xref: /aosp_15_r20/external/mesa3d/src/intel/compiler/brw_schedule_instructions.cpp (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright © 2010 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  * Authors:
24  *    Eric Anholt <[email protected]>
25  *
26  */
27 
28 #include "brw_eu.h"
29 #include "brw_fs.h"
30 #include "brw_fs_live_variables.h"
31 #include "brw_cfg.h"
32 #include <new>
33 
34 using namespace brw;
35 
36 /** @file
37  *
38  * List scheduling of FS instructions.
39  *
40  * The basic model of the list scheduler is to take a basic block,
41  * compute a DAG of the dependencies (RAW ordering with latency, WAW
42  * ordering with latency, WAR ordering), and make a list of the DAG heads.
43  * Heuristically pick a DAG head, then put all the children that are
44  * now DAG heads into the list of things to schedule.
45  *
46  * The heuristic is the important part.  We're trying to be cheap,
47  * since actually computing the optimal scheduling is NP complete.
48  * What we do is track a "current clock".  When we schedule a node, we
49  * update the earliest-unblocked clock time of its children, and
50  * increment the clock.  Then, when trying to schedule, we just pick
51  * the earliest-unblocked instruction to schedule.
52  *
53  * Note that often there will be many things which could execute
54  * immediately, and there are a range of heuristic options to choose
55  * from in picking among those.
56  */
57 
58 static bool debug = false;
59 
60 struct schedule_node_child;
61 
62 class schedule_node : public exec_node
63 {
64 public:
65    void set_latency(const struct brw_isa_info *isa);
66 
67    fs_inst *inst;
68    schedule_node_child *children;
69    int children_count;
70    int children_cap;
71    int initial_parent_count;
72    int initial_unblocked_time;
73    int latency;
74 
75    /**
76     * This is the sum of the instruction's latency plus the maximum delay of
77     * its children, or just the issue_time if it's a leaf node.
78     */
79    int delay;
80 
81    /**
82     * Preferred exit node among the (direct or indirect) successors of this
83     * node.  Among the scheduler nodes blocked by this node, this will be the
84     * one that may cause earliest program termination, or NULL if none of the
85     * successors is an exit node.
86     */
87    schedule_node *exit;
88 
89    /**
90     * How many cycles this instruction takes to issue.
91     *
92     * Instructions in gen hardware are handled one simd4 vector at a time,
93     * with 1 cycle per vector dispatched.  Thus SIMD8 pixel shaders take 2
94     * cycles to dispatch and SIMD16 (compressed) instructions take 4.
95     */
96    int issue_time;
97 
98    /* Temporary data used during the scheduling process. */
99    struct {
100       int parent_count;
101       int unblocked_time;
102 
103       /**
104        * Which iteration of pushing groups of children onto the candidates list
105        * this node was a part of.
106        */
107       unsigned cand_generation;
108    } tmp;
109 };
110 
111 struct schedule_node_child {
112    schedule_node *n;
113    int effective_latency;
114 };
115 
116 static inline void
reset_node_tmp(schedule_node * n)117 reset_node_tmp(schedule_node *n)
118 {
119    n->tmp.parent_count = n->initial_parent_count;
120    n->tmp.unblocked_time = n->initial_unblocked_time;
121    n->tmp.cand_generation = 0;
122 }
123 
124 /**
125  * Lower bound of the scheduling time after which one of the instructions
126  * blocked by this node may lead to program termination.
127  *
128  * exit_unblocked_time() determines a strict partial ordering relation '«' on
129  * the set of scheduler nodes as follows:
130  *
131  *   n « m <-> exit_unblocked_time(n) < exit_unblocked_time(m)
132  *
133  * which can be used to heuristically order nodes according to how early they
134  * can unblock an exit node and lead to program termination.
135  */
136 static inline int
exit_tmp_unblocked_time(const schedule_node * n)137 exit_tmp_unblocked_time(const schedule_node *n)
138 {
139    return n->exit ? n->exit->tmp.unblocked_time : INT_MAX;
140 }
141 
142 static inline int
exit_initial_unblocked_time(const schedule_node * n)143 exit_initial_unblocked_time(const schedule_node *n)
144 {
145    return n->exit ? n->exit->initial_unblocked_time : INT_MAX;
146 }
147 
148 void
set_latency(const struct brw_isa_info * isa)149 schedule_node::set_latency(const struct brw_isa_info *isa)
150 {
151    switch (inst->opcode) {
152    case BRW_OPCODE_MAD:
153       /* 2 cycles
154        *  (since the last two src operands are in different register banks):
155        * mad(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q };
156        *
157        * 3 cycles on IVB, 4 on HSW
158        *  (since the last two src operands are in the same register bank):
159        * mad(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q };
160        *
161        * 18 cycles on IVB, 16 on HSW
162        *  (since the last two src operands are in different register banks):
163        * mad(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q };
164        * mov(8) null   g4<4,5,1>F                     { align16 WE_normal 1Q };
165        *
166        * 20 cycles on IVB, 18 on HSW
167        *  (since the last two src operands are in the same register bank):
168        * mad(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q };
169        * mov(8) null   g4<4,4,1>F                     { align16 WE_normal 1Q };
170        */
171 
172       /* Our register allocator doesn't know about register banks, so use the
173        * higher latency.
174        */
175       latency = 18;
176       break;
177 
178    case BRW_OPCODE_LRP:
179       /* 2 cycles
180        *  (since the last two src operands are in different register banks):
181        * lrp(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q };
182        *
183        * 3 cycles on IVB, 4 on HSW
184        *  (since the last two src operands are in the same register bank):
185        * lrp(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q };
186        *
187        * 16 cycles on IVB, 14 on HSW
188        *  (since the last two src operands are in different register banks):
189        * lrp(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q };
190        * mov(8) null   g4<4,4,1>F                     { align16 WE_normal 1Q };
191        *
192        * 16 cycles
193        *  (since the last two src operands are in the same register bank):
194        * lrp(8) g4<1>F g2.2<4,4,1>F.x  g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q };
195        * mov(8) null   g4<4,4,1>F                     { align16 WE_normal 1Q };
196        */
197 
198       /* Our register allocator doesn't know about register banks, so use the
199        * higher latency.
200        */
201       latency = 14;
202       break;
203 
204    case SHADER_OPCODE_RCP:
205    case SHADER_OPCODE_RSQ:
206    case SHADER_OPCODE_SQRT:
207    case SHADER_OPCODE_LOG2:
208    case SHADER_OPCODE_EXP2:
209    case SHADER_OPCODE_SIN:
210    case SHADER_OPCODE_COS:
211       /* 2 cycles:
212        * math inv(8) g4<1>F g2<0,1,0>F      null       { align1 WE_normal 1Q };
213        *
214        * 18 cycles:
215        * math inv(8) g4<1>F g2<0,1,0>F      null       { align1 WE_normal 1Q };
216        * mov(8)      null   g4<8,8,1>F                 { align1 WE_normal 1Q };
217        *
218        * Same for exp2, log2, rsq, sqrt, sin, cos.
219        */
220       latency = 16;
221       break;
222 
223    case SHADER_OPCODE_POW:
224       /* 2 cycles:
225        * math pow(8) g4<1>F g2<0,1,0>F   g2.1<0,1,0>F  { align1 WE_normal 1Q };
226        *
227        * 26 cycles:
228        * math pow(8) g4<1>F g2<0,1,0>F   g2.1<0,1,0>F  { align1 WE_normal 1Q };
229        * mov(8)      null   g4<8,8,1>F                 { align1 WE_normal 1Q };
230        */
231       latency = 24;
232       break;
233 
234    case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
235       /* testing using varying-index pull constants:
236        *
237        * 16 cycles:
238        * mov(8)  g4<1>D  g2.1<0,1,0>F                  { align1 WE_normal 1Q };
239        * send(8) g4<1>F  g4<8,8,1>D
240        *   data (9, 2, 3) mlen 1 rlen 1                { align1 WE_normal 1Q };
241        *
242        * ~480 cycles:
243        * mov(8)  g4<1>D  g2.1<0,1,0>F                  { align1 WE_normal 1Q };
244        * send(8) g4<1>F  g4<8,8,1>D
245        *   data (9, 2, 3) mlen 1 rlen 1                { align1 WE_normal 1Q };
246        * mov(8)  null    g4<8,8,1>F                    { align1 WE_normal 1Q };
247        *
248        * ~620 cycles:
249        * mov(8)  g4<1>D  g2.1<0,1,0>F                  { align1 WE_normal 1Q };
250        * send(8) g4<1>F  g4<8,8,1>D
251        *   data (9, 2, 3) mlen 1 rlen 1                { align1 WE_normal 1Q };
252        * mov(8)  null    g4<8,8,1>F                    { align1 WE_normal 1Q };
253        * send(8) g4<1>F  g4<8,8,1>D
254        *   data (9, 2, 3) mlen 1 rlen 1                { align1 WE_normal 1Q };
255        * mov(8)  null    g4<8,8,1>F                    { align1 WE_normal 1Q };
256        *
257        * So, if it's cache-hot, it's about 140.  If it's cache cold, it's
258        * about 460.  We expect to mostly be cache hot, so pick something more
259        * in that direction.
260        */
261       latency = 200;
262       break;
263 
264    case SHADER_OPCODE_SEND:
265       switch (inst->sfid) {
266       case BRW_SFID_SAMPLER: {
267          unsigned msg_type = (inst->desc >> 12) & 0x1f;
268          switch (msg_type) {
269          case GFX5_SAMPLER_MESSAGE_SAMPLE_RESINFO:
270          case GFX6_SAMPLER_MESSAGE_SAMPLE_SAMPLEINFO:
271             /* Testing textureSize(sampler2D, 0), one load was 420 +/- 41
272              * cycles (n=15):
273              * mov(8)   g114<1>UD  0D                  { align1 WE_normal 1Q };
274              * send(8)  g6<1>UW    g114<8,8,1>F
275              *   sampler (10, 0, 10, 1) mlen 1 rlen 4  { align1 WE_normal 1Q };
276              * mov(16)  g6<1>F     g6<8,8,1>D                { align1 WE_normal 1Q };
277              *
278              *
279              * Two loads was 535 +/- 30 cycles (n=19):
280              * mov(16)   g114<1>UD  0D                 { align1 WE_normal 1H };
281              * send(16)  g6<1>UW    g114<8,8,1>F
282              *   sampler (10, 0, 10, 2) mlen 2 rlen 8  { align1 WE_normal 1H };
283              * mov(16)   g114<1>UD  0D                 { align1 WE_normal 1H };
284              * mov(16)   g6<1>F     g6<8,8,1>D         { align1 WE_normal 1H };
285              * send(16)  g8<1>UW    g114<8,8,1>F
286              *   sampler (10, 0, 10, 2) mlen 2 rlen 8  { align1 WE_normal 1H };
287              * mov(16)   g8<1>F     g8<8,8,1>D         { align1 WE_normal 1H };
288              * add(16)   g6<1>F     g6<8,8,1>F   g8<8,8,1>F  { align1 WE_normal 1H };
289              *
290              * Since the only caches that should matter are just the
291              * instruction/state cache containing the surface state,
292              * assume that we always have hot caches.
293              */
294             latency = 100;
295             break;
296 
297          default:
298             /* 18 cycles:
299              * mov(8)  g115<1>F   0F                  { align1 WE_normal 1Q };
300              * mov(8)  g114<1>F   0F                  { align1 WE_normal 1Q };
301              * send(8) g4<1>UW    g114<8,8,1>F
302              *   sampler (10, 0, 0, 1) mlen 2 rlen 4  { align1 WE_normal 1Q };
303              *
304              * 697 +/-49 cycles (min 610, n=26):
305              * mov(8)  g115<1>F   0F                  { align1 WE_normal 1Q };
306              * mov(8)  g114<1>F   0F                  { align1 WE_normal 1Q };
307              * send(8) g4<1>UW    g114<8,8,1>F
308              *   sampler (10, 0, 0, 1) mlen 2 rlen 4  { align1 WE_normal 1Q };
309              * mov(8)  null       g4<8,8,1>F          { align1 WE_normal 1Q };
310              *
311              * So the latency on our first texture load of the batchbuffer
312              * takes ~700 cycles, since the caches are cold at that point.
313              *
314              * 840 +/- 92 cycles (min 720, n=25):
315              * mov(8)  g115<1>F   0F                  { align1 WE_normal 1Q };
316              * mov(8)  g114<1>F   0F                  { align1 WE_normal 1Q };
317              * send(8) g4<1>UW    g114<8,8,1>F
318              *   sampler (10, 0, 0, 1) mlen 2 rlen 4  { align1 WE_normal 1Q };
319              * mov(8)  null       g4<8,8,1>F          { align1 WE_normal 1Q };
320              * send(8) g4<1>UW    g114<8,8,1>F
321              *   sampler (10, 0, 0, 1) mlen 2 rlen 4  { align1 WE_normal 1Q };
322              * mov(8)  null       g4<8,8,1>F          { align1 WE_normal 1Q };
323              *
324              * On the second load, it takes just an extra ~140 cycles, and
325              * after accounting for the 14 cycles of the MOV's latency, that
326              * makes ~130.
327              *
328              * 683 +/- 49 cycles (min = 602, n=47):
329              * mov(8)  g115<1>F   0F                  { align1 WE_normal 1Q };
330              * mov(8)  g114<1>F   0F                  { align1 WE_normal 1Q };
331              * send(8) g4<1>UW    g114<8,8,1>F
332              *   sampler (10, 0, 0, 1) mlen 2 rlen 4  { align1 WE_normal 1Q };
333              * send(8) g50<1>UW   g114<8,8,1>F
334              *   sampler (10, 0, 0, 1) mlen 2 rlen 4  { align1 WE_normal 1Q };
335              * mov(8)  null       g4<8,8,1>F          { align1 WE_normal 1Q };
336              *
337              * The unit appears to be pipelined, since this matches up with
338              * the cache-cold case, despite there being two loads here.  If
339              * you replace the g4 in the MOV to null with g50, it's still
340              * 693 +/- 52 (n=39).
341              *
342              * So, take some number between the cache-hot 140 cycles and the
343              * cache-cold 700 cycles.  No particular tuning was done on this.
344              *
345              * I haven't done significant testing of the non-TEX opcodes.
346              * TXL at least looked about the same as TEX.
347              */
348             latency = 200;
349             break;
350          }
351          break;
352       }
353 
354       case GFX6_SFID_DATAPORT_CONSTANT_CACHE:
355          /* See FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD */
356          latency = 200;
357          break;
358 
359       case GFX6_SFID_DATAPORT_RENDER_CACHE:
360          switch (brw_fb_desc_msg_type(isa->devinfo, inst->desc)) {
361          case GFX7_DATAPORT_RC_TYPED_SURFACE_WRITE:
362          case GFX7_DATAPORT_RC_TYPED_SURFACE_READ:
363             /* See also SHADER_OPCODE_TYPED_SURFACE_READ */
364             latency = 600;
365             break;
366 
367          case GFX7_DATAPORT_RC_TYPED_ATOMIC_OP:
368             /* See also SHADER_OPCODE_TYPED_ATOMIC */
369             latency = 14000;
370             break;
371 
372          case GFX6_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE:
373             /* completely fabricated number */
374             latency = 600;
375             break;
376 
377          default:
378             unreachable("Unknown render cache message");
379          }
380          break;
381 
382       case GFX7_SFID_DATAPORT_DATA_CACHE:
383          switch ((inst->desc >> 14) & 0x1f) {
384          case BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ:
385          case GFX7_DATAPORT_DC_UNALIGNED_OWORD_BLOCK_READ:
386          case GFX6_DATAPORT_WRITE_MESSAGE_OWORD_BLOCK_WRITE:
387             /* We have no data for this but assume it's a little faster than
388              * untyped surface read/write.
389              */
390             latency = 200;
391             break;
392 
393          case GFX7_DATAPORT_DC_DWORD_SCATTERED_READ:
394          case GFX6_DATAPORT_WRITE_MESSAGE_DWORD_SCATTERED_WRITE:
395          case HSW_DATAPORT_DC_PORT0_BYTE_SCATTERED_READ:
396          case HSW_DATAPORT_DC_PORT0_BYTE_SCATTERED_WRITE:
397             /* We have no data for this but assume it's roughly the same as
398              * untyped surface read/write.
399              */
400             latency = 300;
401             break;
402 
403          case GFX7_DATAPORT_DC_UNTYPED_SURFACE_READ:
404          case GFX7_DATAPORT_DC_UNTYPED_SURFACE_WRITE:
405             /* Test code:
406              *   mov(8)    g112<1>UD       0x00000000UD       { align1 WE_all 1Q };
407              *   mov(1)    g112.7<1>UD     g1.7<0,1,0>UD      { align1 WE_all };
408              *   mov(8)    g113<1>UD       0x00000000UD       { align1 WE_normal 1Q };
409              *   send(8)   g4<1>UD         g112<8,8,1>UD
410              *             data (38, 6, 5) mlen 2 rlen 1      { align1 WE_normal 1Q };
411              *   .
412              *   . [repeats 8 times]
413              *   .
414              *   mov(8)    g112<1>UD       0x00000000UD       { align1 WE_all 1Q };
415              *   mov(1)    g112.7<1>UD     g1.7<0,1,0>UD      { align1 WE_all };
416              *   mov(8)    g113<1>UD       0x00000000UD       { align1 WE_normal 1Q };
417              *   send(8)   g4<1>UD         g112<8,8,1>UD
418              *             data (38, 6, 5) mlen 2 rlen 1      { align1 WE_normal 1Q };
419              *
420              * Running it 100 times as fragment shader on a 128x128 quad
421              * gives an average latency of 583 cycles per surface read,
422              * standard deviation 0.9%.
423              */
424             latency = 600;
425             break;
426 
427          case GFX7_DATAPORT_DC_UNTYPED_ATOMIC_OP:
428             /* Test code:
429              *   mov(8)    g112<1>ud       0x00000000ud       { align1 WE_all 1Q };
430              *   mov(1)    g112.7<1>ud     g1.7<0,1,0>ud      { align1 WE_all };
431              *   mov(8)    g113<1>ud       0x00000000ud       { align1 WE_normal 1Q };
432              *   send(8)   g4<1>ud         g112<8,8,1>ud
433              *             data (38, 5, 6) mlen 2 rlen 1      { align1 WE_normal 1Q };
434              *
435              * Running it 100 times as fragment shader on a 128x128 quad
436              * gives an average latency of 13867 cycles per atomic op,
437              * standard deviation 3%.  Note that this is a rather
438              * pessimistic estimate, the actual latency in cases with few
439              * collisions between threads and favorable pipelining has been
440              * seen to be reduced by a factor of 100.
441              */
442             latency = 14000;
443             break;
444 
445          default:
446             unreachable("Unknown data cache message");
447          }
448          break;
449 
450       case HSW_SFID_DATAPORT_DATA_CACHE_1:
451          switch (brw_dp_desc_msg_type(isa->devinfo, inst->desc)) {
452          case HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_READ:
453          case HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_WRITE:
454          case HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_READ:
455          case HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_WRITE:
456          case GFX8_DATAPORT_DC_PORT1_A64_UNTYPED_SURFACE_WRITE:
457          case GFX8_DATAPORT_DC_PORT1_A64_UNTYPED_SURFACE_READ:
458          case GFX8_DATAPORT_DC_PORT1_A64_SCATTERED_WRITE:
459          case GFX9_DATAPORT_DC_PORT1_A64_SCATTERED_READ:
460          case GFX9_DATAPORT_DC_PORT1_A64_OWORD_BLOCK_READ:
461          case GFX9_DATAPORT_DC_PORT1_A64_OWORD_BLOCK_WRITE:
462             /* See also GFX7_DATAPORT_DC_UNTYPED_SURFACE_READ */
463             latency = 300;
464             break;
465 
466          case HSW_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_OP:
467          case HSW_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_OP_SIMD4X2:
468          case HSW_DATAPORT_DC_PORT1_TYPED_ATOMIC_OP_SIMD4X2:
469          case HSW_DATAPORT_DC_PORT1_TYPED_ATOMIC_OP:
470          case GFX9_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_FLOAT_OP:
471          case GFX8_DATAPORT_DC_PORT1_A64_UNTYPED_ATOMIC_OP:
472          case GFX9_DATAPORT_DC_PORT1_A64_UNTYPED_ATOMIC_FLOAT_OP:
473          case GFX12_DATAPORT_DC_PORT1_A64_UNTYPED_ATOMIC_HALF_INT_OP:
474          case GFX12_DATAPORT_DC_PORT1_A64_UNTYPED_ATOMIC_HALF_FLOAT_OP:
475             /* See also GFX7_DATAPORT_DC_UNTYPED_ATOMIC_OP */
476             latency = 14000;
477             break;
478 
479          default:
480             unreachable("Unknown data cache message");
481          }
482          break;
483 
484       case GFX7_SFID_PIXEL_INTERPOLATOR:
485          latency = 50; /* TODO */
486          break;
487 
488       case GFX12_SFID_UGM:
489       case GFX12_SFID_TGM:
490       case GFX12_SFID_SLM:
491          switch (lsc_msg_desc_opcode(isa->devinfo, inst->desc)) {
492          case LSC_OP_LOAD:
493          case LSC_OP_STORE:
494          case LSC_OP_LOAD_CMASK:
495          case LSC_OP_STORE_CMASK:
496             latency = 300;
497             break;
498          case LSC_OP_FENCE:
499          case LSC_OP_ATOMIC_INC:
500          case LSC_OP_ATOMIC_DEC:
501          case LSC_OP_ATOMIC_LOAD:
502          case LSC_OP_ATOMIC_STORE:
503          case LSC_OP_ATOMIC_ADD:
504          case LSC_OP_ATOMIC_SUB:
505          case LSC_OP_ATOMIC_MIN:
506          case LSC_OP_ATOMIC_MAX:
507          case LSC_OP_ATOMIC_UMIN:
508          case LSC_OP_ATOMIC_UMAX:
509          case LSC_OP_ATOMIC_CMPXCHG:
510          case LSC_OP_ATOMIC_FADD:
511          case LSC_OP_ATOMIC_FSUB:
512          case LSC_OP_ATOMIC_FMIN:
513          case LSC_OP_ATOMIC_FMAX:
514          case LSC_OP_ATOMIC_FCMPXCHG:
515          case LSC_OP_ATOMIC_AND:
516          case LSC_OP_ATOMIC_OR:
517          case LSC_OP_ATOMIC_XOR:
518             latency = 1400;
519             break;
520          default:
521             unreachable("unsupported new data port message instruction");
522          }
523          break;
524 
525       case BRW_SFID_MESSAGE_GATEWAY:
526       case GEN_RT_SFID_BINDLESS_THREAD_DISPATCH: /* or THREAD_SPAWNER */
527       case GEN_RT_SFID_RAY_TRACE_ACCELERATOR:
528          /* TODO.
529           *
530           * We'll assume for the moment that this is pretty quick as it
531           * doesn't actually return any data.
532           */
533          latency = 200;
534          break;
535 
536       case BRW_SFID_URB:
537          latency = 200;
538          break;
539 
540       default:
541          unreachable("Unknown SFID");
542       }
543       break;
544 
545    case BRW_OPCODE_DPAS:
546       switch (inst->rcount) {
547       case 1:
548          latency = 21;
549          break;
550       case 2:
551          latency = 22;
552          break;
553       case 8:
554       default:
555          latency = 32;
556          break;
557       }
558       break;
559 
560    default:
561       /* 2 cycles:
562        * mul(8) g4<1>F g2<0,1,0>F      0.5F            { align1 WE_normal 1Q };
563        *
564        * 16 cycles:
565        * mul(8) g4<1>F g2<0,1,0>F      0.5F            { align1 WE_normal 1Q };
566        * mov(8) null   g4<8,8,1>F                      { align1 WE_normal 1Q };
567        */
568       latency = 14;
569       break;
570    }
571 }
572 
573 class instruction_scheduler {
574 public:
575    instruction_scheduler(void *mem_ctx, const fs_visitor *s, int grf_count, int hw_reg_count,
576                          int block_count, bool post_reg_alloc);
577 
578    void add_barrier_deps(schedule_node *n);
579    void add_cross_lane_deps(schedule_node *n);
580    void add_dep(schedule_node *before, schedule_node *after, int latency);
581    void add_dep(schedule_node *before, schedule_node *after);
582 
583    void set_current_block(bblock_t *block);
584    void compute_delays();
585    void compute_exits();
586 
587    void schedule(schedule_node *chosen);
588    void update_children(schedule_node *chosen);
589 
590    void calculate_deps();
591    bool is_compressed(const fs_inst *inst);
592    bool register_needs_barrier(const brw_reg &reg);
593    schedule_node *choose_instruction_to_schedule();
594    int calculate_issue_time(const fs_inst *inst);
595 
596    void count_reads_remaining(const fs_inst *inst);
597    void setup_liveness(cfg_t *cfg);
598    void update_register_pressure(const fs_inst *inst);
599    int get_register_pressure_benefit(const fs_inst *inst);
600    void clear_last_grf_write();
601 
602    void schedule_instructions();
603    void run(instruction_scheduler_mode mode);
604 
605    int grf_index(const brw_reg &reg);
606 
607    void *mem_ctx;
608    linear_ctx *lin_ctx;
609 
610    schedule_node *nodes;
611    int nodes_len;
612 
613    /* Current block being processed. */
614    struct {
615       bblock_t *block;
616 
617       /* Range of nodes in the block.  End will point to first node
618        * address after the block, i.e. the range is [start, end).
619        */
620       schedule_node *start;
621       schedule_node *end;
622       int len;
623 
624       int scheduled;
625 
626       unsigned cand_generation;
627       int time;
628       exec_list available;
629    } current;
630 
631    bool post_reg_alloc;
632    int grf_count;
633    const fs_visitor *s;
634 
635    /**
636     * Last instruction to have written the grf (or a channel in the grf, for the
637     * scalar backend)
638     */
639    schedule_node **last_grf_write;
640 
641    unsigned hw_reg_count;
642    int reg_pressure;
643    instruction_scheduler_mode mode;
644 
645    /*
646     * The register pressure at the beginning of each basic block.
647     */
648 
649    int *reg_pressure_in;
650 
651    /*
652     * The virtual GRF's whose range overlaps the beginning of each basic block.
653     */
654 
655    BITSET_WORD **livein;
656 
657    /*
658     * The virtual GRF's whose range overlaps the end of each basic block.
659     */
660 
661    BITSET_WORD **liveout;
662 
663    /*
664     * The hardware GRF's whose range overlaps the end of each basic block.
665     */
666 
667    BITSET_WORD **hw_liveout;
668 
669    /*
670     * Whether we've scheduled a write for this virtual GRF yet.
671     */
672 
673    bool *written;
674 
675    /*
676     * How many reads we haven't scheduled for this virtual GRF yet.
677     */
678 
679    int *reads_remaining;
680 
681    /*
682     * How many reads we haven't scheduled for this hardware GRF yet.
683     */
684 
685    int *hw_reads_remaining;
686 };
687 
instruction_scheduler(void * mem_ctx,const fs_visitor * s,int grf_count,int hw_reg_count,int block_count,bool post_reg_alloc)688 instruction_scheduler::instruction_scheduler(void *mem_ctx, const fs_visitor *s,
689                                              int grf_count, int hw_reg_count,
690                                              int block_count, bool post_reg_alloc)
691    : s(s)
692 {
693    this->mem_ctx = mem_ctx;
694    this->lin_ctx = linear_context(this->mem_ctx);
695    this->grf_count = grf_count;
696    this->post_reg_alloc = post_reg_alloc;
697 
698    const unsigned grf_write_scale = MAX_VGRF_SIZE(s->devinfo);
699    this->last_grf_write = linear_zalloc_array(lin_ctx, schedule_node *, grf_count * grf_write_scale);
700 
701    this->nodes_len = s->cfg->last_block()->end_ip + 1;
702    this->nodes = linear_zalloc_array(lin_ctx, schedule_node, this->nodes_len);
703 
704    const struct brw_isa_info *isa = &s->compiler->isa;
705 
706    schedule_node *n = nodes;
707    foreach_block_and_inst(block, fs_inst, inst, s->cfg) {
708       n->inst = inst;
709 
710       if (!post_reg_alloc)
711          n->latency = 1;
712       else
713          n->set_latency(isa);
714 
715       n++;
716    }
717    assert(n == nodes + nodes_len);
718 
719    current.block = NULL;
720    current.start = NULL;
721    current.end = NULL;
722    current.len = 0;
723    current.time = 0;
724    current.cand_generation = 0;
725    current.available.make_empty();
726 
727    this->hw_reg_count = hw_reg_count;
728    this->mode = SCHEDULE_NONE;
729    this->reg_pressure = 0;
730 
731    if (!post_reg_alloc) {
732       this->reg_pressure_in = linear_zalloc_array(lin_ctx, int, block_count);
733 
734       this->livein = linear_alloc_array(lin_ctx, BITSET_WORD *, block_count);
735       for (int i = 0; i < block_count; i++)
736          this->livein[i] = linear_zalloc_array(lin_ctx, BITSET_WORD,
737                                          BITSET_WORDS(grf_count));
738 
739       this->liveout = linear_alloc_array(lin_ctx, BITSET_WORD *, block_count);
740       for (int i = 0; i < block_count; i++)
741          this->liveout[i] = linear_zalloc_array(lin_ctx, BITSET_WORD,
742                                           BITSET_WORDS(grf_count));
743 
744       this->hw_liveout = linear_alloc_array(lin_ctx, BITSET_WORD *, block_count);
745       for (int i = 0; i < block_count; i++)
746          this->hw_liveout[i] = linear_zalloc_array(lin_ctx, BITSET_WORD,
747                                              BITSET_WORDS(hw_reg_count));
748 
749       setup_liveness(s->cfg);
750 
751       this->written = linear_alloc_array(lin_ctx, bool, grf_count);
752 
753       this->reads_remaining = linear_alloc_array(lin_ctx, int, grf_count);
754 
755       this->hw_reads_remaining = linear_alloc_array(lin_ctx, int, hw_reg_count);
756    } else {
757       this->reg_pressure_in = NULL;
758       this->livein = NULL;
759       this->liveout = NULL;
760       this->hw_liveout = NULL;
761       this->written = NULL;
762       this->reads_remaining = NULL;
763       this->hw_reads_remaining = NULL;
764    }
765 
766    foreach_block(block, s->cfg) {
767       set_current_block(block);
768 
769       for (schedule_node *n = current.start; n < current.end; n++)
770          n->issue_time = calculate_issue_time(n->inst);
771 
772       calculate_deps();
773       compute_delays();
774       compute_exits();
775    }
776 }
777 
778 static bool
is_src_duplicate(const fs_inst * inst,int src)779 is_src_duplicate(const fs_inst *inst, int src)
780 {
781    for (int i = 0; i < src; i++)
782      if (inst->src[i].equals(inst->src[src]))
783        return true;
784 
785   return false;
786 }
787 
788 void
count_reads_remaining(const fs_inst * inst)789 instruction_scheduler::count_reads_remaining(const fs_inst *inst)
790 {
791    assert(reads_remaining);
792 
793    for (int i = 0; i < inst->sources; i++) {
794       if (is_src_duplicate(inst, i))
795          continue;
796 
797       if (inst->src[i].file == VGRF) {
798          reads_remaining[inst->src[i].nr]++;
799       } else if (inst->src[i].file == FIXED_GRF) {
800          if (inst->src[i].nr >= hw_reg_count)
801             continue;
802 
803          for (unsigned j = 0; j < regs_read(inst, i); j++)
804             hw_reads_remaining[inst->src[i].nr + j]++;
805       }
806    }
807 }
808 
809 void
setup_liveness(cfg_t * cfg)810 instruction_scheduler::setup_liveness(cfg_t *cfg)
811 {
812    const fs_live_variables &live = s->live_analysis.require();
813 
814    /* First, compute liveness on a per-GRF level using the in/out sets from
815     * liveness calculation.
816     */
817    for (int block = 0; block < cfg->num_blocks; block++) {
818       for (int i = 0; i < live.num_vars; i++) {
819          if (BITSET_TEST(live.block_data[block].livein, i)) {
820             int vgrf = live.vgrf_from_var[i];
821             if (!BITSET_TEST(livein[block], vgrf)) {
822                reg_pressure_in[block] += s->alloc.sizes[vgrf];
823                BITSET_SET(livein[block], vgrf);
824             }
825          }
826 
827          if (BITSET_TEST(live.block_data[block].liveout, i))
828             BITSET_SET(liveout[block], live.vgrf_from_var[i]);
829       }
830    }
831 
832    /* Now, extend the live in/live out sets for when a range crosses a block
833     * boundary, which matches what our register allocator/interference code
834     * does to account for force_writemask_all and incompatible exec_mask's.
835     */
836    for (int block = 0; block < cfg->num_blocks - 1; block++) {
837       for (int i = 0; i < grf_count; i++) {
838          if (live.vgrf_start[i] <= cfg->blocks[block]->end_ip &&
839              live.vgrf_end[i] >= cfg->blocks[block + 1]->start_ip) {
840             if (!BITSET_TEST(livein[block + 1], i)) {
841                 reg_pressure_in[block + 1] += s->alloc.sizes[i];
842                 BITSET_SET(livein[block + 1], i);
843             }
844 
845             BITSET_SET(liveout[block], i);
846          }
847       }
848    }
849 
850    int payload_last_use_ip[hw_reg_count];
851    s->calculate_payload_ranges(true, hw_reg_count, payload_last_use_ip);
852 
853    for (unsigned i = 0; i < hw_reg_count; i++) {
854       if (payload_last_use_ip[i] == -1)
855          continue;
856 
857       for (int block = 0; block < cfg->num_blocks; block++) {
858          if (cfg->blocks[block]->start_ip <= payload_last_use_ip[i])
859             reg_pressure_in[block]++;
860 
861          if (cfg->blocks[block]->end_ip <= payload_last_use_ip[i])
862             BITSET_SET(hw_liveout[block], i);
863       }
864    }
865 }
866 
867 void
update_register_pressure(const fs_inst * inst)868 instruction_scheduler::update_register_pressure(const fs_inst *inst)
869 {
870    assert(reads_remaining);
871 
872    if (inst->dst.file == VGRF) {
873       written[inst->dst.nr] = true;
874    }
875 
876    for (int i = 0; i < inst->sources; i++) {
877       if (is_src_duplicate(inst, i))
878           continue;
879 
880       if (inst->src[i].file == VGRF) {
881          reads_remaining[inst->src[i].nr]--;
882       } else if (inst->src[i].file == FIXED_GRF &&
883                  inst->src[i].nr < hw_reg_count) {
884          for (unsigned off = 0; off < regs_read(inst, i); off++)
885             hw_reads_remaining[inst->src[i].nr + off]--;
886       }
887    }
888 }
889 
890 int
get_register_pressure_benefit(const fs_inst * inst)891 instruction_scheduler::get_register_pressure_benefit(const fs_inst *inst)
892 {
893    int benefit = 0;
894    const int block_idx = current.block->num;
895 
896    if (inst->dst.file == VGRF) {
897       if (!BITSET_TEST(livein[block_idx], inst->dst.nr) &&
898           !written[inst->dst.nr])
899          benefit -= s->alloc.sizes[inst->dst.nr];
900    }
901 
902    for (int i = 0; i < inst->sources; i++) {
903       if (is_src_duplicate(inst, i))
904          continue;
905 
906       if (inst->src[i].file == VGRF &&
907           !BITSET_TEST(liveout[block_idx], inst->src[i].nr) &&
908           reads_remaining[inst->src[i].nr] == 1)
909          benefit += s->alloc.sizes[inst->src[i].nr];
910 
911       if (inst->src[i].file == FIXED_GRF &&
912           inst->src[i].nr < hw_reg_count) {
913          for (unsigned off = 0; off < regs_read(inst, i); off++) {
914             int reg = inst->src[i].nr + off;
915             if (!BITSET_TEST(hw_liveout[block_idx], reg) &&
916                 hw_reads_remaining[reg] == 1) {
917                benefit++;
918             }
919          }
920       }
921    }
922 
923    return benefit;
924 }
925 
926 void
set_current_block(bblock_t * block)927 instruction_scheduler::set_current_block(bblock_t *block)
928 {
929    current.block = block;
930    current.start = nodes + block->start_ip;
931    current.len = block->end_ip - block->start_ip + 1;
932    current.end = current.start + current.len;
933    current.time = 0;
934    current.scheduled = 0;
935    current.cand_generation = 1;
936 }
937 
938 /** Computation of the delay member of each node. */
939 void
compute_delays()940 instruction_scheduler::compute_delays()
941 {
942    for (schedule_node *n = current.end - 1; n >= current.start; n--) {
943       if (!n->children_count) {
944          n->delay = n->issue_time;
945       } else {
946          for (int i = 0; i < n->children_count; i++) {
947             assert(n->children[i].n->delay);
948             n->delay = MAX2(n->delay, n->latency + n->children[i].n->delay);
949          }
950       }
951    }
952 }
953 
954 void
compute_exits()955 instruction_scheduler::compute_exits()
956 {
957    /* Calculate a lower bound of the scheduling time of each node in the
958     * graph.  This is analogous to the node's critical path but calculated
959     * from the top instead of from the bottom of the block.
960     */
961    for (schedule_node *n = current.start; n < current.end; n++) {
962       for (int i = 0; i < n->children_count; i++) {
963          schedule_node_child *child = &n->children[i];
964          child->n->initial_unblocked_time =
965             MAX2(child->n->initial_unblocked_time,
966                  n->initial_unblocked_time + n->issue_time + child->effective_latency);
967       }
968    }
969 
970    /* Calculate the exit of each node by induction based on the exit nodes of
971     * its children.  The preferred exit of a node is the one among the exit
972     * nodes of its children which can be unblocked first according to the
973     * optimistic unblocked time estimate calculated above.
974     */
975    for (schedule_node *n = current.end - 1; n >= current.start; n--) {
976       n->exit = (n->inst->opcode == BRW_OPCODE_HALT ? n : NULL);
977 
978       for (int i = 0; i < n->children_count; i++) {
979          if (exit_initial_unblocked_time(n->children[i].n) < exit_initial_unblocked_time(n))
980             n->exit = n->children[i].n->exit;
981       }
982    }
983 }
984 
985 /**
986  * Add a dependency between two instruction nodes.
987  *
988  * The @after node will be scheduled after @before.  We will try to
989  * schedule it @latency cycles after @before, but no guarantees there.
990  */
991 void
add_dep(schedule_node * before,schedule_node * after,int latency)992 instruction_scheduler::add_dep(schedule_node *before, schedule_node *after,
993                                int latency)
994 {
995    if (!before || !after)
996       return;
997 
998    assert(before != after);
999 
1000    for (int i = 0; i < before->children_count; i++) {
1001       schedule_node_child *child = &before->children[i];
1002       if (child->n == after) {
1003          child->effective_latency = MAX2(child->effective_latency, latency);
1004          return;
1005       }
1006    }
1007 
1008    if (before->children_cap <= before->children_count) {
1009       if (before->children_cap < 16)
1010          before->children_cap = 16;
1011       else
1012          before->children_cap *= 2;
1013 
1014       before->children = reralloc(mem_ctx, before->children,
1015                                   schedule_node_child,
1016                                   before->children_cap);
1017    }
1018 
1019    schedule_node_child *child = &before->children[before->children_count];
1020    child->n = after;
1021    child->effective_latency = latency;
1022    before->children_count++;
1023    after->initial_parent_count++;
1024 }
1025 
1026 void
add_dep(schedule_node * before,schedule_node * after)1027 instruction_scheduler::add_dep(schedule_node *before, schedule_node *after)
1028 {
1029    if (!before)
1030       return;
1031 
1032    add_dep(before, after, before->latency);
1033 }
1034 
1035 static bool
is_scheduling_barrier(const fs_inst * inst)1036 is_scheduling_barrier(const fs_inst *inst)
1037 {
1038    return inst->opcode == SHADER_OPCODE_HALT_TARGET ||
1039           inst->is_control_flow() ||
1040           inst->has_side_effects();
1041 }
1042 
1043 static bool
has_cross_lane_access(const fs_inst * inst)1044 has_cross_lane_access(const fs_inst *inst)
1045 {
1046    /* FINISHME:
1047     *
1048     * This function is likely incomplete in terms of identify cross lane
1049     * accesses.
1050     */
1051    if (inst->opcode == SHADER_OPCODE_BROADCAST ||
1052        inst->opcode == SHADER_OPCODE_CLUSTER_BROADCAST ||
1053        inst->opcode == SHADER_OPCODE_SHUFFLE ||
1054        inst->opcode == FS_OPCODE_LOAD_LIVE_CHANNELS ||
1055        inst->opcode == SHADER_OPCODE_LOAD_LIVE_CHANNELS ||
1056        inst->opcode == SHADER_OPCODE_FIND_LAST_LIVE_CHANNEL ||
1057        inst->opcode == SHADER_OPCODE_FIND_LIVE_CHANNEL)
1058       return true;
1059 
1060    for (unsigned s = 0; s < inst->sources; s++) {
1061       if (inst->src[s].file == VGRF) {
1062          if (inst->src[s].stride == 0)
1063             return true;
1064       }
1065    }
1066 
1067    return false;
1068 }
1069 
1070 /**
1071  * Some register access need dependencies on other instructions.
1072  */
1073 bool
register_needs_barrier(const brw_reg & reg)1074 instruction_scheduler::register_needs_barrier(const brw_reg &reg)
1075 {
1076    if (reg.file != ARF || reg.is_null())
1077       return false;
1078 
1079    /* If you look at SR register layout, there is nothing in there that
1080     * depends on other instructions. This is just fixed dispatch information.
1081     *
1082     * ATSM PRMs, Volume 9: Render Engine, State Register Fields :
1083     *    sr0.0:
1084     *      - 0:2   TID
1085     *      - 4:13  Slice, DSS, Subslice, EU IDs
1086     *      - 20:22 Priority
1087     *      - 23:23 Priority class
1088     *      - 24:27 FFID
1089     *    sr0.1:
1090     *      - 0:5   IEEE Exception
1091     *      - 21:31 FFTID
1092     *    sr0.2:
1093     *      - 0:31  Dispatch Mask
1094     *    sr0.3:
1095     *      - 0:31  Vector Mask
1096     */
1097    if (reg.nr == BRW_ARF_STATE)
1098       return false;
1099 
1100    return true;
1101 }
1102 
1103 /**
1104  * Sometimes we really want this node to execute after everything that
1105  * was before it and before everything that followed it.  This adds
1106  * the deps to do so.
1107  */
1108 void
add_barrier_deps(schedule_node * n)1109 instruction_scheduler::add_barrier_deps(schedule_node *n)
1110 {
1111    for (schedule_node *prev = n - 1; prev >= current.start; prev--) {
1112       add_dep(prev, n, 0);
1113       if (is_scheduling_barrier(prev->inst))
1114          break;
1115    }
1116 
1117    for (schedule_node *next = n + 1; next < current.end; next++) {
1118       add_dep(n, next, 0);
1119       if (is_scheduling_barrier(next->inst))
1120          break;
1121    }
1122 }
1123 
1124 /**
1125  * Because some instructions like HALT can disable lanes, scheduling prior to
1126  * a cross lane access should not be allowed, otherwise we could end up with
1127  * later instructions accessing uninitialized data.
1128  */
1129 void
add_cross_lane_deps(schedule_node * n)1130 instruction_scheduler::add_cross_lane_deps(schedule_node *n)
1131 {
1132    for (schedule_node *prev = n - 1; prev >= current.start; prev--) {
1133       if (has_cross_lane_access((fs_inst*)prev->inst))
1134          add_dep(prev, n, 0);
1135    }
1136 }
1137 
1138 /* instruction scheduling needs to be aware of when an MRF write
1139  * actually writes 2 MRFs.
1140  */
1141 bool
is_compressed(const fs_inst * inst)1142 instruction_scheduler::is_compressed(const fs_inst *inst)
1143 {
1144    return inst->exec_size == 16;
1145 }
1146 
1147 /* Clears last_grf_write to be ready to start calculating deps for a block
1148  * again.
1149  *
1150  * Since pre-ra grf_count scales with instructions, and instructions scale with
1151  * BBs, we don't want to memset all of last_grf_write per block or you'll end up
1152  * O(n^2) with number of blocks.  For shaders using softfp64, we get a *lot* of
1153  * blocks.
1154  *
1155  * We don't bother being careful for post-ra, since then grf_count doesn't scale
1156  * with instructions.
1157  */
1158 void
clear_last_grf_write()1159 instruction_scheduler::clear_last_grf_write()
1160 {
1161    if (!post_reg_alloc) {
1162       for (schedule_node *n = current.start; n < current.end; n++) {
1163          fs_inst *inst = (fs_inst *)n->inst;
1164 
1165          if (inst->dst.file == VGRF) {
1166             /* Don't bother being careful with regs_written(), quicker to just clear 2 cachelines. */
1167             memset(&last_grf_write[inst->dst.nr * MAX_VGRF_SIZE(s->devinfo)], 0,
1168                    sizeof(*last_grf_write) * MAX_VGRF_SIZE(s->devinfo));
1169          }
1170       }
1171    } else {
1172       memset(last_grf_write, 0,
1173              sizeof(*last_grf_write) * grf_count * MAX_VGRF_SIZE(s->devinfo));
1174    }
1175 }
1176 
1177 int
grf_index(const brw_reg & reg)1178 instruction_scheduler::grf_index(const brw_reg &reg)
1179 {
1180    if (post_reg_alloc)
1181       return reg.nr;
1182    return reg.nr * MAX_VGRF_SIZE(s->devinfo) + reg.offset / REG_SIZE;
1183 }
1184 
1185 void
calculate_deps()1186 instruction_scheduler::calculate_deps()
1187 {
1188    /* Pre-register-allocation, this tracks the last write per VGRF offset.
1189     * After register allocation, reg_offsets are gone and we track individual
1190     * GRF registers.
1191     */
1192    schedule_node *last_conditional_mod[8] = {};
1193    schedule_node *last_accumulator_write = NULL;
1194    /* Fixed HW registers are assumed to be separate from the virtual
1195     * GRFs, so they can be tracked separately.  We don't really write
1196     * to fixed GRFs much, so don't bother tracking them on a more
1197     * granular level.
1198     */
1199    schedule_node *last_fixed_grf_write = NULL;
1200 
1201    /* top-to-bottom dependencies: RAW and WAW. */
1202    for (schedule_node *n = current.start; n < current.end; n++) {
1203       fs_inst *inst = (fs_inst *)n->inst;
1204 
1205       if (is_scheduling_barrier(inst))
1206          add_barrier_deps(n);
1207 
1208       if (inst->opcode == BRW_OPCODE_HALT ||
1209           inst->opcode == SHADER_OPCODE_HALT_TARGET)
1210           add_cross_lane_deps(n);
1211 
1212       /* read-after-write deps. */
1213       for (int i = 0; i < inst->sources; i++) {
1214          if (inst->src[i].file == VGRF) {
1215             for (unsigned r = 0; r < regs_read(inst, i); r++)
1216                add_dep(last_grf_write[grf_index(inst->src[i]) + r], n);
1217          } else if (inst->src[i].file == FIXED_GRF) {
1218             if (post_reg_alloc) {
1219                for (unsigned r = 0; r < regs_read(inst, i); r++)
1220                   add_dep(last_grf_write[inst->src[i].nr + r], n);
1221             } else {
1222                add_dep(last_fixed_grf_write, n);
1223             }
1224          } else if (inst->src[i].is_accumulator()) {
1225             add_dep(last_accumulator_write, n);
1226          } else if (register_needs_barrier(inst->src[i])) {
1227             add_barrier_deps(n);
1228          }
1229       }
1230 
1231 
1232       if (const unsigned mask = inst->flags_read(s->devinfo)) {
1233          assert(mask < (1 << ARRAY_SIZE(last_conditional_mod)));
1234 
1235          for (unsigned i = 0; i < ARRAY_SIZE(last_conditional_mod); i++) {
1236             if (mask & (1 << i))
1237                add_dep(last_conditional_mod[i], n);
1238          }
1239       }
1240 
1241       if (inst->reads_accumulator_implicitly()) {
1242          add_dep(last_accumulator_write, n);
1243       }
1244 
1245       /* write-after-write deps. */
1246       if (inst->dst.file == VGRF) {
1247          int grf_idx = grf_index(inst->dst);
1248          for (unsigned r = 0; r < regs_written(inst); r++) {
1249             add_dep(last_grf_write[grf_idx + r], n);
1250             last_grf_write[grf_idx + r] = n;
1251          }
1252       } else if (inst->dst.file == FIXED_GRF) {
1253          if (post_reg_alloc) {
1254             for (unsigned r = 0; r < regs_written(inst); r++) {
1255                add_dep(last_grf_write[inst->dst.nr + r], n);
1256                last_grf_write[inst->dst.nr + r] = n;
1257             }
1258          } else {
1259             add_dep(last_fixed_grf_write, n);
1260             last_fixed_grf_write = n;
1261          }
1262       } else if (inst->dst.is_accumulator()) {
1263          add_dep(last_accumulator_write, n);
1264          last_accumulator_write = n;
1265       } else if (register_needs_barrier(inst->dst)) {
1266          add_barrier_deps(n);
1267       }
1268 
1269       if (const unsigned mask = inst->flags_written(s->devinfo)) {
1270          assert(mask < (1 << ARRAY_SIZE(last_conditional_mod)));
1271 
1272          for (unsigned i = 0; i < ARRAY_SIZE(last_conditional_mod); i++) {
1273             if (mask & (1 << i)) {
1274                add_dep(last_conditional_mod[i], n, 0);
1275                last_conditional_mod[i] = n;
1276             }
1277          }
1278       }
1279 
1280       if (inst->writes_accumulator_implicitly(s->devinfo) &&
1281           !inst->dst.is_accumulator()) {
1282          add_dep(last_accumulator_write, n);
1283          last_accumulator_write = n;
1284       }
1285    }
1286 
1287    clear_last_grf_write();
1288 
1289    /* bottom-to-top dependencies: WAR */
1290    memset(last_conditional_mod, 0, sizeof(last_conditional_mod));
1291    last_accumulator_write = NULL;
1292    last_fixed_grf_write = NULL;
1293 
1294    for (schedule_node *n = current.end - 1; n >= current.start; n--) {
1295       fs_inst *inst = (fs_inst *)n->inst;
1296 
1297       /* write-after-read deps. */
1298       for (int i = 0; i < inst->sources; i++) {
1299          if (inst->src[i].file == VGRF) {
1300             for (unsigned r = 0; r < regs_read(inst, i); r++)
1301                add_dep(n, last_grf_write[grf_index(inst->src[i]) + r], 0);
1302          } else if (inst->src[i].file == FIXED_GRF) {
1303             if (post_reg_alloc) {
1304                for (unsigned r = 0; r < regs_read(inst, i); r++)
1305                   add_dep(n, last_grf_write[inst->src[i].nr + r], 0);
1306             } else {
1307                add_dep(n, last_fixed_grf_write, 0);
1308             }
1309          } else if (inst->src[i].is_accumulator()) {
1310             add_dep(n, last_accumulator_write, 0);
1311          } else if (register_needs_barrier(inst->src[i])) {
1312             add_barrier_deps(n);
1313          }
1314       }
1315 
1316       if (const unsigned mask = inst->flags_read(s->devinfo)) {
1317          assert(mask < (1 << ARRAY_SIZE(last_conditional_mod)));
1318 
1319          for (unsigned i = 0; i < ARRAY_SIZE(last_conditional_mod); i++) {
1320             if (mask & (1 << i))
1321                add_dep(n, last_conditional_mod[i]);
1322          }
1323       }
1324 
1325       if (inst->reads_accumulator_implicitly()) {
1326          add_dep(n, last_accumulator_write);
1327       }
1328 
1329       /* Update the things this instruction wrote, so earlier reads
1330        * can mark this as WAR dependency.
1331        */
1332       if (inst->dst.file == VGRF) {
1333          for (unsigned r = 0; r < regs_written(inst); r++)
1334             last_grf_write[grf_index(inst->dst) + r] = n;
1335       } else if (inst->dst.file == FIXED_GRF) {
1336          if (post_reg_alloc) {
1337             for (unsigned r = 0; r < regs_written(inst); r++)
1338                last_grf_write[inst->dst.nr + r] = n;
1339          } else {
1340             last_fixed_grf_write = n;
1341          }
1342       } else if (inst->dst.is_accumulator()) {
1343          last_accumulator_write = n;
1344       } else if (register_needs_barrier(inst->dst)) {
1345          add_barrier_deps(n);
1346       }
1347 
1348       if (const unsigned mask = inst->flags_written(s->devinfo)) {
1349          assert(mask < (1 << ARRAY_SIZE(last_conditional_mod)));
1350 
1351          for (unsigned i = 0; i < ARRAY_SIZE(last_conditional_mod); i++) {
1352             if (mask & (1 << i))
1353                last_conditional_mod[i] = n;
1354          }
1355       }
1356 
1357       if (inst->writes_accumulator_implicitly(s->devinfo)) {
1358          last_accumulator_write = n;
1359       }
1360    }
1361 
1362    clear_last_grf_write();
1363 }
1364 
1365 schedule_node *
choose_instruction_to_schedule()1366 instruction_scheduler::choose_instruction_to_schedule()
1367 {
1368    schedule_node *chosen = NULL;
1369 
1370    if (mode == SCHEDULE_PRE || mode == SCHEDULE_POST) {
1371       int chosen_time = 0;
1372 
1373       /* Of the instructions ready to execute or the closest to being ready,
1374        * choose the one most likely to unblock an early program exit, or
1375        * otherwise the oldest one.
1376        */
1377       foreach_in_list(schedule_node, n, &current.available) {
1378          if (!chosen ||
1379              exit_tmp_unblocked_time(n) < exit_tmp_unblocked_time(chosen) ||
1380              (exit_tmp_unblocked_time(n) == exit_tmp_unblocked_time(chosen) &&
1381               n->tmp.unblocked_time < chosen_time)) {
1382             chosen = n;
1383             chosen_time = n->tmp.unblocked_time;
1384          }
1385       }
1386    } else {
1387       int chosen_register_pressure_benefit = 0;
1388 
1389       /* Before register allocation, we don't care about the latencies of
1390        * instructions.  All we care about is reducing live intervals of
1391        * variables so that we can avoid register spilling, or get SIMD16
1392        * shaders which naturally do a better job of hiding instruction
1393        * latency.
1394        */
1395       foreach_in_list(schedule_node, n, &current.available) {
1396          if (!chosen) {
1397             chosen = n;
1398             chosen_register_pressure_benefit =
1399                   get_register_pressure_benefit(chosen->inst);
1400             continue;
1401          }
1402 
1403          /* Most important: If we can definitely reduce register pressure, do
1404           * so immediately.
1405           */
1406          int register_pressure_benefit = get_register_pressure_benefit(n->inst);
1407 
1408          if (register_pressure_benefit > 0 &&
1409              register_pressure_benefit > chosen_register_pressure_benefit) {
1410             chosen = n;
1411             chosen_register_pressure_benefit = register_pressure_benefit;
1412             continue;
1413          } else if (chosen_register_pressure_benefit > 0 &&
1414                     (register_pressure_benefit <
1415                      chosen_register_pressure_benefit)) {
1416             continue;
1417          }
1418 
1419          if (mode == SCHEDULE_PRE_LIFO) {
1420             /* Prefer instructions that recently became available for
1421              * scheduling.  These are the things that are most likely to
1422              * (eventually) make a variable dead and reduce register pressure.
1423              * Typical register pressure estimates don't work for us because
1424              * most of our pressure comes from texturing, where no single
1425              * instruction to schedule will make a vec4 value dead.
1426              */
1427             if (n->tmp.cand_generation > chosen->tmp.cand_generation) {
1428                chosen = n;
1429                chosen_register_pressure_benefit = register_pressure_benefit;
1430                continue;
1431             } else if (n->tmp.cand_generation < chosen->tmp.cand_generation) {
1432                continue;
1433             }
1434          }
1435 
1436          /* For instructions pushed on the cands list at the same time, prefer
1437           * the one with the highest delay to the end of the program.  This is
1438           * most likely to have its values able to be consumed first (such as
1439           * for a large tree of lowered ubo loads, which appear reversed in
1440           * the instruction stream with respect to when they can be consumed).
1441           */
1442          if (n->delay > chosen->delay) {
1443             chosen = n;
1444             chosen_register_pressure_benefit = register_pressure_benefit;
1445             continue;
1446          } else if (n->delay < chosen->delay) {
1447             continue;
1448          }
1449 
1450          /* Prefer the node most likely to unblock an early program exit.
1451           */
1452          if (exit_tmp_unblocked_time(n) < exit_tmp_unblocked_time(chosen)) {
1453             chosen = n;
1454             chosen_register_pressure_benefit = register_pressure_benefit;
1455             continue;
1456          } else if (exit_tmp_unblocked_time(n) > exit_tmp_unblocked_time(chosen)) {
1457             continue;
1458          }
1459 
1460          /* If all other metrics are equal, we prefer the first instruction in
1461           * the list (program execution).
1462           */
1463       }
1464    }
1465 
1466    return chosen;
1467 }
1468 
1469 int
calculate_issue_time(const fs_inst * inst)1470 instruction_scheduler::calculate_issue_time(const fs_inst *inst)
1471 {
1472    const struct brw_isa_info *isa = &s->compiler->isa;
1473    const unsigned overhead = s->grf_used && has_bank_conflict(isa, inst) ?
1474       DIV_ROUND_UP(inst->dst.component_size(inst->exec_size), REG_SIZE) : 0;
1475    if (is_compressed(inst))
1476       return 4 + overhead;
1477    else
1478       return 2 + overhead;
1479 }
1480 
1481 void
schedule(schedule_node * chosen)1482 instruction_scheduler::schedule(schedule_node *chosen)
1483 {
1484    assert(current.scheduled < current.len);
1485    current.scheduled++;
1486 
1487    assert(chosen);
1488    chosen->remove();
1489    current.block->instructions.push_tail(chosen->inst);
1490 
1491    /* If we expected a delay for scheduling, then bump the clock to reflect
1492     * that.  In reality, the hardware will switch to another hyperthread
1493     * and may not return to dispatching our thread for a while even after
1494     * we're unblocked.  After this, we have the time when the chosen
1495     * instruction will start executing.
1496     */
1497    current.time = MAX2(current.time, chosen->tmp.unblocked_time);
1498 
1499    /* Update the clock for how soon an instruction could start after the
1500     * chosen one.
1501     */
1502    current.time += chosen->issue_time;
1503 
1504    if (debug) {
1505       fprintf(stderr, "clock %4d, scheduled: ", current.time);
1506       brw_print_instruction(*s, chosen->inst);
1507    }
1508 }
1509 
1510 void
update_children(schedule_node * chosen)1511 instruction_scheduler::update_children(schedule_node *chosen)
1512 {
1513    /* Now that we've scheduled a new instruction, some of its
1514     * children can be promoted to the list of instructions ready to
1515     * be scheduled.  Update the children's unblocked time for this
1516     * DAG edge as we do so.
1517     */
1518    for (int i = chosen->children_count - 1; i >= 0; i--) {
1519       schedule_node_child *child = &chosen->children[i];
1520 
1521       child->n->tmp.unblocked_time = MAX2(child->n->tmp.unblocked_time,
1522                                           current.time + child->effective_latency);
1523 
1524       if (debug) {
1525          fprintf(stderr, "\tchild %d, %d parents: ", i, child->n->tmp.parent_count);
1526          brw_print_instruction(*s, child->n->inst);
1527       }
1528 
1529       child->n->tmp.cand_generation = current.cand_generation;
1530       child->n->tmp.parent_count--;
1531       if (child->n->tmp.parent_count == 0) {
1532          if (debug) {
1533             fprintf(stderr, "\t\tnow available\n");
1534          }
1535          current.available.push_head(child->n);
1536       }
1537    }
1538    current.cand_generation++;
1539 }
1540 
1541 void
schedule_instructions()1542 instruction_scheduler::schedule_instructions()
1543 {
1544    if (!post_reg_alloc)
1545       reg_pressure = reg_pressure_in[current.block->num];
1546 
1547    assert(current.available.is_empty());
1548    for (schedule_node *n = current.start; n < current.end; n++) {
1549       reset_node_tmp(n);
1550 
1551       /* Add DAG heads to the list of available instructions. */
1552       if (n->tmp.parent_count == 0)
1553          current.available.push_tail(n);
1554    }
1555 
1556    current.block->instructions.make_empty();
1557 
1558    while (!current.available.is_empty()) {
1559       schedule_node *chosen = choose_instruction_to_schedule();
1560       schedule(chosen);
1561 
1562       if (!post_reg_alloc) {
1563          reg_pressure -= get_register_pressure_benefit(chosen->inst);
1564          update_register_pressure(chosen->inst);
1565          if (debug)
1566             fprintf(stderr, "(register pressure %d)\n", reg_pressure);
1567       }
1568 
1569       update_children(chosen);
1570    }
1571 }
1572 
1573 void
run(instruction_scheduler_mode mode)1574 instruction_scheduler::run(instruction_scheduler_mode mode)
1575 {
1576    this->mode = mode;
1577 
1578    if (debug && !post_reg_alloc) {
1579       fprintf(stderr, "\nInstructions before scheduling (reg_alloc %d)\n",
1580               post_reg_alloc);
1581          brw_print_instructions(*s);
1582    }
1583 
1584    if (!post_reg_alloc) {
1585       memset(reads_remaining, 0, grf_count * sizeof(*reads_remaining));
1586       memset(hw_reads_remaining, 0, hw_reg_count * sizeof(*hw_reads_remaining));
1587       memset(written, 0, grf_count * sizeof(*written));
1588    }
1589 
1590    foreach_block(block, s->cfg) {
1591       set_current_block(block);
1592 
1593       if (!post_reg_alloc) {
1594          for (schedule_node *n = current.start; n < current.end; n++)
1595             count_reads_remaining(n->inst);
1596       }
1597 
1598       schedule_instructions();
1599    }
1600 
1601    if (debug && !post_reg_alloc) {
1602       fprintf(stderr, "\nInstructions after scheduling (reg_alloc %d)\n",
1603               post_reg_alloc);
1604       brw_print_instructions(*s);
1605    }
1606 }
1607 
1608 instruction_scheduler *
brw_prepare_scheduler(fs_visitor & s,void * mem_ctx)1609 brw_prepare_scheduler(fs_visitor &s, void *mem_ctx)
1610 {
1611    const int grf_count = s.alloc.count;
1612 
1613    instruction_scheduler *empty = rzalloc(mem_ctx, instruction_scheduler);
1614    return new (empty) instruction_scheduler(mem_ctx, &s, grf_count, s.first_non_payload_grf,
1615                                             s.cfg->num_blocks, /* post_reg_alloc */ false);
1616 }
1617 
1618 void
brw_schedule_instructions_pre_ra(fs_visitor & s,instruction_scheduler * sched,instruction_scheduler_mode mode)1619 brw_schedule_instructions_pre_ra(fs_visitor &s, instruction_scheduler *sched,
1620                                  instruction_scheduler_mode mode)
1621 {
1622    if (mode == SCHEDULE_NONE)
1623       return;
1624 
1625    sched->run(mode);
1626 
1627    s.invalidate_analysis(DEPENDENCY_INSTRUCTIONS);
1628 }
1629 
1630 void
brw_schedule_instructions_post_ra(fs_visitor & s)1631 brw_schedule_instructions_post_ra(fs_visitor &s)
1632 {
1633    const bool post_reg_alloc = true;
1634    const int grf_count = reg_unit(s.devinfo) * s.grf_used;
1635 
1636    void *mem_ctx = ralloc_context(NULL);
1637 
1638    instruction_scheduler sched(mem_ctx, &s, grf_count, s.first_non_payload_grf,
1639                                s.cfg->num_blocks, post_reg_alloc);
1640    sched.run(SCHEDULE_POST);
1641 
1642    ralloc_free(mem_ctx);
1643 
1644    s.invalidate_analysis(DEPENDENCY_INSTRUCTIONS);
1645 }
1646