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 AMD GCN */
30 #include "radv_private.h"
31 #include "radv_shader.h"
34 #include "radv_util.h"
35 #include "main/macros.h"
38 si_write_harvested_raster_configs(struct radv_physical_device
*physical_device
,
39 struct radeon_cmdbuf
*cs
,
40 unsigned raster_config
,
41 unsigned raster_config_1
)
43 unsigned num_se
= MAX2(physical_device
->rad_info
.max_se
, 1);
44 unsigned raster_config_se
[4];
47 ac_get_harvested_configs(&physical_device
->rad_info
,
52 for (se
= 0; se
< num_se
; se
++) {
53 /* GRBM_GFX_INDEX has a different offset on GFX6 and GFX7+ */
54 if (physical_device
->rad_info
.chip_class
< GFX7
)
55 radeon_set_config_reg(cs
, R_00802C_GRBM_GFX_INDEX
,
56 S_00802C_SE_INDEX(se
) |
57 S_00802C_SH_BROADCAST_WRITES(1) |
58 S_00802C_INSTANCE_BROADCAST_WRITES(1));
60 radeon_set_uconfig_reg(cs
, R_030800_GRBM_GFX_INDEX
,
61 S_030800_SE_INDEX(se
) | S_030800_SH_BROADCAST_WRITES(1) |
62 S_030800_INSTANCE_BROADCAST_WRITES(1));
63 radeon_set_context_reg(cs
, R_028350_PA_SC_RASTER_CONFIG
, raster_config_se
[se
]);
66 /* GRBM_GFX_INDEX has a different offset on GFX6 and GFX7+ */
67 if (physical_device
->rad_info
.chip_class
< GFX7
)
68 radeon_set_config_reg(cs
, R_00802C_GRBM_GFX_INDEX
,
69 S_00802C_SE_BROADCAST_WRITES(1) |
70 S_00802C_SH_BROADCAST_WRITES(1) |
71 S_00802C_INSTANCE_BROADCAST_WRITES(1));
73 radeon_set_uconfig_reg(cs
, R_030800_GRBM_GFX_INDEX
,
74 S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
75 S_030800_INSTANCE_BROADCAST_WRITES(1));
77 if (physical_device
->rad_info
.chip_class
>= GFX7
)
78 radeon_set_context_reg(cs
, R_028354_PA_SC_RASTER_CONFIG_1
, raster_config_1
);
82 si_emit_compute(struct radv_physical_device
*physical_device
,
83 struct radeon_cmdbuf
*cs
)
85 radeon_set_sh_reg_seq(cs
, R_00B810_COMPUTE_START_X
, 3);
90 radeon_set_sh_reg_seq(cs
, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0
, 2);
91 /* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1,
92 * renamed COMPUTE_DESTINATION_EN_SEn on gfx10. */
93 radeon_emit(cs
, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
94 radeon_emit(cs
, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
96 if (physical_device
->rad_info
.chip_class
>= GFX7
) {
97 /* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
98 radeon_set_sh_reg_seq(cs
,
99 R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2
, 2);
100 radeon_emit(cs
, S_00B858_SH0_CU_EN(0xffff) |
101 S_00B858_SH1_CU_EN(0xffff));
102 radeon_emit(cs
, S_00B858_SH0_CU_EN(0xffff) |
103 S_00B858_SH1_CU_EN(0xffff));
106 if (physical_device
->rad_info
.chip_class
>= GFX10
)
107 radeon_set_sh_reg(cs
, R_00B8A0_COMPUTE_PGM_RSRC3
, 0);
109 /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
110 * and is now per pipe, so it should be handled in the
111 * kernel if we want to use something other than the default value,
112 * which is now 0x22f.
114 if (physical_device
->rad_info
.chip_class
<= GFX6
) {
115 /* XXX: This should be:
116 * (number of compute units) * 4 * (waves per simd) - 1 */
118 radeon_set_sh_reg(cs
, R_00B82C_COMPUTE_MAX_WAVE_ID
,
119 0x190 /* Default value */);
123 /* 12.4 fixed-point */
124 static unsigned radv_pack_float_12p4(float x
)
127 x
>= 4096 ? 0xffff : x
* 16;
131 si_set_raster_config(struct radv_physical_device
*physical_device
,
132 struct radeon_cmdbuf
*cs
)
134 unsigned num_rb
= MIN2(physical_device
->rad_info
.num_render_backends
, 16);
135 unsigned rb_mask
= physical_device
->rad_info
.enabled_rb_mask
;
136 unsigned raster_config
, raster_config_1
;
138 ac_get_raster_config(&physical_device
->rad_info
,
140 &raster_config_1
, NULL
);
142 /* Always use the default config when all backends are enabled
143 * (or when we failed to determine the enabled backends).
145 if (!rb_mask
|| util_bitcount(rb_mask
) >= num_rb
) {
146 radeon_set_context_reg(cs
, R_028350_PA_SC_RASTER_CONFIG
,
148 if (physical_device
->rad_info
.chip_class
>= GFX7
)
149 radeon_set_context_reg(cs
, R_028354_PA_SC_RASTER_CONFIG_1
,
152 si_write_harvested_raster_configs(physical_device
, cs
,
159 si_emit_graphics(struct radv_physical_device
*physical_device
,
160 struct radeon_cmdbuf
*cs
)
164 /* Only GFX6 can disable CLEAR_STATE for now. */
165 assert(physical_device
->has_clear_state
||
166 physical_device
->rad_info
.chip_class
== GFX6
);
168 radeon_emit(cs
, PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
169 radeon_emit(cs
, CONTEXT_CONTROL_LOAD_ENABLE(1));
170 radeon_emit(cs
, CONTEXT_CONTROL_SHADOW_ENABLE(1));
172 if (physical_device
->has_clear_state
) {
173 radeon_emit(cs
, PKT3(PKT3_CLEAR_STATE
, 0, 0));
177 if (physical_device
->rad_info
.chip_class
<= GFX8
)
178 si_set_raster_config(physical_device
, cs
);
180 radeon_set_context_reg(cs
, R_028A18_VGT_HOS_MAX_TESS_LEVEL
, fui(64));
181 if (!physical_device
->has_clear_state
)
182 radeon_set_context_reg(cs
, R_028A1C_VGT_HOS_MIN_TESS_LEVEL
, fui(0));
184 /* FIXME calculate these values somehow ??? */
185 if (physical_device
->rad_info
.chip_class
<= GFX8
) {
186 radeon_set_context_reg(cs
, R_028A54_VGT_GS_PER_ES
, SI_GS_PER_ES
);
187 radeon_set_context_reg(cs
, R_028A58_VGT_ES_PER_GS
, 0x40);
190 if (!physical_device
->has_clear_state
) {
191 radeon_set_context_reg(cs
, R_028A5C_VGT_GS_PER_VS
, 0x2);
192 radeon_set_context_reg(cs
, R_028A8C_VGT_PRIMITIVEID_RESET
, 0x0);
193 radeon_set_context_reg(cs
, R_028B98_VGT_STRMOUT_BUFFER_CONFIG
, 0x0);
196 radeon_set_context_reg(cs
, R_028AA0_VGT_INSTANCE_STEP_RATE_0
, 1);
197 if (!physical_device
->has_clear_state
)
198 radeon_set_context_reg(cs
, R_028AB8_VGT_VTX_CNT_EN
, 0x0);
199 if (physical_device
->rad_info
.chip_class
< GFX7
)
200 radeon_set_config_reg(cs
, R_008A14_PA_CL_ENHANCE
, S_008A14_NUM_CLIP_SEQ(3) |
201 S_008A14_CLIP_VTX_REORDER_ENA(1));
203 if (!physical_device
->has_clear_state
)
204 radeon_set_context_reg(cs
, R_02882C_PA_SU_PRIM_FILTER_CNTL
, 0);
206 /* CLEAR_STATE doesn't clear these correctly on certain generations.
207 * I don't know why. Deduced by trial and error.
209 if (physical_device
->rad_info
.chip_class
<= GFX7
) {
210 radeon_set_context_reg(cs
, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET
, 0);
211 radeon_set_context_reg(cs
, R_028204_PA_SC_WINDOW_SCISSOR_TL
,
212 S_028204_WINDOW_OFFSET_DISABLE(1));
213 radeon_set_context_reg(cs
, R_028240_PA_SC_GENERIC_SCISSOR_TL
,
214 S_028240_WINDOW_OFFSET_DISABLE(1));
215 radeon_set_context_reg(cs
, R_028244_PA_SC_GENERIC_SCISSOR_BR
,
216 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
217 radeon_set_context_reg(cs
, R_028030_PA_SC_SCREEN_SCISSOR_TL
, 0);
218 radeon_set_context_reg(cs
, R_028034_PA_SC_SCREEN_SCISSOR_BR
,
219 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
222 if (!physical_device
->has_clear_state
) {
223 for (i
= 0; i
< 16; i
++) {
224 radeon_set_context_reg(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+ i
*8, 0);
225 radeon_set_context_reg(cs
, R_0282D4_PA_SC_VPORT_ZMAX_0
+ i
*8, fui(1.0));
229 if (!physical_device
->has_clear_state
) {
230 radeon_set_context_reg(cs
, R_02820C_PA_SC_CLIPRECT_RULE
, 0xFFFF);
231 radeon_set_context_reg(cs
, R_028230_PA_SC_EDGERULE
, 0xAAAAAAAA);
232 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on GFX6 */
233 radeon_set_context_reg(cs
, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET
, 0);
234 radeon_set_context_reg(cs
, R_028820_PA_CL_NANINF_CNTL
, 0);
235 radeon_set_context_reg(cs
, R_028AC0_DB_SRESULTS_COMPARE_STATE0
, 0x0);
236 radeon_set_context_reg(cs
, R_028AC4_DB_SRESULTS_COMPARE_STATE1
, 0x0);
237 radeon_set_context_reg(cs
, R_028AC8_DB_PRELOAD_CONTROL
, 0x0);
240 radeon_set_context_reg(cs
, R_02800C_DB_RENDER_OVERRIDE
,
241 S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE
) |
242 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE
));
244 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
245 radeon_set_uconfig_reg(cs
, R_030964_GE_MAX_VTX_INDX
, ~0);
246 radeon_set_uconfig_reg(cs
, R_030924_GE_MIN_VTX_INDX
, 0);
247 radeon_set_uconfig_reg(cs
, R_030928_GE_INDX_OFFSET
, 0);
248 } else if (physical_device
->rad_info
.chip_class
>= GFX9
) {
249 radeon_set_uconfig_reg(cs
, R_030920_VGT_MAX_VTX_INDX
, ~0);
250 radeon_set_uconfig_reg(cs
, R_030924_VGT_MIN_VTX_INDX
, 0);
251 radeon_set_uconfig_reg(cs
, R_030928_VGT_INDX_OFFSET
, 0);
253 /* These registers, when written, also overwrite the
254 * CLEAR_STATE context, so we can't rely on CLEAR_STATE setting
255 * them. It would be an issue if there was another UMD
258 radeon_set_context_reg(cs
, R_028400_VGT_MAX_VTX_INDX
, ~0);
259 radeon_set_context_reg(cs
, R_028404_VGT_MIN_VTX_INDX
, 0);
260 radeon_set_context_reg(cs
, R_028408_VGT_INDX_OFFSET
, 0);
263 if (physical_device
->rad_info
.chip_class
>= GFX7
) {
264 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
265 /* Logical CUs 16 - 31 */
266 radeon_set_sh_reg(cs
, R_00B404_SPI_SHADER_PGM_RSRC4_HS
,
267 S_00B404_CU_EN(0xffff));
268 radeon_set_sh_reg(cs
, R_00B204_SPI_SHADER_PGM_RSRC4_GS
,
269 S_00B204_CU_EN(0xffff) |
270 S_00B204_SPI_SHADER_LATE_ALLOC_GS_GFX10(0));
271 radeon_set_sh_reg(cs
, R_00B104_SPI_SHADER_PGM_RSRC4_VS
,
272 S_00B104_CU_EN(0xffff));
273 radeon_set_sh_reg(cs
, R_00B004_SPI_SHADER_PGM_RSRC4_PS
,
274 S_00B004_CU_EN(0xffff));
277 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
278 radeon_set_sh_reg(cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
,
279 S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
281 radeon_set_sh_reg(cs
, R_00B51C_SPI_SHADER_PGM_RSRC3_LS
,
282 S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
283 radeon_set_sh_reg(cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
,
284 S_00B41C_WAVE_LIMIT(0x3F));
285 radeon_set_sh_reg(cs
, R_00B31C_SPI_SHADER_PGM_RSRC3_ES
,
286 S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
287 /* If this is 0, Bonaire can hang even if GS isn't being used.
288 * Other chips are unaffected. These are suboptimal values,
289 * but we don't use on-chip GS.
291 radeon_set_context_reg(cs
, R_028A44_VGT_GS_ONCHIP_CNTL
,
292 S_028A44_ES_VERTS_PER_SUBGRP(64) |
293 S_028A44_GS_PRIMS_PER_SUBGRP(4));
295 radeon_set_sh_reg(cs
, R_00B21C_SPI_SHADER_PGM_RSRC3_GS
,
296 S_00B21C_CU_EN(0xffff) | S_00B21C_WAVE_LIMIT(0x3F));
298 if (physical_device
->rad_info
.num_good_cu_per_sh
<= 4) {
299 /* Too few available compute units per SH. Disallowing
300 * VS to run on CU0 could hurt us more than late VS
301 * allocation would help.
303 * LATE_ALLOC_VS = 2 is the highest safe number.
305 radeon_set_sh_reg(cs
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
,
306 S_00B118_CU_EN(0xffff) | S_00B118_WAVE_LIMIT(0x3F) );
307 radeon_set_sh_reg(cs
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
, S_00B11C_LIMIT(2));
309 /* Set LATE_ALLOC_VS == 31. It should be less than
310 * the number of scratch waves. Limitations:
311 * - VS can't execute on CU0.
312 * - If HS writes outputs to LDS, LS can't execute on CU0.
314 radeon_set_sh_reg(cs
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
,
315 S_00B118_CU_EN(0xfffe) | S_00B118_WAVE_LIMIT(0x3F));
316 radeon_set_sh_reg(cs
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
, S_00B11C_LIMIT(31));
319 radeon_set_sh_reg(cs
, R_00B01C_SPI_SHADER_PGM_RSRC3_PS
,
320 S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
323 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
324 radeon_set_context_reg(cs
, R_02835C_PA_SC_TILE_STEERING_OVERRIDE
,
325 physical_device
->rad_info
.pa_sc_tile_steering_override
);
326 radeon_set_context_reg(cs
, R_02807C_DB_RMI_L2_CACHE_CONTROL
,
327 S_02807C_Z_WR_POLICY(V_02807C_CACHE_STREAM_WR
) |
328 S_02807C_S_WR_POLICY(V_02807C_CACHE_STREAM_WR
) |
329 S_02807C_HTILE_WR_POLICY(V_02807C_CACHE_STREAM_WR
) |
330 S_02807C_ZPCPSD_WR_POLICY(V_02807C_CACHE_STREAM_WR
) |
331 S_02807C_Z_RD_POLICY(V_02807C_CACHE_NOA_RD
) |
332 S_02807C_S_RD_POLICY(V_02807C_CACHE_NOA_RD
) |
333 S_02807C_HTILE_RD_POLICY(V_02807C_CACHE_NOA_RD
));
335 radeon_set_context_reg(cs
, R_028410_CB_RMI_GL2_CACHE_CONTROL
,
336 S_028410_CMASK_WR_POLICY(V_028410_CACHE_STREAM_WR
) |
337 S_028410_FMASK_WR_POLICY(V_028410_CACHE_STREAM_WR
) |
338 S_028410_DCC_WR_POLICY(V_028410_CACHE_STREAM_WR
) |
339 S_028410_COLOR_WR_POLICY(V_028410_CACHE_STREAM_WR
) |
340 S_028410_CMASK_RD_POLICY(V_028410_CACHE_NOA_RD
) |
341 S_028410_FMASK_RD_POLICY(V_028410_CACHE_NOA_RD
) |
342 S_028410_DCC_RD_POLICY(V_028410_CACHE_NOA_RD
) |
343 S_028410_COLOR_RD_POLICY(V_028410_CACHE_NOA_RD
));
346 if (physical_device
->rad_info
.chip_class
>= GFX8
) {
347 uint32_t vgt_tess_distribution
;
349 vgt_tess_distribution
= S_028B50_ACCUM_ISOLINE(32) |
350 S_028B50_ACCUM_TRI(11) |
351 S_028B50_ACCUM_QUAD(11) |
352 S_028B50_DONUT_SPLIT(16);
354 if (physical_device
->rad_info
.family
== CHIP_FIJI
||
355 physical_device
->rad_info
.family
>= CHIP_POLARIS10
)
356 vgt_tess_distribution
|= S_028B50_TRAP_SPLIT(3);
358 radeon_set_context_reg(cs
, R_028B50_VGT_TESS_DISTRIBUTION
,
359 vgt_tess_distribution
);
360 } else if (!physical_device
->has_clear_state
) {
361 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 14);
362 radeon_set_context_reg(cs
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 16);
365 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
366 unsigned num_se
= physical_device
->rad_info
.max_se
;
367 unsigned pc_lines
= 0;
368 unsigned max_alloc_count
= 0;
370 switch (physical_device
->rad_info
.family
) {
389 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
390 max_alloc_count
= pc_lines
/ 3;
392 max_alloc_count
= MIN2(128, pc_lines
/ (4 * num_se
));
395 radeon_set_context_reg(cs
, R_028C48_PA_SC_BINNER_CNTL_1
,
396 S_028C48_MAX_ALLOC_COUNT(max_alloc_count
) |
397 S_028C48_MAX_PRIM_PER_BATCH(1023));
398 radeon_set_context_reg(cs
, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL
,
399 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
400 radeon_set_uconfig_reg(cs
, R_030968_VGT_INSTANCE_BASE_ID
, 0);
403 unsigned tmp
= (unsigned)(1.0 * 8.0);
404 radeon_set_context_reg_seq(cs
, R_028A00_PA_SU_POINT_SIZE
, 1);
405 radeon_emit(cs
, S_028A00_HEIGHT(tmp
) | S_028A00_WIDTH(tmp
));
406 radeon_set_context_reg_seq(cs
, R_028A04_PA_SU_POINT_MINMAX
, 1);
407 radeon_emit(cs
, S_028A04_MIN_SIZE(radv_pack_float_12p4(0)) |
408 S_028A04_MAX_SIZE(radv_pack_float_12p4(8192/2)));
410 if (!physical_device
->has_clear_state
) {
411 radeon_set_context_reg(cs
, R_028004_DB_COUNT_CONTROL
,
412 S_028004_ZPASS_INCREMENT_DISABLE(1));
415 /* Enable the Polaris small primitive filter control.
416 * XXX: There is possibly an issue when MSAA is off (see RadeonSI
417 * has_msaa_sample_loc_bug). But this doesn't seem to regress anything,
418 * and AMDVLK doesn't have a workaround as well.
420 if (physical_device
->rad_info
.family
>= CHIP_POLARIS10
) {
421 unsigned small_prim_filter_cntl
=
422 S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
423 /* Workaround for a hw line bug. */
424 S_028830_LINE_FILTER_DISABLE(physical_device
->rad_info
.family
<= CHIP_POLARIS12
);
426 radeon_set_context_reg(cs
, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL
,
427 small_prim_filter_cntl
);
430 si_emit_compute(physical_device
, cs
);
434 cik_create_gfx_config(struct radv_device
*device
)
436 struct radeon_cmdbuf
*cs
= device
->ws
->cs_create(device
->ws
, RING_GFX
);
440 si_emit_graphics(device
->physical_device
, cs
);
442 while (cs
->cdw
& 7) {
443 if (device
->physical_device
->rad_info
.gfx_ib_pad_with_type2
)
444 radeon_emit(cs
, 0x80000000);
446 radeon_emit(cs
, 0xffff1000);
449 device
->gfx_init
= device
->ws
->buffer_create(device
->ws
,
452 RADEON_FLAG_CPU_ACCESS
|
453 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
454 RADEON_FLAG_READ_ONLY
,
455 RADV_BO_PRIORITY_CS
);
456 if (!device
->gfx_init
)
459 void *map
= device
->ws
->buffer_map(device
->gfx_init
);
461 device
->ws
->buffer_destroy(device
->gfx_init
);
462 device
->gfx_init
= NULL
;
465 memcpy(map
, cs
->buf
, cs
->cdw
* 4);
467 device
->ws
->buffer_unmap(device
->gfx_init
);
468 device
->gfx_init_size_dw
= cs
->cdw
;
470 device
->ws
->cs_destroy(cs
);
474 get_viewport_xform(const VkViewport
*viewport
,
475 float scale
[3], float translate
[3])
477 float x
= viewport
->x
;
478 float y
= viewport
->y
;
479 float half_width
= 0.5f
* viewport
->width
;
480 float half_height
= 0.5f
* viewport
->height
;
481 double n
= viewport
->minDepth
;
482 double f
= viewport
->maxDepth
;
484 scale
[0] = half_width
;
485 translate
[0] = half_width
+ x
;
486 scale
[1] = half_height
;
487 translate
[1] = half_height
+ y
;
494 si_write_viewport(struct radeon_cmdbuf
*cs
, int first_vp
,
495 int count
, const VkViewport
*viewports
)
500 radeon_set_context_reg_seq(cs
, R_02843C_PA_CL_VPORT_XSCALE
+
501 first_vp
* 4 * 6, count
* 6);
503 for (i
= 0; i
< count
; i
++) {
504 float scale
[3], translate
[3];
507 get_viewport_xform(&viewports
[i
], scale
, translate
);
508 radeon_emit(cs
, fui(scale
[0]));
509 radeon_emit(cs
, fui(translate
[0]));
510 radeon_emit(cs
, fui(scale
[1]));
511 radeon_emit(cs
, fui(translate
[1]));
512 radeon_emit(cs
, fui(scale
[2]));
513 radeon_emit(cs
, fui(translate
[2]));
516 radeon_set_context_reg_seq(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+
517 first_vp
* 4 * 2, count
* 2);
518 for (i
= 0; i
< count
; i
++) {
519 float zmin
= MIN2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
520 float zmax
= MAX2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
521 radeon_emit(cs
, fui(zmin
));
522 radeon_emit(cs
, fui(zmax
));
526 static VkRect2D
si_scissor_from_viewport(const VkViewport
*viewport
)
528 float scale
[3], translate
[3];
531 get_viewport_xform(viewport
, scale
, translate
);
533 rect
.offset
.x
= translate
[0] - fabs(scale
[0]);
534 rect
.offset
.y
= translate
[1] - fabs(scale
[1]);
535 rect
.extent
.width
= ceilf(translate
[0] + fabs(scale
[0])) - rect
.offset
.x
;
536 rect
.extent
.height
= ceilf(translate
[1] + fabs(scale
[1])) - rect
.offset
.y
;
541 static VkRect2D
si_intersect_scissor(const VkRect2D
*a
, const VkRect2D
*b
) {
543 ret
.offset
.x
= MAX2(a
->offset
.x
, b
->offset
.x
);
544 ret
.offset
.y
= MAX2(a
->offset
.y
, b
->offset
.y
);
545 ret
.extent
.width
= MIN2(a
->offset
.x
+ a
->extent
.width
,
546 b
->offset
.x
+ b
->extent
.width
) - ret
.offset
.x
;
547 ret
.extent
.height
= MIN2(a
->offset
.y
+ a
->extent
.height
,
548 b
->offset
.y
+ b
->extent
.height
) - ret
.offset
.y
;
553 si_write_scissors(struct radeon_cmdbuf
*cs
, int first
,
554 int count
, const VkRect2D
*scissors
,
555 const VkViewport
*viewports
, bool can_use_guardband
)
558 float scale
[3], translate
[3], guardband_x
= INFINITY
, guardband_y
= INFINITY
;
559 const float max_range
= 32767.0f
;
563 radeon_set_context_reg_seq(cs
, R_028250_PA_SC_VPORT_SCISSOR_0_TL
+ first
* 4 * 2, count
* 2);
564 for (i
= 0; i
< count
; i
++) {
565 VkRect2D viewport_scissor
= si_scissor_from_viewport(viewports
+ i
);
566 VkRect2D scissor
= si_intersect_scissor(&scissors
[i
], &viewport_scissor
);
568 get_viewport_xform(viewports
+ i
, scale
, translate
);
569 scale
[0] = fabsf(scale
[0]);
570 scale
[1] = fabsf(scale
[1]);
577 guardband_x
= MIN2(guardband_x
, (max_range
- fabsf(translate
[0])) / scale
[0]);
578 guardband_y
= MIN2(guardband_y
, (max_range
- fabsf(translate
[1])) / scale
[1]);
580 radeon_emit(cs
, S_028250_TL_X(scissor
.offset
.x
) |
581 S_028250_TL_Y(scissor
.offset
.y
) |
582 S_028250_WINDOW_OFFSET_DISABLE(1));
583 radeon_emit(cs
, S_028254_BR_X(scissor
.offset
.x
+ scissor
.extent
.width
) |
584 S_028254_BR_Y(scissor
.offset
.y
+ scissor
.extent
.height
));
586 if (!can_use_guardband
) {
591 radeon_set_context_reg_seq(cs
, R_028BE8_PA_CL_GB_VERT_CLIP_ADJ
, 4);
592 radeon_emit(cs
, fui(guardband_y
));
593 radeon_emit(cs
, fui(1.0));
594 radeon_emit(cs
, fui(guardband_x
));
595 radeon_emit(cs
, fui(1.0));
598 static inline unsigned
599 radv_prims_for_vertices(struct radv_prim_vertex_count
*info
, unsigned num
)
610 return 1 + ((num
- info
->min
) / info
->incr
);
614 si_get_ia_multi_vgt_param(struct radv_cmd_buffer
*cmd_buffer
,
615 bool instanced_draw
, bool indirect_draw
,
616 bool count_from_stream_output
,
617 uint32_t draw_vertex_count
)
619 enum chip_class chip_class
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
;
620 enum radeon_family family
= cmd_buffer
->device
->physical_device
->rad_info
.family
;
621 struct radeon_info
*info
= &cmd_buffer
->device
->physical_device
->rad_info
;
622 const unsigned max_primgroup_in_wave
= 2;
623 /* SWITCH_ON_EOP(0) is always preferable. */
624 bool wd_switch_on_eop
= false;
625 bool ia_switch_on_eop
= false;
626 bool ia_switch_on_eoi
= false;
627 bool partial_vs_wave
= false;
628 bool partial_es_wave
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.partial_es_wave
;
629 bool multi_instances_smaller_than_primgroup
;
631 multi_instances_smaller_than_primgroup
= indirect_draw
;
632 if (!multi_instances_smaller_than_primgroup
&& instanced_draw
) {
633 uint32_t num_prims
= radv_prims_for_vertices(&cmd_buffer
->state
.pipeline
->graphics
.prim_vertex_count
, draw_vertex_count
);
634 if (num_prims
< cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.primgroup_size
)
635 multi_instances_smaller_than_primgroup
= true;
638 ia_switch_on_eoi
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.ia_switch_on_eoi
;
639 partial_vs_wave
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.partial_vs_wave
;
641 if (chip_class
>= GFX7
) {
642 wd_switch_on_eop
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.wd_switch_on_eop
;
644 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
645 * We don't know that for indirect drawing, so treat it as
646 * always problematic. */
647 if (family
== CHIP_HAWAII
&&
648 (instanced_draw
|| indirect_draw
))
649 wd_switch_on_eop
= true;
651 /* Performance recommendation for 4 SE Gfx7-8 parts if
652 * instances are smaller than a primgroup.
653 * Assume indirect draws always use small instances.
654 * This is needed for good VS wave utilization.
656 if (chip_class
<= GFX8
&&
658 multi_instances_smaller_than_primgroup
)
659 wd_switch_on_eop
= true;
661 /* Required on GFX7 and later. */
662 if (info
->max_se
> 2 && !wd_switch_on_eop
)
663 ia_switch_on_eoi
= true;
665 /* Required by Hawaii and, for some special cases, by GFX8. */
666 if (ia_switch_on_eoi
&&
667 (family
== CHIP_HAWAII
||
668 (chip_class
== GFX8
&&
669 /* max primgroup in wave is always 2 - leave this for documentation */
670 (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
) || max_primgroup_in_wave
!= 2))))
671 partial_vs_wave
= true;
673 /* Instancing bug on Bonaire. */
674 if (family
== CHIP_BONAIRE
&& ia_switch_on_eoi
&&
675 (instanced_draw
|| indirect_draw
))
676 partial_vs_wave
= true;
678 /* Hardware requirement when drawing primitives from a stream
681 if (count_from_stream_output
)
682 wd_switch_on_eop
= true;
684 /* If the WD switch is false, the IA switch must be false too. */
685 assert(wd_switch_on_eop
|| !ia_switch_on_eop
);
687 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
688 if (chip_class
<= GFX8
&& ia_switch_on_eoi
)
689 partial_es_wave
= true;
691 if (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
)) {
692 /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
693 * The hw doc says all multi-SE chips are affected, but amdgpu-pro Vulkan
694 * only applies it to Hawaii. Do what amdgpu-pro Vulkan does.
696 if (family
== CHIP_HAWAII
&& ia_switch_on_eoi
) {
697 bool set_vgt_flush
= indirect_draw
;
698 if (!set_vgt_flush
&& instanced_draw
) {
699 uint32_t num_prims
= radv_prims_for_vertices(&cmd_buffer
->state
.pipeline
->graphics
.prim_vertex_count
, draw_vertex_count
);
701 set_vgt_flush
= true;
704 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_VGT_FLUSH
;
708 return cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.base
|
709 S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop
) |
710 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi
) |
711 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave
) |
712 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave
) |
713 S_028AA8_WD_SWITCH_ON_EOP(chip_class
>= GFX7
? wd_switch_on_eop
: 0);
717 void si_cs_emit_write_event_eop(struct radeon_cmdbuf
*cs
,
718 enum chip_class chip_class
,
720 unsigned event
, unsigned event_flags
,
724 uint64_t gfx9_eop_bug_va
)
726 unsigned op
= EVENT_TYPE(event
) |
729 unsigned is_gfx8_mec
= is_mec
&& chip_class
< GFX9
;
730 unsigned sel
= EOP_DATA_SEL(data_sel
);
732 /* Wait for write confirmation before writing data, but don't send
734 if (data_sel
!= EOP_DATA_SEL_DISCARD
)
735 sel
|= EOP_INT_SEL(EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM
);
737 if (chip_class
>= GFX9
|| is_gfx8_mec
) {
738 /* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
739 * counters) must immediately precede every timestamp event to
740 * prevent a GPU hang on GFX9.
742 if (chip_class
== GFX9
&& !is_mec
) {
743 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 2, 0));
744 radeon_emit(cs
, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE
) | EVENT_INDEX(1));
745 radeon_emit(cs
, gfx9_eop_bug_va
);
746 radeon_emit(cs
, gfx9_eop_bug_va
>> 32);
749 radeon_emit(cs
, PKT3(PKT3_RELEASE_MEM
, is_gfx8_mec
? 5 : 6, false));
751 radeon_emit(cs
, sel
);
752 radeon_emit(cs
, va
); /* address lo */
753 radeon_emit(cs
, va
>> 32); /* address hi */
754 radeon_emit(cs
, new_fence
); /* immediate data lo */
755 radeon_emit(cs
, 0); /* immediate data hi */
757 radeon_emit(cs
, 0); /* unused */
759 if (chip_class
== GFX7
||
760 chip_class
== GFX8
) {
761 /* Two EOP events are required to make all engines go idle
762 * (and optional cache flushes executed) before the timestamp
765 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, false));
768 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
769 radeon_emit(cs
, 0); /* immediate data */
770 radeon_emit(cs
, 0); /* unused */
773 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, false));
776 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
777 radeon_emit(cs
, new_fence
); /* immediate data */
778 radeon_emit(cs
, 0); /* unused */
783 radv_cp_wait_mem(struct radeon_cmdbuf
*cs
, uint32_t op
, uint64_t va
,
784 uint32_t ref
, uint32_t mask
)
786 assert(op
== WAIT_REG_MEM_EQUAL
||
787 op
== WAIT_REG_MEM_NOT_EQUAL
||
788 op
== WAIT_REG_MEM_GREATER_OR_EQUAL
);
790 radeon_emit(cs
, PKT3(PKT3_WAIT_REG_MEM
, 5, false));
791 radeon_emit(cs
, op
| WAIT_REG_MEM_MEM_SPACE(1));
793 radeon_emit(cs
, va
>> 32);
794 radeon_emit(cs
, ref
); /* reference value */
795 radeon_emit(cs
, mask
); /* mask */
796 radeon_emit(cs
, 4); /* poll interval */
800 si_emit_acquire_mem(struct radeon_cmdbuf
*cs
,
803 unsigned cp_coher_cntl
)
805 if (is_mec
|| is_gfx9
) {
806 uint32_t hi_val
= is_gfx9
? 0xffffff : 0xff;
807 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 5, false) |
808 PKT3_SHADER_TYPE_S(is_mec
));
809 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
810 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
811 radeon_emit(cs
, hi_val
); /* CP_COHER_SIZE_HI */
812 radeon_emit(cs
, 0); /* CP_COHER_BASE */
813 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
814 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
816 /* ACQUIRE_MEM is only required on a compute ring. */
817 radeon_emit(cs
, PKT3(PKT3_SURFACE_SYNC
, 3, false));
818 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
819 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
820 radeon_emit(cs
, 0); /* CP_COHER_BASE */
821 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
826 si_cs_emit_cache_flush(struct radeon_cmdbuf
*cs
,
827 enum chip_class chip_class
,
831 enum radv_cmd_flush_bits flush_bits
,
832 uint64_t gfx9_eop_bug_va
)
834 unsigned cp_coher_cntl
= 0;
835 uint32_t flush_cb_db
= flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
836 RADV_CMD_FLAG_FLUSH_AND_INV_DB
);
838 if (flush_bits
& RADV_CMD_FLAG_INV_ICACHE
)
839 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
840 if (flush_bits
& RADV_CMD_FLAG_INV_SCACHE
)
841 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
843 if (chip_class
<= GFX8
) {
844 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB
) {
845 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
846 S_0085F0_CB0_DEST_BASE_ENA(1) |
847 S_0085F0_CB1_DEST_BASE_ENA(1) |
848 S_0085F0_CB2_DEST_BASE_ENA(1) |
849 S_0085F0_CB3_DEST_BASE_ENA(1) |
850 S_0085F0_CB4_DEST_BASE_ENA(1) |
851 S_0085F0_CB5_DEST_BASE_ENA(1) |
852 S_0085F0_CB6_DEST_BASE_ENA(1) |
853 S_0085F0_CB7_DEST_BASE_ENA(1);
855 /* Necessary for DCC */
856 if (chip_class
>= GFX8
) {
857 si_cs_emit_write_event_eop(cs
,
860 V_028A90_FLUSH_AND_INV_CB_DATA_TS
,
862 EOP_DATA_SEL_DISCARD
,
867 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB
) {
868 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
869 S_0085F0_DB_DEST_BASE_ENA(1);
873 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
) {
874 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
875 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
878 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
) {
879 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
880 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
883 if (flush_bits
& RADV_CMD_FLAG_PS_PARTIAL_FLUSH
) {
884 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
885 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
886 } else if (flush_bits
& RADV_CMD_FLAG_VS_PARTIAL_FLUSH
) {
887 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
888 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
891 if (flush_bits
& RADV_CMD_FLAG_CS_PARTIAL_FLUSH
) {
892 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
893 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
896 if (chip_class
>= GFX9
&& flush_cb_db
) {
897 unsigned cb_db_event
, tc_flags
;
899 /* Set the CB/DB flush event. */
900 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
902 /* These are the only allowed combinations. If you need to
903 * do multiple operations at once, do them separately.
904 * All operations that invalidate L2 also seem to invalidate
905 * metadata. Volatile (VOL) and WC flushes are not listed here.
907 * TC | TC_WB = writeback & invalidate L2 & L1
908 * TC | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
909 * TC_WB | TC_NC = writeback L2 for MTYPE == NC
910 * TC | TC_NC = invalidate L2 for MTYPE == NC
911 * TC | TC_MD = writeback & invalidate L2 metadata (DCC, etc.)
912 * TCL1 = invalidate L1
914 tc_flags
= EVENT_TC_ACTION_ENA
|
915 EVENT_TC_MD_ACTION_ENA
;
917 /* Ideally flush TC together with CB/DB. */
918 if (flush_bits
& RADV_CMD_FLAG_INV_L2
) {
919 /* Writeback and invalidate everything in L2 & L1. */
920 tc_flags
= EVENT_TC_ACTION_ENA
|
921 EVENT_TC_WB_ACTION_ENA
;
924 /* Clear the flags. */
925 flush_bits
&= ~(RADV_CMD_FLAG_INV_L2
|
926 RADV_CMD_FLAG_WB_L2
|
927 RADV_CMD_FLAG_INV_VCACHE
);
932 si_cs_emit_write_event_eop(cs
, chip_class
, false, cb_db_event
, tc_flags
,
933 EOP_DATA_SEL_VALUE_32BIT
,
934 flush_va
, *flush_cnt
,
936 radv_cp_wait_mem(cs
, WAIT_REG_MEM_EQUAL
, flush_va
,
937 *flush_cnt
, 0xffffffff);
941 if (flush_bits
& RADV_CMD_FLAG_VGT_FLUSH
) {
942 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
943 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
946 /* VGT streamout state sync */
947 if (flush_bits
& RADV_CMD_FLAG_VGT_STREAMOUT_SYNC
) {
948 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
949 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC
) | EVENT_INDEX(0));
952 /* Make sure ME is idle (it executes most packets) before continuing.
953 * This prevents read-after-write hazards between PFP and ME.
955 if ((cp_coher_cntl
||
956 (flush_bits
& (RADV_CMD_FLAG_CS_PARTIAL_FLUSH
|
957 RADV_CMD_FLAG_INV_VCACHE
|
958 RADV_CMD_FLAG_INV_L2
|
959 RADV_CMD_FLAG_WB_L2
))) &&
961 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
965 if ((flush_bits
& RADV_CMD_FLAG_INV_L2
) ||
966 (chip_class
<= GFX7
&& (flush_bits
& RADV_CMD_FLAG_WB_L2
))) {
967 si_emit_acquire_mem(cs
, is_mec
, chip_class
>= GFX9
,
969 S_0085F0_TC_ACTION_ENA(1) |
970 S_0085F0_TCL1_ACTION_ENA(1) |
971 S_0301F0_TC_WB_ACTION_ENA(chip_class
>= GFX8
));
974 if(flush_bits
& RADV_CMD_FLAG_WB_L2
) {
976 * NC = apply to non-coherent MTYPEs
977 * (i.e. MTYPE <= 1, which is what we use everywhere)
979 * WB doesn't work without NC.
981 si_emit_acquire_mem(cs
, is_mec
,
984 S_0301F0_TC_WB_ACTION_ENA(1) |
985 S_0301F0_TC_NC_ACTION_ENA(1));
988 if (flush_bits
& RADV_CMD_FLAG_INV_VCACHE
) {
989 si_emit_acquire_mem(cs
, is_mec
,
992 S_0085F0_TCL1_ACTION_ENA(1));
997 /* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
998 * Therefore, it should be last. Done in PFP.
1001 si_emit_acquire_mem(cs
, is_mec
, chip_class
>= GFX9
, cp_coher_cntl
);
1003 if (flush_bits
& RADV_CMD_FLAG_START_PIPELINE_STATS
) {
1004 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1005 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
1007 } else if (flush_bits
& RADV_CMD_FLAG_STOP_PIPELINE_STATS
) {
1008 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1009 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
1015 si_emit_cache_flush(struct radv_cmd_buffer
*cmd_buffer
)
1017 bool is_compute
= cmd_buffer
->queue_family_index
== RADV_QUEUE_COMPUTE
;
1020 cmd_buffer
->state
.flush_bits
&= ~(RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
1021 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
|
1022 RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
1023 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
|
1024 RADV_CMD_FLAG_PS_PARTIAL_FLUSH
|
1025 RADV_CMD_FLAG_VS_PARTIAL_FLUSH
|
1026 RADV_CMD_FLAG_VGT_FLUSH
|
1027 RADV_CMD_FLAG_START_PIPELINE_STATS
|
1028 RADV_CMD_FLAG_STOP_PIPELINE_STATS
);
1030 if (!cmd_buffer
->state
.flush_bits
)
1033 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 128);
1035 si_cs_emit_cache_flush(cmd_buffer
->cs
,
1036 cmd_buffer
->device
->physical_device
->rad_info
.chip_class
,
1037 &cmd_buffer
->gfx9_fence_idx
,
1038 cmd_buffer
->gfx9_fence_va
,
1039 radv_cmd_buffer_uses_mec(cmd_buffer
),
1040 cmd_buffer
->state
.flush_bits
,
1041 cmd_buffer
->gfx9_eop_bug_va
);
1044 if (unlikely(cmd_buffer
->device
->trace_bo
))
1045 radv_cmd_buffer_trace_emit(cmd_buffer
);
1047 /* Clear the caches that have been flushed to avoid syncing too much
1048 * when there is some pending active queries.
1050 cmd_buffer
->active_query_flush_bits
&= ~cmd_buffer
->state
.flush_bits
;
1052 cmd_buffer
->state
.flush_bits
= 0;
1054 /* If the driver used a compute shader for resetting a query pool, it
1055 * should be finished at this point.
1057 cmd_buffer
->pending_reset_query
= false;
1060 /* sets the CP predication state using a boolean stored at va */
1062 si_emit_set_predication_state(struct radv_cmd_buffer
*cmd_buffer
,
1063 bool draw_visible
, uint64_t va
)
1068 op
= PRED_OP(PREDICATION_OP_BOOL64
);
1070 /* PREDICATION_DRAW_VISIBLE means that if the 32-bit value is
1071 * zero, all rendering commands are discarded. Otherwise, they
1072 * are discarded if the value is non zero.
1074 op
|= draw_visible
? PREDICATION_DRAW_VISIBLE
:
1075 PREDICATION_DRAW_NOT_VISIBLE
;
1077 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1078 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_SET_PREDICATION
, 2, 0));
1079 radeon_emit(cmd_buffer
->cs
, op
);
1080 radeon_emit(cmd_buffer
->cs
, va
);
1081 radeon_emit(cmd_buffer
->cs
, va
>> 32);
1083 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_SET_PREDICATION
, 1, 0));
1084 radeon_emit(cmd_buffer
->cs
, va
);
1085 radeon_emit(cmd_buffer
->cs
, op
| ((va
>> 32) & 0xFF));
1089 /* Set this if you want the 3D engine to wait until CP DMA is done.
1090 * It should be set on the last CP DMA packet. */
1091 #define CP_DMA_SYNC (1 << 0)
1093 /* Set this if the source data was used as a destination in a previous CP DMA
1094 * packet. It's for preventing a read-after-write (RAW) hazard between two
1095 * CP DMA packets. */
1096 #define CP_DMA_RAW_WAIT (1 << 1)
1097 #define CP_DMA_USE_L2 (1 << 2)
1098 #define CP_DMA_CLEAR (1 << 3)
1100 /* Alignment for optimal performance. */
1101 #define SI_CPDMA_ALIGNMENT 32
1103 /* The max number of bytes that can be copied per packet. */
1104 static inline unsigned cp_dma_max_byte_count(struct radv_cmd_buffer
*cmd_buffer
)
1106 unsigned max
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
?
1107 S_414_BYTE_COUNT_GFX9(~0u) :
1108 S_414_BYTE_COUNT_GFX6(~0u);
1110 /* make it aligned for optimal performance */
1111 return max
& ~(SI_CPDMA_ALIGNMENT
- 1);
1114 /* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
1115 * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
1118 static void si_emit_cp_dma(struct radv_cmd_buffer
*cmd_buffer
,
1119 uint64_t dst_va
, uint64_t src_va
,
1120 unsigned size
, unsigned flags
)
1122 struct radeon_cmdbuf
*cs
= cmd_buffer
->cs
;
1123 uint32_t header
= 0, command
= 0;
1125 assert(size
<= cp_dma_max_byte_count(cmd_buffer
));
1127 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 9);
1128 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1129 command
|= S_414_BYTE_COUNT_GFX9(size
);
1131 command
|= S_414_BYTE_COUNT_GFX6(size
);
1134 if (flags
& CP_DMA_SYNC
)
1135 header
|= S_411_CP_SYNC(1);
1137 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1138 command
|= S_414_DISABLE_WR_CONFIRM_GFX9(1);
1140 command
|= S_414_DISABLE_WR_CONFIRM_GFX6(1);
1143 if (flags
& CP_DMA_RAW_WAIT
)
1144 command
|= S_414_RAW_WAIT(1);
1146 /* Src and dst flags. */
1147 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1148 !(flags
& CP_DMA_CLEAR
) &&
1150 header
|= S_411_DST_SEL(V_411_NOWHERE
); /* prefetch only */
1151 else if (flags
& CP_DMA_USE_L2
)
1152 header
|= S_411_DST_SEL(V_411_DST_ADDR_TC_L2
);
1154 if (flags
& CP_DMA_CLEAR
)
1155 header
|= S_411_SRC_SEL(V_411_DATA
);
1156 else if (flags
& CP_DMA_USE_L2
)
1157 header
|= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2
);
1159 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
1160 radeon_emit(cs
, PKT3(PKT3_DMA_DATA
, 5, cmd_buffer
->state
.predicating
));
1161 radeon_emit(cs
, header
);
1162 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
1163 radeon_emit(cs
, src_va
>> 32); /* SRC_ADDR_HI [31:0] */
1164 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
1165 radeon_emit(cs
, dst_va
>> 32); /* DST_ADDR_HI [31:0] */
1166 radeon_emit(cs
, command
);
1168 assert(!(flags
& CP_DMA_USE_L2
));
1169 header
|= S_411_SRC_ADDR_HI(src_va
>> 32);
1170 radeon_emit(cs
, PKT3(PKT3_CP_DMA
, 4, cmd_buffer
->state
.predicating
));
1171 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
1172 radeon_emit(cs
, header
); /* SRC_ADDR_HI [15:0] + flags. */
1173 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
1174 radeon_emit(cs
, (dst_va
>> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
1175 radeon_emit(cs
, command
);
1178 /* CP DMA is executed in ME, but index buffers are read by PFP.
1179 * This ensures that ME (CP DMA) is idle before PFP starts fetching
1180 * indices. If we wanted to execute CP DMA in PFP, this packet
1181 * should precede it.
1183 if (flags
& CP_DMA_SYNC
) {
1184 if (cmd_buffer
->queue_family_index
== RADV_QUEUE_GENERAL
) {
1185 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, cmd_buffer
->state
.predicating
));
1189 /* CP will see the sync flag and wait for all DMAs to complete. */
1190 cmd_buffer
->state
.dma_is_busy
= false;
1193 if (unlikely(cmd_buffer
->device
->trace_bo
))
1194 radv_cmd_buffer_trace_emit(cmd_buffer
);
1197 void si_cp_dma_prefetch(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1200 uint64_t aligned_va
= va
& ~(SI_CPDMA_ALIGNMENT
- 1);
1201 uint64_t aligned_size
= ((va
+ size
+ SI_CPDMA_ALIGNMENT
-1) & ~(SI_CPDMA_ALIGNMENT
- 1)) - aligned_va
;
1203 si_emit_cp_dma(cmd_buffer
, aligned_va
, aligned_va
,
1204 aligned_size
, CP_DMA_USE_L2
);
1207 static void si_cp_dma_prepare(struct radv_cmd_buffer
*cmd_buffer
, uint64_t byte_count
,
1208 uint64_t remaining_size
, unsigned *flags
)
1211 /* Flush the caches for the first copy only.
1212 * Also wait for the previous CP DMA operations.
1214 if (cmd_buffer
->state
.flush_bits
) {
1215 si_emit_cache_flush(cmd_buffer
);
1216 *flags
|= CP_DMA_RAW_WAIT
;
1219 /* Do the synchronization after the last dma, so that all data
1220 * is written to memory.
1222 if (byte_count
== remaining_size
)
1223 *flags
|= CP_DMA_SYNC
;
1226 static void si_cp_dma_realign_engine(struct radv_cmd_buffer
*cmd_buffer
, unsigned size
)
1230 unsigned dma_flags
= 0;
1231 unsigned buf_size
= SI_CPDMA_ALIGNMENT
* 2;
1234 assert(size
< SI_CPDMA_ALIGNMENT
);
1236 radv_cmd_buffer_upload_alloc(cmd_buffer
, buf_size
, SI_CPDMA_ALIGNMENT
, &offset
, &ptr
);
1238 va
= radv_buffer_get_va(cmd_buffer
->upload
.upload_bo
);
1241 si_cp_dma_prepare(cmd_buffer
, size
, size
, &dma_flags
);
1243 si_emit_cp_dma(cmd_buffer
, va
, va
+ SI_CPDMA_ALIGNMENT
, size
,
1247 void si_cp_dma_buffer_copy(struct radv_cmd_buffer
*cmd_buffer
,
1248 uint64_t src_va
, uint64_t dest_va
,
1251 uint64_t main_src_va
, main_dest_va
;
1252 uint64_t skipped_size
= 0, realign_size
= 0;
1254 /* Assume that we are not going to sync after the last DMA operation. */
1255 cmd_buffer
->state
.dma_is_busy
= true;
1257 if (cmd_buffer
->device
->physical_device
->rad_info
.family
<= CHIP_CARRIZO
||
1258 cmd_buffer
->device
->physical_device
->rad_info
.family
== CHIP_STONEY
) {
1259 /* If the size is not aligned, we must add a dummy copy at the end
1260 * just to align the internal counter. Otherwise, the DMA engine
1261 * would slow down by an order of magnitude for following copies.
1263 if (size
% SI_CPDMA_ALIGNMENT
)
1264 realign_size
= SI_CPDMA_ALIGNMENT
- (size
% SI_CPDMA_ALIGNMENT
);
1266 /* If the copy begins unaligned, we must start copying from the next
1267 * aligned block and the skipped part should be copied after everything
1268 * else has been copied. Only the src alignment matters, not dst.
1270 if (src_va
% SI_CPDMA_ALIGNMENT
) {
1271 skipped_size
= SI_CPDMA_ALIGNMENT
- (src_va
% SI_CPDMA_ALIGNMENT
);
1272 /* The main part will be skipped if the size is too small. */
1273 skipped_size
= MIN2(skipped_size
, size
);
1274 size
-= skipped_size
;
1277 main_src_va
= src_va
+ skipped_size
;
1278 main_dest_va
= dest_va
+ skipped_size
;
1281 unsigned dma_flags
= 0;
1282 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(cmd_buffer
));
1284 si_cp_dma_prepare(cmd_buffer
, byte_count
,
1285 size
+ skipped_size
+ realign_size
,
1288 dma_flags
&= ~CP_DMA_SYNC
;
1290 si_emit_cp_dma(cmd_buffer
, main_dest_va
, main_src_va
,
1291 byte_count
, dma_flags
);
1294 main_src_va
+= byte_count
;
1295 main_dest_va
+= byte_count
;
1299 unsigned dma_flags
= 0;
1301 si_cp_dma_prepare(cmd_buffer
, skipped_size
,
1302 size
+ skipped_size
+ realign_size
,
1305 si_emit_cp_dma(cmd_buffer
, dest_va
, src_va
,
1306 skipped_size
, dma_flags
);
1309 si_cp_dma_realign_engine(cmd_buffer
, realign_size
);
1312 void si_cp_dma_clear_buffer(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1313 uint64_t size
, unsigned value
)
1319 assert(va
% 4 == 0 && size
% 4 == 0);
1321 /* Assume that we are not going to sync after the last DMA operation. */
1322 cmd_buffer
->state
.dma_is_busy
= true;
1325 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(cmd_buffer
));
1326 unsigned dma_flags
= CP_DMA_CLEAR
;
1328 si_cp_dma_prepare(cmd_buffer
, byte_count
, size
, &dma_flags
);
1330 /* Emit the clear packet. */
1331 si_emit_cp_dma(cmd_buffer
, va
, value
, byte_count
,
1339 void si_cp_dma_wait_for_idle(struct radv_cmd_buffer
*cmd_buffer
)
1341 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
< GFX7
)
1344 if (!cmd_buffer
->state
.dma_is_busy
)
1347 /* Issue a dummy DMA that copies zero bytes.
1349 * The DMA engine will see that there's no work to do and skip this
1350 * DMA request, however, the CP will see the sync flag and still wait
1351 * for all DMAs to complete.
1353 si_emit_cp_dma(cmd_buffer
, 0, 0, 0, CP_DMA_SYNC
);
1355 cmd_buffer
->state
.dma_is_busy
= false;
1358 /* For MSAA sample positions. */
1359 #define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y) \
1360 ((((unsigned)(s0x) & 0xf) << 0) | (((unsigned)(s0y) & 0xf) << 4) | \
1361 (((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) | \
1362 (((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) | \
1363 (((unsigned)(s3x) & 0xf) << 24) | (((unsigned)(s3y) & 0xf) << 28))
1365 /* For obtaining location coordinates from registers */
1366 #define SEXT4(x) ((int)((x) | ((x) & 0x8 ? 0xfffffff0 : 0)))
1367 #define GET_SFIELD(reg, index) SEXT4(((reg) >> ((index) * 4)) & 0xf)
1368 #define GET_SX(reg, index) GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2)
1369 #define GET_SY(reg, index) GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2 + 1)
1372 static const uint32_t sample_locs_1x
=
1373 FILL_SREG(0, 0, 0, 0, 0, 0, 0, 0);
1374 static const unsigned max_dist_1x
= 0;
1375 static const uint64_t centroid_priority_1x
= 0x0000000000000000ull
;
1378 static const uint32_t sample_locs_2x
=
1379 FILL_SREG(4,4, -4, -4, 0, 0, 0, 0);
1380 static const unsigned max_dist_2x
= 4;
1381 static const uint64_t centroid_priority_2x
= 0x1010101010101010ull
;
1384 static const uint32_t sample_locs_4x
=
1385 FILL_SREG(-2,-6, 6, -2, -6, 2, 2, 6);
1386 static const unsigned max_dist_4x
= 6;
1387 static const uint64_t centroid_priority_4x
= 0x3210321032103210ull
;
1390 static const uint32_t sample_locs_8x
[] = {
1391 FILL_SREG( 1,-3, -1, 3, 5, 1, -3,-5),
1392 FILL_SREG(-5, 5, -7,-1, 3, 7, 7,-7),
1393 /* The following are unused by hardware, but we emit them to IBs
1394 * instead of multiple SET_CONTEXT_REG packets. */
1398 static const unsigned max_dist_8x
= 7;
1399 static const uint64_t centroid_priority_8x
= 0x7654321076543210ull
;
1401 unsigned radv_get_default_max_sample_dist(int log_samples
)
1403 unsigned max_dist
[] = {
1409 return max_dist
[log_samples
];
1412 void radv_emit_default_sample_locations(struct radeon_cmdbuf
*cs
, int nr_samples
)
1414 switch (nr_samples
) {
1417 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1418 radeon_emit(cs
, (uint32_t)centroid_priority_1x
);
1419 radeon_emit(cs
, centroid_priority_1x
>> 32);
1420 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, sample_locs_1x
);
1421 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, sample_locs_1x
);
1422 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, sample_locs_1x
);
1423 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, sample_locs_1x
);
1426 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1427 radeon_emit(cs
, (uint32_t)centroid_priority_2x
);
1428 radeon_emit(cs
, centroid_priority_2x
>> 32);
1429 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, sample_locs_2x
);
1430 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, sample_locs_2x
);
1431 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, sample_locs_2x
);
1432 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, sample_locs_2x
);
1435 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1436 radeon_emit(cs
, (uint32_t)centroid_priority_4x
);
1437 radeon_emit(cs
, centroid_priority_4x
>> 32);
1438 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, sample_locs_4x
);
1439 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, sample_locs_4x
);
1440 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, sample_locs_4x
);
1441 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, sample_locs_4x
);
1444 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1445 radeon_emit(cs
, (uint32_t)centroid_priority_8x
);
1446 radeon_emit(cs
, centroid_priority_8x
>> 32);
1447 radeon_set_context_reg_seq(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 14);
1448 radeon_emit_array(cs
, sample_locs_8x
, 4);
1449 radeon_emit_array(cs
, sample_locs_8x
, 4);
1450 radeon_emit_array(cs
, sample_locs_8x
, 4);
1451 radeon_emit_array(cs
, sample_locs_8x
, 2);
1456 static void radv_get_sample_position(struct radv_device
*device
,
1457 unsigned sample_count
,
1458 unsigned sample_index
, float *out_value
)
1460 const uint32_t *sample_locs
;
1462 switch (sample_count
) {
1465 sample_locs
= &sample_locs_1x
;
1468 sample_locs
= &sample_locs_2x
;
1471 sample_locs
= &sample_locs_4x
;
1474 sample_locs
= sample_locs_8x
;
1478 out_value
[0] = (GET_SX(sample_locs
, sample_index
) + 8) / 16.0f
;
1479 out_value
[1] = (GET_SY(sample_locs
, sample_index
) + 8) / 16.0f
;
1482 void radv_device_init_msaa(struct radv_device
*device
)
1486 radv_get_sample_position(device
, 1, 0, device
->sample_locations_1x
[0]);
1488 for (i
= 0; i
< 2; i
++)
1489 radv_get_sample_position(device
, 2, i
, device
->sample_locations_2x
[i
]);
1490 for (i
= 0; i
< 4; i
++)
1491 radv_get_sample_position(device
, 4, i
, device
->sample_locations_4x
[i
]);
1492 for (i
= 0; i
< 8; i
++)
1493 radv_get_sample_position(device
, 8, i
, device
->sample_locations_8x
[i
]);