• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2012 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * on the rights to use, copy, modify, merge, publish, distribute, sub
9  * license, and/or sell copies of the Software, and to permit persons to whom
10  * the Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
23  */
24 
25 #include "ac_debug.h"
26 #include "si_build_pm4.h"
27 #include "sid.h"
28 #include "util/u_index_modify.h"
29 #include "util/u_log.h"
30 #include "util/u_prim.h"
31 #include "util/u_suballoc.h"
32 #include "util/u_upload_mgr.h"
33 
34 /* special primitive types */
35 #define SI_PRIM_RECTANGLE_LIST PIPE_PRIM_MAX
36 
37 ALWAYS_INLINE
si_conv_pipe_prim(unsigned mode)38 static unsigned si_conv_pipe_prim(unsigned mode)
39 {
40    static const unsigned prim_conv[] = {
41       [PIPE_PRIM_POINTS] = V_008958_DI_PT_POINTLIST,
42       [PIPE_PRIM_LINES] = V_008958_DI_PT_LINELIST,
43       [PIPE_PRIM_LINE_LOOP] = V_008958_DI_PT_LINELOOP,
44       [PIPE_PRIM_LINE_STRIP] = V_008958_DI_PT_LINESTRIP,
45       [PIPE_PRIM_TRIANGLES] = V_008958_DI_PT_TRILIST,
46       [PIPE_PRIM_TRIANGLE_STRIP] = V_008958_DI_PT_TRISTRIP,
47       [PIPE_PRIM_TRIANGLE_FAN] = V_008958_DI_PT_TRIFAN,
48       [PIPE_PRIM_QUADS] = V_008958_DI_PT_QUADLIST,
49       [PIPE_PRIM_QUAD_STRIP] = V_008958_DI_PT_QUADSTRIP,
50       [PIPE_PRIM_POLYGON] = V_008958_DI_PT_POLYGON,
51       [PIPE_PRIM_LINES_ADJACENCY] = V_008958_DI_PT_LINELIST_ADJ,
52       [PIPE_PRIM_LINE_STRIP_ADJACENCY] = V_008958_DI_PT_LINESTRIP_ADJ,
53       [PIPE_PRIM_TRIANGLES_ADJACENCY] = V_008958_DI_PT_TRILIST_ADJ,
54       [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY] = V_008958_DI_PT_TRISTRIP_ADJ,
55       [PIPE_PRIM_PATCHES] = V_008958_DI_PT_PATCH,
56       [SI_PRIM_RECTANGLE_LIST] = V_008958_DI_PT_RECTLIST};
57    assert(mode < ARRAY_SIZE(prim_conv));
58    return prim_conv[mode];
59 }
60 
61 /**
62  * This calculates the LDS size for tessellation shaders (VS, TCS, TES).
63  * LS.LDS_SIZE is shared by all 3 shader stages.
64  *
65  * The information about LDS and other non-compile-time parameters is then
66  * written to userdata SGPRs.
67  */
si_emit_derived_tess_state(struct si_context * sctx,const struct pipe_draw_info * info,unsigned * num_patches)68 static void si_emit_derived_tess_state(struct si_context *sctx, const struct pipe_draw_info *info,
69                                        unsigned *num_patches)
70 {
71    struct radeon_cmdbuf *cs = sctx->gfx_cs;
72    struct si_shader *ls_current;
73    struct si_shader_selector *ls;
74    /* The TES pointer will only be used for sctx->last_tcs.
75     * It would be wrong to think that TCS = TES. */
76    struct si_shader_selector *tcs =
77       sctx->tcs_shader.cso ? sctx->tcs_shader.cso : sctx->tes_shader.cso;
78    unsigned tess_uses_primid = sctx->ia_multi_vgt_param_key.u.tess_uses_prim_id;
79    bool has_primid_instancing_bug = sctx->chip_class == GFX6 && sctx->screen->info.max_se == 1;
80    unsigned tes_sh_base = sctx->shader_pointers.sh_base[PIPE_SHADER_TESS_EVAL];
81    unsigned num_tcs_input_cp = info->vertices_per_patch;
82    unsigned num_tcs_output_cp, num_tcs_inputs, num_tcs_outputs;
83    unsigned num_tcs_patch_outputs;
84    unsigned input_vertex_size, output_vertex_size, pervertex_output_patch_size;
85    unsigned input_patch_size, output_patch_size, output_patch0_offset;
86    unsigned perpatch_output_offset, lds_size;
87    unsigned tcs_in_layout, tcs_out_layout, tcs_out_offsets;
88    unsigned offchip_layout, hardware_lds_size, ls_hs_config;
89 
90    /* Since GFX9 has merged LS-HS in the TCS state, set LS = TCS. */
91    if (sctx->chip_class >= GFX9) {
92       if (sctx->tcs_shader.cso)
93          ls_current = sctx->tcs_shader.current;
94       else
95          ls_current = sctx->fixed_func_tcs_shader.current;
96 
97       ls = ls_current->key.part.tcs.ls;
98    } else {
99       ls_current = sctx->vs_shader.current;
100       ls = sctx->vs_shader.cso;
101    }
102 
103    if (sctx->last_ls == ls_current && sctx->last_tcs == tcs &&
104        sctx->last_tes_sh_base == tes_sh_base && sctx->last_num_tcs_input_cp == num_tcs_input_cp &&
105        (!has_primid_instancing_bug || (sctx->last_tess_uses_primid == tess_uses_primid))) {
106       *num_patches = sctx->last_num_patches;
107       return;
108    }
109 
110    sctx->last_ls = ls_current;
111    sctx->last_tcs = tcs;
112    sctx->last_tes_sh_base = tes_sh_base;
113    sctx->last_num_tcs_input_cp = num_tcs_input_cp;
114    sctx->last_tess_uses_primid = tess_uses_primid;
115 
116    /* This calculates how shader inputs and outputs among VS, TCS, and TES
117     * are laid out in LDS. */
118    num_tcs_inputs = util_last_bit64(ls->outputs_written);
119 
120    if (sctx->tcs_shader.cso) {
121       num_tcs_outputs = util_last_bit64(tcs->outputs_written);
122       num_tcs_output_cp = tcs->info.base.tess.tcs_vertices_out;
123       num_tcs_patch_outputs = util_last_bit64(tcs->patch_outputs_written);
124    } else {
125       /* No TCS. Route varyings from LS to TES. */
126       num_tcs_outputs = num_tcs_inputs;
127       num_tcs_output_cp = num_tcs_input_cp;
128       num_tcs_patch_outputs = 2; /* TESSINNER + TESSOUTER */
129    }
130 
131    input_vertex_size = ls->lshs_vertex_stride;
132    output_vertex_size = num_tcs_outputs * 16;
133 
134    input_patch_size = num_tcs_input_cp * input_vertex_size;
135 
136    pervertex_output_patch_size = num_tcs_output_cp * output_vertex_size;
137    output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
138 
139    /* Ensure that we only need one wave per SIMD so we don't need to check
140     * resource usage. Also ensures that the number of tcs in and out
141     * vertices per threadgroup are at most 256.
142     */
143    unsigned max_verts_per_patch = MAX2(num_tcs_input_cp, num_tcs_output_cp);
144    *num_patches = 256 / max_verts_per_patch;
145 
146    /* Make sure that the data fits in LDS. This assumes the shaders only
147     * use LDS for the inputs and outputs.
148     *
149     * While GFX7 can use 64K per threadgroup, there is a hang on Stoney
150     * with 2 CUs if we use more than 32K. The closed Vulkan driver also
151     * uses 32K at most on all GCN chips.
152     */
153    hardware_lds_size = 32768;
154    *num_patches = MIN2(*num_patches, hardware_lds_size / (input_patch_size + output_patch_size));
155 
156    /* Make sure the output data fits in the offchip buffer */
157    *num_patches =
158       MIN2(*num_patches, (sctx->screen->tess_offchip_block_dw_size * 4) / output_patch_size);
159 
160    /* Not necessary for correctness, but improves performance.
161     * The hardware can do more, but the radeonsi shader constant is
162     * limited to 6 bits.
163     */
164    *num_patches = MIN2(*num_patches, 63); /* triangles: 3 full waves except 3 lanes */
165 
166    /* When distributed tessellation is unsupported, switch between SEs
167     * at a higher frequency to compensate for it.
168     */
169    if (!sctx->screen->info.has_distributed_tess && sctx->screen->info.max_se > 1)
170       *num_patches = MIN2(*num_patches, 16); /* recommended */
171 
172    /* Make sure that vector lanes are reasonably occupied. It probably
173     * doesn't matter much because this is LS-HS, and TES is likely to
174     * occupy significantly more CUs.
175     */
176    unsigned temp_verts_per_tg = *num_patches * max_verts_per_patch;
177    unsigned wave_size = sctx->screen->ge_wave_size;
178 
179    if (temp_verts_per_tg > wave_size && temp_verts_per_tg % wave_size < wave_size * 3 / 4)
180       *num_patches = (temp_verts_per_tg & ~(wave_size - 1)) / max_verts_per_patch;
181 
182    if (sctx->chip_class == GFX6) {
183       /* GFX6 bug workaround, related to power management. Limit LS-HS
184        * threadgroups to only one wave.
185        */
186       unsigned one_wave = wave_size / max_verts_per_patch;
187       *num_patches = MIN2(*num_patches, one_wave);
188    }
189 
190    /* The VGT HS block increments the patch ID unconditionally
191     * within a single threadgroup. This results in incorrect
192     * patch IDs when instanced draws are used.
193     *
194     * The intended solution is to restrict threadgroups to
195     * a single instance by setting SWITCH_ON_EOI, which
196     * should cause IA to split instances up. However, this
197     * doesn't work correctly on GFX6 when there is no other
198     * SE to switch to.
199     */
200    if (has_primid_instancing_bug && tess_uses_primid)
201       *num_patches = 1;
202 
203    sctx->last_num_patches = *num_patches;
204 
205    output_patch0_offset = input_patch_size * *num_patches;
206    perpatch_output_offset = output_patch0_offset + pervertex_output_patch_size;
207 
208    /* Compute userdata SGPRs. */
209    assert(((input_vertex_size / 4) & ~0xff) == 0);
210    assert(((output_vertex_size / 4) & ~0xff) == 0);
211    assert(((input_patch_size / 4) & ~0x1fff) == 0);
212    assert(((output_patch_size / 4) & ~0x1fff) == 0);
213    assert(((output_patch0_offset / 16) & ~0xffff) == 0);
214    assert(((perpatch_output_offset / 16) & ~0xffff) == 0);
215    assert(num_tcs_input_cp <= 32);
216    assert(num_tcs_output_cp <= 32);
217 
218    uint64_t ring_va = (unlikely(sctx->ws->cs_is_secure(sctx->gfx_cs)) ?
219       si_resource(sctx->tess_rings_tmz) : si_resource(sctx->tess_rings))->gpu_address;
220    assert((ring_va & u_bit_consecutive(0, 19)) == 0);
221 
222    tcs_in_layout = S_VS_STATE_LS_OUT_PATCH_SIZE(input_patch_size / 4) |
223                    S_VS_STATE_LS_OUT_VERTEX_SIZE(input_vertex_size / 4);
224    tcs_out_layout = (output_patch_size / 4) | (num_tcs_input_cp << 13) | ring_va;
225    tcs_out_offsets = (output_patch0_offset / 16) | ((perpatch_output_offset / 16) << 16);
226    offchip_layout =
227       *num_patches | (num_tcs_output_cp << 6) | (pervertex_output_patch_size * *num_patches << 12);
228 
229    /* Compute the LDS size. */
230    lds_size = output_patch0_offset + output_patch_size * *num_patches;
231 
232    if (sctx->chip_class >= GFX7) {
233       assert(lds_size <= 65536);
234       lds_size = align(lds_size, 512) / 512;
235    } else {
236       assert(lds_size <= 32768);
237       lds_size = align(lds_size, 256) / 256;
238    }
239 
240    /* Set SI_SGPR_VS_STATE_BITS. */
241    sctx->current_vs_state &= C_VS_STATE_LS_OUT_PATCH_SIZE & C_VS_STATE_LS_OUT_VERTEX_SIZE;
242    sctx->current_vs_state |= tcs_in_layout;
243 
244    /* We should be able to support in-shader LDS use with LLVM >= 9
245     * by just adding the lds_sizes together, but it has never
246     * been tested. */
247    assert(ls_current->config.lds_size == 0);
248 
249    if (sctx->chip_class >= GFX9) {
250       unsigned hs_rsrc2 = ls_current->config.rsrc2;
251 
252       if (sctx->chip_class >= GFX10)
253          hs_rsrc2 |= S_00B42C_LDS_SIZE_GFX10(lds_size);
254       else
255          hs_rsrc2 |= S_00B42C_LDS_SIZE_GFX9(lds_size);
256 
257       radeon_set_sh_reg(cs, R_00B42C_SPI_SHADER_PGM_RSRC2_HS, hs_rsrc2);
258 
259       /* Set userdata SGPRs for merged LS-HS. */
260       radeon_set_sh_reg_seq(
261          cs, R_00B430_SPI_SHADER_USER_DATA_LS_0 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT * 4, 3);
262       radeon_emit(cs, offchip_layout);
263       radeon_emit(cs, tcs_out_offsets);
264       radeon_emit(cs, tcs_out_layout);
265    } else {
266       unsigned ls_rsrc2 = ls_current->config.rsrc2;
267 
268       si_multiwave_lds_size_workaround(sctx->screen, &lds_size);
269       ls_rsrc2 |= S_00B52C_LDS_SIZE(lds_size);
270 
271       /* Due to a hw bug, RSRC2_LS must be written twice with another
272        * LS register written in between. */
273       if (sctx->chip_class == GFX7 && sctx->family != CHIP_HAWAII)
274          radeon_set_sh_reg(cs, R_00B52C_SPI_SHADER_PGM_RSRC2_LS, ls_rsrc2);
275       radeon_set_sh_reg_seq(cs, R_00B528_SPI_SHADER_PGM_RSRC1_LS, 2);
276       radeon_emit(cs, ls_current->config.rsrc1);
277       radeon_emit(cs, ls_rsrc2);
278 
279       /* Set userdata SGPRs for TCS. */
280       radeon_set_sh_reg_seq(
281          cs, R_00B430_SPI_SHADER_USER_DATA_HS_0 + GFX6_SGPR_TCS_OFFCHIP_LAYOUT * 4, 4);
282       radeon_emit(cs, offchip_layout);
283       radeon_emit(cs, tcs_out_offsets);
284       radeon_emit(cs, tcs_out_layout);
285       radeon_emit(cs, tcs_in_layout);
286    }
287 
288    /* Set userdata SGPRs for TES. */
289    radeon_set_sh_reg_seq(cs, tes_sh_base + SI_SGPR_TES_OFFCHIP_LAYOUT * 4, 2);
290    radeon_emit(cs, offchip_layout);
291    radeon_emit(cs, ring_va);
292 
293    ls_hs_config = S_028B58_NUM_PATCHES(*num_patches) | S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp) |
294                   S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp);
295 
296    if (sctx->last_ls_hs_config != ls_hs_config) {
297       if (sctx->chip_class >= GFX7) {
298          radeon_set_context_reg_idx(cs, R_028B58_VGT_LS_HS_CONFIG, 2, ls_hs_config);
299       } else {
300          radeon_set_context_reg(cs, R_028B58_VGT_LS_HS_CONFIG, ls_hs_config);
301       }
302       sctx->last_ls_hs_config = ls_hs_config;
303       sctx->context_roll = true;
304    }
305 }
306 
si_num_prims_for_vertices(enum pipe_prim_type prim,unsigned count,unsigned vertices_per_patch)307 static unsigned si_num_prims_for_vertices(enum pipe_prim_type prim,
308                                           unsigned count, unsigned vertices_per_patch)
309 {
310    switch (prim) {
311    case PIPE_PRIM_PATCHES:
312       return count / vertices_per_patch;
313    case PIPE_PRIM_POLYGON:
314       return count >= 3;
315    case SI_PRIM_RECTANGLE_LIST:
316       return count / 3;
317    default:
318       return u_decomposed_prims_for_vertices(prim, count);
319    }
320 }
321 
si_get_init_multi_vgt_param(struct si_screen * sscreen,union si_vgt_param_key * key)322 static unsigned si_get_init_multi_vgt_param(struct si_screen *sscreen, union si_vgt_param_key *key)
323 {
324    STATIC_ASSERT(sizeof(union si_vgt_param_key) == 4);
325    unsigned max_primgroup_in_wave = 2;
326 
327    /* SWITCH_ON_EOP(0) is always preferable. */
328    bool wd_switch_on_eop = false;
329    bool ia_switch_on_eop = false;
330    bool ia_switch_on_eoi = false;
331    bool partial_vs_wave = false;
332    bool partial_es_wave = false;
333 
334    if (key->u.uses_tess) {
335       /* SWITCH_ON_EOI must be set if PrimID is used. */
336       if (key->u.tess_uses_prim_id)
337          ia_switch_on_eoi = true;
338 
339       /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
340       if ((sscreen->info.family == CHIP_TAHITI || sscreen->info.family == CHIP_PITCAIRN ||
341            sscreen->info.family == CHIP_BONAIRE) &&
342           key->u.uses_gs)
343          partial_vs_wave = true;
344 
345       /* Needed for 028B6C_DISTRIBUTION_MODE != 0. (implies >= GFX8) */
346       if (sscreen->info.has_distributed_tess) {
347          if (key->u.uses_gs) {
348             if (sscreen->info.chip_class == GFX8)
349                partial_es_wave = true;
350          } else {
351             partial_vs_wave = true;
352          }
353       }
354    }
355 
356    /* This is a hardware requirement. */
357    if (key->u.line_stipple_enabled || (sscreen->debug_flags & DBG(SWITCH_ON_EOP))) {
358       ia_switch_on_eop = true;
359       wd_switch_on_eop = true;
360    }
361 
362    if (sscreen->info.chip_class >= GFX7) {
363       /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
364        * 4 shader engines. Set 1 to pass the assertion below.
365        * The other cases are hardware requirements.
366        *
367        * Polaris supports primitive restart with WD_SWITCH_ON_EOP=0
368        * for points, line strips, and tri strips.
369        */
370       if (sscreen->info.max_se <= 2 || key->u.prim == PIPE_PRIM_POLYGON ||
371           key->u.prim == PIPE_PRIM_LINE_LOOP || key->u.prim == PIPE_PRIM_TRIANGLE_FAN ||
372           key->u.prim == PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY ||
373           (key->u.primitive_restart &&
374            (sscreen->info.family < CHIP_POLARIS10 ||
375             (key->u.prim != PIPE_PRIM_POINTS && key->u.prim != PIPE_PRIM_LINE_STRIP &&
376              key->u.prim != PIPE_PRIM_TRIANGLE_STRIP))) ||
377           key->u.count_from_stream_output)
378          wd_switch_on_eop = true;
379 
380       /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
381        * We don't know that for indirect drawing, so treat it as
382        * always problematic. */
383       if (sscreen->info.family == CHIP_HAWAII && key->u.uses_instancing)
384          wd_switch_on_eop = true;
385 
386       /* Performance recommendation for 4 SE Gfx7-8 parts if
387        * instances are smaller than a primgroup.
388        * Assume indirect draws always use small instances.
389        * This is needed for good VS wave utilization.
390        */
391       if (sscreen->info.chip_class <= GFX8 && sscreen->info.max_se == 4 &&
392           key->u.multi_instances_smaller_than_primgroup)
393          wd_switch_on_eop = true;
394 
395       /* Required on GFX7 and later. */
396       if (sscreen->info.max_se == 4 && !wd_switch_on_eop)
397          ia_switch_on_eoi = true;
398 
399       /* HW engineers suggested that PARTIAL_VS_WAVE_ON should be set
400        * to work around a GS hang.
401        */
402       if (key->u.uses_gs &&
403           (sscreen->info.family == CHIP_TONGA || sscreen->info.family == CHIP_FIJI ||
404            sscreen->info.family == CHIP_POLARIS10 || sscreen->info.family == CHIP_POLARIS11 ||
405            sscreen->info.family == CHIP_POLARIS12 || sscreen->info.family == CHIP_VEGAM))
406          partial_vs_wave = true;
407 
408       /* Required by Hawaii and, for some special cases, by GFX8. */
409       if (ia_switch_on_eoi &&
410           (sscreen->info.family == CHIP_HAWAII ||
411            (sscreen->info.chip_class == GFX8 && (key->u.uses_gs || max_primgroup_in_wave != 2))))
412          partial_vs_wave = true;
413 
414       /* Instancing bug on Bonaire. */
415       if (sscreen->info.family == CHIP_BONAIRE && ia_switch_on_eoi && key->u.uses_instancing)
416          partial_vs_wave = true;
417 
418       /* This only applies to Polaris10 and later 4 SE chips.
419        * wd_switch_on_eop is already true on all other chips.
420        */
421       if (!wd_switch_on_eop && key->u.primitive_restart)
422          partial_vs_wave = true;
423 
424       /* If the WD switch is false, the IA switch must be false too. */
425       assert(wd_switch_on_eop || !ia_switch_on_eop);
426    }
427 
428    /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
429    if (sscreen->info.chip_class <= GFX8 && ia_switch_on_eoi)
430       partial_es_wave = true;
431 
432    return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop) | S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi) |
433           S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave) |
434           S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave) |
435           S_028AA8_WD_SWITCH_ON_EOP(sscreen->info.chip_class >= GFX7 ? wd_switch_on_eop : 0) |
436           /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
437           S_028AA8_MAX_PRIMGRP_IN_WAVE(sscreen->info.chip_class == GFX8 ? max_primgroup_in_wave
438                                                                         : 0) |
439           S_030960_EN_INST_OPT_BASIC(sscreen->info.chip_class >= GFX9) |
440           S_030960_EN_INST_OPT_ADV(sscreen->info.chip_class >= GFX9);
441 }
442 
si_init_ia_multi_vgt_param_table(struct si_context * sctx)443 static void si_init_ia_multi_vgt_param_table(struct si_context *sctx)
444 {
445    for (int prim = 0; prim <= SI_PRIM_RECTANGLE_LIST; prim++)
446       for (int uses_instancing = 0; uses_instancing < 2; uses_instancing++)
447          for (int multi_instances = 0; multi_instances < 2; multi_instances++)
448             for (int primitive_restart = 0; primitive_restart < 2; primitive_restart++)
449                for (int count_from_so = 0; count_from_so < 2; count_from_so++)
450                   for (int line_stipple = 0; line_stipple < 2; line_stipple++)
451                      for (int uses_tess = 0; uses_tess < 2; uses_tess++)
452                         for (int tess_uses_primid = 0; tess_uses_primid < 2; tess_uses_primid++)
453                            for (int uses_gs = 0; uses_gs < 2; uses_gs++) {
454                               union si_vgt_param_key key;
455 
456                               key.index = 0;
457                               key.u.prim = prim;
458                               key.u.uses_instancing = uses_instancing;
459                               key.u.multi_instances_smaller_than_primgroup = multi_instances;
460                               key.u.primitive_restart = primitive_restart;
461                               key.u.count_from_stream_output = count_from_so;
462                               key.u.line_stipple_enabled = line_stipple;
463                               key.u.uses_tess = uses_tess;
464                               key.u.tess_uses_prim_id = tess_uses_primid;
465                               key.u.uses_gs = uses_gs;
466 
467                               sctx->ia_multi_vgt_param[key.index] =
468                                  si_get_init_multi_vgt_param(sctx->screen, &key);
469                            }
470 }
471 
si_is_line_stipple_enabled(struct si_context * sctx)472 static bool si_is_line_stipple_enabled(struct si_context *sctx)
473 {
474    struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
475 
476    return rs->line_stipple_enable && sctx->current_rast_prim != PIPE_PRIM_POINTS &&
477           (rs->polygon_mode_is_lines || util_prim_is_lines(sctx->current_rast_prim));
478 }
479 
480 ALWAYS_INLINE
si_get_ia_multi_vgt_param(struct si_context * sctx,const struct pipe_draw_info * info,enum pipe_prim_type prim,unsigned num_patches,unsigned instance_count,bool primitive_restart,unsigned min_vertex_count)481 static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
482                                           const struct pipe_draw_info *info,
483                                           enum pipe_prim_type prim, unsigned num_patches,
484                                           unsigned instance_count, bool primitive_restart,
485                                           unsigned min_vertex_count)
486 {
487    union si_vgt_param_key key = sctx->ia_multi_vgt_param_key;
488    unsigned primgroup_size;
489    unsigned ia_multi_vgt_param;
490 
491    if (sctx->tes_shader.cso) {
492       primgroup_size = num_patches; /* must be a multiple of NUM_PATCHES */
493    } else if (sctx->gs_shader.cso) {
494       primgroup_size = 64; /* recommended with a GS */
495    } else {
496       primgroup_size = 128; /* recommended without a GS and tess */
497    }
498 
499    key.u.prim = prim;
500    key.u.uses_instancing = info->indirect || instance_count > 1;
501    key.u.multi_instances_smaller_than_primgroup =
502       info->indirect ||
503       (instance_count > 1 &&
504        (info->count_from_stream_output ||
505         si_num_prims_for_vertices(prim, min_vertex_count, info->vertices_per_patch) < primgroup_size));
506    key.u.primitive_restart = primitive_restart;
507    key.u.count_from_stream_output = info->count_from_stream_output != NULL;
508    key.u.line_stipple_enabled = si_is_line_stipple_enabled(sctx);
509 
510    ia_multi_vgt_param =
511       sctx->ia_multi_vgt_param[key.index] | S_028AA8_PRIMGROUP_SIZE(primgroup_size - 1);
512 
513    if (sctx->gs_shader.cso) {
514       /* GS requirement. */
515       if (sctx->chip_class <= GFX8 &&
516           SI_GS_PER_ES / primgroup_size >= sctx->screen->gs_table_depth - 3)
517          ia_multi_vgt_param |= S_028AA8_PARTIAL_ES_WAVE_ON(1);
518 
519       /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
520        * The hw doc says all multi-SE chips are affected, but Vulkan
521        * only applies it to Hawaii. Do what Vulkan does.
522        */
523       if (sctx->family == CHIP_HAWAII && G_028AA8_SWITCH_ON_EOI(ia_multi_vgt_param) &&
524           (info->indirect ||
525            (instance_count > 1 &&
526             (info->count_from_stream_output ||
527              si_num_prims_for_vertices(prim, min_vertex_count, info->vertices_per_patch) <= 1))))
528 
529          sctx->flags |= SI_CONTEXT_VGT_FLUSH;
530    }
531 
532    return ia_multi_vgt_param;
533 }
534 
535 ALWAYS_INLINE
si_conv_prim_to_gs_out(unsigned mode)536 static unsigned si_conv_prim_to_gs_out(unsigned mode)
537 {
538    static const int prim_conv[] = {
539       [PIPE_PRIM_POINTS] = V_028A6C_POINTLIST,
540       [PIPE_PRIM_LINES] = V_028A6C_LINESTRIP,
541       [PIPE_PRIM_LINE_LOOP] = V_028A6C_LINESTRIP,
542       [PIPE_PRIM_LINE_STRIP] = V_028A6C_LINESTRIP,
543       [PIPE_PRIM_TRIANGLES] = V_028A6C_TRISTRIP,
544       [PIPE_PRIM_TRIANGLE_STRIP] = V_028A6C_TRISTRIP,
545       [PIPE_PRIM_TRIANGLE_FAN] = V_028A6C_TRISTRIP,
546       [PIPE_PRIM_QUADS] = V_028A6C_TRISTRIP,
547       [PIPE_PRIM_QUAD_STRIP] = V_028A6C_TRISTRIP,
548       [PIPE_PRIM_POLYGON] = V_028A6C_TRISTRIP,
549       [PIPE_PRIM_LINES_ADJACENCY] = V_028A6C_LINESTRIP,
550       [PIPE_PRIM_LINE_STRIP_ADJACENCY] = V_028A6C_LINESTRIP,
551       [PIPE_PRIM_TRIANGLES_ADJACENCY] = V_028A6C_TRISTRIP,
552       [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY] = V_028A6C_TRISTRIP,
553       [PIPE_PRIM_PATCHES] = V_028A6C_POINTLIST,
554       [SI_PRIM_RECTANGLE_LIST] = V_028A6C_RECTLIST,
555    };
556    assert(mode < ARRAY_SIZE(prim_conv));
557 
558    return prim_conv[mode];
559 }
560 
561 /* rast_prim is the primitive type after GS. */
562 ALWAYS_INLINE
si_emit_rasterizer_prim_state(struct si_context * sctx)563 static void si_emit_rasterizer_prim_state(struct si_context *sctx)
564 {
565    struct radeon_cmdbuf *cs = sctx->gfx_cs;
566    enum pipe_prim_type rast_prim = sctx->current_rast_prim;
567    struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
568    unsigned initial_cdw = cs->current.cdw;
569 
570    if (unlikely(si_is_line_stipple_enabled(sctx))) {
571       /* For lines, reset the stipple pattern at each primitive. Otherwise,
572        * reset the stipple pattern at each packet (line strips, line loops).
573        */
574       unsigned value =
575          rs->pa_sc_line_stipple | S_028A0C_AUTO_RESET_CNTL(rast_prim == PIPE_PRIM_LINES ? 1 : 2);
576 
577       radeon_opt_set_context_reg(sctx, R_028A0C_PA_SC_LINE_STIPPLE, SI_TRACKED_PA_SC_LINE_STIPPLE,
578                                  value);
579    }
580 
581    unsigned gs_out_prim = si_conv_prim_to_gs_out(rast_prim);
582    if (unlikely(gs_out_prim != sctx->last_gs_out_prim && (sctx->ngg || sctx->gs_shader.cso))) {
583       radeon_set_context_reg(cs, R_028A6C_VGT_GS_OUT_PRIM_TYPE, gs_out_prim);
584       sctx->last_gs_out_prim = gs_out_prim;
585    }
586 
587    if (initial_cdw != cs->current.cdw)
588       sctx->context_roll = true;
589 
590    if (sctx->ngg) {
591       unsigned vtx_index = rs->flatshade_first ? 0 : gs_out_prim;
592 
593       sctx->current_vs_state &= C_VS_STATE_OUTPRIM & C_VS_STATE_PROVOKING_VTX_INDEX;
594       sctx->current_vs_state |=
595          S_VS_STATE_OUTPRIM(gs_out_prim) | S_VS_STATE_PROVOKING_VTX_INDEX(vtx_index);
596    }
597 }
598 
599 ALWAYS_INLINE
si_emit_vs_state(struct si_context * sctx,const struct pipe_draw_info * info)600 static void si_emit_vs_state(struct si_context *sctx, const struct pipe_draw_info *info)
601 {
602    sctx->current_vs_state &= C_VS_STATE_INDEXED;
603    sctx->current_vs_state |= S_VS_STATE_INDEXED(!!info->index_size);
604 
605    if (sctx->num_vs_blit_sgprs) {
606       /* Re-emit the state after we leave u_blitter. */
607       sctx->last_vs_state = ~0;
608       return;
609    }
610 
611    if (sctx->current_vs_state != sctx->last_vs_state) {
612       struct radeon_cmdbuf *cs = sctx->gfx_cs;
613 
614       /* For the API vertex shader (VS_STATE_INDEXED, LS_OUT_*). */
615       radeon_set_sh_reg(
616          cs, sctx->shader_pointers.sh_base[PIPE_SHADER_VERTEX] + SI_SGPR_VS_STATE_BITS * 4,
617          sctx->current_vs_state);
618 
619       /* Set CLAMP_VERTEX_COLOR and OUTPRIM in the last stage
620        * before the rasterizer.
621        *
622        * For TES or the GS copy shader without NGG:
623        */
624       if (sctx->shader_pointers.sh_base[PIPE_SHADER_VERTEX] != R_00B130_SPI_SHADER_USER_DATA_VS_0) {
625          radeon_set_sh_reg(cs, R_00B130_SPI_SHADER_USER_DATA_VS_0 + SI_SGPR_VS_STATE_BITS * 4,
626                            sctx->current_vs_state);
627       }
628 
629       /* For NGG: */
630       if (sctx->screen->use_ngg &&
631           sctx->shader_pointers.sh_base[PIPE_SHADER_VERTEX] != R_00B230_SPI_SHADER_USER_DATA_GS_0) {
632          radeon_set_sh_reg(cs, R_00B230_SPI_SHADER_USER_DATA_GS_0 + SI_SGPR_VS_STATE_BITS * 4,
633                            sctx->current_vs_state);
634       }
635 
636       sctx->last_vs_state = sctx->current_vs_state;
637    }
638 }
639 
640 ALWAYS_INLINE
si_prim_restart_index_changed(struct si_context * sctx,bool primitive_restart,unsigned restart_index)641 static bool si_prim_restart_index_changed(struct si_context *sctx, bool primitive_restart,
642                                           unsigned restart_index)
643 {
644    return primitive_restart && (restart_index != sctx->last_restart_index ||
645                                 sctx->last_restart_index == SI_RESTART_INDEX_UNKNOWN);
646 }
647 
648 ALWAYS_INLINE
si_emit_ia_multi_vgt_param(struct si_context * sctx,const struct pipe_draw_info * info,enum pipe_prim_type prim,unsigned num_patches,unsigned instance_count,bool primitive_restart,unsigned min_vertex_count)649 static void si_emit_ia_multi_vgt_param(struct si_context *sctx, const struct pipe_draw_info *info,
650                                        enum pipe_prim_type prim, unsigned num_patches,
651                                        unsigned instance_count, bool primitive_restart,
652                                        unsigned min_vertex_count)
653 {
654    struct radeon_cmdbuf *cs = sctx->gfx_cs;
655    unsigned ia_multi_vgt_param;
656 
657    ia_multi_vgt_param =
658       si_get_ia_multi_vgt_param(sctx, info, prim, num_patches, instance_count, primitive_restart,
659                                 min_vertex_count);
660 
661    /* Draw state. */
662    if (ia_multi_vgt_param != sctx->last_multi_vgt_param) {
663       if (sctx->chip_class == GFX9)
664          radeon_set_uconfig_reg_idx(cs, sctx->screen, R_030960_IA_MULTI_VGT_PARAM, 4,
665                                     ia_multi_vgt_param);
666       else if (sctx->chip_class >= GFX7)
667          radeon_set_context_reg_idx(cs, R_028AA8_IA_MULTI_VGT_PARAM, 1, ia_multi_vgt_param);
668       else
669          radeon_set_context_reg(cs, R_028AA8_IA_MULTI_VGT_PARAM, ia_multi_vgt_param);
670 
671       sctx->last_multi_vgt_param = ia_multi_vgt_param;
672    }
673 }
674 
675 /* GFX10 removed IA_MULTI_VGT_PARAM in exchange for GE_CNTL.
676  * We overload last_multi_vgt_param.
677  */
678 ALWAYS_INLINE
gfx10_emit_ge_cntl(struct si_context * sctx,unsigned num_patches)679 static void gfx10_emit_ge_cntl(struct si_context *sctx, unsigned num_patches)
680 {
681    union si_vgt_param_key key = sctx->ia_multi_vgt_param_key;
682    unsigned ge_cntl;
683 
684    if (sctx->ngg) {
685       if (sctx->tes_shader.cso) {
686          ge_cntl = S_03096C_PRIM_GRP_SIZE(num_patches) |
687                    S_03096C_VERT_GRP_SIZE(0) |
688                    S_03096C_BREAK_WAVE_AT_EOI(key.u.tess_uses_prim_id);
689       } else {
690          ge_cntl = si_get_vs_state(sctx)->ge_cntl;
691       }
692    } else {
693       unsigned primgroup_size;
694       unsigned vertgroup_size;
695 
696       if (sctx->tes_shader.cso) {
697          primgroup_size = num_patches; /* must be a multiple of NUM_PATCHES */
698          vertgroup_size = 0;
699       } else if (sctx->gs_shader.cso) {
700          unsigned vgt_gs_onchip_cntl = sctx->gs_shader.current->ctx_reg.gs.vgt_gs_onchip_cntl;
701          primgroup_size = G_028A44_GS_PRIMS_PER_SUBGRP(vgt_gs_onchip_cntl);
702          vertgroup_size = G_028A44_ES_VERTS_PER_SUBGRP(vgt_gs_onchip_cntl);
703       } else {
704          primgroup_size = 128; /* recommended without a GS and tess */
705          vertgroup_size = 0;
706       }
707 
708       ge_cntl = S_03096C_PRIM_GRP_SIZE(primgroup_size) | S_03096C_VERT_GRP_SIZE(vertgroup_size) |
709                 S_03096C_BREAK_WAVE_AT_EOI(key.u.uses_tess && key.u.tess_uses_prim_id);
710    }
711 
712    ge_cntl |= S_03096C_PACKET_TO_ONE_PA(si_is_line_stipple_enabled(sctx));
713 
714    if (ge_cntl != sctx->last_multi_vgt_param) {
715       radeon_set_uconfig_reg(sctx->gfx_cs, R_03096C_GE_CNTL, ge_cntl);
716       sctx->last_multi_vgt_param = ge_cntl;
717    }
718 }
719 
720 ALWAYS_INLINE
si_emit_draw_registers(struct si_context * sctx,const struct pipe_draw_info * info,enum pipe_prim_type prim,unsigned num_patches,unsigned instance_count,bool primitive_restart,unsigned min_vertex_count)721 static void si_emit_draw_registers(struct si_context *sctx, const struct pipe_draw_info *info,
722                                    enum pipe_prim_type prim, unsigned num_patches,
723                                    unsigned instance_count, bool primitive_restart,
724                                    unsigned min_vertex_count)
725 {
726    struct radeon_cmdbuf *cs = sctx->gfx_cs;
727    unsigned vgt_prim = si_conv_pipe_prim(prim);
728 
729    if (sctx->chip_class >= GFX10)
730       gfx10_emit_ge_cntl(sctx, num_patches);
731    else
732       si_emit_ia_multi_vgt_param(sctx, info, prim, num_patches, instance_count, primitive_restart,
733                                  min_vertex_count);
734 
735    if (vgt_prim != sctx->last_prim) {
736       if (sctx->chip_class >= GFX10)
737          radeon_set_uconfig_reg(cs, R_030908_VGT_PRIMITIVE_TYPE, vgt_prim);
738       else if (sctx->chip_class >= GFX7)
739          radeon_set_uconfig_reg_idx(cs, sctx->screen, R_030908_VGT_PRIMITIVE_TYPE, 1, vgt_prim);
740       else
741          radeon_set_config_reg(cs, R_008958_VGT_PRIMITIVE_TYPE, vgt_prim);
742 
743       sctx->last_prim = vgt_prim;
744    }
745 
746    /* Primitive restart. */
747    if (primitive_restart != sctx->last_primitive_restart_en) {
748       if (sctx->chip_class >= GFX9)
749          radeon_set_uconfig_reg(cs, R_03092C_VGT_MULTI_PRIM_IB_RESET_EN, primitive_restart);
750       else
751          radeon_set_context_reg(cs, R_028A94_VGT_MULTI_PRIM_IB_RESET_EN, primitive_restart);
752 
753       sctx->last_primitive_restart_en = primitive_restart;
754    }
755    if (si_prim_restart_index_changed(sctx, primitive_restart, info->restart_index)) {
756       radeon_set_context_reg(cs, R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX, info->restart_index);
757       sctx->last_restart_index = info->restart_index;
758       sctx->context_roll = true;
759    }
760 }
761 
si_emit_draw_packets(struct si_context * sctx,const struct pipe_draw_info * info,const struct pipe_draw_start_count * draws,unsigned num_draws,struct pipe_resource * indexbuf,unsigned index_size,unsigned index_offset,unsigned instance_count,bool dispatch_prim_discard_cs,unsigned original_index_size)762 static void si_emit_draw_packets(struct si_context *sctx, const struct pipe_draw_info *info,
763                                  const struct pipe_draw_start_count *draws,
764                                  unsigned num_draws,
765                                  struct pipe_resource *indexbuf, unsigned index_size,
766                                  unsigned index_offset, unsigned instance_count,
767                                  bool dispatch_prim_discard_cs, unsigned original_index_size)
768 {
769    struct pipe_draw_indirect_info *indirect = info->indirect;
770    struct radeon_cmdbuf *cs = sctx->gfx_cs;
771    unsigned sh_base_reg = sctx->shader_pointers.sh_base[PIPE_SHADER_VERTEX];
772    bool render_cond_bit = sctx->render_cond && !sctx->render_cond_force_off;
773    uint32_t index_max_size = 0;
774    uint64_t index_va = 0;
775 
776    if (info->count_from_stream_output) {
777       struct si_streamout_target *t = (struct si_streamout_target *)info->count_from_stream_output;
778 
779       radeon_set_context_reg(cs, R_028B30_VGT_STRMOUT_DRAW_OPAQUE_VERTEX_STRIDE, t->stride_in_dw);
780       si_cp_copy_data(sctx, sctx->gfx_cs, COPY_DATA_REG, NULL,
781                       R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE >> 2, COPY_DATA_SRC_MEM,
782                       t->buf_filled_size, t->buf_filled_size_offset);
783    }
784 
785    /* draw packet */
786    if (index_size) {
787       /* Register shadowing doesn't shadow INDEX_TYPE. */
788       if (index_size != sctx->last_index_size || sctx->shadowed_regs) {
789          unsigned index_type;
790 
791          /* index type */
792          switch (index_size) {
793          case 1:
794             index_type = V_028A7C_VGT_INDEX_8;
795             break;
796          case 2:
797             index_type =
798                V_028A7C_VGT_INDEX_16 |
799                (SI_BIG_ENDIAN && sctx->chip_class <= GFX7 ? V_028A7C_VGT_DMA_SWAP_16_BIT : 0);
800             break;
801          case 4:
802             index_type =
803                V_028A7C_VGT_INDEX_32 |
804                (SI_BIG_ENDIAN && sctx->chip_class <= GFX7 ? V_028A7C_VGT_DMA_SWAP_32_BIT : 0);
805             break;
806          default:
807             assert(!"unreachable");
808             return;
809          }
810 
811          if (sctx->chip_class >= GFX9) {
812             radeon_set_uconfig_reg_idx(cs, sctx->screen, R_03090C_VGT_INDEX_TYPE, 2, index_type);
813          } else {
814             radeon_emit(cs, PKT3(PKT3_INDEX_TYPE, 0, 0));
815             radeon_emit(cs, index_type);
816          }
817 
818          sctx->last_index_size = index_size;
819       }
820 
821       if (original_index_size) {
822          index_max_size = (indexbuf->width0 - index_offset) / original_index_size;
823          /* Skip draw calls with 0-sized index buffers.
824           * They cause a hang on some chips, like Navi10-14.
825           */
826          if (!index_max_size)
827             return;
828 
829          index_va = si_resource(indexbuf)->gpu_address + index_offset;
830 
831          radeon_add_to_buffer_list(sctx, sctx->gfx_cs, si_resource(indexbuf), RADEON_USAGE_READ,
832                                    RADEON_PRIO_INDEX_BUFFER);
833       }
834    } else {
835       /* On GFX7 and later, non-indexed draws overwrite VGT_INDEX_TYPE,
836        * so the state must be re-emitted before the next indexed draw.
837        */
838       if (sctx->chip_class >= GFX7)
839          sctx->last_index_size = -1;
840    }
841 
842    if (indirect) {
843       assert(num_draws == 1);
844       uint64_t indirect_va = si_resource(indirect->buffer)->gpu_address;
845 
846       assert(indirect_va % 8 == 0);
847 
848       si_invalidate_draw_sh_constants(sctx);
849 
850       radeon_emit(cs, PKT3(PKT3_SET_BASE, 2, 0));
851       radeon_emit(cs, 1);
852       radeon_emit(cs, indirect_va);
853       radeon_emit(cs, indirect_va >> 32);
854 
855       radeon_add_to_buffer_list(sctx, sctx->gfx_cs, si_resource(indirect->buffer),
856                                 RADEON_USAGE_READ, RADEON_PRIO_DRAW_INDIRECT);
857 
858       unsigned di_src_sel = index_size ? V_0287F0_DI_SRC_SEL_DMA : V_0287F0_DI_SRC_SEL_AUTO_INDEX;
859 
860       assert(indirect->offset % 4 == 0);
861 
862       if (index_size) {
863          radeon_emit(cs, PKT3(PKT3_INDEX_BASE, 1, 0));
864          radeon_emit(cs, index_va);
865          radeon_emit(cs, index_va >> 32);
866 
867          radeon_emit(cs, PKT3(PKT3_INDEX_BUFFER_SIZE, 0, 0));
868          radeon_emit(cs, index_max_size);
869       }
870 
871       if (!sctx->screen->has_draw_indirect_multi) {
872          radeon_emit(cs, PKT3(index_size ? PKT3_DRAW_INDEX_INDIRECT : PKT3_DRAW_INDIRECT, 3,
873                               render_cond_bit));
874          radeon_emit(cs, indirect->offset);
875          radeon_emit(cs, (sh_base_reg + SI_SGPR_BASE_VERTEX * 4 - SI_SH_REG_OFFSET) >> 2);
876          radeon_emit(cs, (sh_base_reg + SI_SGPR_START_INSTANCE * 4 - SI_SH_REG_OFFSET) >> 2);
877          radeon_emit(cs, di_src_sel);
878       } else {
879          uint64_t count_va = 0;
880 
881          if (indirect->indirect_draw_count) {
882             struct si_resource *params_buf = si_resource(indirect->indirect_draw_count);
883 
884             radeon_add_to_buffer_list(sctx, sctx->gfx_cs, params_buf, RADEON_USAGE_READ,
885                                       RADEON_PRIO_DRAW_INDIRECT);
886 
887             count_va = params_buf->gpu_address + indirect->indirect_draw_count_offset;
888          }
889 
890          radeon_emit(cs,
891                      PKT3(index_size ? PKT3_DRAW_INDEX_INDIRECT_MULTI : PKT3_DRAW_INDIRECT_MULTI, 8,
892                           render_cond_bit));
893          radeon_emit(cs, indirect->offset);
894          radeon_emit(cs, (sh_base_reg + SI_SGPR_BASE_VERTEX * 4 - SI_SH_REG_OFFSET) >> 2);
895          radeon_emit(cs, (sh_base_reg + SI_SGPR_START_INSTANCE * 4 - SI_SH_REG_OFFSET) >> 2);
896          radeon_emit(cs, ((sh_base_reg + SI_SGPR_DRAWID * 4 - SI_SH_REG_OFFSET) >> 2) |
897                             S_2C3_DRAW_INDEX_ENABLE(1) |
898                             S_2C3_COUNT_INDIRECT_ENABLE(!!indirect->indirect_draw_count));
899          radeon_emit(cs, indirect->draw_count);
900          radeon_emit(cs, count_va);
901          radeon_emit(cs, count_va >> 32);
902          radeon_emit(cs, indirect->stride);
903          radeon_emit(cs, di_src_sel);
904       }
905    } else {
906       int base_vertex;
907 
908       /* Register shadowing requires that we always emit PKT3_NUM_INSTANCES. */
909       if (sctx->shadowed_regs ||
910           sctx->last_instance_count == SI_INSTANCE_COUNT_UNKNOWN ||
911           sctx->last_instance_count != instance_count) {
912          radeon_emit(cs, PKT3(PKT3_NUM_INSTANCES, 0, 0));
913          radeon_emit(cs, instance_count);
914          sctx->last_instance_count = instance_count;
915       }
916 
917       /* Base vertex and start instance. */
918       base_vertex = original_index_size ? info->index_bias : draws[0].start;
919 
920       if (sctx->num_vs_blit_sgprs) {
921          /* Re-emit draw constants after we leave u_blitter. */
922          si_invalidate_draw_sh_constants(sctx);
923 
924          /* Blit VS doesn't use BASE_VERTEX, START_INSTANCE, and DRAWID. */
925          radeon_set_sh_reg_seq(cs, sh_base_reg + SI_SGPR_VS_BLIT_DATA * 4, sctx->num_vs_blit_sgprs);
926          radeon_emit_array(cs, sctx->vs_blit_sh_data, sctx->num_vs_blit_sgprs);
927       } else if (base_vertex != sctx->last_base_vertex ||
928                  sctx->last_base_vertex == SI_BASE_VERTEX_UNKNOWN ||
929                  info->start_instance != sctx->last_start_instance ||
930                  info->drawid != sctx->last_drawid || sh_base_reg != sctx->last_sh_base_reg) {
931          radeon_set_sh_reg_seq(cs, sh_base_reg + SI_SGPR_BASE_VERTEX * 4, 3);
932          radeon_emit(cs, base_vertex);
933          radeon_emit(cs, info->start_instance);
934          radeon_emit(cs, info->drawid);
935 
936          sctx->last_base_vertex = base_vertex;
937          sctx->last_start_instance = info->start_instance;
938          sctx->last_drawid = info->drawid;
939          sctx->last_sh_base_reg = sh_base_reg;
940       }
941 
942       if (index_size) {
943          if (dispatch_prim_discard_cs) {
944             for (unsigned i = 0; i < num_draws; i++) {
945                uint64_t va = index_va + draws[0].start * original_index_size;
946 
947                si_dispatch_prim_discard_cs_and_draw(sctx, info, draws[i].count,
948                                                     original_index_size, base_vertex,
949                                                     va, MIN2(index_max_size, draws[i].count));
950             }
951             return;
952          }
953 
954          for (unsigned i = 0; i < num_draws; i++) {
955             uint64_t va = index_va + draws[i].start * index_size;
956 
957             radeon_emit(cs, PKT3(PKT3_DRAW_INDEX_2, 4, render_cond_bit));
958             radeon_emit(cs, index_max_size);
959             radeon_emit(cs, va);
960             radeon_emit(cs, va >> 32);
961             radeon_emit(cs, draws[i].count);
962             radeon_emit(cs, V_0287F0_DI_SRC_SEL_DMA |
963                         /* NOT_EOP allows merging multiple draws into 1 wave, but only user VGPRs
964                          * can be changed between draws and GS fast launch must be disabled.
965                          * NOT_EOP doesn't work on gfx9 and older.
966                          */
967                         S_0287F0_NOT_EOP(sctx->chip_class >= GFX10 &&
968                                          i < num_draws - 1 &&
969                                          !(sctx->ngg_culling & SI_NGG_CULL_GS_FAST_LAUNCH_ALL)));
970          }
971       } else {
972          for (unsigned i = 0; i < num_draws; i++) {
973             if (i > 0)
974                radeon_set_sh_reg(cs, sh_base_reg + SI_SGPR_BASE_VERTEX * 4, draws[i].start);
975 
976             radeon_emit(cs, PKT3(PKT3_DRAW_INDEX_AUTO, 1, render_cond_bit));
977             radeon_emit(cs, draws[i].count);
978             radeon_emit(cs, V_0287F0_DI_SRC_SEL_AUTO_INDEX |
979                             S_0287F0_USE_OPAQUE(!!info->count_from_stream_output));
980          }
981          if (num_draws > 1 && !sctx->num_vs_blit_sgprs)
982             sctx->last_base_vertex = draws[num_draws - 1].start;
983       }
984    }
985 }
986 
si_emit_surface_sync(struct si_context * sctx,struct radeon_cmdbuf * cs,unsigned cp_coher_cntl)987 void si_emit_surface_sync(struct si_context *sctx, struct radeon_cmdbuf *cs, unsigned cp_coher_cntl)
988 {
989    bool compute_ib = !sctx->has_graphics || cs == sctx->prim_discard_compute_cs;
990 
991    assert(sctx->chip_class <= GFX9);
992 
993    if (sctx->chip_class == GFX9 || compute_ib) {
994       /* Flush caches and wait for the caches to assert idle. */
995       radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, 0));
996       radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
997       radeon_emit(cs, 0xffffffff);    /* CP_COHER_SIZE */
998       radeon_emit(cs, 0xffffff);      /* CP_COHER_SIZE_HI */
999       radeon_emit(cs, 0);             /* CP_COHER_BASE */
1000       radeon_emit(cs, 0);             /* CP_COHER_BASE_HI */
1001       radeon_emit(cs, 0x0000000A);    /* POLL_INTERVAL */
1002    } else {
1003       /* ACQUIRE_MEM is only required on a compute ring. */
1004       radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, 0));
1005       radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
1006       radeon_emit(cs, 0xffffffff);    /* CP_COHER_SIZE */
1007       radeon_emit(cs, 0);             /* CP_COHER_BASE */
1008       radeon_emit(cs, 0x0000000A);    /* POLL_INTERVAL */
1009    }
1010 
1011    /* ACQUIRE_MEM has an implicit context roll if the current context
1012     * is busy. */
1013    if (!compute_ib)
1014       sctx->context_roll = true;
1015 }
1016 
si_prim_discard_signal_next_compute_ib_start(struct si_context * sctx)1017 void si_prim_discard_signal_next_compute_ib_start(struct si_context *sctx)
1018 {
1019    if (!si_compute_prim_discard_enabled(sctx))
1020       return;
1021 
1022    if (!sctx->barrier_buf) {
1023       u_suballocator_alloc(sctx->allocator_zeroed_memory, 4, 4, &sctx->barrier_buf_offset,
1024                            (struct pipe_resource **)&sctx->barrier_buf);
1025    }
1026 
1027    /* Emit a placeholder to signal the next compute IB to start.
1028     * See si_compute_prim_discard.c for explanation.
1029     */
1030    uint32_t signal = 1;
1031    si_cp_write_data(sctx, sctx->barrier_buf, sctx->barrier_buf_offset, 4, V_370_MEM, V_370_ME,
1032                     &signal);
1033 
1034    sctx->last_pkt3_write_data = &sctx->gfx_cs->current.buf[sctx->gfx_cs->current.cdw - 5];
1035 
1036    /* Only the last occurence of WRITE_DATA will be executed.
1037     * The packet will be enabled in si_flush_gfx_cs.
1038     */
1039    *sctx->last_pkt3_write_data = PKT3(PKT3_NOP, 3, 0);
1040 }
1041 
gfx10_emit_cache_flush(struct si_context * ctx)1042 void gfx10_emit_cache_flush(struct si_context *ctx)
1043 {
1044    struct radeon_cmdbuf *cs = ctx->gfx_cs;
1045    uint32_t gcr_cntl = 0;
1046    unsigned cb_db_event = 0;
1047    unsigned flags = ctx->flags;
1048 
1049    if (!ctx->has_graphics) {
1050       /* Only process compute flags. */
1051       flags &= SI_CONTEXT_INV_ICACHE | SI_CONTEXT_INV_SCACHE | SI_CONTEXT_INV_VCACHE |
1052                SI_CONTEXT_INV_L2 | SI_CONTEXT_WB_L2 | SI_CONTEXT_INV_L2_METADATA |
1053                SI_CONTEXT_CS_PARTIAL_FLUSH;
1054    }
1055 
1056    /* We don't need these. */
1057    assert(!(flags & (SI_CONTEXT_VGT_STREAMOUT_SYNC | SI_CONTEXT_FLUSH_AND_INV_DB_META)));
1058 
1059    if (flags & SI_CONTEXT_VGT_FLUSH) {
1060       radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1061       radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
1062    }
1063 
1064    if (flags & SI_CONTEXT_FLUSH_AND_INV_CB)
1065       ctx->num_cb_cache_flushes++;
1066    if (flags & SI_CONTEXT_FLUSH_AND_INV_DB)
1067       ctx->num_db_cache_flushes++;
1068 
1069    if (flags & SI_CONTEXT_INV_ICACHE)
1070       gcr_cntl |= S_586_GLI_INV(V_586_GLI_ALL);
1071    if (flags & SI_CONTEXT_INV_SCACHE) {
1072       /* TODO: When writing to the SMEM L1 cache, we need to set SEQ
1073        * to FORWARD when both L1 and L2 are written out (WB or INV).
1074        */
1075       gcr_cntl |= S_586_GL1_INV(1) | S_586_GLK_INV(1);
1076    }
1077    if (flags & SI_CONTEXT_INV_VCACHE)
1078       gcr_cntl |= S_586_GL1_INV(1) | S_586_GLV_INV(1);
1079 
1080    /* The L2 cache ops are:
1081     * - INV: - invalidate lines that reflect memory (were loaded from memory)
1082     *        - don't touch lines that were overwritten (were stored by gfx clients)
1083     * - WB: - don't touch lines that reflect memory
1084     *       - write back lines that were overwritten
1085     * - WB | INV: - invalidate lines that reflect memory
1086     *             - write back lines that were overwritten
1087     *
1088     * GLM doesn't support WB alone. If WB is set, INV must be set too.
1089     */
1090    if (flags & SI_CONTEXT_INV_L2) {
1091       /* Writeback and invalidate everything in L2. */
1092       gcr_cntl |= S_586_GL2_INV(1) | S_586_GL2_WB(1) | S_586_GLM_INV(1) | S_586_GLM_WB(1);
1093       ctx->num_L2_invalidates++;
1094    } else if (flags & SI_CONTEXT_WB_L2) {
1095       gcr_cntl |= S_586_GL2_WB(1) | S_586_GLM_WB(1) | S_586_GLM_INV(1);
1096    } else if (flags & SI_CONTEXT_INV_L2_METADATA) {
1097       gcr_cntl |= S_586_GLM_INV(1) | S_586_GLM_WB(1);
1098    }
1099 
1100    if (flags & (SI_CONTEXT_FLUSH_AND_INV_CB | SI_CONTEXT_FLUSH_AND_INV_DB)) {
1101       if (flags & SI_CONTEXT_FLUSH_AND_INV_CB) {
1102          /* Flush CMASK/FMASK/DCC. Will wait for idle later. */
1103          radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1104          radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
1105       }
1106       if (flags & SI_CONTEXT_FLUSH_AND_INV_DB) {
1107          /* Flush HTILE. Will wait for idle later. */
1108          radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1109          radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
1110       }
1111 
1112       /* First flush CB/DB, then L1/L2. */
1113       gcr_cntl |= S_586_SEQ(V_586_SEQ_FORWARD);
1114 
1115       if ((flags & (SI_CONTEXT_FLUSH_AND_INV_CB | SI_CONTEXT_FLUSH_AND_INV_DB)) ==
1116           (SI_CONTEXT_FLUSH_AND_INV_CB | SI_CONTEXT_FLUSH_AND_INV_DB)) {
1117          cb_db_event = V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT;
1118       } else if (flags & SI_CONTEXT_FLUSH_AND_INV_CB) {
1119          cb_db_event = V_028A90_FLUSH_AND_INV_CB_DATA_TS;
1120       } else if (flags & SI_CONTEXT_FLUSH_AND_INV_DB) {
1121          cb_db_event = V_028A90_FLUSH_AND_INV_DB_DATA_TS;
1122       } else {
1123          assert(0);
1124       }
1125    } else {
1126       /* Wait for graphics shaders to go idle if requested. */
1127       if (flags & SI_CONTEXT_PS_PARTIAL_FLUSH) {
1128          radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1129          radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
1130          /* Only count explicit shader flushes, not implicit ones. */
1131          ctx->num_vs_flushes++;
1132          ctx->num_ps_flushes++;
1133       } else if (flags & SI_CONTEXT_VS_PARTIAL_FLUSH) {
1134          radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1135          radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
1136          ctx->num_vs_flushes++;
1137       }
1138    }
1139 
1140    if (flags & SI_CONTEXT_CS_PARTIAL_FLUSH && ctx->compute_is_busy) {
1141       radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1142       radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH | EVENT_INDEX(4)));
1143       ctx->num_cs_flushes++;
1144       ctx->compute_is_busy = false;
1145    }
1146 
1147    if (cb_db_event) {
1148       struct si_resource* wait_mem_scratch = unlikely(ctx->ws->cs_is_secure(cs)) ?
1149         ctx->wait_mem_scratch_tmz : ctx->wait_mem_scratch;
1150       /* CB/DB flush and invalidate (or possibly just a wait for a
1151        * meta flush) via RELEASE_MEM.
1152        *
1153        * Combine this with other cache flushes when possible; this
1154        * requires affected shaders to be idle, so do it after the
1155        * CS_PARTIAL_FLUSH before (VS/PS partial flushes are always
1156        * implied).
1157        */
1158       uint64_t va;
1159 
1160       /* Do the flush (enqueue the event and wait for it). */
1161       va = wait_mem_scratch->gpu_address;
1162       ctx->wait_mem_number++;
1163 
1164       /* Get GCR_CNTL fields, because the encoding is different in RELEASE_MEM. */
1165       unsigned glm_wb = G_586_GLM_WB(gcr_cntl);
1166       unsigned glm_inv = G_586_GLM_INV(gcr_cntl);
1167       unsigned glv_inv = G_586_GLV_INV(gcr_cntl);
1168       unsigned gl1_inv = G_586_GL1_INV(gcr_cntl);
1169       assert(G_586_GL2_US(gcr_cntl) == 0);
1170       assert(G_586_GL2_RANGE(gcr_cntl) == 0);
1171       assert(G_586_GL2_DISCARD(gcr_cntl) == 0);
1172       unsigned gl2_inv = G_586_GL2_INV(gcr_cntl);
1173       unsigned gl2_wb = G_586_GL2_WB(gcr_cntl);
1174       unsigned gcr_seq = G_586_SEQ(gcr_cntl);
1175 
1176       gcr_cntl &= C_586_GLM_WB & C_586_GLM_INV & C_586_GLV_INV & C_586_GL1_INV & C_586_GL2_INV &
1177                   C_586_GL2_WB; /* keep SEQ */
1178 
1179       si_cp_release_mem(ctx, cs, cb_db_event,
1180                         S_490_GLM_WB(glm_wb) | S_490_GLM_INV(glm_inv) | S_490_GLV_INV(glv_inv) |
1181                            S_490_GL1_INV(gl1_inv) | S_490_GL2_INV(gl2_inv) | S_490_GL2_WB(gl2_wb) |
1182                            S_490_SEQ(gcr_seq),
1183                         EOP_DST_SEL_MEM, EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM,
1184                         EOP_DATA_SEL_VALUE_32BIT, wait_mem_scratch, va, ctx->wait_mem_number,
1185                         SI_NOT_QUERY);
1186       si_cp_wait_mem(ctx, ctx->gfx_cs, va, ctx->wait_mem_number, 0xffffffff, WAIT_REG_MEM_EQUAL);
1187    }
1188 
1189    /* Ignore fields that only modify the behavior of other fields. */
1190    if (gcr_cntl & C_586_GL1_RANGE & C_586_GL2_RANGE & C_586_SEQ) {
1191       /* Flush caches and wait for the caches to assert idle.
1192        * The cache flush is executed in the ME, but the PFP waits
1193        * for completion.
1194        */
1195       radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 6, 0));
1196       radeon_emit(cs, 0);          /* CP_COHER_CNTL */
1197       radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
1198       radeon_emit(cs, 0xffffff);   /* CP_COHER_SIZE_HI */
1199       radeon_emit(cs, 0);          /* CP_COHER_BASE */
1200       radeon_emit(cs, 0);          /* CP_COHER_BASE_HI */
1201       radeon_emit(cs, 0x0000000A); /* POLL_INTERVAL */
1202       radeon_emit(cs, gcr_cntl);   /* GCR_CNTL */
1203    } else if (cb_db_event || (flags & (SI_CONTEXT_VS_PARTIAL_FLUSH | SI_CONTEXT_PS_PARTIAL_FLUSH |
1204                                        SI_CONTEXT_CS_PARTIAL_FLUSH))) {
1205       /* We need to ensure that PFP waits as well. */
1206       radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
1207       radeon_emit(cs, 0);
1208    }
1209 
1210    if (flags & SI_CONTEXT_START_PIPELINE_STATS) {
1211       radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1212       radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_START) | EVENT_INDEX(0));
1213    } else if (flags & SI_CONTEXT_STOP_PIPELINE_STATS) {
1214       radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1215       radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP) | EVENT_INDEX(0));
1216    }
1217 
1218    ctx->flags = 0;
1219 }
1220 
si_emit_cache_flush(struct si_context * sctx)1221 void si_emit_cache_flush(struct si_context *sctx)
1222 {
1223    struct radeon_cmdbuf *cs = sctx->gfx_cs;
1224    uint32_t flags = sctx->flags;
1225 
1226    if (!sctx->has_graphics) {
1227       /* Only process compute flags. */
1228       flags &= SI_CONTEXT_INV_ICACHE | SI_CONTEXT_INV_SCACHE | SI_CONTEXT_INV_VCACHE |
1229                SI_CONTEXT_INV_L2 | SI_CONTEXT_WB_L2 | SI_CONTEXT_INV_L2_METADATA |
1230                SI_CONTEXT_CS_PARTIAL_FLUSH;
1231    }
1232 
1233    uint32_t cp_coher_cntl = 0;
1234    const uint32_t flush_cb_db = flags & (SI_CONTEXT_FLUSH_AND_INV_CB | SI_CONTEXT_FLUSH_AND_INV_DB);
1235    const bool is_barrier =
1236       flush_cb_db ||
1237       /* INV_ICACHE == beginning of gfx IB. Checking
1238        * INV_ICACHE fixes corruption for DeusExMD with
1239        * compute-based culling, but I don't know why.
1240        */
1241       flags & (SI_CONTEXT_INV_ICACHE | SI_CONTEXT_PS_PARTIAL_FLUSH | SI_CONTEXT_VS_PARTIAL_FLUSH) ||
1242       (flags & SI_CONTEXT_CS_PARTIAL_FLUSH && sctx->compute_is_busy);
1243 
1244    assert(sctx->chip_class <= GFX9);
1245 
1246    if (flags & SI_CONTEXT_FLUSH_AND_INV_CB)
1247       sctx->num_cb_cache_flushes++;
1248    if (flags & SI_CONTEXT_FLUSH_AND_INV_DB)
1249       sctx->num_db_cache_flushes++;
1250 
1251    /* GFX6 has a bug that it always flushes ICACHE and KCACHE if either
1252     * bit is set. An alternative way is to write SQC_CACHES, but that
1253     * doesn't seem to work reliably. Since the bug doesn't affect
1254     * correctness (it only does more work than necessary) and
1255     * the performance impact is likely negligible, there is no plan
1256     * to add a workaround for it.
1257     */
1258 
1259    if (flags & SI_CONTEXT_INV_ICACHE)
1260       cp_coher_cntl |= S_0085F0_SH_ICACHE_ACTION_ENA(1);
1261    if (flags & SI_CONTEXT_INV_SCACHE)
1262       cp_coher_cntl |= S_0085F0_SH_KCACHE_ACTION_ENA(1);
1263 
1264    if (sctx->chip_class <= GFX8) {
1265       if (flags & SI_CONTEXT_FLUSH_AND_INV_CB) {
1266          cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) | S_0085F0_CB0_DEST_BASE_ENA(1) |
1267                           S_0085F0_CB1_DEST_BASE_ENA(1) | S_0085F0_CB2_DEST_BASE_ENA(1) |
1268                           S_0085F0_CB3_DEST_BASE_ENA(1) | S_0085F0_CB4_DEST_BASE_ENA(1) |
1269                           S_0085F0_CB5_DEST_BASE_ENA(1) | S_0085F0_CB6_DEST_BASE_ENA(1) |
1270                           S_0085F0_CB7_DEST_BASE_ENA(1);
1271 
1272          /* Necessary for DCC */
1273          if (sctx->chip_class == GFX8)
1274             si_cp_release_mem(sctx, cs, V_028A90_FLUSH_AND_INV_CB_DATA_TS, 0, EOP_DST_SEL_MEM,
1275                               EOP_INT_SEL_NONE, EOP_DATA_SEL_DISCARD, NULL, 0, 0, SI_NOT_QUERY);
1276       }
1277       if (flags & SI_CONTEXT_FLUSH_AND_INV_DB)
1278          cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) | S_0085F0_DB_DEST_BASE_ENA(1);
1279    }
1280 
1281    if (flags & SI_CONTEXT_FLUSH_AND_INV_CB) {
1282       /* Flush CMASK/FMASK/DCC. SURFACE_SYNC will wait for idle. */
1283       radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1284       radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
1285    }
1286    if (flags & (SI_CONTEXT_FLUSH_AND_INV_DB | SI_CONTEXT_FLUSH_AND_INV_DB_META)) {
1287       /* Flush HTILE. SURFACE_SYNC will wait for idle. */
1288       radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1289       radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
1290    }
1291 
1292    /* Wait for shader engines to go idle.
1293     * VS and PS waits are unnecessary if SURFACE_SYNC is going to wait
1294     * for everything including CB/DB cache flushes.
1295     */
1296    if (!flush_cb_db) {
1297       if (flags & SI_CONTEXT_PS_PARTIAL_FLUSH) {
1298          radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1299          radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
1300          /* Only count explicit shader flushes, not implicit ones
1301           * done by SURFACE_SYNC.
1302           */
1303          sctx->num_vs_flushes++;
1304          sctx->num_ps_flushes++;
1305       } else if (flags & SI_CONTEXT_VS_PARTIAL_FLUSH) {
1306          radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1307          radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
1308          sctx->num_vs_flushes++;
1309       }
1310    }
1311 
1312    if (flags & SI_CONTEXT_CS_PARTIAL_FLUSH && sctx->compute_is_busy) {
1313       radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1314       radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH) | EVENT_INDEX(4));
1315       sctx->num_cs_flushes++;
1316       sctx->compute_is_busy = false;
1317    }
1318 
1319    /* VGT state synchronization. */
1320    if (flags & SI_CONTEXT_VGT_FLUSH) {
1321       radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1322       radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
1323    }
1324    if (flags & SI_CONTEXT_VGT_STREAMOUT_SYNC) {
1325       radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1326       radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC) | EVENT_INDEX(0));
1327    }
1328 
1329    /* GFX9: Wait for idle if we're flushing CB or DB. ACQUIRE_MEM doesn't
1330     * wait for idle on GFX9. We have to use a TS event.
1331     */
1332    if (sctx->chip_class == GFX9 && flush_cb_db) {
1333       uint64_t va;
1334       unsigned tc_flags, cb_db_event;
1335 
1336       /* Set the CB/DB flush event. */
1337       switch (flush_cb_db) {
1338       case SI_CONTEXT_FLUSH_AND_INV_CB:
1339          cb_db_event = V_028A90_FLUSH_AND_INV_CB_DATA_TS;
1340          break;
1341       case SI_CONTEXT_FLUSH_AND_INV_DB:
1342          cb_db_event = V_028A90_FLUSH_AND_INV_DB_DATA_TS;
1343          break;
1344       default:
1345          /* both CB & DB */
1346          cb_db_event = V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT;
1347       }
1348 
1349       /* These are the only allowed combinations. If you need to
1350        * do multiple operations at once, do them separately.
1351        * All operations that invalidate L2 also seem to invalidate
1352        * metadata. Volatile (VOL) and WC flushes are not listed here.
1353        *
1354        * TC    | TC_WB         = writeback & invalidate L2 & L1
1355        * TC    | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
1356        *         TC_WB | TC_NC = writeback L2 for MTYPE == NC
1357        * TC            | TC_NC = invalidate L2 for MTYPE == NC
1358        * TC    | TC_MD         = writeback & invalidate L2 metadata (DCC, etc.)
1359        * TCL1                  = invalidate L1
1360        */
1361       tc_flags = 0;
1362 
1363       if (flags & SI_CONTEXT_INV_L2_METADATA) {
1364          tc_flags = EVENT_TC_ACTION_ENA | EVENT_TC_MD_ACTION_ENA;
1365       }
1366 
1367       /* Ideally flush TC together with CB/DB. */
1368       if (flags & SI_CONTEXT_INV_L2) {
1369          /* Writeback and invalidate everything in L2 & L1. */
1370          tc_flags = EVENT_TC_ACTION_ENA | EVENT_TC_WB_ACTION_ENA;
1371 
1372          /* Clear the flags. */
1373          flags &= ~(SI_CONTEXT_INV_L2 | SI_CONTEXT_WB_L2 | SI_CONTEXT_INV_VCACHE);
1374          sctx->num_L2_invalidates++;
1375       }
1376 
1377       /* Do the flush (enqueue the event and wait for it). */
1378       struct si_resource* wait_mem_scratch = unlikely(sctx->ws->cs_is_secure(cs)) ?
1379         sctx->wait_mem_scratch_tmz : sctx->wait_mem_scratch;
1380       va = wait_mem_scratch->gpu_address;
1381       sctx->wait_mem_number++;
1382 
1383       si_cp_release_mem(sctx, cs, cb_db_event, tc_flags, EOP_DST_SEL_MEM,
1384                         EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM, EOP_DATA_SEL_VALUE_32BIT,
1385                         wait_mem_scratch, va, sctx->wait_mem_number, SI_NOT_QUERY);
1386       si_cp_wait_mem(sctx, cs, va, sctx->wait_mem_number, 0xffffffff, WAIT_REG_MEM_EQUAL);
1387    }
1388 
1389    /* Make sure ME is idle (it executes most packets) before continuing.
1390     * This prevents read-after-write hazards between PFP and ME.
1391     */
1392    if (sctx->has_graphics &&
1393        (cp_coher_cntl || (flags & (SI_CONTEXT_CS_PARTIAL_FLUSH | SI_CONTEXT_INV_VCACHE |
1394                                    SI_CONTEXT_INV_L2 | SI_CONTEXT_WB_L2)))) {
1395       radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
1396       radeon_emit(cs, 0);
1397    }
1398 
1399    /* GFX6-GFX8 only:
1400     *   When one of the CP_COHER_CNTL.DEST_BASE flags is set, SURFACE_SYNC
1401     *   waits for idle, so it should be last. SURFACE_SYNC is done in PFP.
1402     *
1403     * cp_coher_cntl should contain all necessary flags except TC flags
1404     * at this point.
1405     *
1406     * GFX6-GFX7 don't support L2 write-back.
1407     */
1408    if (flags & SI_CONTEXT_INV_L2 || (sctx->chip_class <= GFX7 && (flags & SI_CONTEXT_WB_L2))) {
1409       /* Invalidate L1 & L2. (L1 is always invalidated on GFX6)
1410        * WB must be set on GFX8+ when TC_ACTION is set.
1411        */
1412       si_emit_surface_sync(sctx, sctx->gfx_cs,
1413                            cp_coher_cntl | S_0085F0_TC_ACTION_ENA(1) | S_0085F0_TCL1_ACTION_ENA(1) |
1414                               S_0301F0_TC_WB_ACTION_ENA(sctx->chip_class >= GFX8));
1415       cp_coher_cntl = 0;
1416       sctx->num_L2_invalidates++;
1417    } else {
1418       /* L1 invalidation and L2 writeback must be done separately,
1419        * because both operations can't be done together.
1420        */
1421       if (flags & SI_CONTEXT_WB_L2) {
1422          /* WB = write-back
1423           * NC = apply to non-coherent MTYPEs
1424           *      (i.e. MTYPE <= 1, which is what we use everywhere)
1425           *
1426           * WB doesn't work without NC.
1427           */
1428          si_emit_surface_sync(
1429             sctx, sctx->gfx_cs,
1430             cp_coher_cntl | S_0301F0_TC_WB_ACTION_ENA(1) | S_0301F0_TC_NC_ACTION_ENA(1));
1431          cp_coher_cntl = 0;
1432          sctx->num_L2_writebacks++;
1433       }
1434       if (flags & SI_CONTEXT_INV_VCACHE) {
1435          /* Invalidate per-CU VMEM L1. */
1436          si_emit_surface_sync(sctx, sctx->gfx_cs, cp_coher_cntl | S_0085F0_TCL1_ACTION_ENA(1));
1437          cp_coher_cntl = 0;
1438       }
1439    }
1440 
1441    /* If TC flushes haven't cleared this... */
1442    if (cp_coher_cntl)
1443       si_emit_surface_sync(sctx, sctx->gfx_cs, cp_coher_cntl);
1444 
1445    if (is_barrier)
1446       si_prim_discard_signal_next_compute_ib_start(sctx);
1447 
1448    if (flags & SI_CONTEXT_START_PIPELINE_STATS) {
1449       radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1450       radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_START) | EVENT_INDEX(0));
1451    } else if (flags & SI_CONTEXT_STOP_PIPELINE_STATS) {
1452       radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1453       radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP) | EVENT_INDEX(0));
1454    }
1455 
1456    sctx->flags = 0;
1457 }
1458 
1459 ALWAYS_INLINE
si_upload_vertex_buffer_descriptors(struct si_context * sctx)1460 static bool si_upload_vertex_buffer_descriptors(struct si_context *sctx)
1461 {
1462    unsigned i, count = sctx->num_vertex_elements;
1463    uint32_t *ptr;
1464 
1465    struct si_vertex_elements *velems = sctx->vertex_elements;
1466    unsigned alloc_size = velems->vb_desc_list_alloc_size;
1467 
1468    if (alloc_size) {
1469       /* Vertex buffer descriptors are the only ones which are uploaded
1470        * directly through a staging buffer and don't go through
1471        * the fine-grained upload path.
1472        */
1473       u_upload_alloc(sctx->b.const_uploader, 0, alloc_size,
1474                      si_optimal_tcc_alignment(sctx, alloc_size), &sctx->vb_descriptors_offset,
1475                      (struct pipe_resource **)&sctx->vb_descriptors_buffer, (void **)&ptr);
1476       if (!sctx->vb_descriptors_buffer) {
1477          sctx->vb_descriptors_offset = 0;
1478          sctx->vb_descriptors_gpu_list = NULL;
1479          return false;
1480       }
1481 
1482       sctx->vb_descriptors_gpu_list = ptr;
1483       radeon_add_to_buffer_list(sctx, sctx->gfx_cs, sctx->vb_descriptors_buffer, RADEON_USAGE_READ,
1484                                 RADEON_PRIO_DESCRIPTORS);
1485       sctx->vertex_buffer_pointer_dirty = true;
1486       sctx->prefetch_L2_mask |= SI_PREFETCH_VBO_DESCRIPTORS;
1487    } else {
1488       si_resource_reference(&sctx->vb_descriptors_buffer, NULL);
1489       sctx->vertex_buffer_pointer_dirty = false;
1490       sctx->prefetch_L2_mask &= ~SI_PREFETCH_VBO_DESCRIPTORS;
1491    }
1492 
1493    assert(count <= SI_MAX_ATTRIBS);
1494 
1495    unsigned first_vb_use_mask = velems->first_vb_use_mask;
1496    unsigned num_vbos_in_user_sgprs = sctx->screen->num_vbos_in_user_sgprs;
1497 
1498    for (i = 0; i < count; i++) {
1499       struct pipe_vertex_buffer *vb;
1500       struct si_resource *buf;
1501       unsigned vbo_index = velems->vertex_buffer_index[i];
1502       uint32_t *desc = i < num_vbos_in_user_sgprs ? &sctx->vb_descriptor_user_sgprs[i * 4]
1503                                                   : &ptr[(i - num_vbos_in_user_sgprs) * 4];
1504 
1505       vb = &sctx->vertex_buffer[vbo_index];
1506       buf = si_resource(vb->buffer.resource);
1507       if (!buf) {
1508          memset(desc, 0, 16);
1509          continue;
1510       }
1511 
1512       int64_t offset = (int64_t)((int)vb->buffer_offset) + velems->src_offset[i];
1513 
1514       if (offset >= buf->b.b.width0) {
1515          assert(offset < buf->b.b.width0);
1516          memset(desc, 0, 16);
1517          continue;
1518       }
1519 
1520       uint64_t va = buf->gpu_address + offset;
1521 
1522       int64_t num_records = (int64_t)buf->b.b.width0 - offset;
1523       if (sctx->chip_class != GFX8 && vb->stride) {
1524          /* Round up by rounding down and adding 1 */
1525          num_records = (num_records - velems->format_size[i]) / vb->stride + 1;
1526       }
1527       assert(num_records >= 0 && num_records <= UINT_MAX);
1528 
1529       uint32_t rsrc_word3 = velems->rsrc_word3[i];
1530 
1531       /* OOB_SELECT chooses the out-of-bounds check:
1532        *  - 1: index >= NUM_RECORDS (Structured)
1533        *  - 3: offset >= NUM_RECORDS (Raw)
1534        */
1535       if (sctx->chip_class >= GFX10)
1536          rsrc_word3 |= S_008F0C_OOB_SELECT(vb->stride ? V_008F0C_OOB_SELECT_STRUCTURED
1537                                                       : V_008F0C_OOB_SELECT_RAW);
1538 
1539       desc[0] = va;
1540       desc[1] = S_008F04_BASE_ADDRESS_HI(va >> 32) | S_008F04_STRIDE(vb->stride);
1541       desc[2] = num_records;
1542       desc[3] = rsrc_word3;
1543 
1544       if (first_vb_use_mask & (1 << i)) {
1545          radeon_add_to_buffer_list(sctx, sctx->gfx_cs, si_resource(vb->buffer.resource),
1546                                    RADEON_USAGE_READ, RADEON_PRIO_VERTEX_BUFFER);
1547       }
1548    }
1549 
1550    /* Don't flush the const cache. It would have a very negative effect
1551     * on performance (confirmed by testing). New descriptors are always
1552     * uploaded to a fresh new buffer, so I don't think flushing the const
1553     * cache is needed. */
1554    si_mark_atom_dirty(sctx, &sctx->atoms.s.shader_pointers);
1555    sctx->vertex_buffer_user_sgprs_dirty = num_vbos_in_user_sgprs > 0;
1556    sctx->vertex_buffers_dirty = false;
1557    return true;
1558 }
1559 
si_get_draw_start_count(struct si_context * sctx,const struct pipe_draw_info * info,const struct pipe_draw_start_count * draws,unsigned num_draws,unsigned * start,unsigned * count)1560 static void si_get_draw_start_count(struct si_context *sctx, const struct pipe_draw_info *info,
1561                                     const struct pipe_draw_start_count *draws,
1562                                     unsigned num_draws, unsigned *start, unsigned *count)
1563 {
1564    struct pipe_draw_indirect_info *indirect = info->indirect;
1565 
1566    if (indirect) {
1567       unsigned indirect_count;
1568       struct pipe_transfer *transfer;
1569       unsigned begin, end;
1570       unsigned map_size;
1571       unsigned *data;
1572 
1573       if (indirect->indirect_draw_count) {
1574          data = pipe_buffer_map_range(&sctx->b, indirect->indirect_draw_count,
1575                                       indirect->indirect_draw_count_offset, sizeof(unsigned),
1576                                       PIPE_MAP_READ, &transfer);
1577 
1578          indirect_count = *data;
1579 
1580          pipe_buffer_unmap(&sctx->b, transfer);
1581       } else {
1582          indirect_count = indirect->draw_count;
1583       }
1584 
1585       if (!indirect_count) {
1586          *start = *count = 0;
1587          return;
1588       }
1589 
1590       map_size = (indirect_count - 1) * indirect->stride + 3 * sizeof(unsigned);
1591       data = pipe_buffer_map_range(&sctx->b, indirect->buffer, indirect->offset, map_size,
1592                                    PIPE_MAP_READ, &transfer);
1593 
1594       begin = UINT_MAX;
1595       end = 0;
1596 
1597       for (unsigned i = 0; i < indirect_count; ++i) {
1598          unsigned count = data[0];
1599          unsigned start = data[2];
1600 
1601          if (count > 0) {
1602             begin = MIN2(begin, start);
1603             end = MAX2(end, start + count);
1604          }
1605 
1606          data += indirect->stride / sizeof(unsigned);
1607       }
1608 
1609       pipe_buffer_unmap(&sctx->b, transfer);
1610 
1611       if (begin < end) {
1612          *start = begin;
1613          *count = end - begin;
1614       } else {
1615          *start = *count = 0;
1616       }
1617    } else {
1618       unsigned min_element = UINT_MAX;
1619       unsigned max_element = 0;
1620 
1621       for (unsigned i = 0; i < num_draws; i++) {
1622          min_element = MIN2(min_element, draws[i].start);
1623          max_element = MAX2(max_element, draws[i].start + draws[i].count);
1624       }
1625 
1626       *start = min_element;
1627       *count = max_element - min_element;
1628    }
1629 }
1630 
si_emit_all_states(struct si_context * sctx,const struct pipe_draw_info * info,enum pipe_prim_type prim,unsigned instance_count,unsigned min_vertex_count,bool primitive_restart,unsigned skip_atom_mask)1631 static void si_emit_all_states(struct si_context *sctx, const struct pipe_draw_info *info,
1632                                enum pipe_prim_type prim, unsigned instance_count,
1633                                unsigned min_vertex_count, bool primitive_restart,
1634                                unsigned skip_atom_mask)
1635 {
1636    unsigned num_patches = 0;
1637 
1638    si_emit_rasterizer_prim_state(sctx);
1639    if (sctx->tes_shader.cso)
1640       si_emit_derived_tess_state(sctx, info, &num_patches);
1641 
1642    /* Emit state atoms. */
1643    unsigned mask = sctx->dirty_atoms & ~skip_atom_mask;
1644    while (mask)
1645       sctx->atoms.array[u_bit_scan(&mask)].emit(sctx);
1646 
1647    sctx->dirty_atoms &= skip_atom_mask;
1648 
1649    /* Emit states. */
1650    mask = sctx->dirty_states;
1651    while (mask) {
1652       unsigned i = u_bit_scan(&mask);
1653       struct si_pm4_state *state = sctx->queued.array[i];
1654 
1655       if (!state || sctx->emitted.array[i] == state)
1656          continue;
1657 
1658       si_pm4_emit(sctx, state);
1659       sctx->emitted.array[i] = state;
1660    }
1661    sctx->dirty_states = 0;
1662 
1663    /* Emit draw states. */
1664    si_emit_vs_state(sctx, info);
1665    si_emit_draw_registers(sctx, info, prim, num_patches, instance_count, primitive_restart,
1666                           min_vertex_count);
1667 }
1668 
si_all_vs_resources_read_only(struct si_context * sctx,struct pipe_resource * indexbuf)1669 static bool si_all_vs_resources_read_only(struct si_context *sctx, struct pipe_resource *indexbuf)
1670 {
1671    struct radeon_winsys *ws = sctx->ws;
1672    struct radeon_cmdbuf *cs = sctx->gfx_cs;
1673 
1674    /* Index buffer. */
1675    if (indexbuf && ws->cs_is_buffer_referenced(cs, si_resource(indexbuf)->buf, RADEON_USAGE_WRITE))
1676       goto has_write_reference;
1677 
1678    /* Vertex buffers. */
1679    struct si_vertex_elements *velems = sctx->vertex_elements;
1680    unsigned num_velems = velems->count;
1681 
1682    for (unsigned i = 0; i < num_velems; i++) {
1683       if (!((1 << i) & velems->first_vb_use_mask))
1684          continue;
1685 
1686       unsigned vb_index = velems->vertex_buffer_index[i];
1687       struct pipe_resource *res = sctx->vertex_buffer[vb_index].buffer.resource;
1688       if (!res)
1689          continue;
1690 
1691       if (ws->cs_is_buffer_referenced(cs, si_resource(res)->buf, RADEON_USAGE_WRITE))
1692          goto has_write_reference;
1693    }
1694 
1695    /* Constant and shader buffers. */
1696    struct si_descriptors *buffers =
1697       &sctx->descriptors[si_const_and_shader_buffer_descriptors_idx(PIPE_SHADER_VERTEX)];
1698    for (unsigned i = 0; i < buffers->num_active_slots; i++) {
1699       unsigned index = buffers->first_active_slot + i;
1700       struct pipe_resource *res = sctx->const_and_shader_buffers[PIPE_SHADER_VERTEX].buffers[index];
1701       if (!res)
1702          continue;
1703 
1704       if (ws->cs_is_buffer_referenced(cs, si_resource(res)->buf, RADEON_USAGE_WRITE))
1705          goto has_write_reference;
1706    }
1707 
1708    /* Samplers. */
1709    struct si_shader_selector *vs = sctx->vs_shader.cso;
1710    if (vs->info.base.textures_used) {
1711       unsigned num_samplers = util_last_bit(vs->info.base.textures_used);
1712 
1713       for (unsigned i = 0; i < num_samplers; i++) {
1714          struct pipe_sampler_view *view = sctx->samplers[PIPE_SHADER_VERTEX].views[i];
1715          if (!view)
1716             continue;
1717 
1718          if (ws->cs_is_buffer_referenced(cs, si_resource(view->texture)->buf, RADEON_USAGE_WRITE))
1719             goto has_write_reference;
1720       }
1721    }
1722 
1723    /* Images. */
1724    unsigned num_images = vs->info.base.num_images;
1725    if (num_images) {
1726       for (unsigned i = 0; i < num_images; i++) {
1727          struct pipe_resource *res = sctx->images[PIPE_SHADER_VERTEX].views[i].resource;
1728          if (!res)
1729             continue;
1730 
1731          if (ws->cs_is_buffer_referenced(cs, si_resource(res)->buf, RADEON_USAGE_WRITE))
1732             goto has_write_reference;
1733       }
1734    }
1735 
1736    return true;
1737 
1738 has_write_reference:
1739    /* If the current gfx IB has enough packets, flush it to remove write
1740     * references to buffers.
1741     */
1742    if (cs->prev_dw + cs->current.cdw > 2048) {
1743       si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
1744       assert(si_all_vs_resources_read_only(sctx, indexbuf));
1745       return true;
1746    }
1747    return false;
1748 }
1749 
pd_msg(const char * s)1750 static ALWAYS_INLINE bool pd_msg(const char *s)
1751 {
1752    if (SI_PRIM_DISCARD_DEBUG)
1753       printf("PD failed: %s\n", s);
1754    return false;
1755 }
1756 
si_multi_draw_vbo(struct pipe_context * ctx,const struct pipe_draw_info * info,const struct pipe_draw_start_count * draws,unsigned num_draws)1757 static void si_multi_draw_vbo(struct pipe_context *ctx,
1758                               const struct pipe_draw_info *info,
1759                               const struct pipe_draw_start_count *draws,
1760                               unsigned num_draws)
1761 {
1762    struct si_context *sctx = (struct si_context *)ctx;
1763    struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
1764    struct pipe_resource *indexbuf = info->index.resource;
1765    unsigned dirty_tex_counter, dirty_buf_counter;
1766    enum pipe_prim_type rast_prim, prim = info->mode;
1767    unsigned index_size = info->index_size;
1768    unsigned index_offset = info->indirect ? draws[0].start * index_size : 0;
1769    unsigned instance_count = info->instance_count;
1770    bool primitive_restart =
1771       info->primitive_restart &&
1772       (!sctx->screen->options.prim_restart_tri_strips_only ||
1773        (prim != PIPE_PRIM_TRIANGLE_STRIP && prim != PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY));
1774 
1775    /* GFX6-GFX7 treat instance_count==0 as instance_count==1. There is
1776     * no workaround for indirect draws, but we can at least skip
1777     * direct draws.
1778     */
1779    if (unlikely(!info->indirect && !instance_count))
1780       return;
1781 
1782    struct si_shader_selector *vs = sctx->vs_shader.cso;
1783    if (unlikely(!vs || sctx->num_vertex_elements < vs->num_vs_inputs ||
1784                 (!sctx->ps_shader.cso && !rs->rasterizer_discard) ||
1785                 (!!sctx->tes_shader.cso != (prim == PIPE_PRIM_PATCHES)))) {
1786       assert(0);
1787       return;
1788    }
1789 
1790    /* Recompute and re-emit the texture resource states if needed. */
1791    dirty_tex_counter = p_atomic_read(&sctx->screen->dirty_tex_counter);
1792    if (unlikely(dirty_tex_counter != sctx->last_dirty_tex_counter)) {
1793       sctx->last_dirty_tex_counter = dirty_tex_counter;
1794       sctx->framebuffer.dirty_cbufs |= ((1 << sctx->framebuffer.state.nr_cbufs) - 1);
1795       sctx->framebuffer.dirty_zsbuf = true;
1796       si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
1797       si_update_all_texture_descriptors(sctx);
1798    }
1799 
1800    dirty_buf_counter = p_atomic_read(&sctx->screen->dirty_buf_counter);
1801    if (unlikely(dirty_buf_counter != sctx->last_dirty_buf_counter)) {
1802       sctx->last_dirty_buf_counter = dirty_buf_counter;
1803       /* Rebind all buffers unconditionally. */
1804       si_rebind_buffer(sctx, NULL);
1805    }
1806 
1807    si_decompress_textures(sctx, u_bit_consecutive(0, SI_NUM_GRAPHICS_SHADERS));
1808 
1809    /* Set the rasterization primitive type.
1810     *
1811     * This must be done after si_decompress_textures, which can call
1812     * draw_vbo recursively, and before si_update_shaders, which uses
1813     * current_rast_prim for this draw_vbo call. */
1814    if (sctx->gs_shader.cso) {
1815       /* Only possibilities: POINTS, LINE_STRIP, TRIANGLES */
1816       rast_prim = sctx->gs_shader.cso->rast_prim;
1817    } else if (sctx->tes_shader.cso) {
1818       /* Only possibilities: POINTS, LINE_STRIP, TRIANGLES */
1819       rast_prim = sctx->tes_shader.cso->rast_prim;
1820    } else if (util_rast_prim_is_triangles(prim)) {
1821       rast_prim = PIPE_PRIM_TRIANGLES;
1822    } else {
1823       /* Only possibilities, POINTS, LINE*, RECTANGLES */
1824       rast_prim = prim;
1825    }
1826 
1827    if (rast_prim != sctx->current_rast_prim) {
1828       if (util_prim_is_points_or_lines(sctx->current_rast_prim) !=
1829           util_prim_is_points_or_lines(rast_prim))
1830          si_mark_atom_dirty(sctx, &sctx->atoms.s.guardband);
1831 
1832       sctx->current_rast_prim = rast_prim;
1833       sctx->do_update_shaders = true;
1834    }
1835 
1836    if (sctx->tes_shader.cso && sctx->screen->info.has_ls_vgpr_init_bug) {
1837       /* Determine whether the LS VGPR fix should be applied.
1838        *
1839        * It is only required when num input CPs > num output CPs,
1840        * which cannot happen with the fixed function TCS. We should
1841        * also update this bit when switching from TCS to fixed
1842        * function TCS.
1843        */
1844       struct si_shader_selector *tcs = sctx->tcs_shader.cso;
1845       bool ls_vgpr_fix =
1846          tcs && info->vertices_per_patch > tcs->info.base.tess.tcs_vertices_out;
1847 
1848       if (ls_vgpr_fix != sctx->ls_vgpr_fix) {
1849          sctx->ls_vgpr_fix = ls_vgpr_fix;
1850          sctx->do_update_shaders = true;
1851       }
1852    }
1853 
1854    if (sctx->chip_class <= GFX9 && sctx->gs_shader.cso) {
1855       /* Determine whether the GS triangle strip adjacency fix should
1856        * be applied. Rotate every other triangle if
1857        * - triangle strips with adjacency are fed to the GS and
1858        * - primitive restart is disabled (the rotation doesn't help
1859        *   when the restart occurs after an odd number of triangles).
1860        */
1861       bool gs_tri_strip_adj_fix =
1862          !sctx->tes_shader.cso && prim == PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY && !primitive_restart;
1863 
1864       if (gs_tri_strip_adj_fix != sctx->gs_tri_strip_adj_fix) {
1865          sctx->gs_tri_strip_adj_fix = gs_tri_strip_adj_fix;
1866          sctx->do_update_shaders = true;
1867       }
1868    }
1869 
1870    if (index_size) {
1871       /* Translate or upload, if needed. */
1872       /* 8-bit indices are supported on GFX8. */
1873       if (sctx->chip_class <= GFX7 && index_size == 1) {
1874          unsigned start, count, start_offset, size, offset;
1875          void *ptr;
1876 
1877          si_get_draw_start_count(sctx, info, draws, num_draws, &start, &count);
1878          start_offset = start * 2;
1879          size = count * 2;
1880 
1881          indexbuf = NULL;
1882          u_upload_alloc(ctx->stream_uploader, start_offset, size,
1883                         si_optimal_tcc_alignment(sctx, size), &offset, &indexbuf, &ptr);
1884          if (unlikely(!indexbuf))
1885             return;
1886 
1887          util_shorten_ubyte_elts_to_userptr(&sctx->b, info, 0, 0, index_offset + start, count, ptr);
1888 
1889          /* info->start will be added by the drawing code */
1890          index_offset = offset - start_offset;
1891          index_size = 2;
1892       } else if (info->has_user_indices) {
1893          unsigned start_offset;
1894 
1895          assert(!info->indirect);
1896          assert(num_draws == 1);
1897          start_offset = draws[0].start * index_size;
1898 
1899          indexbuf = NULL;
1900          u_upload_data(ctx->stream_uploader, start_offset, draws[0].count * index_size,
1901                        sctx->screen->info.tcc_cache_line_size,
1902                        (char *)info->index.user + start_offset, &index_offset, &indexbuf);
1903          if (unlikely(!indexbuf))
1904             return;
1905 
1906          /* info->start will be added by the drawing code */
1907          index_offset -= start_offset;
1908       } else if (sctx->chip_class <= GFX7 && si_resource(indexbuf)->TC_L2_dirty) {
1909          /* GFX8 reads index buffers through TC L2, so it doesn't
1910           * need this. */
1911          sctx->flags |= SI_CONTEXT_WB_L2;
1912          si_resource(indexbuf)->TC_L2_dirty = false;
1913       }
1914    }
1915 
1916    bool dispatch_prim_discard_cs = false;
1917    bool prim_discard_cs_instancing = false;
1918    unsigned original_index_size = index_size;
1919    unsigned avg_direct_count = 0;
1920    unsigned min_direct_count = 0;
1921    unsigned total_direct_count = 0;
1922 
1923    if (info->indirect) {
1924       struct pipe_draw_indirect_info *indirect = info->indirect;
1925 
1926       /* Add the buffer size for memory checking in need_cs_space. */
1927       si_context_add_resource_size(sctx, indirect->buffer);
1928 
1929       /* Indirect buffers use TC L2 on GFX9, but not older hw. */
1930       if (sctx->chip_class <= GFX8) {
1931          if (si_resource(indirect->buffer)->TC_L2_dirty) {
1932             sctx->flags |= SI_CONTEXT_WB_L2;
1933             si_resource(indirect->buffer)->TC_L2_dirty = false;
1934          }
1935 
1936          if (indirect->indirect_draw_count &&
1937              si_resource(indirect->indirect_draw_count)->TC_L2_dirty) {
1938             sctx->flags |= SI_CONTEXT_WB_L2;
1939             si_resource(indirect->indirect_draw_count)->TC_L2_dirty = false;
1940          }
1941       }
1942    } else {
1943       min_direct_count = num_draws ? UINT_MAX : 0;
1944       for (unsigned i = 0; i < num_draws; i++) {
1945          unsigned count = draws[i].count;
1946 
1947          total_direct_count += count;
1948          min_direct_count = MIN2(min_direct_count, count);
1949       }
1950       avg_direct_count = (total_direct_count / num_draws) * instance_count;
1951    }
1952 
1953    /* Determine if we can use the primitive discard compute shader. */
1954    if (si_compute_prim_discard_enabled(sctx) &&
1955        (avg_direct_count > sctx->prim_discard_vertex_count_threshold
1956            ? (sctx->compute_num_verts_rejected += total_direct_count, true)
1957            : /* Add, then return true. */
1958            (sctx->compute_num_verts_ineligible += total_direct_count,
1959             false)) && /* Add, then return false. */
1960        (!info->count_from_stream_output || pd_msg("draw_opaque")) &&
1961        (primitive_restart ?
1962                           /* Supported prim types with primitive restart: */
1963            (prim == PIPE_PRIM_TRIANGLE_STRIP || pd_msg("bad prim type with primitive restart")) &&
1964               /* Disallow instancing with primitive restart: */
1965               (instance_count == 1 || pd_msg("instance_count > 1 with primitive restart"))
1966                           :
1967                           /* Supported prim types without primitive restart + allow instancing: */
1968            (1 << prim) & ((1 << PIPE_PRIM_TRIANGLES) | (1 << PIPE_PRIM_TRIANGLE_STRIP) |
1969                           (1 << PIPE_PRIM_TRIANGLE_FAN)) &&
1970               /* Instancing is limited to 16-bit indices, because InstanceID is packed into
1971                  VertexID. */
1972               /* TODO: DrawArraysInstanced doesn't sometimes work, so it's disabled. */
1973               (instance_count == 1 ||
1974                (instance_count <= USHRT_MAX && index_size && index_size <= 2) ||
1975                pd_msg("instance_count too large or index_size == 4 or DrawArraysInstanced"))) &&
1976        (info->drawid == 0 || !sctx->vs_shader.cso->info.uses_drawid || pd_msg("draw_id > 0")) &&
1977        (!sctx->render_cond || pd_msg("render condition")) &&
1978        /* Forced enablement ignores pipeline statistics queries. */
1979        (sctx->screen->debug_flags & (DBG(PD) | DBG(ALWAYS_PD)) ||
1980         (!sctx->num_pipeline_stat_queries && !sctx->streamout.prims_gen_query_enabled) ||
1981         pd_msg("pipestat or primgen query")) &&
1982        (!sctx->vertex_elements->instance_divisor_is_fetched || pd_msg("loads instance divisors")) &&
1983        (!sctx->tes_shader.cso || pd_msg("uses tess")) &&
1984        (!sctx->gs_shader.cso || pd_msg("uses GS")) &&
1985        (!sctx->ps_shader.cso->info.uses_primid || pd_msg("PS uses PrimID")) &&
1986        !rs->polygon_mode_enabled &&
1987 #if SI_PRIM_DISCARD_DEBUG /* same as cso->prim_discard_cs_allowed */
1988        (!sctx->vs_shader.cso->info.uses_bindless_images || pd_msg("uses bindless images")) &&
1989        (!sctx->vs_shader.cso->info.uses_bindless_samplers || pd_msg("uses bindless samplers")) &&
1990        (!sctx->vs_shader.cso->info.writes_memory || pd_msg("writes memory")) &&
1991        (!sctx->vs_shader.cso->info.writes_viewport_index || pd_msg("writes viewport index")) &&
1992        !sctx->vs_shader.cso->info.base.vs.window_space_position &&
1993        !sctx->vs_shader.cso->so.num_outputs &&
1994 #else
1995        (sctx->vs_shader.cso->prim_discard_cs_allowed ||
1996         pd_msg("VS shader uses unsupported features")) &&
1997 #endif
1998        /* Check that all buffers are used for read only, because compute
1999         * dispatches can run ahead. */
2000        (si_all_vs_resources_read_only(sctx, index_size ? indexbuf : NULL) ||
2001         pd_msg("write reference"))) {
2002       switch (si_prepare_prim_discard_or_split_draw(sctx, info, draws, num_draws,
2003                                                     primitive_restart, total_direct_count)) {
2004       case SI_PRIM_DISCARD_ENABLED:
2005          original_index_size = index_size;
2006          prim_discard_cs_instancing = instance_count > 1;
2007          dispatch_prim_discard_cs = true;
2008 
2009          /* The compute shader changes/lowers the following: */
2010          prim = PIPE_PRIM_TRIANGLES;
2011          index_size = 4;
2012          instance_count = 1;
2013          primitive_restart = false;
2014          sctx->compute_num_verts_rejected -= total_direct_count;
2015          sctx->compute_num_verts_accepted += total_direct_count;
2016          break;
2017       case SI_PRIM_DISCARD_DISABLED:
2018          break;
2019       case SI_PRIM_DISCARD_DRAW_SPLIT:
2020          sctx->compute_num_verts_rejected -= total_direct_count;
2021          goto return_cleanup;
2022       case SI_PRIM_DISCARD_MULTI_DRAW_SPLIT:
2023          goto return_cleanup;
2024       }
2025    }
2026 
2027    if (prim_discard_cs_instancing != sctx->prim_discard_cs_instancing) {
2028       sctx->prim_discard_cs_instancing = prim_discard_cs_instancing;
2029       sctx->do_update_shaders = true;
2030    }
2031 
2032    /* Update NGG culling settings. */
2033    uint8_t old_ngg_culling = sctx->ngg_culling;
2034    struct si_shader_selector *hw_vs;
2035    if (sctx->ngg && !dispatch_prim_discard_cs && rast_prim == PIPE_PRIM_TRIANGLES &&
2036        (hw_vs = si_get_vs(sctx)->cso) &&
2037        (avg_direct_count > hw_vs->ngg_cull_vert_threshold ||
2038         (!index_size &&
2039          avg_direct_count > hw_vs->ngg_cull_nonindexed_fast_launch_vert_threshold &&
2040          prim & ((1 << PIPE_PRIM_TRIANGLES) |
2041                  (1 << PIPE_PRIM_TRIANGLE_STRIP))))) {
2042       uint8_t ngg_culling = 0;
2043 
2044       if (rs->rasterizer_discard) {
2045          ngg_culling |= SI_NGG_CULL_FRONT_FACE | SI_NGG_CULL_BACK_FACE;
2046       } else {
2047          /* Polygon mode can't use view and small primitive culling,
2048           * because it draws points or lines where the culling depends
2049           * on the point or line width.
2050           */
2051          if (!rs->polygon_mode_enabled)
2052             ngg_culling |= SI_NGG_CULL_VIEW_SMALLPRIMS;
2053 
2054          if (sctx->viewports.y_inverted ? rs->cull_back : rs->cull_front)
2055             ngg_culling |= SI_NGG_CULL_FRONT_FACE;
2056          if (sctx->viewports.y_inverted ? rs->cull_front : rs->cull_back)
2057             ngg_culling |= SI_NGG_CULL_BACK_FACE;
2058       }
2059 
2060       /* Use NGG fast launch for certain non-indexed primitive types.
2061        * A draw must have at least 1 full primitive.
2062        */
2063       if (ngg_culling && !index_size && min_direct_count >= 3 && !sctx->tes_shader.cso &&
2064           !sctx->gs_shader.cso) {
2065          if (prim == PIPE_PRIM_TRIANGLES)
2066             ngg_culling |= SI_NGG_CULL_GS_FAST_LAUNCH_TRI_LIST;
2067          else if (prim == PIPE_PRIM_TRIANGLE_STRIP)
2068             ngg_culling |= SI_NGG_CULL_GS_FAST_LAUNCH_TRI_STRIP;
2069       }
2070 
2071       if (ngg_culling != old_ngg_culling) {
2072          /* If shader compilation is not ready, this setting will be rejected. */
2073          sctx->ngg_culling = ngg_culling;
2074          sctx->do_update_shaders = true;
2075       }
2076    } else if (old_ngg_culling) {
2077       sctx->ngg_culling = false;
2078       sctx->do_update_shaders = true;
2079    }
2080 
2081    if (sctx->shader_has_inlinable_uniforms_mask &
2082        sctx->inlinable_uniforms_valid_mask &
2083        sctx->inlinable_uniforms_dirty_mask) {
2084       sctx->do_update_shaders = true;
2085       /* If inlinable uniforms are not valid, they are also not dirty, so clear all bits. */
2086       sctx->inlinable_uniforms_dirty_mask = 0;
2087    }
2088 
2089    if (unlikely(sctx->do_update_shaders)) {
2090       if (unlikely(!si_update_shaders(sctx)))
2091          goto return_cleanup;
2092 
2093       /* Insert a VGT_FLUSH when enabling fast launch changes to prevent hangs.
2094        * See issues #2418, #2426, #2434
2095        *
2096        * This is the setting that is used by the draw.
2097        */
2098       uint8_t ngg_culling = si_get_vs(sctx)->current->key.opt.ngg_culling;
2099       if (!(old_ngg_culling & SI_NGG_CULL_GS_FAST_LAUNCH_ALL) &&
2100           ngg_culling & SI_NGG_CULL_GS_FAST_LAUNCH_ALL)
2101          sctx->flags |= SI_CONTEXT_VGT_FLUSH;
2102 
2103       /* Set this to the correct value determined by si_update_shaders. */
2104       sctx->ngg_culling = ngg_culling;
2105    }
2106 
2107    si_need_gfx_cs_space(sctx, num_draws);
2108 
2109    /* If we're using a secure context, determine if cs must be secure or not */
2110    if (unlikely(radeon_uses_secure_bos(sctx->ws))) {
2111       bool secure = si_gfx_resources_check_encrypted(sctx);
2112       if (secure != sctx->ws->cs_is_secure(sctx->gfx_cs)) {
2113          si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW |
2114                                RADEON_FLUSH_TOGGLE_SECURE_SUBMISSION, NULL);
2115       }
2116    }
2117 
2118    /* Since we've called si_context_add_resource_size for vertex buffers,
2119     * this must be called after si_need_cs_space, because we must let
2120     * need_cs_space flush before we add buffers to the buffer list.
2121     */
2122    if (sctx->bo_list_add_all_gfx_resources)
2123       si_gfx_resources_add_all_to_bo_list(sctx);
2124 
2125    if (unlikely(!si_upload_graphics_shader_descriptors(sctx) ||
2126                 (sctx->vertex_buffers_dirty &&
2127                  sctx->num_vertex_elements &&
2128                  !si_upload_vertex_buffer_descriptors(sctx))))
2129       goto return_cleanup;
2130 
2131    /* Vega10/Raven scissor bug workaround. When any context register is
2132     * written (i.e. the GPU rolls the context), PA_SC_VPORT_SCISSOR
2133     * registers must be written too.
2134     */
2135    unsigned masked_atoms = 0;
2136    bool gfx9_scissor_bug = false;
2137 
2138    if (sctx->screen->info.has_gfx9_scissor_bug) {
2139       masked_atoms |= si_get_atom_bit(sctx, &sctx->atoms.s.scissors);
2140       gfx9_scissor_bug = true;
2141 
2142       if (info->count_from_stream_output ||
2143           sctx->dirty_atoms & si_atoms_that_always_roll_context() ||
2144           sctx->dirty_states & si_states_that_always_roll_context())
2145          sctx->context_roll = true;
2146    }
2147 
2148    /* Use optimal packet order based on whether we need to sync the pipeline. */
2149    if (unlikely(sctx->flags & (SI_CONTEXT_FLUSH_AND_INV_CB | SI_CONTEXT_FLUSH_AND_INV_DB |
2150                                SI_CONTEXT_PS_PARTIAL_FLUSH | SI_CONTEXT_CS_PARTIAL_FLUSH))) {
2151       /* If we have to wait for idle, set all states first, so that all
2152        * SET packets are processed in parallel with previous draw calls.
2153        * Then draw and prefetch at the end. This ensures that the time
2154        * the CUs are idle is very short.
2155        */
2156       if (unlikely(sctx->flags & SI_CONTEXT_FLUSH_FOR_RENDER_COND))
2157          masked_atoms |= si_get_atom_bit(sctx, &sctx->atoms.s.render_cond);
2158 
2159       /* Emit all states except possibly render condition. */
2160       si_emit_all_states(sctx, info, prim, instance_count, min_direct_count,
2161                          primitive_restart, masked_atoms);
2162       sctx->emit_cache_flush(sctx);
2163       /* <-- CUs are idle here. */
2164 
2165       if (si_is_atom_dirty(sctx, &sctx->atoms.s.render_cond)) {
2166          sctx->atoms.s.render_cond.emit(sctx);
2167          sctx->dirty_atoms &= ~si_get_atom_bit(sctx, &sctx->atoms.s.render_cond);
2168       }
2169 
2170       if (gfx9_scissor_bug &&
2171           (sctx->context_roll || si_is_atom_dirty(sctx, &sctx->atoms.s.scissors))) {
2172          sctx->atoms.s.scissors.emit(sctx);
2173          sctx->dirty_atoms &= ~si_get_atom_bit(sctx, &sctx->atoms.s.scissors);
2174       }
2175       assert(sctx->dirty_atoms == 0);
2176 
2177       si_emit_draw_packets(sctx, info, draws, num_draws,
2178                            indexbuf, index_size, index_offset, instance_count,
2179                            dispatch_prim_discard_cs, original_index_size);
2180       /* <-- CUs are busy here. */
2181 
2182       /* Start prefetches after the draw has been started. Both will run
2183        * in parallel, but starting the draw first is more important.
2184        */
2185       if (sctx->chip_class >= GFX7 && sctx->prefetch_L2_mask)
2186          cik_emit_prefetch_L2(sctx, false);
2187    } else {
2188       /* If we don't wait for idle, start prefetches first, then set
2189        * states, and draw at the end.
2190        */
2191       if (sctx->flags)
2192          sctx->emit_cache_flush(sctx);
2193 
2194       /* Only prefetch the API VS and VBO descriptors. */
2195       if (sctx->chip_class >= GFX7 && sctx->prefetch_L2_mask)
2196          cik_emit_prefetch_L2(sctx, true);
2197 
2198       si_emit_all_states(sctx, info, prim, instance_count, min_direct_count,
2199                          primitive_restart, masked_atoms);
2200 
2201       if (gfx9_scissor_bug &&
2202           (sctx->context_roll || si_is_atom_dirty(sctx, &sctx->atoms.s.scissors))) {
2203          sctx->atoms.s.scissors.emit(sctx);
2204          sctx->dirty_atoms &= ~si_get_atom_bit(sctx, &sctx->atoms.s.scissors);
2205       }
2206       assert(sctx->dirty_atoms == 0);
2207 
2208       si_emit_draw_packets(sctx, info, draws, num_draws,
2209                            indexbuf, index_size, index_offset, instance_count,
2210                            dispatch_prim_discard_cs, original_index_size);
2211 
2212       /* Prefetch the remaining shaders after the draw has been
2213        * started. */
2214       if (sctx->chip_class >= GFX7 && sctx->prefetch_L2_mask)
2215          cik_emit_prefetch_L2(sctx, false);
2216    }
2217 
2218    /* Clear the context roll flag after the draw call. */
2219    sctx->context_roll = false;
2220 
2221    if (unlikely(sctx->current_saved_cs)) {
2222       si_trace_emit(sctx);
2223       si_log_draw_state(sctx, sctx->log);
2224    }
2225 
2226    /* Workaround for a VGT hang when streamout is enabled.
2227     * It must be done after drawing. */
2228    if ((sctx->family == CHIP_HAWAII || sctx->family == CHIP_TONGA || sctx->family == CHIP_FIJI) &&
2229        si_get_strmout_en(sctx)) {
2230       sctx->flags |= SI_CONTEXT_VGT_STREAMOUT_SYNC;
2231    }
2232 
2233    if (unlikely(sctx->decompression_enabled)) {
2234       sctx->num_decompress_calls++;
2235    } else {
2236       sctx->num_draw_calls++;
2237       if (sctx->framebuffer.state.nr_cbufs > 1)
2238          sctx->num_mrt_draw_calls++;
2239       if (primitive_restart)
2240          sctx->num_prim_restart_calls++;
2241       if (G_0286E8_WAVESIZE(sctx->spi_tmpring_size))
2242          sctx->num_spill_draw_calls++;
2243    }
2244 
2245 return_cleanup:
2246    if (index_size && indexbuf != info->index.resource)
2247       pipe_resource_reference(&indexbuf, NULL);
2248 }
2249 
si_draw_vbo(struct pipe_context * ctx,const struct pipe_draw_info * info)2250 static void si_draw_vbo(struct pipe_context *ctx,
2251                         const struct pipe_draw_info *info)
2252 {
2253    struct pipe_draw_start_count draw = {info->start, info->count};
2254 
2255    si_multi_draw_vbo(ctx, info, &draw, 1);
2256 }
2257 
si_draw_rectangle(struct blitter_context * blitter,void * vertex_elements_cso,blitter_get_vs_func get_vs,int x1,int y1,int x2,int y2,float depth,unsigned num_instances,enum blitter_attrib_type type,const union blitter_attrib * attrib)2258 static void si_draw_rectangle(struct blitter_context *blitter, void *vertex_elements_cso,
2259                               blitter_get_vs_func get_vs, int x1, int y1, int x2, int y2,
2260                               float depth, unsigned num_instances, enum blitter_attrib_type type,
2261                               const union blitter_attrib *attrib)
2262 {
2263    struct pipe_context *pipe = util_blitter_get_pipe(blitter);
2264    struct si_context *sctx = (struct si_context *)pipe;
2265 
2266    /* Pack position coordinates as signed int16. */
2267    sctx->vs_blit_sh_data[0] = (uint32_t)(x1 & 0xffff) | ((uint32_t)(y1 & 0xffff) << 16);
2268    sctx->vs_blit_sh_data[1] = (uint32_t)(x2 & 0xffff) | ((uint32_t)(y2 & 0xffff) << 16);
2269    sctx->vs_blit_sh_data[2] = fui(depth);
2270 
2271    switch (type) {
2272    case UTIL_BLITTER_ATTRIB_COLOR:
2273       memcpy(&sctx->vs_blit_sh_data[3], attrib->color, sizeof(float) * 4);
2274       break;
2275    case UTIL_BLITTER_ATTRIB_TEXCOORD_XY:
2276    case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW:
2277       memcpy(&sctx->vs_blit_sh_data[3], &attrib->texcoord, sizeof(attrib->texcoord));
2278       break;
2279    case UTIL_BLITTER_ATTRIB_NONE:;
2280    }
2281 
2282    pipe->bind_vs_state(pipe, si_get_blitter_vs(sctx, type, num_instances));
2283 
2284    struct pipe_draw_info info = {};
2285    info.mode = SI_PRIM_RECTANGLE_LIST;
2286    info.count = 3;
2287    info.instance_count = num_instances;
2288 
2289    /* Don't set per-stage shader pointers for VS. */
2290    sctx->shader_pointers_dirty &= ~SI_DESCS_SHADER_MASK(VERTEX);
2291    sctx->vertex_buffer_pointer_dirty = false;
2292    sctx->vertex_buffer_user_sgprs_dirty = false;
2293 
2294    si_draw_vbo(pipe, &info);
2295 }
2296 
si_trace_emit(struct si_context * sctx)2297 void si_trace_emit(struct si_context *sctx)
2298 {
2299    struct radeon_cmdbuf *cs = sctx->gfx_cs;
2300    uint32_t trace_id = ++sctx->current_saved_cs->trace_id;
2301 
2302    si_cp_write_data(sctx, sctx->current_saved_cs->trace_buf, 0, 4, V_370_MEM, V_370_ME, &trace_id);
2303 
2304    radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
2305    radeon_emit(cs, AC_ENCODE_TRACE_POINT(trace_id));
2306 
2307    if (sctx->log)
2308       u_log_flush(sctx->log);
2309 }
2310 
si_init_draw_functions(struct si_context * sctx)2311 void si_init_draw_functions(struct si_context *sctx)
2312 {
2313    sctx->b.draw_vbo = si_draw_vbo;
2314    sctx->b.multi_draw = si_multi_draw_vbo;
2315 
2316    sctx->blitter->draw_rectangle = si_draw_rectangle;
2317 
2318    si_init_ia_multi_vgt_param_table(sctx);
2319 }
2320