xref: /aosp_15_r20/external/mesa3d/src/gallium/drivers/radeonsi/si_state_draw.cpp (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright 2012 Advanced Micro Devices, Inc.
3  *
4  * SPDX-License-Identifier: MIT
5  */
6 
7 #include "si_build_pm4.h"
8 
9 #include "util/hash_table.h"
10 #define XXH_INLINE_ALL
11 #include "util/xxhash.h"
12 #include "util/u_cpu_detect.h"
13 #include "util/u_index_modify.h"
14 #include "util/u_prim.h"
15 #include "util/u_upload_mgr.h"
16 #include "ac_rtld.h"
17 #include "si_build_pm4.h"
18 #include "si_tracepoints.h"
19 
20 #if (GFX_VER == 6)
21 #define GFX(name) name##GFX6
22 #elif (GFX_VER == 7)
23 #define GFX(name) name##GFX7
24 #elif (GFX_VER == 8)
25 #define GFX(name) name##GFX8
26 #elif (GFX_VER == 9)
27 #define GFX(name) name##GFX9
28 #elif (GFX_VER == 10)
29 #define GFX(name) name##GFX10
30 #elif (GFX_VER == 103)
31 #define GFX(name) name##GFX10_3
32 #elif (GFX_VER == 11)
33 #define GFX(name) name##GFX11
34 #elif (GFX_VER == 115)
35 #define GFX(name) name##GFX11_5
36 #elif (GFX_VER == 12)
37 #define GFX(name) name##GFX12
38 #else
39 #error "Unknown gfx level"
40 #endif
41 
42 template <amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS, si_has_ngg NGG>
si_update_shaders(struct si_context * sctx)43 static bool si_update_shaders(struct si_context *sctx)
44 {
45    struct pipe_context *ctx = (struct pipe_context *)sctx;
46    struct si_shader *old_vs = si_get_vs_inline(sctx, HAS_TESS, HAS_GS)->current;
47    unsigned old_pa_cl_vs_out_cntl = old_vs ? old_vs->pa_cl_vs_out_cntl : 0;
48    bool old_uses_vs_state_provoking_vertex = old_vs ? old_vs->uses_vs_state_provoking_vertex : false;
49    bool old_uses_gs_state_outprim = old_vs ? old_vs->uses_gs_state_outprim : false;
50    struct si_shader *old_ps = sctx->shader.ps.current;
51    unsigned old_spi_shader_col_format =
52       old_ps ? old_ps->key.ps.part.epilog.spi_shader_col_format : 0;
53    int r;
54 
55    /* Update TCS and TES. */
56    if (HAS_TESS) {
57       if (!sctx->has_tessellation) {
58          si_init_tess_factor_ring(sctx);
59          if (!sctx->has_tessellation)
60             return false;
61       }
62 
63       if (!sctx->is_user_tcs) {
64          if (!si_set_tcs_to_fixed_func_shader(sctx))
65             return false;
66       }
67 
68       r = si_shader_select(ctx, &sctx->shader.tcs);
69       if (r)
70          return false;
71       si_pm4_bind_state(sctx, hs, sctx->shader.tcs.current);
72 
73       if (!HAS_GS || GFX_VERSION <= GFX8) {
74          r = si_shader_select(ctx, &sctx->shader.tes);
75          if (r)
76             return false;
77 
78          if (HAS_GS) {
79             /* TES as ES */
80             assert(GFX_VERSION <= GFX8);
81             si_pm4_bind_state(sctx, es, sctx->shader.tes.current);
82          } else if (NGG) {
83             si_pm4_bind_state(sctx, gs, sctx->shader.tes.current);
84          } else {
85             si_pm4_bind_state(sctx, vs, sctx->shader.tes.current);
86          }
87       }
88    } else {
89       /* Reset TCS to clear fixed function shader. */
90       if (!sctx->is_user_tcs && sctx->shader.tcs.cso) {
91          sctx->shader.tcs.cso = NULL;
92          sctx->shader.tcs.current = NULL;
93       }
94 
95       if (GFX_VERSION <= GFX8) {
96          si_pm4_bind_state(sctx, ls, NULL);
97          sctx->prefetch_L2_mask &= ~SI_PREFETCH_LS;
98       }
99       si_pm4_bind_state(sctx, hs, NULL);
100       sctx->prefetch_L2_mask &= ~SI_PREFETCH_HS;
101    }
102 
103    /* Update GS. */
104    if (HAS_GS) {
105       r = si_shader_select(ctx, &sctx->shader.gs);
106       if (r)
107          return false;
108       si_pm4_bind_state(sctx, gs, sctx->shader.gs.current);
109       if (!NGG) {
110          si_pm4_bind_state(sctx, vs, sctx->shader.gs.current->gs_copy_shader);
111 
112          if (!si_update_gs_ring_buffers(sctx))
113             return false;
114       } else if (GFX_VERSION < GFX11) {
115          si_pm4_bind_state(sctx, vs, NULL);
116          sctx->prefetch_L2_mask &= ~SI_PREFETCH_VS;
117       }
118    } else {
119       if (!NGG) {
120          si_pm4_bind_state(sctx, gs, NULL);
121          sctx->prefetch_L2_mask &= ~SI_PREFETCH_GS;
122          if (GFX_VERSION <= GFX8) {
123             si_pm4_bind_state(sctx, es, NULL);
124             sctx->prefetch_L2_mask &= ~SI_PREFETCH_ES;
125          }
126       }
127    }
128 
129    /* Update VS. */
130    if ((!HAS_TESS && !HAS_GS) || GFX_VERSION <= GFX8) {
131       r = si_shader_select(ctx, &sctx->shader.vs);
132       if (r)
133          return false;
134 
135       if (!HAS_TESS && !HAS_GS) {
136          if (NGG) {
137             si_pm4_bind_state(sctx, gs, sctx->shader.vs.current);
138             if (GFX_VERSION < GFX11) {
139                si_pm4_bind_state(sctx, vs, NULL);
140                sctx->prefetch_L2_mask &= ~SI_PREFETCH_VS;
141             }
142          } else {
143             si_pm4_bind_state(sctx, vs, sctx->shader.vs.current);
144          }
145       } else if (HAS_TESS) {
146          si_pm4_bind_state(sctx, ls, sctx->shader.vs.current);
147       } else {
148          assert(HAS_GS);
149          si_pm4_bind_state(sctx, es, sctx->shader.vs.current);
150       }
151    }
152 
153    if (GFX_VERSION >= GFX9 && HAS_TESS)
154       sctx->vs_uses_base_instance = sctx->queued.named.hs->uses_base_instance;
155    else if (GFX_VERSION >= GFX9 && HAS_GS)
156       sctx->vs_uses_base_instance = sctx->shader.gs.current->uses_base_instance;
157    else
158       sctx->vs_uses_base_instance = sctx->shader.vs.current->uses_base_instance;
159 
160    /* Update VGT_SHADER_STAGES_EN. */
161    uint32_t vgt_stages = 0;
162 
163    if (HAS_TESS) {
164       if (GFX_VERSION >= GFX12) {
165          vgt_stages |= S_028A98_HS_EN(1) |
166                        S_028A98_HS_W32_EN(sctx->queued.named.hs->wave_size == 32);
167       } else {
168          vgt_stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) |
169                        S_028B54_HS_EN(1) |
170                        S_028B54_DYNAMIC_HS(1) |
171                        S_028B54_HS_W32_EN(GFX_VERSION >= GFX10 &&
172                                           sctx->queued.named.hs->wave_size == 32);
173       }
174    }
175 
176    if (NGG) {
177       vgt_stages |= si_get_vs_inline(sctx, HAS_TESS, HAS_GS)->current->ngg.vgt_shader_stages_en;
178    } else {
179       if (HAS_GS) {
180          /* Legacy GS only supports Wave64. */
181          assert(sctx->shader.gs.current->wave_size == 64);
182 
183          vgt_stages |= S_028B54_ES_EN(HAS_TESS ? V_028B54_ES_STAGE_DS : V_028B54_ES_STAGE_REAL) |
184                        S_028B54_GS_EN(1) |
185                        S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER) |
186                        S_028B54_VS_W32_EN(GFX_VERSION >= GFX10 &&
187                                           sctx->shader.gs.current->gs_copy_shader->wave_size == 32);
188       } else if (HAS_TESS) {
189          vgt_stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
190       }
191 
192       vgt_stages |= S_028B54_MAX_PRIMGRP_IN_WAVE(GFX_VERSION >= GFX9 ? 2 : 0) |
193                     S_028B54_VS_W32_EN(!HAS_GS && GFX_VERSION >= GFX10 &&
194                                        si_get_vs_inline(sctx, HAS_TESS, HAS_GS)->current->wave_size == 32);
195    }
196 
197    /* Update GE_CNTL. */
198    uint32_t ge_cntl = 0;
199 
200    if (GFX_VERSION >= GFX10) {
201       union si_vgt_param_key key = sctx->ia_multi_vgt_param_key;
202 
203       if (NGG) {
204          if (HAS_TESS) {
205             if (GFX_VERSION >= GFX11) {
206                ge_cntl = si_get_vs_inline(sctx, HAS_TESS, HAS_GS)->current->ge_cntl |
207                          S_03096C_BREAK_PRIMGRP_AT_EOI(key.u.tess_uses_prim_id);
208             } else {
209                /* PRIM_GRP_SIZE_GFX10 is set by si_emit_vgt_pipeline_state. */
210                ge_cntl = S_03096C_VERT_GRP_SIZE(0) |
211                          S_03096C_BREAK_WAVE_AT_EOI(key.u.tess_uses_prim_id);
212             }
213          } else {
214             ge_cntl = si_get_vs_inline(sctx, HAS_TESS, HAS_GS)->current->ge_cntl;
215          }
216 
217          if (GFX_VERSION >= GFX12)
218             ge_cntl |= S_03096C_DIS_PG_SIZE_ADJUST_FOR_STRIP(1);
219       } else {
220          unsigned primgroup_size;
221          unsigned vertgroup_size;
222          assert(GFX_VERSION < GFX11);
223 
224          if (HAS_TESS) {
225             primgroup_size = 0; /* this is set by si_emit_vgt_pipeline_state */
226             vertgroup_size = 0;
227          } else if (HAS_GS) {
228             unsigned vgt_gs_onchip_cntl = sctx->shader.gs.current->gs.vgt_gs_onchip_cntl;
229             primgroup_size = G_028A44_GS_PRIMS_PER_SUBGRP(vgt_gs_onchip_cntl);
230             vertgroup_size = G_028A44_ES_VERTS_PER_SUBGRP(vgt_gs_onchip_cntl);
231          } else {
232             primgroup_size = 128; /* recommended without a GS and tess */
233             vertgroup_size = 0;
234          }
235 
236          ge_cntl = S_03096C_PRIM_GRP_SIZE_GFX10(primgroup_size) |
237                    S_03096C_VERT_GRP_SIZE(vertgroup_size) |
238                    S_03096C_BREAK_WAVE_AT_EOI(key.u.uses_tess && key.u.tess_uses_prim_id);
239       }
240 
241       /* Note: GE_CNTL.PACKET_TO_ONE_PA should only be set if LINE_STIPPLE_TEX_ENA == 1.
242        * Since we don't use that, we don't have to do anything.
243        */
244    }
245 
246    if (vgt_stages != sctx->vgt_shader_stages_en ||
247        (GFX_VERSION >= GFX10 && ge_cntl != sctx->ge_cntl)) {
248       sctx->vgt_shader_stages_en = vgt_stages;
249       sctx->ge_cntl = ge_cntl;
250       si_mark_atom_dirty(sctx, &sctx->atoms.s.vgt_pipeline_state);
251    }
252 
253    struct si_shader *hw_vs = si_get_vs_inline(sctx, HAS_TESS, HAS_GS)->current;
254 
255    if (old_pa_cl_vs_out_cntl != hw_vs->pa_cl_vs_out_cntl)
256       si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_regs);
257 
258    /* If we start to use any of these, we need to update the SGPR. */
259    if ((hw_vs->uses_vs_state_provoking_vertex && !old_uses_vs_state_provoking_vertex) ||
260        (hw_vs->uses_gs_state_outprim && !old_uses_gs_state_outprim)) {
261       si_update_ngg_sgpr_state_out_prim(sctx, hw_vs, NGG);
262       si_update_ngg_sgpr_state_provoking_vtx(sctx, hw_vs, NGG);
263    }
264 
265    r = si_shader_select(ctx, &sctx->shader.ps);
266    if (r)
267       return false;
268    si_pm4_bind_state(sctx, ps, sctx->shader.ps.current);
269 
270    unsigned db_shader_control = sctx->shader.ps.current->ps.db_shader_control;
271    if (sctx->ps_db_shader_control != db_shader_control) {
272       sctx->ps_db_shader_control = db_shader_control;
273       si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
274       if (sctx->screen->dpbb_allowed)
275          si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
276    }
277 
278    unsigned pa_sc_hisz_control = sctx->shader.ps.current->ps.pa_sc_hisz_control;
279    if (GFX_VERSION >= GFX12 && sctx->screen->dpbb_allowed &&
280        sctx->ps_pa_sc_hisz_control != pa_sc_hisz_control) {
281       sctx->ps_pa_sc_hisz_control = pa_sc_hisz_control;
282       si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
283    }
284 
285    if (si_pm4_state_changed(sctx, ps) ||
286        (!NGG && si_pm4_state_changed(sctx, vs)) ||
287        (NGG && si_pm4_state_changed(sctx, gs))) {
288       sctx->atoms.s.spi_map.emit = sctx->emit_spi_map[sctx->shader.ps.current->ps.num_interp];
289       si_mark_atom_dirty(sctx, &sctx->atoms.s.spi_map);
290    }
291 
292    if ((GFX_VERSION >= GFX10_3 || (GFX_VERSION >= GFX9 && sctx->screen->info.rbplus_allowed)) &&
293        si_pm4_state_changed(sctx, ps) &&
294        (!old_ps || old_spi_shader_col_format !=
295                       sctx->shader.ps.current->key.ps.part.epilog.spi_shader_col_format))
296       si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
297 
298    if (sctx->smoothing_enabled !=
299        sctx->shader.ps.current->key.ps.mono.poly_line_smoothing) {
300       sctx->smoothing_enabled = sctx->shader.ps.current->key.ps.mono.poly_line_smoothing;
301       si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
302 
303       /* NGG cull state uses smoothing_enabled. */
304       if (GFX_VERSION >= GFX10 && sctx->screen->use_ngg_culling)
305          si_mark_atom_dirty(sctx, &sctx->atoms.s.ngg_cull_state);
306 
307       if (GFX_VERSION == GFX11 && sctx->screen->info.has_export_conflict_bug)
308          si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
309 
310       if (sctx->framebuffer.nr_samples <= 1)
311          si_mark_atom_dirty(sctx, &sctx->atoms.s.sample_locations);
312    }
313 
314    if (HAS_TESS)
315       si_update_tess_io_layout_state(sctx);
316 
317    if (GFX_VERSION >= GFX9 && unlikely(sctx->sqtt)) {
318       /* Pretend the bound shaders form a vk pipeline. Include the scratch size in
319        * the hash calculation to force re-emitting the pipeline if the scratch bo
320        * changes.
321        */
322       uint64_t scratch_bo_size = sctx->scratch_buffer ? sctx->scratch_buffer->bo_size : 0;
323       uint32_t total_size = 0;
324 
325       /* Compute pipeline code hash. */
326       XXH64_state_t* xh = XXH64_createState();
327       XXH64_reset(xh, scratch_bo_size);
328 
329       for (int i = 0; i < SI_NUM_GRAPHICS_SHADERS; i++) {
330          struct si_shader *shader = sctx->shaders[i].current;
331          if (sctx->shaders[i].cso && shader) {
332             /* Hash the key. */
333             XXH64_update(xh, &shader->key, sizeof(shader->key));
334             /* Hash the main part binary. */
335             XXH64_update(xh, shader->binary.code_buffer, shader->binary.code_size);
336 
337             total_size += (uint32_t)align_uintptr(shader->binary.uploaded_code_size, 256);
338          }
339       }
340 
341       XXH64_hash_t pipeline_code_hash = XXH64_digest(xh);
342       XXH64_freeState(xh);
343 
344       struct si_sqtt_fake_pipeline *pipeline = NULL;
345       if (!si_sqtt_pipeline_is_registered(sctx->sqtt, pipeline_code_hash)) {
346          /* This is a new pipeline. Allocate a new bo to hold all the shaders. Without
347           * this, shader code export process creates huge rgp files because RGP assumes
348           * the shaders live sequentially in memory (shader N address = shader 0 + offset N)
349           */
350          struct si_resource *bo = si_aligned_buffer_create(
351             &sctx->screen->b,
352             SI_RESOURCE_FLAG_DRIVER_INTERNAL | SI_RESOURCE_FLAG_32BIT,
353             PIPE_USAGE_IMMUTABLE, align(total_size, SI_CPDMA_ALIGNMENT), 256);
354 
355          char *ptr = (char *) (bo ? sctx->screen->ws->buffer_map(sctx->screen->ws,
356                bo->buf, NULL,
357                (enum pipe_map_flags)(PIPE_MAP_READ_WRITE | PIPE_MAP_UNSYNCHRONIZED | RADEON_MAP_TEMPORARY)) :
358              NULL);
359 
360          uint32_t offset = 0;
361          uint64_t scratch_va = sctx->scratch_buffer ? sctx->scratch_buffer->gpu_address : 0;
362 
363          if (ptr) {
364             pipeline = (struct si_sqtt_fake_pipeline *)
365                CALLOC(1, sizeof(struct si_sqtt_fake_pipeline));
366 
367             pipeline->code_hash = pipeline_code_hash;
368             pipeline->bo = bo;
369 
370             /* Re-upload all gfx shaders and init PM4. */
371             si_pm4_clear_state(&pipeline->pm4, sctx->screen, false);
372 
373             uint32_t gfx_sh_offsets[SI_NUM_GRAPHICS_SHADERS] = { 0 };
374             for (int i = 0; i < SI_NUM_GRAPHICS_SHADERS; i++) {
375                struct si_shader *shader = sctx->shaders[i].current;
376                if (sctx->shaders[i].cso && shader) {
377                   struct si_resource *original_bo = shader->bo;
378 
379                   /* Temp override for si_shader_binary_upload_at to work. */
380                   shader->bo = pipeline->bo;
381 
382                   int size = si_shader_binary_upload_at(sctx->screen, shader, scratch_va, offset);
383 
384                   shader->bo = original_bo;
385 
386                   assert(size == (int)shader->binary.uploaded_code_size);
387 
388                   gfx_sh_offsets[i] = offset;
389                   offset += align(size, 256);
390 
391                   struct si_pm4_state *pm4 = &shader->pm4;
392 
393                   /* Patch the SPI_SHADER_PGM_LO_* register to point to the new bo address,
394                    * with the proper offset.
395                    */
396                   uint64_t va_low = shader->gpu_address >> 8;
397                   uint32_t reg = pm4->base.spi_shader_pgm_lo_reg;
398                   ac_pm4_set_reg(&pipeline->pm4.base, reg, va_low);
399                }
400             }
401             ac_pm4_finalize(&pipeline->pm4.base);
402             sctx->screen->ws->buffer_unmap(sctx->screen->ws, bo->buf);
403 
404             _mesa_hash_table_u64_insert(sctx->sqtt->pipeline_bos,
405                                         pipeline_code_hash, pipeline);
406 
407             si_sqtt_register_pipeline(sctx, pipeline, gfx_sh_offsets);
408          } else {
409             if (bo)
410                si_resource_reference(&bo, NULL);
411          }
412       } else {
413          pipeline = (struct si_sqtt_fake_pipeline *)_mesa_hash_table_u64_search(
414              sctx->sqtt->pipeline_bos, pipeline_code_hash);
415       }
416       assert(pipeline);
417 
418       si_sqtt_describe_pipeline_bind(sctx, pipeline_code_hash, 0);
419       si_pm4_bind_state(sctx, sqtt_pipeline, pipeline);
420    }
421 
422    if ((GFX_VERSION <= GFX8 &&
423         (si_pm4_state_enabled_and_changed(sctx, ls) || si_pm4_state_enabled_and_changed(sctx, es))) ||
424        si_pm4_state_enabled_and_changed(sctx, hs) || si_pm4_state_enabled_and_changed(sctx, gs) ||
425        (!NGG && si_pm4_state_enabled_and_changed(sctx, vs)) || si_pm4_state_enabled_and_changed(sctx, ps)) {
426       unsigned scratch_size = 0;
427 
428       if (HAS_TESS) {
429          if (GFX_VERSION <= GFX8) /* LS */
430             scratch_size = MAX2(scratch_size, sctx->shader.vs.current->config.scratch_bytes_per_wave);
431 
432          scratch_size = MAX2(scratch_size, sctx->queued.named.hs->config.scratch_bytes_per_wave);
433 
434          if (HAS_GS) {
435             if (GFX_VERSION <= GFX8) /* ES */
436                scratch_size = MAX2(scratch_size, sctx->shader.tes.current->config.scratch_bytes_per_wave);
437 
438             scratch_size = MAX2(scratch_size, sctx->shader.gs.current->config.scratch_bytes_per_wave);
439          } else {
440             scratch_size = MAX2(scratch_size, sctx->shader.tes.current->config.scratch_bytes_per_wave);
441          }
442       } else if (HAS_GS) {
443          if (GFX_VERSION <= GFX8) /* ES */
444             scratch_size = MAX2(scratch_size, sctx->shader.vs.current->config.scratch_bytes_per_wave);
445 
446          scratch_size = MAX2(scratch_size, sctx->shader.gs.current->config.scratch_bytes_per_wave);
447       } else {
448          scratch_size = MAX2(scratch_size, sctx->shader.vs.current->config.scratch_bytes_per_wave);
449       }
450 
451       scratch_size = MAX2(scratch_size, sctx->shader.ps.current->config.scratch_bytes_per_wave);
452 
453       if (scratch_size && !si_update_spi_tmpring_size(sctx, scratch_size))
454          return false;
455 
456       if (GFX_VERSION >= GFX7) {
457          if (GFX_VERSION <= GFX8 && HAS_TESS && si_pm4_state_enabled_and_changed(sctx, ls))
458             sctx->prefetch_L2_mask |= SI_PREFETCH_LS;
459 
460          if (HAS_TESS && si_pm4_state_enabled_and_changed(sctx, hs))
461             sctx->prefetch_L2_mask |= SI_PREFETCH_HS;
462 
463          if (GFX_VERSION <= GFX8 && HAS_GS && si_pm4_state_enabled_and_changed(sctx, es))
464             sctx->prefetch_L2_mask |= SI_PREFETCH_ES;
465 
466          if ((HAS_GS || NGG) && si_pm4_state_enabled_and_changed(sctx, gs))
467             sctx->prefetch_L2_mask |= SI_PREFETCH_GS;
468 
469          if (!NGG && si_pm4_state_enabled_and_changed(sctx, vs))
470             sctx->prefetch_L2_mask |= SI_PREFETCH_VS;
471 
472          if (si_pm4_state_enabled_and_changed(sctx, ps))
473             sctx->prefetch_L2_mask |= SI_PREFETCH_PS;
474       }
475    }
476 
477    /* si_shader_select_with_key can clear the ngg_culling in the shader key if the shader
478     * compilation hasn't finished. Set it to the same value in si_context.
479     */
480    if (GFX_VERSION >= GFX10 && NGG)
481       sctx->ngg_culling = si_get_vs_inline(sctx, HAS_TESS, HAS_GS)->current->key.ge.opt.ngg_culling;
482 
483    sctx->do_update_shaders = false;
484    return true;
485 }
486 
487 ALWAYS_INLINE
si_conv_pipe_prim(unsigned mode)488 static unsigned si_conv_pipe_prim(unsigned mode)
489 {
490    static const unsigned prim_conv[] = {
491       [MESA_PRIM_POINTS] = V_008958_DI_PT_POINTLIST,
492       [MESA_PRIM_LINES] = V_008958_DI_PT_LINELIST,
493       [MESA_PRIM_LINE_LOOP] = V_008958_DI_PT_LINELOOP,
494       [MESA_PRIM_LINE_STRIP] = V_008958_DI_PT_LINESTRIP,
495       [MESA_PRIM_TRIANGLES] = V_008958_DI_PT_TRILIST,
496       [MESA_PRIM_TRIANGLE_STRIP] = V_008958_DI_PT_TRISTRIP,
497       [MESA_PRIM_TRIANGLE_FAN] = V_008958_DI_PT_TRIFAN,
498       [MESA_PRIM_QUADS] = V_008958_DI_PT_QUADLIST,
499       [MESA_PRIM_QUAD_STRIP] = V_008958_DI_PT_QUADSTRIP,
500       [MESA_PRIM_POLYGON] = V_008958_DI_PT_POLYGON,
501       [MESA_PRIM_LINES_ADJACENCY] = V_008958_DI_PT_LINELIST_ADJ,
502       [MESA_PRIM_LINE_STRIP_ADJACENCY] = V_008958_DI_PT_LINESTRIP_ADJ,
503       [MESA_PRIM_TRIANGLES_ADJACENCY] = V_008958_DI_PT_TRILIST_ADJ,
504       [MESA_PRIM_TRIANGLE_STRIP_ADJACENCY] = V_008958_DI_PT_TRISTRIP_ADJ,
505       [MESA_PRIM_PATCHES] = V_008958_DI_PT_PATCH,
506       [SI_PRIM_RECTANGLE_LIST] = V_008958_DI_PT_RECTLIST};
507    assert(mode < ARRAY_SIZE(prim_conv));
508    return prim_conv[mode];
509 }
510 
511 template<amd_gfx_level GFX_VERSION>
si_cp_dma_prefetch_inline(struct si_context * sctx,uint64_t address,unsigned size)512 static void si_cp_dma_prefetch_inline(struct si_context *sctx, uint64_t address, unsigned size)
513 {
514    assert(GFX_VERSION >= GFX7);
515    assert(sctx->screen->info.has_cp_dma);
516 
517    if (GFX_VERSION >= GFX11)
518       size = MIN2(size, 32768 - SI_CPDMA_ALIGNMENT);
519 
520    /* The prefetch address and size must be aligned, so that we don't have to apply
521     * the complicated hw bug workaround.
522     *
523     * The size should also be less than 2 MB, so that we don't have to use a loop.
524     * Callers shouldn't need to prefetch more than 2 MB.
525     */
526    assert(size % SI_CPDMA_ALIGNMENT == 0);
527    assert(address % SI_CPDMA_ALIGNMENT == 0);
528    assert(size < S_415_BYTE_COUNT_GFX6(~0u));
529 
530    uint32_t header = S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2);
531    uint32_t command = S_415_BYTE_COUNT_GFX6(size);
532 
533    if (GFX_VERSION >= GFX9) {
534       command |= S_415_DISABLE_WR_CONFIRM_GFX9(1);
535       header |= S_411_DST_SEL(V_411_NOWHERE);
536    } else {
537       command |= S_415_DISABLE_WR_CONFIRM_GFX6(1);
538       header |= S_411_DST_SEL(V_411_DST_ADDR_TC_L2);
539    }
540 
541    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
542    radeon_begin(cs);
543    radeon_emit(PKT3(PKT3_DMA_DATA, 5, 0));
544    radeon_emit(header);
545    radeon_emit(address);       /* SRC_ADDR_LO [31:0] */
546    radeon_emit(address >> 32); /* SRC_ADDR_HI [31:0] */
547    radeon_emit(address);       /* DST_ADDR_LO [31:0] */
548    radeon_emit(address >> 32); /* DST_ADDR_HI [31:0] */
549    radeon_emit(command);
550    radeon_end();
551 }
552 
553 #if GFX_VER == 6 /* declare this function only once because it handles all chips. */
554 
si_cp_dma_prefetch(struct si_context * sctx,struct pipe_resource * buf,unsigned offset,unsigned size)555 void si_cp_dma_prefetch(struct si_context *sctx, struct pipe_resource *buf,
556                         unsigned offset, unsigned size)
557 {
558    uint64_t address = si_resource(buf)->gpu_address + offset;
559    switch (sctx->gfx_level) {
560    case GFX7:
561       si_cp_dma_prefetch_inline<GFX7>(sctx, address, size);
562       break;
563    case GFX8:
564       si_cp_dma_prefetch_inline<GFX8>(sctx, address, size);
565       break;
566    case GFX9:
567       si_cp_dma_prefetch_inline<GFX9>(sctx, address, size);
568       break;
569    case GFX10:
570       si_cp_dma_prefetch_inline<GFX10>(sctx, address, size);
571       break;
572    case GFX10_3:
573       si_cp_dma_prefetch_inline<GFX10_3>(sctx, address, size);
574       break;
575    case GFX11:
576       si_cp_dma_prefetch_inline<GFX11>(sctx, address, size);
577       break;
578    case GFX11_5:
579       si_cp_dma_prefetch_inline<GFX11_5>(sctx, address, size);
580       break;
581    case GFX12:
582       si_cp_dma_prefetch_inline<GFX12>(sctx, address, size);
583       break;
584    default:
585       break;
586    }
587 }
588 
589 #endif
590 
591 template<amd_gfx_level GFX_VERSION>
si_prefetch_shader_async(struct si_context * sctx,struct si_shader * shader)592 static void si_prefetch_shader_async(struct si_context *sctx, struct si_shader *shader)
593 {
594    struct pipe_resource *bo = &shader->bo->b.b;
595    si_cp_dma_prefetch_inline<GFX_VERSION>(sctx, shader->gpu_address, bo->width0);
596 }
597 
598 /**
599  * Prefetch shaders.
600  */
601 template<amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS, si_has_ngg NGG>
602 ALWAYS_INLINE
si_prefetch_shaders(struct si_context * sctx)603 static void si_prefetch_shaders(struct si_context *sctx)
604 {
605    unsigned mask = sctx->prefetch_L2_mask;
606 
607    /* GFX6 doesn't support the L2 prefetch. */
608    if (GFX_VERSION < GFX7 || !mask)
609       return;
610 
611    /* Prefetch shaders and VBO descriptors into L2. */
612    if (GFX_VERSION >= GFX11) {
613       if (HAS_TESS && mask & SI_PREFETCH_HS)
614          si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.hs);
615 
616       if (mask & SI_PREFETCH_GS)
617          si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.gs);
618    } else if (GFX_VERSION >= GFX9) {
619       if (HAS_TESS) {
620          if (mask & SI_PREFETCH_HS)
621             si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.hs);
622       }
623       if ((HAS_GS || NGG) && mask & SI_PREFETCH_GS)
624          si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.gs);
625       if (!NGG && mask & SI_PREFETCH_VS)
626             si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.vs);
627    } else {
628       /* GFX6-GFX8 */
629       /* Choose the right spot for the VBO prefetch. */
630       if (HAS_TESS) {
631          if (mask & SI_PREFETCH_LS)
632             si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.ls);
633          if (mask & SI_PREFETCH_HS)
634             si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.hs);
635          if (mask & SI_PREFETCH_ES)
636             si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.es);
637          if (mask & SI_PREFETCH_GS)
638             si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.gs);
639       } else if (HAS_GS) {
640          if (mask & SI_PREFETCH_ES)
641             si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.es);
642          if (mask & SI_PREFETCH_GS)
643             si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.gs);
644       }
645       if (mask & SI_PREFETCH_VS)
646          si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.vs);
647    }
648 
649    if (mask & SI_PREFETCH_PS)
650       si_prefetch_shader_async<GFX_VERSION>(sctx, sctx->queued.named.ps);
651 
652    /* This must be cleared only when AFTER_DRAW is true. */
653    sctx->prefetch_L2_mask = 0;
654 }
655 
si_num_prims_for_vertices(enum mesa_prim prim,unsigned count,unsigned vertices_per_patch)656 static unsigned si_num_prims_for_vertices(enum mesa_prim prim,
657                                           unsigned count, unsigned vertices_per_patch)
658 {
659    switch (prim) {
660    case MESA_PRIM_PATCHES:
661       return count / vertices_per_patch;
662    case MESA_PRIM_POLYGON:
663       /* It's a triangle fan with different edge flags. */
664       return count >= 3 ? count - 2 : 0;
665    case SI_PRIM_RECTANGLE_LIST:
666       return count / 3;
667    default:
668       return u_decomposed_prims_for_vertices(prim, count);
669    }
670 }
671 
si_get_init_multi_vgt_param(struct si_screen * sscreen,union si_vgt_param_key * key)672 static unsigned si_get_init_multi_vgt_param(struct si_screen *sscreen, union si_vgt_param_key *key)
673 {
674    STATIC_ASSERT(sizeof(union si_vgt_param_key) == 2);
675    unsigned max_primgroup_in_wave = 2;
676 
677    /* SWITCH_ON_EOP(0) is always preferable. */
678    bool wd_switch_on_eop = false;
679    bool ia_switch_on_eop = false;
680    bool ia_switch_on_eoi = false;
681    bool partial_vs_wave = false;
682    bool partial_es_wave = false;
683 
684    if (key->u.uses_tess) {
685       /* SWITCH_ON_EOI must be set if PrimID is used. */
686       if (key->u.tess_uses_prim_id)
687          ia_switch_on_eoi = true;
688 
689       /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
690       if ((sscreen->info.family == CHIP_TAHITI || sscreen->info.family == CHIP_PITCAIRN ||
691            sscreen->info.family == CHIP_BONAIRE) &&
692           key->u.uses_gs)
693          partial_vs_wave = true;
694 
695       /* Needed for 028B6C_DISTRIBUTION_MODE != 0. (implies >= GFX8) */
696       if (sscreen->info.has_distributed_tess) {
697          if (key->u.uses_gs) {
698             if (sscreen->info.gfx_level == GFX8)
699                partial_es_wave = true;
700          } else {
701             partial_vs_wave = true;
702          }
703       }
704    }
705 
706    /* This is a hardware requirement. */
707    if (key->u.line_stipple_enabled || (sscreen->debug_flags & DBG(SWITCH_ON_EOP))) {
708       ia_switch_on_eop = true;
709       wd_switch_on_eop = true;
710    }
711 
712    if (sscreen->info.gfx_level >= GFX7) {
713       /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
714        * 4 shader engines. Set 1 to pass the assertion below.
715        * The other cases are hardware requirements.
716        *
717        * Polaris supports primitive restart with WD_SWITCH_ON_EOP=0
718        * for points, line strips, and tri strips.
719        */
720       if (sscreen->info.max_se <= 2 || key->u.prim == MESA_PRIM_POLYGON ||
721           key->u.prim == MESA_PRIM_LINE_LOOP || key->u.prim == MESA_PRIM_TRIANGLE_FAN ||
722           key->u.prim == MESA_PRIM_TRIANGLE_STRIP_ADJACENCY ||
723           (key->u.primitive_restart &&
724            (sscreen->info.family < CHIP_POLARIS10 ||
725             (key->u.prim != MESA_PRIM_POINTS && key->u.prim != MESA_PRIM_LINE_STRIP &&
726              key->u.prim != MESA_PRIM_TRIANGLE_STRIP))) ||
727           key->u.count_from_stream_output)
728          wd_switch_on_eop = true;
729 
730       /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
731        * We don't know that for indirect drawing, so treat it as
732        * always problematic. */
733       if (sscreen->info.family == CHIP_HAWAII && key->u.uses_instancing)
734          wd_switch_on_eop = true;
735 
736       /* Performance recommendation for 4 SE Gfx7-8 parts if
737        * instances are smaller than a primgroup.
738        * Assume indirect draws always use small instances.
739        * This is needed for good VS wave utilization.
740        */
741       if (sscreen->info.gfx_level <= GFX8 && sscreen->info.max_se == 4 &&
742           key->u.multi_instances_smaller_than_primgroup)
743          wd_switch_on_eop = true;
744 
745       /* Required on GFX7 and later. */
746       if (sscreen->info.max_se == 4 && !wd_switch_on_eop)
747          ia_switch_on_eoi = true;
748 
749       /* HW engineers suggested that PARTIAL_VS_WAVE_ON should be set
750        * to work around a GS hang.
751        */
752       if (key->u.uses_gs &&
753           (sscreen->info.family == CHIP_TONGA || sscreen->info.family == CHIP_FIJI ||
754            sscreen->info.family == CHIP_POLARIS10 || sscreen->info.family == CHIP_POLARIS11 ||
755            sscreen->info.family == CHIP_POLARIS12 || sscreen->info.family == CHIP_VEGAM))
756          partial_vs_wave = true;
757 
758       /* Required by Hawaii and, for some special cases, by GFX8. */
759       if (ia_switch_on_eoi &&
760           (sscreen->info.family == CHIP_HAWAII ||
761            (sscreen->info.gfx_level == GFX8 && (key->u.uses_gs || max_primgroup_in_wave != 2))))
762          partial_vs_wave = true;
763 
764       /* Instancing bug on Bonaire. */
765       if (sscreen->info.family == CHIP_BONAIRE && ia_switch_on_eoi && key->u.uses_instancing)
766          partial_vs_wave = true;
767 
768       /* This only applies to Polaris10 and later 4 SE chips.
769        * wd_switch_on_eop is already true on all other chips.
770        */
771       if (!wd_switch_on_eop && key->u.primitive_restart)
772          partial_vs_wave = true;
773 
774       /* If the WD switch is false, the IA switch must be false too. */
775       assert(wd_switch_on_eop || !ia_switch_on_eop);
776    }
777 
778    /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
779    if (sscreen->info.gfx_level <= GFX8 && ia_switch_on_eoi)
780       partial_es_wave = true;
781 
782    return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop) | S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi) |
783           S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave) |
784           S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave) |
785           S_028AA8_WD_SWITCH_ON_EOP(sscreen->info.gfx_level >= GFX7 ? wd_switch_on_eop : 0) |
786           /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
787           S_028AA8_MAX_PRIMGRP_IN_WAVE(sscreen->info.gfx_level == GFX8 ? max_primgroup_in_wave
788                                                                         : 0) |
789           S_030960_EN_INST_OPT_BASIC(sscreen->info.gfx_level >= GFX9) |
790           S_030960_EN_INST_OPT_ADV(sscreen->info.gfx_level >= GFX9);
791 }
792 
si_init_ia_multi_vgt_param_table(struct si_context * sctx)793 static void si_init_ia_multi_vgt_param_table(struct si_context *sctx)
794 {
795    for (int prim = 0; prim <= SI_PRIM_RECTANGLE_LIST; prim++)
796       for (int uses_instancing = 0; uses_instancing < 2; uses_instancing++)
797          for (int multi_instances = 0; multi_instances < 2; multi_instances++)
798             for (int primitive_restart = 0; primitive_restart < 2; primitive_restart++)
799                for (int count_from_so = 0; count_from_so < 2; count_from_so++)
800                   for (int line_stipple = 0; line_stipple < 2; line_stipple++)
801                      for (int uses_tess = 0; uses_tess < 2; uses_tess++)
802                         for (int tess_uses_primid = 0; tess_uses_primid < 2; tess_uses_primid++)
803                            for (int uses_gs = 0; uses_gs < 2; uses_gs++) {
804                               union si_vgt_param_key key;
805 
806                               key.index = 0;
807                               key.u.prim = prim;
808                               key.u.uses_instancing = uses_instancing;
809                               key.u.multi_instances_smaller_than_primgroup = multi_instances;
810                               key.u.primitive_restart = primitive_restart;
811                               key.u.count_from_stream_output = count_from_so;
812                               key.u.line_stipple_enabled = line_stipple;
813                               key.u.uses_tess = uses_tess;
814                               key.u.tess_uses_prim_id = tess_uses_primid;
815                               key.u.uses_gs = uses_gs;
816 
817                               sctx->ia_multi_vgt_param[key.index] =
818                                  si_get_init_multi_vgt_param(sctx->screen, &key);
819                            }
820 }
821 
si_is_line_stipple_enabled(struct si_context * sctx)822 static bool si_is_line_stipple_enabled(struct si_context *sctx)
823 {
824    struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
825 
826    return rs->line_stipple_enable && sctx->current_rast_prim != MESA_PRIM_POINTS &&
827           (rs->polygon_mode_is_lines || util_prim_is_lines(sctx->current_rast_prim));
828 }
829 
830 enum si_is_draw_vertex_state {
831    DRAW_VERTEX_STATE_OFF,
832    DRAW_VERTEX_STATE_ON,
833 };
834 
835 enum si_has_sh_pairs_packed {
836    HAS_SH_PAIRS_PACKED_OFF,
837    HAS_SH_PAIRS_PACKED_ON,
838 };
839 
840 template <si_is_draw_vertex_state IS_DRAW_VERTEX_STATE> ALWAYS_INLINE
num_instanced_prims_less_than(const struct pipe_draw_indirect_info * indirect,enum mesa_prim prim,unsigned min_vertex_count,unsigned instance_count,unsigned num_prims,uint8_t vertices_per_patch)841 static bool num_instanced_prims_less_than(const struct pipe_draw_indirect_info *indirect,
842                                           enum mesa_prim prim,
843                                           unsigned min_vertex_count,
844                                           unsigned instance_count,
845                                           unsigned num_prims,
846                                           uint8_t vertices_per_patch)
847 {
848    if (IS_DRAW_VERTEX_STATE)
849       return 0;
850 
851    if (indirect) {
852       return indirect->buffer ||
853              (instance_count > 1 && indirect->count_from_stream_output);
854    } else {
855       return instance_count > 1 &&
856              si_num_prims_for_vertices(prim, min_vertex_count, vertices_per_patch) < num_prims;
857    }
858 }
859 
860 template <amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS,
861           si_is_draw_vertex_state IS_DRAW_VERTEX_STATE> ALWAYS_INLINE
si_get_ia_multi_vgt_param(struct si_context * sctx,const struct pipe_draw_indirect_info * indirect,enum mesa_prim prim,unsigned instance_count,bool primitive_restart,unsigned min_vertex_count)862 static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
863                                           const struct pipe_draw_indirect_info *indirect,
864                                           enum mesa_prim prim, unsigned instance_count,
865                                           bool primitive_restart, unsigned min_vertex_count)
866 {
867    union si_vgt_param_key key = sctx->ia_multi_vgt_param_key;
868    unsigned primgroup_size;
869    unsigned ia_multi_vgt_param;
870 
871    if (HAS_TESS) {
872       primgroup_size = sctx->num_patches_per_workgroup;
873    } else if (HAS_GS) {
874       primgroup_size = 64; /* recommended with a GS */
875    } else {
876       primgroup_size = 128; /* recommended without a GS and tess */
877    }
878 
879    key.u.prim = prim;
880    key.u.uses_instancing = !IS_DRAW_VERTEX_STATE &&
881                            ((indirect && indirect->buffer) || instance_count > 1);
882    key.u.multi_instances_smaller_than_primgroup =
883       num_instanced_prims_less_than<IS_DRAW_VERTEX_STATE>(indirect, prim, min_vertex_count,
884                                                           instance_count, primgroup_size,
885                                                           sctx->patch_vertices);
886    key.u.primitive_restart = !IS_DRAW_VERTEX_STATE && primitive_restart;
887    key.u.count_from_stream_output = !IS_DRAW_VERTEX_STATE && indirect &&
888                                     indirect->count_from_stream_output;
889    key.u.line_stipple_enabled = si_is_line_stipple_enabled(sctx);
890 
891    ia_multi_vgt_param =
892       sctx->ia_multi_vgt_param[key.index] | S_028AA8_PRIMGROUP_SIZE(primgroup_size - 1);
893 
894    if (HAS_GS) {
895       /* GS requirement. */
896       if (GFX_VERSION <= GFX8 &&
897           SI_GS_PER_ES / primgroup_size >= sctx->screen->gs_table_depth - 3)
898          ia_multi_vgt_param |= S_028AA8_PARTIAL_ES_WAVE_ON(1);
899 
900       /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
901        * The hw doc says all multi-SE chips are affected, but Vulkan
902        * only applies it to Hawaii. Do what Vulkan does.
903        */
904       if (GFX_VERSION == GFX7 &&
905           sctx->family == CHIP_HAWAII && G_028AA8_SWITCH_ON_EOI(ia_multi_vgt_param) &&
906           num_instanced_prims_less_than<IS_DRAW_VERTEX_STATE>(indirect, prim, min_vertex_count,
907                                                               instance_count, 2, sctx->patch_vertices)) {
908          /* The cache flushes should have been emitted already. */
909          assert(sctx->barrier_flags == 0);
910          sctx->barrier_flags = SI_BARRIER_EVENT_VGT_FLUSH;
911          si_emit_barrier_direct(sctx);
912       }
913    }
914 
915    return ia_multi_vgt_param;
916 }
917 
918 /* rast_prim is the primitive type after GS. */
919 template<amd_gfx_level GFX_VERSION, si_has_gs HAS_GS, si_has_ngg NGG> ALWAYS_INLINE
si_emit_rasterizer_prim_state(struct si_context * sctx)920 static void si_emit_rasterizer_prim_state(struct si_context *sctx)
921 {
922    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
923 
924    radeon_begin(cs);
925 
926    if (unlikely(si_is_line_stipple_enabled(sctx))) {
927       /* For lines, reset the stipple pattern at each primitive. Otherwise,
928        * reset the stipple pattern at each packet (line strips, line loops).
929        */
930       enum mesa_prim rast_prim = sctx->current_rast_prim;
931       bool reset_per_prim = rast_prim == MESA_PRIM_LINES ||
932                             rast_prim == MESA_PRIM_LINES_ADJACENCY;
933       /* 0 = no reset, 1 = reset per prim, 2 = reset per packet */
934       if (GFX_VERSION >= GFX12) {
935          radeon_opt_set_context_reg(R_028A44_PA_SC_LINE_STIPPLE_RESET,
936                                     SI_TRACKED_PA_SC_LINE_STIPPLE_RESET,
937                                     S_028A44_AUTO_RESET_CNTL(reset_per_prim ? 1 : 2));
938       } else {
939          struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
940 
941          radeon_opt_set_context_reg(R_028A0C_PA_SC_LINE_STIPPLE,
942                                     SI_TRACKED_PA_SC_LINE_STIPPLE,
943                                     rs->pa_sc_line_stipple |
944                                     S_028A0C_AUTO_RESET_CNTL(reset_per_prim ? 1 : 2));
945       }
946    }
947 
948    if (NGG || HAS_GS) {
949       if (GFX_VERSION >= GFX11) {
950          radeon_opt_set_uconfig_reg(R_030998_VGT_GS_OUT_PRIM_TYPE,
951                                     SI_TRACKED_VGT_GS_OUT_PRIM_TYPE_UCONFIG, sctx->gs_out_prim);
952       } else {
953          radeon_opt_set_context_reg(R_028A6C_VGT_GS_OUT_PRIM_TYPE,
954                                     SI_TRACKED_VGT_GS_OUT_PRIM_TYPE, sctx->gs_out_prim);
955       }
956    }
957 
958    if (GFX_VERSION == GFX9)
959       radeon_end_update_context_roll();
960    else
961       radeon_end();
962 }
963 
964 template <amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS, si_has_ngg NGG,
965           si_is_draw_vertex_state IS_DRAW_VERTEX_STATE, si_has_sh_pairs_packed HAS_SH_PAIRS_PACKED> ALWAYS_INLINE
si_emit_vs_state(struct si_context * sctx,unsigned index_size)966 static void si_emit_vs_state(struct si_context *sctx, unsigned index_size)
967 {
968    if (!IS_DRAW_VERTEX_STATE && sctx->num_vs_blit_sgprs) {
969       /* Re-emit the state after we leave u_blitter. */
970       sctx->last_vs_state = ~0;
971       sctx->last_gs_state = ~0;
972       return;
973    }
974 
975    unsigned vs_state = sctx->current_vs_state; /* all VS bits */
976    unsigned gs_state = sctx->current_gs_state; /* only GS and NGG bits; VS bits will be copied here */
977 
978    if (sctx->shader.vs.cso->info.uses_base_vertex && index_size)
979       vs_state |= ENCODE_FIELD(VS_STATE_INDEXED, 1);
980 
981    /* Copy all state bits from vs_state to gs_state. */
982    gs_state |= vs_state;
983 
984    if (vs_state != sctx->last_vs_state ||
985        ((HAS_GS || NGG) && gs_state != sctx->last_gs_state)) {
986       struct radeon_cmdbuf *cs = &sctx->gfx_cs;
987 
988       /* These are all constant expressions. */
989       unsigned vs_base = si_get_user_data_base(GFX_VERSION, HAS_TESS, HAS_GS, NGG,
990                                                PIPE_SHADER_VERTEX);
991       unsigned tes_base = si_get_user_data_base(GFX_VERSION, HAS_TESS, HAS_GS, NGG,
992                                                 PIPE_SHADER_TESS_EVAL);
993       unsigned gs_base = si_get_user_data_base(GFX_VERSION, HAS_TESS, HAS_GS, NGG,
994                                                PIPE_SHADER_GEOMETRY);
995       unsigned gs_copy_base = R_00B130_SPI_SHADER_USER_DATA_VS_0;
996 
997       radeon_begin(cs);
998       if (HAS_GS) {
999          radeon_set_or_push_gfx_sh_reg(vs_base + SI_SGPR_VS_STATE_BITS * 4, vs_state);
1000 
1001          /* GS always uses the state bits for emulating VGT_ESGS_RING_ITEMSIZE on Gfx9
1002           * (via nir_load_esgs_vertex_stride_amd) and for emulating GS pipeline statistics
1003           * on gfx10.x. NGG GS also has lots of states in there.
1004           */
1005          if (GFX_VERSION >= GFX9)
1006             radeon_set_or_push_gfx_sh_reg(gs_base + SI_SGPR_VS_STATE_BITS * 4, gs_state);
1007 
1008          /* The GS copy shader (for legacy GS) always uses the state bits. */
1009          if (!NGG)
1010             radeon_set_or_push_gfx_sh_reg(gs_copy_base + SI_SGPR_VS_STATE_BITS * 4, gs_state);
1011       } else if (HAS_TESS) {
1012          radeon_set_or_push_gfx_sh_reg(vs_base + SI_SGPR_VS_STATE_BITS * 4, vs_state);
1013          radeon_set_or_push_gfx_sh_reg(tes_base + SI_SGPR_VS_STATE_BITS * 4, NGG ? gs_state : vs_state);
1014       } else {
1015          radeon_set_or_push_gfx_sh_reg(vs_base + SI_SGPR_VS_STATE_BITS * 4, NGG ? gs_state : vs_state);
1016       }
1017       radeon_end();
1018 
1019       sctx->last_vs_state = vs_state;
1020       if (HAS_GS || NGG)
1021          sctx->last_gs_state = gs_state;
1022    }
1023 }
1024 
1025 template <amd_gfx_level GFX_VERSION> ALWAYS_INLINE
si_prim_restart_index_changed(struct si_context * sctx,unsigned index_size,bool primitive_restart,unsigned restart_index)1026 static bool si_prim_restart_index_changed(struct si_context *sctx, unsigned index_size,
1027                                           bool primitive_restart, unsigned restart_index)
1028 {
1029    if (!primitive_restart)
1030       return false;
1031 
1032    if (sctx->last_restart_index == SI_RESTART_INDEX_UNKNOWN)
1033       return true;
1034 
1035    /* GFX8+ only compares the index type number of bits of the restart index, so the unused bits
1036     * are "don't care".
1037     *
1038     * Summary of restart index comparator behavior:
1039     * - GFX6-7: Compare all bits.
1040     * - GFX8: Only compare index type bits.
1041     * - GFX9+: If MATCH_ALL_BITS, compare all bits, else only compare index type bits.
1042     */
1043    if (GFX_VERSION >= GFX8) {
1044       /* This masking eliminates no-op 0xffffffff -> 0xffff restart index changes that cause
1045        * unnecessary context rolls when switching the index type.
1046        */
1047       unsigned index_mask = BITFIELD_MASK(index_size * 8);
1048       return (restart_index & index_mask) != (sctx->last_restart_index & index_mask);
1049    } else {
1050       return restart_index != sctx->last_restart_index;
1051    }
1052 }
1053 
1054 template <amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS,
1055           si_is_draw_vertex_state IS_DRAW_VERTEX_STATE> ALWAYS_INLINE
si_emit_ia_multi_vgt_param(struct si_context * sctx,const struct pipe_draw_indirect_info * indirect,enum mesa_prim prim,unsigned instance_count,bool primitive_restart,unsigned min_vertex_count)1056 static void si_emit_ia_multi_vgt_param(struct si_context *sctx,
1057                                        const struct pipe_draw_indirect_info *indirect,
1058                                        enum mesa_prim prim, unsigned instance_count,
1059                                        bool primitive_restart, unsigned min_vertex_count)
1060 {
1061    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
1062    unsigned ia_multi_vgt_param;
1063 
1064    ia_multi_vgt_param =
1065       si_get_ia_multi_vgt_param<GFX_VERSION, HAS_TESS, HAS_GS, IS_DRAW_VERTEX_STATE>
1066          (sctx, indirect, prim, instance_count, primitive_restart, min_vertex_count);
1067 
1068    radeon_begin(cs);
1069    if (GFX_VERSION == GFX9) {
1070       /* Workaround for SpecviewPerf13 Catia hang on GFX9. */
1071       if (prim != sctx->last_prim)
1072          BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, SI_TRACKED_IA_MULTI_VGT_PARAM_UCONFIG);
1073 
1074       radeon_opt_set_uconfig_reg_idx(R_030960_IA_MULTI_VGT_PARAM,
1075                                      SI_TRACKED_IA_MULTI_VGT_PARAM_UCONFIG,
1076                                      4, ia_multi_vgt_param);
1077    } else if (GFX_VERSION >= GFX7) {
1078       radeon_opt_set_context_reg_idx(R_028AA8_IA_MULTI_VGT_PARAM,
1079                                      SI_TRACKED_IA_MULTI_VGT_PARAM, 1, ia_multi_vgt_param);
1080    } else {
1081       radeon_opt_set_context_reg(R_028AA8_IA_MULTI_VGT_PARAM,
1082                                  SI_TRACKED_IA_MULTI_VGT_PARAM, ia_multi_vgt_param);
1083    }
1084    radeon_end();
1085 }
1086 
1087 template <amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS, si_has_ngg NGG,
1088           si_is_draw_vertex_state IS_DRAW_VERTEX_STATE> ALWAYS_INLINE
si_emit_draw_registers(struct si_context * sctx,const struct pipe_draw_indirect_info * indirect,enum mesa_prim prim,unsigned index_size,unsigned instance_count,bool primitive_restart,unsigned restart_index,unsigned min_vertex_count)1089 static void si_emit_draw_registers(struct si_context *sctx,
1090                                    const struct pipe_draw_indirect_info *indirect,
1091                                    enum mesa_prim prim, unsigned index_size,
1092                                    unsigned instance_count, bool primitive_restart,
1093                                    unsigned restart_index, unsigned min_vertex_count)
1094 {
1095    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
1096 
1097    if (GFX_VERSION <= GFX9) {
1098       si_emit_ia_multi_vgt_param<GFX_VERSION, HAS_TESS, HAS_GS, IS_DRAW_VERTEX_STATE>
1099          (sctx, indirect, prim, instance_count, primitive_restart, min_vertex_count);
1100    }
1101 
1102    radeon_begin(cs);
1103 
1104    if (prim != sctx->last_prim) {
1105       unsigned vgt_prim = HAS_TESS ? V_008958_DI_PT_PATCH : si_conv_pipe_prim(prim);
1106 
1107       if (GFX_VERSION >= GFX12 && HAS_TESS)
1108          vgt_prim |= S_030908_NUM_INPUT_CP(sctx->patch_vertices);
1109 
1110       if (GFX_VERSION >= GFX10)
1111          radeon_set_uconfig_reg(R_030908_VGT_PRIMITIVE_TYPE, vgt_prim);
1112       else if (GFX_VERSION >= GFX7)
1113          radeon_set_uconfig_reg_idx(R_030908_VGT_PRIMITIVE_TYPE, 1, vgt_prim);
1114       else
1115          radeon_set_config_reg(R_008958_VGT_PRIMITIVE_TYPE, vgt_prim);
1116 
1117       sctx->last_prim = prim;
1118    }
1119 
1120    /* Primitive restart. */
1121    if (GFX_VERSION >= GFX11) {
1122       /* GFX11+ can ignore primitive restart for non-indexed draws because it has no effect.
1123        * (it's disabled for non-indexed draws by setting DISABLE_FOR_AUTO_INDEX in the preamble)
1124        */
1125       if (index_size) {
1126          if (primitive_restart != sctx->last_primitive_restart_en) {
1127             radeon_set_uconfig_reg(R_03092C_GE_MULTI_PRIM_IB_RESET_EN,
1128                                    S_03092C_RESET_EN(primitive_restart) |
1129                                    /* This disables primitive restart for non-indexed draws.
1130                                     * By keeping this set, we don't have to unset RESET_EN
1131                                     * for non-indexed draws. */
1132                                    S_03092C_DISABLE_FOR_AUTO_INDEX(1));
1133             sctx->last_primitive_restart_en = primitive_restart;
1134          }
1135          if (si_prim_restart_index_changed<GFX_VERSION>(sctx, index_size, primitive_restart,
1136                                                         restart_index)) {
1137             radeon_set_context_reg(R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX, restart_index);
1138             sctx->last_restart_index = restart_index;
1139          }
1140       }
1141    } else {
1142       if (primitive_restart != sctx->last_primitive_restart_en) {
1143          if (GFX_VERSION >= GFX9)
1144             radeon_set_uconfig_reg(R_03092C_VGT_MULTI_PRIM_IB_RESET_EN, primitive_restart);
1145          else
1146             radeon_set_context_reg(R_028A94_VGT_MULTI_PRIM_IB_RESET_EN, primitive_restart);
1147          sctx->last_primitive_restart_en = primitive_restart;
1148       }
1149       if (si_prim_restart_index_changed<GFX_VERSION>(sctx, index_size, primitive_restart,
1150                                                      restart_index)) {
1151          radeon_set_context_reg(R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX, restart_index);
1152          sctx->last_restart_index = restart_index;
1153          if (GFX_VERSION == GFX9)
1154             sctx->context_roll = true;
1155       }
1156    }
1157    radeon_end();
1158 }
1159 
1160 static ALWAYS_INLINE void
gfx11_emit_buffered_sh_regs_inline(struct si_context * sctx,unsigned * num_regs,struct gfx11_reg_pair * reg_pairs)1161 gfx11_emit_buffered_sh_regs_inline(struct si_context *sctx, unsigned *num_regs,
1162                                    struct gfx11_reg_pair *reg_pairs)
1163 {
1164    unsigned reg_count = *num_regs;
1165 
1166    if (!reg_count)
1167       return;
1168 
1169    *num_regs = 0;
1170 
1171    /* If there is only one register, we can't use the packed SET packet. */
1172    if (reg_count == 1) {
1173       radeon_begin(&sctx->gfx_cs);
1174       radeon_emit(PKT3(PKT3_SET_SH_REG, 1, 0));
1175       radeon_emit(reg_pairs[0].reg_offset[0]);
1176       radeon_emit(reg_pairs[0].reg_value[0]);
1177       radeon_end();
1178       return;
1179    }
1180 
1181    unsigned packet = reg_count <= 14 ? PKT3_SET_SH_REG_PAIRS_PACKED_N :
1182                                        PKT3_SET_SH_REG_PAIRS_PACKED;
1183    unsigned padded_reg_count = align(reg_count, 2);
1184 
1185    radeon_begin(&sctx->gfx_cs);
1186    radeon_emit(PKT3(packet, (padded_reg_count / 2) * 3, 0) | PKT3_RESET_FILTER_CAM_S(1));
1187    radeon_emit(padded_reg_count);
1188    radeon_emit_array(reg_pairs, (reg_count / 2) * 3);
1189 
1190    if (reg_count % 2 == 1) {
1191       unsigned i = reg_count / 2;
1192 
1193       /* Pad the packet by setting the first register again at the end because the register
1194        * count must be even and 2 consecutive offsets must not be equal.
1195        */
1196       radeon_emit(reg_pairs[i].reg_offset[0] | ((uint32_t)reg_pairs[0].reg_offset[0] << 16));
1197       radeon_emit(reg_pairs[i].reg_value[0]);
1198       radeon_emit(reg_pairs[0].reg_value[0]);
1199    }
1200    radeon_end();
1201 }
1202 
1203 #define gfx12_emit_buffered_sh_regs_inline(num_regs, regs) do { \
1204    unsigned __reg_count = *(num_regs); \
1205    if (__reg_count) { \
1206       radeon_emit(PKT3(PKT3_SET_SH_REG_PAIRS, __reg_count * 2 - 1, 0) | PKT3_RESET_FILTER_CAM_S(1)); \
1207       radeon_emit_array(regs, __reg_count * 2); \
1208       *(num_regs) = 0; \
1209    } \
1210 } while (0)
1211 
1212 #if GFX_VER == 6 /* declare this function only once because there is only one variant. */
1213 
si_emit_buffered_compute_sh_regs(struct si_context * sctx)1214 void si_emit_buffered_compute_sh_regs(struct si_context *sctx)
1215 {
1216    if (sctx->gfx_level >= GFX12) {
1217       radeon_begin(&sctx->gfx_cs);
1218       gfx12_emit_buffered_sh_regs_inline(&sctx->num_buffered_compute_sh_regs,
1219                                          sctx->gfx12.buffered_compute_sh_regs);
1220       radeon_end();
1221    } else {
1222       gfx11_emit_buffered_sh_regs_inline(sctx, &sctx->num_buffered_compute_sh_regs,
1223                                          sctx->gfx11.buffered_compute_sh_regs);
1224    }
1225 }
1226 
1227 #endif
1228 
1229 template <amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS, si_has_ngg NGG,
1230           si_is_draw_vertex_state IS_DRAW_VERTEX_STATE, si_has_sh_pairs_packed HAS_SH_PAIRS_PACKED> ALWAYS_INLINE
si_emit_draw_packets(struct si_context * sctx,const struct pipe_draw_info * info,unsigned drawid_base,const struct pipe_draw_indirect_info * indirect,const struct pipe_draw_start_count_bias * draws,unsigned num_draws,struct pipe_resource * indexbuf,unsigned index_size,unsigned index_offset,unsigned instance_count)1231 static void si_emit_draw_packets(struct si_context *sctx, const struct pipe_draw_info *info,
1232                                  unsigned drawid_base,
1233                                  const struct pipe_draw_indirect_info *indirect,
1234                                  const struct pipe_draw_start_count_bias *draws,
1235                                  unsigned num_draws,
1236                                  struct pipe_resource *indexbuf, unsigned index_size,
1237                                  unsigned index_offset, unsigned instance_count)
1238 {
1239    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
1240 
1241    if (unlikely(sctx->sqtt_enabled)) {
1242       si_sqtt_write_event_marker(sctx, &sctx->gfx_cs, sctx->sqtt_next_event,
1243                                  UINT_MAX, UINT_MAX, UINT_MAX);
1244    }
1245 
1246    uint32_t use_opaque = 0;
1247 
1248    if (!IS_DRAW_VERTEX_STATE && indirect && indirect->count_from_stream_output) {
1249       struct si_streamout_target *t = (struct si_streamout_target *)indirect->count_from_stream_output;
1250 
1251       radeon_begin(cs);
1252       radeon_set_context_reg(R_028B30_VGT_STRMOUT_DRAW_OPAQUE_VERTEX_STRIDE, t->stride_in_dw);
1253       radeon_end();
1254 
1255       if (GFX_VERSION >= GFX9) {
1256          /* Use PKT3_LOAD_CONTEXT_REG_INDEX instead of si_cp_copy_data to support state shadowing. */
1257          uint64_t va = t->buf_filled_size->gpu_address + t->buf_filled_size_draw_count_offset;
1258 
1259          radeon_begin(cs);
1260 
1261          // TODO: GFX12: This may be discarded by PFP if the shadow base address is provided by the MQD.
1262          radeon_emit(PKT3(PKT3_LOAD_CONTEXT_REG_INDEX, 3, 0));
1263          radeon_emit(va);
1264          radeon_emit(va >> 32);
1265          radeon_emit((R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE - SI_CONTEXT_REG_OFFSET) >> 2);
1266          radeon_emit(1);
1267 
1268          radeon_end();
1269       } else {
1270          si_cp_copy_data(sctx, &sctx->gfx_cs, COPY_DATA_REG, NULL,
1271                          R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE >> 2, COPY_DATA_SRC_MEM,
1272                          t->buf_filled_size, t->buf_filled_size_draw_count_offset);
1273       }
1274       use_opaque = S_0287F0_USE_OPAQUE(1);
1275       indirect = NULL;
1276    }
1277 
1278    uint32_t index_max_size = 0;
1279    uint64_t index_va = 0;
1280    bool disable_instance_packing = false;
1281 
1282    radeon_begin(cs);
1283 
1284    if (GFX_VERSION == GFX10_3) {
1285       /* Workaround for incorrect stats with adjacent primitive types
1286        * (see PAL's waDisableInstancePacking).
1287        */
1288       if (sctx->num_pipeline_stat_queries &&
1289           sctx->shader.gs.cso == NULL &&
1290           (instance_count > 1 || indirect) &&
1291           (1 << info->mode) & (1 << MESA_PRIM_LINES_ADJACENCY |
1292                                1 << MESA_PRIM_LINE_STRIP_ADJACENCY |
1293                                1 << MESA_PRIM_TRIANGLES_ADJACENCY |
1294                                1 << MESA_PRIM_TRIANGLE_STRIP_ADJACENCY)) {
1295          disable_instance_packing = true;
1296       }
1297    }
1298 
1299    /* draw packet */
1300    if (index_size) {
1301       /* Register shadowing doesn't shadow INDEX_TYPE. */
1302       if (index_size != sctx->last_index_size ||
1303           (GFX_VERSION == GFX10_3 && disable_instance_packing != sctx->disable_instance_packing)) {
1304          unsigned index_type;
1305 
1306          /* Index type computation. When we look at how we need to translate index_size,
1307           * we can see that we just need 2 shifts to get the hw value.
1308           *
1309           * 1 = 001b --> 10b = 2
1310           * 2 = 010b --> 00b = 0
1311           * 4 = 100b --> 01b = 1
1312           */
1313          index_type = (((index_size >> 2) | (index_size << 1)) & 0x3) |
1314                       S_028A7C_DISABLE_INSTANCE_PACKING(disable_instance_packing);
1315 
1316          if (GFX_VERSION <= GFX7 && UTIL_ARCH_BIG_ENDIAN) {
1317             /* GFX7 doesn't support ubyte indices. */
1318             index_type |= index_size == 2 ? V_028A7C_VGT_DMA_SWAP_16_BIT
1319                                           : V_028A7C_VGT_DMA_SWAP_32_BIT;
1320          }
1321 
1322          if (GFX_VERSION >= GFX9) {
1323             radeon_set_uconfig_reg_idx(R_03090C_VGT_INDEX_TYPE, 2, index_type);
1324          } else {
1325             radeon_emit(PKT3(PKT3_INDEX_TYPE, 0, 0));
1326             radeon_emit(index_type);
1327          }
1328 
1329          sctx->last_index_size = index_size;
1330          if (GFX_VERSION == GFX10_3)
1331             sctx->disable_instance_packing = disable_instance_packing;
1332       }
1333 
1334       index_max_size = (indexbuf->width0 - index_offset) >> util_logbase2(index_size);
1335       /* Skip draw calls with 0-sized index buffers.
1336        * They cause a hang on some chips, like Navi10-14.
1337        */
1338       if (!index_max_size) {
1339          radeon_end();
1340          return;
1341       }
1342 
1343       index_va = si_resource(indexbuf)->gpu_address + index_offset;
1344 
1345       radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, si_resource(indexbuf),
1346                                 RADEON_USAGE_READ | RADEON_PRIO_INDEX_BUFFER);
1347    } else {
1348       /* On GFX7 and later, non-indexed draws overwrite VGT_INDEX_TYPE,
1349        * so the state must be re-emitted before the next indexed draw.
1350        */
1351       if (GFX_VERSION >= GFX7)
1352          sctx->last_index_size = -1;
1353       if (GFX_VERSION == GFX10_3 && disable_instance_packing != sctx->disable_instance_packing) {
1354          radeon_set_uconfig_reg_idx(R_03090C_VGT_INDEX_TYPE, 2,
1355                                     S_028A7C_DISABLE_INSTANCE_PACKING(disable_instance_packing));
1356          sctx->disable_instance_packing = disable_instance_packing;
1357       }
1358    }
1359 
1360    unsigned sh_base_reg = si_get_user_data_base(GFX_VERSION, HAS_TESS, HAS_GS, NGG,
1361                                                 PIPE_SHADER_VERTEX);
1362    bool render_cond_bit = sctx->render_cond_enabled;
1363    unsigned tracked_base_vertex_reg;
1364 
1365    if (HAS_TESS)
1366       tracked_base_vertex_reg = SI_TRACKED_SPI_SHADER_USER_DATA_LS__BASE_VERTEX;
1367    else if (HAS_GS || NGG)
1368       tracked_base_vertex_reg = SI_TRACKED_SPI_SHADER_USER_DATA_ES__BASE_VERTEX;
1369    else
1370       tracked_base_vertex_reg = SI_TRACKED_SPI_SHADER_USER_DATA_VS__BASE_VERTEX;
1371 
1372    if (!IS_DRAW_VERTEX_STATE && indirect) {
1373       assert(num_draws == 1);
1374       uint64_t indirect_va = si_resource(indirect->buffer)->gpu_address;
1375 
1376       assert(indirect_va % 8 == 0);
1377 
1378       if (GFX_VERSION >= GFX12) {
1379          gfx12_emit_buffered_sh_regs_inline(&sctx->num_buffered_gfx_sh_regs,
1380                                             sctx->gfx12.buffered_gfx_sh_regs);
1381       } else if (HAS_SH_PAIRS_PACKED) {
1382          radeon_end();
1383          gfx11_emit_buffered_sh_regs_inline(sctx, &sctx->num_buffered_gfx_sh_regs,
1384                                             sctx->gfx11.buffered_gfx_sh_regs);
1385          radeon_begin_again(cs);
1386       }
1387 
1388       /* Invalidate tracked draw constants because DrawIndirect overwrites them. */
1389       BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg); /* BaseVertex */
1390       BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg + 1); /* DrawID */
1391       BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg + 2); /* StartInstance */
1392       sctx->last_instance_count = SI_INSTANCE_COUNT_UNKNOWN;
1393 
1394       radeon_emit(PKT3(PKT3_SET_BASE, 2, 0));
1395       radeon_emit(1);
1396       radeon_emit(indirect_va);
1397       radeon_emit(indirect_va >> 32);
1398 
1399       radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, si_resource(indirect->buffer),
1400                                 RADEON_USAGE_READ | RADEON_PRIO_DRAW_INDIRECT);
1401 
1402       unsigned di_src_sel = index_size ? V_0287F0_DI_SRC_SEL_DMA : V_0287F0_DI_SRC_SEL_AUTO_INDEX;
1403 
1404       assert(indirect->offset % 4 == 0);
1405 
1406       if (index_size) {
1407          radeon_emit(PKT3(PKT3_INDEX_BASE, 1, 0));
1408          radeon_emit(index_va);
1409          radeon_emit(index_va >> 32);
1410 
1411          radeon_emit(PKT3(PKT3_INDEX_BUFFER_SIZE, 0, 0));
1412          radeon_emit(index_max_size);
1413       }
1414 
1415       if (!sctx->screen->has_draw_indirect_multi) {
1416          radeon_emit(PKT3(index_size ? PKT3_DRAW_INDEX_INDIRECT : PKT3_DRAW_INDIRECT, 3,
1417                           render_cond_bit));
1418          radeon_emit(indirect->offset);
1419          radeon_emit((sh_base_reg + SI_SGPR_BASE_VERTEX * 4 - SI_SH_REG_OFFSET) >> 2);
1420          radeon_emit((sh_base_reg + SI_SGPR_START_INSTANCE * 4 - SI_SH_REG_OFFSET) >> 2);
1421          radeon_emit(di_src_sel);
1422       } else {
1423          uint64_t count_va = 0;
1424 
1425          if (indirect->indirect_draw_count) {
1426             struct si_resource *params_buf = si_resource(indirect->indirect_draw_count);
1427 
1428             radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, params_buf,
1429                                       RADEON_USAGE_READ | RADEON_PRIO_DRAW_INDIRECT);
1430 
1431             count_va = params_buf->gpu_address + indirect->indirect_draw_count_offset;
1432          }
1433 
1434          radeon_emit(PKT3(index_size ? PKT3_DRAW_INDEX_INDIRECT_MULTI : PKT3_DRAW_INDIRECT_MULTI, 8,
1435                           render_cond_bit));
1436          radeon_emit(indirect->offset);
1437          radeon_emit((sh_base_reg + SI_SGPR_BASE_VERTEX * 4 - SI_SH_REG_OFFSET) >> 2);
1438          radeon_emit((sh_base_reg + SI_SGPR_START_INSTANCE * 4 - SI_SH_REG_OFFSET) >> 2);
1439          radeon_emit(((sh_base_reg + SI_SGPR_DRAWID * 4 - SI_SH_REG_OFFSET) >> 2) |
1440                      S_2C3_DRAW_INDEX_ENABLE(sctx->shader.vs.cso->info.uses_drawid) |
1441                      S_2C3_COUNT_INDIRECT_ENABLE(!!indirect->indirect_draw_count));
1442          radeon_emit(indirect->draw_count);
1443          radeon_emit(count_va);
1444          radeon_emit(count_va >> 32);
1445          radeon_emit(indirect->stride);
1446          radeon_emit(di_src_sel);
1447       }
1448    } else {
1449       if (sctx->last_instance_count == SI_INSTANCE_COUNT_UNKNOWN ||
1450           sctx->last_instance_count != instance_count) {
1451          radeon_emit(PKT3(PKT3_NUM_INSTANCES, 0, 0));
1452          radeon_emit(instance_count);
1453          sctx->last_instance_count = instance_count;
1454       }
1455 
1456       /* Base vertex and start instance. */
1457       int base_vertex = index_size ? draws[0].index_bias : draws[0].start;
1458 
1459       bool set_draw_id = !IS_DRAW_VERTEX_STATE && sctx->vs_uses_draw_id;
1460       bool set_base_instance = sctx->vs_uses_base_instance;
1461       bool is_blit = !IS_DRAW_VERTEX_STATE && sctx->num_vs_blit_sgprs;
1462 
1463       if (!is_blit) {
1464          /* Prefer SET_SH_REG_PAIRS_PACKED* on Gfx11+. */
1465          if (GFX_VERSION >= GFX12) {
1466             gfx12_opt_push_gfx_sh_reg(sh_base_reg + SI_SGPR_BASE_VERTEX * 4,
1467                                       tracked_base_vertex_reg, base_vertex);
1468             if (set_draw_id) {
1469                gfx12_opt_push_gfx_sh_reg(sh_base_reg + SI_SGPR_DRAWID * 4,
1470                                          tracked_base_vertex_reg + 1, drawid_base);
1471             }
1472             if (set_base_instance) {
1473                gfx12_opt_push_gfx_sh_reg(sh_base_reg + SI_SGPR_START_INSTANCE * 4,
1474                                          tracked_base_vertex_reg + 2, info->start_instance);
1475             }
1476          } else if (HAS_SH_PAIRS_PACKED) {
1477             gfx11_opt_push_gfx_sh_reg(sh_base_reg + SI_SGPR_BASE_VERTEX * 4,
1478                                       tracked_base_vertex_reg, base_vertex);
1479             if (set_draw_id) {
1480                gfx11_opt_push_gfx_sh_reg(sh_base_reg + SI_SGPR_DRAWID * 4,
1481                                          tracked_base_vertex_reg + 1, drawid_base);
1482             }
1483             if (set_base_instance) {
1484                gfx11_opt_push_gfx_sh_reg(sh_base_reg + SI_SGPR_START_INSTANCE * 4,
1485                                          tracked_base_vertex_reg + 2, info->start_instance);
1486             }
1487          } else {
1488             if (set_base_instance) {
1489                radeon_opt_set_sh_reg3(sh_base_reg + SI_SGPR_BASE_VERTEX * 4,
1490                                       tracked_base_vertex_reg, base_vertex, drawid_base,
1491                                       info->start_instance);
1492             } else if (set_draw_id) {
1493                radeon_opt_set_sh_reg2(sh_base_reg + SI_SGPR_BASE_VERTEX * 4,
1494                                       tracked_base_vertex_reg, base_vertex, drawid_base);
1495             } else {
1496                radeon_opt_set_sh_reg(sh_base_reg + SI_SGPR_BASE_VERTEX * 4,
1497                                      tracked_base_vertex_reg, base_vertex);
1498             }
1499          }
1500       }
1501 
1502       if (GFX_VERSION >= GFX12) {
1503          gfx12_emit_buffered_sh_regs_inline(&sctx->num_buffered_gfx_sh_regs,
1504                                             sctx->gfx12.buffered_gfx_sh_regs);
1505       } else if (HAS_SH_PAIRS_PACKED) {
1506          radeon_end();
1507          gfx11_emit_buffered_sh_regs_inline(sctx, &sctx->num_buffered_gfx_sh_regs,
1508                                             sctx->gfx11.buffered_gfx_sh_regs);
1509          radeon_begin_again(cs);
1510       }
1511 
1512       /* Blit SGPRs must be set after gfx1X_emit_buffered_sh_regs_inline because they can
1513        * overwrite them.
1514        */
1515       if (is_blit) {
1516          /* Re-emit draw constants after we leave u_blitter. */
1517          BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg); /* BaseVertex */
1518          BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg + 1); /* DrawID */
1519          BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg + 2); /* StartInstance */
1520 
1521          /* Blit VS doesn't use BASE_VERTEX, START_INSTANCE, and DRAWID. */
1522          radeon_set_sh_reg_seq(sh_base_reg + SI_SGPR_VS_BLIT_DATA * 4, sctx->num_vs_blit_sgprs);
1523          radeon_emit_array(sctx->vs_blit_sh_data, sctx->num_vs_blit_sgprs);
1524       }
1525 
1526       /* Don't update draw_id in the following code if it doesn't increment. */
1527       bool increment_draw_id = !IS_DRAW_VERTEX_STATE && num_draws > 1 &&
1528                                set_draw_id && info->increment_draw_id;
1529 
1530       if (index_size) {
1531          /* NOT_EOP allows merging multiple draws into 1 wave, but only user VGPRs
1532           * can be changed between draws, and GS fast launch must be disabled.
1533           * NOT_EOP doesn't work on gfx9 and older.
1534           *
1535           * Instead of doing this, which evaluates the case conditions repeatedly:
1536           *  for (all draws) {
1537           *    if (case1);
1538           *    else;
1539           *  }
1540           *
1541           * Use this structuring to evaluate the case conditions once:
1542           *  if (case1) for (all draws);
1543           *  else for (all draws);
1544           *
1545           */
1546          bool index_bias_varies = !IS_DRAW_VERTEX_STATE && num_draws > 1 &&
1547                                   info->index_bias_varies;
1548 
1549          if (increment_draw_id) {
1550             if (index_bias_varies) {
1551                for (unsigned i = 0; i < num_draws; i++) {
1552                   uint64_t va = index_va + draws[i].start * index_size;
1553 
1554                   if (i > 0) {
1555                      radeon_set_sh_reg_seq(sh_base_reg + SI_SGPR_BASE_VERTEX * 4, 2);
1556                      radeon_emit(draws[i].index_bias);
1557                      radeon_emit(drawid_base + i);
1558                   }
1559 
1560                   radeon_emit(PKT3(PKT3_DRAW_INDEX_2, 4, render_cond_bit));
1561                   radeon_emit(index_max_size);
1562                   radeon_emit(va);
1563                   radeon_emit(va >> 32);
1564                   radeon_emit(draws[i].count);
1565                   radeon_emit(V_0287F0_DI_SRC_SEL_DMA); /* NOT_EOP disabled */
1566                }
1567                if (num_draws > 1) {
1568                   BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg); /* BaseVertex */
1569                   BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg + 1); /* DrawID */
1570                }
1571             } else {
1572                /* Only DrawID varies. */
1573                for (unsigned i = 0; i < num_draws; i++) {
1574                   uint64_t va = index_va + draws[i].start * index_size;
1575 
1576                   if (i > 0)
1577                      radeon_set_sh_reg(sh_base_reg + SI_SGPR_DRAWID * 4, drawid_base + i);
1578 
1579                   radeon_emit(PKT3(PKT3_DRAW_INDEX_2, 4, render_cond_bit));
1580                   radeon_emit(index_max_size);
1581                   radeon_emit(va);
1582                   radeon_emit(va >> 32);
1583                   radeon_emit(draws[i].count);
1584                   radeon_emit(V_0287F0_DI_SRC_SEL_DMA); /* NOT_EOP disabled */
1585                }
1586                if (num_draws > 1) {
1587                   BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg + 1); /* DrawID */
1588                }
1589             }
1590          } else {
1591             if (index_bias_varies) {
1592                /* Only BaseVertex varies. */
1593                for (unsigned i = 0; i < num_draws; i++) {
1594                   uint64_t va = index_va + draws[i].start * index_size;
1595 
1596                   if (i > 0)
1597                      radeon_set_sh_reg(sh_base_reg + SI_SGPR_BASE_VERTEX * 4, draws[i].index_bias);
1598 
1599                   radeon_emit(PKT3(PKT3_DRAW_INDEX_2, 4, render_cond_bit));
1600                   radeon_emit(index_max_size);
1601                   radeon_emit(va);
1602                   radeon_emit(va >> 32);
1603                   radeon_emit(draws[i].count);
1604                   radeon_emit(V_0287F0_DI_SRC_SEL_DMA); /* NOT_EOP disabled */
1605                }
1606                if (num_draws > 1) {
1607                   BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg); /* BaseVertex */
1608                }
1609             } else {
1610                /* DrawID and BaseVertex are constant. */
1611                if (GFX_VERSION == GFX10) {
1612                   /* GFX10 has a bug that consecutive draw packets with NOT_EOP must not have
1613                    * count == 0 in the last draw (which doesn't set NOT_EOP).
1614                    *
1615                    * So remove all trailing draws with count == 0.
1616                    */
1617                   while (num_draws > 1 && !draws[num_draws - 1].count)
1618                      num_draws--;
1619                }
1620 
1621                for (unsigned i = 0; i < num_draws; i++) {
1622                   uint64_t va = index_va + draws[i].start * index_size;
1623 
1624                   radeon_emit(PKT3(PKT3_DRAW_INDEX_2, 4, render_cond_bit));
1625                   radeon_emit(index_max_size);
1626                   radeon_emit(va);
1627                   radeon_emit(va >> 32);
1628                   radeon_emit(draws[i].count);
1629                   radeon_emit(V_0287F0_DI_SRC_SEL_DMA |
1630                               S_0287F0_NOT_EOP(GFX_VERSION >= GFX10 && GFX_VERSION < GFX12 &&
1631                                                i < num_draws - 1));
1632                }
1633             }
1634          }
1635       } else {
1636          if (GFX_VERSION == GFX12 && !IS_DRAW_VERTEX_STATE &&
1637              indirect && indirect->count_from_stream_output) {
1638             /* DrawTransformFeedback requires 3 SQ_NON_EVENTs after the packet. */
1639             assert(num_draws == 1);
1640 
1641             radeon_emit(PKT3(PKT3_DRAW_INDEX_AUTO, 1, render_cond_bit));
1642             radeon_emit(0);
1643             radeon_emit(V_0287F0_DI_SRC_SEL_AUTO_INDEX | use_opaque);
1644 
1645             for (unsigned i = 0; i < 3; i++)
1646                radeon_event_write(V_028A90_SQ_NON_EVENT);
1647          } else if (increment_draw_id) {
1648             for (unsigned i = 0; i < num_draws; i++) {
1649                if (i > 0) {
1650                   unsigned draw_id = drawid_base + i;
1651 
1652                   radeon_set_sh_reg_seq(sh_base_reg + SI_SGPR_BASE_VERTEX * 4, 2);
1653                   radeon_emit(draws[i].start);
1654                   radeon_emit(draw_id);
1655                }
1656 
1657                radeon_emit(PKT3(PKT3_DRAW_INDEX_AUTO, 1, render_cond_bit));
1658                radeon_emit(draws[i].count);
1659                radeon_emit(V_0287F0_DI_SRC_SEL_AUTO_INDEX | use_opaque);
1660             }
1661             if (num_draws > 1 && (IS_DRAW_VERTEX_STATE || !sctx->num_vs_blit_sgprs)) {
1662                BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg); /* BaseVertex */
1663                BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg + 1); /* DrawID */
1664             }
1665          } else {
1666             for (unsigned i = 0; i < num_draws; i++) {
1667                if (i > 0)
1668                   radeon_set_sh_reg(sh_base_reg + SI_SGPR_BASE_VERTEX * 4, draws[i].start);
1669 
1670                radeon_emit(PKT3(PKT3_DRAW_INDEX_AUTO, 1, render_cond_bit));
1671                radeon_emit(draws[i].count);
1672                radeon_emit(V_0287F0_DI_SRC_SEL_AUTO_INDEX | use_opaque);
1673             }
1674             if (num_draws > 1 && (IS_DRAW_VERTEX_STATE || !sctx->num_vs_blit_sgprs)) {
1675                BITSET_CLEAR(sctx->tracked_regs.reg_saved_mask, tracked_base_vertex_reg); /* BaseVertex */
1676             }
1677          }
1678       }
1679    }
1680 
1681    if (GFX_VERSION >= GFX9 && unlikely(sctx->sqtt_enabled))
1682      radeon_event_write(V_028A90_THREAD_TRACE_MARKER);
1683    radeon_end();
1684 }
1685 
1686 /* Return false if not bound. */
1687 template<amd_gfx_level GFX_VERSION>
si_set_vb_descriptor(struct si_vertex_elements * velems,const struct pipe_vertex_buffer * vb,unsigned index,uint32_t * desc)1688 static void ALWAYS_INLINE si_set_vb_descriptor(struct si_vertex_elements *velems,
1689                                                const struct pipe_vertex_buffer *vb,
1690                                                unsigned index, /* vertex element index */
1691                                                uint32_t *desc) /* where to upload descriptors */
1692 {
1693    struct si_resource *buf = si_resource(vb->buffer.resource);
1694    int64_t offset = (int64_t)((int)vb->buffer_offset) + velems->elem[index].src_offset;
1695 
1696    if (!buf || offset >= buf->b.b.width0) {
1697       memset(desc, 0, 16);
1698       return;
1699    }
1700 
1701    uint64_t va = buf->gpu_address + offset;
1702    unsigned stride = velems->elem[index].stride;
1703 
1704    int64_t num_records = (int64_t)buf->b.b.width0 - offset;
1705    if (GFX_VERSION != GFX8 && stride) {
1706       /* Round up by rounding down and adding 1 */
1707       num_records = (num_records - velems->elem[index].format_size) / stride + 1;
1708    }
1709    assert(num_records >= 0 && num_records <= UINT_MAX);
1710 
1711    desc[0] = va;
1712    desc[1] = S_008F04_BASE_ADDRESS_HI(va >> 32) | S_008F04_STRIDE(stride);
1713    desc[2] = num_records;
1714    desc[3] = velems->elem[index].rsrc_word3;
1715 }
1716 
1717 #if GFX_VER == 6 /* declare this function only once because it supports all chips. */
1718 
si_set_vertex_buffer_descriptor(struct si_screen * sscreen,struct si_vertex_elements * velems,const struct pipe_vertex_buffer * vb,unsigned element_index,uint32_t * out)1719 void si_set_vertex_buffer_descriptor(struct si_screen *sscreen, struct si_vertex_elements *velems,
1720                                      const struct pipe_vertex_buffer *vb, unsigned element_index,
1721                                      uint32_t *out)
1722 {
1723    switch (sscreen->info.gfx_level) {
1724    case GFX6:
1725       si_set_vb_descriptor<GFX6>(velems, vb, element_index, out);
1726       break;
1727    case GFX7:
1728       si_set_vb_descriptor<GFX7>(velems, vb, element_index, out);
1729       break;
1730    case GFX8:
1731       si_set_vb_descriptor<GFX8>(velems, vb, element_index, out);
1732       break;
1733    case GFX9:
1734       si_set_vb_descriptor<GFX9>(velems, vb, element_index, out);
1735       break;
1736    case GFX10:
1737       si_set_vb_descriptor<GFX10>(velems, vb, element_index, out);
1738       break;
1739    case GFX10_3:
1740       si_set_vb_descriptor<GFX10_3>(velems, vb, element_index, out);
1741       break;
1742    case GFX11:
1743       si_set_vb_descriptor<GFX11>(velems, vb, element_index, out);
1744       break;
1745    case GFX11_5:
1746       si_set_vb_descriptor<GFX11_5>(velems, vb, element_index, out);
1747       break;
1748    case GFX12:
1749       si_set_vb_descriptor<GFX12>(velems, vb, element_index, out);
1750       break;
1751    default:
1752       unreachable("unhandled gfx level");
1753    }
1754 }
1755 
1756 #endif
1757 
1758 template<util_popcnt POPCNT>
get_next_vertex_state_elem(struct pipe_vertex_state * state,uint32_t * partial_velem_mask)1759 static ALWAYS_INLINE unsigned get_next_vertex_state_elem(struct pipe_vertex_state *state,
1760                                                          uint32_t *partial_velem_mask)
1761 {
1762    unsigned semantic_index = u_bit_scan(partial_velem_mask);
1763    assert(state->input.full_velem_mask & BITFIELD_BIT(semantic_index));
1764    /* A prefix mask of the full mask gives us the index in pipe_vertex_state. */
1765    return util_bitcount_fast<POPCNT>(state->input.full_velem_mask & BITFIELD_MASK(semantic_index));
1766 }
1767 
1768 template<amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS, si_has_ngg NGG>
get_vb_descriptor_sgpr_ptr_offset(void)1769 static unsigned get_vb_descriptor_sgpr_ptr_offset(void)
1770 {
1771    /* Find the location of the VB descriptor pointer. */
1772    unsigned dw_offset = SI_VS_NUM_USER_SGPR;
1773    if (GFX_VERSION >= GFX9) {
1774       if (HAS_TESS)
1775          dw_offset = GFX9_TCS_NUM_USER_SGPR;
1776       else if (HAS_GS || NGG)
1777          dw_offset = GFX9_GS_NUM_USER_SGPR;
1778    }
1779    return dw_offset * 4;
1780 }
1781 
1782 template <amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS, si_has_ngg NGG,
1783           si_is_draw_vertex_state IS_DRAW_VERTEX_STATE, si_has_sh_pairs_packed HAS_SH_PAIRS_PACKED,
1784           util_popcnt POPCNT> ALWAYS_INLINE
si_upload_and_prefetch_VB_descriptors(struct si_context * sctx,struct pipe_vertex_state * state,uint32_t partial_velem_mask)1785 static bool si_upload_and_prefetch_VB_descriptors(struct si_context *sctx,
1786                                                   struct pipe_vertex_state *state,
1787                                                   uint32_t partial_velem_mask)
1788 {
1789    struct si_vertex_state *vstate = (struct si_vertex_state *)state;
1790    unsigned count = IS_DRAW_VERTEX_STATE ? util_bitcount_fast<POPCNT>(partial_velem_mask) :
1791                                            sctx->num_vertex_elements;
1792    unsigned sh_base = si_get_user_data_base(GFX_VERSION, HAS_TESS, HAS_GS, NGG,
1793                                             PIPE_SHADER_VERTEX);
1794    unsigned num_vbos_in_user_sgprs = si_num_vbos_in_user_sgprs_inline(GFX_VERSION);
1795 
1796    assert(count <= SI_MAX_ATTRIBS);
1797 
1798    if (sctx->vertex_buffers_dirty || IS_DRAW_VERTEX_STATE) {
1799       assert(count || IS_DRAW_VERTEX_STATE);
1800 
1801       struct si_vertex_elements *velems = sctx->vertex_elements;
1802       unsigned alloc_size = IS_DRAW_VERTEX_STATE ?
1803                                vstate->velems.vb_desc_list_alloc_size :
1804                                velems->vb_desc_list_alloc_size;
1805       uint64_t vb_descriptors_address = 0;
1806       uint32_t *ptr;
1807 
1808       if (alloc_size) {
1809          unsigned offset;
1810 
1811          /* Vertex buffer descriptors are the only ones which are uploaded directly
1812           * and don't go through si_upload_graphics_shader_descriptors.
1813           */
1814          u_upload_alloc(sctx->b.const_uploader, 0, alloc_size,
1815                         si_optimal_tcc_alignment(sctx, alloc_size), &offset,
1816                         (struct pipe_resource **)&sctx->last_const_upload_buffer, (void **)&ptr);
1817          if (!sctx->last_const_upload_buffer)
1818             return false;
1819 
1820          radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, sctx->last_const_upload_buffer,
1821                                    RADEON_USAGE_READ | RADEON_PRIO_DESCRIPTORS);
1822          vb_descriptors_address = sctx->last_const_upload_buffer->gpu_address + offset;
1823 
1824          /* GFX6 doesn't support the L2 prefetch. */
1825          if (GFX_VERSION >= GFX7) {
1826             uint64_t address = sctx->last_const_upload_buffer->gpu_address + offset;
1827             si_cp_dma_prefetch_inline<GFX_VERSION>(sctx, address, alloc_size);
1828          }
1829       }
1830 
1831       unsigned count_in_user_sgprs = MIN2(count, num_vbos_in_user_sgprs);
1832       unsigned i = 0;
1833 
1834       if (IS_DRAW_VERTEX_STATE) {
1835          radeon_begin(&sctx->gfx_cs);
1836 
1837          if (count_in_user_sgprs) {
1838             radeon_set_sh_reg_seq(sh_base + SI_SGPR_VS_VB_DESCRIPTOR_FIRST * 4, count_in_user_sgprs * 4);
1839 
1840             /* the first iteration always executes */
1841             do {
1842                unsigned velem_index = get_next_vertex_state_elem<POPCNT>(state, &partial_velem_mask);
1843 
1844                radeon_emit_array(&vstate->descriptors[velem_index * 4], 4);
1845             } while (++i < count_in_user_sgprs);
1846          }
1847 
1848          if (partial_velem_mask) {
1849             assert(alloc_size);
1850 
1851             unsigned vb_desc_offset =
1852                sh_base + get_vb_descriptor_sgpr_ptr_offset<GFX_VERSION, HAS_TESS, HAS_GS, NGG>();
1853 
1854             radeon_set_or_push_gfx_sh_reg(vb_desc_offset, vb_descriptors_address);
1855 
1856             /* the first iteration always executes */
1857             do {
1858                unsigned velem_index = get_next_vertex_state_elem<POPCNT>(state, &partial_velem_mask);
1859                uint32_t *desc = &ptr[(i - num_vbos_in_user_sgprs) * 4];
1860 
1861                memcpy(desc, &vstate->descriptors[velem_index * 4], 16);
1862                i++;
1863             } while (partial_velem_mask);
1864          }
1865          radeon_end();
1866 
1867          if (vstate->b.input.vbuffer.buffer.resource != vstate->b.input.indexbuf) {
1868             radeon_add_to_buffer_list(sctx, &sctx->gfx_cs,
1869                                       si_resource(vstate->b.input.vbuffer.buffer.resource),
1870                                       RADEON_USAGE_READ | RADEON_PRIO_VERTEX_BUFFER);
1871          }
1872 
1873          /* The next draw_vbo should recompute and rebind vertex buffer descriptors. */
1874          sctx->vertex_buffers_dirty = sctx->num_vertex_elements > 0;
1875       } else {
1876          if (count_in_user_sgprs) {
1877             radeon_begin(&sctx->gfx_cs);
1878             radeon_set_sh_reg_seq(sh_base + SI_SGPR_VS_VB_DESCRIPTOR_FIRST * 4,
1879                                   count_in_user_sgprs * 4);
1880 
1881             /* the first iteration always executes */
1882             do {
1883                unsigned vbo_index = velems->vertex_buffer_index[i];
1884                const struct pipe_vertex_buffer *vb = &sctx->vertex_buffer[vbo_index];
1885                uint32_t *desc;
1886 
1887                radeon_emit_array_get_ptr(4, &desc);
1888 
1889                si_set_vb_descriptor<GFX_VERSION>(velems, vb, i, desc);
1890             } while (++i < count_in_user_sgprs);
1891 
1892             radeon_end();
1893          }
1894 
1895          if (alloc_size) {
1896             /* the first iteration always executes */
1897             do {
1898                unsigned vbo_index = velems->vertex_buffer_index[i];
1899                const struct pipe_vertex_buffer *vb = &sctx->vertex_buffer[vbo_index];
1900                uint32_t *desc = &ptr[(i - num_vbos_in_user_sgprs) * 4];
1901 
1902                si_set_vb_descriptor<GFX_VERSION>(velems, vb, i, desc);
1903             } while (++i < count);
1904 
1905             unsigned vb_desc_ptr_offset =
1906                sh_base + get_vb_descriptor_sgpr_ptr_offset<GFX_VERSION, HAS_TESS, HAS_GS, NGG>();
1907             radeon_begin(&sctx->gfx_cs);
1908             radeon_set_or_push_gfx_sh_reg(vb_desc_ptr_offset, vb_descriptors_address);
1909             radeon_end();
1910          }
1911 
1912          sctx->vertex_buffers_dirty = false;
1913       }
1914    }
1915 
1916    return true;
1917 }
1918 
si_get_draw_start_count(struct si_context * sctx,const struct pipe_draw_info * info,const struct pipe_draw_indirect_info * indirect,const struct pipe_draw_start_count_bias * draws,unsigned num_draws,unsigned * start,unsigned * count)1919 static void si_get_draw_start_count(struct si_context *sctx, const struct pipe_draw_info *info,
1920                                     const struct pipe_draw_indirect_info *indirect,
1921                                     const struct pipe_draw_start_count_bias *draws,
1922                                     unsigned num_draws, unsigned *start, unsigned *count)
1923 {
1924    if (indirect && !indirect->count_from_stream_output) {
1925       unsigned indirect_count;
1926       struct pipe_transfer *transfer;
1927       unsigned begin, end;
1928       unsigned map_size;
1929       unsigned *data;
1930 
1931       if (indirect->indirect_draw_count) {
1932          data = (unsigned*)
1933                 pipe_buffer_map_range(&sctx->b, indirect->indirect_draw_count,
1934                                       indirect->indirect_draw_count_offset, sizeof(unsigned),
1935                                       PIPE_MAP_READ, &transfer);
1936 
1937          indirect_count = *data;
1938 
1939          pipe_buffer_unmap(&sctx->b, transfer);
1940       } else {
1941          indirect_count = indirect->draw_count;
1942       }
1943 
1944       if (!indirect_count) {
1945          *start = *count = 0;
1946          return;
1947       }
1948 
1949       map_size = (indirect_count - 1) * indirect->stride + 3 * sizeof(unsigned);
1950       data = (unsigned*)
1951              pipe_buffer_map_range(&sctx->b, indirect->buffer, indirect->offset, map_size,
1952                                    PIPE_MAP_READ, &transfer);
1953 
1954       begin = UINT_MAX;
1955       end = 0;
1956 
1957       for (unsigned i = 0; i < indirect_count; ++i) {
1958          unsigned count = data[0];
1959          unsigned start = data[2];
1960 
1961          if (count > 0) {
1962             begin = MIN2(begin, start);
1963             end = MAX2(end, start + count);
1964          }
1965 
1966          data += indirect->stride / sizeof(unsigned);
1967       }
1968 
1969       pipe_buffer_unmap(&sctx->b, transfer);
1970 
1971       if (begin < end) {
1972          *start = begin;
1973          *count = end - begin;
1974       } else {
1975          *start = *count = 0;
1976       }
1977    } else {
1978       unsigned min_element = UINT_MAX;
1979       unsigned max_element = 0;
1980 
1981       for (unsigned i = 0; i < num_draws; i++) {
1982          min_element = MIN2(min_element, draws[i].start);
1983          max_element = MAX2(max_element, draws[i].start + draws[i].count);
1984       }
1985 
1986       *start = min_element;
1987       *count = max_element - min_element;
1988    }
1989 }
1990 
1991 ALWAYS_INLINE
si_emit_all_states(struct si_context * sctx,uint64_t skip_atom_mask)1992 static void si_emit_all_states(struct si_context *sctx, uint64_t skip_atom_mask)
1993 {
1994    /* Emit states by calling their emit functions. */
1995    uint64_t dirty = sctx->dirty_atoms & ~skip_atom_mask;
1996 
1997    if (dirty) {
1998       sctx->dirty_atoms &= skip_atom_mask;
1999 
2000       /* u_bit_scan64 is too slow on i386. */
2001       if (sizeof(void*) == 8) {
2002          do {
2003             unsigned i = u_bit_scan64(&dirty);
2004             sctx->atoms.array[i].emit(sctx, i);
2005          } while (dirty);
2006       } else {
2007          unsigned dirty_lo = dirty;
2008          unsigned dirty_hi = dirty >> 32;
2009 
2010          while (dirty_lo) {
2011             unsigned i = u_bit_scan(&dirty_lo);
2012             sctx->atoms.array[i].emit(sctx, i);
2013          }
2014          while (dirty_hi) {
2015             unsigned i = 32 + u_bit_scan(&dirty_hi);
2016             sctx->atoms.array[i].emit(sctx, i);
2017          }
2018       }
2019    }
2020 }
2021 
2022 #define DRAW_CLEANUP do {                                 \
2023       if (index_size && indexbuf != info->index.resource) \
2024          pipe_resource_reference(&indexbuf, NULL);        \
2025    } while (0)
2026 
2027 template <amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS, si_has_ngg NGG,
2028           si_is_draw_vertex_state IS_DRAW_VERTEX_STATE, si_has_sh_pairs_packed HAS_SH_PAIRS_PACKED,
2029           util_popcnt POPCNT> ALWAYS_INLINE
si_draw(struct pipe_context * ctx,const struct pipe_draw_info * info,unsigned drawid_offset,const struct pipe_draw_indirect_info * indirect,const struct pipe_draw_start_count_bias * draws,unsigned num_draws,struct pipe_vertex_state * state,uint32_t partial_velem_mask)2030 static void si_draw(struct pipe_context *ctx,
2031                     const struct pipe_draw_info *info,
2032                     unsigned drawid_offset,
2033                     const struct pipe_draw_indirect_info *indirect,
2034                     const struct pipe_draw_start_count_bias *draws,
2035                     unsigned num_draws,
2036                     struct pipe_vertex_state *state,
2037                     uint32_t partial_velem_mask)
2038 {
2039    /* Keep code that uses the least number of local variables as close to the beginning
2040     * of this function as possible to minimize register pressure.
2041     *
2042     * It doesn't matter where we return due to invalid parameters because such cases
2043     * shouldn't occur in practice.
2044     */
2045    struct si_context *sctx = (struct si_context *)ctx;
2046 
2047    si_check_dirty_buffers_textures(sctx);
2048 
2049    if (GFX_VERSION < GFX11)
2050       gfx6_decompress_textures(sctx, u_bit_consecutive(0, SI_NUM_GRAPHICS_SHADERS));
2051    else if (GFX_VERSION < GFX12)
2052       gfx11_decompress_textures(sctx, u_bit_consecutive(0, SI_NUM_GRAPHICS_SHADERS));
2053 
2054    si_need_gfx_cs_space(sctx, num_draws);
2055 
2056    if (u_trace_perfetto_active(&sctx->ds.trace_context))
2057       trace_si_begin_draw(&sctx->trace);
2058 
2059    unsigned instance_count = info->instance_count;
2060 
2061    /* GFX6-GFX7 treat instance_count==0 as instance_count==1. There is
2062     * no workaround for indirect draws, but we can at least skip
2063     * direct draws.
2064     * 'instance_count == 0' seems to be problematic on Renoir chips (#4866),
2065     * so simplify the condition and drop these draws for all <= GFX9 chips.
2066     */
2067    if (GFX_VERSION <= GFX9 && unlikely(!IS_DRAW_VERTEX_STATE && !indirect && !instance_count))
2068       return;
2069 
2070    struct si_shader_selector *vs = sctx->shader.vs.cso;
2071    struct si_vertex_state *vstate = (struct si_vertex_state *)state;
2072    if (unlikely(!vs ||
2073                 (!IS_DRAW_VERTEX_STATE && sctx->num_vertex_elements < vs->info.num_vs_inputs) ||
2074                 (IS_DRAW_VERTEX_STATE && vstate->velems.count < vs->info.num_vs_inputs) ||
2075                 !sctx->shader.ps.cso || (HAS_TESS != (info->mode == MESA_PRIM_PATCHES)))) {
2076       assert(0);
2077       return;
2078    }
2079 
2080    enum mesa_prim prim = HAS_TESS ? MESA_PRIM_PATCHES : (enum mesa_prim)info->mode;
2081 
2082    if (GFX_VERSION <= GFX9 && HAS_GS) {
2083       /* Determine whether the GS triangle strip adjacency fix should
2084        * be applied. Rotate every other triangle if triangle strips with
2085        * adjacency are fed to the GS. This doesn't work if primitive
2086        * restart occurs after an odd number of triangles.
2087        */
2088       bool gs_tri_strip_adj_fix =
2089          !HAS_TESS && prim == MESA_PRIM_TRIANGLE_STRIP_ADJACENCY;
2090 
2091       if (gs_tri_strip_adj_fix != sctx->shader.gs.key.ge.mono.u.gs_tri_strip_adj_fix) {
2092          sctx->shader.gs.key.ge.mono.u.gs_tri_strip_adj_fix = gs_tri_strip_adj_fix;
2093          sctx->do_update_shaders = true;
2094       }
2095    }
2096 
2097    struct pipe_resource *indexbuf = info->index.resource;
2098    unsigned index_size = info->index_size;
2099    unsigned index_offset = indirect && indirect->buffer ? draws[0].start * index_size : 0;
2100 
2101    if (index_size) {
2102       /* Translate or upload, if needed. */
2103       /* 8-bit indices are supported on GFX8. */
2104       if (!IS_DRAW_VERTEX_STATE && GFX_VERSION <= GFX7 && index_size == 1) {
2105          unsigned start, count, start_offset, size;
2106 
2107          si_get_draw_start_count(sctx, info, indirect, draws, num_draws, &start, &count);
2108          start_offset = start * 2;
2109          size = count * 2;
2110 
2111          /* Don't use u_upload_alloc because we don't need to map the buffer for CPU access. */
2112          indexbuf = pipe_buffer_create(&sctx->screen->b, 0, PIPE_USAGE_IMMUTABLE, start_offset + size);
2113          if (unlikely(!indexbuf))
2114             return;
2115 
2116          si_compute_shorten_ubyte_buffer(sctx, indexbuf, info->index.resource, start_offset,
2117                                          index_offset + start, count, sctx->render_cond_enabled);
2118          si_barrier_after_simple_buffer_op(sctx, 0, indexbuf, info->index.resource);
2119 
2120          index_offset = 0;
2121          index_size = 2;
2122 
2123          /* GFX6-7 don't read index buffers through L2. */
2124          sctx->barrier_flags |= SI_BARRIER_WB_L2 | SI_BARRIER_PFP_SYNC_ME;
2125          si_mark_atom_dirty(sctx, &sctx->atoms.s.barrier);
2126          si_resource(indexbuf)->L2_cache_dirty = false;
2127       } else if (!IS_DRAW_VERTEX_STATE && info->has_user_indices) {
2128          unsigned start_offset;
2129 
2130          assert(!indirect);
2131          assert(num_draws == 1);
2132          start_offset = draws[0].start * index_size;
2133 
2134          indexbuf = NULL;
2135          u_upload_data(ctx->stream_uploader, start_offset, draws[0].count * index_size,
2136                        sctx->screen->info.tcc_cache_line_size,
2137                        (char *)info->index.user + start_offset, &index_offset, &indexbuf);
2138          if (unlikely(!indexbuf))
2139             return;
2140 
2141          /* info->start will be added by the drawing code */
2142          index_offset -= start_offset;
2143       } else if ((GFX_VERSION <= GFX7 || GFX_VERSION == GFX12) &&
2144                  si_resource(indexbuf)->L2_cache_dirty) {
2145          /* GFX8-GFX11 reads index buffers through L2, so it doesn't
2146           * need this. */
2147          sctx->barrier_flags |= SI_BARRIER_WB_L2 | SI_BARRIER_PFP_SYNC_ME;
2148          si_mark_atom_dirty(sctx, &sctx->atoms.s.barrier);
2149          si_resource(indexbuf)->L2_cache_dirty = false;
2150       }
2151    }
2152 
2153    unsigned min_direct_count = 0;
2154    unsigned total_direct_count = 0;
2155 
2156    if (!IS_DRAW_VERTEX_STATE && indirect) {
2157       /* Indirect buffers use L2 on GFX9-GFX11, but not other hw. */
2158       if (GFX_VERSION <= GFX8 || GFX_VERSION == GFX12) {
2159          if (indirect->buffer && si_resource(indirect->buffer)->L2_cache_dirty) {
2160             sctx->barrier_flags |= SI_BARRIER_WB_L2 | SI_BARRIER_PFP_SYNC_ME;
2161             si_mark_atom_dirty(sctx, &sctx->atoms.s.barrier);
2162             si_resource(indirect->buffer)->L2_cache_dirty = false;
2163          }
2164 
2165          if (indirect->indirect_draw_count &&
2166              si_resource(indirect->indirect_draw_count)->L2_cache_dirty) {
2167             sctx->barrier_flags |= SI_BARRIER_WB_L2 | SI_BARRIER_PFP_SYNC_ME;
2168             si_mark_atom_dirty(sctx, &sctx->atoms.s.barrier);
2169             si_resource(indirect->indirect_draw_count)->L2_cache_dirty = false;
2170          }
2171       }
2172       total_direct_count = INT_MAX; /* just set something other than 0 to enable shader culling */
2173    } else {
2174       total_direct_count = min_direct_count = draws[0].count;
2175 
2176       for (unsigned i = 1; i < num_draws; i++) {
2177          unsigned count = draws[i].count;
2178 
2179          total_direct_count += count;
2180          min_direct_count = MIN2(min_direct_count, count);
2181       }
2182    }
2183 
2184    /* Set the rasterization primitive type.
2185     *
2186     * This must be done after si_decompress_textures, which can call
2187     * draw_vbo recursively, and before si_update_shaders, which uses
2188     * current_rast_prim for this draw_vbo call.
2189     */
2190    if (!HAS_GS && !HAS_TESS) {
2191       enum mesa_prim rast_prim;
2192 
2193       if (util_rast_prim_is_triangles(prim)) {
2194          rast_prim = MESA_PRIM_TRIANGLES;
2195       } else {
2196          /* Only possibilities, POINTS, LINE*, RECTANGLES */
2197          rast_prim = prim;
2198       }
2199 
2200       si_set_rasterized_prim(sctx, rast_prim, si_get_vs_inline(sctx, HAS_TESS, HAS_GS)->current,
2201                              NGG);
2202    }
2203 
2204    if (IS_DRAW_VERTEX_STATE) {
2205       /* draw_vertex_state doesn't use the current vertex buffers and vertex elements,
2206        * so disable all VS input lowering.
2207        */
2208       if (!sctx->force_trivial_vs_inputs) {
2209          sctx->force_trivial_vs_inputs = true;
2210 
2211          /* Update shaders to disable VS input lowering. */
2212          if (sctx->uses_nontrivial_vs_inputs) {
2213             si_vs_key_update_inputs(sctx);
2214             sctx->do_update_shaders = true;
2215          }
2216       }
2217    } else {
2218       if (sctx->force_trivial_vs_inputs) {
2219          sctx->force_trivial_vs_inputs = false;
2220 
2221          /* Update shaders to possibly enable VS input lowering. */
2222          if (sctx->uses_nontrivial_vs_inputs) {
2223             si_vs_key_update_inputs(sctx);
2224             sctx->do_update_shaders = true;
2225          }
2226       }
2227    }
2228 
2229    /* Update NGG culling settings. */
2230    uint16_t old_ngg_culling = sctx->ngg_culling;
2231    if (GFX_VERSION >= GFX10) {
2232       struct si_shader_selector *hw_vs = si_get_vs_inline(sctx, HAS_TESS, HAS_GS)->cso;
2233 
2234       if (NGG &&
2235           /* Tessellation and GS set ngg_cull_vert_threshold to UINT_MAX if the prim type
2236            * is not points, so this check is only needed for VS. */
2237           (HAS_TESS || HAS_GS || util_rast_prim_is_lines_or_triangles(sctx->current_rast_prim)) &&
2238           /* Only the first draw for a shader starts with culling disabled and it's disabled
2239            * until we pass the total_direct_count check and then it stays enabled until
2240            * the shader is changed. This eliminates most culling on/off state changes. */
2241           (old_ngg_culling || total_direct_count > hw_vs->ngg_cull_vert_threshold)) {
2242          struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
2243 
2244          /* Check that the current shader allows culling. */
2245          assert(hw_vs->ngg_cull_vert_threshold != UINT_MAX);
2246 
2247          uint16_t ngg_culling;
2248 
2249          if (util_prim_is_lines(sctx->current_rast_prim)) {
2250             /* Overwrite it to mask out face cull flags. */
2251             ngg_culling = rs->ngg_cull_flags_lines;
2252          } else {
2253             ngg_culling = sctx->viewport0_y_inverted ? rs->ngg_cull_flags_tris_y_inverted :
2254                                                        rs->ngg_cull_flags_tris;
2255             assert(ngg_culling); /* rasterizer state should always set this to non-zero */
2256          }
2257 
2258          if (ngg_culling != old_ngg_culling) {
2259             /* If shader compilation is not ready, this setting will be rejected. */
2260             sctx->ngg_culling = ngg_culling;
2261             sctx->do_update_shaders = true;
2262          }
2263       } else if (old_ngg_culling) {
2264          sctx->ngg_culling = 0;
2265          sctx->do_update_shaders = true;
2266       }
2267    }
2268 
2269    if (unlikely(sctx->do_update_shaders)) {
2270       if (unlikely(!(si_update_shaders<GFX_VERSION, HAS_TESS, HAS_GS, NGG>(sctx)))) {
2271          DRAW_CLEANUP;
2272          return;
2273       }
2274    }
2275 
2276    /* This is the optimal packet order:
2277     * Set all states first, so that all SET packets are processed in parallel with previous
2278     * draw calls. Then flush caches and wait if needed. Then draw and prefetch at the end.
2279     * It's better to draw before prefetches because we want to start fetching indices before
2280     * shaders. The idea is to minimize the time when the CUs are idle.
2281     */
2282 
2283    /* Vega10/Raven scissor bug workaround. When any context register is
2284     * written (i.e. the GPU rolls the context), PA_SC_VPORT_SCISSOR
2285     * registers must be written too.
2286     */
2287    bool gfx9_scissor_bug = false;
2288    uint64_t masked_atoms = 0;
2289 
2290    if (GFX_VERSION == GFX9 && sctx->screen->info.has_gfx9_scissor_bug) {
2291       masked_atoms |= si_get_atom_bit(sctx, &sctx->atoms.s.scissors);
2292       gfx9_scissor_bug = true;
2293 
2294       if ((!IS_DRAW_VERTEX_STATE && indirect && indirect->count_from_stream_output) ||
2295           sctx->dirty_atoms & si_atoms_that_always_roll_context())
2296          sctx->context_roll = true;
2297    }
2298 
2299    bool primitive_restart = !IS_DRAW_VERTEX_STATE && info->primitive_restart;
2300 
2301    /* Emit states. */
2302    si_emit_rasterizer_prim_state<GFX_VERSION, HAS_GS, NGG>(sctx);
2303    /* This emits states and flushes caches. */
2304    si_emit_all_states(sctx, masked_atoms);
2305    /* This can be done after si_emit_all_states because it doesn't set barrier flags. */
2306    si_emit_draw_registers<GFX_VERSION, HAS_TESS, HAS_GS, NGG, IS_DRAW_VERTEX_STATE>
2307          (sctx, indirect, prim, index_size, instance_count, primitive_restart,
2308           info->restart_index, min_direct_count);
2309 
2310    /* <-- CUs are idle here if the barrier atom waited. */
2311 
2312    /* This must be done after si_emit_all_states, which can affect this. */
2313    si_emit_vs_state<GFX_VERSION, HAS_TESS, HAS_GS, NGG, IS_DRAW_VERTEX_STATE, HAS_SH_PAIRS_PACKED>
2314          (sctx, index_size);
2315 
2316    /* This needs to be done after cache flushes because ACQUIRE_MEM rolls the context. */
2317    if (GFX_VERSION == GFX9 && gfx9_scissor_bug &&
2318        (sctx->context_roll || si_is_atom_dirty(sctx, &sctx->atoms.s.scissors))) {
2319       sctx->atoms.s.scissors.emit(sctx, -1);
2320       sctx->dirty_atoms &= ~si_get_atom_bit(sctx, &sctx->atoms.s.scissors);
2321    }
2322    assert(sctx->dirty_atoms == 0);
2323 
2324    /* This uploads VBO descriptors, sets user SGPRs, and executes the L2 prefetch.
2325     * It should done after cache flushing.
2326     */
2327    if (unlikely((!si_upload_and_prefetch_VB_descriptors
2328                      <GFX_VERSION, HAS_TESS, HAS_GS, NGG, IS_DRAW_VERTEX_STATE, HAS_SH_PAIRS_PACKED, POPCNT>
2329                      (sctx, state, partial_velem_mask)))) {
2330       DRAW_CLEANUP;
2331       return;
2332    }
2333 
2334    si_emit_draw_packets<GFX_VERSION, HAS_TESS, HAS_GS, NGG, IS_DRAW_VERTEX_STATE, HAS_SH_PAIRS_PACKED>
2335          (sctx, info, drawid_offset, indirect, draws, num_draws, indexbuf,
2336           index_size, index_offset, instance_count);
2337    /* <-- CUs start to get busy here if we waited. */
2338 
2339    /* Start prefetches after the draw has been started. Both will run
2340     * in parallel, but starting the draw first is more important.
2341     */
2342    si_prefetch_shaders<GFX_VERSION, HAS_TESS, HAS_GS, NGG>(sctx);
2343 
2344    /* Clear the context roll flag after the draw call.
2345     * Only used by the gfx9 scissor bug.
2346     */
2347    if (GFX_VERSION == GFX9)
2348       sctx->context_roll = false;
2349 
2350    if (unlikely(sctx->current_saved_cs)) {
2351       si_trace_emit(sctx);
2352       si_log_draw_state(sctx, sctx->log);
2353    }
2354 
2355    /* Workaround for a VGT hang when streamout is enabled.
2356     * It must be done after drawing. */
2357    if (((GFX_VERSION == GFX7 && sctx->family == CHIP_HAWAII) ||
2358         (GFX_VERSION == GFX8 && (sctx->family == CHIP_TONGA || sctx->family == CHIP_FIJI))) &&
2359        si_get_strmout_en(sctx)) {
2360       radeon_begin(&sctx->gfx_cs);
2361       radeon_event_write(V_028A90_VGT_STREAMOUT_SYNC);
2362       radeon_end();
2363    }
2364 
2365    if (unlikely(GFX_VERSION < GFX12 && sctx->decompression_enabled)) {
2366       sctx->num_decompress_calls++;
2367    } else {
2368       sctx->num_draw_calls += num_draws;
2369    }
2370 
2371    /* On Gfx12, this is only used to detect whether a depth texture is in the cleared state. */
2372    if (sctx->framebuffer.state.zsbuf) {
2373       struct si_texture *zstex = (struct si_texture *)sctx->framebuffer.state.zsbuf->texture;
2374       zstex->depth_cleared_level_mask &= ~BITFIELD_BIT(sctx->framebuffer.state.zsbuf->u.tex.level);
2375    }
2376 
2377    if (u_trace_perfetto_active(&sctx->ds.trace_context)) {
2378       trace_si_end_draw(&sctx->trace, total_direct_count);
2379    }
2380 
2381    DRAW_CLEANUP;
2382 }
2383 
2384 template <amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS, si_has_ngg NGG,
2385           si_has_sh_pairs_packed HAS_SH_PAIRS_PACKED>
si_draw_vbo(struct pipe_context * ctx,const struct pipe_draw_info * info,unsigned drawid_offset,const struct pipe_draw_indirect_info * indirect,const struct pipe_draw_start_count_bias * draws,unsigned num_draws)2386 static void si_draw_vbo(struct pipe_context *ctx,
2387                         const struct pipe_draw_info *info,
2388                         unsigned drawid_offset,
2389                         const struct pipe_draw_indirect_info *indirect,
2390                         const struct pipe_draw_start_count_bias *draws,
2391                         unsigned num_draws)
2392 {
2393    si_draw<GFX_VERSION, HAS_TESS, HAS_GS, NGG, DRAW_VERTEX_STATE_OFF, HAS_SH_PAIRS_PACKED, POPCNT_NO>
2394       (ctx, info, drawid_offset, indirect, draws, num_draws, NULL, 0);
2395 }
2396 
2397 template <amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS, si_has_ngg NGG,
2398           si_has_sh_pairs_packed HAS_SH_PAIRS_PACKED, util_popcnt POPCNT>
si_draw_vertex_state(struct pipe_context * ctx,struct pipe_vertex_state * vstate,uint32_t partial_velem_mask,struct pipe_draw_vertex_state_info info,const struct pipe_draw_start_count_bias * draws,unsigned num_draws)2399 static void si_draw_vertex_state(struct pipe_context *ctx,
2400                                  struct pipe_vertex_state *vstate,
2401                                  uint32_t partial_velem_mask,
2402                                  struct pipe_draw_vertex_state_info info,
2403                                  const struct pipe_draw_start_count_bias *draws,
2404                                  unsigned num_draws)
2405 {
2406    struct si_vertex_state *state = (struct si_vertex_state *)vstate;
2407    struct pipe_draw_info dinfo = {};
2408 
2409    dinfo.mode = info.mode;
2410    dinfo.index_size = 4;
2411    dinfo.instance_count = 1;
2412    dinfo.index.resource = state->b.input.indexbuf;
2413 
2414    si_draw<GFX_VERSION, HAS_TESS, HAS_GS, NGG, DRAW_VERTEX_STATE_ON, HAS_SH_PAIRS_PACKED, POPCNT>
2415       (ctx, &dinfo, 0, NULL, draws, num_draws, vstate, partial_velem_mask);
2416 
2417    if (info.take_vertex_state_ownership)
2418       pipe_vertex_state_reference(&vstate, NULL);
2419 }
2420 
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)2421 static void si_draw_rectangle(struct blitter_context *blitter, void *vertex_elements_cso,
2422                               blitter_get_vs_func get_vs, int x1, int y1, int x2, int y2,
2423                               float depth, unsigned num_instances, enum blitter_attrib_type type,
2424                               const union blitter_attrib *attrib)
2425 {
2426    struct pipe_context *pipe = util_blitter_get_pipe(blitter);
2427    struct si_context *sctx = (struct si_context *)pipe;
2428    uint32_t attribute_ring_address_lo =
2429       sctx->gfx_level >= GFX11 ? sctx->screen->attribute_pos_prim_ring->gpu_address : 0;
2430 
2431    /* Pack position coordinates as signed int16. */
2432    sctx->vs_blit_sh_data[0] = (uint32_t)(x1 & 0xffff) | ((uint32_t)(y1 & 0xffff) << 16);
2433    sctx->vs_blit_sh_data[1] = (uint32_t)(x2 & 0xffff) | ((uint32_t)(y2 & 0xffff) << 16);
2434    sctx->vs_blit_sh_data[2] = fui(depth);
2435 
2436    switch (type) {
2437    case UTIL_BLITTER_ATTRIB_COLOR:
2438       memcpy(&sctx->vs_blit_sh_data[3], attrib->color, sizeof(float) * 4);
2439       sctx->vs_blit_sh_data[7] = attribute_ring_address_lo;
2440       break;
2441    case UTIL_BLITTER_ATTRIB_TEXCOORD_XY:
2442    case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW:
2443       memcpy(&sctx->vs_blit_sh_data[3], &attrib->texcoord, sizeof(attrib->texcoord));
2444       sctx->vs_blit_sh_data[9] = attribute_ring_address_lo;
2445       break;
2446    case UTIL_BLITTER_ATTRIB_NONE:;
2447    }
2448 
2449    pipe->bind_vs_state(pipe, si_get_blitter_vs(sctx, type, num_instances));
2450 
2451    struct pipe_draw_info info = {};
2452    struct pipe_draw_start_count_bias draw;
2453 
2454    info.mode = SI_PRIM_RECTANGLE_LIST;
2455    info.instance_count = num_instances;
2456 
2457    draw.start = 0;
2458    draw.count = 3;
2459 
2460    /* Blits don't use vertex buffers. */
2461    sctx->vertex_buffers_dirty = false;
2462 
2463    pipe->draw_vbo(pipe, &info, 0, NULL, &draw, 1);
2464 }
2465 
2466 template <amd_gfx_level GFX_VERSION, si_has_tess HAS_TESS, si_has_gs HAS_GS, si_has_ngg NGG,
2467           util_popcnt POPCNT>
si_init_draw_vbo(struct si_context * sctx)2468 static void si_init_draw_vbo(struct si_context *sctx)
2469 {
2470    if (NGG && GFX_VERSION < GFX10)
2471       return;
2472 
2473    if (!NGG && GFX_VERSION >= GFX11)
2474       return;
2475 
2476    if (GFX_VERSION >= GFX11 && GFX_VERSION < GFX12 && sctx->screen->info.has_set_sh_pairs_packed) {
2477       sctx->draw_vbo[HAS_TESS][HAS_GS][NGG] =
2478          si_draw_vbo<GFX_VERSION, HAS_TESS, HAS_GS, NGG, HAS_SH_PAIRS_PACKED_ON>;
2479 
2480       sctx->draw_vertex_state[HAS_TESS][HAS_GS][NGG] =
2481          si_draw_vertex_state<GFX_VERSION, HAS_TESS, HAS_GS, NGG, HAS_SH_PAIRS_PACKED_ON, POPCNT>;
2482    } else {
2483       sctx->draw_vbo[HAS_TESS][HAS_GS][NGG] =
2484          si_draw_vbo<GFX_VERSION, HAS_TESS, HAS_GS, NGG, HAS_SH_PAIRS_PACKED_OFF>;
2485 
2486       sctx->draw_vertex_state[HAS_TESS][HAS_GS][NGG] =
2487          si_draw_vertex_state<GFX_VERSION, HAS_TESS, HAS_GS, NGG, HAS_SH_PAIRS_PACKED_OFF, POPCNT>;
2488    }
2489 }
2490 
2491 template <amd_gfx_level GFX_VERSION>
si_init_draw_vbo_all_pipeline_options(struct si_context * sctx)2492 static void si_init_draw_vbo_all_pipeline_options(struct si_context *sctx)
2493 {
2494    if (util_get_cpu_caps()->has_popcnt) {
2495       si_init_draw_vbo<GFX_VERSION, TESS_OFF, GS_OFF, NGG_OFF, POPCNT_YES>(sctx);
2496       si_init_draw_vbo<GFX_VERSION, TESS_OFF, GS_ON,  NGG_OFF, POPCNT_YES>(sctx);
2497       si_init_draw_vbo<GFX_VERSION, TESS_ON,  GS_OFF, NGG_OFF, POPCNT_YES>(sctx);
2498       si_init_draw_vbo<GFX_VERSION, TESS_ON,  GS_ON,  NGG_OFF, POPCNT_YES>(sctx);
2499       si_init_draw_vbo<GFX_VERSION, TESS_OFF, GS_OFF, NGG_ON, POPCNT_YES>(sctx);
2500       si_init_draw_vbo<GFX_VERSION, TESS_OFF, GS_ON,  NGG_ON, POPCNT_YES>(sctx);
2501       si_init_draw_vbo<GFX_VERSION, TESS_ON,  GS_OFF, NGG_ON, POPCNT_YES>(sctx);
2502       si_init_draw_vbo<GFX_VERSION, TESS_ON,  GS_ON,  NGG_ON, POPCNT_YES>(sctx);
2503    } else {
2504       si_init_draw_vbo<GFX_VERSION, TESS_OFF, GS_OFF, NGG_OFF, POPCNT_NO>(sctx);
2505       si_init_draw_vbo<GFX_VERSION, TESS_OFF, GS_ON,  NGG_OFF, POPCNT_NO>(sctx);
2506       si_init_draw_vbo<GFX_VERSION, TESS_ON,  GS_OFF, NGG_OFF, POPCNT_NO>(sctx);
2507       si_init_draw_vbo<GFX_VERSION, TESS_ON,  GS_ON,  NGG_OFF, POPCNT_NO>(sctx);
2508       si_init_draw_vbo<GFX_VERSION, TESS_OFF, GS_OFF, NGG_ON, POPCNT_NO>(sctx);
2509       si_init_draw_vbo<GFX_VERSION, TESS_OFF, GS_ON,  NGG_ON, POPCNT_NO>(sctx);
2510       si_init_draw_vbo<GFX_VERSION, TESS_ON,  GS_OFF, NGG_ON, POPCNT_NO>(sctx);
2511       si_init_draw_vbo<GFX_VERSION, TESS_ON,  GS_ON,  NGG_ON, POPCNT_NO>(sctx);
2512    }
2513 }
2514 
si_invalid_draw_vbo(struct pipe_context * pipe,const struct pipe_draw_info * info,unsigned drawid_offset,const struct pipe_draw_indirect_info * indirect,const struct pipe_draw_start_count_bias * draws,unsigned num_draws)2515 static void si_invalid_draw_vbo(struct pipe_context *pipe,
2516                                 const struct pipe_draw_info *info,
2517                                 unsigned drawid_offset,
2518                                 const struct pipe_draw_indirect_info *indirect,
2519                                 const struct pipe_draw_start_count_bias *draws,
2520                                 unsigned num_draws)
2521 {
2522    unreachable("vertex shader not bound");
2523 }
2524 
si_invalid_draw_vertex_state(struct pipe_context * ctx,struct pipe_vertex_state * vstate,uint32_t partial_velem_mask,struct pipe_draw_vertex_state_info info,const struct pipe_draw_start_count_bias * draws,unsigned num_draws)2525 static void si_invalid_draw_vertex_state(struct pipe_context *ctx,
2526                                          struct pipe_vertex_state *vstate,
2527                                          uint32_t partial_velem_mask,
2528                                          struct pipe_draw_vertex_state_info info,
2529                                          const struct pipe_draw_start_count_bias *draws,
2530                                          unsigned num_draws)
2531 {
2532    unreachable("vertex shader not bound");
2533 }
2534 
2535 extern "C"
GFX(si_init_draw_functions_)2536 void GFX(si_init_draw_functions_)(struct si_context *sctx)
2537 {
2538    assert(sctx->gfx_level == GFX());
2539 
2540    si_init_draw_vbo_all_pipeline_options<GFX()>(sctx);
2541 
2542    /* Bind a fake draw_vbo, so that draw_vbo isn't NULL, which would skip
2543     * initialization of callbacks in upper layers (such as u_threaded_context).
2544     */
2545    sctx->b.draw_vbo = si_invalid_draw_vbo;
2546    sctx->b.draw_vertex_state = si_invalid_draw_vertex_state;
2547    sctx->blitter->draw_rectangle = si_draw_rectangle;
2548 
2549    si_init_ia_multi_vgt_param_table(sctx);
2550 }
2551