1/*
2 * Copyright © 2011 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 "main/macros.h"
25#include "brw_batch.h"
26#include "brw_context.h"
27#include "brw_state.h"
28#include "brw_defines.h"
29
30#include "common/intel_l3_config.h"
31
32/**
33 * The following diagram shows how we partition the URB:
34 *
35 *        16kB or 32kB               Rest of the URB space
36 *   __________-__________   _________________-_________________
37 *  /                     \ /                                   \
38 * +-------------------------------------------------------------+
39 * |  VS/HS/DS/GS/FS Push  |           VS/HS/DS/GS URB           |
40 * |       Constants       |               Entries               |
41 * +-------------------------------------------------------------+
42 *
43 * Notably, push constants must be stored at the beginning of the URB
44 * space, while entries can be stored anywhere.  Ivybridge and Haswell
45 * GT1/GT2 have a maximum constant buffer size of 16kB, while Haswell GT3
46 * doubles this (32kB).
47 *
48 * Ivybridge and Haswell GT1/GT2 allow push constants to be located (and
49 * sized) in increments of 1kB.  Haswell GT3 requires them to be located and
50 * sized in increments of 2kB.
51 *
52 * Currently we split the constant buffer space evenly among whatever stages
53 * are active.  This is probably not ideal, but simple.
54 *
55 * Ivybridge GT1 and Haswell GT1 have 128kB of URB space.
56 * Ivybridge GT2 and Haswell GT2 have 256kB of URB space.
57 * Haswell GT3 has 512kB of URB space.
58 *
59 * See "Volume 2a: 3D Pipeline," section 1.8, "Volume 1b: Configurations",
60 * and the documentation for 3DSTATE_PUSH_CONSTANT_ALLOC_xS.
61 */
62static void
63gfx7_allocate_push_constants(struct brw_context *brw)
64{
65   const struct intel_device_info *devinfo = &brw->screen->devinfo;
66
67   /* BRW_NEW_GEOMETRY_PROGRAM */
68   bool gs_present = brw->programs[MESA_SHADER_GEOMETRY];
69
70   /* BRW_NEW_TESS_PROGRAMS */
71   bool tess_present = brw->programs[MESA_SHADER_TESS_EVAL];
72
73   unsigned avail_size = 16;
74   unsigned multiplier = devinfo->max_constant_urb_size_kb / 16;
75
76   int stages = 2 + gs_present + 2 * tess_present;
77
78   /* Divide up the available space equally between stages.  Because we
79    * round down (using floor division), there may be some left over
80    * space.  We allocate that to the pixel shader stage.
81    */
82   unsigned size_per_stage = avail_size / stages;
83
84   unsigned vs_size = size_per_stage;
85   unsigned hs_size = tess_present ? size_per_stage : 0;
86   unsigned ds_size = tess_present ? size_per_stage : 0;
87   unsigned gs_size = gs_present ? size_per_stage : 0;
88   unsigned fs_size = avail_size - size_per_stage * (stages - 1);
89
90   gfx7_emit_push_constant_state(brw, multiplier * vs_size,
91                                 multiplier * hs_size, multiplier * ds_size,
92                                 multiplier * gs_size, multiplier * fs_size);
93
94   /* From p115 of the Ivy Bridge PRM (3.2.1.4 3DSTATE_PUSH_CONSTANT_ALLOC_VS):
95    *
96    *     Programming Restriction:
97    *
98    *     The 3DSTATE_CONSTANT_VS must be reprogrammed prior to the next
99    *     3DPRIMITIVE command after programming the
100    *     3DSTATE_PUSH_CONSTANT_ALLOC_VS.
101    *
102    * Similar text exists for the other 3DSTATE_PUSH_CONSTANT_ALLOC_*
103    * commands.
104    */
105   brw->vs.base.push_constants_dirty = true;
106   brw->tcs.base.push_constants_dirty = true;
107   brw->tes.base.push_constants_dirty = true;
108   brw->gs.base.push_constants_dirty = true;
109   brw->wm.base.push_constants_dirty = true;
110}
111
112void
113gfx7_emit_push_constant_state(struct brw_context *brw, unsigned vs_size,
114                              unsigned hs_size, unsigned ds_size,
115                              unsigned gs_size, unsigned fs_size)
116{
117   const struct intel_device_info *devinfo = &brw->screen->devinfo;
118   unsigned offset = 0;
119
120   /* From the SKL PRM, Workarounds section (#878):
121    *
122    *    Push constant buffer corruption possible. WA: Insert 2 zero-length
123    *    PushConst_PS before every intended PushConst_PS update, issue a
124    *    NULLPRIM after each of the zero len PC update to make sure CS commits
125    *    them.
126    *
127    * This workaround is attempting to solve a pixel shader push constant
128    * synchronization issue.
129    *
130    * There's an unpublished WA that involves re-emitting
131    * 3DSTATE_PUSH_CONSTANT_ALLOC_PS for every 500-ish 3DSTATE_CONSTANT_PS
132    * packets. Since our counting methods may not be reliable due to
133    * context-switching and pre-emption, we instead choose to approximate this
134    * behavior by re-emitting the packet at the top of the batch.
135    */
136   if (brw->ctx.NewDriverState == BRW_NEW_BATCH) {
137       /* SKL GT2 and GLK 2x6 have reliably demonstrated this issue thus far.
138        * We've also seen some intermittent failures from SKL GT4 and BXT in
139        * the past.
140        */
141      if (!devinfo->is_skylake &&
142          !devinfo->is_broxton &&
143          !devinfo->is_geminilake)
144         return;
145   }
146
147   BEGIN_BATCH(10);
148   OUT_BATCH(_3DSTATE_PUSH_CONSTANT_ALLOC_VS << 16 | (2 - 2));
149   OUT_BATCH(vs_size | offset << GFX7_PUSH_CONSTANT_BUFFER_OFFSET_SHIFT);
150   offset += vs_size;
151
152   OUT_BATCH(_3DSTATE_PUSH_CONSTANT_ALLOC_HS << 16 | (2 - 2));
153   OUT_BATCH(hs_size | offset << GFX7_PUSH_CONSTANT_BUFFER_OFFSET_SHIFT);
154   offset += hs_size;
155
156   OUT_BATCH(_3DSTATE_PUSH_CONSTANT_ALLOC_DS << 16 | (2 - 2));
157   OUT_BATCH(ds_size | offset << GFX7_PUSH_CONSTANT_BUFFER_OFFSET_SHIFT);
158   offset += ds_size;
159
160   OUT_BATCH(_3DSTATE_PUSH_CONSTANT_ALLOC_GS << 16 | (2 - 2));
161   OUT_BATCH(gs_size | offset << GFX7_PUSH_CONSTANT_BUFFER_OFFSET_SHIFT);
162   offset += gs_size;
163
164   OUT_BATCH(_3DSTATE_PUSH_CONSTANT_ALLOC_PS << 16 | (2 - 2));
165   OUT_BATCH(fs_size | offset << GFX7_PUSH_CONSTANT_BUFFER_OFFSET_SHIFT);
166   ADVANCE_BATCH();
167
168   /* From p292 of the Ivy Bridge PRM (11.2.4 3DSTATE_PUSH_CONSTANT_ALLOC_PS):
169    *
170    *     A PIPE_CONTROL command with the CS Stall bit set must be programmed
171    *     in the ring after this instruction.
172    *
173    * No such restriction exists for Haswell or Baytrail.
174    */
175   if (devinfo->verx10 <= 70 && !devinfo->is_baytrail)
176      gfx7_emit_cs_stall_flush(brw);
177}
178
179const struct brw_tracked_state gfx7_push_constant_space = {
180   .dirty = {
181      .mesa = 0,
182      .brw = BRW_NEW_CONTEXT |
183             BRW_NEW_BATCH | /* Push constant workaround */
184             BRW_NEW_GEOMETRY_PROGRAM |
185             BRW_NEW_TESS_PROGRAMS,
186   },
187   .emit = gfx7_allocate_push_constants,
188};
189
190static void
191upload_urb(struct brw_context *brw)
192{
193   /* BRW_NEW_VS_PROG_DATA */
194   const struct brw_vue_prog_data *vs_vue_prog_data =
195      brw_vue_prog_data(brw->vs.base.prog_data);
196   const unsigned vs_size = MAX2(vs_vue_prog_data->urb_entry_size, 1);
197   /* BRW_NEW_GS_PROG_DATA */
198   const bool gs_present = brw->gs.base.prog_data;
199   /* BRW_NEW_TES_PROG_DATA */
200   const bool tess_present = brw->tes.base.prog_data;
201
202   gfx7_upload_urb(brw, vs_size, gs_present, tess_present);
203}
204
205void
206gfx7_upload_urb(struct brw_context *brw, unsigned vs_size,
207                bool gs_present, bool tess_present)
208{
209   const struct intel_device_info *devinfo = &brw->screen->devinfo;
210
211   /* BRW_NEW_{VS,TCS,TES,GS}_PROG_DATA */
212   struct brw_vue_prog_data *prog_data[4] = {
213      [MESA_SHADER_VERTEX] =
214         brw_vue_prog_data(brw->vs.base.prog_data),
215      [MESA_SHADER_TESS_CTRL] =
216         tess_present ? brw_vue_prog_data(brw->tcs.base.prog_data) : NULL,
217      [MESA_SHADER_TESS_EVAL] =
218         tess_present ? brw_vue_prog_data(brw->tes.base.prog_data) : NULL,
219      [MESA_SHADER_GEOMETRY] =
220         gs_present ? brw_vue_prog_data(brw->gs.base.prog_data) : NULL,
221   };
222
223   unsigned entry_size[4];
224   entry_size[MESA_SHADER_VERTEX] = vs_size;
225   for (int i = MESA_SHADER_TESS_CTRL; i <= MESA_SHADER_GEOMETRY; i++) {
226      entry_size[i] = prog_data[i] ? prog_data[i]->urb_entry_size : 1;
227   }
228
229   /* If we're just switching between programs with the same URB requirements,
230    * skip the rest of the logic.
231    */
232   if (brw->urb.vsize == entry_size[MESA_SHADER_VERTEX] &&
233       brw->urb.gs_present == gs_present &&
234       brw->urb.gsize == entry_size[MESA_SHADER_GEOMETRY] &&
235       brw->urb.tess_present == tess_present &&
236       brw->urb.hsize == entry_size[MESA_SHADER_TESS_CTRL] &&
237       brw->urb.dsize == entry_size[MESA_SHADER_TESS_EVAL]) {
238      return;
239   }
240   brw->urb.vsize = entry_size[MESA_SHADER_VERTEX];
241   brw->urb.gs_present = gs_present;
242   brw->urb.gsize = entry_size[MESA_SHADER_GEOMETRY];
243   brw->urb.tess_present = tess_present;
244   brw->urb.hsize = entry_size[MESA_SHADER_TESS_CTRL];
245   brw->urb.dsize = entry_size[MESA_SHADER_TESS_EVAL];
246
247   unsigned entries[4];
248   unsigned start[4];
249   bool constrained;
250   intel_get_urb_config(devinfo, brw->l3.config,
251                        tess_present, gs_present, entry_size,
252                        entries, start, NULL, &constrained);
253
254   if (devinfo->verx10 == 70 && !devinfo->is_baytrail)
255      gfx7_emit_vs_workaround_flush(brw);
256
257   BEGIN_BATCH(8);
258   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
259      assert(devinfo->ver != 10 || entry_size[i] % 3);
260      OUT_BATCH((_3DSTATE_URB_VS + i) << 16 | (2 - 2));
261      OUT_BATCH(entries[i] |
262                ((entry_size[i] - 1) << GFX7_URB_ENTRY_SIZE_SHIFT) |
263                (start[i] << GFX7_URB_STARTING_ADDRESS_SHIFT));
264   }
265   ADVANCE_BATCH();
266}
267
268const struct brw_tracked_state gfx7_urb = {
269   .dirty = {
270      .mesa = 0,
271      .brw = BRW_NEW_BLORP |
272             BRW_NEW_CONTEXT |
273             BRW_NEW_URB_SIZE |
274             BRW_NEW_GS_PROG_DATA |
275             BRW_NEW_TCS_PROG_DATA |
276             BRW_NEW_TES_PROG_DATA |
277             BRW_NEW_VS_PROG_DATA,
278   },
279   .emit = upload_urb,
280};
281