projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
freedreno/ir3: move binning_pass out of shader variant key
[mesa.git] / src / gallium / drivers / freedreno / a6xx / fd6_emit.c
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/u_format.h"
33 #include "util/u_viewport.h"
34
35 #include "freedreno_resource.h"
36 #include "freedreno_query_hw.h"
37
38 #include "fd6_emit.h"
39 #include "fd6_blend.h"
40 #include "fd6_context.h"
41 #include "fd6_image.h"
42 #include "fd6_program.h"
43 #include "fd6_rasterizer.h"
44 #include "fd6_texture.h"
45 #include "fd6_format.h"
46 #include "fd6_zsa.h"
47
48 static uint32_t
49 shader_t_to_opcode(enum shader_t type)
50 {
51 switch (type) {
52 case SHADER_VERTEX:
53 case SHADER_TCS:
54 case SHADER_TES:
55 case SHADER_GEOM:
56 return CP_LOAD_STATE6_GEOM;
57 case SHADER_FRAGMENT:
58 case SHADER_COMPUTE:
59 return CP_LOAD_STATE6_FRAG;
60 default:
61 unreachable("bad shader type");
62 }
63 }
64
65 /* regid: base const register
66 * prsc or dwords: buffer containing constant values
67 * sizedwords: size of const value buffer
68 */
69 static void
70 fd6_emit_const(struct fd_ringbuffer *ring, enum shader_t type,
71 uint32_t regid, uint32_t offset, uint32_t sizedwords,
72 const uint32_t *dwords, struct pipe_resource *prsc)
73 {
74 uint32_t i, sz;
75 enum a6xx_state_src src;
76
77 debug_assert((regid % 4) == 0);
78 debug_assert((sizedwords % 4) == 0);
79
80 if (prsc) {
81 sz = 0;
82 src = SS6_INDIRECT;
83 } else {
84 sz = sizedwords;
85 src = SS6_DIRECT;
86 }
87
88 OUT_PKT7(ring, shader_t_to_opcode(type), 3 + sz);
89 OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(regid/4) |
90 CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
91 CP_LOAD_STATE6_0_STATE_SRC(src) |
92 CP_LOAD_STATE6_0_STATE_BLOCK(fd6_stage2shadersb(type)) |
93 CP_LOAD_STATE6_0_NUM_UNIT(sizedwords/4));
94 if (prsc) {
95 struct fd_bo *bo = fd_resource(prsc)->bo;
96 OUT_RELOC(ring, bo, offset, 0, 0);
97 } else {
98 OUT_RING(ring, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
99 OUT_RING(ring, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
100 dwords = (uint32_t *)&((uint8_t *)dwords)[offset];
101 }
102 for (i = 0; i < sz; i++) {
103 OUT_RING(ring, dwords[i]);
104 }
105 }
106
107 static void
108 fd6_emit_const_bo(struct fd_ringbuffer *ring, enum shader_t type, boolean write,
109 uint32_t regid, uint32_t num, struct pipe_resource **prscs, uint32_t *offsets)
110 {
111 uint32_t anum = align(num, 2);
112 uint32_t i;
113
114 debug_assert((regid % 4) == 0);
115
116 OUT_PKT7(ring, shader_t_to_opcode(type), 3 + (2 * anum));
117 OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(regid/4) |
118 CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS)|
119 CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
120 CP_LOAD_STATE6_0_STATE_BLOCK(fd6_stage2shadersb(type)) |
121 CP_LOAD_STATE6_0_NUM_UNIT(anum/2));
122 OUT_RING(ring, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
123 OUT_RING(ring, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
124
125 for (i = 0; i < num; i++) {
126 if (prscs[i]) {
127 if (write) {
128 OUT_RELOCW(ring, fd_resource(prscs[i])->bo, offsets[i], 0, 0);
129 } else {
130 OUT_RELOC(ring, fd_resource(prscs[i])->bo, offsets[i], 0, 0);
131 }
132 } else {
133 OUT_RING(ring, 0xbad00000 | (i << 16));
134 OUT_RING(ring, 0xbad00000 | (i << 16));
135 }
136 }
137
138 for (; i < anum; i++) {
139 OUT_RING(ring, 0xffffffff);
140 OUT_RING(ring, 0xffffffff);
141 }
142 }
143
144 /* Border color layout is diff from a4xx/a5xx.. if it turns out to be
145 * the same as a6xx then move this somewhere common ;-)
146 *
147 * Entry layout looks like (total size, 0x60 bytes):
148 */
149
150 struct PACKED bcolor_entry {
151 uint32_t fp32[4];
152 uint16_t ui16[4];
153 int16_t si16[4];
154 uint16_t fp16[4];
155 uint16_t rgb565;
156 uint16_t rgb5a1;
157 uint16_t rgba4;
158 uint8_t __pad0[2];
159 uint8_t ui8[4];
160 int8_t si8[4];
161 uint32_t rgb10a2;
162 uint32_t z24; /* also s8? */
163 uint16_t srgb[4]; /* appears to duplicate fp16[], but clamped, used for srgb */
164 uint8_t __pad1[24];
165 };
166
167 #define FD6_BORDER_COLOR_SIZE 0x60
168 #define FD6_BORDER_COLOR_UPLOAD_SIZE (2 * PIPE_MAX_SAMPLERS * FD6_BORDER_COLOR_SIZE)
169
170 static void
171 setup_border_colors(struct fd_texture_stateobj *tex, struct bcolor_entry *entries)
172 {
173 unsigned i, j;
174 STATIC_ASSERT(sizeof(struct bcolor_entry) == FD6_BORDER_COLOR_SIZE);
175
176 for (i = 0; i < tex->num_samplers; i++) {
177 struct bcolor_entry *e = &entries[i];
178 struct pipe_sampler_state *sampler = tex->samplers[i];
179 union pipe_color_union *bc;
180
181 if (!sampler)
182 continue;
183
184 bc = &sampler->border_color;
185
186 /*
187 * XXX HACK ALERT XXX
188 *
189 * The border colors need to be swizzled in a particular
190 * format-dependent order. Even though samplers don't know about
191 * formats, we can assume that with a GL state tracker, there's a
192 * 1:1 correspondence between sampler and texture. Take advantage
193 * of that knowledge.
194 */
195 if ((i >= tex->num_textures) || !tex->textures[i])
196 continue;
197
198 enum pipe_format format = tex->textures[i]->format;
199 const struct util_format_description *desc =
200 util_format_description(format);
201
202 e->rgb565 = 0;
203 e->rgb5a1 = 0;
204 e->rgba4 = 0;
205 e->rgb10a2 = 0;
206 e->z24 = 0;
207
208 for (j = 0; j < 4; j++) {
209 int c = desc->swizzle[j];
210 int cd = c;
211
212 /*
213 * HACK: for PIPE_FORMAT_X24S8_UINT we end up w/ the
214 * stencil border color value in bc->ui[0] but according
215 * to desc->swizzle and desc->channel, the .x component
216 * is NONE and the stencil value is in the y component.
217 * Meanwhile the hardware wants this in the .x componetn.
218 */
219 if ((format == PIPE_FORMAT_X24S8_UINT) ||
220 (format == PIPE_FORMAT_X32_S8X24_UINT)) {
221 if (j == 0) {
222 c = 1;
223 cd = 0;
224 } else {
225 continue;
226 }
227 }
228
229 if (c >= 4)
230 continue;
231
232 if (desc->channel[c].pure_integer) {
233 uint16_t clamped;
234 switch (desc->channel[c].size) {
235 case 2:
236 assert(desc->channel[c].type == UTIL_FORMAT_TYPE_UNSIGNED);
237 clamped = CLAMP(bc->ui[j], 0, 0x3);
238 break;
239 case 8:
240 if (desc->channel[c].type == UTIL_FORMAT_TYPE_SIGNED)
241 clamped = CLAMP(bc->i[j], -128, 127);
242 else
243 clamped = CLAMP(bc->ui[j], 0, 255);
244 break;
245 case 10:
246 assert(desc->channel[c].type == UTIL_FORMAT_TYPE_UNSIGNED);
247 clamped = CLAMP(bc->ui[j], 0, 0x3ff);
248 break;
249 case 16:
250 if (desc->channel[c].type == UTIL_FORMAT_TYPE_SIGNED)
251 clamped = CLAMP(bc->i[j], -32768, 32767);
252 else
253 clamped = CLAMP(bc->ui[j], 0, 65535);
254 break;
255 default:
256 assert(!"Unexpected bit size");
257 case 32:
258 clamped = 0;
259 break;
260 }
261 e->fp32[cd] = bc->ui[j];
262 e->fp16[cd] = clamped;
263 } else {
264 float f = bc->f[j];
265 float f_u = CLAMP(f, 0, 1);
266 float f_s = CLAMP(f, -1, 1);
267
268 e->fp32[c] = fui(f);
269 e->fp16[c] = util_float_to_half(f);
270 e->srgb[c] = util_float_to_half(f_u);
271 e->ui16[c] = f_u * 0xffff;
272 e->si16[c] = f_s * 0x7fff;
273 e->ui8[c] = f_u * 0xff;
274 e->si8[c] = f_s * 0x7f;
275 if (c == 1)
276 e->rgb565 |= (int)(f_u * 0x3f) << 5;
277 else if (c < 3)
278 e->rgb565 |= (int)(f_u * 0x1f) << (c ? 11 : 0);
279 if (c == 3)
280 e->rgb5a1 |= (f_u > 0.5) ? 0x8000 : 0;
281 else
282 e->rgb5a1 |= (int)(f_u * 0x1f) << (c * 5);
283 if (c == 3)
284 e->rgb10a2 |= (int)(f_u * 0x3) << 30;
285 else
286 e->rgb10a2 |= (int)(f_u * 0x3ff) << (c * 10);
287 e->rgba4 |= (int)(f_u * 0xf) << (c * 4);
288 if (c == 0)
289 e->z24 = f_u * 0xffffff;
290 }
291 }
292
293 #ifdef DEBUG
294 memset(&e->__pad0, 0, sizeof(e->__pad0));
295 memset(&e->__pad1, 0, sizeof(e->__pad1));
296 #endif
297 }
298 }
299
300 static void
301 emit_border_color(struct fd_context *ctx, struct fd_ringbuffer *ring)
302 {
303 struct fd6_context *fd6_ctx = fd6_context(ctx);
304 struct bcolor_entry *entries;
305 unsigned off;
306 void *ptr;
307
308 STATIC_ASSERT(sizeof(struct bcolor_entry) == FD6_BORDER_COLOR_SIZE);
309
310 u_upload_alloc(fd6_ctx->border_color_uploader,
311 0, FD6_BORDER_COLOR_UPLOAD_SIZE,
312 FD6_BORDER_COLOR_UPLOAD_SIZE, &off,
313 &fd6_ctx->border_color_buf,
314 &ptr);
315
316 entries = ptr;
317
318 setup_border_colors(&ctx->tex[PIPE_SHADER_VERTEX], &entries[0]);
319 setup_border_colors(&ctx->tex[PIPE_SHADER_FRAGMENT],
320 &entries[ctx->tex[PIPE_SHADER_VERTEX].num_samplers]);
321
322 OUT_PKT4(ring, REG_A6XX_SP_TP_BORDER_COLOR_BASE_ADDR_LO, 2);
323 OUT_RELOC(ring, fd_resource(fd6_ctx->border_color_buf)->bo, off, 0, 0);
324
325 u_upload_unmap(fd6_ctx->border_color_uploader);
326 }
327
328 bool
329 fd6_emit_textures(struct fd_pipe *pipe, struct fd_ringbuffer *ring,
330 enum a6xx_state_block sb, struct fd_texture_stateobj *tex,
331 unsigned bcolor_offset)
332 {
333 bool needs_border = false;
334 unsigned opcode, tex_samp_reg, tex_const_reg, tex_count_reg;
335
336 switch (sb) {
337 case SB6_VS_TEX:
338 opcode = CP_LOAD_STATE6_GEOM;
339 tex_samp_reg = REG_A6XX_SP_VS_TEX_SAMP_LO;
340 tex_const_reg = REG_A6XX_SP_VS_TEX_CONST_LO;
341 tex_count_reg = REG_A6XX_SP_VS_TEX_COUNT;
342 break;
343 case SB6_FS_TEX:
344 opcode = CP_LOAD_STATE6_FRAG;
345 tex_samp_reg = REG_A6XX_SP_FS_TEX_SAMP_LO;
346 tex_const_reg = REG_A6XX_SP_FS_TEX_CONST_LO;
347 tex_count_reg = REG_A6XX_SP_FS_TEX_COUNT;
348 break;
349 case SB6_CS_TEX:
350 opcode = CP_LOAD_STATE6_FRAG;
351 tex_samp_reg = REG_A6XX_SP_CS_TEX_SAMP_LO;
352 tex_const_reg = REG_A6XX_SP_CS_TEX_CONST_LO;
353 tex_count_reg = 0; //REG_A6XX_SP_CS_TEX_COUNT;
354 break;
355 default:
356 unreachable("bad state block");
357 }
358
359
360 if (tex->num_samplers > 0) {
361 struct fd_ringbuffer *state =
362 fd_ringbuffer_new_flags(pipe, tex->num_samplers * 4 * 4,
363 FD_RINGBUFFER_OBJECT);
364 for (unsigned i = 0; i < tex->num_samplers; i++) {
365 static const struct fd6_sampler_stateobj dummy_sampler = {};
366 const struct fd6_sampler_stateobj *sampler = tex->samplers[i] ?
367 fd6_sampler_stateobj(tex->samplers[i]) : &dummy_sampler;
368 OUT_RING(state, sampler->texsamp0);
369 OUT_RING(state, sampler->texsamp1);
370 OUT_RING(state, sampler->texsamp2 |
371 A6XX_TEX_SAMP_2_BCOLOR_OFFSET(bcolor_offset));
372 OUT_RING(state, sampler->texsamp3);
373 needs_border |= sampler->needs_border;
374 }
375
376 /* output sampler state: */
377 OUT_PKT7(ring, opcode, 3);
378 OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(0) |
379 CP_LOAD_STATE6_0_STATE_TYPE(ST6_SHADER) |
380 CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
381 CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
382 CP_LOAD_STATE6_0_NUM_UNIT(tex->num_samplers));
383 OUT_RB(ring, state); /* SRC_ADDR_LO/HI */
384
385 OUT_PKT4(ring, tex_samp_reg, 2);
386 OUT_RB(ring, state); /* SRC_ADDR_LO/HI */
387
388 fd_ringbuffer_del(state);
389 }
390
391 if (tex->num_textures > 0) {
392 struct fd_ringbuffer *state =
393 fd_ringbuffer_new_flags(pipe, tex->num_textures * 16 * 4,
394 FD_RINGBUFFER_OBJECT);
395 for (unsigned i = 0; i < tex->num_textures; i++) {
396 static const struct fd6_pipe_sampler_view dummy_view = {};
397 const struct fd6_pipe_sampler_view *view = tex->textures[i] ?
398 fd6_pipe_sampler_view(tex->textures[i]) : &dummy_view;
399 enum a6xx_tile_mode tile_mode = TILE6_LINEAR;
400
401 if (view->base.texture)
402 tile_mode = fd_resource(view->base.texture)->tile_mode;
403
404 OUT_RING(state, view->texconst0 |
405 A6XX_TEX_CONST_0_TILE_MODE(tile_mode));
406 OUT_RING(state, view->texconst1);
407 OUT_RING(state, view->texconst2);
408 OUT_RING(state, view->texconst3);
409
410 if (view->base.texture) {
411 struct fd_resource *rsc = fd_resource(view->base.texture);
412 if (view->base.format == PIPE_FORMAT_X32_S8X24_UINT)
413 rsc = rsc->stencil;
414 OUT_RELOC(state, rsc->bo, view->offset,
415 (uint64_t)view->texconst5 << 32, 0);
416 } else {
417 OUT_RING(state, 0x00000000);
418 OUT_RING(state, view->texconst5);
419 }
420
421 OUT_RING(state, view->texconst6);
422 OUT_RING(state, view->texconst7);
423 OUT_RING(state, view->texconst8);
424 OUT_RING(state, view->texconst9);
425 OUT_RING(state, view->texconst10);
426 OUT_RING(state, view->texconst11);
427 OUT_RING(state, 0);
428 OUT_RING(state, 0);
429 OUT_RING(state, 0);
430 OUT_RING(state, 0);
431 }
432
433 /* emit texture state: */
434 OUT_PKT7(ring, opcode, 3);
435 OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(0) |
436 CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
437 CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
438 CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
439 CP_LOAD_STATE6_0_NUM_UNIT(tex->num_textures));
440 OUT_RB(ring, state); /* SRC_ADDR_LO/HI */
441
442 OUT_PKT4(ring, tex_const_reg, 2);
443 OUT_RB(ring, state); /* SRC_ADDR_LO/HI */
444
445 fd_ringbuffer_del(state);
446 }
447
448 if (tex_count_reg) {
449 OUT_PKT4(ring, tex_count_reg, 1);
450 OUT_RING(ring, tex->num_textures);
451 }
452
453 return needs_border;
454 }
455
456 static void
457 emit_ssbos(struct fd_context *ctx, struct fd_ringbuffer *ring,
458 enum a6xx_state_block sb, struct fd_shaderbuf_stateobj *so)
459 {
460 unsigned count = util_last_bit(so->enabled_mask);
461 unsigned opcode;
462
463 if (count == 0)
464 return;
465
466 switch (sb) {
467 case SB6_SSBO:
468 case SB6_CS_SSBO:
469 opcode = CP_LOAD_STATE6_GEOM;
470 break;
471 default:
472 unreachable("bad state block");
473 }
474
475 OUT_PKT7(ring, opcode, 3 + (4 * count));
476 OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(0) |
477 CP_LOAD_STATE6_0_STATE_TYPE(0) |
478 CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
479 CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
480 CP_LOAD_STATE6_0_NUM_UNIT(count));
481 OUT_RING(ring, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
482 OUT_RING(ring, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
483 for (unsigned i = 0; i < count; i++) {
484 OUT_RING(ring, 0x00000000);
485 OUT_RING(ring, 0x00000000);
486 OUT_RING(ring, 0x00000000);
487 OUT_RING(ring, 0x00000000);
488 }
489
490 #if 0
491 OUT_PKT7(ring, opcode, 3 + (2 * count));
492 OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(0) |
493 CP_LOAD_STATE6_0_STATE_TYPE(1) |
494 CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
495 CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
496 CP_LOAD_STATE6_0_NUM_UNIT(count));
497 OUT_RING(ring, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
498 OUT_RING(ring, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
499 for (unsigned i = 0; i < count; i++) {
500 struct pipe_shader_buffer *buf = &so->sb[i];
501 unsigned sz = buf->buffer_size;
502
503 /* width is in dwords, overflows into height: */
504 sz /= 4;
505
506 OUT_RING(ring, A6XX_SSBO_1_0_WIDTH(sz));
507 OUT_RING(ring, A6XX_SSBO_1_1_HEIGHT(sz >> 16));
508 }
509 #endif
510
511 OUT_PKT7(ring, opcode, 3 + (2 * count));
512 OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(0) |
513 CP_LOAD_STATE6_0_STATE_TYPE(2) |
514 CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
515 CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
516 CP_LOAD_STATE6_0_NUM_UNIT(count));
517 OUT_RING(ring, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
518 OUT_RING(ring, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
519 for (unsigned i = 0; i < count; i++) {
520 struct pipe_shader_buffer *buf = &so->sb[i];
521 if (buf->buffer) {
522 struct fd_resource *rsc = fd_resource(buf->buffer);
523 OUT_RELOCW(ring, rsc->bo, buf->buffer_offset, 0, 0);
524 } else {
525 OUT_RING(ring, 0x00000000);
526 OUT_RING(ring, 0x00000000);
527 }
528 }
529 }
530
531 void
532 fd6_emit_vertex_bufs(struct fd_ringbuffer *ring, struct fd6_emit *emit)
533 {
534 int32_t i, j;
535 const struct fd_vertex_state *vtx = emit->vtx;
536 const struct ir3_shader_variant *vp = fd6_emit_get_vp(emit);
537
538 for (i = 0, j = 0; i <= vp->inputs_count; i++) {
539 if (vp->inputs[i].sysval)
540 continue;
541 if (vp->inputs[i].compmask) {
542 struct pipe_vertex_element *elem = &vtx->vtx->pipe[i];
543 const struct pipe_vertex_buffer *vb =
544 &vtx->vertexbuf.vb[elem->vertex_buffer_index];
545 struct fd_resource *rsc = fd_resource(vb->buffer.resource);
546 enum pipe_format pfmt = elem->src_format;
547 enum a6xx_vtx_fmt fmt = fd6_pipe2vtx(pfmt);
548 bool isint = util_format_is_pure_integer(pfmt);
549 uint32_t off = vb->buffer_offset + elem->src_offset;
550 uint32_t size = fd_bo_size(rsc->bo) - off;
551 debug_assert(fmt != ~0);
552
553 #ifdef DEBUG
554 /* see dEQP-GLES31.stress.vertex_attribute_binding.buffer_bounds.bind_vertex_buffer_offset_near_wrap_10
555 */
556 if (off > fd_bo_size(rsc->bo))
557 continue;
558 #endif
559
560 OUT_PKT4(ring, REG_A6XX_VFD_FETCH(j), 4);
561 OUT_RELOC(ring, rsc->bo, off, 0, 0);
562 OUT_RING(ring, size); /* VFD_FETCH[j].SIZE */
563 OUT_RING(ring, vb->stride); /* VFD_FETCH[j].STRIDE */
564
565 OUT_PKT4(ring, REG_A6XX_VFD_DECODE(j), 2);
566 OUT_RING(ring, A6XX_VFD_DECODE_INSTR_IDX(j) |
567 A6XX_VFD_DECODE_INSTR_FORMAT(fmt) |
568 COND(elem->instance_divisor, A6XX_VFD_DECODE_INSTR_INSTANCED) |
569 A6XX_VFD_DECODE_INSTR_SWAP(fd6_pipe2swap(pfmt)) |
570 A6XX_VFD_DECODE_INSTR_UNK30 |
571 COND(!isint, A6XX_VFD_DECODE_INSTR_FLOAT));
572 OUT_RING(ring, MAX2(1, elem->instance_divisor)); /* VFD_DECODE[j].STEP_RATE */
573
574 OUT_PKT4(ring, REG_A6XX_VFD_DEST_CNTL(j), 1);
575 OUT_RING(ring, A6XX_VFD_DEST_CNTL_INSTR_WRITEMASK(vp->inputs[i].compmask) |
576 A6XX_VFD_DEST_CNTL_INSTR_REGID(vp->inputs[i].regid));
577
578 j++;
579 }
580 }
581
582 OUT_PKT4(ring, REG_A6XX_VFD_CONTROL_0, 1);
583 OUT_RING(ring, A6XX_VFD_CONTROL_0_VTXCNT(j) | (j << 8));
584 }
585
586 void
587 fd6_emit_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
588 struct fd6_emit *emit)
589 {
590 struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
591 const struct ir3_shader_variant *vp = fd6_emit_get_vp(emit);
592 const struct ir3_shader_variant *fp = fd6_emit_get_fp(emit);
593 const enum fd_dirty_3d_state dirty = emit->dirty;
594 bool needs_border = false;
595
596 emit_marker6(ring, 5);
597
598 if ((dirty & FD_DIRTY_FRAMEBUFFER) && !emit->binning_pass) {
599 unsigned char mrt_comp[A6XX_MAX_RENDER_TARGETS] = {0};
600
601 for (unsigned i = 0; i < pfb->nr_cbufs; i++) {
602 mrt_comp[i] = ((i < pfb->nr_cbufs) && pfb->cbufs[i]) ? 0xf : 0;
603 }
604
605 OUT_PKT4(ring, REG_A6XX_RB_RENDER_COMPONENTS, 1);
606 OUT_RING(ring, A6XX_RB_RENDER_COMPONENTS_RT0(mrt_comp[0]) |
607 A6XX_RB_RENDER_COMPONENTS_RT1(mrt_comp[1]) |
608 A6XX_RB_RENDER_COMPONENTS_RT2(mrt_comp[2]) |
609 A6XX_RB_RENDER_COMPONENTS_RT3(mrt_comp[3]) |
610 A6XX_RB_RENDER_COMPONENTS_RT4(mrt_comp[4]) |
611 A6XX_RB_RENDER_COMPONENTS_RT5(mrt_comp[5]) |
612 A6XX_RB_RENDER_COMPONENTS_RT6(mrt_comp[6]) |
613 A6XX_RB_RENDER_COMPONENTS_RT7(mrt_comp[7]));
614
615 OUT_PKT4(ring, REG_A6XX_SP_FS_RENDER_COMPONENTS, 1);
616 OUT_RING(ring,
617 A6XX_SP_FS_RENDER_COMPONENTS_RT0(mrt_comp[0]) |
618 A6XX_SP_FS_RENDER_COMPONENTS_RT1(mrt_comp[1]) |
619 A6XX_SP_FS_RENDER_COMPONENTS_RT2(mrt_comp[2]) |
620 A6XX_SP_FS_RENDER_COMPONENTS_RT3(mrt_comp[3]) |
621 A6XX_SP_FS_RENDER_COMPONENTS_RT4(mrt_comp[4]) |
622 A6XX_SP_FS_RENDER_COMPONENTS_RT5(mrt_comp[5]) |
623 A6XX_SP_FS_RENDER_COMPONENTS_RT6(mrt_comp[6]) |
624 A6XX_SP_FS_RENDER_COMPONENTS_RT7(mrt_comp[7]));
625 }
626
627 if (dirty & FD_DIRTY_ZSA) {
628 struct fd6_zsa_stateobj *zsa = fd6_zsa_stateobj(ctx->zsa);
629 uint32_t rb_alpha_control = zsa->rb_alpha_control;
630
631 if (util_format_is_pure_integer(pipe_surface_format(pfb->cbufs[0])))
632 rb_alpha_control &= ~A6XX_RB_ALPHA_CONTROL_ALPHA_TEST;
633
634 OUT_PKT4(ring, REG_A6XX_RB_ALPHA_CONTROL, 1);
635 OUT_RING(ring, rb_alpha_control);
636
637 OUT_PKT4(ring, REG_A6XX_RB_STENCIL_CONTROL, 1);
638 OUT_RING(ring, zsa->rb_stencil_control);
639 }
640
641 if (dirty & (FD_DIRTY_ZSA | FD_DIRTY_BLEND | FD_DIRTY_PROG)) {
642 struct fd6_zsa_stateobj *zsa = fd6_zsa_stateobj(ctx->zsa);
643
644 if (pfb->zsbuf) {
645 struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
646 uint32_t gras_lrz_cntl = zsa->gras_lrz_cntl;
647 uint32_t rb_lrz_cntl = zsa->rb_lrz_cntl;
648
649 if (emit->no_lrz_write || !rsc->lrz || !rsc->lrz_valid) {
650 gras_lrz_cntl = 0;
651 rb_lrz_cntl = 0;
652 } else if (emit->binning_pass && zsa->lrz_write) {
653 gras_lrz_cntl |= A6XX_GRAS_LRZ_CNTL_LRZ_WRITE;
654 }
655
656 OUT_PKT4(ring, REG_A6XX_GRAS_LRZ_CNTL, 1);
657 OUT_RING(ring, gras_lrz_cntl);
658
659 OUT_PKT4(ring, REG_A6XX_RB_LRZ_CNTL, 1);
660 OUT_RING(ring, rb_lrz_cntl);
661 }
662 }
663
664 if (dirty & (FD_DIRTY_ZSA | FD_DIRTY_STENCIL_REF)) {
665 struct fd6_zsa_stateobj *zsa = fd6_zsa_stateobj(ctx->zsa);
666 struct pipe_stencil_ref *sr = &ctx->stencil_ref;
667
668 OUT_PKT4(ring, REG_A6XX_RB_STENCILREF, 3);
669 OUT_RING(ring, A6XX_RB_STENCILREF_REF(sr->ref_value[0]) |
670 A6XX_RB_STENCILREF_BFREF(sr->ref_value[1]));
671 OUT_RING(ring, zsa->rb_stencilmask);
672 OUT_RING(ring, zsa->rb_stencilwrmask);
673 }
674
675 if (dirty & (FD_DIRTY_ZSA | FD_DIRTY_RASTERIZER | FD_DIRTY_PROG)) {
676 struct fd6_zsa_stateobj *zsa = fd6_zsa_stateobj(ctx->zsa);
677 bool fragz = fp->has_kill | fp->writes_pos;
678
679 OUT_PKT4(ring, REG_A6XX_RB_DEPTH_CNTL, 1);
680 OUT_RING(ring, zsa->rb_depth_cntl);
681
682 OUT_PKT4(ring, REG_A6XX_RB_DEPTH_PLANE_CNTL, 1);
683 OUT_RING(ring, COND(fragz, A6XX_RB_DEPTH_PLANE_CNTL_FRAG_WRITES_Z));
684
685 OUT_PKT4(ring, REG_A6XX_GRAS_SU_DEPTH_PLANE_CNTL, 1);
686 OUT_RING(ring, COND(fragz, A6XX_GRAS_SU_DEPTH_PLANE_CNTL_FRAG_WRITES_Z));
687 }
688
689 /* NOTE: scissor enabled bit is part of rasterizer state: */
690 if (dirty & (FD_DIRTY_SCISSOR | FD_DIRTY_RASTERIZER)) {
691 struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
692
693 OUT_PKT4(ring, REG_A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0, 2);
694 OUT_RING(ring, A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0_X(scissor->minx) |
695 A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0_Y(scissor->miny));
696 OUT_RING(ring, A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0_X(scissor->maxx - 1) |
697 A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0_Y(scissor->maxy - 1));
698
699 OUT_PKT4(ring, REG_A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0, 2);
700 OUT_RING(ring, A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0_X(scissor->minx) |
701 A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0_Y(scissor->miny));
702 OUT_RING(ring, A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0_X(scissor->maxx - 1) |
703 A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0_Y(scissor->maxy - 1));
704
705 ctx->batch->max_scissor.minx = MIN2(ctx->batch->max_scissor.minx, scissor->minx);
706 ctx->batch->max_scissor.miny = MIN2(ctx->batch->max_scissor.miny, scissor->miny);
707 ctx->batch->max_scissor.maxx = MAX2(ctx->batch->max_scissor.maxx, scissor->maxx);
708 ctx->batch->max_scissor.maxy = MAX2(ctx->batch->max_scissor.maxy, scissor->maxy);
709 }
710
711 if (dirty & FD_DIRTY_VIEWPORT) {
712 fd_wfi(ctx->batch, ring);
713 OUT_PKT4(ring, REG_A6XX_GRAS_CL_VPORT_XOFFSET_0, 6);
714 OUT_RING(ring, A6XX_GRAS_CL_VPORT_XOFFSET_0(ctx->viewport.translate[0]));
715 OUT_RING(ring, A6XX_GRAS_CL_VPORT_XSCALE_0(ctx->viewport.scale[0]));
716 OUT_RING(ring, A6XX_GRAS_CL_VPORT_YOFFSET_0(ctx->viewport.translate[1]));
717 OUT_RING(ring, A6XX_GRAS_CL_VPORT_YSCALE_0(ctx->viewport.scale[1]));
718 OUT_RING(ring, A6XX_GRAS_CL_VPORT_ZOFFSET_0(ctx->viewport.translate[2]));
719 OUT_RING(ring, A6XX_GRAS_CL_VPORT_ZSCALE_0(ctx->viewport.scale[2]));
720 }
721
722 if (dirty & FD_DIRTY_PROG)
723 fd6_program_emit(ctx, ring, emit);
724
725 if (dirty & FD_DIRTY_RASTERIZER) {
726 struct fd6_rasterizer_stateobj *rasterizer =
727 fd6_rasterizer_stateobj(ctx->rasterizer);
728
729 OUT_PKT4(ring, REG_A6XX_GRAS_UNKNOWN_8000, 1);
730 OUT_RING(ring, 0x80);
731 OUT_PKT4(ring, REG_A6XX_GRAS_UNKNOWN_8001, 1);
732 OUT_RING(ring, 0x0);
733 OUT_PKT4(ring, REG_A6XX_GRAS_UNKNOWN_8004, 1);
734 OUT_RING(ring, 0x0);
735
736 OUT_PKT4(ring, REG_A6XX_GRAS_SU_CNTL, 1);
737 OUT_RING(ring, rasterizer->gras_su_cntl);
738
739 OUT_PKT4(ring, REG_A6XX_GRAS_SU_POINT_MINMAX, 2);
740 OUT_RING(ring, rasterizer->gras_su_point_minmax);
741 OUT_RING(ring, rasterizer->gras_su_point_size);
742
743 OUT_PKT4(ring, REG_A6XX_GRAS_SU_POLY_OFFSET_SCALE, 3);
744 OUT_RING(ring, rasterizer->gras_su_poly_offset_scale);
745 OUT_RING(ring, rasterizer->gras_su_poly_offset_offset);
746 OUT_RING(ring, rasterizer->gras_su_poly_offset_clamp);
747
748 #if 0
749 OUT_PKT4(ring, REG_A6XX_PC_RASTER_CNTL, 1);
750 OUT_RING(ring, rasterizer->pc_raster_cntl);
751
752 OUT_PKT4(ring, REG_A6XX_GRAS_CL_CNTL, 1);
753 OUT_RING(ring, rasterizer->gras_cl_clip_cntl);
754 #endif
755 }
756
757 /* note: must come after program emit.. because there is some overlap
758 * in registers, ex. PC_PRIMITIVE_CNTL and we rely on some cached
759 * values from fd6_program_emit() to avoid having to re-emit the prog
760 * every time rast state changes.
761 *
762 * Since the primitive restart state is not part of a tracked object, we
763 * re-emit this register every time.
764 */
765 if (emit->info && ctx->rasterizer) {
766 struct fd6_rasterizer_stateobj *rasterizer =
767 fd6_rasterizer_stateobj(ctx->rasterizer);
768 OUT_PKT4(ring, REG_A6XX_PC_UNKNOWN_9806, 1);
769 OUT_RING(ring, 0);
770 OUT_PKT4(ring, REG_A6XX_PC_UNKNOWN_9990, 1);
771 OUT_RING(ring, 0);
772 OUT_PKT4(ring, REG_A6XX_VFD_UNKNOWN_A008, 1);
773 OUT_RING(ring, 0);
774
775
776 OUT_PKT4(ring, REG_A6XX_PC_PRIMITIVE_CNTL_0, 1);
777 OUT_RING(ring, rasterizer->pc_primitive_cntl |
778 COND(emit->info->primitive_restart && emit->info->index_size,
779 A6XX_PC_PRIMITIVE_CNTL_0_PRIMITIVE_RESTART));
780 }
781
782 if (dirty & (FD_DIRTY_FRAMEBUFFER | FD_DIRTY_RASTERIZER | FD_DIRTY_PROG)) {
783 uint32_t posz_regid = ir3_find_output_regid(fp, FRAG_RESULT_DEPTH);
784 unsigned nr = pfb->nr_cbufs;
785
786 if (emit->binning_pass)
787 nr = 0;
788 else if (ctx->rasterizer->rasterizer_discard)
789 nr = 0;
790
791 OUT_PKT4(ring, REG_A6XX_RB_FS_OUTPUT_CNTL0, 2);
792 OUT_RING(ring, COND(fp->writes_pos, A6XX_RB_FS_OUTPUT_CNTL0_FRAG_WRITES_Z));
793 OUT_RING(ring, A6XX_RB_FS_OUTPUT_CNTL1_MRT(nr));
794
795 OUT_PKT4(ring, REG_A6XX_SP_FS_OUTPUT_CNTL0, 2);
796 OUT_RING(ring, A6XX_SP_FS_OUTPUT_CNTL0_DEPTH_REGID(posz_regid) |
797 0xfcfc0000);
798 OUT_RING(ring, A6XX_SP_FS_OUTPUT_CNTL1_MRT(nr));
799 }
800
801 #define DIRTY_CONST (FD_DIRTY_SHADER_PROG | FD_DIRTY_SHADER_CONST | \
802 FD_DIRTY_SHADER_SSBO | FD_DIRTY_SHADER_IMAGE)
803
804 if (ctx->dirty_shader[PIPE_SHADER_VERTEX] & DIRTY_CONST) {
805 struct fd_ringbuffer *vsconstobj =
806 fd_ringbuffer_new_flags(ctx->pipe, 0x1000,
807 FD_RINGBUFFER_OBJECT | FD_RINGBUFFER_STREAMING);
808
809 ir3_emit_vs_consts(vp, vsconstobj, ctx, emit->info);
810 fd6_emit_add_group(emit, vsconstobj, FD6_GROUP_VS_CONST, 0x7);
811 fd_ringbuffer_del(vsconstobj);
812 }
813
814 if ((ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & DIRTY_CONST) &&
815 !emit->binning_pass) {
816 struct fd_ringbuffer *fsconstobj =
817 fd_ringbuffer_new_flags(ctx->pipe, 0x1000,
818 FD_RINGBUFFER_OBJECT | FD_RINGBUFFER_STREAMING);
819
820 ir3_emit_fs_consts(fp, fsconstobj, ctx);
821 fd6_emit_add_group(emit, fsconstobj, FD6_GROUP_FS_CONST, 0x7);
822 fd_ringbuffer_del(fsconstobj);
823 }
824
825 struct pipe_stream_output_info *info = &vp->shader->stream_output;
826 if (info->num_outputs) {
827 struct fd_streamout_stateobj *so = &ctx->streamout;
828
829 emit->streamout_mask = 0;
830
831 for (unsigned i = 0; i < so->num_targets; i++) {
832 struct pipe_stream_output_target *target = so->targets[i];
833
834 if (!target)
835 continue;
836
837 unsigned offset = (so->offsets[i] * info->stride[i] * 4) +
838 target->buffer_offset;
839
840 OUT_PKT4(ring, REG_A6XX_VPC_SO_BUFFER_BASE_LO(i), 3);
841 /* VPC_SO[i].BUFFER_BASE_LO: */
842 OUT_RELOCW(ring, fd_resource(target->buffer)->bo, 0, 0, 0);
843 OUT_RING(ring, target->buffer_size + offset);
844
845 OUT_PKT4(ring, REG_A6XX_VPC_SO_BUFFER_OFFSET(i), 3);
846 OUT_RING(ring, offset);
847 /* VPC_SO[i].FLUSH_BASE_LO/HI: */
848 // TODO just give hw a dummy addr for now.. we should
849 // be using this an then CP_MEM_TO_REG to set the
850 // VPC_SO[i].BUFFER_OFFSET for the next draw..
851 OUT_RELOCW(ring, fd6_context(ctx)->blit_mem, 0x100, 0, 0);
852
853 emit->streamout_mask |= (1 << i);
854 }
855
856 if (emit->streamout_mask) {
857 struct fd6_streamout_state *tf = &fd6_context(ctx)->tf;
858
859 OUT_PKT7(ring, CP_CONTEXT_REG_BUNCH, 12 + (2 * tf->prog_count));
860 OUT_RING(ring, REG_A6XX_VPC_SO_BUF_CNTL);
861 OUT_RING(ring, tf->vpc_so_buf_cntl);
862 OUT_RING(ring, REG_A6XX_VPC_SO_NCOMP(0));
863 OUT_RING(ring, tf->ncomp[0]);
864 OUT_RING(ring, REG_A6XX_VPC_SO_NCOMP(1));
865 OUT_RING(ring, tf->ncomp[1]);
866 OUT_RING(ring, REG_A6XX_VPC_SO_NCOMP(2));
867 OUT_RING(ring, tf->ncomp[2]);
868 OUT_RING(ring, REG_A6XX_VPC_SO_NCOMP(3));
869 OUT_RING(ring, tf->ncomp[3]);
870 OUT_RING(ring, REG_A6XX_VPC_SO_CNTL);
871 OUT_RING(ring, A6XX_VPC_SO_CNTL_ENABLE);
872 for (unsigned i = 0; i < tf->prog_count; i++) {
873 OUT_RING(ring, REG_A6XX_VPC_SO_PROG);
874 OUT_RING(ring, tf->prog[i]);
875 }
876
877 OUT_PKT4(ring, REG_A6XX_VPC_SO_OVERRIDE, 1);
878 OUT_RING(ring, 0x0);
879 } else {
880 OUT_PKT7(ring, CP_CONTEXT_REG_BUNCH, 4);
881 OUT_RING(ring, REG_A6XX_VPC_SO_CNTL);
882 OUT_RING(ring, 0);
883 OUT_RING(ring, REG_A6XX_VPC_SO_BUF_CNTL);
884 OUT_RING(ring, 0);
885
886 OUT_PKT4(ring, REG_A6XX_VPC_SO_OVERRIDE, 1);
887 OUT_RING(ring, A6XX_VPC_SO_OVERRIDE_SO_DISABLE);
888 }
889 }
890
891 if ((dirty & FD_DIRTY_BLEND)) {
892 struct fd6_blend_stateobj *blend = fd6_blend_stateobj(ctx->blend);
893 uint32_t i;
894
895 for (i = 0; i < A6XX_MAX_RENDER_TARGETS; i++) {
896 enum pipe_format format = pipe_surface_format(pfb->cbufs[i]);
897 bool is_int = util_format_is_pure_integer(format);
898 bool has_alpha = util_format_has_alpha(format);
899 uint32_t control = blend->rb_mrt[i].control;
900 uint32_t blend_control = blend->rb_mrt[i].blend_control_alpha;
901
902 if (is_int) {
903 control &= A6XX_RB_MRT_CONTROL_COMPONENT_ENABLE__MASK;
904 control |= A6XX_RB_MRT_CONTROL_ROP_CODE(ROP_COPY);
905 }
906
907 if (has_alpha) {
908 blend_control |= blend->rb_mrt[i].blend_control_rgb;
909 } else {
910 blend_control |= blend->rb_mrt[i].blend_control_no_alpha_rgb;
911 control &= ~A6XX_RB_MRT_CONTROL_BLEND2;
912 }
913
914 OUT_PKT4(ring, REG_A6XX_RB_MRT_CONTROL(i), 1);
915 OUT_RING(ring, control);
916
917 OUT_PKT4(ring, REG_A6XX_RB_MRT_BLEND_CONTROL(i), 1);
918 OUT_RING(ring, blend_control);
919 }
920
921 OUT_PKT4(ring, REG_A6XX_RB_BLEND_CNTL, 1);
922 OUT_RING(ring, blend->rb_blend_cntl |
923 A6XX_RB_BLEND_CNTL_SAMPLE_MASK(0xffff));
924
925 OUT_PKT4(ring, REG_A6XX_SP_BLEND_CNTL, 1);
926 OUT_RING(ring, blend->sp_blend_cntl);
927 }
928
929 if (dirty & FD_DIRTY_BLEND_COLOR) {
930 struct pipe_blend_color *bcolor = &ctx->blend_color;
931
932 OUT_PKT4(ring, REG_A6XX_RB_BLEND_RED_F32, 4);
933 OUT_RING(ring, A6XX_RB_BLEND_RED_F32(bcolor->color[0]));
934 OUT_RING(ring, A6XX_RB_BLEND_GREEN_F32(bcolor->color[1]));
935 OUT_RING(ring, A6XX_RB_BLEND_BLUE_F32(bcolor->color[2]));
936 OUT_RING(ring, A6XX_RB_BLEND_ALPHA_F32(bcolor->color[3]));
937 }
938
939 if ((ctx->dirty_shader[PIPE_SHADER_VERTEX] & FD_DIRTY_SHADER_TEX) &&
940 ctx->tex[PIPE_SHADER_VERTEX].num_textures > 0) {
941 struct fd6_texture_state *tex = fd6_texture_state(ctx,
942 SB6_VS_TEX, &ctx->tex[PIPE_SHADER_VERTEX]);
943
944 needs_border |= tex->needs_border;
945
946 fd6_emit_add_group(emit, tex->stateobj, FD6_GROUP_VS_TEX, 0x7);
947 }
948
949 if ((ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_TEX) &&
950 ctx->tex[PIPE_SHADER_FRAGMENT].num_textures > 0) {
951 struct fd6_texture_state *tex = fd6_texture_state(ctx,
952 SB6_FS_TEX, &ctx->tex[PIPE_SHADER_FRAGMENT]);
953
954 needs_border |= tex->needs_border;
955
956 fd6_emit_add_group(emit, tex->stateobj, FD6_GROUP_FS_TEX, 0x7);
957 }
958
959 if (needs_border)
960 emit_border_color(ctx, ring);
961
962 if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_SSBO)
963 emit_ssbos(ctx, ring, SB6_SSBO, &ctx->shaderbuf[PIPE_SHADER_FRAGMENT]);
964
965 if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_IMAGE)
966 fd6_emit_images(ctx, ring, PIPE_SHADER_FRAGMENT);
967
968 if (emit->num_groups > 0) {
969 OUT_PKT7(ring, CP_SET_DRAW_STATE, 3 * emit->num_groups);
970 for (unsigned i = 0; i < emit->num_groups; i++) {
971 struct fd6_state_group *g = &emit->groups[i];
972 unsigned n = fd_ringbuffer_size(g->stateobj) / 4;
973
974 OUT_RING(ring, CP_SET_DRAW_STATE__0_COUNT(n) |
975 CP_SET_DRAW_STATE__0_ENABLE_MASK(g->enable_mask) |
976 CP_SET_DRAW_STATE__0_GROUP_ID(g->group_id));
977 OUT_RB(ring, g->stateobj);
978
979 fd_ringbuffer_del(g->stateobj);
980 }
981 emit->num_groups = 0;
982 }
983 }
984
985 void
986 fd6_emit_cs_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
987 struct ir3_shader_variant *cp)
988 {
989 enum fd_dirty_shader_state dirty = ctx->dirty_shader[PIPE_SHADER_COMPUTE];
990
991 if (dirty & FD_DIRTY_SHADER_TEX) {
992 bool needs_border = false;
993 needs_border |= fd6_emit_textures(ctx->pipe, ring, SB6_CS_TEX,
994 &ctx->tex[PIPE_SHADER_COMPUTE], 0);
995
996 if (needs_border)
997 emit_border_color(ctx, ring);
998
999 #if 0
1000 OUT_PKT4(ring, REG_A6XX_TPL1_VS_TEX_COUNT, 1);
1001 OUT_RING(ring, 0);
1002
1003 OUT_PKT4(ring, REG_A6XX_TPL1_HS_TEX_COUNT, 1);
1004 OUT_RING(ring, 0);
1005
1006 OUT_PKT4(ring, REG_A6XX_TPL1_DS_TEX_COUNT, 1);
1007 OUT_RING(ring, 0);
1008
1009 OUT_PKT4(ring, REG_A6XX_TPL1_GS_TEX_COUNT, 1);
1010 OUT_RING(ring, 0);
1011
1012 OUT_PKT4(ring, REG_A6XX_TPL1_FS_TEX_COUNT, 1);
1013 OUT_RING(ring, 0);
1014 #endif
1015 }
1016
1017 #if 0
1018 OUT_PKT4(ring, REG_A6XX_TPL1_CS_TEX_COUNT, 1);
1019 OUT_RING(ring, ctx->shaderimg[PIPE_SHADER_COMPUTE].enabled_mask ?
1020 ~0 : ctx->tex[PIPE_SHADER_COMPUTE].num_textures);
1021 #endif
1022
1023 if (dirty & FD_DIRTY_SHADER_SSBO)
1024 emit_ssbos(ctx, ring, SB6_CS_SSBO, &ctx->shaderbuf[PIPE_SHADER_COMPUTE]);
1025
1026 if (dirty & FD_DIRTY_SHADER_IMAGE)
1027 fd6_emit_images(ctx, ring, PIPE_SHADER_COMPUTE);
1028 }
1029
1030
1031 /* emit setup at begin of new cmdstream buffer (don't rely on previous
1032 * state, there could have been a context switch between ioctls):
1033 */
1034 void
1035 fd6_emit_restore(struct fd_batch *batch, struct fd_ringbuffer *ring)
1036 {
1037 //struct fd_context *ctx = batch->ctx;
1038
1039 fd6_cache_flush(batch, ring);
1040
1041 OUT_PKT4(ring, REG_A6XX_HLSQ_UPDATE_CNTL, 1);
1042 OUT_RING(ring, 0xfffff);
1043
1044 /*
1045 t7 opcode: CP_PERFCOUNTER_ACTION (50) (4 dwords)
1046 0000000500024048: 70d08003 00000000 001c5000 00000005
1047 t7 opcode: CP_PERFCOUNTER_ACTION (50) (4 dwords)
1048 0000000500024058: 70d08003 00000010 001c7000 00000005
1049
1050 t7 opcode: CP_WAIT_FOR_IDLE (26) (1 dwords)
1051 0000000500024068: 70268000
1052 */
1053
1054 WRITE(REG_A6XX_RB_CCU_CNTL, 0x7c400004);
1055 WRITE(REG_A6XX_RB_UNKNOWN_8E04, 0x00100000);
1056 WRITE(REG_A6XX_SP_UNKNOWN_AE04, 0x8);
1057 WRITE(REG_A6XX_SP_UNKNOWN_AE00, 0);
1058 WRITE(REG_A6XX_SP_UNKNOWN_AE0F, 0x3f);
1059 WRITE(REG_A6XX_SP_UNKNOWN_B605, 0x44);
1060 WRITE(REG_A6XX_SP_UNKNOWN_B600, 0x100000);
1061 WRITE(REG_A6XX_HLSQ_UNKNOWN_BE00, 0x80);
1062 WRITE(REG_A6XX_HLSQ_UNKNOWN_BE01, 0);
1063
1064 WRITE(REG_A6XX_VPC_UNKNOWN_9600, 0);
1065 WRITE(REG_A6XX_GRAS_UNKNOWN_8600, 0x880);
1066 WRITE(REG_A6XX_HLSQ_UNKNOWN_BE04, 0);
1067 WRITE(REG_A6XX_SP_UNKNOWN_AE03, 0x00000410);
1068 WRITE(REG_A6XX_SP_UNKNOWN_AB20, 0);
1069 WRITE(REG_A6XX_SP_UNKNOWN_B182, 0);
1070 WRITE(REG_A6XX_HLSQ_UNKNOWN_BB11, 0);
1071 WRITE(REG_A6XX_UCHE_UNKNOWN_0E12, 0x3200000);
1072 WRITE(REG_A6XX_UCHE_CLIENT_PF, 4);
1073 WRITE(REG_A6XX_RB_UNKNOWN_8E01, 0x0);
1074 WRITE(REG_A6XX_SP_UNKNOWN_AB00, 0x5);
1075 WRITE(REG_A6XX_VFD_UNKNOWN_A009, 0x00000001);
1076 WRITE(REG_A6XX_RB_UNKNOWN_8811, 0x00000010);
1077 WRITE(REG_A6XX_PC_MODE_CNTL, 0x1f);
1078
1079 OUT_PKT4(ring, REG_A6XX_RB_SRGB_CNTL, 1);
1080 OUT_RING(ring, 0);
1081
1082 WRITE(REG_A6XX_GRAS_UNKNOWN_8101, 0);
1083 WRITE(REG_A6XX_GRAS_UNKNOWN_8109, 0);
1084 WRITE(REG_A6XX_GRAS_UNKNOWN_8110, 0);
1085
1086 WRITE(REG_A6XX_RB_RENDER_CONTROL0, 0x401);
1087 WRITE(REG_A6XX_RB_RENDER_CONTROL1, 0);
1088 WRITE(REG_A6XX_RB_FS_OUTPUT_CNTL0, 0);
1089 WRITE(REG_A6XX_RB_UNKNOWN_8810, 0);
1090 WRITE(REG_A6XX_RB_UNKNOWN_8818, 0);
1091 WRITE(REG_A6XX_RB_UNKNOWN_8819, 0);
1092 WRITE(REG_A6XX_RB_UNKNOWN_881A, 0);
1093 WRITE(REG_A6XX_RB_UNKNOWN_881B, 0);
1094 WRITE(REG_A6XX_RB_UNKNOWN_881C, 0);
1095 WRITE(REG_A6XX_RB_UNKNOWN_881D, 0);
1096 WRITE(REG_A6XX_RB_UNKNOWN_881E, 0);
1097 WRITE(REG_A6XX_RB_UNKNOWN_88F0, 0);
1098
1099 WRITE(REG_A6XX_VPC_UNKNOWN_9101, 0xffff00);
1100 WRITE(REG_A6XX_VPC_UNKNOWN_9107, 0);
1101
1102 WRITE(REG_A6XX_VPC_UNKNOWN_9236, 1);
1103 WRITE(REG_A6XX_VPC_UNKNOWN_9300, 0);
1104
1105 WRITE(REG_A6XX_VPC_SO_OVERRIDE, A6XX_VPC_SO_OVERRIDE_SO_DISABLE);
1106
1107 WRITE(REG_A6XX_PC_UNKNOWN_9801, 0);
1108 WRITE(REG_A6XX_PC_UNKNOWN_9806, 0);
1109 WRITE(REG_A6XX_PC_UNKNOWN_9980, 0);
1110
1111 WRITE(REG_A6XX_PC_UNKNOWN_9B06, 0);
1112 WRITE(REG_A6XX_PC_UNKNOWN_9B06, 0);
1113
1114 WRITE(REG_A6XX_SP_UNKNOWN_A81B, 0);
1115
1116 WRITE(REG_A6XX_SP_UNKNOWN_B183, 0);
1117
1118 WRITE(REG_A6XX_GRAS_UNKNOWN_8099, 0);
1119 WRITE(REG_A6XX_GRAS_UNKNOWN_809B, 0);
1120 WRITE(REG_A6XX_GRAS_UNKNOWN_80A0, 2);
1121 WRITE(REG_A6XX_GRAS_UNKNOWN_80AF, 0);
1122 WRITE(REG_A6XX_VPC_UNKNOWN_9210, 0);
1123 WRITE(REG_A6XX_VPC_UNKNOWN_9211, 0);
1124 WRITE(REG_A6XX_VPC_UNKNOWN_9602, 0);
1125 WRITE(REG_A6XX_PC_UNKNOWN_9981, 0x3);
1126 WRITE(REG_A6XX_PC_UNKNOWN_9E72, 0);
1127 WRITE(REG_A6XX_VPC_UNKNOWN_9108, 0x3);
1128 WRITE(REG_A6XX_SP_TP_UNKNOWN_B304, 0);
1129 WRITE(REG_A6XX_SP_TP_UNKNOWN_B309, 0x000000a2);
1130 WRITE(REG_A6XX_RB_UNKNOWN_8804, 0);
1131 WRITE(REG_A6XX_GRAS_UNKNOWN_80A4, 0);
1132 WRITE(REG_A6XX_GRAS_UNKNOWN_80A5, 0);
1133 WRITE(REG_A6XX_GRAS_UNKNOWN_80A6, 0);
1134 WRITE(REG_A6XX_RB_UNKNOWN_8805, 0);
1135 WRITE(REG_A6XX_RB_UNKNOWN_8806, 0);
1136 WRITE(REG_A6XX_RB_UNKNOWN_8878, 0);
1137 WRITE(REG_A6XX_RB_UNKNOWN_8879, 0);
1138 WRITE(REG_A6XX_HLSQ_CONTROL_5_REG, 0xfc);
1139
1140 emit_marker6(ring, 7);
1141
1142 OUT_PKT4(ring, REG_A6XX_VFD_MODE_CNTL, 1);
1143 OUT_RING(ring, 0x00000000); /* VFD_MODE_CNTL */
1144
1145 WRITE(REG_A6XX_VFD_UNKNOWN_A008, 0);
1146
1147 OUT_PKT4(ring, REG_A6XX_PC_MODE_CNTL, 1);
1148 OUT_RING(ring, 0x0000001f); /* PC_MODE_CNTL */
1149
1150 /* we don't use this yet.. probably best to disable.. */
1151 OUT_PKT7(ring, CP_SET_DRAW_STATE, 3);
1152 OUT_RING(ring, CP_SET_DRAW_STATE__0_COUNT(0) |
1153 CP_SET_DRAW_STATE__0_DISABLE_ALL_GROUPS |
1154 CP_SET_DRAW_STATE__0_GROUP_ID(0));
1155 OUT_RING(ring, CP_SET_DRAW_STATE__1_ADDR_LO(0));
1156 OUT_RING(ring, CP_SET_DRAW_STATE__2_ADDR_HI(0));
1157
1158 OUT_PKT4(ring, REG_A6XX_VPC_SO_BUFFER_BASE_LO(0), 3);
1159 OUT_RING(ring, 0x00000000); /* VPC_SO_BUFFER_BASE_LO_0 */
1160 OUT_RING(ring, 0x00000000); /* VPC_SO_BUFFER_BASE_HI_0 */
1161 OUT_RING(ring, 0x00000000); /* VPC_SO_BUFFER_SIZE_0 */
1162
1163 OUT_PKT4(ring, REG_A6XX_VPC_SO_FLUSH_BASE_LO(0), 2);
1164 OUT_RING(ring, 0x00000000); /* VPC_SO_FLUSH_BASE_LO_0 */
1165 OUT_RING(ring, 0x00000000); /* VPC_SO_FLUSH_BASE_HI_0 */
1166
1167 OUT_PKT4(ring, REG_A6XX_VPC_SO_BUF_CNTL, 1);
1168 OUT_RING(ring, 0x00000000); /* VPC_SO_BUF_CNTL */
1169
1170 OUT_PKT4(ring, REG_A6XX_VPC_SO_BUFFER_OFFSET(0), 1);
1171 OUT_RING(ring, 0x00000000); /* UNKNOWN_E2AB */
1172
1173 OUT_PKT4(ring, REG_A6XX_VPC_SO_BUFFER_BASE_LO(1), 3);
1174 OUT_RING(ring, 0x00000000);
1175 OUT_RING(ring, 0x00000000);
1176 OUT_RING(ring, 0x00000000);
1177
1178 OUT_PKT4(ring, REG_A6XX_VPC_SO_BUFFER_OFFSET(1), 6);
1179 OUT_RING(ring, 0x00000000);
1180 OUT_RING(ring, 0x00000000);
1181 OUT_RING(ring, 0x00000000);
1182 OUT_RING(ring, 0x00000000);
1183 OUT_RING(ring, 0x00000000);
1184 OUT_RING(ring, 0x00000000);
1185
1186 OUT_PKT4(ring, REG_A6XX_VPC_SO_BUFFER_OFFSET(2), 6);
1187 OUT_RING(ring, 0x00000000);
1188 OUT_RING(ring, 0x00000000);
1189 OUT_RING(ring, 0x00000000);
1190 OUT_RING(ring, 0x00000000);
1191 OUT_RING(ring, 0x00000000);
1192 OUT_RING(ring, 0x00000000);
1193
1194 OUT_PKT4(ring, REG_A6XX_VPC_SO_BUFFER_OFFSET(3), 3);
1195 OUT_RING(ring, 0x00000000);
1196 OUT_RING(ring, 0x00000000);
1197 OUT_RING(ring, 0x00000000);
1198
1199 OUT_PKT4(ring, REG_A6XX_SP_HS_CTRL_REG0, 1);
1200 OUT_RING(ring, 0x00000000);
1201
1202 OUT_PKT4(ring, REG_A6XX_SP_GS_CTRL_REG0, 1);
1203 OUT_RING(ring, 0x00000000);
1204
1205 OUT_PKT4(ring, REG_A6XX_GRAS_LRZ_CNTL, 1);
1206 OUT_RING(ring, 0x00000000);
1207
1208 OUT_PKT4(ring, REG_A6XX_RB_LRZ_CNTL, 1);
1209 OUT_RING(ring, 0x00000000);
1210 }
1211
1212 static void
1213 fd6_emit_ib(struct fd_ringbuffer *ring, struct fd_ringbuffer *target)
1214 {
1215 emit_marker6(ring, 6);
1216 __OUT_IB5(ring, target);
1217 emit_marker6(ring, 6);
1218 }
1219
1220 static void
1221 fd6_mem_to_mem(struct fd_ringbuffer *ring, struct pipe_resource *dst,
1222 unsigned dst_off, struct pipe_resource *src, unsigned src_off,
1223 unsigned sizedwords)
1224 {
1225 struct fd_bo *src_bo = fd_resource(src)->bo;
1226 struct fd_bo *dst_bo = fd_resource(dst)->bo;
1227 unsigned i;
1228
1229 for (i = 0; i < sizedwords; i++) {
1230 OUT_PKT7(ring, CP_MEM_TO_MEM, 5);
1231 OUT_RING(ring, 0x00000000);
1232 OUT_RELOCW(ring, dst_bo, dst_off, 0, 0);
1233 OUT_RELOC (ring, src_bo, src_off, 0, 0);
1234
1235 dst_off += 4;
1236 src_off += 4;
1237 }
1238 }
1239
1240 void
1241 fd6_emit_init(struct pipe_context *pctx)
1242 {
1243 struct fd_context *ctx = fd_context(pctx);
1244 ctx->emit_const = fd6_emit_const;
1245 ctx->emit_const_bo = fd6_emit_const_bo;
1246 ctx->emit_ib = fd6_emit_ib;
1247 ctx->mem_to_mem = fd6_mem_to_mem;
1248 }