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 "si_shader.h"
29 #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"
35 #include "util/u_memory.h"
37 static unsigned si_conv_pipe_prim(unsigned mode
)
39 static const unsigned prim_conv
[] = {
40 [PIPE_PRIM_POINTS
] = V_008958_DI_PT_POINTLIST
,
41 [PIPE_PRIM_LINES
] = V_008958_DI_PT_LINELIST
,
42 [PIPE_PRIM_LINE_LOOP
] = V_008958_DI_PT_LINELOOP
,
43 [PIPE_PRIM_LINE_STRIP
] = V_008958_DI_PT_LINESTRIP
,
44 [PIPE_PRIM_TRIANGLES
] = V_008958_DI_PT_TRILIST
,
45 [PIPE_PRIM_TRIANGLE_STRIP
] = V_008958_DI_PT_TRISTRIP
,
46 [PIPE_PRIM_TRIANGLE_FAN
] = V_008958_DI_PT_TRIFAN
,
47 [PIPE_PRIM_QUADS
] = V_008958_DI_PT_QUADLIST
,
48 [PIPE_PRIM_QUAD_STRIP
] = V_008958_DI_PT_QUADSTRIP
,
49 [PIPE_PRIM_POLYGON
] = V_008958_DI_PT_POLYGON
,
50 [PIPE_PRIM_LINES_ADJACENCY
] = V_008958_DI_PT_LINELIST_ADJ
,
51 [PIPE_PRIM_LINE_STRIP_ADJACENCY
] = V_008958_DI_PT_LINESTRIP_ADJ
,
52 [PIPE_PRIM_TRIANGLES_ADJACENCY
] = V_008958_DI_PT_TRILIST_ADJ
,
53 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
] = V_008958_DI_PT_TRISTRIP_ADJ
,
54 [PIPE_PRIM_PATCHES
] = V_008958_DI_PT_PATCH
,
55 [R600_PRIM_RECTANGLE_LIST
] = V_008958_DI_PT_RECTLIST
57 assert(mode
< ARRAY_SIZE(prim_conv
));
58 return prim_conv
[mode
];
61 static unsigned si_conv_prim_to_gs_out(unsigned mode
)
63 static const int prim_conv
[] = {
64 [PIPE_PRIM_POINTS
] = V_028A6C_OUTPRIM_TYPE_POINTLIST
,
65 [PIPE_PRIM_LINES
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
66 [PIPE_PRIM_LINE_LOOP
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
67 [PIPE_PRIM_LINE_STRIP
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
68 [PIPE_PRIM_TRIANGLES
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
69 [PIPE_PRIM_TRIANGLE_STRIP
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
70 [PIPE_PRIM_TRIANGLE_FAN
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
71 [PIPE_PRIM_QUADS
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
72 [PIPE_PRIM_QUAD_STRIP
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
73 [PIPE_PRIM_POLYGON
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
74 [PIPE_PRIM_LINES_ADJACENCY
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
75 [PIPE_PRIM_LINE_STRIP_ADJACENCY
] = V_028A6C_OUTPRIM_TYPE_LINESTRIP
,
76 [PIPE_PRIM_TRIANGLES_ADJACENCY
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
77 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
,
78 [PIPE_PRIM_PATCHES
] = V_028A6C_OUTPRIM_TYPE_POINTLIST
,
79 [R600_PRIM_RECTANGLE_LIST
] = V_028A6C_OUTPRIM_TYPE_TRISTRIP
81 assert(mode
< ARRAY_SIZE(prim_conv
));
83 return prim_conv
[mode
];
87 * This calculates the LDS size for tessellation shaders (VS, TCS, TES).
88 * LS.LDS_SIZE is shared by all 3 shader stages.
90 * The information about LDS and other non-compile-time parameters is then
91 * written to userdata SGPRs.
93 static void si_emit_derived_tess_state(struct si_context
*sctx
,
94 const struct pipe_draw_info
*info
,
95 unsigned *num_patches
)
97 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
98 struct si_shader_ctx_state
*ls
= &sctx
->vs_shader
;
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 tes_sh_base
= sctx
->shader_userdata
.sh_base
[PIPE_SHADER_TESS_EVAL
];
104 unsigned num_tcs_input_cp
= info
->vertices_per_patch
;
105 unsigned num_tcs_output_cp
, num_tcs_inputs
, num_tcs_outputs
;
106 unsigned num_tcs_patch_outputs
;
107 unsigned input_vertex_size
, output_vertex_size
, pervertex_output_patch_size
;
108 unsigned input_patch_size
, output_patch_size
, output_patch0_offset
;
109 unsigned perpatch_output_offset
, lds_size
, ls_rsrc2
;
110 unsigned tcs_in_layout
, tcs_out_layout
, tcs_out_offsets
;
111 unsigned offchip_layout
, hardware_lds_size
;
113 /* This calculates how shader inputs and outputs among VS, TCS, and TES
114 * are laid out in LDS. */
115 num_tcs_inputs
= util_last_bit64(ls
->cso
->outputs_written
);
117 if (sctx
->tcs_shader
.cso
) {
118 num_tcs_outputs
= util_last_bit64(tcs
->outputs_written
);
119 num_tcs_output_cp
= tcs
->info
.properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
];
120 num_tcs_patch_outputs
= util_last_bit64(tcs
->patch_outputs_written
);
122 /* No TCS. Route varyings from LS to TES. */
123 num_tcs_outputs
= num_tcs_inputs
;
124 num_tcs_output_cp
= num_tcs_input_cp
;
125 num_tcs_patch_outputs
= 2; /* TESSINNER + TESSOUTER */
128 input_vertex_size
= num_tcs_inputs
* 16;
129 output_vertex_size
= num_tcs_outputs
* 16;
131 input_patch_size
= num_tcs_input_cp
* input_vertex_size
;
133 pervertex_output_patch_size
= num_tcs_output_cp
* output_vertex_size
;
134 output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
136 /* Ensure that we only need one wave per SIMD so we don't need to check
137 * resource usage. Also ensures that the number of tcs in and out
138 * vertices per threadgroup are at most 256.
140 *num_patches
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
) * 4;
142 /* Make sure that the data fits in LDS. This assumes the shaders only
143 * use LDS for the inputs and outputs.
145 hardware_lds_size
= sctx
->b
.chip_class
>= CIK
? 65536 : 32768;
146 *num_patches
= MIN2(*num_patches
, hardware_lds_size
/ (input_patch_size
+
149 /* Make sure the output data fits in the offchip buffer */
150 *num_patches
= MIN2(*num_patches
, SI_TESS_OFFCHIP_BLOCK_SIZE
/
153 /* Not necessary for correctness, but improves performance. The
154 * specific value is taken from the proprietary driver.
156 *num_patches
= MIN2(*num_patches
, 40);
158 output_patch0_offset
= input_patch_size
* *num_patches
;
159 perpatch_output_offset
= output_patch0_offset
+ pervertex_output_patch_size
;
161 lds_size
= output_patch0_offset
+ output_patch_size
* *num_patches
;
162 ls_rsrc2
= ls
->current
->config
.rsrc2
;
164 if (sctx
->b
.chip_class
>= CIK
) {
165 assert(lds_size
<= 65536);
166 ls_rsrc2
|= S_00B52C_LDS_SIZE(align(lds_size
, 512) / 512);
168 assert(lds_size
<= 32768);
169 ls_rsrc2
|= S_00B52C_LDS_SIZE(align(lds_size
, 256) / 256);
172 if (sctx
->last_ls
== ls
->current
&&
173 sctx
->last_tcs
== tcs
&&
174 sctx
->last_tes_sh_base
== tes_sh_base
&&
175 sctx
->last_num_tcs_input_cp
== num_tcs_input_cp
)
178 sctx
->last_ls
= ls
->current
;
179 sctx
->last_tcs
= tcs
;
180 sctx
->last_tes_sh_base
= tes_sh_base
;
181 sctx
->last_num_tcs_input_cp
= num_tcs_input_cp
;
183 /* Due to a hw bug, RSRC2_LS must be written twice with another
184 * LS register written in between. */
185 if (sctx
->b
.chip_class
== CIK
&& sctx
->b
.family
!= CHIP_HAWAII
)
186 radeon_set_sh_reg(cs
, R_00B52C_SPI_SHADER_PGM_RSRC2_LS
, ls_rsrc2
);
187 radeon_set_sh_reg_seq(cs
, R_00B528_SPI_SHADER_PGM_RSRC1_LS
, 2);
188 radeon_emit(cs
, ls
->current
->config
.rsrc1
);
189 radeon_emit(cs
, ls_rsrc2
);
191 /* Compute userdata SGPRs. */
192 assert(((input_vertex_size
/ 4) & ~0xff) == 0);
193 assert(((output_vertex_size
/ 4) & ~0xff) == 0);
194 assert(((input_patch_size
/ 4) & ~0x1fff) == 0);
195 assert(((output_patch_size
/ 4) & ~0x1fff) == 0);
196 assert(((output_patch0_offset
/ 16) & ~0xffff) == 0);
197 assert(((perpatch_output_offset
/ 16) & ~0xffff) == 0);
198 assert(num_tcs_input_cp
<= 32);
199 assert(num_tcs_output_cp
<= 32);
201 tcs_in_layout
= (input_patch_size
/ 4) |
202 ((input_vertex_size
/ 4) << 13);
203 tcs_out_layout
= (output_patch_size
/ 4) |
204 ((output_vertex_size
/ 4) << 13);
205 tcs_out_offsets
= (output_patch0_offset
/ 16) |
206 ((perpatch_output_offset
/ 16) << 16);
207 offchip_layout
= (pervertex_output_patch_size
* *num_patches
<< 16) |
208 (num_tcs_output_cp
<< 9) | *num_patches
;
210 /* Set them for LS. */
211 radeon_set_sh_reg(cs
,
212 R_00B530_SPI_SHADER_USER_DATA_LS_0
+ SI_SGPR_LS_OUT_LAYOUT
* 4,
215 /* Set them for TCS. */
216 radeon_set_sh_reg_seq(cs
,
217 R_00B430_SPI_SHADER_USER_DATA_HS_0
+ SI_SGPR_TCS_OFFCHIP_LAYOUT
* 4, 4);
218 radeon_emit(cs
, offchip_layout
);
219 radeon_emit(cs
, tcs_out_offsets
);
220 radeon_emit(cs
, tcs_out_layout
| (num_tcs_input_cp
<< 26));
221 radeon_emit(cs
, tcs_in_layout
);
223 /* Set them for TES. */
224 radeon_set_sh_reg_seq(cs
, tes_sh_base
+ SI_SGPR_TCS_OFFCHIP_LAYOUT
* 4, 1);
225 radeon_emit(cs
, offchip_layout
);
228 static unsigned si_num_prims_for_vertices(const struct pipe_draw_info
*info
)
230 switch (info
->mode
) {
231 case PIPE_PRIM_PATCHES
:
232 return info
->count
/ info
->vertices_per_patch
;
233 case R600_PRIM_RECTANGLE_LIST
:
234 return info
->count
/ 3;
236 return u_prims_for_vertices(info
->mode
, info
->count
);
240 static unsigned si_get_ia_multi_vgt_param(struct si_context
*sctx
,
241 const struct pipe_draw_info
*info
,
242 unsigned num_patches
)
244 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
245 unsigned prim
= info
->mode
;
246 unsigned primgroup_size
= 128; /* recommended without a GS */
247 unsigned max_primgroup_in_wave
= 2;
249 /* SWITCH_ON_EOP(0) is always preferable. */
250 bool wd_switch_on_eop
= false;
251 bool ia_switch_on_eop
= false;
252 bool ia_switch_on_eoi
= false;
253 bool partial_vs_wave
= false;
254 bool partial_es_wave
= false;
256 if (sctx
->gs_shader
.cso
)
257 primgroup_size
= 64; /* recommended with a GS */
259 if (sctx
->tes_shader
.cso
) {
260 /* primgroup_size must be set to a multiple of NUM_PATCHES */
261 primgroup_size
= num_patches
;
263 /* SWITCH_ON_EOI must be set if PrimID is used. */
264 if ((sctx
->tcs_shader
.cso
&& sctx
->tcs_shader
.cso
->info
.uses_primid
) ||
265 sctx
->tes_shader
.cso
->info
.uses_primid
)
266 ia_switch_on_eoi
= true;
268 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
269 if ((sctx
->b
.family
== CHIP_TAHITI
||
270 sctx
->b
.family
== CHIP_PITCAIRN
||
271 sctx
->b
.family
== CHIP_BONAIRE
) &&
273 partial_vs_wave
= true;
275 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
276 if (sctx
->b
.chip_class
>= VI
) {
277 if (sctx
->gs_shader
.cso
)
278 partial_es_wave
= true;
280 partial_vs_wave
= true;
284 /* This is a hardware requirement. */
285 if ((rs
&& rs
->line_stipple_enable
) ||
286 (sctx
->b
.screen
->debug_flags
& DBG_SWITCH_ON_EOP
)) {
287 ia_switch_on_eop
= true;
288 wd_switch_on_eop
= true;
291 if (sctx
->b
.chip_class
>= CIK
) {
292 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
293 * 4 shader engines. Set 1 to pass the assertion below.
294 * The other cases are hardware requirements. */
295 if (sctx
->b
.screen
->info
.max_se
< 4 ||
296 prim
== PIPE_PRIM_POLYGON
||
297 prim
== PIPE_PRIM_LINE_LOOP
||
298 prim
== PIPE_PRIM_TRIANGLE_FAN
||
299 prim
== PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
||
300 info
->primitive_restart
||
301 info
->count_from_stream_output
)
302 wd_switch_on_eop
= true;
304 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
305 * We don't know that for indirect drawing, so treat it as
306 * always problematic. */
307 if (sctx
->b
.family
== CHIP_HAWAII
&&
308 (info
->indirect
|| info
->instance_count
> 1))
309 wd_switch_on_eop
= true;
311 /* Performance recommendation for 4 SE Gfx7-8 parts if
312 * instances are smaller than a primgroup. Ignore the fact
313 * primgroup_size is a primitive count, not vertex count.
314 * Don't do anything for indirect draws.
316 if (sctx
->b
.chip_class
<= VI
&&
317 sctx
->b
.screen
->info
.max_se
>= 4 &&
319 info
->instance_count
> 1 && info
->count
< primgroup_size
)
320 wd_switch_on_eop
= true;
322 /* Required on CIK and later. */
323 if (sctx
->b
.screen
->info
.max_se
> 2 && !wd_switch_on_eop
)
324 ia_switch_on_eoi
= true;
326 /* Required by Hawaii and, for some special cases, by VI. */
327 if (ia_switch_on_eoi
&&
328 (sctx
->b
.family
== CHIP_HAWAII
||
329 (sctx
->b
.chip_class
== VI
&&
330 (sctx
->gs_shader
.cso
|| max_primgroup_in_wave
!= 2))))
331 partial_vs_wave
= true;
333 /* Instancing bug on Bonaire. */
334 if (sctx
->b
.family
== CHIP_BONAIRE
&& ia_switch_on_eoi
&&
335 (info
->indirect
|| info
->instance_count
> 1))
336 partial_vs_wave
= true;
338 /* If the WD switch is false, the IA switch must be false too. */
339 assert(wd_switch_on_eop
|| !ia_switch_on_eop
);
342 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
343 if (ia_switch_on_eoi
)
344 partial_es_wave
= true;
346 /* GS requirement. */
347 if (SI_GS_PER_ES
/ primgroup_size
>= sctx
->screen
->gs_table_depth
- 3)
348 partial_es_wave
= true;
350 /* Hw bug with single-primitive instances and SWITCH_ON_EOI
351 * on multi-SE chips. */
352 if (sctx
->b
.screen
->info
.max_se
>= 2 && ia_switch_on_eoi
&&
354 (info
->instance_count
> 1 &&
355 si_num_prims_for_vertices(info
) <= 1)))
356 sctx
->b
.flags
|= SI_CONTEXT_VGT_FLUSH
;
358 return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop
) |
359 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi
) |
360 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave
) |
361 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave
) |
362 S_028AA8_PRIMGROUP_SIZE(primgroup_size
- 1) |
363 S_028AA8_WD_SWITCH_ON_EOP(sctx
->b
.chip_class
>= CIK
? wd_switch_on_eop
: 0) |
364 S_028AA8_MAX_PRIMGRP_IN_WAVE(sctx
->b
.chip_class
>= VI
?
365 max_primgroup_in_wave
: 0);
368 static unsigned si_get_ls_hs_config(struct si_context
*sctx
,
369 const struct pipe_draw_info
*info
,
370 unsigned num_patches
)
372 unsigned num_output_cp
;
374 if (!sctx
->tes_shader
.cso
)
377 num_output_cp
= sctx
->tcs_shader
.cso
?
378 sctx
->tcs_shader
.cso
->info
.properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
] :
379 info
->vertices_per_patch
;
381 return S_028B58_NUM_PATCHES(num_patches
) |
382 S_028B58_HS_NUM_INPUT_CP(info
->vertices_per_patch
) |
383 S_028B58_HS_NUM_OUTPUT_CP(num_output_cp
);
386 static void si_emit_scratch_reloc(struct si_context
*sctx
)
388 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
390 if (!sctx
->emit_scratch_reloc
)
393 radeon_set_context_reg(cs
, R_0286E8_SPI_TMPRING_SIZE
,
394 sctx
->spi_tmpring_size
);
396 if (sctx
->scratch_buffer
) {
397 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
398 sctx
->scratch_buffer
, RADEON_USAGE_READWRITE
,
399 RADEON_PRIO_SCRATCH_BUFFER
);
402 sctx
->emit_scratch_reloc
= false;
405 /* rast_prim is the primitive type after GS. */
406 static void si_emit_rasterizer_prim_state(struct si_context
*sctx
)
408 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
409 unsigned rast_prim
= sctx
->current_rast_prim
;
410 struct si_state_rasterizer
*rs
= sctx
->emitted
.named
.rasterizer
;
412 /* Skip this if not rendering lines. */
413 if (rast_prim
!= PIPE_PRIM_LINES
&&
414 rast_prim
!= PIPE_PRIM_LINE_LOOP
&&
415 rast_prim
!= PIPE_PRIM_LINE_STRIP
&&
416 rast_prim
!= PIPE_PRIM_LINES_ADJACENCY
&&
417 rast_prim
!= PIPE_PRIM_LINE_STRIP_ADJACENCY
)
420 if (rast_prim
== sctx
->last_rast_prim
&&
421 rs
->pa_sc_line_stipple
== sctx
->last_sc_line_stipple
)
424 radeon_set_context_reg(cs
, R_028A0C_PA_SC_LINE_STIPPLE
,
425 rs
->pa_sc_line_stipple
|
426 S_028A0C_AUTO_RESET_CNTL(rast_prim
== PIPE_PRIM_LINES
? 1 :
427 rast_prim
== PIPE_PRIM_LINE_STRIP
? 2 : 0));
429 sctx
->last_rast_prim
= rast_prim
;
430 sctx
->last_sc_line_stipple
= rs
->pa_sc_line_stipple
;
433 static void si_emit_draw_registers(struct si_context
*sctx
,
434 const struct pipe_draw_info
*info
)
436 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
437 unsigned prim
= si_conv_pipe_prim(info
->mode
);
438 unsigned gs_out_prim
= si_conv_prim_to_gs_out(sctx
->current_rast_prim
);
439 unsigned ia_multi_vgt_param
, ls_hs_config
, num_patches
= 0;
441 /* Polaris needs different VTX_REUSE_DEPTH settings depending on
442 * whether the "fractional odd" tessellation spacing is used.
444 if (sctx
->b
.family
>= CHIP_POLARIS10
) {
445 struct si_shader_selector
*tes
= sctx
->tes_shader
.cso
;
446 unsigned vtx_reuse_depth
= 30;
449 tes
->info
.properties
[TGSI_PROPERTY_TES_SPACING
] ==
450 PIPE_TESS_SPACING_FRACTIONAL_ODD
)
451 vtx_reuse_depth
= 14;
453 if (vtx_reuse_depth
!= sctx
->last_vtx_reuse_depth
) {
454 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
456 sctx
->last_vtx_reuse_depth
= vtx_reuse_depth
;
460 if (sctx
->tes_shader
.cso
)
461 si_emit_derived_tess_state(sctx
, info
, &num_patches
);
463 ia_multi_vgt_param
= si_get_ia_multi_vgt_param(sctx
, info
, num_patches
);
464 ls_hs_config
= si_get_ls_hs_config(sctx
, info
, num_patches
);
467 if (prim
!= sctx
->last_prim
||
468 ia_multi_vgt_param
!= sctx
->last_multi_vgt_param
||
469 ls_hs_config
!= sctx
->last_ls_hs_config
) {
470 if (sctx
->b
.family
>= CHIP_POLARIS10
) {
471 radeon_set_context_reg_idx(cs
, R_028AA8_IA_MULTI_VGT_PARAM
, 1, ia_multi_vgt_param
);
472 radeon_set_context_reg_idx(cs
, R_028B58_VGT_LS_HS_CONFIG
, 2, ls_hs_config
);
473 radeon_set_uconfig_reg_idx(cs
, R_030908_VGT_PRIMITIVE_TYPE
, 1, prim
);
474 } else if (sctx
->b
.chip_class
>= CIK
) {
475 radeon_emit(cs
, PKT3(PKT3_DRAW_PREAMBLE
, 2, 0));
476 radeon_emit(cs
, prim
); /* VGT_PRIMITIVE_TYPE */
477 radeon_emit(cs
, ia_multi_vgt_param
); /* IA_MULTI_VGT_PARAM */
478 radeon_emit(cs
, ls_hs_config
); /* VGT_LS_HS_CONFIG */
480 radeon_set_config_reg(cs
, R_008958_VGT_PRIMITIVE_TYPE
, prim
);
481 radeon_set_context_reg(cs
, R_028AA8_IA_MULTI_VGT_PARAM
, ia_multi_vgt_param
);
482 radeon_set_context_reg(cs
, R_028B58_VGT_LS_HS_CONFIG
, ls_hs_config
);
485 sctx
->last_prim
= prim
;
486 sctx
->last_multi_vgt_param
= ia_multi_vgt_param
;
487 sctx
->last_ls_hs_config
= ls_hs_config
;
490 if (gs_out_prim
!= sctx
->last_gs_out_prim
) {
491 radeon_set_context_reg(cs
, R_028A6C_VGT_GS_OUT_PRIM_TYPE
, gs_out_prim
);
492 sctx
->last_gs_out_prim
= gs_out_prim
;
495 /* Primitive restart. */
496 if (info
->primitive_restart
!= sctx
->last_primitive_restart_en
) {
497 radeon_set_context_reg(cs
, R_028A94_VGT_MULTI_PRIM_IB_RESET_EN
, info
->primitive_restart
);
498 sctx
->last_primitive_restart_en
= info
->primitive_restart
;
500 if (info
->primitive_restart
&&
501 (info
->restart_index
!= sctx
->last_restart_index
||
502 sctx
->last_restart_index
== SI_RESTART_INDEX_UNKNOWN
)) {
503 radeon_set_context_reg(cs
, R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX
,
504 info
->restart_index
);
505 sctx
->last_restart_index
= info
->restart_index
;
510 static void si_emit_draw_packets(struct si_context
*sctx
,
511 const struct pipe_draw_info
*info
,
512 const struct pipe_index_buffer
*ib
)
514 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
515 unsigned sh_base_reg
= sctx
->shader_userdata
.sh_base
[PIPE_SHADER_VERTEX
];
516 bool render_cond_bit
= sctx
->b
.render_cond
&& !sctx
->b
.render_cond_force_off
;
518 if (info
->count_from_stream_output
) {
519 struct r600_so_target
*t
=
520 (struct r600_so_target
*)info
->count_from_stream_output
;
521 uint64_t va
= t
->buf_filled_size
->gpu_address
+
522 t
->buf_filled_size_offset
;
524 radeon_set_context_reg(cs
, R_028B30_VGT_STRMOUT_DRAW_OPAQUE_VERTEX_STRIDE
,
527 radeon_emit(cs
, PKT3(PKT3_COPY_DATA
, 4, 0));
528 radeon_emit(cs
, COPY_DATA_SRC_SEL(COPY_DATA_MEM
) |
529 COPY_DATA_DST_SEL(COPY_DATA_REG
) |
530 COPY_DATA_WR_CONFIRM
);
531 radeon_emit(cs
, va
); /* src address lo */
532 radeon_emit(cs
, va
>> 32); /* src address hi */
533 radeon_emit(cs
, R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE
>> 2);
534 radeon_emit(cs
, 0); /* unused */
536 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
537 t
->buf_filled_size
, RADEON_USAGE_READ
,
538 RADEON_PRIO_SO_FILLED_SIZE
);
543 radeon_emit(cs
, PKT3(PKT3_INDEX_TYPE
, 0, 0));
546 switch (ib
->index_size
) {
548 radeon_emit(cs
, V_028A7C_VGT_INDEX_8
);
551 radeon_emit(cs
, V_028A7C_VGT_INDEX_16
|
552 (SI_BIG_ENDIAN
&& sctx
->b
.chip_class
<= CIK
?
553 V_028A7C_VGT_DMA_SWAP_16_BIT
: 0));
556 radeon_emit(cs
, V_028A7C_VGT_INDEX_32
|
557 (SI_BIG_ENDIAN
&& sctx
->b
.chip_class
<= CIK
?
558 V_028A7C_VGT_DMA_SWAP_32_BIT
: 0));
561 assert(!"unreachable");
566 if (!info
->indirect
) {
569 radeon_emit(cs
, PKT3(PKT3_NUM_INSTANCES
, 0, 0));
570 radeon_emit(cs
, info
->instance_count
);
572 /* Base vertex and start instance. */
573 base_vertex
= info
->indexed
? info
->index_bias
: info
->start
;
575 if (base_vertex
!= sctx
->last_base_vertex
||
576 sctx
->last_base_vertex
== SI_BASE_VERTEX_UNKNOWN
||
577 info
->start_instance
!= sctx
->last_start_instance
||
578 sh_base_reg
!= sctx
->last_sh_base_reg
) {
579 radeon_set_sh_reg_seq(cs
, sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4, 2);
580 radeon_emit(cs
, base_vertex
);
581 radeon_emit(cs
, info
->start_instance
);
583 sctx
->last_base_vertex
= base_vertex
;
584 sctx
->last_start_instance
= info
->start_instance
;
585 sctx
->last_sh_base_reg
= sh_base_reg
;
588 si_invalidate_draw_sh_constants(sctx
);
590 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
591 (struct r600_resource
*)info
->indirect
,
592 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
596 uint32_t index_max_size
= (ib
->buffer
->width0
- ib
->offset
) /
598 uint64_t index_va
= r600_resource(ib
->buffer
)->gpu_address
+ ib
->offset
;
600 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
601 (struct r600_resource
*)ib
->buffer
,
602 RADEON_USAGE_READ
, RADEON_PRIO_INDEX_BUFFER
);
604 if (info
->indirect
) {
605 uint64_t indirect_va
= r600_resource(info
->indirect
)->gpu_address
;
607 assert(indirect_va
% 8 == 0);
608 assert(index_va
% 2 == 0);
609 assert(info
->indirect_offset
% 4 == 0);
611 radeon_emit(cs
, PKT3(PKT3_SET_BASE
, 2, 0));
613 radeon_emit(cs
, indirect_va
);
614 radeon_emit(cs
, indirect_va
>> 32);
616 radeon_emit(cs
, PKT3(PKT3_INDEX_BASE
, 1, 0));
617 radeon_emit(cs
, index_va
);
618 radeon_emit(cs
, index_va
>> 32);
620 radeon_emit(cs
, PKT3(PKT3_INDEX_BUFFER_SIZE
, 0, 0));
621 radeon_emit(cs
, index_max_size
);
623 if (sctx
->b
.family
< CHIP_POLARIS10
) {
624 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_INDIRECT
, 3, render_cond_bit
));
625 radeon_emit(cs
, info
->indirect_offset
);
626 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
627 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
628 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_DMA
);
630 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_INDIRECT_MULTI
, 8, render_cond_bit
));
631 radeon_emit(cs
, info
->indirect_offset
);
632 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
633 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
634 radeon_emit(cs
, 0); /* draw_index */
635 radeon_emit(cs
, 1); /* count */
636 radeon_emit(cs
, 0); /* count_addr -- disabled */
638 radeon_emit(cs
, 16); /* stride */
639 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_DMA
);
642 index_va
+= info
->start
* ib
->index_size
;
644 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_2
, 4, render_cond_bit
));
645 radeon_emit(cs
, index_max_size
);
646 radeon_emit(cs
, index_va
);
647 radeon_emit(cs
, (index_va
>> 32UL) & 0xFF);
648 radeon_emit(cs
, info
->count
);
649 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_DMA
);
652 if (info
->indirect
) {
653 uint64_t indirect_va
= r600_resource(info
->indirect
)->gpu_address
;
655 assert(indirect_va
% 8 == 0);
656 assert(info
->indirect_offset
% 4 == 0);
658 radeon_emit(cs
, PKT3(PKT3_SET_BASE
, 2, 0));
660 radeon_emit(cs
, indirect_va
);
661 radeon_emit(cs
, indirect_va
>> 32);
663 if (sctx
->b
.family
< CHIP_POLARIS10
) {
664 radeon_emit(cs
, PKT3(PKT3_DRAW_INDIRECT
, 3, render_cond_bit
));
665 radeon_emit(cs
, info
->indirect_offset
);
666 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
667 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
668 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_AUTO_INDEX
);
670 radeon_emit(cs
, PKT3(PKT3_DRAW_INDIRECT_MULTI
, 8, render_cond_bit
));
671 radeon_emit(cs
, info
->indirect_offset
);
672 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_BASE_VERTEX
* 4 - SI_SH_REG_OFFSET
) >> 2);
673 radeon_emit(cs
, (sh_base_reg
+ SI_SGPR_START_INSTANCE
* 4 - SI_SH_REG_OFFSET
) >> 2);
674 radeon_emit(cs
, 0); /* draw_index */
675 radeon_emit(cs
, 1); /* count */
676 radeon_emit(cs
, 0); /* count_addr -- disabled */
678 radeon_emit(cs
, 16); /* stride */
679 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_AUTO_INDEX
);
682 radeon_emit(cs
, PKT3(PKT3_DRAW_INDEX_AUTO
, 1, render_cond_bit
));
683 radeon_emit(cs
, info
->count
);
684 radeon_emit(cs
, V_0287F0_DI_SRC_SEL_AUTO_INDEX
|
685 S_0287F0_USE_OPAQUE(!!info
->count_from_stream_output
));
690 void si_emit_cache_flush(struct si_context
*si_ctx
, struct r600_atom
*atom
)
692 struct r600_common_context
*sctx
= &si_ctx
->b
;
693 struct radeon_winsys_cs
*cs
= sctx
->gfx
.cs
;
694 uint32_t cp_coher_cntl
= 0;
696 /* SI has a bug that it always flushes ICACHE and KCACHE if either
697 * bit is set. An alternative way is to write SQC_CACHES, but that
698 * doesn't seem to work reliably. Since the bug doesn't affect
699 * correctness (it only does more work than necessary) and
700 * the performance impact is likely negligible, there is no plan
701 * to add a workaround for it.
704 if (sctx
->flags
& SI_CONTEXT_INV_ICACHE
)
705 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
706 if (sctx
->flags
& SI_CONTEXT_INV_SMEM_L1
)
707 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
709 if (sctx
->flags
& SI_CONTEXT_INV_VMEM_L1
)
710 cp_coher_cntl
|= S_0085F0_TCL1_ACTION_ENA(1);
711 if (sctx
->flags
& SI_CONTEXT_INV_GLOBAL_L2
) {
712 cp_coher_cntl
|= S_0085F0_TC_ACTION_ENA(1);
714 if (sctx
->chip_class
>= VI
)
715 cp_coher_cntl
|= S_0301F0_TC_WB_ACTION_ENA(1);
718 if (sctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_CB
) {
719 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
720 S_0085F0_CB0_DEST_BASE_ENA(1) |
721 S_0085F0_CB1_DEST_BASE_ENA(1) |
722 S_0085F0_CB2_DEST_BASE_ENA(1) |
723 S_0085F0_CB3_DEST_BASE_ENA(1) |
724 S_0085F0_CB4_DEST_BASE_ENA(1) |
725 S_0085F0_CB5_DEST_BASE_ENA(1) |
726 S_0085F0_CB6_DEST_BASE_ENA(1) |
727 S_0085F0_CB7_DEST_BASE_ENA(1);
729 /* Necessary for DCC */
730 if (sctx
->chip_class
>= VI
) {
731 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, 0));
732 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_DATA_TS
) |
740 if (sctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_DB
) {
741 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
742 S_0085F0_DB_DEST_BASE_ENA(1);
745 if (sctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_CB_META
) {
746 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
747 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
748 /* needed for wait for idle in SURFACE_SYNC */
749 assert(sctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_CB
);
751 if (sctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_DB_META
) {
752 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
753 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
754 /* needed for wait for idle in SURFACE_SYNC */
755 assert(sctx
->flags
& SI_CONTEXT_FLUSH_AND_INV_DB
);
758 /* Wait for shader engines to go idle.
759 * VS and PS waits are unnecessary if SURFACE_SYNC is going to wait
760 * for everything including CB/DB cache flushes.
762 if (!(sctx
->flags
& (SI_CONTEXT_FLUSH_AND_INV_CB
|
763 SI_CONTEXT_FLUSH_AND_INV_DB
))) {
764 if (sctx
->flags
& SI_CONTEXT_PS_PARTIAL_FLUSH
) {
765 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
766 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
767 } else if (sctx
->flags
& SI_CONTEXT_VS_PARTIAL_FLUSH
) {
768 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
769 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
772 if (sctx
->flags
& SI_CONTEXT_CS_PARTIAL_FLUSH
) {
773 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
774 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
| EVENT_INDEX(4)));
777 /* VGT state synchronization. */
778 if (sctx
->flags
& SI_CONTEXT_VGT_FLUSH
) {
779 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
780 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
782 if (sctx
->flags
& SI_CONTEXT_VGT_STREAMOUT_SYNC
) {
783 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
784 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC
) | EVENT_INDEX(0));
787 /* Make sure ME is idle (it executes most packets) before continuing.
788 * This prevents read-after-write hazards between PFP and ME.
790 if (cp_coher_cntl
|| (sctx
->flags
& SI_CONTEXT_CS_PARTIAL_FLUSH
)) {
791 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
795 /* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
796 * Therefore, it should be last. Done in PFP.
799 /* ACQUIRE_MEM is only required on a compute ring. */
800 radeon_emit(cs
, PKT3(PKT3_SURFACE_SYNC
, 3, 0));
801 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
802 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
803 radeon_emit(cs
, 0); /* CP_COHER_BASE */
804 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
807 if (sctx
->flags
& R600_CONTEXT_START_PIPELINE_STATS
) {
808 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
809 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
811 } else if (sctx
->flags
& R600_CONTEXT_STOP_PIPELINE_STATS
) {
812 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
813 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
820 static void si_get_draw_start_count(struct si_context
*sctx
,
821 const struct pipe_draw_info
*info
,
822 unsigned *start
, unsigned *count
)
824 if (info
->indirect
) {
825 struct r600_resource
*indirect
=
826 (struct r600_resource
*)info
->indirect
;
827 int *data
= r600_buffer_map_sync_with_rings(&sctx
->b
,
828 indirect
, PIPE_TRANSFER_READ
);
829 data
+= info
->indirect_offset
/sizeof(int);
833 *start
= info
->start
;
834 *count
= info
->count
;
838 void si_ce_pre_draw_synchronization(struct si_context
*sctx
)
840 if (sctx
->ce_need_synchronization
) {
841 radeon_emit(sctx
->ce_ib
, PKT3(PKT3_INCREMENT_CE_COUNTER
, 0, 0));
842 radeon_emit(sctx
->ce_ib
, 1);
844 radeon_emit(sctx
->b
.gfx
.cs
, PKT3(PKT3_WAIT_ON_CE_COUNTER
, 0, 0));
845 radeon_emit(sctx
->b
.gfx
.cs
, 1);
849 void si_ce_post_draw_synchronization(struct si_context
*sctx
)
851 if (sctx
->ce_need_synchronization
) {
852 radeon_emit(sctx
->b
.gfx
.cs
, PKT3(PKT3_INCREMENT_DE_COUNTER
, 0, 0));
853 radeon_emit(sctx
->b
.gfx
.cs
, 0);
855 sctx
->ce_need_synchronization
= false;
859 void si_draw_vbo(struct pipe_context
*ctx
, const struct pipe_draw_info
*info
)
861 struct si_context
*sctx
= (struct si_context
*)ctx
;
862 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
863 struct pipe_index_buffer ib
= {};
864 unsigned mask
, dirty_fb_counter
, dirty_tex_counter
;
866 if (!info
->count
&& !info
->indirect
&&
867 (info
->indexed
|| !info
->count_from_stream_output
))
870 if (!sctx
->vs_shader
.cso
) {
874 if (!sctx
->ps_shader
.cso
&& (!rs
|| !rs
->rasterizer_discard
)) {
878 if (!!sctx
->tes_shader
.cso
!= (info
->mode
== PIPE_PRIM_PATCHES
)) {
883 /* Re-emit the framebuffer state if needed. */
884 dirty_fb_counter
= p_atomic_read(&sctx
->b
.screen
->dirty_fb_counter
);
885 if (dirty_fb_counter
!= sctx
->b
.last_dirty_fb_counter
) {
886 sctx
->b
.last_dirty_fb_counter
= dirty_fb_counter
;
887 sctx
->framebuffer
.dirty_cbufs
|=
888 ((1 << sctx
->framebuffer
.state
.nr_cbufs
) - 1);
889 sctx
->framebuffer
.dirty_zsbuf
= true;
890 si_mark_atom_dirty(sctx
, &sctx
->framebuffer
.atom
);
893 /* Invalidate & recompute texture descriptors if needed. */
894 dirty_tex_counter
= p_atomic_read(&sctx
->b
.screen
->dirty_tex_descriptor_counter
);
895 if (dirty_tex_counter
!= sctx
->b
.last_dirty_tex_descriptor_counter
) {
896 sctx
->b
.last_dirty_tex_descriptor_counter
= dirty_tex_counter
;
897 si_update_all_texture_descriptors(sctx
);
900 si_decompress_graphics_textures(sctx
);
902 /* Set the rasterization primitive type.
904 * This must be done after si_decompress_textures, which can call
905 * draw_vbo recursively, and before si_update_shaders, which uses
906 * current_rast_prim for this draw_vbo call. */
907 if (sctx
->gs_shader
.cso
)
908 sctx
->current_rast_prim
= sctx
->gs_shader
.cso
->gs_output_prim
;
909 else if (sctx
->tes_shader
.cso
)
910 sctx
->current_rast_prim
=
911 sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
913 sctx
->current_rast_prim
= info
->mode
;
915 if (!si_update_shaders(sctx
) ||
916 !si_upload_graphics_shader_descriptors(sctx
))
920 /* Initialize the index buffer struct. */
921 pipe_resource_reference(&ib
.buffer
, sctx
->index_buffer
.buffer
);
922 ib
.user_buffer
= sctx
->index_buffer
.user_buffer
;
923 ib
.index_size
= sctx
->index_buffer
.index_size
;
924 ib
.offset
= sctx
->index_buffer
.offset
;
926 /* Translate or upload, if needed. */
927 /* 8-bit indices are supported on VI. */
928 if (sctx
->b
.chip_class
<= CIK
&& ib
.index_size
== 1) {
929 struct pipe_resource
*out_buffer
= NULL
;
930 unsigned out_offset
, start
, count
, start_offset
;
933 si_get_draw_start_count(sctx
, info
, &start
, &count
);
934 start_offset
= start
* ib
.index_size
;
936 u_upload_alloc(sctx
->b
.uploader
, start_offset
, count
* 2, 256,
937 &out_offset
, &out_buffer
, &ptr
);
939 pipe_resource_reference(&ib
.buffer
, NULL
);
943 util_shorten_ubyte_elts_to_userptr(&sctx
->b
.b
, &ib
, 0,
944 ib
.offset
+ start_offset
,
947 pipe_resource_reference(&ib
.buffer
, NULL
);
948 ib
.user_buffer
= NULL
;
949 ib
.buffer
= out_buffer
;
950 /* info->start will be added by the drawing code */
951 ib
.offset
= out_offset
- start_offset
;
953 } else if (ib
.user_buffer
&& !ib
.buffer
) {
954 unsigned start
, count
, start_offset
;
956 si_get_draw_start_count(sctx
, info
, &start
, &count
);
957 start_offset
= start
* ib
.index_size
;
959 u_upload_data(sctx
->b
.uploader
, start_offset
, count
* ib
.index_size
,
960 256, (char*)ib
.user_buffer
+ start_offset
,
961 &ib
.offset
, &ib
.buffer
);
964 /* info->start will be added by the drawing code */
965 ib
.offset
-= start_offset
;
969 /* VI reads index buffers through TC L2. */
970 if (info
->indexed
&& sctx
->b
.chip_class
<= CIK
&&
971 r600_resource(ib
.buffer
)->TC_L2_dirty
) {
972 sctx
->b
.flags
|= SI_CONTEXT_INV_GLOBAL_L2
;
973 r600_resource(ib
.buffer
)->TC_L2_dirty
= false;
976 /* Check flush flags. */
978 si_mark_atom_dirty(sctx
, sctx
->atoms
.s
.cache_flush
);
980 si_need_cs_space(sctx
);
983 mask
= sctx
->dirty_atoms
;
985 struct r600_atom
*atom
= sctx
->atoms
.array
[u_bit_scan(&mask
)];
987 atom
->emit(&sctx
->b
, atom
);
989 sctx
->dirty_atoms
= 0;
991 si_pm4_emit_dirty(sctx
);
992 si_emit_scratch_reloc(sctx
);
993 si_emit_rasterizer_prim_state(sctx
);
994 si_emit_draw_registers(sctx
, info
);
996 si_ce_pre_draw_synchronization(sctx
);
998 si_emit_draw_packets(sctx
, info
, &ib
);
1000 si_ce_post_draw_synchronization(sctx
);
1002 if (sctx
->trace_buf
)
1003 si_trace_emit(sctx
);
1005 /* Workaround for a VGT hang when streamout is enabled.
1006 * It must be done after drawing. */
1007 if ((sctx
->b
.family
== CHIP_HAWAII
||
1008 sctx
->b
.family
== CHIP_TONGA
||
1009 sctx
->b
.family
== CHIP_FIJI
) &&
1010 r600_get_strmout_en(&sctx
->b
)) {
1011 sctx
->b
.flags
|= SI_CONTEXT_VGT_STREAMOUT_SYNC
;
1014 /* Set the depth buffer as dirty. */
1015 if (sctx
->framebuffer
.state
.zsbuf
) {
1016 struct pipe_surface
*surf
= sctx
->framebuffer
.state
.zsbuf
;
1017 struct r600_texture
*rtex
= (struct r600_texture
*)surf
->texture
;
1019 rtex
->dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
1021 if (rtex
->surface
.flags
& RADEON_SURF_SBUFFER
)
1022 rtex
->stencil_dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
1024 if (sctx
->framebuffer
.compressed_cb_mask
) {
1025 struct pipe_surface
*surf
;
1026 struct r600_texture
*rtex
;
1027 unsigned mask
= sctx
->framebuffer
.compressed_cb_mask
;
1030 unsigned i
= u_bit_scan(&mask
);
1031 surf
= sctx
->framebuffer
.state
.cbufs
[i
];
1032 rtex
= (struct r600_texture
*)surf
->texture
;
1034 rtex
->dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
1038 pipe_resource_reference(&ib
.buffer
, NULL
);
1039 sctx
->b
.num_draw_calls
++;
1040 if (G_0286E8_WAVESIZE(sctx
->spi_tmpring_size
))
1041 sctx
->b
.num_spill_draw_calls
++;
1044 void si_trace_emit(struct si_context
*sctx
)
1046 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
1049 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
, sctx
->trace_buf
,
1050 RADEON_USAGE_READWRITE
, RADEON_PRIO_TRACE
);
1051 radeon_emit(cs
, PKT3(PKT3_WRITE_DATA
, 3, 0));
1052 radeon_emit(cs
, S_370_DST_SEL(V_370_MEMORY_SYNC
) |
1053 S_370_WR_CONFIRM(1) |
1054 S_370_ENGINE_SEL(V_370_ME
));
1055 radeon_emit(cs
, sctx
->trace_buf
->gpu_address
);
1056 radeon_emit(cs
, sctx
->trace_buf
->gpu_address
>> 32);
1057 radeon_emit(cs
, sctx
->trace_id
);
1058 radeon_emit(cs
, PKT3(PKT3_NOP
, 0, 0));
1059 radeon_emit(cs
, SI_ENCODE_TRACE_POINT(sctx
->trace_id
));