1/*
2 * Copyright (c) 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 <stdlib.h>
25#include <math.h>
26
27#include "util/macros.h"
28#include "main/macros.h"
29#include "compiler/shader_enums.h"
30
31#include "gen_l3_config.h"
32
33/**
34 * The following diagram shows how we partition the URB:
35 *
36 *        16kb or 32kb               Rest of the URB space
37 *   __________-__________   _________________-_________________
38 *  /                     \ /                                   \
39 * +-------------------------------------------------------------+
40 * |  VS/HS/DS/GS/FS Push  |           VS/HS/DS/GS URB           |
41 * |       Constants       |               Entries               |
42 * +-------------------------------------------------------------+
43 *
44 * Push constants must be stored at the beginning of the URB space,
45 * while URB entries can be stored anywhere.  We choose to lay them
46 * out in pipeline order (VS -> HS -> DS -> GS).
47 */
48
49/**
50 * Decide how to partition the URB among the various stages.
51 *
52 * \param[in] push_constant_bytes - space allocate for push constants.
53 * \param[in] urb_size_bytes - total size of the URB (from L3 config).
54 * \param[in] tess_present - are tessellation shaders active?
55 * \param[in] gs_present - are geometry shaders active?
56 * \param[in] entry_size - the URB entry size (from the shader compiler)
57 * \param[out] entries - the number of URB entries for each stage
58 * \param[out] start - the starting offset for each stage
59 */
60void
61gen_get_urb_config(const struct gen_device_info *devinfo,
62                   unsigned push_constant_bytes, unsigned urb_size_bytes,
63                   bool tess_present, bool gs_present,
64                   const unsigned entry_size[4],
65                   unsigned entries[4], unsigned start[4])
66{
67   const bool active[4] = { true, tess_present, tess_present, gs_present };
68
69   /* URB allocations must be done in 8k chunks. */
70   const unsigned chunk_size_bytes = 8192;
71
72   const unsigned push_constant_chunks =
73      push_constant_bytes / chunk_size_bytes;
74   const unsigned urb_chunks = urb_size_bytes / chunk_size_bytes;
75
76   /* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
77    *
78    *     VS Number of URB Entries must be divisible by 8 if the VS URB Entry
79    *     Allocation Size is less than 9 512-bit URB entries.
80    *
81    * Similar text exists for HS, DS and GS.
82    */
83   unsigned granularity[4];
84   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
85      granularity[i] = (entry_size[i] < 9) ? 8 : 1;
86   }
87
88   unsigned min_entries[4] = {
89      /* VS has a lower limit on the number of URB entries.
90       *
91       * From the Broadwell PRM, 3DSTATE_URB_VS instruction:
92       * "When tessellation is enabled, the VS Number of URB Entries must be
93       *  greater than or equal to 192."
94       */
95      [MESA_SHADER_VERTEX] = tess_present && devinfo->gen == 8 ?
96         192 : devinfo->urb.min_entries[MESA_SHADER_VERTEX],
97
98      /* There are two constraints on the minimum amount of URB space we can
99       * allocate:
100       *
101       * (1) We need room for at least 2 URB entries, since we always operate
102       * the GS in DUAL_OBJECT mode.
103       *
104       * (2) We can't allocate less than nr_gs_entries_granularity.
105       */
106      [MESA_SHADER_GEOMETRY] = gs_present ? 2 : 0,
107
108      [MESA_SHADER_TESS_CTRL] = tess_present ? 1 : 0,
109
110      [MESA_SHADER_TESS_EVAL] = tess_present ?
111         devinfo->urb.min_entries[MESA_SHADER_TESS_EVAL] : 0,
112   };
113
114   /* Min VS Entries isn't a multiple of 8 on Cherryview/Broxton; round up.
115    * Round them all up.
116    */
117   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
118      min_entries[i] = ALIGN(min_entries[i], granularity[i]);
119   }
120
121   unsigned entry_size_bytes[4];
122   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
123      entry_size_bytes[i] = 64 * entry_size[i];
124   }
125
126   /* Initially, assign each stage the minimum amount of URB space it needs,
127    * and make a note of how much additional space it "wants" (the amount of
128    * additional space it could actually make use of).
129    */
130   unsigned chunks[4];
131   unsigned wants[4];
132   unsigned total_needs = push_constant_chunks;
133   unsigned total_wants = 0;
134
135   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
136      if (active[i]) {
137         chunks[i] = DIV_ROUND_UP(min_entries[i] * entry_size_bytes[i],
138                                  chunk_size_bytes);
139
140         wants[i] =
141            DIV_ROUND_UP(devinfo->urb.max_entries[i] * entry_size_bytes[i],
142                         chunk_size_bytes) - chunks[i];
143      } else {
144         chunks[i] = 0;
145         wants[i] = 0;
146      }
147
148      total_needs += chunks[i];
149      total_wants += wants[i];
150   }
151
152   assert(total_needs <= urb_chunks);
153
154   /* Mete out remaining space (if any) in proportion to "wants". */
155   unsigned remaining_space = MIN2(urb_chunks - total_needs, total_wants);
156
157   if (remaining_space > 0) {
158      for (int i = MESA_SHADER_VERTEX;
159           total_wants > 0 && i <= MESA_SHADER_TESS_EVAL; i++) {
160         unsigned additional = (unsigned)
161            roundf(wants[i] * (((float) remaining_space) / total_wants));
162         chunks[i] += additional;
163         remaining_space -= additional;
164         total_wants -= wants[i];
165      }
166
167      chunks[MESA_SHADER_GEOMETRY] += remaining_space;
168   }
169
170   /* Sanity check that we haven't over-allocated. */
171   unsigned total_chunks = push_constant_chunks;
172   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
173      total_chunks += chunks[i];
174   }
175   assert(total_chunks <= urb_chunks);
176
177   /* Finally, compute the number of entries that can fit in the space
178    * allocated to each stage.
179    */
180   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
181      entries[i] = chunks[i] * chunk_size_bytes / entry_size_bytes[i];
182
183      /* Since we rounded up when computing wants[], this may be slightly
184       * more than the maximum allowed amount, so correct for that.
185       */
186      entries[i] = MIN2(entries[i], devinfo->urb.max_entries[i]);
187
188      /* Ensure that we program a multiple of the granularity. */
189      entries[i] = ROUND_DOWN_TO(entries[i], granularity[i]);
190
191      /* Finally, sanity check to make sure we have at least the minimum
192       * number of entries needed for each stage.
193       */
194      assert(entries[i] >= min_entries[i]);
195   }
196
197   /* Lay out the URB in pipeline order: push constants, VS, HS, DS, GS. */
198   int next = push_constant_chunks;
199   for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
200      if (entries[i]) {
201         start[i] = next;
202         next += chunks[i];
203      } else {
204         /* Just put disabled stages at the beginning. */
205         start[i] = 0;
206      }
207   }
208}
209