projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: fix the VGT performance tweak for small instances
[mesa.git] / src / gallium / drivers / radeonsi / si_state_draw.c
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 */
26
27 #include "si_pipe.h"
28 #include "si_shader.h"
29 #include "radeon/r600_cs.h"
30 #include "sid.h"
31
32 #include "util/u_index_modify.h"
33 #include "util/u_upload_mgr.h"
34 #include "util/u_prim.h"
35 #include "util/u_memory.h"
36
37 static unsigned si_conv_pipe_prim(unsigned mode)
38 {
39 static const unsigned prim_conv[] = {
40 [PIPE_PRIM_POINTS] = V_008958_DI_PT_POINTLIST,
41 [PIPE_PRIM_LINES] = V_008958_DI_PT_LINELIST,
42 [PIPE_PRIM_LINE_LOOP] = V_008958_DI_PT_LINELOOP,
43 [PIPE_PRIM_LINE_STRIP] = V_008958_DI_PT_LINESTRIP,
44 [PIPE_PRIM_TRIANGLES] = V_008958_DI_PT_TRILIST,
45 [PIPE_PRIM_TRIANGLE_STRIP] = V_008958_DI_PT_TRISTRIP,
46 [PIPE_PRIM_TRIANGLE_FAN] = V_008958_DI_PT_TRIFAN,
47 [PIPE_PRIM_QUADS] = V_008958_DI_PT_QUADLIST,
48 [PIPE_PRIM_QUAD_STRIP] = V_008958_DI_PT_QUADSTRIP,
49 [PIPE_PRIM_POLYGON] = V_008958_DI_PT_POLYGON,
50 [PIPE_PRIM_LINES_ADJACENCY] = V_008958_DI_PT_LINELIST_ADJ,
51 [PIPE_PRIM_LINE_STRIP_ADJACENCY] = V_008958_DI_PT_LINESTRIP_ADJ,
52 [PIPE_PRIM_TRIANGLES_ADJACENCY] = V_008958_DI_PT_TRILIST_ADJ,
53 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY] = V_008958_DI_PT_TRISTRIP_ADJ,
54 [PIPE_PRIM_PATCHES] = V_008958_DI_PT_PATCH,
55 [R600_PRIM_RECTANGLE_LIST] = V_008958_DI_PT_RECTLIST
56 };
57 assert(mode < ARRAY_SIZE(prim_conv));
58 return prim_conv[mode];
59 }
60
61 static unsigned si_conv_prim_to_gs_out(unsigned mode)
62 {
63 static const int prim_conv[] = {
64 [PIPE_PRIM_POINTS] = V_028A6C_OUTPRIM_TYPE_POINTLIST,
65 [PIPE_PRIM_LINES] = V_028A6C_OUTPRIM_TYPE_LINESTRIP,
66 [PIPE_PRIM_LINE_LOOP] = V_028A6C_OUTPRIM_TYPE_LINESTRIP,
67 [PIPE_PRIM_LINE_STRIP] = V_028A6C_OUTPRIM_TYPE_LINESTRIP,
68 [PIPE_PRIM_TRIANGLES] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
69 [PIPE_PRIM_TRIANGLE_STRIP] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
70 [PIPE_PRIM_TRIANGLE_FAN] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
71 [PIPE_PRIM_QUADS] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
72 [PIPE_PRIM_QUAD_STRIP] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
73 [PIPE_PRIM_POLYGON] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
74 [PIPE_PRIM_LINES_ADJACENCY] = V_028A6C_OUTPRIM_TYPE_LINESTRIP,
75 [PIPE_PRIM_LINE_STRIP_ADJACENCY] = V_028A6C_OUTPRIM_TYPE_LINESTRIP,
76 [PIPE_PRIM_TRIANGLES_ADJACENCY] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
77 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
78 [PIPE_PRIM_PATCHES] = V_028A6C_OUTPRIM_TYPE_POINTLIST,
79 [R600_PRIM_RECTANGLE_LIST] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
80 };
81 assert(mode < ARRAY_SIZE(prim_conv));
82
83 return prim_conv[mode];
84 }
85
86 /**
87 * This calculates the LDS size for tessellation shaders (VS, TCS, TES).
88 * LS.LDS_SIZE is shared by all 3 shader stages.
89 *
90 * The information about LDS and other non-compile-time parameters is then
91 * written to userdata SGPRs.
92 */
93 static void si_emit_derived_tess_state(struct si_context *sctx,
94 const struct pipe_draw_info *info,
95 unsigned *num_patches)
96 {
97 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
98 struct si_shader_ctx_state *ls = &sctx->vs_shader;
99 /* The TES pointer will only be used for sctx->last_tcs.
100 * It would be wrong to think that TCS = TES. */
101 struct si_shader_selector *tcs =
102 sctx->tcs_shader.cso ? sctx->tcs_shader.cso : sctx->tes_shader.cso;
103 unsigned tes_sh_base = sctx->shader_userdata.sh_base[PIPE_SHADER_TESS_EVAL];
104 unsigned num_tcs_input_cp = info->vertices_per_patch;
105 unsigned num_tcs_output_cp, num_tcs_inputs, num_tcs_outputs;
106 unsigned num_tcs_patch_outputs;
107 unsigned input_vertex_size, output_vertex_size, pervertex_output_patch_size;
108 unsigned input_patch_size, output_patch_size, output_patch0_offset;
109 unsigned perpatch_output_offset, lds_size, ls_rsrc2;
110 unsigned tcs_in_layout, tcs_out_layout, tcs_out_offsets;
111 unsigned offchip_layout, hardware_lds_size;
112
113 /* This calculates how shader inputs and outputs among VS, TCS, and TES
114 * are laid out in LDS. */
115 num_tcs_inputs = util_last_bit64(ls->cso->outputs_written);
116
117 if (sctx->tcs_shader.cso) {
118 num_tcs_outputs = util_last_bit64(tcs->outputs_written);
119 num_tcs_output_cp = tcs->info.properties[TGSI_PROPERTY_TCS_VERTICES_OUT];
120 num_tcs_patch_outputs = util_last_bit64(tcs->patch_outputs_written);
121 } else {
122 /* No TCS. Route varyings from LS to TES. */
123 num_tcs_outputs = num_tcs_inputs;
124 num_tcs_output_cp = num_tcs_input_cp;
125 num_tcs_patch_outputs = 2; /* TESSINNER + TESSOUTER */
126 }
127
128 input_vertex_size = num_tcs_inputs * 16;
129 output_vertex_size = num_tcs_outputs * 16;
130
131 input_patch_size = num_tcs_input_cp * input_vertex_size;
132
133 pervertex_output_patch_size = num_tcs_output_cp * output_vertex_size;
134 output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
135
136 /* Ensure that we only need one wave per SIMD so we don't need to check
137 * resource usage. Also ensures that the number of tcs in and out
138 * vertices per threadgroup are at most 256.
139 */
140 *num_patches = 64 / MAX2(num_tcs_input_cp, num_tcs_output_cp) * 4;
141
142 /* Make sure that the data fits in LDS. This assumes the shaders only
143 * use LDS for the inputs and outputs.
144 */
145 hardware_lds_size = sctx->b.chip_class >= CIK ? 65536 : 32768;
146 *num_patches = MIN2(*num_patches, hardware_lds_size / (input_patch_size +
147 output_patch_size));
148
149 /* Make sure the output data fits in the offchip buffer */
150 *num_patches = MIN2(*num_patches,
151 (sctx->screen->tess_offchip_block_dw_size * 4) /
152 output_patch_size);
153
154 /* Not necessary for correctness, but improves performance. The
155 * specific value is taken from the proprietary driver.
156 */
157 *num_patches = MIN2(*num_patches, 40);
158
159 output_patch0_offset = input_patch_size * *num_patches;
160 perpatch_output_offset = output_patch0_offset + pervertex_output_patch_size;
161
162 lds_size = output_patch0_offset + output_patch_size * *num_patches;
163 ls_rsrc2 = ls->current->config.rsrc2;
164
165 if (sctx->b.chip_class >= CIK) {
166 assert(lds_size <= 65536);
167 ls_rsrc2 |= S_00B52C_LDS_SIZE(align(lds_size, 512) / 512);
168 } else {
169 assert(lds_size <= 32768);
170 ls_rsrc2 |= S_00B52C_LDS_SIZE(align(lds_size, 256) / 256);
171 }
172
173 if (sctx->last_ls == ls->current &&
174 sctx->last_tcs == tcs &&
175 sctx->last_tes_sh_base == tes_sh_base &&
176 sctx->last_num_tcs_input_cp == num_tcs_input_cp)
177 return;
178
179 sctx->last_ls = ls->current;
180 sctx->last_tcs = tcs;
181 sctx->last_tes_sh_base = tes_sh_base;
182 sctx->last_num_tcs_input_cp = num_tcs_input_cp;
183
184 /* Due to a hw bug, RSRC2_LS must be written twice with another
185 * LS register written in between. */
186 if (sctx->b.chip_class == CIK && sctx->b.family != CHIP_HAWAII)
187 radeon_set_sh_reg(cs, R_00B52C_SPI_SHADER_PGM_RSRC2_LS, ls_rsrc2);
188 radeon_set_sh_reg_seq(cs, R_00B528_SPI_SHADER_PGM_RSRC1_LS, 2);
189 radeon_emit(cs, ls->current->config.rsrc1);
190 radeon_emit(cs, ls_rsrc2);
191
192 /* Compute userdata SGPRs. */
193 assert(((input_vertex_size / 4) & ~0xff) == 0);
194 assert(((output_vertex_size / 4) & ~0xff) == 0);
195 assert(((input_patch_size / 4) & ~0x1fff) == 0);
196 assert(((output_patch_size / 4) & ~0x1fff) == 0);
197 assert(((output_patch0_offset / 16) & ~0xffff) == 0);
198 assert(((perpatch_output_offset / 16) & ~0xffff) == 0);
199 assert(num_tcs_input_cp <= 32);
200 assert(num_tcs_output_cp <= 32);
201
202 tcs_in_layout = (input_patch_size / 4) |
203 ((input_vertex_size / 4) << 13);
204 tcs_out_layout = (output_patch_size / 4) |
205 ((output_vertex_size / 4) << 13);
206 tcs_out_offsets = (output_patch0_offset / 16) |
207 ((perpatch_output_offset / 16) << 16);
208 offchip_layout = (pervertex_output_patch_size * *num_patches << 16) |
209 (num_tcs_output_cp << 9) | *num_patches;
210
211 /* Set them for LS. */
212 radeon_set_sh_reg(cs,
213 R_00B530_SPI_SHADER_USER_DATA_LS_0 + SI_SGPR_LS_OUT_LAYOUT * 4,
214 tcs_in_layout);
215
216 /* Set them for TCS. */
217 radeon_set_sh_reg_seq(cs,
218 R_00B430_SPI_SHADER_USER_DATA_HS_0 + SI_SGPR_TCS_OFFCHIP_LAYOUT * 4, 4);
219 radeon_emit(cs, offchip_layout);
220 radeon_emit(cs, tcs_out_offsets);
221 radeon_emit(cs, tcs_out_layout | (num_tcs_input_cp << 26));
222 radeon_emit(cs, tcs_in_layout);
223
224 /* Set them for TES. */
225 radeon_set_sh_reg_seq(cs, tes_sh_base + SI_SGPR_TCS_OFFCHIP_LAYOUT * 4, 1);
226 radeon_emit(cs, offchip_layout);
227 }
228
229 static unsigned si_num_prims_for_vertices(const struct pipe_draw_info *info)
230 {
231 switch (info->mode) {
232 case PIPE_PRIM_PATCHES:
233 return info->count / info->vertices_per_patch;
234 case R600_PRIM_RECTANGLE_LIST:
235 return info->count / 3;
236 default:
237 return u_prims_for_vertices(info->mode, info->count);
238 }
239 }
240
241 static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
242 const struct pipe_draw_info *info,
243 unsigned num_patches)
244 {
245 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
246 unsigned prim = info->mode;
247 unsigned primgroup_size = 128; /* recommended without a GS */
248 unsigned max_primgroup_in_wave = 2;
249
250 /* SWITCH_ON_EOP(0) is always preferable. */
251 bool wd_switch_on_eop = false;
252 bool ia_switch_on_eop = false;
253 bool ia_switch_on_eoi = false;
254 bool partial_vs_wave = false;
255 bool partial_es_wave = false;
256
257 if (sctx->gs_shader.cso)
258 primgroup_size = 64; /* recommended with a GS */
259
260 if (sctx->tes_shader.cso) {
261 /* primgroup_size must be set to a multiple of NUM_PATCHES */
262 primgroup_size = num_patches;
263
264 /* SWITCH_ON_EOI must be set if PrimID is used. */
265 if ((sctx->tcs_shader.cso && sctx->tcs_shader.cso->info.uses_primid) ||
266 sctx->tes_shader.cso->info.uses_primid)
267 ia_switch_on_eoi = true;
268
269 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
270 if ((sctx->b.family == CHIP_TAHITI ||
271 sctx->b.family == CHIP_PITCAIRN ||
272 sctx->b.family == CHIP_BONAIRE) &&
273 sctx->gs_shader.cso)
274 partial_vs_wave = true;
275
276 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
277 if (sctx->screen->has_distributed_tess) {
278 if (sctx->gs_shader.cso)
279 partial_es_wave = true;
280 else
281 partial_vs_wave = true;
282 }
283 }
284
285 /* This is a hardware requirement. */
286 if ((rs && rs->line_stipple_enable) ||
287 (sctx->b.screen->debug_flags & DBG_SWITCH_ON_EOP)) {
288 ia_switch_on_eop = true;
289 wd_switch_on_eop = true;
290 }
291
292 if (sctx->b.chip_class >= CIK) {
293 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
294 * 4 shader engines. Set 1 to pass the assertion below.
295 * The other cases are hardware requirements.
296 *
297 * Polaris supports primitive restart with WD_SWITCH_ON_EOP=0
298 * for points, line strips, and tri strips.
299 */
300 if (sctx->b.screen->info.max_se < 4 ||
301 prim == PIPE_PRIM_POLYGON ||
302 prim == PIPE_PRIM_LINE_LOOP ||
303 prim == PIPE_PRIM_TRIANGLE_FAN ||
304 prim == PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY ||
305 (info->primitive_restart &&
306 (sctx->b.family < CHIP_POLARIS10 ||
307 (prim != PIPE_PRIM_POINTS &&
308 prim != PIPE_PRIM_LINE_STRIP &&
309 prim != PIPE_PRIM_TRIANGLE_STRIP))) ||
310 info->count_from_stream_output)
311 wd_switch_on_eop = true;
312
313 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
314 * We don't know that for indirect drawing, so treat it as
315 * always problematic. */
316 if (sctx->b.family == CHIP_HAWAII &&
317 (info->indirect || info->instance_count > 1))
318 wd_switch_on_eop = true;
319
320 /* Performance recommendation for 4 SE Gfx7-8 parts if
321 * instances are smaller than a primgroup.
322 * Assume indirect draws always use small instances.
323 * This is needed for good VS wave utilization.
324 */
325 if (sctx->b.chip_class <= VI &&
326 sctx->b.screen->info.max_se >= 4 &&
327 (info->indirect ||
328 (info->instance_count > 1 &&
329 si_num_prims_for_vertices(info) < primgroup_size)))
330 wd_switch_on_eop = true;
331
332 /* Required on CIK and later. */
333 if (sctx->b.screen->info.max_se > 2 && !wd_switch_on_eop)
334 ia_switch_on_eoi = true;
335
336 /* Required by Hawaii and, for some special cases, by VI. */
337 if (ia_switch_on_eoi &&
338 (sctx->b.family == CHIP_HAWAII ||
339 (sctx->b.chip_class == VI &&
340 (sctx->gs_shader.cso || max_primgroup_in_wave != 2))))
341 partial_vs_wave = true;
342
343 /* Instancing bug on Bonaire. */
344 if (sctx->b.family == CHIP_BONAIRE && ia_switch_on_eoi &&
345 (info->indirect || info->instance_count > 1))
346 partial_vs_wave = true;
347
348 /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
349 * The hw doc says all multi-SE chips are affected, but Vulkan
350 * only applies it to Hawaii. Do what Vulkan does.
351 */
352 if (sctx->b.family == CHIP_HAWAII &&
353 sctx->gs_shader.cso &&
354 ia_switch_on_eoi &&
355 (info->indirect ||
356 (info->instance_count > 1 &&
357 si_num_prims_for_vertices(info) <= 1)))
358 sctx->b.flags |= SI_CONTEXT_VGT_FLUSH;
359
360
361 /* If the WD switch is false, the IA switch must be false too. */
362 assert(wd_switch_on_eop || !ia_switch_on_eop);
363 }
364
365 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
366 if (ia_switch_on_eoi)
367 partial_es_wave = true;
368
369 /* GS requirement. */
370 if (SI_GS_PER_ES / primgroup_size >= sctx->screen->gs_table_depth - 3)
371 partial_es_wave = true;
372
373 return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop) |
374 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi) |
375 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave) |
376 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave) |
377 S_028AA8_PRIMGROUP_SIZE(primgroup_size - 1) |
378 S_028AA8_WD_SWITCH_ON_EOP(sctx->b.chip_class >= CIK ? wd_switch_on_eop : 0) |
379 S_028AA8_MAX_PRIMGRP_IN_WAVE(sctx->b.chip_class >= VI ?
380 max_primgroup_in_wave : 0);
381 }
382
383 static unsigned si_get_ls_hs_config(struct si_context *sctx,
384 const struct pipe_draw_info *info,
385 unsigned num_patches)
386 {
387 unsigned num_output_cp;
388
389 if (!sctx->tes_shader.cso)
390 return 0;
391
392 num_output_cp = sctx->tcs_shader.cso ?
393 sctx->tcs_shader.cso->info.properties[TGSI_PROPERTY_TCS_VERTICES_OUT] :
394 info->vertices_per_patch;
395
396 return S_028B58_NUM_PATCHES(num_patches) |
397 S_028B58_HS_NUM_INPUT_CP(info->vertices_per_patch) |
398 S_028B58_HS_NUM_OUTPUT_CP(num_output_cp);
399 }
400
401 static void si_emit_scratch_reloc(struct si_context *sctx)
402 {
403 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
404
405 if (!sctx->emit_scratch_reloc)
406 return;
407
408 radeon_set_context_reg(cs, R_0286E8_SPI_TMPRING_SIZE,
409 sctx->spi_tmpring_size);
410
411 if (sctx->scratch_buffer) {
412 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
413 sctx->scratch_buffer, RADEON_USAGE_READWRITE,
414 RADEON_PRIO_SCRATCH_BUFFER);
415
416 }
417 sctx->emit_scratch_reloc = false;
418 }
419
420 /* rast_prim is the primitive type after GS. */
421 static void si_emit_rasterizer_prim_state(struct si_context *sctx)
422 {
423 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
424 unsigned rast_prim = sctx->current_rast_prim;
425 struct si_state_rasterizer *rs = sctx->emitted.named.rasterizer;
426
427 /* Skip this if not rendering lines. */
428 if (rast_prim != PIPE_PRIM_LINES &&
429 rast_prim != PIPE_PRIM_LINE_LOOP &&
430 rast_prim != PIPE_PRIM_LINE_STRIP &&
431 rast_prim != PIPE_PRIM_LINES_ADJACENCY &&
432 rast_prim != PIPE_PRIM_LINE_STRIP_ADJACENCY)
433 return;
434
435 if (rast_prim == sctx->last_rast_prim &&
436 rs->pa_sc_line_stipple == sctx->last_sc_line_stipple)
437 return;
438
439 /* For lines, reset the stipple pattern at each primitive. Otherwise,
440 * reset the stipple pattern at each packet (line strips, line loops).
441 */
442 radeon_set_context_reg(cs, R_028A0C_PA_SC_LINE_STIPPLE,
443 rs->pa_sc_line_stipple |
444 S_028A0C_AUTO_RESET_CNTL(rast_prim == PIPE_PRIM_LINES ? 1 : 2));
445
446 sctx->last_rast_prim = rast_prim;
447 sctx->last_sc_line_stipple = rs->pa_sc_line_stipple;
448 }
449
450 static void si_emit_draw_registers(struct si_context *sctx,
451 const struct pipe_draw_info *info)
452 {
453 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
454 unsigned prim = si_conv_pipe_prim(info->mode);
455 unsigned gs_out_prim = si_conv_prim_to_gs_out(sctx->current_rast_prim);
456 unsigned ia_multi_vgt_param, ls_hs_config, num_patches = 0;
457
458 /* Polaris needs different VTX_REUSE_DEPTH settings depending on
459 * whether the "fractional odd" tessellation spacing is used.
460 */
461 if (sctx->b.family >= CHIP_POLARIS10) {
462 struct si_shader_selector *tes = sctx->tes_shader.cso;
463 unsigned vtx_reuse_depth = 30;
464
465 if (tes &&
466 tes->info.properties[TGSI_PROPERTY_TES_SPACING] ==
467 PIPE_TESS_SPACING_FRACTIONAL_ODD)
468 vtx_reuse_depth = 14;
469
470 if (vtx_reuse_depth != sctx->last_vtx_reuse_depth) {
471 radeon_set_context_reg(cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
472 vtx_reuse_depth);
473 sctx->last_vtx_reuse_depth = vtx_reuse_depth;
474 }
475 }
476
477 if (sctx->tes_shader.cso)
478 si_emit_derived_tess_state(sctx, info, &num_patches);
479
480 ia_multi_vgt_param = si_get_ia_multi_vgt_param(sctx, info, num_patches);
481 ls_hs_config = si_get_ls_hs_config(sctx, info, num_patches);
482
483 /* Draw state. */
484 if (prim != sctx->last_prim ||
485 ia_multi_vgt_param != sctx->last_multi_vgt_param ||
486 ls_hs_config != sctx->last_ls_hs_config) {
487 if (sctx->b.chip_class >= CIK) {
488 radeon_set_context_reg_idx(cs, R_028AA8_IA_MULTI_VGT_PARAM, 1, ia_multi_vgt_param);
489 radeon_set_context_reg_idx(cs, R_028B58_VGT_LS_HS_CONFIG, 2, ls_hs_config);
490 radeon_set_uconfig_reg_idx(cs, R_030908_VGT_PRIMITIVE_TYPE, 1, prim);
491 } else {
492 radeon_set_config_reg(cs, R_008958_VGT_PRIMITIVE_TYPE, prim);
493 radeon_set_context_reg(cs, R_028AA8_IA_MULTI_VGT_PARAM, ia_multi_vgt_param);
494 radeon_set_context_reg(cs, R_028B58_VGT_LS_HS_CONFIG, ls_hs_config);
495 }
496
497 sctx->last_prim = prim;
498 sctx->last_multi_vgt_param = ia_multi_vgt_param;
499 sctx->last_ls_hs_config = ls_hs_config;
500 }
501
502 if (gs_out_prim != sctx->last_gs_out_prim) {
503 radeon_set_context_reg(cs, R_028A6C_VGT_GS_OUT_PRIM_TYPE, gs_out_prim);
504 sctx->last_gs_out_prim = gs_out_prim;
505 }
506
507 /* Primitive restart. */
508 if (info->primitive_restart != sctx->last_primitive_restart_en) {
509 radeon_set_context_reg(cs, R_028A94_VGT_MULTI_PRIM_IB_RESET_EN, info->primitive_restart);
510 sctx->last_primitive_restart_en = info->primitive_restart;
511
512 if (info->primitive_restart &&
513 (info->restart_index != sctx->last_restart_index ||
514 sctx->last_restart_index == SI_RESTART_INDEX_UNKNOWN)) {
515 radeon_set_context_reg(cs, R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX,
516 info->restart_index);
517 sctx->last_restart_index = info->restart_index;
518 }
519 }
520 }
521
522 static void si_emit_draw_packets(struct si_context *sctx,
523 const struct pipe_draw_info *info,
524 const struct pipe_index_buffer *ib)
525 {
526 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
527 unsigned sh_base_reg = sctx->shader_userdata.sh_base[PIPE_SHADER_VERTEX];
528 bool render_cond_bit = sctx->b.render_cond && !sctx->b.render_cond_force_off;
529 uint32_t index_max_size = 0;
530 uint64_t index_va = 0;
531
532 if (info->count_from_stream_output) {
533 struct r600_so_target *t =
534 (struct r600_so_target*)info->count_from_stream_output;
535 uint64_t va = t->buf_filled_size->gpu_address +
536 t->buf_filled_size_offset;
537
538 radeon_set_context_reg(cs, R_028B30_VGT_STRMOUT_DRAW_OPAQUE_VERTEX_STRIDE,
539 t->stride_in_dw);
540
541 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
542 radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
543 COPY_DATA_DST_SEL(COPY_DATA_REG) |
544 COPY_DATA_WR_CONFIRM);
545 radeon_emit(cs, va); /* src address lo */
546 radeon_emit(cs, va >> 32); /* src address hi */
547 radeon_emit(cs, R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE >> 2);
548 radeon_emit(cs, 0); /* unused */
549
550 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
551 t->buf_filled_size, RADEON_USAGE_READ,
552 RADEON_PRIO_SO_FILLED_SIZE);
553 }
554
555 /* draw packet */
556 if (info->indexed) {
557 if (ib->index_size != sctx->last_index_size) {
558 radeon_emit(cs, PKT3(PKT3_INDEX_TYPE, 0, 0));
559
560 /* index type */
561 switch (ib->index_size) {
562 case 1:
563 radeon_emit(cs, V_028A7C_VGT_INDEX_8);
564 break;
565 case 2:
566 radeon_emit(cs, V_028A7C_VGT_INDEX_16 |
567 (SI_BIG_ENDIAN && sctx->b.chip_class <= CIK ?
568 V_028A7C_VGT_DMA_SWAP_16_BIT : 0));
569 break;
570 case 4:
571 radeon_emit(cs, V_028A7C_VGT_INDEX_32 |
572 (SI_BIG_ENDIAN && sctx->b.chip_class <= CIK ?
573 V_028A7C_VGT_DMA_SWAP_32_BIT : 0));
574 break;
575 default:
576 assert(!"unreachable");
577 return;
578 }
579
580 sctx->last_index_size = ib->index_size;
581 }
582
583 index_max_size = (ib->buffer->width0 - ib->offset) /
584 ib->index_size;
585 index_va = r600_resource(ib->buffer)->gpu_address + ib->offset;
586
587 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
588 (struct r600_resource *)ib->buffer,
589 RADEON_USAGE_READ, RADEON_PRIO_INDEX_BUFFER);
590 } else {
591 /* On CI and later, non-indexed draws overwrite VGT_INDEX_TYPE,
592 * so the state must be re-emitted before the next indexed draw.
593 */
594 if (sctx->b.chip_class >= CIK)
595 sctx->last_index_size = -1;
596 }
597
598 if (!info->indirect) {
599 int base_vertex;
600
601 radeon_emit(cs, PKT3(PKT3_NUM_INSTANCES, 0, 0));
602 radeon_emit(cs, info->instance_count);
603
604 /* Base vertex and start instance. */
605 base_vertex = info->indexed ? info->index_bias : info->start;
606
607 if (base_vertex != sctx->last_base_vertex ||
608 sctx->last_base_vertex == SI_BASE_VERTEX_UNKNOWN ||
609 info->start_instance != sctx->last_start_instance ||
610 info->drawid != sctx->last_drawid ||
611 sh_base_reg != sctx->last_sh_base_reg) {
612 radeon_set_sh_reg_seq(cs, sh_base_reg + SI_SGPR_BASE_VERTEX * 4, 3);
613 radeon_emit(cs, base_vertex);
614 radeon_emit(cs, info->start_instance);
615 radeon_emit(cs, info->drawid);
616
617 sctx->last_base_vertex = base_vertex;
618 sctx->last_start_instance = info->start_instance;
619 sctx->last_drawid = info->drawid;
620 sctx->last_sh_base_reg = sh_base_reg;
621 }
622 } else {
623 uint64_t indirect_va = r600_resource(info->indirect)->gpu_address;
624
625 assert(indirect_va % 8 == 0);
626
627 si_invalidate_draw_sh_constants(sctx);
628
629 radeon_emit(cs, PKT3(PKT3_SET_BASE, 2, 0));
630 radeon_emit(cs, 1);
631 radeon_emit(cs, indirect_va);
632 radeon_emit(cs, indirect_va >> 32);
633
634 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
635 (struct r600_resource *)info->indirect,
636 RADEON_USAGE_READ, RADEON_PRIO_DRAW_INDIRECT);
637 }
638
639 if (info->indirect) {
640 unsigned di_src_sel = info->indexed ? V_0287F0_DI_SRC_SEL_DMA
641 : V_0287F0_DI_SRC_SEL_AUTO_INDEX;
642
643 assert(info->indirect_offset % 4 == 0);
644
645 if (info->indexed) {
646 radeon_emit(cs, PKT3(PKT3_INDEX_BASE, 1, 0));
647 radeon_emit(cs, index_va);
648 radeon_emit(cs, index_va >> 32);
649
650 radeon_emit(cs, PKT3(PKT3_INDEX_BUFFER_SIZE, 0, 0));
651 radeon_emit(cs, index_max_size);
652 }
653
654 if (!sctx->screen->has_draw_indirect_multi) {
655 radeon_emit(cs, PKT3(info->indexed ? PKT3_DRAW_INDEX_INDIRECT
656 : PKT3_DRAW_INDIRECT,
657 3, render_cond_bit));
658 radeon_emit(cs, info->indirect_offset);
659 radeon_emit(cs, (sh_base_reg + SI_SGPR_BASE_VERTEX * 4 - SI_SH_REG_OFFSET) >> 2);
660 radeon_emit(cs, (sh_base_reg + SI_SGPR_START_INSTANCE * 4 - SI_SH_REG_OFFSET) >> 2);
661 radeon_emit(cs, di_src_sel);
662 } else {
663 uint64_t count_va = 0;
664
665 if (info->indirect_params) {
666 struct r600_resource *params_buf =
667 (struct r600_resource *)info->indirect_params;
668
669 radeon_add_to_buffer_list(
670 &sctx->b, &sctx->b.gfx, params_buf,
671 RADEON_USAGE_READ, RADEON_PRIO_DRAW_INDIRECT);
672
673 count_va = params_buf->gpu_address + info->indirect_params_offset;
674 }
675
676 radeon_emit(cs, PKT3(info->indexed ? PKT3_DRAW_INDEX_INDIRECT_MULTI :
677 PKT3_DRAW_INDIRECT_MULTI,
678 8, render_cond_bit));
679 radeon_emit(cs, info->indirect_offset);
680 radeon_emit(cs, (sh_base_reg + SI_SGPR_BASE_VERTEX * 4 - SI_SH_REG_OFFSET) >> 2);
681 radeon_emit(cs, (sh_base_reg + SI_SGPR_START_INSTANCE * 4 - SI_SH_REG_OFFSET) >> 2);
682 radeon_emit(cs, ((sh_base_reg + SI_SGPR_DRAWID * 4 - SI_SH_REG_OFFSET) >> 2) |
683 S_2C3_DRAW_INDEX_ENABLE(1) |
684 S_2C3_COUNT_INDIRECT_ENABLE(!!info->indirect_params));
685 radeon_emit(cs, info->indirect_count);
686 radeon_emit(cs, count_va);
687 radeon_emit(cs, count_va >> 32);
688 radeon_emit(cs, info->indirect_stride);
689 radeon_emit(cs, di_src_sel);
690 }
691 } else {
692 if (info->indexed) {
693 index_va += info->start * ib->index_size;
694
695 radeon_emit(cs, PKT3(PKT3_DRAW_INDEX_2, 4, render_cond_bit));
696 radeon_emit(cs, index_max_size);
697 radeon_emit(cs, index_va);
698 radeon_emit(cs, (index_va >> 32UL) & 0xFF);
699 radeon_emit(cs, info->count);
700 radeon_emit(cs, V_0287F0_DI_SRC_SEL_DMA);
701 } else {
702 radeon_emit(cs, PKT3(PKT3_DRAW_INDEX_AUTO, 1, render_cond_bit));
703 radeon_emit(cs, info->count);
704 radeon_emit(cs, V_0287F0_DI_SRC_SEL_AUTO_INDEX |
705 S_0287F0_USE_OPAQUE(!!info->count_from_stream_output));
706 }
707 }
708 }
709
710 void si_emit_cache_flush(struct si_context *sctx)
711 {
712 struct r600_common_context *rctx = &sctx->b;
713 struct radeon_winsys_cs *cs = rctx->gfx.cs;
714 uint32_t cp_coher_cntl = 0;
715
716 /* SI has a bug that it always flushes ICACHE and KCACHE if either
717 * bit is set. An alternative way is to write SQC_CACHES, but that
718 * doesn't seem to work reliably. Since the bug doesn't affect
719 * correctness (it only does more work than necessary) and
720 * the performance impact is likely negligible, there is no plan
721 * to add a workaround for it.
722 */
723
724 if (rctx->flags & SI_CONTEXT_INV_ICACHE)
725 cp_coher_cntl |= S_0085F0_SH_ICACHE_ACTION_ENA(1);
726 if (rctx->flags & SI_CONTEXT_INV_SMEM_L1)
727 cp_coher_cntl |= S_0085F0_SH_KCACHE_ACTION_ENA(1);
728
729 if (rctx->flags & SI_CONTEXT_INV_VMEM_L1)
730 cp_coher_cntl |= S_0085F0_TCL1_ACTION_ENA(1);
731 if (rctx->flags & SI_CONTEXT_INV_GLOBAL_L2) {
732 cp_coher_cntl |= S_0085F0_TC_ACTION_ENA(1);
733
734 if (rctx->chip_class >= VI)
735 cp_coher_cntl |= S_0301F0_TC_WB_ACTION_ENA(1);
736 }
737
738 if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_CB) {
739 cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) |
740 S_0085F0_CB0_DEST_BASE_ENA(1) |
741 S_0085F0_CB1_DEST_BASE_ENA(1) |
742 S_0085F0_CB2_DEST_BASE_ENA(1) |
743 S_0085F0_CB3_DEST_BASE_ENA(1) |
744 S_0085F0_CB4_DEST_BASE_ENA(1) |
745 S_0085F0_CB5_DEST_BASE_ENA(1) |
746 S_0085F0_CB6_DEST_BASE_ENA(1) |
747 S_0085F0_CB7_DEST_BASE_ENA(1);
748
749 /* Necessary for DCC */
750 if (rctx->chip_class >= VI) {
751 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
752 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_DATA_TS) |
753 EVENT_INDEX(5));
754 radeon_emit(cs, 0);
755 radeon_emit(cs, 0);
756 radeon_emit(cs, 0);
757 radeon_emit(cs, 0);
758 }
759 }
760 if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_DB) {
761 cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) |
762 S_0085F0_DB_DEST_BASE_ENA(1);
763 }
764
765 if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_CB_META) {
766 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
767 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
768 /* needed for wait for idle in SURFACE_SYNC */
769 assert(rctx->flags & SI_CONTEXT_FLUSH_AND_INV_CB);
770 }
771 if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_DB_META) {
772 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
773 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
774 /* needed for wait for idle in SURFACE_SYNC */
775 assert(rctx->flags & SI_CONTEXT_FLUSH_AND_INV_DB);
776 }
777
778 /* Wait for shader engines to go idle.
779 * VS and PS waits are unnecessary if SURFACE_SYNC is going to wait
780 * for everything including CB/DB cache flushes.
781 */
782 if (!(rctx->flags & (SI_CONTEXT_FLUSH_AND_INV_CB |
783 SI_CONTEXT_FLUSH_AND_INV_DB))) {
784 if (rctx->flags & SI_CONTEXT_PS_PARTIAL_FLUSH) {
785 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
786 radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
787 /* Only count explicit shader flushes, not implicit ones
788 * done by SURFACE_SYNC.
789 */
790 rctx->num_vs_flushes++;
791 rctx->num_ps_flushes++;
792 } else if (rctx->flags & SI_CONTEXT_VS_PARTIAL_FLUSH) {
793 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
794 radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
795 rctx->num_vs_flushes++;
796 }
797 }
798
799 if (rctx->flags & SI_CONTEXT_CS_PARTIAL_FLUSH &&
800 sctx->compute_is_busy) {
801 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
802 radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH | EVENT_INDEX(4)));
803 rctx->num_cs_flushes++;
804 sctx->compute_is_busy = false;
805 }
806
807 /* VGT state synchronization. */
808 if (rctx->flags & SI_CONTEXT_VGT_FLUSH) {
809 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
810 radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
811 }
812 if (rctx->flags & SI_CONTEXT_VGT_STREAMOUT_SYNC) {
813 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
814 radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC) | EVENT_INDEX(0));
815 }
816
817 /* Make sure ME is idle (it executes most packets) before continuing.
818 * This prevents read-after-write hazards between PFP and ME.
819 */
820 if (cp_coher_cntl || (rctx->flags & SI_CONTEXT_CS_PARTIAL_FLUSH)) {
821 radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
822 radeon_emit(cs, 0);
823 }
824
825 /* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
826 * Therefore, it should be last. Done in PFP.
827 */
828 if (cp_coher_cntl) {
829 /* ACQUIRE_MEM is only required on a compute ring. */
830 radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, 0));
831 radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
832 radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
833 radeon_emit(cs, 0); /* CP_COHER_BASE */
834 radeon_emit(cs, 0x0000000A); /* POLL_INTERVAL */
835 }
836
837 if (rctx->flags & R600_CONTEXT_START_PIPELINE_STATS) {
838 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
839 radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_START) |
840 EVENT_INDEX(0));
841 } else if (rctx->flags & R600_CONTEXT_STOP_PIPELINE_STATS) {
842 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
843 radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP) |
844 EVENT_INDEX(0));
845 }
846
847 rctx->flags = 0;
848 }
849
850 static void si_get_draw_start_count(struct si_context *sctx,
851 const struct pipe_draw_info *info,
852 unsigned *start, unsigned *count)
853 {
854 if (info->indirect) {
855 struct r600_resource *indirect =
856 (struct r600_resource*)info->indirect;
857 int *data = r600_buffer_map_sync_with_rings(&sctx->b,
858 indirect, PIPE_TRANSFER_READ);
859 data += info->indirect_offset/sizeof(int);
860 *start = data[2];
861 *count = data[0];
862 } else {
863 *start = info->start;
864 *count = info->count;
865 }
866 }
867
868 void si_ce_pre_draw_synchronization(struct si_context *sctx)
869 {
870 if (sctx->ce_need_synchronization) {
871 radeon_emit(sctx->ce_ib, PKT3(PKT3_INCREMENT_CE_COUNTER, 0, 0));
872 radeon_emit(sctx->ce_ib, 1);
873
874 radeon_emit(sctx->b.gfx.cs, PKT3(PKT3_WAIT_ON_CE_COUNTER, 0, 0));
875 radeon_emit(sctx->b.gfx.cs, 1);
876 }
877 }
878
879 void si_ce_post_draw_synchronization(struct si_context *sctx)
880 {
881 if (sctx->ce_need_synchronization) {
882 radeon_emit(sctx->b.gfx.cs, PKT3(PKT3_INCREMENT_DE_COUNTER, 0, 0));
883 radeon_emit(sctx->b.gfx.cs, 0);
884
885 sctx->ce_need_synchronization = false;
886 }
887 }
888
889 void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
890 {
891 struct si_context *sctx = (struct si_context *)ctx;
892 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
893 struct pipe_index_buffer ib = {};
894 unsigned mask, dirty_fb_counter, dirty_tex_counter, rast_prim;
895
896 if (likely(!info->indirect)) {
897 /* SI-CI treat instance_count==0 as instance_count==1. There is
898 * no workaround for indirect draws, but we can at least skip
899 * direct draws.
900 */
901 if (unlikely(!info->instance_count))
902 return;
903
904 /* Handle count == 0. */
905 if (unlikely(!info->count &&
906 (info->indexed || !info->count_from_stream_output)))
907 return;
908 }
909
910 if (unlikely(!sctx->vs_shader.cso)) {
911 assert(0);
912 return;
913 }
914 if (unlikely(!sctx->ps_shader.cso && (!rs || !rs->rasterizer_discard))) {
915 assert(0);
916 return;
917 }
918 if (unlikely(!!sctx->tes_shader.cso != (info->mode == PIPE_PRIM_PATCHES))) {
919 assert(0);
920 return;
921 }
922
923 /* Re-emit the framebuffer state if needed. */
924 dirty_fb_counter = p_atomic_read(&sctx->b.screen->dirty_fb_counter);
925 if (unlikely(dirty_fb_counter != sctx->b.last_dirty_fb_counter)) {
926 sctx->b.last_dirty_fb_counter = dirty_fb_counter;
927 sctx->framebuffer.dirty_cbufs |=
928 ((1 << sctx->framebuffer.state.nr_cbufs) - 1);
929 sctx->framebuffer.dirty_zsbuf = true;
930 si_mark_atom_dirty(sctx, &sctx->framebuffer.atom);
931 }
932
933 /* Invalidate & recompute texture descriptors if needed. */
934 dirty_tex_counter = p_atomic_read(&sctx->b.screen->dirty_tex_descriptor_counter);
935 if (unlikely(dirty_tex_counter != sctx->b.last_dirty_tex_descriptor_counter)) {
936 sctx->b.last_dirty_tex_descriptor_counter = dirty_tex_counter;
937 si_update_all_texture_descriptors(sctx);
938 }
939
940 si_decompress_graphics_textures(sctx);
941
942 /* Set the rasterization primitive type.
943 *
944 * This must be done after si_decompress_textures, which can call
945 * draw_vbo recursively, and before si_update_shaders, which uses
946 * current_rast_prim for this draw_vbo call. */
947 if (sctx->gs_shader.cso)
948 rast_prim = sctx->gs_shader.cso->gs_output_prim;
949 else if (sctx->tes_shader.cso)
950 rast_prim = sctx->tes_shader.cso->info.properties[TGSI_PROPERTY_TES_PRIM_MODE];
951 else
952 rast_prim = info->mode;
953
954 if (rast_prim != sctx->current_rast_prim) {
955 sctx->current_rast_prim = rast_prim;
956 sctx->do_update_shaders = true;
957 }
958
959 if (sctx->do_update_shaders && !si_update_shaders(sctx))
960 return;
961
962 if (!si_upload_graphics_shader_descriptors(sctx))
963 return;
964
965 if (info->indexed) {
966 /* Initialize the index buffer struct. */
967 pipe_resource_reference(&ib.buffer, sctx->index_buffer.buffer);
968 ib.user_buffer = sctx->index_buffer.user_buffer;
969 ib.index_size = sctx->index_buffer.index_size;
970 ib.offset = sctx->index_buffer.offset;
971
972 /* Translate or upload, if needed. */
973 /* 8-bit indices are supported on VI. */
974 if (sctx->b.chip_class <= CIK && ib.index_size == 1) {
975 struct pipe_resource *out_buffer = NULL;
976 unsigned out_offset, start, count, start_offset;
977 void *ptr;
978
979 si_get_draw_start_count(sctx, info, &start, &count);
980 start_offset = start * ib.index_size;
981
982 u_upload_alloc(sctx->b.uploader, start_offset, count * 2, 256,
983 &out_offset, &out_buffer, &ptr);
984 if (!out_buffer) {
985 pipe_resource_reference(&ib.buffer, NULL);
986 return;
987 }
988
989 util_shorten_ubyte_elts_to_userptr(&sctx->b.b, &ib, 0,
990 ib.offset + start_offset,
991 count, ptr);
992
993 pipe_resource_reference(&ib.buffer, NULL);
994 ib.user_buffer = NULL;
995 ib.buffer = out_buffer;
996 /* info->start will be added by the drawing code */
997 ib.offset = out_offset - start_offset;
998 ib.index_size = 2;
999 } else if (ib.user_buffer && !ib.buffer) {
1000 unsigned start, count, start_offset;
1001
1002 si_get_draw_start_count(sctx, info, &start, &count);
1003 start_offset = start * ib.index_size;
1004
1005 u_upload_data(sctx->b.uploader, start_offset, count * ib.index_size,
1006 256, (char*)ib.user_buffer + start_offset,
1007 &ib.offset, &ib.buffer);
1008 if (!ib.buffer)
1009 return;
1010 /* info->start will be added by the drawing code */
1011 ib.offset -= start_offset;
1012 }
1013 }
1014
1015 /* VI reads index buffers through TC L2. */
1016 if (info->indexed && sctx->b.chip_class <= CIK &&
1017 r600_resource(ib.buffer)->TC_L2_dirty) {
1018 sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
1019 r600_resource(ib.buffer)->TC_L2_dirty = false;
1020 }
1021
1022 if (info->indirect && r600_resource(info->indirect)->TC_L2_dirty) {
1023 sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
1024 r600_resource(info->indirect)->TC_L2_dirty = false;
1025 }
1026
1027 if (info->indirect_params &&
1028 r600_resource(info->indirect_params)->TC_L2_dirty) {
1029 sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
1030 r600_resource(info->indirect_params)->TC_L2_dirty = false;
1031 }
1032
1033 /* Add buffer sizes for memory checking in need_cs_space. */
1034 if (sctx->emit_scratch_reloc && sctx->scratch_buffer)
1035 r600_context_add_resource_size(ctx, &sctx->scratch_buffer->b.b);
1036 if (info->indirect)
1037 r600_context_add_resource_size(ctx, info->indirect);
1038
1039 si_need_cs_space(sctx);
1040
1041 /* Since we've called r600_context_add_resource_size for vertex buffers,
1042 * this must be called after si_need_cs_space, because we must let
1043 * need_cs_space flush before we add buffers to the buffer list.
1044 */
1045 if (!si_upload_vertex_buffer_descriptors(sctx))
1046 return;
1047
1048 /* Flushed caches prior to emitting states. */
1049 if (sctx->b.flags)
1050 si_emit_cache_flush(sctx);
1051
1052 /* Emit states. */
1053 mask = sctx->dirty_atoms;
1054 while (mask) {
1055 struct r600_atom *atom = sctx->atoms.array[u_bit_scan(&mask)];
1056
1057 atom->emit(&sctx->b, atom);
1058 }
1059 sctx->dirty_atoms = 0;
1060
1061 si_pm4_emit_dirty(sctx);
1062 si_emit_scratch_reloc(sctx);
1063 si_emit_rasterizer_prim_state(sctx);
1064 si_emit_draw_registers(sctx, info);
1065
1066 si_ce_pre_draw_synchronization(sctx);
1067
1068 si_emit_draw_packets(sctx, info, &ib);
1069
1070 si_ce_post_draw_synchronization(sctx);
1071
1072 if (sctx->trace_buf)
1073 si_trace_emit(sctx);
1074
1075 /* Workaround for a VGT hang when streamout is enabled.
1076 * It must be done after drawing. */
1077 if ((sctx->b.family == CHIP_HAWAII ||
1078 sctx->b.family == CHIP_TONGA ||
1079 sctx->b.family == CHIP_FIJI) &&
1080 r600_get_strmout_en(&sctx->b)) {
1081 sctx->b.flags |= SI_CONTEXT_VGT_STREAMOUT_SYNC;
1082 }
1083
1084 /* Set the depth buffer as dirty. */
1085 if (sctx->framebuffer.state.zsbuf) {
1086 struct pipe_surface *surf = sctx->framebuffer.state.zsbuf;
1087 struct r600_texture *rtex = (struct r600_texture *)surf->texture;
1088
1089 rtex->dirty_level_mask |= 1 << surf->u.tex.level;
1090
1091 if (rtex->surface.flags & RADEON_SURF_SBUFFER)
1092 rtex->stencil_dirty_level_mask |= 1 << surf->u.tex.level;
1093 }
1094 if (sctx->framebuffer.compressed_cb_mask) {
1095 struct pipe_surface *surf;
1096 struct r600_texture *rtex;
1097 unsigned mask = sctx->framebuffer.compressed_cb_mask;
1098
1099 do {
1100 unsigned i = u_bit_scan(&mask);
1101 surf = sctx->framebuffer.state.cbufs[i];
1102 rtex = (struct r600_texture*)surf->texture;
1103
1104 if (rtex->fmask.size)
1105 rtex->dirty_level_mask |= 1 << surf->u.tex.level;
1106 if (rtex->dcc_gather_statistics)
1107 rtex->separate_dcc_dirty = true;
1108 } while (mask);
1109 }
1110
1111 pipe_resource_reference(&ib.buffer, NULL);
1112 sctx->b.num_draw_calls++;
1113 if (G_0286E8_WAVESIZE(sctx->spi_tmpring_size))
1114 sctx->b.num_spill_draw_calls++;
1115 }
1116
1117 void si_trace_emit(struct si_context *sctx)
1118 {
1119 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
1120
1121 sctx->trace_id++;
1122 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, sctx->trace_buf,
1123 RADEON_USAGE_READWRITE, RADEON_PRIO_TRACE);
1124 radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 3, 0));
1125 radeon_emit(cs, S_370_DST_SEL(V_370_MEMORY_SYNC) |
1126 S_370_WR_CONFIRM(1) |
1127 S_370_ENGINE_SEL(V_370_ME));
1128 radeon_emit(cs, sctx->trace_buf->gpu_address);
1129 radeon_emit(cs, sctx->trace_buf->gpu_address >> 32);
1130 radeon_emit(cs, sctx->trace_id);
1131 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
1132 radeon_emit(cs, SI_ENCODE_TRACE_POINT(sctx->trace_id));
1133 }