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.
25 #include "radeon/r600_cs.h"
29 #include "util/u_index_modify.h"
30 #include "util/u_log.h"
31 #include "util/u_upload_mgr.h"
32 #include "util/u_prim.h"
36 static unsigned si_conv_pipe_prim(unsigned mode
)
38 static const unsigned prim_conv
[] = {
39 [PIPE_PRIM_POINTS
] = V_008958_DI_PT_POINTLIST
,
40 [PIPE_PRIM_LINES
] = V_008958_DI_PT_LINELIST
,
41 [PIPE_PRIM_LINE_LOOP
] = V_008958_DI_PT_LINELOOP
,
42 [PIPE_PRIM_LINE_STRIP
] = V_008958_DI_PT_LINESTRIP
,
43 [PIPE_PRIM_TRIANGLES
] = V_008958_DI_PT_TRILIST
,
44 [PIPE_PRIM_TRIANGLE_STRIP
] = V_008958_DI_PT_TRISTRIP
,
45 [PIPE_PRIM_TRIANGLE_FAN
] = V_008958_DI_PT_TRIFAN
,
46 [PIPE_PRIM_QUADS
] = V_008958_DI_PT_QUADLIST
,
47 [PIPE_PRIM_QUAD_STRIP
] = V_008958_DI_PT_QUADSTRIP
,
48 [PIPE_PRIM_POLYGON
] = V_008958_DI_PT_POLYGON
,
49 [PIPE_PRIM_LINES_ADJACENCY
] = V_008958_DI_PT_LINELIST_ADJ
,
50 [PIPE_PRIM_LINE_STRIP_ADJACENCY
] = V_008958_DI_PT_LINESTRIP_ADJ
,
51 [PIPE_PRIM_TRIANGLES_ADJACENCY
] = V_008958_DI_PT_TRILIST_ADJ
,
52 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
] = V_008958_DI_PT_TRISTRIP_ADJ
,
53 [PIPE_PRIM_PATCHES
] = V_008958_DI_PT_PATCH
,
54 [R600_PRIM_RECTANGLE_LIST
] = V_008958_DI_PT_RECTLIST
56 assert(mode
< ARRAY_SIZE(prim_conv
));
57 return prim_conv
[mode
];
60 static unsigned si_conv_prim_to_gs_out(unsigned mode
)
62 static const int prim_conv
[] = {
63 [PIPE_PRIM_POINTS
] = V_028A6C_OUTPRIM_TYPE_POINTLIST
,
64 [PIPE_PRIM_LINES
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
65 [PIPE_PRIM_LINE_LOOP
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
66 [PIPE_PRIM_LINE_STRIP
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
67 [PIPE_PRIM_TRIANGLES
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
68 [PIPE_PRIM_TRIANGLE_STRIP
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
69 [PIPE_PRIM_TRIANGLE_FAN
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
70 [PIPE_PRIM_QUADS
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
71 [PIPE_PRIM_QUAD_STRIP
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
72 [PIPE_PRIM_POLYGON
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
73 [PIPE_PRIM_LINES_ADJACENCY
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
74 [PIPE_PRIM_LINE_STRIP_ADJACENCY
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
75 [PIPE_PRIM_TRIANGLES_ADJACENCY
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
76 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
77 [PIPE_PRIM_PATCHES
] = V_028A6C_OUTPRIM_TYPE_POINTLIST
,
78 [R600_PRIM_RECTANGLE_LIST
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
80 assert(mode
< ARRAY_SIZE(prim_conv
));
82 return prim_conv
[mode
];
86 * This calculates the LDS size for tessellation shaders (VS, TCS, TES).
87 * LS.LDS_SIZE is shared by all 3 shader stages.
89 * The information about LDS and other non-compile-time parameters is then
90 * written to userdata SGPRs.
92 static void si_emit_derived_tess_state(struct si_context
*sctx
,
93 const struct pipe_draw_info
*info
,
94 unsigned *num_patches
)
96 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
97 struct si_shader
*ls_current
;
98 struct si_shader_selector
*ls
;
99 /* The TES pointer will only be used for sctx->last_tcs.
100 * It would be wrong to think that TCS = TES. */
101 struct si_shader_selector
*tcs
=
102 sctx
->tcs_shader
.cso
? sctx
->tcs_shader
.cso
: sctx
->tes_shader
.cso
;
103 unsigned tess_uses_primid
= sctx
->ia_multi_vgt_param_key
.u
.tess_uses_prim_id
;
104 bool has_primid_instancing_bug
= sctx
->b
.chip_class
== SI
&&
105 sctx
->b
.screen
->info
.max_se
== 1;
106 unsigned tes_sh_base
= sctx
->shader_pointers
.sh_base
[PIPE_SHADER_TESS_EVAL
];
107 unsigned num_tcs_input_cp
= info
->vertices_per_patch
;
108 unsigned num_tcs_output_cp
, num_tcs_inputs
, num_tcs_outputs
;
109 unsigned num_tcs_patch_outputs
;
110 unsigned input_vertex_size
, output_vertex_size
, pervertex_output_patch_size
;
111 unsigned input_patch_size
, output_patch_size
, output_patch0_offset
;
112 unsigned perpatch_output_offset
, lds_size
;
113 unsigned tcs_in_layout
, tcs_out_layout
, tcs_out_offsets
;
114 unsigned offchip_layout
, hardware_lds_size
, ls_hs_config
;
116 /* Since GFX9 has merged LS-HS in the TCS state, set LS = TCS. */
117 if (sctx
->b
.chip_class
>= GFX9
) {
118 if (sctx
->tcs_shader
.cso
)
119 ls_current
= sctx
->tcs_shader
.current
;
121 ls_current
= sctx
->fixed_func_tcs_shader
.current
;
123 ls
= ls_current
->key
.part
.tcs
.ls
;
125 ls_current
= sctx
->vs_shader
.current
;
126 ls
= sctx
->vs_shader
.cso
;
129 if (sctx
->last_ls
== ls_current
&&
130 sctx
->last_tcs
== tcs
&&
131 sctx
->last_tes_sh_base
== tes_sh_base
&&
132 sctx
->last_num_tcs_input_cp
== num_tcs_input_cp
&&
133 (!has_primid_instancing_bug
||
134 (sctx
->last_tess_uses_primid
== tess_uses_primid
))) {
135 *num_patches
= sctx
->last_num_patches
;
139 sctx
->last_ls
= ls_current
;
140 sctx
->last_tcs
= tcs
;
141 sctx
->last_tes_sh_base
= tes_sh_base
;
142 sctx
->last_num_tcs_input_cp
= num_tcs_input_cp
;
143 sctx
->last_tess_uses_primid
= tess_uses_primid
;
145 /* This calculates how shader inputs and outputs among VS, TCS, and TES
146 * are laid out in LDS. */
147 num_tcs_inputs
= util_last_bit64(ls
->outputs_written
);
149 if (sctx
->tcs_shader
.cso
) {
150 num_tcs_outputs
= util_last_bit64(tcs
->outputs_written
);
151 num_tcs_output_cp
= tcs
->info
.properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
];
152 num_tcs_patch_outputs
= util_last_bit64(tcs
->patch_outputs_written
);
154 /* No TCS. Route varyings from LS to TES. */
155 num_tcs_outputs
= num_tcs_inputs
;
156 num_tcs_output_cp
= num_tcs_input_cp
;
157 num_tcs_patch_outputs
= 2; /* TESSINNER + TESSOUTER */
160 input_vertex_size
= num_tcs_inputs
* 16;
161 output_vertex_size
= num_tcs_outputs
* 16;
163 input_patch_size
= num_tcs_input_cp
* input_vertex_size
;
165 pervertex_output_patch_size
= num_tcs_output_cp
* output_vertex_size
;
166 output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
168 /* Ensure that we only need one wave per SIMD so we don't need to check
169 * resource usage. Also ensures that the number of tcs in and out
170 * vertices per threadgroup are at most 256.
172 *num_patches
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
) * 4;
174 /* Make sure that the data fits in LDS. This assumes the shaders only
175 * use LDS for the inputs and outputs.
177 * While CIK can use 64K per threadgroup, there is a hang on Stoney
178 * with 2 CUs if we use more than 32K. The closed Vulkan driver also
179 * uses 32K at most on all GCN chips.
181 hardware_lds_size
= 32768;
182 *num_patches
= MIN2(*num_patches
, hardware_lds_size
/ (input_patch_size
+
185 /* Make sure the output data fits in the offchip buffer */
186 *num_patches
= MIN2(*num_patches
,
187 (sctx
->screen
->tess_offchip_block_dw_size
* 4) /
190 /* Not necessary for correctness, but improves performance. The
191 * specific value is taken from the proprietary driver.
193 *num_patches
= MIN2(*num_patches
, 40);
195 if (sctx
->b
.chip_class
== SI
) {
196 /* SI bug workaround, related to power management. Limit LS-HS
197 * threadgroups to only one wave.
199 unsigned one_wave
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
);
200 *num_patches
= MIN2(*num_patches
, one_wave
);
203 /* The VGT HS block increments the patch ID unconditionally
204 * within a single threadgroup. This results in incorrect
205 * patch IDs when instanced draws are used.
207 * The intended solution is to restrict threadgroups to
208 * a single instance by setting SWITCH_ON_EOI, which
209 * should cause IA to split instances up. However, this
210 * doesn't work correctly on SI when there is no other
213 if (has_primid_instancing_bug
&& tess_uses_primid
)
216 sctx
->last_num_patches
= *num_patches
;
218 output_patch0_offset
= input_patch_size
* *num_patches
;
219 perpatch_output_offset
= output_patch0_offset
+ pervertex_output_patch_size
;
221 /* Compute userdata SGPRs. */
222 assert(((input_vertex_size
/ 4) & ~0xff) == 0);
223 assert(((output_vertex_size
/ 4) & ~0xff) == 0);
224 assert(((input_patch_size
/ 4) & ~0x1fff) == 0);
225 assert(((output_patch_size
/ 4) & ~0x1fff) == 0);
226 assert(((output_patch0_offset
/ 16) & ~0xffff) == 0);
227 assert(((perpatch_output_offset
/ 16) & ~0xffff) == 0);
228 assert(num_tcs_input_cp
<= 32);
229 assert(num_tcs_output_cp
<= 32);
231 tcs_in_layout
= S_VS_STATE_LS_OUT_PATCH_SIZE(input_patch_size
/ 4) |
232 S_VS_STATE_LS_OUT_VERTEX_SIZE(input_vertex_size
/ 4);
233 tcs_out_layout
= output_patch_size
/ 4;
234 tcs_out_offsets
= (output_patch0_offset
/ 16) |
235 ((perpatch_output_offset
/ 16) << 16);
236 offchip_layout
= *num_patches
|
237 (num_tcs_output_cp
<< 6) |
238 (pervertex_output_patch_size
* *num_patches
<< 12);
240 /* Compute the LDS size. */
241 lds_size
= output_patch0_offset
+ output_patch_size
* *num_patches
;
243 if (sctx
->b
.chip_class
>= CIK
) {
244 assert(lds_size
<= 65536);
245 lds_size
= align(lds_size
, 512) / 512;
247 assert(lds_size
<= 32768);
248 lds_size
= align(lds_size
, 256) / 256;
251 /* Set SI_SGPR_VS_STATE_BITS. */
252 sctx
->current_vs_state
&= C_VS_STATE_LS_OUT_PATCH_SIZE
&
253 C_VS_STATE_LS_OUT_VERTEX_SIZE
;
254 sctx
->current_vs_state
|= tcs_in_layout
;
256 if (sctx
->b
.chip_class
>= GFX9
) {
257 unsigned hs_rsrc2
= ls_current
->config
.rsrc2
|
258 S_00B42C_LDS_SIZE(lds_size
);
260 radeon_set_sh_reg(cs
, R_00B42C_SPI_SHADER_PGM_RSRC2_HS
, hs_rsrc2
);
262 /* Set userdata SGPRs for merged LS-HS. */
263 radeon_set_sh_reg_seq(cs
,
264 R_00B430_SPI_SHADER_USER_DATA_LS_0
+
265 GFX9_SGPR_TCS_OFFCHIP_LAYOUT
* 4, 3);
266 radeon_emit(cs
, offchip_layout
);
267 radeon_emit(cs
, tcs_out_offsets
);
268 radeon_emit(cs
, tcs_out_layout
| (num_tcs_input_cp
<< 26));
270 unsigned ls_rsrc2
= ls_current
->config
.rsrc2
;
272 si_multiwave_lds_size_workaround(sctx
->screen
, &lds_size
);
273 ls_rsrc2
|= S_00B52C_LDS_SIZE(lds_size
);
275 /* Due to a hw bug, RSRC2_LS must be written twice with another
276 * LS register written in between. */
277 if (sctx
->b
.chip_class
== CIK
&& sctx
->b
.family
!= CHIP_HAWAII
)
278 radeon_set_sh_reg(cs
, R_00B52C_SPI_SHADER_PGM_RSRC2_LS
, ls_rsrc2
);
279 radeon_set_sh_reg_seq(cs
, R_00B528_SPI_SHADER_PGM_RSRC1_LS
, 2);
280 radeon_emit(cs
, ls_current
->config
.rsrc1
);
281 radeon_emit(cs
, ls_rsrc2
);
283 /* Set userdata SGPRs for TCS. */
284 radeon_set_sh_reg_seq(cs
,
285 R_00B430_SPI_SHADER_USER_DATA_HS_0
+ GFX6_SGPR_TCS_OFFCHIP_LAYOUT
* 4, 4);
286 radeon_emit(cs
, offchip_layout
);
287 radeon_emit(cs
, tcs_out_offsets
);
288 radeon_emit(cs
, tcs_out_layout
| (num_tcs_input_cp
<< 26));
289 radeon_emit(cs
, tcs_in_layout
);
292 /* Set userdata SGPRs for TES. */
293 radeon_set_sh_reg_seq(cs
, tes_sh_base
+ SI_SGPR_TES_OFFCHIP_LAYOUT
* 4, 2);
294 radeon_emit(cs
, offchip_layout
);
295 radeon_emit(cs
, r600_resource(sctx
->tess_offchip_ring
)->gpu_address
>> 16);
297 ls_hs_config
= S_028B58_NUM_PATCHES(*num_patches
) |
298 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp
) |
299 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp
);
301 if (sctx
->b
.chip_class
>= CIK
)
302 radeon_set_context_reg_idx(cs
, R_028B58_VGT_LS_HS_CONFIG
, 2,
305 radeon_set_context_reg(cs
, R_028B58_VGT_LS_HS_CONFIG
,
309 static unsigned si_num_prims_for_vertices(const struct pipe_draw_info
*info
)
311 switch (info
->mode
) {
312 case PIPE_PRIM_PATCHES
:
313 return info
->count
/ info
->vertices_per_patch
;
314 case R600_PRIM_RECTANGLE_LIST
:
315 return info
->count
/ 3;
317 return u_prims_for_vertices(info
->mode
, info
->count
);
322 si_get_init_multi_vgt_param(struct si_screen
*sscreen
,
323 union si_vgt_param_key
*key
)
325 STATIC_ASSERT(sizeof(union si_vgt_param_key
) == 4);
326 unsigned max_primgroup_in_wave
= 2;
328 /* SWITCH_ON_EOP(0) is always preferable. */
329 bool wd_switch_on_eop
= false;
330 bool ia_switch_on_eop
= false;
331 bool ia_switch_on_eoi
= false;
332 bool partial_vs_wave
= false;
333 bool partial_es_wave
= false;
335 if (key
->u
.uses_tess
) {
336 /* SWITCH_ON_EOI must be set if PrimID is used. */
337 if (key
->u
.tess_uses_prim_id
)
338 ia_switch_on_eoi
= true;
340 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
341 if ((sscreen
->b
.family
== CHIP_TAHITI
||
342 sscreen
->b
.family
== CHIP_PITCAIRN
||
343 sscreen
->b
.family
== CHIP_BONAIRE
) &&
345 partial_vs_wave
= true;
347 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
348 if (sscreen
->has_distributed_tess
) {
349 if (key
->u
.uses_gs
) {
350 if (sscreen
->b
.chip_class
<= VI
)
351 partial_es_wave
= true;
353 /* GPU hang workaround. */
354 if (sscreen
->b
.family
== CHIP_TONGA
||
355 sscreen
->b
.family
== CHIP_FIJI
||
356 sscreen
->b
.family
== CHIP_POLARIS10
||
357 sscreen
->b
.family
== CHIP_POLARIS11
||
358 sscreen
->b
.family
== CHIP_POLARIS12
)
359 partial_vs_wave
= true;
361 partial_vs_wave
= true;
366 /* This is a hardware requirement. */
367 if (key
->u
.line_stipple_enabled
||
368 (sscreen
->b
.debug_flags
& DBG(SWITCH_ON_EOP
))) {
369 ia_switch_on_eop
= true;
370 wd_switch_on_eop
= true;
373 if (sscreen
->b
.chip_class
>= CIK
) {
374 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
375 * 4 shader engines. Set 1 to pass the assertion below.
376 * The other cases are hardware requirements.
378 * Polaris supports primitive restart with WD_SWITCH_ON_EOP=0
379 * for points, line strips, and tri strips.
381 if (sscreen
->b
.info
.max_se
< 4 ||
382 key
->u
.prim
== PIPE_PRIM_POLYGON
||
383 key
->u
.prim
== PIPE_PRIM_LINE_LOOP
||
384 key
->u
.prim
== PIPE_PRIM_TRIANGLE_FAN
||
385 key
->u
.prim
== PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
||
386 (key
->u
.primitive_restart
&&
387 (sscreen
->b
.family
< CHIP_POLARIS10
||
388 (key
->u
.prim
!= PIPE_PRIM_POINTS
&&
389 key
->u
.prim
!= PIPE_PRIM_LINE_STRIP
&&
390 key
->u
.prim
!= PIPE_PRIM_TRIANGLE_STRIP
))) ||
391 key
->u
.count_from_stream_output
)
392 wd_switch_on_eop
= true;
394 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
395 * We don't know that for indirect drawing, so treat it as
396 * always problematic. */
397 if (sscreen
->b
.family
== CHIP_HAWAII
&&
398 key
->u
.uses_instancing
)
399 wd_switch_on_eop
= true;
401 /* Performance recommendation for 4 SE Gfx7-8 parts if
402 * instances are smaller than a primgroup.
403 * Assume indirect draws always use small instances.
404 * This is needed for good VS wave utilization.
406 if (sscreen
->b
.chip_class
<= VI
&&
407 sscreen
->b
.info
.max_se
== 4 &&
408 key
->u
.multi_instances_smaller_than_primgroup
)
409 wd_switch_on_eop
= true;
411 /* Required on CIK and later. */
412 if (sscreen
->b
.info
.max_se
> 2 && !wd_switch_on_eop
)
413 ia_switch_on_eoi
= true;
415 /* Required by Hawaii and, for some special cases, by VI. */
416 if (ia_switch_on_eoi
&&
417 (sscreen
->b
.family
== CHIP_HAWAII
||
418 (sscreen
->b
.chip_class
== VI
&&
419 (key
->u
.uses_gs
|| max_primgroup_in_wave
!= 2))))
420 partial_vs_wave
= true;
422 /* Instancing bug on Bonaire. */
423 if (sscreen
->b
.family
== CHIP_BONAIRE
&& ia_switch_on_eoi
&&
424 key
->u
.uses_instancing
)
425 partial_vs_wave
= true;
427 /* If the WD switch is false, the IA switch must be false too. */
428 assert(wd_switch_on_eop
|| !ia_switch_on_eop
);
431 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
432 if (sscreen
->b
.chip_class
<= VI
&& ia_switch_on_eoi
)
433 partial_es_wave
= true;
435 return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop
) |
436 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi
) |
437 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave
) |
438 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave
) |
439 S_028AA8_WD_SWITCH_ON_EOP(sscreen
->b
.chip_class
>= CIK
? wd_switch_on_eop
: 0) |
440 /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
441 S_028AA8_MAX_PRIMGRP_IN_WAVE(sscreen
->b
.chip_class
== VI
?
442 max_primgroup_in_wave
: 0) |
443 S_030960_EN_INST_OPT_BASIC(sscreen
->b
.chip_class
>= GFX9
) |
444 S_030960_EN_INST_OPT_ADV(sscreen
->b
.chip_class
>= GFX9
);
447 void si_init_ia_multi_vgt_param_table(struct si_context
*sctx
)
449 for (int prim
= 0; prim
<= R600_PRIM_RECTANGLE_LIST
; prim
++)
450 for (int uses_instancing
= 0; uses_instancing
< 2; uses_instancing
++)
451 for (int multi_instances
= 0; multi_instances
< 2; multi_instances
++)
452 for (int primitive_restart
= 0; primitive_restart
< 2; primitive_restart
++)
453 for (int count_from_so
= 0; count_from_so
< 2; count_from_so
++)
454 for (int line_stipple
= 0; line_stipple
< 2; line_stipple
++)
455 for (int uses_tess
= 0; uses_tess
< 2; uses_tess
++)
456 for (int tess_uses_primid
= 0; tess_uses_primid
< 2; tess_uses_primid
++)
457 for (int uses_gs
= 0; uses_gs
< 2; uses_gs
++) {
458 union si_vgt_param_key key
;
462 key
.u
.uses_instancing
= uses_instancing
;
463 key
.u
.multi_instances_smaller_than_primgroup
= multi_instances
;
464 key
.u
.primitive_restart
= primitive_restart
;
465 key
.u
.count_from_stream_output
= count_from_so
;
466 key
.u
.line_stipple_enabled
= line_stipple
;
467 key
.u
.uses_tess
= uses_tess
;
468 key
.u
.tess_uses_prim_id
= tess_uses_primid
;
469 key
.u
.uses_gs
= uses_gs
;
471 sctx
->ia_multi_vgt_param
[key
.index
] =
472 si_get_init_multi_vgt_param(sctx
->screen
, &key
);
476 static unsigned si_get_ia_multi_vgt_param(struct si_context
*sctx
,
477 const struct pipe_draw_info
*info
,
478 unsigned num_patches
)
480 union si_vgt_param_key key
= sctx
->ia_multi_vgt_param_key
;
481 unsigned primgroup_size
;
482 unsigned ia_multi_vgt_param
;
484 if (sctx
->tes_shader
.cso
) {
485 primgroup_size
= num_patches
; /* must be a multiple of NUM_PATCHES */
486 } else if (sctx
->gs_shader
.cso
) {
487 primgroup_size
= 64; /* recommended with a GS */
489 primgroup_size
= 128; /* recommended without a GS and tess */
492 key
.u
.prim
= info
->mode
;
493 key
.u
.uses_instancing
= info
->indirect
|| info
->instance_count
> 1;
494 key
.u
.multi_instances_smaller_than_primgroup
=
496 (info
->instance_count
> 1 &&
497 (info
->count_from_stream_output
||
498 si_num_prims_for_vertices(info
) < primgroup_size
));
499 key
.u
.primitive_restart
= info
->primitive_restart
;
500 key
.u
.count_from_stream_output
= info
->count_from_stream_output
!= NULL
;
502 ia_multi_vgt_param
= sctx
->ia_multi_vgt_param
[key
.index
] |
503 S_028AA8_PRIMGROUP_SIZE(primgroup_size
- 1);
505 if (sctx
->gs_shader
.cso
) {
506 /* GS requirement. */
507 if (sctx
->b
.chip_class
<= VI
&&
508 SI_GS_PER_ES
/ primgroup_size
>= sctx
->screen
->gs_table_depth
- 3)
509 ia_multi_vgt_param
|= S_028AA8_PARTIAL_ES_WAVE_ON(1);
511 /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
512 * The hw doc says all multi-SE chips are affected, but Vulkan
513 * only applies it to Hawaii. Do what Vulkan does.
515 if (sctx
->b
.family
== CHIP_HAWAII
&&
516 G_028AA8_SWITCH_ON_EOI(ia_multi_vgt_param
) &&
518 (info
->instance_count
> 1 &&
519 (info
->count_from_stream_output
||
520 si_num_prims_for_vertices(info
) <= 1))))
521 sctx
->b
.flags
|= SI_CONTEXT_VGT_FLUSH
;
524 return ia_multi_vgt_param
;
527 /* rast_prim is the primitive type after GS. */
528 static void si_emit_rasterizer_prim_state(struct si_context
*sctx
)
530 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
531 enum pipe_prim_type rast_prim
= sctx
->current_rast_prim
;
532 struct si_state_rasterizer
*rs
= sctx
->emitted
.named
.rasterizer
;
534 /* Skip this if not rendering lines. */
535 if (rast_prim
!= PIPE_PRIM_LINES
&&
536 rast_prim
!= PIPE_PRIM_LINE_LOOP
&&
537 rast_prim
!= PIPE_PRIM_LINE_STRIP
&&
538 rast_prim
!= PIPE_PRIM_LINES_ADJACENCY
&&
539 rast_prim
!= PIPE_PRIM_LINE_STRIP_ADJACENCY
)
542 if (rast_prim
== sctx
->last_rast_prim
&&
543 rs
->pa_sc_line_stipple
== sctx
->last_sc_line_stipple
)
546 /* For lines, reset the stipple pattern at each primitive. Otherwise,
547 * reset the stipple pattern at each packet (line strips, line loops).
549 radeon_set_context_reg(cs
, R_028A0C_PA_SC_LINE_STIPPLE
,
550 rs
->pa_sc_line_stipple
|
551 S_028A0C_AUTO_RESET_CNTL(rast_prim
== PIPE_PRIM_LINES
? 1 : 2));
553 sctx
->last_rast_prim
= rast_prim
;
554 sctx
->last_sc_line_stipple
= rs
->pa_sc_line_stipple
;
557 static void si_emit_vs_state(struct si_context
*sctx
,
558 const struct pipe_draw_info
*info
)
560 sctx
->current_vs_state
&= C_VS_STATE_INDEXED
;
561 sctx
->current_vs_state
|= S_VS_STATE_INDEXED(!!info
->index_size
);
563 if (sctx
->num_vs_blit_sgprs
) {
564 /* Re-emit the state after we leave u_blitter. */
565 sctx
->last_vs_state
= ~0;
569 if (sctx
->current_vs_state
!= sctx
->last_vs_state
) {
570 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
572 radeon_set_sh_reg(cs
,
573 sctx
->shader_pointers
.sh_base
[PIPE_SHADER_VERTEX
] +
574 SI_SGPR_VS_STATE_BITS
* 4,
575 sctx
->current_vs_state
);
577 sctx
->last_vs_state
= sctx
->current_vs_state
;
581 static void si_emit_draw_registers(struct si_context
*sctx
,
582 const struct pipe_draw_info
*info
,
583 unsigned num_patches
)
585 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
586 unsigned prim
= si_conv_pipe_prim(info
->mode
);
587 unsigned gs_out_prim
= si_conv_prim_to_gs_out(sctx
->current_rast_prim
);
588 unsigned ia_multi_vgt_param
;
590 ia_multi_vgt_param
= si_get_ia_multi_vgt_param(sctx
, info
, num_patches
);
593 if (ia_multi_vgt_param
!= sctx
->last_multi_vgt_param
) {
594 if (sctx
->b
.chip_class
>= GFX9
)
595 radeon_set_uconfig_reg_idx(cs
, R_030960_IA_MULTI_VGT_PARAM
, 4, ia_multi_vgt_param
);
596 else if (sctx
->b
.chip_class
>= CIK
)
597 radeon_set_context_reg_idx(cs
, R_028AA8_IA_MULTI_VGT_PARAM
, 1, ia_multi_vgt_param
);
599 radeon_set_context_reg(cs
, R_028AA8_IA_MULTI_VGT_PARAM
, ia_multi_vgt_param
);
601 sctx
->last_multi_vgt_param
= ia_multi_vgt_param
;
603 if (prim
!= sctx
->last_prim
) {
604 if (sctx
->b
.chip_class
>= CIK
)
605 radeon_set_uconfig_reg_idx(cs
, R_030908_VGT_PRIMITIVE_TYPE
, 1, prim
);
607 radeon_set_config_reg(cs
, R_008958_VGT_PRIMITIVE_TYPE
, prim
);
609 sctx
->last_prim
= prim
;
612 if (gs_out_prim
!= sctx
->last_gs_out_prim
) {
613 radeon_set_context_reg(cs
, R_028A6C_VGT_GS_OUT_PRIM_TYPE
, gs_out_prim
);
614 sctx
->last_gs_out_prim
= gs_out_prim
;
617 /* Primitive restart. */
618 if (info
->primitive_restart
!= sctx
->last_primitive_restart_en
) {
619 if (sctx
->b
.chip_class
>= GFX9
)
620 radeon_set_uconfig_reg(cs
, R_03092C_VGT_MULTI_PRIM_IB_RESET_EN
,
621 info
->primitive_restart
);
623 radeon_set_context_reg(cs
, R_028A94_VGT_MULTI_PRIM_IB_RESET_EN
,
624 info
->primitive_restart
);
626 sctx
->last_primitive_restart_en
= info
->primitive_restart
;
629 if (info
->primitive_restart
&&
630 (info
->restart_index
!= sctx
->last_restart_index
||
631 sctx
->last_restart_index
== SI_RESTART_INDEX_UNKNOWN
)) {
632 radeon_set_context_reg(cs
, R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX
,
633 info
->restart_index
);
634 sctx
->last_restart_index
= info
->restart_index
;
638 static void si_emit_draw_packets(struct si_context
*sctx
,
639 const struct pipe_draw_info
*info
,
640 struct pipe_resource
*indexbuf
,
642 unsigned index_offset
)
644 struct pipe_draw_indirect_info
*indirect
= info
->indirect
;
645 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
646 unsigned sh_base_reg
= sctx
->shader_pointers
.sh_base
[PIPE_SHADER_VERTEX
];
647 bool render_cond_bit
= sctx
->b
.render_cond
&& !sctx
->b
.render_cond_force_off
;
648 uint32_t index_max_size
= 0;
649 uint64_t index_va
= 0;
651 if (info
->count_from_stream_output
) {
652 struct si_streamout_target
*t
=
653 (struct si_streamout_target
*)info
->count_from_stream_output
;
654 uint64_t va
= t
->buf_filled_size
->gpu_address
+
655 t
->buf_filled_size_offset
;
657 radeon_set_context_reg(cs
, R_028B30_VGT_STRMOUT_DRAW_OPAQUE_VERTEX_STRIDE
,
660 radeon_emit(cs
, PKT3(PKT3_COPY_DATA
, 4, 0));
661 radeon_emit(cs
, COPY_DATA_SRC_SEL(COPY_DATA_MEM
) |
662 COPY_DATA_DST_SEL(COPY_DATA_REG
) |
663 COPY_DATA_WR_CONFIRM
);
664 radeon_emit(cs
, va
); /* src address lo */
665 radeon_emit(cs
, va
>> 32); /* src address hi */
666 radeon_emit(cs
, R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE
>> 2);
667 radeon_emit(cs
, 0); /* unused */
669 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
670 t
->buf_filled_size
, RADEON_USAGE_READ
,
671 RADEON_PRIO_SO_FILLED_SIZE
);
676 if (index_size
!= sctx
->last_index_size
) {
680 switch (index_size
) {
682 index_type
= V_028A7C_VGT_INDEX_8
;
685 index_type
= V_028A7C_VGT_INDEX_16
|
686 (SI_BIG_ENDIAN
&& sctx
->b
.chip_class
<= CIK
?
687 V_028A7C_VGT_DMA_SWAP_16_BIT
: 0);
690 index_type
= V_028A7C_VGT_INDEX_32
|
691 (SI_BIG_ENDIAN
&& sctx
->b
.chip_class
<= CIK
?
692 V_028A7C_VGT_DMA_SWAP_32_BIT
: 0);
695 assert(!"unreachable");
699 if (sctx
->b
.chip_class
>= GFX9
) {
700 radeon_set_uconfig_reg_idx(cs
, R_03090C_VGT_INDEX_TYPE
,
703 radeon_emit(cs
, PKT3(PKT3_INDEX_TYPE
, 0, 0));
704 radeon_emit(cs
, index_type
);
707 sctx
->last_index_size
= index_size
;
710 index_max_size
= (indexbuf
->width0
- index_offset
) /
712 index_va
= r600_resource(indexbuf
)->gpu_address
+ index_offset
;
714 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
715 (struct r600_resource
*)indexbuf
,
716 RADEON_USAGE_READ
, RADEON_PRIO_INDEX_BUFFER
);
718 /* On CI and later, non-indexed draws overwrite VGT_INDEX_TYPE,
719 * so the state must be re-emitted before the next indexed draw.
721 if (sctx
->b
.chip_class
>= CIK
)
722 sctx
->last_index_size
= -1;
726 uint64_t indirect_va
= r600_resource(indirect
->buffer
)->gpu_address
;
728 assert(indirect_va
% 8 == 0);
730 si_invalidate_draw_sh_constants(sctx
);
732 radeon_emit(cs
, PKT3(PKT3_SET_BASE
, 2, 0));
734 radeon_emit(cs
, indirect_va
);
735 radeon_emit(cs
, indirect_va
>> 32);
737 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
738 (struct r600_resource
*)indirect
->buffer
,
739 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
741 unsigned di_src_sel
= index_size
? V_0287F0_DI_SRC_SEL_DMA
742 : V_0287F0_DI_SRC_SEL_AUTO_INDEX
;
744 assert(indirect
->offset
% 4 == 0);
747 radeon_emit(cs
, PKT3(PKT3_INDEX_BASE
, 1, 0));
748 radeon_emit(cs
, index_va
);
749 radeon_emit(cs
, index_va
>> 32);
751 radeon_emit(cs
, PKT3(PKT3_INDEX_BUFFER_SIZE
, 0, 0));
752 radeon_emit(cs
, index_max_size
);
755 if (!sctx
->screen
->has_draw_indirect_multi
) {
756 radeon_emit(cs
, PKT3(index_size
? PKT3_DRAW_INDEX_INDIRECT
757 : PKT3_DRAW_INDIRECT
,
758 3, render_cond_bit
));
759 radeon_emit(cs
, indirect
->offset
);
760 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
761 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
762 radeon_emit(cs
, di_src_sel
);
764 uint64_t count_va
= 0;
766 if (indirect
->indirect_draw_count
) {
767 struct r600_resource
*params_buf
=
768 (struct r600_resource
*)indirect
->indirect_draw_count
;
770 radeon_add_to_buffer_list(
771 &sctx
->b
, &sctx
->b
.gfx
, params_buf
,
772 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
774 count_va
= params_buf
->gpu_address
+ indirect
->indirect_draw_count_offset
;
777 radeon_emit(cs
, PKT3(index_size
? PKT3_DRAW_INDEX_INDIRECT_MULTI
:
778 PKT3_DRAW_INDIRECT_MULTI
,
779 8, render_cond_bit
));
780 radeon_emit(cs
, indirect
->offset
);
781 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
782 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
783 radeon_emit(cs
, ((sh_base_reg
+ SI_SGPR_DRAWID
* 4 - SI_SH_REG_OFFSET
) >> 2) |
784 S_2C3_DRAW_INDEX_ENABLE(1) |
785 S_2C3_COUNT_INDIRECT_ENABLE(!!indirect
->indirect_draw_count
));
786 radeon_emit(cs
, indirect
->draw_count
);
787 radeon_emit(cs
, count_va
);
788 radeon_emit(cs
, count_va
>> 32);
789 radeon_emit(cs
, indirect
->stride
);
790 radeon_emit(cs
, di_src_sel
);
795 radeon_emit(cs
, PKT3(PKT3_NUM_INSTANCES
, 0, 0));
796 radeon_emit(cs
, info
->instance_count
);
798 /* Base vertex and start instance. */
799 base_vertex
= index_size
? info
->index_bias
: info
->start
;
801 if (sctx
->num_vs_blit_sgprs
) {
802 /* Re-emit draw constants after we leave u_blitter. */
803 si_invalidate_draw_sh_constants(sctx
);
805 /* Blit VS doesn't use BASE_VERTEX, START_INSTANCE, and DRAWID. */
806 radeon_set_sh_reg_seq(cs
, sh_base_reg
+ SI_SGPR_VS_BLIT_DATA
* 4,
807 sctx
->num_vs_blit_sgprs
);
808 radeon_emit_array(cs
, sctx
->vs_blit_sh_data
,
809 sctx
->num_vs_blit_sgprs
);
810 } else if (base_vertex
!= sctx
->last_base_vertex
||
811 sctx
->last_base_vertex
== SI_BASE_VERTEX_UNKNOWN
||
812 info
->start_instance
!= sctx
->last_start_instance
||
813 info
->drawid
!= sctx
->last_drawid
||
814 sh_base_reg
!= sctx
->last_sh_base_reg
) {
815 radeon_set_sh_reg_seq(cs
, sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4, 3);
816 radeon_emit(cs
, base_vertex
);
817 radeon_emit(cs
, info
->start_instance
);
818 radeon_emit(cs
, info
->drawid
);
820 sctx
->last_base_vertex
= base_vertex
;
821 sctx
->last_start_instance
= info
->start_instance
;
822 sctx
->last_drawid
= info
->drawid
;
823 sctx
->last_sh_base_reg
= sh_base_reg
;
827 index_va
+= info
->start
* index_size
;
829 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_2
, 4, render_cond_bit
));
830 radeon_emit(cs
, index_max_size
);
831 radeon_emit(cs
, index_va
);
832 radeon_emit(cs
, index_va
>> 32);
833 radeon_emit(cs
, info
->count
);
834 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_DMA
);
836 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_AUTO
, 1, render_cond_bit
));
837 radeon_emit(cs
, info
->count
);
838 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_AUTO_INDEX
|
839 S_0287F0_USE_OPAQUE(!!info
->count_from_stream_output
));
844 static void si_emit_surface_sync(struct r600_common_context
*rctx
,
845 unsigned cp_coher_cntl
)
847 struct radeon_winsys_cs
*cs
= rctx
->gfx
.cs
;
849 if (rctx
->chip_class
>= GFX9
) {
850 /* Flush caches and wait for the caches to assert idle. */
851 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 5, 0));
852 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
853 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
854 radeon_emit(cs
, 0xffffff); /* CP_COHER_SIZE_HI */
855 radeon_emit(cs
, 0); /* CP_COHER_BASE */
856 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
857 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
859 /* ACQUIRE_MEM is only required on a compute ring. */
860 radeon_emit(cs
, PKT3(PKT3_SURFACE_SYNC
, 3, 0));
861 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
862 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
863 radeon_emit(cs
, 0); /* CP_COHER_BASE */
864 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
868 void si_emit_cache_flush(struct si_context
*sctx
)
870 struct r600_common_context
*rctx
= &sctx
->b
;
871 struct radeon_winsys_cs
*cs
= rctx
->gfx
.cs
;
872 uint32_t cp_coher_cntl
= 0;
873 uint32_t flush_cb_db
= rctx
->flags
& (SI_CONTEXT_FLUSH_AND_INV_CB
|
874 SI_CONTEXT_FLUSH_AND_INV_DB
);
876 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_CB
)
877 sctx
->b
.num_cb_cache_flushes
++;
878 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_DB
)
879 sctx
->b
.num_db_cache_flushes
++;
881 /* SI has a bug that it always flushes ICACHE and KCACHE if either
882 * bit is set. An alternative way is to write SQC_CACHES, but that
883 * doesn't seem to work reliably. Since the bug doesn't affect
884 * correctness (it only does more work than necessary) and
885 * the performance impact is likely negligible, there is no plan
886 * to add a workaround for it.
889 if (rctx
->flags
& SI_CONTEXT_INV_ICACHE
)
890 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
891 if (rctx
->flags
& SI_CONTEXT_INV_SMEM_L1
)
892 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
894 if (rctx
->chip_class
<= VI
) {
895 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_CB
) {
896 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
897 S_0085F0_CB0_DEST_BASE_ENA(1) |
898 S_0085F0_CB1_DEST_BASE_ENA(1) |
899 S_0085F0_CB2_DEST_BASE_ENA(1) |
900 S_0085F0_CB3_DEST_BASE_ENA(1) |
901 S_0085F0_CB4_DEST_BASE_ENA(1) |
902 S_0085F0_CB5_DEST_BASE_ENA(1) |
903 S_0085F0_CB6_DEST_BASE_ENA(1) |
904 S_0085F0_CB7_DEST_BASE_ENA(1);
906 /* Necessary for DCC */
907 if (rctx
->chip_class
== VI
)
908 si_gfx_write_event_eop(rctx
, V_028A90_FLUSH_AND_INV_CB_DATA_TS
,
909 0, EOP_DATA_SEL_DISCARD
, NULL
,
910 0, 0, R600_NOT_QUERY
);
912 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_DB
)
913 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
914 S_0085F0_DB_DEST_BASE_ENA(1);
917 if (rctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_CB
) {
918 /* Flush CMASK/FMASK/DCC. SURFACE_SYNC will wait for idle. */
919 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
920 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
922 if (rctx
->flags
& (SI_CONTEXT_FLUSH_AND_INV_DB
|
923 SI_CONTEXT_FLUSH_AND_INV_DB_META
)) {
924 /* Flush HTILE. SURFACE_SYNC will wait for idle. */
925 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
926 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
929 /* Wait for shader engines to go idle.
930 * VS and PS waits are unnecessary if SURFACE_SYNC is going to wait
931 * for everything including CB/DB cache flushes.
934 if (rctx
->flags
& SI_CONTEXT_PS_PARTIAL_FLUSH
) {
935 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
936 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
937 /* Only count explicit shader flushes, not implicit ones
938 * done by SURFACE_SYNC.
940 rctx
->num_vs_flushes
++;
941 rctx
->num_ps_flushes
++;
942 } else if (rctx
->flags
& SI_CONTEXT_VS_PARTIAL_FLUSH
) {
943 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
944 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
945 rctx
->num_vs_flushes
++;
949 if (rctx
->flags
& SI_CONTEXT_CS_PARTIAL_FLUSH
&&
950 sctx
->compute_is_busy
) {
951 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
952 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
| EVENT_INDEX(4)));
953 rctx
->num_cs_flushes
++;
954 sctx
->compute_is_busy
= false;
957 /* VGT state synchronization. */
958 if (rctx
->flags
& SI_CONTEXT_VGT_FLUSH
) {
959 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
960 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
962 if (rctx
->flags
& SI_CONTEXT_VGT_STREAMOUT_SYNC
) {
963 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
964 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC
) | EVENT_INDEX(0));
967 /* GFX9: Wait for idle if we're flushing CB or DB. ACQUIRE_MEM doesn't
968 * wait for idle on GFX9. We have to use a TS event.
970 if (sctx
->b
.chip_class
>= GFX9
&& flush_cb_db
) {
972 unsigned tc_flags
, cb_db_event
;
974 /* Set the CB/DB flush event. */
975 switch (flush_cb_db
) {
976 case SI_CONTEXT_FLUSH_AND_INV_CB
:
977 cb_db_event
= V_028A90_FLUSH_AND_INV_CB_DATA_TS
;
979 case SI_CONTEXT_FLUSH_AND_INV_DB
:
980 cb_db_event
= V_028A90_FLUSH_AND_INV_DB_DATA_TS
;
984 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
987 /* These are the only allowed combinations. If you need to
988 * do multiple operations at once, do them separately.
989 * All operations that invalidate L2 also seem to invalidate
990 * metadata. Volatile (VOL) and WC flushes are not listed here.
992 * TC | TC_WB = writeback & invalidate L2 & L1
993 * TC | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
994 * TC_WB | TC_NC = writeback L2 for MTYPE == NC
995 * TC | TC_NC = invalidate L2 for MTYPE == NC
996 * TC | TC_MD = writeback & invalidate L2 metadata (DCC, etc.)
997 * TCL1 = invalidate L1
1001 if (rctx
->flags
& SI_CONTEXT_INV_L2_METADATA
) {
1002 tc_flags
= EVENT_TC_ACTION_ENA
|
1003 EVENT_TC_MD_ACTION_ENA
;
1006 /* Ideally flush TC together with CB/DB. */
1007 if (rctx
->flags
& SI_CONTEXT_INV_GLOBAL_L2
) {
1008 /* Writeback and invalidate everything in L2 & L1. */
1009 tc_flags
= EVENT_TC_ACTION_ENA
|
1010 EVENT_TC_WB_ACTION_ENA
;
1012 /* Clear the flags. */
1013 rctx
->flags
&= ~(SI_CONTEXT_INV_GLOBAL_L2
|
1014 SI_CONTEXT_WRITEBACK_GLOBAL_L2
|
1015 SI_CONTEXT_INV_VMEM_L1
);
1016 sctx
->b
.num_L2_invalidates
++;
1019 /* Do the flush (enqueue the event and wait for it). */
1020 va
= sctx
->wait_mem_scratch
->gpu_address
;
1021 sctx
->wait_mem_number
++;
1023 si_gfx_write_event_eop(rctx
, cb_db_event
, tc_flags
,
1024 EOP_DATA_SEL_VALUE_32BIT
,
1025 sctx
->wait_mem_scratch
, va
,
1026 sctx
->wait_mem_number
, R600_NOT_QUERY
);
1027 si_gfx_wait_fence(rctx
, va
, sctx
->wait_mem_number
, 0xffffffff);
1030 /* Make sure ME is idle (it executes most packets) before continuing.
1031 * This prevents read-after-write hazards between PFP and ME.
1033 if (cp_coher_cntl
||
1034 (rctx
->flags
& (SI_CONTEXT_CS_PARTIAL_FLUSH
|
1035 SI_CONTEXT_INV_VMEM_L1
|
1036 SI_CONTEXT_INV_GLOBAL_L2
|
1037 SI_CONTEXT_WRITEBACK_GLOBAL_L2
))) {
1038 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
1043 * When one of the CP_COHER_CNTL.DEST_BASE flags is set, SURFACE_SYNC
1044 * waits for idle, so it should be last. SURFACE_SYNC is done in PFP.
1046 * cp_coher_cntl should contain all necessary flags except TC flags
1049 * SI-CIK don't support L2 write-back.
1051 if (rctx
->flags
& SI_CONTEXT_INV_GLOBAL_L2
||
1052 (rctx
->chip_class
<= CIK
&&
1053 (rctx
->flags
& SI_CONTEXT_WRITEBACK_GLOBAL_L2
))) {
1054 /* Invalidate L1 & L2. (L1 is always invalidated on SI)
1055 * WB must be set on VI+ when TC_ACTION is set.
1057 si_emit_surface_sync(rctx
, cp_coher_cntl
|
1058 S_0085F0_TC_ACTION_ENA(1) |
1059 S_0085F0_TCL1_ACTION_ENA(1) |
1060 S_0301F0_TC_WB_ACTION_ENA(rctx
->chip_class
>= VI
));
1062 sctx
->b
.num_L2_invalidates
++;
1064 /* L1 invalidation and L2 writeback must be done separately,
1065 * because both operations can't be done together.
1067 if (rctx
->flags
& SI_CONTEXT_WRITEBACK_GLOBAL_L2
) {
1069 * NC = apply to non-coherent MTYPEs
1070 * (i.e. MTYPE <= 1, which is what we use everywhere)
1072 * WB doesn't work without NC.
1074 si_emit_surface_sync(rctx
, cp_coher_cntl
|
1075 S_0301F0_TC_WB_ACTION_ENA(1) |
1076 S_0301F0_TC_NC_ACTION_ENA(1));
1078 sctx
->b
.num_L2_writebacks
++;
1080 if (rctx
->flags
& SI_CONTEXT_INV_VMEM_L1
) {
1081 /* Invalidate per-CU VMEM L1. */
1082 si_emit_surface_sync(rctx
, cp_coher_cntl
|
1083 S_0085F0_TCL1_ACTION_ENA(1));
1088 /* If TC flushes haven't cleared this... */
1090 si_emit_surface_sync(rctx
, cp_coher_cntl
);
1092 if (rctx
->flags
& R600_CONTEXT_START_PIPELINE_STATS
) {
1093 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1094 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
1096 } else if (rctx
->flags
& R600_CONTEXT_STOP_PIPELINE_STATS
) {
1097 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1098 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
1105 static void si_get_draw_start_count(struct si_context
*sctx
,
1106 const struct pipe_draw_info
*info
,
1107 unsigned *start
, unsigned *count
)
1109 struct pipe_draw_indirect_info
*indirect
= info
->indirect
;
1112 unsigned indirect_count
;
1113 struct pipe_transfer
*transfer
;
1114 unsigned begin
, end
;
1118 if (indirect
->indirect_draw_count
) {
1119 data
= pipe_buffer_map_range(&sctx
->b
.b
,
1120 indirect
->indirect_draw_count
,
1121 indirect
->indirect_draw_count_offset
,
1123 PIPE_TRANSFER_READ
, &transfer
);
1125 indirect_count
= *data
;
1127 pipe_buffer_unmap(&sctx
->b
.b
, transfer
);
1129 indirect_count
= indirect
->draw_count
;
1132 if (!indirect_count
) {
1133 *start
= *count
= 0;
1137 map_size
= (indirect_count
- 1) * indirect
->stride
+ 3 * sizeof(unsigned);
1138 data
= pipe_buffer_map_range(&sctx
->b
.b
, indirect
->buffer
,
1139 indirect
->offset
, map_size
,
1140 PIPE_TRANSFER_READ
, &transfer
);
1145 for (unsigned i
= 0; i
< indirect_count
; ++i
) {
1146 unsigned count
= data
[0];
1147 unsigned start
= data
[2];
1150 begin
= MIN2(begin
, start
);
1151 end
= MAX2(end
, start
+ count
);
1154 data
+= indirect
->stride
/ sizeof(unsigned);
1157 pipe_buffer_unmap(&sctx
->b
.b
, transfer
);
1161 *count
= end
- begin
;
1163 *start
= *count
= 0;
1166 *start
= info
->start
;
1167 *count
= info
->count
;
1171 static void si_emit_all_states(struct si_context
*sctx
, const struct pipe_draw_info
*info
,
1172 unsigned skip_atom_mask
)
1174 /* Emit state atoms. */
1175 unsigned mask
= sctx
->dirty_atoms
& ~skip_atom_mask
;
1177 struct r600_atom
*atom
= sctx
->atoms
.array
[u_bit_scan(&mask
)];
1179 atom
->emit(&sctx
->b
, atom
);
1181 sctx
->dirty_atoms
&= skip_atom_mask
;
1184 mask
= sctx
->dirty_states
;
1186 unsigned i
= u_bit_scan(&mask
);
1187 struct si_pm4_state
*state
= sctx
->queued
.array
[i
];
1189 if (!state
|| sctx
->emitted
.array
[i
] == state
)
1192 si_pm4_emit(sctx
, state
);
1193 sctx
->emitted
.array
[i
] = state
;
1195 sctx
->dirty_states
= 0;
1197 /* Emit draw states. */
1198 unsigned num_patches
= 0;
1200 si_emit_rasterizer_prim_state(sctx
);
1201 if (sctx
->tes_shader
.cso
)
1202 si_emit_derived_tess_state(sctx
, info
, &num_patches
);
1203 si_emit_vs_state(sctx
, info
);
1204 si_emit_draw_registers(sctx
, info
, num_patches
);
1207 void si_draw_vbo(struct pipe_context
*ctx
, const struct pipe_draw_info
*info
)
1209 struct si_context
*sctx
= (struct si_context
*)ctx
;
1210 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1211 struct pipe_resource
*indexbuf
= info
->index
.resource
;
1212 unsigned dirty_tex_counter
;
1213 enum pipe_prim_type rast_prim
;
1214 unsigned index_size
= info
->index_size
;
1215 unsigned index_offset
= info
->indirect
? info
->start
* index_size
: 0;
1217 if (likely(!info
->indirect
)) {
1218 /* SI-CI treat instance_count==0 as instance_count==1. There is
1219 * no workaround for indirect draws, but we can at least skip
1222 if (unlikely(!info
->instance_count
))
1225 /* Handle count == 0. */
1226 if (unlikely(!info
->count
&&
1227 (index_size
|| !info
->count_from_stream_output
)))
1231 if (unlikely(!sctx
->vs_shader
.cso
)) {
1235 if (unlikely(!sctx
->ps_shader
.cso
&& (!rs
|| !rs
->rasterizer_discard
))) {
1239 if (unlikely(!!sctx
->tes_shader
.cso
!= (info
->mode
== PIPE_PRIM_PATCHES
))) {
1244 /* Recompute and re-emit the texture resource states if needed. */
1245 dirty_tex_counter
= p_atomic_read(&sctx
->b
.screen
->dirty_tex_counter
);
1246 if (unlikely(dirty_tex_counter
!= sctx
->b
.last_dirty_tex_counter
)) {
1247 sctx
->b
.last_dirty_tex_counter
= dirty_tex_counter
;
1248 sctx
->framebuffer
.dirty_cbufs
|=
1249 ((1 << sctx
->framebuffer
.state
.nr_cbufs
) - 1);
1250 sctx
->framebuffer
.dirty_zsbuf
= true;
1251 si_mark_atom_dirty(sctx
, &sctx
->framebuffer
.atom
);
1252 si_update_all_texture_descriptors(sctx
);
1255 si_decompress_textures(sctx
, u_bit_consecutive(0, SI_NUM_GRAPHICS_SHADERS
));
1257 /* Set the rasterization primitive type.
1259 * This must be done after si_decompress_textures, which can call
1260 * draw_vbo recursively, and before si_update_shaders, which uses
1261 * current_rast_prim for this draw_vbo call. */
1262 if (sctx
->gs_shader
.cso
)
1263 rast_prim
= sctx
->gs_shader
.cso
->gs_output_prim
;
1264 else if (sctx
->tes_shader
.cso
) {
1265 if (sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_POINT_MODE
])
1266 rast_prim
= PIPE_PRIM_POINTS
;
1268 rast_prim
= sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
1270 rast_prim
= info
->mode
;
1272 if (rast_prim
!= sctx
->current_rast_prim
) {
1273 bool old_is_poly
= sctx
->current_rast_prim
>= PIPE_PRIM_TRIANGLES
;
1274 bool new_is_poly
= rast_prim
>= PIPE_PRIM_TRIANGLES
;
1275 if (old_is_poly
!= new_is_poly
) {
1276 sctx
->scissors
.dirty_mask
= (1 << SI_MAX_VIEWPORTS
) - 1;
1277 si_mark_atom_dirty(sctx
, &sctx
->scissors
.atom
);
1280 sctx
->current_rast_prim
= rast_prim
;
1281 sctx
->do_update_shaders
= true;
1284 if (sctx
->tes_shader
.cso
&&
1285 (sctx
->b
.family
== CHIP_VEGA10
|| sctx
->b
.family
== CHIP_RAVEN
)) {
1286 /* Determine whether the LS VGPR fix should be applied.
1288 * It is only required when num input CPs > num output CPs,
1289 * which cannot happen with the fixed function TCS. We should
1290 * also update this bit when switching from TCS to fixed
1293 struct si_shader_selector
*tcs
= sctx
->tcs_shader
.cso
;
1296 info
->vertices_per_patch
>
1297 tcs
->info
.properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
];
1299 if (ls_vgpr_fix
!= sctx
->ls_vgpr_fix
) {
1300 sctx
->ls_vgpr_fix
= ls_vgpr_fix
;
1301 sctx
->do_update_shaders
= true;
1305 if (sctx
->gs_shader
.cso
) {
1306 /* Determine whether the GS triangle strip adjacency fix should
1307 * be applied. Rotate every other triangle if
1308 * - triangle strips with adjacency are fed to the GS and
1309 * - primitive restart is disabled (the rotation doesn't help
1310 * when the restart occurs after an odd number of triangles).
1312 bool gs_tri_strip_adj_fix
=
1313 !sctx
->tes_shader
.cso
&&
1314 info
->mode
== PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
&&
1315 !info
->primitive_restart
;
1317 if (gs_tri_strip_adj_fix
!= sctx
->gs_tri_strip_adj_fix
) {
1318 sctx
->gs_tri_strip_adj_fix
= gs_tri_strip_adj_fix
;
1319 sctx
->do_update_shaders
= true;
1323 if (sctx
->do_update_shaders
&& !si_update_shaders(sctx
))
1327 /* Translate or upload, if needed. */
1328 /* 8-bit indices are supported on VI. */
1329 if (sctx
->b
.chip_class
<= CIK
&& index_size
== 1) {
1330 unsigned start
, count
, start_offset
, size
, offset
;
1333 si_get_draw_start_count(sctx
, info
, &start
, &count
);
1334 start_offset
= start
* 2;
1338 u_upload_alloc(ctx
->stream_uploader
, start_offset
,
1340 si_optimal_tcc_alignment(sctx
, size
),
1341 &offset
, &indexbuf
, &ptr
);
1345 util_shorten_ubyte_elts_to_userptr(&sctx
->b
.b
, info
, 0, 0,
1346 index_offset
+ start
,
1349 /* info->start will be added by the drawing code */
1350 index_offset
= offset
- start_offset
;
1352 } else if (info
->has_user_indices
) {
1353 unsigned start_offset
;
1355 assert(!info
->indirect
);
1356 start_offset
= info
->start
* index_size
;
1359 u_upload_data(ctx
->stream_uploader
, start_offset
,
1360 info
->count
* index_size
,
1361 sctx
->screen
->b
.info
.tcc_cache_line_size
,
1362 (char*)info
->index
.user
+ start_offset
,
1363 &index_offset
, &indexbuf
);
1367 /* info->start will be added by the drawing code */
1368 index_offset
-= start_offset
;
1369 } else if (sctx
->b
.chip_class
<= CIK
&&
1370 r600_resource(indexbuf
)->TC_L2_dirty
) {
1371 /* VI reads index buffers through TC L2, so it doesn't
1373 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1374 r600_resource(indexbuf
)->TC_L2_dirty
= false;
1378 if (info
->indirect
) {
1379 struct pipe_draw_indirect_info
*indirect
= info
->indirect
;
1381 /* Add the buffer size for memory checking in need_cs_space. */
1382 r600_context_add_resource_size(ctx
, indirect
->buffer
);
1384 /* Indirect buffers use TC L2 on GFX9, but not older hw. */
1385 if (sctx
->b
.chip_class
<= VI
) {
1386 if (r600_resource(indirect
->buffer
)->TC_L2_dirty
) {
1387 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1388 r600_resource(indirect
->buffer
)->TC_L2_dirty
= false;
1391 if (indirect
->indirect_draw_count
&&
1392 r600_resource(indirect
->indirect_draw_count
)->TC_L2_dirty
) {
1393 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1394 r600_resource(indirect
->indirect_draw_count
)->TC_L2_dirty
= false;
1399 si_need_cs_space(sctx
);
1401 /* Since we've called r600_context_add_resource_size for vertex buffers,
1402 * this must be called after si_need_cs_space, because we must let
1403 * need_cs_space flush before we add buffers to the buffer list.
1405 if (!si_upload_vertex_buffer_descriptors(sctx
))
1408 /* Vega10/Raven scissor bug workaround. This must be done before VPORT
1409 * scissor registers are changed. There is also a more efficient but
1410 * more involved alternative workaround.
1412 if ((sctx
->b
.family
== CHIP_VEGA10
|| sctx
->b
.family
== CHIP_RAVEN
) &&
1413 si_is_atom_dirty(sctx
, &sctx
->scissors
.atom
)) {
1414 sctx
->b
.flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
;
1415 si_emit_cache_flush(sctx
);
1418 /* Use optimal packet order based on whether we need to sync the pipeline. */
1419 if (unlikely(sctx
->b
.flags
& (SI_CONTEXT_FLUSH_AND_INV_CB
|
1420 SI_CONTEXT_FLUSH_AND_INV_DB
|
1421 SI_CONTEXT_PS_PARTIAL_FLUSH
|
1422 SI_CONTEXT_CS_PARTIAL_FLUSH
))) {
1423 /* If we have to wait for idle, set all states first, so that all
1424 * SET packets are processed in parallel with previous draw calls.
1425 * Then upload descriptors, set shader pointers, and draw, and
1426 * prefetch at the end. This ensures that the time the CUs
1427 * are idle is very short. (there are only SET_SH packets between
1428 * the wait and the draw)
1430 struct r600_atom
*shader_pointers
= &sctx
->shader_pointers
.atom
;
1431 unsigned masked_atoms
= 1u << shader_pointers
->id
;
1433 if (unlikely(sctx
->b
.flags
& R600_CONTEXT_FLUSH_FOR_RENDER_COND
))
1434 masked_atoms
|= 1u << sctx
->b
.render_cond_atom
.id
;
1436 /* Emit all states except shader pointers and render condition. */
1437 si_emit_all_states(sctx
, info
, masked_atoms
);
1438 si_emit_cache_flush(sctx
);
1440 /* <-- CUs are idle here. */
1441 if (!si_upload_graphics_shader_descriptors(sctx
))
1444 /* Set shader pointers after descriptors are uploaded. */
1445 if (si_is_atom_dirty(sctx
, shader_pointers
))
1446 shader_pointers
->emit(&sctx
->b
, NULL
);
1447 if (si_is_atom_dirty(sctx
, &sctx
->b
.render_cond_atom
))
1448 sctx
->b
.render_cond_atom
.emit(&sctx
->b
, NULL
);
1449 sctx
->dirty_atoms
= 0;
1451 si_emit_draw_packets(sctx
, info
, indexbuf
, index_size
, index_offset
);
1452 /* <-- CUs are busy here. */
1454 /* Start prefetches after the draw has been started. Both will run
1455 * in parallel, but starting the draw first is more important.
1457 if (sctx
->b
.chip_class
>= CIK
&& sctx
->prefetch_L2_mask
)
1458 cik_emit_prefetch_L2(sctx
);
1460 /* If we don't wait for idle, start prefetches first, then set
1461 * states, and draw at the end.
1464 si_emit_cache_flush(sctx
);
1466 if (sctx
->b
.chip_class
>= CIK
&& sctx
->prefetch_L2_mask
)
1467 cik_emit_prefetch_L2(sctx
);
1469 if (!si_upload_graphics_shader_descriptors(sctx
))
1472 si_emit_all_states(sctx
, info
, 0);
1473 si_emit_draw_packets(sctx
, info
, indexbuf
, index_size
, index_offset
);
1476 if (unlikely(sctx
->current_saved_cs
)) {
1477 si_trace_emit(sctx
);
1478 si_log_draw_state(sctx
, sctx
->b
.log
);
1481 /* Workaround for a VGT hang when streamout is enabled.
1482 * It must be done after drawing. */
1483 if ((sctx
->b
.family
== CHIP_HAWAII
||
1484 sctx
->b
.family
== CHIP_TONGA
||
1485 sctx
->b
.family
== CHIP_FIJI
) &&
1486 si_get_strmout_en(sctx
)) {
1487 sctx
->b
.flags
|= SI_CONTEXT_VGT_STREAMOUT_SYNC
;
1490 if (unlikely(sctx
->decompression_enabled
)) {
1491 sctx
->b
.num_decompress_calls
++;
1493 sctx
->b
.num_draw_calls
++;
1494 if (sctx
->framebuffer
.state
.nr_cbufs
> 1)
1495 sctx
->b
.num_mrt_draw_calls
++;
1496 if (info
->primitive_restart
)
1497 sctx
->b
.num_prim_restart_calls
++;
1498 if (G_0286E8_WAVESIZE(sctx
->spi_tmpring_size
))
1499 sctx
->b
.num_spill_draw_calls
++;
1501 if (index_size
&& indexbuf
!= info
->index
.resource
)
1502 pipe_resource_reference(&indexbuf
, NULL
);
1505 void si_draw_rectangle(struct blitter_context
*blitter
,
1506 void *vertex_elements_cso
,
1507 blitter_get_vs_func get_vs
,
1508 int x1
, int y1
, int x2
, int y2
,
1509 float depth
, unsigned num_instances
,
1510 enum blitter_attrib_type type
,
1511 const union blitter_attrib
*attrib
)
1513 struct pipe_context
*pipe
= util_blitter_get_pipe(blitter
);
1514 struct si_context
*sctx
= (struct si_context
*)pipe
;
1516 /* Pack position coordinates as signed int16. */
1517 sctx
->vs_blit_sh_data
[0] = (uint32_t)(x1
& 0xffff) |
1518 ((uint32_t)(y1
& 0xffff) << 16);
1519 sctx
->vs_blit_sh_data
[1] = (uint32_t)(x2
& 0xffff) |
1520 ((uint32_t)(y2
& 0xffff) << 16);
1521 sctx
->vs_blit_sh_data
[2] = fui(depth
);
1524 case UTIL_BLITTER_ATTRIB_COLOR
:
1525 memcpy(&sctx
->vs_blit_sh_data
[3], attrib
->color
,
1528 case UTIL_BLITTER_ATTRIB_TEXCOORD_XY
:
1529 case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW
:
1530 memcpy(&sctx
->vs_blit_sh_data
[3], &attrib
->texcoord
,
1531 sizeof(attrib
->texcoord
));
1533 case UTIL_BLITTER_ATTRIB_NONE
:;
1536 pipe
->bind_vs_state(pipe
, si_get_blit_vs(sctx
, type
, num_instances
));
1538 struct pipe_draw_info info
= {};
1539 info
.mode
= R600_PRIM_RECTANGLE_LIST
;
1541 info
.instance_count
= num_instances
;
1543 /* Don't set per-stage shader pointers for VS. */
1544 sctx
->shader_pointers_dirty
&= ~SI_DESCS_SHADER_MASK(VERTEX
);
1545 sctx
->vertex_buffer_pointer_dirty
= false;
1547 si_draw_vbo(pipe
, &info
);
1550 void si_trace_emit(struct si_context
*sctx
)
1552 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
1553 uint64_t va
= sctx
->current_saved_cs
->trace_buf
->gpu_address
;
1554 uint32_t trace_id
= ++sctx
->current_saved_cs
->trace_id
;
1556 radeon_emit(cs
, PKT3(PKT3_WRITE_DATA
, 3, 0));
1557 radeon_emit(cs
, S_370_DST_SEL(V_370_MEMORY_SYNC
) |
1558 S_370_WR_CONFIRM(1) |
1559 S_370_ENGINE_SEL(V_370_ME
));
1560 radeon_emit(cs
, va
);
1561 radeon_emit(cs
, va
>> 32);
1562 radeon_emit(cs
, trace_id
);
1563 radeon_emit(cs
, PKT3(PKT3_NOP
, 0, 0));
1564 radeon_emit(cs
, AC_ENCODE_TRACE_POINT(trace_id
));
1567 u_log_flush(sctx
->b
.log
);