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 *num_patches
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
) * 4;
151 /* Make sure that the data fits in LDS. This assumes the shaders only
152 * use LDS for the inputs and outputs.
154 * While CIK can use 64K per threadgroup, there is a hang on Stoney
155 * with 2 CUs if we use more than 32K. The closed Vulkan driver also
156 * uses 32K at most on all GCN chips.
158 hardware_lds_size
= 32768;
159 *num_patches
= MIN2(*num_patches
, hardware_lds_size
/ (input_patch_size
+
162 /* Make sure the output data fits in the offchip buffer */
163 *num_patches
= MIN2(*num_patches
,
164 (sctx
->screen
->tess_offchip_block_dw_size
* 4) /
167 /* Not necessary for correctness, but improves performance. The
168 * specific value is taken from the proprietary driver.
170 *num_patches
= MIN2(*num_patches
, 40);
172 if (sctx
->chip_class
== SI
) {
173 /* SI bug workaround, related to power management. Limit LS-HS
174 * threadgroups to only one wave.
176 unsigned one_wave
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
);
177 *num_patches
= MIN2(*num_patches
, one_wave
);
180 /* The VGT HS block increments the patch ID unconditionally
181 * within a single threadgroup. This results in incorrect
182 * patch IDs when instanced draws are used.
184 * The intended solution is to restrict threadgroups to
185 * a single instance by setting SWITCH_ON_EOI, which
186 * should cause IA to split instances up. However, this
187 * doesn't work correctly on SI when there is no other
190 if (has_primid_instancing_bug
&& tess_uses_primid
)
193 sctx
->last_num_patches
= *num_patches
;
195 output_patch0_offset
= input_patch_size
* *num_patches
;
196 perpatch_output_offset
= output_patch0_offset
+ pervertex_output_patch_size
;
198 /* Compute userdata SGPRs. */
199 assert(((input_vertex_size
/ 4) & ~0xff) == 0);
200 assert(((output_vertex_size
/ 4) & ~0xff) == 0);
201 assert(((input_patch_size
/ 4) & ~0x1fff) == 0);
202 assert(((output_patch_size
/ 4) & ~0x1fff) == 0);
203 assert(((output_patch0_offset
/ 16) & ~0xffff) == 0);
204 assert(((perpatch_output_offset
/ 16) & ~0xffff) == 0);
205 assert(num_tcs_input_cp
<= 32);
206 assert(num_tcs_output_cp
<= 32);
208 uint64_t ring_va
= r600_resource(sctx
->tess_rings
)->gpu_address
;
209 assert((ring_va
& u_bit_consecutive(0, 19)) == 0);
211 tcs_in_layout
= S_VS_STATE_LS_OUT_PATCH_SIZE(input_patch_size
/ 4) |
212 S_VS_STATE_LS_OUT_VERTEX_SIZE(input_vertex_size
/ 4);
213 tcs_out_layout
= (output_patch_size
/ 4) |
214 (num_tcs_input_cp
<< 13) |
216 tcs_out_offsets
= (output_patch0_offset
/ 16) |
217 ((perpatch_output_offset
/ 16) << 16);
218 offchip_layout
= *num_patches
|
219 (num_tcs_output_cp
<< 6) |
220 (pervertex_output_patch_size
* *num_patches
<< 12);
222 /* Compute the LDS size. */
223 lds_size
= output_patch0_offset
+ output_patch_size
* *num_patches
;
225 if (sctx
->chip_class
>= CIK
) {
226 assert(lds_size
<= 65536);
227 lds_size
= align(lds_size
, 512) / 512;
229 assert(lds_size
<= 32768);
230 lds_size
= align(lds_size
, 256) / 256;
233 /* Set SI_SGPR_VS_STATE_BITS. */
234 sctx
->current_vs_state
&= C_VS_STATE_LS_OUT_PATCH_SIZE
&
235 C_VS_STATE_LS_OUT_VERTEX_SIZE
;
236 sctx
->current_vs_state
|= tcs_in_layout
;
238 if (sctx
->chip_class
>= GFX9
) {
239 unsigned hs_rsrc2
= ls_current
->config
.rsrc2
|
240 S_00B42C_LDS_SIZE(lds_size
);
242 radeon_set_sh_reg(cs
, R_00B42C_SPI_SHADER_PGM_RSRC2_HS
, hs_rsrc2
);
244 /* Set userdata SGPRs for merged LS-HS. */
245 radeon_set_sh_reg_seq(cs
,
246 R_00B430_SPI_SHADER_USER_DATA_LS_0
+
247 GFX9_SGPR_TCS_OFFCHIP_LAYOUT
* 4, 3);
248 radeon_emit(cs
, offchip_layout
);
249 radeon_emit(cs
, tcs_out_offsets
);
250 radeon_emit(cs
, tcs_out_layout
);
252 unsigned ls_rsrc2
= ls_current
->config
.rsrc2
;
254 si_multiwave_lds_size_workaround(sctx
->screen
, &lds_size
);
255 ls_rsrc2
|= S_00B52C_LDS_SIZE(lds_size
);
257 /* Due to a hw bug, RSRC2_LS must be written twice with another
258 * LS register written in between. */
259 if (sctx
->chip_class
== CIK
&& sctx
->family
!= CHIP_HAWAII
)
260 radeon_set_sh_reg(cs
, R_00B52C_SPI_SHADER_PGM_RSRC2_LS
, ls_rsrc2
);
261 radeon_set_sh_reg_seq(cs
, R_00B528_SPI_SHADER_PGM_RSRC1_LS
, 2);
262 radeon_emit(cs
, ls_current
->config
.rsrc1
);
263 radeon_emit(cs
, ls_rsrc2
);
265 /* Set userdata SGPRs for TCS. */
266 radeon_set_sh_reg_seq(cs
,
267 R_00B430_SPI_SHADER_USER_DATA_HS_0
+ GFX6_SGPR_TCS_OFFCHIP_LAYOUT
* 4, 4);
268 radeon_emit(cs
, offchip_layout
);
269 radeon_emit(cs
, tcs_out_offsets
);
270 radeon_emit(cs
, tcs_out_layout
);
271 radeon_emit(cs
, tcs_in_layout
);
274 /* Set userdata SGPRs for TES. */
275 radeon_set_sh_reg_seq(cs
, tes_sh_base
+ SI_SGPR_TES_OFFCHIP_LAYOUT
* 4, 2);
276 radeon_emit(cs
, offchip_layout
);
277 radeon_emit(cs
, ring_va
);
279 ls_hs_config
= S_028B58_NUM_PATCHES(*num_patches
) |
280 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp
) |
281 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp
);
283 if (sctx
->last_ls_hs_config
!= ls_hs_config
) {
284 if (sctx
->chip_class
>= CIK
) {
285 radeon_set_context_reg_idx(cs
, R_028B58_VGT_LS_HS_CONFIG
, 2,
288 radeon_set_context_reg(cs
, R_028B58_VGT_LS_HS_CONFIG
,
291 sctx
->last_ls_hs_config
= ls_hs_config
;
292 return true; /* true if the context rolls */
297 static unsigned si_num_prims_for_vertices(const struct pipe_draw_info
*info
)
299 switch (info
->mode
) {
300 case PIPE_PRIM_PATCHES
:
301 return info
->count
/ info
->vertices_per_patch
;
302 case SI_PRIM_RECTANGLE_LIST
:
303 return info
->count
/ 3;
305 return u_prims_for_vertices(info
->mode
, info
->count
);
310 si_get_init_multi_vgt_param(struct si_screen
*sscreen
,
311 union si_vgt_param_key
*key
)
313 STATIC_ASSERT(sizeof(union si_vgt_param_key
) == 4);
314 unsigned max_primgroup_in_wave
= 2;
316 /* SWITCH_ON_EOP(0) is always preferable. */
317 bool wd_switch_on_eop
= false;
318 bool ia_switch_on_eop
= false;
319 bool ia_switch_on_eoi
= false;
320 bool partial_vs_wave
= false;
321 bool partial_es_wave
= false;
323 if (key
->u
.uses_tess
) {
324 /* SWITCH_ON_EOI must be set if PrimID is used. */
325 if (key
->u
.tess_uses_prim_id
)
326 ia_switch_on_eoi
= true;
328 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
329 if ((sscreen
->info
.family
== CHIP_TAHITI
||
330 sscreen
->info
.family
== CHIP_PITCAIRN
||
331 sscreen
->info
.family
== CHIP_BONAIRE
) &&
333 partial_vs_wave
= true;
335 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
336 if (sscreen
->has_distributed_tess
) {
337 if (key
->u
.uses_gs
) {
338 if (sscreen
->info
.chip_class
<= VI
)
339 partial_es_wave
= true;
341 /* GPU hang workaround. */
342 if (sscreen
->info
.family
== CHIP_TONGA
||
343 sscreen
->info
.family
== CHIP_FIJI
||
344 sscreen
->info
.family
== CHIP_POLARIS10
||
345 sscreen
->info
.family
== CHIP_POLARIS11
||
346 sscreen
->info
.family
== CHIP_POLARIS12
||
347 sscreen
->info
.family
== CHIP_VEGAM
)
348 partial_vs_wave
= true;
350 partial_vs_wave
= true;
355 /* This is a hardware requirement. */
356 if (key
->u
.line_stipple_enabled
||
357 (sscreen
->debug_flags
& DBG(SWITCH_ON_EOP
))) {
358 ia_switch_on_eop
= true;
359 wd_switch_on_eop
= true;
362 if (sscreen
->info
.chip_class
>= CIK
) {
363 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
364 * 4 shader engines. Set 1 to pass the assertion below.
365 * The other cases are hardware requirements.
367 * Polaris supports primitive restart with WD_SWITCH_ON_EOP=0
368 * for points, line strips, and tri strips.
370 if (sscreen
->info
.max_se
< 4 ||
371 key
->u
.prim
== PIPE_PRIM_POLYGON
||
372 key
->u
.prim
== PIPE_PRIM_LINE_LOOP
||
373 key
->u
.prim
== PIPE_PRIM_TRIANGLE_FAN
||
374 key
->u
.prim
== PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
||
375 (key
->u
.primitive_restart
&&
376 (sscreen
->info
.family
< CHIP_POLARIS10
||
377 (key
->u
.prim
!= PIPE_PRIM_POINTS
&&
378 key
->u
.prim
!= PIPE_PRIM_LINE_STRIP
&&
379 key
->u
.prim
!= PIPE_PRIM_TRIANGLE_STRIP
))) ||
380 key
->u
.count_from_stream_output
)
381 wd_switch_on_eop
= true;
383 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
384 * We don't know that for indirect drawing, so treat it as
385 * always problematic. */
386 if (sscreen
->info
.family
== CHIP_HAWAII
&&
387 key
->u
.uses_instancing
)
388 wd_switch_on_eop
= true;
390 /* Performance recommendation for 4 SE Gfx7-8 parts if
391 * instances are smaller than a primgroup.
392 * Assume indirect draws always use small instances.
393 * This is needed for good VS wave utilization.
395 if (sscreen
->info
.chip_class
<= VI
&&
396 sscreen
->info
.max_se
== 4 &&
397 key
->u
.multi_instances_smaller_than_primgroup
)
398 wd_switch_on_eop
= true;
400 /* Required on CIK and later. */
401 if (sscreen
->info
.max_se
> 2 && !wd_switch_on_eop
)
402 ia_switch_on_eoi
= true;
404 /* Required by Hawaii and, for some special cases, by VI. */
405 if (ia_switch_on_eoi
&&
406 (sscreen
->info
.family
== CHIP_HAWAII
||
407 (sscreen
->info
.chip_class
== VI
&&
408 (key
->u
.uses_gs
|| max_primgroup_in_wave
!= 2))))
409 partial_vs_wave
= true;
411 /* Instancing bug on Bonaire. */
412 if (sscreen
->info
.family
== CHIP_BONAIRE
&& ia_switch_on_eoi
&&
413 key
->u
.uses_instancing
)
414 partial_vs_wave
= true;
416 /* If the WD switch is false, the IA switch must be false too. */
417 assert(wd_switch_on_eop
|| !ia_switch_on_eop
);
420 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
421 if (sscreen
->info
.chip_class
<= VI
&& ia_switch_on_eoi
)
422 partial_es_wave
= true;
424 return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop
) |
425 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi
) |
426 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave
) |
427 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave
) |
428 S_028AA8_WD_SWITCH_ON_EOP(sscreen
->info
.chip_class
>= CIK
? wd_switch_on_eop
: 0) |
429 /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
430 S_028AA8_MAX_PRIMGRP_IN_WAVE(sscreen
->info
.chip_class
== VI
?
431 max_primgroup_in_wave
: 0) |
432 S_030960_EN_INST_OPT_BASIC(sscreen
->info
.chip_class
>= GFX9
) |
433 S_030960_EN_INST_OPT_ADV(sscreen
->info
.chip_class
>= GFX9
);
436 void si_init_ia_multi_vgt_param_table(struct si_context
*sctx
)
438 for (int prim
= 0; prim
<= SI_PRIM_RECTANGLE_LIST
; prim
++)
439 for (int uses_instancing
= 0; uses_instancing
< 2; uses_instancing
++)
440 for (int multi_instances
= 0; multi_instances
< 2; multi_instances
++)
441 for (int primitive_restart
= 0; primitive_restart
< 2; primitive_restart
++)
442 for (int count_from_so
= 0; count_from_so
< 2; count_from_so
++)
443 for (int line_stipple
= 0; line_stipple
< 2; line_stipple
++)
444 for (int uses_tess
= 0; uses_tess
< 2; uses_tess
++)
445 for (int tess_uses_primid
= 0; tess_uses_primid
< 2; tess_uses_primid
++)
446 for (int uses_gs
= 0; uses_gs
< 2; uses_gs
++) {
447 union si_vgt_param_key key
;
451 key
.u
.uses_instancing
= uses_instancing
;
452 key
.u
.multi_instances_smaller_than_primgroup
= multi_instances
;
453 key
.u
.primitive_restart
= primitive_restart
;
454 key
.u
.count_from_stream_output
= count_from_so
;
455 key
.u
.line_stipple_enabled
= line_stipple
;
456 key
.u
.uses_tess
= uses_tess
;
457 key
.u
.tess_uses_prim_id
= tess_uses_primid
;
458 key
.u
.uses_gs
= uses_gs
;
460 sctx
->ia_multi_vgt_param
[key
.index
] =
461 si_get_init_multi_vgt_param(sctx
->screen
, &key
);
465 static unsigned si_get_ia_multi_vgt_param(struct si_context
*sctx
,
466 const struct pipe_draw_info
*info
,
467 unsigned num_patches
)
469 union si_vgt_param_key key
= sctx
->ia_multi_vgt_param_key
;
470 unsigned primgroup_size
;
471 unsigned ia_multi_vgt_param
;
473 if (sctx
->tes_shader
.cso
) {
474 primgroup_size
= num_patches
; /* must be a multiple of NUM_PATCHES */
475 } else if (sctx
->gs_shader
.cso
) {
476 primgroup_size
= 64; /* recommended with a GS */
478 primgroup_size
= 128; /* recommended without a GS and tess */
481 key
.u
.prim
= info
->mode
;
482 key
.u
.uses_instancing
= info
->indirect
|| info
->instance_count
> 1;
483 key
.u
.multi_instances_smaller_than_primgroup
=
485 (info
->instance_count
> 1 &&
486 (info
->count_from_stream_output
||
487 si_num_prims_for_vertices(info
) < primgroup_size
));
488 key
.u
.primitive_restart
= info
->primitive_restart
;
489 key
.u
.count_from_stream_output
= info
->count_from_stream_output
!= NULL
;
491 ia_multi_vgt_param
= sctx
->ia_multi_vgt_param
[key
.index
] |
492 S_028AA8_PRIMGROUP_SIZE(primgroup_size
- 1);
494 if (sctx
->gs_shader
.cso
) {
495 /* GS requirement. */
496 if (sctx
->chip_class
<= VI
&&
497 SI_GS_PER_ES
/ primgroup_size
>= sctx
->screen
->gs_table_depth
- 3)
498 ia_multi_vgt_param
|= S_028AA8_PARTIAL_ES_WAVE_ON(1);
500 /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
501 * The hw doc says all multi-SE chips are affected, but Vulkan
502 * only applies it to Hawaii. Do what Vulkan does.
504 if (sctx
->family
== CHIP_HAWAII
&&
505 G_028AA8_SWITCH_ON_EOI(ia_multi_vgt_param
) &&
507 (info
->instance_count
> 1 &&
508 (info
->count_from_stream_output
||
509 si_num_prims_for_vertices(info
) <= 1))))
510 sctx
->flags
|= SI_CONTEXT_VGT_FLUSH
;
513 return ia_multi_vgt_param
;
516 /* rast_prim is the primitive type after GS. */
517 static bool si_emit_rasterizer_prim_state(struct si_context
*sctx
)
519 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
520 enum pipe_prim_type rast_prim
= sctx
->current_rast_prim
;
521 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
523 /* Skip this if not rendering lines. */
524 if (rast_prim
!= PIPE_PRIM_LINES
&&
525 rast_prim
!= PIPE_PRIM_LINE_LOOP
&&
526 rast_prim
!= PIPE_PRIM_LINE_STRIP
&&
527 rast_prim
!= PIPE_PRIM_LINES_ADJACENCY
&&
528 rast_prim
!= PIPE_PRIM_LINE_STRIP_ADJACENCY
)
531 if (rast_prim
== sctx
->last_rast_prim
&&
532 rs
->pa_sc_line_stipple
== sctx
->last_sc_line_stipple
)
535 /* For lines, reset the stipple pattern at each primitive. Otherwise,
536 * reset the stipple pattern at each packet (line strips, line loops).
538 radeon_set_context_reg(cs
, R_028A0C_PA_SC_LINE_STIPPLE
,
539 rs
->pa_sc_line_stipple
|
540 S_028A0C_AUTO_RESET_CNTL(rast_prim
== PIPE_PRIM_LINES
? 1 : 2));
542 sctx
->last_rast_prim
= rast_prim
;
543 sctx
->last_sc_line_stipple
= rs
->pa_sc_line_stipple
;
544 return true; /* true if the context rolls */
547 static void si_emit_vs_state(struct si_context
*sctx
,
548 const struct pipe_draw_info
*info
)
550 sctx
->current_vs_state
&= C_VS_STATE_INDEXED
;
551 sctx
->current_vs_state
|= S_VS_STATE_INDEXED(!!info
->index_size
);
553 if (sctx
->num_vs_blit_sgprs
) {
554 /* Re-emit the state after we leave u_blitter. */
555 sctx
->last_vs_state
= ~0;
559 if (sctx
->current_vs_state
!= sctx
->last_vs_state
) {
560 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
562 radeon_set_sh_reg(cs
,
563 sctx
->shader_pointers
.sh_base
[PIPE_SHADER_VERTEX
] +
564 SI_SGPR_VS_STATE_BITS
* 4,
565 sctx
->current_vs_state
);
567 sctx
->last_vs_state
= sctx
->current_vs_state
;
571 static inline bool si_prim_restart_index_changed(struct si_context
*sctx
,
572 const struct pipe_draw_info
*info
)
574 return info
->primitive_restart
&&
575 (info
->restart_index
!= sctx
->last_restart_index
||
576 sctx
->last_restart_index
== SI_RESTART_INDEX_UNKNOWN
);
579 static void si_emit_draw_registers(struct si_context
*sctx
,
580 const struct pipe_draw_info
*info
,
581 unsigned num_patches
)
583 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
584 unsigned prim
= si_conv_pipe_prim(info
->mode
);
585 unsigned ia_multi_vgt_param
;
587 ia_multi_vgt_param
= si_get_ia_multi_vgt_param(sctx
, info
, num_patches
);
590 if (ia_multi_vgt_param
!= sctx
->last_multi_vgt_param
) {
591 if (sctx
->chip_class
>= GFX9
)
592 radeon_set_uconfig_reg_idx(cs
, R_030960_IA_MULTI_VGT_PARAM
, 4, ia_multi_vgt_param
);
593 else if (sctx
->chip_class
>= CIK
)
594 radeon_set_context_reg_idx(cs
, R_028AA8_IA_MULTI_VGT_PARAM
, 1, ia_multi_vgt_param
);
596 radeon_set_context_reg(cs
, R_028AA8_IA_MULTI_VGT_PARAM
, ia_multi_vgt_param
);
598 sctx
->last_multi_vgt_param
= ia_multi_vgt_param
;
600 if (prim
!= sctx
->last_prim
) {
601 if (sctx
->chip_class
>= CIK
)
602 radeon_set_uconfig_reg_idx(cs
, R_030908_VGT_PRIMITIVE_TYPE
, 1, prim
);
604 radeon_set_config_reg(cs
, R_008958_VGT_PRIMITIVE_TYPE
, prim
);
606 sctx
->last_prim
= prim
;
609 /* Primitive restart. */
610 if (info
->primitive_restart
!= sctx
->last_primitive_restart_en
) {
611 if (sctx
->chip_class
>= GFX9
)
612 radeon_set_uconfig_reg(cs
, R_03092C_VGT_MULTI_PRIM_IB_RESET_EN
,
613 info
->primitive_restart
);
615 radeon_set_context_reg(cs
, R_028A94_VGT_MULTI_PRIM_IB_RESET_EN
,
616 info
->primitive_restart
);
618 sctx
->last_primitive_restart_en
= info
->primitive_restart
;
621 if (si_prim_restart_index_changed(sctx
, info
)) {
622 radeon_set_context_reg(cs
, R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX
,
623 info
->restart_index
);
624 sctx
->last_restart_index
= info
->restart_index
;
628 static void si_emit_draw_packets(struct si_context
*sctx
,
629 const struct pipe_draw_info
*info
,
630 struct pipe_resource
*indexbuf
,
632 unsigned index_offset
)
634 struct pipe_draw_indirect_info
*indirect
= info
->indirect
;
635 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
636 unsigned sh_base_reg
= sctx
->shader_pointers
.sh_base
[PIPE_SHADER_VERTEX
];
637 bool render_cond_bit
= sctx
->render_cond
&& !sctx
->render_cond_force_off
;
638 uint32_t index_max_size
= 0;
639 uint64_t index_va
= 0;
641 if (info
->count_from_stream_output
) {
642 struct si_streamout_target
*t
=
643 (struct si_streamout_target
*)info
->count_from_stream_output
;
644 uint64_t va
= t
->buf_filled_size
->gpu_address
+
645 t
->buf_filled_size_offset
;
647 radeon_set_context_reg(cs
, R_028B30_VGT_STRMOUT_DRAW_OPAQUE_VERTEX_STRIDE
,
650 radeon_emit(cs
, PKT3(PKT3_COPY_DATA
, 4, 0));
651 radeon_emit(cs
, COPY_DATA_SRC_SEL(COPY_DATA_MEM
) |
652 COPY_DATA_DST_SEL(COPY_DATA_REG
) |
653 COPY_DATA_WR_CONFIRM
);
654 radeon_emit(cs
, va
); /* src address lo */
655 radeon_emit(cs
, va
>> 32); /* src address hi */
656 radeon_emit(cs
, R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE
>> 2);
657 radeon_emit(cs
, 0); /* unused */
659 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
660 t
->buf_filled_size
, RADEON_USAGE_READ
,
661 RADEON_PRIO_SO_FILLED_SIZE
);
666 if (index_size
!= sctx
->last_index_size
) {
670 switch (index_size
) {
672 index_type
= V_028A7C_VGT_INDEX_8
;
675 index_type
= V_028A7C_VGT_INDEX_16
|
676 (SI_BIG_ENDIAN
&& sctx
->chip_class
<= CIK
?
677 V_028A7C_VGT_DMA_SWAP_16_BIT
: 0);
680 index_type
= V_028A7C_VGT_INDEX_32
|
681 (SI_BIG_ENDIAN
&& sctx
->chip_class
<= CIK
?
682 V_028A7C_VGT_DMA_SWAP_32_BIT
: 0);
685 assert(!"unreachable");
689 if (sctx
->chip_class
>= GFX9
) {
690 radeon_set_uconfig_reg_idx(cs
, R_03090C_VGT_INDEX_TYPE
,
693 radeon_emit(cs
, PKT3(PKT3_INDEX_TYPE
, 0, 0));
694 radeon_emit(cs
, index_type
);
697 sctx
->last_index_size
= index_size
;
700 index_max_size
= (indexbuf
->width0
- index_offset
) /
702 index_va
= r600_resource(indexbuf
)->gpu_address
+ index_offset
;
704 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
705 r600_resource(indexbuf
),
706 RADEON_USAGE_READ
, RADEON_PRIO_INDEX_BUFFER
);
708 /* On CI and later, non-indexed draws overwrite VGT_INDEX_TYPE,
709 * so the state must be re-emitted before the next indexed draw.
711 if (sctx
->chip_class
>= CIK
)
712 sctx
->last_index_size
= -1;
716 uint64_t indirect_va
= r600_resource(indirect
->buffer
)->gpu_address
;
718 assert(indirect_va
% 8 == 0);
720 si_invalidate_draw_sh_constants(sctx
);
722 radeon_emit(cs
, PKT3(PKT3_SET_BASE
, 2, 0));
724 radeon_emit(cs
, indirect_va
);
725 radeon_emit(cs
, indirect_va
>> 32);
727 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
728 r600_resource(indirect
->buffer
),
729 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
731 unsigned di_src_sel
= index_size
? V_0287F0_DI_SRC_SEL_DMA
732 : V_0287F0_DI_SRC_SEL_AUTO_INDEX
;
734 assert(indirect
->offset
% 4 == 0);
737 radeon_emit(cs
, PKT3(PKT3_INDEX_BASE
, 1, 0));
738 radeon_emit(cs
, index_va
);
739 radeon_emit(cs
, index_va
>> 32);
741 radeon_emit(cs
, PKT3(PKT3_INDEX_BUFFER_SIZE
, 0, 0));
742 radeon_emit(cs
, index_max_size
);
745 if (!sctx
->screen
->has_draw_indirect_multi
) {
746 radeon_emit(cs
, PKT3(index_size
? PKT3_DRAW_INDEX_INDIRECT
747 : PKT3_DRAW_INDIRECT
,
748 3, render_cond_bit
));
749 radeon_emit(cs
, indirect
->offset
);
750 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
751 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
752 radeon_emit(cs
, di_src_sel
);
754 uint64_t count_va
= 0;
756 if (indirect
->indirect_draw_count
) {
757 struct r600_resource
*params_buf
=
758 r600_resource(indirect
->indirect_draw_count
);
760 radeon_add_to_buffer_list(
761 sctx
, sctx
->gfx_cs
, params_buf
,
762 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
764 count_va
= params_buf
->gpu_address
+ indirect
->indirect_draw_count_offset
;
767 radeon_emit(cs
, PKT3(index_size
? PKT3_DRAW_INDEX_INDIRECT_MULTI
:
768 PKT3_DRAW_INDIRECT_MULTI
,
769 8, render_cond_bit
));
770 radeon_emit(cs
, indirect
->offset
);
771 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
772 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
773 radeon_emit(cs
, ((sh_base_reg
+ SI_SGPR_DRAWID
* 4 - SI_SH_REG_OFFSET
) >> 2) |
774 S_2C3_DRAW_INDEX_ENABLE(1) |
775 S_2C3_COUNT_INDIRECT_ENABLE(!!indirect
->indirect_draw_count
));
776 radeon_emit(cs
, indirect
->draw_count
);
777 radeon_emit(cs
, count_va
);
778 radeon_emit(cs
, count_va
>> 32);
779 radeon_emit(cs
, indirect
->stride
);
780 radeon_emit(cs
, di_src_sel
);
785 radeon_emit(cs
, PKT3(PKT3_NUM_INSTANCES
, 0, 0));
786 radeon_emit(cs
, info
->instance_count
);
788 /* Base vertex and start instance. */
789 base_vertex
= index_size
? info
->index_bias
: info
->start
;
791 if (sctx
->num_vs_blit_sgprs
) {
792 /* Re-emit draw constants after we leave u_blitter. */
793 si_invalidate_draw_sh_constants(sctx
);
795 /* Blit VS doesn't use BASE_VERTEX, START_INSTANCE, and DRAWID. */
796 radeon_set_sh_reg_seq(cs
, sh_base_reg
+ SI_SGPR_VS_BLIT_DATA
* 4,
797 sctx
->num_vs_blit_sgprs
);
798 radeon_emit_array(cs
, sctx
->vs_blit_sh_data
,
799 sctx
->num_vs_blit_sgprs
);
800 } else if (base_vertex
!= sctx
->last_base_vertex
||
801 sctx
->last_base_vertex
== SI_BASE_VERTEX_UNKNOWN
||
802 info
->start_instance
!= sctx
->last_start_instance
||
803 info
->drawid
!= sctx
->last_drawid
||
804 sh_base_reg
!= sctx
->last_sh_base_reg
) {
805 radeon_set_sh_reg_seq(cs
, sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4, 3);
806 radeon_emit(cs
, base_vertex
);
807 radeon_emit(cs
, info
->start_instance
);
808 radeon_emit(cs
, info
->drawid
);
810 sctx
->last_base_vertex
= base_vertex
;
811 sctx
->last_start_instance
= info
->start_instance
;
812 sctx
->last_drawid
= info
->drawid
;
813 sctx
->last_sh_base_reg
= sh_base_reg
;
817 index_va
+= info
->start
* index_size
;
819 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_2
, 4, render_cond_bit
));
820 radeon_emit(cs
, index_max_size
);
821 radeon_emit(cs
, index_va
);
822 radeon_emit(cs
, index_va
>> 32);
823 radeon_emit(cs
, info
->count
);
824 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_DMA
);
826 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_AUTO
, 1, render_cond_bit
));
827 radeon_emit(cs
, info
->count
);
828 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_AUTO_INDEX
|
829 S_0287F0_USE_OPAQUE(!!info
->count_from_stream_output
));
834 static void si_emit_surface_sync(struct si_context
*sctx
,
835 unsigned cp_coher_cntl
)
837 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
839 if (sctx
->chip_class
>= GFX9
) {
840 /* Flush caches and wait for the caches to assert idle. */
841 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 5, 0));
842 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
843 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
844 radeon_emit(cs
, 0xffffff); /* CP_COHER_SIZE_HI */
845 radeon_emit(cs
, 0); /* CP_COHER_BASE */
846 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
847 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
849 /* ACQUIRE_MEM is only required on a compute ring. */
850 radeon_emit(cs
, PKT3(PKT3_SURFACE_SYNC
, 3, 0));
851 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
852 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
853 radeon_emit(cs
, 0); /* CP_COHER_BASE */
854 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
858 void si_emit_cache_flush(struct si_context
*sctx
)
860 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
861 uint32_t flags
= sctx
->flags
;
862 uint32_t cp_coher_cntl
= 0;
863 uint32_t flush_cb_db
= flags
& (SI_CONTEXT_FLUSH_AND_INV_CB
|
864 SI_CONTEXT_FLUSH_AND_INV_DB
);
866 if (flags
& SI_CONTEXT_FLUSH_AND_INV_CB
)
867 sctx
->num_cb_cache_flushes
++;
868 if (flags
& SI_CONTEXT_FLUSH_AND_INV_DB
)
869 sctx
->num_db_cache_flushes
++;
871 /* SI has a bug that it always flushes ICACHE and KCACHE if either
872 * bit is set. An alternative way is to write SQC_CACHES, but that
873 * doesn't seem to work reliably. Since the bug doesn't affect
874 * correctness (it only does more work than necessary) and
875 * the performance impact is likely negligible, there is no plan
876 * to add a workaround for it.
879 if (flags
& SI_CONTEXT_INV_ICACHE
)
880 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
881 if (flags
& SI_CONTEXT_INV_SMEM_L1
)
882 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
884 if (sctx
->chip_class
<= VI
) {
885 if (flags
& SI_CONTEXT_FLUSH_AND_INV_CB
) {
886 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
887 S_0085F0_CB0_DEST_BASE_ENA(1) |
888 S_0085F0_CB1_DEST_BASE_ENA(1) |
889 S_0085F0_CB2_DEST_BASE_ENA(1) |
890 S_0085F0_CB3_DEST_BASE_ENA(1) |
891 S_0085F0_CB4_DEST_BASE_ENA(1) |
892 S_0085F0_CB5_DEST_BASE_ENA(1) |
893 S_0085F0_CB6_DEST_BASE_ENA(1) |
894 S_0085F0_CB7_DEST_BASE_ENA(1);
896 /* Necessary for DCC */
897 if (sctx
->chip_class
== VI
)
898 si_gfx_write_event_eop(sctx
, V_028A90_FLUSH_AND_INV_CB_DATA_TS
,
899 0, EOP_DATA_SEL_DISCARD
, NULL
,
902 if (flags
& SI_CONTEXT_FLUSH_AND_INV_DB
)
903 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
904 S_0085F0_DB_DEST_BASE_ENA(1);
907 if (flags
& SI_CONTEXT_FLUSH_AND_INV_CB
) {
908 /* Flush CMASK/FMASK/DCC. SURFACE_SYNC will wait for idle. */
909 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
910 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
912 if (flags
& (SI_CONTEXT_FLUSH_AND_INV_DB
|
913 SI_CONTEXT_FLUSH_AND_INV_DB_META
)) {
914 /* Flush HTILE. SURFACE_SYNC will wait for idle. */
915 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
916 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
919 /* Wait for shader engines to go idle.
920 * VS and PS waits are unnecessary if SURFACE_SYNC is going to wait
921 * for everything including CB/DB cache flushes.
924 if (flags
& SI_CONTEXT_PS_PARTIAL_FLUSH
) {
925 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
926 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
927 /* Only count explicit shader flushes, not implicit ones
928 * done by SURFACE_SYNC.
930 sctx
->num_vs_flushes
++;
931 sctx
->num_ps_flushes
++;
932 } else if (flags
& SI_CONTEXT_VS_PARTIAL_FLUSH
) {
933 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
934 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
935 sctx
->num_vs_flushes
++;
939 if (flags
& SI_CONTEXT_CS_PARTIAL_FLUSH
&&
940 sctx
->compute_is_busy
) {
941 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
942 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
| EVENT_INDEX(4)));
943 sctx
->num_cs_flushes
++;
944 sctx
->compute_is_busy
= false;
947 /* VGT state synchronization. */
948 if (flags
& SI_CONTEXT_VGT_FLUSH
) {
949 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
950 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
952 if (flags
& SI_CONTEXT_VGT_STREAMOUT_SYNC
) {
953 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
954 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC
) | EVENT_INDEX(0));
957 /* GFX9: Wait for idle if we're flushing CB or DB. ACQUIRE_MEM doesn't
958 * wait for idle on GFX9. We have to use a TS event.
960 if (sctx
->chip_class
>= GFX9
&& flush_cb_db
) {
962 unsigned tc_flags
, cb_db_event
;
964 /* Set the CB/DB flush event. */
965 switch (flush_cb_db
) {
966 case SI_CONTEXT_FLUSH_AND_INV_CB
:
967 cb_db_event
= V_028A90_FLUSH_AND_INV_CB_DATA_TS
;
969 case SI_CONTEXT_FLUSH_AND_INV_DB
:
970 cb_db_event
= V_028A90_FLUSH_AND_INV_DB_DATA_TS
;
974 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
977 /* These are the only allowed combinations. If you need to
978 * do multiple operations at once, do them separately.
979 * All operations that invalidate L2 also seem to invalidate
980 * metadata. Volatile (VOL) and WC flushes are not listed here.
982 * TC | TC_WB = writeback & invalidate L2 & L1
983 * TC | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
984 * TC_WB | TC_NC = writeback L2 for MTYPE == NC
985 * TC | TC_NC = invalidate L2 for MTYPE == NC
986 * TC | TC_MD = writeback & invalidate L2 metadata (DCC, etc.)
987 * TCL1 = invalidate L1
991 if (flags
& SI_CONTEXT_INV_L2_METADATA
) {
992 tc_flags
= EVENT_TC_ACTION_ENA
|
993 EVENT_TC_MD_ACTION_ENA
;
996 /* Ideally flush TC together with CB/DB. */
997 if (flags
& SI_CONTEXT_INV_GLOBAL_L2
) {
998 /* Writeback and invalidate everything in L2 & L1. */
999 tc_flags
= EVENT_TC_ACTION_ENA
|
1000 EVENT_TC_WB_ACTION_ENA
;
1002 /* Clear the flags. */
1003 flags
&= ~(SI_CONTEXT_INV_GLOBAL_L2
|
1004 SI_CONTEXT_WRITEBACK_GLOBAL_L2
|
1005 SI_CONTEXT_INV_VMEM_L1
);
1006 sctx
->num_L2_invalidates
++;
1009 /* Do the flush (enqueue the event and wait for it). */
1010 va
= sctx
->wait_mem_scratch
->gpu_address
;
1011 sctx
->wait_mem_number
++;
1013 si_gfx_write_event_eop(sctx
, cb_db_event
, tc_flags
,
1014 EOP_DATA_SEL_VALUE_32BIT
,
1015 sctx
->wait_mem_scratch
, va
,
1016 sctx
->wait_mem_number
, SI_NOT_QUERY
);
1017 si_gfx_wait_fence(sctx
, va
, sctx
->wait_mem_number
, 0xffffffff);
1020 /* Make sure ME is idle (it executes most packets) before continuing.
1021 * This prevents read-after-write hazards between PFP and ME.
1023 if (cp_coher_cntl
||
1024 (flags
& (SI_CONTEXT_CS_PARTIAL_FLUSH
|
1025 SI_CONTEXT_INV_VMEM_L1
|
1026 SI_CONTEXT_INV_GLOBAL_L2
|
1027 SI_CONTEXT_WRITEBACK_GLOBAL_L2
))) {
1028 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
1033 * When one of the CP_COHER_CNTL.DEST_BASE flags is set, SURFACE_SYNC
1034 * waits for idle, so it should be last. SURFACE_SYNC is done in PFP.
1036 * cp_coher_cntl should contain all necessary flags except TC flags
1039 * SI-CIK don't support L2 write-back.
1041 if (flags
& SI_CONTEXT_INV_GLOBAL_L2
||
1042 (sctx
->chip_class
<= CIK
&&
1043 (flags
& SI_CONTEXT_WRITEBACK_GLOBAL_L2
))) {
1044 /* Invalidate L1 & L2. (L1 is always invalidated on SI)
1045 * WB must be set on VI+ when TC_ACTION is set.
1047 si_emit_surface_sync(sctx
, cp_coher_cntl
|
1048 S_0085F0_TC_ACTION_ENA(1) |
1049 S_0085F0_TCL1_ACTION_ENA(1) |
1050 S_0301F0_TC_WB_ACTION_ENA(sctx
->chip_class
>= VI
));
1052 sctx
->num_L2_invalidates
++;
1054 /* L1 invalidation and L2 writeback must be done separately,
1055 * because both operations can't be done together.
1057 if (flags
& SI_CONTEXT_WRITEBACK_GLOBAL_L2
) {
1059 * NC = apply to non-coherent MTYPEs
1060 * (i.e. MTYPE <= 1, which is what we use everywhere)
1062 * WB doesn't work without NC.
1064 si_emit_surface_sync(sctx
, cp_coher_cntl
|
1065 S_0301F0_TC_WB_ACTION_ENA(1) |
1066 S_0301F0_TC_NC_ACTION_ENA(1));
1068 sctx
->num_L2_writebacks
++;
1070 if (flags
& SI_CONTEXT_INV_VMEM_L1
) {
1071 /* Invalidate per-CU VMEM L1. */
1072 si_emit_surface_sync(sctx
, cp_coher_cntl
|
1073 S_0085F0_TCL1_ACTION_ENA(1));
1078 /* If TC flushes haven't cleared this... */
1080 si_emit_surface_sync(sctx
, cp_coher_cntl
);
1082 if (flags
& SI_CONTEXT_START_PIPELINE_STATS
) {
1083 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1084 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
1086 } else if (flags
& SI_CONTEXT_STOP_PIPELINE_STATS
) {
1087 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1088 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
1095 static void si_get_draw_start_count(struct si_context
*sctx
,
1096 const struct pipe_draw_info
*info
,
1097 unsigned *start
, unsigned *count
)
1099 struct pipe_draw_indirect_info
*indirect
= info
->indirect
;
1102 unsigned indirect_count
;
1103 struct pipe_transfer
*transfer
;
1104 unsigned begin
, end
;
1108 if (indirect
->indirect_draw_count
) {
1109 data
= pipe_buffer_map_range(&sctx
->b
,
1110 indirect
->indirect_draw_count
,
1111 indirect
->indirect_draw_count_offset
,
1113 PIPE_TRANSFER_READ
, &transfer
);
1115 indirect_count
= *data
;
1117 pipe_buffer_unmap(&sctx
->b
, transfer
);
1119 indirect_count
= indirect
->draw_count
;
1122 if (!indirect_count
) {
1123 *start
= *count
= 0;
1127 map_size
= (indirect_count
- 1) * indirect
->stride
+ 3 * sizeof(unsigned);
1128 data
= pipe_buffer_map_range(&sctx
->b
, indirect
->buffer
,
1129 indirect
->offset
, map_size
,
1130 PIPE_TRANSFER_READ
, &transfer
);
1135 for (unsigned i
= 0; i
< indirect_count
; ++i
) {
1136 unsigned count
= data
[0];
1137 unsigned start
= data
[2];
1140 begin
= MIN2(begin
, start
);
1141 end
= MAX2(end
, start
+ count
);
1144 data
+= indirect
->stride
/ sizeof(unsigned);
1147 pipe_buffer_unmap(&sctx
->b
, transfer
);
1151 *count
= end
- begin
;
1153 *start
= *count
= 0;
1156 *start
= info
->start
;
1157 *count
= info
->count
;
1161 static void si_emit_all_states(struct si_context
*sctx
, const struct pipe_draw_info
*info
,
1162 unsigned skip_atom_mask
)
1164 unsigned num_patches
= 0;
1165 bool context_roll
= false; /* set correctly for GFX9 only */
1167 context_roll
|= si_emit_rasterizer_prim_state(sctx
);
1168 if (sctx
->tes_shader
.cso
)
1169 context_roll
|= si_emit_derived_tess_state(sctx
, info
, &num_patches
);
1170 if (info
->count_from_stream_output
)
1171 context_roll
= true;
1173 /* Vega10/Raven scissor bug workaround. When any context register is
1174 * written (i.e. the GPU rolls the context), PA_SC_VPORT_SCISSOR
1175 * registers must be written too.
1177 if ((sctx
->family
== CHIP_VEGA10
|| sctx
->family
== CHIP_RAVEN
) &&
1179 sctx
->dirty_atoms
& si_atoms_that_roll_context() ||
1180 sctx
->dirty_states
& si_states_that_roll_context() ||
1181 si_prim_restart_index_changed(sctx
, info
))) {
1182 sctx
->scissors
.dirty_mask
= (1 << SI_MAX_VIEWPORTS
) - 1;
1183 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.scissors
);
1186 /* Emit state atoms. */
1187 unsigned mask
= sctx
->dirty_atoms
& ~skip_atom_mask
;
1189 sctx
->atoms
.array
[u_bit_scan(&mask
)].emit(sctx
);
1191 sctx
->dirty_atoms
&= skip_atom_mask
;
1194 mask
= sctx
->dirty_states
;
1196 unsigned i
= u_bit_scan(&mask
);
1197 struct si_pm4_state
*state
= sctx
->queued
.array
[i
];
1199 if (!state
|| sctx
->emitted
.array
[i
] == state
)
1202 si_pm4_emit(sctx
, state
);
1203 sctx
->emitted
.array
[i
] = state
;
1205 sctx
->dirty_states
= 0;
1207 /* Emit draw states. */
1208 si_emit_vs_state(sctx
, info
);
1209 si_emit_draw_registers(sctx
, info
, num_patches
);
1212 void si_draw_vbo(struct pipe_context
*ctx
, const struct pipe_draw_info
*info
)
1214 struct si_context
*sctx
= (struct si_context
*)ctx
;
1215 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1216 struct pipe_resource
*indexbuf
= info
->index
.resource
;
1217 unsigned dirty_tex_counter
;
1218 enum pipe_prim_type rast_prim
;
1219 unsigned index_size
= info
->index_size
;
1220 unsigned index_offset
= info
->indirect
? info
->start
* index_size
: 0;
1222 if (likely(!info
->indirect
)) {
1223 /* SI-CI treat instance_count==0 as instance_count==1. There is
1224 * no workaround for indirect draws, but we can at least skip
1227 if (unlikely(!info
->instance_count
))
1230 /* Handle count == 0. */
1231 if (unlikely(!info
->count
&&
1232 (index_size
|| !info
->count_from_stream_output
)))
1236 if (unlikely(!sctx
->vs_shader
.cso
)) {
1240 if (unlikely(!sctx
->ps_shader
.cso
&& (!rs
|| !rs
->rasterizer_discard
))) {
1244 if (unlikely(!!sctx
->tes_shader
.cso
!= (info
->mode
== PIPE_PRIM_PATCHES
))) {
1249 /* Recompute and re-emit the texture resource states if needed. */
1250 dirty_tex_counter
= p_atomic_read(&sctx
->screen
->dirty_tex_counter
);
1251 if (unlikely(dirty_tex_counter
!= sctx
->last_dirty_tex_counter
)) {
1252 sctx
->last_dirty_tex_counter
= dirty_tex_counter
;
1253 sctx
->framebuffer
.dirty_cbufs
|=
1254 ((1 << sctx
->framebuffer
.state
.nr_cbufs
) - 1);
1255 sctx
->framebuffer
.dirty_zsbuf
= true;
1256 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.framebuffer
);
1257 si_update_all_texture_descriptors(sctx
);
1260 si_decompress_textures(sctx
, u_bit_consecutive(0, SI_NUM_GRAPHICS_SHADERS
));
1262 /* Set the rasterization primitive type.
1264 * This must be done after si_decompress_textures, which can call
1265 * draw_vbo recursively, and before si_update_shaders, which uses
1266 * current_rast_prim for this draw_vbo call. */
1267 if (sctx
->gs_shader
.cso
)
1268 rast_prim
= sctx
->gs_shader
.cso
->gs_output_prim
;
1269 else if (sctx
->tes_shader
.cso
) {
1270 if (sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_POINT_MODE
])
1271 rast_prim
= PIPE_PRIM_POINTS
;
1273 rast_prim
= sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
1275 rast_prim
= info
->mode
;
1277 if (rast_prim
!= sctx
->current_rast_prim
) {
1278 bool old_is_poly
= sctx
->current_rast_prim
>= PIPE_PRIM_TRIANGLES
;
1279 bool new_is_poly
= rast_prim
>= PIPE_PRIM_TRIANGLES
;
1280 if (old_is_poly
!= new_is_poly
) {
1281 sctx
->scissors
.dirty_mask
= (1 << SI_MAX_VIEWPORTS
) - 1;
1282 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.scissors
);
1285 sctx
->current_rast_prim
= rast_prim
;
1286 sctx
->do_update_shaders
= true;
1289 if (sctx
->tes_shader
.cso
&&
1290 sctx
->screen
->has_ls_vgpr_init_bug
) {
1291 /* Determine whether the LS VGPR fix should be applied.
1293 * It is only required when num input CPs > num output CPs,
1294 * which cannot happen with the fixed function TCS. We should
1295 * also update this bit when switching from TCS to fixed
1298 struct si_shader_selector
*tcs
= sctx
->tcs_shader
.cso
;
1301 info
->vertices_per_patch
>
1302 tcs
->info
.properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
];
1304 if (ls_vgpr_fix
!= sctx
->ls_vgpr_fix
) {
1305 sctx
->ls_vgpr_fix
= ls_vgpr_fix
;
1306 sctx
->do_update_shaders
= true;
1310 if (sctx
->gs_shader
.cso
) {
1311 /* Determine whether the GS triangle strip adjacency fix should
1312 * be applied. Rotate every other triangle if
1313 * - triangle strips with adjacency are fed to the GS and
1314 * - primitive restart is disabled (the rotation doesn't help
1315 * when the restart occurs after an odd number of triangles).
1317 bool gs_tri_strip_adj_fix
=
1318 !sctx
->tes_shader
.cso
&&
1319 info
->mode
== PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
&&
1320 !info
->primitive_restart
;
1322 if (gs_tri_strip_adj_fix
!= sctx
->gs_tri_strip_adj_fix
) {
1323 sctx
->gs_tri_strip_adj_fix
= gs_tri_strip_adj_fix
;
1324 sctx
->do_update_shaders
= true;
1328 if (sctx
->do_update_shaders
&& !si_update_shaders(sctx
))
1332 /* Translate or upload, if needed. */
1333 /* 8-bit indices are supported on VI. */
1334 if (sctx
->chip_class
<= CIK
&& index_size
== 1) {
1335 unsigned start
, count
, start_offset
, size
, offset
;
1338 si_get_draw_start_count(sctx
, info
, &start
, &count
);
1339 start_offset
= start
* 2;
1343 u_upload_alloc(ctx
->stream_uploader
, start_offset
,
1345 si_optimal_tcc_alignment(sctx
, size
),
1346 &offset
, &indexbuf
, &ptr
);
1350 util_shorten_ubyte_elts_to_userptr(&sctx
->b
, info
, 0, 0,
1351 index_offset
+ start
,
1354 /* info->start will be added by the drawing code */
1355 index_offset
= offset
- start_offset
;
1357 } else if (info
->has_user_indices
) {
1358 unsigned start_offset
;
1360 assert(!info
->indirect
);
1361 start_offset
= info
->start
* index_size
;
1364 u_upload_data(ctx
->stream_uploader
, start_offset
,
1365 info
->count
* index_size
,
1366 sctx
->screen
->info
.tcc_cache_line_size
,
1367 (char*)info
->index
.user
+ start_offset
,
1368 &index_offset
, &indexbuf
);
1372 /* info->start will be added by the drawing code */
1373 index_offset
-= start_offset
;
1374 } else if (sctx
->chip_class
<= CIK
&&
1375 r600_resource(indexbuf
)->TC_L2_dirty
) {
1376 /* VI reads index buffers through TC L2, so it doesn't
1378 sctx
->flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1379 r600_resource(indexbuf
)->TC_L2_dirty
= false;
1383 if (info
->indirect
) {
1384 struct pipe_draw_indirect_info
*indirect
= info
->indirect
;
1386 /* Add the buffer size for memory checking in need_cs_space. */
1387 si_context_add_resource_size(sctx
, indirect
->buffer
);
1389 /* Indirect buffers use TC L2 on GFX9, but not older hw. */
1390 if (sctx
->chip_class
<= VI
) {
1391 if (r600_resource(indirect
->buffer
)->TC_L2_dirty
) {
1392 sctx
->flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1393 r600_resource(indirect
->buffer
)->TC_L2_dirty
= false;
1396 if (indirect
->indirect_draw_count
&&
1397 r600_resource(indirect
->indirect_draw_count
)->TC_L2_dirty
) {
1398 sctx
->flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1399 r600_resource(indirect
->indirect_draw_count
)->TC_L2_dirty
= false;
1404 si_need_gfx_cs_space(sctx
);
1406 /* Since we've called si_context_add_resource_size for vertex buffers,
1407 * this must be called after si_need_cs_space, because we must let
1408 * need_cs_space flush before we add buffers to the buffer list.
1410 if (!si_upload_vertex_buffer_descriptors(sctx
))
1413 /* Use optimal packet order based on whether we need to sync the pipeline. */
1414 if (unlikely(sctx
->flags
& (SI_CONTEXT_FLUSH_AND_INV_CB
|
1415 SI_CONTEXT_FLUSH_AND_INV_DB
|
1416 SI_CONTEXT_PS_PARTIAL_FLUSH
|
1417 SI_CONTEXT_CS_PARTIAL_FLUSH
))) {
1418 /* If we have to wait for idle, set all states first, so that all
1419 * SET packets are processed in parallel with previous draw calls.
1420 * Then draw and prefetch at the end. This ensures that the time
1421 * the CUs are idle is very short.
1423 unsigned masked_atoms
= 0;
1425 if (unlikely(sctx
->flags
& SI_CONTEXT_FLUSH_FOR_RENDER_COND
))
1426 masked_atoms
|= si_get_atom_bit(sctx
, &sctx
->atoms
.s
.render_cond
);
1428 if (!si_upload_graphics_shader_descriptors(sctx
))
1431 /* Emit all states except possibly render condition. */
1432 si_emit_all_states(sctx
, info
, masked_atoms
);
1433 si_emit_cache_flush(sctx
);
1434 /* <-- CUs are idle here. */
1436 if (si_is_atom_dirty(sctx
, &sctx
->atoms
.s
.render_cond
))
1437 sctx
->atoms
.s
.render_cond
.emit(sctx
);
1438 sctx
->dirty_atoms
= 0;
1440 si_emit_draw_packets(sctx
, info
, indexbuf
, index_size
, index_offset
);
1441 /* <-- CUs are busy here. */
1443 /* Start prefetches after the draw has been started. Both will run
1444 * in parallel, but starting the draw first is more important.
1446 if (sctx
->chip_class
>= CIK
&& sctx
->prefetch_L2_mask
)
1447 cik_emit_prefetch_L2(sctx
, false);
1449 /* If we don't wait for idle, start prefetches first, then set
1450 * states, and draw at the end.
1453 si_emit_cache_flush(sctx
);
1455 /* Only prefetch the API VS and VBO descriptors. */
1456 if (sctx
->chip_class
>= CIK
&& sctx
->prefetch_L2_mask
)
1457 cik_emit_prefetch_L2(sctx
, true);
1459 if (!si_upload_graphics_shader_descriptors(sctx
))
1462 si_emit_all_states(sctx
, info
, 0);
1463 si_emit_draw_packets(sctx
, info
, indexbuf
, index_size
, index_offset
);
1465 /* Prefetch the remaining shaders after the draw has been
1467 if (sctx
->chip_class
>= CIK
&& sctx
->prefetch_L2_mask
)
1468 cik_emit_prefetch_L2(sctx
, false);
1471 if (unlikely(sctx
->current_saved_cs
)) {
1472 si_trace_emit(sctx
);
1473 si_log_draw_state(sctx
, sctx
->log
);
1476 /* Workaround for a VGT hang when streamout is enabled.
1477 * It must be done after drawing. */
1478 if ((sctx
->family
== CHIP_HAWAII
||
1479 sctx
->family
== CHIP_TONGA
||
1480 sctx
->family
== CHIP_FIJI
) &&
1481 si_get_strmout_en(sctx
)) {
1482 sctx
->flags
|= SI_CONTEXT_VGT_STREAMOUT_SYNC
;
1485 if (unlikely(sctx
->decompression_enabled
)) {
1486 sctx
->num_decompress_calls
++;
1488 sctx
->num_draw_calls
++;
1489 if (sctx
->framebuffer
.state
.nr_cbufs
> 1)
1490 sctx
->num_mrt_draw_calls
++;
1491 if (info
->primitive_restart
)
1492 sctx
->num_prim_restart_calls
++;
1493 if (G_0286E8_WAVESIZE(sctx
->spi_tmpring_size
))
1494 sctx
->num_spill_draw_calls
++;
1496 if (index_size
&& indexbuf
!= info
->index
.resource
)
1497 pipe_resource_reference(&indexbuf
, NULL
);
1500 void si_draw_rectangle(struct blitter_context
*blitter
,
1501 void *vertex_elements_cso
,
1502 blitter_get_vs_func get_vs
,
1503 int x1
, int y1
, int x2
, int y2
,
1504 float depth
, unsigned num_instances
,
1505 enum blitter_attrib_type type
,
1506 const union blitter_attrib
*attrib
)
1508 struct pipe_context
*pipe
= util_blitter_get_pipe(blitter
);
1509 struct si_context
*sctx
= (struct si_context
*)pipe
;
1511 /* Pack position coordinates as signed int16. */
1512 sctx
->vs_blit_sh_data
[0] = (uint32_t)(x1
& 0xffff) |
1513 ((uint32_t)(y1
& 0xffff) << 16);
1514 sctx
->vs_blit_sh_data
[1] = (uint32_t)(x2
& 0xffff) |
1515 ((uint32_t)(y2
& 0xffff) << 16);
1516 sctx
->vs_blit_sh_data
[2] = fui(depth
);
1519 case UTIL_BLITTER_ATTRIB_COLOR
:
1520 memcpy(&sctx
->vs_blit_sh_data
[3], attrib
->color
,
1523 case UTIL_BLITTER_ATTRIB_TEXCOORD_XY
:
1524 case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW
:
1525 memcpy(&sctx
->vs_blit_sh_data
[3], &attrib
->texcoord
,
1526 sizeof(attrib
->texcoord
));
1528 case UTIL_BLITTER_ATTRIB_NONE
:;
1531 pipe
->bind_vs_state(pipe
, si_get_blit_vs(sctx
, type
, num_instances
));
1533 struct pipe_draw_info info
= {};
1534 info
.mode
= SI_PRIM_RECTANGLE_LIST
;
1536 info
.instance_count
= num_instances
;
1538 /* Don't set per-stage shader pointers for VS. */
1539 sctx
->shader_pointers_dirty
&= ~SI_DESCS_SHADER_MASK(VERTEX
);
1540 sctx
->vertex_buffer_pointer_dirty
= false;
1542 si_draw_vbo(pipe
, &info
);
1545 void si_trace_emit(struct si_context
*sctx
)
1547 struct radeon_winsys_cs
*cs
= sctx
->gfx_cs
;
1548 uint64_t va
= sctx
->current_saved_cs
->trace_buf
->gpu_address
;
1549 uint32_t trace_id
= ++sctx
->current_saved_cs
->trace_id
;
1551 radeon_emit(cs
, PKT3(PKT3_WRITE_DATA
, 3, 0));
1552 radeon_emit(cs
, S_370_DST_SEL(V_370_MEMORY_SYNC
) |
1553 S_370_WR_CONFIRM(1) |
1554 S_370_ENGINE_SEL(V_370_ME
));
1555 radeon_emit(cs
, va
);
1556 radeon_emit(cs
, va
>> 32);
1557 radeon_emit(cs
, trace_id
);
1558 radeon_emit(cs
, PKT3(PKT3_NOP
, 0, 0));
1559 radeon_emit(cs
, AC_ENCODE_TRACE_POINT(trace_id
));
1562 u_log_flush(sctx
->log
);