2 * Copyright 2012 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 * Christian König <christian.koenig@amd.com>
28 #include "radeon/r600_cs.h"
32 #include "util/u_index_modify.h"
33 #include "util/u_upload_mgr.h"
34 #include "util/u_prim.h"
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 [R600_PRIM_RECTANGLE_LIST
] = V_008958_DI_PT_RECTLIST
58 assert(mode
< ARRAY_SIZE(prim_conv
));
59 return prim_conv
[mode
];
62 static unsigned si_conv_prim_to_gs_out(unsigned mode
)
64 static const int prim_conv
[] = {
65 [PIPE_PRIM_POINTS
] = V_028A6C_OUTPRIM_TYPE_POINTLIST
,
66 [PIPE_PRIM_LINES
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
67 [PIPE_PRIM_LINE_LOOP
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
68 [PIPE_PRIM_LINE_STRIP
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
69 [PIPE_PRIM_TRIANGLES
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
70 [PIPE_PRIM_TRIANGLE_STRIP
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
71 [PIPE_PRIM_TRIANGLE_FAN
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
72 [PIPE_PRIM_QUADS
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
73 [PIPE_PRIM_QUAD_STRIP
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
74 [PIPE_PRIM_POLYGON
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
75 [PIPE_PRIM_LINES_ADJACENCY
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
76 [PIPE_PRIM_LINE_STRIP_ADJACENCY
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
77 [PIPE_PRIM_TRIANGLES_ADJACENCY
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
78 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
79 [PIPE_PRIM_PATCHES
] = V_028A6C_OUTPRIM_TYPE_POINTLIST
,
80 [R600_PRIM_RECTANGLE_LIST
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
82 assert(mode
< ARRAY_SIZE(prim_conv
));
84 return prim_conv
[mode
];
88 * This calculates the LDS size for tessellation shaders (VS, TCS, TES).
89 * LS.LDS_SIZE is shared by all 3 shader stages.
91 * The information about LDS and other non-compile-time parameters is then
92 * written to userdata SGPRs.
94 static void si_emit_derived_tess_state(struct si_context
*sctx
,
95 const struct pipe_draw_info
*info
,
96 unsigned *num_patches
)
98 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
99 struct si_shader_ctx_state
*ls
= &sctx
->vs_shader
;
100 /* The TES pointer will only be used for sctx->last_tcs.
101 * It would be wrong to think that TCS = TES. */
102 struct si_shader_selector
*tcs
=
103 sctx
->tcs_shader
.cso
? sctx
->tcs_shader
.cso
: sctx
->tes_shader
.cso
;
104 unsigned tes_sh_base
= sctx
->shader_userdata
.sh_base
[PIPE_SHADER_TESS_EVAL
];
105 unsigned num_tcs_input_cp
= info
->vertices_per_patch
;
106 unsigned num_tcs_output_cp
, num_tcs_inputs
, num_tcs_outputs
;
107 unsigned num_tcs_patch_outputs
;
108 unsigned input_vertex_size
, output_vertex_size
, pervertex_output_patch_size
;
109 unsigned input_patch_size
, output_patch_size
, output_patch0_offset
;
110 unsigned perpatch_output_offset
, lds_size
, ls_rsrc2
;
111 unsigned tcs_in_layout
, tcs_out_layout
, tcs_out_offsets
;
112 unsigned offchip_layout
, hardware_lds_size
, ls_hs_config
;
114 if (sctx
->last_ls
== ls
->current
&&
115 sctx
->last_tcs
== tcs
&&
116 sctx
->last_tes_sh_base
== tes_sh_base
&&
117 sctx
->last_num_tcs_input_cp
== num_tcs_input_cp
) {
118 *num_patches
= sctx
->last_num_patches
;
122 sctx
->last_ls
= ls
->current
;
123 sctx
->last_tcs
= tcs
;
124 sctx
->last_tes_sh_base
= tes_sh_base
;
125 sctx
->last_num_tcs_input_cp
= num_tcs_input_cp
;
127 /* This calculates how shader inputs and outputs among VS, TCS, and TES
128 * are laid out in LDS. */
129 num_tcs_inputs
= util_last_bit64(ls
->cso
->outputs_written
);
131 if (sctx
->tcs_shader
.cso
) {
132 num_tcs_outputs
= util_last_bit64(tcs
->outputs_written
);
133 num_tcs_output_cp
= tcs
->info
.properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
];
134 num_tcs_patch_outputs
= util_last_bit64(tcs
->patch_outputs_written
);
136 /* No TCS. Route varyings from LS to TES. */
137 num_tcs_outputs
= num_tcs_inputs
;
138 num_tcs_output_cp
= num_tcs_input_cp
;
139 num_tcs_patch_outputs
= 2; /* TESSINNER + TESSOUTER */
142 input_vertex_size
= num_tcs_inputs
* 16;
143 output_vertex_size
= num_tcs_outputs
* 16;
145 input_patch_size
= num_tcs_input_cp
* input_vertex_size
;
147 pervertex_output_patch_size
= num_tcs_output_cp
* output_vertex_size
;
148 output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
150 /* Ensure that we only need one wave per SIMD so we don't need to check
151 * resource usage. Also ensures that the number of tcs in and out
152 * vertices per threadgroup are at most 256.
154 *num_patches
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
) * 4;
156 /* Make sure that the data fits in LDS. This assumes the shaders only
157 * use LDS for the inputs and outputs.
159 hardware_lds_size
= sctx
->b
.chip_class
>= CIK
? 65536 : 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 /* SI bug workaround - limit LS-HS threadgroups to only one wave. */
174 if (sctx
->b
.chip_class
== SI
) {
175 unsigned one_wave
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
);
176 *num_patches
= MIN2(*num_patches
, one_wave
);
179 sctx
->last_num_patches
= *num_patches
;
181 output_patch0_offset
= input_patch_size
* *num_patches
;
182 perpatch_output_offset
= output_patch0_offset
+ pervertex_output_patch_size
;
184 lds_size
= output_patch0_offset
+ output_patch_size
* *num_patches
;
185 ls_rsrc2
= ls
->current
->config
.rsrc2
;
187 if (sctx
->b
.chip_class
>= CIK
) {
188 assert(lds_size
<= 65536);
189 lds_size
= align(lds_size
, 512) / 512;
191 assert(lds_size
<= 32768);
192 lds_size
= align(lds_size
, 256) / 256;
194 si_multiwave_lds_size_workaround(sctx
->screen
, &lds_size
);
195 ls_rsrc2
|= S_00B52C_LDS_SIZE(lds_size
);
197 /* Due to a hw bug, RSRC2_LS must be written twice with another
198 * LS register written in between. */
199 if (sctx
->b
.chip_class
== CIK
&& sctx
->b
.family
!= CHIP_HAWAII
)
200 radeon_set_sh_reg(cs
, R_00B52C_SPI_SHADER_PGM_RSRC2_LS
, ls_rsrc2
);
201 radeon_set_sh_reg_seq(cs
, R_00B528_SPI_SHADER_PGM_RSRC1_LS
, 2);
202 radeon_emit(cs
, ls
->current
->config
.rsrc1
);
203 radeon_emit(cs
, ls_rsrc2
);
205 /* Compute userdata SGPRs. */
206 assert(((input_vertex_size
/ 4) & ~0xff) == 0);
207 assert(((output_vertex_size
/ 4) & ~0xff) == 0);
208 assert(((input_patch_size
/ 4) & ~0x1fff) == 0);
209 assert(((output_patch_size
/ 4) & ~0x1fff) == 0);
210 assert(((output_patch0_offset
/ 16) & ~0xffff) == 0);
211 assert(((perpatch_output_offset
/ 16) & ~0xffff) == 0);
212 assert(num_tcs_input_cp
<= 32);
213 assert(num_tcs_output_cp
<= 32);
215 tcs_in_layout
= (input_patch_size
/ 4) |
216 ((input_vertex_size
/ 4) << 13);
217 tcs_out_layout
= (output_patch_size
/ 4) |
218 ((output_vertex_size
/ 4) << 13);
219 tcs_out_offsets
= (output_patch0_offset
/ 16) |
220 ((perpatch_output_offset
/ 16) << 16);
221 offchip_layout
= (pervertex_output_patch_size
* *num_patches
<< 16) |
222 (num_tcs_output_cp
<< 9) | *num_patches
;
224 /* Set them for LS. */
225 radeon_set_sh_reg(cs
,
226 R_00B530_SPI_SHADER_USER_DATA_LS_0
+ SI_SGPR_LS_OUT_LAYOUT
* 4,
229 /* Set them for TCS. */
230 radeon_set_sh_reg_seq(cs
,
231 R_00B430_SPI_SHADER_USER_DATA_HS_0
+ SI_SGPR_TCS_OFFCHIP_LAYOUT
* 4, 4);
232 radeon_emit(cs
, offchip_layout
);
233 radeon_emit(cs
, tcs_out_offsets
);
234 radeon_emit(cs
, tcs_out_layout
| (num_tcs_input_cp
<< 26));
235 radeon_emit(cs
, tcs_in_layout
);
237 /* Set them for TES. */
238 radeon_set_sh_reg_seq(cs
, tes_sh_base
+ SI_SGPR_TCS_OFFCHIP_LAYOUT
* 4, 1);
239 radeon_emit(cs
, offchip_layout
);
241 ls_hs_config
= S_028B58_NUM_PATCHES(*num_patches
) |
242 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp
) |
243 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp
);
245 if (sctx
->b
.chip_class
>= CIK
)
246 radeon_set_context_reg_idx(cs
, R_028B58_VGT_LS_HS_CONFIG
, 2,
249 radeon_set_context_reg(cs
, R_028B58_VGT_LS_HS_CONFIG
,
253 static unsigned si_num_prims_for_vertices(const struct pipe_draw_info
*info
)
255 switch (info
->mode
) {
256 case PIPE_PRIM_PATCHES
:
257 return info
->count
/ info
->vertices_per_patch
;
258 case R600_PRIM_RECTANGLE_LIST
:
259 return info
->count
/ 3;
261 return u_prims_for_vertices(info
->mode
, info
->count
);
266 si_get_init_multi_vgt_param(struct si_screen
*sscreen
,
267 union si_vgt_param_key
*key
)
269 STATIC_ASSERT(sizeof(union si_vgt_param_key
) == 4);
270 unsigned max_primgroup_in_wave
= 2;
272 /* SWITCH_ON_EOP(0) is always preferable. */
273 bool wd_switch_on_eop
= false;
274 bool ia_switch_on_eop
= false;
275 bool ia_switch_on_eoi
= false;
276 bool partial_vs_wave
= false;
277 bool partial_es_wave
= false;
279 if (key
->u
.uses_tess
) {
280 /* SWITCH_ON_EOI must be set if PrimID is used. */
281 if (key
->u
.tcs_tes_uses_prim_id
)
282 ia_switch_on_eoi
= true;
284 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
285 if ((sscreen
->b
.family
== CHIP_TAHITI
||
286 sscreen
->b
.family
== CHIP_PITCAIRN
||
287 sscreen
->b
.family
== CHIP_BONAIRE
) &&
289 partial_vs_wave
= true;
291 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
292 if (sscreen
->has_distributed_tess
) {
293 if (key
->u
.uses_gs
) {
294 partial_es_wave
= true;
296 /* GPU hang workaround. */
297 if (sscreen
->b
.family
== CHIP_TONGA
||
298 sscreen
->b
.family
== CHIP_FIJI
||
299 sscreen
->b
.family
== CHIP_POLARIS10
||
300 sscreen
->b
.family
== CHIP_POLARIS11
)
301 partial_vs_wave
= true;
303 partial_vs_wave
= true;
308 /* This is a hardware requirement. */
309 if (key
->u
.line_stipple_enabled
||
310 (sscreen
->b
.debug_flags
& DBG_SWITCH_ON_EOP
)) {
311 ia_switch_on_eop
= true;
312 wd_switch_on_eop
= true;
315 if (sscreen
->b
.chip_class
>= CIK
) {
316 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
317 * 4 shader engines. Set 1 to pass the assertion below.
318 * The other cases are hardware requirements.
320 * Polaris supports primitive restart with WD_SWITCH_ON_EOP=0
321 * for points, line strips, and tri strips.
323 if (sscreen
->b
.info
.max_se
< 4 ||
324 key
->u
.prim
== PIPE_PRIM_POLYGON
||
325 key
->u
.prim
== PIPE_PRIM_LINE_LOOP
||
326 key
->u
.prim
== PIPE_PRIM_TRIANGLE_FAN
||
327 key
->u
.prim
== PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
||
328 (key
->u
.primitive_restart
&&
329 (sscreen
->b
.family
< CHIP_POLARIS10
||
330 (key
->u
.prim
!= PIPE_PRIM_POINTS
&&
331 key
->u
.prim
!= PIPE_PRIM_LINE_STRIP
&&
332 key
->u
.prim
!= PIPE_PRIM_TRIANGLE_STRIP
))) ||
333 key
->u
.count_from_stream_output
)
334 wd_switch_on_eop
= true;
336 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
337 * We don't know that for indirect drawing, so treat it as
338 * always problematic. */
339 if (sscreen
->b
.family
== CHIP_HAWAII
&&
340 key
->u
.uses_instancing
)
341 wd_switch_on_eop
= true;
343 /* Performance recommendation for 4 SE Gfx7-8 parts if
344 * instances are smaller than a primgroup.
345 * Assume indirect draws always use small instances.
346 * This is needed for good VS wave utilization.
348 if (sscreen
->b
.chip_class
<= VI
&&
349 sscreen
->b
.info
.max_se
== 4 &&
350 key
->u
.multi_instances_smaller_than_primgroup
)
351 wd_switch_on_eop
= true;
353 /* Required on CIK and later. */
354 if (sscreen
->b
.info
.max_se
> 2 && !wd_switch_on_eop
)
355 ia_switch_on_eoi
= true;
357 /* Required by Hawaii and, for some special cases, by VI. */
358 if (ia_switch_on_eoi
&&
359 (sscreen
->b
.family
== CHIP_HAWAII
||
360 (sscreen
->b
.chip_class
== VI
&&
361 (key
->u
.uses_gs
|| max_primgroup_in_wave
!= 2))))
362 partial_vs_wave
= true;
364 /* Instancing bug on Bonaire. */
365 if (sscreen
->b
.family
== CHIP_BONAIRE
&& ia_switch_on_eoi
&&
366 key
->u
.uses_instancing
)
367 partial_vs_wave
= true;
369 /* If the WD switch is false, the IA switch must be false too. */
370 assert(wd_switch_on_eop
|| !ia_switch_on_eop
);
373 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
374 if (ia_switch_on_eoi
)
375 partial_es_wave
= true;
377 return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop
) |
378 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi
) |
379 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave
) |
380 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave
) |
381 S_028AA8_WD_SWITCH_ON_EOP(sscreen
->b
.chip_class
>= CIK
? wd_switch_on_eop
: 0) |
382 S_028AA8_MAX_PRIMGRP_IN_WAVE(sscreen
->b
.chip_class
>= VI
?
383 max_primgroup_in_wave
: 0) |
384 S_030960_EN_INST_OPT_BASIC(sscreen
->b
.chip_class
>= GFX9
) |
385 S_030960_EN_INST_OPT_ADV(sscreen
->b
.chip_class
>= GFX9
);
388 void si_init_ia_multi_vgt_param_table(struct si_context
*sctx
)
390 for (int prim
= 0; prim
<= R600_PRIM_RECTANGLE_LIST
; prim
++)
391 for (int uses_instancing
= 0; uses_instancing
< 2; uses_instancing
++)
392 for (int multi_instances
= 0; multi_instances
< 2; multi_instances
++)
393 for (int primitive_restart
= 0; primitive_restart
< 2; primitive_restart
++)
394 for (int count_from_so
= 0; count_from_so
< 2; count_from_so
++)
395 for (int line_stipple
= 0; line_stipple
< 2; line_stipple
++)
396 for (int uses_tess
= 0; uses_tess
< 2; uses_tess
++)
397 for (int tess_uses_primid
= 0; tess_uses_primid
< 2; tess_uses_primid
++)
398 for (int uses_gs
= 0; uses_gs
< 2; uses_gs
++) {
399 union si_vgt_param_key key
;
403 key
.u
.uses_instancing
= uses_instancing
;
404 key
.u
.multi_instances_smaller_than_primgroup
= multi_instances
;
405 key
.u
.primitive_restart
= primitive_restart
;
406 key
.u
.count_from_stream_output
= count_from_so
;
407 key
.u
.line_stipple_enabled
= line_stipple
;
408 key
.u
.uses_tess
= uses_tess
;
409 key
.u
.tcs_tes_uses_prim_id
= tess_uses_primid
;
410 key
.u
.uses_gs
= uses_gs
;
412 sctx
->ia_multi_vgt_param
[key
.index
] =
413 si_get_init_multi_vgt_param(sctx
->screen
, &key
);
417 static unsigned si_get_ia_multi_vgt_param(struct si_context
*sctx
,
418 const struct pipe_draw_info
*info
,
419 unsigned num_patches
)
421 union si_vgt_param_key key
= sctx
->ia_multi_vgt_param_key
;
422 unsigned primgroup_size
;
423 unsigned ia_multi_vgt_param
;
425 if (sctx
->tes_shader
.cso
) {
426 primgroup_size
= num_patches
; /* must be a multiple of NUM_PATCHES */
427 } else if (sctx
->gs_shader
.cso
) {
428 primgroup_size
= 64; /* recommended with a GS */
430 primgroup_size
= 128; /* recommended without a GS and tess */
433 key
.u
.prim
= info
->mode
;
434 key
.u
.uses_instancing
= info
->indirect
|| info
->instance_count
> 1;
435 key
.u
.multi_instances_smaller_than_primgroup
=
437 (info
->instance_count
> 1 &&
438 (info
->count_from_stream_output
||
439 si_num_prims_for_vertices(info
) < primgroup_size
));
440 key
.u
.primitive_restart
= info
->primitive_restart
;
441 key
.u
.count_from_stream_output
= info
->count_from_stream_output
!= NULL
;
443 ia_multi_vgt_param
= sctx
->ia_multi_vgt_param
[key
.index
] |
444 S_028AA8_PRIMGROUP_SIZE(primgroup_size
- 1);
446 if (sctx
->gs_shader
.cso
) {
447 /* GS requirement. */
448 if (SI_GS_PER_ES
/ primgroup_size
>= sctx
->screen
->gs_table_depth
- 3)
449 ia_multi_vgt_param
|= S_028AA8_PARTIAL_ES_WAVE_ON(1);
451 /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
452 * The hw doc says all multi-SE chips are affected, but Vulkan
453 * only applies it to Hawaii. Do what Vulkan does.
455 if (sctx
->b
.family
== CHIP_HAWAII
&&
456 G_028AA8_SWITCH_ON_EOI(ia_multi_vgt_param
) &&
458 (info
->instance_count
> 1 &&
459 (info
->count_from_stream_output
||
460 si_num_prims_for_vertices(info
) <= 1))))
461 sctx
->b
.flags
|= SI_CONTEXT_VGT_FLUSH
;
464 return ia_multi_vgt_param
;
467 /* rast_prim is the primitive type after GS. */
468 static void si_emit_rasterizer_prim_state(struct si_context
*sctx
)
470 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
471 enum pipe_prim_type rast_prim
= sctx
->current_rast_prim
;
472 struct si_state_rasterizer
*rs
= sctx
->emitted
.named
.rasterizer
;
474 /* Skip this if not rendering lines. */
475 if (rast_prim
!= PIPE_PRIM_LINES
&&
476 rast_prim
!= PIPE_PRIM_LINE_LOOP
&&
477 rast_prim
!= PIPE_PRIM_LINE_STRIP
&&
478 rast_prim
!= PIPE_PRIM_LINES_ADJACENCY
&&
479 rast_prim
!= PIPE_PRIM_LINE_STRIP_ADJACENCY
)
482 if (rast_prim
== sctx
->last_rast_prim
&&
483 rs
->pa_sc_line_stipple
== sctx
->last_sc_line_stipple
)
486 /* For lines, reset the stipple pattern at each primitive. Otherwise,
487 * reset the stipple pattern at each packet (line strips, line loops).
489 radeon_set_context_reg(cs
, R_028A0C_PA_SC_LINE_STIPPLE
,
490 rs
->pa_sc_line_stipple
|
491 S_028A0C_AUTO_RESET_CNTL(rast_prim
== PIPE_PRIM_LINES
? 1 : 2));
493 sctx
->last_rast_prim
= rast_prim
;
494 sctx
->last_sc_line_stipple
= rs
->pa_sc_line_stipple
;
497 static void si_emit_draw_registers(struct si_context
*sctx
,
498 const struct pipe_draw_info
*info
)
500 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
501 unsigned prim
= si_conv_pipe_prim(info
->mode
);
502 unsigned gs_out_prim
= si_conv_prim_to_gs_out(sctx
->current_rast_prim
);
503 unsigned ia_multi_vgt_param
, num_patches
= 0;
505 if (sctx
->tes_shader
.cso
)
506 si_emit_derived_tess_state(sctx
, info
, &num_patches
);
508 ia_multi_vgt_param
= si_get_ia_multi_vgt_param(sctx
, info
, num_patches
);
511 if (ia_multi_vgt_param
!= sctx
->last_multi_vgt_param
) {
512 if (sctx
->b
.chip_class
>= GFX9
)
513 radeon_set_uconfig_reg_idx(cs
, R_030960_IA_MULTI_VGT_PARAM
, 4, ia_multi_vgt_param
);
514 else if (sctx
->b
.chip_class
>= CIK
)
515 radeon_set_context_reg_idx(cs
, R_028AA8_IA_MULTI_VGT_PARAM
, 1, ia_multi_vgt_param
);
517 radeon_set_context_reg(cs
, R_028AA8_IA_MULTI_VGT_PARAM
, ia_multi_vgt_param
);
519 sctx
->last_multi_vgt_param
= ia_multi_vgt_param
;
521 if (prim
!= sctx
->last_prim
) {
522 if (sctx
->b
.chip_class
>= CIK
)
523 radeon_set_uconfig_reg_idx(cs
, R_030908_VGT_PRIMITIVE_TYPE
, 1, prim
);
525 radeon_set_config_reg(cs
, R_008958_VGT_PRIMITIVE_TYPE
, prim
);
527 sctx
->last_prim
= prim
;
530 if (gs_out_prim
!= sctx
->last_gs_out_prim
) {
531 radeon_set_context_reg(cs
, R_028A6C_VGT_GS_OUT_PRIM_TYPE
, gs_out_prim
);
532 sctx
->last_gs_out_prim
= gs_out_prim
;
535 /* Primitive restart. */
536 if (info
->primitive_restart
!= sctx
->last_primitive_restart_en
) {
537 if (sctx
->b
.chip_class
>= GFX9
)
538 radeon_set_uconfig_reg(cs
, R_03092C_VGT_MULTI_PRIM_IB_RESET_EN
,
539 info
->primitive_restart
);
541 radeon_set_context_reg(cs
, R_028A94_VGT_MULTI_PRIM_IB_RESET_EN
,
542 info
->primitive_restart
);
544 sctx
->last_primitive_restart_en
= info
->primitive_restart
;
547 if (info
->primitive_restart
&&
548 (info
->restart_index
!= sctx
->last_restart_index
||
549 sctx
->last_restart_index
== SI_RESTART_INDEX_UNKNOWN
)) {
550 radeon_set_context_reg(cs
, R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX
,
551 info
->restart_index
);
552 sctx
->last_restart_index
= info
->restart_index
;
556 static void si_emit_draw_packets(struct si_context
*sctx
,
557 const struct pipe_draw_info
*info
,
558 const struct pipe_index_buffer
*ib
)
560 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
561 unsigned sh_base_reg
= sctx
->shader_userdata
.sh_base
[PIPE_SHADER_VERTEX
];
562 bool render_cond_bit
= sctx
->b
.render_cond
&& !sctx
->b
.render_cond_force_off
;
563 uint32_t index_max_size
= 0;
564 uint64_t index_va
= 0;
566 if (info
->count_from_stream_output
) {
567 struct r600_so_target
*t
=
568 (struct r600_so_target
*)info
->count_from_stream_output
;
569 uint64_t va
= t
->buf_filled_size
->gpu_address
+
570 t
->buf_filled_size_offset
;
572 radeon_set_context_reg(cs
, R_028B30_VGT_STRMOUT_DRAW_OPAQUE_VERTEX_STRIDE
,
575 radeon_emit(cs
, PKT3(PKT3_COPY_DATA
, 4, 0));
576 radeon_emit(cs
, COPY_DATA_SRC_SEL(COPY_DATA_MEM
) |
577 COPY_DATA_DST_SEL(COPY_DATA_REG
) |
578 COPY_DATA_WR_CONFIRM
);
579 radeon_emit(cs
, va
); /* src address lo */
580 radeon_emit(cs
, va
>> 32); /* src address hi */
581 radeon_emit(cs
, R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE
>> 2);
582 radeon_emit(cs
, 0); /* unused */
584 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
585 t
->buf_filled_size
, RADEON_USAGE_READ
,
586 RADEON_PRIO_SO_FILLED_SIZE
);
591 if (ib
->index_size
!= sctx
->last_index_size
) {
595 switch (ib
->index_size
) {
597 index_type
= V_028A7C_VGT_INDEX_8
;
600 index_type
= V_028A7C_VGT_INDEX_16
|
601 (SI_BIG_ENDIAN
&& sctx
->b
.chip_class
<= CIK
?
602 V_028A7C_VGT_DMA_SWAP_16_BIT
: 0);
605 index_type
= V_028A7C_VGT_INDEX_32
|
606 (SI_BIG_ENDIAN
&& sctx
->b
.chip_class
<= CIK
?
607 V_028A7C_VGT_DMA_SWAP_32_BIT
: 0);
610 assert(!"unreachable");
614 if (sctx
->b
.chip_class
>= GFX9
) {
615 radeon_set_uconfig_reg_idx(cs
, R_03090C_VGT_INDEX_TYPE
,
618 radeon_emit(cs
, PKT3(PKT3_INDEX_TYPE
, 0, 0));
619 radeon_emit(cs
, index_type
);
622 sctx
->last_index_size
= ib
->index_size
;
625 index_max_size
= (ib
->buffer
->width0
- ib
->offset
) /
627 index_va
= r600_resource(ib
->buffer
)->gpu_address
+ ib
->offset
;
629 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
630 (struct r600_resource
*)ib
->buffer
,
631 RADEON_USAGE_READ
, RADEON_PRIO_INDEX_BUFFER
);
633 /* On CI and later, non-indexed draws overwrite VGT_INDEX_TYPE,
634 * so the state must be re-emitted before the next indexed draw.
636 if (sctx
->b
.chip_class
>= CIK
)
637 sctx
->last_index_size
= -1;
640 if (!info
->indirect
) {
643 radeon_emit(cs
, PKT3(PKT3_NUM_INSTANCES
, 0, 0));
644 radeon_emit(cs
, info
->instance_count
);
646 /* Base vertex and start instance. */
647 base_vertex
= info
->indexed
? info
->index_bias
: info
->start
;
649 if (base_vertex
!= sctx
->last_base_vertex
||
650 sctx
->last_base_vertex
== SI_BASE_VERTEX_UNKNOWN
||
651 info
->start_instance
!= sctx
->last_start_instance
||
652 info
->drawid
!= sctx
->last_drawid
||
653 sh_base_reg
!= sctx
->last_sh_base_reg
) {
654 radeon_set_sh_reg_seq(cs
, sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4, 3);
655 radeon_emit(cs
, base_vertex
);
656 radeon_emit(cs
, info
->start_instance
);
657 radeon_emit(cs
, info
->drawid
);
659 sctx
->last_base_vertex
= base_vertex
;
660 sctx
->last_start_instance
= info
->start_instance
;
661 sctx
->last_drawid
= info
->drawid
;
662 sctx
->last_sh_base_reg
= sh_base_reg
;
665 uint64_t indirect_va
= r600_resource(info
->indirect
)->gpu_address
;
667 assert(indirect_va
% 8 == 0);
669 si_invalidate_draw_sh_constants(sctx
);
671 radeon_emit(cs
, PKT3(PKT3_SET_BASE
, 2, 0));
673 radeon_emit(cs
, indirect_va
);
674 radeon_emit(cs
, indirect_va
>> 32);
676 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
677 (struct r600_resource
*)info
->indirect
,
678 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
681 if (info
->indirect
) {
682 unsigned di_src_sel
= info
->indexed
? V_0287F0_DI_SRC_SEL_DMA
683 : V_0287F0_DI_SRC_SEL_AUTO_INDEX
;
685 assert(info
->indirect_offset
% 4 == 0);
688 radeon_emit(cs
, PKT3(PKT3_INDEX_BASE
, 1, 0));
689 radeon_emit(cs
, index_va
);
690 radeon_emit(cs
, index_va
>> 32);
692 radeon_emit(cs
, PKT3(PKT3_INDEX_BUFFER_SIZE
, 0, 0));
693 radeon_emit(cs
, index_max_size
);
696 if (!sctx
->screen
->has_draw_indirect_multi
) {
697 radeon_emit(cs
, PKT3(info
->indexed
? PKT3_DRAW_INDEX_INDIRECT
698 : PKT3_DRAW_INDIRECT
,
699 3, render_cond_bit
));
700 radeon_emit(cs
, info
->indirect_offset
);
701 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
702 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
703 radeon_emit(cs
, di_src_sel
);
705 uint64_t count_va
= 0;
707 if (info
->indirect_params
) {
708 struct r600_resource
*params_buf
=
709 (struct r600_resource
*)info
->indirect_params
;
711 radeon_add_to_buffer_list(
712 &sctx
->b
, &sctx
->b
.gfx
, params_buf
,
713 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
715 count_va
= params_buf
->gpu_address
+ info
->indirect_params_offset
;
718 radeon_emit(cs
, PKT3(info
->indexed
? PKT3_DRAW_INDEX_INDIRECT_MULTI
:
719 PKT3_DRAW_INDIRECT_MULTI
,
720 8, render_cond_bit
));
721 radeon_emit(cs
, info
->indirect_offset
);
722 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
723 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
724 radeon_emit(cs
, ((sh_base_reg
+ SI_SGPR_DRAWID
* 4 - SI_SH_REG_OFFSET
) >> 2) |
725 S_2C3_DRAW_INDEX_ENABLE(1) |
726 S_2C3_COUNT_INDIRECT_ENABLE(!!info
->indirect_params
));
727 radeon_emit(cs
, info
->indirect_count
);
728 radeon_emit(cs
, count_va
);
729 radeon_emit(cs
, count_va
>> 32);
730 radeon_emit(cs
, info
->indirect_stride
);
731 radeon_emit(cs
, di_src_sel
);
735 index_va
+= info
->start
* ib
->index_size
;
737 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_2
, 4, render_cond_bit
));
738 radeon_emit(cs
, index_max_size
);
739 radeon_emit(cs
, index_va
);
740 radeon_emit(cs
, index_va
>> 32);
741 radeon_emit(cs
, info
->count
);
742 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_DMA
);
744 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_AUTO
, 1, render_cond_bit
));
745 radeon_emit(cs
, info
->count
);
746 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_AUTO_INDEX
|
747 S_0287F0_USE_OPAQUE(!!info
->count_from_stream_output
));
752 static void si_emit_surface_sync(struct r600_common_context
*rctx
,
753 unsigned cp_coher_cntl
)
755 struct radeon_winsys_cs
*cs
= rctx
->gfx
.cs
;
757 if (rctx
->chip_class
>= GFX9
) {
758 /* Flush caches and wait for the caches to assert idle. */
759 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 5, 0));
760 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
761 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
762 radeon_emit(cs
, 0xffffff); /* CP_COHER_SIZE_HI */
763 radeon_emit(cs
, 0); /* CP_COHER_BASE */
764 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
765 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
767 /* ACQUIRE_MEM is only required on a compute ring. */
768 radeon_emit(cs
, PKT3(PKT3_SURFACE_SYNC
, 3, 0));
769 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
770 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
771 radeon_emit(cs
, 0); /* CP_COHER_BASE */
772 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
776 void si_emit_cache_flush(struct si_context
*sctx
)
778 struct r600_common_context
*rctx
= &sctx
->b
;
779 struct radeon_winsys_cs
*cs
= rctx
->gfx
.cs
;
780 uint32_t cp_coher_cntl
= 0;
781 uint32_t flush_cb_db
= rctx
->flags
& (SI_CONTEXT_FLUSH_AND_INV_CB
|
782 SI_CONTEXT_FLUSH_AND_INV_DB
);
784 if (rctx
->flags
& (SI_CONTEXT_FLUSH_AND_INV_CB
|
785 SI_CONTEXT_FLUSH_AND_INV_DB
))
786 sctx
->b
.num_fb_cache_flushes
++;
788 /* SI has a bug that it always flushes ICACHE and KCACHE if either
789 * bit is set. An alternative way is to write SQC_CACHES, but that
790 * doesn't seem to work reliably. Since the bug doesn't affect
791 * correctness (it only does more work than necessary) and
792 * the performance impact is likely negligible, there is no plan
793 * to add a workaround for it.
796 if (rctx
->flags
& SI_CONTEXT_INV_ICACHE
)
797 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
798 if (rctx
->flags
& SI_CONTEXT_INV_SMEM_L1
)
799 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
801 if (rctx
->chip_class
<= VI
) {
802 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_CB
) {
803 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
804 S_0085F0_CB0_DEST_BASE_ENA(1) |
805 S_0085F0_CB1_DEST_BASE_ENA(1) |
806 S_0085F0_CB2_DEST_BASE_ENA(1) |
807 S_0085F0_CB3_DEST_BASE_ENA(1) |
808 S_0085F0_CB4_DEST_BASE_ENA(1) |
809 S_0085F0_CB5_DEST_BASE_ENA(1) |
810 S_0085F0_CB6_DEST_BASE_ENA(1) |
811 S_0085F0_CB7_DEST_BASE_ENA(1);
813 /* Necessary for DCC */
814 if (rctx
->chip_class
== VI
)
815 r600_gfx_write_event_eop(rctx
, V_028A90_FLUSH_AND_INV_CB_DATA_TS
,
816 0, 0, NULL
, 0, 0, 0);
818 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_DB
)
819 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
820 S_0085F0_DB_DEST_BASE_ENA(1);
823 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_CB
) {
824 /* Flush CMASK/FMASK/DCC. SURFACE_SYNC will wait for idle. */
825 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
826 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
828 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_DB
) {
829 /* Flush HTILE. SURFACE_SYNC will wait for idle. */
830 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
831 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
834 /* Wait for shader engines to go idle.
835 * VS and PS waits are unnecessary if SURFACE_SYNC is going to wait
836 * for everything including CB/DB cache flushes.
839 if (rctx
->flags
& SI_CONTEXT_PS_PARTIAL_FLUSH
) {
840 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
841 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
842 /* Only count explicit shader flushes, not implicit ones
843 * done by SURFACE_SYNC.
845 rctx
->num_vs_flushes
++;
846 rctx
->num_ps_flushes
++;
847 } else if (rctx
->flags
& SI_CONTEXT_VS_PARTIAL_FLUSH
) {
848 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
849 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
850 rctx
->num_vs_flushes
++;
854 if (rctx
->flags
& SI_CONTEXT_CS_PARTIAL_FLUSH
&&
855 sctx
->compute_is_busy
) {
856 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
857 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
| EVENT_INDEX(4)));
858 rctx
->num_cs_flushes
++;
859 sctx
->compute_is_busy
= false;
862 /* VGT state synchronization. */
863 if (rctx
->flags
& SI_CONTEXT_VGT_FLUSH
) {
864 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
865 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
867 if (rctx
->flags
& SI_CONTEXT_VGT_STREAMOUT_SYNC
) {
868 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
869 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC
) | EVENT_INDEX(0));
872 /* GFX9: Wait for idle if we're flushing CB or DB. ACQUIRE_MEM doesn't
873 * wait for idle on GFX9. We have to use a TS event.
875 if (sctx
->b
.chip_class
>= GFX9
&& flush_cb_db
) {
876 struct r600_resource
*rbuf
= NULL
;
878 unsigned offset
= 0, tc_flags
, cb_db_event
;
880 /* Set the CB/DB flush event. */
881 switch (flush_cb_db
) {
882 case SI_CONTEXT_FLUSH_AND_INV_CB
:
883 cb_db_event
= V_028A90_FLUSH_AND_INV_CB_DATA_TS
;
885 case SI_CONTEXT_FLUSH_AND_INV_DB
:
886 cb_db_event
= V_028A90_FLUSH_AND_INV_DB_DATA_TS
;
890 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
893 /* TC | TC_WB = invalidate L2 data
894 * TC_MD | TC_WB = invalidate L2 metadata
895 * TC | TC_WB | TC_MD = invalidate L2 data & metadata
897 * The metadata cache must always be invalidated for coherency
898 * between CB/DB and shaders. (metadata = HTILE, CMASK, DCC)
900 * TC must be invalidated on GFX9 only if the CB/DB surface is
901 * not pipe-aligned. If the surface is RB-aligned, it might not
902 * strictly be pipe-aligned since RB alignment takes precendence.
904 tc_flags
= EVENT_TC_WB_ACTION_ENA
|
905 EVENT_TC_MD_ACTION_ENA
;
907 /* Ideally flush TC together with CB/DB. */
908 if (rctx
->flags
& SI_CONTEXT_INV_GLOBAL_L2
) {
909 tc_flags
|= EVENT_TC_ACTION_ENA
|
910 EVENT_TCL1_ACTION_ENA
;
912 /* Clear the flags. */
913 rctx
->flags
&= ~(SI_CONTEXT_INV_GLOBAL_L2
|
914 SI_CONTEXT_WRITEBACK_GLOBAL_L2
|
915 SI_CONTEXT_INV_VMEM_L1
);
918 /* Allocate memory for the fence. */
919 u_suballocator_alloc(rctx
->allocator_zeroed_memory
, 4, 4,
920 &offset
, (struct pipe_resource
**)&rbuf
);
921 va
= rbuf
->gpu_address
+ offset
;
923 r600_gfx_write_event_eop(rctx
, cb_db_event
, tc_flags
, 1,
925 r600_gfx_wait_fence(rctx
, va
, 1, 0xffffffff);
928 /* Make sure ME is idle (it executes most packets) before continuing.
929 * This prevents read-after-write hazards between PFP and ME.
932 (rctx
->flags
& (SI_CONTEXT_CS_PARTIAL_FLUSH
|
933 SI_CONTEXT_INV_VMEM_L1
|
934 SI_CONTEXT_INV_GLOBAL_L2
|
935 SI_CONTEXT_WRITEBACK_GLOBAL_L2
))) {
936 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
941 * When one of the CP_COHER_CNTL.DEST_BASE flags is set, SURFACE_SYNC
942 * waits for idle, so it should be last. SURFACE_SYNC is done in PFP.
944 * cp_coher_cntl should contain all necessary flags except TC flags
947 * SI-CIK don't support L2 write-back.
949 if (rctx
->flags
& SI_CONTEXT_INV_GLOBAL_L2
||
950 (rctx
->chip_class
<= CIK
&&
951 (rctx
->flags
& SI_CONTEXT_WRITEBACK_GLOBAL_L2
))) {
952 /* Invalidate L1 & L2. (L1 is always invalidated on SI)
953 * WB must be set on VI+ when TC_ACTION is set.
955 si_emit_surface_sync(rctx
, cp_coher_cntl
|
956 S_0085F0_TC_ACTION_ENA(1) |
957 S_0085F0_TCL1_ACTION_ENA(1) |
958 S_0301F0_TC_WB_ACTION_ENA(rctx
->chip_class
>= VI
));
960 sctx
->b
.num_L2_invalidates
++;
962 /* L1 invalidation and L2 writeback must be done separately,
963 * because both operations can't be done together.
965 if (rctx
->flags
& SI_CONTEXT_WRITEBACK_GLOBAL_L2
) {
967 * NC = apply to non-coherent MTYPEs
968 * (i.e. MTYPE <= 1, which is what we use everywhere)
970 * WB doesn't work without NC.
972 si_emit_surface_sync(rctx
, cp_coher_cntl
|
973 S_0301F0_TC_WB_ACTION_ENA(1) |
974 S_0301F0_TC_NC_ACTION_ENA(1));
976 sctx
->b
.num_L2_writebacks
++;
978 if (rctx
->flags
& SI_CONTEXT_INV_VMEM_L1
) {
979 /* Invalidate per-CU VMEM L1. */
980 si_emit_surface_sync(rctx
, cp_coher_cntl
|
981 S_0085F0_TCL1_ACTION_ENA(1));
986 /* If TC flushes haven't cleared this... */
988 si_emit_surface_sync(rctx
, cp_coher_cntl
);
990 if (rctx
->flags
& R600_CONTEXT_START_PIPELINE_STATS
) {
991 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
992 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
994 } else if (rctx
->flags
& R600_CONTEXT_STOP_PIPELINE_STATS
) {
995 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
996 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
1003 static void si_get_draw_start_count(struct si_context
*sctx
,
1004 const struct pipe_draw_info
*info
,
1005 unsigned *start
, unsigned *count
)
1007 if (info
->indirect
) {
1008 unsigned indirect_count
;
1009 struct pipe_transfer
*transfer
;
1010 unsigned begin
, end
;
1014 if (info
->indirect_params
) {
1015 data
= pipe_buffer_map_range(&sctx
->b
.b
,
1016 info
->indirect_params
,
1017 info
->indirect_params_offset
,
1019 PIPE_TRANSFER_READ
, &transfer
);
1021 indirect_count
= *data
;
1023 pipe_buffer_unmap(&sctx
->b
.b
, transfer
);
1025 indirect_count
= info
->indirect_count
;
1028 if (!indirect_count
) {
1029 *start
= *count
= 0;
1033 map_size
= (indirect_count
- 1) * info
->indirect_stride
+ 3 * sizeof(unsigned);
1034 data
= pipe_buffer_map_range(&sctx
->b
.b
, info
->indirect
,
1035 info
->indirect_offset
, map_size
,
1036 PIPE_TRANSFER_READ
, &transfer
);
1041 for (unsigned i
= 0; i
< indirect_count
; ++i
) {
1042 unsigned count
= data
[0];
1043 unsigned start
= data
[2];
1046 begin
= MIN2(begin
, start
);
1047 end
= MAX2(end
, start
+ count
);
1050 data
+= info
->indirect_stride
/ sizeof(unsigned);
1053 pipe_buffer_unmap(&sctx
->b
.b
, transfer
);
1057 *count
= end
- begin
;
1059 *start
= *count
= 0;
1062 *start
= info
->start
;
1063 *count
= info
->count
;
1067 void si_ce_pre_draw_synchronization(struct si_context
*sctx
)
1069 if (sctx
->ce_need_synchronization
) {
1070 radeon_emit(sctx
->ce_ib
, PKT3(PKT3_INCREMENT_CE_COUNTER
, 0, 0));
1071 radeon_emit(sctx
->ce_ib
, 1);
1073 radeon_emit(sctx
->b
.gfx
.cs
, PKT3(PKT3_WAIT_ON_CE_COUNTER
, 0, 0));
1074 radeon_emit(sctx
->b
.gfx
.cs
, 1);
1078 void si_ce_post_draw_synchronization(struct si_context
*sctx
)
1080 if (sctx
->ce_need_synchronization
) {
1081 radeon_emit(sctx
->b
.gfx
.cs
, PKT3(PKT3_INCREMENT_DE_COUNTER
, 0, 0));
1082 radeon_emit(sctx
->b
.gfx
.cs
, 0);
1084 sctx
->ce_need_synchronization
= false;
1088 void si_draw_vbo(struct pipe_context
*ctx
, const struct pipe_draw_info
*info
)
1090 struct si_context
*sctx
= (struct si_context
*)ctx
;
1091 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1092 const struct pipe_index_buffer
*ib
= &sctx
->index_buffer
;
1093 struct pipe_index_buffer ib_tmp
; /* for index buffer uploads only */
1094 unsigned mask
, dirty_tex_counter
;
1095 enum pipe_prim_type rast_prim
;
1097 if (likely(!info
->indirect
)) {
1098 /* SI-CI treat instance_count==0 as instance_count==1. There is
1099 * no workaround for indirect draws, but we can at least skip
1102 if (unlikely(!info
->instance_count
))
1105 /* Handle count == 0. */
1106 if (unlikely(!info
->count
&&
1107 (info
->indexed
|| !info
->count_from_stream_output
)))
1111 if (unlikely(!sctx
->vs_shader
.cso
)) {
1115 if (unlikely(!sctx
->ps_shader
.cso
&& (!rs
|| !rs
->rasterizer_discard
))) {
1119 if (unlikely(!!sctx
->tes_shader
.cso
!= (info
->mode
== PIPE_PRIM_PATCHES
))) {
1124 /* Recompute and re-emit the texture resource states if needed. */
1125 dirty_tex_counter
= p_atomic_read(&sctx
->b
.screen
->dirty_tex_counter
);
1126 if (unlikely(dirty_tex_counter
!= sctx
->b
.last_dirty_tex_counter
)) {
1127 sctx
->b
.last_dirty_tex_counter
= dirty_tex_counter
;
1128 sctx
->framebuffer
.dirty_cbufs
|=
1129 ((1 << sctx
->framebuffer
.state
.nr_cbufs
) - 1);
1130 sctx
->framebuffer
.dirty_zsbuf
= true;
1131 sctx
->framebuffer
.do_update_surf_dirtiness
= true;
1132 si_mark_atom_dirty(sctx
, &sctx
->framebuffer
.atom
);
1133 si_update_all_texture_descriptors(sctx
);
1136 si_decompress_graphics_textures(sctx
);
1138 /* Set the rasterization primitive type.
1140 * This must be done after si_decompress_textures, which can call
1141 * draw_vbo recursively, and before si_update_shaders, which uses
1142 * current_rast_prim for this draw_vbo call. */
1143 if (sctx
->gs_shader
.cso
)
1144 rast_prim
= sctx
->gs_shader
.cso
->gs_output_prim
;
1145 else if (sctx
->tes_shader
.cso
)
1146 rast_prim
= sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
1148 rast_prim
= info
->mode
;
1150 if (rast_prim
!= sctx
->current_rast_prim
) {
1151 sctx
->current_rast_prim
= rast_prim
;
1152 sctx
->do_update_shaders
= true;
1155 if (sctx
->gs_shader
.cso
) {
1156 /* Determine whether the GS triangle strip adjacency fix should
1157 * be applied. Rotate every other triangle if
1158 * - triangle strips with adjacency are fed to the GS and
1159 * - primitive restart is disabled (the rotation doesn't help
1160 * when the restart occurs after an odd number of triangles).
1162 bool gs_tri_strip_adj_fix
=
1163 !sctx
->tes_shader
.cso
&&
1164 info
->mode
== PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
&&
1165 !info
->primitive_restart
;
1167 if (gs_tri_strip_adj_fix
!= sctx
->gs_tri_strip_adj_fix
) {
1168 sctx
->gs_tri_strip_adj_fix
= gs_tri_strip_adj_fix
;
1169 sctx
->do_update_shaders
= true;
1173 if (sctx
->do_update_shaders
&& !si_update_shaders(sctx
))
1176 if (!si_upload_graphics_shader_descriptors(sctx
))
1179 ib_tmp
.buffer
= NULL
;
1181 if (info
->indexed
) {
1182 /* Translate or upload, if needed. */
1183 /* 8-bit indices are supported on VI. */
1184 if (sctx
->b
.chip_class
<= CIK
&& ib
->index_size
== 1) {
1185 unsigned start
, count
, start_offset
, size
;
1188 si_get_draw_start_count(sctx
, info
, &start
, &count
);
1189 start_offset
= start
* 2;
1192 u_upload_alloc(ctx
->stream_uploader
, start_offset
,
1194 si_optimal_tcc_alignment(sctx
, size
),
1195 &ib_tmp
.offset
, &ib_tmp
.buffer
, &ptr
);
1199 util_shorten_ubyte_elts_to_userptr(&sctx
->b
.b
, ib
, 0, 0,
1203 /* info->start will be added by the drawing code */
1204 ib_tmp
.offset
-= start_offset
;
1205 ib_tmp
.index_size
= 2;
1207 } else if (ib
->user_buffer
&& !ib
->buffer
) {
1208 unsigned start
, count
, start_offset
;
1210 si_get_draw_start_count(sctx
, info
, &start
, &count
);
1211 start_offset
= start
* ib
->index_size
;
1213 u_upload_data(ctx
->stream_uploader
, start_offset
,
1214 count
* ib
->index_size
,
1215 sctx
->screen
->b
.info
.tcc_cache_line_size
,
1216 (char*)ib
->user_buffer
+ start_offset
,
1217 &ib_tmp
.offset
, &ib_tmp
.buffer
);
1221 /* info->start will be added by the drawing code */
1222 ib_tmp
.offset
-= start_offset
;
1223 ib_tmp
.index_size
= ib
->index_size
;
1225 } else if (sctx
->b
.chip_class
<= CIK
&&
1226 r600_resource(ib
->buffer
)->TC_L2_dirty
) {
1227 /* VI reads index buffers through TC L2, so it doesn't
1229 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1230 r600_resource(ib
->buffer
)->TC_L2_dirty
= false;
1234 if (info
->indirect
) {
1235 /* Add the buffer size for memory checking in need_cs_space. */
1236 r600_context_add_resource_size(ctx
, info
->indirect
);
1238 if (r600_resource(info
->indirect
)->TC_L2_dirty
) {
1239 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1240 r600_resource(info
->indirect
)->TC_L2_dirty
= false;
1243 if (info
->indirect_params
&&
1244 r600_resource(info
->indirect_params
)->TC_L2_dirty
) {
1245 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1246 r600_resource(info
->indirect_params
)->TC_L2_dirty
= false;
1250 si_need_cs_space(sctx
);
1252 /* Since we've called r600_context_add_resource_size for vertex buffers,
1253 * this must be called after si_need_cs_space, because we must let
1254 * need_cs_space flush before we add buffers to the buffer list.
1256 if (!si_upload_vertex_buffer_descriptors(sctx
))
1259 /* GFX9 scissor bug workaround. There is also a more efficient but
1260 * more involved alternative workaround. */
1261 if (sctx
->b
.chip_class
== GFX9
&&
1262 si_is_atom_dirty(sctx
, &sctx
->b
.scissors
.atom
))
1263 sctx
->b
.flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
;
1265 /* Flush caches before the first state atom, which does L2 prefetches. */
1267 si_emit_cache_flush(sctx
);
1269 /* Emit state atoms. */
1270 mask
= sctx
->dirty_atoms
;
1272 struct r600_atom
*atom
= sctx
->atoms
.array
[u_bit_scan(&mask
)];
1274 atom
->emit(&sctx
->b
, atom
);
1276 sctx
->dirty_atoms
= 0;
1279 mask
= sctx
->dirty_states
;
1281 unsigned i
= u_bit_scan(&mask
);
1282 struct si_pm4_state
*state
= sctx
->queued
.array
[i
];
1284 if (!state
|| sctx
->emitted
.array
[i
] == state
)
1287 si_pm4_emit(sctx
, state
);
1288 sctx
->emitted
.array
[i
] = state
;
1290 sctx
->dirty_states
= 0;
1292 si_emit_rasterizer_prim_state(sctx
);
1293 si_emit_draw_registers(sctx
, info
);
1295 si_ce_pre_draw_synchronization(sctx
);
1296 si_emit_draw_packets(sctx
, info
, ib
);
1297 si_ce_post_draw_synchronization(sctx
);
1299 if (sctx
->trace_buf
)
1300 si_trace_emit(sctx
);
1302 /* Workaround for a VGT hang when streamout is enabled.
1303 * It must be done after drawing. */
1304 if ((sctx
->b
.family
== CHIP_HAWAII
||
1305 sctx
->b
.family
== CHIP_TONGA
||
1306 sctx
->b
.family
== CHIP_FIJI
) &&
1307 r600_get_strmout_en(&sctx
->b
)) {
1308 sctx
->b
.flags
|= SI_CONTEXT_VGT_STREAMOUT_SYNC
;
1311 if (sctx
->framebuffer
.do_update_surf_dirtiness
) {
1312 /* Set the depth buffer as dirty. */
1313 if (sctx
->framebuffer
.state
.zsbuf
) {
1314 struct pipe_surface
*surf
= sctx
->framebuffer
.state
.zsbuf
;
1315 struct r600_texture
*rtex
= (struct r600_texture
*)surf
->texture
;
1317 if (!rtex
->tc_compatible_htile
)
1318 rtex
->dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
1320 if (rtex
->surface
.flags
& RADEON_SURF_SBUFFER
)
1321 rtex
->stencil_dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
1323 if (sctx
->framebuffer
.compressed_cb_mask
) {
1324 struct pipe_surface
*surf
;
1325 struct r600_texture
*rtex
;
1326 unsigned mask
= sctx
->framebuffer
.compressed_cb_mask
;
1329 unsigned i
= u_bit_scan(&mask
);
1330 surf
= sctx
->framebuffer
.state
.cbufs
[i
];
1331 rtex
= (struct r600_texture
*)surf
->texture
;
1333 if (rtex
->fmask
.size
)
1334 rtex
->dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
1335 if (rtex
->dcc_gather_statistics
)
1336 rtex
->separate_dcc_dirty
= true;
1339 sctx
->framebuffer
.do_update_surf_dirtiness
= false;
1342 pipe_resource_reference(&ib_tmp
.buffer
, NULL
);
1343 sctx
->b
.num_draw_calls
++;
1344 if (G_0286E8_WAVESIZE(sctx
->spi_tmpring_size
))
1345 sctx
->b
.num_spill_draw_calls
++;
1348 void si_trace_emit(struct si_context
*sctx
)
1350 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
1353 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
, sctx
->trace_buf
,
1354 RADEON_USAGE_READWRITE
, RADEON_PRIO_TRACE
);
1355 radeon_emit(cs
, PKT3(PKT3_WRITE_DATA
, 3, 0));
1356 radeon_emit(cs
, S_370_DST_SEL(V_370_MEMORY_SYNC
) |
1357 S_370_WR_CONFIRM(1) |
1358 S_370_ENGINE_SEL(V_370_ME
));
1359 radeon_emit(cs
, sctx
->trace_buf
->gpu_address
);
1360 radeon_emit(cs
, sctx
->trace_buf
->gpu_address
>> 32);
1361 radeon_emit(cs
, sctx
->trace_id
);
1362 radeon_emit(cs
, PKT3(PKT3_NOP
, 0, 0));
1363 radeon_emit(cs
, AC_ENCODE_TRACE_POINT(sctx
->trace_id
));