1/*
2 * Copyright © 2015 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 <assert.h>
25#include <stdbool.h>
26#include <string.h>
27#include <unistd.h>
28#include <fcntl.h>
29
30#include "anv_private.h"
31
32#include "common/intel_aux_map.h"
33#include "common/intel_sample_positions.h"
34#include "genxml/gen_macros.h"
35#include "genxml/genX_pack.h"
36
37#include "vk_util.h"
38
39/**
40 * Compute an \p n x \p m pixel hashing table usable as slice, subslice or
41 * pixel pipe hashing table.  The resulting table is the cyclic repetition of
42 * a fixed pattern with periodicity equal to \p period.
43 *
44 * If \p index is specified to be equal to \p period, a 2-way hashing table
45 * will be generated such that indices 0 and 1 are returned for the following
46 * fractions of entries respectively:
47 *
48 *   p_0 = ceil(period / 2) / period
49 *   p_1 = floor(period / 2) / period
50 *
51 * If \p index is even and less than \p period, a 3-way hashing table will be
52 * generated such that indices 0, 1 and 2 are returned for the following
53 * fractions of entries:
54 *
55 *   p_0 = (ceil(period / 2) - 1) / period
56 *   p_1 = floor(period / 2) / period
57 *   p_2 = 1 / period
58 *
59 * The equations above apply if \p flip is equal to 0, if it is equal to 1 p_0
60 * and p_1 will be swapped for the result.  Note that in the context of pixel
61 * pipe hashing this can be always 0 on Gfx12 platforms, since the hardware
62 * transparently remaps logical indices found on the table to physical pixel
63 * pipe indices from the highest to lowest EU count.
64 */
65UNUSED static void
66calculate_pixel_hashing_table(unsigned n, unsigned m,
67                              unsigned period, unsigned index, bool flip,
68                              uint32_t *p)
69{
70   for (unsigned i = 0; i < n; i++) {
71      for (unsigned j = 0; j < m; j++) {
72         const unsigned k = (i + j) % period;
73         p[j + m * i] = (k == index ? 2 : (k & 1) ^ flip);
74      }
75   }
76}
77
78static void
79genX(emit_slice_hashing_state)(struct anv_device *device,
80                               struct anv_batch *batch)
81{
82#if GFX_VER == 11
83   assert(device->info.ppipe_subslices[2] == 0);
84
85   if (device->info.ppipe_subslices[0] == device->info.ppipe_subslices[1])
86     return;
87
88   if (!device->slice_hash.alloc_size) {
89      unsigned size = GENX(SLICE_HASH_TABLE_length) * 4;
90      device->slice_hash =
91         anv_state_pool_alloc(&device->dynamic_state_pool, size, 64);
92
93      const bool flip = device->info.ppipe_subslices[0] <
94                     device->info.ppipe_subslices[1];
95      struct GENX(SLICE_HASH_TABLE) table;
96      calculate_pixel_hashing_table(16, 16, 3, 3, flip, table.Entry[0]);
97
98      GENX(SLICE_HASH_TABLE_pack)(NULL, device->slice_hash.map, &table);
99   }
100
101   anv_batch_emit(batch, GENX(3DSTATE_SLICE_TABLE_STATE_POINTERS), ptr) {
102      ptr.SliceHashStatePointerValid = true;
103      ptr.SliceHashTableStatePointer = device->slice_hash.offset;
104   }
105
106   anv_batch_emit(batch, GENX(3DSTATE_3D_MODE), mode) {
107      mode.SliceHashingTableEnable = true;
108   }
109#elif GFX_VERx10 == 120
110   /* For each n calculate ppipes_of[n], equal to the number of pixel pipes
111    * present with n active dual subslices.
112    */
113   unsigned ppipes_of[3] = {};
114
115   for (unsigned n = 0; n < ARRAY_SIZE(ppipes_of); n++) {
116      for (unsigned p = 0; p < ARRAY_SIZE(device->info.ppipe_subslices); p++)
117         ppipes_of[n] += (device->info.ppipe_subslices[p] == n);
118   }
119
120   /* Gfx12 has three pixel pipes. */
121   assert(ppipes_of[0] + ppipes_of[1] + ppipes_of[2] == 3);
122
123   if (ppipes_of[2] == 3 || ppipes_of[0] == 2) {
124      /* All three pixel pipes have the maximum number of active dual
125       * subslices, or there is only one active pixel pipe: Nothing to do.
126       */
127      return;
128   }
129
130   anv_batch_emit(batch, GENX(3DSTATE_SUBSLICE_HASH_TABLE), p) {
131      p.SliceHashControl[0] = TABLE_0;
132
133      if (ppipes_of[2] == 2 && ppipes_of[0] == 1)
134         calculate_pixel_hashing_table(8, 16, 2, 2, 0, p.TwoWayTableEntry[0]);
135      else if (ppipes_of[2] == 1 && ppipes_of[1] == 1 && ppipes_of[0] == 1)
136         calculate_pixel_hashing_table(8, 16, 3, 3, 0, p.TwoWayTableEntry[0]);
137
138      if (ppipes_of[2] == 2 && ppipes_of[1] == 1)
139         calculate_pixel_hashing_table(8, 16, 5, 4, 0, p.ThreeWayTableEntry[0]);
140      else if (ppipes_of[2] == 2 && ppipes_of[0] == 1)
141         calculate_pixel_hashing_table(8, 16, 2, 2, 0, p.ThreeWayTableEntry[0]);
142      else if (ppipes_of[2] == 1 && ppipes_of[1] == 1 && ppipes_of[0] == 1)
143         calculate_pixel_hashing_table(8, 16, 3, 3, 0, p.ThreeWayTableEntry[0]);
144      else
145         unreachable("Illegal fusing.");
146   }
147
148   anv_batch_emit(batch, GENX(3DSTATE_3D_MODE), p) {
149      p.SubsliceHashingTableEnable = true;
150      p.SubsliceHashingTableEnableMask = true;
151   }
152#endif
153}
154
155static VkResult
156init_render_queue_state(struct anv_queue *queue)
157{
158   struct anv_device *device = queue->device;
159   uint32_t cmds[64];
160   struct anv_batch batch = {
161      .start = cmds,
162      .next = cmds,
163      .end = (void *) cmds + sizeof(cmds),
164   };
165
166   anv_batch_emit(&batch, GENX(PIPELINE_SELECT), ps) {
167#if GFX_VER >= 9
168      ps.MaskBits = GFX_VER >= 12 ? 0x13 : 3;
169      ps.MediaSamplerDOPClockGateEnable = GFX_VER >= 12;
170#endif
171      ps.PipelineSelection = _3D;
172   }
173
174#if GFX_VER == 9
175   anv_batch_write_reg(&batch, GENX(CACHE_MODE_1), cm1) {
176      cm1.FloatBlendOptimizationEnable = true;
177      cm1.FloatBlendOptimizationEnableMask = true;
178      cm1.MSCRAWHazardAvoidanceBit = true;
179      cm1.MSCRAWHazardAvoidanceBitMask = true;
180      cm1.PartialResolveDisableInVC = true;
181      cm1.PartialResolveDisableInVCMask = true;
182   }
183#endif
184
185   anv_batch_emit(&batch, GENX(3DSTATE_AA_LINE_PARAMETERS), aa);
186
187   anv_batch_emit(&batch, GENX(3DSTATE_DRAWING_RECTANGLE), rect) {
188      rect.ClippedDrawingRectangleYMin = 0;
189      rect.ClippedDrawingRectangleXMin = 0;
190      rect.ClippedDrawingRectangleYMax = UINT16_MAX;
191      rect.ClippedDrawingRectangleXMax = UINT16_MAX;
192      rect.DrawingRectangleOriginY = 0;
193      rect.DrawingRectangleOriginX = 0;
194   }
195
196#if GFX_VER >= 8
197   anv_batch_emit(&batch, GENX(3DSTATE_WM_CHROMAKEY), ck);
198
199   genX(emit_sample_pattern)(&batch, 0, NULL);
200
201   /* The BDW+ docs describe how to use the 3DSTATE_WM_HZ_OP instruction in the
202    * section titled, "Optimized Depth Buffer Clear and/or Stencil Buffer
203    * Clear." It mentions that the packet overrides GPU state for the clear
204    * operation and needs to be reset to 0s to clear the overrides. Depending
205    * on the kernel, we may not get a context with the state for this packet
206    * zeroed. Do it ourselves just in case. We've observed this to prevent a
207    * number of GPU hangs on ICL.
208    */
209   anv_batch_emit(&batch, GENX(3DSTATE_WM_HZ_OP), hzp);
210#endif
211
212#if GFX_VER == 11
213   /* The default behavior of bit 5 "Headerless Message for Pre-emptable
214    * Contexts" in SAMPLER MODE register is set to 0, which means
215    * headerless sampler messages are not allowed for pre-emptable
216    * contexts. Set the bit 5 to 1 to allow them.
217    */
218   anv_batch_write_reg(&batch, GENX(SAMPLER_MODE), sm) {
219      sm.HeaderlessMessageforPreemptableContexts = true;
220      sm.HeaderlessMessageforPreemptableContextsMask = true;
221   }
222
223   /* Bit 1 "Enabled Texel Offset Precision Fix" must be set in
224    * HALF_SLICE_CHICKEN7 register.
225    */
226   anv_batch_write_reg(&batch, GENX(HALF_SLICE_CHICKEN7), hsc7) {
227      hsc7.EnabledTexelOffsetPrecisionFix = true;
228      hsc7.EnabledTexelOffsetPrecisionFixMask = true;
229   }
230
231   anv_batch_write_reg(&batch, GENX(TCCNTLREG), tcc) {
232      tcc.L3DataPartialWriteMergingEnable = true;
233      tcc.ColorZPartialWriteMergingEnable = true;
234      tcc.URBPartialWriteMergingEnable = true;
235      tcc.TCDisable = true;
236   }
237#endif
238   genX(emit_slice_hashing_state)(device, &batch);
239
240#if GFX_VER >= 11
241   /* hardware specification recommends disabling repacking for
242    * the compatibility with decompression mechanism in display controller.
243    */
244   if (device->info.disable_ccs_repack) {
245      anv_batch_write_reg(&batch, GENX(CACHE_MODE_0), cm0) {
246         cm0.DisableRepackingforCompression = true;
247         cm0.DisableRepackingforCompressionMask = true;
248      }
249   }
250
251   /* an unknown issue is causing vs push constants to become
252    * corrupted during object-level preemption. For now, restrict
253    * to command buffer level preemption to avoid rendering
254    * corruption.
255    */
256   anv_batch_write_reg(&batch, GENX(CS_CHICKEN1), cc1) {
257      cc1.ReplayMode = MidcmdbufferPreemption;
258      cc1.ReplayModeMask = true;
259   }
260
261#if GFX_VERx10 < 125
262#define AA_LINE_QUALITY_REG GENX(3D_CHICKEN3)
263#else
264#define AA_LINE_QUALITY_REG GENX(CHICKEN_RASTER_1)
265#endif
266
267   /* Enable the new line drawing algorithm that produces higher quality
268    * lines.
269    */
270   anv_batch_write_reg(&batch, AA_LINE_QUALITY_REG, c3) {
271      c3.AALineQualityFix = true;
272      c3.AALineQualityFixMask = true;
273   }
274#endif
275
276#if GFX_VER == 12
277   if (device->info.has_aux_map) {
278      uint64_t aux_base_addr = intel_aux_map_get_base(device->aux_map_ctx);
279      assert(aux_base_addr % (32 * 1024) == 0);
280      anv_batch_emit(&batch, GENX(MI_LOAD_REGISTER_IMM), lri) {
281         lri.RegisterOffset = GENX(GFX_AUX_TABLE_BASE_ADDR_num);
282         lri.DataDWord = aux_base_addr & 0xffffffff;
283      }
284      anv_batch_emit(&batch, GENX(MI_LOAD_REGISTER_IMM), lri) {
285         lri.RegisterOffset = GENX(GFX_AUX_TABLE_BASE_ADDR_num) + 4;
286         lri.DataDWord = aux_base_addr >> 32;
287      }
288   }
289#endif
290
291   /* Set the "CONSTANT_BUFFER Address Offset Disable" bit, so
292    * 3DSTATE_CONSTANT_XS buffer 0 is an absolute address.
293    *
294    * This is only safe on kernels with context isolation support.
295    */
296   if (GFX_VER >= 8 && device->physical->has_context_isolation) {
297#if GFX_VER >= 9
298      anv_batch_write_reg(&batch, GENX(CS_DEBUG_MODE2), csdm2) {
299         csdm2.CONSTANT_BUFFERAddressOffsetDisable = true;
300         csdm2.CONSTANT_BUFFERAddressOffsetDisableMask = true;
301      }
302#elif GFX_VER == 8
303      anv_batch_write_reg(&batch, GENX(INSTPM), instpm) {
304         instpm.CONSTANT_BUFFERAddressOffsetDisable = true;
305         instpm.CONSTANT_BUFFERAddressOffsetDisableMask = true;
306      }
307#endif
308   }
309
310#if GFX_VER >= 11
311   /* Starting with GFX version 11, SLM is no longer part of the L3$ config
312    * so it never changes throughout the lifetime of the VkDevice.
313    */
314   const struct intel_l3_config *cfg = intel_get_default_l3_config(&device->info);
315   genX(emit_l3_config)(&batch, device, cfg);
316   device->l3_config = cfg;
317#endif
318
319   anv_batch_emit(&batch, GENX(MI_BATCH_BUFFER_END), bbe);
320
321   assert(batch.next <= batch.end);
322
323   return anv_queue_submit_simple_batch(queue, &batch);
324}
325
326void
327genX(init_physical_device_state)(ASSERTED struct anv_physical_device *device)
328{
329   assert(device->info.verx10 == GFX_VERx10);
330}
331
332VkResult
333genX(init_device_state)(struct anv_device *device)
334{
335   VkResult res;
336
337   device->slice_hash = (struct anv_state) { 0 };
338   for (uint32_t i = 0; i < device->queue_count; i++) {
339      struct anv_queue *queue = &device->queues[i];
340      switch (queue->family->engine_class) {
341      case I915_ENGINE_CLASS_RENDER:
342         res = init_render_queue_state(queue);
343         break;
344      default:
345         res = vk_error(device, VK_ERROR_INITIALIZATION_FAILED);
346         break;
347      }
348      if (res != VK_SUCCESS)
349         return res;
350   }
351
352   return res;
353}
354
355void
356genX(emit_l3_config)(struct anv_batch *batch,
357                     const struct anv_device *device,
358                     const struct intel_l3_config *cfg)
359{
360   UNUSED const struct intel_device_info *devinfo = &device->info;
361
362#if GFX_VER >= 8
363
364#if GFX_VER >= 12
365#define L3_ALLOCATION_REG GENX(L3ALLOC)
366#define L3_ALLOCATION_REG_num GENX(L3ALLOC_num)
367#else
368#define L3_ALLOCATION_REG GENX(L3CNTLREG)
369#define L3_ALLOCATION_REG_num GENX(L3CNTLREG_num)
370#endif
371
372   anv_batch_write_reg(batch, L3_ALLOCATION_REG, l3cr) {
373      if (cfg == NULL) {
374#if GFX_VER >= 12
375         l3cr.L3FullWayAllocationEnable = true;
376#else
377         unreachable("Invalid L3$ config");
378#endif
379      } else {
380#if GFX_VER < 11
381         l3cr.SLMEnable = cfg->n[INTEL_L3P_SLM];
382#endif
383#if GFX_VER == 11
384         /* Wa_1406697149: Bit 9 "Error Detection Behavior Control" must be
385          * set in L3CNTLREG register. The default setting of the bit is not
386          * the desirable behavior.
387          */
388         l3cr.ErrorDetectionBehaviorControl = true;
389         l3cr.UseFullWays = true;
390#endif /* GFX_VER == 11 */
391         assert(cfg->n[INTEL_L3P_IS] == 0);
392         assert(cfg->n[INTEL_L3P_C] == 0);
393         assert(cfg->n[INTEL_L3P_T] == 0);
394         l3cr.URBAllocation = cfg->n[INTEL_L3P_URB];
395         l3cr.ROAllocation = cfg->n[INTEL_L3P_RO];
396         l3cr.DCAllocation = cfg->n[INTEL_L3P_DC];
397         l3cr.AllAllocation = cfg->n[INTEL_L3P_ALL];
398      }
399   }
400
401#else /* GFX_VER < 8 */
402
403   const bool has_dc = cfg->n[INTEL_L3P_DC] || cfg->n[INTEL_L3P_ALL];
404   const bool has_is = cfg->n[INTEL_L3P_IS] || cfg->n[INTEL_L3P_RO] ||
405                       cfg->n[INTEL_L3P_ALL];
406   const bool has_c = cfg->n[INTEL_L3P_C] || cfg->n[INTEL_L3P_RO] ||
407                      cfg->n[INTEL_L3P_ALL];
408   const bool has_t = cfg->n[INTEL_L3P_T] || cfg->n[INTEL_L3P_RO] ||
409                      cfg->n[INTEL_L3P_ALL];
410
411   assert(!cfg->n[INTEL_L3P_ALL]);
412
413   /* When enabled SLM only uses a portion of the L3 on half of the banks,
414    * the matching space on the remaining banks has to be allocated to a
415    * client (URB for all validated configurations) set to the
416    * lower-bandwidth 2-bank address hashing mode.
417    */
418   const bool urb_low_bw = cfg->n[INTEL_L3P_SLM] && !devinfo->is_baytrail;
419   assert(!urb_low_bw || cfg->n[INTEL_L3P_URB] == cfg->n[INTEL_L3P_SLM]);
420
421   /* Minimum number of ways that can be allocated to the URB. */
422   const unsigned n0_urb = devinfo->is_baytrail ? 32 : 0;
423   assert(cfg->n[INTEL_L3P_URB] >= n0_urb);
424
425   anv_batch_write_reg(batch, GENX(L3SQCREG1), l3sqc) {
426      l3sqc.ConvertDC_UC = !has_dc;
427      l3sqc.ConvertIS_UC = !has_is;
428      l3sqc.ConvertC_UC = !has_c;
429      l3sqc.ConvertT_UC = !has_t;
430#if GFX_VERx10 == 75
431      l3sqc.L3SQGeneralPriorityCreditInitialization = SQGPCI_DEFAULT;
432#else
433      l3sqc.L3SQGeneralPriorityCreditInitialization =
434         devinfo->is_baytrail ? BYT_SQGPCI_DEFAULT : SQGPCI_DEFAULT;
435#endif
436      l3sqc.L3SQHighPriorityCreditInitialization = SQHPCI_DEFAULT;
437   }
438
439   anv_batch_write_reg(batch, GENX(L3CNTLREG2), l3cr2) {
440      l3cr2.SLMEnable = cfg->n[INTEL_L3P_SLM];
441      l3cr2.URBLowBandwidth = urb_low_bw;
442      l3cr2.URBAllocation = cfg->n[INTEL_L3P_URB] - n0_urb;
443#if !GFX_VERx10 == 75
444      l3cr2.ALLAllocation = cfg->n[INTEL_L3P_ALL];
445#endif
446      l3cr2.ROAllocation = cfg->n[INTEL_L3P_RO];
447      l3cr2.DCAllocation = cfg->n[INTEL_L3P_DC];
448   }
449
450   anv_batch_write_reg(batch, GENX(L3CNTLREG3), l3cr3) {
451      l3cr3.ISAllocation = cfg->n[INTEL_L3P_IS];
452      l3cr3.ISLowBandwidth = 0;
453      l3cr3.CAllocation = cfg->n[INTEL_L3P_C];
454      l3cr3.CLowBandwidth = 0;
455      l3cr3.TAllocation = cfg->n[INTEL_L3P_T];
456      l3cr3.TLowBandwidth = 0;
457   }
458
459#if GFX_VERx10 == 75
460   if (device->physical->cmd_parser_version >= 4) {
461      /* Enable L3 atomics on HSW if we have a DC partition, otherwise keep
462       * them disabled to avoid crashing the system hard.
463       */
464      anv_batch_write_reg(batch, GENX(SCRATCH1), s1) {
465         s1.L3AtomicDisable = !has_dc;
466      }
467      anv_batch_write_reg(batch, GENX(CHICKEN3), c3) {
468         c3.L3AtomicDisableMask = true;
469         c3.L3AtomicDisable = !has_dc;
470      }
471   }
472#endif /* GFX_VERx10 == 75 */
473
474#endif /* GFX_VER < 8 */
475}
476
477void
478genX(emit_multisample)(struct anv_batch *batch, uint32_t samples,
479                       const VkSampleLocationEXT *locations)
480{
481   anv_batch_emit(batch, GENX(3DSTATE_MULTISAMPLE), ms) {
482      ms.NumberofMultisamples       = __builtin_ffs(samples) - 1;
483
484      ms.PixelLocation              = CENTER;
485#if GFX_VER >= 8
486      /* The PRM says that this bit is valid only for DX9:
487       *
488       *    SW can choose to set this bit only for DX9 API. DX10/OGL API's
489       *    should not have any effect by setting or not setting this bit.
490       */
491      ms.PixelPositionOffsetEnable  = false;
492#else
493
494      if (locations) {
495         switch (samples) {
496         case 1:
497            INTEL_SAMPLE_POS_1X_ARRAY(ms.Sample, locations);
498            break;
499         case 2:
500            INTEL_SAMPLE_POS_2X_ARRAY(ms.Sample, locations);
501            break;
502         case 4:
503            INTEL_SAMPLE_POS_4X_ARRAY(ms.Sample, locations);
504            break;
505         case 8:
506            INTEL_SAMPLE_POS_8X_ARRAY(ms.Sample, locations);
507            break;
508         default:
509            break;
510         }
511      } else {
512         switch (samples) {
513         case 1:
514            INTEL_SAMPLE_POS_1X(ms.Sample);
515            break;
516         case 2:
517            INTEL_SAMPLE_POS_2X(ms.Sample);
518            break;
519         case 4:
520            INTEL_SAMPLE_POS_4X(ms.Sample);
521            break;
522         case 8:
523            INTEL_SAMPLE_POS_8X(ms.Sample);
524            break;
525         default:
526            break;
527         }
528      }
529#endif
530   }
531}
532
533#if GFX_VER >= 8
534void
535genX(emit_sample_pattern)(struct anv_batch *batch, uint32_t samples,
536                          const VkSampleLocationEXT *locations)
537{
538   /* See the Vulkan 1.0 spec Table 24.1 "Standard sample locations" and
539    * VkPhysicalDeviceFeatures::standardSampleLocations.
540    */
541   anv_batch_emit(batch, GENX(3DSTATE_SAMPLE_PATTERN), sp) {
542      if (locations) {
543         /* The Skylake PRM Vol. 2a "3DSTATE_SAMPLE_PATTERN" says:
544          *
545          *    "When programming the sample offsets (for NUMSAMPLES_4 or _8
546          *    and MSRASTMODE_xxx_PATTERN), the order of the samples 0 to 3
547          *    (or 7 for 8X, or 15 for 16X) must have monotonically increasing
548          *    distance from the pixel center. This is required to get the
549          *    correct centroid computation in the device."
550          *
551          * However, the Vulkan spec seems to require that the the samples
552          * occur in the order provided through the API. The standard sample
553          * patterns have the above property that they have monotonically
554          * increasing distances from the center but client-provided ones do
555          * not. As long as this only affects centroid calculations as the
556          * docs say, we should be ok because OpenGL and Vulkan only require
557          * that the centroid be some lit sample and that it's the same for
558          * all samples in a pixel; they have no requirement that it be the
559          * one closest to center.
560          */
561         switch (samples) {
562         case 1:
563            INTEL_SAMPLE_POS_1X_ARRAY(sp._1xSample, locations);
564            break;
565         case 2:
566            INTEL_SAMPLE_POS_2X_ARRAY(sp._2xSample, locations);
567            break;
568         case 4:
569            INTEL_SAMPLE_POS_4X_ARRAY(sp._4xSample, locations);
570            break;
571         case 8:
572            INTEL_SAMPLE_POS_8X_ARRAY(sp._8xSample, locations);
573            break;
574#if GFX_VER >= 9
575         case 16:
576            INTEL_SAMPLE_POS_16X_ARRAY(sp._16xSample, locations);
577            break;
578#endif
579         default:
580            break;
581         }
582      } else {
583         INTEL_SAMPLE_POS_1X(sp._1xSample);
584         INTEL_SAMPLE_POS_2X(sp._2xSample);
585         INTEL_SAMPLE_POS_4X(sp._4xSample);
586         INTEL_SAMPLE_POS_8X(sp._8xSample);
587#if GFX_VER >= 9
588         INTEL_SAMPLE_POS_16X(sp._16xSample);
589#endif
590      }
591   }
592}
593#endif
594
595#if GFX_VER >= 11
596void
597genX(emit_shading_rate)(struct anv_batch *batch,
598                        const struct anv_graphics_pipeline *pipeline,
599                        struct anv_state cps_states,
600                        struct anv_dynamic_state *dynamic_state)
601{
602   const struct brw_wm_prog_data *wm_prog_data = get_wm_prog_data(pipeline);
603   const bool cps_enable = wm_prog_data && wm_prog_data->per_coarse_pixel_dispatch;
604
605#if GFX_VER == 11
606   anv_batch_emit(batch, GENX(3DSTATE_CPS), cps) {
607      cps.CoarsePixelShadingMode = cps_enable ? CPS_MODE_CONSTANT : CPS_MODE_NONE;
608      if (cps_enable) {
609         cps.MinCPSizeX = dynamic_state->fragment_shading_rate.width;
610         cps.MinCPSizeY = dynamic_state->fragment_shading_rate.height;
611      }
612   }
613#elif GFX_VER == 12
614   for (uint32_t i = 0; i < dynamic_state->viewport.count; i++) {
615      uint32_t *cps_state_dwords =
616         cps_states.map + GENX(CPS_STATE_length) * 4 * i;
617      struct GENX(CPS_STATE) cps_state = {
618         .CoarsePixelShadingMode = cps_enable ? CPS_MODE_CONSTANT : CPS_MODE_NONE,
619      };
620
621      if (cps_enable) {
622         cps_state.MinCPSizeX = dynamic_state->fragment_shading_rate.width;
623         cps_state.MinCPSizeY = dynamic_state->fragment_shading_rate.height;
624      }
625
626      GENX(CPS_STATE_pack)(NULL, cps_state_dwords, &cps_state);
627   }
628
629   anv_batch_emit(batch, GENX(3DSTATE_CPS_POINTERS), cps) {
630      cps.CoarsePixelShadingStateArrayPointer = cps_states.offset;
631   }
632#endif
633}
634#endif /* GFX_VER >= 11 */
635
636static uint32_t
637vk_to_intel_tex_filter(VkFilter filter, bool anisotropyEnable)
638{
639   switch (filter) {
640   default:
641      assert(!"Invalid filter");
642   case VK_FILTER_NEAREST:
643      return anisotropyEnable ? MAPFILTER_ANISOTROPIC : MAPFILTER_NEAREST;
644   case VK_FILTER_LINEAR:
645      return anisotropyEnable ? MAPFILTER_ANISOTROPIC : MAPFILTER_LINEAR;
646   }
647}
648
649static uint32_t
650vk_to_intel_max_anisotropy(float ratio)
651{
652   return (anv_clamp_f(ratio, 2, 16) - 2) / 2;
653}
654
655static const uint32_t vk_to_intel_mipmap_mode[] = {
656   [VK_SAMPLER_MIPMAP_MODE_NEAREST]          = MIPFILTER_NEAREST,
657   [VK_SAMPLER_MIPMAP_MODE_LINEAR]           = MIPFILTER_LINEAR
658};
659
660static const uint32_t vk_to_intel_tex_address[] = {
661   [VK_SAMPLER_ADDRESS_MODE_REPEAT]          = TCM_WRAP,
662   [VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT] = TCM_MIRROR,
663   [VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE]   = TCM_CLAMP,
664   [VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE] = TCM_MIRROR_ONCE,
665   [VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER] = TCM_CLAMP_BORDER,
666};
667
668/* Vulkan specifies the result of shadow comparisons as:
669 *     1     if   ref <op> texel,
670 *     0     otherwise.
671 *
672 * The hardware does:
673 *     0     if texel <op> ref,
674 *     1     otherwise.
675 *
676 * So, these look a bit strange because there's both a negation
677 * and swapping of the arguments involved.
678 */
679static const uint32_t vk_to_intel_shadow_compare_op[] = {
680   [VK_COMPARE_OP_NEVER]                        = PREFILTEROP_ALWAYS,
681   [VK_COMPARE_OP_LESS]                         = PREFILTEROP_LEQUAL,
682   [VK_COMPARE_OP_EQUAL]                        = PREFILTEROP_NOTEQUAL,
683   [VK_COMPARE_OP_LESS_OR_EQUAL]                = PREFILTEROP_LESS,
684   [VK_COMPARE_OP_GREATER]                      = PREFILTEROP_GEQUAL,
685   [VK_COMPARE_OP_NOT_EQUAL]                    = PREFILTEROP_EQUAL,
686   [VK_COMPARE_OP_GREATER_OR_EQUAL]             = PREFILTEROP_GREATER,
687   [VK_COMPARE_OP_ALWAYS]                       = PREFILTEROP_NEVER,
688};
689
690#if GFX_VER >= 9
691static const uint32_t vk_to_intel_sampler_reduction_mode[] = {
692   [VK_SAMPLER_REDUCTION_MODE_WEIGHTED_AVERAGE_EXT] = STD_FILTER,
693   [VK_SAMPLER_REDUCTION_MODE_MIN_EXT]              = MINIMUM,
694   [VK_SAMPLER_REDUCTION_MODE_MAX_EXT]              = MAXIMUM,
695};
696#endif
697
698VkResult genX(CreateSampler)(
699    VkDevice                                    _device,
700    const VkSamplerCreateInfo*                  pCreateInfo,
701    const VkAllocationCallbacks*                pAllocator,
702    VkSampler*                                  pSampler)
703{
704   ANV_FROM_HANDLE(anv_device, device, _device);
705   struct anv_sampler *sampler;
706
707   assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO);
708
709   sampler = vk_object_zalloc(&device->vk, pAllocator, sizeof(*sampler),
710                              VK_OBJECT_TYPE_SAMPLER);
711   if (!sampler)
712      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
713
714   sampler->n_planes = 1;
715
716   uint32_t border_color_stride = GFX_VERx10 == 75 ? 512 : 64;
717   uint32_t border_color_offset;
718   ASSERTED bool has_custom_color = false;
719   if (pCreateInfo->borderColor <= VK_BORDER_COLOR_INT_OPAQUE_WHITE) {
720      border_color_offset = device->border_colors.offset +
721                            pCreateInfo->borderColor *
722                            border_color_stride;
723   } else {
724      assert(GFX_VER >= 8);
725      sampler->custom_border_color =
726         anv_state_reserved_pool_alloc(&device->custom_border_colors);
727      border_color_offset = sampler->custom_border_color.offset;
728   }
729
730#if GFX_VER >= 9
731   unsigned sampler_reduction_mode = STD_FILTER;
732   bool enable_sampler_reduction = false;
733#endif
734
735   vk_foreach_struct(ext, pCreateInfo->pNext) {
736      switch (ext->sType) {
737      case VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_INFO: {
738         VkSamplerYcbcrConversionInfo *pSamplerConversion =
739            (VkSamplerYcbcrConversionInfo *) ext;
740         ANV_FROM_HANDLE(anv_ycbcr_conversion, conversion,
741                         pSamplerConversion->conversion);
742
743         /* Ignore conversion for non-YUV formats. This fulfills a requirement
744          * for clients that want to utilize same code path for images with
745          * external formats (VK_FORMAT_UNDEFINED) and "regular" RGBA images
746          * where format is known.
747          */
748         if (conversion == NULL || !conversion->format->can_ycbcr)
749            break;
750
751         sampler->n_planes = conversion->format->n_planes;
752         sampler->conversion = conversion;
753         break;
754      }
755#if GFX_VER >= 9
756      case VK_STRUCTURE_TYPE_SAMPLER_REDUCTION_MODE_CREATE_INFO: {
757         VkSamplerReductionModeCreateInfo *sampler_reduction =
758            (VkSamplerReductionModeCreateInfo *) ext;
759         sampler_reduction_mode =
760            vk_to_intel_sampler_reduction_mode[sampler_reduction->reductionMode];
761         enable_sampler_reduction = true;
762         break;
763      }
764#endif
765      case VK_STRUCTURE_TYPE_SAMPLER_CUSTOM_BORDER_COLOR_CREATE_INFO_EXT: {
766         VkSamplerCustomBorderColorCreateInfoEXT *custom_border_color =
767            (VkSamplerCustomBorderColorCreateInfoEXT *) ext;
768         if (sampler->custom_border_color.map == NULL)
769            break;
770         struct gfx8_border_color *cbc = sampler->custom_border_color.map;
771         if (custom_border_color->format == VK_FORMAT_B4G4R4A4_UNORM_PACK16) {
772            /* B4G4R4A4_UNORM_PACK16 is treated as R4G4B4A4_UNORM_PACK16 with
773             * a swizzle, but this does not carry over to the sampler for
774             * border colors, so we need to do the swizzle ourselves here.
775             */
776            cbc->uint32[0] = custom_border_color->customBorderColor.uint32[2];
777            cbc->uint32[1] = custom_border_color->customBorderColor.uint32[1];
778            cbc->uint32[2] = custom_border_color->customBorderColor.uint32[0];
779            cbc->uint32[3] = custom_border_color->customBorderColor.uint32[3];
780         } else {
781            /* Both structs share the same layout, so just copy them over. */
782            memcpy(cbc, &custom_border_color->customBorderColor,
783                   sizeof(VkClearColorValue));
784         }
785         has_custom_color = true;
786         break;
787      }
788      default:
789         anv_debug_ignored_stype(ext->sType);
790         break;
791      }
792   }
793
794   assert((sampler->custom_border_color.map == NULL) || has_custom_color);
795
796   if (device->physical->has_bindless_samplers) {
797      /* If we have bindless, allocate enough samplers.  We allocate 32 bytes
798       * for each sampler instead of 16 bytes because we want all bindless
799       * samplers to be 32-byte aligned so we don't have to use indirect
800       * sampler messages on them.
801       */
802      sampler->bindless_state =
803         anv_state_pool_alloc(&device->dynamic_state_pool,
804                              sampler->n_planes * 32, 32);
805   }
806
807   for (unsigned p = 0; p < sampler->n_planes; p++) {
808      const bool plane_has_chroma =
809         sampler->conversion && sampler->conversion->format->planes[p].has_chroma;
810      const VkFilter min_filter =
811         plane_has_chroma ? sampler->conversion->chroma_filter : pCreateInfo->minFilter;
812      const VkFilter mag_filter =
813         plane_has_chroma ? sampler->conversion->chroma_filter : pCreateInfo->magFilter;
814      const bool enable_min_filter_addr_rounding = min_filter != VK_FILTER_NEAREST;
815      const bool enable_mag_filter_addr_rounding = mag_filter != VK_FILTER_NEAREST;
816      /* From Broadwell PRM, SAMPLER_STATE:
817       *   "Mip Mode Filter must be set to MIPFILTER_NONE for Planar YUV surfaces."
818       */
819      const bool isl_format_is_planar_yuv = sampler->conversion &&
820         isl_format_is_yuv(sampler->conversion->format->planes[0].isl_format) &&
821         isl_format_is_planar(sampler->conversion->format->planes[0].isl_format);
822
823      const uint32_t mip_filter_mode =
824         isl_format_is_planar_yuv ?
825         MIPFILTER_NONE : vk_to_intel_mipmap_mode[pCreateInfo->mipmapMode];
826
827      struct GENX(SAMPLER_STATE) sampler_state = {
828         .SamplerDisable = false,
829         .TextureBorderColorMode = DX10OGL,
830
831#if GFX_VER >= 11
832         .CPSLODCompensationEnable = true,
833#endif
834
835#if GFX_VER >= 8
836         .LODPreClampMode = CLAMP_MODE_OGL,
837#else
838         .LODPreClampEnable = CLAMP_ENABLE_OGL,
839#endif
840
841#if GFX_VER == 8
842         .BaseMipLevel = 0.0,
843#endif
844         .MipModeFilter = mip_filter_mode,
845         .MagModeFilter = vk_to_intel_tex_filter(mag_filter, pCreateInfo->anisotropyEnable),
846         .MinModeFilter = vk_to_intel_tex_filter(min_filter, pCreateInfo->anisotropyEnable),
847         .TextureLODBias = anv_clamp_f(pCreateInfo->mipLodBias, -16, 15.996),
848         .AnisotropicAlgorithm =
849            pCreateInfo->anisotropyEnable ? EWAApproximation : LEGACY,
850         .MinLOD = anv_clamp_f(pCreateInfo->minLod, 0, 14),
851         .MaxLOD = anv_clamp_f(pCreateInfo->maxLod, 0, 14),
852         .ChromaKeyEnable = 0,
853         .ChromaKeyIndex = 0,
854         .ChromaKeyMode = 0,
855         .ShadowFunction =
856            vk_to_intel_shadow_compare_op[pCreateInfo->compareEnable ?
857                                        pCreateInfo->compareOp : VK_COMPARE_OP_NEVER],
858         .CubeSurfaceControlMode = OVERRIDE,
859
860         .BorderColorPointer = border_color_offset,
861
862#if GFX_VER >= 8
863         .LODClampMagnificationMode = MIPNONE,
864#endif
865
866         .MaximumAnisotropy = vk_to_intel_max_anisotropy(pCreateInfo->maxAnisotropy),
867         .RAddressMinFilterRoundingEnable = enable_min_filter_addr_rounding,
868         .RAddressMagFilterRoundingEnable = enable_mag_filter_addr_rounding,
869         .VAddressMinFilterRoundingEnable = enable_min_filter_addr_rounding,
870         .VAddressMagFilterRoundingEnable = enable_mag_filter_addr_rounding,
871         .UAddressMinFilterRoundingEnable = enable_min_filter_addr_rounding,
872         .UAddressMagFilterRoundingEnable = enable_mag_filter_addr_rounding,
873         .TrilinearFilterQuality = 0,
874         .NonnormalizedCoordinateEnable = pCreateInfo->unnormalizedCoordinates,
875         .TCXAddressControlMode = vk_to_intel_tex_address[pCreateInfo->addressModeU],
876         .TCYAddressControlMode = vk_to_intel_tex_address[pCreateInfo->addressModeV],
877         .TCZAddressControlMode = vk_to_intel_tex_address[pCreateInfo->addressModeW],
878
879#if GFX_VER >= 9
880         .ReductionType = sampler_reduction_mode,
881         .ReductionTypeEnable = enable_sampler_reduction,
882#endif
883      };
884
885      GENX(SAMPLER_STATE_pack)(NULL, sampler->state[p], &sampler_state);
886
887      if (sampler->bindless_state.map) {
888         memcpy(sampler->bindless_state.map + p * 32,
889                sampler->state[p], GENX(SAMPLER_STATE_length) * 4);
890      }
891   }
892
893   *pSampler = anv_sampler_to_handle(sampler);
894
895   return VK_SUCCESS;
896}
897