freedreno/a5xx: fix batch leak in fd5 blitter path
[mesa.git] / src / gallium / drivers / freedreno / a5xx / fd5_blitter.c
1 /*
2 * Copyright (C) 2017 Rob Clark <robclark@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 * 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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 * Rob Clark <robclark@freedesktop.org>
25 */
26
27 #include "freedreno_blitter.h"
28 #include "freedreno_resource.h"
29
30 #include "fd5_blitter.h"
31 #include "fd5_format.h"
32 #include "fd5_emit.h"
33
34 /* Make sure none of the requested dimensions extend beyond the size of the
35 * resource. Not entirely sure why this happens, but sometimes it does, and
36 * w/ 2d blt doesn't have wrap modes like a sampler, so force those cases
37 * back to u_blitter
38 */
39 static bool
40 ok_dims(const struct pipe_resource *r, const struct pipe_box *b, int lvl)
41 {
42 return (b->x >= 0) && (b->x + b->width <= u_minify(r->width0, lvl)) &&
43 (b->y >= 0) && (b->y + b->height <= u_minify(r->height0, lvl)) &&
44 (b->z >= 0) && (b->z + b->depth <= u_minify(r->depth0, lvl));
45 }
46
47 /* Not sure if format restrictions differ for src and dst, or if
48 * they only matter when src fmt != dst fmt.. but there appear to
49 * be *some* limitations so let's just start blacklisting stuff that
50 * piglit complains about
51 */
52 static bool
53 ok_format(enum pipe_format fmt)
54 {
55 if (util_format_is_compressed(fmt))
56 return false;
57
58 switch (fmt) {
59 case PIPE_FORMAT_R10G10B10A2_SSCALED:
60 case PIPE_FORMAT_R10G10B10A2_SNORM:
61 case PIPE_FORMAT_B10G10R10A2_USCALED:
62 case PIPE_FORMAT_B10G10R10A2_SSCALED:
63 case PIPE_FORMAT_B10G10R10A2_SNORM:
64 case PIPE_FORMAT_R10G10B10A2_UNORM:
65 case PIPE_FORMAT_R10G10B10A2_USCALED:
66 case PIPE_FORMAT_B10G10R10A2_UNORM:
67 case PIPE_FORMAT_R10SG10SB10SA2U_NORM:
68 case PIPE_FORMAT_B10G10R10A2_UINT:
69 case PIPE_FORMAT_R10G10B10A2_UINT:
70 return false;
71 default:
72 break;
73 }
74
75 if (fd5_pipe2color(fmt) == ~0)
76 return false;
77
78 return true;
79 }
80
81 static bool
82 can_do_blit(const struct pipe_blit_info *info)
83 {
84 /* I think we can do scaling, but not in z dimension since that would
85 * require blending..
86 */
87 if (info->dst.box.depth != info->src.box.depth)
88 return false;
89
90 if (!ok_format(info->dst.format))
91 return false;
92
93 if (!ok_format(info->src.format))
94 return false;
95
96 /* hw ignores {SRC,DST}_INFO.COLOR_SWAP if {SRC,DST}_INFO.TILE_MODE
97 * is set (not linear). We can kind of get around that when tiling/
98 * untiling by setting both src and dst COLOR_SWAP=WZYX, but that
99 * means the formats must match:
100 */
101 if ((fd_resource(info->dst.resource)->tile_mode ||
102 fd_resource(info->src.resource)->tile_mode) &&
103 info->dst.format != info->src.format)
104 return false;
105
106 /* until we figure out a few more registers: */
107 if ((info->dst.box.width != info->src.box.width) ||
108 (info->dst.box.height != info->src.box.height))
109 return false;
110
111 /* src box can be inverted, which we don't support.. dst box cannot: */
112 if ((info->src.box.width < 0) || (info->src.box.height < 0))
113 return false;
114
115 if (!ok_dims(info->src.resource, &info->src.box, info->src.level))
116 return false;
117
118 if (!ok_dims(info->dst.resource, &info->dst.box, info->dst.level))
119 return false;
120
121 debug_assert(info->dst.box.width >= 0);
122 debug_assert(info->dst.box.height >= 0);
123 debug_assert(info->dst.box.depth >= 0);
124
125 if ((info->dst.resource->nr_samples > 1) ||
126 (info->src.resource->nr_samples > 1))
127 return false;
128
129 if (info->scissor_enable)
130 return false;
131
132 if (info->window_rectangle_include)
133 return false;
134
135 if (info->render_condition_enable)
136 return false;
137
138 if (info->alpha_blend)
139 return false;
140
141 if (info->filter != PIPE_TEX_FILTER_NEAREST)
142 return false;
143
144 if (info->mask != util_format_get_mask(info->src.format))
145 return false;
146
147 if (info->mask != util_format_get_mask(info->dst.format))
148 return false;
149
150 return true;
151 }
152
153 static void
154 emit_setup(struct fd_ringbuffer *ring)
155 {
156 OUT_PKT7(ring, CP_EVENT_WRITE, 1);
157 OUT_RING(ring, LRZ_FLUSH);
158
159 OUT_PKT7(ring, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
160 OUT_RING(ring, 0x0);
161
162 OUT_PKT4(ring, REG_A5XX_PC_POWER_CNTL, 1);
163 OUT_RING(ring, 0x00000003); /* PC_POWER_CNTL */
164
165 OUT_PKT4(ring, REG_A5XX_VFD_POWER_CNTL, 1);
166 OUT_RING(ring, 0x00000003); /* VFD_POWER_CNTL */
167
168 /* 0x10000000 for BYPASS.. 0x7c13c080 for GMEM: */
169 OUT_WFI5(ring);
170 OUT_PKT4(ring, REG_A5XX_RB_CCU_CNTL, 1);
171 OUT_RING(ring, 0x10000000); /* RB_CCU_CNTL */
172
173 OUT_PKT4(ring, REG_A5XX_RB_RENDER_CNTL, 1);
174 OUT_RING(ring, 0x00000008);
175
176 OUT_PKT4(ring, REG_A5XX_UNKNOWN_2100, 1);
177 OUT_RING(ring, 0x86000000); /* UNKNOWN_2100 */
178
179 OUT_PKT4(ring, REG_A5XX_UNKNOWN_2180, 1);
180 OUT_RING(ring, 0x86000000); /* UNKNOWN_2180 */
181
182 OUT_PKT4(ring, REG_A5XX_UNKNOWN_2184, 1);
183 OUT_RING(ring, 0x00000009); /* UNKNOWN_2184 */
184
185 OUT_PKT4(ring, REG_A5XX_RB_CNTL, 1);
186 OUT_RING(ring, A5XX_RB_CNTL_BYPASS);
187
188 OUT_PKT4(ring, REG_A5XX_RB_MODE_CNTL, 1);
189 OUT_RING(ring, 0x00000004); /* RB_MODE_CNTL */
190
191 OUT_PKT4(ring, REG_A5XX_SP_MODE_CNTL, 1);
192 OUT_RING(ring, 0x0000000c); /* SP_MODE_CNTL */
193
194 OUT_PKT4(ring, REG_A5XX_TPL1_MODE_CNTL, 1);
195 OUT_RING(ring, 0x00000344); /* TPL1_MODE_CNTL */
196
197 OUT_PKT4(ring, REG_A5XX_HLSQ_MODE_CNTL, 1);
198 OUT_RING(ring, 0x00000002); /* HLSQ_MODE_CNTL */
199
200 OUT_PKT4(ring, REG_A5XX_GRAS_CL_CNTL, 1);
201 OUT_RING(ring, 0x00000181); /* GRAS_CL_CNTL */
202 }
203
204 /* buffers need to be handled specially since x/width can exceed the bounds
205 * supported by hw.. if necessary decompose into (potentially) two 2D blits
206 */
207 static void
208 emit_blit_buffer(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
209 {
210 const struct pipe_box *sbox = &info->src.box;
211 const struct pipe_box *dbox = &info->dst.box;
212 struct fd_resource *src, *dst;
213 unsigned sshift, dshift;
214
215 src = fd_resource(info->src.resource);
216 dst = fd_resource(info->dst.resource);
217
218 debug_assert(src->cpp == 1);
219 debug_assert(dst->cpp == 1);
220 debug_assert(info->src.resource->format == info->dst.resource->format);
221 debug_assert((sbox->y == 0) && (sbox->height == 1));
222 debug_assert((dbox->y == 0) && (dbox->height == 1));
223 debug_assert((sbox->z == 0) && (sbox->depth == 1));
224 debug_assert((dbox->z == 0) && (dbox->depth == 1));
225 debug_assert(sbox->width == dbox->width);
226 debug_assert(info->src.level == 0);
227 debug_assert(info->dst.level == 0);
228
229 /*
230 * Buffers can have dimensions bigger than max width, remap into
231 * multiple 1d blits to fit within max dimension
232 *
233 * Note that blob uses .ARRAY_PITCH=128 for blitting buffers, which
234 * seems to prevent overfetch related faults. Not quite sure what
235 * the deal is there.
236 *
237 * Low 6 bits of SRC/DST addresses need to be zero (ie. address
238 * aligned to 64) so we need to shift src/dst x1/x2 to make up the
239 * difference. On top of already splitting up the blit so width
240 * isn't > 16k.
241 *
242 * We perhaps could do a bit better, if src and dst are aligned but
243 * in the worst case this means we have to split the copy up into
244 * 16k (0x4000) minus 64 (0x40).
245 */
246
247 sshift = sbox->x & 0x3f;
248 dshift = dbox->x & 0x3f;
249
250 for (unsigned off = 0; off < sbox->width; off += (0x4000 - 0x40)) {
251 unsigned soff, doff, w, p;
252
253 soff = (sbox->x + off) & ~0x3f;
254 doff = (dbox->x + off) & ~0x3f;
255
256 w = MIN2(sbox->width - off, (0x4000 - 0x40));
257 p = align(w, 64);
258
259 debug_assert((soff + w) <= fd_bo_size(src->bo));
260 debug_assert((doff + w) <= fd_bo_size(dst->bo));
261
262 OUT_PKT7(ring, CP_SET_RENDER_MODE, 1);
263 OUT_RING(ring, CP_SET_RENDER_MODE_0_MODE(BLIT2D));
264
265 /*
266 * Emit source:
267 */
268 OUT_PKT4(ring, REG_A5XX_RB_2D_SRC_INFO, 9);
269 OUT_RING(ring, A5XX_RB_2D_SRC_INFO_COLOR_FORMAT(RB5_R8_UNORM) |
270 A5XX_RB_2D_SRC_INFO_TILE_MODE(TILE5_LINEAR) |
271 A5XX_RB_2D_SRC_INFO_COLOR_SWAP(WZYX));
272 OUT_RELOC(ring, src->bo, soff, 0, 0); /* RB_2D_SRC_LO/HI */
273 OUT_RING(ring, A5XX_RB_2D_SRC_SIZE_PITCH(p) |
274 A5XX_RB_2D_SRC_SIZE_ARRAY_PITCH(128));
275 OUT_RING(ring, 0x00000000);
276 OUT_RING(ring, 0x00000000);
277 OUT_RING(ring, 0x00000000);
278 OUT_RING(ring, 0x00000000);
279 OUT_RING(ring, 0x00000000);
280
281 OUT_PKT4(ring, REG_A5XX_GRAS_2D_SRC_INFO, 1);
282 OUT_RING(ring, A5XX_GRAS_2D_SRC_INFO_COLOR_FORMAT(RB5_R8_UNORM) |
283 A5XX_GRAS_2D_SRC_INFO_COLOR_SWAP(WZYX));
284
285 /*
286 * Emit destination:
287 */
288 OUT_PKT4(ring, REG_A5XX_RB_2D_DST_INFO, 9);
289 OUT_RING(ring, A5XX_RB_2D_DST_INFO_COLOR_FORMAT(RB5_R8_UNORM) |
290 A5XX_RB_2D_DST_INFO_TILE_MODE(TILE5_LINEAR) |
291 A5XX_RB_2D_DST_INFO_COLOR_SWAP(WZYX));
292 OUT_RELOCW(ring, dst->bo, doff, 0, 0); /* RB_2D_DST_LO/HI */
293 OUT_RING(ring, A5XX_RB_2D_DST_SIZE_PITCH(p) |
294 A5XX_RB_2D_DST_SIZE_ARRAY_PITCH(128));
295 OUT_RING(ring, 0x00000000);
296 OUT_RING(ring, 0x00000000);
297 OUT_RING(ring, 0x00000000);
298 OUT_RING(ring, 0x00000000);
299 OUT_RING(ring, 0x00000000);
300
301 OUT_PKT4(ring, REG_A5XX_GRAS_2D_DST_INFO, 1);
302 OUT_RING(ring, A5XX_GRAS_2D_DST_INFO_COLOR_FORMAT(RB5_R8_UNORM) |
303 A5XX_GRAS_2D_DST_INFO_COLOR_SWAP(WZYX));
304
305 /*
306 * Blit command:
307 */
308 OUT_PKT7(ring, CP_BLIT, 5);
309 OUT_RING(ring, CP_BLIT_0_OP(BLIT_OP_COPY));
310 OUT_RING(ring, CP_BLIT_1_SRC_X1(sshift) | CP_BLIT_1_SRC_Y1(0));
311 OUT_RING(ring, CP_BLIT_2_SRC_X2(sshift+w-1) | CP_BLIT_2_SRC_Y2(0));
312 OUT_RING(ring, CP_BLIT_3_DST_X1(dshift) | CP_BLIT_3_DST_Y1(0));
313 OUT_RING(ring, CP_BLIT_4_DST_X2(dshift+w-1) | CP_BLIT_4_DST_Y2(0));
314
315 OUT_PKT7(ring, CP_SET_RENDER_MODE, 1);
316 OUT_RING(ring, CP_SET_RENDER_MODE_0_MODE(END2D));
317
318 OUT_WFI5(ring);
319 }
320 }
321
322 static void
323 emit_blit(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
324 {
325 const struct pipe_box *sbox = &info->src.box;
326 const struct pipe_box *dbox = &info->dst.box;
327 struct fd_resource *src, *dst;
328 struct fd_resource_slice *sslice, *dslice;
329 enum a5xx_color_fmt sfmt, dfmt;
330 enum a5xx_tile_mode stile, dtile;
331 enum a3xx_color_swap sswap, dswap;
332 unsigned ssize, dsize, spitch, dpitch;
333 unsigned sx1, sy1, sx2, sy2;
334 unsigned dx1, dy1, dx2, dy2;
335
336 src = fd_resource(info->src.resource);
337 dst = fd_resource(info->dst.resource);
338
339 sslice = fd_resource_slice(src, info->src.level);
340 dslice = fd_resource_slice(dst, info->dst.level);
341
342 sfmt = fd5_pipe2color(info->src.format);
343 dfmt = fd5_pipe2color(info->dst.format);
344
345 stile = fd_resource_level_linear(info->src.resource, info->src.level) ?
346 TILE5_LINEAR : src->tile_mode;
347 dtile = fd_resource_level_linear(info->dst.resource, info->dst.level) ?
348 TILE5_LINEAR : dst->tile_mode;
349
350 sswap = fd5_pipe2swap(info->src.format);
351 dswap = fd5_pipe2swap(info->dst.format);
352
353 spitch = sslice->pitch * src->cpp;
354 dpitch = dslice->pitch * dst->cpp;
355
356 /* if dtile, then dswap ignored by hw, and likewise if stile then sswap
357 * ignored by hw.. but in this case we have already rejected the blit
358 * if src and dst formats differ, so juse use WZYX for both src and
359 * dst swap mode (so we don't change component order)
360 */
361 if (stile || dtile) {
362 debug_assert(info->src.format == info->dst.format);
363 sswap = dswap = WZYX;
364 }
365
366 sx1 = sbox->x;
367 sy1 = sbox->y;
368 sx2 = sbox->x + sbox->width - 1;
369 sy2 = sbox->y + sbox->height - 1;
370
371 dx1 = dbox->x;
372 dy1 = dbox->y;
373 dx2 = dbox->x + dbox->width - 1;
374 dy2 = dbox->y + dbox->height - 1;
375
376 if (info->src.resource->target == PIPE_TEXTURE_3D)
377 ssize = sslice->size0;
378 else
379 ssize = src->layer_size;
380
381 if (info->dst.resource->target == PIPE_TEXTURE_3D)
382 dsize = dslice->size0;
383 else
384 dsize = dst->layer_size;
385
386 for (unsigned i = 0; i < info->dst.box.depth; i++) {
387 unsigned soff = fd_resource_offset(src, info->src.level, sbox->z + i);
388 unsigned doff = fd_resource_offset(dst, info->dst.level, dbox->z + i);
389
390 debug_assert((soff + (sbox->height * spitch)) <= fd_bo_size(src->bo));
391 debug_assert((doff + (dbox->height * dpitch)) <= fd_bo_size(dst->bo));
392
393 OUT_PKT7(ring, CP_SET_RENDER_MODE, 1);
394 OUT_RING(ring, CP_SET_RENDER_MODE_0_MODE(BLIT2D));
395
396 /*
397 * Emit source:
398 */
399 OUT_PKT4(ring, REG_A5XX_RB_2D_SRC_INFO, 9);
400 OUT_RING(ring, A5XX_RB_2D_SRC_INFO_COLOR_FORMAT(sfmt) |
401 A5XX_RB_2D_SRC_INFO_TILE_MODE(stile) |
402 A5XX_RB_2D_SRC_INFO_COLOR_SWAP(sswap));
403 OUT_RELOC(ring, src->bo, soff, 0, 0); /* RB_2D_SRC_LO/HI */
404 OUT_RING(ring, A5XX_RB_2D_SRC_SIZE_PITCH(spitch) |
405 A5XX_RB_2D_SRC_SIZE_ARRAY_PITCH(ssize));
406 OUT_RING(ring, 0x00000000);
407 OUT_RING(ring, 0x00000000);
408 OUT_RING(ring, 0x00000000);
409 OUT_RING(ring, 0x00000000);
410 OUT_RING(ring, 0x00000000);
411
412 OUT_PKT4(ring, REG_A5XX_GRAS_2D_SRC_INFO, 1);
413 OUT_RING(ring, A5XX_GRAS_2D_SRC_INFO_COLOR_FORMAT(sfmt) |
414 A5XX_GRAS_2D_SRC_INFO_TILE_MODE(stile) |
415 A5XX_GRAS_2D_SRC_INFO_COLOR_SWAP(sswap));
416
417 /*
418 * Emit destination:
419 */
420 OUT_PKT4(ring, REG_A5XX_RB_2D_DST_INFO, 9);
421 OUT_RING(ring, A5XX_RB_2D_DST_INFO_COLOR_FORMAT(dfmt) |
422 A5XX_RB_2D_DST_INFO_TILE_MODE(dtile) |
423 A5XX_RB_2D_DST_INFO_COLOR_SWAP(dswap));
424 OUT_RELOCW(ring, dst->bo, doff, 0, 0); /* RB_2D_DST_LO/HI */
425 OUT_RING(ring, A5XX_RB_2D_DST_SIZE_PITCH(dpitch) |
426 A5XX_RB_2D_DST_SIZE_ARRAY_PITCH(dsize));
427 OUT_RING(ring, 0x00000000);
428 OUT_RING(ring, 0x00000000);
429 OUT_RING(ring, 0x00000000);
430 OUT_RING(ring, 0x00000000);
431 OUT_RING(ring, 0x00000000);
432
433 OUT_PKT4(ring, REG_A5XX_GRAS_2D_DST_INFO, 1);
434 OUT_RING(ring, A5XX_GRAS_2D_DST_INFO_COLOR_FORMAT(dfmt) |
435 A5XX_GRAS_2D_DST_INFO_TILE_MODE(dtile) |
436 A5XX_GRAS_2D_DST_INFO_COLOR_SWAP(dswap));
437
438 /*
439 * Blit command:
440 */
441 OUT_PKT7(ring, CP_BLIT, 5);
442 OUT_RING(ring, CP_BLIT_0_OP(BLIT_OP_COPY));
443 OUT_RING(ring, CP_BLIT_1_SRC_X1(sx1) | CP_BLIT_1_SRC_Y1(sy1));
444 OUT_RING(ring, CP_BLIT_2_SRC_X2(sx2) | CP_BLIT_2_SRC_Y2(sy2));
445 OUT_RING(ring, CP_BLIT_3_DST_X1(dx1) | CP_BLIT_3_DST_Y1(dy1));
446 OUT_RING(ring, CP_BLIT_4_DST_X2(dx2) | CP_BLIT_4_DST_Y2(dy2));
447
448 OUT_PKT7(ring, CP_SET_RENDER_MODE, 1);
449 OUT_RING(ring, CP_SET_RENDER_MODE_0_MODE(END2D));
450 }
451 }
452
453 bool
454 fd5_blitter_blit(struct fd_context *ctx, const struct pipe_blit_info *info)
455 {
456 struct fd_batch *batch;
457
458 if (!can_do_blit(info)) {
459 return false;
460 }
461
462 batch = fd_bc_alloc_batch(&ctx->screen->batch_cache, ctx, true);
463
464 fd5_emit_restore(batch, batch->draw);
465 fd5_emit_lrz_flush(batch->draw);
466
467 emit_setup(batch->draw);
468
469 if ((info->src.resource->target == PIPE_BUFFER) &&
470 (info->dst.resource->target == PIPE_BUFFER)) {
471 assert(fd_resource(info->src.resource)->tile_mode == TILE5_LINEAR);
472 assert(fd_resource(info->dst.resource)->tile_mode == TILE5_LINEAR);
473 emit_blit_buffer(batch->draw, info);
474 } else {
475 /* I don't *think* we need to handle blits between buffer <-> !buffer */
476 debug_assert(info->src.resource->target != PIPE_BUFFER);
477 debug_assert(info->dst.resource->target != PIPE_BUFFER);
478 emit_blit(batch->draw, info);
479 }
480
481 fd_resource(info->dst.resource)->valid = true;
482 batch->needs_flush = true;
483
484 fd_batch_flush(batch, false, false);
485 fd_batch_reference(&batch, NULL);
486
487 return true;
488 }
489
490 unsigned
491 fd5_tile_mode(const struct pipe_resource *tmpl)
492 {
493 /* basically just has to be a format we can blit, so uploads/downloads
494 * via linear staging buffer works:
495 */
496 if (ok_format(tmpl->format))
497 return TILE5_3;
498
499 return TILE5_LINEAR;
500 }