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_vs_state(struct si_context
*sctx
)
499 if (sctx
->current_vs_state
!= sctx
->last_vs_state
) {
500 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
502 radeon_set_sh_reg(cs
,
503 R_00B130_SPI_SHADER_USER_DATA_VS_0
+ SI_SGPR_VS_STATE_BITS
* 4,
504 sctx
->current_vs_state
);
506 sctx
->last_vs_state
= sctx
->current_vs_state
;
510 static void si_emit_draw_registers(struct si_context
*sctx
,
511 const struct pipe_draw_info
*info
,
512 unsigned num_patches
)
514 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
515 unsigned prim
= si_conv_pipe_prim(info
->mode
);
516 unsigned gs_out_prim
= si_conv_prim_to_gs_out(sctx
->current_rast_prim
);
517 unsigned ia_multi_vgt_param
;
519 ia_multi_vgt_param
= si_get_ia_multi_vgt_param(sctx
, info
, num_patches
);
522 if (ia_multi_vgt_param
!= sctx
->last_multi_vgt_param
) {
523 if (sctx
->b
.chip_class
>= GFX9
)
524 radeon_set_uconfig_reg_idx(cs
, R_030960_IA_MULTI_VGT_PARAM
, 4, ia_multi_vgt_param
);
525 else if (sctx
->b
.chip_class
>= CIK
)
526 radeon_set_context_reg_idx(cs
, R_028AA8_IA_MULTI_VGT_PARAM
, 1, ia_multi_vgt_param
);
528 radeon_set_context_reg(cs
, R_028AA8_IA_MULTI_VGT_PARAM
, ia_multi_vgt_param
);
530 sctx
->last_multi_vgt_param
= ia_multi_vgt_param
;
532 if (prim
!= sctx
->last_prim
) {
533 if (sctx
->b
.chip_class
>= CIK
)
534 radeon_set_uconfig_reg_idx(cs
, R_030908_VGT_PRIMITIVE_TYPE
, 1, prim
);
536 radeon_set_config_reg(cs
, R_008958_VGT_PRIMITIVE_TYPE
, prim
);
538 sctx
->last_prim
= prim
;
541 if (gs_out_prim
!= sctx
->last_gs_out_prim
) {
542 radeon_set_context_reg(cs
, R_028A6C_VGT_GS_OUT_PRIM_TYPE
, gs_out_prim
);
543 sctx
->last_gs_out_prim
= gs_out_prim
;
546 /* Primitive restart. */
547 if (info
->primitive_restart
!= sctx
->last_primitive_restart_en
) {
548 if (sctx
->b
.chip_class
>= GFX9
)
549 radeon_set_uconfig_reg(cs
, R_03092C_VGT_MULTI_PRIM_IB_RESET_EN
,
550 info
->primitive_restart
);
552 radeon_set_context_reg(cs
, R_028A94_VGT_MULTI_PRIM_IB_RESET_EN
,
553 info
->primitive_restart
);
555 sctx
->last_primitive_restart_en
= info
->primitive_restart
;
558 if (info
->primitive_restart
&&
559 (info
->restart_index
!= sctx
->last_restart_index
||
560 sctx
->last_restart_index
== SI_RESTART_INDEX_UNKNOWN
)) {
561 radeon_set_context_reg(cs
, R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX
,
562 info
->restart_index
);
563 sctx
->last_restart_index
= info
->restart_index
;
567 static void si_emit_draw_packets(struct si_context
*sctx
,
568 const struct pipe_draw_info
*info
,
569 const struct pipe_index_buffer
*ib
)
571 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
572 unsigned sh_base_reg
= sctx
->shader_userdata
.sh_base
[PIPE_SHADER_VERTEX
];
573 bool render_cond_bit
= sctx
->b
.render_cond
&& !sctx
->b
.render_cond_force_off
;
574 uint32_t index_max_size
= 0;
575 uint64_t index_va
= 0;
577 if (info
->count_from_stream_output
) {
578 struct r600_so_target
*t
=
579 (struct r600_so_target
*)info
->count_from_stream_output
;
580 uint64_t va
= t
->buf_filled_size
->gpu_address
+
581 t
->buf_filled_size_offset
;
583 radeon_set_context_reg(cs
, R_028B30_VGT_STRMOUT_DRAW_OPAQUE_VERTEX_STRIDE
,
586 radeon_emit(cs
, PKT3(PKT3_COPY_DATA
, 4, 0));
587 radeon_emit(cs
, COPY_DATA_SRC_SEL(COPY_DATA_MEM
) |
588 COPY_DATA_DST_SEL(COPY_DATA_REG
) |
589 COPY_DATA_WR_CONFIRM
);
590 radeon_emit(cs
, va
); /* src address lo */
591 radeon_emit(cs
, va
>> 32); /* src address hi */
592 radeon_emit(cs
, R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE
>> 2);
593 radeon_emit(cs
, 0); /* unused */
595 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
596 t
->buf_filled_size
, RADEON_USAGE_READ
,
597 RADEON_PRIO_SO_FILLED_SIZE
);
602 if (ib
->index_size
!= sctx
->last_index_size
) {
606 switch (ib
->index_size
) {
608 index_type
= V_028A7C_VGT_INDEX_8
;
611 index_type
= V_028A7C_VGT_INDEX_16
|
612 (SI_BIG_ENDIAN
&& sctx
->b
.chip_class
<= CIK
?
613 V_028A7C_VGT_DMA_SWAP_16_BIT
: 0);
616 index_type
= V_028A7C_VGT_INDEX_32
|
617 (SI_BIG_ENDIAN
&& sctx
->b
.chip_class
<= CIK
?
618 V_028A7C_VGT_DMA_SWAP_32_BIT
: 0);
621 assert(!"unreachable");
625 if (sctx
->b
.chip_class
>= GFX9
) {
626 radeon_set_uconfig_reg_idx(cs
, R_03090C_VGT_INDEX_TYPE
,
629 radeon_emit(cs
, PKT3(PKT3_INDEX_TYPE
, 0, 0));
630 radeon_emit(cs
, index_type
);
633 sctx
->last_index_size
= ib
->index_size
;
636 index_max_size
= (ib
->buffer
->width0
- ib
->offset
) /
638 index_va
= r600_resource(ib
->buffer
)->gpu_address
+ ib
->offset
;
640 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
641 (struct r600_resource
*)ib
->buffer
,
642 RADEON_USAGE_READ
, RADEON_PRIO_INDEX_BUFFER
);
644 /* On CI and later, non-indexed draws overwrite VGT_INDEX_TYPE,
645 * so the state must be re-emitted before the next indexed draw.
647 if (sctx
->b
.chip_class
>= CIK
)
648 sctx
->last_index_size
= -1;
651 if (!info
->indirect
) {
654 radeon_emit(cs
, PKT3(PKT3_NUM_INSTANCES
, 0, 0));
655 radeon_emit(cs
, info
->instance_count
);
657 /* Base vertex and start instance. */
658 base_vertex
= info
->indexed
? info
->index_bias
: info
->start
;
660 if (base_vertex
!= sctx
->last_base_vertex
||
661 sctx
->last_base_vertex
== SI_BASE_VERTEX_UNKNOWN
||
662 info
->start_instance
!= sctx
->last_start_instance
||
663 info
->drawid
!= sctx
->last_drawid
||
664 sh_base_reg
!= sctx
->last_sh_base_reg
) {
665 radeon_set_sh_reg_seq(cs
, sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4, 3);
666 radeon_emit(cs
, base_vertex
);
667 radeon_emit(cs
, info
->start_instance
);
668 radeon_emit(cs
, info
->drawid
);
670 sctx
->last_base_vertex
= base_vertex
;
671 sctx
->last_start_instance
= info
->start_instance
;
672 sctx
->last_drawid
= info
->drawid
;
673 sctx
->last_sh_base_reg
= sh_base_reg
;
676 uint64_t indirect_va
= r600_resource(info
->indirect
)->gpu_address
;
678 assert(indirect_va
% 8 == 0);
680 si_invalidate_draw_sh_constants(sctx
);
682 radeon_emit(cs
, PKT3(PKT3_SET_BASE
, 2, 0));
684 radeon_emit(cs
, indirect_va
);
685 radeon_emit(cs
, indirect_va
>> 32);
687 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
688 (struct r600_resource
*)info
->indirect
,
689 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
692 if (info
->indirect
) {
693 unsigned di_src_sel
= info
->indexed
? V_0287F0_DI_SRC_SEL_DMA
694 : V_0287F0_DI_SRC_SEL_AUTO_INDEX
;
696 assert(info
->indirect_offset
% 4 == 0);
699 radeon_emit(cs
, PKT3(PKT3_INDEX_BASE
, 1, 0));
700 radeon_emit(cs
, index_va
);
701 radeon_emit(cs
, index_va
>> 32);
703 radeon_emit(cs
, PKT3(PKT3_INDEX_BUFFER_SIZE
, 0, 0));
704 radeon_emit(cs
, index_max_size
);
707 if (!sctx
->screen
->has_draw_indirect_multi
) {
708 radeon_emit(cs
, PKT3(info
->indexed
? PKT3_DRAW_INDEX_INDIRECT
709 : PKT3_DRAW_INDIRECT
,
710 3, render_cond_bit
));
711 radeon_emit(cs
, info
->indirect_offset
);
712 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
713 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
714 radeon_emit(cs
, di_src_sel
);
716 uint64_t count_va
= 0;
718 if (info
->indirect_params
) {
719 struct r600_resource
*params_buf
=
720 (struct r600_resource
*)info
->indirect_params
;
722 radeon_add_to_buffer_list(
723 &sctx
->b
, &sctx
->b
.gfx
, params_buf
,
724 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
726 count_va
= params_buf
->gpu_address
+ info
->indirect_params_offset
;
729 radeon_emit(cs
, PKT3(info
->indexed
? PKT3_DRAW_INDEX_INDIRECT_MULTI
:
730 PKT3_DRAW_INDIRECT_MULTI
,
731 8, render_cond_bit
));
732 radeon_emit(cs
, info
->indirect_offset
);
733 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
734 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
735 radeon_emit(cs
, ((sh_base_reg
+ SI_SGPR_DRAWID
* 4 - SI_SH_REG_OFFSET
) >> 2) |
736 S_2C3_DRAW_INDEX_ENABLE(1) |
737 S_2C3_COUNT_INDIRECT_ENABLE(!!info
->indirect_params
));
738 radeon_emit(cs
, info
->indirect_count
);
739 radeon_emit(cs
, count_va
);
740 radeon_emit(cs
, count_va
>> 32);
741 radeon_emit(cs
, info
->indirect_stride
);
742 radeon_emit(cs
, di_src_sel
);
746 index_va
+= info
->start
* ib
->index_size
;
748 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_2
, 4, render_cond_bit
));
749 radeon_emit(cs
, index_max_size
);
750 radeon_emit(cs
, index_va
);
751 radeon_emit(cs
, index_va
>> 32);
752 radeon_emit(cs
, info
->count
);
753 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_DMA
);
755 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_AUTO
, 1, render_cond_bit
));
756 radeon_emit(cs
, info
->count
);
757 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_AUTO_INDEX
|
758 S_0287F0_USE_OPAQUE(!!info
->count_from_stream_output
));
763 static void si_emit_surface_sync(struct r600_common_context
*rctx
,
764 unsigned cp_coher_cntl
)
766 struct radeon_winsys_cs
*cs
= rctx
->gfx
.cs
;
768 if (rctx
->chip_class
>= GFX9
) {
769 /* Flush caches and wait for the caches to assert idle. */
770 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 5, 0));
771 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
772 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
773 radeon_emit(cs
, 0xffffff); /* CP_COHER_SIZE_HI */
774 radeon_emit(cs
, 0); /* CP_COHER_BASE */
775 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
776 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
778 /* ACQUIRE_MEM is only required on a compute ring. */
779 radeon_emit(cs
, PKT3(PKT3_SURFACE_SYNC
, 3, 0));
780 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
781 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
782 radeon_emit(cs
, 0); /* CP_COHER_BASE */
783 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
787 void si_emit_cache_flush(struct si_context
*sctx
)
789 struct r600_common_context
*rctx
= &sctx
->b
;
790 struct radeon_winsys_cs
*cs
= rctx
->gfx
.cs
;
791 uint32_t cp_coher_cntl
= 0;
792 uint32_t flush_cb_db
= rctx
->flags
& (SI_CONTEXT_FLUSH_AND_INV_CB
|
793 SI_CONTEXT_FLUSH_AND_INV_DB
);
795 if (rctx
->flags
& (SI_CONTEXT_FLUSH_AND_INV_CB
|
796 SI_CONTEXT_FLUSH_AND_INV_DB
))
797 sctx
->b
.num_fb_cache_flushes
++;
799 /* SI has a bug that it always flushes ICACHE and KCACHE if either
800 * bit is set. An alternative way is to write SQC_CACHES, but that
801 * doesn't seem to work reliably. Since the bug doesn't affect
802 * correctness (it only does more work than necessary) and
803 * the performance impact is likely negligible, there is no plan
804 * to add a workaround for it.
807 if (rctx
->flags
& SI_CONTEXT_INV_ICACHE
)
808 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
809 if (rctx
->flags
& SI_CONTEXT_INV_SMEM_L1
)
810 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
812 if (rctx
->chip_class
<= VI
) {
813 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_CB
) {
814 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
815 S_0085F0_CB0_DEST_BASE_ENA(1) |
816 S_0085F0_CB1_DEST_BASE_ENA(1) |
817 S_0085F0_CB2_DEST_BASE_ENA(1) |
818 S_0085F0_CB3_DEST_BASE_ENA(1) |
819 S_0085F0_CB4_DEST_BASE_ENA(1) |
820 S_0085F0_CB5_DEST_BASE_ENA(1) |
821 S_0085F0_CB6_DEST_BASE_ENA(1) |
822 S_0085F0_CB7_DEST_BASE_ENA(1);
824 /* Necessary for DCC */
825 if (rctx
->chip_class
== VI
)
826 r600_gfx_write_event_eop(rctx
, V_028A90_FLUSH_AND_INV_CB_DATA_TS
,
827 0, 0, NULL
, 0, 0, 0);
829 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_DB
)
830 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
831 S_0085F0_DB_DEST_BASE_ENA(1);
834 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_CB
) {
835 /* Flush CMASK/FMASK/DCC. SURFACE_SYNC will wait for idle. */
836 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
837 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
839 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_DB
) {
840 /* Flush HTILE. SURFACE_SYNC will wait for idle. */
841 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
842 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
845 /* Wait for shader engines to go idle.
846 * VS and PS waits are unnecessary if SURFACE_SYNC is going to wait
847 * for everything including CB/DB cache flushes.
850 if (rctx
->flags
& SI_CONTEXT_PS_PARTIAL_FLUSH
) {
851 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
852 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
853 /* Only count explicit shader flushes, not implicit ones
854 * done by SURFACE_SYNC.
856 rctx
->num_vs_flushes
++;
857 rctx
->num_ps_flushes
++;
858 } else if (rctx
->flags
& SI_CONTEXT_VS_PARTIAL_FLUSH
) {
859 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
860 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
861 rctx
->num_vs_flushes
++;
865 if (rctx
->flags
& SI_CONTEXT_CS_PARTIAL_FLUSH
&&
866 sctx
->compute_is_busy
) {
867 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
868 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
| EVENT_INDEX(4)));
869 rctx
->num_cs_flushes
++;
870 sctx
->compute_is_busy
= false;
873 /* VGT state synchronization. */
874 if (rctx
->flags
& SI_CONTEXT_VGT_FLUSH
) {
875 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
876 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
878 if (rctx
->flags
& SI_CONTEXT_VGT_STREAMOUT_SYNC
) {
879 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
880 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC
) | EVENT_INDEX(0));
883 /* GFX9: Wait for idle if we're flushing CB or DB. ACQUIRE_MEM doesn't
884 * wait for idle on GFX9. We have to use a TS event.
886 if (sctx
->b
.chip_class
>= GFX9
&& flush_cb_db
) {
887 struct r600_resource
*rbuf
= NULL
;
889 unsigned offset
= 0, tc_flags
, cb_db_event
;
891 /* Set the CB/DB flush event. */
892 switch (flush_cb_db
) {
893 case SI_CONTEXT_FLUSH_AND_INV_CB
:
894 cb_db_event
= V_028A90_FLUSH_AND_INV_CB_DATA_TS
;
896 case SI_CONTEXT_FLUSH_AND_INV_DB
:
897 cb_db_event
= V_028A90_FLUSH_AND_INV_DB_DATA_TS
;
901 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
904 /* TC | TC_WB = invalidate L2 data
905 * TC_MD | TC_WB = invalidate L2 metadata
906 * TC | TC_WB | TC_MD = invalidate L2 data & metadata
908 * The metadata cache must always be invalidated for coherency
909 * between CB/DB and shaders. (metadata = HTILE, CMASK, DCC)
911 * TC must be invalidated on GFX9 only if the CB/DB surface is
912 * not pipe-aligned. If the surface is RB-aligned, it might not
913 * strictly be pipe-aligned since RB alignment takes precendence.
915 tc_flags
= EVENT_TC_WB_ACTION_ENA
|
916 EVENT_TC_MD_ACTION_ENA
;
918 /* Ideally flush TC together with CB/DB. */
919 if (rctx
->flags
& SI_CONTEXT_INV_GLOBAL_L2
) {
920 tc_flags
|= EVENT_TC_ACTION_ENA
|
921 EVENT_TCL1_ACTION_ENA
;
923 /* Clear the flags. */
924 rctx
->flags
&= ~(SI_CONTEXT_INV_GLOBAL_L2
|
925 SI_CONTEXT_WRITEBACK_GLOBAL_L2
|
926 SI_CONTEXT_INV_VMEM_L1
);
929 /* Allocate memory for the fence. */
930 u_suballocator_alloc(rctx
->allocator_zeroed_memory
, 4, 4,
931 &offset
, (struct pipe_resource
**)&rbuf
);
932 va
= rbuf
->gpu_address
+ offset
;
934 r600_gfx_write_event_eop(rctx
, cb_db_event
, tc_flags
, 1,
936 r600_gfx_wait_fence(rctx
, va
, 1, 0xffffffff);
939 /* Make sure ME is idle (it executes most packets) before continuing.
940 * This prevents read-after-write hazards between PFP and ME.
943 (rctx
->flags
& (SI_CONTEXT_CS_PARTIAL_FLUSH
|
944 SI_CONTEXT_INV_VMEM_L1
|
945 SI_CONTEXT_INV_GLOBAL_L2
|
946 SI_CONTEXT_WRITEBACK_GLOBAL_L2
))) {
947 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
952 * When one of the CP_COHER_CNTL.DEST_BASE flags is set, SURFACE_SYNC
953 * waits for idle, so it should be last. SURFACE_SYNC is done in PFP.
955 * cp_coher_cntl should contain all necessary flags except TC flags
958 * SI-CIK don't support L2 write-back.
960 if (rctx
->flags
& SI_CONTEXT_INV_GLOBAL_L2
||
961 (rctx
->chip_class
<= CIK
&&
962 (rctx
->flags
& SI_CONTEXT_WRITEBACK_GLOBAL_L2
))) {
963 /* Invalidate L1 & L2. (L1 is always invalidated on SI)
964 * WB must be set on VI+ when TC_ACTION is set.
966 si_emit_surface_sync(rctx
, cp_coher_cntl
|
967 S_0085F0_TC_ACTION_ENA(1) |
968 S_0085F0_TCL1_ACTION_ENA(1) |
969 S_0301F0_TC_WB_ACTION_ENA(rctx
->chip_class
>= VI
));
971 sctx
->b
.num_L2_invalidates
++;
973 /* L1 invalidation and L2 writeback must be done separately,
974 * because both operations can't be done together.
976 if (rctx
->flags
& SI_CONTEXT_WRITEBACK_GLOBAL_L2
) {
978 * NC = apply to non-coherent MTYPEs
979 * (i.e. MTYPE <= 1, which is what we use everywhere)
981 * WB doesn't work without NC.
983 si_emit_surface_sync(rctx
, cp_coher_cntl
|
984 S_0301F0_TC_WB_ACTION_ENA(1) |
985 S_0301F0_TC_NC_ACTION_ENA(1));
987 sctx
->b
.num_L2_writebacks
++;
989 if (rctx
->flags
& SI_CONTEXT_INV_VMEM_L1
) {
990 /* Invalidate per-CU VMEM L1. */
991 si_emit_surface_sync(rctx
, cp_coher_cntl
|
992 S_0085F0_TCL1_ACTION_ENA(1));
997 /* If TC flushes haven't cleared this... */
999 si_emit_surface_sync(rctx
, cp_coher_cntl
);
1001 if (rctx
->flags
& R600_CONTEXT_START_PIPELINE_STATS
) {
1002 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1003 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
1005 } else if (rctx
->flags
& R600_CONTEXT_STOP_PIPELINE_STATS
) {
1006 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1007 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
1014 static void si_get_draw_start_count(struct si_context
*sctx
,
1015 const struct pipe_draw_info
*info
,
1016 unsigned *start
, unsigned *count
)
1018 if (info
->indirect
) {
1019 unsigned indirect_count
;
1020 struct pipe_transfer
*transfer
;
1021 unsigned begin
, end
;
1025 if (info
->indirect_params
) {
1026 data
= pipe_buffer_map_range(&sctx
->b
.b
,
1027 info
->indirect_params
,
1028 info
->indirect_params_offset
,
1030 PIPE_TRANSFER_READ
, &transfer
);
1032 indirect_count
= *data
;
1034 pipe_buffer_unmap(&sctx
->b
.b
, transfer
);
1036 indirect_count
= info
->indirect_count
;
1039 if (!indirect_count
) {
1040 *start
= *count
= 0;
1044 map_size
= (indirect_count
- 1) * info
->indirect_stride
+ 3 * sizeof(unsigned);
1045 data
= pipe_buffer_map_range(&sctx
->b
.b
, info
->indirect
,
1046 info
->indirect_offset
, map_size
,
1047 PIPE_TRANSFER_READ
, &transfer
);
1052 for (unsigned i
= 0; i
< indirect_count
; ++i
) {
1053 unsigned count
= data
[0];
1054 unsigned start
= data
[2];
1057 begin
= MIN2(begin
, start
);
1058 end
= MAX2(end
, start
+ count
);
1061 data
+= info
->indirect_stride
/ sizeof(unsigned);
1064 pipe_buffer_unmap(&sctx
->b
.b
, transfer
);
1068 *count
= end
- begin
;
1070 *start
= *count
= 0;
1073 *start
= info
->start
;
1074 *count
= info
->count
;
1078 void si_ce_pre_draw_synchronization(struct si_context
*sctx
)
1080 if (sctx
->ce_need_synchronization
) {
1081 radeon_emit(sctx
->ce_ib
, PKT3(PKT3_INCREMENT_CE_COUNTER
, 0, 0));
1082 radeon_emit(sctx
->ce_ib
, 1);
1084 radeon_emit(sctx
->b
.gfx
.cs
, PKT3(PKT3_WAIT_ON_CE_COUNTER
, 0, 0));
1085 radeon_emit(sctx
->b
.gfx
.cs
, 1);
1089 void si_ce_post_draw_synchronization(struct si_context
*sctx
)
1091 if (sctx
->ce_need_synchronization
) {
1092 radeon_emit(sctx
->b
.gfx
.cs
, PKT3(PKT3_INCREMENT_DE_COUNTER
, 0, 0));
1093 radeon_emit(sctx
->b
.gfx
.cs
, 0);
1095 sctx
->ce_need_synchronization
= false;
1099 void si_draw_vbo(struct pipe_context
*ctx
, const struct pipe_draw_info
*info
)
1101 struct si_context
*sctx
= (struct si_context
*)ctx
;
1102 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1103 const struct pipe_index_buffer
*ib
= &sctx
->index_buffer
;
1104 struct pipe_index_buffer ib_tmp
; /* for index buffer uploads only */
1105 unsigned mask
, dirty_tex_counter
;
1106 enum pipe_prim_type rast_prim
;
1107 unsigned num_patches
= 0;
1109 if (likely(!info
->indirect
)) {
1110 /* SI-CI treat instance_count==0 as instance_count==1. There is
1111 * no workaround for indirect draws, but we can at least skip
1114 if (unlikely(!info
->instance_count
))
1117 /* Handle count == 0. */
1118 if (unlikely(!info
->count
&&
1119 (info
->indexed
|| !info
->count_from_stream_output
)))
1123 if (unlikely(!sctx
->vs_shader
.cso
)) {
1127 if (unlikely(!sctx
->ps_shader
.cso
&& (!rs
|| !rs
->rasterizer_discard
))) {
1131 if (unlikely(!!sctx
->tes_shader
.cso
!= (info
->mode
== PIPE_PRIM_PATCHES
))) {
1136 /* Recompute and re-emit the texture resource states if needed. */
1137 dirty_tex_counter
= p_atomic_read(&sctx
->b
.screen
->dirty_tex_counter
);
1138 if (unlikely(dirty_tex_counter
!= sctx
->b
.last_dirty_tex_counter
)) {
1139 sctx
->b
.last_dirty_tex_counter
= dirty_tex_counter
;
1140 sctx
->framebuffer
.dirty_cbufs
|=
1141 ((1 << sctx
->framebuffer
.state
.nr_cbufs
) - 1);
1142 sctx
->framebuffer
.dirty_zsbuf
= true;
1143 sctx
->framebuffer
.do_update_surf_dirtiness
= true;
1144 si_mark_atom_dirty(sctx
, &sctx
->framebuffer
.atom
);
1145 si_update_all_texture_descriptors(sctx
);
1148 si_decompress_graphics_textures(sctx
);
1150 /* Set the rasterization primitive type.
1152 * This must be done after si_decompress_textures, which can call
1153 * draw_vbo recursively, and before si_update_shaders, which uses
1154 * current_rast_prim for this draw_vbo call. */
1155 if (sctx
->gs_shader
.cso
)
1156 rast_prim
= sctx
->gs_shader
.cso
->gs_output_prim
;
1157 else if (sctx
->tes_shader
.cso
)
1158 rast_prim
= sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
1160 rast_prim
= info
->mode
;
1162 if (rast_prim
!= sctx
->current_rast_prim
) {
1163 sctx
->current_rast_prim
= rast_prim
;
1164 sctx
->do_update_shaders
= true;
1167 if (sctx
->gs_shader
.cso
) {
1168 /* Determine whether the GS triangle strip adjacency fix should
1169 * be applied. Rotate every other triangle if
1170 * - triangle strips with adjacency are fed to the GS and
1171 * - primitive restart is disabled (the rotation doesn't help
1172 * when the restart occurs after an odd number of triangles).
1174 bool gs_tri_strip_adj_fix
=
1175 !sctx
->tes_shader
.cso
&&
1176 info
->mode
== PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
&&
1177 !info
->primitive_restart
;
1179 if (gs_tri_strip_adj_fix
!= sctx
->gs_tri_strip_adj_fix
) {
1180 sctx
->gs_tri_strip_adj_fix
= gs_tri_strip_adj_fix
;
1181 sctx
->do_update_shaders
= true;
1185 if (sctx
->do_update_shaders
&& !si_update_shaders(sctx
))
1188 if (!si_upload_graphics_shader_descriptors(sctx
))
1191 ib_tmp
.buffer
= NULL
;
1193 if (info
->indexed
) {
1194 /* Translate or upload, if needed. */
1195 /* 8-bit indices are supported on VI. */
1196 if (sctx
->b
.chip_class
<= CIK
&& ib
->index_size
== 1) {
1197 unsigned start
, count
, start_offset
, size
;
1200 si_get_draw_start_count(sctx
, info
, &start
, &count
);
1201 start_offset
= start
* 2;
1204 u_upload_alloc(ctx
->stream_uploader
, start_offset
,
1206 si_optimal_tcc_alignment(sctx
, size
),
1207 &ib_tmp
.offset
, &ib_tmp
.buffer
, &ptr
);
1211 util_shorten_ubyte_elts_to_userptr(&sctx
->b
.b
, ib
, 0, 0,
1215 /* info->start will be added by the drawing code */
1216 ib_tmp
.offset
-= start_offset
;
1217 ib_tmp
.index_size
= 2;
1219 } else if (ib
->user_buffer
&& !ib
->buffer
) {
1220 unsigned start
, count
, start_offset
;
1222 si_get_draw_start_count(sctx
, info
, &start
, &count
);
1223 start_offset
= start
* ib
->index_size
;
1225 u_upload_data(ctx
->stream_uploader
, start_offset
,
1226 count
* ib
->index_size
,
1227 sctx
->screen
->b
.info
.tcc_cache_line_size
,
1228 (char*)ib
->user_buffer
+ start_offset
,
1229 &ib_tmp
.offset
, &ib_tmp
.buffer
);
1233 /* info->start will be added by the drawing code */
1234 ib_tmp
.offset
-= start_offset
;
1235 ib_tmp
.index_size
= ib
->index_size
;
1237 } else if (sctx
->b
.chip_class
<= CIK
&&
1238 r600_resource(ib
->buffer
)->TC_L2_dirty
) {
1239 /* VI reads index buffers through TC L2, so it doesn't
1241 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1242 r600_resource(ib
->buffer
)->TC_L2_dirty
= false;
1246 if (info
->indirect
) {
1247 /* Add the buffer size for memory checking in need_cs_space. */
1248 r600_context_add_resource_size(ctx
, info
->indirect
);
1250 if (r600_resource(info
->indirect
)->TC_L2_dirty
) {
1251 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1252 r600_resource(info
->indirect
)->TC_L2_dirty
= false;
1255 if (info
->indirect_params
&&
1256 r600_resource(info
->indirect_params
)->TC_L2_dirty
) {
1257 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1258 r600_resource(info
->indirect_params
)->TC_L2_dirty
= false;
1262 si_need_cs_space(sctx
);
1264 /* Since we've called r600_context_add_resource_size for vertex buffers,
1265 * this must be called after si_need_cs_space, because we must let
1266 * need_cs_space flush before we add buffers to the buffer list.
1268 if (!si_upload_vertex_buffer_descriptors(sctx
))
1271 /* GFX9 scissor bug workaround. There is also a more efficient but
1272 * more involved alternative workaround. */
1273 if (sctx
->b
.chip_class
== GFX9
&&
1274 si_is_atom_dirty(sctx
, &sctx
->b
.scissors
.atom
))
1275 sctx
->b
.flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
;
1277 /* Flush caches before the first state atom, which does L2 prefetches. */
1279 si_emit_cache_flush(sctx
);
1281 /* Emit state atoms. */
1282 mask
= sctx
->dirty_atoms
;
1284 struct r600_atom
*atom
= sctx
->atoms
.array
[u_bit_scan(&mask
)];
1286 atom
->emit(&sctx
->b
, atom
);
1288 sctx
->dirty_atoms
= 0;
1291 mask
= sctx
->dirty_states
;
1293 unsigned i
= u_bit_scan(&mask
);
1294 struct si_pm4_state
*state
= sctx
->queued
.array
[i
];
1296 if (!state
|| sctx
->emitted
.array
[i
] == state
)
1299 si_pm4_emit(sctx
, state
);
1300 sctx
->emitted
.array
[i
] = state
;
1302 sctx
->dirty_states
= 0;
1304 si_emit_rasterizer_prim_state(sctx
);
1305 if (sctx
->tes_shader
.cso
)
1306 si_emit_derived_tess_state(sctx
, info
, &num_patches
);
1307 si_emit_vs_state(sctx
);
1308 si_emit_draw_registers(sctx
, info
, num_patches
);
1310 si_ce_pre_draw_synchronization(sctx
);
1311 si_emit_draw_packets(sctx
, info
, ib
);
1312 si_ce_post_draw_synchronization(sctx
);
1314 if (sctx
->trace_buf
)
1315 si_trace_emit(sctx
);
1317 /* Workaround for a VGT hang when streamout is enabled.
1318 * It must be done after drawing. */
1319 if ((sctx
->b
.family
== CHIP_HAWAII
||
1320 sctx
->b
.family
== CHIP_TONGA
||
1321 sctx
->b
.family
== CHIP_FIJI
) &&
1322 r600_get_strmout_en(&sctx
->b
)) {
1323 sctx
->b
.flags
|= SI_CONTEXT_VGT_STREAMOUT_SYNC
;
1326 if (sctx
->framebuffer
.do_update_surf_dirtiness
) {
1327 /* Set the depth buffer as dirty. */
1328 if (sctx
->framebuffer
.state
.zsbuf
) {
1329 struct pipe_surface
*surf
= sctx
->framebuffer
.state
.zsbuf
;
1330 struct r600_texture
*rtex
= (struct r600_texture
*)surf
->texture
;
1332 if (!rtex
->tc_compatible_htile
)
1333 rtex
->dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
1335 if (rtex
->surface
.flags
& RADEON_SURF_SBUFFER
)
1336 rtex
->stencil_dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
1338 if (sctx
->framebuffer
.compressed_cb_mask
) {
1339 struct pipe_surface
*surf
;
1340 struct r600_texture
*rtex
;
1341 unsigned mask
= sctx
->framebuffer
.compressed_cb_mask
;
1344 unsigned i
= u_bit_scan(&mask
);
1345 surf
= sctx
->framebuffer
.state
.cbufs
[i
];
1346 rtex
= (struct r600_texture
*)surf
->texture
;
1348 if (rtex
->fmask
.size
)
1349 rtex
->dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
1350 if (rtex
->dcc_gather_statistics
)
1351 rtex
->separate_dcc_dirty
= true;
1354 sctx
->framebuffer
.do_update_surf_dirtiness
= false;
1357 pipe_resource_reference(&ib_tmp
.buffer
, NULL
);
1358 sctx
->b
.num_draw_calls
++;
1359 if (G_0286E8_WAVESIZE(sctx
->spi_tmpring_size
))
1360 sctx
->b
.num_spill_draw_calls
++;
1363 void si_trace_emit(struct si_context
*sctx
)
1365 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
1368 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
, sctx
->trace_buf
,
1369 RADEON_USAGE_READWRITE
, RADEON_PRIO_TRACE
);
1370 radeon_emit(cs
, PKT3(PKT3_WRITE_DATA
, 3, 0));
1371 radeon_emit(cs
, S_370_DST_SEL(V_370_MEMORY_SYNC
) |
1372 S_370_WR_CONFIRM(1) |
1373 S_370_ENGINE_SEL(V_370_ME
));
1374 radeon_emit(cs
, sctx
->trace_buf
->gpu_address
);
1375 radeon_emit(cs
, sctx
->trace_buf
->gpu_address
>> 32);
1376 radeon_emit(cs
, sctx
->trace_id
);
1377 radeon_emit(cs
, PKT3(PKT3_NOP
, 0, 0));
1378 radeon_emit(cs
, AC_ENCODE_TRACE_POINT(sctx
->trace_id
));