2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
6 * Copyright © 2015 Advanced Micro Devices, Inc.
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 /* command buffer handling for SI */
30 #include "radv_private.h"
34 #include "radv_util.h"
35 #include "main/macros.h"
37 #define SI_GS_PER_ES 128
40 si_write_harvested_raster_configs(struct radv_physical_device
*physical_device
,
41 struct radeon_winsys_cs
*cs
,
42 unsigned raster_config
,
43 unsigned raster_config_1
)
45 unsigned sh_per_se
= MAX2(physical_device
->rad_info
.max_sh_per_se
, 1);
46 unsigned num_se
= MAX2(physical_device
->rad_info
.max_se
, 1);
47 unsigned rb_mask
= physical_device
->rad_info
.enabled_rb_mask
;
48 unsigned num_rb
= MIN2(physical_device
->rad_info
.num_render_backends
, 16);
49 unsigned rb_per_pkr
= MIN2(num_rb
/ num_se
/ sh_per_se
, 2);
50 unsigned rb_per_se
= num_rb
/ num_se
;
54 se_mask
[0] = ((1 << rb_per_se
) - 1) & rb_mask
;
55 se_mask
[1] = (se_mask
[0] << rb_per_se
) & rb_mask
;
56 se_mask
[2] = (se_mask
[1] << rb_per_se
) & rb_mask
;
57 se_mask
[3] = (se_mask
[2] << rb_per_se
) & rb_mask
;
59 assert(num_se
== 1 || num_se
== 2 || num_se
== 4);
60 assert(sh_per_se
== 1 || sh_per_se
== 2);
61 assert(rb_per_pkr
== 1 || rb_per_pkr
== 2);
63 /* XXX: I can't figure out what the *_XSEL and *_YSEL
64 * fields are for, so I'm leaving them as their default
67 if ((num_se
> 2) && ((!se_mask
[0] && !se_mask
[1]) ||
68 (!se_mask
[2] && !se_mask
[3]))) {
69 raster_config_1
&= C_028354_SE_PAIR_MAP
;
71 if (!se_mask
[0] && !se_mask
[1]) {
73 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_3
);
76 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_0
);
80 for (se
= 0; se
< num_se
; se
++) {
81 unsigned raster_config_se
= raster_config
;
82 unsigned pkr0_mask
= ((1 << rb_per_pkr
) - 1) << (se
* rb_per_se
);
83 unsigned pkr1_mask
= pkr0_mask
<< rb_per_pkr
;
84 int idx
= (se
/ 2) * 2;
86 if ((num_se
> 1) && (!se_mask
[idx
] || !se_mask
[idx
+ 1])) {
87 raster_config_se
&= C_028350_SE_MAP
;
91 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_3
);
94 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_0
);
100 if (rb_per_se
> 2 && (!pkr0_mask
|| !pkr1_mask
)) {
101 raster_config_se
&= C_028350_PKR_MAP
;
105 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_3
);
108 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_0
);
112 if (rb_per_se
>= 2) {
113 unsigned rb0_mask
= 1 << (se
* rb_per_se
);
114 unsigned rb1_mask
= rb0_mask
<< 1;
118 if (!rb0_mask
|| !rb1_mask
) {
119 raster_config_se
&= C_028350_RB_MAP_PKR0
;
123 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_3
);
126 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_0
);
131 rb0_mask
= 1 << (se
* rb_per_se
+ rb_per_pkr
);
132 rb1_mask
= rb0_mask
<< 1;
135 if (!rb0_mask
|| !rb1_mask
) {
136 raster_config_se
&= C_028350_RB_MAP_PKR1
;
140 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_3
);
143 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_0
);
149 /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
150 if (physical_device
->rad_info
.chip_class
< CIK
)
151 radeon_set_config_reg(cs
, GRBM_GFX_INDEX
,
152 SE_INDEX(se
) | SH_BROADCAST_WRITES
|
153 INSTANCE_BROADCAST_WRITES
);
155 radeon_set_uconfig_reg(cs
, R_030800_GRBM_GFX_INDEX
,
156 S_030800_SE_INDEX(se
) | S_030800_SH_BROADCAST_WRITES(1) |
157 S_030800_INSTANCE_BROADCAST_WRITES(1));
158 radeon_set_context_reg(cs
, R_028350_PA_SC_RASTER_CONFIG
, raster_config_se
);
159 if (physical_device
->rad_info
.chip_class
>= CIK
)
160 radeon_set_context_reg(cs
, R_028354_PA_SC_RASTER_CONFIG_1
, raster_config_1
);
163 /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
164 if (physical_device
->rad_info
.chip_class
< CIK
)
165 radeon_set_config_reg(cs
, GRBM_GFX_INDEX
,
166 SE_BROADCAST_WRITES
| SH_BROADCAST_WRITES
|
167 INSTANCE_BROADCAST_WRITES
);
169 radeon_set_uconfig_reg(cs
, R_030800_GRBM_GFX_INDEX
,
170 S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
171 S_030800_INSTANCE_BROADCAST_WRITES(1));
175 si_emit_compute(struct radv_physical_device
*physical_device
,
176 struct radeon_winsys_cs
*cs
)
178 radeon_set_sh_reg_seq(cs
, R_00B810_COMPUTE_START_X
, 3);
183 radeon_set_sh_reg_seq(cs
, R_00B854_COMPUTE_RESOURCE_LIMITS
, 3);
185 /* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1 */
186 radeon_emit(cs
, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
187 radeon_emit(cs
, S_00B85C_SH0_CU_EN(0xffff) | S_00B85C_SH1_CU_EN(0xffff));
189 if (physical_device
->rad_info
.chip_class
>= CIK
) {
190 /* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
191 radeon_set_sh_reg_seq(cs
,
192 R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2
, 2);
193 radeon_emit(cs
, S_00B864_SH0_CU_EN(0xffff) |
194 S_00B864_SH1_CU_EN(0xffff));
195 radeon_emit(cs
, S_00B868_SH0_CU_EN(0xffff) |
196 S_00B868_SH1_CU_EN(0xffff));
199 /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
200 * and is now per pipe, so it should be handled in the
201 * kernel if we want to use something other than the default value,
202 * which is now 0x22f.
204 if (physical_device
->rad_info
.chip_class
<= SI
) {
205 /* XXX: This should be:
206 * (number of compute units) * 4 * (waves per simd) - 1 */
208 radeon_set_sh_reg(cs
, R_00B82C_COMPUTE_MAX_WAVE_ID
,
209 0x190 /* Default value */);
214 si_init_compute(struct radv_cmd_buffer
*cmd_buffer
)
216 struct radv_physical_device
*physical_device
= cmd_buffer
->device
->physical_device
;
217 si_emit_compute(physical_device
, cmd_buffer
->cs
);
221 si_emit_config(struct radv_physical_device
*physical_device
,
222 struct radeon_winsys_cs
*cs
)
224 unsigned num_rb
= MIN2(physical_device
->rad_info
.num_render_backends
, 16);
225 unsigned rb_mask
= physical_device
->rad_info
.enabled_rb_mask
;
226 unsigned raster_config
, raster_config_1
;
229 radeon_emit(cs
, PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
230 radeon_emit(cs
, CONTEXT_CONTROL_LOAD_ENABLE(1));
231 radeon_emit(cs
, CONTEXT_CONTROL_SHADOW_ENABLE(1));
233 radeon_set_context_reg(cs
, R_028A18_VGT_HOS_MAX_TESS_LEVEL
, fui(64));
234 radeon_set_context_reg(cs
, R_028A1C_VGT_HOS_MIN_TESS_LEVEL
, fui(0));
236 /* FIXME calculate these values somehow ??? */
237 radeon_set_context_reg(cs
, R_028A54_VGT_GS_PER_ES
, SI_GS_PER_ES
);
238 radeon_set_context_reg(cs
, R_028A58_VGT_ES_PER_GS
, 0x40);
239 radeon_set_context_reg(cs
, R_028A5C_VGT_GS_PER_VS
, 0x2);
241 radeon_set_context_reg(cs
, R_028A8C_VGT_PRIMITIVEID_RESET
, 0x0);
242 radeon_set_context_reg(cs
, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET
, 0);
244 radeon_set_context_reg(cs
, R_028B98_VGT_STRMOUT_BUFFER_CONFIG
, 0x0);
245 radeon_set_context_reg(cs
, R_028AA0_VGT_INSTANCE_STEP_RATE_0
, 1);
246 if (physical_device
->rad_info
.chip_class
>= GFX9
)
247 radeon_set_context_reg(cs
, R_028AB4_VGT_REUSE_OFF
, 0);
248 radeon_set_context_reg(cs
, R_028AB8_VGT_VTX_CNT_EN
, 0x0);
249 if (physical_device
->rad_info
.chip_class
< CIK
)
250 radeon_set_config_reg(cs
, R_008A14_PA_CL_ENHANCE
, S_008A14_NUM_CLIP_SEQ(3) |
251 S_008A14_CLIP_VTX_REORDER_ENA(1));
253 radeon_set_context_reg(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 0x76543210);
254 radeon_set_context_reg(cs
, R_028BD8_PA_SC_CENTROID_PRIORITY_1
, 0xfedcba98);
256 radeon_set_context_reg(cs
, R_02882C_PA_SU_PRIM_FILTER_CNTL
, 0);
258 for (i
= 0; i
< 16; i
++) {
259 radeon_set_context_reg(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+ i
*8, 0);
260 radeon_set_context_reg(cs
, R_0282D4_PA_SC_VPORT_ZMAX_0
+ i
*8, fui(1.0));
263 switch (physical_device
->rad_info
.family
) {
266 raster_config
= 0x2a00126a;
267 raster_config_1
= 0x00000000;
270 raster_config
= 0x0000124a;
271 raster_config_1
= 0x00000000;
274 raster_config
= 0x00000082;
275 raster_config_1
= 0x00000000;
278 raster_config
= 0x00000000;
279 raster_config_1
= 0x00000000;
282 raster_config
= 0x16000012;
283 raster_config_1
= 0x00000000;
286 raster_config
= 0x3a00161a;
287 raster_config_1
= 0x0000002e;
290 if (physical_device
->rad_info
.cik_macrotile_mode_array
[0] == 0x000000e8) {
291 /* old kernels with old tiling config */
292 raster_config
= 0x16000012;
293 raster_config_1
= 0x0000002a;
295 raster_config
= 0x3a00161a;
296 raster_config_1
= 0x0000002e;
300 raster_config
= 0x16000012;
301 raster_config_1
= 0x0000002a;
305 raster_config
= 0x16000012;
306 raster_config_1
= 0x00000000;
309 raster_config
= 0x16000012;
310 raster_config_1
= 0x0000002a;
314 raster_config
= 0x00000000;
316 raster_config
= 0x00000002;
317 raster_config_1
= 0x00000000;
320 raster_config
= 0x00000002;
321 raster_config_1
= 0x00000000;
324 /* KV should be 0x00000002, but that causes problems with radeon */
325 raster_config
= 0x00000000; /* 0x00000002 */
326 raster_config_1
= 0x00000000;
331 raster_config
= 0x00000000;
332 raster_config_1
= 0x00000000;
335 if (physical_device
->rad_info
.chip_class
<= VI
) {
337 "radeonsi: Unknown GPU, using 0 for raster_config\n");
338 raster_config
= 0x00000000;
339 raster_config_1
= 0x00000000;
344 /* Always use the default config when all backends are enabled
345 * (or when we failed to determine the enabled backends).
347 if (physical_device
->rad_info
.chip_class
<= VI
) {
348 if (!rb_mask
|| util_bitcount(rb_mask
) >= num_rb
) {
349 radeon_set_context_reg(cs
, R_028350_PA_SC_RASTER_CONFIG
,
351 if (physical_device
->rad_info
.chip_class
>= CIK
)
352 radeon_set_context_reg(cs
, R_028354_PA_SC_RASTER_CONFIG_1
,
355 si_write_harvested_raster_configs(physical_device
, cs
, raster_config
, raster_config_1
);
359 radeon_set_context_reg(cs
, R_028204_PA_SC_WINDOW_SCISSOR_TL
, S_028204_WINDOW_OFFSET_DISABLE(1));
360 radeon_set_context_reg(cs
, R_028240_PA_SC_GENERIC_SCISSOR_TL
, S_028240_WINDOW_OFFSET_DISABLE(1));
361 radeon_set_context_reg(cs
, R_028244_PA_SC_GENERIC_SCISSOR_BR
,
362 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
363 radeon_set_context_reg(cs
, R_028030_PA_SC_SCREEN_SCISSOR_TL
, 0);
364 radeon_set_context_reg(cs
, R_028034_PA_SC_SCREEN_SCISSOR_BR
,
365 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
367 radeon_set_context_reg(cs
, R_02820C_PA_SC_CLIPRECT_RULE
, 0xFFFF);
368 radeon_set_context_reg(cs
, R_028230_PA_SC_EDGERULE
, 0xAAAAAAAA);
369 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
370 radeon_set_context_reg(cs
, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET
, 0);
371 radeon_set_context_reg(cs
, R_028820_PA_CL_NANINF_CNTL
, 0);
373 radeon_set_context_reg(cs
, R_028AC0_DB_SRESULTS_COMPARE_STATE0
, 0x0);
374 radeon_set_context_reg(cs
, R_028AC4_DB_SRESULTS_COMPARE_STATE1
, 0x0);
375 radeon_set_context_reg(cs
, R_028AC8_DB_PRELOAD_CONTROL
, 0x0);
376 radeon_set_context_reg(cs
, R_02800C_DB_RENDER_OVERRIDE
,
377 S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE
) |
378 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE
));
380 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
381 radeon_set_context_reg(cs
, R_030920_VGT_MAX_VTX_INDX
, ~0);
382 radeon_set_context_reg(cs
, R_030924_VGT_MIN_VTX_INDX
, 0);
383 radeon_set_context_reg(cs
, R_030928_VGT_INDX_OFFSET
, 0);
385 radeon_set_context_reg(cs
, R_028400_VGT_MAX_VTX_INDX
, ~0);
386 radeon_set_context_reg(cs
, R_028404_VGT_MIN_VTX_INDX
, 0);
387 radeon_set_context_reg(cs
, R_028408_VGT_INDX_OFFSET
, 0);
390 if (physical_device
->rad_info
.chip_class
>= CIK
) {
391 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
392 radeon_set_sh_reg(cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
, S_00B41C_CU_EN(0xffff));
394 radeon_set_sh_reg(cs
, R_00B51C_SPI_SHADER_PGM_RSRC3_LS
, S_00B51C_CU_EN(0xffff));
395 radeon_set_sh_reg(cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
, 0);
396 radeon_set_sh_reg(cs
, R_00B31C_SPI_SHADER_PGM_RSRC3_ES
, S_00B31C_CU_EN(0xffff));
397 /* If this is 0, Bonaire can hang even if GS isn't being used.
398 * Other chips are unaffected. These are suboptimal values,
399 * but we don't use on-chip GS.
401 radeon_set_context_reg(cs
, R_028A44_VGT_GS_ONCHIP_CNTL
,
402 S_028A44_ES_VERTS_PER_SUBGRP(64) |
403 S_028A44_GS_PRIMS_PER_SUBGRP(4));
405 radeon_set_sh_reg(cs
, R_00B21C_SPI_SHADER_PGM_RSRC3_GS
, S_00B21C_CU_EN(0xffff));
407 if (physical_device
->rad_info
.num_good_compute_units
/
408 (physical_device
->rad_info
.max_se
* physical_device
->rad_info
.max_sh_per_se
) <= 4) {
409 /* Too few available compute units per SH. Disallowing
410 * VS to run on CU0 could hurt us more than late VS
411 * allocation would help.
413 * LATE_ALLOC_VS = 2 is the highest safe number.
415 radeon_set_sh_reg(cs
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
, S_00B118_CU_EN(0xffff));
416 radeon_set_sh_reg(cs
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
, S_00B11C_LIMIT(2));
418 /* Set LATE_ALLOC_VS == 31. It should be less than
419 * the number of scratch waves. Limitations:
420 * - VS can't execute on CU0.
421 * - If HS writes outputs to LDS, LS can't execute on CU0.
423 radeon_set_sh_reg(cs
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
, S_00B118_CU_EN(0xfffe));
424 radeon_set_sh_reg(cs
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
, S_00B11C_LIMIT(31));
427 radeon_set_sh_reg(cs
, R_00B01C_SPI_SHADER_PGM_RSRC3_PS
, S_00B01C_CU_EN(0xffff));
430 if (physical_device
->rad_info
.chip_class
>= VI
) {
431 uint32_t vgt_tess_distribution
;
432 radeon_set_context_reg(cs
, R_028424_CB_DCC_CONTROL
,
433 S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
434 S_028424_OVERWRITE_COMBINER_WATERMARK(4));
435 if (physical_device
->rad_info
.family
< CHIP_POLARIS10
)
436 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 30);
437 radeon_set_context_reg(cs
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 32);
439 vgt_tess_distribution
= S_028B50_ACCUM_ISOLINE(32) |
440 S_028B50_ACCUM_TRI(11) |
441 S_028B50_ACCUM_QUAD(11) |
442 S_028B50_DONUT_SPLIT(16);
444 if (physical_device
->rad_info
.family
== CHIP_FIJI
||
445 physical_device
->rad_info
.family
>= CHIP_POLARIS10
)
446 vgt_tess_distribution
|= S_028B50_TRAP_SPLIT(3);
448 radeon_set_context_reg(cs
, R_028B50_VGT_TESS_DISTRIBUTION
,
449 vgt_tess_distribution
);
451 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 14);
452 radeon_set_context_reg(cs
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 16);
455 if (physical_device
->has_rbplus
)
456 radeon_set_context_reg(cs
, R_028C40_PA_SC_SHADER_CONTROL
, 0);
458 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
459 unsigned num_se
= physical_device
->rad_info
.max_se
;
460 unsigned pc_lines
= 0;
462 switch (physical_device
->rad_info
.family
) {
473 radeon_set_context_reg(cs
, R_028060_DB_DFSM_CONTROL
,
474 S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF
));
475 radeon_set_context_reg(cs
, R_028064_DB_RENDER_FILTER
, 0);
476 /* TODO: We can use this to disable RBs for rendering to GART: */
477 radeon_set_context_reg(cs
, R_02835C_PA_SC_TILE_STEERING_OVERRIDE
, 0);
478 radeon_set_context_reg(cs
, R_02883C_PA_SU_OVER_RASTERIZATION_CNTL
, 0);
479 /* TODO: Enable the binner: */
480 radeon_set_context_reg(cs
, R_028C44_PA_SC_BINNER_CNTL_0
,
481 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC
) |
482 S_028C44_DISABLE_START_OF_PRIM(1));
483 radeon_set_context_reg(cs
, R_028C48_PA_SC_BINNER_CNTL_1
,
484 S_028C48_MAX_ALLOC_COUNT(MIN2(128, pc_lines
/ (4 * num_se
))) |
485 S_028C48_MAX_PRIM_PER_BATCH(1023));
486 radeon_set_context_reg(cs
, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL
,
487 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
488 radeon_set_context_reg(cs
, R_030968_VGT_INSTANCE_BASE_ID
, 0);
490 si_emit_compute(physical_device
, cs
);
493 void si_init_config(struct radv_cmd_buffer
*cmd_buffer
)
495 struct radv_physical_device
*physical_device
= cmd_buffer
->device
->physical_device
;
497 si_emit_config(physical_device
, cmd_buffer
->cs
);
501 cik_create_gfx_config(struct radv_device
*device
)
503 struct radeon_winsys_cs
*cs
= device
->ws
->cs_create(device
->ws
, RING_GFX
);
507 si_emit_config(device
->physical_device
, cs
);
509 while (cs
->cdw
& 7) {
510 if (device
->physical_device
->rad_info
.gfx_ib_pad_with_type2
)
511 radeon_emit(cs
, 0x80000000);
513 radeon_emit(cs
, 0xffff1000);
516 device
->gfx_init
= device
->ws
->buffer_create(device
->ws
,
519 RADEON_FLAG_CPU_ACCESS
);
520 if (!device
->gfx_init
)
523 void *map
= device
->ws
->buffer_map(device
->gfx_init
);
525 device
->ws
->buffer_destroy(device
->gfx_init
);
526 device
->gfx_init
= NULL
;
529 memcpy(map
, cs
->buf
, cs
->cdw
* 4);
531 device
->ws
->buffer_unmap(device
->gfx_init
);
532 device
->gfx_init_size_dw
= cs
->cdw
;
534 device
->ws
->cs_destroy(cs
);
538 get_viewport_xform(const VkViewport
*viewport
,
539 float scale
[3], float translate
[3])
541 float x
= viewport
->x
;
542 float y
= viewport
->y
;
543 float half_width
= 0.5f
* viewport
->width
;
544 float half_height
= 0.5f
* viewport
->height
;
545 double n
= viewport
->minDepth
;
546 double f
= viewport
->maxDepth
;
548 scale
[0] = half_width
;
549 translate
[0] = half_width
+ x
;
550 scale
[1] = half_height
;
551 translate
[1] = half_height
+ y
;
558 si_write_viewport(struct radeon_winsys_cs
*cs
, int first_vp
,
559 int count
, const VkViewport
*viewports
)
564 radeon_set_context_reg_seq(cs
, R_02843C_PA_CL_VPORT_XSCALE
+
565 first_vp
* 4 * 6, count
* 6);
567 for (i
= 0; i
< count
; i
++) {
568 float scale
[3], translate
[3];
571 get_viewport_xform(&viewports
[i
], scale
, translate
);
572 radeon_emit(cs
, fui(scale
[0]));
573 radeon_emit(cs
, fui(translate
[0]));
574 radeon_emit(cs
, fui(scale
[1]));
575 radeon_emit(cs
, fui(translate
[1]));
576 radeon_emit(cs
, fui(scale
[2]));
577 radeon_emit(cs
, fui(translate
[2]));
580 radeon_set_context_reg_seq(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+
581 first_vp
* 4 * 2, count
* 2);
582 for (i
= 0; i
< count
; i
++) {
583 float zmin
= MIN2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
584 float zmax
= MAX2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
585 radeon_emit(cs
, fui(zmin
));
586 radeon_emit(cs
, fui(zmax
));
590 static VkRect2D
si_scissor_from_viewport(const VkViewport
*viewport
)
592 float scale
[3], translate
[3];
595 get_viewport_xform(viewport
, scale
, translate
);
597 rect
.offset
.x
= translate
[0] - abs(scale
[0]);
598 rect
.offset
.y
= translate
[1] - abs(scale
[1]);
599 rect
.extent
.width
= ceilf(translate
[0] + abs(scale
[0])) - rect
.offset
.x
;
600 rect
.extent
.height
= ceilf(translate
[1] + abs(scale
[1])) - rect
.offset
.y
;
605 static VkRect2D
si_intersect_scissor(const VkRect2D
*a
, const VkRect2D
*b
) {
607 ret
.offset
.x
= MAX2(a
->offset
.x
, b
->offset
.x
);
608 ret
.offset
.y
= MAX2(a
->offset
.y
, b
->offset
.y
);
609 ret
.extent
.width
= MIN2(a
->offset
.x
+ a
->extent
.width
,
610 b
->offset
.x
+ b
->extent
.width
) - ret
.offset
.x
;
611 ret
.extent
.height
= MIN2(a
->offset
.y
+ a
->extent
.height
,
612 b
->offset
.y
+ b
->extent
.height
) - ret
.offset
.y
;
617 si_write_scissors(struct radeon_winsys_cs
*cs
, int first
,
618 int count
, const VkRect2D
*scissors
,
619 const VkViewport
*viewports
, bool can_use_guardband
)
622 float scale
[3], translate
[3], guardband_x
= INFINITY
, guardband_y
= INFINITY
;
623 const float max_range
= 32767.0f
;
626 radeon_set_context_reg_seq(cs
, R_028250_PA_SC_VPORT_SCISSOR_0_TL
+ first
* 4 * 2, count
* 2);
627 for (i
= 0; i
< count
; i
++) {
628 VkRect2D viewport_scissor
= si_scissor_from_viewport(viewports
+ i
);
629 VkRect2D scissor
= si_intersect_scissor(&scissors
[i
], &viewport_scissor
);
631 get_viewport_xform(viewports
+ i
, scale
, translate
);
632 scale
[0] = abs(scale
[0]);
633 scale
[1] = abs(scale
[1]);
640 guardband_x
= MIN2(guardband_x
, (max_range
- abs(translate
[0])) / scale
[0]);
641 guardband_y
= MIN2(guardband_y
, (max_range
- abs(translate
[1])) / scale
[1]);
643 radeon_emit(cs
, S_028250_TL_X(scissor
.offset
.x
) |
644 S_028250_TL_Y(scissor
.offset
.y
) |
645 S_028250_WINDOW_OFFSET_DISABLE(1));
646 radeon_emit(cs
, S_028254_BR_X(scissor
.offset
.x
+ scissor
.extent
.width
) |
647 S_028254_BR_Y(scissor
.offset
.y
+ scissor
.extent
.height
));
649 if (!can_use_guardband
) {
654 radeon_set_context_reg_seq(cs
, R_028BE8_PA_CL_GB_VERT_CLIP_ADJ
, 4);
655 radeon_emit(cs
, fui(guardband_y
));
656 radeon_emit(cs
, fui(1.0));
657 radeon_emit(cs
, fui(guardband_x
));
658 radeon_emit(cs
, fui(1.0));
661 static inline unsigned
662 radv_prims_for_vertices(struct radv_prim_vertex_count
*info
, unsigned num
)
673 return 1 + ((num
- info
->min
) / info
->incr
);
677 si_get_ia_multi_vgt_param(struct radv_cmd_buffer
*cmd_buffer
,
678 bool instanced_draw
, bool indirect_draw
,
679 uint32_t draw_vertex_count
)
681 enum chip_class chip_class
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
;
682 enum radeon_family family
= cmd_buffer
->device
->physical_device
->rad_info
.family
;
683 struct radeon_info
*info
= &cmd_buffer
->device
->physical_device
->rad_info
;
684 unsigned prim
= cmd_buffer
->state
.pipeline
->graphics
.prim
;
685 unsigned primgroup_size
= 128; /* recommended without a GS */
686 unsigned max_primgroup_in_wave
= 2;
687 /* SWITCH_ON_EOP(0) is always preferable. */
688 bool wd_switch_on_eop
= false;
689 bool ia_switch_on_eop
= false;
690 bool ia_switch_on_eoi
= false;
691 bool partial_vs_wave
= false;
692 bool partial_es_wave
= false;
693 uint32_t num_prims
= radv_prims_for_vertices(&cmd_buffer
->state
.pipeline
->graphics
.prim_vertex_count
, draw_vertex_count
);
694 bool multi_instances_smaller_than_primgroup
;
696 if (radv_pipeline_has_tess(cmd_buffer
->state
.pipeline
))
697 primgroup_size
= cmd_buffer
->state
.pipeline
->graphics
.tess
.num_patches
;
698 else if (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
))
699 primgroup_size
= 64; /* recommended with a GS */
701 multi_instances_smaller_than_primgroup
= indirect_draw
|| (instanced_draw
&&
702 num_prims
< primgroup_size
);
703 if (radv_pipeline_has_tess(cmd_buffer
->state
.pipeline
)) {
704 /* SWITCH_ON_EOI must be set if PrimID is used. */
705 if (cmd_buffer
->state
.pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.uses_prim_id
||
706 cmd_buffer
->state
.pipeline
->shaders
[MESA_SHADER_TESS_EVAL
]->info
.tes
.uses_prim_id
)
707 ia_switch_on_eoi
= true;
709 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
710 if ((family
== CHIP_TAHITI
||
711 family
== CHIP_PITCAIRN
||
712 family
== CHIP_BONAIRE
) &&
713 radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
))
714 partial_vs_wave
= true;
716 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
717 if (cmd_buffer
->device
->has_distributed_tess
) {
718 if (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
)) {
719 partial_es_wave
= true;
721 if (family
== CHIP_TONGA
||
722 family
== CHIP_FIJI
||
723 family
== CHIP_POLARIS10
||
724 family
== CHIP_POLARIS11
||
725 family
== CHIP_POLARIS12
)
726 partial_vs_wave
= true;
728 partial_vs_wave
= true;
732 /* TODO linestipple */
734 if (chip_class
>= CIK
) {
735 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
736 * 4 shader engines. Set 1 to pass the assertion below.
737 * The other cases are hardware requirements. */
738 if (info
->max_se
< 4 ||
739 prim
== V_008958_DI_PT_POLYGON
||
740 prim
== V_008958_DI_PT_LINELOOP
||
741 prim
== V_008958_DI_PT_TRIFAN
||
742 prim
== V_008958_DI_PT_TRISTRIP_ADJ
||
743 (cmd_buffer
->state
.pipeline
->graphics
.prim_restart_enable
&&
744 (family
< CHIP_POLARIS10
||
745 (prim
!= V_008958_DI_PT_POINTLIST
&&
746 prim
!= V_008958_DI_PT_LINESTRIP
&&
747 prim
!= V_008958_DI_PT_TRISTRIP
))))
748 wd_switch_on_eop
= true;
750 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
751 * We don't know that for indirect drawing, so treat it as
752 * always problematic. */
753 if (family
== CHIP_HAWAII
&&
754 (instanced_draw
|| indirect_draw
))
755 wd_switch_on_eop
= true;
757 /* Performance recommendation for 4 SE Gfx7-8 parts if
758 * instances are smaller than a primgroup.
759 * Assume indirect draws always use small instances.
760 * This is needed for good VS wave utilization.
762 if (chip_class
<= VI
&&
764 multi_instances_smaller_than_primgroup
)
765 wd_switch_on_eop
= true;
767 /* Required on CIK and later. */
768 if (info
->max_se
> 2 && !wd_switch_on_eop
)
769 ia_switch_on_eoi
= true;
771 /* Required by Hawaii and, for some special cases, by VI. */
772 if (ia_switch_on_eoi
&&
773 (family
== CHIP_HAWAII
||
775 (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
) || max_primgroup_in_wave
!= 2))))
776 partial_vs_wave
= true;
778 /* Instancing bug on Bonaire. */
779 if (family
== CHIP_BONAIRE
&& ia_switch_on_eoi
&&
780 (instanced_draw
|| indirect_draw
))
781 partial_vs_wave
= true;
783 /* If the WD switch is false, the IA switch must be false too. */
784 assert(wd_switch_on_eop
|| !ia_switch_on_eop
);
786 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
787 if (ia_switch_on_eoi
)
788 partial_es_wave
= true;
790 if (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
)) {
791 /* GS requirement. */
792 if (SI_GS_PER_ES
/ primgroup_size
>= cmd_buffer
->device
->gs_table_depth
- 3)
793 partial_es_wave
= true;
795 /* Hw bug with single-primitive instances and SWITCH_ON_EOI
796 * on multi-SE chips. */
797 if (info
->max_se
>= 2 && ia_switch_on_eoi
&&
798 ((instanced_draw
|| indirect_draw
) &&
800 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_VGT_FLUSH
;
803 return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop
) |
804 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi
) |
805 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave
) |
806 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave
) |
807 S_028AA8_PRIMGROUP_SIZE(primgroup_size
- 1) |
808 S_028AA8_WD_SWITCH_ON_EOP(chip_class
>= CIK
? wd_switch_on_eop
: 0) |
809 S_028AA8_MAX_PRIMGRP_IN_WAVE(chip_class
>= VI
?
810 max_primgroup_in_wave
: 0);
814 void si_cs_emit_write_event_eop(struct radeon_winsys_cs
*cs
,
815 enum chip_class chip_class
,
817 unsigned event
, unsigned event_flags
,
823 unsigned op
= EVENT_TYPE(event
) |
826 unsigned is_gfx8_mec
= is_mec
&& chip_class
< GFX9
;
828 if (chip_class
>= GFX9
|| is_gfx8_mec
) {
829 radeon_emit(cs
, PKT3(PKT3_RELEASE_MEM
, is_gfx8_mec
? 5 : 6, 0));
831 radeon_emit(cs
, EOP_DATA_SEL(data_sel
));
832 radeon_emit(cs
, va
); /* address lo */
833 radeon_emit(cs
, va
>> 32); /* address hi */
834 radeon_emit(cs
, new_fence
); /* immediate data lo */
835 radeon_emit(cs
, 0); /* immediate data hi */
837 radeon_emit(cs
, 0); /* unused */
839 if (chip_class
== CIK
||
841 /* Two EOP events are required to make all engines go idle
842 * (and optional cache flushes executed) before the timestamp
845 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, 0));
848 radeon_emit(cs
, ((va
>> 32) & 0xffff) | EOP_DATA_SEL(data_sel
));
849 radeon_emit(cs
, old_fence
); /* immediate data */
850 radeon_emit(cs
, 0); /* unused */
853 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, 0));
856 radeon_emit(cs
, ((va
>> 32) & 0xffff) | EOP_DATA_SEL(data_sel
));
857 radeon_emit(cs
, new_fence
); /* immediate data */
858 radeon_emit(cs
, 0); /* unused */
863 si_emit_wait_fence(struct radeon_winsys_cs
*cs
,
864 uint64_t va
, uint32_t ref
,
867 radeon_emit(cs
, PKT3(PKT3_WAIT_REG_MEM
, 5, 0));
868 radeon_emit(cs
, WAIT_REG_MEM_EQUAL
| WAIT_REG_MEM_MEM_SPACE(1));
870 radeon_emit(cs
, va
>> 32);
871 radeon_emit(cs
, ref
); /* reference value */
872 radeon_emit(cs
, mask
); /* mask */
873 radeon_emit(cs
, 4); /* poll interval */
877 si_emit_acquire_mem(struct radeon_winsys_cs
*cs
,
878 bool is_mec
, bool is_gfx9
,
879 unsigned cp_coher_cntl
)
881 if (is_mec
|| is_gfx9
) {
882 uint32_t hi_val
= is_gfx9
? 0xffffff : 0xff;
883 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 5, 0) |
884 PKT3_SHADER_TYPE_S(is_mec
));
885 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
886 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
887 radeon_emit(cs
, hi_val
); /* CP_COHER_SIZE_HI */
888 radeon_emit(cs
, 0); /* CP_COHER_BASE */
889 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
890 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
892 /* ACQUIRE_MEM is only required on a compute ring. */
893 radeon_emit(cs
, PKT3(PKT3_SURFACE_SYNC
, 3, 0));
894 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
895 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
896 radeon_emit(cs
, 0); /* CP_COHER_BASE */
897 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
902 si_cs_emit_cache_flush(struct radeon_winsys_cs
*cs
,
903 enum chip_class chip_class
,
907 enum radv_cmd_flush_bits flush_bits
)
909 unsigned cp_coher_cntl
= 0;
910 uint32_t flush_cb_db
= flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
911 RADV_CMD_FLAG_FLUSH_AND_INV_DB
);
913 if (flush_bits
& RADV_CMD_FLAG_INV_ICACHE
)
914 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
915 if (flush_bits
& RADV_CMD_FLAG_INV_SMEM_L1
)
916 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
918 if (chip_class
<= VI
) {
919 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB
) {
920 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
921 S_0085F0_CB0_DEST_BASE_ENA(1) |
922 S_0085F0_CB1_DEST_BASE_ENA(1) |
923 S_0085F0_CB2_DEST_BASE_ENA(1) |
924 S_0085F0_CB3_DEST_BASE_ENA(1) |
925 S_0085F0_CB4_DEST_BASE_ENA(1) |
926 S_0085F0_CB5_DEST_BASE_ENA(1) |
927 S_0085F0_CB6_DEST_BASE_ENA(1) |
928 S_0085F0_CB7_DEST_BASE_ENA(1);
930 /* Necessary for DCC */
931 if (chip_class
>= VI
) {
932 si_cs_emit_write_event_eop(cs
,
935 V_028A90_FLUSH_AND_INV_CB_DATA_TS
,
939 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB
) {
940 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
941 S_0085F0_DB_DEST_BASE_ENA(1);
945 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
) {
946 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
947 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
950 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
) {
951 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
952 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
956 if (flush_bits
& RADV_CMD_FLAG_PS_PARTIAL_FLUSH
) {
957 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
958 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
959 } else if (flush_bits
& RADV_CMD_FLAG_VS_PARTIAL_FLUSH
) {
960 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
961 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
965 if (flush_bits
& RADV_CMD_FLAG_CS_PARTIAL_FLUSH
) {
966 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
967 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
970 if (chip_class
>= GFX9
&& flush_cb_db
) {
971 unsigned cb_db_event
, tc_flags
;
973 /* Set the CB/DB flush event. */
974 switch (flush_cb_db
) {
975 case RADV_CMD_FLAG_FLUSH_AND_INV_CB
:
976 cb_db_event
= V_028A90_FLUSH_AND_INV_CB_DATA_TS
;
978 case RADV_CMD_FLAG_FLUSH_AND_INV_DB
:
979 cb_db_event
= V_028A90_FLUSH_AND_INV_DB_DATA_TS
;
983 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
986 /* TC | TC_WB = invalidate L2 data
987 * TC_MD | TC_WB = invalidate L2 metadata
988 * TC | TC_WB | TC_MD = invalidate L2 data & metadata
990 * The metadata cache must always be invalidated for coherency
991 * between CB/DB and shaders. (metadata = HTILE, CMASK, DCC)
993 * TC must be invalidated on GFX9 only if the CB/DB surface is
994 * not pipe-aligned. If the surface is RB-aligned, it might not
995 * strictly be pipe-aligned since RB alignment takes precendence.
997 tc_flags
= EVENT_TC_WB_ACTION_ENA
|
998 EVENT_TC_MD_ACTION_ENA
;
1000 /* Ideally flush TC together with CB/DB. */
1001 if (flush_bits
& RADV_CMD_FLAG_INV_GLOBAL_L2
) {
1002 tc_flags
|= EVENT_TC_ACTION_ENA
|
1003 EVENT_TCL1_ACTION_ENA
;
1005 /* Clear the flags. */
1006 flush_bits
&= ~(RADV_CMD_FLAG_INV_GLOBAL_L2
|
1007 RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2
|
1008 RADV_CMD_FLAG_INV_VMEM_L1
);
1011 uint32_t old_fence
= (*flush_cnt
)++;
1013 si_cs_emit_write_event_eop(cs
, chip_class
, false, cb_db_event
, tc_flags
, 1,
1014 flush_va
, old_fence
, *flush_cnt
);
1015 si_emit_wait_fence(cs
, flush_va
, *flush_cnt
, 0xffffffff);
1018 /* VGT state sync */
1019 if (flush_bits
& RADV_CMD_FLAG_VGT_FLUSH
) {
1020 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1021 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
1024 /* Make sure ME is idle (it executes most packets) before continuing.
1025 * This prevents read-after-write hazards between PFP and ME.
1027 if ((cp_coher_cntl
||
1028 (flush_bits
& (RADV_CMD_FLAG_CS_PARTIAL_FLUSH
|
1029 RADV_CMD_FLAG_INV_VMEM_L1
|
1030 RADV_CMD_FLAG_INV_GLOBAL_L2
|
1031 RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2
))) &&
1033 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
1037 if ((flush_bits
& RADV_CMD_FLAG_INV_GLOBAL_L2
) ||
1038 (chip_class
<= CIK
&& (flush_bits
& RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2
))) {
1039 si_emit_acquire_mem(cs
, is_mec
, chip_class
>= GFX9
,
1041 S_0085F0_TC_ACTION_ENA(1) |
1042 S_0085F0_TCL1_ACTION_ENA(1) |
1043 S_0301F0_TC_WB_ACTION_ENA(chip_class
>= VI
));
1046 if(flush_bits
& RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2
) {
1048 * NC = apply to non-coherent MTYPEs
1049 * (i.e. MTYPE <= 1, which is what we use everywhere)
1051 * WB doesn't work without NC.
1053 si_emit_acquire_mem(cs
, is_mec
, chip_class
>= GFX9
,
1055 S_0301F0_TC_WB_ACTION_ENA(1) |
1056 S_0301F0_TC_NC_ACTION_ENA(1));
1059 if (flush_bits
& RADV_CMD_FLAG_INV_VMEM_L1
) {
1060 si_emit_acquire_mem(cs
, is_mec
, chip_class
>= GFX9
,
1062 S_0085F0_TCL1_ACTION_ENA(1));
1067 /* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
1068 * Therefore, it should be last. Done in PFP.
1071 si_emit_acquire_mem(cs
, is_mec
, chip_class
>= GFX9
, cp_coher_cntl
);
1075 si_emit_cache_flush(struct radv_cmd_buffer
*cmd_buffer
)
1077 bool is_compute
= cmd_buffer
->queue_family_index
== RADV_QUEUE_COMPUTE
;
1078 enum chip_class chip_class
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
;
1080 cmd_buffer
->state
.flush_bits
&= ~(RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
1081 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
|
1082 RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
1083 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
|
1084 RADV_CMD_FLAG_PS_PARTIAL_FLUSH
|
1085 RADV_CMD_FLAG_VS_PARTIAL_FLUSH
|
1086 RADV_CMD_FLAG_VGT_FLUSH
);
1088 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 128);
1090 uint32_t *ptr
= NULL
;
1092 if (chip_class
== GFX9
) {
1093 va
= cmd_buffer
->device
->ws
->buffer_get_va(cmd_buffer
->gfx9_fence_bo
) + cmd_buffer
->gfx9_fence_offset
;
1094 ptr
= &cmd_buffer
->gfx9_fence_idx
;
1096 si_cs_emit_cache_flush(cmd_buffer
->cs
,
1097 cmd_buffer
->device
->physical_device
->rad_info
.chip_class
,
1099 radv_cmd_buffer_uses_mec(cmd_buffer
),
1100 cmd_buffer
->state
.flush_bits
);
1103 if (cmd_buffer
->state
.flush_bits
)
1104 radv_cmd_buffer_trace_emit(cmd_buffer
);
1105 cmd_buffer
->state
.flush_bits
= 0;
1109 /* Set this if you want the 3D engine to wait until CP DMA is done.
1110 * It should be set on the last CP DMA packet. */
1111 #define CP_DMA_SYNC (1 << 0)
1113 /* Set this if the source data was used as a destination in a previous CP DMA
1114 * packet. It's for preventing a read-after-write (RAW) hazard between two
1115 * CP DMA packets. */
1116 #define CP_DMA_RAW_WAIT (1 << 1)
1117 #define CP_DMA_USE_L2 (1 << 2)
1118 #define CP_DMA_CLEAR (1 << 3)
1120 /* Alignment for optimal performance. */
1121 #define SI_CPDMA_ALIGNMENT 32
1123 /* The max number of bytes that can be copied per packet. */
1124 static inline unsigned cp_dma_max_byte_count(struct radv_cmd_buffer
*cmd_buffer
)
1126 unsigned max
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
?
1127 S_414_BYTE_COUNT_GFX9(~0u) :
1128 S_414_BYTE_COUNT_GFX6(~0u);
1130 /* make it aligned for optimal performance */
1131 return max
& ~(SI_CPDMA_ALIGNMENT
- 1);
1134 /* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
1135 * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
1138 static void si_emit_cp_dma(struct radv_cmd_buffer
*cmd_buffer
,
1139 uint64_t dst_va
, uint64_t src_va
,
1140 unsigned size
, unsigned flags
)
1142 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
1143 uint32_t header
= 0, command
= 0;
1146 assert(size
<= cp_dma_max_byte_count(cmd_buffer
));
1148 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 9);
1149 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1150 command
|= S_414_BYTE_COUNT_GFX9(size
);
1152 command
|= S_414_BYTE_COUNT_GFX6(size
);
1155 if (flags
& CP_DMA_SYNC
)
1156 header
|= S_411_CP_SYNC(1);
1158 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1159 command
|= S_414_DISABLE_WR_CONFIRM_GFX9(1);
1161 command
|= S_414_DISABLE_WR_CONFIRM_GFX6(1);
1164 if (flags
& CP_DMA_RAW_WAIT
)
1165 command
|= S_414_RAW_WAIT(1);
1167 /* Src and dst flags. */
1168 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1169 !(flags
& CP_DMA_CLEAR
) &&
1171 header
|= S_411_DSL_SEL(V_411_NOWHERE
); /* prefetch only */
1172 else if (flags
& CP_DMA_USE_L2
)
1173 header
|= S_411_DSL_SEL(V_411_DST_ADDR_TC_L2
);
1175 if (flags
& CP_DMA_CLEAR
)
1176 header
|= S_411_SRC_SEL(V_411_DATA
);
1177 else if (flags
& CP_DMA_USE_L2
)
1178 header
|= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2
);
1180 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1181 radeon_emit(cs
, PKT3(PKT3_DMA_DATA
, 5, 0));
1182 radeon_emit(cs
, header
);
1183 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
1184 radeon_emit(cs
, src_va
>> 32); /* SRC_ADDR_HI [31:0] */
1185 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
1186 radeon_emit(cs
, dst_va
>> 32); /* DST_ADDR_HI [31:0] */
1187 radeon_emit(cs
, command
);
1189 header
|= S_411_SRC_ADDR_HI(src_va
>> 32);
1190 radeon_emit(cs
, PKT3(PKT3_CP_DMA
, 4, 0));
1191 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
1192 radeon_emit(cs
, header
); /* SRC_ADDR_HI [15:0] + flags. */
1193 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
1194 radeon_emit(cs
, (dst_va
>> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
1195 radeon_emit(cs
, command
);
1198 /* CP DMA is executed in ME, but index buffers are read by PFP.
1199 * This ensures that ME (CP DMA) is idle before PFP starts fetching
1200 * indices. If we wanted to execute CP DMA in PFP, this packet
1201 * should precede it.
1203 if ((flags
& CP_DMA_SYNC
) && cmd_buffer
->queue_family_index
== RADV_QUEUE_GENERAL
) {
1204 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
1208 radv_cmd_buffer_trace_emit(cmd_buffer
);
1211 void si_cp_dma_prefetch(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1214 uint64_t aligned_va
= va
& ~(SI_CPDMA_ALIGNMENT
- 1);
1215 uint64_t aligned_size
= ((va
+ size
+ SI_CPDMA_ALIGNMENT
-1) & ~(SI_CPDMA_ALIGNMENT
- 1)) - aligned_va
;
1217 si_emit_cp_dma(cmd_buffer
, aligned_va
, aligned_va
,
1218 aligned_size
, CP_DMA_USE_L2
);
1221 static void si_cp_dma_prepare(struct radv_cmd_buffer
*cmd_buffer
, uint64_t byte_count
,
1222 uint64_t remaining_size
, unsigned *flags
)
1225 /* Flush the caches for the first copy only.
1226 * Also wait for the previous CP DMA operations.
1228 if (cmd_buffer
->state
.flush_bits
) {
1229 si_emit_cache_flush(cmd_buffer
);
1230 *flags
|= CP_DMA_RAW_WAIT
;
1233 /* Do the synchronization after the last dma, so that all data
1234 * is written to memory.
1236 if (byte_count
== remaining_size
)
1237 *flags
|= CP_DMA_SYNC
;
1240 static void si_cp_dma_realign_engine(struct radv_cmd_buffer
*cmd_buffer
, unsigned size
)
1244 unsigned dma_flags
= 0;
1245 unsigned buf_size
= SI_CPDMA_ALIGNMENT
* 2;
1248 assert(size
< SI_CPDMA_ALIGNMENT
);
1250 radv_cmd_buffer_upload_alloc(cmd_buffer
, buf_size
, SI_CPDMA_ALIGNMENT
, &offset
, &ptr
);
1252 va
= cmd_buffer
->device
->ws
->buffer_get_va(cmd_buffer
->upload
.upload_bo
);
1255 si_cp_dma_prepare(cmd_buffer
, size
, size
, &dma_flags
);
1257 si_emit_cp_dma(cmd_buffer
, va
, va
+ SI_CPDMA_ALIGNMENT
, size
,
1261 void si_cp_dma_buffer_copy(struct radv_cmd_buffer
*cmd_buffer
,
1262 uint64_t src_va
, uint64_t dest_va
,
1265 uint64_t main_src_va
, main_dest_va
;
1266 uint64_t skipped_size
= 0, realign_size
= 0;
1269 if (cmd_buffer
->device
->physical_device
->rad_info
.family
<= CHIP_CARRIZO
||
1270 cmd_buffer
->device
->physical_device
->rad_info
.family
== CHIP_STONEY
) {
1271 /* If the size is not aligned, we must add a dummy copy at the end
1272 * just to align the internal counter. Otherwise, the DMA engine
1273 * would slow down by an order of magnitude for following copies.
1275 if (size
% SI_CPDMA_ALIGNMENT
)
1276 realign_size
= SI_CPDMA_ALIGNMENT
- (size
% SI_CPDMA_ALIGNMENT
);
1278 /* If the copy begins unaligned, we must start copying from the next
1279 * aligned block and the skipped part should be copied after everything
1280 * else has been copied. Only the src alignment matters, not dst.
1282 if (src_va
% SI_CPDMA_ALIGNMENT
) {
1283 skipped_size
= SI_CPDMA_ALIGNMENT
- (src_va
% SI_CPDMA_ALIGNMENT
);
1284 /* The main part will be skipped if the size is too small. */
1285 skipped_size
= MIN2(skipped_size
, size
);
1286 size
-= skipped_size
;
1289 main_src_va
= src_va
+ skipped_size
;
1290 main_dest_va
= dest_va
+ skipped_size
;
1293 unsigned dma_flags
= 0;
1294 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(cmd_buffer
));
1296 si_cp_dma_prepare(cmd_buffer
, byte_count
,
1297 size
+ skipped_size
+ realign_size
,
1300 si_emit_cp_dma(cmd_buffer
, main_dest_va
, main_src_va
,
1301 byte_count
, dma_flags
);
1304 main_src_va
+= byte_count
;
1305 main_dest_va
+= byte_count
;
1309 unsigned dma_flags
= 0;
1311 si_cp_dma_prepare(cmd_buffer
, skipped_size
,
1312 size
+ skipped_size
+ realign_size
,
1315 si_emit_cp_dma(cmd_buffer
, dest_va
, src_va
,
1316 skipped_size
, dma_flags
);
1319 si_cp_dma_realign_engine(cmd_buffer
, realign_size
);
1322 void si_cp_dma_clear_buffer(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1323 uint64_t size
, unsigned value
)
1329 assert(va
% 4 == 0 && size
% 4 == 0);
1332 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(cmd_buffer
));
1333 unsigned dma_flags
= CP_DMA_CLEAR
;
1335 si_cp_dma_prepare(cmd_buffer
, byte_count
, size
, &dma_flags
);
1337 /* Emit the clear packet. */
1338 si_emit_cp_dma(cmd_buffer
, va
, value
, byte_count
,
1346 /* For MSAA sample positions. */
1347 #define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y) \
1348 (((s0x) & 0xf) | (((unsigned)(s0y) & 0xf) << 4) | \
1349 (((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) | \
1350 (((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) | \
1351 (((unsigned)(s3x) & 0xf) << 24) | (((unsigned)(s3y) & 0xf) << 28))
1355 * There are two locations (4, 4), (-4, -4). */
1356 const uint32_t eg_sample_locs_2x
[4] = {
1357 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1358 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1359 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1360 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1362 const unsigned eg_max_dist_2x
= 4;
1364 * There are 4 locations: (-2, 6), (6, -2), (-6, 2), (2, 6). */
1365 const uint32_t eg_sample_locs_4x
[4] = {
1366 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1367 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1368 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1369 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1371 const unsigned eg_max_dist_4x
= 6;
1374 static const uint32_t cm_sample_locs_8x
[] = {
1375 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1376 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1377 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1378 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1379 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1380 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1381 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1382 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1384 static const unsigned cm_max_dist_8x
= 8;
1385 /* Cayman 16xMSAA */
1386 static const uint32_t cm_sample_locs_16x
[] = {
1387 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1388 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1389 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1390 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1391 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1392 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1393 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1394 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1395 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1396 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1397 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1398 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1399 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1400 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1401 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1402 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1404 static const unsigned cm_max_dist_16x
= 8;
1406 unsigned radv_cayman_get_maxdist(int log_samples
)
1408 unsigned max_dist
[] = {
1415 return max_dist
[log_samples
];
1418 void radv_cayman_emit_msaa_sample_locs(struct radeon_winsys_cs
*cs
, int nr_samples
)
1420 switch (nr_samples
) {
1423 radeon_set_context_reg(cs
, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 0);
1424 radeon_set_context_reg(cs
, CM_R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, 0);
1425 radeon_set_context_reg(cs
, CM_R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, 0);
1426 radeon_set_context_reg(cs
, CM_R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, 0);
1429 radeon_set_context_reg(cs
, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, eg_sample_locs_2x
[0]);
1430 radeon_set_context_reg(cs
, CM_R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, eg_sample_locs_2x
[1]);
1431 radeon_set_context_reg(cs
, CM_R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, eg_sample_locs_2x
[2]);
1432 radeon_set_context_reg(cs
, CM_R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, eg_sample_locs_2x
[3]);
1435 radeon_set_context_reg(cs
, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, eg_sample_locs_4x
[0]);
1436 radeon_set_context_reg(cs
, CM_R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, eg_sample_locs_4x
[1]);
1437 radeon_set_context_reg(cs
, CM_R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, eg_sample_locs_4x
[2]);
1438 radeon_set_context_reg(cs
, CM_R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, eg_sample_locs_4x
[3]);
1441 radeon_set_context_reg_seq(cs
, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 14);
1442 radeon_emit(cs
, cm_sample_locs_8x
[0]);
1443 radeon_emit(cs
, cm_sample_locs_8x
[4]);
1446 radeon_emit(cs
, cm_sample_locs_8x
[1]);
1447 radeon_emit(cs
, cm_sample_locs_8x
[5]);
1450 radeon_emit(cs
, cm_sample_locs_8x
[2]);
1451 radeon_emit(cs
, cm_sample_locs_8x
[6]);
1454 radeon_emit(cs
, cm_sample_locs_8x
[3]);
1455 radeon_emit(cs
, cm_sample_locs_8x
[7]);
1458 radeon_set_context_reg_seq(cs
, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 16);
1459 radeon_emit(cs
, cm_sample_locs_16x
[0]);
1460 radeon_emit(cs
, cm_sample_locs_16x
[4]);
1461 radeon_emit(cs
, cm_sample_locs_16x
[8]);
1462 radeon_emit(cs
, cm_sample_locs_16x
[12]);
1463 radeon_emit(cs
, cm_sample_locs_16x
[1]);
1464 radeon_emit(cs
, cm_sample_locs_16x
[5]);
1465 radeon_emit(cs
, cm_sample_locs_16x
[9]);
1466 radeon_emit(cs
, cm_sample_locs_16x
[13]);
1467 radeon_emit(cs
, cm_sample_locs_16x
[2]);
1468 radeon_emit(cs
, cm_sample_locs_16x
[6]);
1469 radeon_emit(cs
, cm_sample_locs_16x
[10]);
1470 radeon_emit(cs
, cm_sample_locs_16x
[14]);
1471 radeon_emit(cs
, cm_sample_locs_16x
[3]);
1472 radeon_emit(cs
, cm_sample_locs_16x
[7]);
1473 radeon_emit(cs
, cm_sample_locs_16x
[11]);
1474 radeon_emit(cs
, cm_sample_locs_16x
[15]);
1479 static void radv_cayman_get_sample_position(struct radv_device
*device
,
1480 unsigned sample_count
,
1481 unsigned sample_index
, float *out_value
)
1487 switch (sample_count
) {
1490 out_value
[0] = out_value
[1] = 0.5;
1493 offset
= 4 * (sample_index
* 2);
1494 val
.idx
= (eg_sample_locs_2x
[0] >> offset
) & 0xf;
1495 out_value
[0] = (float)(val
.idx
+ 8) / 16.0f
;
1496 val
.idx
= (eg_sample_locs_2x
[0] >> (offset
+ 4)) & 0xf;
1497 out_value
[1] = (float)(val
.idx
+ 8) / 16.0f
;
1500 offset
= 4 * (sample_index
* 2);
1501 val
.idx
= (eg_sample_locs_4x
[0] >> offset
) & 0xf;
1502 out_value
[0] = (float)(val
.idx
+ 8) / 16.0f
;
1503 val
.idx
= (eg_sample_locs_4x
[0] >> (offset
+ 4)) & 0xf;
1504 out_value
[1] = (float)(val
.idx
+ 8) / 16.0f
;
1507 offset
= 4 * (sample_index
% 4 * 2);
1508 index
= (sample_index
/ 4) * 4;
1509 val
.idx
= (cm_sample_locs_8x
[index
] >> offset
) & 0xf;
1510 out_value
[0] = (float)(val
.idx
+ 8) / 16.0f
;
1511 val
.idx
= (cm_sample_locs_8x
[index
] >> (offset
+ 4)) & 0xf;
1512 out_value
[1] = (float)(val
.idx
+ 8) / 16.0f
;
1515 offset
= 4 * (sample_index
% 4 * 2);
1516 index
= (sample_index
/ 4) * 4;
1517 val
.idx
= (cm_sample_locs_16x
[index
] >> offset
) & 0xf;
1518 out_value
[0] = (float)(val
.idx
+ 8) / 16.0f
;
1519 val
.idx
= (cm_sample_locs_16x
[index
] >> (offset
+ 4)) & 0xf;
1520 out_value
[1] = (float)(val
.idx
+ 8) / 16.0f
;
1525 void radv_device_init_msaa(struct radv_device
*device
)
1528 radv_cayman_get_sample_position(device
, 1, 0, device
->sample_locations_1x
[0]);
1530 for (i
= 0; i
< 2; i
++)
1531 radv_cayman_get_sample_position(device
, 2, i
, device
->sample_locations_2x
[i
]);
1532 for (i
= 0; i
< 4; i
++)
1533 radv_cayman_get_sample_position(device
, 4, i
, device
->sample_locations_4x
[i
]);
1534 for (i
= 0; i
< 8; i
++)
1535 radv_cayman_get_sample_position(device
, 8, i
, device
->sample_locations_8x
[i
]);
1536 for (i
= 0; i
< 16; i
++)
1537 radv_cayman_get_sample_position(device
, 16, i
, device
->sample_locations_16x
[i
]);