1 // SPDX-License-Identifier: MIT
2 /*
3 * Copyright © 2022 Intel Corporation
4 */
5
6 #include "xe_ring_ops.h"
7
8 #include <generated/xe_wa_oob.h>
9
10 #include "instructions/xe_gpu_commands.h"
11 #include "instructions/xe_mi_commands.h"
12 #include "regs/xe_engine_regs.h"
13 #include "regs/xe_gt_regs.h"
14 #include "regs/xe_lrc_layout.h"
15 #include "xe_exec_queue_types.h"
16 #include "xe_gt.h"
17 #include "xe_lrc.h"
18 #include "xe_macros.h"
19 #include "xe_sched_job.h"
20 #include "xe_sriov.h"
21 #include "xe_vm_types.h"
22 #include "xe_vm.h"
23 #include "xe_wa.h"
24
25 /*
26 * 3D-related flags that can't be set on _engines_ that lack access to the 3D
27 * pipeline (i.e., CCS engines).
28 */
29 #define PIPE_CONTROL_3D_ENGINE_FLAGS (\
30 PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH | \
31 PIPE_CONTROL_DEPTH_CACHE_FLUSH | \
32 PIPE_CONTROL_TILE_CACHE_FLUSH | \
33 PIPE_CONTROL_DEPTH_STALL | \
34 PIPE_CONTROL_STALL_AT_SCOREBOARD | \
35 PIPE_CONTROL_PSD_SYNC | \
36 PIPE_CONTROL_AMFS_FLUSH | \
37 PIPE_CONTROL_VF_CACHE_INVALIDATE | \
38 PIPE_CONTROL_GLOBAL_SNAPSHOT_RESET)
39
40 /* 3D-related flags that can't be set on _platforms_ that lack a 3D pipeline */
41 #define PIPE_CONTROL_3D_ARCH_FLAGS ( \
42 PIPE_CONTROL_3D_ENGINE_FLAGS | \
43 PIPE_CONTROL_INDIRECT_STATE_DISABLE | \
44 PIPE_CONTROL_FLUSH_ENABLE | \
45 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE | \
46 PIPE_CONTROL_DC_FLUSH_ENABLE)
47
preparser_disable(bool state)48 static u32 preparser_disable(bool state)
49 {
50 return MI_ARB_CHECK | BIT(8) | state;
51 }
52
emit_aux_table_inv(struct xe_gt * gt,struct xe_reg reg,u32 * dw,int i)53 static int emit_aux_table_inv(struct xe_gt *gt, struct xe_reg reg,
54 u32 *dw, int i)
55 {
56 dw[i++] = MI_LOAD_REGISTER_IMM | MI_LRI_NUM_REGS(1) | MI_LRI_MMIO_REMAP_EN;
57 dw[i++] = reg.addr + gt->mmio.adj_offset;
58 dw[i++] = AUX_INV;
59 dw[i++] = MI_NOOP;
60
61 return i;
62 }
63
emit_user_interrupt(u32 * dw,int i)64 static int emit_user_interrupt(u32 *dw, int i)
65 {
66 dw[i++] = MI_USER_INTERRUPT;
67 dw[i++] = MI_ARB_ON_OFF | MI_ARB_ENABLE;
68 dw[i++] = MI_ARB_CHECK;
69
70 return i;
71 }
72
emit_store_imm_ggtt(u32 addr,u32 value,u32 * dw,int i)73 static int emit_store_imm_ggtt(u32 addr, u32 value, u32 *dw, int i)
74 {
75 dw[i++] = MI_STORE_DATA_IMM | MI_SDI_GGTT | MI_SDI_NUM_DW(1);
76 dw[i++] = addr;
77 dw[i++] = 0;
78 dw[i++] = value;
79
80 return i;
81 }
82
emit_flush_dw(u32 * dw,int i)83 static int emit_flush_dw(u32 *dw, int i)
84 {
85 dw[i++] = MI_FLUSH_DW | MI_FLUSH_IMM_DW;
86 dw[i++] = 0;
87 dw[i++] = 0;
88 dw[i++] = 0;
89
90 return i;
91 }
92
emit_flush_imm_ggtt(u32 addr,u32 value,bool invalidate_tlb,u32 * dw,int i)93 static int emit_flush_imm_ggtt(u32 addr, u32 value, bool invalidate_tlb,
94 u32 *dw, int i)
95 {
96 dw[i++] = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_IMM_DW |
97 (invalidate_tlb ? MI_INVALIDATE_TLB : 0);
98 dw[i++] = addr | MI_FLUSH_DW_USE_GTT;
99 dw[i++] = 0;
100 dw[i++] = value;
101
102 return i;
103 }
104
emit_bb_start(u64 batch_addr,u32 ppgtt_flag,u32 * dw,int i)105 static int emit_bb_start(u64 batch_addr, u32 ppgtt_flag, u32 *dw, int i)
106 {
107 dw[i++] = MI_BATCH_BUFFER_START | ppgtt_flag | XE_INSTR_NUM_DW(3);
108 dw[i++] = lower_32_bits(batch_addr);
109 dw[i++] = upper_32_bits(batch_addr);
110
111 return i;
112 }
113
emit_flush_invalidate(u32 flag,u32 * dw,int i)114 static int emit_flush_invalidate(u32 flag, u32 *dw, int i)
115 {
116 dw[i] = MI_FLUSH_DW;
117 dw[i] |= flag;
118 dw[i++] |= MI_INVALIDATE_TLB | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_IMM_DW |
119 MI_FLUSH_DW_STORE_INDEX;
120
121 dw[i++] = LRC_PPHWSP_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT;
122 dw[i++] = 0;
123 dw[i++] = ~0U;
124
125 return i;
126 }
127
128 static int
emit_pipe_control(u32 * dw,int i,u32 bit_group_0,u32 bit_group_1,u32 offset,u32 value)129 emit_pipe_control(u32 *dw, int i, u32 bit_group_0, u32 bit_group_1, u32 offset, u32 value)
130 {
131 dw[i++] = GFX_OP_PIPE_CONTROL(6) | bit_group_0;
132 dw[i++] = bit_group_1;
133 dw[i++] = offset;
134 dw[i++] = 0;
135 dw[i++] = value;
136 dw[i++] = 0;
137
138 return i;
139 }
140
emit_pipe_invalidate(u32 mask_flags,bool invalidate_tlb,u32 * dw,int i)141 static int emit_pipe_invalidate(u32 mask_flags, bool invalidate_tlb, u32 *dw,
142 int i)
143 {
144 u32 flags = PIPE_CONTROL_CS_STALL |
145 PIPE_CONTROL_COMMAND_CACHE_INVALIDATE |
146 PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE |
147 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
148 PIPE_CONTROL_VF_CACHE_INVALIDATE |
149 PIPE_CONTROL_CONST_CACHE_INVALIDATE |
150 PIPE_CONTROL_STATE_CACHE_INVALIDATE |
151 PIPE_CONTROL_QW_WRITE |
152 PIPE_CONTROL_STORE_DATA_INDEX;
153
154 if (invalidate_tlb)
155 flags |= PIPE_CONTROL_TLB_INVALIDATE;
156
157 flags &= ~mask_flags;
158
159 return emit_pipe_control(dw, i, 0, flags, LRC_PPHWSP_SCRATCH_ADDR, 0);
160 }
161
emit_store_imm_ppgtt_posted(u64 addr,u64 value,u32 * dw,int i)162 static int emit_store_imm_ppgtt_posted(u64 addr, u64 value,
163 u32 *dw, int i)
164 {
165 dw[i++] = MI_STORE_DATA_IMM | MI_SDI_NUM_QW(1);
166 dw[i++] = lower_32_bits(addr);
167 dw[i++] = upper_32_bits(addr);
168 dw[i++] = lower_32_bits(value);
169 dw[i++] = upper_32_bits(value);
170
171 return i;
172 }
173
emit_render_cache_flush(struct xe_sched_job * job,u32 * dw,int i)174 static int emit_render_cache_flush(struct xe_sched_job *job, u32 *dw, int i)
175 {
176 struct xe_gt *gt = job->q->gt;
177 bool lacks_render = !(gt->info.engine_mask & XE_HW_ENGINE_RCS_MASK);
178 u32 flags;
179
180 flags = (PIPE_CONTROL_CS_STALL |
181 PIPE_CONTROL_TILE_CACHE_FLUSH |
182 PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
183 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
184 PIPE_CONTROL_DC_FLUSH_ENABLE |
185 PIPE_CONTROL_FLUSH_ENABLE);
186
187 if (XE_WA(gt, 1409600907))
188 flags |= PIPE_CONTROL_DEPTH_STALL;
189
190 if (lacks_render)
191 flags &= ~PIPE_CONTROL_3D_ARCH_FLAGS;
192 else if (job->q->class == XE_ENGINE_CLASS_COMPUTE)
193 flags &= ~PIPE_CONTROL_3D_ENGINE_FLAGS;
194
195 return emit_pipe_control(dw, i, PIPE_CONTROL0_HDC_PIPELINE_FLUSH, flags, 0, 0);
196 }
197
emit_pipe_control_to_ring_end(struct xe_hw_engine * hwe,u32 * dw,int i)198 static int emit_pipe_control_to_ring_end(struct xe_hw_engine *hwe, u32 *dw, int i)
199 {
200 if (hwe->class != XE_ENGINE_CLASS_RENDER)
201 return i;
202
203 if (XE_WA(hwe->gt, 16020292621))
204 i = emit_pipe_control(dw, i, 0, PIPE_CONTROL_LRI_POST_SYNC,
205 RING_NOPID(hwe->mmio_base).addr, 0);
206
207 return i;
208 }
209
emit_pipe_imm_ggtt(u32 addr,u32 value,bool stall_only,u32 * dw,int i)210 static int emit_pipe_imm_ggtt(u32 addr, u32 value, bool stall_only, u32 *dw,
211 int i)
212 {
213 u32 flags = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_GLOBAL_GTT_IVB |
214 PIPE_CONTROL_QW_WRITE;
215
216 if (!stall_only)
217 flags |= PIPE_CONTROL_FLUSH_ENABLE;
218
219 return emit_pipe_control(dw, i, 0, flags, addr, value);
220 }
221
get_ppgtt_flag(struct xe_sched_job * job)222 static u32 get_ppgtt_flag(struct xe_sched_job *job)
223 {
224 if (job->q->vm && !job->ggtt)
225 return BIT(8);
226
227 return 0;
228 }
229
emit_copy_timestamp(struct xe_lrc * lrc,u32 * dw,int i)230 static int emit_copy_timestamp(struct xe_lrc *lrc, u32 *dw, int i)
231 {
232 dw[i++] = MI_COPY_MEM_MEM | MI_COPY_MEM_MEM_SRC_GGTT |
233 MI_COPY_MEM_MEM_DST_GGTT;
234 dw[i++] = xe_lrc_ctx_job_timestamp_ggtt_addr(lrc);
235 dw[i++] = 0;
236 dw[i++] = xe_lrc_ctx_timestamp_ggtt_addr(lrc);
237 dw[i++] = 0;
238 dw[i++] = MI_NOOP;
239
240 return i;
241 }
242
243 /* for engines that don't require any special HW handling (no EUs, no aux inval, etc) */
__emit_job_gen12_simple(struct xe_sched_job * job,struct xe_lrc * lrc,u64 batch_addr,u32 seqno)244 static void __emit_job_gen12_simple(struct xe_sched_job *job, struct xe_lrc *lrc,
245 u64 batch_addr, u32 seqno)
246 {
247 u32 dw[MAX_JOB_SIZE_DW], i = 0;
248 u32 ppgtt_flag = get_ppgtt_flag(job);
249 struct xe_gt *gt = job->q->gt;
250
251 i = emit_copy_timestamp(lrc, dw, i);
252
253 if (job->ring_ops_flush_tlb) {
254 dw[i++] = preparser_disable(true);
255 i = emit_flush_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
256 seqno, true, dw, i);
257 dw[i++] = preparser_disable(false);
258 } else {
259 i = emit_store_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
260 seqno, dw, i);
261 }
262
263 i = emit_bb_start(batch_addr, ppgtt_flag, dw, i);
264
265 if (job->user_fence.used) {
266 i = emit_flush_dw(dw, i);
267 i = emit_store_imm_ppgtt_posted(job->user_fence.addr,
268 job->user_fence.value,
269 dw, i);
270 }
271
272 i = emit_flush_imm_ggtt(xe_lrc_seqno_ggtt_addr(lrc), seqno, false, dw, i);
273
274 i = emit_user_interrupt(dw, i);
275
276 xe_gt_assert(gt, i <= MAX_JOB_SIZE_DW);
277
278 xe_lrc_write_ring(lrc, dw, i * sizeof(*dw));
279 }
280
has_aux_ccs(struct xe_device * xe)281 static bool has_aux_ccs(struct xe_device *xe)
282 {
283 /*
284 * PVC is a special case that has no compression of either type
285 * (FlatCCS or AuxCCS). Also, AuxCCS is no longer used from Xe2
286 * onward, so any future platforms with no FlatCCS will not have
287 * AuxCCS either.
288 */
289 if (GRAPHICS_VER(xe) >= 20 || xe->info.platform == XE_PVC)
290 return false;
291
292 return !xe->info.has_flat_ccs;
293 }
294
__emit_job_gen12_video(struct xe_sched_job * job,struct xe_lrc * lrc,u64 batch_addr,u32 seqno)295 static void __emit_job_gen12_video(struct xe_sched_job *job, struct xe_lrc *lrc,
296 u64 batch_addr, u32 seqno)
297 {
298 u32 dw[MAX_JOB_SIZE_DW], i = 0;
299 u32 ppgtt_flag = get_ppgtt_flag(job);
300 struct xe_gt *gt = job->q->gt;
301 struct xe_device *xe = gt_to_xe(gt);
302 bool decode = job->q->class == XE_ENGINE_CLASS_VIDEO_DECODE;
303
304 i = emit_copy_timestamp(lrc, dw, i);
305
306 dw[i++] = preparser_disable(true);
307
308 /* hsdes: 1809175790 */
309 if (has_aux_ccs(xe)) {
310 if (decode)
311 i = emit_aux_table_inv(gt, VD0_AUX_INV, dw, i);
312 else
313 i = emit_aux_table_inv(gt, VE0_AUX_INV, dw, i);
314 }
315
316 if (job->ring_ops_flush_tlb)
317 i = emit_flush_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
318 seqno, true, dw, i);
319
320 dw[i++] = preparser_disable(false);
321
322 if (!job->ring_ops_flush_tlb)
323 i = emit_store_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
324 seqno, dw, i);
325
326 i = emit_bb_start(batch_addr, ppgtt_flag, dw, i);
327
328 if (job->user_fence.used) {
329 i = emit_flush_dw(dw, i);
330 i = emit_store_imm_ppgtt_posted(job->user_fence.addr,
331 job->user_fence.value,
332 dw, i);
333 }
334
335 i = emit_flush_imm_ggtt(xe_lrc_seqno_ggtt_addr(lrc), seqno, false, dw, i);
336
337 i = emit_user_interrupt(dw, i);
338
339 xe_gt_assert(gt, i <= MAX_JOB_SIZE_DW);
340
341 xe_lrc_write_ring(lrc, dw, i * sizeof(*dw));
342 }
343
__emit_job_gen12_render_compute(struct xe_sched_job * job,struct xe_lrc * lrc,u64 batch_addr,u32 seqno)344 static void __emit_job_gen12_render_compute(struct xe_sched_job *job,
345 struct xe_lrc *lrc,
346 u64 batch_addr, u32 seqno)
347 {
348 u32 dw[MAX_JOB_SIZE_DW], i = 0;
349 u32 ppgtt_flag = get_ppgtt_flag(job);
350 struct xe_gt *gt = job->q->gt;
351 struct xe_device *xe = gt_to_xe(gt);
352 bool lacks_render = !(gt->info.engine_mask & XE_HW_ENGINE_RCS_MASK);
353 u32 mask_flags = 0;
354
355 i = emit_copy_timestamp(lrc, dw, i);
356
357 dw[i++] = preparser_disable(true);
358 if (lacks_render)
359 mask_flags = PIPE_CONTROL_3D_ARCH_FLAGS;
360 else if (job->q->class == XE_ENGINE_CLASS_COMPUTE)
361 mask_flags = PIPE_CONTROL_3D_ENGINE_FLAGS;
362
363 /* See __xe_pt_bind_vma() for a discussion on TLB invalidations. */
364 i = emit_pipe_invalidate(mask_flags, job->ring_ops_flush_tlb, dw, i);
365
366 /* hsdes: 1809175790 */
367 if (has_aux_ccs(xe))
368 i = emit_aux_table_inv(gt, CCS_AUX_INV, dw, i);
369
370 dw[i++] = preparser_disable(false);
371
372 i = emit_store_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
373 seqno, dw, i);
374
375 i = emit_bb_start(batch_addr, ppgtt_flag, dw, i);
376
377 i = emit_render_cache_flush(job, dw, i);
378
379 if (job->user_fence.used)
380 i = emit_store_imm_ppgtt_posted(job->user_fence.addr,
381 job->user_fence.value,
382 dw, i);
383
384 i = emit_pipe_imm_ggtt(xe_lrc_seqno_ggtt_addr(lrc), seqno, lacks_render, dw, i);
385
386 i = emit_user_interrupt(dw, i);
387
388 i = emit_pipe_control_to_ring_end(job->q->hwe, dw, i);
389
390 xe_gt_assert(gt, i <= MAX_JOB_SIZE_DW);
391
392 xe_lrc_write_ring(lrc, dw, i * sizeof(*dw));
393 }
394
emit_migration_job_gen12(struct xe_sched_job * job,struct xe_lrc * lrc,u32 seqno)395 static void emit_migration_job_gen12(struct xe_sched_job *job,
396 struct xe_lrc *lrc, u32 seqno)
397 {
398 u32 dw[MAX_JOB_SIZE_DW], i = 0;
399
400 i = emit_copy_timestamp(lrc, dw, i);
401
402 i = emit_store_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
403 seqno, dw, i);
404
405 dw[i++] = MI_ARB_ON_OFF | MI_ARB_DISABLE; /* Enabled again below */
406
407 i = emit_bb_start(job->ptrs[0].batch_addr, BIT(8), dw, i);
408
409 if (!IS_SRIOV_VF(gt_to_xe(job->q->gt))) {
410 /* XXX: Do we need this? Leaving for now. */
411 dw[i++] = preparser_disable(true);
412 i = emit_flush_invalidate(0, dw, i);
413 dw[i++] = preparser_disable(false);
414 }
415
416 i = emit_bb_start(job->ptrs[1].batch_addr, BIT(8), dw, i);
417
418 dw[i++] = MI_FLUSH_DW | MI_INVALIDATE_TLB | job->migrate_flush_flags |
419 MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_IMM_DW;
420 dw[i++] = xe_lrc_seqno_ggtt_addr(lrc) | MI_FLUSH_DW_USE_GTT;
421 dw[i++] = 0;
422 dw[i++] = seqno; /* value */
423
424 i = emit_user_interrupt(dw, i);
425
426 xe_gt_assert(job->q->gt, i <= MAX_JOB_SIZE_DW);
427
428 xe_lrc_write_ring(lrc, dw, i * sizeof(*dw));
429 }
430
emit_job_gen12_gsc(struct xe_sched_job * job)431 static void emit_job_gen12_gsc(struct xe_sched_job *job)
432 {
433 struct xe_gt *gt = job->q->gt;
434
435 xe_gt_assert(gt, job->q->width <= 1); /* no parallel submission for GSCCS */
436
437 __emit_job_gen12_simple(job, job->q->lrc[0],
438 job->ptrs[0].batch_addr,
439 xe_sched_job_lrc_seqno(job));
440 }
441
emit_job_gen12_copy(struct xe_sched_job * job)442 static void emit_job_gen12_copy(struct xe_sched_job *job)
443 {
444 int i;
445
446 if (xe_sched_job_is_migration(job->q)) {
447 emit_migration_job_gen12(job, job->q->lrc[0],
448 xe_sched_job_lrc_seqno(job));
449 return;
450 }
451
452 for (i = 0; i < job->q->width; ++i)
453 __emit_job_gen12_simple(job, job->q->lrc[i],
454 job->ptrs[i].batch_addr,
455 xe_sched_job_lrc_seqno(job));
456 }
457
emit_job_gen12_video(struct xe_sched_job * job)458 static void emit_job_gen12_video(struct xe_sched_job *job)
459 {
460 int i;
461
462 /* FIXME: Not doing parallel handshake for now */
463 for (i = 0; i < job->q->width; ++i)
464 __emit_job_gen12_video(job, job->q->lrc[i],
465 job->ptrs[i].batch_addr,
466 xe_sched_job_lrc_seqno(job));
467 }
468
emit_job_gen12_render_compute(struct xe_sched_job * job)469 static void emit_job_gen12_render_compute(struct xe_sched_job *job)
470 {
471 int i;
472
473 for (i = 0; i < job->q->width; ++i)
474 __emit_job_gen12_render_compute(job, job->q->lrc[i],
475 job->ptrs[i].batch_addr,
476 xe_sched_job_lrc_seqno(job));
477 }
478
479 static const struct xe_ring_ops ring_ops_gen12_gsc = {
480 .emit_job = emit_job_gen12_gsc,
481 };
482
483 static const struct xe_ring_ops ring_ops_gen12_copy = {
484 .emit_job = emit_job_gen12_copy,
485 };
486
487 static const struct xe_ring_ops ring_ops_gen12_video = {
488 .emit_job = emit_job_gen12_video,
489 };
490
491 static const struct xe_ring_ops ring_ops_gen12_render_compute = {
492 .emit_job = emit_job_gen12_render_compute,
493 };
494
495 const struct xe_ring_ops *
xe_ring_ops_get(struct xe_gt * gt,enum xe_engine_class class)496 xe_ring_ops_get(struct xe_gt *gt, enum xe_engine_class class)
497 {
498 switch (class) {
499 case XE_ENGINE_CLASS_OTHER:
500 return &ring_ops_gen12_gsc;
501 case XE_ENGINE_CLASS_COPY:
502 return &ring_ops_gen12_copy;
503 case XE_ENGINE_CLASS_VIDEO_DECODE:
504 case XE_ENGINE_CLASS_VIDEO_ENHANCE:
505 return &ring_ops_gen12_video;
506 case XE_ENGINE_CLASS_RENDER:
507 case XE_ENGINE_CLASS_COMPUTE:
508 return &ring_ops_gen12_render_compute;
509 default:
510 return NULL;
511 }
512 }
513