of a "HiZ resolve".  It sets the CCS values
1052 * for a given layer/level of a surface to 0x0 which is the "uncompressed"
1053 * state which tells the sampler to go look at the main surface.
1054 */
1055void
1056blorp_ccs_ambiguate(struct blorp_batch *batch,
1057                    struct blorp_surf *surf,
1058                    uint32_t level, uint32_t layer)
1059{
1060   struct blorp_params params;
1061   blorp_params_init(&params);
1062
1063   assert(ISL_DEV_GEN(batch->blorp->isl_dev) >= 7);
1064
1065   const struct isl_format_layout *aux_fmtl =
1066      isl_format_get_layout(surf->aux_surf->format);
1067   assert(aux_fmtl->txc == ISL_TXC_CCS);
1068
1069   params.dst = (struct brw_blorp_surface_info) {
1070      .enabled = true,
1071      .addr = surf->aux_addr,
1072      .view = {
1073         .usage = ISL_SURF_USAGE_RENDER_TARGET_BIT,
1074         .format = ISL_FORMAT_R32G32B32A32_UINT,
1075         .base_level = 0,
1076         .base_array_layer = 0,
1077         .levels = 1,
1078         .array_len = 1,
1079         .swizzle = ISL_SWIZZLE_IDENTITY,
1080      },
1081   };
1082
1083   uint32_t z = 0;
1084   if (surf->surf->dim == ISL_SURF_DIM_3D) {
1085      z = layer;
1086      layer = 0;
1087   }
1088
1089   uint32_t offset_B, x_offset_el, y_offset_el;
1090   isl_surf_get_image_offset_el(surf->aux_surf, level, layer, z,
1091                                &x_offset_el, &y_offset_el);
1092   isl_tiling_get_intratile_offset_el(surf->aux_surf->tiling, aux_fmtl->bpb,
1093                                      surf->aux_surf->row_pitch_B,
1094                                      x_offset_el, y_offset_el,
1095                                      &offset_B, &x_offset_el, &y_offset_el);
1096   params.dst.addr.offset += offset_B;
1097
1098   const uint32_t width_px =
1099      minify(surf->aux_surf->logical_level0_px.width, level);
1100   const uint32_t height_px =
1101      minify(surf->aux_surf->logical_level0_px.height, level);
1102   const uint32_t width_el = DIV_ROUND_UP(width_px, aux_fmtl->bw);
1103   const uint32_t height_el = DIV_ROUND_UP(height_px, aux_fmtl->bh);
1104
1105   struct isl_tile_info ccs_tile_info;
1106   isl_surf_get_tile_info(surf->aux_surf, &ccs_tile_info);
1107
1108   /* We're going to map it as a regular RGBA32_UINT surface.  We need to
1109    * downscale a good deal.  We start by computing the area on the CCS to
1110    * clear in units of Y-tiled cache lines.
1111    */
1112   uint32_t x_offset_cl, y_offset_cl, width_cl, height_cl;
1113   if (ISL_DEV_GEN(batch->blorp->isl_dev) >= 8) {
1114      /* From the Sky Lake PRM Vol. 12 in the section on planes:
1115       *
1116       *    "The Color Control Surface (CCS) contains the compression status
1117       *    of the cache-line pairs. The compression state of the cache-line
1118       *    pair is specified by 2 bits in the CCS.  Each CCS cache-line
1119       *    represents an area on the main surface of 16x16 sets of 128 byte
1120       *    Y-tiled cache-line-pairs. CCS is always Y tiled."
1121       *
1122       * Each 2-bit surface element in the CCS corresponds to a single
1123       * cache-line pair in the main surface.  This means that 16x16 el block
1124       * in the CCS maps to a Y-tiled cache line.  Fortunately, CCS layouts
1125       * are calculated with a very large alignment so we can round up to a
1126       * whole cache line without worrying about overdraw.
1127       */
1128
1129      /* On Broadwell and above, a CCS tile is the same as a Y tile when
1130       * viewed at the cache-line granularity.  Fortunately, the horizontal
1131       * and vertical alignment requirements of the CCS are such that we can
1132       * align to an entire cache line without worrying about crossing over
1133       * from one LOD to another.
1134       */
1135      const uint32_t x_el_per_cl = ccs_tile_info.logical_extent_el.w / 8;
1136      const uint32_t y_el_per_cl = ccs_tile_info.logical_extent_el.h / 8;
1137      assert(surf->aux_surf->image_alignment_el.w % x_el_per_cl == 0);
1138      assert(surf->aux_surf->image_alignment_el.h % y_el_per_cl == 0);
1139
1140      assert(x_offset_el % x_el_per_cl == 0);
1141      assert(y_offset_el % y_el_per_cl == 0);
1142      x_offset_cl = x_offset_el / x_el_per_cl;
1143      y_offset_cl = y_offset_el / y_el_per_cl;
1144      width_cl = DIV_ROUND_UP(width_el, x_el_per_cl);
1145      height_cl = DIV_ROUND_UP(height_el, y_el_per_cl);
1146   } else {
1147      /* On gen7, the CCS tiling is not so nice.  However, there we are
1148       * guaranteed that we only have a single level and slice so we don't
1149       * have to worry about it and can just align to a whole tile.
1150       */
1151      assert(surf->aux_surf->logical_level0_px.depth == 1);
1152      assert(surf->aux_surf->logical_level0_px.array_len == 1);
1153      assert(x_offset_el == 0 && y_offset_el == 0);
1154      const uint32_t width_tl =
1155         DIV_ROUND_UP(width_el, ccs_tile_info.logical_extent_el.w);
1156      const uint32_t height_tl =
1157         DIV_ROUND_UP(height_el, ccs_tile_info.logical_extent_el.h);
1158      x_offset_cl = 0;
1159      y_offset_cl = 0;
1160      width_cl = width_tl * 8;
1161      height_cl = height_tl * 8;
1162   }
1163
1164   /* We're going to use a RGBA32 format so as to write data as quickly as
1165    * possible.  A y-tiled cache line will then be 1x4 px.
1166    */
1167   const uint32_t x_offset_rgba_px = x_offset_cl;
1168   const uint32_t y_offset_rgba_px = y_offset_cl * 4;
1169   const uint32_t width_rgba_px = width_cl;
1170   const uint32_t height_rgba_px = height_cl * 4;
1171
1172   MAYBE_UNUSED bool ok =
1173      isl_surf_init(batch->blorp->isl_dev, &params.dst.surf,
1174                    .dim = ISL_SURF_DIM_2D,
1175                    .format = ISL_FORMAT_R32G32B32A32_UINT,
1176                    .width = width_rgba_px + x_offset_rgba_px,
1177                    .height = height_rgba_px + y_offset_rgba_px,
1178                    .depth = 1,
1179                    .levels = 1,
1180                    .array_len = 1,
1181                    .samples = 1,
1182                    .row_pitch_B = surf->aux_surf->row_pitch_B,
1183                    .usage = ISL_SURF_USAGE_RENDER_TARGET_BIT,
1184                    .tiling_flags = ISL_TILING_Y0_BIT);
1185   assert(ok);
1186
1187   params.x0 = x_offset_rgba_px;
1188   params.y0 = y_offset_rgba_px;
1189   params.x1 = x_offset_rgba_px + width_rgba_px;
1190   params.y1 = y_offset_rgba_px + height_rgba_px;
1191
1192   /* A CCS value of 0 means "uncompressed." */
1193   memset(&params.wm_inputs.clear_color, 0,
1194          sizeof(params.wm_inputs.clear_color));
1195
1196   if (!blorp_params_get_clear_kernel(batch, &params, true, false))
1197      return;
1198
1199   batch->blorp->exec(batch, &params);
1200}
1201