1 /*
2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 *
23 * Authors:
24 * Jerome Glisse
25 */
26 #include "r600_pipe.h"
27 #include "r600d.h"
28 #include "util/u_memory.h"
29 #include <errno.h>
30 #include <unistd.h>
31
32
r600_need_cs_space(struct r600_context * ctx,unsigned num_dw,boolean count_draw_in)33 void r600_need_cs_space(struct r600_context *ctx, unsigned num_dw,
34 boolean count_draw_in)
35 {
36 /* Flush the DMA IB if it's not empty. */
37 if (radeon_emitted(ctx->b.dma.cs, 0))
38 ctx->b.dma.flush(ctx, PIPE_FLUSH_ASYNC, NULL);
39
40 if (!radeon_cs_memory_below_limit(ctx->b.screen, ctx->b.gfx.cs,
41 ctx->b.vram, ctx->b.gtt)) {
42 ctx->b.gtt = 0;
43 ctx->b.vram = 0;
44 ctx->b.gfx.flush(ctx, PIPE_FLUSH_ASYNC, NULL);
45 return;
46 }
47 /* all will be accounted once relocation are emited */
48 ctx->b.gtt = 0;
49 ctx->b.vram = 0;
50
51 /* Check available space in CS. */
52 if (count_draw_in) {
53 uint64_t mask;
54
55 /* The number of dwords all the dirty states would take. */
56 mask = ctx->dirty_atoms;
57 while (mask != 0)
58 num_dw += ctx->atoms[u_bit_scan64(&mask)]->num_dw;
59
60 /* The upper-bound of how much space a draw command would take. */
61 num_dw += R600_MAX_FLUSH_CS_DWORDS + R600_MAX_DRAW_CS_DWORDS;
62 }
63
64 /* Count in r600_suspend_queries. */
65 num_dw += ctx->b.num_cs_dw_queries_suspend;
66
67 /* Count in streamout_end at the end of CS. */
68 if (ctx->b.streamout.begin_emitted) {
69 num_dw += ctx->b.streamout.num_dw_for_end;
70 }
71
72 /* SX_MISC */
73 if (ctx->b.chip_class == R600) {
74 num_dw += 3;
75 }
76
77 /* Count in framebuffer cache flushes at the end of CS. */
78 num_dw += R600_MAX_FLUSH_CS_DWORDS;
79
80 /* The fence at the end of CS. */
81 num_dw += 10;
82
83 /* Flush if there's not enough space. */
84 if (!ctx->b.ws->cs_check_space(ctx->b.gfx.cs, num_dw)) {
85 ctx->b.gfx.flush(ctx, PIPE_FLUSH_ASYNC, NULL);
86 }
87 }
88
r600_flush_emit(struct r600_context * rctx)89 void r600_flush_emit(struct r600_context *rctx)
90 {
91 struct radeon_winsys_cs *cs = rctx->b.gfx.cs;
92 unsigned cp_coher_cntl = 0;
93 unsigned wait_until = 0;
94
95 if (!rctx->b.flags) {
96 return;
97 }
98
99 /* Ensure coherency between streamout and shaders. */
100 if (rctx->b.flags & R600_CONTEXT_STREAMOUT_FLUSH)
101 rctx->b.flags |= r600_get_flush_flags(R600_COHERENCY_SHADER);
102
103 if (rctx->b.flags & R600_CONTEXT_WAIT_3D_IDLE) {
104 wait_until |= S_008040_WAIT_3D_IDLE(1);
105 }
106 if (rctx->b.flags & R600_CONTEXT_WAIT_CP_DMA_IDLE) {
107 wait_until |= S_008040_WAIT_CP_DMA_IDLE(1);
108 }
109
110 if (wait_until) {
111 /* Use of WAIT_UNTIL is deprecated on Cayman+ */
112 if (rctx->b.family >= CHIP_CAYMAN) {
113 /* emit a PS partial flush on Cayman/TN */
114 rctx->b.flags |= R600_CONTEXT_PS_PARTIAL_FLUSH;
115 }
116 }
117
118 /* Wait packets must be executed first, because SURFACE_SYNC doesn't
119 * wait for shaders if it's not flushing CB or DB.
120 */
121 if (rctx->b.flags & R600_CONTEXT_PS_PARTIAL_FLUSH) {
122 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
123 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
124 }
125
126 if (wait_until) {
127 /* Use of WAIT_UNTIL is deprecated on Cayman+ */
128 if (rctx->b.family < CHIP_CAYMAN) {
129 /* wait for things to settle */
130 radeon_set_config_reg(cs, R_008040_WAIT_UNTIL, wait_until);
131 }
132 }
133
134 if (rctx->b.chip_class >= R700 &&
135 (rctx->b.flags & R600_CONTEXT_FLUSH_AND_INV_CB_META)) {
136 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
137 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
138 }
139
140 if (rctx->b.chip_class >= R700 &&
141 (rctx->b.flags & R600_CONTEXT_FLUSH_AND_INV_DB_META)) {
142 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
143 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
144
145 /* Set FULL_CACHE_ENA for DB META flushes on r7xx and later.
146 *
147 * This hack predates use of FLUSH_AND_INV_DB_META, so it's
148 * unclear whether it's still needed or even whether it has
149 * any effect.
150 */
151 cp_coher_cntl |= S_0085F0_FULL_CACHE_ENA(1);
152 }
153
154 if (rctx->b.flags & R600_CONTEXT_FLUSH_AND_INV ||
155 (rctx->b.chip_class == R600 && rctx->b.flags & R600_CONTEXT_STREAMOUT_FLUSH)) {
156 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
157 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_EVENT) | EVENT_INDEX(0));
158 }
159
160 if (rctx->b.flags & R600_CONTEXT_INV_CONST_CACHE) {
161 /* Direct constant addressing uses the shader cache.
162 * Indirect contant addressing uses the vertex cache. */
163 cp_coher_cntl |= S_0085F0_SH_ACTION_ENA(1) |
164 (rctx->has_vertex_cache ? S_0085F0_VC_ACTION_ENA(1)
165 : S_0085F0_TC_ACTION_ENA(1));
166 }
167 if (rctx->b.flags & R600_CONTEXT_INV_VERTEX_CACHE) {
168 cp_coher_cntl |= rctx->has_vertex_cache ? S_0085F0_VC_ACTION_ENA(1)
169 : S_0085F0_TC_ACTION_ENA(1);
170 }
171 if (rctx->b.flags & R600_CONTEXT_INV_TEX_CACHE) {
172 /* Textures use the texture cache.
173 * Texture buffer objects use the vertex cache. */
174 cp_coher_cntl |= S_0085F0_TC_ACTION_ENA(1) |
175 (rctx->has_vertex_cache ? S_0085F0_VC_ACTION_ENA(1) : 0);
176 }
177
178 /* Don't use the DB CP COHER logic on r6xx.
179 * There are hw bugs.
180 */
181 if (rctx->b.chip_class >= R700 &&
182 (rctx->b.flags & R600_CONTEXT_FLUSH_AND_INV_DB)) {
183 cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) |
184 S_0085F0_DB_DEST_BASE_ENA(1) |
185 S_0085F0_SMX_ACTION_ENA(1);
186 }
187
188 /* Don't use the CB CP COHER logic on r6xx.
189 * There are hw bugs.
190 */
191 if (rctx->b.chip_class >= R700 &&
192 (rctx->b.flags & R600_CONTEXT_FLUSH_AND_INV_CB)) {
193 cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) |
194 S_0085F0_CB0_DEST_BASE_ENA(1) |
195 S_0085F0_CB1_DEST_BASE_ENA(1) |
196 S_0085F0_CB2_DEST_BASE_ENA(1) |
197 S_0085F0_CB3_DEST_BASE_ENA(1) |
198 S_0085F0_CB4_DEST_BASE_ENA(1) |
199 S_0085F0_CB5_DEST_BASE_ENA(1) |
200 S_0085F0_CB6_DEST_BASE_ENA(1) |
201 S_0085F0_CB7_DEST_BASE_ENA(1) |
202 S_0085F0_SMX_ACTION_ENA(1);
203 if (rctx->b.chip_class >= EVERGREEN)
204 cp_coher_cntl |= S_0085F0_CB8_DEST_BASE_ENA(1) |
205 S_0085F0_CB9_DEST_BASE_ENA(1) |
206 S_0085F0_CB10_DEST_BASE_ENA(1) |
207 S_0085F0_CB11_DEST_BASE_ENA(1);
208 }
209
210 if (rctx->b.chip_class >= R700 &&
211 rctx->b.flags & R600_CONTEXT_STREAMOUT_FLUSH) {
212 cp_coher_cntl |= S_0085F0_SO0_DEST_BASE_ENA(1) |
213 S_0085F0_SO1_DEST_BASE_ENA(1) |
214 S_0085F0_SO2_DEST_BASE_ENA(1) |
215 S_0085F0_SO3_DEST_BASE_ENA(1) |
216 S_0085F0_SMX_ACTION_ENA(1);
217 }
218
219 /* Workaround for buggy flushing on some R6xx chipsets. */
220 if ((rctx->b.flags & (R600_CONTEXT_FLUSH_AND_INV |
221 R600_CONTEXT_STREAMOUT_FLUSH)) &&
222 (rctx->b.family == CHIP_RV670 ||
223 rctx->b.family == CHIP_RS780 ||
224 rctx->b.family == CHIP_RS880)) {
225 cp_coher_cntl |= S_0085F0_CB1_DEST_BASE_ENA(1) |
226 S_0085F0_DEST_BASE_0_ENA(1);
227 }
228
229 if (cp_coher_cntl) {
230 radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, 0));
231 radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
232 radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
233 radeon_emit(cs, 0); /* CP_COHER_BASE */
234 radeon_emit(cs, 0x0000000A); /* POLL_INTERVAL */
235 }
236
237 if (rctx->b.flags & R600_CONTEXT_START_PIPELINE_STATS) {
238 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
239 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_PIPELINESTAT_START) |
240 EVENT_INDEX(0));
241 } else if (rctx->b.flags & R600_CONTEXT_STOP_PIPELINE_STATS) {
242 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
243 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_PIPELINESTAT_STOP) |
244 EVENT_INDEX(0));
245 }
246
247 /* everything is properly flushed */
248 rctx->b.flags = 0;
249 }
250
r600_context_gfx_flush(void * context,unsigned flags,struct pipe_fence_handle ** fence)251 void r600_context_gfx_flush(void *context, unsigned flags,
252 struct pipe_fence_handle **fence)
253 {
254 struct r600_context *ctx = context;
255 struct radeon_winsys_cs *cs = ctx->b.gfx.cs;
256 struct radeon_winsys *ws = ctx->b.ws;
257
258 if (!radeon_emitted(cs, ctx->b.initial_gfx_cs_size))
259 return;
260
261 if (r600_check_device_reset(&ctx->b))
262 return;
263
264 r600_preflush_suspend_features(&ctx->b);
265
266 /* flush the framebuffer cache */
267 ctx->b.flags |= R600_CONTEXT_FLUSH_AND_INV |
268 R600_CONTEXT_FLUSH_AND_INV_CB |
269 R600_CONTEXT_FLUSH_AND_INV_DB |
270 R600_CONTEXT_FLUSH_AND_INV_CB_META |
271 R600_CONTEXT_FLUSH_AND_INV_DB_META |
272 R600_CONTEXT_WAIT_3D_IDLE |
273 R600_CONTEXT_WAIT_CP_DMA_IDLE;
274
275 r600_flush_emit(ctx);
276
277 if (ctx->trace_buf)
278 eg_trace_emit(ctx);
279 /* old kernels and userspace don't set SX_MISC, so we must reset it to 0 here */
280 if (ctx->b.chip_class == R600) {
281 radeon_set_context_reg(cs, R_028350_SX_MISC, 0);
282 }
283
284 if (ctx->is_debug) {
285 /* Save the IB for debug contexts. */
286 radeon_clear_saved_cs(&ctx->last_gfx);
287 radeon_save_cs(ws, cs, &ctx->last_gfx, true);
288 r600_resource_reference(&ctx->last_trace_buf, ctx->trace_buf);
289 r600_resource_reference(&ctx->trace_buf, NULL);
290 }
291 /* Flush the CS. */
292 ws->cs_flush(cs, flags, &ctx->b.last_gfx_fence);
293 if (fence)
294 ws->fence_reference(fence, ctx->b.last_gfx_fence);
295 ctx->b.num_gfx_cs_flushes++;
296
297 if (ctx->is_debug) {
298 if (!ws->fence_wait(ws, ctx->b.last_gfx_fence, 10000000)) {
299 const char *fname = getenv("R600_TRACE");
300 if (!fname)
301 exit(-1);
302 FILE *fl = fopen(fname, "w+");
303 if (fl) {
304 eg_dump_debug_state(&ctx->b.b, fl, 0);
305 fclose(fl);
306 } else
307 perror(fname);
308 exit(-1);
309 }
310 }
311 r600_begin_new_cs(ctx);
312 }
313
r600_begin_new_cs(struct r600_context * ctx)314 void r600_begin_new_cs(struct r600_context *ctx)
315 {
316 unsigned shader;
317
318 if (ctx->is_debug) {
319 uint32_t zero = 0;
320
321 /* Create a buffer used for writing trace IDs and initialize it to 0. */
322 assert(!ctx->trace_buf);
323 ctx->trace_buf = (struct r600_resource*)
324 pipe_buffer_create(ctx->b.b.screen, 0,
325 PIPE_USAGE_STAGING, 4);
326 if (ctx->trace_buf)
327 pipe_buffer_write_nooverlap(&ctx->b.b, &ctx->trace_buf->b.b,
328 0, sizeof(zero), &zero);
329 ctx->trace_id = 0;
330 }
331
332 if (ctx->trace_buf)
333 eg_trace_emit(ctx);
334
335 ctx->b.flags = 0;
336 ctx->b.gtt = 0;
337 ctx->b.vram = 0;
338
339 /* Begin a new CS. */
340 r600_emit_command_buffer(ctx->b.gfx.cs, &ctx->start_cs_cmd);
341
342 /* Re-emit states. */
343 r600_mark_atom_dirty(ctx, &ctx->alphatest_state.atom);
344 r600_mark_atom_dirty(ctx, &ctx->blend_color.atom);
345 r600_mark_atom_dirty(ctx, &ctx->cb_misc_state.atom);
346 r600_mark_atom_dirty(ctx, &ctx->clip_misc_state.atom);
347 r600_mark_atom_dirty(ctx, &ctx->clip_state.atom);
348 r600_mark_atom_dirty(ctx, &ctx->db_misc_state.atom);
349 r600_mark_atom_dirty(ctx, &ctx->db_state.atom);
350 r600_mark_atom_dirty(ctx, &ctx->framebuffer.atom);
351 if (ctx->b.chip_class >= EVERGREEN) {
352 r600_mark_atom_dirty(ctx, &ctx->fragment_images.atom);
353 r600_mark_atom_dirty(ctx, &ctx->fragment_buffers.atom);
354 r600_mark_atom_dirty(ctx, &ctx->compute_images.atom);
355 r600_mark_atom_dirty(ctx, &ctx->compute_buffers.atom);
356 }
357 r600_mark_atom_dirty(ctx, &ctx->hw_shader_stages[R600_HW_STAGE_PS].atom);
358 r600_mark_atom_dirty(ctx, &ctx->poly_offset_state.atom);
359 r600_mark_atom_dirty(ctx, &ctx->vgt_state.atom);
360 r600_mark_atom_dirty(ctx, &ctx->sample_mask.atom);
361 ctx->b.scissors.dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
362 r600_mark_atom_dirty(ctx, &ctx->b.scissors.atom);
363 ctx->b.viewports.dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
364 ctx->b.viewports.depth_range_dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
365 r600_mark_atom_dirty(ctx, &ctx->b.viewports.atom);
366 if (ctx->b.chip_class <= EVERGREEN) {
367 r600_mark_atom_dirty(ctx, &ctx->config_state.atom);
368 }
369 r600_mark_atom_dirty(ctx, &ctx->stencil_ref.atom);
370 r600_mark_atom_dirty(ctx, &ctx->vertex_fetch_shader.atom);
371 r600_mark_atom_dirty(ctx, &ctx->hw_shader_stages[R600_HW_STAGE_ES].atom);
372 r600_mark_atom_dirty(ctx, &ctx->shader_stages.atom);
373 if (ctx->gs_shader) {
374 r600_mark_atom_dirty(ctx, &ctx->hw_shader_stages[R600_HW_STAGE_GS].atom);
375 r600_mark_atom_dirty(ctx, &ctx->gs_rings.atom);
376 }
377 if (ctx->tes_shader) {
378 r600_mark_atom_dirty(ctx, &ctx->hw_shader_stages[EG_HW_STAGE_HS].atom);
379 r600_mark_atom_dirty(ctx, &ctx->hw_shader_stages[EG_HW_STAGE_LS].atom);
380 }
381 r600_mark_atom_dirty(ctx, &ctx->hw_shader_stages[R600_HW_STAGE_VS].atom);
382 r600_mark_atom_dirty(ctx, &ctx->b.streamout.enable_atom);
383 r600_mark_atom_dirty(ctx, &ctx->b.render_cond_atom);
384
385 if (ctx->blend_state.cso)
386 r600_mark_atom_dirty(ctx, &ctx->blend_state.atom);
387 if (ctx->dsa_state.cso)
388 r600_mark_atom_dirty(ctx, &ctx->dsa_state.atom);
389 if (ctx->rasterizer_state.cso)
390 r600_mark_atom_dirty(ctx, &ctx->rasterizer_state.atom);
391
392 if (ctx->b.chip_class <= R700) {
393 r600_mark_atom_dirty(ctx, &ctx->seamless_cube_map.atom);
394 }
395
396 ctx->vertex_buffer_state.dirty_mask = ctx->vertex_buffer_state.enabled_mask;
397 r600_vertex_buffers_dirty(ctx);
398
399 /* Re-emit shader resources. */
400 for (shader = 0; shader < PIPE_SHADER_TYPES; shader++) {
401 struct r600_constbuf_state *constbuf = &ctx->constbuf_state[shader];
402 struct r600_textures_info *samplers = &ctx->samplers[shader];
403
404 constbuf->dirty_mask = constbuf->enabled_mask;
405 samplers->views.dirty_mask = samplers->views.enabled_mask;
406 samplers->states.dirty_mask = samplers->states.enabled_mask;
407
408 r600_constant_buffers_dirty(ctx, constbuf);
409 r600_sampler_views_dirty(ctx, &samplers->views);
410 r600_sampler_states_dirty(ctx, &samplers->states);
411 }
412
413 r600_postflush_resume_features(&ctx->b);
414
415 /* Re-emit the draw state. */
416 ctx->last_primitive_type = -1;
417 ctx->last_start_instance = -1;
418 ctx->last_rast_prim = -1;
419 ctx->current_rast_prim = -1;
420
421 assert(!ctx->b.gfx.cs->prev_dw);
422 ctx->b.initial_gfx_cs_size = ctx->b.gfx.cs->current.cdw;
423 }
424
r600_emit_pfp_sync_me(struct r600_context * rctx)425 void r600_emit_pfp_sync_me(struct r600_context *rctx)
426 {
427 struct radeon_winsys_cs *cs = rctx->b.gfx.cs;
428
429 if (rctx->b.chip_class >= EVERGREEN &&
430 rctx->b.screen->info.drm_minor >= 46) {
431 radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
432 radeon_emit(cs, 0);
433 } else {
434 /* Emulate PFP_SYNC_ME by writing a value to memory in ME and
435 * waiting for it in PFP.
436 */
437 struct r600_resource *buf = NULL;
438 unsigned offset, reloc;
439 uint64_t va;
440
441 /* 16-byte address alignment is required by WAIT_REG_MEM. */
442 u_suballocator_alloc(rctx->b.allocator_zeroed_memory, 4, 16,
443 &offset, (struct pipe_resource**)&buf);
444 if (!buf) {
445 /* This is too heavyweight, but will work. */
446 rctx->b.gfx.flush(rctx, PIPE_FLUSH_ASYNC, NULL);
447 return;
448 }
449
450 reloc = radeon_add_to_buffer_list(&rctx->b, &rctx->b.gfx, buf,
451 RADEON_USAGE_READWRITE,
452 RADEON_PRIO_FENCE);
453
454 va = buf->gpu_address + offset;
455 assert(va % 16 == 0);
456
457 /* Write 1 to memory in ME. */
458 radeon_emit(cs, PKT3(PKT3_MEM_WRITE, 3, 0));
459 radeon_emit(cs, va);
460 radeon_emit(cs, ((va >> 32) & 0xff) | MEM_WRITE_32_BITS);
461 radeon_emit(cs, 1);
462 radeon_emit(cs, 0);
463
464 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
465 radeon_emit(cs, reloc);
466
467 /* Wait in PFP (PFP can only do GEQUAL against memory). */
468 radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, 0));
469 radeon_emit(cs, WAIT_REG_MEM_GEQUAL |
470 WAIT_REG_MEM_MEMORY |
471 WAIT_REG_MEM_PFP);
472 radeon_emit(cs, va);
473 radeon_emit(cs, va >> 32);
474 radeon_emit(cs, 1); /* reference value */
475 radeon_emit(cs, 0xffffffff); /* mask */
476 radeon_emit(cs, 4); /* poll interval */
477
478 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
479 radeon_emit(cs, reloc);
480
481 r600_resource_reference(&buf, NULL);
482 }
483 }
484
485 /* The max number of bytes to copy per packet. */
486 #define CP_DMA_MAX_BYTE_COUNT ((1 << 21) - 8)
487
r600_cp_dma_copy_buffer(struct r600_context * rctx,struct pipe_resource * dst,uint64_t dst_offset,struct pipe_resource * src,uint64_t src_offset,unsigned size)488 void r600_cp_dma_copy_buffer(struct r600_context *rctx,
489 struct pipe_resource *dst, uint64_t dst_offset,
490 struct pipe_resource *src, uint64_t src_offset,
491 unsigned size)
492 {
493 struct radeon_winsys_cs *cs = rctx->b.gfx.cs;
494
495 assert(size);
496 assert(rctx->screen->b.has_cp_dma);
497
498 /* Mark the buffer range of destination as valid (initialized),
499 * so that transfer_map knows it should wait for the GPU when mapping
500 * that range. */
501 util_range_add(&r600_resource(dst)->valid_buffer_range, dst_offset,
502 dst_offset + size);
503
504 dst_offset += r600_resource(dst)->gpu_address;
505 src_offset += r600_resource(src)->gpu_address;
506
507 /* Flush the caches where the resources are bound. */
508 rctx->b.flags |= r600_get_flush_flags(R600_COHERENCY_SHADER) |
509 R600_CONTEXT_WAIT_3D_IDLE;
510
511 /* There are differences between R700 and EG in CP DMA,
512 * but we only use the common bits here. */
513 while (size) {
514 unsigned sync = 0;
515 unsigned byte_count = MIN2(size, CP_DMA_MAX_BYTE_COUNT);
516 unsigned src_reloc, dst_reloc;
517
518 r600_need_cs_space(rctx,
519 10 + (rctx->b.flags ? R600_MAX_FLUSH_CS_DWORDS : 0) +
520 3 + R600_MAX_PFP_SYNC_ME_DWORDS, FALSE);
521
522 /* Flush the caches for the first copy only. */
523 if (rctx->b.flags) {
524 r600_flush_emit(rctx);
525 }
526
527 /* Do the synchronization after the last copy, so that all data is written to memory. */
528 if (size == byte_count) {
529 sync = PKT3_CP_DMA_CP_SYNC;
530 }
531
532 /* This must be done after r600_need_cs_space. */
533 src_reloc = radeon_add_to_buffer_list(&rctx->b, &rctx->b.gfx, (struct r600_resource*)src,
534 RADEON_USAGE_READ, RADEON_PRIO_CP_DMA);
535 dst_reloc = radeon_add_to_buffer_list(&rctx->b, &rctx->b.gfx, (struct r600_resource*)dst,
536 RADEON_USAGE_WRITE, RADEON_PRIO_CP_DMA);
537
538 radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, 0));
539 radeon_emit(cs, src_offset); /* SRC_ADDR_LO [31:0] */
540 radeon_emit(cs, sync | ((src_offset >> 32) & 0xff)); /* CP_SYNC [31] | SRC_ADDR_HI [7:0] */
541 radeon_emit(cs, dst_offset); /* DST_ADDR_LO [31:0] */
542 radeon_emit(cs, (dst_offset >> 32) & 0xff); /* DST_ADDR_HI [7:0] */
543 radeon_emit(cs, byte_count); /* COMMAND [29:22] | BYTE_COUNT [20:0] */
544
545 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
546 radeon_emit(cs, src_reloc);
547 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
548 radeon_emit(cs, dst_reloc);
549
550 size -= byte_count;
551 src_offset += byte_count;
552 dst_offset += byte_count;
553 }
554
555 /* CP_DMA_CP_SYNC doesn't wait for idle on R6xx, but this does. */
556 if (rctx->b.chip_class == R600)
557 radeon_set_config_reg(cs, R_008040_WAIT_UNTIL,
558 S_008040_WAIT_CP_DMA_IDLE(1));
559
560 /* CP DMA is executed in ME, but index buffers are read by PFP.
561 * This ensures that ME (CP DMA) is idle before PFP starts fetching
562 * indices. If we wanted to execute CP DMA in PFP, this packet
563 * should precede it.
564 */
565 r600_emit_pfp_sync_me(rctx);
566 }
567
r600_dma_copy_buffer(struct r600_context * rctx,struct pipe_resource * dst,struct pipe_resource * src,uint64_t dst_offset,uint64_t src_offset,uint64_t size)568 void r600_dma_copy_buffer(struct r600_context *rctx,
569 struct pipe_resource *dst,
570 struct pipe_resource *src,
571 uint64_t dst_offset,
572 uint64_t src_offset,
573 uint64_t size)
574 {
575 struct radeon_winsys_cs *cs = rctx->b.dma.cs;
576 unsigned i, ncopy, csize;
577 struct r600_resource *rdst = (struct r600_resource*)dst;
578 struct r600_resource *rsrc = (struct r600_resource*)src;
579
580 /* Mark the buffer range of destination as valid (initialized),
581 * so that transfer_map knows it should wait for the GPU when mapping
582 * that range. */
583 util_range_add(&rdst->valid_buffer_range, dst_offset,
584 dst_offset + size);
585
586 size >>= 2; /* convert to dwords */
587 ncopy = (size / R600_DMA_COPY_MAX_SIZE_DW) + !!(size % R600_DMA_COPY_MAX_SIZE_DW);
588
589 r600_need_dma_space(&rctx->b, ncopy * 5, rdst, rsrc);
590 for (i = 0; i < ncopy; i++) {
591 csize = size < R600_DMA_COPY_MAX_SIZE_DW ? size : R600_DMA_COPY_MAX_SIZE_DW;
592 /* emit reloc before writing cs so that cs is always in consistent state */
593 radeon_add_to_buffer_list(&rctx->b, &rctx->b.dma, rsrc, RADEON_USAGE_READ,
594 RADEON_PRIO_SDMA_BUFFER);
595 radeon_add_to_buffer_list(&rctx->b, &rctx->b.dma, rdst, RADEON_USAGE_WRITE,
596 RADEON_PRIO_SDMA_BUFFER);
597 radeon_emit(cs, DMA_PACKET(DMA_PACKET_COPY, 0, 0, csize));
598 radeon_emit(cs, dst_offset & 0xfffffffc);
599 radeon_emit(cs, src_offset & 0xfffffffc);
600 radeon_emit(cs, (dst_offset >> 32UL) & 0xff);
601 radeon_emit(cs, (src_offset >> 32UL) & 0xff);
602 dst_offset += csize << 2;
603 src_offset += csize << 2;
604 size -= csize;
605 }
606 }
607