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"
33 #include "radv_util.h"
34 #include "main/macros.h"
36 #define SI_GS_PER_ES 128
39 si_write_harvested_raster_configs(struct radv_physical_device
*physical_device
,
40 struct radeon_winsys_cs
*cs
,
41 unsigned raster_config
,
42 unsigned raster_config_1
)
44 unsigned sh_per_se
= MAX2(physical_device
->rad_info
.max_sh_per_se
, 1);
45 unsigned num_se
= MAX2(physical_device
->rad_info
.max_se
, 1);
46 unsigned rb_mask
= physical_device
->rad_info
.enabled_rb_mask
;
47 unsigned num_rb
= MIN2(physical_device
->rad_info
.num_render_backends
, 16);
48 unsigned rb_per_pkr
= MIN2(num_rb
/ num_se
/ sh_per_se
, 2);
49 unsigned rb_per_se
= num_rb
/ num_se
;
53 se_mask
[0] = ((1 << rb_per_se
) - 1) & rb_mask
;
54 se_mask
[1] = (se_mask
[0] << rb_per_se
) & rb_mask
;
55 se_mask
[2] = (se_mask
[1] << rb_per_se
) & rb_mask
;
56 se_mask
[3] = (se_mask
[2] << rb_per_se
) & rb_mask
;
58 assert(num_se
== 1 || num_se
== 2 || num_se
== 4);
59 assert(sh_per_se
== 1 || sh_per_se
== 2);
60 assert(rb_per_pkr
== 1 || rb_per_pkr
== 2);
62 /* XXX: I can't figure out what the *_XSEL and *_YSEL
63 * fields are for, so I'm leaving them as their default
66 if ((num_se
> 2) && ((!se_mask
[0] && !se_mask
[1]) ||
67 (!se_mask
[2] && !se_mask
[3]))) {
68 raster_config_1
&= C_028354_SE_PAIR_MAP
;
70 if (!se_mask
[0] && !se_mask
[1]) {
72 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_3
);
75 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_0
);
79 for (se
= 0; se
< num_se
; se
++) {
80 unsigned raster_config_se
= raster_config
;
81 unsigned pkr0_mask
= ((1 << rb_per_pkr
) - 1) << (se
* rb_per_se
);
82 unsigned pkr1_mask
= pkr0_mask
<< rb_per_pkr
;
83 int idx
= (se
/ 2) * 2;
85 if ((num_se
> 1) && (!se_mask
[idx
] || !se_mask
[idx
+ 1])) {
86 raster_config_se
&= C_028350_SE_MAP
;
90 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_3
);
93 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_0
);
99 if (rb_per_se
> 2 && (!pkr0_mask
|| !pkr1_mask
)) {
100 raster_config_se
&= C_028350_PKR_MAP
;
104 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_3
);
107 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_0
);
111 if (rb_per_se
>= 2) {
112 unsigned rb0_mask
= 1 << (se
* rb_per_se
);
113 unsigned rb1_mask
= rb0_mask
<< 1;
117 if (!rb0_mask
|| !rb1_mask
) {
118 raster_config_se
&= C_028350_RB_MAP_PKR0
;
122 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_3
);
125 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_0
);
130 rb0_mask
= 1 << (se
* rb_per_se
+ rb_per_pkr
);
131 rb1_mask
= rb0_mask
<< 1;
134 if (!rb0_mask
|| !rb1_mask
) {
135 raster_config_se
&= C_028350_RB_MAP_PKR1
;
139 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_3
);
142 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_0
);
148 /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
149 if (physical_device
->rad_info
.chip_class
< CIK
)
150 radeon_set_config_reg(cs
, GRBM_GFX_INDEX
,
151 SE_INDEX(se
) | SH_BROADCAST_WRITES
|
152 INSTANCE_BROADCAST_WRITES
);
154 radeon_set_uconfig_reg(cs
, R_030800_GRBM_GFX_INDEX
,
155 S_030800_SE_INDEX(se
) | S_030800_SH_BROADCAST_WRITES(1) |
156 S_030800_INSTANCE_BROADCAST_WRITES(1));
157 radeon_set_context_reg(cs
, R_028350_PA_SC_RASTER_CONFIG
, raster_config_se
);
158 if (physical_device
->rad_info
.chip_class
>= CIK
)
159 radeon_set_context_reg(cs
, R_028354_PA_SC_RASTER_CONFIG_1
, raster_config_1
);
162 /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
163 if (physical_device
->rad_info
.chip_class
< CIK
)
164 radeon_set_config_reg(cs
, GRBM_GFX_INDEX
,
165 SE_BROADCAST_WRITES
| SH_BROADCAST_WRITES
|
166 INSTANCE_BROADCAST_WRITES
);
168 radeon_set_uconfig_reg(cs
, R_030800_GRBM_GFX_INDEX
,
169 S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
170 S_030800_INSTANCE_BROADCAST_WRITES(1));
174 si_emit_compute(struct radv_physical_device
*physical_device
,
175 struct radeon_winsys_cs
*cs
)
177 radeon_set_sh_reg_seq(cs
, R_00B810_COMPUTE_START_X
, 3);
182 radeon_set_sh_reg_seq(cs
, R_00B854_COMPUTE_RESOURCE_LIMITS
, 3);
184 /* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1 */
185 radeon_emit(cs
, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
186 radeon_emit(cs
, S_00B85C_SH0_CU_EN(0xffff) | S_00B85C_SH1_CU_EN(0xffff));
188 if (physical_device
->rad_info
.chip_class
>= CIK
) {
189 /* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
190 radeon_set_sh_reg_seq(cs
,
191 R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2
, 2);
192 radeon_emit(cs
, S_00B864_SH0_CU_EN(0xffff) |
193 S_00B864_SH1_CU_EN(0xffff));
194 radeon_emit(cs
, S_00B868_SH0_CU_EN(0xffff) |
195 S_00B868_SH1_CU_EN(0xffff));
198 /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
199 * and is now per pipe, so it should be handled in the
200 * kernel if we want to use something other than the default value,
201 * which is now 0x22f.
203 if (physical_device
->rad_info
.chip_class
<= SI
) {
204 /* XXX: This should be:
205 * (number of compute units) * 4 * (waves per simd) - 1 */
207 radeon_set_sh_reg(cs
, R_00B82C_COMPUTE_MAX_WAVE_ID
,
208 0x190 /* Default value */);
213 si_init_compute(struct radv_cmd_buffer
*cmd_buffer
)
215 struct radv_physical_device
*physical_device
= cmd_buffer
->device
->physical_device
;
216 si_emit_compute(physical_device
, cmd_buffer
->cs
);
220 si_emit_config(struct radv_physical_device
*physical_device
,
221 struct radeon_winsys_cs
*cs
)
223 unsigned num_rb
= MIN2(physical_device
->rad_info
.num_render_backends
, 16);
224 unsigned rb_mask
= physical_device
->rad_info
.enabled_rb_mask
;
225 unsigned raster_config
, raster_config_1
;
228 radeon_emit(cs
, PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
229 radeon_emit(cs
, CONTEXT_CONTROL_LOAD_ENABLE(1));
230 radeon_emit(cs
, CONTEXT_CONTROL_SHADOW_ENABLE(1));
232 radeon_set_context_reg(cs
, R_028A18_VGT_HOS_MAX_TESS_LEVEL
, fui(64));
233 radeon_set_context_reg(cs
, R_028A1C_VGT_HOS_MIN_TESS_LEVEL
, fui(0));
235 /* FIXME calculate these values somehow ??? */
236 radeon_set_context_reg(cs
, R_028A54_VGT_GS_PER_ES
, SI_GS_PER_ES
);
237 radeon_set_context_reg(cs
, R_028A58_VGT_ES_PER_GS
, 0x40);
238 radeon_set_context_reg(cs
, R_028A5C_VGT_GS_PER_VS
, 0x2);
240 radeon_set_context_reg(cs
, R_028A8C_VGT_PRIMITIVEID_RESET
, 0x0);
241 radeon_set_context_reg(cs
, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET
, 0);
243 radeon_set_context_reg(cs
, R_028B98_VGT_STRMOUT_BUFFER_CONFIG
, 0x0);
244 radeon_set_context_reg(cs
, R_028AB8_VGT_VTX_CNT_EN
, 0x0);
245 if (physical_device
->rad_info
.chip_class
< CIK
)
246 radeon_set_config_reg(cs
, R_008A14_PA_CL_ENHANCE
, S_008A14_NUM_CLIP_SEQ(3) |
247 S_008A14_CLIP_VTX_REORDER_ENA(1));
249 radeon_set_context_reg(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 0x76543210);
250 radeon_set_context_reg(cs
, R_028BD8_PA_SC_CENTROID_PRIORITY_1
, 0xfedcba98);
252 radeon_set_context_reg(cs
, R_02882C_PA_SU_PRIM_FILTER_CNTL
, 0);
254 for (i
= 0; i
< 16; i
++) {
255 radeon_set_context_reg(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+ i
*8, 0);
256 radeon_set_context_reg(cs
, R_0282D4_PA_SC_VPORT_ZMAX_0
+ i
*8, fui(1.0));
259 switch (physical_device
->rad_info
.family
) {
262 raster_config
= 0x2a00126a;
263 raster_config_1
= 0x00000000;
266 raster_config
= 0x0000124a;
267 raster_config_1
= 0x00000000;
270 raster_config
= 0x00000082;
271 raster_config_1
= 0x00000000;
274 raster_config
= 0x00000000;
275 raster_config_1
= 0x00000000;
278 raster_config
= 0x16000012;
279 raster_config_1
= 0x00000000;
282 raster_config
= 0x3a00161a;
283 raster_config_1
= 0x0000002e;
286 if (physical_device
->rad_info
.cik_macrotile_mode_array
[0] == 0x000000e8) {
287 /* old kernels with old tiling config */
288 raster_config
= 0x16000012;
289 raster_config_1
= 0x0000002a;
291 raster_config
= 0x3a00161a;
292 raster_config_1
= 0x0000002e;
296 raster_config
= 0x16000012;
297 raster_config_1
= 0x0000002a;
300 raster_config
= 0x16000012;
301 raster_config_1
= 0x00000000;
304 raster_config
= 0x16000012;
305 raster_config_1
= 0x0000002a;
309 raster_config
= 0x00000000;
311 raster_config
= 0x00000002;
312 raster_config_1
= 0x00000000;
315 raster_config
= 0x00000002;
316 raster_config_1
= 0x00000000;
319 /* KV should be 0x00000002, but that causes problems with radeon */
320 raster_config
= 0x00000000; /* 0x00000002 */
321 raster_config_1
= 0x00000000;
326 raster_config
= 0x00000000;
327 raster_config_1
= 0x00000000;
331 "radeonsi: Unknown GPU, using 0 for raster_config\n");
332 raster_config
= 0x00000000;
333 raster_config_1
= 0x00000000;
337 /* Always use the default config when all backends are enabled
338 * (or when we failed to determine the enabled backends).
340 if (!rb_mask
|| util_bitcount(rb_mask
) >= num_rb
) {
341 radeon_set_context_reg(cs
, R_028350_PA_SC_RASTER_CONFIG
,
343 if (physical_device
->rad_info
.chip_class
>= CIK
)
344 radeon_set_context_reg(cs
, R_028354_PA_SC_RASTER_CONFIG_1
,
347 si_write_harvested_raster_configs(physical_device
, cs
, raster_config
, raster_config_1
);
350 radeon_set_context_reg(cs
, R_028204_PA_SC_WINDOW_SCISSOR_TL
, S_028204_WINDOW_OFFSET_DISABLE(1));
351 radeon_set_context_reg(cs
, R_028240_PA_SC_GENERIC_SCISSOR_TL
, S_028240_WINDOW_OFFSET_DISABLE(1));
352 radeon_set_context_reg(cs
, R_028244_PA_SC_GENERIC_SCISSOR_BR
,
353 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
354 radeon_set_context_reg(cs
, R_028030_PA_SC_SCREEN_SCISSOR_TL
, 0);
355 radeon_set_context_reg(cs
, R_028034_PA_SC_SCREEN_SCISSOR_BR
,
356 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
358 radeon_set_context_reg(cs
, R_02820C_PA_SC_CLIPRECT_RULE
, 0xFFFF);
359 radeon_set_context_reg(cs
, R_028230_PA_SC_EDGERULE
, 0xAAAAAAAA);
360 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
361 radeon_set_context_reg(cs
, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET
, 0);
362 radeon_set_context_reg(cs
, R_028820_PA_CL_NANINF_CNTL
, 0);
364 radeon_set_context_reg(cs
, R_028BE8_PA_CL_GB_VERT_CLIP_ADJ
, fui(1.0));
365 radeon_set_context_reg(cs
, R_028BEC_PA_CL_GB_VERT_DISC_ADJ
, fui(1.0));
366 radeon_set_context_reg(cs
, R_028BF0_PA_CL_GB_HORZ_CLIP_ADJ
, fui(1.0));
367 radeon_set_context_reg(cs
, R_028BF4_PA_CL_GB_HORZ_DISC_ADJ
, fui(1.0));
369 radeon_set_context_reg(cs
, R_028AC0_DB_SRESULTS_COMPARE_STATE0
, 0x0);
370 radeon_set_context_reg(cs
, R_028AC4_DB_SRESULTS_COMPARE_STATE1
, 0x0);
371 radeon_set_context_reg(cs
, R_028AC8_DB_PRELOAD_CONTROL
, 0x0);
372 radeon_set_context_reg(cs
, R_02800C_DB_RENDER_OVERRIDE
,
373 S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE
) |
374 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE
));
376 radeon_set_context_reg(cs
, R_028400_VGT_MAX_VTX_INDX
, ~0);
377 radeon_set_context_reg(cs
, R_028404_VGT_MIN_VTX_INDX
, 0);
378 radeon_set_context_reg(cs
, R_028408_VGT_INDX_OFFSET
, 0);
380 if (physical_device
->rad_info
.chip_class
>= CIK
) {
381 /* If this is 0, Bonaire can hang even if GS isn't being used.
382 * Other chips are unaffected. These are suboptimal values,
383 * but we don't use on-chip GS.
385 radeon_set_context_reg(cs
, R_028A44_VGT_GS_ONCHIP_CNTL
,
386 S_028A44_ES_VERTS_PER_SUBGRP(64) |
387 S_028A44_GS_PRIMS_PER_SUBGRP(4));
389 radeon_set_sh_reg(cs
, R_00B51C_SPI_SHADER_PGM_RSRC3_LS
, S_00B51C_CU_EN(0xffff));
390 radeon_set_sh_reg(cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
, 0);
391 radeon_set_sh_reg(cs
, R_00B31C_SPI_SHADER_PGM_RSRC3_ES
, S_00B31C_CU_EN(0xffff));
392 radeon_set_sh_reg(cs
, R_00B21C_SPI_SHADER_PGM_RSRC3_GS
, S_00B21C_CU_EN(0xffff));
394 if (physical_device
->rad_info
.num_good_compute_units
/
395 (physical_device
->rad_info
.max_se
* physical_device
->rad_info
.max_sh_per_se
) <= 4) {
396 /* Too few available compute units per SH. Disallowing
397 * VS to run on CU0 could hurt us more than late VS
398 * allocation would help.
400 * LATE_ALLOC_VS = 2 is the highest safe number.
402 radeon_set_sh_reg(cs
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
, S_00B118_CU_EN(0xffff));
403 radeon_set_sh_reg(cs
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
, S_00B11C_LIMIT(2));
405 /* Set LATE_ALLOC_VS == 31. It should be less than
406 * the number of scratch waves. Limitations:
407 * - VS can't execute on CU0.
408 * - If HS writes outputs to LDS, LS can't execute on CU0.
410 radeon_set_sh_reg(cs
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
, S_00B118_CU_EN(0xfffe));
411 radeon_set_sh_reg(cs
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
, S_00B11C_LIMIT(31));
414 radeon_set_sh_reg(cs
, R_00B01C_SPI_SHADER_PGM_RSRC3_PS
, S_00B01C_CU_EN(0xffff));
417 if (physical_device
->rad_info
.chip_class
>= VI
) {
418 radeon_set_context_reg(cs
, R_028424_CB_DCC_CONTROL
,
419 S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
420 S_028424_OVERWRITE_COMBINER_WATERMARK(4));
421 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 30);
422 radeon_set_context_reg(cs
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 32);
423 radeon_set_context_reg(cs
, R_028B50_VGT_TESS_DISTRIBUTION
,
424 S_028B50_ACCUM_ISOLINE(32) |
425 S_028B50_ACCUM_TRI(11) |
426 S_028B50_ACCUM_QUAD(11) |
427 S_028B50_DONUT_SPLIT(16));
429 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 14);
430 radeon_set_context_reg(cs
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 16);
433 if (physical_device
->rad_info
.family
== CHIP_STONEY
)
434 radeon_set_context_reg(cs
, R_028C40_PA_SC_SHADER_CONTROL
, 0);
436 si_emit_compute(physical_device
, cs
);
439 void si_init_config(struct radv_cmd_buffer
*cmd_buffer
)
441 struct radv_physical_device
*physical_device
= cmd_buffer
->device
->physical_device
;
443 si_emit_config(physical_device
, cmd_buffer
->cs
);
447 cik_create_gfx_config(struct radv_device
*device
)
449 struct radeon_winsys_cs
*cs
= device
->ws
->cs_create(device
->ws
, RING_GFX
);
453 si_emit_config(device
->physical_device
, cs
);
455 while (cs
->cdw
& 7) {
456 if (device
->physical_device
->rad_info
.gfx_ib_pad_with_type2
)
457 radeon_emit(cs
, 0x80000000);
459 radeon_emit(cs
, 0xffff1000);
462 device
->gfx_init
= device
->ws
->buffer_create(device
->ws
,
465 RADEON_FLAG_CPU_ACCESS
);
466 if (!device
->gfx_init
)
469 void *map
= device
->ws
->buffer_map(device
->gfx_init
);
471 device
->ws
->buffer_destroy(device
->gfx_init
);
472 device
->gfx_init
= NULL
;
475 memcpy(map
, cs
->buf
, cs
->cdw
* 4);
477 device
->ws
->buffer_unmap(device
->gfx_init
);
478 device
->gfx_init_size_dw
= cs
->cdw
;
480 device
->ws
->cs_destroy(cs
);
484 get_viewport_xform(const VkViewport
*viewport
,
485 float scale
[3], float translate
[3])
487 float x
= viewport
->x
;
488 float y
= viewport
->y
;
489 float half_width
= 0.5f
* viewport
->width
;
490 float half_height
= 0.5f
* viewport
->height
;
491 double n
= viewport
->minDepth
;
492 double f
= viewport
->maxDepth
;
494 scale
[0] = half_width
;
495 translate
[0] = half_width
+ x
;
496 scale
[1] = half_height
;
497 translate
[1] = half_height
+ y
;
504 si_write_viewport(struct radeon_winsys_cs
*cs
, int first_vp
,
505 int count
, const VkViewport
*viewports
)
510 radeon_set_context_reg_seq(cs
, R_02843C_PA_CL_VPORT_XSCALE
, 6);
511 radeon_emit(cs
, fui(1.0));
512 radeon_emit(cs
, fui(0.0));
513 radeon_emit(cs
, fui(1.0));
514 radeon_emit(cs
, fui(0.0));
515 radeon_emit(cs
, fui(1.0));
516 radeon_emit(cs
, fui(0.0));
518 radeon_set_context_reg_seq(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
, 2);
519 radeon_emit(cs
, fui(0.0));
520 radeon_emit(cs
, fui(1.0));
524 radeon_set_context_reg_seq(cs
, R_02843C_PA_CL_VPORT_XSCALE
+
525 first_vp
* 4 * 6, count
* 6);
527 for (i
= 0; i
< count
; i
++) {
528 float scale
[3], translate
[3];
531 get_viewport_xform(&viewports
[i
], scale
, translate
);
532 radeon_emit(cs
, fui(scale
[0]));
533 radeon_emit(cs
, fui(translate
[0]));
534 radeon_emit(cs
, fui(scale
[1]));
535 radeon_emit(cs
, fui(translate
[1]));
536 radeon_emit(cs
, fui(scale
[2]));
537 radeon_emit(cs
, fui(translate
[2]));
540 radeon_set_context_reg_seq(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+
541 first_vp
* 4 * 2, count
* 2);
542 for (i
= 0; i
< count
; i
++) {
543 float zmin
= MIN2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
544 float zmax
= MAX2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
545 radeon_emit(cs
, fui(zmin
));
546 radeon_emit(cs
, fui(zmax
));
551 si_write_scissors(struct radeon_winsys_cs
*cs
, int first
,
552 int count
, const VkRect2D
*scissors
)
558 radeon_set_context_reg_seq(cs
, R_028250_PA_SC_VPORT_SCISSOR_0_TL
+ first
* 4 * 2, count
* 2);
559 for (i
= 0; i
< count
; i
++) {
560 radeon_emit(cs
, S_028250_TL_X(scissors
[i
].offset
.x
) |
561 S_028250_TL_Y(scissors
[i
].offset
.y
) |
562 S_028250_WINDOW_OFFSET_DISABLE(1));
563 radeon_emit(cs
, S_028254_BR_X(scissors
[i
].offset
.x
+ scissors
[i
].extent
.width
) |
564 S_028254_BR_Y(scissors
[i
].offset
.y
+ scissors
[i
].extent
.height
));
568 static inline unsigned
569 radv_prims_for_vertices(struct radv_prim_vertex_count
*info
, unsigned num
)
580 return 1 + ((num
- info
->min
) / info
->incr
);
584 si_get_ia_multi_vgt_param(struct radv_cmd_buffer
*cmd_buffer
,
585 bool instanced_or_indirect_draw
,
586 uint32_t draw_vertex_count
)
588 enum chip_class chip_class
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
;
589 enum radeon_family family
= cmd_buffer
->device
->physical_device
->rad_info
.family
;
590 struct radeon_info
*info
= &cmd_buffer
->device
->physical_device
->rad_info
;
591 unsigned prim
= cmd_buffer
->state
.pipeline
->graphics
.prim
;
592 unsigned primgroup_size
= 128; /* recommended without a GS */
593 unsigned max_primgroup_in_wave
= 2;
594 /* SWITCH_ON_EOP(0) is always preferable. */
595 bool wd_switch_on_eop
= false;
596 bool ia_switch_on_eop
= false;
597 bool ia_switch_on_eoi
= false;
598 bool partial_vs_wave
= false;
599 bool partial_es_wave
= false;
600 uint32_t num_prims
= radv_prims_for_vertices(&cmd_buffer
->state
.pipeline
->graphics
.prim_vertex_count
, draw_vertex_count
);
601 bool multi_instances_smaller_than_primgroup
;
603 if (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
))
604 primgroup_size
= 64; /* recommended with a GS */
606 multi_instances_smaller_than_primgroup
= (instanced_or_indirect_draw
||
607 num_prims
< primgroup_size
);
610 /* TODO linestipple */
612 if (chip_class
>= CIK
) {
613 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
614 * 4 shader engines. Set 1 to pass the assertion below.
615 * The other cases are hardware requirements. */
616 if (info
->max_se
< 4 ||
617 prim
== V_008958_DI_PT_POLYGON
||
618 prim
== V_008958_DI_PT_LINELOOP
||
619 prim
== V_008958_DI_PT_TRIFAN
||
620 prim
== V_008958_DI_PT_TRISTRIP_ADJ
||
621 (cmd_buffer
->state
.pipeline
->graphics
.prim_restart_enable
&&
622 (family
< CHIP_POLARIS10
||
623 (prim
!= V_008958_DI_PT_POINTLIST
&&
624 prim
!= V_008958_DI_PT_LINESTRIP
&&
625 prim
!= V_008958_DI_PT_TRISTRIP
))))
626 wd_switch_on_eop
= true;
628 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
629 * We don't know that for indirect drawing, so treat it as
630 * always problematic. */
631 if (family
== CHIP_HAWAII
&&
632 instanced_or_indirect_draw
)
633 wd_switch_on_eop
= true;
635 /* Performance recommendation for 4 SE Gfx7-8 parts if
636 * instances are smaller than a primgroup.
637 * Assume indirect draws always use small instances.
638 * This is needed for good VS wave utilization.
640 if (chip_class
<= VI
&&
642 multi_instances_smaller_than_primgroup
)
643 wd_switch_on_eop
= true;
645 /* Required on CIK and later. */
646 if (info
->max_se
> 2 && !wd_switch_on_eop
)
647 ia_switch_on_eoi
= true;
649 /* Required by Hawaii and, for some special cases, by VI. */
650 if (ia_switch_on_eoi
&&
651 (family
== CHIP_HAWAII
||
653 (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
) || max_primgroup_in_wave
!= 2))))
654 partial_vs_wave
= true;
656 /* Instancing bug on Bonaire. */
657 if (family
== CHIP_BONAIRE
&& ia_switch_on_eoi
&&
658 instanced_or_indirect_draw
)
659 partial_vs_wave
= true;
661 /* If the WD switch is false, the IA switch must be false too. */
662 assert(wd_switch_on_eop
|| !ia_switch_on_eop
);
664 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
665 if (ia_switch_on_eoi
)
666 partial_es_wave
= true;
668 if (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
)) {
669 /* GS requirement. */
670 if (SI_GS_PER_ES
/ primgroup_size
>= cmd_buffer
->device
->gs_table_depth
- 3)
671 partial_es_wave
= true;
673 /* Hw bug with single-primitive instances and SWITCH_ON_EOI
674 * on multi-SE chips. */
675 if (info
->max_se
>= 2 && ia_switch_on_eoi
&&
676 (instanced_or_indirect_draw
&&
678 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_VGT_FLUSH
;
681 return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop
) |
682 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi
) |
683 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave
) |
684 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave
) |
685 S_028AA8_PRIMGROUP_SIZE(primgroup_size
- 1) |
686 S_028AA8_WD_SWITCH_ON_EOP(chip_class
>= CIK
? wd_switch_on_eop
: 0) |
687 S_028AA8_MAX_PRIMGRP_IN_WAVE(chip_class
>= VI
?
688 max_primgroup_in_wave
: 0);
693 si_emit_cache_flush(struct radv_cmd_buffer
*cmd_buffer
)
695 enum chip_class chip_class
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
;
696 unsigned cp_coher_cntl
= 0;
697 bool is_compute
= cmd_buffer
->queue_family_index
== RADV_QUEUE_COMPUTE
;
700 cmd_buffer
->state
.flush_bits
&= ~(RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
701 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
|
702 RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
703 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
|
704 RADV_CMD_FLAG_PS_PARTIAL_FLUSH
|
705 RADV_CMD_FLAG_VS_PARTIAL_FLUSH
|
706 RADV_CMD_FLAG_VGT_FLUSH
);
708 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 128);
710 if (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_INV_ICACHE
)
711 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
712 if (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_INV_SMEM_L1
)
713 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
714 if (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_INV_VMEM_L1
)
715 cp_coher_cntl
|= S_0085F0_TCL1_ACTION_ENA(1);
716 if (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_INV_GLOBAL_L2
) {
717 cp_coher_cntl
|= S_0085F0_TC_ACTION_ENA(1);
718 if (chip_class
>= VI
)
719 cp_coher_cntl
|= S_0301F0_TC_WB_ACTION_ENA(1);
722 if (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB
) {
723 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
724 S_0085F0_CB0_DEST_BASE_ENA(1) |
725 S_0085F0_CB1_DEST_BASE_ENA(1) |
726 S_0085F0_CB2_DEST_BASE_ENA(1) |
727 S_0085F0_CB3_DEST_BASE_ENA(1) |
728 S_0085F0_CB4_DEST_BASE_ENA(1) |
729 S_0085F0_CB5_DEST_BASE_ENA(1) |
730 S_0085F0_CB6_DEST_BASE_ENA(1) |
731 S_0085F0_CB7_DEST_BASE_ENA(1);
733 /* Necessary for DCC */
734 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= VI
) {
735 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, 0));
736 radeon_emit(cmd_buffer
->cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_DATA_TS
) |
738 radeon_emit(cmd_buffer
->cs
, 0);
739 radeon_emit(cmd_buffer
->cs
, 0);
740 radeon_emit(cmd_buffer
->cs
, 0);
741 radeon_emit(cmd_buffer
->cs
, 0);
745 if (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB
) {
746 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
747 S_0085F0_DB_DEST_BASE_ENA(1);
750 if (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
) {
751 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
752 radeon_emit(cmd_buffer
->cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
755 if (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
) {
756 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
757 radeon_emit(cmd_buffer
->cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
760 if (!(cmd_buffer
->state
.flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
761 RADV_CMD_FLAG_FLUSH_AND_INV_DB
))) {
762 if (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_PS_PARTIAL_FLUSH
) {
763 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
764 radeon_emit(cmd_buffer
->cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
765 } else if (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_VS_PARTIAL_FLUSH
) {
766 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
767 radeon_emit(cmd_buffer
->cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
771 if (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_CS_PARTIAL_FLUSH
) {
772 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
773 radeon_emit(cmd_buffer
->cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
777 if (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_VGT_FLUSH
) {
778 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
779 radeon_emit(cmd_buffer
->cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
782 /* Make sure ME is idle (it executes most packets) before continuing.
783 * This prevents read-after-write hazards between PFP and ME.
785 if ((cp_coher_cntl
|| (cmd_buffer
->state
.flush_bits
& RADV_CMD_FLAG_CS_PARTIAL_FLUSH
)) &&
786 !radv_cmd_buffer_uses_mec(cmd_buffer
)) {
787 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
788 radeon_emit(cmd_buffer
->cs
, 0);
791 /* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
792 * Therefore, it should be last. Done in PFP.
795 if (radv_cmd_buffer_uses_mec(cmd_buffer
)) {
796 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_ACQUIRE_MEM
, 5, 0) |
797 PKT3_SHADER_TYPE_S(1));
798 radeon_emit(cmd_buffer
->cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
799 radeon_emit(cmd_buffer
->cs
, 0xffffffff); /* CP_COHER_SIZE */
800 radeon_emit(cmd_buffer
->cs
, 0xff); /* CP_COHER_SIZE_HI */
801 radeon_emit(cmd_buffer
->cs
, 0); /* CP_COHER_BASE */
802 radeon_emit(cmd_buffer
->cs
, 0); /* CP_COHER_BASE_HI */
803 radeon_emit(cmd_buffer
->cs
, 0x0000000A); /* POLL_INTERVAL */
805 /* ACQUIRE_MEM is only required on a compute ring. */
806 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_SURFACE_SYNC
, 3, 0));
807 radeon_emit(cmd_buffer
->cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
808 radeon_emit(cmd_buffer
->cs
, 0xffffffff); /* CP_COHER_SIZE */
809 radeon_emit(cmd_buffer
->cs
, 0); /* CP_COHER_BASE */
810 radeon_emit(cmd_buffer
->cs
, 0x0000000A); /* POLL_INTERVAL */
814 if (cmd_buffer
->state
.flush_bits
)
815 radv_cmd_buffer_trace_emit(cmd_buffer
);
816 cmd_buffer
->state
.flush_bits
= 0;
820 /* Set this if you want the 3D engine to wait until CP DMA is done.
821 * It should be set on the last CP DMA packet. */
822 #define R600_CP_DMA_SYNC (1 << 0) /* R600+ */
824 /* Set this if the source data was used as a destination in a previous CP DMA
825 * packet. It's for preventing a read-after-write (RAW) hazard between two
827 #define SI_CP_DMA_RAW_WAIT (1 << 1) /* SI+ */
828 #define CIK_CP_DMA_USE_L2 (1 << 2)
830 /* Alignment for optimal performance. */
831 #define CP_DMA_ALIGNMENT 32
832 /* The max number of bytes to copy per packet. */
833 #define CP_DMA_MAX_BYTE_COUNT ((1 << 21) - CP_DMA_ALIGNMENT)
835 static void si_emit_cp_dma_copy_buffer(struct radv_cmd_buffer
*cmd_buffer
,
836 uint64_t dst_va
, uint64_t src_va
,
837 unsigned size
, unsigned flags
)
839 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
840 uint32_t sync_flag
= flags
& R600_CP_DMA_SYNC
? S_411_CP_SYNC(1) : 0;
841 uint32_t wr_confirm
= !(flags
& R600_CP_DMA_SYNC
) ? S_414_DISABLE_WR_CONFIRM(1) : 0;
842 uint32_t raw_wait
= flags
& SI_CP_DMA_RAW_WAIT
? S_414_RAW_WAIT(1) : 0;
843 uint32_t sel
= flags
& CIK_CP_DMA_USE_L2
?
844 S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2
) |
845 S_411_DSL_SEL(V_411_DST_ADDR_TC_L2
) : 0;
848 assert((size
& ((1<<21)-1)) == size
);
850 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 9);
852 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
853 radeon_emit(cs
, PKT3(PKT3_DMA_DATA
, 5, 0));
854 radeon_emit(cs
, sync_flag
| sel
); /* CP_SYNC [31] */
855 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
856 radeon_emit(cs
, src_va
>> 32); /* SRC_ADDR_HI [31:0] */
857 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
858 radeon_emit(cs
, dst_va
>> 32); /* DST_ADDR_HI [31:0] */
859 radeon_emit(cs
, size
| wr_confirm
| raw_wait
); /* COMMAND [29:22] | BYTE_COUNT [20:0] */
861 radeon_emit(cs
, PKT3(PKT3_CP_DMA
, 4, 0));
862 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
863 radeon_emit(cs
, sync_flag
| ((src_va
>> 32) & 0xffff)); /* CP_SYNC [31] | SRC_ADDR_HI [15:0] */
864 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
865 radeon_emit(cs
, (dst_va
>> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
866 radeon_emit(cs
, size
| wr_confirm
| raw_wait
); /* COMMAND [29:22] | BYTE_COUNT [20:0] */
869 /* CP DMA is executed in ME, but index buffers are read by PFP.
870 * This ensures that ME (CP DMA) is idle before PFP starts fetching
871 * indices. If we wanted to execute CP DMA in PFP, this packet
874 if (sync_flag
&& cmd_buffer
->queue_family_index
== RADV_QUEUE_GENERAL
) {
875 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
879 radv_cmd_buffer_trace_emit(cmd_buffer
);
882 /* Emit a CP DMA packet to clear a buffer. The size must fit in bits [20:0]. */
883 static void si_emit_cp_dma_clear_buffer(struct radv_cmd_buffer
*cmd_buffer
,
884 uint64_t dst_va
, unsigned size
,
885 uint32_t clear_value
, unsigned flags
)
887 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
888 uint32_t sync_flag
= flags
& R600_CP_DMA_SYNC
? S_411_CP_SYNC(1) : 0;
889 uint32_t wr_confirm
= !(flags
& R600_CP_DMA_SYNC
) ? S_414_DISABLE_WR_CONFIRM(1) : 0;
890 uint32_t raw_wait
= flags
& SI_CP_DMA_RAW_WAIT
? S_414_RAW_WAIT(1) : 0;
891 uint32_t dst_sel
= flags
& CIK_CP_DMA_USE_L2
? S_411_DSL_SEL(V_411_DST_ADDR_TC_L2
) : 0;
894 assert((size
& ((1<<21)-1)) == size
);
896 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 9);
898 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
899 radeon_emit(cs
, PKT3(PKT3_DMA_DATA
, 5, 0));
900 radeon_emit(cs
, sync_flag
| dst_sel
| S_411_SRC_SEL(V_411_DATA
)); /* CP_SYNC [31] | SRC_SEL[30:29] */
901 radeon_emit(cs
, clear_value
); /* DATA [31:0] */
903 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
904 radeon_emit(cs
, dst_va
>> 32); /* DST_ADDR_HI [15:0] */
905 radeon_emit(cs
, size
| wr_confirm
| raw_wait
); /* COMMAND [29:22] | BYTE_COUNT [20:0] */
907 radeon_emit(cs
, PKT3(PKT3_CP_DMA
, 4, 0));
908 radeon_emit(cs
, clear_value
); /* DATA [31:0] */
909 radeon_emit(cs
, sync_flag
| S_411_SRC_SEL(V_411_DATA
)); /* CP_SYNC [31] | SRC_SEL[30:29] */
910 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
911 radeon_emit(cs
, (dst_va
>> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
912 radeon_emit(cs
, size
| wr_confirm
| raw_wait
); /* COMMAND [29:22] | BYTE_COUNT [20:0] */
915 /* See "copy_buffer" for explanation. */
916 if (sync_flag
&& cmd_buffer
->queue_family_index
== RADV_QUEUE_GENERAL
) {
917 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
920 radv_cmd_buffer_trace_emit(cmd_buffer
);
923 static void si_cp_dma_prepare(struct radv_cmd_buffer
*cmd_buffer
, uint64_t byte_count
,
924 uint64_t remaining_size
, unsigned *flags
)
926 cmd_buffer
->no_draws
= false;
927 /* Flush the caches for the first copy only.
928 * Also wait for the previous CP DMA operations.
930 if (cmd_buffer
->state
.flush_bits
) {
931 si_emit_cache_flush(cmd_buffer
);
932 *flags
|= SI_CP_DMA_RAW_WAIT
;
935 /* Do the synchronization after the last dma, so that all data
936 * is written to memory.
938 if (byte_count
== remaining_size
)
939 *flags
|= R600_CP_DMA_SYNC
;
942 static void si_cp_dma_realign_engine(struct radv_cmd_buffer
*cmd_buffer
, unsigned size
)
946 unsigned dma_flags
= 0;
947 unsigned buf_size
= CP_DMA_ALIGNMENT
* 2;
950 assert(size
< CP_DMA_ALIGNMENT
);
952 radv_cmd_buffer_upload_alloc(cmd_buffer
, buf_size
, CP_DMA_ALIGNMENT
, &offset
, &ptr
);
954 va
= cmd_buffer
->device
->ws
->buffer_get_va(cmd_buffer
->upload
.upload_bo
);
957 si_cp_dma_prepare(cmd_buffer
, size
, size
, &dma_flags
);
959 si_emit_cp_dma_copy_buffer(cmd_buffer
, va
, va
+ CP_DMA_ALIGNMENT
, size
,
963 void si_cp_dma_buffer_copy(struct radv_cmd_buffer
*cmd_buffer
,
964 uint64_t src_va
, uint64_t dest_va
,
967 uint64_t main_src_va
, main_dest_va
;
968 uint64_t skipped_size
= 0, realign_size
= 0;
971 if (cmd_buffer
->device
->physical_device
->rad_info
.family
<= CHIP_CARRIZO
||
972 cmd_buffer
->device
->physical_device
->rad_info
.family
== CHIP_STONEY
) {
973 /* If the size is not aligned, we must add a dummy copy at the end
974 * just to align the internal counter. Otherwise, the DMA engine
975 * would slow down by an order of magnitude for following copies.
977 if (size
% CP_DMA_ALIGNMENT
)
978 realign_size
= CP_DMA_ALIGNMENT
- (size
% CP_DMA_ALIGNMENT
);
980 /* If the copy begins unaligned, we must start copying from the next
981 * aligned block and the skipped part should be copied after everything
982 * else has been copied. Only the src alignment matters, not dst.
984 if (src_va
% CP_DMA_ALIGNMENT
) {
985 skipped_size
= CP_DMA_ALIGNMENT
- (src_va
% CP_DMA_ALIGNMENT
);
986 /* The main part will be skipped if the size is too small. */
987 skipped_size
= MIN2(skipped_size
, size
);
988 size
-= skipped_size
;
991 main_src_va
= src_va
+ skipped_size
;
992 main_dest_va
= dest_va
+ skipped_size
;
995 unsigned dma_flags
= 0;
996 unsigned byte_count
= MIN2(size
, CP_DMA_MAX_BYTE_COUNT
);
998 si_cp_dma_prepare(cmd_buffer
, byte_count
,
999 size
+ skipped_size
+ realign_size
,
1002 si_emit_cp_dma_copy_buffer(cmd_buffer
, main_dest_va
, main_src_va
,
1003 byte_count
, dma_flags
);
1006 main_src_va
+= byte_count
;
1007 main_dest_va
+= byte_count
;
1011 unsigned dma_flags
= 0;
1013 si_cp_dma_prepare(cmd_buffer
, skipped_size
,
1014 size
+ skipped_size
+ realign_size
,
1017 si_emit_cp_dma_copy_buffer(cmd_buffer
, dest_va
, src_va
,
1018 skipped_size
, dma_flags
);
1021 si_cp_dma_realign_engine(cmd_buffer
, realign_size
);
1024 void si_cp_dma_clear_buffer(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1025 uint64_t size
, unsigned value
)
1031 assert(va
% 4 == 0 && size
% 4 == 0);
1034 unsigned byte_count
= MIN2(size
, CP_DMA_MAX_BYTE_COUNT
);
1035 unsigned dma_flags
= 0;
1037 si_cp_dma_prepare(cmd_buffer
, byte_count
, size
, &dma_flags
);
1039 /* Emit the clear packet. */
1040 si_emit_cp_dma_clear_buffer(cmd_buffer
, va
, byte_count
, value
,
1048 /* For MSAA sample positions. */
1049 #define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y) \
1050 (((s0x) & 0xf) | (((unsigned)(s0y) & 0xf) << 4) | \
1051 (((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) | \
1052 (((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) | \
1053 (((unsigned)(s3x) & 0xf) << 24) | (((unsigned)(s3y) & 0xf) << 28))
1057 * There are two locations (4, 4), (-4, -4). */
1058 const uint32_t eg_sample_locs_2x
[4] = {
1059 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1060 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1061 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1062 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1064 const unsigned eg_max_dist_2x
= 4;
1066 * There are 4 locations: (-2, 6), (6, -2), (-6, 2), (2, 6). */
1067 const uint32_t eg_sample_locs_4x
[4] = {
1068 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1069 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1070 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1071 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1073 const unsigned eg_max_dist_4x
= 6;
1076 static const uint32_t cm_sample_locs_8x
[] = {
1077 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1078 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1079 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1080 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1081 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1082 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1083 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1084 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1086 static const unsigned cm_max_dist_8x
= 8;
1087 /* Cayman 16xMSAA */
1088 static const uint32_t cm_sample_locs_16x
[] = {
1089 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1090 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1091 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1092 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1093 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1094 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1095 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1096 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1097 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1098 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1099 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1100 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1101 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1102 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1103 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1104 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1106 static const unsigned cm_max_dist_16x
= 8;
1108 unsigned radv_cayman_get_maxdist(int log_samples
)
1110 unsigned max_dist
[] = {
1117 return max_dist
[log_samples
];
1120 void radv_cayman_emit_msaa_sample_locs(struct radeon_winsys_cs
*cs
, int nr_samples
)
1122 switch (nr_samples
) {
1125 radeon_set_context_reg(cs
, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 0);
1126 radeon_set_context_reg(cs
, CM_R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, 0);
1127 radeon_set_context_reg(cs
, CM_R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, 0);
1128 radeon_set_context_reg(cs
, CM_R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, 0);
1131 radeon_set_context_reg(cs
, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, eg_sample_locs_2x
[0]);
1132 radeon_set_context_reg(cs
, CM_R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, eg_sample_locs_2x
[1]);
1133 radeon_set_context_reg(cs
, CM_R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, eg_sample_locs_2x
[2]);
1134 radeon_set_context_reg(cs
, CM_R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, eg_sample_locs_2x
[3]);
1137 radeon_set_context_reg(cs
, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, eg_sample_locs_4x
[0]);
1138 radeon_set_context_reg(cs
, CM_R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, eg_sample_locs_4x
[1]);
1139 radeon_set_context_reg(cs
, CM_R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, eg_sample_locs_4x
[2]);
1140 radeon_set_context_reg(cs
, CM_R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, eg_sample_locs_4x
[3]);
1143 radeon_set_context_reg_seq(cs
, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 14);
1144 radeon_emit(cs
, cm_sample_locs_8x
[0]);
1145 radeon_emit(cs
, cm_sample_locs_8x
[4]);
1148 radeon_emit(cs
, cm_sample_locs_8x
[1]);
1149 radeon_emit(cs
, cm_sample_locs_8x
[5]);
1152 radeon_emit(cs
, cm_sample_locs_8x
[2]);
1153 radeon_emit(cs
, cm_sample_locs_8x
[6]);
1156 radeon_emit(cs
, cm_sample_locs_8x
[3]);
1157 radeon_emit(cs
, cm_sample_locs_8x
[7]);
1160 radeon_set_context_reg_seq(cs
, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 16);
1161 radeon_emit(cs
, cm_sample_locs_16x
[0]);
1162 radeon_emit(cs
, cm_sample_locs_16x
[4]);
1163 radeon_emit(cs
, cm_sample_locs_16x
[8]);
1164 radeon_emit(cs
, cm_sample_locs_16x
[12]);
1165 radeon_emit(cs
, cm_sample_locs_16x
[1]);
1166 radeon_emit(cs
, cm_sample_locs_16x
[5]);
1167 radeon_emit(cs
, cm_sample_locs_16x
[9]);
1168 radeon_emit(cs
, cm_sample_locs_16x
[13]);
1169 radeon_emit(cs
, cm_sample_locs_16x
[2]);
1170 radeon_emit(cs
, cm_sample_locs_16x
[6]);
1171 radeon_emit(cs
, cm_sample_locs_16x
[10]);
1172 radeon_emit(cs
, cm_sample_locs_16x
[14]);
1173 radeon_emit(cs
, cm_sample_locs_16x
[3]);
1174 radeon_emit(cs
, cm_sample_locs_16x
[7]);
1175 radeon_emit(cs
, cm_sample_locs_16x
[11]);
1176 radeon_emit(cs
, cm_sample_locs_16x
[15]);
1181 static void radv_cayman_get_sample_position(struct radv_device
*device
,
1182 unsigned sample_count
,
1183 unsigned sample_index
, float *out_value
)
1189 switch (sample_count
) {
1192 out_value
[0] = out_value
[1] = 0.5;
1195 offset
= 4 * (sample_index
* 2);
1196 val
.idx
= (eg_sample_locs_2x
[0] >> offset
) & 0xf;
1197 out_value
[0] = (float)(val
.idx
+ 8) / 16.0f
;
1198 val
.idx
= (eg_sample_locs_2x
[0] >> (offset
+ 4)) & 0xf;
1199 out_value
[1] = (float)(val
.idx
+ 8) / 16.0f
;
1202 offset
= 4 * (sample_index
* 2);
1203 val
.idx
= (eg_sample_locs_4x
[0] >> offset
) & 0xf;
1204 out_value
[0] = (float)(val
.idx
+ 8) / 16.0f
;
1205 val
.idx
= (eg_sample_locs_4x
[0] >> (offset
+ 4)) & 0xf;
1206 out_value
[1] = (float)(val
.idx
+ 8) / 16.0f
;
1209 offset
= 4 * (sample_index
% 4 * 2);
1210 index
= (sample_index
/ 4) * 4;
1211 val
.idx
= (cm_sample_locs_8x
[index
] >> offset
) & 0xf;
1212 out_value
[0] = (float)(val
.idx
+ 8) / 16.0f
;
1213 val
.idx
= (cm_sample_locs_8x
[index
] >> (offset
+ 4)) & 0xf;
1214 out_value
[1] = (float)(val
.idx
+ 8) / 16.0f
;
1217 offset
= 4 * (sample_index
% 4 * 2);
1218 index
= (sample_index
/ 4) * 4;
1219 val
.idx
= (cm_sample_locs_16x
[index
] >> offset
) & 0xf;
1220 out_value
[0] = (float)(val
.idx
+ 8) / 16.0f
;
1221 val
.idx
= (cm_sample_locs_16x
[index
] >> (offset
+ 4)) & 0xf;
1222 out_value
[1] = (float)(val
.idx
+ 8) / 16.0f
;
1227 void radv_device_init_msaa(struct radv_device
*device
)
1230 radv_cayman_get_sample_position(device
, 1, 0, device
->sample_locations_1x
[0]);
1232 for (i
= 0; i
< 2; i
++)
1233 radv_cayman_get_sample_position(device
, 2, i
, device
->sample_locations_2x
[i
]);
1234 for (i
= 0; i
< 4; i
++)
1235 radv_cayman_get_sample_position(device
, 4, i
, device
->sample_locations_4x
[i
]);
1236 for (i
= 0; i
< 8; i
++)
1237 radv_cayman_get_sample_position(device
, 8, i
, device
->sample_locations_8x
[i
]);
1238 for (i
= 0; i
< 16; i
++)
1239 radv_cayman_get_sample_position(device
, 16, i
, device
->sample_locations_16x
[i
]);