2 * Copyright 2012 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 #include "si_build_pm4.h"
28 #include "util/u_index_modify.h"
29 #include "util/u_log.h"
30 #include "util/u_upload_mgr.h"
31 #include "util/u_prim.h"
35 /* special primitive types */
36 #define SI_PRIM_RECTANGLE_LIST PIPE_PRIM_MAX
38 static unsigned si_conv_pipe_prim(unsigned mode
)
40 static const unsigned prim_conv
[] = {
41 [PIPE_PRIM_POINTS
] = V_008958_DI_PT_POINTLIST
,
42 [PIPE_PRIM_LINES
] = V_008958_DI_PT_LINELIST
,
43 [PIPE_PRIM_LINE_LOOP
] = V_008958_DI_PT_LINELOOP
,
44 [PIPE_PRIM_LINE_STRIP
] = V_008958_DI_PT_LINESTRIP
,
45 [PIPE_PRIM_TRIANGLES
] = V_008958_DI_PT_TRILIST
,
46 [PIPE_PRIM_TRIANGLE_STRIP
] = V_008958_DI_PT_TRISTRIP
,
47 [PIPE_PRIM_TRIANGLE_FAN
] = V_008958_DI_PT_TRIFAN
,
48 [PIPE_PRIM_QUADS
] = V_008958_DI_PT_QUADLIST
,
49 [PIPE_PRIM_QUAD_STRIP
] = V_008958_DI_PT_QUADSTRIP
,
50 [PIPE_PRIM_POLYGON
] = V_008958_DI_PT_POLYGON
,
51 [PIPE_PRIM_LINES_ADJACENCY
] = V_008958_DI_PT_LINELIST_ADJ
,
52 [PIPE_PRIM_LINE_STRIP_ADJACENCY
] = V_008958_DI_PT_LINESTRIP_ADJ
,
53 [PIPE_PRIM_TRIANGLES_ADJACENCY
] = V_008958_DI_PT_TRILIST_ADJ
,
54 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
] = V_008958_DI_PT_TRISTRIP_ADJ
,
55 [PIPE_PRIM_PATCHES
] = V_008958_DI_PT_PATCH
,
56 [SI_PRIM_RECTANGLE_LIST
] = V_008958_DI_PT_RECTLIST
58 assert(mode
< ARRAY_SIZE(prim_conv
));
59 return prim_conv
[mode
];
63 * This calculates the LDS size for tessellation shaders (VS, TCS, TES).
64 * LS.LDS_SIZE is shared by all 3 shader stages.
66 * The information about LDS and other non-compile-time parameters is then
67 * written to userdata SGPRs.
69 static bool si_emit_derived_tess_state(struct si_context
*sctx
,
70 const struct pipe_draw_info
*info
,
71 unsigned *num_patches
)
73 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
74 struct si_shader
*ls_current
;
75 struct si_shader_selector
*ls
;
76 /* The TES pointer will only be used for sctx->last_tcs.
77 * It would be wrong to think that TCS = TES. */
78 struct si_shader_selector
*tcs
=
79 sctx
->tcs_shader
.cso
? sctx
->tcs_shader
.cso
: sctx
->tes_shader
.cso
;
80 unsigned tess_uses_primid
= sctx
->ia_multi_vgt_param_key
.u
.tess_uses_prim_id
;
81 bool has_primid_instancing_bug
= sctx
->chip_class
== SI
&&
82 sctx
->screen
->info
.max_se
== 1;
83 unsigned tes_sh_base
= sctx
->shader_pointers
.sh_base
[PIPE_SHADER_TESS_EVAL
];
84 unsigned num_tcs_input_cp
= info
->vertices_per_patch
;
85 unsigned num_tcs_output_cp
, num_tcs_inputs
, num_tcs_outputs
;
86 unsigned num_tcs_patch_outputs
;
87 unsigned input_vertex_size
, output_vertex_size
, pervertex_output_patch_size
;
88 unsigned input_patch_size
, output_patch_size
, output_patch0_offset
;
89 unsigned perpatch_output_offset
, lds_size
;
90 unsigned tcs_in_layout
, tcs_out_layout
, tcs_out_offsets
;
91 unsigned offchip_layout
, hardware_lds_size
, ls_hs_config
;
93 /* Since GFX9 has merged LS-HS in the TCS state, set LS = TCS. */
94 if (sctx
->chip_class
>= GFX9
) {
95 if (sctx
->tcs_shader
.cso
)
96 ls_current
= sctx
->tcs_shader
.current
;
98 ls_current
= sctx
->fixed_func_tcs_shader
.current
;
100 ls
= ls_current
->key
.part
.tcs
.ls
;
102 ls_current
= sctx
->vs_shader
.current
;
103 ls
= sctx
->vs_shader
.cso
;
106 if (sctx
->last_ls
== ls_current
&&
107 sctx
->last_tcs
== tcs
&&
108 sctx
->last_tes_sh_base
== tes_sh_base
&&
109 sctx
->last_num_tcs_input_cp
== num_tcs_input_cp
&&
110 (!has_primid_instancing_bug
||
111 (sctx
->last_tess_uses_primid
== tess_uses_primid
))) {
112 *num_patches
= sctx
->last_num_patches
;
116 sctx
->last_ls
= ls_current
;
117 sctx
->last_tcs
= tcs
;
118 sctx
->last_tes_sh_base
= tes_sh_base
;
119 sctx
->last_num_tcs_input_cp
= num_tcs_input_cp
;
120 sctx
->last_tess_uses_primid
= tess_uses_primid
;
122 /* This calculates how shader inputs and outputs among VS, TCS, and TES
123 * are laid out in LDS. */
124 num_tcs_inputs
= util_last_bit64(ls
->outputs_written
);
126 if (sctx
->tcs_shader
.cso
) {
127 num_tcs_outputs
= util_last_bit64(tcs
->outputs_written
);
128 num_tcs_output_cp
= tcs
->info
.properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
];
129 num_tcs_patch_outputs
= util_last_bit64(tcs
->patch_outputs_written
);
131 /* No TCS. Route varyings from LS to TES. */
132 num_tcs_outputs
= num_tcs_inputs
;
133 num_tcs_output_cp
= num_tcs_input_cp
;
134 num_tcs_patch_outputs
= 2; /* TESSINNER + TESSOUTER */
137 input_vertex_size
= num_tcs_inputs
* 16;
138 output_vertex_size
= num_tcs_outputs
* 16;
140 input_patch_size
= num_tcs_input_cp
* input_vertex_size
;
142 pervertex_output_patch_size
= num_tcs_output_cp
* output_vertex_size
;
143 output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
145 /* Ensure that we only need one wave per SIMD so we don't need to check
146 * resource usage. Also ensures that the number of tcs in and out
147 * vertices per threadgroup are at most 256.
149 unsigned max_verts_per_patch
= MAX2(num_tcs_input_cp
, num_tcs_output_cp
);
150 *num_patches
= 256 / max_verts_per_patch
;
152 /* Make sure that the data fits in LDS. This assumes the shaders only
153 * use LDS for the inputs and outputs.
155 * While CIK can use 64K per threadgroup, there is a hang on Stoney
156 * with 2 CUs if we use more than 32K. The closed Vulkan driver also
157 * uses 32K at most on all GCN chips.
159 hardware_lds_size
= 32768;
160 *num_patches
= MIN2(*num_patches
, hardware_lds_size
/ (input_patch_size
+
163 /* Make sure the output data fits in the offchip buffer */
164 *num_patches
= MIN2(*num_patches
,
165 (sctx
->screen
->tess_offchip_block_dw_size
* 4) /
168 /* Not necessary for correctness, but improves performance. The
169 * specific value is taken from the proprietary driver.
171 *num_patches
= MIN2(*num_patches
, 40);
173 if (sctx
->chip_class
== SI
) {
174 /* SI bug workaround, related to power management. Limit LS-HS
175 * threadgroups to only one wave.
177 unsigned one_wave
= 64 / max_verts_per_patch
;
178 *num_patches
= MIN2(*num_patches
, one_wave
);
181 /* The VGT HS block increments the patch ID unconditionally
182 * within a single threadgroup. This results in incorrect
183 * patch IDs when instanced draws are used.
185 * The intended solution is to restrict threadgroups to
186 * a single instance by setting SWITCH_ON_EOI, which
187 * should cause IA to split instances up. However, this
188 * doesn't work correctly on SI when there is no other
191 if (has_primid_instancing_bug
&& tess_uses_primid
)
194 sctx
->last_num_patches
= *num_patches
;
196 output_patch0_offset
= input_patch_size
* *num_patches
;
197 perpatch_output_offset
= output_patch0_offset
+ pervertex_output_patch_size
;
199 /* Compute userdata SGPRs. */
200 assert(((input_vertex_size
/ 4) & ~0xff) == 0);
201 assert(((output_vertex_size
/ 4) & ~0xff) == 0);
202 assert(((input_patch_size
/ 4) & ~0x1fff) == 0);
203 assert(((output_patch_size
/ 4) & ~0x1fff) == 0);
204 assert(((output_patch0_offset
/ 16) & ~0xffff) == 0);
205 assert(((perpatch_output_offset
/ 16) & ~0xffff) == 0);
206 assert(num_tcs_input_cp
<= 32);
207 assert(num_tcs_output_cp
<= 32);
209 uint64_t ring_va
= r600_resource(sctx
->tess_rings
)->gpu_address
;
210 assert((ring_va
& u_bit_consecutive(0, 19)) == 0);
212 tcs_in_layout
= S_VS_STATE_LS_OUT_PATCH_SIZE(input_patch_size
/ 4) |
213 S_VS_STATE_LS_OUT_VERTEX_SIZE(input_vertex_size
/ 4);
214 tcs_out_layout
= (output_patch_size
/ 4) |
215 (num_tcs_input_cp
<< 13) |
217 tcs_out_offsets
= (output_patch0_offset
/ 16) |
218 ((perpatch_output_offset
/ 16) << 16);
219 offchip_layout
= *num_patches
|
220 (num_tcs_output_cp
<< 6) |
221 (pervertex_output_patch_size
* *num_patches
<< 12);
223 /* Compute the LDS size. */
224 lds_size
= output_patch0_offset
+ output_patch_size
* *num_patches
;
226 if (sctx
->chip_class
>= CIK
) {
227 assert(lds_size
<= 65536);
228 lds_size
= align(lds_size
, 512) / 512;
230 assert(lds_size
<= 32768);
231 lds_size
= align(lds_size
, 256) / 256;
234 /* Set SI_SGPR_VS_STATE_BITS. */
235 sctx
->current_vs_state
&= C_VS_STATE_LS_OUT_PATCH_SIZE
&
236 C_VS_STATE_LS_OUT_VERTEX_SIZE
;
237 sctx
->current_vs_state
|= tcs_in_layout
;
239 if (sctx
->chip_class
>= GFX9
) {
240 unsigned hs_rsrc2
= ls_current
->config
.rsrc2
|
241 S_00B42C_LDS_SIZE(lds_size
);
243 radeon_set_sh_reg(cs
, R_00B42C_SPI_SHADER_PGM_RSRC2_HS
, hs_rsrc2
);
245 /* Set userdata SGPRs for merged LS-HS. */
246 radeon_set_sh_reg_seq(cs
,
247 R_00B430_SPI_SHADER_USER_DATA_LS_0
+
248 GFX9_SGPR_TCS_OFFCHIP_LAYOUT
* 4, 3);
249 radeon_emit(cs
, offchip_layout
);
250 radeon_emit(cs
, tcs_out_offsets
);
251 radeon_emit(cs
, tcs_out_layout
);
253 unsigned ls_rsrc2
= ls_current
->config
.rsrc2
;
255 si_multiwave_lds_size_workaround(sctx
->screen
, &lds_size
);
256 ls_rsrc2
|= S_00B52C_LDS_SIZE(lds_size
);
258 /* Due to a hw bug, RSRC2_LS must be written twice with another
259 * LS register written in between. */
260 if (sctx
->chip_class
== CIK
&& sctx
->family
!= CHIP_HAWAII
)
261 radeon_set_sh_reg(cs
, R_00B52C_SPI_SHADER_PGM_RSRC2_LS
, ls_rsrc2
);
262 radeon_set_sh_reg_seq(cs
, R_00B528_SPI_SHADER_PGM_RSRC1_LS
, 2);
263 radeon_emit(cs
, ls_current
->config
.rsrc1
);
264 radeon_emit(cs
, ls_rsrc2
);
266 /* Set userdata SGPRs for TCS. */
267 radeon_set_sh_reg_seq(cs
,
268 R_00B430_SPI_SHADER_USER_DATA_HS_0
+ GFX6_SGPR_TCS_OFFCHIP_LAYOUT
* 4, 4);
269 radeon_emit(cs
, offchip_layout
);
270 radeon_emit(cs
, tcs_out_offsets
);
271 radeon_emit(cs
, tcs_out_layout
);
272 radeon_emit(cs
, tcs_in_layout
);
275 /* Set userdata SGPRs for TES. */
276 radeon_set_sh_reg_seq(cs
, tes_sh_base
+ SI_SGPR_TES_OFFCHIP_LAYOUT
* 4, 2);
277 radeon_emit(cs
, offchip_layout
);
278 radeon_emit(cs
, ring_va
);
280 ls_hs_config
= S_028B58_NUM_PATCHES(*num_patches
) |
281 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp
) |
282 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp
);
284 if (sctx
->last_ls_hs_config
!= ls_hs_config
) {
285 if (sctx
->chip_class
>= CIK
) {
286 radeon_set_context_reg_idx(cs
, R_028B58_VGT_LS_HS_CONFIG
, 2,
289 radeon_set_context_reg(cs
, R_028B58_VGT_LS_HS_CONFIG
,
292 sctx
->last_ls_hs_config
= ls_hs_config
;
293 return true; /* true if the context rolls */
298 static unsigned si_num_prims_for_vertices(const struct pipe_draw_info
*info
)
300 switch (info
->mode
) {
301 case PIPE_PRIM_PATCHES
:
302 return info
->count
/ info
->vertices_per_patch
;
303 case SI_PRIM_RECTANGLE_LIST
:
304 return info
->count
/ 3;
306 return u_prims_for_vertices(info
->mode
, info
->count
);
311 si_get_init_multi_vgt_param(struct si_screen
*sscreen
,
312 union si_vgt_param_key
*key
)
314 STATIC_ASSERT(sizeof(union si_vgt_param_key
) == 4);
315 unsigned max_primgroup_in_wave
= 2;
317 /* SWITCH_ON_EOP(0) is always preferable. */
318 bool wd_switch_on_eop
= false;
319 bool ia_switch_on_eop
= false;
320 bool ia_switch_on_eoi
= false;
321 bool partial_vs_wave
= false;
322 bool partial_es_wave
= false;
324 if (key
->u
.uses_tess
) {
325 /* SWITCH_ON_EOI must be set if PrimID is used. */
326 if (key
->u
.tess_uses_prim_id
)
327 ia_switch_on_eoi
= true;
329 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
330 if ((sscreen
->info
.family
== CHIP_TAHITI
||
331 sscreen
->info
.family
== CHIP_PITCAIRN
||
332 sscreen
->info
.family
== CHIP_BONAIRE
) &&
334 partial_vs_wave
= true;
336 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
337 if (sscreen
->has_distributed_tess
) {
338 if (key
->u
.uses_gs
) {
339 if (sscreen
->info
.chip_class
<= VI
)
340 partial_es_wave
= true;
342 /* GPU hang workaround. */
343 if (sscreen
->info
.family
== CHIP_TONGA
||
344 sscreen
->info
.family
== CHIP_FIJI
||
345 sscreen
->info
.family
== CHIP_POLARIS10
||
346 sscreen
->info
.family
== CHIP_POLARIS11
||
347 sscreen
->info
.family
== CHIP_POLARIS12
||
348 sscreen
->info
.family
== CHIP_VEGAM
)
349 partial_vs_wave
= true;
351 partial_vs_wave
= true;
356 /* This is a hardware requirement. */
357 if (key
->u
.line_stipple_enabled
||
358 (sscreen
->debug_flags
& DBG(SWITCH_ON_EOP
))) {
359 ia_switch_on_eop
= true;
360 wd_switch_on_eop
= true;
363 if (sscreen
->info
.chip_class
>= CIK
) {
364 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
365 * 4 shader engines. Set 1 to pass the assertion below.
366 * The other cases are hardware requirements.
368 * Polaris supports primitive restart with WD_SWITCH_ON_EOP=0
369 * for points, line strips, and tri strips.
371 if (sscreen
->info
.max_se
< 4 ||
372 key
->u
.prim
== PIPE_PRIM_POLYGON
||
373 key
->u
.prim
== PIPE_PRIM_LINE_LOOP
||
374 key
->u
.prim
== PIPE_PRIM_TRIANGLE_FAN
||
375 key
->u
.prim
== PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
||
376 (key
->u
.primitive_restart
&&
377 (sscreen
->info
.family
< CHIP_POLARIS10
||
378 (key
->u
.prim
!= PIPE_PRIM_POINTS
&&
379 key
->u
.prim
!= PIPE_PRIM_LINE_STRIP
&&
380 key
->u
.prim
!= PIPE_PRIM_TRIANGLE_STRIP
))) ||
381 key
->u
.count_from_stream_output
)
382 wd_switch_on_eop
= true;
384 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
385 * We don't know that for indirect drawing, so treat it as
386 * always problematic. */
387 if (sscreen
->info
.family
== CHIP_HAWAII
&&
388 key
->u
.uses_instancing
)
389 wd_switch_on_eop
= true;
391 /* Performance recommendation for 4 SE Gfx7-8 parts if
392 * instances are smaller than a primgroup.
393 * Assume indirect draws always use small instances.
394 * This is needed for good VS wave utilization.
396 if (sscreen
->info
.chip_class
<= VI
&&
397 sscreen
->info
.max_se
== 4 &&
398 key
->u
.multi_instances_smaller_than_primgroup
)
399 wd_switch_on_eop
= true;
401 /* Required on CIK and later. */
402 if (sscreen
->info
.max_se
> 2 && !wd_switch_on_eop
)
403 ia_switch_on_eoi
= true;
405 /* Required by Hawaii and, for some special cases, by VI. */
406 if (ia_switch_on_eoi
&&
407 (sscreen
->info
.family
== CHIP_HAWAII
||
408 (sscreen
->info
.chip_class
== VI
&&
409 (key
->u
.uses_gs
|| max_primgroup_in_wave
!= 2))))
410 partial_vs_wave
= true;
412 /* Instancing bug on Bonaire. */
413 if (sscreen
->info
.family
== CHIP_BONAIRE
&& ia_switch_on_eoi
&&
414 key
->u
.uses_instancing
)
415 partial_vs_wave
= true;
417 /* If the WD switch is false, the IA switch must be false too. */
418 assert(wd_switch_on_eop
|| !ia_switch_on_eop
);
421 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
422 if (sscreen
->info
.chip_class
<= VI
&& ia_switch_on_eoi
)
423 partial_es_wave
= true;
425 return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop
) |
426 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi
) |
427 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave
) |
428 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave
) |
429 S_028AA8_WD_SWITCH_ON_EOP(sscreen
->info
.chip_class
>= CIK
? wd_switch_on_eop
: 0) |
430 /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
431 S_028AA8_MAX_PRIMGRP_IN_WAVE(sscreen
->info
.chip_class
== VI
?
432 max_primgroup_in_wave
: 0) |
433 S_030960_EN_INST_OPT_BASIC(sscreen
->info
.chip_class
>= GFX9
) |
434 S_030960_EN_INST_OPT_ADV(sscreen
->info
.chip_class
>= GFX9
);
437 void si_init_ia_multi_vgt_param_table(struct si_context
*sctx
)
439 for (int prim
= 0; prim
<= SI_PRIM_RECTANGLE_LIST
; prim
++)
440 for (int uses_instancing
= 0; uses_instancing
< 2; uses_instancing
++)
441 for (int multi_instances
= 0; multi_instances
< 2; multi_instances
++)
442 for (int primitive_restart
= 0; primitive_restart
< 2; primitive_restart
++)
443 for (int count_from_so
= 0; count_from_so
< 2; count_from_so
++)
444 for (int line_stipple
= 0; line_stipple
< 2; line_stipple
++)
445 for (int uses_tess
= 0; uses_tess
< 2; uses_tess
++)
446 for (int tess_uses_primid
= 0; tess_uses_primid
< 2; tess_uses_primid
++)
447 for (int uses_gs
= 0; uses_gs
< 2; uses_gs
++) {
448 union si_vgt_param_key key
;
452 key
.u
.uses_instancing
= uses_instancing
;
453 key
.u
.multi_instances_smaller_than_primgroup
= multi_instances
;
454 key
.u
.primitive_restart
= primitive_restart
;
455 key
.u
.count_from_stream_output
= count_from_so
;
456 key
.u
.line_stipple_enabled
= line_stipple
;
457 key
.u
.uses_tess
= uses_tess
;
458 key
.u
.tess_uses_prim_id
= tess_uses_primid
;
459 key
.u
.uses_gs
= uses_gs
;
461 sctx
->ia_multi_vgt_param
[key
.index
] =
462 si_get_init_multi_vgt_param(sctx
->screen
, &key
);
466 static unsigned si_get_ia_multi_vgt_param(struct si_context
*sctx
,
467 const struct pipe_draw_info
*info
,
468 unsigned num_patches
)
470 union si_vgt_param_key key
= sctx
->ia_multi_vgt_param_key
;
471 unsigned primgroup_size
;
472 unsigned ia_multi_vgt_param
;
474 if (sctx
->tes_shader
.cso
) {
475 primgroup_size
= num_patches
; /* must be a multiple of NUM_PATCHES */
476 } else if (sctx
->gs_shader
.cso
) {
477 primgroup_size
= 64; /* recommended with a GS */
479 primgroup_size
= 128; /* recommended without a GS and tess */
482 key
.u
.prim
= info
->mode
;
483 key
.u
.uses_instancing
= info
->indirect
|| info
->instance_count
> 1;
484 key
.u
.multi_instances_smaller_than_primgroup
=
486 (info
->instance_count
> 1 &&
487 (info
->count_from_stream_output
||
488 si_num_prims_for_vertices(info
) < primgroup_size
));
489 key
.u
.primitive_restart
= info
->primitive_restart
;
490 key
.u
.count_from_stream_output
= info
->count_from_stream_output
!= NULL
;
492 ia_multi_vgt_param
= sctx
->ia_multi_vgt_param
[key
.index
] |
493 S_028AA8_PRIMGROUP_SIZE(primgroup_size
- 1);
495 if (sctx
->gs_shader
.cso
) {
496 /* GS requirement. */
497 if (sctx
->chip_class
<= VI
&&
498 SI_GS_PER_ES
/ primgroup_size
>= sctx
->screen
->gs_table_depth
- 3)
499 ia_multi_vgt_param
|= S_028AA8_PARTIAL_ES_WAVE_ON(1);
501 /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
502 * The hw doc says all multi-SE chips are affected, but Vulkan
503 * only applies it to Hawaii. Do what Vulkan does.
505 if (sctx
->family
== CHIP_HAWAII
&&
506 G_028AA8_SWITCH_ON_EOI(ia_multi_vgt_param
) &&
508 (info
->instance_count
> 1 &&
509 (info
->count_from_stream_output
||
510 si_num_prims_for_vertices(info
) <= 1))))
511 sctx
->flags
|= SI_CONTEXT_VGT_FLUSH
;
514 return ia_multi_vgt_param
;
517 /* rast_prim is the primitive type after GS. */
518 static bool si_emit_rasterizer_prim_state(struct si_context
*sctx
)
520 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
521 enum pipe_prim_type rast_prim
= sctx
->current_rast_prim
;
522 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
524 /* Skip this if not rendering lines. */
525 if (!util_prim_is_lines(rast_prim
))
528 if (rast_prim
== sctx
->last_rast_prim
&&
529 rs
->pa_sc_line_stipple
== sctx
->last_sc_line_stipple
)
532 /* For lines, reset the stipple pattern at each primitive. Otherwise,
533 * reset the stipple pattern at each packet (line strips, line loops).
535 radeon_set_context_reg(cs
, R_028A0C_PA_SC_LINE_STIPPLE
,
536 rs
->pa_sc_line_stipple
|
537 S_028A0C_AUTO_RESET_CNTL(rast_prim
== PIPE_PRIM_LINES
? 1 : 2));
539 sctx
->last_rast_prim
= rast_prim
;
540 sctx
->last_sc_line_stipple
= rs
->pa_sc_line_stipple
;
541 return true; /* true if the context rolls */
544 static void si_emit_vs_state(struct si_context
*sctx
,
545 const struct pipe_draw_info
*info
)
547 sctx
->current_vs_state
&= C_VS_STATE_INDEXED
;
548 sctx
->current_vs_state
|= S_VS_STATE_INDEXED(!!info
->index_size
);
550 if (sctx
->num_vs_blit_sgprs
) {
551 /* Re-emit the state after we leave u_blitter. */
552 sctx
->last_vs_state
= ~0;
556 if (sctx
->current_vs_state
!= sctx
->last_vs_state
) {
557 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
559 radeon_set_sh_reg(cs
,
560 sctx
->shader_pointers
.sh_base
[PIPE_SHADER_VERTEX
] +
561 SI_SGPR_VS_STATE_BITS
* 4,
562 sctx
->current_vs_state
);
564 sctx
->last_vs_state
= sctx
->current_vs_state
;
568 static inline bool si_prim_restart_index_changed(struct si_context
*sctx
,
569 const struct pipe_draw_info
*info
)
571 return info
->primitive_restart
&&
572 (info
->restart_index
!= sctx
->last_restart_index
||
573 sctx
->last_restart_index
== SI_RESTART_INDEX_UNKNOWN
);
576 static void si_emit_draw_registers(struct si_context
*sctx
,
577 const struct pipe_draw_info
*info
,
578 unsigned num_patches
)
580 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
581 unsigned prim
= si_conv_pipe_prim(info
->mode
);
582 unsigned ia_multi_vgt_param
;
584 ia_multi_vgt_param
= si_get_ia_multi_vgt_param(sctx
, info
, num_patches
);
587 if (ia_multi_vgt_param
!= sctx
->last_multi_vgt_param
) {
588 if (sctx
->chip_class
>= GFX9
)
589 radeon_set_uconfig_reg_idx(cs
, R_030960_IA_MULTI_VGT_PARAM
, 4, ia_multi_vgt_param
);
590 else if (sctx
->chip_class
>= CIK
)
591 radeon_set_context_reg_idx(cs
, R_028AA8_IA_MULTI_VGT_PARAM
, 1, ia_multi_vgt_param
);
593 radeon_set_context_reg(cs
, R_028AA8_IA_MULTI_VGT_PARAM
, ia_multi_vgt_param
);
595 sctx
->last_multi_vgt_param
= ia_multi_vgt_param
;
597 if (prim
!= sctx
->last_prim
) {
598 if (sctx
->chip_class
>= CIK
)
599 radeon_set_uconfig_reg_idx(cs
, R_030908_VGT_PRIMITIVE_TYPE
, 1, prim
);
601 radeon_set_config_reg(cs
, R_008958_VGT_PRIMITIVE_TYPE
, prim
);
603 sctx
->last_prim
= prim
;
606 /* Primitive restart. */
607 if (info
->primitive_restart
!= sctx
->last_primitive_restart_en
) {
608 if (sctx
->chip_class
>= GFX9
)
609 radeon_set_uconfig_reg(cs
, R_03092C_VGT_MULTI_PRIM_IB_RESET_EN
,
610 info
->primitive_restart
);
612 radeon_set_context_reg(cs
, R_028A94_VGT_MULTI_PRIM_IB_RESET_EN
,
613 info
->primitive_restart
);
615 sctx
->last_primitive_restart_en
= info
->primitive_restart
;
618 if (si_prim_restart_index_changed(sctx
, info
)) {
619 radeon_set_context_reg(cs
, R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX
,
620 info
->restart_index
);
621 sctx
->last_restart_index
= info
->restart_index
;
625 static void si_emit_draw_packets(struct si_context
*sctx
,
626 const struct pipe_draw_info
*info
,
627 struct pipe_resource
*indexbuf
,
629 unsigned index_offset
)
631 struct pipe_draw_indirect_info
*indirect
= info
->indirect
;
632 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
633 unsigned sh_base_reg
= sctx
->shader_pointers
.sh_base
[PIPE_SHADER_VERTEX
];
634 bool render_cond_bit
= sctx
->render_cond
&& !sctx
->render_cond_force_off
;
635 uint32_t index_max_size
= 0;
636 uint64_t index_va
= 0;
638 if (info
->count_from_stream_output
) {
639 struct si_streamout_target
*t
=
640 (struct si_streamout_target
*)info
->count_from_stream_output
;
641 uint64_t va
= t
->buf_filled_size
->gpu_address
+
642 t
->buf_filled_size_offset
;
644 radeon_set_context_reg(cs
, R_028B30_VGT_STRMOUT_DRAW_OPAQUE_VERTEX_STRIDE
,
647 radeon_emit(cs
, PKT3(PKT3_COPY_DATA
, 4, 0));
648 radeon_emit(cs
, COPY_DATA_SRC_SEL(COPY_DATA_MEM
) |
649 COPY_DATA_DST_SEL(COPY_DATA_REG
) |
650 COPY_DATA_WR_CONFIRM
);
651 radeon_emit(cs
, va
); /* src address lo */
652 radeon_emit(cs
, va
>> 32); /* src address hi */
653 radeon_emit(cs
, R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE
>> 2);
654 radeon_emit(cs
, 0); /* unused */
656 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
657 t
->buf_filled_size
, RADEON_USAGE_READ
,
658 RADEON_PRIO_SO_FILLED_SIZE
);
663 if (index_size
!= sctx
->last_index_size
) {
667 switch (index_size
) {
669 index_type
= V_028A7C_VGT_INDEX_8
;
672 index_type
= V_028A7C_VGT_INDEX_16
|
673 (SI_BIG_ENDIAN
&& sctx
->chip_class
<= CIK
?
674 V_028A7C_VGT_DMA_SWAP_16_BIT
: 0);
677 index_type
= V_028A7C_VGT_INDEX_32
|
678 (SI_BIG_ENDIAN
&& sctx
->chip_class
<= CIK
?
679 V_028A7C_VGT_DMA_SWAP_32_BIT
: 0);
682 assert(!"unreachable");
686 if (sctx
->chip_class
>= GFX9
) {
687 radeon_set_uconfig_reg_idx(cs
, R_03090C_VGT_INDEX_TYPE
,
690 radeon_emit(cs
, PKT3(PKT3_INDEX_TYPE
, 0, 0));
691 radeon_emit(cs
, index_type
);
694 sctx
->last_index_size
= index_size
;
697 index_max_size
= (indexbuf
->width0
- index_offset
) /
699 index_va
= r600_resource(indexbuf
)->gpu_address
+ index_offset
;
701 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
702 r600_resource(indexbuf
),
703 RADEON_USAGE_READ
, RADEON_PRIO_INDEX_BUFFER
);
705 /* On CI and later, non-indexed draws overwrite VGT_INDEX_TYPE,
706 * so the state must be re-emitted before the next indexed draw.
708 if (sctx
->chip_class
>= CIK
)
709 sctx
->last_index_size
= -1;
713 uint64_t indirect_va
= r600_resource(indirect
->buffer
)->gpu_address
;
715 assert(indirect_va
% 8 == 0);
717 si_invalidate_draw_sh_constants(sctx
);
719 radeon_emit(cs
, PKT3(PKT3_SET_BASE
, 2, 0));
721 radeon_emit(cs
, indirect_va
);
722 radeon_emit(cs
, indirect_va
>> 32);
724 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
725 r600_resource(indirect
->buffer
),
726 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
728 unsigned di_src_sel
= index_size
? V_0287F0_DI_SRC_SEL_DMA
729 : V_0287F0_DI_SRC_SEL_AUTO_INDEX
;
731 assert(indirect
->offset
% 4 == 0);
734 radeon_emit(cs
, PKT3(PKT3_INDEX_BASE
, 1, 0));
735 radeon_emit(cs
, index_va
);
736 radeon_emit(cs
, index_va
>> 32);
738 radeon_emit(cs
, PKT3(PKT3_INDEX_BUFFER_SIZE
, 0, 0));
739 radeon_emit(cs
, index_max_size
);
742 if (!sctx
->screen
->has_draw_indirect_multi
) {
743 radeon_emit(cs
, PKT3(index_size
? PKT3_DRAW_INDEX_INDIRECT
744 : PKT3_DRAW_INDIRECT
,
745 3, render_cond_bit
));
746 radeon_emit(cs
, indirect
->offset
);
747 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
748 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
749 radeon_emit(cs
, di_src_sel
);
751 uint64_t count_va
= 0;
753 if (indirect
->indirect_draw_count
) {
754 struct r600_resource
*params_buf
=
755 r600_resource(indirect
->indirect_draw_count
);
757 radeon_add_to_buffer_list(
758 sctx
, sctx
->gfx_cs
, params_buf
,
759 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
761 count_va
= params_buf
->gpu_address
+ indirect
->indirect_draw_count_offset
;
764 radeon_emit(cs
, PKT3(index_size
? PKT3_DRAW_INDEX_INDIRECT_MULTI
:
765 PKT3_DRAW_INDIRECT_MULTI
,
766 8, render_cond_bit
));
767 radeon_emit(cs
, indirect
->offset
);
768 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
769 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
770 radeon_emit(cs
, ((sh_base_reg
+ SI_SGPR_DRAWID
* 4 - SI_SH_REG_OFFSET
) >> 2) |
771 S_2C3_DRAW_INDEX_ENABLE(1) |
772 S_2C3_COUNT_INDIRECT_ENABLE(!!indirect
->indirect_draw_count
));
773 radeon_emit(cs
, indirect
->draw_count
);
774 radeon_emit(cs
, count_va
);
775 radeon_emit(cs
, count_va
>> 32);
776 radeon_emit(cs
, indirect
->stride
);
777 radeon_emit(cs
, di_src_sel
);
782 radeon_emit(cs
, PKT3(PKT3_NUM_INSTANCES
, 0, 0));
783 radeon_emit(cs
, info
->instance_count
);
785 /* Base vertex and start instance. */
786 base_vertex
= index_size
? info
->index_bias
: info
->start
;
788 if (sctx
->num_vs_blit_sgprs
) {
789 /* Re-emit draw constants after we leave u_blitter. */
790 si_invalidate_draw_sh_constants(sctx
);
792 /* Blit VS doesn't use BASE_VERTEX, START_INSTANCE, and DRAWID. */
793 radeon_set_sh_reg_seq(cs
, sh_base_reg
+ SI_SGPR_VS_BLIT_DATA
* 4,
794 sctx
->num_vs_blit_sgprs
);
795 radeon_emit_array(cs
, sctx
->vs_blit_sh_data
,
796 sctx
->num_vs_blit_sgprs
);
797 } else if (base_vertex
!= sctx
->last_base_vertex
||
798 sctx
->last_base_vertex
== SI_BASE_VERTEX_UNKNOWN
||
799 info
->start_instance
!= sctx
->last_start_instance
||
800 info
->drawid
!= sctx
->last_drawid
||
801 sh_base_reg
!= sctx
->last_sh_base_reg
) {
802 radeon_set_sh_reg_seq(cs
, sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4, 3);
803 radeon_emit(cs
, base_vertex
);
804 radeon_emit(cs
, info
->start_instance
);
805 radeon_emit(cs
, info
->drawid
);
807 sctx
->last_base_vertex
= base_vertex
;
808 sctx
->last_start_instance
= info
->start_instance
;
809 sctx
->last_drawid
= info
->drawid
;
810 sctx
->last_sh_base_reg
= sh_base_reg
;
814 index_va
+= info
->start
* index_size
;
816 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_2
, 4, render_cond_bit
));
817 radeon_emit(cs
, index_max_size
);
818 radeon_emit(cs
, index_va
);
819 radeon_emit(cs
, index_va
>> 32);
820 radeon_emit(cs
, info
->count
);
821 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_DMA
);
823 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_AUTO
, 1, render_cond_bit
));
824 radeon_emit(cs
, info
->count
);
825 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_AUTO_INDEX
|
826 S_0287F0_USE_OPAQUE(!!info
->count_from_stream_output
));
831 static void si_emit_surface_sync(struct si_context
*sctx
,
832 unsigned cp_coher_cntl
)
834 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
836 if (sctx
->chip_class
>= GFX9
) {
837 /* Flush caches and wait for the caches to assert idle. */
838 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 5, 0));
839 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
840 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
841 radeon_emit(cs
, 0xffffff); /* CP_COHER_SIZE_HI */
842 radeon_emit(cs
, 0); /* CP_COHER_BASE */
843 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
844 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
846 /* ACQUIRE_MEM is only required on a compute ring. */
847 radeon_emit(cs
, PKT3(PKT3_SURFACE_SYNC
, 3, 0));
848 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
849 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
850 radeon_emit(cs
, 0); /* CP_COHER_BASE */
851 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
855 void si_emit_cache_flush(struct si_context
*sctx
)
857 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
858 uint32_t flags
= sctx
->flags
;
859 uint32_t cp_coher_cntl
= 0;
860 uint32_t flush_cb_db
= flags
& (SI_CONTEXT_FLUSH_AND_INV_CB
|
861 SI_CONTEXT_FLUSH_AND_INV_DB
);
863 if (flags
& SI_CONTEXT_FLUSH_AND_INV_CB
)
864 sctx
->num_cb_cache_flushes
++;
865 if (flags
& SI_CONTEXT_FLUSH_AND_INV_DB
)
866 sctx
->num_db_cache_flushes
++;
868 /* SI has a bug that it always flushes ICACHE and KCACHE if either
869 * bit is set. An alternative way is to write SQC_CACHES, but that
870 * doesn't seem to work reliably. Since the bug doesn't affect
871 * correctness (it only does more work than necessary) and
872 * the performance impact is likely negligible, there is no plan
873 * to add a workaround for it.
876 if (flags
& SI_CONTEXT_INV_ICACHE
)
877 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
878 if (flags
& SI_CONTEXT_INV_SMEM_L1
)
879 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
881 if (sctx
->chip_class
<= VI
) {
882 if (flags
& SI_CONTEXT_FLUSH_AND_INV_CB
) {
883 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
884 S_0085F0_CB0_DEST_BASE_ENA(1) |
885 S_0085F0_CB1_DEST_BASE_ENA(1) |
886 S_0085F0_CB2_DEST_BASE_ENA(1) |
887 S_0085F0_CB3_DEST_BASE_ENA(1) |
888 S_0085F0_CB4_DEST_BASE_ENA(1) |
889 S_0085F0_CB5_DEST_BASE_ENA(1) |
890 S_0085F0_CB6_DEST_BASE_ENA(1) |
891 S_0085F0_CB7_DEST_BASE_ENA(1);
893 /* Necessary for DCC */
894 if (sctx
->chip_class
== VI
)
895 si_gfx_write_event_eop(sctx
, V_028A90_FLUSH_AND_INV_CB_DATA_TS
,
896 0, EOP_DATA_SEL_DISCARD
, NULL
,
899 if (flags
& SI_CONTEXT_FLUSH_AND_INV_DB
)
900 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
901 S_0085F0_DB_DEST_BASE_ENA(1);
904 if (flags
& SI_CONTEXT_FLUSH_AND_INV_CB
) {
905 /* Flush CMASK/FMASK/DCC. SURFACE_SYNC will wait for idle. */
906 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
907 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
909 if (flags
& (SI_CONTEXT_FLUSH_AND_INV_DB
|
910 SI_CONTEXT_FLUSH_AND_INV_DB_META
)) {
911 /* Flush HTILE. SURFACE_SYNC will wait for idle. */
912 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
913 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
916 /* Wait for shader engines to go idle.
917 * VS and PS waits are unnecessary if SURFACE_SYNC is going to wait
918 * for everything including CB/DB cache flushes.
921 if (flags
& SI_CONTEXT_PS_PARTIAL_FLUSH
) {
922 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
923 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
924 /* Only count explicit shader flushes, not implicit ones
925 * done by SURFACE_SYNC.
927 sctx
->num_vs_flushes
++;
928 sctx
->num_ps_flushes
++;
929 } else if (flags
& SI_CONTEXT_VS_PARTIAL_FLUSH
) {
930 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
931 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
932 sctx
->num_vs_flushes
++;
936 if (flags
& SI_CONTEXT_CS_PARTIAL_FLUSH
&&
937 sctx
->compute_is_busy
) {
938 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
939 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
| EVENT_INDEX(4)));
940 sctx
->num_cs_flushes
++;
941 sctx
->compute_is_busy
= false;
944 /* VGT state synchronization. */
945 if (flags
& SI_CONTEXT_VGT_FLUSH
) {
946 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
947 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
949 if (flags
& SI_CONTEXT_VGT_STREAMOUT_SYNC
) {
950 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
951 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC
) | EVENT_INDEX(0));
954 /* GFX9: Wait for idle if we're flushing CB or DB. ACQUIRE_MEM doesn't
955 * wait for idle on GFX9. We have to use a TS event.
957 if (sctx
->chip_class
>= GFX9
&& flush_cb_db
) {
959 unsigned tc_flags
, cb_db_event
;
961 /* Set the CB/DB flush event. */
962 switch (flush_cb_db
) {
963 case SI_CONTEXT_FLUSH_AND_INV_CB
:
964 cb_db_event
= V_028A90_FLUSH_AND_INV_CB_DATA_TS
;
966 case SI_CONTEXT_FLUSH_AND_INV_DB
:
967 cb_db_event
= V_028A90_FLUSH_AND_INV_DB_DATA_TS
;
971 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
974 /* These are the only allowed combinations. If you need to
975 * do multiple operations at once, do them separately.
976 * All operations that invalidate L2 also seem to invalidate
977 * metadata. Volatile (VOL) and WC flushes are not listed here.
979 * TC | TC_WB = writeback & invalidate L2 & L1
980 * TC | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
981 * TC_WB | TC_NC = writeback L2 for MTYPE == NC
982 * TC | TC_NC = invalidate L2 for MTYPE == NC
983 * TC | TC_MD = writeback & invalidate L2 metadata (DCC, etc.)
984 * TCL1 = invalidate L1
988 if (flags
& SI_CONTEXT_INV_L2_METADATA
) {
989 tc_flags
= EVENT_TC_ACTION_ENA
|
990 EVENT_TC_MD_ACTION_ENA
;
993 /* Ideally flush TC together with CB/DB. */
994 if (flags
& SI_CONTEXT_INV_GLOBAL_L2
) {
995 /* Writeback and invalidate everything in L2 & L1. */
996 tc_flags
= EVENT_TC_ACTION_ENA
|
997 EVENT_TC_WB_ACTION_ENA
;
999 /* Clear the flags. */
1000 flags
&= ~(SI_CONTEXT_INV_GLOBAL_L2
|
1001 SI_CONTEXT_WRITEBACK_GLOBAL_L2
|
1002 SI_CONTEXT_INV_VMEM_L1
);
1003 sctx
->num_L2_invalidates
++;
1006 /* Do the flush (enqueue the event and wait for it). */
1007 va
= sctx
->wait_mem_scratch
->gpu_address
;
1008 sctx
->wait_mem_number
++;
1010 si_gfx_write_event_eop(sctx
, cb_db_event
, tc_flags
,
1011 EOP_DATA_SEL_VALUE_32BIT
,
1012 sctx
->wait_mem_scratch
, va
,
1013 sctx
->wait_mem_number
, SI_NOT_QUERY
);
1014 si_gfx_wait_fence(sctx
, va
, sctx
->wait_mem_number
, 0xffffffff);
1017 /* Make sure ME is idle (it executes most packets) before continuing.
1018 * This prevents read-after-write hazards between PFP and ME.
1020 if (cp_coher_cntl
||
1021 (flags
& (SI_CONTEXT_CS_PARTIAL_FLUSH
|
1022 SI_CONTEXT_INV_VMEM_L1
|
1023 SI_CONTEXT_INV_GLOBAL_L2
|
1024 SI_CONTEXT_WRITEBACK_GLOBAL_L2
))) {
1025 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
1030 * When one of the CP_COHER_CNTL.DEST_BASE flags is set, SURFACE_SYNC
1031 * waits for idle, so it should be last. SURFACE_SYNC is done in PFP.
1033 * cp_coher_cntl should contain all necessary flags except TC flags
1036 * SI-CIK don't support L2 write-back.
1038 if (flags
& SI_CONTEXT_INV_GLOBAL_L2
||
1039 (sctx
->chip_class
<= CIK
&&
1040 (flags
& SI_CONTEXT_WRITEBACK_GLOBAL_L2
))) {
1041 /* Invalidate L1 & L2. (L1 is always invalidated on SI)
1042 * WB must be set on VI+ when TC_ACTION is set.
1044 si_emit_surface_sync(sctx
, cp_coher_cntl
|
1045 S_0085F0_TC_ACTION_ENA(1) |
1046 S_0085F0_TCL1_ACTION_ENA(1) |
1047 S_0301F0_TC_WB_ACTION_ENA(sctx
->chip_class
>= VI
));
1049 sctx
->num_L2_invalidates
++;
1051 /* L1 invalidation and L2 writeback must be done separately,
1052 * because both operations can't be done together.
1054 if (flags
& SI_CONTEXT_WRITEBACK_GLOBAL_L2
) {
1056 * NC = apply to non-coherent MTYPEs
1057 * (i.e. MTYPE <= 1, which is what we use everywhere)
1059 * WB doesn't work without NC.
1061 si_emit_surface_sync(sctx
, cp_coher_cntl
|
1062 S_0301F0_TC_WB_ACTION_ENA(1) |
1063 S_0301F0_TC_NC_ACTION_ENA(1));
1065 sctx
->num_L2_writebacks
++;
1067 if (flags
& SI_CONTEXT_INV_VMEM_L1
) {
1068 /* Invalidate per-CU VMEM L1. */
1069 si_emit_surface_sync(sctx
, cp_coher_cntl
|
1070 S_0085F0_TCL1_ACTION_ENA(1));
1075 /* If TC flushes haven't cleared this... */
1077 si_emit_surface_sync(sctx
, cp_coher_cntl
);
1079 if (flags
& SI_CONTEXT_START_PIPELINE_STATS
) {
1080 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1081 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
1083 } else if (flags
& SI_CONTEXT_STOP_PIPELINE_STATS
) {
1084 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1085 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
1092 static void si_get_draw_start_count(struct si_context
*sctx
,
1093 const struct pipe_draw_info
*info
,
1094 unsigned *start
, unsigned *count
)
1096 struct pipe_draw_indirect_info
*indirect
= info
->indirect
;
1099 unsigned indirect_count
;
1100 struct pipe_transfer
*transfer
;
1101 unsigned begin
, end
;
1105 if (indirect
->indirect_draw_count
) {
1106 data
= pipe_buffer_map_range(&sctx
->b
,
1107 indirect
->indirect_draw_count
,
1108 indirect
->indirect_draw_count_offset
,
1110 PIPE_TRANSFER_READ
, &transfer
);
1112 indirect_count
= *data
;
1114 pipe_buffer_unmap(&sctx
->b
, transfer
);
1116 indirect_count
= indirect
->draw_count
;
1119 if (!indirect_count
) {
1120 *start
= *count
= 0;
1124 map_size
= (indirect_count
- 1) * indirect
->stride
+ 3 * sizeof(unsigned);
1125 data
= pipe_buffer_map_range(&sctx
->b
, indirect
->buffer
,
1126 indirect
->offset
, map_size
,
1127 PIPE_TRANSFER_READ
, &transfer
);
1132 for (unsigned i
= 0; i
< indirect_count
; ++i
) {
1133 unsigned count
= data
[0];
1134 unsigned start
= data
[2];
1137 begin
= MIN2(begin
, start
);
1138 end
= MAX2(end
, start
+ count
);
1141 data
+= indirect
->stride
/ sizeof(unsigned);
1144 pipe_buffer_unmap(&sctx
->b
, transfer
);
1148 *count
= end
- begin
;
1150 *start
= *count
= 0;
1153 *start
= info
->start
;
1154 *count
= info
->count
;
1158 static void si_emit_all_states(struct si_context
*sctx
, const struct pipe_draw_info
*info
,
1159 unsigned skip_atom_mask
)
1161 unsigned num_patches
= 0;
1162 bool context_roll
= false; /* set correctly for GFX9 only */
1164 context_roll
|= si_emit_rasterizer_prim_state(sctx
);
1165 if (sctx
->tes_shader
.cso
)
1166 context_roll
|= si_emit_derived_tess_state(sctx
, info
, &num_patches
);
1167 if (info
->count_from_stream_output
)
1168 context_roll
= true;
1170 /* Vega10/Raven scissor bug workaround. When any context register is
1171 * written (i.e. the GPU rolls the context), PA_SC_VPORT_SCISSOR
1172 * registers must be written too.
1174 if ((sctx
->family
== CHIP_VEGA10
|| sctx
->family
== CHIP_RAVEN
) &&
1176 sctx
->dirty_atoms
& si_atoms_that_roll_context() ||
1177 sctx
->dirty_states
& si_states_that_roll_context() ||
1178 si_prim_restart_index_changed(sctx
, info
))) {
1179 sctx
->scissors
.dirty_mask
= (1 << SI_MAX_VIEWPORTS
) - 1;
1180 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.scissors
);
1183 /* Emit state atoms. */
1184 unsigned mask
= sctx
->dirty_atoms
& ~skip_atom_mask
;
1186 sctx
->atoms
.array
[u_bit_scan(&mask
)].emit(sctx
);
1188 sctx
->dirty_atoms
&= skip_atom_mask
;
1191 mask
= sctx
->dirty_states
;
1193 unsigned i
= u_bit_scan(&mask
);
1194 struct si_pm4_state
*state
= sctx
->queued
.array
[i
];
1196 if (!state
|| sctx
->emitted
.array
[i
] == state
)
1199 si_pm4_emit(sctx
, state
);
1200 sctx
->emitted
.array
[i
] = state
;
1202 sctx
->dirty_states
= 0;
1204 /* Emit draw states. */
1205 si_emit_vs_state(sctx
, info
);
1206 si_emit_draw_registers(sctx
, info
, num_patches
);
1209 void si_draw_vbo(struct pipe_context
*ctx
, const struct pipe_draw_info
*info
)
1211 struct si_context
*sctx
= (struct si_context
*)ctx
;
1212 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1213 struct pipe_resource
*indexbuf
= info
->index
.resource
;
1214 unsigned dirty_tex_counter
;
1215 enum pipe_prim_type rast_prim
;
1216 unsigned index_size
= info
->index_size
;
1217 unsigned index_offset
= info
->indirect
? info
->start
* index_size
: 0;
1219 if (likely(!info
->indirect
)) {
1220 /* SI-CI treat instance_count==0 as instance_count==1. There is
1221 * no workaround for indirect draws, but we can at least skip
1224 if (unlikely(!info
->instance_count
))
1227 /* Handle count == 0. */
1228 if (unlikely(!info
->count
&&
1229 (index_size
|| !info
->count_from_stream_output
)))
1233 if (unlikely(!sctx
->vs_shader
.cso
||
1235 (!sctx
->ps_shader
.cso
&& !rs
->rasterizer_discard
) ||
1236 (!!sctx
->tes_shader
.cso
!= (info
->mode
== PIPE_PRIM_PATCHES
)))) {
1241 /* Recompute and re-emit the texture resource states if needed. */
1242 dirty_tex_counter
= p_atomic_read(&sctx
->screen
->dirty_tex_counter
);
1243 if (unlikely(dirty_tex_counter
!= sctx
->last_dirty_tex_counter
)) {
1244 sctx
->last_dirty_tex_counter
= dirty_tex_counter
;
1245 sctx
->framebuffer
.dirty_cbufs
|=
1246 ((1 << sctx
->framebuffer
.state
.nr_cbufs
) - 1);
1247 sctx
->framebuffer
.dirty_zsbuf
= true;
1248 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.framebuffer
);
1249 si_update_all_texture_descriptors(sctx
);
1252 si_decompress_textures(sctx
, u_bit_consecutive(0, SI_NUM_GRAPHICS_SHADERS
));
1254 /* Set the rasterization primitive type.
1256 * This must be done after si_decompress_textures, which can call
1257 * draw_vbo recursively, and before si_update_shaders, which uses
1258 * current_rast_prim for this draw_vbo call. */
1259 if (sctx
->gs_shader
.cso
)
1260 rast_prim
= sctx
->gs_shader
.cso
->gs_output_prim
;
1261 else if (sctx
->tes_shader
.cso
) {
1262 if (sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_POINT_MODE
])
1263 rast_prim
= PIPE_PRIM_POINTS
;
1265 rast_prim
= sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
1267 rast_prim
= info
->mode
;
1269 if (rast_prim
!= sctx
->current_rast_prim
) {
1270 if (util_prim_is_points_or_lines(sctx
->current_rast_prim
) !=
1271 util_prim_is_points_or_lines(rast_prim
))
1272 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.guardband
);
1274 sctx
->current_rast_prim
= rast_prim
;
1275 sctx
->do_update_shaders
= true;
1278 if (sctx
->tes_shader
.cso
&&
1279 sctx
->screen
->has_ls_vgpr_init_bug
) {
1280 /* Determine whether the LS VGPR fix should be applied.
1282 * It is only required when num input CPs > num output CPs,
1283 * which cannot happen with the fixed function TCS. We should
1284 * also update this bit when switching from TCS to fixed
1287 struct si_shader_selector
*tcs
= sctx
->tcs_shader
.cso
;
1290 info
->vertices_per_patch
>
1291 tcs
->info
.properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
];
1293 if (ls_vgpr_fix
!= sctx
->ls_vgpr_fix
) {
1294 sctx
->ls_vgpr_fix
= ls_vgpr_fix
;
1295 sctx
->do_update_shaders
= true;
1299 if (sctx
->gs_shader
.cso
) {
1300 /* Determine whether the GS triangle strip adjacency fix should
1301 * be applied. Rotate every other triangle if
1302 * - triangle strips with adjacency are fed to the GS and
1303 * - primitive restart is disabled (the rotation doesn't help
1304 * when the restart occurs after an odd number of triangles).
1306 bool gs_tri_strip_adj_fix
=
1307 !sctx
->tes_shader
.cso
&&
1308 info
->mode
== PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
&&
1309 !info
->primitive_restart
;
1311 if (gs_tri_strip_adj_fix
!= sctx
->gs_tri_strip_adj_fix
) {
1312 sctx
->gs_tri_strip_adj_fix
= gs_tri_strip_adj_fix
;
1313 sctx
->do_update_shaders
= true;
1317 if (sctx
->do_update_shaders
&& !si_update_shaders(sctx
))
1321 /* Translate or upload, if needed. */
1322 /* 8-bit indices are supported on VI. */
1323 if (sctx
->chip_class
<= CIK
&& index_size
== 1) {
1324 unsigned start
, count
, start_offset
, size
, offset
;
1327 si_get_draw_start_count(sctx
, info
, &start
, &count
);
1328 start_offset
= start
* 2;
1332 u_upload_alloc(ctx
->stream_uploader
, start_offset
,
1334 si_optimal_tcc_alignment(sctx
, size
),
1335 &offset
, &indexbuf
, &ptr
);
1339 util_shorten_ubyte_elts_to_userptr(&sctx
->b
, info
, 0, 0,
1340 index_offset
+ start
,
1343 /* info->start will be added by the drawing code */
1344 index_offset
= offset
- start_offset
;
1346 } else if (info
->has_user_indices
) {
1347 unsigned start_offset
;
1349 assert(!info
->indirect
);
1350 start_offset
= info
->start
* index_size
;
1353 u_upload_data(ctx
->stream_uploader
, start_offset
,
1354 info
->count
* index_size
,
1355 sctx
->screen
->info
.tcc_cache_line_size
,
1356 (char*)info
->index
.user
+ start_offset
,
1357 &index_offset
, &indexbuf
);
1361 /* info->start will be added by the drawing code */
1362 index_offset
-= start_offset
;
1363 } else if (sctx
->chip_class
<= CIK
&&
1364 r600_resource(indexbuf
)->TC_L2_dirty
) {
1365 /* VI reads index buffers through TC L2, so it doesn't
1367 sctx
->flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1368 r600_resource(indexbuf
)->TC_L2_dirty
= false;
1372 if (info
->indirect
) {
1373 struct pipe_draw_indirect_info
*indirect
= info
->indirect
;
1375 /* Add the buffer size for memory checking in need_cs_space. */
1376 si_context_add_resource_size(sctx
, indirect
->buffer
);
1378 /* Indirect buffers use TC L2 on GFX9, but not older hw. */
1379 if (sctx
->chip_class
<= VI
) {
1380 if (r600_resource(indirect
->buffer
)->TC_L2_dirty
) {
1381 sctx
->flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1382 r600_resource(indirect
->buffer
)->TC_L2_dirty
= false;
1385 if (indirect
->indirect_draw_count
&&
1386 r600_resource(indirect
->indirect_draw_count
)->TC_L2_dirty
) {
1387 sctx
->flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1388 r600_resource(indirect
->indirect_draw_count
)->TC_L2_dirty
= false;
1393 si_need_gfx_cs_space(sctx
);
1395 /* Since we've called si_context_add_resource_size for vertex buffers,
1396 * this must be called after si_need_cs_space, because we must let
1397 * need_cs_space flush before we add buffers to the buffer list.
1399 if (!si_upload_vertex_buffer_descriptors(sctx
))
1402 /* Use optimal packet order based on whether we need to sync the pipeline. */
1403 if (unlikely(sctx
->flags
& (SI_CONTEXT_FLUSH_AND_INV_CB
|
1404 SI_CONTEXT_FLUSH_AND_INV_DB
|
1405 SI_CONTEXT_PS_PARTIAL_FLUSH
|
1406 SI_CONTEXT_CS_PARTIAL_FLUSH
))) {
1407 /* If we have to wait for idle, set all states first, so that all
1408 * SET packets are processed in parallel with previous draw calls.
1409 * Then draw and prefetch at the end. This ensures that the time
1410 * the CUs are idle is very short.
1412 unsigned masked_atoms
= 0;
1414 if (unlikely(sctx
->flags
& SI_CONTEXT_FLUSH_FOR_RENDER_COND
))
1415 masked_atoms
|= si_get_atom_bit(sctx
, &sctx
->atoms
.s
.render_cond
);
1417 if (!si_upload_graphics_shader_descriptors(sctx
))
1420 /* Emit all states except possibly render condition. */
1421 si_emit_all_states(sctx
, info
, masked_atoms
);
1422 si_emit_cache_flush(sctx
);
1423 /* <-- CUs are idle here. */
1425 if (si_is_atom_dirty(sctx
, &sctx
->atoms
.s
.render_cond
))
1426 sctx
->atoms
.s
.render_cond
.emit(sctx
);
1427 sctx
->dirty_atoms
= 0;
1429 si_emit_draw_packets(sctx
, info
, indexbuf
, index_size
, index_offset
);
1430 /* <-- CUs are busy here. */
1432 /* Start prefetches after the draw has been started. Both will run
1433 * in parallel, but starting the draw first is more important.
1435 if (sctx
->chip_class
>= CIK
&& sctx
->prefetch_L2_mask
)
1436 cik_emit_prefetch_L2(sctx
, false);
1438 /* If we don't wait for idle, start prefetches first, then set
1439 * states, and draw at the end.
1442 si_emit_cache_flush(sctx
);
1444 /* Only prefetch the API VS and VBO descriptors. */
1445 if (sctx
->chip_class
>= CIK
&& sctx
->prefetch_L2_mask
)
1446 cik_emit_prefetch_L2(sctx
, true);
1448 if (!si_upload_graphics_shader_descriptors(sctx
))
1451 si_emit_all_states(sctx
, info
, 0);
1452 si_emit_draw_packets(sctx
, info
, indexbuf
, index_size
, index_offset
);
1454 /* Prefetch the remaining shaders after the draw has been
1456 if (sctx
->chip_class
>= CIK
&& sctx
->prefetch_L2_mask
)
1457 cik_emit_prefetch_L2(sctx
, false);
1460 if (unlikely(sctx
->current_saved_cs
)) {
1461 si_trace_emit(sctx
);
1462 si_log_draw_state(sctx
, sctx
->log
);
1465 /* Workaround for a VGT hang when streamout is enabled.
1466 * It must be done after drawing. */
1467 if ((sctx
->family
== CHIP_HAWAII
||
1468 sctx
->family
== CHIP_TONGA
||
1469 sctx
->family
== CHIP_FIJI
) &&
1470 si_get_strmout_en(sctx
)) {
1471 sctx
->flags
|= SI_CONTEXT_VGT_STREAMOUT_SYNC
;
1474 if (unlikely(sctx
->decompression_enabled
)) {
1475 sctx
->num_decompress_calls
++;
1477 sctx
->num_draw_calls
++;
1478 if (sctx
->framebuffer
.state
.nr_cbufs
> 1)
1479 sctx
->num_mrt_draw_calls
++;
1480 if (info
->primitive_restart
)
1481 sctx
->num_prim_restart_calls
++;
1482 if (G_0286E8_WAVESIZE(sctx
->spi_tmpring_size
))
1483 sctx
->num_spill_draw_calls
++;
1485 if (index_size
&& indexbuf
!= info
->index
.resource
)
1486 pipe_resource_reference(&indexbuf
, NULL
);
1489 void si_draw_rectangle(struct blitter_context
*blitter
,
1490 void *vertex_elements_cso
,
1491 blitter_get_vs_func get_vs
,
1492 int x1
, int y1
, int x2
, int y2
,
1493 float depth
, unsigned num_instances
,
1494 enum blitter_attrib_type type
,
1495 const union blitter_attrib
*attrib
)
1497 struct pipe_context
*pipe
= util_blitter_get_pipe(blitter
);
1498 struct si_context
*sctx
= (struct si_context
*)pipe
;
1500 /* Pack position coordinates as signed int16. */
1501 sctx
->vs_blit_sh_data
[0] = (uint32_t)(x1
& 0xffff) |
1502 ((uint32_t)(y1
& 0xffff) << 16);
1503 sctx
->vs_blit_sh_data
[1] = (uint32_t)(x2
& 0xffff) |
1504 ((uint32_t)(y2
& 0xffff) << 16);
1505 sctx
->vs_blit_sh_data
[2] = fui(depth
);
1508 case UTIL_BLITTER_ATTRIB_COLOR
:
1509 memcpy(&sctx
->vs_blit_sh_data
[3], attrib
->color
,
1512 case UTIL_BLITTER_ATTRIB_TEXCOORD_XY
:
1513 case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW
:
1514 memcpy(&sctx
->vs_blit_sh_data
[3], &attrib
->texcoord
,
1515 sizeof(attrib
->texcoord
));
1517 case UTIL_BLITTER_ATTRIB_NONE
:;
1520 pipe
->bind_vs_state(pipe
, si_get_blit_vs(sctx
, type
, num_instances
));
1522 struct pipe_draw_info info
= {};
1523 info
.mode
= SI_PRIM_RECTANGLE_LIST
;
1525 info
.instance_count
= num_instances
;
1527 /* Don't set per-stage shader pointers for VS. */
1528 sctx
->shader_pointers_dirty
&= ~SI_DESCS_SHADER_MASK(VERTEX
);
1529 sctx
->vertex_buffer_pointer_dirty
= false;
1531 si_draw_vbo(pipe
, &info
);
1534 void si_trace_emit(struct si_context
*sctx
)
1536 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
1537 uint64_t va
= sctx
->current_saved_cs
->trace_buf
->gpu_address
;
1538 uint32_t trace_id
= ++sctx
->current_saved_cs
->trace_id
;
1540 radeon_emit(cs
, PKT3(PKT3_WRITE_DATA
, 3, 0));
1541 radeon_emit(cs
, S_370_DST_SEL(V_370_MEMORY_SYNC
) |
1542 S_370_WR_CONFIRM(1) |
1543 S_370_ENGINE_SEL(V_370_ME
));
1544 radeon_emit(cs
, va
);
1545 radeon_emit(cs
, va
>> 32);
1546 radeon_emit(cs
, trace_id
);
1547 radeon_emit(cs
, PKT3(PKT3_NOP
, 0, 0));
1548 radeon_emit(cs
, AC_ENCODE_TRACE_POINT(trace_id
));
1551 u_log_flush(sctx
->log
);