1 /*
2 * Copyright (C) 2016 Rob Clark <robclark@freedesktop.org>
3 * Copyright © 2018 Google, Inc.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 *
24 * Authors:
25 * Rob Clark <robclark@freedesktop.org>
26 */
27
28 #include "pipe/p_state.h"
29 #include "util/u_string.h"
30 #include "util/u_memory.h"
31 #include "util/u_helpers.h"
32 #include "util/format/u_format.h"
33 #include "util/u_viewport.h"
34
35 #include "freedreno_log.h"
36 #include "freedreno_resource.h"
37 #include "freedreno_state.h"
38 #include "freedreno_query_hw.h"
39 #include "common/freedreno_guardband.h"
40
41 #include "fd6_emit.h"
42 #include "fd6_blend.h"
43 #include "fd6_const.h"
44 #include "fd6_context.h"
45 #include "fd6_image.h"
46 #include "fd6_pack.h"
47 #include "fd6_program.h"
48 #include "fd6_rasterizer.h"
49 #include "fd6_texture.h"
50 #include "fd6_format.h"
51 #include "fd6_zsa.h"
52
53 /* Border color layout is diff from a4xx/a5xx.. if it turns out to be
54 * the same as a6xx then move this somewhere common ;-)
55 *
56 * Entry layout looks like (total size, 0x60 bytes):
57 */
58
59 struct PACKED bcolor_entry {
60 uint32_t fp32[4];
61 uint16_t ui16[4];
62 int16_t si16[4];
63 uint16_t fp16[4];
64 uint16_t rgb565;
65 uint16_t rgb5a1;
66 uint16_t rgba4;
67 uint8_t __pad0[2];
68 uint8_t ui8[4];
69 int8_t si8[4];
70 uint32_t rgb10a2;
71 uint32_t z24; /* also s8? */
72 uint16_t srgb[4]; /* appears to duplicate fp16[], but clamped, used for srgb */
73 uint8_t __pad1[56];
74 };
75
76 #define FD6_BORDER_COLOR_SIZE sizeof(struct bcolor_entry)
77 #define FD6_BORDER_COLOR_UPLOAD_SIZE (2 * PIPE_MAX_SAMPLERS * FD6_BORDER_COLOR_SIZE)
78
79 static void
setup_border_colors(struct fd_texture_stateobj * tex,struct bcolor_entry * entries)80 setup_border_colors(struct fd_texture_stateobj *tex, struct bcolor_entry *entries)
81 {
82 unsigned i, j;
83 STATIC_ASSERT(sizeof(struct bcolor_entry) == FD6_BORDER_COLOR_SIZE);
84
85 for (i = 0; i < tex->num_samplers; i++) {
86 struct bcolor_entry *e = &entries[i];
87 struct pipe_sampler_state *sampler = tex->samplers[i];
88 union pipe_color_union *bc;
89
90 if (!sampler)
91 continue;
92
93 bc = &sampler->border_color;
94
95 /*
96 * XXX HACK ALERT XXX
97 *
98 * The border colors need to be swizzled in a particular
99 * format-dependent order. Even though samplers don't know about
100 * formats, we can assume that with a GL state tracker, there's a
101 * 1:1 correspondence between sampler and texture. Take advantage
102 * of that knowledge.
103 */
104 if ((i >= tex->num_textures) || !tex->textures[i])
105 continue;
106
107 struct pipe_sampler_view *view = tex->textures[i];
108 enum pipe_format format = view->format;
109 const struct util_format_description *desc =
110 util_format_description(format);
111
112 e->rgb565 = 0;
113 e->rgb5a1 = 0;
114 e->rgba4 = 0;
115 e->rgb10a2 = 0;
116 e->z24 = 0;
117
118 unsigned char swiz[4];
119
120 fd6_tex_swiz(format, swiz,
121 view->swizzle_r, view->swizzle_g,
122 view->swizzle_b, view->swizzle_a);
123
124 for (j = 0; j < 4; j++) {
125 int c = swiz[j];
126 int cd = c;
127
128 /*
129 * HACK: for PIPE_FORMAT_X24S8_UINT we end up w/ the
130 * stencil border color value in bc->ui[0] but according
131 * to desc->swizzle and desc->channel, the .x/.w component
132 * is NONE and the stencil value is in the y component.
133 * Meanwhile the hardware wants this in the .w component
134 * for x24s8 and the .x component for x32_s8x24.
135 */
136 if ((format == PIPE_FORMAT_X24S8_UINT) ||
137 (format == PIPE_FORMAT_X32_S8X24_UINT)) {
138 if (j == 0) {
139 c = 1;
140 cd = (format == PIPE_FORMAT_X32_S8X24_UINT) ? 0 : 3;
141 } else {
142 continue;
143 }
144 }
145
146 if (c >= 4)
147 continue;
148
149 if (desc->channel[c].pure_integer) {
150 uint16_t clamped;
151 switch (desc->channel[c].size) {
152 case 2:
153 assert(desc->channel[c].type == UTIL_FORMAT_TYPE_UNSIGNED);
154 clamped = CLAMP(bc->ui[j], 0, 0x3);
155 break;
156 case 8:
157 if (desc->channel[c].type == UTIL_FORMAT_TYPE_SIGNED)
158 clamped = CLAMP(bc->i[j], -128, 127);
159 else
160 clamped = CLAMP(bc->ui[j], 0, 255);
161 break;
162 case 10:
163 assert(desc->channel[c].type == UTIL_FORMAT_TYPE_UNSIGNED);
164 clamped = CLAMP(bc->ui[j], 0, 0x3ff);
165 break;
166 case 16:
167 if (desc->channel[c].type == UTIL_FORMAT_TYPE_SIGNED)
168 clamped = CLAMP(bc->i[j], -32768, 32767);
169 else
170 clamped = CLAMP(bc->ui[j], 0, 65535);
171 break;
172 default:
173 assert(!"Unexpected bit size");
174 case 32:
175 clamped = 0;
176 break;
177 }
178 e->fp32[cd] = bc->ui[j];
179 e->fp16[cd] = clamped;
180 } else {
181 float f = bc->f[j];
182 float f_u = CLAMP(f, 0, 1);
183 float f_s = CLAMP(f, -1, 1);
184
185 e->fp32[c] = fui(f);
186 e->fp16[c] = _mesa_float_to_half(f);
187 e->srgb[c] = _mesa_float_to_half(f_u);
188 e->ui16[c] = f_u * 0xffff;
189 e->si16[c] = f_s * 0x7fff;
190 e->ui8[c] = f_u * 0xff;
191 e->si8[c] = f_s * 0x7f;
192 if (c == 1)
193 e->rgb565 |= (int)(f_u * 0x3f) << 5;
194 else if (c < 3)
195 e->rgb565 |= (int)(f_u * 0x1f) << (c ? 11 : 0);
196 if (c == 3)
197 e->rgb5a1 |= (f_u > 0.5) ? 0x8000 : 0;
198 else
199 e->rgb5a1 |= (int)(f_u * 0x1f) << (c * 5);
200 if (c == 3)
201 e->rgb10a2 |= (int)(f_u * 0x3) << 30;
202 else
203 e->rgb10a2 |= (int)(f_u * 0x3ff) << (c * 10);
204 e->rgba4 |= (int)(f_u * 0xf) << (c * 4);
205 if (c == 0)
206 e->z24 = f_u * 0xffffff;
207 }
208 }
209
210 #ifdef DEBUG
211 memset(&e->__pad0, 0, sizeof(e->__pad0));
212 memset(&e->__pad1, 0, sizeof(e->__pad1));
213 #endif
214 }
215 }
216
217 static void
emit_border_color(struct fd_context * ctx,struct fd_ringbuffer * ring)218 emit_border_color(struct fd_context *ctx, struct fd_ringbuffer *ring)
219 {
220 struct fd6_context *fd6_ctx = fd6_context(ctx);
221 struct bcolor_entry *entries;
222 unsigned off;
223 void *ptr;
224
225 STATIC_ASSERT(sizeof(struct bcolor_entry) == FD6_BORDER_COLOR_SIZE);
226
227 u_upload_alloc(fd6_ctx->border_color_uploader,
228 0, FD6_BORDER_COLOR_UPLOAD_SIZE,
229 FD6_BORDER_COLOR_UPLOAD_SIZE, &off,
230 &fd6_ctx->border_color_buf,
231 &ptr);
232
233 entries = ptr;
234
235 setup_border_colors(&ctx->tex[PIPE_SHADER_VERTEX], &entries[0]);
236 setup_border_colors(&ctx->tex[PIPE_SHADER_FRAGMENT],
237 &entries[ctx->tex[PIPE_SHADER_VERTEX].num_samplers]);
238
239 OUT_PKT4(ring, REG_A6XX_SP_TP_BORDER_COLOR_BASE_ADDR_LO, 2);
240 OUT_RELOC(ring, fd_resource(fd6_ctx->border_color_buf)->bo, off, 0, 0);
241
242 u_upload_unmap(fd6_ctx->border_color_uploader);
243 }
244
245 static void
fd6_emit_fb_tex(struct fd_ringbuffer * state,struct fd_context * ctx)246 fd6_emit_fb_tex(struct fd_ringbuffer *state, struct fd_context *ctx)
247 {
248 struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
249 struct pipe_surface *psurf = pfb->cbufs[0];
250 struct fd_resource *rsc = fd_resource(psurf->texture);
251
252 uint32_t texconst0 = fd6_tex_const_0(psurf->texture, psurf->u.tex.level,
253 psurf->format, PIPE_SWIZZLE_X, PIPE_SWIZZLE_Y,
254 PIPE_SWIZZLE_Z, PIPE_SWIZZLE_W);
255
256 /* always TILE6_2 mode in GMEM.. which also means no swap: */
257 texconst0 &= ~(A6XX_TEX_CONST_0_SWAP__MASK | A6XX_TEX_CONST_0_TILE_MODE__MASK);
258 texconst0 |= A6XX_TEX_CONST_0_TILE_MODE(TILE6_2);
259
260 OUT_RING(state, texconst0);
261 OUT_RING(state, A6XX_TEX_CONST_1_WIDTH(pfb->width) |
262 A6XX_TEX_CONST_1_HEIGHT(pfb->height));
263 OUT_RINGP(state, A6XX_TEX_CONST_2_TYPE(A6XX_TEX_2D),
264 &ctx->batch->fb_read_patches);
265 OUT_RING(state, A6XX_TEX_CONST_3_ARRAY_PITCH(rsc->layout.layer_size));
266
267 OUT_RING(state, A6XX_TEX_CONST_4_BASE_LO(ctx->screen->gmem_base));
268 OUT_RING(state, A6XX_TEX_CONST_5_BASE_HI(ctx->screen->gmem_base >> 32) |
269 A6XX_TEX_CONST_5_DEPTH(1));
270 OUT_RING(state, 0); /* texconst6 */
271 OUT_RING(state, 0); /* texconst7 */
272 OUT_RING(state, 0); /* texconst8 */
273 OUT_RING(state, 0); /* texconst9 */
274 OUT_RING(state, 0); /* texconst10 */
275 OUT_RING(state, 0); /* texconst11 */
276 OUT_RING(state, 0);
277 OUT_RING(state, 0);
278 OUT_RING(state, 0);
279 OUT_RING(state, 0);
280 }
281
282 bool
fd6_emit_textures(struct fd_pipe * pipe,struct fd_ringbuffer * ring,enum pipe_shader_type type,struct fd_texture_stateobj * tex,unsigned bcolor_offset,const struct ir3_shader_variant * v,struct fd_context * ctx)283 fd6_emit_textures(struct fd_pipe *pipe, struct fd_ringbuffer *ring,
284 enum pipe_shader_type type, struct fd_texture_stateobj *tex,
285 unsigned bcolor_offset,
286 /* can be NULL if no image/SSBO/fb state to merge in: */
287 const struct ir3_shader_variant *v, struct fd_context *ctx)
288 {
289 bool needs_border = false;
290 unsigned opcode, tex_samp_reg, tex_const_reg, tex_count_reg;
291 enum a6xx_state_block sb;
292
293 switch (type) {
294 case PIPE_SHADER_VERTEX:
295 sb = SB6_VS_TEX;
296 opcode = CP_LOAD_STATE6_GEOM;
297 tex_samp_reg = REG_A6XX_SP_VS_TEX_SAMP_LO;
298 tex_const_reg = REG_A6XX_SP_VS_TEX_CONST_LO;
299 tex_count_reg = REG_A6XX_SP_VS_TEX_COUNT;
300 break;
301 case PIPE_SHADER_TESS_CTRL:
302 sb = SB6_HS_TEX;
303 opcode = CP_LOAD_STATE6_GEOM;
304 tex_samp_reg = REG_A6XX_SP_HS_TEX_SAMP_LO;
305 tex_const_reg = REG_A6XX_SP_HS_TEX_CONST_LO;
306 tex_count_reg = REG_A6XX_SP_HS_TEX_COUNT;
307 break;
308 case PIPE_SHADER_TESS_EVAL:
309 sb = SB6_DS_TEX;
310 opcode = CP_LOAD_STATE6_GEOM;
311 tex_samp_reg = REG_A6XX_SP_DS_TEX_SAMP_LO;
312 tex_const_reg = REG_A6XX_SP_DS_TEX_CONST_LO;
313 tex_count_reg = REG_A6XX_SP_DS_TEX_COUNT;
314 break;
315 case PIPE_SHADER_GEOMETRY:
316 sb = SB6_GS_TEX;
317 opcode = CP_LOAD_STATE6_GEOM;
318 tex_samp_reg = REG_A6XX_SP_GS_TEX_SAMP_LO;
319 tex_const_reg = REG_A6XX_SP_GS_TEX_CONST_LO;
320 tex_count_reg = REG_A6XX_SP_GS_TEX_COUNT;
321 break;
322 case PIPE_SHADER_FRAGMENT:
323 sb = SB6_FS_TEX;
324 opcode = CP_LOAD_STATE6_FRAG;
325 tex_samp_reg = REG_A6XX_SP_FS_TEX_SAMP_LO;
326 tex_const_reg = REG_A6XX_SP_FS_TEX_CONST_LO;
327 tex_count_reg = REG_A6XX_SP_FS_TEX_COUNT;
328 break;
329 case PIPE_SHADER_COMPUTE:
330 sb = SB6_CS_TEX;
331 opcode = CP_LOAD_STATE6_FRAG;
332 tex_samp_reg = REG_A6XX_SP_CS_TEX_SAMP_LO;
333 tex_const_reg = REG_A6XX_SP_CS_TEX_CONST_LO;
334 tex_count_reg = REG_A6XX_SP_CS_TEX_COUNT;
335 break;
336 default:
337 unreachable("bad state block");
338 }
339
340 if (tex->num_samplers > 0) {
341 struct fd_ringbuffer *state =
342 fd_ringbuffer_new_object(pipe, tex->num_samplers * 4 * 4);
343 for (unsigned i = 0; i < tex->num_samplers; i++) {
344 static const struct fd6_sampler_stateobj dummy_sampler = {};
345 const struct fd6_sampler_stateobj *sampler = tex->samplers[i] ?
346 fd6_sampler_stateobj(tex->samplers[i]) : &dummy_sampler;
347 OUT_RING(state, sampler->texsamp0);
348 OUT_RING(state, sampler->texsamp1);
349 OUT_RING(state, sampler->texsamp2 |
350 A6XX_TEX_SAMP_2_BCOLOR(i + bcolor_offset));
351 OUT_RING(state, sampler->texsamp3);
352 needs_border |= sampler->needs_border;
353 }
354
355 /* output sampler state: */
356 OUT_PKT7(ring, opcode, 3);
357 OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(0) |
358 CP_LOAD_STATE6_0_STATE_TYPE(ST6_SHADER) |
359 CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
360 CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
361 CP_LOAD_STATE6_0_NUM_UNIT(tex->num_samplers));
362 OUT_RB(ring, state); /* SRC_ADDR_LO/HI */
363
364 OUT_PKT4(ring, tex_samp_reg, 2);
365 OUT_RB(ring, state); /* SRC_ADDR_LO/HI */
366
367 fd_ringbuffer_del(state);
368 }
369
370 unsigned num_merged_textures = tex->num_textures;
371 unsigned num_textures = tex->num_textures;
372 if (v) {
373 num_merged_textures += v->image_mapping.num_tex;
374
375 if (v->fb_read)
376 num_merged_textures++;
377
378 /* There could be more bound textures than what the shader uses.
379 * Which isn't known at shader compile time. So in the case we
380 * are merging tex state, only emit the textures that the shader
381 * uses (since the image/SSBO related tex state comes immediately
382 * after)
383 */
384 num_textures = v->image_mapping.tex_base;
385 }
386
387 if (num_merged_textures > 0) {
388 struct fd_ringbuffer *state =
389 fd_ringbuffer_new_object(pipe, num_merged_textures * 16 * 4);
390 for (unsigned i = 0; i < num_textures; i++) {
391 static const struct fd6_pipe_sampler_view dummy_view = {};
392 const struct fd6_pipe_sampler_view *view = tex->textures[i] ?
393 fd6_pipe_sampler_view(tex->textures[i]) : &dummy_view;
394
395 OUT_RING(state, view->texconst0);
396 OUT_RING(state, view->texconst1);
397 OUT_RING(state, view->texconst2);
398 OUT_RING(state, view->texconst3);
399
400 if (view->ptr1) {
401 OUT_RELOC(state, view->ptr1->bo, view->offset1,
402 (uint64_t)view->texconst5 << 32, 0);
403 } else {
404 OUT_RING(state, 0x00000000);
405 OUT_RING(state, view->texconst5);
406 }
407
408 OUT_RING(state, view->texconst6);
409
410 if (view->ptr2) {
411 OUT_RELOC(state, view->ptr2->bo, view->offset2, 0, 0);
412 } else {
413 OUT_RING(state, 0);
414 OUT_RING(state, 0);
415 }
416
417 OUT_RING(state, view->texconst9);
418 OUT_RING(state, view->texconst10);
419 OUT_RING(state, view->texconst11);
420 OUT_RING(state, 0);
421 OUT_RING(state, 0);
422 OUT_RING(state, 0);
423 OUT_RING(state, 0);
424 }
425
426 if (v) {
427 const struct ir3_ibo_mapping *mapping = &v->image_mapping;
428 struct fd_shaderbuf_stateobj *buf = &ctx->shaderbuf[type];
429 struct fd_shaderimg_stateobj *img = &ctx->shaderimg[type];
430
431 for (unsigned i = 0; i < mapping->num_tex; i++) {
432 unsigned idx = mapping->tex_to_image[i];
433 if (idx & IBO_SSBO) {
434 fd6_emit_ssbo_tex(state, &buf->sb[idx & ~IBO_SSBO]);
435 } else {
436 fd6_emit_image_tex(state, &img->si[idx]);
437 }
438 }
439
440 if (v->fb_read) {
441 fd6_emit_fb_tex(state, ctx);
442 }
443 }
444
445 /* emit texture state: */
446 OUT_PKT7(ring, opcode, 3);
447 OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(0) |
448 CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
449 CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
450 CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
451 CP_LOAD_STATE6_0_NUM_UNIT(num_merged_textures));
452 OUT_RB(ring, state); /* SRC_ADDR_LO/HI */
453
454 OUT_PKT4(ring, tex_const_reg, 2);
455 OUT_RB(ring, state); /* SRC_ADDR_LO/HI */
456
457 fd_ringbuffer_del(state);
458 }
459
460 OUT_PKT4(ring, tex_count_reg, 1);
461 OUT_RING(ring, num_merged_textures);
462
463 return needs_border;
464 }
465
466 /* Emits combined texture state, which also includes any Image/SSBO
467 * related texture state merged in (because we must have all texture
468 * state for a given stage in a single buffer). In the fast-path, if
469 * we don't need to merge in any image/ssbo related texture state, we
470 * just use cached texture stateobj. Otherwise we generate a single-
471 * use stateobj.
472 *
473 * TODO Is there some sane way we can still use cached texture stateobj
474 * with image/ssbo in use?
475 *
476 * returns whether border_color is required:
477 */
478 static bool
fd6_emit_combined_textures(struct fd_ringbuffer * ring,struct fd6_emit * emit,enum pipe_shader_type type,const struct ir3_shader_variant * v)479 fd6_emit_combined_textures(struct fd_ringbuffer *ring, struct fd6_emit *emit,
480 enum pipe_shader_type type, const struct ir3_shader_variant *v)
481 {
482 struct fd_context *ctx = emit->ctx;
483 bool needs_border = false;
484
485 static const struct {
486 enum fd6_state_id state_id;
487 unsigned enable_mask;
488 } s[PIPE_SHADER_TYPES] = {
489 [PIPE_SHADER_VERTEX] = { FD6_GROUP_VS_TEX, ENABLE_ALL },
490 [PIPE_SHADER_TESS_CTRL] = { FD6_GROUP_HS_TEX, ENABLE_ALL },
491 [PIPE_SHADER_TESS_EVAL] = { FD6_GROUP_DS_TEX, ENABLE_ALL },
492 [PIPE_SHADER_GEOMETRY] = { FD6_GROUP_GS_TEX, ENABLE_ALL },
493 [PIPE_SHADER_FRAGMENT] = { FD6_GROUP_FS_TEX, ENABLE_DRAW },
494 };
495
496 debug_assert(s[type].state_id);
497
498 if (!v->image_mapping.num_tex && !v->fb_read) {
499 /* in the fast-path, when we don't have to mix in any image/SSBO
500 * related texture state, we can just lookup the stateobj and
501 * re-emit that:
502 *
503 * Also, framebuffer-read is a slow-path because an extra
504 * texture needs to be inserted.
505 *
506 * TODO we can probably simmplify things if we also treated
507 * border_color as a slow-path.. this way the tex state key
508 * wouldn't depend on bcolor_offset.. but fb_read might rather
509 * be *somehow* a fast-path if we eventually used it for PLS.
510 * I suppose there would be no harm in just *always* inserting
511 * an fb_read texture?
512 */
513 if ((ctx->dirty_shader[type] & FD_DIRTY_SHADER_TEX) &&
514 ctx->tex[type].num_textures > 0) {
515 struct fd6_texture_state *tex = fd6_texture_state(ctx,
516 type, &ctx->tex[type]);
517
518 needs_border |= tex->needs_border;
519
520 fd6_emit_add_group(emit, tex->stateobj, s[type].state_id,
521 s[type].enable_mask);
522 }
523 } else {
524 /* In the slow-path, create a one-shot texture state object
525 * if either TEX|PROG|SSBO|IMAGE state is dirty:
526 */
527 if ((ctx->dirty_shader[type] &
528 (FD_DIRTY_SHADER_TEX | FD_DIRTY_SHADER_PROG |
529 FD_DIRTY_SHADER_IMAGE | FD_DIRTY_SHADER_SSBO)) ||
530 v->fb_read) {
531 struct fd_texture_stateobj *tex = &ctx->tex[type];
532 struct fd_ringbuffer *stateobj =
533 fd_submit_new_ringbuffer(ctx->batch->submit,
534 0x1000, FD_RINGBUFFER_STREAMING);
535 unsigned bcolor_offset =
536 fd6_border_color_offset(ctx, type, tex);
537
538 needs_border |= fd6_emit_textures(ctx->pipe, stateobj, type, tex,
539 bcolor_offset, v, ctx);
540
541 fd6_emit_take_group(emit, stateobj, s[type].state_id,
542 s[type].enable_mask);
543 }
544 }
545
546 return needs_border;
547 }
548
549 static struct fd_ringbuffer *
build_vbo_state(struct fd6_emit * emit)550 build_vbo_state(struct fd6_emit *emit)
551 {
552 const struct fd_vertex_state *vtx = emit->vtx;
553
554 struct fd_ringbuffer *ring = fd_submit_new_ringbuffer(emit->ctx->batch->submit,
555 4 * (1 + vtx->vertexbuf.count * 4), FD_RINGBUFFER_STREAMING);
556
557 OUT_PKT4(ring, REG_A6XX_VFD_FETCH(0), 4 * vtx->vertexbuf.count);
558 for (int32_t j = 0; j < vtx->vertexbuf.count; j++) {
559 const struct pipe_vertex_buffer *vb = &vtx->vertexbuf.vb[j];
560 struct fd_resource *rsc = fd_resource(vb->buffer.resource);
561 if (rsc == NULL) {
562 OUT_RING(ring, 0);
563 OUT_RING(ring, 0);
564 OUT_RING(ring, 0);
565 OUT_RING(ring, 0);
566 } else {
567 uint32_t off = vb->buffer_offset;
568 uint32_t size = fd_bo_size(rsc->bo) - off;
569
570 OUT_RELOC(ring, rsc->bo, off, 0, 0);
571 OUT_RING(ring, size); /* VFD_FETCH[j].SIZE */
572 OUT_RING(ring, vb->stride); /* VFD_FETCH[j].STRIDE */
573 }
574 }
575
576 return ring;
577 }
578
579 static enum a6xx_ztest_mode
compute_ztest_mode(struct fd6_emit * emit,bool lrz_valid)580 compute_ztest_mode(struct fd6_emit *emit, bool lrz_valid)
581 {
582 struct fd_context *ctx = emit->ctx;
583 struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
584 struct fd6_zsa_stateobj *zsa = fd6_zsa_stateobj(ctx->zsa);
585 const struct ir3_shader_variant *fs = emit->fs;
586
587 if (fs->shader->nir->info.fs.early_fragment_tests)
588 return A6XX_EARLY_Z;
589
590 if (fs->no_earlyz || fs->writes_pos || !zsa->base.depth.enabled) {
591 return A6XX_LATE_Z;
592 } else if ((fs->has_kill || zsa->alpha_test) &&
593 (zsa->base.depth.writemask || !pfb->zsbuf)) {
594 /* Slightly odd, but seems like the hw wants us to select
595 * LATE_Z mode if there is no depth buffer + discard. Either
596 * that, or when occlusion query is enabled. See:
597 *
598 * dEQP-GLES31.functional.fbo.no_attachments.*
599 */
600 return lrz_valid ? A6XX_EARLY_LRZ_LATE_Z : A6XX_LATE_Z;
601 } else {
602 return A6XX_EARLY_Z;
603 }
604 }
605
606 /**
607 * Calculate normalized LRZ state based on zsa/prog/blend state, updating
608 * the zsbuf's lrz state as necessary to detect the cases where we need
609 * to invalidate lrz.
610 */
611 static struct fd6_lrz_state
compute_lrz_state(struct fd6_emit * emit,bool binning_pass)612 compute_lrz_state(struct fd6_emit *emit, bool binning_pass)
613 {
614 struct fd_context *ctx = emit->ctx;
615 struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
616 const struct ir3_shader_variant *fs = emit->fs;
617 struct fd6_lrz_state lrz;
618
619 if (!pfb->zsbuf) {
620 memset(&lrz, 0, sizeof(lrz));
621 if (!binning_pass) {
622 lrz.z_mode = compute_ztest_mode(emit, false);
623 }
624 return lrz;
625 }
626
627 struct fd6_blend_stateobj *blend = fd6_blend_stateobj(ctx->blend);
628 struct fd6_zsa_stateobj *zsa = fd6_zsa_stateobj(ctx->zsa);
629 struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
630
631 lrz = zsa->lrz;
632
633 /* normalize lrz state: */
634 if (blend->reads_dest || fs->writes_pos || fs->no_earlyz || fs->has_kill) {
635 lrz.write = false;
636 if (binning_pass)
637 lrz.enable = false;
638 }
639
640 /* if we change depthfunc direction, bail out on using LRZ. The
641 * LRZ buffer encodes a min/max depth value per block, but if
642 * we switch from GT/GE <-> LT/LE, those values cannot be
643 * interpreted properly.
644 */
645 if (zsa->base.depth.enabled &&
646 (rsc->lrz_direction != FD_LRZ_UNKNOWN) &&
647 (rsc->lrz_direction != lrz.direction)) {
648 rsc->lrz_valid = false;
649 }
650
651 if (zsa->invalidate_lrz || !rsc->lrz_valid) {
652 rsc->lrz_valid = false;
653 memset(&lrz, 0, sizeof(lrz));
654 }
655
656 if (fs->no_earlyz || fs->writes_pos) {
657 lrz.enable = false;
658 lrz.write = false;
659 lrz.test = false;
660 }
661
662 if (!binning_pass) {
663 lrz.z_mode = compute_ztest_mode(emit, rsc->lrz_valid);
664 }
665
666 /* Once we start writing to the real depth buffer, we lock in the
667 * direction for LRZ.. if we have to skip a LRZ write for any
668 * reason, it is still safe to have LRZ until there is a direction
669 * reversal. Prior to the reversal, since we disabled LRZ writes
670 * in the "unsafe" cases, this just means that the LRZ test may
671 * not early-discard some things that end up not passing a later
672 * test (ie. be overly concervative). But once you have a reversal
673 * of direction, it is possible to increase/decrease the z value
674 * to the point where the overly-conservative test is incorrect.
675 */
676 if (zsa->base.depth.writemask) {
677 rsc->lrz_direction = lrz.direction;
678 }
679
680 return lrz;
681 }
682
683 static struct fd_ringbuffer *
build_lrz(struct fd6_emit * emit,bool binning_pass)684 build_lrz(struct fd6_emit *emit, bool binning_pass)
685 {
686 struct fd_context *ctx = emit->ctx;
687 struct fd6_context *fd6_ctx = fd6_context(ctx);
688 struct fd6_lrz_state lrz =
689 compute_lrz_state(emit, binning_pass);
690
691 /* If the LRZ state has not changed, we can skip the emit: */
692 if (!ctx->last.dirty &&
693 !memcmp(&fd6_ctx->last.lrz[binning_pass], &lrz, sizeof(lrz)))
694 return NULL;
695
696 fd6_ctx->last.lrz[binning_pass] = lrz;
697
698 struct fd_ringbuffer *ring = fd_submit_new_ringbuffer(ctx->batch->submit,
699 8*4, FD_RINGBUFFER_STREAMING);
700
701 OUT_REG(ring, A6XX_GRAS_LRZ_CNTL(
702 .enable = lrz.enable,
703 .lrz_write = lrz.write,
704 .greater = lrz.direction == FD_LRZ_GREATER,
705 .z_test_enable = lrz.test,
706 ));
707 OUT_REG(ring, A6XX_RB_LRZ_CNTL(
708 .enable = lrz.enable,
709 ));
710
711 OUT_REG(ring, A6XX_RB_DEPTH_PLANE_CNTL(
712 .z_mode = lrz.z_mode,
713 ));
714
715 OUT_REG(ring, A6XX_GRAS_SU_DEPTH_PLANE_CNTL(
716 .z_mode = lrz.z_mode,
717 ));
718
719 return ring;
720 }
721
722 static void
fd6_emit_streamout(struct fd_ringbuffer * ring,struct fd6_emit * emit,struct ir3_stream_output_info * info)723 fd6_emit_streamout(struct fd_ringbuffer *ring, struct fd6_emit *emit, struct ir3_stream_output_info *info)
724 {
725 struct fd_context *ctx = emit->ctx;
726 const struct fd6_program_state *prog = fd6_emit_get_prog(emit);
727 struct fd_streamout_stateobj *so = &ctx->streamout;
728
729 emit->streamout_mask = 0;
730
731 for (unsigned i = 0; i < so->num_targets; i++) {
732 struct pipe_stream_output_target *target = so->targets[i];
733
734 if (!target)
735 continue;
736
737 OUT_PKT4(ring, REG_A6XX_VPC_SO_BUFFER_BASE_LO(i), 3);
738 /* VPC_SO[i].BUFFER_BASE_LO: */
739 OUT_RELOC(ring, fd_resource(target->buffer)->bo, target->buffer_offset, 0, 0);
740 OUT_RING(ring, target->buffer_size - target->buffer_offset);
741
742 if (so->reset & (1 << i)) {
743 unsigned offset = (so->offsets[i] * info->stride[i] * 4);
744 OUT_PKT4(ring, REG_A6XX_VPC_SO_BUFFER_OFFSET(i), 1);
745 OUT_RING(ring, offset);
746 } else {
747 OUT_PKT7(ring, CP_MEM_TO_REG, 3);
748 OUT_RING(ring, CP_MEM_TO_REG_0_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(i)) |
749 CP_MEM_TO_REG_0_SHIFT_BY_2 | CP_MEM_TO_REG_0_UNK31 |
750 CP_MEM_TO_REG_0_CNT(0));
751 OUT_RELOC(ring, control_ptr(fd6_context(ctx), flush_base[i].offset));
752 }
753
754 OUT_PKT4(ring, REG_A6XX_VPC_SO_FLUSH_BASE_LO(i), 2);
755 OUT_RELOC(ring, control_ptr(fd6_context(ctx), flush_base[i]));
756
757 so->reset &= ~(1 << i);
758
759 emit->streamout_mask |= (1 << i);
760 }
761
762 if (emit->streamout_mask) {
763 fd6_emit_add_group(emit, prog->streamout_stateobj, FD6_GROUP_SO, ENABLE_ALL);
764 } else {
765 /* If we transition from a draw with streamout to one without, turn
766 * off streamout.
767 */
768 if (ctx->last.streamout_mask != 0) {
769 struct fd_ringbuffer *obj = fd_submit_new_ringbuffer(emit->ctx->batch->submit,
770 5 * 4, FD_RINGBUFFER_STREAMING);
771
772 OUT_PKT7(obj, CP_CONTEXT_REG_BUNCH, 4);
773 OUT_RING(obj, REG_A6XX_VPC_SO_CNTL);
774 OUT_RING(obj, 0);
775 OUT_RING(obj, REG_A6XX_VPC_SO_STREAM_CNTL);
776 OUT_RING(obj, 0);
777
778 fd6_emit_take_group(emit, obj, FD6_GROUP_SO, ENABLE_ALL);
779 }
780 }
781
782 ctx->last.streamout_mask = emit->streamout_mask;
783 }
784
785 void
fd6_emit_state(struct fd_ringbuffer * ring,struct fd6_emit * emit)786 fd6_emit_state(struct fd_ringbuffer *ring, struct fd6_emit *emit)
787 {
788 struct fd_context *ctx = emit->ctx;
789 struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
790 const struct fd6_program_state *prog = fd6_emit_get_prog(emit);
791 const struct ir3_shader_variant *vs = emit->vs;
792 const struct ir3_shader_variant *hs = emit->hs;
793 const struct ir3_shader_variant *ds = emit->ds;
794 const struct ir3_shader_variant *gs = emit->gs;
795 const struct ir3_shader_variant *fs = emit->fs;
796 const enum fd_dirty_3d_state dirty = emit->dirty;
797 bool needs_border = false;
798
799 emit_marker6(ring, 5);
800
801 /* NOTE: we track fb_read differently than _BLEND_ENABLED since
802 * we might at some point decide to do sysmem in some cases when
803 * blend is enabled:
804 */
805 if (fs->fb_read)
806 ctx->batch->gmem_reason |= FD_GMEM_FB_READ;
807
808 if (emit->dirty & FD_DIRTY_VTXSTATE) {
809 struct fd6_vertex_stateobj *vtx = fd6_vertex_stateobj(ctx->vtx.vtx);
810
811 fd6_emit_add_group(emit, vtx->stateobj, FD6_GROUP_VTXSTATE, ENABLE_ALL);
812 }
813
814 if (emit->dirty & FD_DIRTY_VTXBUF) {
815 struct fd_ringbuffer *state;
816
817 state = build_vbo_state(emit);
818 fd6_emit_take_group(emit, state, FD6_GROUP_VBO, ENABLE_ALL);
819 }
820
821 if (dirty & (FD_DIRTY_ZSA | FD_DIRTY_RASTERIZER)) {
822 struct fd_ringbuffer *state =
823 fd6_zsa_state(ctx,
824 util_format_is_pure_integer(pipe_surface_format(pfb->cbufs[0])),
825 fd_depth_clamp_enabled(ctx));
826
827 fd6_emit_add_group(emit, state, FD6_GROUP_ZSA, ENABLE_ALL);
828 }
829
830 if (dirty & (FD_DIRTY_ZSA | FD_DIRTY_BLEND | FD_DIRTY_PROG)) {
831 struct fd_ringbuffer *state;
832
833 state = build_lrz(emit, false);
834 if (state) {
835 fd6_emit_take_group(emit, state, FD6_GROUP_LRZ, ENABLE_DRAW);
836 }
837
838 state = build_lrz(emit, true);
839 if (state) {
840 fd6_emit_take_group(emit, state,
841 FD6_GROUP_LRZ_BINNING, CP_SET_DRAW_STATE__0_BINNING);
842 }
843 }
844
845 if (dirty & FD_DIRTY_STENCIL_REF) {
846 struct pipe_stencil_ref *sr = &ctx->stencil_ref;
847
848 OUT_PKT4(ring, REG_A6XX_RB_STENCILREF, 1);
849 OUT_RING(ring, A6XX_RB_STENCILREF_REF(sr->ref_value[0]) |
850 A6XX_RB_STENCILREF_BFREF(sr->ref_value[1]));
851 }
852
853 /* NOTE: scissor enabled bit is part of rasterizer state, but
854 * fd_rasterizer_state_bind() will mark scissor dirty if needed:
855 */
856 if (dirty & FD_DIRTY_SCISSOR) {
857 struct fd_ringbuffer *ring = fd_submit_new_ringbuffer(
858 emit->ctx->batch->submit, 3*4, FD_RINGBUFFER_STREAMING);
859 struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
860
861 OUT_REG(ring,
862 A6XX_GRAS_SC_SCREEN_SCISSOR_TL(0,
863 .x = scissor->minx,
864 .y = scissor->miny
865 ),
866 A6XX_GRAS_SC_SCREEN_SCISSOR_BR(0,
867 .x = MAX2(scissor->maxx, 1) - 1,
868 .y = MAX2(scissor->maxy, 1) - 1
869 )
870 );
871
872 fd6_emit_take_group(emit, ring, FD6_GROUP_SCISSOR, ENABLE_ALL);
873
874 ctx->batch->max_scissor.minx = MIN2(ctx->batch->max_scissor.minx, scissor->minx);
875 ctx->batch->max_scissor.miny = MIN2(ctx->batch->max_scissor.miny, scissor->miny);
876 ctx->batch->max_scissor.maxx = MAX2(ctx->batch->max_scissor.maxx, scissor->maxx);
877 ctx->batch->max_scissor.maxy = MAX2(ctx->batch->max_scissor.maxy, scissor->maxy);
878 }
879
880 if (dirty & FD_DIRTY_VIEWPORT) {
881 struct pipe_scissor_state *scissor = &ctx->viewport_scissor;
882
883 OUT_REG(ring,
884 A6XX_GRAS_CL_VPORT_XOFFSET(0, ctx->viewport.translate[0]),
885 A6XX_GRAS_CL_VPORT_XSCALE(0, ctx->viewport.scale[0]),
886 A6XX_GRAS_CL_VPORT_YOFFSET(0, ctx->viewport.translate[1]),
887 A6XX_GRAS_CL_VPORT_YSCALE(0, ctx->viewport.scale[1]),
888 A6XX_GRAS_CL_VPORT_ZOFFSET(0, ctx->viewport.translate[2]),
889 A6XX_GRAS_CL_VPORT_ZSCALE(0, ctx->viewport.scale[2])
890 );
891
892 OUT_REG(ring,
893 A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL(0,
894 .x = scissor->minx,
895 .y = scissor->miny
896 ),
897 A6XX_GRAS_SC_VIEWPORT_SCISSOR_BR(0,
898 .x = MAX2(scissor->maxx, 1) - 1,
899 .y = MAX2(scissor->maxy, 1) - 1
900 )
901 );
902
903 unsigned guardband_x =
904 fd_calc_guardband(ctx->viewport.translate[0], ctx->viewport.scale[0],
905 false);
906 unsigned guardband_y =
907 fd_calc_guardband(ctx->viewport.translate[1], ctx->viewport.scale[1],
908 false);
909
910 OUT_REG(ring, A6XX_GRAS_CL_GUARDBAND_CLIP_ADJ(
911 .horz = guardband_x,
912 .vert = guardband_y
913 )
914 );
915 }
916
917 /* The clamp ranges are only used when the rasterizer wants depth
918 * clamping.
919 */
920 if ((dirty & (FD_DIRTY_VIEWPORT | FD_DIRTY_RASTERIZER)) &&
921 fd_depth_clamp_enabled(ctx)) {
922 float zmin, zmax;
923 util_viewport_zmin_zmax(&ctx->viewport, ctx->rasterizer->clip_halfz,
924 &zmin, &zmax);
925
926 OUT_REG(ring,
927 A6XX_GRAS_CL_Z_CLAMP_MIN(0, zmin),
928 A6XX_GRAS_CL_Z_CLAMP_MAX(0, zmax));
929
930 OUT_REG(ring,
931 A6XX_RB_Z_CLAMP_MIN(zmin),
932 A6XX_RB_Z_CLAMP_MAX(zmax));
933 }
934
935 if (dirty & FD_DIRTY_PROG) {
936 fd6_emit_add_group(emit, prog->config_stateobj, FD6_GROUP_PROG_CONFIG, ENABLE_ALL);
937 fd6_emit_add_group(emit, prog->stateobj, FD6_GROUP_PROG, ENABLE_DRAW);
938 fd6_emit_add_group(emit, prog->binning_stateobj,
939 FD6_GROUP_PROG_BINNING, CP_SET_DRAW_STATE__0_BINNING);
940
941 /* emit remaining streaming program state, ie. what depends on
942 * other emit state, so cannot be pre-baked.
943 */
944 struct fd_ringbuffer *streaming = fd6_program_interp_state(emit);
945
946 fd6_emit_take_group(emit, streaming, FD6_GROUP_PROG_INTERP, ENABLE_DRAW);
947 }
948
949 if (dirty & FD_DIRTY_RASTERIZER) {
950 struct fd_ringbuffer *stateobj =
951 fd6_rasterizer_state(ctx, emit->primitive_restart);
952 fd6_emit_add_group(emit, stateobj,
953 FD6_GROUP_RASTERIZER, ENABLE_ALL);
954 }
955
956 if (dirty & (FD_DIRTY_FRAMEBUFFER | FD_DIRTY_RASTERIZER_DISCARD | FD_DIRTY_PROG)) {
957 struct fd_ringbuffer *ring = fd_submit_new_ringbuffer(
958 emit->ctx->batch->submit, 5 * 4, FD_RINGBUFFER_STREAMING);
959
960 unsigned nr = pfb->nr_cbufs;
961
962 if (ctx->rasterizer->rasterizer_discard)
963 nr = 0;
964
965 OUT_PKT4(ring, REG_A6XX_RB_FS_OUTPUT_CNTL0, 2);
966 OUT_RING(ring, COND(fs->writes_pos, A6XX_RB_FS_OUTPUT_CNTL0_FRAG_WRITES_Z) |
967 COND(fs->writes_smask && pfb->samples > 1,
968 A6XX_RB_FS_OUTPUT_CNTL0_FRAG_WRITES_SAMPMASK));
969 OUT_RING(ring, A6XX_RB_FS_OUTPUT_CNTL1_MRT(nr));
970
971 OUT_PKT4(ring, REG_A6XX_SP_FS_OUTPUT_CNTL1, 1);
972 OUT_RING(ring, A6XX_SP_FS_OUTPUT_CNTL1_MRT(nr));
973
974 fd6_emit_take_group(emit, ring, FD6_GROUP_PROG_FB_RAST, ENABLE_DRAW);
975 }
976
977 fd6_emit_consts(emit);
978
979 struct ir3_stream_output_info *info = &fd6_last_shader(prog)->shader->stream_output;
980 if (info->num_outputs)
981 fd6_emit_streamout(ring, emit, info);
982
983 if (dirty & (FD_DIRTY_BLEND | FD_DIRTY_SAMPLE_MASK)) {
984 struct fd6_blend_variant *blend = fd6_blend_variant(ctx->blend,
985 pfb->samples, ctx->sample_mask);
986 fd6_emit_add_group(emit, blend->stateobj, FD6_GROUP_BLEND, ENABLE_DRAW);
987 }
988
989 if (dirty & FD_DIRTY_BLEND_COLOR) {
990 struct pipe_blend_color *bcolor = &ctx->blend_color;
991 struct fd_ringbuffer *ring = fd_submit_new_ringbuffer(
992 emit->ctx->batch->submit, 5*4, FD_RINGBUFFER_STREAMING);
993
994 OUT_REG(ring,
995 A6XX_RB_BLEND_RED_F32(bcolor->color[0]),
996 A6XX_RB_BLEND_GREEN_F32(bcolor->color[1]),
997 A6XX_RB_BLEND_BLUE_F32(bcolor->color[2]),
998 A6XX_RB_BLEND_ALPHA_F32(bcolor->color[3])
999 );
1000
1001 fd6_emit_take_group(emit, ring, FD6_GROUP_BLEND_COLOR, ENABLE_DRAW);
1002 }
1003
1004 needs_border |= fd6_emit_combined_textures(ring, emit, PIPE_SHADER_VERTEX, vs);
1005 if (hs) {
1006 needs_border |= fd6_emit_combined_textures(ring, emit, PIPE_SHADER_TESS_CTRL, hs);
1007 needs_border |= fd6_emit_combined_textures(ring, emit, PIPE_SHADER_TESS_EVAL, ds);
1008 }
1009 if (gs) {
1010 needs_border |= fd6_emit_combined_textures(ring, emit, PIPE_SHADER_GEOMETRY, gs);
1011 }
1012 needs_border |= fd6_emit_combined_textures(ring, emit, PIPE_SHADER_FRAGMENT, fs);
1013
1014 if (needs_border)
1015 emit_border_color(ctx, ring);
1016
1017 if (hs) {
1018 debug_assert(ir3_shader_nibo(hs) == 0);
1019 debug_assert(ir3_shader_nibo(ds) == 0);
1020 }
1021 if (gs) {
1022 debug_assert(ir3_shader_nibo(gs) == 0);
1023 }
1024
1025 #define DIRTY_IBO (FD_DIRTY_SHADER_SSBO | FD_DIRTY_SHADER_IMAGE | \
1026 FD_DIRTY_SHADER_PROG)
1027 if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & DIRTY_IBO) {
1028 struct fd_ringbuffer *state =
1029 fd6_build_ibo_state(ctx, fs, PIPE_SHADER_FRAGMENT);
1030 struct fd_ringbuffer *obj = fd_submit_new_ringbuffer(
1031 ctx->batch->submit, 0x100, FD_RINGBUFFER_STREAMING);
1032
1033 OUT_PKT7(obj, CP_LOAD_STATE6, 3);
1034 OUT_RING(obj, CP_LOAD_STATE6_0_DST_OFF(0) |
1035 CP_LOAD_STATE6_0_STATE_TYPE(ST6_SHADER) |
1036 CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
1037 CP_LOAD_STATE6_0_STATE_BLOCK(SB6_IBO) |
1038 CP_LOAD_STATE6_0_NUM_UNIT(ir3_shader_nibo(fs)));
1039 OUT_RB(obj, state);
1040
1041 OUT_PKT4(obj, REG_A6XX_SP_IBO_LO, 2);
1042 OUT_RB(obj, state);
1043
1044 /* TODO if we used CP_SET_DRAW_STATE for compute shaders, we could
1045 * de-duplicate this from program->config_stateobj
1046 */
1047 OUT_PKT4(obj, REG_A6XX_SP_IBO_COUNT, 1);
1048 OUT_RING(obj, ir3_shader_nibo(fs));
1049
1050 fd6_emit_ibo_consts(emit, fs, PIPE_SHADER_FRAGMENT, ring);
1051
1052 fd6_emit_take_group(emit, obj, FD6_GROUP_IBO, ENABLE_DRAW);
1053 fd_ringbuffer_del(state);
1054 }
1055
1056 if (emit->num_groups > 0) {
1057 OUT_PKT7(ring, CP_SET_DRAW_STATE, 3 * emit->num_groups);
1058 for (unsigned i = 0; i < emit->num_groups; i++) {
1059 struct fd6_state_group *g = &emit->groups[i];
1060 unsigned n = g->stateobj ?
1061 fd_ringbuffer_size(g->stateobj) / 4 : 0;
1062
1063 debug_assert((g->enable_mask & ~ENABLE_ALL) == 0);
1064
1065 if (n == 0) {
1066 OUT_RING(ring, CP_SET_DRAW_STATE__0_COUNT(0) |
1067 CP_SET_DRAW_STATE__0_DISABLE |
1068 g->enable_mask |
1069 CP_SET_DRAW_STATE__0_GROUP_ID(g->group_id));
1070 OUT_RING(ring, 0x00000000);
1071 OUT_RING(ring, 0x00000000);
1072 } else {
1073 OUT_RING(ring, CP_SET_DRAW_STATE__0_COUNT(n) |
1074 g->enable_mask |
1075 CP_SET_DRAW_STATE__0_GROUP_ID(g->group_id));
1076 OUT_RB(ring, g->stateobj);
1077 }
1078
1079 if (g->stateobj)
1080 fd_ringbuffer_del(g->stateobj);
1081 }
1082 emit->num_groups = 0;
1083 }
1084 }
1085
1086 void
fd6_emit_cs_state(struct fd_context * ctx,struct fd_ringbuffer * ring,struct ir3_shader_variant * cp)1087 fd6_emit_cs_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
1088 struct ir3_shader_variant *cp)
1089 {
1090 enum fd_dirty_shader_state dirty = ctx->dirty_shader[PIPE_SHADER_COMPUTE];
1091
1092 if (dirty & (FD_DIRTY_SHADER_TEX | FD_DIRTY_SHADER_PROG |
1093 FD_DIRTY_SHADER_IMAGE | FD_DIRTY_SHADER_SSBO)) {
1094 struct fd_texture_stateobj *tex = &ctx->tex[PIPE_SHADER_COMPUTE];
1095 unsigned bcolor_offset = fd6_border_color_offset(ctx, PIPE_SHADER_COMPUTE, tex);
1096
1097 bool needs_border = fd6_emit_textures(ctx->pipe, ring, PIPE_SHADER_COMPUTE, tex,
1098 bcolor_offset, cp, ctx);
1099
1100 if (needs_border)
1101 emit_border_color(ctx, ring);
1102
1103 OUT_PKT4(ring, REG_A6XX_SP_VS_TEX_COUNT, 1);
1104 OUT_RING(ring, 0);
1105
1106 OUT_PKT4(ring, REG_A6XX_SP_HS_TEX_COUNT, 1);
1107 OUT_RING(ring, 0);
1108
1109 OUT_PKT4(ring, REG_A6XX_SP_DS_TEX_COUNT, 1);
1110 OUT_RING(ring, 0);
1111
1112 OUT_PKT4(ring, REG_A6XX_SP_GS_TEX_COUNT, 1);
1113 OUT_RING(ring, 0);
1114
1115 OUT_PKT4(ring, REG_A6XX_SP_FS_TEX_COUNT, 1);
1116 OUT_RING(ring, 0);
1117 }
1118
1119 if (dirty & (FD_DIRTY_SHADER_SSBO | FD_DIRTY_SHADER_IMAGE)) {
1120 struct fd_ringbuffer *state =
1121 fd6_build_ibo_state(ctx, cp, PIPE_SHADER_COMPUTE);
1122
1123 OUT_PKT7(ring, CP_LOAD_STATE6_FRAG, 3);
1124 OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(0) |
1125 CP_LOAD_STATE6_0_STATE_TYPE(ST6_IBO) |
1126 CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
1127 CP_LOAD_STATE6_0_STATE_BLOCK(SB6_CS_SHADER) |
1128 CP_LOAD_STATE6_0_NUM_UNIT(ir3_shader_nibo(cp)));
1129 OUT_RB(ring, state);
1130
1131 OUT_PKT4(ring, REG_A6XX_SP_CS_IBO_LO, 2);
1132 OUT_RB(ring, state);
1133
1134 OUT_PKT4(ring, REG_A6XX_SP_CS_IBO_COUNT, 1);
1135 OUT_RING(ring, ir3_shader_nibo(cp));
1136
1137 fd_ringbuffer_del(state);
1138 }
1139 }
1140
1141
1142 /* emit setup at begin of new cmdstream buffer (don't rely on previous
1143 * state, there could have been a context switch between ioctls):
1144 */
1145 void
fd6_emit_restore(struct fd_batch * batch,struct fd_ringbuffer * ring)1146 fd6_emit_restore(struct fd_batch *batch, struct fd_ringbuffer *ring)
1147 {
1148 //struct fd_context *ctx = batch->ctx;
1149
1150 fd_log(batch, "START RESTORE");
1151
1152 fd6_cache_inv(batch, ring);
1153
1154 OUT_REG(ring, A6XX_HLSQ_INVALIDATE_CMD(
1155 .vs_state = true,
1156 .hs_state = true,
1157 .ds_state = true,
1158 .gs_state = true,
1159 .fs_state = true,
1160 .cs_state = true,
1161 .gfx_ibo = true,
1162 .cs_ibo = true,
1163 .gfx_shared_const = true,
1164 .cs_shared_const = true,
1165 .gfx_bindless = 0x1f,
1166 .cs_bindless = 0x1f
1167 ));
1168
1169 OUT_WFI5(ring);
1170
1171 WRITE(REG_A6XX_RB_UNKNOWN_8E04, 0x0);
1172 WRITE(REG_A6XX_SP_UNKNOWN_AE04, 0x8);
1173 WRITE(REG_A6XX_SP_UNKNOWN_AE00, 0);
1174 WRITE(REG_A6XX_SP_UNKNOWN_AE0F, 0x3f);
1175 WRITE(REG_A6XX_SP_UNKNOWN_B605, 0x44);
1176 WRITE(REG_A6XX_SP_UNKNOWN_B600, 0x100000);
1177 WRITE(REG_A6XX_HLSQ_UNKNOWN_BE00, 0x80);
1178 WRITE(REG_A6XX_HLSQ_UNKNOWN_BE01, 0);
1179
1180 WRITE(REG_A6XX_VPC_UNKNOWN_9600, 0);
1181 WRITE(REG_A6XX_GRAS_UNKNOWN_8600, 0x880);
1182 WRITE(REG_A6XX_HLSQ_UNKNOWN_BE04, 0x80000);
1183 WRITE(REG_A6XX_SP_UNKNOWN_AE03, 0x1430);
1184 WRITE(REG_A6XX_SP_IBO_COUNT, 0);
1185 WRITE(REG_A6XX_SP_UNKNOWN_B182, 0);
1186 WRITE(REG_A6XX_HLSQ_SHARED_CONSTS, 0);
1187 WRITE(REG_A6XX_UCHE_UNKNOWN_0E12, 0x3200000);
1188 WRITE(REG_A6XX_UCHE_CLIENT_PF, 4);
1189 WRITE(REG_A6XX_RB_UNKNOWN_8E01, 0x1);
1190 WRITE(REG_A6XX_SP_MODE_CONTROL, A6XX_SP_MODE_CONTROL_CONSTANT_DEMOTION_ENABLE | 4);
1191 WRITE(REG_A6XX_VFD_ADD_OFFSET, A6XX_VFD_ADD_OFFSET_VERTEX);
1192 WRITE(REG_A6XX_RB_UNKNOWN_8811, 0x00000010);
1193 WRITE(REG_A6XX_PC_MODE_CNTL, 0x1f);
1194
1195 WRITE(REG_A6XX_GRAS_UNKNOWN_8101, 0);
1196 WRITE(REG_A6XX_GRAS_SAMPLE_CNTL, 0);
1197 WRITE(REG_A6XX_GRAS_UNKNOWN_8110, 0x2);
1198
1199 WRITE(REG_A6XX_RB_UNKNOWN_8818, 0);
1200 WRITE(REG_A6XX_RB_UNKNOWN_8819, 0);
1201 WRITE(REG_A6XX_RB_UNKNOWN_881A, 0);
1202 WRITE(REG_A6XX_RB_UNKNOWN_881B, 0);
1203 WRITE(REG_A6XX_RB_UNKNOWN_881C, 0);
1204 WRITE(REG_A6XX_RB_UNKNOWN_881D, 0);
1205 WRITE(REG_A6XX_RB_UNKNOWN_881E, 0);
1206 WRITE(REG_A6XX_RB_UNKNOWN_88F0, 0);
1207
1208 WRITE(REG_A6XX_VPC_POINT_COORD_INVERT,
1209 A6XX_VPC_POINT_COORD_INVERT(0).value);
1210 WRITE(REG_A6XX_VPC_UNKNOWN_9300, 0);
1211
1212 WRITE(REG_A6XX_VPC_SO_DISABLE, A6XX_VPC_SO_DISABLE(true).value);
1213
1214 WRITE(REG_A6XX_PC_RASTER_CNTL, 0);
1215
1216 WRITE(REG_A6XX_PC_MULTIVIEW_CNTL, 0);
1217
1218 WRITE(REG_A6XX_SP_UNKNOWN_A81B, 0);
1219
1220 WRITE(REG_A6XX_SP_UNKNOWN_B183, 0);
1221
1222 WRITE(REG_A6XX_GRAS_UNKNOWN_8099, 0);
1223 WRITE(REG_A6XX_GRAS_VS_LAYER_CNTL, 0);
1224 WRITE(REG_A6XX_GRAS_UNKNOWN_80A0, 2);
1225 WRITE(REG_A6XX_GRAS_UNKNOWN_80AF, 0);
1226 WRITE(REG_A6XX_VPC_UNKNOWN_9210, 0);
1227 WRITE(REG_A6XX_VPC_UNKNOWN_9211, 0);
1228 WRITE(REG_A6XX_VPC_UNKNOWN_9602, 0);
1229 WRITE(REG_A6XX_PC_UNKNOWN_9E72, 0);
1230 WRITE(REG_A6XX_SP_TP_SAMPLE_CONFIG, 0);
1231 /* NOTE blob seems to (mostly?) use 0xb2 for SP_TP_UNKNOWN_B309
1232 * but this seems to kill texture gather offsets.
1233 */
1234 WRITE(REG_A6XX_SP_TP_UNKNOWN_B309, 0xa2);
1235 WRITE(REG_A6XX_RB_SAMPLE_CONFIG, 0);
1236 WRITE(REG_A6XX_GRAS_SAMPLE_CONFIG, 0);
1237 WRITE(REG_A6XX_RB_Z_BOUNDS_MIN, 0);
1238 WRITE(REG_A6XX_RB_Z_BOUNDS_MAX, 0);
1239 WRITE(REG_A6XX_HLSQ_CONTROL_5_REG, 0xfc);
1240
1241 emit_marker6(ring, 7);
1242
1243 OUT_PKT4(ring, REG_A6XX_VFD_MODE_CNTL, 1);
1244 OUT_RING(ring, 0x00000000); /* VFD_MODE_CNTL */
1245
1246 WRITE(REG_A6XX_VFD_MULTIVIEW_CNTL, 0);
1247
1248 OUT_PKT4(ring, REG_A6XX_PC_MODE_CNTL, 1);
1249 OUT_RING(ring, 0x0000001f); /* PC_MODE_CNTL */
1250
1251 /* we don't use this yet.. probably best to disable.. */
1252 OUT_PKT7(ring, CP_SET_DRAW_STATE, 3);
1253 OUT_RING(ring, CP_SET_DRAW_STATE__0_COUNT(0) |
1254 CP_SET_DRAW_STATE__0_DISABLE_ALL_GROUPS |
1255 CP_SET_DRAW_STATE__0_GROUP_ID(0));
1256 OUT_RING(ring, CP_SET_DRAW_STATE__1_ADDR_LO(0));
1257 OUT_RING(ring, CP_SET_DRAW_STATE__2_ADDR_HI(0));
1258
1259 OUT_PKT4(ring, REG_A6XX_VPC_SO_STREAM_CNTL, 1);
1260 OUT_RING(ring, 0x00000000); /* VPC_SO_STREAM_CNTL */
1261
1262 OUT_PKT4(ring, REG_A6XX_GRAS_LRZ_CNTL, 1);
1263 OUT_RING(ring, 0x00000000);
1264
1265 OUT_PKT4(ring, REG_A6XX_RB_LRZ_CNTL, 1);
1266 OUT_RING(ring, 0x00000000);
1267
1268 fd_log(batch, "END RESTORE");
1269 }
1270
1271 static void
fd6_mem_to_mem(struct fd_ringbuffer * ring,struct pipe_resource * dst,unsigned dst_off,struct pipe_resource * src,unsigned src_off,unsigned sizedwords)1272 fd6_mem_to_mem(struct fd_ringbuffer *ring, struct pipe_resource *dst,
1273 unsigned dst_off, struct pipe_resource *src, unsigned src_off,
1274 unsigned sizedwords)
1275 {
1276 struct fd_bo *src_bo = fd_resource(src)->bo;
1277 struct fd_bo *dst_bo = fd_resource(dst)->bo;
1278 unsigned i;
1279
1280 for (i = 0; i < sizedwords; i++) {
1281 OUT_PKT7(ring, CP_MEM_TO_MEM, 5);
1282 OUT_RING(ring, 0x00000000);
1283 OUT_RELOC(ring, dst_bo, dst_off, 0, 0);
1284 OUT_RELOC(ring, src_bo, src_off, 0, 0);
1285
1286 dst_off += 4;
1287 src_off += 4;
1288 }
1289 }
1290
1291 /* this is *almost* the same as fd6_cache_flush().. which I guess
1292 * could be re-worked to be something a bit more generic w/ param
1293 * indicating what needs to be flushed.. although that would mean
1294 * figuring out which events trigger what state to flush..
1295 */
1296 static void
fd6_framebuffer_barrier(struct fd_context * ctx)1297 fd6_framebuffer_barrier(struct fd_context *ctx)
1298 {
1299 struct fd6_context *fd6_ctx = fd6_context(ctx);
1300 struct fd_batch *batch = ctx->batch;
1301 struct fd_ringbuffer *ring = batch->draw;
1302 unsigned seqno;
1303
1304 seqno = fd6_event_write(batch, ring, RB_DONE_TS, true);
1305
1306 OUT_PKT7(ring, CP_WAIT_REG_MEM, 6);
1307 OUT_RING(ring, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ) |
1308 CP_WAIT_REG_MEM_0_POLL_MEMORY);
1309 OUT_RELOC(ring, control_ptr(fd6_ctx, seqno));
1310 OUT_RING(ring, CP_WAIT_REG_MEM_3_REF(seqno));
1311 OUT_RING(ring, CP_WAIT_REG_MEM_4_MASK(~0));
1312 OUT_RING(ring, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(16));
1313
1314 fd6_event_write(batch, ring, PC_CCU_FLUSH_COLOR_TS, true);
1315 fd6_event_write(batch, ring, PC_CCU_FLUSH_DEPTH_TS, true);
1316
1317 seqno = fd6_event_write(batch, ring, CACHE_FLUSH_TS, true);
1318
1319 fd6_event_write(batch, ring, 0x31, false);
1320
1321 OUT_PKT7(ring, CP_WAIT_MEM_GTE, 4);
1322 OUT_RING(ring, CP_WAIT_MEM_GTE_0_RESERVED(0));
1323 OUT_RELOC(ring, control_ptr(fd6_ctx, seqno));
1324 OUT_RING(ring, CP_WAIT_MEM_GTE_3_REF(seqno));
1325 }
1326
1327 void
fd6_emit_init_screen(struct pipe_screen * pscreen)1328 fd6_emit_init_screen(struct pipe_screen *pscreen)
1329 {
1330 struct fd_screen *screen = fd_screen(pscreen);
1331 screen->emit_ib = fd6_emit_ib;
1332 screen->mem_to_mem = fd6_mem_to_mem;
1333 }
1334
1335 void
fd6_emit_init(struct pipe_context * pctx)1336 fd6_emit_init(struct pipe_context *pctx)
1337 {
1338 struct fd_context *ctx = fd_context(pctx);
1339 ctx->framebuffer_barrier = fd6_framebuffer_barrier;
1340 }
1341