projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: apply the streamout workaround to Fiji as well
[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
36 static void si_decompress_textures(struct si_context *sctx)
37 {
38 if (!sctx->blitter->running) {
39 /* Flush depth textures which need to be flushed. */
40 for (int i = 0; i < SI_NUM_SHADERS; i++) {
41 if (sctx->samplers[i].depth_texture_mask) {
42 si_flush_depth_textures(sctx, &sctx->samplers[i]);
43 }
44 if (sctx->samplers[i].compressed_colortex_mask) {
45 si_decompress_color_textures(sctx, &sctx->samplers[i]);
46 }
47 }
48 }
49 }
50
51 static unsigned si_conv_pipe_prim(unsigned mode)
52 {
53 static const unsigned prim_conv[] = {
54 [PIPE_PRIM_POINTS] = V_008958_DI_PT_POINTLIST,
55 [PIPE_PRIM_LINES] = V_008958_DI_PT_LINELIST,
56 [PIPE_PRIM_LINE_LOOP] = V_008958_DI_PT_LINELOOP,
57 [PIPE_PRIM_LINE_STRIP] = V_008958_DI_PT_LINESTRIP,
58 [PIPE_PRIM_TRIANGLES] = V_008958_DI_PT_TRILIST,
59 [PIPE_PRIM_TRIANGLE_STRIP] = V_008958_DI_PT_TRISTRIP,
60 [PIPE_PRIM_TRIANGLE_FAN] = V_008958_DI_PT_TRIFAN,
61 [PIPE_PRIM_QUADS] = V_008958_DI_PT_QUADLIST,
62 [PIPE_PRIM_QUAD_STRIP] = V_008958_DI_PT_QUADSTRIP,
63 [PIPE_PRIM_POLYGON] = V_008958_DI_PT_POLYGON,
64 [PIPE_PRIM_LINES_ADJACENCY] = V_008958_DI_PT_LINELIST_ADJ,
65 [PIPE_PRIM_LINE_STRIP_ADJACENCY] = V_008958_DI_PT_LINESTRIP_ADJ,
66 [PIPE_PRIM_TRIANGLES_ADJACENCY] = V_008958_DI_PT_TRILIST_ADJ,
67 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY] = V_008958_DI_PT_TRISTRIP_ADJ,
68 [PIPE_PRIM_PATCHES] = V_008958_DI_PT_PATCH,
69 [R600_PRIM_RECTANGLE_LIST] = V_008958_DI_PT_RECTLIST
70 };
71 assert(mode < Elements(prim_conv));
72 return prim_conv[mode];
73 }
74
75 static unsigned si_conv_prim_to_gs_out(unsigned mode)
76 {
77 static const int prim_conv[] = {
78 [PIPE_PRIM_POINTS] = V_028A6C_OUTPRIM_TYPE_POINTLIST,
79 [PIPE_PRIM_LINES] = V_028A6C_OUTPRIM_TYPE_LINESTRIP,
80 [PIPE_PRIM_LINE_LOOP] = V_028A6C_OUTPRIM_TYPE_LINESTRIP,
81 [PIPE_PRIM_LINE_STRIP] = V_028A6C_OUTPRIM_TYPE_LINESTRIP,
82 [PIPE_PRIM_TRIANGLES] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
83 [PIPE_PRIM_TRIANGLE_STRIP] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
84 [PIPE_PRIM_TRIANGLE_FAN] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
85 [PIPE_PRIM_QUADS] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
86 [PIPE_PRIM_QUAD_STRIP] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
87 [PIPE_PRIM_POLYGON] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
88 [PIPE_PRIM_LINES_ADJACENCY] = V_028A6C_OUTPRIM_TYPE_LINESTRIP,
89 [PIPE_PRIM_LINE_STRIP_ADJACENCY] = V_028A6C_OUTPRIM_TYPE_LINESTRIP,
90 [PIPE_PRIM_TRIANGLES_ADJACENCY] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
91 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY] = V_028A6C_OUTPRIM_TYPE_TRISTRIP,
92 [PIPE_PRIM_PATCHES] = V_028A6C_OUTPRIM_TYPE_POINTLIST,
93 [R600_PRIM_RECTANGLE_LIST] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
94 };
95 assert(mode < Elements(prim_conv));
96
97 return prim_conv[mode];
98 }
99
100 /**
101 * This calculates the LDS size for tessellation shaders (VS, TCS, TES).
102 * LS.LDS_SIZE is shared by all 3 shader stages.
103 *
104 * The information about LDS and other non-compile-time parameters is then
105 * written to userdata SGPRs.
106 */
107 static void si_emit_derived_tess_state(struct si_context *sctx,
108 const struct pipe_draw_info *info,
109 unsigned *num_patches)
110 {
111 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
112 struct si_shader_ctx_state *ls = &sctx->vs_shader;
113 /* The TES pointer will only be used for sctx->last_tcs.
114 * It would be wrong to think that TCS = TES. */
115 struct si_shader_selector *tcs =
116 sctx->tcs_shader.cso ? sctx->tcs_shader.cso : sctx->tes_shader.cso;
117 unsigned tes_sh_base = sctx->shader_userdata.sh_base[PIPE_SHADER_TESS_EVAL];
118 unsigned num_tcs_input_cp = info->vertices_per_patch;
119 unsigned num_tcs_output_cp, num_tcs_inputs, num_tcs_outputs;
120 unsigned num_tcs_patch_outputs;
121 unsigned input_vertex_size, output_vertex_size, pervertex_output_patch_size;
122 unsigned input_patch_size, output_patch_size, output_patch0_offset;
123 unsigned perpatch_output_offset, lds_size, ls_rsrc2;
124 unsigned tcs_in_layout, tcs_out_layout, tcs_out_offsets;
125
126 *num_patches = 1; /* TODO: calculate this */
127
128 if (sctx->last_ls == ls->current &&
129 sctx->last_tcs == tcs &&
130 sctx->last_tes_sh_base == tes_sh_base &&
131 sctx->last_num_tcs_input_cp == num_tcs_input_cp)
132 return;
133
134 sctx->last_ls = ls->current;
135 sctx->last_tcs = tcs;
136 sctx->last_tes_sh_base = tes_sh_base;
137 sctx->last_num_tcs_input_cp = num_tcs_input_cp;
138
139 /* This calculates how shader inputs and outputs among VS, TCS, and TES
140 * are laid out in LDS. */
141 num_tcs_inputs = util_last_bit64(ls->cso->outputs_written);
142
143 if (sctx->tcs_shader.cso) {
144 num_tcs_outputs = util_last_bit64(tcs->outputs_written);
145 num_tcs_output_cp = tcs->info.properties[TGSI_PROPERTY_TCS_VERTICES_OUT];
146 num_tcs_patch_outputs = util_last_bit64(tcs->patch_outputs_written);
147 } else {
148 /* No TCS. Route varyings from LS to TES. */
149 num_tcs_outputs = num_tcs_inputs;
150 num_tcs_output_cp = num_tcs_input_cp;
151 num_tcs_patch_outputs = 2; /* TESSINNER + TESSOUTER */
152 }
153
154 input_vertex_size = num_tcs_inputs * 16;
155 output_vertex_size = num_tcs_outputs * 16;
156
157 input_patch_size = num_tcs_input_cp * input_vertex_size;
158
159 pervertex_output_patch_size = num_tcs_output_cp * output_vertex_size;
160 output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
161
162 output_patch0_offset = sctx->tcs_shader.cso ? input_patch_size * *num_patches : 0;
163 perpatch_output_offset = output_patch0_offset + pervertex_output_patch_size;
164
165 lds_size = output_patch0_offset + output_patch_size * *num_patches;
166 ls_rsrc2 = ls->current->rsrc2;
167
168 if (sctx->b.chip_class >= CIK) {
169 assert(lds_size <= 65536);
170 ls_rsrc2 |= S_00B52C_LDS_SIZE(align(lds_size, 512) / 512);
171 } else {
172 assert(lds_size <= 32768);
173 ls_rsrc2 |= S_00B52C_LDS_SIZE(align(lds_size, 256) / 256);
174 }
175
176 /* Due to a hw bug, RSRC2_LS must be written twice with another
177 * LS register written in between. */
178 if (sctx->b.chip_class == CIK && sctx->b.family != CHIP_HAWAII)
179 radeon_set_sh_reg(cs, R_00B52C_SPI_SHADER_PGM_RSRC2_LS, ls_rsrc2);
180 radeon_set_sh_reg_seq(cs, R_00B528_SPI_SHADER_PGM_RSRC1_LS, 2);
181 radeon_emit(cs, ls->current->rsrc1);
182 radeon_emit(cs, ls_rsrc2);
183
184 /* Compute userdata SGPRs. */
185 assert(((input_vertex_size / 4) & ~0xff) == 0);
186 assert(((output_vertex_size / 4) & ~0xff) == 0);
187 assert(((input_patch_size / 4) & ~0x1fff) == 0);
188 assert(((output_patch_size / 4) & ~0x1fff) == 0);
189 assert(((output_patch0_offset / 16) & ~0xffff) == 0);
190 assert(((perpatch_output_offset / 16) & ~0xffff) == 0);
191 assert(num_tcs_input_cp <= 32);
192 assert(num_tcs_output_cp <= 32);
193
194 tcs_in_layout = (input_patch_size / 4) |
195 ((input_vertex_size / 4) << 13);
196 tcs_out_layout = (output_patch_size / 4) |
197 ((output_vertex_size / 4) << 13);
198 tcs_out_offsets = (output_patch0_offset / 16) |
199 ((perpatch_output_offset / 16) << 16);
200
201 /* Set them for LS. */
202 radeon_set_sh_reg(cs,
203 R_00B530_SPI_SHADER_USER_DATA_LS_0 + SI_SGPR_LS_OUT_LAYOUT * 4,
204 tcs_in_layout);
205
206 /* Set them for TCS. */
207 radeon_set_sh_reg_seq(cs,
208 R_00B430_SPI_SHADER_USER_DATA_HS_0 + SI_SGPR_TCS_OUT_OFFSETS * 4, 3);
209 radeon_emit(cs, tcs_out_offsets);
210 radeon_emit(cs, tcs_out_layout | (num_tcs_input_cp << 26));
211 radeon_emit(cs, tcs_in_layout);
212
213 /* Set them for TES. */
214 radeon_set_sh_reg_seq(cs, tes_sh_base + SI_SGPR_TCS_OUT_OFFSETS * 4, 2);
215 radeon_emit(cs, tcs_out_offsets);
216 radeon_emit(cs, tcs_out_layout | (num_tcs_output_cp << 26));
217 }
218
219 static unsigned si_num_prims_for_vertices(const struct pipe_draw_info *info)
220 {
221 switch (info->mode) {
222 case PIPE_PRIM_PATCHES:
223 return info->count / info->vertices_per_patch;
224 case R600_PRIM_RECTANGLE_LIST:
225 return info->count / 3;
226 default:
227 return u_prims_for_vertices(info->mode, info->count);
228 }
229 }
230
231 static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
232 const struct pipe_draw_info *info,
233 unsigned num_patches)
234 {
235 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
236 unsigned prim = info->mode;
237 unsigned primgroup_size = 128; /* recommended without a GS */
238 unsigned max_primgroup_in_wave = 2;
239
240 /* SWITCH_ON_EOP(0) is always preferable. */
241 bool wd_switch_on_eop = false;
242 bool ia_switch_on_eop = false;
243 bool ia_switch_on_eoi = false;
244 bool partial_vs_wave = false;
245 bool partial_es_wave = false;
246
247 if (sctx->gs_shader.cso)
248 primgroup_size = 64; /* recommended with a GS */
249
250 if (sctx->tes_shader.cso) {
251 unsigned num_cp_out =
252 sctx->tcs_shader.cso ?
253 sctx->tcs_shader.cso->info.properties[TGSI_PROPERTY_TCS_VERTICES_OUT] :
254 info->vertices_per_patch;
255 unsigned max_size = 256 / MAX2(info->vertices_per_patch, num_cp_out);
256
257 primgroup_size = MIN2(primgroup_size, max_size);
258
259 /* primgroup_size must be set to a multiple of NUM_PATCHES */
260 primgroup_size = (primgroup_size / num_patches) * num_patches;
261
262 /* SWITCH_ON_EOI must be set if PrimID is used. */
263 if ((sctx->tcs_shader.cso && sctx->tcs_shader.cso->info.uses_primid) ||
264 sctx->tes_shader.cso->info.uses_primid)
265 ia_switch_on_eoi = true;
266
267 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
268 if ((sctx->b.family == CHIP_TAHITI ||
269 sctx->b.family == CHIP_PITCAIRN ||
270 sctx->b.family == CHIP_BONAIRE) &&
271 sctx->gs_shader.cso)
272 partial_vs_wave = true;
273 }
274
275 /* This is a hardware requirement. */
276 if ((rs && rs->line_stipple_enable) ||
277 (sctx->b.screen->debug_flags & DBG_SWITCH_ON_EOP)) {
278 ia_switch_on_eop = true;
279 wd_switch_on_eop = true;
280 }
281
282 if (sctx->b.chip_class >= CIK) {
283 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
284 * 4 shader engines. Set 1 to pass the assertion below.
285 * The other cases are hardware requirements. */
286 if (sctx->b.screen->info.max_se < 4 ||
287 prim == PIPE_PRIM_POLYGON ||
288 prim == PIPE_PRIM_LINE_LOOP ||
289 prim == PIPE_PRIM_TRIANGLE_FAN ||
290 prim == PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY ||
291 info->primitive_restart ||
292 info->count_from_stream_output)
293 wd_switch_on_eop = true;
294
295 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
296 * We don't know that for indirect drawing, so treat it as
297 * always problematic. */
298 if (sctx->b.family == CHIP_HAWAII &&
299 (info->indirect || info->instance_count > 1))
300 wd_switch_on_eop = true;
301
302 /* Required on CIK and later. */
303 if (sctx->b.screen->info.max_se > 2 && !wd_switch_on_eop)
304 ia_switch_on_eoi = true;
305
306 /* Required by Hawaii and, for some special cases, by VI. */
307 if (ia_switch_on_eoi &&
308 (sctx->b.family == CHIP_HAWAII ||
309 (sctx->b.chip_class == VI &&
310 (sctx->gs_shader.cso || max_primgroup_in_wave != 2))))
311 partial_vs_wave = true;
312
313 /* Instancing bug on Bonaire. */
314 if (sctx->b.family == CHIP_BONAIRE && ia_switch_on_eoi &&
315 (info->indirect || info->instance_count > 1))
316 partial_vs_wave = true;
317
318 /* If the WD switch is false, the IA switch must be false too. */
319 assert(wd_switch_on_eop || !ia_switch_on_eop);
320 }
321
322 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
323 if (ia_switch_on_eoi)
324 partial_es_wave = true;
325
326 /* GS requirement. */
327 if (SI_GS_PER_ES / primgroup_size >= sctx->screen->gs_table_depth - 3)
328 partial_es_wave = true;
329
330 /* Hw bug with single-primitive instances and SWITCH_ON_EOI
331 * on multi-SE chips. */
332 if (sctx->b.screen->info.max_se >= 2 && ia_switch_on_eoi &&
333 (info->indirect ||
334 (info->instance_count > 1 &&
335 si_num_prims_for_vertices(info) <= 1)))
336 sctx->b.flags |= SI_CONTEXT_VGT_FLUSH;
337
338 return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop) |
339 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi) |
340 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave) |
341 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave) |
342 S_028AA8_PRIMGROUP_SIZE(primgroup_size - 1) |
343 S_028AA8_WD_SWITCH_ON_EOP(sctx->b.chip_class >= CIK ? wd_switch_on_eop : 0) |
344 S_028AA8_MAX_PRIMGRP_IN_WAVE(sctx->b.chip_class >= VI ?
345 max_primgroup_in_wave : 0);
346 }
347
348 static unsigned si_get_ls_hs_config(struct si_context *sctx,
349 const struct pipe_draw_info *info,
350 unsigned num_patches)
351 {
352 unsigned num_output_cp;
353
354 if (!sctx->tes_shader.cso)
355 return 0;
356
357 num_output_cp = sctx->tcs_shader.cso ?
358 sctx->tcs_shader.cso->info.properties[TGSI_PROPERTY_TCS_VERTICES_OUT] :
359 info->vertices_per_patch;
360
361 return S_028B58_NUM_PATCHES(num_patches) |
362 S_028B58_HS_NUM_INPUT_CP(info->vertices_per_patch) |
363 S_028B58_HS_NUM_OUTPUT_CP(num_output_cp);
364 }
365
366 static void si_emit_scratch_reloc(struct si_context *sctx)
367 {
368 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
369
370 if (!sctx->emit_scratch_reloc)
371 return;
372
373 radeon_set_context_reg(cs, R_0286E8_SPI_TMPRING_SIZE,
374 sctx->spi_tmpring_size);
375
376 if (sctx->scratch_buffer) {
377 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
378 sctx->scratch_buffer, RADEON_USAGE_READWRITE,
379 RADEON_PRIO_SCRATCH_BUFFER);
380
381 }
382 sctx->emit_scratch_reloc = false;
383 }
384
385 /* rast_prim is the primitive type after GS. */
386 static void si_emit_rasterizer_prim_state(struct si_context *sctx)
387 {
388 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
389 unsigned rast_prim = sctx->current_rast_prim;
390 struct si_state_rasterizer *rs = sctx->emitted.named.rasterizer;
391
392 /* Skip this if not rendering lines. */
393 if (rast_prim != PIPE_PRIM_LINES &&
394 rast_prim != PIPE_PRIM_LINE_LOOP &&
395 rast_prim != PIPE_PRIM_LINE_STRIP &&
396 rast_prim != PIPE_PRIM_LINES_ADJACENCY &&
397 rast_prim != PIPE_PRIM_LINE_STRIP_ADJACENCY)
398 return;
399
400 if (rast_prim == sctx->last_rast_prim &&
401 rs->pa_sc_line_stipple == sctx->last_sc_line_stipple)
402 return;
403
404 radeon_set_context_reg(cs, R_028A0C_PA_SC_LINE_STIPPLE,
405 rs->pa_sc_line_stipple |
406 S_028A0C_AUTO_RESET_CNTL(rast_prim == PIPE_PRIM_LINES ? 1 :
407 rast_prim == PIPE_PRIM_LINE_STRIP ? 2 : 0));
408
409 sctx->last_rast_prim = rast_prim;
410 sctx->last_sc_line_stipple = rs->pa_sc_line_stipple;
411 }
412
413 static void si_emit_draw_registers(struct si_context *sctx,
414 const struct pipe_draw_info *info)
415 {
416 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
417 unsigned prim = si_conv_pipe_prim(info->mode);
418 unsigned gs_out_prim = si_conv_prim_to_gs_out(sctx->current_rast_prim);
419 unsigned ia_multi_vgt_param, ls_hs_config, num_patches = 0;
420
421 if (sctx->tes_shader.cso)
422 si_emit_derived_tess_state(sctx, info, &num_patches);
423
424 ia_multi_vgt_param = si_get_ia_multi_vgt_param(sctx, info, num_patches);
425 ls_hs_config = si_get_ls_hs_config(sctx, info, num_patches);
426
427 /* Draw state. */
428 if (prim != sctx->last_prim ||
429 ia_multi_vgt_param != sctx->last_multi_vgt_param ||
430 ls_hs_config != sctx->last_ls_hs_config) {
431 if (sctx->b.chip_class >= CIK) {
432 radeon_emit(cs, PKT3(PKT3_DRAW_PREAMBLE, 2, 0));
433 radeon_emit(cs, prim); /* VGT_PRIMITIVE_TYPE */
434 radeon_emit(cs, ia_multi_vgt_param); /* IA_MULTI_VGT_PARAM */
435 radeon_emit(cs, ls_hs_config); /* VGT_LS_HS_CONFIG */
436 } else {
437 radeon_set_config_reg(cs, R_008958_VGT_PRIMITIVE_TYPE, prim);
438 radeon_set_context_reg(cs, R_028AA8_IA_MULTI_VGT_PARAM, ia_multi_vgt_param);
439 radeon_set_context_reg(cs, R_028B58_VGT_LS_HS_CONFIG, ls_hs_config);
440 }
441 sctx->last_prim = prim;
442 sctx->last_multi_vgt_param = ia_multi_vgt_param;
443 sctx->last_ls_hs_config = ls_hs_config;
444 }
445
446 if (gs_out_prim != sctx->last_gs_out_prim) {
447 radeon_set_context_reg(cs, R_028A6C_VGT_GS_OUT_PRIM_TYPE, gs_out_prim);
448 sctx->last_gs_out_prim = gs_out_prim;
449 }
450
451 /* Primitive restart. */
452 if (info->primitive_restart != sctx->last_primitive_restart_en) {
453 radeon_set_context_reg(cs, R_028A94_VGT_MULTI_PRIM_IB_RESET_EN, info->primitive_restart);
454 sctx->last_primitive_restart_en = info->primitive_restart;
455
456 if (info->primitive_restart &&
457 (info->restart_index != sctx->last_restart_index ||
458 sctx->last_restart_index == SI_RESTART_INDEX_UNKNOWN)) {
459 radeon_set_context_reg(cs, R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX,
460 info->restart_index);
461 sctx->last_restart_index = info->restart_index;
462 }
463 }
464 }
465
466 static void si_emit_draw_packets(struct si_context *sctx,
467 const struct pipe_draw_info *info,
468 const struct pipe_index_buffer *ib)
469 {
470 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
471 unsigned sh_base_reg = sctx->shader_userdata.sh_base[PIPE_SHADER_VERTEX];
472 bool render_cond_bit = sctx->b.render_cond && !sctx->b.render_cond_force_off;
473
474 if (info->count_from_stream_output) {
475 struct r600_so_target *t =
476 (struct r600_so_target*)info->count_from_stream_output;
477 uint64_t va = t->buf_filled_size->gpu_address +
478 t->buf_filled_size_offset;
479
480 radeon_set_context_reg(cs, R_028B30_VGT_STRMOUT_DRAW_OPAQUE_VERTEX_STRIDE,
481 t->stride_in_dw);
482
483 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
484 radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
485 COPY_DATA_DST_SEL(COPY_DATA_REG) |
486 COPY_DATA_WR_CONFIRM);
487 radeon_emit(cs, va); /* src address lo */
488 radeon_emit(cs, va >> 32); /* src address hi */
489 radeon_emit(cs, R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE >> 2);
490 radeon_emit(cs, 0); /* unused */
491
492 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
493 t->buf_filled_size, RADEON_USAGE_READ,
494 RADEON_PRIO_SO_FILLED_SIZE);
495 }
496
497 /* draw packet */
498 if (info->indexed) {
499 radeon_emit(cs, PKT3(PKT3_INDEX_TYPE, 0, 0));
500
501 /* index type */
502 switch (ib->index_size) {
503 case 1:
504 radeon_emit(cs, V_028A7C_VGT_INDEX_8);
505 break;
506 case 2:
507 radeon_emit(cs, V_028A7C_VGT_INDEX_16 |
508 (SI_BIG_ENDIAN && sctx->b.chip_class <= CIK ?
509 V_028A7C_VGT_DMA_SWAP_16_BIT : 0));
510 break;
511 case 4:
512 radeon_emit(cs, V_028A7C_VGT_INDEX_32 |
513 (SI_BIG_ENDIAN && sctx->b.chip_class <= CIK ?
514 V_028A7C_VGT_DMA_SWAP_32_BIT : 0));
515 break;
516 default:
517 assert(!"unreachable");
518 return;
519 }
520 }
521
522 if (!info->indirect) {
523 int base_vertex;
524
525 radeon_emit(cs, PKT3(PKT3_NUM_INSTANCES, 0, 0));
526 radeon_emit(cs, info->instance_count);
527
528 /* Base vertex and start instance. */
529 base_vertex = info->indexed ? info->index_bias : info->start;
530
531 if (base_vertex != sctx->last_base_vertex ||
532 sctx->last_base_vertex == SI_BASE_VERTEX_UNKNOWN ||
533 info->start_instance != sctx->last_start_instance ||
534 sh_base_reg != sctx->last_sh_base_reg) {
535 radeon_set_sh_reg_seq(cs, sh_base_reg + SI_SGPR_BASE_VERTEX * 4, 2);
536 radeon_emit(cs, base_vertex);
537 radeon_emit(cs, info->start_instance);
538
539 sctx->last_base_vertex = base_vertex;
540 sctx->last_start_instance = info->start_instance;
541 sctx->last_sh_base_reg = sh_base_reg;
542 }
543 } else {
544 si_invalidate_draw_sh_constants(sctx);
545
546 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
547 (struct r600_resource *)info->indirect,
548 RADEON_USAGE_READ, RADEON_PRIO_DRAW_INDIRECT);
549 }
550
551 if (info->indexed) {
552 uint32_t index_max_size = (ib->buffer->width0 - ib->offset) /
553 ib->index_size;
554 uint64_t index_va = r600_resource(ib->buffer)->gpu_address + ib->offset;
555
556 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
557 (struct r600_resource *)ib->buffer,
558 RADEON_USAGE_READ, RADEON_PRIO_INDEX_BUFFER);
559
560 if (info->indirect) {
561 uint64_t indirect_va = r600_resource(info->indirect)->gpu_address;
562
563 assert(indirect_va % 8 == 0);
564 assert(index_va % 2 == 0);
565 assert(info->indirect_offset % 4 == 0);
566
567 radeon_emit(cs, PKT3(PKT3_SET_BASE, 2, 0));
568 radeon_emit(cs, 1);
569 radeon_emit(cs, indirect_va);
570 radeon_emit(cs, indirect_va >> 32);
571
572 radeon_emit(cs, PKT3(PKT3_INDEX_BASE, 1, 0));
573 radeon_emit(cs, index_va);
574 radeon_emit(cs, index_va >> 32);
575
576 radeon_emit(cs, PKT3(PKT3_INDEX_BUFFER_SIZE, 0, 0));
577 radeon_emit(cs, index_max_size);
578
579 radeon_emit(cs, PKT3(PKT3_DRAW_INDEX_INDIRECT, 3, render_cond_bit));
580 radeon_emit(cs, info->indirect_offset);
581 radeon_emit(cs, (sh_base_reg + SI_SGPR_BASE_VERTEX * 4 - SI_SH_REG_OFFSET) >> 2);
582 radeon_emit(cs, (sh_base_reg + SI_SGPR_START_INSTANCE * 4 - SI_SH_REG_OFFSET) >> 2);
583 radeon_emit(cs, V_0287F0_DI_SRC_SEL_DMA);
584 } else {
585 index_va += info->start * ib->index_size;
586
587 radeon_emit(cs, PKT3(PKT3_DRAW_INDEX_2, 4, render_cond_bit));
588 radeon_emit(cs, index_max_size);
589 radeon_emit(cs, index_va);
590 radeon_emit(cs, (index_va >> 32UL) & 0xFF);
591 radeon_emit(cs, info->count);
592 radeon_emit(cs, V_0287F0_DI_SRC_SEL_DMA);
593 }
594 } else {
595 if (info->indirect) {
596 uint64_t indirect_va = r600_resource(info->indirect)->gpu_address;
597
598 assert(indirect_va % 8 == 0);
599 assert(info->indirect_offset % 4 == 0);
600
601 radeon_emit(cs, PKT3(PKT3_SET_BASE, 2, 0));
602 radeon_emit(cs, 1);
603 radeon_emit(cs, indirect_va);
604 radeon_emit(cs, indirect_va >> 32);
605
606 radeon_emit(cs, PKT3(PKT3_DRAW_INDIRECT, 3, render_cond_bit));
607 radeon_emit(cs, info->indirect_offset);
608 radeon_emit(cs, (sh_base_reg + SI_SGPR_BASE_VERTEX * 4 - SI_SH_REG_OFFSET) >> 2);
609 radeon_emit(cs, (sh_base_reg + SI_SGPR_START_INSTANCE * 4 - SI_SH_REG_OFFSET) >> 2);
610 radeon_emit(cs, V_0287F0_DI_SRC_SEL_AUTO_INDEX);
611 } else {
612 radeon_emit(cs, PKT3(PKT3_DRAW_INDEX_AUTO, 1, render_cond_bit));
613 radeon_emit(cs, info->count);
614 radeon_emit(cs, V_0287F0_DI_SRC_SEL_AUTO_INDEX |
615 S_0287F0_USE_OPAQUE(!!info->count_from_stream_output));
616 }
617 }
618 }
619
620 void si_emit_cache_flush(struct si_context *si_ctx, struct r600_atom *atom)
621 {
622 struct r600_common_context *sctx = &si_ctx->b;
623 struct radeon_winsys_cs *cs = sctx->gfx.cs;
624 uint32_t cp_coher_cntl = 0;
625 uint32_t compute =
626 PKT3_SHADER_TYPE_S(!!(sctx->flags & SI_CONTEXT_FLAG_COMPUTE));
627
628 /* SI has a bug that it always flushes ICACHE and KCACHE if either
629 * bit is set. An alternative way is to write SQC_CACHES, but that
630 * doesn't seem to work reliably. Since the bug doesn't affect
631 * correctness (it only does more work than necessary) and
632 * the performance impact is likely negligible, there is no plan
633 * to fix it.
634 */
635
636 if (sctx->flags & SI_CONTEXT_INV_ICACHE)
637 cp_coher_cntl |= S_0085F0_SH_ICACHE_ACTION_ENA(1);
638 if (sctx->flags & SI_CONTEXT_INV_SMEM_L1)
639 cp_coher_cntl |= S_0085F0_SH_KCACHE_ACTION_ENA(1);
640
641 if (sctx->flags & SI_CONTEXT_INV_VMEM_L1)
642 cp_coher_cntl |= S_0085F0_TCL1_ACTION_ENA(1);
643 if (sctx->flags & SI_CONTEXT_INV_GLOBAL_L2) {
644 cp_coher_cntl |= S_0085F0_TC_ACTION_ENA(1);
645
646 /* TODO: this might not be needed. */
647 if (sctx->chip_class >= VI)
648 cp_coher_cntl |= S_0301F0_TC_WB_ACTION_ENA(1);
649 }
650
651 if (sctx->flags & SI_CONTEXT_FLUSH_AND_INV_CB) {
652 cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) |
653 S_0085F0_CB0_DEST_BASE_ENA(1) |
654 S_0085F0_CB1_DEST_BASE_ENA(1) |
655 S_0085F0_CB2_DEST_BASE_ENA(1) |
656 S_0085F0_CB3_DEST_BASE_ENA(1) |
657 S_0085F0_CB4_DEST_BASE_ENA(1) |
658 S_0085F0_CB5_DEST_BASE_ENA(1) |
659 S_0085F0_CB6_DEST_BASE_ENA(1) |
660 S_0085F0_CB7_DEST_BASE_ENA(1);
661
662 /* Necessary for DCC */
663 if (sctx->chip_class >= VI) {
664 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0) | compute);
665 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_DATA_TS) |
666 EVENT_INDEX(5));
667 radeon_emit(cs, 0);
668 radeon_emit(cs, 0);
669 radeon_emit(cs, 0);
670 radeon_emit(cs, 0);
671 }
672 }
673 if (sctx->flags & SI_CONTEXT_FLUSH_AND_INV_DB) {
674 cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) |
675 S_0085F0_DB_DEST_BASE_ENA(1);
676 }
677
678 if (sctx->flags & SI_CONTEXT_FLUSH_AND_INV_CB_META) {
679 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
680 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
681 }
682 if (sctx->flags & SI_CONTEXT_FLUSH_AND_INV_DB_META) {
683 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
684 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
685 }
686 if (sctx->flags & SI_CONTEXT_FLUSH_WITH_INV_L2) {
687 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
688 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH) | EVENT_INDEX(7) |
689 EVENT_WRITE_INV_L2);
690 }
691
692 /* FLUSH_AND_INV events must be emitted before PS_PARTIAL_FLUSH.
693 * Otherwise, clearing CMASK (CB meta) with CP DMA isn't reliable.
694 *
695 * I think the reason is that FLUSH_AND_INV is only added to a queue
696 * and it is PS_PARTIAL_FLUSH that waits for it to complete.
697 */
698 if (sctx->flags & SI_CONTEXT_PS_PARTIAL_FLUSH) {
699 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
700 radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
701 } else if (sctx->flags & SI_CONTEXT_VS_PARTIAL_FLUSH) {
702 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
703 radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
704 }
705 if (sctx->flags & SI_CONTEXT_CS_PARTIAL_FLUSH) {
706 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
707 radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH | EVENT_INDEX(4)));
708 }
709 if (sctx->flags & SI_CONTEXT_VGT_FLUSH) {
710 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
711 radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
712 }
713 if (sctx->flags & SI_CONTEXT_VGT_STREAMOUT_SYNC) {
714 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
715 radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC) | EVENT_INDEX(0));
716 }
717
718 /* SURFACE_SYNC must be emitted after partial flushes.
719 * It looks like SURFACE_SYNC flushes caches immediately and doesn't
720 * wait for any engines. This should be last.
721 */
722 if (cp_coher_cntl) {
723 if (sctx->chip_class >= CIK) {
724 radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, 0) | compute);
725 radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
726 radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
727 radeon_emit(cs, 0xff); /* CP_COHER_SIZE_HI */
728 radeon_emit(cs, 0); /* CP_COHER_BASE */
729 radeon_emit(cs, 0); /* CP_COHER_BASE_HI */
730 radeon_emit(cs, 0x0000000A); /* POLL_INTERVAL */
731 } else {
732 radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, 0) | compute);
733 radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
734 radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
735 radeon_emit(cs, 0); /* CP_COHER_BASE */
736 radeon_emit(cs, 0x0000000A); /* POLL_INTERVAL */
737 }
738 }
739
740 sctx->flags = 0;
741 }
742
743 static void si_get_draw_start_count(struct si_context *sctx,
744 const struct pipe_draw_info *info,
745 unsigned *start, unsigned *count)
746 {
747 if (info->indirect) {
748 struct r600_resource *indirect =
749 (struct r600_resource*)info->indirect;
750 int *data = r600_buffer_map_sync_with_rings(&sctx->b,
751 indirect, PIPE_TRANSFER_READ);
752 data += info->indirect_offset/sizeof(int);
753 *start = data[2];
754 *count = data[0];
755 } else {
756 *start = info->start;
757 *count = info->count;
758 }
759 }
760
761 void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
762 {
763 struct si_context *sctx = (struct si_context *)ctx;
764 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
765 struct pipe_index_buffer ib = {};
766 unsigned mask;
767
768 if (!info->count && !info->indirect &&
769 (info->indexed || !info->count_from_stream_output))
770 return;
771
772 if (!sctx->vs_shader.cso) {
773 assert(0);
774 return;
775 }
776 if (!sctx->ps_shader.cso && (!rs || !rs->rasterizer_discard)) {
777 assert(0);
778 return;
779 }
780 if (!!sctx->tes_shader.cso != (info->mode == PIPE_PRIM_PATCHES)) {
781 assert(0);
782 return;
783 }
784
785 si_decompress_textures(sctx);
786
787 /* Set the rasterization primitive type.
788 *
789 * This must be done after si_decompress_textures, which can call
790 * draw_vbo recursively, and before si_update_shaders, which uses
791 * current_rast_prim for this draw_vbo call. */
792 if (sctx->gs_shader.cso)
793 sctx->current_rast_prim = sctx->gs_shader.cso->gs_output_prim;
794 else if (sctx->tes_shader.cso)
795 sctx->current_rast_prim =
796 sctx->tes_shader.cso->info.properties[TGSI_PROPERTY_TES_PRIM_MODE];
797 else
798 sctx->current_rast_prim = info->mode;
799
800 if (!si_update_shaders(sctx) ||
801 !si_upload_shader_descriptors(sctx))
802 return;
803
804 if (info->indexed) {
805 /* Initialize the index buffer struct. */
806 pipe_resource_reference(&ib.buffer, sctx->index_buffer.buffer);
807 ib.user_buffer = sctx->index_buffer.user_buffer;
808 ib.index_size = sctx->index_buffer.index_size;
809 ib.offset = sctx->index_buffer.offset;
810
811 /* Translate or upload, if needed. */
812 /* 8-bit indices are supported on VI. */
813 if (sctx->b.chip_class <= CIK && ib.index_size == 1) {
814 struct pipe_resource *out_buffer = NULL;
815 unsigned out_offset, start, count, start_offset;
816 void *ptr;
817
818 si_get_draw_start_count(sctx, info, &start, &count);
819 start_offset = start * ib.index_size;
820
821 u_upload_alloc(sctx->b.uploader, start_offset, count * 2,
822 &out_offset, &out_buffer, &ptr);
823 if (!out_buffer) {
824 pipe_resource_reference(&ib.buffer, NULL);
825 return;
826 }
827
828 util_shorten_ubyte_elts_to_userptr(&sctx->b.b, &ib, 0,
829 ib.offset + start_offset,
830 count, ptr);
831
832 pipe_resource_reference(&ib.buffer, NULL);
833 ib.user_buffer = NULL;
834 ib.buffer = out_buffer;
835 /* info->start will be added by the drawing code */
836 ib.offset = out_offset - start_offset;
837 ib.index_size = 2;
838 } else if (ib.user_buffer && !ib.buffer) {
839 unsigned start, count, start_offset;
840
841 si_get_draw_start_count(sctx, info, &start, &count);
842 start_offset = start * ib.index_size;
843
844 u_upload_data(sctx->b.uploader, start_offset, count * ib.index_size,
845 (char*)ib.user_buffer + start_offset,
846 &ib.offset, &ib.buffer);
847 if (!ib.buffer)
848 return;
849 /* info->start will be added by the drawing code */
850 ib.offset -= start_offset;
851 }
852 }
853
854 /* VI reads index buffers through TC L2. */
855 if (info->indexed && sctx->b.chip_class <= CIK &&
856 r600_resource(ib.buffer)->TC_L2_dirty) {
857 sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
858 r600_resource(ib.buffer)->TC_L2_dirty = false;
859 }
860
861 /* Check flush flags. */
862 if (sctx->b.flags)
863 si_mark_atom_dirty(sctx, sctx->atoms.s.cache_flush);
864
865 si_need_cs_space(sctx);
866
867 /* Emit states. */
868 mask = sctx->dirty_atoms;
869 while (mask) {
870 struct r600_atom *atom = sctx->atoms.array[u_bit_scan(&mask)];
871
872 atom->emit(&sctx->b, atom);
873 }
874 sctx->dirty_atoms = 0;
875
876 si_pm4_emit_dirty(sctx);
877 si_emit_scratch_reloc(sctx);
878 si_emit_rasterizer_prim_state(sctx);
879 si_emit_draw_registers(sctx, info);
880 si_emit_draw_packets(sctx, info, &ib);
881
882 if (sctx->trace_buf)
883 si_trace_emit(sctx);
884
885 /* Workaround for a VGT hang when streamout is enabled.
886 * It must be done after drawing. */
887 if ((sctx->b.family == CHIP_HAWAII ||
888 sctx->b.family == CHIP_TONGA ||
889 sctx->b.family == CHIP_FIJI) &&
890 (sctx->b.streamout.streamout_enabled ||
891 sctx->b.streamout.prims_gen_query_enabled)) {
892 sctx->b.flags |= SI_CONTEXT_VGT_STREAMOUT_SYNC;
893 }
894
895 /* Set the depth buffer as dirty. */
896 if (sctx->framebuffer.state.zsbuf) {
897 struct pipe_surface *surf = sctx->framebuffer.state.zsbuf;
898 struct r600_texture *rtex = (struct r600_texture *)surf->texture;
899
900 rtex->dirty_level_mask |= 1 << surf->u.tex.level;
901
902 if (rtex->surface.flags & RADEON_SURF_SBUFFER)
903 rtex->stencil_dirty_level_mask |= 1 << surf->u.tex.level;
904 }
905 if (sctx->framebuffer.compressed_cb_mask) {
906 struct pipe_surface *surf;
907 struct r600_texture *rtex;
908 unsigned mask = sctx->framebuffer.compressed_cb_mask;
909
910 do {
911 unsigned i = u_bit_scan(&mask);
912 surf = sctx->framebuffer.state.cbufs[i];
913 rtex = (struct r600_texture*)surf->texture;
914
915 rtex->dirty_level_mask |= 1 << surf->u.tex.level;
916 } while (mask);
917 }
918
919 pipe_resource_reference(&ib.buffer, NULL);
920 sctx->b.num_draw_calls++;
921 }
922
923 void si_trace_emit(struct si_context *sctx)
924 {
925 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
926
927 sctx->trace_id++;
928 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, sctx->trace_buf,
929 RADEON_USAGE_READWRITE, RADEON_PRIO_TRACE);
930 radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 3, 0));
931 radeon_emit(cs, S_370_DST_SEL(V_370_MEMORY_SYNC) |
932 S_370_WR_CONFIRM(1) |
933 S_370_ENGINE_SEL(V_370_ME));
934 radeon_emit(cs, sctx->trace_buf->gpu_address);
935 radeon_emit(cs, sctx->trace_buf->gpu_address >> 32);
936 radeon_emit(cs, sctx->trace_id);
937 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
938 radeon_emit(cs, SI_ENCODE_TRACE_POINT(sctx->trace_id));
939 }