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#include "dev/intel_debug.h"
33
34#include "blorp_nir_builder.h"
35
36#define FILE_DEBUG_FLAG DEBUG_BLORP
37
38#pragma pack(push, 1)
39struct brw_blorp_const_color_prog_key
40{
41   struct brw_blorp_base_key base;
42   bool use_simd16_replicated_data;
43   bool clear_rgb_as_red;
44   uint8_t local_y;
45};
46#pragma pack(pop)
47
48static bool
49blorp_params_get_clear_kernel_fs(struct blorp_batch *batch,
50                                 struct blorp_params *params,
51                                 bool use_replicated_data,
52                                 bool clear_rgb_as_red)
53{
54   struct blorp_context *blorp = batch->blorp;
55
56   const struct brw_blorp_const_color_prog_key blorp_key = {
57      .base = BRW_BLORP_BASE_KEY_INIT(BLORP_SHADER_TYPE_CLEAR),
58      .base.shader_pipeline = BLORP_SHADER_PIPELINE_RENDER,
59      .use_simd16_replicated_data = use_replicated_data,
60      .clear_rgb_as_red = clear_rgb_as_red,
61      .local_y = 0,
62   };
63
64   if (blorp->lookup_shader(batch, &blorp_key, sizeof(blorp_key),
65                            &params->wm_prog_kernel, &params->wm_prog_data))
66      return true;
67
68   void *mem_ctx = ralloc_context(NULL);
69
70   nir_builder b;
71   blorp_nir_init_shader(&b, mem_ctx, MESA_SHADER_FRAGMENT,
72                         blorp_shader_type_to_name(blorp_key.base.shader_type));
73
74   nir_variable *v_color =
75      BLORP_CREATE_NIR_INPUT(b.shader, clear_color, glsl_vec4_type());
76   nir_ssa_def *color = nir_load_var(&b, v_color);
77
78   if (clear_rgb_as_red) {
79      nir_ssa_def *pos = nir_f2i32(&b, nir_load_frag_coord(&b));
80      nir_ssa_def *comp = nir_umod(&b, nir_channel(&b, pos, 0),
81                                       nir_imm_int(&b, 3));
82      color = nir_pad_vec4(&b, nir_vector_extract(&b, color, comp));
83   }
84
85   nir_variable *frag_color = nir_variable_create(b.shader, nir_var_shader_out,
86                                                  glsl_vec4_type(),
87                                                  "gl_FragColor");
88   frag_color->data.location = FRAG_RESULT_COLOR;
89   nir_store_var(&b, frag_color, color, 0xf);
90
91   struct brw_wm_prog_key wm_key;
92   brw_blorp_init_wm_prog_key(&wm_key);
93
94   struct brw_wm_prog_data prog_data;
95   const unsigned *program =
96      blorp_compile_fs(blorp, mem_ctx, b.shader, &wm_key, use_replicated_data,
97                       &prog_data);
98
99   bool result =
100      blorp->upload_shader(batch, MESA_SHADER_FRAGMENT,
101                           &blorp_key, sizeof(blorp_key),
102                           program, prog_data.base.program_size,
103                           &prog_data.base, sizeof(prog_data),
104                           &params->wm_prog_kernel, &params->wm_prog_data);
105
106   ralloc_free(mem_ctx);
107   return result;
108}
109
110static bool
111blorp_params_get_clear_kernel_cs(struct blorp_batch *batch,
112                                 struct blorp_params *params,
113                                 bool clear_rgb_as_red)
114{
115   struct blorp_context *blorp = batch->blorp;
116
117   const struct brw_blorp_const_color_prog_key blorp_key = {
118      .base = BRW_BLORP_BASE_KEY_INIT(BLORP_SHADER_TYPE_CLEAR),
119      .base.shader_pipeline = BLORP_SHADER_PIPELINE_COMPUTE,
120      .use_simd16_replicated_data = false,
121      .clear_rgb_as_red = clear_rgb_as_red,
122      .local_y = blorp_get_cs_local_y(params),
123   };
124
125   if (blorp->lookup_shader(batch, &blorp_key, sizeof(blorp_key),
126                            &params->cs_prog_kernel, &params->cs_prog_data))
127      return true;
128
129   void *mem_ctx = ralloc_context(NULL);
130
131   nir_builder b;
132   blorp_nir_init_shader(&b, mem_ctx, MESA_SHADER_COMPUTE, "BLORP-gpgpu-clear");
133   blorp_set_cs_dims(b.shader, blorp_key.local_y);
134
135   nir_ssa_def *dst_pos = nir_load_global_invocation_id(&b, 32);
136
137   nir_variable *v_color =
138      BLORP_CREATE_NIR_INPUT(b.shader, clear_color, glsl_vec4_type());
139   nir_ssa_def *color = nir_load_var(&b, v_color);
140
141   nir_variable *v_bounds_rect =
142      BLORP_CREATE_NIR_INPUT(b.shader, bounds_rect, glsl_vec4_type());
143   nir_ssa_def *bounds_rect = nir_load_var(&b, v_bounds_rect);
144   nir_ssa_def *in_bounds = blorp_check_in_bounds(&b, bounds_rect, dst_pos);
145
146   if (clear_rgb_as_red) {
147      nir_ssa_def *comp = nir_umod(&b, nir_channel(&b, dst_pos, 0),
148                                       nir_imm_int(&b, 3));
149      color = nir_pad_vec4(&b, nir_vector_extract(&b, color, comp));
150   }
151
152   nir_push_if(&b, in_bounds);
153
154   nir_image_store(&b, nir_imm_int(&b, 0),
155                   nir_pad_vector_imm_int(&b, dst_pos, 0, 4),
156                   nir_imm_int(&b, 0),
157                   nir_pad_vector_imm_int(&b, color, 0, 4),
158                   nir_imm_int(&b, 0),
159                   .image_dim = GLSL_SAMPLER_DIM_2D,
160                   .image_array = true,
161                   .access = ACCESS_NON_READABLE);
162
163   nir_pop_if(&b, NULL);
164
165   struct brw_cs_prog_key cs_key;
166   brw_blorp_init_cs_prog_key(&cs_key);
167
168   struct brw_cs_prog_data prog_data;
169   const unsigned *program =
170      blorp_compile_cs(blorp, mem_ctx, b.shader, &cs_key, &prog_data);
171
172   bool result =
173      blorp->upload_shader(batch, MESA_SHADER_COMPUTE,
174                           &blorp_key, sizeof(blorp_key),
175                           program, prog_data.base.program_size,
176                           &prog_data.base, sizeof(prog_data),
177                           &params->cs_prog_kernel, &params->cs_prog_data);
178
179   ralloc_free(mem_ctx);
180   return result;
181}
182
183static bool
184blorp_params_get_clear_kernel(struct blorp_batch *batch,
185                              struct blorp_params *params,
186                              bool use_replicated_data,
187                              bool clear_rgb_as_red)
188{
189   if (batch->flags & BLORP_BATCH_USE_COMPUTE) {
190      assert(!use_replicated_data);
191      return blorp_params_get_clear_kernel_cs(batch, params, clear_rgb_as_red);
192   } else {
193      return blorp_params_get_clear_kernel_fs(batch, params,
194                                              use_replicated_data,
195                                              clear_rgb_as_red);
196   }
197}
198
199#pragma pack(push, 1)
200struct layer_offset_vs_key {
201   struct brw_blorp_base_key base;
202   unsigned num_inputs;
203};
204#pragma pack(pop)
205
206/* In the case of doing attachment clears, we are using a surface state that
207 * is handed to us so we can't set (and don't even know) the base array layer.
208 * In order to do a layered clear in this scenario, we need some way of adding
209 * the base array layer to the instance id.  Unfortunately, our hardware has
210 * no real concept of "base instance", so we have to do it manually in a
211 * vertex shader.
212 */
213static bool
214blorp_params_get_layer_offset_vs(struct blorp_batch *batch,
215                                 struct blorp_params *params)
216{
217   struct blorp_context *blorp = batch->blorp;
218   struct layer_offset_vs_key blorp_key = {
219      .base = BRW_BLORP_BASE_KEY_INIT(BLORP_SHADER_TYPE_LAYER_OFFSET_VS),
220   };
221
222   if (params->wm_prog_data)
223      blorp_key.num_inputs = params->wm_prog_data->num_varying_inputs;
224
225   if (blorp->lookup_shader(batch, &blorp_key, sizeof(blorp_key),
226                            &params->vs_prog_kernel, &params->vs_prog_data))
227      return true;
228
229   void *mem_ctx = ralloc_context(NULL);
230
231   nir_builder b;
232   blorp_nir_init_shader(&b, mem_ctx, MESA_SHADER_VERTEX,
233                         blorp_shader_type_to_name(blorp_key.base.shader_type));
234
235   const struct glsl_type *uvec4_type = glsl_vector_type(GLSL_TYPE_UINT, 4);
236
237   /* First we deal with the header which has instance and base instance */
238   nir_variable *a_header = nir_variable_create(b.shader, nir_var_shader_in,
239                                                uvec4_type, "header");
240   a_header->data.location = VERT_ATTRIB_GENERIC0;
241
242   nir_variable *v_layer = nir_variable_create(b.shader, nir_var_shader_out,
243                                               glsl_int_type(), "layer_id");
244   v_layer->data.location = VARYING_SLOT_LAYER;
245
246   /* Compute the layer id */
247   nir_ssa_def *header = nir_load_var(&b, a_header);
248   nir_ssa_def *base_layer = nir_channel(&b, header, 0);
249   nir_ssa_def *instance = nir_channel(&b, header, 1);
250   nir_store_var(&b, v_layer, nir_iadd(&b, instance, base_layer), 0x1);
251
252   /* Then we copy the vertex from the next slot to VARYING_SLOT_POS */
253   nir_variable *a_vertex = nir_variable_create(b.shader, nir_var_shader_in,
254                                                glsl_vec4_type(), "a_vertex");
255   a_vertex->data.location = VERT_ATTRIB_GENERIC1;
256
257   nir_variable *v_pos = nir_variable_create(b.shader, nir_var_shader_out,
258                                             glsl_vec4_type(), "v_pos");
259   v_pos->data.location = VARYING_SLOT_POS;
260
261   nir_copy_var(&b, v_pos, a_vertex);
262
263   /* Then we copy everything else */
264   for (unsigned i = 0; i < blorp_key.num_inputs; i++) {
265      nir_variable *a_in = nir_variable_create(b.shader, nir_var_shader_in,
266                                               uvec4_type, "input");
267      a_in->data.location = VERT_ATTRIB_GENERIC2 + i;
268
269      nir_variable *v_out = nir_variable_create(b.shader, nir_var_shader_out,
270                                                uvec4_type, "output");
271      v_out->data.location = VARYING_SLOT_VAR0 + i;
272
273      nir_copy_var(&b, v_out, a_in);
274   }
275
276   struct brw_vs_prog_data vs_prog_data;
277   memset(&vs_prog_data, 0, sizeof(vs_prog_data));
278
279   const unsigned *program =
280      blorp_compile_vs(blorp, mem_ctx, b.shader, &vs_prog_data);
281
282   bool result =
283      blorp->upload_shader(batch, MESA_SHADER_VERTEX,
284                           &blorp_key, sizeof(blorp_key),
285                           program, vs_prog_data.base.base.program_size,
286                           &vs_prog_data.base.base, sizeof(vs_prog_data),
287                           &params->vs_prog_kernel, &params->vs_prog_data);
288
289   ralloc_free(mem_ctx);
290   return result;
291}
292
293/* The x0, y0, x1, and y1 parameters must already be populated with the render
294 * area of the framebuffer to be cleared.
295 */
296static void
297get_fast_clear_rect(const struct isl_device *dev,
298                    const struct isl_surf *aux_surf,
299                    unsigned *x0, unsigned *y0,
300                    unsigned *x1, unsigned *y1)
301{
302   unsigned int x_align, y_align;
303   unsigned int x_scaledown, y_scaledown;
304
305   /* Only single sampled surfaces need to (and actually can) be resolved. */
306   if (aux_surf->usage == ISL_SURF_USAGE_CCS_BIT) {
307      /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
308       * Target(s)", beneath the "Fast Color Clear" bullet (p327):
309       *
310       *     Clear pass must have a clear rectangle that must follow
311       *     alignment rules in terms of pixels and lines as shown in the
312       *     table below. Further, the clear-rectangle height and width
313       *     must be multiple of the following dimensions. If the height
314       *     and width of the render target being cleared do not meet these
315       *     requirements, an MCS buffer can be created such that it
316       *     follows the requirement and covers the RT.
317       *
318       * The alignment size in the table that follows is related to the
319       * alignment size that is baked into the CCS surface format but with X
320       * alignment multiplied by 16 and Y alignment multiplied by 32.
321       */
322      x_align = isl_format_get_layout(aux_surf->format)->bw;
323      y_align = isl_format_get_layout(aux_surf->format)->bh;
324
325      x_align *= 16;
326
327      /* The line alignment requirement for Y-tiled is halved at SKL and again
328       * at TGL.
329       */
330      if (dev->info->ver >= 12)
331         y_align *= 8;
332      else if (dev->info->ver >= 9)
333         y_align *= 16;
334      else
335         y_align *= 32;
336
337      /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
338       * Target(s)", beneath the "Fast Color Clear" bullet (p327):
339       *
340       *     In order to optimize the performance MCS buffer (when bound to
341       *     1X RT) clear similarly to MCS buffer clear for MSRT case,
342       *     clear rect is required to be scaled by the following factors
343       *     in the horizontal and vertical directions:
344       *
345       * The X and Y scale down factors in the table that follows are each
346       * equal to half the alignment value computed above.
347       */
348      x_scaledown = x_align / 2;
349      y_scaledown = y_align / 2;
350
351      if (ISL_DEV_IS_HASWELL(dev)) {
352         /* From BSpec: 3D-Media-GPGPU Engine > 3D Pipeline > Pixel > Pixel
353          * Backend > MCS Buffer for Render Target(s) [DevIVB+] > Table "Color
354          * Clear of Non-MultiSampled Render Target Restrictions":
355          *
356          *   Clear rectangle must be aligned to two times the number of
357          *   pixels in the table shown below due to 16x16 hashing across the
358          *   slice.
359          *
360          * This restriction is only documented to exist on HSW GT3 but
361          * empirical evidence suggests that it's also needed GT2.
362          */
363         x_align *= 2;
364         y_align *= 2;
365      }
366   } else {
367      assert(aux_surf->usage == ISL_SURF_USAGE_MCS_BIT);
368
369      /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
370       * Target(s)", beneath the "MSAA Compression" bullet (p326):
371       *
372       *     Clear pass for this case requires that scaled down primitive
373       *     is sent down with upper left co-ordinate to coincide with
374       *     actual rectangle being cleared. For MSAA, clear rectangle’s
375       *     height and width need to as show in the following table in
376       *     terms of (width,height) of the RT.
377       *
378       *     MSAA  Width of Clear Rect  Height of Clear Rect
379       *      2X     Ceil(1/8*width)      Ceil(1/2*height)
380       *      4X     Ceil(1/8*width)      Ceil(1/2*height)
381       *      8X     Ceil(1/2*width)      Ceil(1/2*height)
382       *     16X         width            Ceil(1/2*height)
383       *
384       * The text "with upper left co-ordinate to coincide with actual
385       * rectangle being cleared" is a little confusing--it seems to imply
386       * that to clear a rectangle from (x,y) to (x+w,y+h), one needs to
387       * feed the pipeline using the rectangle (x,y) to
388       * (x+Ceil(w/N),y+Ceil(h/2)), where N is either 2 or 8 depending on
389       * the number of samples.  Experiments indicate that this is not
390       * quite correct; actually, what the hardware appears to do is to
391       * align whatever rectangle is sent down the pipeline to the nearest
392       * multiple of 2x2 blocks, and then scale it up by a factor of N
393       * horizontally and 2 vertically.  So the resulting alignment is 4
394       * vertically and either 4 or 16 horizontally, and the scaledown
395       * factor is 2 vertically and either 2 or 8 horizontally.
396       */
397      switch (aux_surf->format) {
398      case ISL_FORMAT_MCS_2X:
399      case ISL_FORMAT_MCS_4X:
400         x_scaledown = 8;
401         break;
402      case ISL_FORMAT_MCS_8X:
403         x_scaledown = 2;
404         break;
405      case ISL_FORMAT_MCS_16X:
406         x_scaledown = 1;
407         break;
408      default:
409         unreachable("Unexpected MCS format for fast clear");
410      }
411      y_scaledown = 2;
412      x_align = x_scaledown * 2;
413      y_align = y_scaledown * 2;
414   }
415
416   *x0 = ROUND_DOWN_TO(*x0,  x_align) / x_scaledown;
417   *y0 = ROUND_DOWN_TO(*y0, y_align) / y_scaledown;
418   *x1 = ALIGN(*x1, x_align) / x_scaledown;
419   *y1 = ALIGN(*y1, y_align) / y_scaledown;
420}
421
422void
423blorp_fast_clear(struct blorp_batch *batch,
424                 const struct blorp_surf *surf,
425                 enum isl_format format, struct isl_swizzle swizzle,
426                 uint32_t level, uint32_t start_layer, uint32_t num_layers,
427                 uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1)
428{
429   struct blorp_params params;
430   blorp_params_init(&params);
431   params.num_layers = num_layers;
432   assert((batch->flags & BLORP_BATCH_USE_COMPUTE) == 0);
433
434   params.x0 = x0;
435   params.y0 = y0;
436   params.x1 = x1;
437   params.y1 = y1;
438
439   memset(&params.wm_inputs.clear_color, 0xff, 4*sizeof(float));
440   params.fast_clear_op = ISL_AUX_OP_FAST_CLEAR;
441
442   get_fast_clear_rect(batch->blorp->isl_dev, surf->aux_surf,
443                       &params.x0, &params.y0, &params.x1, &params.y1);
444
445   if (!blorp_params_get_clear_kernel(batch, &params, true, false))
446      return;
447
448   brw_blorp_surface_info_init(batch, &params.dst, surf, level,
449                               start_layer, format, true);
450   params.num_samples = params.dst.surf.samples;
451
452   assert(params.num_samples != 0);
453   if (params.num_samples == 1)
454      params.snapshot_type = INTEL_SNAPSHOT_CCS_COLOR_CLEAR;
455   else
456      params.snapshot_type = INTEL_SNAPSHOT_MCS_COLOR_CLEAR;
457
458   /* If a swizzle was provided, we need to swizzle the clear color so that
459    * the hardware color format conversion will work properly.
460    */
461   params.dst.clear_color =
462      isl_color_value_swizzle_inv(params.dst.clear_color, swizzle);
463
464   batch->blorp->exec(batch, &params);
465}
466
467bool
468blorp_clear_supports_compute(struct blorp_context *blorp,
469                             uint8_t color_write_disable, bool blend_enabled,
470                             enum isl_aux_usage aux_usage)
471{
472   if (blorp->isl_dev->info->ver < 7)
473      return false;
474   if (color_write_disable != 0 || blend_enabled)
475      return false;
476   if (blorp->isl_dev->info->ver >= 12) {
477      return aux_usage == ISL_AUX_USAGE_GFX12_CCS_E ||
478             aux_usage == ISL_AUX_USAGE_CCS_E ||
479             aux_usage == ISL_AUX_USAGE_NONE;
480   } else {
481      return aux_usage == ISL_AUX_USAGE_NONE;
482   }
483}
484
485void
486blorp_clear(struct blorp_batch *batch,
487            const struct blorp_surf *surf,
488            enum isl_format format, struct isl_swizzle swizzle,
489            uint32_t level, uint32_t start_layer, uint32_t num_layers,
490            uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1,
491            union isl_color_value clear_color,
492            uint8_t color_write_disable)
493{
494   struct blorp_params params;
495   blorp_params_init(&params);
496   params.snapshot_type = INTEL_SNAPSHOT_SLOW_COLOR_CLEAR;
497
498   const bool compute = batch->flags & BLORP_BATCH_USE_COMPUTE;
499   if (compute)
500      assert(blorp_clear_supports_compute(batch->blorp, color_write_disable,
501                                          false, surf->aux_usage));
502
503   /* Manually apply the clear destination swizzle.  This way swizzled clears
504    * will work for swizzles which we can't normally use for rendering and it
505    * also ensures that they work on pre-Haswell hardware which can't swizlle
506    * at all.
507    */
508   clear_color = isl_color_value_swizzle_inv(clear_color, swizzle);
509   swizzle = ISL_SWIZZLE_IDENTITY;
510
511   bool clear_rgb_as_red = false;
512   if (format == ISL_FORMAT_R9G9B9E5_SHAREDEXP) {
513      clear_color.u32[0] = float3_to_rgb9e5(clear_color.f32);
514      format = ISL_FORMAT_R32_UINT;
515   } else if (format == ISL_FORMAT_L8_UNORM_SRGB) {
516      clear_color.f32[0] = util_format_linear_to_srgb_float(clear_color.f32[0]);
517      format = ISL_FORMAT_R8_UNORM;
518   } else if (format == ISL_FORMAT_A4B4G4R4_UNORM) {
519      /* Broadwell and earlier cannot render to this format so we need to work
520       * around it by swapping the colors around and using B4G4R4A4 instead.
521       */
522      const struct isl_swizzle ARGB = ISL_SWIZZLE(ALPHA, RED, GREEN, BLUE);
523      clear_color = isl_color_value_swizzle_inv(clear_color, ARGB);
524      format = ISL_FORMAT_B4G4R4A4_UNORM;
525   } else if (isl_format_get_layout(format)->bpb % 3 == 0) {
526      clear_rgb_as_red = true;
527      if (format == ISL_FORMAT_R8G8B8_UNORM_SRGB) {
528         clear_color.f32[0] = util_format_linear_to_srgb_float(clear_color.f32[0]);
529         clear_color.f32[1] = util_format_linear_to_srgb_float(clear_color.f32[1]);
530         clear_color.f32[2] = util_format_linear_to_srgb_float(clear_color.f32[2]);
531      }
532   }
533
534   memcpy(&params.wm_inputs.clear_color, clear_color.f32, sizeof(float) * 4);
535
536   bool use_simd16_replicated_data = true;
537
538   /* From the SNB PRM (Vol4_Part1):
539    *
540    *     "Replicated data (Message Type = 111) is only supported when
541    *      accessing tiled memory.  Using this Message Type to access linear
542    *      (untiled) memory is UNDEFINED."
543    */
544   if (surf->surf->tiling == ISL_TILING_LINEAR)
545      use_simd16_replicated_data = false;
546
547   /* Replicated clears don't work yet before gfx6 */
548   if (batch->blorp->isl_dev->info->ver < 6)
549      use_simd16_replicated_data = false;
550
551   if (compute)
552      use_simd16_replicated_data = false;
553
554   /* Constant color writes ignore everyting in blend and color calculator
555    * state.  This is not documented.
556    */
557   params.color_write_disable = color_write_disable & BITFIELD_MASK(4);
558   if (color_write_disable)
559      use_simd16_replicated_data = false;
560
561   if (!blorp_params_get_clear_kernel(batch, &params,
562                                      use_simd16_replicated_data,
563                                      clear_rgb_as_red))
564      return;
565
566   if (!compute && !blorp_ensure_sf_program(batch, &params))
567      return;
568
569   while (num_layers > 0) {
570      brw_blorp_surface_info_init(batch, &params.dst, surf, level,
571                                  start_layer, format, true);
572      params.dst.view.swizzle = swizzle;
573
574      params.x0 = x0;
575      params.y0 = y0;
576      params.x1 = x1;
577      params.y1 = y1;
578
579      if (compute) {
580         params.wm_inputs.bounds_rect.x0 = x0;
581         params.wm_inputs.bounds_rect.y0 = y0;
582         params.wm_inputs.bounds_rect.x1 = x1;
583         params.wm_inputs.bounds_rect.y1 = y1;
584      }
585
586      if (params.dst.tile_x_sa || params.dst.tile_y_sa) {
587         assert(params.dst.surf.samples == 1);
588         assert(num_layers == 1);
589         params.x0 += params.dst.tile_x_sa;
590         params.y0 += params.dst.tile_y_sa;
591         params.x1 += params.dst.tile_x_sa;
592         params.y1 += params.dst.tile_y_sa;
593      }
594
595      /* The MinLOD and MinimumArrayElement don't work properly for cube maps.
596       * Convert them to a single slice on gfx4.
597       */
598      if (batch->blorp->isl_dev->info->ver == 4 &&
599          (params.dst.surf.usage & ISL_SURF_USAGE_CUBE_BIT)) {
600         blorp_surf_convert_to_single_slice(batch->blorp->isl_dev, &params.dst);
601      }
602
603      if (clear_rgb_as_red) {
604         surf_fake_rgb_with_red(batch->blorp->isl_dev, &params.dst);
605         params.x0 *= 3;
606         params.x1 *= 3;
607      }
608
609      if (isl_format_is_compressed(params.dst.surf.format)) {
610         blorp_surf_convert_to_uncompressed(batch->blorp->isl_dev, &params.dst,
611                                            NULL, NULL, NULL, NULL);
612                                            //&dst_x, &dst_y, &dst_w, &dst_h);
613      }
614
615      if (params.dst.tile_x_sa || params.dst.tile_y_sa) {
616         /* Either we're on gfx4 where there is no multisampling or the
617          * surface is compressed which also implies no multisampling.
618          * Therefore, sa == px and we don't need to do a conversion.
619          */
620         assert(params.dst.surf.samples == 1);
621         params.x0 += params.dst.tile_x_sa;
622         params.y0 += params.dst.tile_y_sa;
623         params.x1 += params.dst.tile_x_sa;
624         params.y1 += params.dst.tile_y_sa;
625      }
626
627      params.num_samples = params.dst.surf.samples;
628
629      /* We may be restricted on the number of layers we can bind at any one
630       * time.  In particular, Sandy Bridge has a maximum number of layers of
631       * 512 but a maximum 3D texture size is much larger.
632       */
633      params.num_layers = MIN2(params.dst.view.array_len, num_layers);
634
635      const unsigned max_image_width = 16 * 1024;
636      if (params.dst.surf.logical_level0_px.width > max_image_width) {
637         /* Clearing an RGB image as red multiplies the surface width by 3
638          * so it may now be too wide for the hardware surface limits.  We
639          * have to break the clear up into pieces in order to clear wide
640          * images.
641          */
642         assert(clear_rgb_as_red);
643         assert(params.dst.surf.dim == ISL_SURF_DIM_2D);
644         assert(params.dst.surf.tiling == ISL_TILING_LINEAR);
645         assert(params.dst.surf.logical_level0_px.depth == 1);
646         assert(params.dst.surf.logical_level0_px.array_len == 1);
647         assert(params.dst.surf.levels == 1);
648         assert(params.dst.surf.samples == 1);
649         assert(params.dst.tile_x_sa == 0 || params.dst.tile_y_sa == 0);
650         assert(params.dst.aux_usage == ISL_AUX_USAGE_NONE);
651
652         /* max_image_width rounded down to a multiple of 3 */
653         const unsigned max_fake_rgb_width = (max_image_width / 3) * 3;
654         const unsigned cpp =
655            isl_format_get_layout(params.dst.surf.format)->bpb / 8;
656
657         params.dst.surf.logical_level0_px.width = max_fake_rgb_width;
658         params.dst.surf.phys_level0_sa.width = max_fake_rgb_width;
659
660         uint32_t orig_x0 = params.x0, orig_x1 = params.x1;
661         uint64_t orig_offset = params.dst.addr.offset;
662         for (uint32_t x = orig_x0; x < orig_x1; x += max_fake_rgb_width) {
663            /* Offset to the surface.  It's easy because we're linear */
664            params.dst.addr.offset = orig_offset + x * cpp;
665
666            params.x0 = 0;
667            params.x1 = MIN2(orig_x1 - x, max_image_width);
668
669            batch->blorp->exec(batch, &params);
670         }
671      } else {
672         batch->blorp->exec(batch, &params);
673      }
674
675      start_layer += params.num_layers;
676      num_layers -= params.num_layers;
677   }
678}
679
680static bool
681blorp_clear_stencil_as_rgba(struct blorp_batch *batch,
682                            const struct blorp_surf *surf,
683                            uint32_t level, uint32_t start_layer,
684                            uint32_t num_layers,
685                            uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1,
686                            uint8_t stencil_mask, uint8_t stencil_value)
687{
688   assert((batch->flags & BLORP_BATCH_USE_COMPUTE) == 0);
689
690   /* We only support separate W-tiled stencil for now */
691   if (surf->surf->format != ISL_FORMAT_R8_UINT ||
692       surf->surf->tiling != ISL_TILING_W)
693      return false;
694
695   /* Stencil mask support would require piles of shader magic */
696   if (stencil_mask != 0xff)
697      return false;
698
699   if (surf->surf->samples > 1) {
700      /* Adjust x0, y0, x1, and y1 to be in units of samples */
701      assert(surf->surf->msaa_layout == ISL_MSAA_LAYOUT_INTERLEAVED);
702      struct isl_extent2d msaa_px_size_sa =
703         isl_get_interleaved_msaa_px_size_sa(surf->surf->samples);
704
705      x0 *= msaa_px_size_sa.w;
706      y0 *= msaa_px_size_sa.h;
707      x1 *= msaa_px_size_sa.w;
708      y1 *= msaa_px_size_sa.h;
709   }
710
711   /* W-tiles and Y-tiles have the same layout as far as cache lines are
712    * concerned: both are 8x8 cache lines laid out Y-major.  The difference is
713    * entirely in how the data is arranged withing the cache line.  W-tiling
714    * is 8x8 pixels in a swizzled pattern while Y-tiling is 16B by 4 rows
715    * regardless of image format size.  As long as everything is aligned to 8,
716    * we can just treat the W-tiled image as Y-tiled, ignore the layout
717    * difference within a cache line, and blast out data.
718    */
719   if (x0 % 8 != 0 || y0 % 8 != 0 || x1 % 8 != 0 || y1 % 8 != 0)
720      return false;
721
722   struct blorp_params params;
723   blorp_params_init(&params);
724   params.snapshot_type = INTEL_SNAPSHOT_SLOW_DEPTH_CLEAR;
725
726   if (!blorp_params_get_clear_kernel(batch, &params, true, false))
727      return false;
728
729   memset(&params.wm_inputs.clear_color, stencil_value,
730          sizeof(params.wm_inputs.clear_color));
731
732   /* The Sandy Bridge PRM Vol. 4 Pt. 2, section 2.11.2.1.1 has the
733    * following footnote to the format table:
734    *
735    *    128 BPE Formats cannot be Tiled Y when used as render targets
736    *
737    * We have to use RGBA16_UINT on SNB.
738    */
739   enum isl_format wide_format;
740   if (ISL_GFX_VER(batch->blorp->isl_dev) <= 6) {
741      wide_format = ISL_FORMAT_R16G16B16A16_UINT;
742
743      /* For RGBA16_UINT, we need to mask the stencil value otherwise, we risk
744       * clamping giving us the wrong values
745       */
746      for (unsigned i = 0; i < 4; i++)
747         params.wm_inputs.clear_color[i] &= 0xffff;
748   } else {
749      wide_format = ISL_FORMAT_R32G32B32A32_UINT;
750   }
751
752   for (uint32_t a = 0; a < num_layers; a++) {
753      uint32_t layer = start_layer + a;
754
755      brw_blorp_surface_info_init(batch, &params.dst, surf, level,
756                                  layer, ISL_FORMAT_UNSUPPORTED, true);
757
758      if (surf->surf->samples > 1)
759         blorp_surf_fake_interleaved_msaa(batch->blorp->isl_dev, &params.dst);
760
761      /* Make it Y-tiled */
762      blorp_surf_retile_w_to_y(batch->blorp->isl_dev, &params.dst);
763
764      unsigned wide_Bpp =
765         isl_format_get_layout(wide_format)->bpb / 8;
766
767      params.dst.view.format = params.dst.surf.format = wide_format;
768      assert(params.dst.surf.logical_level0_px.width % wide_Bpp == 0);
769      params.dst.surf.logical_level0_px.width /= wide_Bpp;
770      assert(params.dst.tile_x_sa % wide_Bpp == 0);
771      params.dst.tile_x_sa /= wide_Bpp;
772
773      params.x0 = params.dst.tile_x_sa + x0 / (wide_Bpp / 2);
774      params.y0 = params.dst.tile_y_sa + y0 / 2;
775      params.x1 = params.dst.tile_x_sa + x1 / (wide_Bpp / 2);
776      params.y1 = params.dst.tile_y_sa + y1 / 2;
777
778      batch->blorp->exec(batch, &params);
779   }
780
781   return true;
782}
783
784void
785blorp_clear_depth_stencil(struct blorp_batch *batch,
786                          const struct blorp_surf *depth,
787                          const struct blorp_surf *stencil,
788                          uint32_t level, uint32_t start_layer,
789                          uint32_t num_layers,
790                          uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1,
791                          bool clear_depth, float depth_value,
792                          uint8_t stencil_mask, uint8_t stencil_value)
793{
794   assert((batch->flags & BLORP_BATCH_USE_COMPUTE) == 0);
795
796   if (!clear_depth && blorp_clear_stencil_as_rgba(batch, stencil, level,
797                                                   start_layer, num_layers,
798                                                   x0, y0, x1, y1,
799                                                   stencil_mask,
800                                                   stencil_value))
801      return;
802
803   struct blorp_params params;
804   blorp_params_init(&params);
805   params.snapshot_type = INTEL_SNAPSHOT_SLOW_DEPTH_CLEAR;
806
807   params.x0 = x0;
808   params.y0 = y0;
809   params.x1 = x1;
810   params.y1 = y1;
811
812   if (ISL_GFX_VER(batch->blorp->isl_dev) == 6) {
813      /* For some reason, Sandy Bridge gets occlusion queries wrong if we
814       * don't have a shader.  In particular, it records samples even though
815       * we disable statistics in 3DSTATE_WM.  Give it the usual clear shader
816       * to work around the issue.
817       */
818      if (!blorp_params_get_clear_kernel(batch, &params, false, false))
819         return;
820   }
821
822   while (num_layers > 0) {
823      params.num_layers = num_layers;
824
825      if (stencil_mask) {
826         brw_blorp_surface_info_init(batch, &params.stencil, stencil,
827                                     level, start_layer,
828                                     ISL_FORMAT_UNSUPPORTED, true);
829         params.stencil_mask = stencil_mask;
830         params.stencil_ref = stencil_value;
831
832         params.dst.surf.samples = params.stencil.surf.samples;
833         params.dst.surf.logical_level0_px =
834            params.stencil.surf.logical_level0_px;
835         params.dst.view = params.stencil.view;
836
837         params.num_samples = params.stencil.surf.samples;
838
839         /* We may be restricted on the number of layers we can bind at any
840          * one time.  In particular, Sandy Bridge has a maximum number of
841          * layers of 512 but a maximum 3D texture size is much larger.
842          */
843         if (params.stencil.view.array_len < params.num_layers)
844            params.num_layers = params.stencil.view.array_len;
845      }
846
847      if (clear_depth) {
848         brw_blorp_surface_info_init(batch, &params.depth, depth,
849                                     level, start_layer,
850                                     ISL_FORMAT_UNSUPPORTED, true);
851         params.z = depth_value;
852         params.depth_format =
853            isl_format_get_depth_format(depth->surf->format, false);
854
855         params.dst.surf.samples = params.depth.surf.samples;
856         params.dst.surf.logical_level0_px =
857            params.depth.surf.logical_level0_px;
858         params.dst.view = params.depth.view;
859
860         params.num_samples = params.depth.surf.samples;
861
862         /* We may be restricted on the number of layers we can bind at any
863          * one time.  In particular, Sandy Bridge has a maximum number of
864          * layers of 512 but a maximum 3D texture size is much larger.
865          */
866         if (params.depth.view.array_len < params.num_layers)
867            params.num_layers = params.depth.view.array_len;
868      }
869
870      batch->blorp->exec(batch, &params);
871
872      start_layer += params.num_layers;
873      num_layers -= params.num_layers;
874   }
875}
876
877bool
878blorp_can_hiz_clear_depth(const struct intel_device_info *devinfo,
879                          const struct isl_surf *surf,
880                          enum isl_aux_usage aux_usage,
881                          uint32_t level, uint32_t layer,
882                          uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1)
883{
884   /* This function currently doesn't support any gen prior to gfx8 */
885   assert(devinfo->ver >= 8);
886
887   if (devinfo->ver == 8 && surf->format == ISL_FORMAT_R16_UNORM) {
888      /* Apply the D16 alignment restrictions. On BDW, HiZ has an 8x4 sample
889       * block with the following property: as the number of samples increases,
890       * the number of pixels representable by this block decreases by a factor
891       * of the sample dimensions. Sample dimensions scale following the MSAA
892       * interleaved pattern.
893       *
894       * Sample|Sample|Pixel
895       * Count |Dim   |Dim
896       * ===================
897       *    1  | 1x1  | 8x4
898       *    2  | 2x1  | 4x4
899       *    4  | 2x2  | 4x2
900       *    8  | 4x2  | 2x2
901       *   16  | 4x4  | 2x1
902       *
903       * Table: Pixel Dimensions in a HiZ Sample Block Pre-SKL
904       */
905      const struct isl_extent2d sa_block_dim =
906         isl_get_interleaved_msaa_px_size_sa(surf->samples);
907      const uint8_t align_px_w = 8 / sa_block_dim.w;
908      const uint8_t align_px_h = 4 / sa_block_dim.h;
909
910      /* Fast depth clears clear an entire sample block at a time. As a result,
911       * the rectangle must be aligned to the dimensions of the encompassing
912       * pixel block for a successful operation.
913       *
914       * Fast clears can still work if the upper-left corner is aligned and the
915       * bottom-rigtht corner touches the edge of a depth buffer whose extent
916       * is unaligned. This is because each miplevel in the depth buffer is
917       * padded by the Pixel Dim (similar to a standard compressed texture).
918       * In this case, the clear rectangle could be padded by to match the full
919       * depth buffer extent but to support multiple clearing techniques, we
920       * chose to be unaware of the depth buffer's extent and thus don't handle
921       * this case.
922       */
923      if (x0 % align_px_w || y0 % align_px_h ||
924          x1 % align_px_w || y1 % align_px_h)
925         return false;
926   } else if (aux_usage == ISL_AUX_USAGE_HIZ_CCS_WT) {
927      /* We have to set the WM_HZ_OP::FullSurfaceDepthandStencilClear bit
928       * whenever we clear an uninitialized HIZ buffer (as some drivers
929       * currently do). However, this bit seems liable to clear 16x8 pixels in
930       * the ZCS on Gfx12 - greater than the slice alignments for depth
931       * buffers.
932       */
933      assert(surf->image_alignment_el.w % 16 != 0 ||
934             surf->image_alignment_el.h % 8 != 0);
935
936      /* This is the hypothesis behind some corruption that was seen with the
937       * amd_vertex_shader_layer-layered-depth-texture-render piglit test.
938       *
939       * From the Compressed Depth Buffers section of the Bspec, under the
940       * Gfx12 texture performant and ZCS columns:
941       *
942       *    Update with clear at either 16x8 or 8x4 granularity, based on
943       *    fs_clr or otherwise.
944       *
945       * There are a number of ways to avoid full surface CCS clears that
946       * overlap other slices, but for now we choose to disable fast-clears
947       * when an initializing clear could hit another miplevel.
948       *
949       * NOTE: Because the CCS compresses the depth buffer and not a version
950       * of it that has been rearranged with different alignments (like Gfx8+
951       * HIZ), we have to make sure that the x0 and y0 are at least 16x8
952       * aligned in the context of the entire surface.
953       */
954      uint32_t slice_x0, slice_y0, slice_z0, slice_a0;
955      isl_surf_get_image_offset_el(surf, level,
956                                   surf->dim == ISL_SURF_DIM_3D ? 0 : layer,
957                                   surf->dim == ISL_SURF_DIM_3D ? layer: 0,
958                                   &slice_x0, &slice_y0, &slice_z0, &slice_a0);
959      assert(slice_z0 == 0 && slice_a0 == 0);
960      const bool max_x1_y1 =
961         x1 == minify(surf->logical_level0_px.width, level) &&
962         y1 == minify(surf->logical_level0_px.height, level);
963      const uint32_t haligned_x1 = ALIGN(x1, surf->image_alignment_el.w);
964      const uint32_t valigned_y1 = ALIGN(y1, surf->image_alignment_el.h);
965      const bool unaligned = (slice_x0 + x0) % 16 || (slice_y0 + y0) % 8 ||
966                             (max_x1_y1 ? haligned_x1 % 16 || valigned_y1 % 8 :
967                              x1 % 16 || y1 % 8);
968      const bool partial_clear = x0 > 0 || y0 > 0 || !max_x1_y1;
969      const bool multislice_surf = surf->levels > 1 ||
970                                   surf->logical_level0_px.depth > 1 ||
971                                   surf->logical_level0_px.array_len > 1;
972
973      if (unaligned && (partial_clear || multislice_surf))
974         return false;
975   }
976
977   return isl_aux_usage_has_hiz(aux_usage);
978}
979
980static bool
981blorp_can_clear_full_surface(const struct blorp_surf *depth,
982                             const struct blorp_surf *stencil,
983                             uint32_t level,
984                             uint32_t x0, uint32_t y0,
985                             uint32_t x1, uint32_t y1,
986                             bool clear_depth,
987                             bool clear_stencil)
988{
989   uint32_t width = 0, height = 0;
990   if (clear_stencil) {
991      width = minify(stencil->surf->logical_level0_px.width, level);
992      height = minify(stencil->surf->logical_level0_px.height, level);
993   }
994
995   if (clear_depth && !(width || height)) {
996      width = minify(depth->surf->logical_level0_px.width, level);
997      height = minify(depth->surf->logical_level0_px.height, level);
998   }
999
1000   return x0 == 0 && y0 == 0 && width == x1 && height == y1;
1001}
1002
1003void
1004blorp_hiz_clear_depth_stencil(struct blorp_batch *batch,
1005                              const struct blorp_surf *depth,
1006                              const struct blorp_surf *stencil,
1007                              uint32_t level,
1008                              uint32_t start_layer, uint32_t num_layers,
1009                              uint32_t x0, uint32_t y0,
1010                              uint32_t x1, uint32_t y1,
1011                              bool clear_depth, float depth_value,
1012                              bool clear_stencil, uint8_t stencil_value)
1013{
1014   struct blorp_params params;
1015   blorp_params_init(&params);
1016   params.snapshot_type = INTEL_SNAPSHOT_HIZ_CLEAR;
1017
1018   /* This requires WM_HZ_OP which only exists on gfx8+ */
1019   assert(ISL_GFX_VER(batch->blorp->isl_dev) >= 8);
1020
1021   params.hiz_op = ISL_AUX_OP_FAST_CLEAR;
1022   /* From BSpec: 3DSTATE_WM_HZ_OP_BODY >> Full Surface Depth and Stencil Clear
1023    *
1024    *    "Software must set this only when the APP requires the entire Depth
1025    *    surface to be cleared."
1026    */
1027   params.full_surface_hiz_op =
1028      blorp_can_clear_full_surface(depth, stencil, level, x0, y0, x1, y1,
1029                                   clear_depth, clear_stencil);
1030   params.num_layers = 1;
1031
1032   params.x0 = x0;
1033   params.y0 = y0;
1034   params.x1 = x1;
1035   params.y1 = y1;
1036
1037   for (uint32_t l = 0; l < num_layers; l++) {
1038      const uint32_t layer = start_layer + l;
1039      if (clear_stencil) {
1040         brw_blorp_surface_info_init(batch, &params.stencil, stencil,
1041                                     level, layer,
1042                                     ISL_FORMAT_UNSUPPORTED, true);
1043         params.stencil_mask = 0xff;
1044         params.stencil_ref = stencil_value;
1045         params.num_samples = params.stencil.surf.samples;
1046      }
1047
1048      if (clear_depth) {
1049         /* If we're clearing depth, we must have HiZ */
1050         assert(depth && isl_aux_usage_has_hiz(depth->aux_usage));
1051
1052         brw_blorp_surface_info_init(batch, &params.depth, depth,
1053                                     level, layer,
1054                                     ISL_FORMAT_UNSUPPORTED, true);
1055         params.depth.clear_color.f32[0] = depth_value;
1056         params.depth_format =
1057            isl_format_get_depth_format(depth->surf->format, false);
1058         params.num_samples = params.depth.surf.samples;
1059      }
1060
1061      batch->blorp->exec(batch, &params);
1062   }
1063}
1064
1065/* Given a depth stencil attachment, this function performs a fast depth clear
1066 * on a depth portion and a regular clear on the stencil portion. When
1067 * performing a fast depth clear on the depth portion, the HiZ buffer is simply
1068 * tagged as cleared so the depth clear value is not actually needed.
1069 */
1070void
1071blorp_gfx8_hiz_clear_attachments(struct blorp_batch *batch,
1072                                 uint32_t num_samples,
1073                                 uint32_t x0, uint32_t y0,
1074                                 uint32_t x1, uint32_t y1,
1075                                 bool clear_depth, bool clear_stencil,
1076                                 uint8_t stencil_value)
1077{
1078   assert(batch->flags & BLORP_BATCH_NO_EMIT_DEPTH_STENCIL);
1079
1080   struct blorp_params params;
1081   blorp_params_init(&params);
1082   params.snapshot_type = INTEL_SNAPSHOT_HIZ_CLEAR;
1083   params.num_layers = 1;
1084   params.hiz_op = ISL_AUX_OP_FAST_CLEAR;
1085   params.x0 = x0;
1086   params.y0 = y0;
1087   params.x1 = x1;
1088   params.y1 = y1;
1089   params.num_samples = num_samples;
1090   params.depth.enabled = clear_depth;
1091   params.stencil.enabled = clear_stencil;
1092   params.stencil_ref = stencil_value;
1093   batch->blorp->exec(batch, &params);
1094}
1095
1096/** Clear active color/depth/stencili attachments
1097 *
1098 * This function performs a clear operation on the currently bound
1099 * color/depth/stencil attachments.  It is assumed that any information passed
1100 * in here is valid, consistent, and in-bounds relative to the currently
1101 * attached depth/stencil.  The binding_table_offset parameter is the 32-bit
1102 * offset relative to surface state base address where pre-baked binding table
1103 * that we are to use lives.  If clear_color is false, binding_table_offset
1104 * must point to a binding table with one entry which is a valid null surface
1105 * that matches the currently bound depth and stencil.
1106 */
1107void
1108blorp_clear_attachments(struct blorp_batch *batch,
1109                        uint32_t binding_table_offset,
1110                        enum isl_format depth_format,
1111                        uint32_t num_samples,
1112                        uint32_t start_layer, uint32_t num_layers,
1113                        uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1,
1114                        bool clear_color, union isl_color_value color_value,
1115                        bool clear_depth, float depth_value,
1116                        uint8_t stencil_mask, uint8_t stencil_value)
1117{
1118   struct blorp_params params;
1119   blorp_params_init(&params);
1120
1121   assert((batch->flags & BLORP_BATCH_USE_COMPUTE) == 0);
1122   assert(batch->flags & BLORP_BATCH_NO_EMIT_DEPTH_STENCIL);
1123
1124   params.x0 = x0;
1125   params.y0 = y0;
1126   params.x1 = x1;
1127   params.y1 = y1;
1128
1129   params.use_pre_baked_binding_table = true;
1130   params.pre_baked_binding_table_offset = binding_table_offset;
1131
1132   params.num_layers = num_layers;
1133   params.num_samples = num_samples;
1134
1135   if (clear_color) {
1136      params.dst.enabled = true;
1137      params.snapshot_type = INTEL_SNAPSHOT_SLOW_COLOR_CLEAR;
1138
1139      memcpy(&params.wm_inputs.clear_color, color_value.f32, sizeof(float) * 4);
1140
1141      /* Unfortunately, without knowing whether or not our destination surface
1142       * is tiled or not, we have to assume it may be linear.  This means no
1143       * SIMD16_REPDATA for us. :-(
1144       */
1145      if (!blorp_params_get_clear_kernel(batch, &params, false, false))
1146         return;
1147   }
1148
1149   if (clear_depth) {
1150      params.depth.enabled = true;
1151      params.snapshot_type = INTEL_SNAPSHOT_SLOW_DEPTH_CLEAR;
1152
1153      params.z = depth_value;
1154      params.depth_format = isl_format_get_depth_format(depth_format, false);
1155   }
1156
1157   if (stencil_mask) {
1158      params.stencil.enabled = true;
1159      params.snapshot_type = INTEL_SNAPSHOT_SLOW_DEPTH_CLEAR;
1160
1161      params.stencil_mask = stencil_mask;
1162      params.stencil_ref = stencil_value;
1163   }
1164
1165   if (!blorp_params_get_layer_offset_vs(batch, &params))
1166      return;
1167
1168   params.vs_inputs.base_layer = start_layer;
1169
1170   batch->blorp->exec(batch, &params);
1171}
1172
1173void
1174blorp_ccs_resolve(struct blorp_batch *batch,
1175                  struct blorp_surf *surf, uint32_t level,
1176                  uint32_t start_layer, uint32_t num_layers,
1177                  enum isl_format format,
1178                  enum isl_aux_op resolve_op)
1179{
1180   assert((batch->flags & BLORP_BATCH_USE_COMPUTE) == 0);
1181   struct blorp_params params;
1182
1183   blorp_params_init(&params);
1184   switch(resolve_op) {
1185   case ISL_AUX_OP_AMBIGUATE:
1186      params.snapshot_type = INTEL_SNAPSHOT_CCS_AMBIGUATE;
1187      break;
1188   case ISL_AUX_OP_FULL_RESOLVE:
1189      params.snapshot_type = INTEL_SNAPSHOT_CCS_RESOLVE;
1190      break;
1191   case ISL_AUX_OP_PARTIAL_RESOLVE:
1192      params.snapshot_type = INTEL_SNAPSHOT_CCS_PARTIAL_RESOLVE;
1193      break;
1194   default:
1195      assert(false);
1196   }
1197   brw_blorp_surface_info_init(batch, &params.dst, surf,
1198                               level, start_layer, format, true);
1199
1200   /* From the Ivy Bridge PRM, Vol2 Part1 11.9 "Render Target Resolve":
1201    *
1202    *     A rectangle primitive must be scaled down by the following factors
1203    *     with respect to render target being resolved.
1204    *
1205    * The scaledown factors in the table that follows are related to the block
1206    * size of the CCS format.  For IVB and HSW, we divide by two, for BDW we
1207    * multiply by 8 and 16. On Sky Lake, we multiply by 8.
1208    */
1209   const struct isl_format_layout *aux_fmtl =
1210      isl_format_get_layout(params.dst.aux_surf.format);
1211   assert(aux_fmtl->txc == ISL_TXC_CCS);
1212
1213   unsigned x_scaledown, y_scaledown;
1214   if (ISL_GFX_VER(batch->blorp->isl_dev) >= 12) {
1215      x_scaledown = aux_fmtl->bw * 8;
1216      y_scaledown = aux_fmtl->bh * 4;
1217   } else if (ISL_GFX_VER(batch->blorp->isl_dev) >= 9) {
1218      x_scaledown = aux_fmtl->bw * 8;
1219      y_scaledown = aux_fmtl->bh * 8;
1220   } else if (ISL_GFX_VER(batch->blorp->isl_dev) >= 8) {
1221      x_scaledown = aux_fmtl->bw * 8;
1222      y_scaledown = aux_fmtl->bh * 16;
1223   } else {
1224      x_scaledown = aux_fmtl->bw / 2;
1225      y_scaledown = aux_fmtl->bh / 2;
1226   }
1227   params.x0 = params.y0 = 0;
1228   params.x1 = minify(params.dst.surf.logical_level0_px.width, level);
1229   params.y1 = minify(params.dst.surf.logical_level0_px.height, level);
1230   params.x1 = ALIGN(params.x1, x_scaledown) / x_scaledown;
1231   params.y1 = ALIGN(params.y1, y_scaledown) / y_scaledown;
1232
1233   if (batch->blorp->isl_dev->info->ver >= 10) {
1234      assert(resolve_op == ISL_AUX_OP_FULL_RESOLVE ||
1235             resolve_op == ISL_AUX_OP_PARTIAL_RESOLVE ||
1236             resolve_op == ISL_AUX_OP_AMBIGUATE);
1237   } else if (batch->blorp->isl_dev->info->ver >= 9) {
1238      assert(resolve_op == ISL_AUX_OP_FULL_RESOLVE ||
1239             resolve_op == ISL_AUX_OP_PARTIAL_RESOLVE);
1240   } else {
1241      /* Broadwell and earlier do not have a partial resolve */
1242      assert(resolve_op == ISL_AUX_OP_FULL_RESOLVE);
1243   }
1244   params.fast_clear_op = resolve_op;
1245   params.num_layers = num_layers;
1246
1247   /* Note: there is no need to initialize push constants because it doesn't
1248    * matter what data gets dispatched to the render target.  However, we must
1249    * ensure that the fragment shader delivers the data using the "replicated
1250    * color" message.
1251    */
1252
1253   if (!blorp_params_get_clear_kernel(batch, &params, true, false))
1254      return;
1255
1256   batch->blorp->exec(batch, &params);
1257}
1258
1259static nir_ssa_def *
1260blorp_nir_bit(nir_builder *b, nir_ssa_def *src, unsigned bit)
1261{
1262   return nir_iand(b, nir_ushr(b, src, nir_imm_int(b, bit)),
1263                      nir_imm_int(b, 1));
1264}
1265
1266#pragma pack(push, 1)
1267struct blorp_mcs_partial_resolve_key
1268{
1269   struct brw_blorp_base_key base;
1270   bool indirect_clear_color;
1271   bool int_format;
1272   uint32_t num_samples;
1273};
1274#pragma pack(pop)
1275
1276static bool
1277blorp_params_get_mcs_partial_resolve_kernel(struct blorp_batch *batch,
1278                                            struct blorp_params *params)
1279{
1280   struct blorp_context *blorp = batch->blorp;
1281   const struct blorp_mcs_partial_resolve_key blorp_key = {
1282      .base = BRW_BLORP_BASE_KEY_INIT(BLORP_SHADER_TYPE_MCS_PARTIAL_RESOLVE),
1283      .indirect_clear_color = params->dst.clear_color_addr.buffer != NULL,
1284      .int_format = isl_format_has_int_channel(params->dst.view.format),
1285      .num_samples = params->num_samples,
1286   };
1287
1288   if (blorp->lookup_shader(batch, &blorp_key, sizeof(blorp_key),
1289                            &params->wm_prog_kernel, &params->wm_prog_data))
1290      return true;
1291
1292   void *mem_ctx = ralloc_context(NULL);
1293
1294   nir_builder b;
1295   blorp_nir_init_shader(&b, mem_ctx, MESA_SHADER_FRAGMENT,
1296                         blorp_shader_type_to_name(blorp_key.base.shader_type));
1297
1298   nir_variable *v_color =
1299      BLORP_CREATE_NIR_INPUT(b.shader, clear_color, glsl_vec4_type());
1300
1301   nir_variable *frag_color =
1302      nir_variable_create(b.shader, nir_var_shader_out,
1303                          glsl_vec4_type(), "gl_FragColor");
1304   frag_color->data.location = FRAG_RESULT_COLOR;
1305
1306   /* Do an MCS fetch and check if it is equal to the magic clear value */
1307   nir_ssa_def *mcs =
1308      blorp_nir_txf_ms_mcs(&b, nir_f2i32(&b, nir_load_frag_coord(&b)),
1309                               nir_load_layer_id(&b));
1310   nir_ssa_def *is_clear =
1311      blorp_nir_mcs_is_clear_color(&b, mcs, blorp_key.num_samples);
1312
1313   /* If we aren't the clear value, discard. */
1314   nir_discard_if(&b, nir_inot(&b, is_clear));
1315
1316   nir_ssa_def *clear_color = nir_load_var(&b, v_color);
1317   if (blorp_key.indirect_clear_color && blorp->isl_dev->info->ver <= 8) {
1318      /* Gfx7-8 clear colors are stored as single 0/1 bits */
1319      clear_color = nir_vec4(&b, blorp_nir_bit(&b, clear_color, 31),
1320                                 blorp_nir_bit(&b, clear_color, 30),
1321                                 blorp_nir_bit(&b, clear_color, 29),
1322                                 blorp_nir_bit(&b, clear_color, 28));
1323
1324      if (!blorp_key.int_format)
1325         clear_color = nir_i2f32(&b, clear_color);
1326   }
1327   nir_store_var(&b, frag_color, clear_color, 0xf);
1328
1329   struct brw_wm_prog_key wm_key;
1330   brw_blorp_init_wm_prog_key(&wm_key);
1331   wm_key.base.tex.compressed_multisample_layout_mask = 1;
1332   wm_key.base.tex.msaa_16 = blorp_key.num_samples == 16;
1333   wm_key.multisample_fbo = true;
1334
1335   struct brw_wm_prog_data prog_data;
1336   const unsigned *program =
1337      blorp_compile_fs(blorp, mem_ctx, b.shader, &wm_key, false,
1338                       &prog_data);
1339
1340   bool result =
1341      blorp->upload_shader(batch, MESA_SHADER_FRAGMENT,
1342                           &blorp_key, sizeof(blorp_key),
1343                           program, prog_data.base.program_size,
1344                           &prog_data.base, sizeof(prog_data),
1345                           &params->wm_prog_kernel, &params->wm_prog_data);
1346
1347   ralloc_free(mem_ctx);
1348   return result;
1349}
1350
1351void
1352blorp_mcs_partial_resolve(struct blorp_batch *batch,
1353                          struct blorp_surf *surf,
1354                          enum isl_format format,
1355                          uint32_t start_layer, uint32_t num_layers)
1356{
1357   struct blorp_params params;
1358   blorp_params_init(&params);
1359   params.snapshot_type = INTEL_SNAPSHOT_MCS_PARTIAL_RESOLVE;
1360
1361   assert(batch->blorp->isl_dev->info->ver >= 7);
1362
1363   params.x0 = 0;
1364   params.y0 = 0;
1365   params.x1 = surf->surf->logical_level0_px.width;
1366   params.y1 = surf->surf->logical_level0_px.height;
1367
1368   brw_blorp_surface_info_init(batch, &params.src, surf, 0,
1369                               start_layer, format, false);
1370   brw_blorp_surface_info_init(batch, &params.dst, surf, 0,
1371                               start_layer, format, true);
1372
1373   params.num_samples = params.dst.surf.samples;
1374   params.num_layers = num_layers;
1375   params.dst_clear_color_as_input = surf->clear_color_addr.buffer != NULL;
1376
1377   memcpy(&params.wm_inputs.clear_color,
1378          surf->clear_color.f32, sizeof(float) * 4);
1379
1380   if (!blorp_params_get_mcs_partial_resolve_kernel(batch, &params))
1381      return;
1382
1383   batch->blorp->exec(batch, &params);
1384}
1385
1386/** Clear a CCS to the "uncompressed" state
1387 *
1388 * This pass is the CCS equivalent of a "HiZ resolve".  It sets the CCS values
1389 * for a given layer/level of a surface to 0x0 which is the "uncompressed"
1390 * state which tells the sampler to go look at the main surface.
1391 */
1392void
1393blorp_ccs_ambiguate(struct blorp_batch *batch,
1394                    struct blorp_surf *surf,
1395                    uint32_t level, uint32_t layer)
1396{
1397   assert((batch->flags & BLORP_BATCH_USE_COMPUTE) == 0);
1398
1399   if (ISL_GFX_VER(batch->blorp->isl_dev) >= 10) {
1400      /* On gfx10 and above, we have a hardware resolve op for this */
1401      return blorp_ccs_resolve(batch, surf, level, layer, 1,
1402                               surf->surf->format, ISL_AUX_OP_AMBIGUATE);
1403   }
1404
1405   struct blorp_params params;
1406   blorp_params_init(&params);
1407   params.snapshot_type = INTEL_SNAPSHOT_CCS_AMBIGUATE;
1408
1409   assert(ISL_GFX_VER(batch->blorp->isl_dev) >= 7);
1410
1411   const struct isl_format_layout *aux_fmtl =
1412      isl_format_get_layout(surf->aux_surf->format);
1413   assert(aux_fmtl->txc == ISL_TXC_CCS);
1414
1415   params.dst = (struct brw_blorp_surface_info) {
1416      .enabled = true,
1417      .addr = surf->aux_addr,
1418      .view = {
1419         .usage = ISL_SURF_USAGE_RENDER_TARGET_BIT,
1420         .format = ISL_FORMAT_R32G32B32A32_UINT,
1421         .base_level = 0,
1422         .base_array_layer = 0,
1423         .levels = 1,
1424         .array_len = 1,
1425         .swizzle = ISL_SWIZZLE_IDENTITY,
1426      },
1427   };
1428
1429   uint32_t z = 0;
1430   if (surf->surf->dim == ISL_SURF_DIM_3D) {
1431      z = layer;
1432      layer = 0;
1433   }
1434
1435   uint64_t offset_B;
1436   uint32_t x_offset_el, y_offset_el;
1437   isl_surf_get_image_offset_B_tile_el(surf->aux_surf, level, layer, z,
1438                                       &offset_B, &x_offset_el, &y_offset_el);
1439   params.dst.addr.offset += offset_B;
1440
1441   const uint32_t width_px =
1442      minify(surf->aux_surf->logical_level0_px.width, level);
1443   const uint32_t height_px =
1444      minify(surf->aux_surf->logical_level0_px.height, level);
1445   const uint32_t width_el = DIV_ROUND_UP(width_px, aux_fmtl->bw);
1446   const uint32_t height_el = DIV_ROUND_UP(height_px, aux_fmtl->bh);
1447
1448   struct isl_tile_info ccs_tile_info;
1449   isl_surf_get_tile_info(surf->aux_surf, &ccs_tile_info);
1450
1451   /* We're going to map it as a regular RGBA32_UINT surface.  We need to
1452    * downscale a good deal.  We start by computing the area on the CCS to
1453    * clear in units of Y-tiled cache lines.
1454    */
1455   uint32_t x_offset_cl, y_offset_cl, width_cl, height_cl;
1456   if (ISL_GFX_VER(batch->blorp->isl_dev) >= 8) {
1457      /* From the Sky Lake PRM Vol. 12 in the section on planes:
1458       *
1459       *    "The Color Control Surface (CCS) contains the compression status
1460       *    of the cache-line pairs. The compression state of the cache-line
1461       *    pair is specified by 2 bits in the CCS.  Each CCS cache-line
1462       *    represents an area on the main surface of 16x16 sets of 128 byte
1463       *    Y-tiled cache-line-pairs. CCS is always Y tiled."
1464       *
1465       * Each 2-bit surface element in the CCS corresponds to a single
1466       * cache-line pair in the main surface.  This means that 16x16 el block
1467       * in the CCS maps to a Y-tiled cache line.  Fortunately, CCS layouts
1468       * are calculated with a very large alignment so we can round up to a
1469       * whole cache line without worrying about overdraw.
1470       */
1471
1472      /* On Broadwell and above, a CCS tile is the same as a Y tile when
1473       * viewed at the cache-line granularity.  Fortunately, the horizontal
1474       * and vertical alignment requirements of the CCS are such that we can
1475       * align to an entire cache line without worrying about crossing over
1476       * from one LOD to another.
1477       */
1478      const uint32_t x_el_per_cl = ccs_tile_info.logical_extent_el.w / 8;
1479      const uint32_t y_el_per_cl = ccs_tile_info.logical_extent_el.h / 8;
1480      assert(surf->aux_surf->image_alignment_el.w % x_el_per_cl == 0);
1481      assert(surf->aux_surf->image_alignment_el.h % y_el_per_cl == 0);
1482
1483      assert(x_offset_el % x_el_per_cl == 0);
1484      assert(y_offset_el % y_el_per_cl == 0);
1485      x_offset_cl = x_offset_el / x_el_per_cl;
1486      y_offset_cl = y_offset_el / y_el_per_cl;
1487      width_cl = DIV_ROUND_UP(width_el, x_el_per_cl);
1488      height_cl = DIV_ROUND_UP(height_el, y_el_per_cl);
1489   } else {
1490      /* On gfx7, the CCS tiling is not so nice.  However, there we are
1491       * guaranteed that we only have a single level and slice so we don't
1492       * have to worry about it and can just align to a whole tile.
1493       */
1494      assert(surf->aux_surf->logical_level0_px.depth == 1);
1495      assert(surf->aux_surf->logical_level0_px.array_len == 1);
1496      assert(x_offset_el == 0 && y_offset_el == 0);
1497      const uint32_t width_tl =
1498         DIV_ROUND_UP(width_el, ccs_tile_info.logical_extent_el.w);
1499      const uint32_t height_tl =
1500         DIV_ROUND_UP(height_el, ccs_tile_info.logical_extent_el.h);
1501      x_offset_cl = 0;
1502      y_offset_cl = 0;
1503      width_cl = width_tl * 8;
1504      height_cl = height_tl * 8;
1505   }
1506
1507   /* We're going to use a RGBA32 format so as to write data as quickly as
1508    * possible.  A y-tiled cache line will then be 1x4 px.
1509    */
1510   const uint32_t x_offset_rgba_px = x_offset_cl;
1511   const uint32_t y_offset_rgba_px = y_offset_cl * 4;
1512   const uint32_t width_rgba_px = width_cl;
1513   const uint32_t height_rgba_px = height_cl * 4;
1514
1515   ASSERTED bool ok =
1516      isl_surf_init(batch->blorp->isl_dev, &params.dst.surf,
1517                    .dim = ISL_SURF_DIM_2D,
1518                    .format = ISL_FORMAT_R32G32B32A32_UINT,
1519                    .width = width_rgba_px + x_offset_rgba_px,
1520                    .height = height_rgba_px + y_offset_rgba_px,
1521                    .depth = 1,
1522                    .levels = 1,
1523                    .array_len = 1,
1524                    .samples = 1,
1525                    .row_pitch_B = surf->aux_surf->row_pitch_B,
1526                    .usage = ISL_SURF_USAGE_RENDER_TARGET_BIT,
1527                    .tiling_flags = ISL_TILING_Y0_BIT);
1528   assert(ok);
1529
1530   params.x0 = x_offset_rgba_px;
1531   params.y0 = y_offset_rgba_px;
1532   params.x1 = x_offset_rgba_px + width_rgba_px;
1533   params.y1 = y_offset_rgba_px + height_rgba_px;
1534
1535   /* A CCS value of 0 means "uncompressed." */
1536   memset(&params.wm_inputs.clear_color, 0,
1537          sizeof(params.wm_inputs.clear_color));
1538
1539   if (!blorp_params_get_clear_kernel(batch, &params, true, false))
1540      return;
1541
1542   batch->blorp->exec(batch, &params);
1543}
1544