1/*
2 * Copyright 2012 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 *
23 * Authors:
24 *      Christian König <christian.koenig@amd.com>
25 *      Marek Olšák <maraeo@gmail.com>
26 */
27
28#include "si_pipe.h"
29#include "sid.h"
30#include "radeon/r600_cs.h"
31
32#include "tgsi/tgsi_parse.h"
33#include "tgsi/tgsi_ureg.h"
34#include "util/hash_table.h"
35#include "util/crc32.h"
36#include "util/u_memory.h"
37#include "util/u_prim.h"
38
39/* SHADER_CACHE */
40
41/**
42 * Return the TGSI binary in a buffer. The first 4 bytes contain its size as
43 * integer.
44 */
45static void *si_get_tgsi_binary(struct si_shader_selector *sel)
46{
47	unsigned tgsi_size = tgsi_num_tokens(sel->tokens) *
48			     sizeof(struct tgsi_token);
49	unsigned size = 4 + tgsi_size + sizeof(sel->so);
50	char *result = (char*)MALLOC(size);
51
52	if (!result)
53		return NULL;
54
55	*((uint32_t*)result) = size;
56	memcpy(result + 4, sel->tokens, tgsi_size);
57	memcpy(result + 4 + tgsi_size, &sel->so, sizeof(sel->so));
58	return result;
59}
60
61/** Copy "data" to "ptr" and return the next dword following copied data. */
62static uint32_t *write_data(uint32_t *ptr, const void *data, unsigned size)
63{
64	/* data may be NULL if size == 0 */
65	if (size)
66		memcpy(ptr, data, size);
67	ptr += DIV_ROUND_UP(size, 4);
68	return ptr;
69}
70
71/** Read data from "ptr". Return the next dword following the data. */
72static uint32_t *read_data(uint32_t *ptr, void *data, unsigned size)
73{
74	memcpy(data, ptr, size);
75	ptr += DIV_ROUND_UP(size, 4);
76	return ptr;
77}
78
79/**
80 * Write the size as uint followed by the data. Return the next dword
81 * following the copied data.
82 */
83static uint32_t *write_chunk(uint32_t *ptr, const void *data, unsigned size)
84{
85	*ptr++ = size;
86	return write_data(ptr, data, size);
87}
88
89/**
90 * Read the size as uint followed by the data. Return both via parameters.
91 * Return the next dword following the data.
92 */
93static uint32_t *read_chunk(uint32_t *ptr, void **data, unsigned *size)
94{
95	*size = *ptr++;
96	assert(*data == NULL);
97	if (!*size)
98		return ptr;
99	*data = malloc(*size);
100	return read_data(ptr, *data, *size);
101}
102
103/**
104 * Return the shader binary in a buffer. The first 4 bytes contain its size
105 * as integer.
106 */
107static void *si_get_shader_binary(struct si_shader *shader)
108{
109	/* There is always a size of data followed by the data itself. */
110	unsigned relocs_size = shader->binary.reloc_count *
111			       sizeof(shader->binary.relocs[0]);
112	unsigned disasm_size = strlen(shader->binary.disasm_string) + 1;
113	unsigned llvm_ir_size = shader->binary.llvm_ir_string ?
114				strlen(shader->binary.llvm_ir_string) + 1 : 0;
115	unsigned size =
116		4 + /* total size */
117		4 + /* CRC32 of the data below */
118		align(sizeof(shader->config), 4) +
119		align(sizeof(shader->info), 4) +
120		4 + align(shader->binary.code_size, 4) +
121		4 + align(shader->binary.rodata_size, 4) +
122		4 + align(relocs_size, 4) +
123		4 + align(disasm_size, 4) +
124		4 + align(llvm_ir_size, 4);
125	void *buffer = CALLOC(1, size);
126	uint32_t *ptr = (uint32_t*)buffer;
127
128	if (!buffer)
129		return NULL;
130
131	*ptr++ = size;
132	ptr++; /* CRC32 is calculated at the end. */
133
134	ptr = write_data(ptr, &shader->config, sizeof(shader->config));
135	ptr = write_data(ptr, &shader->info, sizeof(shader->info));
136	ptr = write_chunk(ptr, shader->binary.code, shader->binary.code_size);
137	ptr = write_chunk(ptr, shader->binary.rodata, shader->binary.rodata_size);
138	ptr = write_chunk(ptr, shader->binary.relocs, relocs_size);
139	ptr = write_chunk(ptr, shader->binary.disasm_string, disasm_size);
140	ptr = write_chunk(ptr, shader->binary.llvm_ir_string, llvm_ir_size);
141	assert((char *)ptr - (char *)buffer == size);
142
143	/* Compute CRC32. */
144	ptr = (uint32_t*)buffer;
145	ptr++;
146	*ptr = util_hash_crc32(ptr + 1, size - 8);
147
148	return buffer;
149}
150
151static bool si_load_shader_binary(struct si_shader *shader, void *binary)
152{
153	uint32_t *ptr = (uint32_t*)binary;
154	uint32_t size = *ptr++;
155	uint32_t crc32 = *ptr++;
156	unsigned chunk_size;
157
158	if (util_hash_crc32(ptr, size - 8) != crc32) {
159		fprintf(stderr, "radeonsi: binary shader has invalid CRC32\n");
160		return false;
161	}
162
163	ptr = read_data(ptr, &shader->config, sizeof(shader->config));
164	ptr = read_data(ptr, &shader->info, sizeof(shader->info));
165	ptr = read_chunk(ptr, (void**)&shader->binary.code,
166			 &shader->binary.code_size);
167	ptr = read_chunk(ptr, (void**)&shader->binary.rodata,
168			 &shader->binary.rodata_size);
169	ptr = read_chunk(ptr, (void**)&shader->binary.relocs, &chunk_size);
170	shader->binary.reloc_count = chunk_size / sizeof(shader->binary.relocs[0]);
171	ptr = read_chunk(ptr, (void**)&shader->binary.disasm_string, &chunk_size);
172	ptr = read_chunk(ptr, (void**)&shader->binary.llvm_ir_string, &chunk_size);
173
174	return true;
175}
176
177/**
178 * Insert a shader into the cache. It's assumed the shader is not in the cache.
179 * Use si_shader_cache_load_shader before calling this.
180 *
181 * Returns false on failure, in which case the tgsi_binary should be freed.
182 */
183static bool si_shader_cache_insert_shader(struct si_screen *sscreen,
184					  void *tgsi_binary,
185					  struct si_shader *shader)
186{
187	void *hw_binary;
188	struct hash_entry *entry;
189
190	entry = _mesa_hash_table_search(sscreen->shader_cache, tgsi_binary);
191	if (entry)
192		return false; /* already added */
193
194	hw_binary = si_get_shader_binary(shader);
195	if (!hw_binary)
196		return false;
197
198	if (_mesa_hash_table_insert(sscreen->shader_cache, tgsi_binary,
199				    hw_binary) == NULL) {
200		FREE(hw_binary);
201		return false;
202	}
203
204	return true;
205}
206
207static bool si_shader_cache_load_shader(struct si_screen *sscreen,
208					void *tgsi_binary,
209				        struct si_shader *shader)
210{
211	struct hash_entry *entry =
212		_mesa_hash_table_search(sscreen->shader_cache, tgsi_binary);
213	if (!entry)
214		return false;
215
216	if (!si_load_shader_binary(shader, entry->data))
217		return false;
218
219	p_atomic_inc(&sscreen->b.num_shader_cache_hits);
220	return true;
221}
222
223static uint32_t si_shader_cache_key_hash(const void *key)
224{
225	/* The first dword is the key size. */
226	return util_hash_crc32(key, *(uint32_t*)key);
227}
228
229static bool si_shader_cache_key_equals(const void *a, const void *b)
230{
231	uint32_t *keya = (uint32_t*)a;
232	uint32_t *keyb = (uint32_t*)b;
233
234	/* The first dword is the key size. */
235	if (*keya != *keyb)
236		return false;
237
238	return memcmp(keya, keyb, *keya) == 0;
239}
240
241static void si_destroy_shader_cache_entry(struct hash_entry *entry)
242{
243	FREE((void*)entry->key);
244	FREE(entry->data);
245}
246
247bool si_init_shader_cache(struct si_screen *sscreen)
248{
249	pipe_mutex_init(sscreen->shader_cache_mutex);
250	sscreen->shader_cache =
251		_mesa_hash_table_create(NULL,
252					si_shader_cache_key_hash,
253					si_shader_cache_key_equals);
254	return sscreen->shader_cache != NULL;
255}
256
257void si_destroy_shader_cache(struct si_screen *sscreen)
258{
259	if (sscreen->shader_cache)
260		_mesa_hash_table_destroy(sscreen->shader_cache,
261					 si_destroy_shader_cache_entry);
262	pipe_mutex_destroy(sscreen->shader_cache_mutex);
263}
264
265/* SHADER STATES */
266
267static void si_set_tesseval_regs(struct si_screen *sscreen,
268				 struct si_shader *shader,
269				 struct si_pm4_state *pm4)
270{
271	struct tgsi_shader_info *info = &shader->selector->info;
272	unsigned tes_prim_mode = info->properties[TGSI_PROPERTY_TES_PRIM_MODE];
273	unsigned tes_spacing = info->properties[TGSI_PROPERTY_TES_SPACING];
274	bool tes_vertex_order_cw = info->properties[TGSI_PROPERTY_TES_VERTEX_ORDER_CW];
275	bool tes_point_mode = info->properties[TGSI_PROPERTY_TES_POINT_MODE];
276	unsigned type, partitioning, topology, distribution_mode;
277
278	switch (tes_prim_mode) {
279	case PIPE_PRIM_LINES:
280		type = V_028B6C_TESS_ISOLINE;
281		break;
282	case PIPE_PRIM_TRIANGLES:
283		type = V_028B6C_TESS_TRIANGLE;
284		break;
285	case PIPE_PRIM_QUADS:
286		type = V_028B6C_TESS_QUAD;
287		break;
288	default:
289		assert(0);
290		return;
291	}
292
293	switch (tes_spacing) {
294	case PIPE_TESS_SPACING_FRACTIONAL_ODD:
295		partitioning = V_028B6C_PART_FRAC_ODD;
296		break;
297	case PIPE_TESS_SPACING_FRACTIONAL_EVEN:
298		partitioning = V_028B6C_PART_FRAC_EVEN;
299		break;
300	case PIPE_TESS_SPACING_EQUAL:
301		partitioning = V_028B6C_PART_INTEGER;
302		break;
303	default:
304		assert(0);
305		return;
306	}
307
308	if (tes_point_mode)
309		topology = V_028B6C_OUTPUT_POINT;
310	else if (tes_prim_mode == PIPE_PRIM_LINES)
311		topology = V_028B6C_OUTPUT_LINE;
312	else if (tes_vertex_order_cw)
313		/* for some reason, this must be the other way around */
314		topology = V_028B6C_OUTPUT_TRIANGLE_CCW;
315	else
316		topology = V_028B6C_OUTPUT_TRIANGLE_CW;
317
318	if (sscreen->has_distributed_tess) {
319		if (sscreen->b.family == CHIP_FIJI ||
320		    sscreen->b.family >= CHIP_POLARIS10)
321			distribution_mode = V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS;
322		else
323			distribution_mode = V_028B6C_DISTRIBUTION_MODE_DONUTS;
324	} else
325		distribution_mode = V_028B6C_DISTRIBUTION_MODE_NO_DIST;
326
327	si_pm4_set_reg(pm4, R_028B6C_VGT_TF_PARAM,
328		       S_028B6C_TYPE(type) |
329		       S_028B6C_PARTITIONING(partitioning) |
330		       S_028B6C_TOPOLOGY(topology) |
331		       S_028B6C_DISTRIBUTION_MODE(distribution_mode));
332}
333
334static struct si_pm4_state *si_get_shader_pm4_state(struct si_shader *shader)
335{
336	if (shader->pm4)
337		si_pm4_clear_state(shader->pm4);
338	else
339		shader->pm4 = CALLOC_STRUCT(si_pm4_state);
340
341	return shader->pm4;
342}
343
344static void si_shader_ls(struct si_shader *shader)
345{
346	struct si_pm4_state *pm4;
347	unsigned vgpr_comp_cnt;
348	uint64_t va;
349
350	pm4 = si_get_shader_pm4_state(shader);
351	if (!pm4)
352		return;
353
354	va = shader->bo->gpu_address;
355	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
356
357	/* We need at least 2 components for LS.
358	 * VGPR0-3: (VertexID, RelAutoindex, ???, InstanceID). */
359	vgpr_comp_cnt = shader->info.uses_instanceid ? 3 : 1;
360
361	si_pm4_set_reg(pm4, R_00B520_SPI_SHADER_PGM_LO_LS, va >> 8);
362	si_pm4_set_reg(pm4, R_00B524_SPI_SHADER_PGM_HI_LS, va >> 40);
363
364	shader->config.rsrc1 = S_00B528_VGPRS((shader->config.num_vgprs - 1) / 4) |
365			   S_00B528_SGPRS((shader->config.num_sgprs - 1) / 8) |
366		           S_00B528_VGPR_COMP_CNT(vgpr_comp_cnt) |
367			   S_00B528_DX10_CLAMP(1) |
368			   S_00B528_FLOAT_MODE(shader->config.float_mode);
369	shader->config.rsrc2 = S_00B52C_USER_SGPR(SI_LS_NUM_USER_SGPR) |
370			   S_00B52C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
371}
372
373static void si_shader_hs(struct si_shader *shader)
374{
375	struct si_pm4_state *pm4;
376	uint64_t va;
377
378	pm4 = si_get_shader_pm4_state(shader);
379	if (!pm4)
380		return;
381
382	va = shader->bo->gpu_address;
383	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
384
385	si_pm4_set_reg(pm4, R_00B420_SPI_SHADER_PGM_LO_HS, va >> 8);
386	si_pm4_set_reg(pm4, R_00B424_SPI_SHADER_PGM_HI_HS, va >> 40);
387	si_pm4_set_reg(pm4, R_00B428_SPI_SHADER_PGM_RSRC1_HS,
388		       S_00B428_VGPRS((shader->config.num_vgprs - 1) / 4) |
389		       S_00B428_SGPRS((shader->config.num_sgprs - 1) / 8) |
390		       S_00B428_DX10_CLAMP(1) |
391		       S_00B428_FLOAT_MODE(shader->config.float_mode));
392	si_pm4_set_reg(pm4, R_00B42C_SPI_SHADER_PGM_RSRC2_HS,
393		       S_00B42C_USER_SGPR(SI_TCS_NUM_USER_SGPR) |
394		       S_00B42C_OC_LDS_EN(1) |
395		       S_00B42C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
396}
397
398static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
399{
400	struct si_pm4_state *pm4;
401	unsigned num_user_sgprs;
402	unsigned vgpr_comp_cnt;
403	uint64_t va;
404	unsigned oc_lds_en;
405
406	pm4 = si_get_shader_pm4_state(shader);
407	if (!pm4)
408		return;
409
410	va = shader->bo->gpu_address;
411	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
412
413	if (shader->selector->type == PIPE_SHADER_VERTEX) {
414		vgpr_comp_cnt = shader->info.uses_instanceid ? 3 : 0;
415		num_user_sgprs = SI_ES_NUM_USER_SGPR;
416	} else if (shader->selector->type == PIPE_SHADER_TESS_EVAL) {
417		vgpr_comp_cnt = 3; /* all components are needed for TES */
418		num_user_sgprs = SI_TES_NUM_USER_SGPR;
419	} else
420		unreachable("invalid shader selector type");
421
422	oc_lds_en = shader->selector->type == PIPE_SHADER_TESS_EVAL ? 1 : 0;
423
424	si_pm4_set_reg(pm4, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
425		       shader->selector->esgs_itemsize / 4);
426	si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
427	si_pm4_set_reg(pm4, R_00B324_SPI_SHADER_PGM_HI_ES, va >> 40);
428	si_pm4_set_reg(pm4, R_00B328_SPI_SHADER_PGM_RSRC1_ES,
429		       S_00B328_VGPRS((shader->config.num_vgprs - 1) / 4) |
430		       S_00B328_SGPRS((shader->config.num_sgprs - 1) / 8) |
431		       S_00B328_VGPR_COMP_CNT(vgpr_comp_cnt) |
432		       S_00B328_DX10_CLAMP(1) |
433		       S_00B328_FLOAT_MODE(shader->config.float_mode));
434	si_pm4_set_reg(pm4, R_00B32C_SPI_SHADER_PGM_RSRC2_ES,
435		       S_00B32C_USER_SGPR(num_user_sgprs) |
436		       S_00B32C_OC_LDS_EN(oc_lds_en) |
437		       S_00B32C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
438
439	if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
440		si_set_tesseval_regs(sscreen, shader, pm4);
441}
442
443/**
444 * Calculate the appropriate setting of VGT_GS_MODE when \p shader is a
445 * geometry shader.
446 */
447static uint32_t si_vgt_gs_mode(struct si_shader_selector *sel)
448{
449	unsigned gs_max_vert_out = sel->gs_max_out_vertices;
450	unsigned cut_mode;
451
452	if (gs_max_vert_out <= 128) {
453		cut_mode = V_028A40_GS_CUT_128;
454	} else if (gs_max_vert_out <= 256) {
455		cut_mode = V_028A40_GS_CUT_256;
456	} else if (gs_max_vert_out <= 512) {
457		cut_mode = V_028A40_GS_CUT_512;
458	} else {
459		assert(gs_max_vert_out <= 1024);
460		cut_mode = V_028A40_GS_CUT_1024;
461	}
462
463	return S_028A40_MODE(V_028A40_GS_SCENARIO_G) |
464	       S_028A40_CUT_MODE(cut_mode)|
465	       S_028A40_ES_WRITE_OPTIMIZE(1) |
466	       S_028A40_GS_WRITE_OPTIMIZE(1);
467}
468
469static void si_shader_gs(struct si_shader *shader)
470{
471	struct si_shader_selector *sel = shader->selector;
472	const ubyte *num_components = sel->info.num_stream_output_components;
473	unsigned gs_num_invocations = sel->gs_num_invocations;
474	struct si_pm4_state *pm4;
475	uint64_t va;
476	unsigned max_stream = sel->max_gs_stream;
477	unsigned offset;
478
479	pm4 = si_get_shader_pm4_state(shader);
480	if (!pm4)
481		return;
482
483	si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE, si_vgt_gs_mode(shader->selector));
484
485	offset = num_components[0] * sel->gs_max_out_vertices;
486	si_pm4_set_reg(pm4, R_028A60_VGT_GSVS_RING_OFFSET_1, offset);
487	if (max_stream >= 1)
488		offset += num_components[1] * sel->gs_max_out_vertices;
489	si_pm4_set_reg(pm4, R_028A64_VGT_GSVS_RING_OFFSET_2, offset);
490	if (max_stream >= 2)
491		offset += num_components[2] * sel->gs_max_out_vertices;
492	si_pm4_set_reg(pm4, R_028A68_VGT_GSVS_RING_OFFSET_3, offset);
493	if (max_stream >= 3)
494		offset += num_components[3] * sel->gs_max_out_vertices;
495	si_pm4_set_reg(pm4, R_028AB0_VGT_GSVS_RING_ITEMSIZE, offset);
496
497	/* The GSVS_RING_ITEMSIZE register takes 15 bits */
498	assert(offset < (1 << 15));
499
500	si_pm4_set_reg(pm4, R_028B38_VGT_GS_MAX_VERT_OUT, shader->selector->gs_max_out_vertices);
501
502	si_pm4_set_reg(pm4, R_028B5C_VGT_GS_VERT_ITEMSIZE, num_components[0]);
503	si_pm4_set_reg(pm4, R_028B60_VGT_GS_VERT_ITEMSIZE_1, (max_stream >= 1) ? num_components[1] : 0);
504	si_pm4_set_reg(pm4, R_028B64_VGT_GS_VERT_ITEMSIZE_2, (max_stream >= 2) ? num_components[2] : 0);
505	si_pm4_set_reg(pm4, R_028B68_VGT_GS_VERT_ITEMSIZE_3, (max_stream >= 3) ? num_components[3] : 0);
506
507	si_pm4_set_reg(pm4, R_028B90_VGT_GS_INSTANCE_CNT,
508		       S_028B90_CNT(MIN2(gs_num_invocations, 127)) |
509		       S_028B90_ENABLE(gs_num_invocations > 0));
510
511	va = shader->bo->gpu_address;
512	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
513	si_pm4_set_reg(pm4, R_00B220_SPI_SHADER_PGM_LO_GS, va >> 8);
514	si_pm4_set_reg(pm4, R_00B224_SPI_SHADER_PGM_HI_GS, va >> 40);
515
516	si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
517		       S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
518		       S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8) |
519		       S_00B228_DX10_CLAMP(1) |
520		       S_00B228_FLOAT_MODE(shader->config.float_mode));
521	si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
522		       S_00B22C_USER_SGPR(SI_GS_NUM_USER_SGPR) |
523		       S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
524}
525
526/**
527 * Compute the state for \p shader, which will run as a vertex shader on the
528 * hardware.
529 *
530 * If \p gs is non-NULL, it points to the geometry shader for which this shader
531 * is the copy shader.
532 */
533static void si_shader_vs(struct si_screen *sscreen, struct si_shader *shader,
534                         struct si_shader_selector *gs)
535{
536	struct si_pm4_state *pm4;
537	unsigned num_user_sgprs;
538	unsigned nparams, vgpr_comp_cnt;
539	uint64_t va;
540	unsigned oc_lds_en;
541	unsigned window_space =
542	   shader->selector->info.properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
543	bool enable_prim_id = si_vs_exports_prim_id(shader);
544
545	pm4 = si_get_shader_pm4_state(shader);
546	if (!pm4)
547		return;
548
549	/* We always write VGT_GS_MODE in the VS state, because every switch
550	 * between different shader pipelines involving a different GS or no
551	 * GS at all involves a switch of the VS (different GS use different
552	 * copy shaders). On the other hand, when the API switches from a GS to
553	 * no GS and then back to the same GS used originally, the GS state is
554	 * not sent again.
555	 */
556	if (!gs) {
557		si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE,
558			       S_028A40_MODE(enable_prim_id ? V_028A40_GS_SCENARIO_A : 0));
559		si_pm4_set_reg(pm4, R_028A84_VGT_PRIMITIVEID_EN, enable_prim_id);
560	} else {
561		si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE, si_vgt_gs_mode(gs));
562		si_pm4_set_reg(pm4, R_028A84_VGT_PRIMITIVEID_EN, 0);
563	}
564
565	va = shader->bo->gpu_address;
566	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
567
568	if (gs) {
569		vgpr_comp_cnt = 0; /* only VertexID is needed for GS-COPY. */
570		num_user_sgprs = SI_GSCOPY_NUM_USER_SGPR;
571	} else if (shader->selector->type == PIPE_SHADER_VERTEX) {
572		vgpr_comp_cnt = shader->info.uses_instanceid ? 3 : (enable_prim_id ? 2 : 0);
573		num_user_sgprs = SI_VS_NUM_USER_SGPR;
574	} else if (shader->selector->type == PIPE_SHADER_TESS_EVAL) {
575		vgpr_comp_cnt = 3; /* all components are needed for TES */
576		num_user_sgprs = SI_TES_NUM_USER_SGPR;
577	} else
578		unreachable("invalid shader selector type");
579
580	/* VS is required to export at least one param. */
581	nparams = MAX2(shader->info.nr_param_exports, 1);
582	si_pm4_set_reg(pm4, R_0286C4_SPI_VS_OUT_CONFIG,
583		       S_0286C4_VS_EXPORT_COUNT(nparams - 1));
584
585	si_pm4_set_reg(pm4, R_02870C_SPI_SHADER_POS_FORMAT,
586		       S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
587		       S_02870C_POS1_EXPORT_FORMAT(shader->info.nr_pos_exports > 1 ?
588						   V_02870C_SPI_SHADER_4COMP :
589						   V_02870C_SPI_SHADER_NONE) |
590		       S_02870C_POS2_EXPORT_FORMAT(shader->info.nr_pos_exports > 2 ?
591						   V_02870C_SPI_SHADER_4COMP :
592						   V_02870C_SPI_SHADER_NONE) |
593		       S_02870C_POS3_EXPORT_FORMAT(shader->info.nr_pos_exports > 3 ?
594						   V_02870C_SPI_SHADER_4COMP :
595						   V_02870C_SPI_SHADER_NONE));
596
597	oc_lds_en = shader->selector->type == PIPE_SHADER_TESS_EVAL ? 1 : 0;
598
599	si_pm4_set_reg(pm4, R_00B120_SPI_SHADER_PGM_LO_VS, va >> 8);
600	si_pm4_set_reg(pm4, R_00B124_SPI_SHADER_PGM_HI_VS, va >> 40);
601	si_pm4_set_reg(pm4, R_00B128_SPI_SHADER_PGM_RSRC1_VS,
602		       S_00B128_VGPRS((shader->config.num_vgprs - 1) / 4) |
603		       S_00B128_SGPRS((shader->config.num_sgprs - 1) / 8) |
604		       S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt) |
605		       S_00B128_DX10_CLAMP(1) |
606		       S_00B128_FLOAT_MODE(shader->config.float_mode));
607	si_pm4_set_reg(pm4, R_00B12C_SPI_SHADER_PGM_RSRC2_VS,
608		       S_00B12C_USER_SGPR(num_user_sgprs) |
609		       S_00B12C_OC_LDS_EN(oc_lds_en) |
610		       S_00B12C_SO_BASE0_EN(!!shader->selector->so.stride[0]) |
611		       S_00B12C_SO_BASE1_EN(!!shader->selector->so.stride[1]) |
612		       S_00B12C_SO_BASE2_EN(!!shader->selector->so.stride[2]) |
613		       S_00B12C_SO_BASE3_EN(!!shader->selector->so.stride[3]) |
614		       S_00B12C_SO_EN(!!shader->selector->so.num_outputs) |
615		       S_00B12C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
616	if (window_space)
617		si_pm4_set_reg(pm4, R_028818_PA_CL_VTE_CNTL,
618			       S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1));
619	else
620		si_pm4_set_reg(pm4, R_028818_PA_CL_VTE_CNTL,
621			       S_028818_VTX_W0_FMT(1) |
622			       S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
623			       S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
624			       S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
625
626	if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
627		si_set_tesseval_regs(sscreen, shader, pm4);
628}
629
630static unsigned si_get_ps_num_interp(struct si_shader *ps)
631{
632	struct tgsi_shader_info *info = &ps->selector->info;
633	unsigned num_colors = !!(info->colors_read & 0x0f) +
634			      !!(info->colors_read & 0xf0);
635	unsigned num_interp = ps->selector->info.num_inputs +
636			      (ps->key.part.ps.prolog.color_two_side ? num_colors : 0);
637
638	assert(num_interp <= 32);
639	return MIN2(num_interp, 32);
640}
641
642static unsigned si_get_spi_shader_col_format(struct si_shader *shader)
643{
644	unsigned value = shader->key.part.ps.epilog.spi_shader_col_format;
645	unsigned i, num_targets = (util_last_bit(value) + 3) / 4;
646
647	/* If the i-th target format is set, all previous target formats must
648	 * be non-zero to avoid hangs.
649	 */
650	for (i = 0; i < num_targets; i++)
651		if (!(value & (0xf << (i * 4))))
652			value |= V_028714_SPI_SHADER_32_R << (i * 4);
653
654	return value;
655}
656
657static unsigned si_get_cb_shader_mask(unsigned spi_shader_col_format)
658{
659	unsigned i, cb_shader_mask = 0;
660
661	for (i = 0; i < 8; i++) {
662		switch ((spi_shader_col_format >> (i * 4)) & 0xf) {
663		case V_028714_SPI_SHADER_ZERO:
664			break;
665		case V_028714_SPI_SHADER_32_R:
666			cb_shader_mask |= 0x1 << (i * 4);
667			break;
668		case V_028714_SPI_SHADER_32_GR:
669			cb_shader_mask |= 0x3 << (i * 4);
670			break;
671		case V_028714_SPI_SHADER_32_AR:
672			cb_shader_mask |= 0x9 << (i * 4);
673			break;
674		case V_028714_SPI_SHADER_FP16_ABGR:
675		case V_028714_SPI_SHADER_UNORM16_ABGR:
676		case V_028714_SPI_SHADER_SNORM16_ABGR:
677		case V_028714_SPI_SHADER_UINT16_ABGR:
678		case V_028714_SPI_SHADER_SINT16_ABGR:
679		case V_028714_SPI_SHADER_32_ABGR:
680			cb_shader_mask |= 0xf << (i * 4);
681			break;
682		default:
683			assert(0);
684		}
685	}
686	return cb_shader_mask;
687}
688
689static void si_shader_ps(struct si_shader *shader)
690{
691	struct tgsi_shader_info *info = &shader->selector->info;
692	struct si_pm4_state *pm4;
693	unsigned spi_ps_in_control, spi_shader_col_format, cb_shader_mask;
694	unsigned spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
695	uint64_t va;
696	unsigned input_ena = shader->config.spi_ps_input_ena;
697
698	/* we need to enable at least one of them, otherwise we hang the GPU */
699	assert(G_0286CC_PERSP_SAMPLE_ENA(input_ena) ||
700	       G_0286CC_PERSP_CENTER_ENA(input_ena) ||
701	       G_0286CC_PERSP_CENTROID_ENA(input_ena) ||
702	       G_0286CC_PERSP_PULL_MODEL_ENA(input_ena) ||
703	       G_0286CC_LINEAR_SAMPLE_ENA(input_ena) ||
704	       G_0286CC_LINEAR_CENTER_ENA(input_ena) ||
705	       G_0286CC_LINEAR_CENTROID_ENA(input_ena) ||
706	       G_0286CC_LINE_STIPPLE_TEX_ENA(input_ena));
707	/* POS_W_FLOAT_ENA requires one of the perspective weights. */
708	assert(!G_0286CC_POS_W_FLOAT_ENA(input_ena) ||
709	       G_0286CC_PERSP_SAMPLE_ENA(input_ena) ||
710	       G_0286CC_PERSP_CENTER_ENA(input_ena) ||
711	       G_0286CC_PERSP_CENTROID_ENA(input_ena) ||
712	       G_0286CC_PERSP_PULL_MODEL_ENA(input_ena));
713
714	/* Validate interpolation optimization flags (read as implications). */
715	assert(!shader->key.part.ps.prolog.bc_optimize_for_persp ||
716	       (G_0286CC_PERSP_CENTER_ENA(input_ena) &&
717		G_0286CC_PERSP_CENTROID_ENA(input_ena)));
718	assert(!shader->key.part.ps.prolog.bc_optimize_for_linear ||
719	       (G_0286CC_LINEAR_CENTER_ENA(input_ena) &&
720		G_0286CC_LINEAR_CENTROID_ENA(input_ena)));
721	assert(!shader->key.part.ps.prolog.force_persp_center_interp ||
722	       (!G_0286CC_PERSP_SAMPLE_ENA(input_ena) &&
723		!G_0286CC_PERSP_CENTROID_ENA(input_ena)));
724	assert(!shader->key.part.ps.prolog.force_linear_center_interp ||
725	       (!G_0286CC_LINEAR_SAMPLE_ENA(input_ena) &&
726		!G_0286CC_LINEAR_CENTROID_ENA(input_ena)));
727	assert(!shader->key.part.ps.prolog.force_persp_sample_interp ||
728	       (!G_0286CC_PERSP_CENTER_ENA(input_ena) &&
729		!G_0286CC_PERSP_CENTROID_ENA(input_ena)));
730	assert(!shader->key.part.ps.prolog.force_linear_sample_interp ||
731	       (!G_0286CC_LINEAR_CENTER_ENA(input_ena) &&
732		!G_0286CC_LINEAR_CENTROID_ENA(input_ena)));
733
734	/* Validate cases when the optimizations are off (read as implications). */
735	assert(shader->key.part.ps.prolog.bc_optimize_for_persp ||
736	       !G_0286CC_PERSP_CENTER_ENA(input_ena) ||
737	       !G_0286CC_PERSP_CENTROID_ENA(input_ena));
738	assert(shader->key.part.ps.prolog.bc_optimize_for_linear ||
739	       !G_0286CC_LINEAR_CENTER_ENA(input_ena) ||
740	       !G_0286CC_LINEAR_CENTROID_ENA(input_ena));
741
742	pm4 = si_get_shader_pm4_state(shader);
743	if (!pm4)
744		return;
745
746	/* SPI_BARYC_CNTL.POS_FLOAT_LOCATION
747	 * Possible vaules:
748	 * 0 -> Position = pixel center
749	 * 1 -> Position = pixel centroid
750	 * 2 -> Position = at sample position
751	 *
752	 * From GLSL 4.5 specification, section 7.1:
753	 *   "The variable gl_FragCoord is available as an input variable from
754	 *    within fragment shaders and it holds the window relative coordinates
755	 *    (x, y, z, 1/w) values for the fragment. If multi-sampling, this
756	 *    value can be for any location within the pixel, or one of the
757	 *    fragment samples. The use of centroid does not further restrict
758	 *    this value to be inside the current primitive."
759	 *
760	 * Meaning that centroid has no effect and we can return anything within
761	 * the pixel. Thus, return the value at sample position, because that's
762	 * the most accurate one shaders can get.
763	 */
764	spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2);
765
766	if (info->properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] ==
767	    TGSI_FS_COORD_PIXEL_CENTER_INTEGER)
768		spi_baryc_cntl |= S_0286E0_POS_FLOAT_ULC(1);
769
770	spi_shader_col_format = si_get_spi_shader_col_format(shader);
771	cb_shader_mask = si_get_cb_shader_mask(spi_shader_col_format);
772
773	/* Ensure that some export memory is always allocated, for two reasons:
774	 *
775	 * 1) Correctness: The hardware ignores the EXEC mask if no export
776	 *    memory is allocated, so KILL and alpha test do not work correctly
777	 *    without this.
778	 * 2) Performance: Every shader needs at least a NULL export, even when
779	 *    it writes no color/depth output. The NULL export instruction
780	 *    stalls without this setting.
781	 *
782	 * Don't add this to CB_SHADER_MASK.
783	 */
784	if (!spi_shader_col_format &&
785	    !info->writes_z && !info->writes_stencil && !info->writes_samplemask)
786		spi_shader_col_format = V_028714_SPI_SHADER_32_R;
787
788	si_pm4_set_reg(pm4, R_0286CC_SPI_PS_INPUT_ENA, input_ena);
789	si_pm4_set_reg(pm4, R_0286D0_SPI_PS_INPUT_ADDR,
790		       shader->config.spi_ps_input_addr);
791
792	/* Set interpolation controls. */
793	spi_ps_in_control = S_0286D8_NUM_INTERP(si_get_ps_num_interp(shader));
794
795	/* Set registers. */
796	si_pm4_set_reg(pm4, R_0286E0_SPI_BARYC_CNTL, spi_baryc_cntl);
797	si_pm4_set_reg(pm4, R_0286D8_SPI_PS_IN_CONTROL, spi_ps_in_control);
798
799	si_pm4_set_reg(pm4, R_028710_SPI_SHADER_Z_FORMAT,
800		       si_get_spi_shader_z_format(info->writes_z,
801						  info->writes_stencil,
802						  info->writes_samplemask));
803
804	si_pm4_set_reg(pm4, R_028714_SPI_SHADER_COL_FORMAT, spi_shader_col_format);
805	si_pm4_set_reg(pm4, R_02823C_CB_SHADER_MASK, cb_shader_mask);
806
807	va = shader->bo->gpu_address;
808	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
809	si_pm4_set_reg(pm4, R_00B020_SPI_SHADER_PGM_LO_PS, va >> 8);
810	si_pm4_set_reg(pm4, R_00B024_SPI_SHADER_PGM_HI_PS, va >> 40);
811
812	si_pm4_set_reg(pm4, R_00B028_SPI_SHADER_PGM_RSRC1_PS,
813		       S_00B028_VGPRS((shader->config.num_vgprs - 1) / 4) |
814		       S_00B028_SGPRS((shader->config.num_sgprs - 1) / 8) |
815		       S_00B028_DX10_CLAMP(1) |
816		       S_00B028_FLOAT_MODE(shader->config.float_mode));
817	si_pm4_set_reg(pm4, R_00B02C_SPI_SHADER_PGM_RSRC2_PS,
818		       S_00B02C_EXTRA_LDS_SIZE(shader->config.lds_size) |
819		       S_00B02C_USER_SGPR(SI_PS_NUM_USER_SGPR) |
820		       S_00B32C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
821}
822
823static void si_shader_init_pm4_state(struct si_screen *sscreen,
824                                     struct si_shader *shader)
825{
826	switch (shader->selector->type) {
827	case PIPE_SHADER_VERTEX:
828		if (shader->key.as_ls)
829			si_shader_ls(shader);
830		else if (shader->key.as_es)
831			si_shader_es(sscreen, shader);
832		else
833			si_shader_vs(sscreen, shader, NULL);
834		break;
835	case PIPE_SHADER_TESS_CTRL:
836		si_shader_hs(shader);
837		break;
838	case PIPE_SHADER_TESS_EVAL:
839		if (shader->key.as_es)
840			si_shader_es(sscreen, shader);
841		else
842			si_shader_vs(sscreen, shader, NULL);
843		break;
844	case PIPE_SHADER_GEOMETRY:
845		si_shader_gs(shader);
846		break;
847	case PIPE_SHADER_FRAGMENT:
848		si_shader_ps(shader);
849		break;
850	default:
851		assert(0);
852	}
853}
854
855static unsigned si_get_alpha_test_func(struct si_context *sctx)
856{
857	/* Alpha-test should be disabled if colorbuffer 0 is integer. */
858	if (sctx->queued.named.dsa)
859		return sctx->queued.named.dsa->alpha_func;
860
861	return PIPE_FUNC_ALWAYS;
862}
863
864static void si_shader_selector_key_hw_vs(struct si_context *sctx,
865					 struct si_shader_selector *vs,
866					 struct si_shader_key *key)
867{
868	struct si_shader_selector *ps = sctx->ps_shader.cso;
869
870	key->opt.hw_vs.clip_disable =
871		sctx->queued.named.rasterizer->clip_plane_enable == 0 &&
872		(vs->info.clipdist_writemask ||
873		 vs->info.writes_clipvertex) &&
874		!vs->info.culldist_writemask;
875
876	/* Find out if PS is disabled. */
877	bool ps_disabled = true;
878	if (ps) {
879		bool ps_modifies_zs = ps->info.uses_kill ||
880				      ps->info.writes_z ||
881				      ps->info.writes_stencil ||
882				      ps->info.writes_samplemask ||
883				      si_get_alpha_test_func(sctx) != PIPE_FUNC_ALWAYS;
884
885		unsigned ps_colormask = sctx->framebuffer.colorbuf_enabled_4bit &
886					sctx->queued.named.blend->cb_target_mask;
887		if (!ps->info.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS])
888			ps_colormask &= ps->colors_written_4bit;
889
890		ps_disabled = sctx->queued.named.rasterizer->rasterizer_discard ||
891			      (!ps_colormask &&
892			       !ps_modifies_zs &&
893			       !ps->info.writes_memory);
894	}
895
896	/* Find out which VS outputs aren't used by the PS. */
897	uint64_t outputs_written = vs->outputs_written;
898	uint32_t outputs_written2 = vs->outputs_written2;
899	uint64_t inputs_read = 0;
900	uint32_t inputs_read2 = 0;
901
902	outputs_written &= ~0x3; /* ignore POSITION, PSIZE */
903
904	if (!ps_disabled) {
905		inputs_read = ps->inputs_read;
906		inputs_read2 = ps->inputs_read2;
907	}
908
909	uint64_t linked = outputs_written & inputs_read;
910	uint32_t linked2 = outputs_written2 & inputs_read2;
911
912	key->opt.hw_vs.kill_outputs = ~linked & outputs_written;
913	key->opt.hw_vs.kill_outputs2 = ~linked2 & outputs_written2;
914}
915
916/* Compute the key for the hw shader variant */
917static inline void si_shader_selector_key(struct pipe_context *ctx,
918					  struct si_shader_selector *sel,
919					  struct si_shader_key *key)
920{
921	struct si_context *sctx = (struct si_context *)ctx;
922	unsigned i;
923
924	memset(key, 0, sizeof(*key));
925
926	switch (sel->type) {
927	case PIPE_SHADER_VERTEX:
928		if (sctx->vertex_elements) {
929			unsigned count = MIN2(sel->info.num_inputs,
930					      sctx->vertex_elements->count);
931			for (i = 0; i < count; ++i)
932				key->part.vs.prolog.instance_divisors[i] =
933					sctx->vertex_elements->elements[i].instance_divisor;
934
935			key->mono.vs.fix_fetch =
936				sctx->vertex_elements->fix_fetch &
937				u_bit_consecutive64(0, 4 * count);
938		}
939		if (sctx->tes_shader.cso)
940			key->as_ls = 1;
941		else if (sctx->gs_shader.cso)
942			key->as_es = 1;
943		else {
944			si_shader_selector_key_hw_vs(sctx, sel, key);
945
946			if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
947				key->part.vs.epilog.export_prim_id = 1;
948		}
949		break;
950	case PIPE_SHADER_TESS_CTRL:
951		key->part.tcs.epilog.prim_mode =
952			sctx->tes_shader.cso->info.properties[TGSI_PROPERTY_TES_PRIM_MODE];
953
954		if (sel == sctx->fixed_func_tcs_shader.cso)
955			key->mono.tcs.inputs_to_copy = sctx->vs_shader.cso->outputs_written;
956		break;
957	case PIPE_SHADER_TESS_EVAL:
958		if (sctx->gs_shader.cso)
959			key->as_es = 1;
960		else {
961			si_shader_selector_key_hw_vs(sctx, sel, key);
962
963			if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
964				key->part.tes.epilog.export_prim_id = 1;
965		}
966		break;
967	case PIPE_SHADER_GEOMETRY:
968		key->part.gs.prolog.tri_strip_adj_fix = sctx->gs_tri_strip_adj_fix;
969		break;
970	case PIPE_SHADER_FRAGMENT: {
971		struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
972		struct si_state_blend *blend = sctx->queued.named.blend;
973
974		if (sel->info.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
975		    sel->info.colors_written == 0x1)
976			key->part.ps.epilog.last_cbuf = MAX2(sctx->framebuffer.state.nr_cbufs, 1) - 1;
977
978		if (blend) {
979			/* Select the shader color format based on whether
980			 * blending or alpha are needed.
981			 */
982			key->part.ps.epilog.spi_shader_col_format =
983				(blend->blend_enable_4bit & blend->need_src_alpha_4bit &
984				 sctx->framebuffer.spi_shader_col_format_blend_alpha) |
985				(blend->blend_enable_4bit & ~blend->need_src_alpha_4bit &
986				 sctx->framebuffer.spi_shader_col_format_blend) |
987				(~blend->blend_enable_4bit & blend->need_src_alpha_4bit &
988				 sctx->framebuffer.spi_shader_col_format_alpha) |
989				(~blend->blend_enable_4bit & ~blend->need_src_alpha_4bit &
990				 sctx->framebuffer.spi_shader_col_format);
991
992			/* The output for dual source blending should have
993			 * the same format as the first output.
994			 */
995			if (blend->dual_src_blend)
996				key->part.ps.epilog.spi_shader_col_format |=
997					(key->part.ps.epilog.spi_shader_col_format & 0xf) << 4;
998		} else
999			key->part.ps.epilog.spi_shader_col_format = sctx->framebuffer.spi_shader_col_format;
1000
1001		/* If alpha-to-coverage is enabled, we have to export alpha
1002		 * even if there is no color buffer.
1003		 */
1004		if (!(key->part.ps.epilog.spi_shader_col_format & 0xf) &&
1005		    blend && blend->alpha_to_coverage)
1006			key->part.ps.epilog.spi_shader_col_format |= V_028710_SPI_SHADER_32_AR;
1007
1008		/* On SI and CIK except Hawaii, the CB doesn't clamp outputs
1009		 * to the range supported by the type if a channel has less
1010		 * than 16 bits and the export format is 16_ABGR.
1011		 */
1012		if (sctx->b.chip_class <= CIK && sctx->b.family != CHIP_HAWAII) {
1013			key->part.ps.epilog.color_is_int8 = sctx->framebuffer.color_is_int8;
1014			key->part.ps.epilog.color_is_int10 = sctx->framebuffer.color_is_int10;
1015		}
1016
1017		/* Disable unwritten outputs (if WRITE_ALL_CBUFS isn't enabled). */
1018		if (!key->part.ps.epilog.last_cbuf) {
1019			key->part.ps.epilog.spi_shader_col_format &= sel->colors_written_4bit;
1020			key->part.ps.epilog.color_is_int8 &= sel->info.colors_written;
1021			key->part.ps.epilog.color_is_int10 &= sel->info.colors_written;
1022		}
1023
1024		if (rs) {
1025			bool is_poly = (sctx->current_rast_prim >= PIPE_PRIM_TRIANGLES &&
1026					sctx->current_rast_prim <= PIPE_PRIM_POLYGON) ||
1027				       sctx->current_rast_prim >= PIPE_PRIM_TRIANGLES_ADJACENCY;
1028			bool is_line = !is_poly && sctx->current_rast_prim != PIPE_PRIM_POINTS;
1029
1030			key->part.ps.prolog.color_two_side = rs->two_side && sel->info.colors_read;
1031			key->part.ps.prolog.flatshade_colors = rs->flatshade && sel->info.colors_read;
1032
1033			if (sctx->queued.named.blend) {
1034				key->part.ps.epilog.alpha_to_one = sctx->queued.named.blend->alpha_to_one &&
1035							      rs->multisample_enable;
1036			}
1037
1038			key->part.ps.prolog.poly_stipple = rs->poly_stipple_enable && is_poly;
1039			key->part.ps.epilog.poly_line_smoothing = ((is_poly && rs->poly_smooth) ||
1040							      (is_line && rs->line_smooth)) &&
1041							     sctx->framebuffer.nr_samples <= 1;
1042			key->part.ps.epilog.clamp_color = rs->clamp_fragment_color;
1043
1044			if (rs->force_persample_interp &&
1045			    rs->multisample_enable &&
1046			    sctx->framebuffer.nr_samples > 1 &&
1047			    sctx->ps_iter_samples > 1) {
1048				key->part.ps.prolog.force_persp_sample_interp =
1049					sel->info.uses_persp_center ||
1050					sel->info.uses_persp_centroid;
1051
1052				key->part.ps.prolog.force_linear_sample_interp =
1053					sel->info.uses_linear_center ||
1054					sel->info.uses_linear_centroid;
1055			} else if (rs->multisample_enable &&
1056				   sctx->framebuffer.nr_samples > 1) {
1057				key->part.ps.prolog.bc_optimize_for_persp =
1058					sel->info.uses_persp_center &&
1059					sel->info.uses_persp_centroid;
1060				key->part.ps.prolog.bc_optimize_for_linear =
1061					sel->info.uses_linear_center &&
1062					sel->info.uses_linear_centroid;
1063			} else {
1064				/* Make sure SPI doesn't compute more than 1 pair
1065				 * of (i,j), which is the optimization here. */
1066				key->part.ps.prolog.force_persp_center_interp =
1067					sel->info.uses_persp_center +
1068					sel->info.uses_persp_centroid +
1069					sel->info.uses_persp_sample > 1;
1070
1071				key->part.ps.prolog.force_linear_center_interp =
1072					sel->info.uses_linear_center +
1073					sel->info.uses_linear_centroid +
1074					sel->info.uses_linear_sample > 1;
1075			}
1076		}
1077
1078		key->part.ps.epilog.alpha_func = si_get_alpha_test_func(sctx);
1079		break;
1080	}
1081	default:
1082		assert(0);
1083	}
1084}
1085
1086static void si_build_shader_variant(void *job, int thread_index)
1087{
1088	struct si_shader *shader = (struct si_shader *)job;
1089	struct si_shader_selector *sel = shader->selector;
1090	struct si_screen *sscreen = sel->screen;
1091	LLVMTargetMachineRef tm;
1092	struct pipe_debug_callback *debug = &shader->compiler_ctx_state.debug;
1093	int r;
1094
1095	if (thread_index >= 0) {
1096		assert(thread_index < ARRAY_SIZE(sscreen->tm));
1097		tm = sscreen->tm[thread_index];
1098		if (!debug->async)
1099			debug = NULL;
1100	} else {
1101		tm = shader->compiler_ctx_state.tm;
1102	}
1103
1104	r = si_shader_create(sscreen, tm, shader, debug);
1105	if (unlikely(r)) {
1106		R600_ERR("Failed to build shader variant (type=%u) %d\n",
1107			 sel->type, r);
1108		shader->compilation_failed = true;
1109		return;
1110	}
1111
1112	if (shader->compiler_ctx_state.is_debug_context) {
1113		FILE *f = open_memstream(&shader->shader_log,
1114					 &shader->shader_log_size);
1115		if (f) {
1116			si_shader_dump(sscreen, shader, NULL, sel->type, f, false);
1117			fclose(f);
1118		}
1119	}
1120
1121	si_shader_init_pm4_state(sscreen, shader);
1122}
1123
1124/* Select the hw shader variant depending on the current state. */
1125static int si_shader_select_with_key(struct si_screen *sscreen,
1126				     struct si_shader_ctx_state *state,
1127				     struct si_compiler_ctx_state *compiler_state,
1128				     struct si_shader_key *key,
1129				     int thread_index)
1130{
1131	static const struct si_shader_key zeroed;
1132	struct si_shader_selector *sel = state->cso;
1133	struct si_shader *current = state->current;
1134	struct si_shader *iter, *shader = NULL;
1135
1136	if (unlikely(sscreen->b.debug_flags & DBG_NO_OPT_VARIANT)) {
1137		memset(&key->opt, 0, sizeof(key->opt));
1138	}
1139
1140again:
1141	/* Check if we don't need to change anything.
1142	 * This path is also used for most shaders that don't need multiple
1143	 * variants, it will cost just a computation of the key and this
1144	 * test. */
1145	if (likely(current &&
1146		   memcmp(&current->key, key, sizeof(*key)) == 0 &&
1147		   (!current->is_optimized ||
1148		    util_queue_fence_is_signalled(&current->optimized_ready))))
1149		return current->compilation_failed ? -1 : 0;
1150
1151	/* This must be done before the mutex is locked, because async GS
1152	 * compilation calls this function too, and therefore must enter
1153	 * the mutex first.
1154	 *
1155	 * Only wait if we are in a draw call. Don't wait if we are
1156	 * in a compiler thread.
1157	 */
1158	if (thread_index < 0)
1159		util_queue_job_wait(&sel->ready);
1160
1161	pipe_mutex_lock(sel->mutex);
1162
1163	/* Find the shader variant. */
1164	for (iter = sel->first_variant; iter; iter = iter->next_variant) {
1165		/* Don't check the "current" shader. We checked it above. */
1166		if (current != iter &&
1167		    memcmp(&iter->key, key, sizeof(*key)) == 0) {
1168			/* If it's an optimized shader and its compilation has
1169			 * been started but isn't done, use the unoptimized
1170			 * shader so as not to cause a stall due to compilation.
1171			 */
1172			if (iter->is_optimized &&
1173			    !util_queue_fence_is_signalled(&iter->optimized_ready)) {
1174				memset(&key->opt, 0, sizeof(key->opt));
1175				pipe_mutex_unlock(sel->mutex);
1176				goto again;
1177			}
1178
1179			if (iter->compilation_failed) {
1180				pipe_mutex_unlock(sel->mutex);
1181				return -1; /* skip the draw call */
1182			}
1183
1184			state->current = iter;
1185			pipe_mutex_unlock(sel->mutex);
1186			return 0;
1187		}
1188	}
1189
1190	/* Build a new shader. */
1191	shader = CALLOC_STRUCT(si_shader);
1192	if (!shader) {
1193		pipe_mutex_unlock(sel->mutex);
1194		return -ENOMEM;
1195	}
1196	shader->selector = sel;
1197	shader->key = *key;
1198	shader->compiler_ctx_state = *compiler_state;
1199
1200	/* Monolithic-only shaders don't make a distinction between optimized
1201	 * and unoptimized. */
1202	shader->is_monolithic =
1203		!sel->main_shader_part ||
1204		sel->main_shader_part->key.as_ls != key->as_ls ||
1205		sel->main_shader_part->key.as_es != key->as_es ||
1206		memcmp(&key->opt, &zeroed.opt, sizeof(key->opt)) != 0 ||
1207		memcmp(&key->mono, &zeroed.mono, sizeof(key->mono)) != 0;
1208
1209	shader->is_optimized =
1210		!sscreen->use_monolithic_shaders &&
1211		memcmp(&key->opt, &zeroed.opt, sizeof(key->opt)) != 0;
1212	if (shader->is_optimized)
1213		util_queue_fence_init(&shader->optimized_ready);
1214
1215	if (!sel->last_variant) {
1216		sel->first_variant = shader;
1217		sel->last_variant = shader;
1218	} else {
1219		sel->last_variant->next_variant = shader;
1220		sel->last_variant = shader;
1221	}
1222
1223	/* If it's an optimized shader, compile it asynchronously. */
1224	if (shader->is_optimized &&
1225	    thread_index < 0) {
1226		/* Compile it asynchronously. */
1227		util_queue_add_job(&sscreen->shader_compiler_queue,
1228				   shader, &shader->optimized_ready,
1229				   si_build_shader_variant, NULL);
1230
1231		/* Use the default (unoptimized) shader for now. */
1232		memset(&key->opt, 0, sizeof(key->opt));
1233		pipe_mutex_unlock(sel->mutex);
1234		goto again;
1235	}
1236
1237	assert(!shader->is_optimized);
1238	si_build_shader_variant(shader, thread_index);
1239
1240	if (!shader->compilation_failed)
1241		state->current = shader;
1242
1243	pipe_mutex_unlock(sel->mutex);
1244	return shader->compilation_failed ? -1 : 0;
1245}
1246
1247static int si_shader_select(struct pipe_context *ctx,
1248			    struct si_shader_ctx_state *state,
1249			    struct si_compiler_ctx_state *compiler_state)
1250{
1251	struct si_context *sctx = (struct si_context *)ctx;
1252	struct si_shader_key key;
1253
1254	si_shader_selector_key(ctx, state->cso, &key);
1255	return si_shader_select_with_key(sctx->screen, state, compiler_state,
1256					 &key, -1);
1257}
1258
1259static void si_parse_next_shader_property(const struct tgsi_shader_info *info,
1260					  struct si_shader_key *key)
1261{
1262	unsigned next_shader = info->properties[TGSI_PROPERTY_NEXT_SHADER];
1263
1264	switch (info->processor) {
1265	case PIPE_SHADER_VERTEX:
1266		switch (next_shader) {
1267		case PIPE_SHADER_GEOMETRY:
1268			key->as_es = 1;
1269			break;
1270		case PIPE_SHADER_TESS_CTRL:
1271		case PIPE_SHADER_TESS_EVAL:
1272			key->as_ls = 1;
1273			break;
1274		default:
1275			/* If POSITION isn't written, it can't be a HW VS.
1276			 * Assume that it's a HW LS. (the next shader is TCS)
1277			 * This heuristic is needed for separate shader objects.
1278			 */
1279			if (!info->writes_position)
1280				key->as_ls = 1;
1281		}
1282		break;
1283
1284	case PIPE_SHADER_TESS_EVAL:
1285		if (next_shader == PIPE_SHADER_GEOMETRY ||
1286		    !info->writes_position)
1287			key->as_es = 1;
1288		break;
1289	}
1290}
1291
1292/**
1293 * Compile the main shader part or the monolithic shader as part of
1294 * si_shader_selector initialization. Since it can be done asynchronously,
1295 * there is no way to report compile failures to applications.
1296 */
1297void si_init_shader_selector_async(void *job, int thread_index)
1298{
1299	struct si_shader_selector *sel = (struct si_shader_selector *)job;
1300	struct si_screen *sscreen = sel->screen;
1301	LLVMTargetMachineRef tm;
1302	struct pipe_debug_callback *debug = &sel->compiler_ctx_state.debug;
1303	unsigned i;
1304
1305	if (thread_index >= 0) {
1306		assert(thread_index < ARRAY_SIZE(sscreen->tm));
1307		tm = sscreen->tm[thread_index];
1308		if (!debug->async)
1309			debug = NULL;
1310	} else {
1311		tm = sel->compiler_ctx_state.tm;
1312	}
1313
1314	/* Compile the main shader part for use with a prolog and/or epilog.
1315	 * If this fails, the driver will try to compile a monolithic shader
1316	 * on demand.
1317	 */
1318	if (!sscreen->use_monolithic_shaders) {
1319		struct si_shader *shader = CALLOC_STRUCT(si_shader);
1320		void *tgsi_binary;
1321
1322		if (!shader) {
1323			fprintf(stderr, "radeonsi: can't allocate a main shader part\n");
1324			return;
1325		}
1326
1327		shader->selector = sel;
1328		si_parse_next_shader_property(&sel->info, &shader->key);
1329
1330		tgsi_binary = si_get_tgsi_binary(sel);
1331
1332		/* Try to load the shader from the shader cache. */
1333		pipe_mutex_lock(sscreen->shader_cache_mutex);
1334
1335		if (tgsi_binary &&
1336		    si_shader_cache_load_shader(sscreen, tgsi_binary, shader)) {
1337			FREE(tgsi_binary);
1338			pipe_mutex_unlock(sscreen->shader_cache_mutex);
1339		} else {
1340			pipe_mutex_unlock(sscreen->shader_cache_mutex);
1341
1342			/* Compile the shader if it hasn't been loaded from the cache. */
1343			if (si_compile_tgsi_shader(sscreen, tm, shader, false,
1344						   debug) != 0) {
1345				FREE(shader);
1346				FREE(tgsi_binary);
1347				fprintf(stderr, "radeonsi: can't compile a main shader part\n");
1348				return;
1349			}
1350
1351			if (tgsi_binary) {
1352				pipe_mutex_lock(sscreen->shader_cache_mutex);
1353				if (!si_shader_cache_insert_shader(sscreen, tgsi_binary, shader))
1354					FREE(tgsi_binary);
1355				pipe_mutex_unlock(sscreen->shader_cache_mutex);
1356			}
1357		}
1358
1359		sel->main_shader_part = shader;
1360
1361		/* Unset "outputs_written" flags for outputs converted to
1362		 * DEFAULT_VAL, so that later inter-shader optimizations don't
1363		 * try to eliminate outputs that don't exist in the final
1364		 * shader.
1365		 *
1366		 * This is only done if non-monolithic shaders are enabled.
1367		 */
1368		if ((sel->type == PIPE_SHADER_VERTEX ||
1369		     sel->type == PIPE_SHADER_TESS_EVAL) &&
1370		    !shader->key.as_ls &&
1371		    !shader->key.as_es) {
1372			unsigned i;
1373
1374			for (i = 0; i < sel->info.num_outputs; i++) {
1375				unsigned offset = shader->info.vs_output_param_offset[i];
1376
1377				if (offset <= EXP_PARAM_OFFSET_31)
1378					continue;
1379
1380				unsigned name = sel->info.output_semantic_name[i];
1381				unsigned index = sel->info.output_semantic_index[i];
1382				unsigned id;
1383
1384				switch (name) {
1385				case TGSI_SEMANTIC_GENERIC:
1386					/* don't process indices the function can't handle */
1387					if (index >= 60)
1388						break;
1389					/* fall through */
1390				case TGSI_SEMANTIC_CLIPDIST:
1391					id = si_shader_io_get_unique_index(name, index);
1392					sel->outputs_written &= ~(1ull << id);
1393					break;
1394				case TGSI_SEMANTIC_POSITION: /* ignore these */
1395				case TGSI_SEMANTIC_PSIZE:
1396				case TGSI_SEMANTIC_CLIPVERTEX:
1397				case TGSI_SEMANTIC_EDGEFLAG:
1398					break;
1399				default:
1400					id = si_shader_io_get_unique_index2(name, index);
1401					sel->outputs_written2 &= ~(1u << id);
1402				}
1403			}
1404		}
1405	}
1406
1407	/* Pre-compilation. */
1408	if (sscreen->b.debug_flags & DBG_PRECOMPILE) {
1409		struct si_shader_ctx_state state = {sel};
1410		struct si_shader_key key;
1411
1412		memset(&key, 0, sizeof(key));
1413		si_parse_next_shader_property(&sel->info, &key);
1414
1415		/* Set reasonable defaults, so that the shader key doesn't
1416		 * cause any code to be eliminated.
1417		 */
1418		switch (sel->type) {
1419		case PIPE_SHADER_TESS_CTRL:
1420			key.part.tcs.epilog.prim_mode = PIPE_PRIM_TRIANGLES;
1421			break;
1422		case PIPE_SHADER_FRAGMENT:
1423			key.part.ps.prolog.bc_optimize_for_persp =
1424				sel->info.uses_persp_center &&
1425				sel->info.uses_persp_centroid;
1426			key.part.ps.prolog.bc_optimize_for_linear =
1427				sel->info.uses_linear_center &&
1428				sel->info.uses_linear_centroid;
1429			key.part.ps.epilog.alpha_func = PIPE_FUNC_ALWAYS;
1430			for (i = 0; i < 8; i++)
1431				if (sel->info.colors_written & (1 << i))
1432					key.part.ps.epilog.spi_shader_col_format |=
1433						V_028710_SPI_SHADER_FP16_ABGR << (i * 4);
1434			break;
1435		}
1436
1437		if (si_shader_select_with_key(sscreen, &state,
1438					      &sel->compiler_ctx_state, &key,
1439					      thread_index))
1440			fprintf(stderr, "radeonsi: can't create a monolithic shader\n");
1441	}
1442
1443	/* The GS copy shader is always pre-compiled. */
1444	if (sel->type == PIPE_SHADER_GEOMETRY) {
1445		sel->gs_copy_shader = si_generate_gs_copy_shader(sscreen, tm, sel, debug);
1446		if (!sel->gs_copy_shader) {
1447			fprintf(stderr, "radeonsi: can't create GS copy shader\n");
1448			return;
1449		}
1450
1451		si_shader_vs(sscreen, sel->gs_copy_shader, sel);
1452	}
1453}
1454
1455static void *si_create_shader_selector(struct pipe_context *ctx,
1456				       const struct pipe_shader_state *state)
1457{
1458	struct si_screen *sscreen = (struct si_screen *)ctx->screen;
1459	struct si_context *sctx = (struct si_context*)ctx;
1460	struct si_shader_selector *sel = CALLOC_STRUCT(si_shader_selector);
1461	int i;
1462
1463	if (!sel)
1464		return NULL;
1465
1466	sel->screen = sscreen;
1467	sel->compiler_ctx_state.tm = sctx->tm;
1468	sel->compiler_ctx_state.debug = sctx->b.debug;
1469	sel->compiler_ctx_state.is_debug_context = sctx->is_debug;
1470	sel->tokens = tgsi_dup_tokens(state->tokens);
1471	if (!sel->tokens) {
1472		FREE(sel);
1473		return NULL;
1474	}
1475
1476	sel->so = state->stream_output;
1477	tgsi_scan_shader(state->tokens, &sel->info);
1478	sel->type = sel->info.processor;
1479	p_atomic_inc(&sscreen->b.num_shaders_created);
1480
1481	/* Set which opcode uses which (i,j) pair. */
1482	if (sel->info.uses_persp_opcode_interp_centroid)
1483		sel->info.uses_persp_centroid = true;
1484
1485	if (sel->info.uses_linear_opcode_interp_centroid)
1486		sel->info.uses_linear_centroid = true;
1487
1488	if (sel->info.uses_persp_opcode_interp_offset ||
1489	    sel->info.uses_persp_opcode_interp_sample)
1490		sel->info.uses_persp_center = true;
1491
1492	if (sel->info.uses_linear_opcode_interp_offset ||
1493	    sel->info.uses_linear_opcode_interp_sample)
1494		sel->info.uses_linear_center = true;
1495
1496	switch (sel->type) {
1497	case PIPE_SHADER_GEOMETRY:
1498		sel->gs_output_prim =
1499			sel->info.properties[TGSI_PROPERTY_GS_OUTPUT_PRIM];
1500		sel->gs_max_out_vertices =
1501			sel->info.properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
1502		sel->gs_num_invocations =
1503			sel->info.properties[TGSI_PROPERTY_GS_INVOCATIONS];
1504		sel->gsvs_vertex_size = sel->info.num_outputs * 16;
1505		sel->max_gsvs_emit_size = sel->gsvs_vertex_size *
1506					  sel->gs_max_out_vertices;
1507
1508		sel->max_gs_stream = 0;
1509		for (i = 0; i < sel->so.num_outputs; i++)
1510			sel->max_gs_stream = MAX2(sel->max_gs_stream,
1511						  sel->so.output[i].stream);
1512
1513		sel->gs_input_verts_per_prim =
1514			u_vertices_per_prim(sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM]);
1515		break;
1516
1517	case PIPE_SHADER_TESS_CTRL:
1518		/* Always reserve space for these. */
1519		sel->patch_outputs_written |=
1520			(1llu << si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSINNER, 0)) |
1521			(1llu << si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSOUTER, 0));
1522		/* fall through */
1523	case PIPE_SHADER_VERTEX:
1524	case PIPE_SHADER_TESS_EVAL:
1525		for (i = 0; i < sel->info.num_outputs; i++) {
1526			unsigned name = sel->info.output_semantic_name[i];
1527			unsigned index = sel->info.output_semantic_index[i];
1528
1529			switch (name) {
1530			case TGSI_SEMANTIC_TESSINNER:
1531			case TGSI_SEMANTIC_TESSOUTER:
1532			case TGSI_SEMANTIC_PATCH:
1533				sel->patch_outputs_written |=
1534					1llu << si_shader_io_get_unique_index(name, index);
1535				break;
1536
1537			case TGSI_SEMANTIC_GENERIC:
1538				/* don't process indices the function can't handle */
1539				if (index >= 60)
1540					break;
1541				/* fall through */
1542			case TGSI_SEMANTIC_POSITION:
1543			case TGSI_SEMANTIC_PSIZE:
1544			case TGSI_SEMANTIC_CLIPDIST:
1545				sel->outputs_written |=
1546					1llu << si_shader_io_get_unique_index(name, index);
1547				break;
1548			case TGSI_SEMANTIC_CLIPVERTEX: /* ignore these */
1549			case TGSI_SEMANTIC_EDGEFLAG:
1550				break;
1551			default:
1552				sel->outputs_written2 |=
1553					1u << si_shader_io_get_unique_index2(name, index);
1554			}
1555		}
1556		sel->esgs_itemsize = util_last_bit64(sel->outputs_written) * 16;
1557		break;
1558
1559	case PIPE_SHADER_FRAGMENT:
1560		for (i = 0; i < sel->info.num_inputs; i++) {
1561			unsigned name = sel->info.input_semantic_name[i];
1562			unsigned index = sel->info.input_semantic_index[i];
1563
1564			switch (name) {
1565			case TGSI_SEMANTIC_CLIPDIST:
1566			case TGSI_SEMANTIC_GENERIC:
1567				sel->inputs_read |=
1568					1llu << si_shader_io_get_unique_index(name, index);
1569				break;
1570			case TGSI_SEMANTIC_PCOORD: /* ignore this */
1571				break;
1572			default:
1573				sel->inputs_read2 |=
1574					1u << si_shader_io_get_unique_index2(name, index);
1575			}
1576		}
1577
1578		for (i = 0; i < 8; i++)
1579			if (sel->info.colors_written & (1 << i))
1580				sel->colors_written_4bit |= 0xf << (4 * i);
1581
1582		for (i = 0; i < sel->info.num_inputs; i++) {
1583			if (sel->info.input_semantic_name[i] == TGSI_SEMANTIC_COLOR) {
1584				int index = sel->info.input_semantic_index[i];
1585				sel->color_attr_index[index] = i;
1586			}
1587		}
1588		break;
1589	}
1590
1591	/* DB_SHADER_CONTROL */
1592	sel->db_shader_control =
1593		S_02880C_Z_EXPORT_ENABLE(sel->info.writes_z) |
1594		S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(sel->info.writes_stencil) |
1595		S_02880C_MASK_EXPORT_ENABLE(sel->info.writes_samplemask) |
1596		S_02880C_KILL_ENABLE(sel->info.uses_kill);
1597
1598	switch (sel->info.properties[TGSI_PROPERTY_FS_DEPTH_LAYOUT]) {
1599	case TGSI_FS_DEPTH_LAYOUT_GREATER:
1600		sel->db_shader_control |=
1601			S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z);
1602		break;
1603	case TGSI_FS_DEPTH_LAYOUT_LESS:
1604		sel->db_shader_control |=
1605			S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z);
1606		break;
1607	}
1608
1609	/* Z_ORDER, EXEC_ON_HIER_FAIL and EXEC_ON_NOOP should be set as following:
1610	 *
1611	 *   | early Z/S | writes_mem | allow_ReZ? |      Z_ORDER       | EXEC_ON_HIER_FAIL | EXEC_ON_NOOP
1612	 * --|-----------|------------|------------|--------------------|-------------------|-------------
1613	 * 1a|   false   |   false    |   true     | EarlyZ_Then_ReZ    |         0         |     0
1614	 * 1b|   false   |   false    |   false    | EarlyZ_Then_LateZ  |         0         |     0
1615	 * 2 |   false   |   true     |   n/a      |       LateZ        |         1         |     0
1616	 * 3 |   true    |   false    |   n/a      | EarlyZ_Then_LateZ  |         0         |     0
1617	 * 4 |   true    |   true     |   n/a      | EarlyZ_Then_LateZ  |         0         |     1
1618	 *
1619	 * In cases 3 and 4, HW will force Z_ORDER to EarlyZ regardless of what's set in the register.
1620	 * In case 2, NOOP_CULL is a don't care field. In case 2, 3 and 4, ReZ doesn't make sense.
1621	 *
1622	 * Don't use ReZ without profiling !!!
1623	 *
1624	 * ReZ decreases performance by 15% in DiRT: Showdown on Ultra settings, which has pretty complex
1625	 * shaders.
1626	 */
1627	if (sel->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL]) {
1628		/* Cases 3, 4. */
1629		sel->db_shader_control |= S_02880C_DEPTH_BEFORE_SHADER(1) |
1630					  S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z) |
1631					  S_02880C_EXEC_ON_NOOP(sel->info.writes_memory);
1632	} else if (sel->info.writes_memory) {
1633		/* Case 2. */
1634		sel->db_shader_control |= S_02880C_Z_ORDER(V_02880C_LATE_Z) |
1635					  S_02880C_EXEC_ON_HIER_FAIL(1);
1636	} else {
1637		/* Case 1. */
1638		sel->db_shader_control |= S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z);
1639	}
1640
1641	pipe_mutex_init(sel->mutex);
1642	util_queue_fence_init(&sel->ready);
1643
1644	if ((sctx->b.debug.debug_message && !sctx->b.debug.async) ||
1645	    sctx->is_debug ||
1646	    r600_can_dump_shader(&sscreen->b, sel->info.processor) ||
1647	    !util_queue_is_initialized(&sscreen->shader_compiler_queue))
1648		si_init_shader_selector_async(sel, -1);
1649	else
1650		util_queue_add_job(&sscreen->shader_compiler_queue, sel,
1651                                   &sel->ready, si_init_shader_selector_async,
1652                                   NULL);
1653
1654	return sel;
1655}
1656
1657static void si_bind_vs_shader(struct pipe_context *ctx, void *state)
1658{
1659	struct si_context *sctx = (struct si_context *)ctx;
1660	struct si_shader_selector *sel = state;
1661
1662	if (sctx->vs_shader.cso == sel)
1663		return;
1664
1665	sctx->vs_shader.cso = sel;
1666	sctx->vs_shader.current = sel ? sel->first_variant : NULL;
1667	sctx->do_update_shaders = true;
1668	si_mark_atom_dirty(sctx, &sctx->clip_regs);
1669	r600_update_vs_writes_viewport_index(&sctx->b, si_get_vs_info(sctx));
1670}
1671
1672static void si_bind_gs_shader(struct pipe_context *ctx, void *state)
1673{
1674	struct si_context *sctx = (struct si_context *)ctx;
1675	struct si_shader_selector *sel = state;
1676	bool enable_changed = !!sctx->gs_shader.cso != !!sel;
1677
1678	if (sctx->gs_shader.cso == sel)
1679		return;
1680
1681	sctx->gs_shader.cso = sel;
1682	sctx->gs_shader.current = sel ? sel->first_variant : NULL;
1683	sctx->do_update_shaders = true;
1684	si_mark_atom_dirty(sctx, &sctx->clip_regs);
1685	sctx->last_rast_prim = -1; /* reset this so that it gets updated */
1686
1687	if (enable_changed)
1688		si_shader_change_notify(sctx);
1689	r600_update_vs_writes_viewport_index(&sctx->b, si_get_vs_info(sctx));
1690}
1691
1692static void si_bind_tcs_shader(struct pipe_context *ctx, void *state)
1693{
1694	struct si_context *sctx = (struct si_context *)ctx;
1695	struct si_shader_selector *sel = state;
1696	bool enable_changed = !!sctx->tcs_shader.cso != !!sel;
1697
1698	if (sctx->tcs_shader.cso == sel)
1699		return;
1700
1701	sctx->tcs_shader.cso = sel;
1702	sctx->tcs_shader.current = sel ? sel->first_variant : NULL;
1703	sctx->do_update_shaders = true;
1704
1705	if (enable_changed)
1706		sctx->last_tcs = NULL; /* invalidate derived tess state */
1707}
1708
1709static void si_bind_tes_shader(struct pipe_context *ctx, void *state)
1710{
1711	struct si_context *sctx = (struct si_context *)ctx;
1712	struct si_shader_selector *sel = state;
1713	bool enable_changed = !!sctx->tes_shader.cso != !!sel;
1714
1715	if (sctx->tes_shader.cso == sel)
1716		return;
1717
1718	sctx->tes_shader.cso = sel;
1719	sctx->tes_shader.current = sel ? sel->first_variant : NULL;
1720	sctx->do_update_shaders = true;
1721	si_mark_atom_dirty(sctx, &sctx->clip_regs);
1722	sctx->last_rast_prim = -1; /* reset this so that it gets updated */
1723
1724	if (enable_changed) {
1725		si_shader_change_notify(sctx);
1726		sctx->last_tes_sh_base = -1; /* invalidate derived tess state */
1727	}
1728	r600_update_vs_writes_viewport_index(&sctx->b, si_get_vs_info(sctx));
1729}
1730
1731static void si_bind_ps_shader(struct pipe_context *ctx, void *state)
1732{
1733	struct si_context *sctx = (struct si_context *)ctx;
1734	struct si_shader_selector *sel = state;
1735
1736	/* skip if supplied shader is one already in use */
1737	if (sctx->ps_shader.cso == sel)
1738		return;
1739
1740	sctx->ps_shader.cso = sel;
1741	sctx->ps_shader.current = sel ? sel->first_variant : NULL;
1742	sctx->do_update_shaders = true;
1743	si_mark_atom_dirty(sctx, &sctx->cb_render_state);
1744}
1745
1746static void si_delete_shader(struct si_context *sctx, struct si_shader *shader)
1747{
1748	if (shader->is_optimized) {
1749		util_queue_job_wait(&shader->optimized_ready);
1750		util_queue_fence_destroy(&shader->optimized_ready);
1751	}
1752
1753	if (shader->pm4) {
1754		switch (shader->selector->type) {
1755		case PIPE_SHADER_VERTEX:
1756			if (shader->key.as_ls)
1757				si_pm4_delete_state(sctx, ls, shader->pm4);
1758			else if (shader->key.as_es)
1759				si_pm4_delete_state(sctx, es, shader->pm4);
1760			else
1761				si_pm4_delete_state(sctx, vs, shader->pm4);
1762			break;
1763		case PIPE_SHADER_TESS_CTRL:
1764			si_pm4_delete_state(sctx, hs, shader->pm4);
1765			break;
1766		case PIPE_SHADER_TESS_EVAL:
1767			if (shader->key.as_es)
1768				si_pm4_delete_state(sctx, es, shader->pm4);
1769			else
1770				si_pm4_delete_state(sctx, vs, shader->pm4);
1771			break;
1772		case PIPE_SHADER_GEOMETRY:
1773			if (shader->is_gs_copy_shader)
1774				si_pm4_delete_state(sctx, vs, shader->pm4);
1775			else
1776				si_pm4_delete_state(sctx, gs, shader->pm4);
1777			break;
1778		case PIPE_SHADER_FRAGMENT:
1779			si_pm4_delete_state(sctx, ps, shader->pm4);
1780			break;
1781		}
1782	}
1783
1784	si_shader_destroy(shader);
1785	free(shader);
1786}
1787
1788static void si_delete_shader_selector(struct pipe_context *ctx, void *state)
1789{
1790	struct si_context *sctx = (struct si_context *)ctx;
1791	struct si_shader_selector *sel = (struct si_shader_selector *)state;
1792	struct si_shader *p = sel->first_variant, *c;
1793	struct si_shader_ctx_state *current_shader[SI_NUM_SHADERS] = {
1794		[PIPE_SHADER_VERTEX] = &sctx->vs_shader,
1795		[PIPE_SHADER_TESS_CTRL] = &sctx->tcs_shader,
1796		[PIPE_SHADER_TESS_EVAL] = &sctx->tes_shader,
1797		[PIPE_SHADER_GEOMETRY] = &sctx->gs_shader,
1798		[PIPE_SHADER_FRAGMENT] = &sctx->ps_shader,
1799	};
1800
1801	util_queue_job_wait(&sel->ready);
1802
1803	if (current_shader[sel->type]->cso == sel) {
1804		current_shader[sel->type]->cso = NULL;
1805		current_shader[sel->type]->current = NULL;
1806	}
1807
1808	while (p) {
1809		c = p->next_variant;
1810		si_delete_shader(sctx, p);
1811		p = c;
1812	}
1813
1814	if (sel->main_shader_part)
1815		si_delete_shader(sctx, sel->main_shader_part);
1816	if (sel->gs_copy_shader)
1817		si_delete_shader(sctx, sel->gs_copy_shader);
1818
1819	util_queue_fence_destroy(&sel->ready);
1820	pipe_mutex_destroy(sel->mutex);
1821	free(sel->tokens);
1822	free(sel);
1823}
1824
1825static unsigned si_get_ps_input_cntl(struct si_context *sctx,
1826				     struct si_shader *vs, unsigned name,
1827				     unsigned index, unsigned interpolate)
1828{
1829	struct tgsi_shader_info *vsinfo = &vs->selector->info;
1830	unsigned j, offset, ps_input_cntl = 0;
1831
1832	if (interpolate == TGSI_INTERPOLATE_CONSTANT ||
1833	    (interpolate == TGSI_INTERPOLATE_COLOR && sctx->flatshade))
1834		ps_input_cntl |= S_028644_FLAT_SHADE(1);
1835
1836	if (name == TGSI_SEMANTIC_PCOORD ||
1837	    (name == TGSI_SEMANTIC_TEXCOORD &&
1838	     sctx->sprite_coord_enable & (1 << index))) {
1839		ps_input_cntl |= S_028644_PT_SPRITE_TEX(1);
1840	}
1841
1842	for (j = 0; j < vsinfo->num_outputs; j++) {
1843		if (name == vsinfo->output_semantic_name[j] &&
1844		    index == vsinfo->output_semantic_index[j]) {
1845			offset = vs->info.vs_output_param_offset[j];
1846
1847			if (offset <= EXP_PARAM_OFFSET_31) {
1848				/* The input is loaded from parameter memory. */
1849				ps_input_cntl |= S_028644_OFFSET(offset);
1850			} else if (!G_028644_PT_SPRITE_TEX(ps_input_cntl)) {
1851				if (offset == EXP_PARAM_UNDEFINED) {
1852					/* This can happen with depth-only rendering. */
1853					offset = 0;
1854				} else {
1855					/* The input is a DEFAULT_VAL constant. */
1856					assert(offset >= EXP_PARAM_DEFAULT_VAL_0000 &&
1857					       offset <= EXP_PARAM_DEFAULT_VAL_1111);
1858					offset -= EXP_PARAM_DEFAULT_VAL_0000;
1859				}
1860
1861				ps_input_cntl = S_028644_OFFSET(0x20) |
1862						S_028644_DEFAULT_VAL(offset);
1863			}
1864			break;
1865		}
1866	}
1867
1868	if (name == TGSI_SEMANTIC_PRIMID)
1869		/* PrimID is written after the last output. */
1870		ps_input_cntl |= S_028644_OFFSET(vs->info.vs_output_param_offset[vsinfo->num_outputs]);
1871	else if (j == vsinfo->num_outputs && !G_028644_PT_SPRITE_TEX(ps_input_cntl)) {
1872		/* No corresponding output found, load defaults into input.
1873		 * Don't set any other bits.
1874		 * (FLAT_SHADE=1 completely changes behavior) */
1875		ps_input_cntl = S_028644_OFFSET(0x20);
1876		/* D3D 9 behaviour. GL is undefined */
1877		if (name == TGSI_SEMANTIC_COLOR && index == 0)
1878			ps_input_cntl |= S_028644_DEFAULT_VAL(3);
1879	}
1880	return ps_input_cntl;
1881}
1882
1883static void si_emit_spi_map(struct si_context *sctx, struct r600_atom *atom)
1884{
1885	struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
1886	struct si_shader *ps = sctx->ps_shader.current;
1887	struct si_shader *vs = si_get_vs_state(sctx);
1888	struct tgsi_shader_info *psinfo = ps ? &ps->selector->info : NULL;
1889	unsigned i, num_interp, num_written = 0, bcol_interp[2];
1890
1891	if (!ps || !ps->selector->info.num_inputs)
1892		return;
1893
1894	num_interp = si_get_ps_num_interp(ps);
1895	assert(num_interp > 0);
1896	radeon_set_context_reg_seq(cs, R_028644_SPI_PS_INPUT_CNTL_0, num_interp);
1897
1898	for (i = 0; i < psinfo->num_inputs; i++) {
1899		unsigned name = psinfo->input_semantic_name[i];
1900		unsigned index = psinfo->input_semantic_index[i];
1901		unsigned interpolate = psinfo->input_interpolate[i];
1902
1903		radeon_emit(cs, si_get_ps_input_cntl(sctx, vs, name, index,
1904						     interpolate));
1905		num_written++;
1906
1907		if (name == TGSI_SEMANTIC_COLOR) {
1908			assert(index < ARRAY_SIZE(bcol_interp));
1909			bcol_interp[index] = interpolate;
1910		}
1911	}
1912
1913	if (ps->key.part.ps.prolog.color_two_side) {
1914		unsigned bcol = TGSI_SEMANTIC_BCOLOR;
1915
1916		for (i = 0; i < 2; i++) {
1917			if (!(psinfo->colors_read & (0xf << (i * 4))))
1918				continue;
1919
1920			radeon_emit(cs, si_get_ps_input_cntl(sctx, vs, bcol,
1921							     i, bcol_interp[i]));
1922			num_written++;
1923		}
1924	}
1925	assert(num_interp == num_written);
1926}
1927
1928/**
1929 * Writing CONFIG or UCONFIG VGT registers requires VGT_FLUSH before that.
1930 */
1931static void si_init_config_add_vgt_flush(struct si_context *sctx)
1932{
1933	if (sctx->init_config_has_vgt_flush)
1934		return;
1935
1936	/* Done by Vulkan before VGT_FLUSH. */
1937	si_pm4_cmd_begin(sctx->init_config, PKT3_EVENT_WRITE);
1938	si_pm4_cmd_add(sctx->init_config,
1939		       EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
1940	si_pm4_cmd_end(sctx->init_config, false);
1941
1942	/* VGT_FLUSH is required even if VGT is idle. It resets VGT pointers. */
1943	si_pm4_cmd_begin(sctx->init_config, PKT3_EVENT_WRITE);
1944	si_pm4_cmd_add(sctx->init_config, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
1945	si_pm4_cmd_end(sctx->init_config, false);
1946	sctx->init_config_has_vgt_flush = true;
1947}
1948
1949/* Initialize state related to ESGS / GSVS ring buffers */
1950static bool si_update_gs_ring_buffers(struct si_context *sctx)
1951{
1952	struct si_shader_selector *es =
1953		sctx->tes_shader.cso ? sctx->tes_shader.cso : sctx->vs_shader.cso;
1954	struct si_shader_selector *gs = sctx->gs_shader.cso;
1955	struct si_pm4_state *pm4;
1956
1957	/* Chip constants. */
1958	unsigned num_se = sctx->screen->b.info.max_se;
1959	unsigned wave_size = 64;
1960	unsigned max_gs_waves = 32 * num_se; /* max 32 per SE on GCN */
1961	unsigned gs_vertex_reuse = 16 * num_se; /* GS_VERTEX_REUSE register (per SE) */
1962	unsigned alignment = 256 * num_se;
1963	/* The maximum size is 63.999 MB per SE. */
1964	unsigned max_size = ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se;
1965
1966	/* Calculate the minimum size. */
1967	unsigned min_esgs_ring_size = align(es->esgs_itemsize * gs_vertex_reuse *
1968					    wave_size, alignment);
1969
1970	/* These are recommended sizes, not minimum sizes. */
1971	unsigned esgs_ring_size = max_gs_waves * 2 * wave_size *
1972				  es->esgs_itemsize * gs->gs_input_verts_per_prim;
1973	unsigned gsvs_ring_size = max_gs_waves * 2 * wave_size *
1974				  gs->max_gsvs_emit_size;
1975
1976	min_esgs_ring_size = align(min_esgs_ring_size, alignment);
1977	esgs_ring_size = align(esgs_ring_size, alignment);
1978	gsvs_ring_size = align(gsvs_ring_size, alignment);
1979
1980	esgs_ring_size = CLAMP(esgs_ring_size, min_esgs_ring_size, max_size);
1981	gsvs_ring_size = MIN2(gsvs_ring_size, max_size);
1982
1983	/* Some rings don't have to be allocated if shaders don't use them.
1984	 * (e.g. no varyings between ES and GS or GS and VS)
1985	 */
1986	bool update_esgs = esgs_ring_size &&
1987			   (!sctx->esgs_ring ||
1988			    sctx->esgs_ring->width0 < esgs_ring_size);
1989	bool update_gsvs = gsvs_ring_size &&
1990			   (!sctx->gsvs_ring ||
1991			    sctx->gsvs_ring->width0 < gsvs_ring_size);
1992
1993	if (!update_esgs && !update_gsvs)
1994		return true;
1995
1996	if (update_esgs) {
1997		pipe_resource_reference(&sctx->esgs_ring, NULL);
1998		sctx->esgs_ring = pipe_buffer_create(sctx->b.b.screen, 0,
1999						     PIPE_USAGE_DEFAULT,
2000						     esgs_ring_size);
2001		if (!sctx->esgs_ring)
2002			return false;
2003	}
2004
2005	if (update_gsvs) {
2006		pipe_resource_reference(&sctx->gsvs_ring, NULL);
2007		sctx->gsvs_ring = pipe_buffer_create(sctx->b.b.screen, 0,
2008						     PIPE_USAGE_DEFAULT,
2009						     gsvs_ring_size);
2010		if (!sctx->gsvs_ring)
2011			return false;
2012	}
2013
2014	/* Create the "init_config_gs_rings" state. */
2015	pm4 = CALLOC_STRUCT(si_pm4_state);
2016	if (!pm4)
2017		return false;
2018
2019	if (sctx->b.chip_class >= CIK) {
2020		if (sctx->esgs_ring)
2021			si_pm4_set_reg(pm4, R_030900_VGT_ESGS_RING_SIZE,
2022				       sctx->esgs_ring->width0 / 256);
2023		if (sctx->gsvs_ring)
2024			si_pm4_set_reg(pm4, R_030904_VGT_GSVS_RING_SIZE,
2025				       sctx->gsvs_ring->width0 / 256);
2026	} else {
2027		if (sctx->esgs_ring)
2028			si_pm4_set_reg(pm4, R_0088C8_VGT_ESGS_RING_SIZE,
2029				       sctx->esgs_ring->width0 / 256);
2030		if (sctx->gsvs_ring)
2031			si_pm4_set_reg(pm4, R_0088CC_VGT_GSVS_RING_SIZE,
2032				       sctx->gsvs_ring->width0 / 256);
2033	}
2034
2035	/* Set the state. */
2036	if (sctx->init_config_gs_rings)
2037		si_pm4_free_state(sctx, sctx->init_config_gs_rings, ~0);
2038	sctx->init_config_gs_rings = pm4;
2039
2040	if (!sctx->init_config_has_vgt_flush) {
2041		si_init_config_add_vgt_flush(sctx);
2042		si_pm4_upload_indirect_buffer(sctx, sctx->init_config);
2043	}
2044
2045	/* Flush the context to re-emit both init_config states. */
2046	sctx->b.initial_gfx_cs_size = 0; /* force flush */
2047	si_context_gfx_flush(sctx, RADEON_FLUSH_ASYNC, NULL);
2048
2049	/* Set ring bindings. */
2050	if (sctx->esgs_ring) {
2051		si_set_ring_buffer(&sctx->b.b, SI_ES_RING_ESGS,
2052				   sctx->esgs_ring, 0, sctx->esgs_ring->width0,
2053				   true, true, 4, 64, 0);
2054		si_set_ring_buffer(&sctx->b.b, SI_GS_RING_ESGS,
2055				   sctx->esgs_ring, 0, sctx->esgs_ring->width0,
2056				   false, false, 0, 0, 0);
2057	}
2058	if (sctx->gsvs_ring) {
2059		si_set_ring_buffer(&sctx->b.b, SI_RING_GSVS,
2060				   sctx->gsvs_ring, 0, sctx->gsvs_ring->width0,
2061				   false, false, 0, 0, 0);
2062	}
2063
2064	return true;
2065}
2066
2067/**
2068 * @returns 1 if \p sel has been updated to use a new scratch buffer
2069 *          0 if not
2070 *          < 0 if there was a failure
2071 */
2072static int si_update_scratch_buffer(struct si_context *sctx,
2073				    struct si_shader *shader)
2074{
2075	uint64_t scratch_va = sctx->scratch_buffer->gpu_address;
2076	int r;
2077
2078	if (!shader)
2079		return 0;
2080
2081	/* This shader doesn't need a scratch buffer */
2082	if (shader->config.scratch_bytes_per_wave == 0)
2083		return 0;
2084
2085	/* This shader is already configured to use the current
2086	 * scratch buffer. */
2087	if (shader->scratch_bo == sctx->scratch_buffer)
2088		return 0;
2089
2090	assert(sctx->scratch_buffer);
2091
2092	si_shader_apply_scratch_relocs(sctx, shader, &shader->config, scratch_va);
2093
2094	/* Replace the shader bo with a new bo that has the relocs applied. */
2095	r = si_shader_binary_upload(sctx->screen, shader);
2096	if (r)
2097		return r;
2098
2099	/* Update the shader state to use the new shader bo. */
2100	si_shader_init_pm4_state(sctx->screen, shader);
2101
2102	r600_resource_reference(&shader->scratch_bo, sctx->scratch_buffer);
2103
2104	return 1;
2105}
2106
2107static unsigned si_get_current_scratch_buffer_size(struct si_context *sctx)
2108{
2109	return sctx->scratch_buffer ? sctx->scratch_buffer->b.b.width0 : 0;
2110}
2111
2112static unsigned si_get_scratch_buffer_bytes_per_wave(struct si_shader *shader)
2113{
2114	return shader ? shader->config.scratch_bytes_per_wave : 0;
2115}
2116
2117static unsigned si_get_max_scratch_bytes_per_wave(struct si_context *sctx)
2118{
2119	unsigned bytes = 0;
2120
2121	bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->ps_shader.current));
2122	bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->gs_shader.current));
2123	bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->vs_shader.current));
2124	bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->tcs_shader.current));
2125	bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->tes_shader.current));
2126	return bytes;
2127}
2128
2129static bool si_update_spi_tmpring_size(struct si_context *sctx)
2130{
2131	unsigned current_scratch_buffer_size =
2132		si_get_current_scratch_buffer_size(sctx);
2133	unsigned scratch_bytes_per_wave =
2134		si_get_max_scratch_bytes_per_wave(sctx);
2135	unsigned scratch_needed_size = scratch_bytes_per_wave *
2136		sctx->scratch_waves;
2137	unsigned spi_tmpring_size;
2138	int r;
2139
2140	if (scratch_needed_size > 0) {
2141		if (scratch_needed_size > current_scratch_buffer_size) {
2142			/* Create a bigger scratch buffer */
2143			r600_resource_reference(&sctx->scratch_buffer, NULL);
2144
2145			sctx->scratch_buffer = (struct r600_resource*)
2146					pipe_buffer_create(&sctx->screen->b.b, 0,
2147	                                PIPE_USAGE_DEFAULT, scratch_needed_size);
2148			if (!sctx->scratch_buffer)
2149				return false;
2150			sctx->emit_scratch_reloc = true;
2151		}
2152
2153		/* Update the shaders, so they are using the latest scratch.  The
2154		 * scratch buffer may have been changed since these shaders were
2155		 * last used, so we still need to try to update them, even if
2156		 * they require scratch buffers smaller than the current size.
2157		 */
2158		r = si_update_scratch_buffer(sctx, sctx->ps_shader.current);
2159		if (r < 0)
2160			return false;
2161		if (r == 1)
2162			si_pm4_bind_state(sctx, ps, sctx->ps_shader.current->pm4);
2163
2164		r = si_update_scratch_buffer(sctx, sctx->gs_shader.current);
2165		if (r < 0)
2166			return false;
2167		if (r == 1)
2168			si_pm4_bind_state(sctx, gs, sctx->gs_shader.current->pm4);
2169
2170		r = si_update_scratch_buffer(sctx, sctx->tcs_shader.current);
2171		if (r < 0)
2172			return false;
2173		if (r == 1)
2174			si_pm4_bind_state(sctx, hs, sctx->tcs_shader.current->pm4);
2175
2176		/* VS can be bound as LS, ES, or VS. */
2177		r = si_update_scratch_buffer(sctx, sctx->vs_shader.current);
2178		if (r < 0)
2179			return false;
2180		if (r == 1) {
2181			if (sctx->tes_shader.current)
2182				si_pm4_bind_state(sctx, ls, sctx->vs_shader.current->pm4);
2183			else if (sctx->gs_shader.current)
2184				si_pm4_bind_state(sctx, es, sctx->vs_shader.current->pm4);
2185			else
2186				si_pm4_bind_state(sctx, vs, sctx->vs_shader.current->pm4);
2187		}
2188
2189		/* TES can be bound as ES or VS. */
2190		r = si_update_scratch_buffer(sctx, sctx->tes_shader.current);
2191		if (r < 0)
2192			return false;
2193		if (r == 1) {
2194			if (sctx->gs_shader.current)
2195				si_pm4_bind_state(sctx, es, sctx->tes_shader.current->pm4);
2196			else
2197				si_pm4_bind_state(sctx, vs, sctx->tes_shader.current->pm4);
2198		}
2199	}
2200
2201	/* The LLVM shader backend should be reporting aligned scratch_sizes. */
2202	assert((scratch_needed_size & ~0x3FF) == scratch_needed_size &&
2203		"scratch size should already be aligned correctly.");
2204
2205	spi_tmpring_size = S_0286E8_WAVES(sctx->scratch_waves) |
2206			   S_0286E8_WAVESIZE(scratch_bytes_per_wave >> 10);
2207	if (spi_tmpring_size != sctx->spi_tmpring_size) {
2208		sctx->spi_tmpring_size = spi_tmpring_size;
2209		sctx->emit_scratch_reloc = true;
2210	}
2211	return true;
2212}
2213
2214static void si_init_tess_factor_ring(struct si_context *sctx)
2215{
2216	bool double_offchip_buffers = sctx->b.chip_class >= CIK &&
2217				      sctx->b.family != CHIP_CARRIZO &&
2218				      sctx->b.family != CHIP_STONEY;
2219	unsigned max_offchip_buffers_per_se = double_offchip_buffers ? 128 : 64;
2220	unsigned max_offchip_buffers = max_offchip_buffers_per_se *
2221				       sctx->screen->b.info.max_se;
2222	unsigned offchip_granularity;
2223
2224	switch (sctx->screen->tess_offchip_block_dw_size) {
2225	default:
2226		assert(0);
2227		/* fall through */
2228	case 8192:
2229		offchip_granularity = V_03093C_X_8K_DWORDS;
2230		break;
2231	case 4096:
2232		offchip_granularity = V_03093C_X_4K_DWORDS;
2233		break;
2234	}
2235
2236	switch (sctx->b.chip_class) {
2237	case SI:
2238		max_offchip_buffers = MIN2(max_offchip_buffers, 126);
2239		break;
2240	case CIK:
2241		max_offchip_buffers = MIN2(max_offchip_buffers, 508);
2242		break;
2243	case VI:
2244	default:
2245		max_offchip_buffers = MIN2(max_offchip_buffers, 512);
2246		break;
2247	}
2248
2249	assert(!sctx->tf_ring);
2250	sctx->tf_ring = pipe_buffer_create(sctx->b.b.screen, 0,
2251					   PIPE_USAGE_DEFAULT,
2252					   32768 * sctx->screen->b.info.max_se);
2253	if (!sctx->tf_ring)
2254		return;
2255
2256	assert(((sctx->tf_ring->width0 / 4) & C_030938_SIZE) == 0);
2257
2258	sctx->tess_offchip_ring = pipe_buffer_create(sctx->b.b.screen, 0,
2259	                                             PIPE_USAGE_DEFAULT,
2260	                                             max_offchip_buffers *
2261	                                             sctx->screen->tess_offchip_block_dw_size * 4);
2262	if (!sctx->tess_offchip_ring)
2263		return;
2264
2265	si_init_config_add_vgt_flush(sctx);
2266
2267	/* Append these registers to the init config state. */
2268	if (sctx->b.chip_class >= CIK) {
2269		if (sctx->b.chip_class >= VI)
2270			--max_offchip_buffers;
2271
2272		si_pm4_set_reg(sctx->init_config, R_030938_VGT_TF_RING_SIZE,
2273			       S_030938_SIZE(sctx->tf_ring->width0 / 4));
2274		si_pm4_set_reg(sctx->init_config, R_030940_VGT_TF_MEMORY_BASE,
2275			       r600_resource(sctx->tf_ring)->gpu_address >> 8);
2276		si_pm4_set_reg(sctx->init_config, R_03093C_VGT_HS_OFFCHIP_PARAM,
2277		             S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers) |
2278		             S_03093C_OFFCHIP_GRANULARITY(offchip_granularity));
2279	} else {
2280		assert(offchip_granularity == V_03093C_X_8K_DWORDS);
2281		si_pm4_set_reg(sctx->init_config, R_008988_VGT_TF_RING_SIZE,
2282			       S_008988_SIZE(sctx->tf_ring->width0 / 4));
2283		si_pm4_set_reg(sctx->init_config, R_0089B8_VGT_TF_MEMORY_BASE,
2284			       r600_resource(sctx->tf_ring)->gpu_address >> 8);
2285		si_pm4_set_reg(sctx->init_config, R_0089B0_VGT_HS_OFFCHIP_PARAM,
2286		               S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers));
2287	}
2288
2289	/* Flush the context to re-emit the init_config state.
2290	 * This is done only once in a lifetime of a context.
2291	 */
2292	si_pm4_upload_indirect_buffer(sctx, sctx->init_config);
2293	sctx->b.initial_gfx_cs_size = 0; /* force flush */
2294	si_context_gfx_flush(sctx, RADEON_FLUSH_ASYNC, NULL);
2295
2296	si_set_ring_buffer(&sctx->b.b, SI_HS_RING_TESS_FACTOR, sctx->tf_ring,
2297			   0, sctx->tf_ring->width0, false, false, 0, 0, 0);
2298
2299	si_set_ring_buffer(&sctx->b.b, SI_HS_RING_TESS_OFFCHIP,
2300	                   sctx->tess_offchip_ring, 0,
2301	                   sctx->tess_offchip_ring->width0, false, false, 0, 0, 0);
2302}
2303
2304/**
2305 * This is used when TCS is NULL in the VS->TCS->TES chain. In this case,
2306 * VS passes its outputs to TES directly, so the fixed-function shader only
2307 * has to write TESSOUTER and TESSINNER.
2308 */
2309static void si_generate_fixed_func_tcs(struct si_context *sctx)
2310{
2311	struct ureg_src outer, inner;
2312	struct ureg_dst tessouter, tessinner;
2313	struct ureg_program *ureg = ureg_create(PIPE_SHADER_TESS_CTRL);
2314
2315	if (!ureg)
2316		return; /* if we get here, we're screwed */
2317
2318	assert(!sctx->fixed_func_tcs_shader.cso);
2319
2320	outer = ureg_DECL_system_value(ureg,
2321				       TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI, 0);
2322	inner = ureg_DECL_system_value(ureg,
2323				       TGSI_SEMANTIC_DEFAULT_TESSINNER_SI, 0);
2324
2325	tessouter = ureg_DECL_output(ureg, TGSI_SEMANTIC_TESSOUTER, 0);
2326	tessinner = ureg_DECL_output(ureg, TGSI_SEMANTIC_TESSINNER, 0);
2327
2328	ureg_MOV(ureg, tessouter, outer);
2329	ureg_MOV(ureg, tessinner, inner);
2330	ureg_END(ureg);
2331
2332	sctx->fixed_func_tcs_shader.cso =
2333		ureg_create_shader_and_destroy(ureg, &sctx->b.b);
2334}
2335
2336static void si_update_vgt_shader_config(struct si_context *sctx)
2337{
2338	/* Calculate the index of the config.
2339	 * 0 = VS, 1 = VS+GS, 2 = VS+Tess, 3 = VS+Tess+GS */
2340	unsigned index = 2*!!sctx->tes_shader.cso + !!sctx->gs_shader.cso;
2341	struct si_pm4_state **pm4 = &sctx->vgt_shader_config[index];
2342
2343	if (!*pm4) {
2344		uint32_t stages = 0;
2345
2346		*pm4 = CALLOC_STRUCT(si_pm4_state);
2347
2348		if (sctx->tes_shader.cso) {
2349			stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) |
2350				  S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
2351
2352			if (sctx->gs_shader.cso)
2353				stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) |
2354					  S_028B54_GS_EN(1) |
2355				          S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
2356			else
2357				stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
2358		} else if (sctx->gs_shader.cso) {
2359			stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) |
2360				  S_028B54_GS_EN(1) |
2361			          S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
2362		}
2363
2364		si_pm4_set_reg(*pm4, R_028B54_VGT_SHADER_STAGES_EN, stages);
2365	}
2366	si_pm4_bind_state(sctx, vgt_shader_config, *pm4);
2367}
2368
2369static void si_update_so(struct si_context *sctx, struct si_shader_selector *shader)
2370{
2371	struct pipe_stream_output_info *so = &shader->so;
2372	uint32_t enabled_stream_buffers_mask = 0;
2373	int i;
2374
2375	for (i = 0; i < so->num_outputs; i++)
2376		enabled_stream_buffers_mask |= (1 << so->output[i].output_buffer) << (so->output[i].stream * 4);
2377	sctx->b.streamout.enabled_stream_buffers_mask = enabled_stream_buffers_mask;
2378	sctx->b.streamout.stride_in_dw = shader->so.stride;
2379}
2380
2381bool si_update_shaders(struct si_context *sctx)
2382{
2383	struct pipe_context *ctx = (struct pipe_context*)sctx;
2384	struct si_compiler_ctx_state compiler_state;
2385	struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
2386	struct si_shader *old_vs = si_get_vs_state(sctx);
2387	bool old_clip_disable = old_vs ? old_vs->key.opt.hw_vs.clip_disable : false;
2388	int r;
2389
2390	compiler_state.tm = sctx->tm;
2391	compiler_state.debug = sctx->b.debug;
2392	compiler_state.is_debug_context = sctx->is_debug;
2393
2394	/* Update stages before GS. */
2395	if (sctx->tes_shader.cso) {
2396		if (!sctx->tf_ring) {
2397			si_init_tess_factor_ring(sctx);
2398			if (!sctx->tf_ring)
2399				return false;
2400		}
2401
2402		/* VS as LS */
2403		r = si_shader_select(ctx, &sctx->vs_shader, &compiler_state);
2404		if (r)
2405			return false;
2406		si_pm4_bind_state(sctx, ls, sctx->vs_shader.current->pm4);
2407
2408		if (sctx->tcs_shader.cso) {
2409			r = si_shader_select(ctx, &sctx->tcs_shader,
2410					     &compiler_state);
2411			if (r)
2412				return false;
2413			si_pm4_bind_state(sctx, hs, sctx->tcs_shader.current->pm4);
2414		} else {
2415			if (!sctx->fixed_func_tcs_shader.cso) {
2416				si_generate_fixed_func_tcs(sctx);
2417				if (!sctx->fixed_func_tcs_shader.cso)
2418					return false;
2419			}
2420
2421			r = si_shader_select(ctx, &sctx->fixed_func_tcs_shader,
2422					     &compiler_state);
2423			if (r)
2424				return false;
2425			si_pm4_bind_state(sctx, hs,
2426					  sctx->fixed_func_tcs_shader.current->pm4);
2427		}
2428
2429		r = si_shader_select(ctx, &sctx->tes_shader, &compiler_state);
2430		if (r)
2431			return false;
2432
2433		if (sctx->gs_shader.cso) {
2434			/* TES as ES */
2435			si_pm4_bind_state(sctx, es, sctx->tes_shader.current->pm4);
2436		} else {
2437			/* TES as VS */
2438			si_pm4_bind_state(sctx, vs, sctx->tes_shader.current->pm4);
2439			si_update_so(sctx, sctx->tes_shader.cso);
2440		}
2441	} else if (sctx->gs_shader.cso) {
2442		/* VS as ES */
2443		r = si_shader_select(ctx, &sctx->vs_shader, &compiler_state);
2444		if (r)
2445			return false;
2446		si_pm4_bind_state(sctx, es, sctx->vs_shader.current->pm4);
2447	} else {
2448		/* VS as VS */
2449		r = si_shader_select(ctx, &sctx->vs_shader, &compiler_state);
2450		if (r)
2451			return false;
2452		si_pm4_bind_state(sctx, vs, sctx->vs_shader.current->pm4);
2453		si_update_so(sctx, sctx->vs_shader.cso);
2454	}
2455
2456	/* Update GS. */
2457	if (sctx->gs_shader.cso) {
2458		r = si_shader_select(ctx, &sctx->gs_shader, &compiler_state);
2459		if (r)
2460			return false;
2461		si_pm4_bind_state(sctx, gs, sctx->gs_shader.current->pm4);
2462		si_pm4_bind_state(sctx, vs, sctx->gs_shader.cso->gs_copy_shader->pm4);
2463		si_update_so(sctx, sctx->gs_shader.cso);
2464
2465		if (!si_update_gs_ring_buffers(sctx))
2466			return false;
2467	} else {
2468		si_pm4_bind_state(sctx, gs, NULL);
2469		si_pm4_bind_state(sctx, es, NULL);
2470	}
2471
2472	si_update_vgt_shader_config(sctx);
2473
2474	if (old_clip_disable != si_get_vs_state(sctx)->key.opt.hw_vs.clip_disable)
2475		si_mark_atom_dirty(sctx, &sctx->clip_regs);
2476
2477	if (sctx->ps_shader.cso) {
2478		unsigned db_shader_control;
2479
2480		r = si_shader_select(ctx, &sctx->ps_shader, &compiler_state);
2481		if (r)
2482			return false;
2483		si_pm4_bind_state(sctx, ps, sctx->ps_shader.current->pm4);
2484
2485		db_shader_control =
2486			sctx->ps_shader.cso->db_shader_control |
2487			S_02880C_KILL_ENABLE(si_get_alpha_test_func(sctx) != PIPE_FUNC_ALWAYS);
2488
2489		if (si_pm4_state_changed(sctx, ps) || si_pm4_state_changed(sctx, vs) ||
2490		    sctx->sprite_coord_enable != rs->sprite_coord_enable ||
2491		    sctx->flatshade != rs->flatshade) {
2492			sctx->sprite_coord_enable = rs->sprite_coord_enable;
2493			sctx->flatshade = rs->flatshade;
2494			si_mark_atom_dirty(sctx, &sctx->spi_map);
2495		}
2496
2497		if (sctx->b.family == CHIP_STONEY && si_pm4_state_changed(sctx, ps))
2498			si_mark_atom_dirty(sctx, &sctx->cb_render_state);
2499
2500		if (sctx->ps_db_shader_control != db_shader_control) {
2501			sctx->ps_db_shader_control = db_shader_control;
2502			si_mark_atom_dirty(sctx, &sctx->db_render_state);
2503		}
2504
2505		if (sctx->smoothing_enabled != sctx->ps_shader.current->key.part.ps.epilog.poly_line_smoothing) {
2506			sctx->smoothing_enabled = sctx->ps_shader.current->key.part.ps.epilog.poly_line_smoothing;
2507			si_mark_atom_dirty(sctx, &sctx->msaa_config);
2508
2509			if (sctx->b.chip_class == SI)
2510				si_mark_atom_dirty(sctx, &sctx->db_render_state);
2511
2512			if (sctx->framebuffer.nr_samples <= 1)
2513				si_mark_atom_dirty(sctx, &sctx->msaa_sample_locs.atom);
2514		}
2515	}
2516
2517	if (si_pm4_state_changed(sctx, ls) ||
2518	    si_pm4_state_changed(sctx, hs) ||
2519	    si_pm4_state_changed(sctx, es) ||
2520	    si_pm4_state_changed(sctx, gs) ||
2521	    si_pm4_state_changed(sctx, vs) ||
2522	    si_pm4_state_changed(sctx, ps)) {
2523		if (!si_update_spi_tmpring_size(sctx))
2524			return false;
2525	}
2526
2527	sctx->do_update_shaders = false;
2528	return true;
2529}
2530
2531void si_init_shader_functions(struct si_context *sctx)
2532{
2533	si_init_atom(sctx, &sctx->spi_map, &sctx->atoms.s.spi_map, si_emit_spi_map);
2534
2535	sctx->b.b.create_vs_state = si_create_shader_selector;
2536	sctx->b.b.create_tcs_state = si_create_shader_selector;
2537	sctx->b.b.create_tes_state = si_create_shader_selector;
2538	sctx->b.b.create_gs_state = si_create_shader_selector;
2539	sctx->b.b.create_fs_state = si_create_shader_selector;
2540
2541	sctx->b.b.bind_vs_state = si_bind_vs_shader;
2542	sctx->b.b.bind_tcs_state = si_bind_tcs_shader;
2543	sctx->b.b.bind_tes_state = si_bind_tes_shader;
2544	sctx->b.b.bind_gs_state = si_bind_gs_shader;
2545	sctx->b.b.bind_fs_state = si_bind_ps_shader;
2546
2547	sctx->b.b.delete_vs_state = si_delete_shader_selector;
2548	sctx->b.b.delete_tcs_state = si_delete_shader_selector;
2549	sctx->b.b.delete_tes_state = si_delete_shader_selector;
2550	sctx->b.b.delete_gs_state = si_delete_shader_selector;
2551	sctx->b.b.delete_fs_state = si_delete_shader_selector;
2552}
2553