projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
freedreno/ir3: split kill from no_earlyz
[mesa.git] / src / freedreno / vulkan / tu_pipeline.c
1 /*
2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
4 *
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
17 * Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
25 * DEALINGS IN THE SOFTWARE.
26 */
27
28 #include "tu_private.h"
29
30 #include "ir3/ir3_nir.h"
31 #include "main/menums.h"
32 #include "nir/nir.h"
33 #include "nir/nir_builder.h"
34 #include "spirv/nir_spirv.h"
35 #include "util/debug.h"
36 #include "util/mesa-sha1.h"
37 #include "util/u_atomic.h"
38 #include "vk_format.h"
39 #include "vk_util.h"
40
41 #include "tu_cs.h"
42
43 /* Emit IB that preloads the descriptors that the shader uses */
44
45 static inline uint32_t
46 tu6_vkstage2opcode(VkShaderStageFlags stage)
47 {
48 switch (stage) {
49 case VK_SHADER_STAGE_VERTEX_BIT:
50 case VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT:
51 case VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT:
52 case VK_SHADER_STAGE_GEOMETRY_BIT:
53 return CP_LOAD_STATE6_GEOM;
54 case VK_SHADER_STAGE_FRAGMENT_BIT:
55 case VK_SHADER_STAGE_COMPUTE_BIT:
56 return CP_LOAD_STATE6_FRAG;
57 default:
58 unreachable("bad shader type");
59 }
60 }
61
62 static enum a6xx_state_block
63 tu6_tex_stage2sb(VkShaderStageFlags stage)
64 {
65 switch (stage) {
66 case VK_SHADER_STAGE_VERTEX_BIT:
67 return SB6_VS_TEX;
68 case VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT:
69 return SB6_HS_TEX;
70 case VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT:
71 return SB6_DS_TEX;
72 case VK_SHADER_STAGE_GEOMETRY_BIT:
73 return SB6_GS_TEX;
74 case VK_SHADER_STAGE_FRAGMENT_BIT:
75 return SB6_FS_TEX;
76 case VK_SHADER_STAGE_COMPUTE_BIT:
77 return SB6_CS_TEX;
78 default:
79 unreachable("bad shader stage");
80 }
81 }
82
83 static enum a6xx_state_block
84 tu6_ubo_stage2sb(VkShaderStageFlags stage)
85 {
86 switch (stage) {
87 case VK_SHADER_STAGE_VERTEX_BIT:
88 return SB6_VS_SHADER;
89 case VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT:
90 return SB6_HS_SHADER;
91 case VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT:
92 return SB6_DS_SHADER;
93 case VK_SHADER_STAGE_GEOMETRY_BIT:
94 return SB6_GS_SHADER;
95 case VK_SHADER_STAGE_FRAGMENT_BIT:
96 return SB6_FS_SHADER;
97 case VK_SHADER_STAGE_COMPUTE_BIT:
98 return SB6_CS_SHADER;
99 default:
100 unreachable("bad shader stage");
101 }
102 }
103
104 static void
105 emit_load_state(struct tu_cs *cs, unsigned opcode, enum a6xx_state_type st,
106 enum a6xx_state_block sb, unsigned base, unsigned offset,
107 unsigned count)
108 {
109 /* Note: just emit one packet, even if count overflows NUM_UNIT. It's not
110 * clear if emitting more packets will even help anything. Presumably the
111 * descriptor cache is relatively small, and these packets stop doing
112 * anything when there are too many descriptors.
113 */
114 tu_cs_emit_pkt7(cs, opcode, 3);
115 tu_cs_emit(cs,
116 CP_LOAD_STATE6_0_STATE_TYPE(st) |
117 CP_LOAD_STATE6_0_STATE_SRC(SS6_BINDLESS) |
118 CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
119 CP_LOAD_STATE6_0_NUM_UNIT(MIN2(count, 1024-1)));
120 tu_cs_emit_qw(cs, offset | (base << 28));
121 }
122
123 static unsigned
124 tu6_load_state_size(struct tu_pipeline_layout *layout, bool compute)
125 {
126 const unsigned load_state_size = 4;
127 unsigned size = 0;
128 for (unsigned i = 0; i < layout->num_sets; i++) {
129 struct tu_descriptor_set_layout *set_layout = layout->set[i].layout;
130 for (unsigned j = 0; j < set_layout->binding_count; j++) {
131 struct tu_descriptor_set_binding_layout *binding = &set_layout->binding[j];
132 unsigned count = 0;
133 /* Note: some users, like amber for example, pass in
134 * VK_SHADER_STAGE_ALL which includes a bunch of extra bits, so
135 * filter these out by using VK_SHADER_STAGE_ALL_GRAPHICS explicitly.
136 */
137 VkShaderStageFlags stages = compute ?
138 binding->shader_stages & VK_SHADER_STAGE_COMPUTE_BIT :
139 binding->shader_stages & VK_SHADER_STAGE_ALL_GRAPHICS;
140 unsigned stage_count = util_bitcount(stages);
141 switch (binding->type) {
142 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
143 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
144 case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
145 case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
146 /* IBO-backed resources only need one packet for all graphics stages */
147 if (stages & ~VK_SHADER_STAGE_COMPUTE_BIT)
148 count += 1;
149 if (stages & VK_SHADER_STAGE_COMPUTE_BIT)
150 count += 1;
151 break;
152 case VK_DESCRIPTOR_TYPE_SAMPLER:
153 case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
154 case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
155 case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
156 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
157 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
158 /* Textures and UBO's needs a packet for each stage */
159 count = stage_count;
160 break;
161 case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
162 /* Because of how we pack combined images and samplers, we
163 * currently can't use one packet for the whole array.
164 */
165 count = stage_count * binding->array_size * 2;
166 break;
167 default:
168 unreachable("bad descriptor type");
169 }
170 size += count * load_state_size;
171 }
172 }
173 return size;
174 }
175
176 static void
177 tu6_emit_load_state(struct tu_pipeline *pipeline, bool compute)
178 {
179 unsigned size = tu6_load_state_size(pipeline->layout, compute);
180 if (size == 0)
181 return;
182
183 struct tu_cs cs;
184 tu_cs_begin_sub_stream(&pipeline->cs, size, &cs);
185
186 struct tu_pipeline_layout *layout = pipeline->layout;
187 for (unsigned i = 0; i < layout->num_sets; i++) {
188 struct tu_descriptor_set_layout *set_layout = layout->set[i].layout;
189 for (unsigned j = 0; j < set_layout->binding_count; j++) {
190 struct tu_descriptor_set_binding_layout *binding = &set_layout->binding[j];
191 unsigned base = i;
192 unsigned offset = binding->offset / 4;
193 /* Note: some users, like amber for example, pass in
194 * VK_SHADER_STAGE_ALL which includes a bunch of extra bits, so
195 * filter these out by using VK_SHADER_STAGE_ALL_GRAPHICS explicitly.
196 */
197 VkShaderStageFlags stages = compute ?
198 binding->shader_stages & VK_SHADER_STAGE_COMPUTE_BIT :
199 binding->shader_stages & VK_SHADER_STAGE_ALL_GRAPHICS;
200 unsigned count = binding->array_size;
201 if (count == 0 || stages == 0)
202 continue;
203 switch (binding->type) {
204 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
205 base = MAX_SETS;
206 offset = (layout->input_attachment_count +
207 layout->set[i].dynamic_offset_start +
208 binding->dynamic_offset_offset) * A6XX_TEX_CONST_DWORDS;
209 /* fallthrough */
210 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
211 case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
212 case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
213 /* IBO-backed resources only need one packet for all graphics stages */
214 if (stages & ~VK_SHADER_STAGE_COMPUTE_BIT) {
215 emit_load_state(&cs, CP_LOAD_STATE6, ST6_SHADER, SB6_IBO,
216 base, offset, count);
217 }
218 if (stages & VK_SHADER_STAGE_COMPUTE_BIT) {
219 emit_load_state(&cs, CP_LOAD_STATE6_FRAG, ST6_IBO, SB6_CS_SHADER,
220 base, offset, count);
221 }
222 break;
223 case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
224 base = MAX_SETS;
225 offset = (layout->set[i].input_attachment_start +
226 binding->input_attachment_offset) * A6XX_TEX_CONST_DWORDS;
227 case VK_DESCRIPTOR_TYPE_SAMPLER:
228 case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
229 case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER: {
230 unsigned stage_log2;
231 for_each_bit(stage_log2, stages) {
232 VkShaderStageFlags stage = 1 << stage_log2;
233 emit_load_state(&cs, tu6_vkstage2opcode(stage),
234 binding->type == VK_DESCRIPTOR_TYPE_SAMPLER ?
235 ST6_SHADER : ST6_CONSTANTS,
236 tu6_tex_stage2sb(stage), base, offset, count);
237 }
238 break;
239 }
240 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
241 base = MAX_SETS;
242 offset = (layout->input_attachment_count +
243 layout->set[i].dynamic_offset_start +
244 binding->dynamic_offset_offset) * A6XX_TEX_CONST_DWORDS;
245 /* fallthrough */
246 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER: {
247 unsigned stage_log2;
248 for_each_bit(stage_log2, stages) {
249 VkShaderStageFlags stage = 1 << stage_log2;
250 emit_load_state(&cs, tu6_vkstage2opcode(stage), ST6_UBO,
251 tu6_ubo_stage2sb(stage), base, offset, count);
252 }
253 break;
254 }
255 case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: {
256 unsigned stage_log2;
257 for_each_bit(stage_log2, stages) {
258 VkShaderStageFlags stage = 1 << stage_log2;
259 /* TODO: We could emit less CP_LOAD_STATE6 if we used
260 * struct-of-arrays instead of array-of-structs.
261 */
262 for (unsigned i = 0; i < count; i++) {
263 unsigned tex_offset = offset + 2 * i * A6XX_TEX_CONST_DWORDS;
264 unsigned sam_offset = offset + (2 * i + 1) * A6XX_TEX_CONST_DWORDS;
265 emit_load_state(&cs, tu6_vkstage2opcode(stage),
266 ST6_CONSTANTS, tu6_tex_stage2sb(stage),
267 base, tex_offset, 1);
268 emit_load_state(&cs, tu6_vkstage2opcode(stage),
269 ST6_SHADER, tu6_tex_stage2sb(stage),
270 base, sam_offset, 1);
271 }
272 }
273 break;
274 }
275 default:
276 unreachable("bad descriptor type");
277 }
278 }
279 }
280
281 pipeline->load_state.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &cs);
282 }
283
284 struct tu_pipeline_builder
285 {
286 struct tu_device *device;
287 struct tu_pipeline_cache *cache;
288 struct tu_pipeline_layout *layout;
289 const VkAllocationCallbacks *alloc;
290 const VkGraphicsPipelineCreateInfo *create_info;
291
292 struct tu_shader *shaders[MESA_SHADER_STAGES];
293 uint32_t shader_offsets[MESA_SHADER_STAGES];
294 uint32_t binning_vs_offset;
295 uint32_t shader_total_size;
296
297 bool rasterizer_discard;
298 /* these states are affectd by rasterizer_discard */
299 VkSampleCountFlagBits samples;
300 bool use_color_attachments;
301 bool use_dual_src_blend;
302 uint32_t color_attachment_count;
303 VkFormat color_attachment_formats[MAX_RTS];
304 VkFormat depth_attachment_format;
305 uint32_t render_components;
306 };
307
308 static enum tu_dynamic_state_bits
309 tu_dynamic_state_bit(VkDynamicState state)
310 {
311 switch (state) {
312 case VK_DYNAMIC_STATE_VIEWPORT:
313 return TU_DYNAMIC_VIEWPORT;
314 case VK_DYNAMIC_STATE_SCISSOR:
315 return TU_DYNAMIC_SCISSOR;
316 case VK_DYNAMIC_STATE_LINE_WIDTH:
317 return TU_DYNAMIC_LINE_WIDTH;
318 case VK_DYNAMIC_STATE_DEPTH_BIAS:
319 return TU_DYNAMIC_DEPTH_BIAS;
320 case VK_DYNAMIC_STATE_BLEND_CONSTANTS:
321 return TU_DYNAMIC_BLEND_CONSTANTS;
322 case VK_DYNAMIC_STATE_DEPTH_BOUNDS:
323 return TU_DYNAMIC_DEPTH_BOUNDS;
324 case VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK:
325 return TU_DYNAMIC_STENCIL_COMPARE_MASK;
326 case VK_DYNAMIC_STATE_STENCIL_WRITE_MASK:
327 return TU_DYNAMIC_STENCIL_WRITE_MASK;
328 case VK_DYNAMIC_STATE_STENCIL_REFERENCE:
329 return TU_DYNAMIC_STENCIL_REFERENCE;
330 case VK_DYNAMIC_STATE_SAMPLE_LOCATIONS_EXT:
331 return TU_DYNAMIC_SAMPLE_LOCATIONS;
332 default:
333 unreachable("invalid dynamic state");
334 return 0;
335 }
336 }
337
338 static gl_shader_stage
339 tu_shader_stage(VkShaderStageFlagBits stage)
340 {
341 switch (stage) {
342 case VK_SHADER_STAGE_VERTEX_BIT:
343 return MESA_SHADER_VERTEX;
344 case VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT:
345 return MESA_SHADER_TESS_CTRL;
346 case VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT:
347 return MESA_SHADER_TESS_EVAL;
348 case VK_SHADER_STAGE_GEOMETRY_BIT:
349 return MESA_SHADER_GEOMETRY;
350 case VK_SHADER_STAGE_FRAGMENT_BIT:
351 return MESA_SHADER_FRAGMENT;
352 case VK_SHADER_STAGE_COMPUTE_BIT:
353 return MESA_SHADER_COMPUTE;
354 default:
355 unreachable("invalid VkShaderStageFlagBits");
356 return MESA_SHADER_NONE;
357 }
358 }
359
360 static bool
361 tu_logic_op_reads_dst(VkLogicOp op)
362 {
363 switch (op) {
364 case VK_LOGIC_OP_CLEAR:
365 case VK_LOGIC_OP_COPY:
366 case VK_LOGIC_OP_COPY_INVERTED:
367 case VK_LOGIC_OP_SET:
368 return false;
369 default:
370 return true;
371 }
372 }
373
374 static VkBlendFactor
375 tu_blend_factor_no_dst_alpha(VkBlendFactor factor)
376 {
377 /* treat dst alpha as 1.0 and avoid reading it */
378 switch (factor) {
379 case VK_BLEND_FACTOR_DST_ALPHA:
380 return VK_BLEND_FACTOR_ONE;
381 case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA:
382 return VK_BLEND_FACTOR_ZERO;
383 default:
384 return factor;
385 }
386 }
387
388 static bool tu_blend_factor_is_dual_src(VkBlendFactor factor)
389 {
390 switch (factor) {
391 case VK_BLEND_FACTOR_SRC1_COLOR:
392 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR:
393 case VK_BLEND_FACTOR_SRC1_ALPHA:
394 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA:
395 return true;
396 default:
397 return false;
398 }
399 }
400
401 static bool
402 tu_blend_state_is_dual_src(const VkPipelineColorBlendStateCreateInfo *info)
403 {
404 if (!info)
405 return false;
406
407 for (unsigned i = 0; i < info->attachmentCount; i++) {
408 const VkPipelineColorBlendAttachmentState *blend = &info->pAttachments[i];
409 if (tu_blend_factor_is_dual_src(blend->srcColorBlendFactor) ||
410 tu_blend_factor_is_dual_src(blend->dstColorBlendFactor) ||
411 tu_blend_factor_is_dual_src(blend->srcAlphaBlendFactor) ||
412 tu_blend_factor_is_dual_src(blend->dstAlphaBlendFactor))
413 return true;
414 }
415
416 return false;
417 }
418
419 static enum pc_di_primtype
420 tu6_primtype(VkPrimitiveTopology topology)
421 {
422 switch (topology) {
423 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
424 return DI_PT_POINTLIST;
425 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
426 return DI_PT_LINELIST;
427 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
428 return DI_PT_LINESTRIP;
429 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
430 return DI_PT_TRILIST;
431 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
432 return DI_PT_TRISTRIP;
433 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
434 return DI_PT_TRIFAN;
435 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
436 return DI_PT_LINE_ADJ;
437 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
438 return DI_PT_LINESTRIP_ADJ;
439 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
440 return DI_PT_TRI_ADJ;
441 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
442 return DI_PT_TRISTRIP_ADJ;
443 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST:
444 default:
445 unreachable("invalid primitive topology");
446 return DI_PT_NONE;
447 }
448 }
449
450 static enum adreno_compare_func
451 tu6_compare_func(VkCompareOp op)
452 {
453 switch (op) {
454 case VK_COMPARE_OP_NEVER:
455 return FUNC_NEVER;
456 case VK_COMPARE_OP_LESS:
457 return FUNC_LESS;
458 case VK_COMPARE_OP_EQUAL:
459 return FUNC_EQUAL;
460 case VK_COMPARE_OP_LESS_OR_EQUAL:
461 return FUNC_LEQUAL;
462 case VK_COMPARE_OP_GREATER:
463 return FUNC_GREATER;
464 case VK_COMPARE_OP_NOT_EQUAL:
465 return FUNC_NOTEQUAL;
466 case VK_COMPARE_OP_GREATER_OR_EQUAL:
467 return FUNC_GEQUAL;
468 case VK_COMPARE_OP_ALWAYS:
469 return FUNC_ALWAYS;
470 default:
471 unreachable("invalid VkCompareOp");
472 return FUNC_NEVER;
473 }
474 }
475
476 static enum adreno_stencil_op
477 tu6_stencil_op(VkStencilOp op)
478 {
479 switch (op) {
480 case VK_STENCIL_OP_KEEP:
481 return STENCIL_KEEP;
482 case VK_STENCIL_OP_ZERO:
483 return STENCIL_ZERO;
484 case VK_STENCIL_OP_REPLACE:
485 return STENCIL_REPLACE;
486 case VK_STENCIL_OP_INCREMENT_AND_CLAMP:
487 return STENCIL_INCR_CLAMP;
488 case VK_STENCIL_OP_DECREMENT_AND_CLAMP:
489 return STENCIL_DECR_CLAMP;
490 case VK_STENCIL_OP_INVERT:
491 return STENCIL_INVERT;
492 case VK_STENCIL_OP_INCREMENT_AND_WRAP:
493 return STENCIL_INCR_WRAP;
494 case VK_STENCIL_OP_DECREMENT_AND_WRAP:
495 return STENCIL_DECR_WRAP;
496 default:
497 unreachable("invalid VkStencilOp");
498 return STENCIL_KEEP;
499 }
500 }
501
502 static enum a3xx_rop_code
503 tu6_rop(VkLogicOp op)
504 {
505 switch (op) {
506 case VK_LOGIC_OP_CLEAR:
507 return ROP_CLEAR;
508 case VK_LOGIC_OP_AND:
509 return ROP_AND;
510 case VK_LOGIC_OP_AND_REVERSE:
511 return ROP_AND_REVERSE;
512 case VK_LOGIC_OP_COPY:
513 return ROP_COPY;
514 case VK_LOGIC_OP_AND_INVERTED:
515 return ROP_AND_INVERTED;
516 case VK_LOGIC_OP_NO_OP:
517 return ROP_NOOP;
518 case VK_LOGIC_OP_XOR:
519 return ROP_XOR;
520 case VK_LOGIC_OP_OR:
521 return ROP_OR;
522 case VK_LOGIC_OP_NOR:
523 return ROP_NOR;
524 case VK_LOGIC_OP_EQUIVALENT:
525 return ROP_EQUIV;
526 case VK_LOGIC_OP_INVERT:
527 return ROP_INVERT;
528 case VK_LOGIC_OP_OR_REVERSE:
529 return ROP_OR_REVERSE;
530 case VK_LOGIC_OP_COPY_INVERTED:
531 return ROP_COPY_INVERTED;
532 case VK_LOGIC_OP_OR_INVERTED:
533 return ROP_OR_INVERTED;
534 case VK_LOGIC_OP_NAND:
535 return ROP_NAND;
536 case VK_LOGIC_OP_SET:
537 return ROP_SET;
538 default:
539 unreachable("invalid VkLogicOp");
540 return ROP_NOOP;
541 }
542 }
543
544 static enum adreno_rb_blend_factor
545 tu6_blend_factor(VkBlendFactor factor)
546 {
547 switch (factor) {
548 case VK_BLEND_FACTOR_ZERO:
549 return FACTOR_ZERO;
550 case VK_BLEND_FACTOR_ONE:
551 return FACTOR_ONE;
552 case VK_BLEND_FACTOR_SRC_COLOR:
553 return FACTOR_SRC_COLOR;
554 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR:
555 return FACTOR_ONE_MINUS_SRC_COLOR;
556 case VK_BLEND_FACTOR_DST_COLOR:
557 return FACTOR_DST_COLOR;
558 case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR:
559 return FACTOR_ONE_MINUS_DST_COLOR;
560 case VK_BLEND_FACTOR_SRC_ALPHA:
561 return FACTOR_SRC_ALPHA;
562 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA:
563 return FACTOR_ONE_MINUS_SRC_ALPHA;
564 case VK_BLEND_FACTOR_DST_ALPHA:
565 return FACTOR_DST_ALPHA;
566 case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA:
567 return FACTOR_ONE_MINUS_DST_ALPHA;
568 case VK_BLEND_FACTOR_CONSTANT_COLOR:
569 return FACTOR_CONSTANT_COLOR;
570 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR:
571 return FACTOR_ONE_MINUS_CONSTANT_COLOR;
572 case VK_BLEND_FACTOR_CONSTANT_ALPHA:
573 return FACTOR_CONSTANT_ALPHA;
574 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA:
575 return FACTOR_ONE_MINUS_CONSTANT_ALPHA;
576 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE:
577 return FACTOR_SRC_ALPHA_SATURATE;
578 case VK_BLEND_FACTOR_SRC1_COLOR:
579 return FACTOR_SRC1_COLOR;
580 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR:
581 return FACTOR_ONE_MINUS_SRC1_COLOR;
582 case VK_BLEND_FACTOR_SRC1_ALPHA:
583 return FACTOR_SRC1_ALPHA;
584 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA:
585 return FACTOR_ONE_MINUS_SRC1_ALPHA;
586 default:
587 unreachable("invalid VkBlendFactor");
588 return FACTOR_ZERO;
589 }
590 }
591
592 static enum a3xx_rb_blend_opcode
593 tu6_blend_op(VkBlendOp op)
594 {
595 switch (op) {
596 case VK_BLEND_OP_ADD:
597 return BLEND_DST_PLUS_SRC;
598 case VK_BLEND_OP_SUBTRACT:
599 return BLEND_SRC_MINUS_DST;
600 case VK_BLEND_OP_REVERSE_SUBTRACT:
601 return BLEND_DST_MINUS_SRC;
602 case VK_BLEND_OP_MIN:
603 return BLEND_MIN_DST_SRC;
604 case VK_BLEND_OP_MAX:
605 return BLEND_MAX_DST_SRC;
606 default:
607 unreachable("invalid VkBlendOp");
608 return BLEND_DST_PLUS_SRC;
609 }
610 }
611
612 static uint32_t
613 emit_xs_config(const struct ir3_shader_variant *sh)
614 {
615 if (sh->instrlen) {
616 return A6XX_SP_VS_CONFIG_ENABLED |
617 COND(sh->bindless_tex, A6XX_SP_VS_CONFIG_BINDLESS_TEX) |
618 COND(sh->bindless_samp, A6XX_SP_VS_CONFIG_BINDLESS_SAMP) |
619 COND(sh->bindless_ibo, A6XX_SP_VS_CONFIG_BINDLESS_IBO) |
620 COND(sh->bindless_ubo, A6XX_SP_VS_CONFIG_BINDLESS_UBO);
621 } else {
622 return 0;
623 }
624 }
625
626 static void
627 tu6_emit_vs_config(struct tu_cs *cs, struct tu_shader *shader,
628 const struct ir3_shader_variant *vs)
629 {
630 uint32_t sp_vs_ctrl =
631 A6XX_SP_VS_CTRL_REG0_THREADSIZE(FOUR_QUADS) |
632 A6XX_SP_VS_CTRL_REG0_FULLREGFOOTPRINT(vs->info.max_reg + 1) |
633 A6XX_SP_VS_CTRL_REG0_MERGEDREGS |
634 A6XX_SP_VS_CTRL_REG0_BRANCHSTACK(vs->branchstack);
635 if (vs->need_pixlod)
636 sp_vs_ctrl |= A6XX_SP_VS_CTRL_REG0_PIXLODENABLE;
637 if (vs->need_fine_derivatives)
638 sp_vs_ctrl |= A6XX_SP_VS_CTRL_REG0_DIFF_FINE;
639
640 tu_cs_emit_pkt4(cs, REG_A6XX_SP_VS_CTRL_REG0, 1);
641 tu_cs_emit(cs, sp_vs_ctrl);
642
643 tu_cs_emit_pkt4(cs, REG_A6XX_SP_VS_CONFIG, 2);
644 tu_cs_emit(cs, emit_xs_config(vs));
645 tu_cs_emit(cs, vs->instrlen);
646
647 tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_VS_CNTL, 1);
648 tu_cs_emit(cs, A6XX_HLSQ_VS_CNTL_CONSTLEN(align(vs->constlen, 4)) |
649 A6XX_HLSQ_VS_CNTL_ENABLED);
650 }
651
652 static void
653 tu6_emit_hs_config(struct tu_cs *cs, struct tu_shader *shader,
654 const struct ir3_shader_variant *hs)
655 {
656 tu_cs_emit_pkt4(cs, REG_A6XX_SP_HS_UNKNOWN_A831, 1);
657 tu_cs_emit(cs, 0);
658
659 tu_cs_emit_pkt4(cs, REG_A6XX_SP_HS_CONFIG, 2);
660 tu_cs_emit(cs, emit_xs_config(hs));
661 tu_cs_emit(cs, hs->instrlen);
662
663 tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_HS_CNTL, 1);
664 tu_cs_emit(cs, A6XX_HLSQ_HS_CNTL_CONSTLEN(align(hs->constlen, 4)));
665 }
666
667 static void
668 tu6_emit_ds_config(struct tu_cs *cs, struct tu_shader *shader,
669 const struct ir3_shader_variant *ds)
670 {
671 tu_cs_emit_pkt4(cs, REG_A6XX_SP_DS_CONFIG, 2);
672 tu_cs_emit(cs, emit_xs_config(ds));
673 tu_cs_emit(cs, ds->instrlen);
674
675 tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_DS_CNTL, 1);
676 tu_cs_emit(cs, A6XX_HLSQ_DS_CNTL_CONSTLEN(align(ds->constlen, 4)));
677 }
678
679 static void
680 tu6_emit_gs_config(struct tu_cs *cs, struct tu_shader *shader,
681 const struct ir3_shader_variant *gs)
682 {
683 bool has_gs = gs->type != MESA_SHADER_NONE;
684 tu_cs_emit_pkt4(cs, REG_A6XX_SP_GS_PRIM_SIZE, 1);
685 tu_cs_emit(cs, 0);
686
687 tu_cs_emit_pkt4(cs, REG_A6XX_SP_GS_CONFIG, 2);
688 tu_cs_emit(cs, emit_xs_config(gs));
689 tu_cs_emit(cs, gs->instrlen);
690
691 tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_GS_CNTL, 1);
692 tu_cs_emit(cs, COND(has_gs, A6XX_HLSQ_GS_CNTL_ENABLED) |
693 A6XX_HLSQ_GS_CNTL_CONSTLEN(align(gs->constlen, 4)));
694 }
695
696 static void
697 tu6_emit_fs_config(struct tu_cs *cs, struct tu_shader *shader,
698 const struct ir3_shader_variant *fs)
699 {
700 uint32_t sp_fs_ctrl =
701 A6XX_SP_FS_CTRL_REG0_THREADSIZE(FOUR_QUADS) | 0x1000000 |
702 A6XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(fs->info.max_reg + 1) |
703 A6XX_SP_FS_CTRL_REG0_MERGEDREGS |
704 A6XX_SP_FS_CTRL_REG0_BRANCHSTACK(fs->branchstack);
705 if (fs->total_in > 0)
706 sp_fs_ctrl |= A6XX_SP_FS_CTRL_REG0_VARYING;
707 if (fs->need_pixlod)
708 sp_fs_ctrl |= A6XX_SP_FS_CTRL_REG0_PIXLODENABLE;
709 if (fs->need_fine_derivatives)
710 sp_fs_ctrl |= A6XX_SP_FS_CTRL_REG0_DIFF_FINE;
711
712 tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_CTRL_REG0, 1);
713 tu_cs_emit(cs, sp_fs_ctrl);
714
715 tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_CONFIG, 2);
716 tu_cs_emit(cs, emit_xs_config(fs));
717 tu_cs_emit(cs, fs->instrlen);
718
719 tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_FS_CNTL, 1);
720 tu_cs_emit(cs, A6XX_HLSQ_FS_CNTL_CONSTLEN(align(fs->constlen, 4)) |
721 A6XX_HLSQ_FS_CNTL_ENABLED);
722 }
723
724 static void
725 tu6_emit_cs_config(struct tu_cs *cs, const struct tu_shader *shader,
726 const struct ir3_shader_variant *v)
727 {
728 tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_UPDATE_CNTL, 1);
729 tu_cs_emit(cs, 0xff);
730
731 unsigned constlen = align(v->constlen, 4);
732 tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_CS_CNTL, 1);
733 tu_cs_emit(cs, A6XX_HLSQ_CS_CNTL_CONSTLEN(constlen) |
734 A6XX_HLSQ_CS_CNTL_ENABLED);
735
736 tu_cs_emit_pkt4(cs, REG_A6XX_SP_CS_CONFIG, 2);
737 tu_cs_emit(cs, emit_xs_config(v));
738 tu_cs_emit(cs, v->instrlen);
739
740 tu_cs_emit_pkt4(cs, REG_A6XX_SP_CS_CTRL_REG0, 1);
741 tu_cs_emit(cs, A6XX_SP_CS_CTRL_REG0_THREADSIZE(FOUR_QUADS) |
742 A6XX_SP_CS_CTRL_REG0_FULLREGFOOTPRINT(v->info.max_reg + 1) |
743 A6XX_SP_CS_CTRL_REG0_MERGEDREGS |
744 A6XX_SP_CS_CTRL_REG0_BRANCHSTACK(v->branchstack) |
745 COND(v->need_pixlod, A6XX_SP_CS_CTRL_REG0_PIXLODENABLE) |
746 COND(v->need_fine_derivatives, A6XX_SP_CS_CTRL_REG0_DIFF_FINE));
747
748 tu_cs_emit_pkt4(cs, REG_A6XX_SP_CS_UNKNOWN_A9B1, 1);
749 tu_cs_emit(cs, 0x41);
750
751 uint32_t local_invocation_id =
752 ir3_find_sysval_regid(v, SYSTEM_VALUE_LOCAL_INVOCATION_ID);
753 uint32_t work_group_id =
754 ir3_find_sysval_regid(v, SYSTEM_VALUE_WORK_GROUP_ID);
755
756 tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_CS_CNTL_0, 2);
757 tu_cs_emit(cs,
758 A6XX_HLSQ_CS_CNTL_0_WGIDCONSTID(work_group_id) |
759 A6XX_HLSQ_CS_CNTL_0_UNK0(regid(63, 0)) |
760 A6XX_HLSQ_CS_CNTL_0_UNK1(regid(63, 0)) |
761 A6XX_HLSQ_CS_CNTL_0_LOCALIDREGID(local_invocation_id));
762 tu_cs_emit(cs, 0x2fc); /* HLSQ_CS_UNKNOWN_B998 */
763 }
764
765 static void
766 tu6_emit_vs_system_values(struct tu_cs *cs,
767 const struct ir3_shader_variant *vs,
768 const struct ir3_shader_variant *gs,
769 bool primid_passthru)
770 {
771 const uint32_t vertexid_regid =
772 ir3_find_sysval_regid(vs, SYSTEM_VALUE_VERTEX_ID);
773 const uint32_t instanceid_regid =
774 ir3_find_sysval_regid(vs, SYSTEM_VALUE_INSTANCE_ID);
775 const uint32_t primitiveid_regid = gs->type != MESA_SHADER_NONE ?
776 ir3_find_sysval_regid(gs, SYSTEM_VALUE_PRIMITIVE_ID) :
777 regid(63, 0);
778 const uint32_t gsheader_regid = gs->type != MESA_SHADER_NONE ?
779 ir3_find_sysval_regid(gs, SYSTEM_VALUE_GS_HEADER_IR3) :
780 regid(63, 0);
781
782 tu_cs_emit_pkt4(cs, REG_A6XX_VFD_CONTROL_1, 6);
783 tu_cs_emit(cs, A6XX_VFD_CONTROL_1_REGID4VTX(vertexid_regid) |
784 A6XX_VFD_CONTROL_1_REGID4INST(instanceid_regid) |
785 A6XX_VFD_CONTROL_1_REGID4PRIMID(primitiveid_regid) |
786 0xfc000000);
787 tu_cs_emit(cs, 0x0000fcfc); /* VFD_CONTROL_2 */
788 tu_cs_emit(cs, 0xfcfcfcfc); /* VFD_CONTROL_3 */
789 tu_cs_emit(cs, 0x000000fc); /* VFD_CONTROL_4 */
790 tu_cs_emit(cs, A6XX_VFD_CONTROL_5_REGID_GSHEADER(gsheader_regid) |
791 0xfc00); /* VFD_CONTROL_5 */
792 tu_cs_emit(cs, COND(primid_passthru, A6XX_VFD_CONTROL_6_PRIMID_PASSTHRU)); /* VFD_CONTROL_6 */
793 }
794
795 /* Add any missing varyings needed for stream-out. Otherwise varyings not
796 * used by fragment shader will be stripped out.
797 */
798 static void
799 tu6_link_streamout(struct ir3_shader_linkage *l,
800 const struct ir3_shader_variant *v)
801 {
802 const struct ir3_stream_output_info *info = &v->shader->stream_output;
803
804 /*
805 * First, any stream-out varyings not already in linkage map (ie. also
806 * consumed by frag shader) need to be added:
807 */
808 for (unsigned i = 0; i < info->num_outputs; i++) {
809 const struct ir3_stream_output *out = &info->output[i];
810 unsigned compmask =
811 (1 << (out->num_components + out->start_component)) - 1;
812 unsigned k = out->register_index;
813 unsigned idx, nextloc = 0;
814
815 /* psize/pos need to be the last entries in linkage map, and will
816 * get added link_stream_out, so skip over them:
817 */
818 if (v->outputs[k].slot == VARYING_SLOT_PSIZ ||
819 v->outputs[k].slot == VARYING_SLOT_POS)
820 continue;
821
822 for (idx = 0; idx < l->cnt; idx++) {
823 if (l->var[idx].regid == v->outputs[k].regid)
824 break;
825 nextloc = MAX2(nextloc, l->var[idx].loc + 4);
826 }
827
828 /* add if not already in linkage map: */
829 if (idx == l->cnt)
830 ir3_link_add(l, v->outputs[k].regid, compmask, nextloc);
831
832 /* expand component-mask if needed, ie streaming out all components
833 * but frag shader doesn't consume all components:
834 */
835 if (compmask & ~l->var[idx].compmask) {
836 l->var[idx].compmask |= compmask;
837 l->max_loc = MAX2(l->max_loc, l->var[idx].loc +
838 util_last_bit(l->var[idx].compmask));
839 }
840 }
841 }
842
843 static void
844 tu6_setup_streamout(const struct ir3_shader_variant *v,
845 struct ir3_shader_linkage *l, struct tu_streamout_state *tf)
846 {
847 const struct ir3_stream_output_info *info = &v->shader->stream_output;
848
849 memset(tf, 0, sizeof(*tf));
850
851 tf->prog_count = align(l->max_loc, 2) / 2;
852
853 debug_assert(tf->prog_count < ARRAY_SIZE(tf->prog));
854
855 /* set stride info to the streamout state */
856 for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++)
857 tf->stride[i] = info->stride[i];
858
859 for (unsigned i = 0; i < info->num_outputs; i++) {
860 const struct ir3_stream_output *out = &info->output[i];
861 unsigned k = out->register_index;
862 unsigned idx;
863
864 /* Skip it, if there's an unused reg in the middle of outputs. */
865 if (v->outputs[k].regid == INVALID_REG)
866 continue;
867
868 tf->ncomp[out->output_buffer] += out->num_components;
869
870 /* linkage map sorted by order frag shader wants things, so
871 * a bit less ideal here..
872 */
873 for (idx = 0; idx < l->cnt; idx++)
874 if (l->var[idx].regid == v->outputs[k].regid)
875 break;
876
877 debug_assert(idx < l->cnt);
878
879 for (unsigned j = 0; j < out->num_components; j++) {
880 unsigned c = j + out->start_component;
881 unsigned loc = l->var[idx].loc + c;
882 unsigned off = j + out->dst_offset; /* in dwords */
883
884 if (loc & 1) {
885 tf->prog[loc/2] |= A6XX_VPC_SO_PROG_B_EN |
886 A6XX_VPC_SO_PROG_B_BUF(out->output_buffer) |
887 A6XX_VPC_SO_PROG_B_OFF(off * 4);
888 } else {
889 tf->prog[loc/2] |= A6XX_VPC_SO_PROG_A_EN |
890 A6XX_VPC_SO_PROG_A_BUF(out->output_buffer) |
891 A6XX_VPC_SO_PROG_A_OFF(off * 4);
892 }
893 }
894 }
895
896 tf->vpc_so_buf_cntl = A6XX_VPC_SO_BUF_CNTL_ENABLE |
897 COND(tf->ncomp[0] > 0, A6XX_VPC_SO_BUF_CNTL_BUF0) |
898 COND(tf->ncomp[1] > 0, A6XX_VPC_SO_BUF_CNTL_BUF1) |
899 COND(tf->ncomp[2] > 0, A6XX_VPC_SO_BUF_CNTL_BUF2) |
900 COND(tf->ncomp[3] > 0, A6XX_VPC_SO_BUF_CNTL_BUF3);
901 }
902
903 static void
904 tu6_emit_const(struct tu_cs *cs, uint32_t opcode, uint32_t base,
905 enum a6xx_state_block block, uint32_t offset,
906 uint32_t size, uint32_t *dwords) {
907 assert(size % 4 == 0);
908
909 tu_cs_emit_pkt7(cs, opcode, 3 + size);
910 tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(base) |
911 CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
912 CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
913 CP_LOAD_STATE6_0_STATE_BLOCK(block) |
914 CP_LOAD_STATE6_0_NUM_UNIT(size / 4));
915
916 tu_cs_emit(cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
917 tu_cs_emit(cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
918 dwords = (uint32_t *)&((uint8_t *)dwords)[offset];
919
920 tu_cs_emit_array(cs, dwords, size);
921 }
922
923 static void
924 tu6_emit_link_map(struct tu_cs *cs,
925 const struct ir3_shader_variant *producer,
926 const struct ir3_shader_variant *consumer) {
927 const struct ir3_const_state *const_state = &consumer->shader->const_state;
928 uint32_t base = const_state->offsets.primitive_map;
929 uint32_t patch_locs[MAX_VARYING] = { }, num_loc;
930 num_loc = ir3_link_geometry_stages(producer, consumer, patch_locs);
931 int size = DIV_ROUND_UP(num_loc, 4);
932
933 size = (MIN2(size + base, consumer->constlen) - base) * 4;
934 if (size <= 0)
935 return;
936
937 tu6_emit_const(cs, CP_LOAD_STATE6_GEOM, base, SB6_GS_SHADER, 0, size,
938 patch_locs);
939 }
940
941 static uint16_t
942 gl_primitive_to_tess(uint16_t primitive) {
943 switch (primitive) {
944 case GL_POINTS:
945 return TESS_POINTS;
946 case GL_LINE_STRIP:
947 return TESS_LINES;
948 case GL_TRIANGLE_STRIP:
949 return TESS_CW_TRIS;
950 default:
951 unreachable("");
952 }
953 }
954
955 static void
956 tu6_emit_vpc(struct tu_cs *cs,
957 const struct ir3_shader_variant *vs,
958 const struct ir3_shader_variant *gs,
959 const struct ir3_shader_variant *fs,
960 bool binning_pass,
961 struct tu_streamout_state *tf)
962 {
963 bool has_gs = gs->type != MESA_SHADER_NONE;
964 const struct ir3_shader_variant *last_shader = has_gs ? gs : vs;
965 struct ir3_shader_linkage linkage = { 0 };
966 ir3_link_shaders(&linkage, last_shader, fs, true);
967
968 if (last_shader->shader->stream_output.num_outputs)
969 tu6_link_streamout(&linkage, last_shader);
970
971 /* We do this after linking shaders in order to know whether PrimID
972 * passthrough needs to be enabled.
973 */
974 bool primid_passthru = linkage.primid_loc != 0xff;
975 tu6_emit_vs_system_values(cs, vs, gs, primid_passthru);
976
977 tu_cs_emit_pkt4(cs, REG_A6XX_VPC_VAR_DISABLE(0), 4);
978 tu_cs_emit(cs, ~linkage.varmask[0]);
979 tu_cs_emit(cs, ~linkage.varmask[1]);
980 tu_cs_emit(cs, ~linkage.varmask[2]);
981 tu_cs_emit(cs, ~linkage.varmask[3]);
982
983 /* a6xx finds position/pointsize at the end */
984 const uint32_t position_regid =
985 ir3_find_output_regid(last_shader, VARYING_SLOT_POS);
986 const uint32_t pointsize_regid =
987 ir3_find_output_regid(last_shader, VARYING_SLOT_PSIZ);
988 const uint32_t layer_regid = has_gs ?
989 ir3_find_output_regid(gs, VARYING_SLOT_LAYER) : regid(63, 0);
990
991 uint32_t pointsize_loc = 0xff, position_loc = 0xff, layer_loc = 0xff;
992 if (layer_regid != regid(63, 0)) {
993 layer_loc = linkage.max_loc;
994 ir3_link_add(&linkage, layer_regid, 0x1, linkage.max_loc);
995 }
996 if (position_regid != regid(63, 0)) {
997 position_loc = linkage.max_loc;
998 ir3_link_add(&linkage, position_regid, 0xf, linkage.max_loc);
999 }
1000 if (pointsize_regid != regid(63, 0)) {
1001 pointsize_loc = linkage.max_loc;
1002 ir3_link_add(&linkage, pointsize_regid, 0x1, linkage.max_loc);
1003 }
1004
1005 if (last_shader->shader->stream_output.num_outputs)
1006 tu6_setup_streamout(last_shader, &linkage, tf);
1007
1008 /* map outputs of the last shader to VPC */
1009 assert(linkage.cnt <= 32);
1010 const uint32_t sp_out_count = DIV_ROUND_UP(linkage.cnt, 2);
1011 const uint32_t sp_vpc_dst_count = DIV_ROUND_UP(linkage.cnt, 4);
1012 uint32_t sp_out[16];
1013 uint32_t sp_vpc_dst[8];
1014 for (uint32_t i = 0; i < linkage.cnt; i++) {
1015 ((uint16_t *) sp_out)[i] =
1016 A6XX_SP_VS_OUT_REG_A_REGID(linkage.var[i].regid) |
1017 A6XX_SP_VS_OUT_REG_A_COMPMASK(linkage.var[i].compmask);
1018 ((uint8_t *) sp_vpc_dst)[i] =
1019 A6XX_SP_VS_VPC_DST_REG_OUTLOC0(linkage.var[i].loc);
1020 }
1021
1022 if (has_gs)
1023 tu_cs_emit_pkt4(cs, REG_A6XX_SP_GS_OUT_REG(0), sp_out_count);
1024 else
1025 tu_cs_emit_pkt4(cs, REG_A6XX_SP_VS_OUT_REG(0), sp_out_count);
1026 tu_cs_emit_array(cs, sp_out, sp_out_count);
1027
1028 if (has_gs)
1029 tu_cs_emit_pkt4(cs, REG_A6XX_SP_GS_VPC_DST_REG(0), sp_vpc_dst_count);
1030 else
1031 tu_cs_emit_pkt4(cs, REG_A6XX_SP_VS_VPC_DST_REG(0), sp_vpc_dst_count);
1032 tu_cs_emit_array(cs, sp_vpc_dst, sp_vpc_dst_count);
1033
1034 tu_cs_emit_pkt4(cs, REG_A6XX_PC_PRIMID_CNTL, 1);
1035 tu_cs_emit(cs, COND(primid_passthru, A6XX_PC_PRIMID_CNTL_PRIMID_PASSTHRU));
1036
1037 tu_cs_emit_pkt4(cs, REG_A6XX_VPC_CNTL_0, 1);
1038 tu_cs_emit(cs, A6XX_VPC_CNTL_0_NUMNONPOSVAR(fs->total_in) |
1039 (fs->total_in > 0 ? A6XX_VPC_CNTL_0_VARYING : 0) |
1040 A6XX_VPC_CNTL_0_PRIMIDLOC(linkage.primid_loc) |
1041 A6XX_VPC_CNTL_0_UNKLOC(0xff));
1042
1043 tu_cs_emit_pkt4(cs, REG_A6XX_VPC_PACK, 1);
1044 tu_cs_emit(cs, A6XX_VPC_PACK_POSITIONLOC(position_loc) |
1045 A6XX_VPC_PACK_PSIZELOC(pointsize_loc) |
1046 A6XX_VPC_PACK_STRIDE_IN_VPC(linkage.max_loc));
1047
1048 if (has_gs) {
1049 tu_cs_emit_pkt4(cs, REG_A6XX_SP_GS_CTRL_REG0, 1);
1050 tu_cs_emit(cs, A6XX_SP_GS_CTRL_REG0_THREADSIZE(TWO_QUADS) |
1051 A6XX_SP_GS_CTRL_REG0_FULLREGFOOTPRINT(gs->info.max_reg + 1) |
1052 A6XX_SP_GS_CTRL_REG0_BRANCHSTACK(gs->branchstack) |
1053 COND(gs->need_pixlod, A6XX_SP_GS_CTRL_REG0_PIXLODENABLE));
1054
1055 tu6_emit_link_map(cs, vs, gs);
1056
1057 uint32_t primitive_regid =
1058 ir3_find_sysval_regid(gs, SYSTEM_VALUE_PRIMITIVE_ID);
1059 tu_cs_emit_pkt4(cs, REG_A6XX_VPC_PACK_GS, 1);
1060 tu_cs_emit(cs, A6XX_VPC_PACK_GS_POSITIONLOC(position_loc) |
1061 A6XX_VPC_PACK_GS_PSIZELOC(pointsize_loc) |
1062 A6XX_VPC_PACK_GS_STRIDE_IN_VPC(linkage.max_loc));
1063
1064 tu_cs_emit_pkt4(cs, REG_A6XX_VPC_UNKNOWN_9105, 1);
1065 tu_cs_emit(cs, A6XX_VPC_UNKNOWN_9105_LAYERLOC(layer_loc) | 0xff00);
1066
1067 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_UNKNOWN_809C, 1);
1068 tu_cs_emit(cs, CONDREG(layer_regid,
1069 A6XX_GRAS_UNKNOWN_809C_GS_WRITES_LAYER));
1070
1071 uint32_t flags_regid = ir3_find_output_regid(gs,
1072 VARYING_SLOT_GS_VERTEX_FLAGS_IR3);
1073
1074 tu_cs_emit_pkt4(cs, REG_A6XX_SP_PRIMITIVE_CNTL_GS, 1);
1075 tu_cs_emit(cs, A6XX_SP_PRIMITIVE_CNTL_GS_GSOUT(linkage.cnt) |
1076 A6XX_SP_PRIMITIVE_CNTL_GS_FLAGS_REGID(flags_regid));
1077
1078 tu_cs_emit_pkt4(cs, REG_A6XX_PC_PRIMITIVE_CNTL_2, 1);
1079 tu_cs_emit(cs, A6XX_PC_PRIMITIVE_CNTL_2_STRIDE_IN_VPC(linkage.max_loc) |
1080 CONDREG(pointsize_regid, A6XX_PC_PRIMITIVE_CNTL_2_PSIZE) |
1081 CONDREG(layer_regid, A6XX_PC_PRIMITIVE_CNTL_2_LAYER) |
1082 CONDREG(primitive_regid, A6XX_PC_PRIMITIVE_CNTL_2_PRIMITIVE_ID));
1083
1084 uint32_t vertices_out = gs->shader->nir->info.gs.vertices_out - 1;
1085 uint16_t output =
1086 gl_primitive_to_tess(gs->shader->nir->info.gs.output_primitive);
1087 uint32_t invocations = gs->shader->nir->info.gs.invocations - 1;
1088 tu_cs_emit_pkt4(cs, REG_A6XX_PC_PRIMITIVE_CNTL_5, 1);
1089 tu_cs_emit(cs,
1090 A6XX_PC_PRIMITIVE_CNTL_5_GS_VERTICES_OUT(vertices_out) |
1091 A6XX_PC_PRIMITIVE_CNTL_5_GS_OUTPUT(output) |
1092 A6XX_PC_PRIMITIVE_CNTL_5_GS_INVOCATIONS(invocations));
1093
1094 tu_cs_emit_pkt4(cs, REG_A6XX_PC_PRIMITIVE_CNTL_3, 1);
1095 tu_cs_emit(cs, 0);
1096
1097 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_UNKNOWN_8003, 1);
1098 tu_cs_emit(cs, 0);
1099
1100 tu_cs_emit_pkt4(cs, REG_A6XX_VPC_UNKNOWN_9100, 1);
1101 tu_cs_emit(cs, 0xff);
1102
1103 tu_cs_emit_pkt4(cs, REG_A6XX_VPC_UNKNOWN_9102, 1);
1104 tu_cs_emit(cs, 0xffff00);
1105
1106 /* Size of per-primitive alloction in ldlw memory in vec4s. */
1107 uint32_t vec4_size =
1108 gs->shader->nir->info.gs.vertices_in *
1109 DIV_ROUND_UP(vs->shader->output_size, 4);
1110 tu_cs_emit_pkt4(cs, REG_A6XX_PC_PRIMITIVE_CNTL_6, 1);
1111 tu_cs_emit(cs, A6XX_PC_PRIMITIVE_CNTL_6_STRIDE_IN_VPC(vec4_size));
1112
1113 tu_cs_emit_pkt4(cs, REG_A6XX_PC_UNKNOWN_9B07, 1);
1114 tu_cs_emit(cs, 0);
1115
1116 tu_cs_emit_pkt4(cs, REG_A6XX_SP_GS_PRIM_SIZE, 1);
1117 tu_cs_emit(cs, vs->shader->output_size);
1118 }
1119
1120 tu_cs_emit_pkt4(cs, REG_A6XX_SP_PRIMITIVE_CNTL, 1);
1121 tu_cs_emit(cs, A6XX_SP_PRIMITIVE_CNTL_VSOUT(linkage.cnt));
1122
1123 tu_cs_emit_pkt4(cs, REG_A6XX_PC_PRIMITIVE_CNTL_1, 1);
1124 tu_cs_emit(cs, A6XX_PC_PRIMITIVE_CNTL_1_STRIDE_IN_VPC(linkage.max_loc) |
1125 (last_shader->writes_psize ? A6XX_PC_PRIMITIVE_CNTL_1_PSIZE : 0));
1126 }
1127
1128 static int
1129 tu6_vpc_varying_mode(const struct ir3_shader_variant *fs,
1130 uint32_t index,
1131 uint8_t *interp_mode,
1132 uint8_t *ps_repl_mode)
1133 {
1134 enum
1135 {
1136 INTERP_SMOOTH = 0,
1137 INTERP_FLAT = 1,
1138 INTERP_ZERO = 2,
1139 INTERP_ONE = 3,
1140 };
1141 enum
1142 {
1143 PS_REPL_NONE = 0,
1144 PS_REPL_S = 1,
1145 PS_REPL_T = 2,
1146 PS_REPL_ONE_MINUS_T = 3,
1147 };
1148
1149 const uint32_t compmask = fs->inputs[index].compmask;
1150
1151 /* NOTE: varyings are packed, so if compmask is 0xb then first, second, and
1152 * fourth component occupy three consecutive varying slots
1153 */
1154 int shift = 0;
1155 *interp_mode = 0;
1156 *ps_repl_mode = 0;
1157 if (fs->inputs[index].slot == VARYING_SLOT_PNTC) {
1158 if (compmask & 0x1) {
1159 *ps_repl_mode |= PS_REPL_S << shift;
1160 shift += 2;
1161 }
1162 if (compmask & 0x2) {
1163 *ps_repl_mode |= PS_REPL_T << shift;
1164 shift += 2;
1165 }
1166 if (compmask & 0x4) {
1167 *interp_mode |= INTERP_ZERO << shift;
1168 shift += 2;
1169 }
1170 if (compmask & 0x8) {
1171 *interp_mode |= INTERP_ONE << 6;
1172 shift += 2;
1173 }
1174 } else if ((fs->inputs[index].interpolate == INTERP_MODE_FLAT) ||
1175 fs->inputs[index].rasterflat) {
1176 for (int i = 0; i < 4; i++) {
1177 if (compmask & (1 << i)) {
1178 *interp_mode |= INTERP_FLAT << shift;
1179 shift += 2;
1180 }
1181 }
1182 }
1183
1184 return shift;
1185 }
1186
1187 static void
1188 tu6_emit_vpc_varying_modes(struct tu_cs *cs,
1189 const struct ir3_shader_variant *fs,
1190 bool binning_pass)
1191 {
1192 uint32_t interp_modes[8] = { 0 };
1193 uint32_t ps_repl_modes[8] = { 0 };
1194
1195 if (!binning_pass) {
1196 for (int i = -1;
1197 (i = ir3_next_varying(fs, i)) < (int) fs->inputs_count;) {
1198
1199 /* get the mode for input i */
1200 uint8_t interp_mode;
1201 uint8_t ps_repl_mode;
1202 const int bits =
1203 tu6_vpc_varying_mode(fs, i, &interp_mode, &ps_repl_mode);
1204
1205 /* OR the mode into the array */
1206 const uint32_t inloc = fs->inputs[i].inloc * 2;
1207 uint32_t n = inloc / 32;
1208 uint32_t shift = inloc % 32;
1209 interp_modes[n] |= interp_mode << shift;
1210 ps_repl_modes[n] |= ps_repl_mode << shift;
1211 if (shift + bits > 32) {
1212 n++;
1213 shift = 32 - shift;
1214
1215 interp_modes[n] |= interp_mode >> shift;
1216 ps_repl_modes[n] |= ps_repl_mode >> shift;
1217 }
1218 }
1219 }
1220
1221 tu_cs_emit_pkt4(cs, REG_A6XX_VPC_VARYING_INTERP_MODE(0), 8);
1222 tu_cs_emit_array(cs, interp_modes, 8);
1223
1224 tu_cs_emit_pkt4(cs, REG_A6XX_VPC_VARYING_PS_REPL_MODE(0), 8);
1225 tu_cs_emit_array(cs, ps_repl_modes, 8);
1226 }
1227
1228 static void
1229 tu6_emit_fs_inputs(struct tu_cs *cs, const struct ir3_shader_variant *fs)
1230 {
1231 uint32_t face_regid, coord_regid, zwcoord_regid, samp_id_regid;
1232 uint32_t ij_pix_regid, ij_samp_regid, ij_cent_regid, ij_size_regid;
1233 uint32_t smask_in_regid;
1234
1235 bool sample_shading = fs->per_samp; /* TODO | key->sample_shading; */
1236 bool enable_varyings = fs->total_in > 0;
1237
1238 samp_id_regid = ir3_find_sysval_regid(fs, SYSTEM_VALUE_SAMPLE_ID);
1239 smask_in_regid = ir3_find_sysval_regid(fs, SYSTEM_VALUE_SAMPLE_MASK_IN);
1240 face_regid = ir3_find_sysval_regid(fs, SYSTEM_VALUE_FRONT_FACE);
1241 coord_regid = ir3_find_sysval_regid(fs, SYSTEM_VALUE_FRAG_COORD);
1242 zwcoord_regid = VALIDREG(coord_regid) ? coord_regid + 2 : regid(63, 0);
1243 ij_pix_regid = ir3_find_sysval_regid(fs, SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL);
1244 ij_samp_regid = ir3_find_sysval_regid(fs, SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE);
1245 ij_cent_regid = ir3_find_sysval_regid(fs, SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID);
1246 ij_size_regid = ir3_find_sysval_regid(fs, SYSTEM_VALUE_BARYCENTRIC_PERSP_SIZE);
1247
1248 if (fs->num_sampler_prefetch > 0) {
1249 assert(VALIDREG(ij_pix_regid));
1250 /* also, it seems like ij_pix is *required* to be r0.x */
1251 assert(ij_pix_regid == regid(0, 0));
1252 }
1253
1254 tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_PREFETCH_CNTL, 1 + fs->num_sampler_prefetch);
1255 tu_cs_emit(cs, A6XX_SP_FS_PREFETCH_CNTL_COUNT(fs->num_sampler_prefetch) |
1256 A6XX_SP_FS_PREFETCH_CNTL_UNK4(regid(63, 0)) |
1257 0x7000); // XXX);
1258 for (int i = 0; i < fs->num_sampler_prefetch; i++) {
1259 const struct ir3_sampler_prefetch *prefetch = &fs->sampler_prefetch[i];
1260 tu_cs_emit(cs, A6XX_SP_FS_PREFETCH_CMD_SRC(prefetch->src) |
1261 A6XX_SP_FS_PREFETCH_CMD_SAMP_ID(prefetch->samp_id) |
1262 A6XX_SP_FS_PREFETCH_CMD_TEX_ID(prefetch->tex_id) |
1263 A6XX_SP_FS_PREFETCH_CMD_DST(prefetch->dst) |
1264 A6XX_SP_FS_PREFETCH_CMD_WRMASK(prefetch->wrmask) |
1265 COND(prefetch->half_precision, A6XX_SP_FS_PREFETCH_CMD_HALF) |
1266 A6XX_SP_FS_PREFETCH_CMD_CMD(prefetch->cmd));
1267 }
1268
1269 if (fs->num_sampler_prefetch > 0) {
1270 tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_BINDLESS_PREFETCH_CMD(0), fs->num_sampler_prefetch);
1271 for (int i = 0; i < fs->num_sampler_prefetch; i++) {
1272 const struct ir3_sampler_prefetch *prefetch = &fs->sampler_prefetch[i];
1273 tu_cs_emit(cs,
1274 A6XX_SP_FS_BINDLESS_PREFETCH_CMD_SAMP_ID(prefetch->samp_bindless_id) |
1275 A6XX_SP_FS_BINDLESS_PREFETCH_CMD_TEX_ID(prefetch->tex_bindless_id));
1276 }
1277 }
1278
1279 tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_CONTROL_1_REG, 5);
1280 tu_cs_emit(cs, 0x7);
1281 tu_cs_emit(cs, A6XX_HLSQ_CONTROL_2_REG_FACEREGID(face_regid) |
1282 A6XX_HLSQ_CONTROL_2_REG_SAMPLEID(samp_id_regid) |
1283 A6XX_HLSQ_CONTROL_2_REG_SAMPLEMASK(smask_in_regid) |
1284 A6XX_HLSQ_CONTROL_2_REG_SIZE(ij_size_regid));
1285 tu_cs_emit(cs, A6XX_HLSQ_CONTROL_3_REG_BARY_IJ_PIXEL(ij_pix_regid) |
1286 A6XX_HLSQ_CONTROL_3_REG_BARY_IJ_CENTROID(ij_cent_regid) |
1287 0xfc00fc00);
1288 tu_cs_emit(cs, A6XX_HLSQ_CONTROL_4_REG_XYCOORDREGID(coord_regid) |
1289 A6XX_HLSQ_CONTROL_4_REG_ZWCOORDREGID(zwcoord_regid) |
1290 A6XX_HLSQ_CONTROL_4_REG_BARY_IJ_PIXEL_PERSAMP(ij_samp_regid) |
1291 0x0000fc00);
1292 tu_cs_emit(cs, 0xfc);
1293
1294 tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_UNKNOWN_B980, 1);
1295 tu_cs_emit(cs, enable_varyings ? 3 : 1);
1296
1297 tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_UPDATE_CNTL, 1);
1298 tu_cs_emit(cs, 0xff); /* XXX */
1299
1300 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_CNTL, 1);
1301 tu_cs_emit(cs,
1302 CONDREG(ij_pix_regid, A6XX_GRAS_CNTL_VARYING) |
1303 CONDREG(ij_cent_regid, A6XX_GRAS_CNTL_CENTROID) |
1304 CONDREG(ij_samp_regid, A6XX_GRAS_CNTL_PERSAMP_VARYING) |
1305 COND(VALIDREG(ij_size_regid) && !sample_shading, A6XX_GRAS_CNTL_SIZE) |
1306 COND(VALIDREG(ij_size_regid) && sample_shading, A6XX_GRAS_CNTL_SIZE_PERSAMP) |
1307 COND(fs->fragcoord_compmask != 0, A6XX_GRAS_CNTL_SIZE |
1308 A6XX_GRAS_CNTL_COORD_MASK(fs->fragcoord_compmask)) |
1309 COND(fs->frag_face, A6XX_GRAS_CNTL_SIZE));
1310
1311 tu_cs_emit_pkt4(cs, REG_A6XX_RB_RENDER_CONTROL0, 2);
1312 tu_cs_emit(cs,
1313 CONDREG(ij_pix_regid, A6XX_RB_RENDER_CONTROL0_VARYING) |
1314 CONDREG(ij_cent_regid, A6XX_RB_RENDER_CONTROL0_CENTROID) |
1315 CONDREG(ij_samp_regid, A6XX_RB_RENDER_CONTROL0_PERSAMP_VARYING) |
1316 COND(enable_varyings, A6XX_RB_RENDER_CONTROL0_UNK10) |
1317 COND(VALIDREG(ij_size_regid) && !sample_shading, A6XX_RB_RENDER_CONTROL0_SIZE) |
1318 COND(VALIDREG(ij_size_regid) && sample_shading, A6XX_RB_RENDER_CONTROL0_SIZE_PERSAMP) |
1319 COND(fs->fragcoord_compmask != 0, A6XX_RB_RENDER_CONTROL0_SIZE |
1320 A6XX_RB_RENDER_CONTROL0_COORD_MASK(fs->fragcoord_compmask)) |
1321 COND(fs->frag_face, A6XX_RB_RENDER_CONTROL0_SIZE));
1322 tu_cs_emit(cs,
1323 CONDREG(smask_in_regid, A6XX_RB_RENDER_CONTROL1_SAMPLEMASK) |
1324 CONDREG(samp_id_regid, A6XX_RB_RENDER_CONTROL1_SAMPLEID) |
1325 CONDREG(ij_size_regid, A6XX_RB_RENDER_CONTROL1_SIZE) |
1326 COND(fs->frag_face, A6XX_RB_RENDER_CONTROL1_FACENESS));
1327
1328 tu_cs_emit_pkt4(cs, REG_A6XX_RB_SAMPLE_CNTL, 1);
1329 tu_cs_emit(cs, COND(sample_shading, A6XX_RB_SAMPLE_CNTL_PER_SAMP_MODE));
1330
1331 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_UNKNOWN_8101, 1);
1332 tu_cs_emit(cs, COND(sample_shading, 0x6)); // XXX
1333
1334 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SAMPLE_CNTL, 1);
1335 tu_cs_emit(cs, COND(sample_shading, A6XX_GRAS_SAMPLE_CNTL_PER_SAMP_MODE));
1336 }
1337
1338 static void
1339 tu6_emit_fs_outputs(struct tu_cs *cs,
1340 const struct ir3_shader_variant *fs,
1341 uint32_t mrt_count, bool dual_src_blend,
1342 uint32_t render_components)
1343 {
1344 uint32_t smask_regid, posz_regid;
1345
1346 posz_regid = ir3_find_output_regid(fs, FRAG_RESULT_DEPTH);
1347 smask_regid = ir3_find_output_regid(fs, FRAG_RESULT_SAMPLE_MASK);
1348
1349 uint32_t fragdata_regid[8];
1350 if (fs->color0_mrt) {
1351 fragdata_regid[0] = ir3_find_output_regid(fs, FRAG_RESULT_COLOR);
1352 for (uint32_t i = 1; i < ARRAY_SIZE(fragdata_regid); i++)
1353 fragdata_regid[i] = fragdata_regid[0];
1354 } else {
1355 for (uint32_t i = 0; i < ARRAY_SIZE(fragdata_regid); i++)
1356 fragdata_regid[i] = ir3_find_output_regid(fs, FRAG_RESULT_DATA0 + i);
1357 }
1358
1359 tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_OUTPUT_CNTL0, 2);
1360 tu_cs_emit(cs, A6XX_SP_FS_OUTPUT_CNTL0_DEPTH_REGID(posz_regid) |
1361 A6XX_SP_FS_OUTPUT_CNTL0_SAMPMASK_REGID(smask_regid) |
1362 COND(dual_src_blend, A6XX_SP_FS_OUTPUT_CNTL0_DUAL_COLOR_IN_ENABLE) |
1363 0xfc000000);
1364 tu_cs_emit(cs, A6XX_SP_FS_OUTPUT_CNTL1_MRT(mrt_count));
1365
1366 tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_OUTPUT_REG(0), 8);
1367 for (uint32_t i = 0; i < ARRAY_SIZE(fragdata_regid); i++) {
1368 // TODO we could have a mix of half and full precision outputs,
1369 // we really need to figure out half-precision from IR3_REG_HALF
1370 tu_cs_emit(cs, A6XX_SP_FS_OUTPUT_REG_REGID(fragdata_regid[i]) |
1371 (false ? A6XX_SP_FS_OUTPUT_REG_HALF_PRECISION : 0));
1372 }
1373
1374 tu_cs_emit_regs(cs,
1375 A6XX_SP_FS_RENDER_COMPONENTS(.dword = render_components));
1376
1377 tu_cs_emit_pkt4(cs, REG_A6XX_RB_FS_OUTPUT_CNTL0, 2);
1378 tu_cs_emit(cs, COND(fs->writes_pos, A6XX_RB_FS_OUTPUT_CNTL0_FRAG_WRITES_Z) |
1379 COND(fs->writes_smask, A6XX_RB_FS_OUTPUT_CNTL0_FRAG_WRITES_SAMPMASK) |
1380 COND(dual_src_blend, A6XX_RB_FS_OUTPUT_CNTL0_DUAL_COLOR_IN_ENABLE));
1381 tu_cs_emit(cs, A6XX_RB_FS_OUTPUT_CNTL1_MRT(mrt_count));
1382
1383 tu_cs_emit_regs(cs,
1384 A6XX_RB_RENDER_COMPONENTS(.dword = render_components));
1385
1386 uint32_t gras_su_depth_plane_cntl = 0;
1387 uint32_t rb_depth_plane_cntl = 0;
1388 if (fs->no_earlyz || fs->has_kill || fs->writes_pos) {
1389 gras_su_depth_plane_cntl |= A6XX_GRAS_SU_DEPTH_PLANE_CNTL_FRAG_WRITES_Z;
1390 rb_depth_plane_cntl |= A6XX_RB_DEPTH_PLANE_CNTL_FRAG_WRITES_Z;
1391 }
1392
1393 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SU_DEPTH_PLANE_CNTL, 1);
1394 tu_cs_emit(cs, gras_su_depth_plane_cntl);
1395
1396 tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_PLANE_CNTL, 1);
1397 tu_cs_emit(cs, rb_depth_plane_cntl);
1398 }
1399
1400 static void
1401 tu6_emit_shader_object(struct tu_cs *cs,
1402 gl_shader_stage stage,
1403 const struct ir3_shader_variant *variant,
1404 const struct tu_bo *binary_bo,
1405 uint32_t binary_offset)
1406 {
1407 uint16_t reg;
1408 uint8_t opcode;
1409 enum a6xx_state_block sb;
1410 switch (stage) {
1411 case MESA_SHADER_VERTEX:
1412 reg = REG_A6XX_SP_VS_OBJ_START_LO;
1413 opcode = CP_LOAD_STATE6_GEOM;
1414 sb = SB6_VS_SHADER;
1415 break;
1416 case MESA_SHADER_TESS_CTRL:
1417 reg = REG_A6XX_SP_HS_OBJ_START_LO;
1418 opcode = CP_LOAD_STATE6_GEOM;
1419 sb = SB6_HS_SHADER;
1420 break;
1421 case MESA_SHADER_TESS_EVAL:
1422 reg = REG_A6XX_SP_DS_OBJ_START_LO;
1423 opcode = CP_LOAD_STATE6_GEOM;
1424 sb = SB6_DS_SHADER;
1425 break;
1426 case MESA_SHADER_GEOMETRY:
1427 reg = REG_A6XX_SP_GS_OBJ_START_LO;
1428 opcode = CP_LOAD_STATE6_GEOM;
1429 sb = SB6_GS_SHADER;
1430 break;
1431 case MESA_SHADER_FRAGMENT:
1432 reg = REG_A6XX_SP_FS_OBJ_START_LO;
1433 opcode = CP_LOAD_STATE6_FRAG;
1434 sb = SB6_FS_SHADER;
1435 break;
1436 case MESA_SHADER_COMPUTE:
1437 reg = REG_A6XX_SP_CS_OBJ_START_LO;
1438 opcode = CP_LOAD_STATE6_FRAG;
1439 sb = SB6_CS_SHADER;
1440 break;
1441 default:
1442 unreachable("invalid gl_shader_stage");
1443 opcode = CP_LOAD_STATE6_GEOM;
1444 sb = SB6_VS_SHADER;
1445 break;
1446 }
1447
1448 if (!variant->instrlen) {
1449 tu_cs_emit_pkt4(cs, reg, 2);
1450 tu_cs_emit_qw(cs, 0);
1451 return;
1452 }
1453
1454 assert(variant->type == stage);
1455
1456 const uint64_t binary_iova = binary_bo->iova + binary_offset;
1457 assert((binary_iova & 0xf) == 0);
1458 /* note: it looks like HW might try to read a few instructions beyond the instrlen size
1459 * of the shader. this could be a potential source of problems at some point
1460 * possibly this doesn't happen if shader iova is aligned enough (to 4k for example)
1461 */
1462
1463 tu_cs_emit_pkt4(cs, reg, 2);
1464 tu_cs_emit_qw(cs, binary_iova);
1465
1466 /* always indirect */
1467 const bool indirect = true;
1468 if (indirect) {
1469 tu_cs_emit_pkt7(cs, opcode, 3);
1470 tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(0) |
1471 CP_LOAD_STATE6_0_STATE_TYPE(ST6_SHADER) |
1472 CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
1473 CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
1474 CP_LOAD_STATE6_0_NUM_UNIT(variant->instrlen));
1475 tu_cs_emit_qw(cs, binary_iova);
1476 } else {
1477 const void *binary = binary_bo->map + binary_offset;
1478
1479 tu_cs_emit_pkt7(cs, opcode, 3 + variant->info.sizedwords);
1480 tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(0) |
1481 CP_LOAD_STATE6_0_STATE_TYPE(ST6_SHADER) |
1482 CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
1483 CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
1484 CP_LOAD_STATE6_0_NUM_UNIT(variant->instrlen));
1485 tu_cs_emit_qw(cs, 0);
1486 tu_cs_emit_array(cs, binary, variant->info.sizedwords);
1487 }
1488 }
1489
1490 static void
1491 tu6_emit_immediates(struct tu_cs *cs, const struct ir3_shader_variant *v,
1492 uint32_t opcode, enum a6xx_state_block block)
1493 {
1494 /* dummy variant */
1495 if (!v->shader)
1496 return;
1497
1498 const struct ir3_const_state *const_state = &v->shader->const_state;
1499 uint32_t base = const_state->offsets.immediate;
1500 int size = const_state->immediates_count;
1501
1502 /* truncate size to avoid writing constants that shader
1503 * does not use:
1504 */
1505 size = MIN2(size + base, v->constlen) - base;
1506
1507 if (size <= 0)
1508 return;
1509
1510 tu_cs_emit_pkt7(cs, opcode, 3 + size * 4);
1511 tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(base) |
1512 CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
1513 CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
1514 CP_LOAD_STATE6_0_STATE_BLOCK(block) |
1515 CP_LOAD_STATE6_0_NUM_UNIT(size));
1516 tu_cs_emit(cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
1517 tu_cs_emit(cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
1518
1519 for (unsigned i = 0; i < size; i++) {
1520 tu_cs_emit(cs, const_state->immediates[i].val[0]);
1521 tu_cs_emit(cs, const_state->immediates[i].val[1]);
1522 tu_cs_emit(cs, const_state->immediates[i].val[2]);
1523 tu_cs_emit(cs, const_state->immediates[i].val[3]);
1524 }
1525 }
1526
1527 static void
1528 tu6_emit_geometry_consts(struct tu_cs *cs,
1529 const struct ir3_shader_variant *vs,
1530 const struct ir3_shader_variant *gs) {
1531 unsigned num_vertices = gs->shader->nir->info.gs.vertices_in;
1532
1533 uint32_t params[4] = {
1534 vs->shader->output_size * num_vertices * 4, /* primitive stride */
1535 vs->shader->output_size * 4, /* vertex stride */
1536 0,
1537 0,
1538 };
1539 uint32_t vs_base = vs->shader->const_state.offsets.primitive_param;
1540 tu6_emit_const(cs, CP_LOAD_STATE6_GEOM, vs_base, SB6_VS_SHADER, 0,
1541 ARRAY_SIZE(params), params);
1542
1543 uint32_t gs_base = gs->shader->const_state.offsets.primitive_param;
1544 tu6_emit_const(cs, CP_LOAD_STATE6_GEOM, gs_base, SB6_GS_SHADER, 0,
1545 ARRAY_SIZE(params), params);
1546 }
1547
1548 static void
1549 tu6_emit_program(struct tu_cs *cs,
1550 const struct tu_pipeline_builder *builder,
1551 const struct tu_bo *binary_bo,
1552 bool binning_pass,
1553 struct tu_streamout_state *tf)
1554 {
1555 static const struct ir3_shader_variant dummy_variant = {
1556 .type = MESA_SHADER_NONE
1557 };
1558 assert(builder->shaders[MESA_SHADER_VERTEX]);
1559 const struct ir3_shader_variant *vs =
1560 &builder->shaders[MESA_SHADER_VERTEX]->variants[0];
1561 const struct ir3_shader_variant *hs =
1562 builder->shaders[MESA_SHADER_TESS_CTRL]
1563 ? &builder->shaders[MESA_SHADER_TESS_CTRL]->variants[0]
1564 : &dummy_variant;
1565 const struct ir3_shader_variant *ds =
1566 builder->shaders[MESA_SHADER_TESS_EVAL]
1567 ? &builder->shaders[MESA_SHADER_TESS_EVAL]->variants[0]
1568 : &dummy_variant;
1569 const struct ir3_shader_variant *gs =
1570 builder->shaders[MESA_SHADER_GEOMETRY]
1571 ? &builder->shaders[MESA_SHADER_GEOMETRY]->variants[0]
1572 : &dummy_variant;
1573 const struct ir3_shader_variant *fs =
1574 builder->shaders[MESA_SHADER_FRAGMENT]
1575 ? &builder->shaders[MESA_SHADER_FRAGMENT]->variants[0]
1576 : &dummy_variant;
1577 bool has_gs = gs->type != MESA_SHADER_NONE;
1578
1579 if (binning_pass) {
1580 /* if we have streamout, use full VS in binning pass, as the
1581 * binning pass VS will have outputs on other than position/psize
1582 * stripped out:
1583 */
1584 if (vs->shader->stream_output.num_outputs == 0)
1585 vs = &builder->shaders[MESA_SHADER_VERTEX]->variants[1];
1586 fs = &dummy_variant;
1587 }
1588
1589 tu6_emit_vs_config(cs, builder->shaders[MESA_SHADER_VERTEX], vs);
1590 tu6_emit_hs_config(cs, builder->shaders[MESA_SHADER_TESS_CTRL], hs);
1591 tu6_emit_ds_config(cs, builder->shaders[MESA_SHADER_TESS_EVAL], ds);
1592 tu6_emit_gs_config(cs, builder->shaders[MESA_SHADER_GEOMETRY], gs);
1593 tu6_emit_fs_config(cs, builder->shaders[MESA_SHADER_FRAGMENT], fs);
1594
1595 tu6_emit_vpc(cs, vs, gs, fs, binning_pass, tf);
1596 tu6_emit_vpc_varying_modes(cs, fs, binning_pass);
1597 tu6_emit_fs_inputs(cs, fs);
1598 tu6_emit_fs_outputs(cs, fs, builder->color_attachment_count,
1599 builder->use_dual_src_blend,
1600 builder->render_components);
1601
1602 tu6_emit_shader_object(cs, MESA_SHADER_VERTEX, vs, binary_bo,
1603 binning_pass ? builder->binning_vs_offset : builder->shader_offsets[MESA_SHADER_VERTEX]);
1604 if (has_gs)
1605 tu6_emit_shader_object(cs, MESA_SHADER_GEOMETRY, gs, binary_bo,
1606 builder->shader_offsets[MESA_SHADER_GEOMETRY]);
1607 tu6_emit_shader_object(cs, MESA_SHADER_FRAGMENT, fs, binary_bo,
1608 builder->shader_offsets[MESA_SHADER_FRAGMENT]);
1609
1610 tu6_emit_immediates(cs, vs, CP_LOAD_STATE6_GEOM, SB6_VS_SHADER);
1611 if (has_gs) {
1612 tu6_emit_immediates(cs, gs, CP_LOAD_STATE6_GEOM, SB6_GS_SHADER);
1613 tu6_emit_geometry_consts(cs, vs, gs);
1614 }
1615 if (!binning_pass)
1616 tu6_emit_immediates(cs, fs, CP_LOAD_STATE6_FRAG, SB6_FS_SHADER);
1617 }
1618
1619 static void
1620 tu6_emit_vertex_input(struct tu_cs *cs,
1621 const struct ir3_shader_variant *vs,
1622 const VkPipelineVertexInputStateCreateInfo *info,
1623 uint8_t bindings[MAX_VERTEX_ATTRIBS],
1624 uint32_t *count)
1625 {
1626 uint32_t vfd_fetch_idx = 0;
1627 uint32_t vfd_decode_idx = 0;
1628 uint32_t binding_instanced = 0; /* bitmask of instanced bindings */
1629
1630 for (uint32_t i = 0; i < info->vertexBindingDescriptionCount; i++) {
1631 const VkVertexInputBindingDescription *binding =
1632 &info->pVertexBindingDescriptions[i];
1633
1634 tu_cs_emit_regs(cs,
1635 A6XX_VFD_FETCH_STRIDE(vfd_fetch_idx, binding->stride));
1636
1637 if (binding->inputRate == VK_VERTEX_INPUT_RATE_INSTANCE)
1638 binding_instanced |= 1 << binding->binding;
1639
1640 bindings[vfd_fetch_idx] = binding->binding;
1641 vfd_fetch_idx++;
1642 }
1643
1644 /* TODO: emit all VFD_DECODE/VFD_DEST_CNTL in same (two) pkt4 */
1645
1646 for (uint32_t i = 0; i < info->vertexAttributeDescriptionCount; i++) {
1647 const VkVertexInputAttributeDescription *attr =
1648 &info->pVertexAttributeDescriptions[i];
1649 uint32_t binding_idx, input_idx;
1650
1651 for (binding_idx = 0; binding_idx < vfd_fetch_idx; binding_idx++) {
1652 if (bindings[binding_idx] == attr->binding)
1653 break;
1654 }
1655 assert(binding_idx < vfd_fetch_idx);
1656
1657 for (input_idx = 0; input_idx < vs->inputs_count; input_idx++) {
1658 if ((vs->inputs[input_idx].slot - VERT_ATTRIB_GENERIC0) == attr->location)
1659 break;
1660 }
1661
1662 /* attribute not used, skip it */
1663 if (input_idx == vs->inputs_count)
1664 continue;
1665
1666 const struct tu_native_format format = tu6_format_vtx(attr->format);
1667 tu_cs_emit_regs(cs,
1668 A6XX_VFD_DECODE_INSTR(vfd_decode_idx,
1669 .idx = binding_idx,
1670 .offset = attr->offset,
1671 .instanced = binding_instanced & (1 << attr->binding),
1672 .format = format.fmt,
1673 .swap = format.swap,
1674 .unk30 = 1,
1675 ._float = !vk_format_is_int(attr->format)),
1676 A6XX_VFD_DECODE_STEP_RATE(vfd_decode_idx, 1));
1677
1678 tu_cs_emit_regs(cs,
1679 A6XX_VFD_DEST_CNTL_INSTR(vfd_decode_idx,
1680 .writemask = vs->inputs[input_idx].compmask,
1681 .regid = vs->inputs[input_idx].regid));
1682
1683 vfd_decode_idx++;
1684 }
1685
1686 tu_cs_emit_regs(cs,
1687 A6XX_VFD_CONTROL_0(
1688 .fetch_cnt = vfd_fetch_idx,
1689 .decode_cnt = vfd_decode_idx));
1690
1691 *count = vfd_fetch_idx;
1692 }
1693
1694 static uint32_t
1695 tu6_guardband_adj(uint32_t v)
1696 {
1697 if (v > 256)
1698 return (uint32_t)(511.0 - 65.0 * (log2(v) - 8.0));
1699 else
1700 return 511;
1701 }
1702
1703 void
1704 tu6_emit_viewport(struct tu_cs *cs, const VkViewport *viewport)
1705 {
1706 float offsets[3];
1707 float scales[3];
1708 scales[0] = viewport->width / 2.0f;
1709 scales[1] = viewport->height / 2.0f;
1710 scales[2] = viewport->maxDepth - viewport->minDepth;
1711 offsets[0] = viewport->x + scales[0];
1712 offsets[1] = viewport->y + scales[1];
1713 offsets[2] = viewport->minDepth;
1714
1715 VkOffset2D min;
1716 VkOffset2D max;
1717 min.x = (int32_t) viewport->x;
1718 max.x = (int32_t) ceilf(viewport->x + viewport->width);
1719 if (viewport->height >= 0.0f) {
1720 min.y = (int32_t) viewport->y;
1721 max.y = (int32_t) ceilf(viewport->y + viewport->height);
1722 } else {
1723 min.y = (int32_t)(viewport->y + viewport->height);
1724 max.y = (int32_t) ceilf(viewport->y);
1725 }
1726 /* the spec allows viewport->height to be 0.0f */
1727 if (min.y == max.y)
1728 max.y++;
1729 assert(min.x >= 0 && min.x < max.x);
1730 assert(min.y >= 0 && min.y < max.y);
1731
1732 VkExtent2D guardband_adj;
1733 guardband_adj.width = tu6_guardband_adj(max.x - min.x);
1734 guardband_adj.height = tu6_guardband_adj(max.y - min.y);
1735
1736 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_CL_VPORT_XOFFSET_0, 6);
1737 tu_cs_emit(cs, A6XX_GRAS_CL_VPORT_XOFFSET_0(offsets[0]).value);
1738 tu_cs_emit(cs, A6XX_GRAS_CL_VPORT_XSCALE_0(scales[0]).value);
1739 tu_cs_emit(cs, A6XX_GRAS_CL_VPORT_YOFFSET_0(offsets[1]).value);
1740 tu_cs_emit(cs, A6XX_GRAS_CL_VPORT_YSCALE_0(scales[1]).value);
1741 tu_cs_emit(cs, A6XX_GRAS_CL_VPORT_ZOFFSET_0(offsets[2]).value);
1742 tu_cs_emit(cs, A6XX_GRAS_CL_VPORT_ZSCALE_0(scales[2]).value);
1743
1744 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0, 2);
1745 tu_cs_emit(cs, A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0_X(min.x) |
1746 A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0_Y(min.y));
1747 tu_cs_emit(cs, A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0_X(max.x - 1) |
1748 A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0_Y(max.y - 1));
1749
1750 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_CL_GUARDBAND_CLIP_ADJ, 1);
1751 tu_cs_emit(cs,
1752 A6XX_GRAS_CL_GUARDBAND_CLIP_ADJ_HORZ(guardband_adj.width) |
1753 A6XX_GRAS_CL_GUARDBAND_CLIP_ADJ_VERT(guardband_adj.height));
1754
1755 float z_clamp_min = MIN2(viewport->minDepth, viewport->maxDepth);
1756 float z_clamp_max = MAX2(viewport->minDepth, viewport->maxDepth);
1757
1758 tu_cs_emit_regs(cs,
1759 A6XX_GRAS_CL_Z_CLAMP_MIN(z_clamp_min),
1760 A6XX_GRAS_CL_Z_CLAMP_MAX(z_clamp_max));
1761
1762 tu_cs_emit_regs(cs,
1763 A6XX_RB_Z_CLAMP_MIN(z_clamp_min),
1764 A6XX_RB_Z_CLAMP_MAX(z_clamp_max));
1765 }
1766
1767 void
1768 tu6_emit_scissor(struct tu_cs *cs, const VkRect2D *scissor)
1769 {
1770 const VkOffset2D min = scissor->offset;
1771 const VkOffset2D max = {
1772 scissor->offset.x + scissor->extent.width,
1773 scissor->offset.y + scissor->extent.height,
1774 };
1775
1776 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0, 2);
1777 tu_cs_emit(cs, A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0_X(min.x) |
1778 A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0_Y(min.y));
1779 tu_cs_emit(cs, A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0_X(max.x - 1) |
1780 A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0_Y(max.y - 1));
1781 }
1782
1783 void
1784 tu6_emit_sample_locations(struct tu_cs *cs, const VkSampleLocationsInfoEXT *samp_loc)
1785 {
1786 if (!samp_loc) {
1787 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SAMPLE_CONFIG, 1);
1788 tu_cs_emit(cs, 0);
1789
1790 tu_cs_emit_pkt4(cs, REG_A6XX_RB_SAMPLE_CONFIG, 1);
1791 tu_cs_emit(cs, 0);
1792
1793 tu_cs_emit_pkt4(cs, REG_A6XX_SP_TP_SAMPLE_CONFIG, 1);
1794 tu_cs_emit(cs, 0);
1795 return;
1796 }
1797
1798 assert(samp_loc->sampleLocationsPerPixel == samp_loc->sampleLocationsCount);
1799 assert(samp_loc->sampleLocationGridSize.width == 1);
1800 assert(samp_loc->sampleLocationGridSize.height == 1);
1801
1802 uint32_t sample_config =
1803 A6XX_RB_SAMPLE_CONFIG_LOCATION_ENABLE;
1804 uint32_t sample_locations = 0;
1805 for (uint32_t i = 0; i < samp_loc->sampleLocationsCount; i++) {
1806 sample_locations |=
1807 (A6XX_RB_SAMPLE_LOCATION_0_SAMPLE_0_X(samp_loc->pSampleLocations[i].x) |
1808 A6XX_RB_SAMPLE_LOCATION_0_SAMPLE_0_Y(samp_loc->pSampleLocations[i].y)) << i*8;
1809 }
1810
1811 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SAMPLE_CONFIG, 2);
1812 tu_cs_emit(cs, sample_config);
1813 tu_cs_emit(cs, sample_locations);
1814
1815 tu_cs_emit_pkt4(cs, REG_A6XX_RB_SAMPLE_CONFIG, 2);
1816 tu_cs_emit(cs, sample_config);
1817 tu_cs_emit(cs, sample_locations);
1818
1819 tu_cs_emit_pkt4(cs, REG_A6XX_SP_TP_SAMPLE_CONFIG, 2);
1820 tu_cs_emit(cs, sample_config);
1821 tu_cs_emit(cs, sample_locations);
1822 }
1823
1824 static void
1825 tu6_emit_gras_unknowns(struct tu_cs *cs)
1826 {
1827 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_UNKNOWN_8001, 1);
1828 tu_cs_emit(cs, 0x0);
1829 }
1830
1831 static void
1832 tu6_emit_point_size(struct tu_cs *cs)
1833 {
1834 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SU_POINT_MINMAX, 2);
1835 tu_cs_emit(cs, A6XX_GRAS_SU_POINT_MINMAX_MIN(1.0f / 16.0f) |
1836 A6XX_GRAS_SU_POINT_MINMAX_MAX(4092.0f));
1837 tu_cs_emit(cs, A6XX_GRAS_SU_POINT_SIZE(1.0f).value);
1838 }
1839
1840 static uint32_t
1841 tu6_gras_su_cntl(const VkPipelineRasterizationStateCreateInfo *rast_info,
1842 VkSampleCountFlagBits samples)
1843 {
1844 uint32_t gras_su_cntl = 0;
1845
1846 if (rast_info->cullMode & VK_CULL_MODE_FRONT_BIT)
1847 gras_su_cntl |= A6XX_GRAS_SU_CNTL_CULL_FRONT;
1848 if (rast_info->cullMode & VK_CULL_MODE_BACK_BIT)
1849 gras_su_cntl |= A6XX_GRAS_SU_CNTL_CULL_BACK;
1850
1851 if (rast_info->frontFace == VK_FRONT_FACE_CLOCKWISE)
1852 gras_su_cntl |= A6XX_GRAS_SU_CNTL_FRONT_CW;
1853
1854 /* don't set A6XX_GRAS_SU_CNTL_LINEHALFWIDTH */
1855
1856 if (rast_info->depthBiasEnable)
1857 gras_su_cntl |= A6XX_GRAS_SU_CNTL_POLY_OFFSET;
1858
1859 if (samples > VK_SAMPLE_COUNT_1_BIT)
1860 gras_su_cntl |= A6XX_GRAS_SU_CNTL_MSAA_ENABLE;
1861
1862 return gras_su_cntl;
1863 }
1864
1865 void
1866 tu6_emit_gras_su_cntl(struct tu_cs *cs,
1867 uint32_t gras_su_cntl,
1868 float line_width)
1869 {
1870 assert((gras_su_cntl & A6XX_GRAS_SU_CNTL_LINEHALFWIDTH__MASK) == 0);
1871 gras_su_cntl |= A6XX_GRAS_SU_CNTL_LINEHALFWIDTH(line_width / 2.0f);
1872
1873 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SU_CNTL, 1);
1874 tu_cs_emit(cs, gras_su_cntl);
1875 }
1876
1877 void
1878 tu6_emit_depth_bias(struct tu_cs *cs,
1879 float constant_factor,
1880 float clamp,
1881 float slope_factor)
1882 {
1883 tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SU_POLY_OFFSET_SCALE, 3);
1884 tu_cs_emit(cs, A6XX_GRAS_SU_POLY_OFFSET_SCALE(slope_factor).value);
1885 tu_cs_emit(cs, A6XX_GRAS_SU_POLY_OFFSET_OFFSET(constant_factor).value);
1886 tu_cs_emit(cs, A6XX_GRAS_SU_POLY_OFFSET_OFFSET_CLAMP(clamp).value);
1887 }
1888
1889 static void
1890 tu6_emit_alpha_control_disable(struct tu_cs *cs)
1891 {
1892 tu_cs_emit_pkt4(cs, REG_A6XX_RB_ALPHA_CONTROL, 1);
1893 tu_cs_emit(cs, 0);
1894 }
1895
1896 static void
1897 tu6_emit_depth_control(struct tu_cs *cs,
1898 const VkPipelineDepthStencilStateCreateInfo *ds_info,
1899 const VkPipelineRasterizationStateCreateInfo *rast_info)
1900 {
1901 assert(!ds_info->depthBoundsTestEnable);
1902
1903 uint32_t rb_depth_cntl = 0;
1904 if (ds_info->depthTestEnable) {
1905 rb_depth_cntl |=
1906 A6XX_RB_DEPTH_CNTL_Z_ENABLE |
1907 A6XX_RB_DEPTH_CNTL_ZFUNC(tu6_compare_func(ds_info->depthCompareOp)) |
1908 A6XX_RB_DEPTH_CNTL_Z_TEST_ENABLE;
1909
1910 if (rast_info->depthClampEnable)
1911 rb_depth_cntl |= A6XX_RB_DEPTH_CNTL_Z_CLAMP_ENABLE;
1912
1913 if (ds_info->depthWriteEnable)
1914 rb_depth_cntl |= A6XX_RB_DEPTH_CNTL_Z_WRITE_ENABLE;
1915 }
1916
1917 tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_CNTL, 1);
1918 tu_cs_emit(cs, rb_depth_cntl);
1919 }
1920
1921 static void
1922 tu6_emit_stencil_control(struct tu_cs *cs,
1923 const VkPipelineDepthStencilStateCreateInfo *ds_info)
1924 {
1925 uint32_t rb_stencil_control = 0;
1926 if (ds_info->stencilTestEnable) {
1927 const VkStencilOpState *front = &ds_info->front;
1928 const VkStencilOpState *back = &ds_info->back;
1929 rb_stencil_control |=
1930 A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE |
1931 A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE_BF |
1932 A6XX_RB_STENCIL_CONTROL_STENCIL_READ |
1933 A6XX_RB_STENCIL_CONTROL_FUNC(tu6_compare_func(front->compareOp)) |
1934 A6XX_RB_STENCIL_CONTROL_FAIL(tu6_stencil_op(front->failOp)) |
1935 A6XX_RB_STENCIL_CONTROL_ZPASS(tu6_stencil_op(front->passOp)) |
1936 A6XX_RB_STENCIL_CONTROL_ZFAIL(tu6_stencil_op(front->depthFailOp)) |
1937 A6XX_RB_STENCIL_CONTROL_FUNC_BF(tu6_compare_func(back->compareOp)) |
1938 A6XX_RB_STENCIL_CONTROL_FAIL_BF(tu6_stencil_op(back->failOp)) |
1939 A6XX_RB_STENCIL_CONTROL_ZPASS_BF(tu6_stencil_op(back->passOp)) |
1940 A6XX_RB_STENCIL_CONTROL_ZFAIL_BF(tu6_stencil_op(back->depthFailOp));
1941 }
1942
1943 tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCIL_CONTROL, 1);
1944 tu_cs_emit(cs, rb_stencil_control);
1945 }
1946
1947 void
1948 tu6_emit_stencil_compare_mask(struct tu_cs *cs, uint32_t front, uint32_t back)
1949 {
1950 tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCILMASK, 1);
1951 tu_cs_emit(
1952 cs, A6XX_RB_STENCILMASK_MASK(front) | A6XX_RB_STENCILMASK_BFMASK(back));
1953 }
1954
1955 void
1956 tu6_emit_stencil_write_mask(struct tu_cs *cs, uint32_t front, uint32_t back)
1957 {
1958 tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCILWRMASK, 1);
1959 tu_cs_emit(cs, A6XX_RB_STENCILWRMASK_WRMASK(front) |
1960 A6XX_RB_STENCILWRMASK_BFWRMASK(back));
1961 }
1962
1963 void
1964 tu6_emit_stencil_reference(struct tu_cs *cs, uint32_t front, uint32_t back)
1965 {
1966 tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCILREF, 1);
1967 tu_cs_emit(cs,
1968 A6XX_RB_STENCILREF_REF(front) | A6XX_RB_STENCILREF_BFREF(back));
1969 }
1970
1971 static uint32_t
1972 tu6_rb_mrt_blend_control(const VkPipelineColorBlendAttachmentState *att,
1973 bool has_alpha)
1974 {
1975 const enum a3xx_rb_blend_opcode color_op = tu6_blend_op(att->colorBlendOp);
1976 const enum adreno_rb_blend_factor src_color_factor = tu6_blend_factor(
1977 has_alpha ? att->srcColorBlendFactor
1978 : tu_blend_factor_no_dst_alpha(att->srcColorBlendFactor));
1979 const enum adreno_rb_blend_factor dst_color_factor = tu6_blend_factor(
1980 has_alpha ? att->dstColorBlendFactor
1981 : tu_blend_factor_no_dst_alpha(att->dstColorBlendFactor));
1982 const enum a3xx_rb_blend_opcode alpha_op = tu6_blend_op(att->alphaBlendOp);
1983 const enum adreno_rb_blend_factor src_alpha_factor =
1984 tu6_blend_factor(att->srcAlphaBlendFactor);
1985 const enum adreno_rb_blend_factor dst_alpha_factor =
1986 tu6_blend_factor(att->dstAlphaBlendFactor);
1987
1988 return A6XX_RB_MRT_BLEND_CONTROL_RGB_SRC_FACTOR(src_color_factor) |
1989 A6XX_RB_MRT_BLEND_CONTROL_RGB_BLEND_OPCODE(color_op) |
1990 A6XX_RB_MRT_BLEND_CONTROL_RGB_DEST_FACTOR(dst_color_factor) |
1991 A6XX_RB_MRT_BLEND_CONTROL_ALPHA_SRC_FACTOR(src_alpha_factor) |
1992 A6XX_RB_MRT_BLEND_CONTROL_ALPHA_BLEND_OPCODE(alpha_op) |
1993 A6XX_RB_MRT_BLEND_CONTROL_ALPHA_DEST_FACTOR(dst_alpha_factor);
1994 }
1995
1996 static uint32_t
1997 tu6_rb_mrt_control(const VkPipelineColorBlendAttachmentState *att,
1998 uint32_t rb_mrt_control_rop,
1999 bool is_int,
2000 bool has_alpha)
2001 {
2002 uint32_t rb_mrt_control =
2003 A6XX_RB_MRT_CONTROL_COMPONENT_ENABLE(att->colorWriteMask);
2004
2005 /* ignore blending and logic op for integer attachments */
2006 if (is_int) {
2007 rb_mrt_control |= A6XX_RB_MRT_CONTROL_ROP_CODE(ROP_COPY);
2008 return rb_mrt_control;
2009 }
2010
2011 rb_mrt_control |= rb_mrt_control_rop;
2012
2013 if (att->blendEnable) {
2014 rb_mrt_control |= A6XX_RB_MRT_CONTROL_BLEND;
2015
2016 if (has_alpha)
2017 rb_mrt_control |= A6XX_RB_MRT_CONTROL_BLEND2;
2018 }
2019
2020 return rb_mrt_control;
2021 }
2022
2023 static void
2024 tu6_emit_rb_mrt_controls(struct tu_cs *cs,
2025 const VkPipelineColorBlendStateCreateInfo *blend_info,
2026 const VkFormat attachment_formats[MAX_RTS],
2027 uint32_t *blend_enable_mask)
2028 {
2029 *blend_enable_mask = 0;
2030
2031 bool rop_reads_dst = false;
2032 uint32_t rb_mrt_control_rop = 0;
2033 if (blend_info->logicOpEnable) {
2034 rop_reads_dst = tu_logic_op_reads_dst(blend_info->logicOp);
2035 rb_mrt_control_rop =
2036 A6XX_RB_MRT_CONTROL_ROP_ENABLE |
2037 A6XX_RB_MRT_CONTROL_ROP_CODE(tu6_rop(blend_info->logicOp));
2038 }
2039
2040 for (uint32_t i = 0; i < blend_info->attachmentCount; i++) {
2041 const VkPipelineColorBlendAttachmentState *att =
2042 &blend_info->pAttachments[i];
2043 const VkFormat format = attachment_formats[i];
2044
2045 uint32_t rb_mrt_control = 0;
2046 uint32_t rb_mrt_blend_control = 0;
2047 if (format != VK_FORMAT_UNDEFINED) {
2048 const bool is_int = vk_format_is_int(format);
2049 const bool has_alpha = vk_format_has_alpha(format);
2050
2051 rb_mrt_control =
2052 tu6_rb_mrt_control(att, rb_mrt_control_rop, is_int, has_alpha);
2053 rb_mrt_blend_control = tu6_rb_mrt_blend_control(att, has_alpha);
2054
2055 if (att->blendEnable || rop_reads_dst)
2056 *blend_enable_mask |= 1 << i;
2057 }
2058
2059 tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_CONTROL(i), 2);
2060 tu_cs_emit(cs, rb_mrt_control);
2061 tu_cs_emit(cs, rb_mrt_blend_control);
2062 }
2063 }
2064
2065 static void
2066 tu6_emit_blend_control(struct tu_cs *cs,
2067 uint32_t blend_enable_mask,
2068 bool dual_src_blend,
2069 const VkPipelineMultisampleStateCreateInfo *msaa_info)
2070 {
2071 assert(!msaa_info->alphaToOneEnable);
2072
2073 uint32_t sp_blend_cntl = A6XX_SP_BLEND_CNTL_UNK8;
2074 if (blend_enable_mask)
2075 sp_blend_cntl |= A6XX_SP_BLEND_CNTL_ENABLED;
2076 if (dual_src_blend)
2077 sp_blend_cntl |= A6XX_SP_BLEND_CNTL_DUAL_COLOR_IN_ENABLE;
2078 if (msaa_info->alphaToCoverageEnable)
2079 sp_blend_cntl |= A6XX_SP_BLEND_CNTL_ALPHA_TO_COVERAGE;
2080
2081 const uint32_t sample_mask =
2082 msaa_info->pSampleMask ? *msaa_info->pSampleMask
2083 : ((1 << msaa_info->rasterizationSamples) - 1);
2084
2085 /* set A6XX_RB_BLEND_CNTL_INDEPENDENT_BLEND only when enabled? */
2086 uint32_t rb_blend_cntl =
2087 A6XX_RB_BLEND_CNTL_ENABLE_BLEND(blend_enable_mask) |
2088 A6XX_RB_BLEND_CNTL_INDEPENDENT_BLEND |
2089 A6XX_RB_BLEND_CNTL_SAMPLE_MASK(sample_mask);
2090 if (dual_src_blend)
2091 rb_blend_cntl |= A6XX_RB_BLEND_CNTL_DUAL_COLOR_IN_ENABLE;
2092 if (msaa_info->alphaToCoverageEnable)
2093 rb_blend_cntl |= A6XX_RB_BLEND_CNTL_ALPHA_TO_COVERAGE;
2094
2095 tu_cs_emit_pkt4(cs, REG_A6XX_SP_BLEND_CNTL, 1);
2096 tu_cs_emit(cs, sp_blend_cntl);
2097
2098 tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLEND_CNTL, 1);
2099 tu_cs_emit(cs, rb_blend_cntl);
2100 }
2101
2102 void
2103 tu6_emit_blend_constants(struct tu_cs *cs, const float constants[4])
2104 {
2105 tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLEND_RED_F32, 4);
2106 tu_cs_emit_array(cs, (const uint32_t *) constants, 4);
2107 }
2108
2109 static VkResult
2110 tu_pipeline_create(struct tu_device *dev,
2111 struct tu_pipeline_layout *layout,
2112 bool compute,
2113 const VkAllocationCallbacks *pAllocator,
2114 struct tu_pipeline **out_pipeline)
2115 {
2116 struct tu_pipeline *pipeline =
2117 vk_zalloc2(&dev->alloc, pAllocator, sizeof(*pipeline), 8,
2118 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2119 if (!pipeline)
2120 return VK_ERROR_OUT_OF_HOST_MEMORY;
2121
2122 tu_cs_init(&pipeline->cs, dev, TU_CS_MODE_SUB_STREAM, 2048);
2123
2124 /* Reserve the space now such that tu_cs_begin_sub_stream never fails. Note
2125 * that LOAD_STATE can potentially take up a large amount of space so we
2126 * calculate its size explicitly.
2127 */
2128 unsigned load_state_size = tu6_load_state_size(layout, compute);
2129 VkResult result = tu_cs_reserve_space(&pipeline->cs, 2048 + load_state_size);
2130 if (result != VK_SUCCESS) {
2131 vk_free2(&dev->alloc, pAllocator, pipeline);
2132 return result;
2133 }
2134
2135 *out_pipeline = pipeline;
2136
2137 return VK_SUCCESS;
2138 }
2139
2140 static VkResult
2141 tu_pipeline_builder_compile_shaders(struct tu_pipeline_builder *builder)
2142 {
2143 const VkPipelineShaderStageCreateInfo *stage_infos[MESA_SHADER_STAGES] = {
2144 NULL
2145 };
2146 for (uint32_t i = 0; i < builder->create_info->stageCount; i++) {
2147 gl_shader_stage stage =
2148 tu_shader_stage(builder->create_info->pStages[i].stage);
2149 stage_infos[stage] = &builder->create_info->pStages[i];
2150 }
2151
2152 struct tu_shader_compile_options options;
2153 tu_shader_compile_options_init(&options, builder->create_info);
2154
2155 /* compile shaders in reverse order */
2156 struct tu_shader *next_stage_shader = NULL;
2157 for (gl_shader_stage stage = MESA_SHADER_STAGES - 1;
2158 stage > MESA_SHADER_NONE; stage--) {
2159 const VkPipelineShaderStageCreateInfo *stage_info = stage_infos[stage];
2160 if (!stage_info && stage != MESA_SHADER_FRAGMENT)
2161 continue;
2162
2163 struct tu_shader *shader =
2164 tu_shader_create(builder->device, stage, stage_info, builder->layout,
2165 builder->alloc);
2166 if (!shader)
2167 return VK_ERROR_OUT_OF_HOST_MEMORY;
2168
2169 VkResult result =
2170 tu_shader_compile(builder->device, shader, next_stage_shader,
2171 &options, builder->alloc);
2172 if (result != VK_SUCCESS)
2173 return result;
2174
2175 builder->shaders[stage] = shader;
2176 builder->shader_offsets[stage] = builder->shader_total_size;
2177 builder->shader_total_size +=
2178 sizeof(uint32_t) * shader->variants[0].info.sizedwords;
2179
2180 next_stage_shader = shader;
2181 }
2182
2183 if (builder->shaders[MESA_SHADER_VERTEX]->has_binning_pass) {
2184 const struct tu_shader *vs = builder->shaders[MESA_SHADER_VERTEX];
2185 const struct ir3_shader_variant *variant;
2186
2187 if (vs->ir3_shader.stream_output.num_outputs)
2188 variant = &vs->variants[0];
2189 else
2190 variant = &vs->variants[1];
2191
2192 builder->binning_vs_offset = builder->shader_total_size;
2193 builder->shader_total_size +=
2194 sizeof(uint32_t) * variant->info.sizedwords;
2195 }
2196
2197 return VK_SUCCESS;
2198 }
2199
2200 static VkResult
2201 tu_pipeline_builder_upload_shaders(struct tu_pipeline_builder *builder,
2202 struct tu_pipeline *pipeline)
2203 {
2204 struct tu_bo *bo = &pipeline->program.binary_bo;
2205
2206 VkResult result =
2207 tu_bo_init_new(builder->device, bo, builder->shader_total_size);
2208 if (result != VK_SUCCESS)
2209 return result;
2210
2211 result = tu_bo_map(builder->device, bo);
2212 if (result != VK_SUCCESS)
2213 return result;
2214
2215 for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++) {
2216 const struct tu_shader *shader = builder->shaders[i];
2217 if (!shader)
2218 continue;
2219
2220 memcpy(bo->map + builder->shader_offsets[i], shader->binary,
2221 sizeof(uint32_t) * shader->variants[0].info.sizedwords);
2222 }
2223
2224 if (builder->shaders[MESA_SHADER_VERTEX]->has_binning_pass) {
2225 const struct tu_shader *vs = builder->shaders[MESA_SHADER_VERTEX];
2226 const struct ir3_shader_variant *variant;
2227 void *bin;
2228
2229 if (vs->ir3_shader.stream_output.num_outputs) {
2230 variant = &vs->variants[0];
2231 bin = vs->binary;
2232 } else {
2233 variant = &vs->variants[1];
2234 bin = vs->binning_binary;
2235 }
2236
2237 memcpy(bo->map + builder->binning_vs_offset, bin,
2238 sizeof(uint32_t) * variant->info.sizedwords);
2239 }
2240
2241 return VK_SUCCESS;
2242 }
2243
2244 static void
2245 tu_pipeline_builder_parse_dynamic(struct tu_pipeline_builder *builder,
2246 struct tu_pipeline *pipeline)
2247 {
2248 const VkPipelineDynamicStateCreateInfo *dynamic_info =
2249 builder->create_info->pDynamicState;
2250
2251 if (!dynamic_info)
2252 return;
2253
2254 for (uint32_t i = 0; i < dynamic_info->dynamicStateCount; i++) {
2255 pipeline->dynamic_state.mask |=
2256 tu_dynamic_state_bit(dynamic_info->pDynamicStates[i]);
2257 }
2258 }
2259
2260 static void
2261 tu_pipeline_set_linkage(struct tu_program_descriptor_linkage *link,
2262 struct tu_shader *shader,
2263 struct ir3_shader_variant *v)
2264 {
2265 link->ubo_state = v->shader->ubo_state;
2266 link->const_state = v->shader->const_state;
2267 link->constlen = v->constlen;
2268 link->push_consts = shader->push_consts;
2269 }
2270
2271 static void
2272 tu_pipeline_builder_parse_shader_stages(struct tu_pipeline_builder *builder,
2273 struct tu_pipeline *pipeline)
2274 {
2275 struct tu_cs prog_cs;
2276 tu_cs_begin_sub_stream(&pipeline->cs, 512, &prog_cs);
2277 tu6_emit_program(&prog_cs, builder, &pipeline->program.binary_bo, false, &pipeline->streamout);
2278 pipeline->program.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &prog_cs);
2279
2280 tu_cs_begin_sub_stream(&pipeline->cs, 512, &prog_cs);
2281 tu6_emit_program(&prog_cs, builder, &pipeline->program.binary_bo, true, &pipeline->streamout);
2282 pipeline->program.binning_state_ib =
2283 tu_cs_end_sub_stream(&pipeline->cs, &prog_cs);
2284
2285 VkShaderStageFlags stages = 0;
2286 for (unsigned i = 0; i < builder->create_info->stageCount; i++) {
2287 stages |= builder->create_info->pStages[i].stage;
2288 }
2289 pipeline->active_stages = stages;
2290
2291 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
2292 if (!builder->shaders[i])
2293 continue;
2294
2295 tu_pipeline_set_linkage(&pipeline->program.link[i],
2296 builder->shaders[i],
2297 &builder->shaders[i]->variants[0]);
2298 }
2299
2300 if (builder->shaders[MESA_SHADER_FRAGMENT]) {
2301 memcpy(pipeline->program.input_attachment_idx,
2302 builder->shaders[MESA_SHADER_FRAGMENT]->attachment_idx,
2303 sizeof(pipeline->program.input_attachment_idx));
2304 }
2305 }
2306
2307 static void
2308 tu_pipeline_builder_parse_vertex_input(struct tu_pipeline_builder *builder,
2309 struct tu_pipeline *pipeline)
2310 {
2311 const VkPipelineVertexInputStateCreateInfo *vi_info =
2312 builder->create_info->pVertexInputState;
2313 const struct tu_shader *vs = builder->shaders[MESA_SHADER_VERTEX];
2314
2315 struct tu_cs vi_cs;
2316 tu_cs_begin_sub_stream(&pipeline->cs,
2317 MAX_VERTEX_ATTRIBS * 7 + 2, &vi_cs);
2318 tu6_emit_vertex_input(&vi_cs, &vs->variants[0], vi_info,
2319 pipeline->vi.bindings, &pipeline->vi.count);
2320 pipeline->vi.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &vi_cs);
2321
2322 if (vs->has_binning_pass) {
2323 tu_cs_begin_sub_stream(&pipeline->cs,
2324 MAX_VERTEX_ATTRIBS * 7 + 2, &vi_cs);
2325 tu6_emit_vertex_input(
2326 &vi_cs, &vs->variants[1], vi_info, pipeline->vi.binning_bindings,
2327 &pipeline->vi.binning_count);
2328 pipeline->vi.binning_state_ib =
2329 tu_cs_end_sub_stream(&pipeline->cs, &vi_cs);
2330 }
2331 }
2332
2333 static void
2334 tu_pipeline_builder_parse_input_assembly(struct tu_pipeline_builder *builder,
2335 struct tu_pipeline *pipeline)
2336 {
2337 const VkPipelineInputAssemblyStateCreateInfo *ia_info =
2338 builder->create_info->pInputAssemblyState;
2339
2340 pipeline->ia.primtype = tu6_primtype(ia_info->topology);
2341 pipeline->ia.primitive_restart = ia_info->primitiveRestartEnable;
2342 }
2343
2344 static void
2345 tu_pipeline_builder_parse_viewport(struct tu_pipeline_builder *builder,
2346 struct tu_pipeline *pipeline)
2347 {
2348 /* The spec says:
2349 *
2350 * pViewportState is a pointer to an instance of the
2351 * VkPipelineViewportStateCreateInfo structure, and is ignored if the
2352 * pipeline has rasterization disabled."
2353 *
2354 * We leave the relevant registers stale in that case.
2355 */
2356 if (builder->rasterizer_discard)
2357 return;
2358
2359 const VkPipelineViewportStateCreateInfo *vp_info =
2360 builder->create_info->pViewportState;
2361
2362 struct tu_cs vp_cs;
2363 tu_cs_begin_sub_stream(&pipeline->cs, 21, &vp_cs);
2364
2365 if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_VIEWPORT)) {
2366 assert(vp_info->viewportCount == 1);
2367 tu6_emit_viewport(&vp_cs, vp_info->pViewports);
2368 }
2369
2370 if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_SCISSOR)) {
2371 assert(vp_info->scissorCount == 1);
2372 tu6_emit_scissor(&vp_cs, vp_info->pScissors);
2373 }
2374
2375 pipeline->vp.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &vp_cs);
2376 }
2377
2378 static void
2379 tu_pipeline_builder_parse_rasterization(struct tu_pipeline_builder *builder,
2380 struct tu_pipeline *pipeline)
2381 {
2382 const VkPipelineRasterizationStateCreateInfo *rast_info =
2383 builder->create_info->pRasterizationState;
2384
2385 assert(rast_info->polygonMode == VK_POLYGON_MODE_FILL);
2386
2387 struct tu_cs rast_cs;
2388 tu_cs_begin_sub_stream(&pipeline->cs, 20, &rast_cs);
2389
2390
2391 tu_cs_emit_regs(&rast_cs,
2392 A6XX_GRAS_CL_CNTL(
2393 .znear_clip_disable = rast_info->depthClampEnable,
2394 .zfar_clip_disable = rast_info->depthClampEnable,
2395 .unk5 = rast_info->depthClampEnable,
2396 .zero_gb_scale_z = 1,
2397 .vp_clip_code_ignore = 1));
2398 /* move to hw ctx init? */
2399 tu6_emit_gras_unknowns(&rast_cs);
2400 tu6_emit_point_size(&rast_cs);
2401
2402 const uint32_t gras_su_cntl =
2403 tu6_gras_su_cntl(rast_info, builder->samples);
2404
2405 if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_LINE_WIDTH))
2406 tu6_emit_gras_su_cntl(&rast_cs, gras_su_cntl, rast_info->lineWidth);
2407
2408 if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_DEPTH_BIAS)) {
2409 tu6_emit_depth_bias(&rast_cs, rast_info->depthBiasConstantFactor,
2410 rast_info->depthBiasClamp,
2411 rast_info->depthBiasSlopeFactor);
2412 }
2413
2414 pipeline->rast.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &rast_cs);
2415
2416 pipeline->rast.gras_su_cntl = gras_su_cntl;
2417 }
2418
2419 static void
2420 tu_pipeline_builder_parse_depth_stencil(struct tu_pipeline_builder *builder,
2421 struct tu_pipeline *pipeline)
2422 {
2423 /* The spec says:
2424 *
2425 * pDepthStencilState is a pointer to an instance of the
2426 * VkPipelineDepthStencilStateCreateInfo structure, and is ignored if
2427 * the pipeline has rasterization disabled or if the subpass of the
2428 * render pass the pipeline is created against does not use a
2429 * depth/stencil attachment.
2430 *
2431 * Disable both depth and stencil tests if there is no ds attachment,
2432 * Disable depth test if ds attachment is S8_UINT, since S8_UINT defines
2433 * only the separate stencil attachment
2434 */
2435 static const VkPipelineDepthStencilStateCreateInfo dummy_ds_info;
2436 const VkPipelineDepthStencilStateCreateInfo *ds_info =
2437 builder->depth_attachment_format != VK_FORMAT_UNDEFINED
2438 ? builder->create_info->pDepthStencilState
2439 : &dummy_ds_info;
2440 const VkPipelineDepthStencilStateCreateInfo *ds_info_depth =
2441 builder->depth_attachment_format != VK_FORMAT_S8_UINT
2442 ? ds_info : &dummy_ds_info;
2443
2444 struct tu_cs ds_cs;
2445 tu_cs_begin_sub_stream(&pipeline->cs, 12, &ds_cs);
2446
2447 /* move to hw ctx init? */
2448 tu6_emit_alpha_control_disable(&ds_cs);
2449
2450 tu6_emit_depth_control(&ds_cs, ds_info_depth,
2451 builder->create_info->pRasterizationState);
2452 tu6_emit_stencil_control(&ds_cs, ds_info);
2453
2454 if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_COMPARE_MASK)) {
2455 tu6_emit_stencil_compare_mask(&ds_cs, ds_info->front.compareMask,
2456 ds_info->back.compareMask);
2457 }
2458 if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_WRITE_MASK)) {
2459 tu6_emit_stencil_write_mask(&ds_cs, ds_info->front.writeMask,
2460 ds_info->back.writeMask);
2461 }
2462 if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_REFERENCE)) {
2463 tu6_emit_stencil_reference(&ds_cs, ds_info->front.reference,
2464 ds_info->back.reference);
2465 }
2466
2467 pipeline->ds.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &ds_cs);
2468 }
2469
2470 static void
2471 tu_pipeline_builder_parse_multisample_and_color_blend(
2472 struct tu_pipeline_builder *builder, struct tu_pipeline *pipeline)
2473 {
2474 /* The spec says:
2475 *
2476 * pMultisampleState is a pointer to an instance of the
2477 * VkPipelineMultisampleStateCreateInfo, and is ignored if the pipeline
2478 * has rasterization disabled.
2479 *
2480 * Also,
2481 *
2482 * pColorBlendState is a pointer to an instance of the
2483 * VkPipelineColorBlendStateCreateInfo structure, and is ignored if the
2484 * pipeline has rasterization disabled or if the subpass of the render
2485 * pass the pipeline is created against does not use any color
2486 * attachments.
2487 *
2488 * We leave the relevant registers stale when rasterization is disabled.
2489 */
2490 if (builder->rasterizer_discard)
2491 return;
2492
2493 static const VkPipelineColorBlendStateCreateInfo dummy_blend_info;
2494 const VkPipelineMultisampleStateCreateInfo *msaa_info =
2495 builder->create_info->pMultisampleState;
2496 const VkPipelineColorBlendStateCreateInfo *blend_info =
2497 builder->use_color_attachments ? builder->create_info->pColorBlendState
2498 : &dummy_blend_info;
2499
2500 struct tu_cs blend_cs;
2501 tu_cs_begin_sub_stream(&pipeline->cs, MAX_RTS * 3 + 18, &blend_cs);
2502
2503 uint32_t blend_enable_mask;
2504 tu6_emit_rb_mrt_controls(&blend_cs, blend_info,
2505 builder->color_attachment_formats,
2506 &blend_enable_mask);
2507
2508 if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_BLEND_CONSTANTS))
2509 tu6_emit_blend_constants(&blend_cs, blend_info->blendConstants);
2510
2511 if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_SAMPLE_LOCATIONS)) {
2512 const struct VkPipelineSampleLocationsStateCreateInfoEXT *sample_locations =
2513 vk_find_struct_const(msaa_info->pNext, PIPELINE_SAMPLE_LOCATIONS_STATE_CREATE_INFO_EXT);
2514 const VkSampleLocationsInfoEXT *samp_loc = NULL;
2515
2516 if (sample_locations && sample_locations->sampleLocationsEnable)
2517 samp_loc = &sample_locations->sampleLocationsInfo;
2518
2519 tu6_emit_sample_locations(&blend_cs, samp_loc);
2520 }
2521
2522 tu6_emit_blend_control(&blend_cs, blend_enable_mask,
2523 builder->use_dual_src_blend, msaa_info);
2524
2525 pipeline->blend.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &blend_cs);
2526 }
2527
2528 static void
2529 tu_pipeline_finish(struct tu_pipeline *pipeline,
2530 struct tu_device *dev,
2531 const VkAllocationCallbacks *alloc)
2532 {
2533 tu_cs_finish(&pipeline->cs);
2534
2535 if (pipeline->program.binary_bo.gem_handle)
2536 tu_bo_finish(dev, &pipeline->program.binary_bo);
2537 }
2538
2539 static VkResult
2540 tu_pipeline_builder_build(struct tu_pipeline_builder *builder,
2541 struct tu_pipeline **pipeline)
2542 {
2543 VkResult result = tu_pipeline_create(builder->device, builder->layout,
2544 false, builder->alloc, pipeline);
2545 if (result != VK_SUCCESS)
2546 return result;
2547
2548 (*pipeline)->layout = builder->layout;
2549
2550 /* compile and upload shaders */
2551 result = tu_pipeline_builder_compile_shaders(builder);
2552 if (result == VK_SUCCESS)
2553 result = tu_pipeline_builder_upload_shaders(builder, *pipeline);
2554 if (result != VK_SUCCESS) {
2555 tu_pipeline_finish(*pipeline, builder->device, builder->alloc);
2556 vk_free2(&builder->device->alloc, builder->alloc, *pipeline);
2557 *pipeline = VK_NULL_HANDLE;
2558
2559 return result;
2560 }
2561
2562 tu_pipeline_builder_parse_dynamic(builder, *pipeline);
2563 tu_pipeline_builder_parse_shader_stages(builder, *pipeline);
2564 tu_pipeline_builder_parse_vertex_input(builder, *pipeline);
2565 tu_pipeline_builder_parse_input_assembly(builder, *pipeline);
2566 tu_pipeline_builder_parse_viewport(builder, *pipeline);
2567 tu_pipeline_builder_parse_rasterization(builder, *pipeline);
2568 tu_pipeline_builder_parse_depth_stencil(builder, *pipeline);
2569 tu_pipeline_builder_parse_multisample_and_color_blend(builder, *pipeline);
2570 tu6_emit_load_state(*pipeline, false);
2571
2572 /* we should have reserved enough space upfront such that the CS never
2573 * grows
2574 */
2575 assert((*pipeline)->cs.bo_count == 1);
2576
2577 return VK_SUCCESS;
2578 }
2579
2580 static void
2581 tu_pipeline_builder_finish(struct tu_pipeline_builder *builder)
2582 {
2583 for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++) {
2584 if (!builder->shaders[i])
2585 continue;
2586 tu_shader_destroy(builder->device, builder->shaders[i], builder->alloc);
2587 }
2588 }
2589
2590 static void
2591 tu_pipeline_builder_init_graphics(
2592 struct tu_pipeline_builder *builder,
2593 struct tu_device *dev,
2594 struct tu_pipeline_cache *cache,
2595 const VkGraphicsPipelineCreateInfo *create_info,
2596 const VkAllocationCallbacks *alloc)
2597 {
2598 TU_FROM_HANDLE(tu_pipeline_layout, layout, create_info->layout);
2599
2600 *builder = (struct tu_pipeline_builder) {
2601 .device = dev,
2602 .cache = cache,
2603 .create_info = create_info,
2604 .alloc = alloc,
2605 .layout = layout,
2606 };
2607
2608 builder->rasterizer_discard =
2609 create_info->pRasterizationState->rasterizerDiscardEnable;
2610
2611 if (builder->rasterizer_discard) {
2612 builder->samples = VK_SAMPLE_COUNT_1_BIT;
2613 } else {
2614 builder->samples = create_info->pMultisampleState->rasterizationSamples;
2615
2616 const struct tu_render_pass *pass =
2617 tu_render_pass_from_handle(create_info->renderPass);
2618 const struct tu_subpass *subpass =
2619 &pass->subpasses[create_info->subpass];
2620
2621 const uint32_t a = subpass->depth_stencil_attachment.attachment;
2622 builder->depth_attachment_format = (a != VK_ATTACHMENT_UNUSED) ?
2623 pass->attachments[a].format : VK_FORMAT_UNDEFINED;
2624
2625 assert(subpass->color_count == 0 ||
2626 !create_info->pColorBlendState ||
2627 subpass->color_count == create_info->pColorBlendState->attachmentCount);
2628 builder->color_attachment_count = subpass->color_count;
2629 for (uint32_t i = 0; i < subpass->color_count; i++) {
2630 const uint32_t a = subpass->color_attachments[i].attachment;
2631 if (a == VK_ATTACHMENT_UNUSED)
2632 continue;
2633
2634 builder->color_attachment_formats[i] = pass->attachments[a].format;
2635 builder->use_color_attachments = true;
2636 builder->render_components |= 0xf << (i * 4);
2637 }
2638
2639 if (tu_blend_state_is_dual_src(create_info->pColorBlendState)) {
2640 builder->color_attachment_count++;
2641 builder->use_dual_src_blend = true;
2642 /* dual source blending has an extra fs output in the 2nd slot */
2643 if (subpass->color_attachments[0].attachment != VK_ATTACHMENT_UNUSED)
2644 builder->render_components |= 0xf << 4;
2645 }
2646 }
2647 }
2648
2649 static VkResult
2650 tu_graphics_pipeline_create(VkDevice device,
2651 VkPipelineCache pipelineCache,
2652 const VkGraphicsPipelineCreateInfo *pCreateInfo,
2653 const VkAllocationCallbacks *pAllocator,
2654 VkPipeline *pPipeline)
2655 {
2656 TU_FROM_HANDLE(tu_device, dev, device);
2657 TU_FROM_HANDLE(tu_pipeline_cache, cache, pipelineCache);
2658
2659 struct tu_pipeline_builder builder;
2660 tu_pipeline_builder_init_graphics(&builder, dev, cache,
2661 pCreateInfo, pAllocator);
2662
2663 struct tu_pipeline *pipeline = NULL;
2664 VkResult result = tu_pipeline_builder_build(&builder, &pipeline);
2665 tu_pipeline_builder_finish(&builder);
2666
2667 if (result == VK_SUCCESS)
2668 *pPipeline = tu_pipeline_to_handle(pipeline);
2669 else
2670 *pPipeline = VK_NULL_HANDLE;
2671
2672 return result;
2673 }
2674
2675 VkResult
2676 tu_CreateGraphicsPipelines(VkDevice device,
2677 VkPipelineCache pipelineCache,
2678 uint32_t count,
2679 const VkGraphicsPipelineCreateInfo *pCreateInfos,
2680 const VkAllocationCallbacks *pAllocator,
2681 VkPipeline *pPipelines)
2682 {
2683 VkResult final_result = VK_SUCCESS;
2684
2685 for (uint32_t i = 0; i < count; i++) {
2686 VkResult result = tu_graphics_pipeline_create(device, pipelineCache,
2687 &pCreateInfos[i], pAllocator,
2688 &pPipelines[i]);
2689
2690 if (result != VK_SUCCESS)
2691 final_result = result;
2692 }
2693
2694 return final_result;
2695 }
2696
2697 static void
2698 tu6_emit_compute_program(struct tu_cs *cs,
2699 struct tu_shader *shader,
2700 const struct tu_bo *binary_bo)
2701 {
2702 const struct ir3_shader_variant *v = &shader->variants[0];
2703
2704 tu6_emit_cs_config(cs, shader, v);
2705
2706 /* The compute program is the only one in the pipeline, so 0 offset. */
2707 tu6_emit_shader_object(cs, MESA_SHADER_COMPUTE, v, binary_bo, 0);
2708
2709 tu6_emit_immediates(cs, v, CP_LOAD_STATE6_FRAG, SB6_CS_SHADER);
2710 }
2711
2712 static VkResult
2713 tu_compute_upload_shader(VkDevice device,
2714 struct tu_pipeline *pipeline,
2715 struct tu_shader *shader)
2716 {
2717 TU_FROM_HANDLE(tu_device, dev, device);
2718 struct tu_bo *bo = &pipeline->program.binary_bo;
2719 struct ir3_shader_variant *v = &shader->variants[0];
2720
2721 uint32_t shader_size = sizeof(uint32_t) * v->info.sizedwords;
2722 VkResult result =
2723 tu_bo_init_new(dev, bo, shader_size);
2724 if (result != VK_SUCCESS)
2725 return result;
2726
2727 result = tu_bo_map(dev, bo);
2728 if (result != VK_SUCCESS)
2729 return result;
2730
2731 memcpy(bo->map, shader->binary, shader_size);
2732
2733 return VK_SUCCESS;
2734 }
2735
2736
2737 static VkResult
2738 tu_compute_pipeline_create(VkDevice device,
2739 VkPipelineCache _cache,
2740 const VkComputePipelineCreateInfo *pCreateInfo,
2741 const VkAllocationCallbacks *pAllocator,
2742 VkPipeline *pPipeline)
2743 {
2744 TU_FROM_HANDLE(tu_device, dev, device);
2745 TU_FROM_HANDLE(tu_pipeline_layout, layout, pCreateInfo->layout);
2746 const VkPipelineShaderStageCreateInfo *stage_info = &pCreateInfo->stage;
2747 VkResult result;
2748
2749 struct tu_pipeline *pipeline;
2750
2751 *pPipeline = VK_NULL_HANDLE;
2752
2753 result = tu_pipeline_create(dev, layout, true, pAllocator, &pipeline);
2754 if (result != VK_SUCCESS)
2755 return result;
2756
2757 pipeline->layout = layout;
2758
2759 struct tu_shader_compile_options options;
2760 tu_shader_compile_options_init(&options, NULL);
2761
2762 struct tu_shader *shader =
2763 tu_shader_create(dev, MESA_SHADER_COMPUTE, stage_info, layout, pAllocator);
2764 if (!shader) {
2765 result = VK_ERROR_OUT_OF_HOST_MEMORY;
2766 goto fail;
2767 }
2768
2769 result = tu_shader_compile(dev, shader, NULL, &options, pAllocator);
2770 if (result != VK_SUCCESS)
2771 goto fail;
2772
2773 struct ir3_shader_variant *v = &shader->variants[0];
2774
2775 tu_pipeline_set_linkage(&pipeline->program.link[MESA_SHADER_COMPUTE],
2776 shader, v);
2777
2778 result = tu_compute_upload_shader(device, pipeline, shader);
2779 if (result != VK_SUCCESS)
2780 goto fail;
2781
2782 for (int i = 0; i < 3; i++)
2783 pipeline->compute.local_size[i] = v->shader->nir->info.cs.local_size[i];
2784
2785 struct tu_cs prog_cs;
2786 tu_cs_begin_sub_stream(&pipeline->cs, 512, &prog_cs);
2787 tu6_emit_compute_program(&prog_cs, shader, &pipeline->program.binary_bo);
2788 pipeline->program.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &prog_cs);
2789
2790 tu6_emit_load_state(pipeline, true);
2791
2792 *pPipeline = tu_pipeline_to_handle(pipeline);
2793 return VK_SUCCESS;
2794
2795 fail:
2796 if (shader)
2797 tu_shader_destroy(dev, shader, pAllocator);
2798
2799 tu_pipeline_finish(pipeline, dev, pAllocator);
2800 vk_free2(&dev->alloc, pAllocator, pipeline);
2801
2802 return result;
2803 }
2804
2805 VkResult
2806 tu_CreateComputePipelines(VkDevice device,
2807 VkPipelineCache pipelineCache,
2808 uint32_t count,
2809 const VkComputePipelineCreateInfo *pCreateInfos,
2810 const VkAllocationCallbacks *pAllocator,
2811 VkPipeline *pPipelines)
2812 {
2813 VkResult final_result = VK_SUCCESS;
2814
2815 for (uint32_t i = 0; i < count; i++) {
2816 VkResult result = tu_compute_pipeline_create(device, pipelineCache,
2817 &pCreateInfos[i],
2818 pAllocator, &pPipelines[i]);
2819 if (result != VK_SUCCESS)
2820 final_result = result;
2821 }
2822
2823 return final_result;
2824 }
2825
2826 void
2827 tu_DestroyPipeline(VkDevice _device,
2828 VkPipeline _pipeline,
2829 const VkAllocationCallbacks *pAllocator)
2830 {
2831 TU_FROM_HANDLE(tu_device, dev, _device);
2832 TU_FROM_HANDLE(tu_pipeline, pipeline, _pipeline);
2833
2834 if (!_pipeline)
2835 return;
2836
2837 tu_pipeline_finish(pipeline, dev, pAllocator);
2838 vk_free2(&dev->alloc, pAllocator, pipeline);
2839 }