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"
37 si_write_harvested_raster_configs(struct radv_physical_device
*physical_device
,
38 struct radeon_cmdbuf
*cs
,
39 unsigned raster_config
,
40 unsigned raster_config_1
)
42 unsigned num_se
= MAX2(physical_device
->rad_info
.max_se
, 1);
43 unsigned raster_config_se
[4];
46 ac_get_harvested_configs(&physical_device
->rad_info
,
51 for (se
= 0; se
< num_se
; se
++) {
52 /* GRBM_GFX_INDEX has a different offset on GFX6 and GFX7+ */
53 if (physical_device
->rad_info
.chip_class
< GFX7
)
54 radeon_set_config_reg(cs
, R_00802C_GRBM_GFX_INDEX
,
55 S_00802C_SE_INDEX(se
) |
56 S_00802C_SH_BROADCAST_WRITES(1) |
57 S_00802C_INSTANCE_BROADCAST_WRITES(1));
59 radeon_set_uconfig_reg(cs
, R_030800_GRBM_GFX_INDEX
,
60 S_030800_SE_INDEX(se
) | S_030800_SH_BROADCAST_WRITES(1) |
61 S_030800_INSTANCE_BROADCAST_WRITES(1));
62 radeon_set_context_reg(cs
, R_028350_PA_SC_RASTER_CONFIG
, raster_config_se
[se
]);
65 /* GRBM_GFX_INDEX has a different offset on GFX6 and GFX7+ */
66 if (physical_device
->rad_info
.chip_class
< GFX7
)
67 radeon_set_config_reg(cs
, R_00802C_GRBM_GFX_INDEX
,
68 S_00802C_SE_BROADCAST_WRITES(1) |
69 S_00802C_SH_BROADCAST_WRITES(1) |
70 S_00802C_INSTANCE_BROADCAST_WRITES(1));
72 radeon_set_uconfig_reg(cs
, R_030800_GRBM_GFX_INDEX
,
73 S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
74 S_030800_INSTANCE_BROADCAST_WRITES(1));
76 if (physical_device
->rad_info
.chip_class
>= GFX7
)
77 radeon_set_context_reg(cs
, R_028354_PA_SC_RASTER_CONFIG_1
, raster_config_1
);
81 si_emit_compute(struct radv_physical_device
*physical_device
,
82 struct radeon_cmdbuf
*cs
)
84 radeon_set_sh_reg_seq(cs
, R_00B810_COMPUTE_START_X
, 3);
89 radeon_set_sh_reg_seq(cs
, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0
, 2);
90 /* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1,
91 * renamed COMPUTE_DESTINATION_EN_SEn on gfx10. */
92 radeon_emit(cs
, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
93 radeon_emit(cs
, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
95 if (physical_device
->rad_info
.chip_class
>= GFX7
) {
96 /* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
97 radeon_set_sh_reg_seq(cs
,
98 R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2
, 2);
99 radeon_emit(cs
, S_00B858_SH0_CU_EN(0xffff) |
100 S_00B858_SH1_CU_EN(0xffff));
101 radeon_emit(cs
, S_00B858_SH0_CU_EN(0xffff) |
102 S_00B858_SH1_CU_EN(0xffff));
105 if (physical_device
->rad_info
.chip_class
>= GFX10
)
106 radeon_set_sh_reg(cs
, R_00B8A0_COMPUTE_PGM_RSRC3
, 0);
108 /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
109 * and is now per pipe, so it should be handled in the
110 * kernel if we want to use something other than the default value,
111 * which is now 0x22f.
113 if (physical_device
->rad_info
.chip_class
<= GFX6
) {
114 /* XXX: This should be:
115 * (number of compute units) * 4 * (waves per simd) - 1 */
117 radeon_set_sh_reg(cs
, R_00B82C_COMPUTE_MAX_WAVE_ID
,
118 0x190 /* Default value */);
122 /* 12.4 fixed-point */
123 static unsigned radv_pack_float_12p4(float x
)
126 x
>= 4096 ? 0xffff : x
* 16;
130 si_set_raster_config(struct radv_physical_device
*physical_device
,
131 struct radeon_cmdbuf
*cs
)
133 unsigned num_rb
= MIN2(physical_device
->rad_info
.num_render_backends
, 16);
134 unsigned rb_mask
= physical_device
->rad_info
.enabled_rb_mask
;
135 unsigned raster_config
, raster_config_1
;
137 ac_get_raster_config(&physical_device
->rad_info
,
139 &raster_config_1
, NULL
);
141 /* Always use the default config when all backends are enabled
142 * (or when we failed to determine the enabled backends).
144 if (!rb_mask
|| util_bitcount(rb_mask
) >= num_rb
) {
145 radeon_set_context_reg(cs
, R_028350_PA_SC_RASTER_CONFIG
,
147 if (physical_device
->rad_info
.chip_class
>= GFX7
)
148 radeon_set_context_reg(cs
, R_028354_PA_SC_RASTER_CONFIG_1
,
151 si_write_harvested_raster_configs(physical_device
, cs
,
158 si_emit_graphics(struct radv_device
*device
,
159 struct radeon_cmdbuf
*cs
)
161 struct radv_physical_device
*physical_device
= device
->physical_device
;
163 bool has_clear_state
= physical_device
->rad_info
.has_clear_state
;
166 radeon_emit(cs
, PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
167 radeon_emit(cs
, CC0_UPDATE_LOAD_ENABLES(1));
168 radeon_emit(cs
, CC1_UPDATE_SHADOW_ENABLES(1));
170 if (has_clear_state
) {
171 radeon_emit(cs
, PKT3(PKT3_CLEAR_STATE
, 0, 0));
175 if (physical_device
->rad_info
.chip_class
<= GFX8
)
176 si_set_raster_config(physical_device
, cs
);
178 radeon_set_context_reg(cs
, R_028A18_VGT_HOS_MAX_TESS_LEVEL
, fui(64));
179 if (!has_clear_state
)
180 radeon_set_context_reg(cs
, R_028A1C_VGT_HOS_MIN_TESS_LEVEL
, fui(0));
182 /* FIXME calculate these values somehow ??? */
183 if (physical_device
->rad_info
.chip_class
<= GFX8
) {
184 radeon_set_context_reg(cs
, R_028A54_VGT_GS_PER_ES
, SI_GS_PER_ES
);
185 radeon_set_context_reg(cs
, R_028A58_VGT_ES_PER_GS
, 0x40);
188 if (!has_clear_state
) {
189 radeon_set_context_reg(cs
, R_028A5C_VGT_GS_PER_VS
, 0x2);
190 radeon_set_context_reg(cs
, R_028A8C_VGT_PRIMITIVEID_RESET
, 0x0);
191 radeon_set_context_reg(cs
, R_028B98_VGT_STRMOUT_BUFFER_CONFIG
, 0x0);
194 if (physical_device
->rad_info
.chip_class
<= GFX9
)
195 radeon_set_context_reg(cs
, R_028AA0_VGT_INSTANCE_STEP_RATE_0
, 1);
196 if (!has_clear_state
)
197 radeon_set_context_reg(cs
, R_028AB8_VGT_VTX_CNT_EN
, 0x0);
198 if (physical_device
->rad_info
.chip_class
< GFX7
)
199 radeon_set_config_reg(cs
, R_008A14_PA_CL_ENHANCE
, S_008A14_NUM_CLIP_SEQ(3) |
200 S_008A14_CLIP_VTX_REORDER_ENA(1));
202 if (!has_clear_state
)
203 radeon_set_context_reg(cs
, R_02882C_PA_SU_PRIM_FILTER_CNTL
, 0);
205 /* CLEAR_STATE doesn't clear these correctly on certain generations.
206 * I don't know why. Deduced by trial and error.
208 if (physical_device
->rad_info
.chip_class
<= GFX7
|| !has_clear_state
) {
209 radeon_set_context_reg(cs
, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET
, 0);
210 radeon_set_context_reg(cs
, R_028204_PA_SC_WINDOW_SCISSOR_TL
,
211 S_028204_WINDOW_OFFSET_DISABLE(1));
212 radeon_set_context_reg(cs
, R_028240_PA_SC_GENERIC_SCISSOR_TL
,
213 S_028240_WINDOW_OFFSET_DISABLE(1));
214 radeon_set_context_reg(cs
, R_028244_PA_SC_GENERIC_SCISSOR_BR
,
215 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
216 radeon_set_context_reg(cs
, R_028030_PA_SC_SCREEN_SCISSOR_TL
, 0);
217 radeon_set_context_reg(cs
, R_028034_PA_SC_SCREEN_SCISSOR_BR
,
218 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
221 if (!has_clear_state
) {
222 for (i
= 0; i
< 16; i
++) {
223 radeon_set_context_reg(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+ i
*8, 0);
224 radeon_set_context_reg(cs
, R_0282D4_PA_SC_VPORT_ZMAX_0
+ i
*8, fui(1.0));
228 if (!has_clear_state
) {
229 radeon_set_context_reg(cs
, R_02820C_PA_SC_CLIPRECT_RULE
, 0xFFFF);
230 radeon_set_context_reg(cs
, R_028230_PA_SC_EDGERULE
, 0xAAAAAAAA);
231 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on GFX6 */
232 radeon_set_context_reg(cs
, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET
, 0);
233 radeon_set_context_reg(cs
, R_028820_PA_CL_NANINF_CNTL
, 0);
234 radeon_set_context_reg(cs
, R_028AC0_DB_SRESULTS_COMPARE_STATE0
, 0x0);
235 radeon_set_context_reg(cs
, R_028AC4_DB_SRESULTS_COMPARE_STATE1
, 0x0);
236 radeon_set_context_reg(cs
, R_028AC8_DB_PRELOAD_CONTROL
, 0x0);
239 radeon_set_context_reg(cs
, R_02800C_DB_RENDER_OVERRIDE
,
240 S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE
) |
241 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE
));
243 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
244 radeon_set_context_reg(cs
, R_028A98_VGT_DRAW_PAYLOAD_CNTL
, 0);
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 radeon_set_uconfig_reg(cs
, R_03097C_GE_STEREO_CNTL
, 0);
249 radeon_set_uconfig_reg(cs
, R_030988_GE_USER_VGPR_EN
, 0);
250 } else if (physical_device
->rad_info
.chip_class
== GFX9
) {
251 radeon_set_uconfig_reg(cs
, R_030920_VGT_MAX_VTX_INDX
, ~0);
252 radeon_set_uconfig_reg(cs
, R_030924_VGT_MIN_VTX_INDX
, 0);
253 radeon_set_uconfig_reg(cs
, R_030928_VGT_INDX_OFFSET
, 0);
255 /* These registers, when written, also overwrite the
256 * CLEAR_STATE context, so we can't rely on CLEAR_STATE setting
257 * them. It would be an issue if there was another UMD
260 radeon_set_context_reg(cs
, R_028400_VGT_MAX_VTX_INDX
, ~0);
261 radeon_set_context_reg(cs
, R_028404_VGT_MIN_VTX_INDX
, 0);
262 radeon_set_context_reg(cs
, R_028408_VGT_INDX_OFFSET
, 0);
265 if (physical_device
->rad_info
.chip_class
>= GFX7
) {
266 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
267 /* Logical CUs 16 - 31 */
268 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B404_SPI_SHADER_PGM_RSRC4_HS
,
269 3, S_00B404_CU_EN(0xffff));
270 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B104_SPI_SHADER_PGM_RSRC4_VS
,
271 3, S_00B104_CU_EN(0xffff));
272 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B004_SPI_SHADER_PGM_RSRC4_PS
,
273 3, S_00B004_CU_EN(0xffff));
276 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
277 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
,
278 3, S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
280 radeon_set_sh_reg(cs
, R_00B51C_SPI_SHADER_PGM_RSRC3_LS
,
281 S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
282 radeon_set_sh_reg(cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
,
283 S_00B41C_WAVE_LIMIT(0x3F));
284 radeon_set_sh_reg(cs
, R_00B31C_SPI_SHADER_PGM_RSRC3_ES
,
285 S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
286 /* If this is 0, Bonaire can hang even if GS isn't being used.
287 * Other chips are unaffected. These are suboptimal values,
288 * but we don't use on-chip GS.
290 radeon_set_context_reg(cs
, R_028A44_VGT_GS_ONCHIP_CNTL
,
291 S_028A44_ES_VERTS_PER_SUBGRP(64) |
292 S_028A44_GS_PRIMS_PER_SUBGRP(4));
295 /* Compute LATE_ALLOC_VS.LIMIT. */
296 unsigned num_cu_per_sh
= physical_device
->rad_info
.min_good_cu_per_sa
;
297 unsigned late_alloc_wave64
= 0; /* The limit is per SA. */
298 unsigned late_alloc_wave64_gs
= 0;
299 unsigned cu_mask_vs
= 0xffff;
300 unsigned cu_mask_gs
= 0xffff;
302 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
303 /* For Wave32, the hw will launch twice the number of late
304 * alloc waves, so 1 == 2x wave32.
306 if (!physical_device
->rad_info
.use_late_alloc
) {
307 late_alloc_wave64
= 0;
308 } else if (num_cu_per_sh
<= 6) {
309 late_alloc_wave64
= num_cu_per_sh
- 2;
311 late_alloc_wave64
= (num_cu_per_sh
- 2) * 4;
313 /* CU2 & CU3 disabled because of the dual CU design */
315 cu_mask_gs
= 0xfff3; /* NGG only */
318 late_alloc_wave64_gs
= late_alloc_wave64
;
320 /* Don't use late alloc for NGG on Navi14 due to a hw
321 * bug. If NGG is never used, enable all CUs.
323 if (!physical_device
->use_ngg
||
324 physical_device
->rad_info
.family
== CHIP_NAVI14
) {
325 late_alloc_wave64_gs
= 0;
329 if (!physical_device
->rad_info
.use_late_alloc
) {
330 late_alloc_wave64
= 0;
331 } else if (num_cu_per_sh
<= 4) {
332 /* Too few available compute units per SA.
333 * Disallowing VS to run on one CU could hurt
334 * us more than late VS allocation would help.
336 * 2 is the highest safe number that allows us
337 * to keep all CUs enabled.
339 late_alloc_wave64
= 2;
341 /* This is a good initial value, allowing 1
342 * late_alloc wave per SIMD on num_cu - 2.
344 late_alloc_wave64
= (num_cu_per_sh
- 2) * 4;
347 if (late_alloc_wave64
> 2)
348 cu_mask_vs
= 0xfffe; /* 1 CU disabled */
351 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
,
352 3, S_00B118_CU_EN(cu_mask_vs
) |
353 S_00B118_WAVE_LIMIT(0x3F));
354 radeon_set_sh_reg(cs
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
,
355 S_00B11C_LIMIT(late_alloc_wave64
));
357 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B21C_SPI_SHADER_PGM_RSRC3_GS
,
358 3, S_00B21C_CU_EN(cu_mask_gs
) | S_00B21C_WAVE_LIMIT(0x3F));
360 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
361 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B204_SPI_SHADER_PGM_RSRC4_GS
,
362 3, S_00B204_CU_EN(0xffff) |
363 S_00B204_SPI_SHADER_LATE_ALLOC_GS_GFX10(late_alloc_wave64_gs
));
366 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B01C_SPI_SHADER_PGM_RSRC3_PS
,
367 3, S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
370 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
371 /* Break up a pixel wave if it contains deallocs for more than
372 * half the parameter cache.
374 * To avoid a deadlock where pixel waves aren't launched
375 * because they're waiting for more pixels while the frontend
376 * is stuck waiting for PC space, the maximum allowed value is
377 * the size of the PC minus the largest possible allocation for
378 * a single primitive shader subgroup.
380 radeon_set_context_reg(cs
, R_028C50_PA_SC_NGG_MODE_CNTL
,
381 S_028C50_MAX_DEALLOCS_IN_WAVE(512));
382 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 14);
384 /* Enable CMASK/FMASK/HTILE/DCC caching in L2 for small chips. */
385 unsigned meta_write_policy
, meta_read_policy
;
387 /* TODO: investigate whether LRU improves performance on other chips too */
388 if (physical_device
->rad_info
.num_render_backends
<= 4) {
389 meta_write_policy
= V_02807C_CACHE_LRU_WR
; /* cache writes */
390 meta_read_policy
= V_02807C_CACHE_LRU_RD
; /* cache reads */
392 meta_write_policy
= V_02807C_CACHE_STREAM_WR
; /* write combine */
393 meta_read_policy
= V_02807C_CACHE_NOA_RD
; /* don't cache reads */
396 radeon_set_context_reg(cs
, R_02807C_DB_RMI_L2_CACHE_CONTROL
,
397 S_02807C_Z_WR_POLICY(V_02807C_CACHE_STREAM_WR
) |
398 S_02807C_S_WR_POLICY(V_02807C_CACHE_STREAM_WR
) |
399 S_02807C_HTILE_WR_POLICY(meta_write_policy
) |
400 S_02807C_ZPCPSD_WR_POLICY(V_02807C_CACHE_STREAM_WR
) |
401 S_02807C_Z_RD_POLICY(V_02807C_CACHE_NOA_RD
) |
402 S_02807C_S_RD_POLICY(V_02807C_CACHE_NOA_RD
) |
403 S_02807C_HTILE_RD_POLICY(meta_read_policy
));
405 radeon_set_context_reg(cs
, R_028410_CB_RMI_GL2_CACHE_CONTROL
,
406 S_028410_CMASK_WR_POLICY(meta_write_policy
) |
407 S_028410_FMASK_WR_POLICY(meta_write_policy
) |
408 S_028410_DCC_WR_POLICY(meta_write_policy
) |
409 S_028410_COLOR_WR_POLICY(V_028410_CACHE_STREAM_WR
) |
410 S_028410_CMASK_RD_POLICY(meta_read_policy
) |
411 S_028410_FMASK_RD_POLICY(meta_read_policy
) |
412 S_028410_DCC_RD_POLICY(meta_read_policy
) |
413 S_028410_COLOR_RD_POLICY(V_028410_CACHE_NOA_RD
));
414 radeon_set_context_reg(cs
, R_028428_CB_COVERAGE_OUT_CONTROL
, 0);
416 radeon_set_sh_reg(cs
, R_00B0C0_SPI_SHADER_REQ_CTRL_PS
,
417 S_00B0C0_SOFT_GROUPING_EN(1) |
418 S_00B0C0_NUMBER_OF_REQUESTS_PER_CU(4 - 1));
419 radeon_set_sh_reg(cs
, R_00B1C0_SPI_SHADER_REQ_CTRL_VS
, 0);
421 if (physical_device
->rad_info
.chip_class
>= GFX10_3
) {
422 radeon_set_context_reg(cs
, R_028750_SX_PS_DOWNCONVERT_CONTROL_GFX103
, 0xff);
423 radeon_set_context_reg(cs
, 0x28848, 1 << 9); /* This fixes sample shading. */
426 if (physical_device
->rad_info
.chip_class
== GFX10
) {
427 /* SQ_NON_EVENT must be emitted before GE_PC_ALLOC is written. */
428 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
429 radeon_emit(cs
, EVENT_TYPE(V_028A90_SQ_NON_EVENT
) | EVENT_INDEX(0));
432 /* TODO: For culling, replace 128 with 256. */
433 radeon_set_uconfig_reg(cs
, R_030980_GE_PC_ALLOC
,
434 S_030980_OVERSUB_EN(physical_device
->rad_info
.use_late_alloc
) |
435 S_030980_NUM_PC_LINES(128 * physical_device
->rad_info
.max_se
- 1));
438 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
439 radeon_set_context_reg(cs
, R_028B50_VGT_TESS_DISTRIBUTION
,
440 S_028B50_ACCUM_ISOLINE(40) |
441 S_028B50_ACCUM_TRI(30) |
442 S_028B50_ACCUM_QUAD(24) |
443 S_028B50_DONUT_SPLIT(24) |
444 S_028B50_TRAP_SPLIT(6));
445 } else if (physical_device
->rad_info
.chip_class
>= GFX8
) {
446 uint32_t vgt_tess_distribution
;
448 vgt_tess_distribution
= S_028B50_ACCUM_ISOLINE(32) |
449 S_028B50_ACCUM_TRI(11) |
450 S_028B50_ACCUM_QUAD(11) |
451 S_028B50_DONUT_SPLIT(16);
453 if (physical_device
->rad_info
.family
== CHIP_FIJI
||
454 physical_device
->rad_info
.family
>= CHIP_POLARIS10
)
455 vgt_tess_distribution
|= S_028B50_TRAP_SPLIT(3);
457 radeon_set_context_reg(cs
, R_028B50_VGT_TESS_DISTRIBUTION
,
458 vgt_tess_distribution
);
459 } else if (!has_clear_state
) {
460 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 14);
461 radeon_set_context_reg(cs
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 16);
464 if (device
->border_color_data
.bo
) {
465 uint64_t border_color_va
= radv_buffer_get_va(device
->border_color_data
.bo
);
467 radeon_set_context_reg(cs
, R_028080_TA_BC_BASE_ADDR
, border_color_va
>> 8);
468 if (physical_device
->rad_info
.chip_class
>= GFX7
) {
469 radeon_set_context_reg(cs
, R_028084_TA_BC_BASE_ADDR_HI
,
470 S_028084_ADDRESS(border_color_va
>> 40));
474 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
475 radeon_set_context_reg(cs
, R_028C48_PA_SC_BINNER_CNTL_1
,
476 S_028C48_MAX_ALLOC_COUNT(physical_device
->rad_info
.pbb_max_alloc_count
- 1) |
477 S_028C48_MAX_PRIM_PER_BATCH(1023));
478 radeon_set_context_reg(cs
, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL
,
479 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
480 radeon_set_uconfig_reg(cs
, R_030968_VGT_INSTANCE_BASE_ID
, 0);
483 unsigned tmp
= (unsigned)(1.0 * 8.0);
484 radeon_set_context_reg_seq(cs
, R_028A00_PA_SU_POINT_SIZE
, 1);
485 radeon_emit(cs
, S_028A00_HEIGHT(tmp
) | S_028A00_WIDTH(tmp
));
486 radeon_set_context_reg_seq(cs
, R_028A04_PA_SU_POINT_MINMAX
, 1);
487 radeon_emit(cs
, S_028A04_MIN_SIZE(radv_pack_float_12p4(0)) |
488 S_028A04_MAX_SIZE(radv_pack_float_12p4(8191.875/2)));
490 if (!has_clear_state
) {
491 radeon_set_context_reg(cs
, R_028004_DB_COUNT_CONTROL
,
492 S_028004_ZPASS_INCREMENT_DISABLE(1));
495 /* Enable the Polaris small primitive filter control.
496 * XXX: There is possibly an issue when MSAA is off (see RadeonSI
497 * has_msaa_sample_loc_bug). But this doesn't seem to regress anything,
498 * and AMDVLK doesn't have a workaround as well.
500 if (physical_device
->rad_info
.family
>= CHIP_POLARIS10
) {
501 unsigned small_prim_filter_cntl
=
502 S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
503 /* Workaround for a hw line bug. */
504 S_028830_LINE_FILTER_DISABLE(physical_device
->rad_info
.family
<= CHIP_POLARIS12
);
506 radeon_set_context_reg(cs
, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL
,
507 small_prim_filter_cntl
);
510 radeon_set_context_reg(cs
, R_0286D4_SPI_INTERP_CONTROL_0
,
511 S_0286D4_FLAT_SHADE_ENA(1) |
512 S_0286D4_PNT_SPRITE_ENA(1) |
513 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S
) |
514 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T
) |
515 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0
) |
516 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1
) |
517 S_0286D4_PNT_SPRITE_TOP_1(0)); /* vulkan is top to bottom - 1.0 at bottom */
519 radeon_set_context_reg(cs
, R_028BE4_PA_SU_VTX_CNTL
,
520 S_028BE4_PIX_CENTER(1) |
521 S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN
) |
522 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH
));
524 radeon_set_context_reg(cs
, R_028818_PA_CL_VTE_CNTL
,
525 S_028818_VTX_W0_FMT(1) |
526 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
527 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
528 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
530 si_emit_compute(physical_device
, cs
);
534 cik_create_gfx_config(struct radv_device
*device
)
536 struct radeon_cmdbuf
*cs
= device
->ws
->cs_create(device
->ws
, RING_GFX
);
540 si_emit_graphics(device
, cs
);
542 while (cs
->cdw
& 7) {
543 if (device
->physical_device
->rad_info
.gfx_ib_pad_with_type2
)
544 radeon_emit(cs
, PKT2_NOP_PAD
);
546 radeon_emit(cs
, PKT3_NOP_PAD
);
549 device
->gfx_init
= device
->ws
->buffer_create(device
->ws
,
552 RADEON_FLAG_CPU_ACCESS
|
553 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
554 RADEON_FLAG_READ_ONLY
|
556 RADV_BO_PRIORITY_CS
);
557 if (!device
->gfx_init
)
560 void *map
= device
->ws
->buffer_map(device
->gfx_init
);
562 device
->ws
->buffer_destroy(device
->gfx_init
);
563 device
->gfx_init
= NULL
;
566 memcpy(map
, cs
->buf
, cs
->cdw
* 4);
568 device
->ws
->buffer_unmap(device
->gfx_init
);
569 device
->gfx_init_size_dw
= cs
->cdw
;
571 device
->ws
->cs_destroy(cs
);
575 get_viewport_xform(const VkViewport
*viewport
,
576 float scale
[3], float translate
[3])
578 float x
= viewport
->x
;
579 float y
= viewport
->y
;
580 float half_width
= 0.5f
* viewport
->width
;
581 float half_height
= 0.5f
* viewport
->height
;
582 double n
= viewport
->minDepth
;
583 double f
= viewport
->maxDepth
;
585 scale
[0] = half_width
;
586 translate
[0] = half_width
+ x
;
587 scale
[1] = half_height
;
588 translate
[1] = half_height
+ y
;
595 si_write_viewport(struct radeon_cmdbuf
*cs
, int first_vp
,
596 int count
, const VkViewport
*viewports
)
601 radeon_set_context_reg_seq(cs
, R_02843C_PA_CL_VPORT_XSCALE
+
602 first_vp
* 4 * 6, count
* 6);
604 for (i
= 0; i
< count
; i
++) {
605 float scale
[3], translate
[3];
608 get_viewport_xform(&viewports
[i
], scale
, translate
);
609 radeon_emit(cs
, fui(scale
[0]));
610 radeon_emit(cs
, fui(translate
[0]));
611 radeon_emit(cs
, fui(scale
[1]));
612 radeon_emit(cs
, fui(translate
[1]));
613 radeon_emit(cs
, fui(scale
[2]));
614 radeon_emit(cs
, fui(translate
[2]));
617 radeon_set_context_reg_seq(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+
618 first_vp
* 4 * 2, count
* 2);
619 for (i
= 0; i
< count
; i
++) {
620 float zmin
= MIN2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
621 float zmax
= MAX2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
622 radeon_emit(cs
, fui(zmin
));
623 radeon_emit(cs
, fui(zmax
));
627 static VkRect2D
si_scissor_from_viewport(const VkViewport
*viewport
)
629 float scale
[3], translate
[3];
632 get_viewport_xform(viewport
, scale
, translate
);
634 rect
.offset
.x
= translate
[0] - fabsf(scale
[0]);
635 rect
.offset
.y
= translate
[1] - fabsf(scale
[1]);
636 rect
.extent
.width
= ceilf(translate
[0] + fabsf(scale
[0])) - rect
.offset
.x
;
637 rect
.extent
.height
= ceilf(translate
[1] + fabsf(scale
[1])) - rect
.offset
.y
;
642 static VkRect2D
si_intersect_scissor(const VkRect2D
*a
, const VkRect2D
*b
) {
644 ret
.offset
.x
= MAX2(a
->offset
.x
, b
->offset
.x
);
645 ret
.offset
.y
= MAX2(a
->offset
.y
, b
->offset
.y
);
646 ret
.extent
.width
= MIN2(a
->offset
.x
+ a
->extent
.width
,
647 b
->offset
.x
+ b
->extent
.width
) - ret
.offset
.x
;
648 ret
.extent
.height
= MIN2(a
->offset
.y
+ a
->extent
.height
,
649 b
->offset
.y
+ b
->extent
.height
) - ret
.offset
.y
;
654 si_write_scissors(struct radeon_cmdbuf
*cs
, int first
,
655 int count
, const VkRect2D
*scissors
,
656 const VkViewport
*viewports
, bool can_use_guardband
)
659 float scale
[3], translate
[3], guardband_x
= INFINITY
, guardband_y
= INFINITY
;
660 const float max_range
= 32767.0f
;
664 radeon_set_context_reg_seq(cs
, R_028250_PA_SC_VPORT_SCISSOR_0_TL
+ first
* 4 * 2, count
* 2);
665 for (i
= 0; i
< count
; i
++) {
666 VkRect2D viewport_scissor
= si_scissor_from_viewport(viewports
+ i
);
667 VkRect2D scissor
= si_intersect_scissor(&scissors
[i
], &viewport_scissor
);
669 get_viewport_xform(viewports
+ i
, scale
, translate
);
670 scale
[0] = fabsf(scale
[0]);
671 scale
[1] = fabsf(scale
[1]);
678 guardband_x
= MIN2(guardband_x
, (max_range
- fabsf(translate
[0])) / scale
[0]);
679 guardband_y
= MIN2(guardband_y
, (max_range
- fabsf(translate
[1])) / scale
[1]);
681 radeon_emit(cs
, S_028250_TL_X(scissor
.offset
.x
) |
682 S_028250_TL_Y(scissor
.offset
.y
) |
683 S_028250_WINDOW_OFFSET_DISABLE(1));
684 radeon_emit(cs
, S_028254_BR_X(scissor
.offset
.x
+ scissor
.extent
.width
) |
685 S_028254_BR_Y(scissor
.offset
.y
+ scissor
.extent
.height
));
687 if (!can_use_guardband
) {
692 radeon_set_context_reg_seq(cs
, R_028BE8_PA_CL_GB_VERT_CLIP_ADJ
, 4);
693 radeon_emit(cs
, fui(guardband_y
));
694 radeon_emit(cs
, fui(1.0));
695 radeon_emit(cs
, fui(guardband_x
));
696 radeon_emit(cs
, fui(1.0));
699 static inline unsigned
700 radv_prims_for_vertices(struct radv_prim_vertex_count
*info
, unsigned num
)
711 return 1 + ((num
- info
->min
) / info
->incr
);
714 static const struct radv_prim_vertex_count prim_size_table
[] = {
715 [V_008958_DI_PT_NONE
] = {0, 0},
716 [V_008958_DI_PT_POINTLIST
] = {1, 1},
717 [V_008958_DI_PT_LINELIST
] = {2, 2},
718 [V_008958_DI_PT_LINESTRIP
] = {2, 1},
719 [V_008958_DI_PT_TRILIST
] = {3, 3},
720 [V_008958_DI_PT_TRIFAN
] = {3, 1},
721 [V_008958_DI_PT_TRISTRIP
] = {3, 1},
722 [V_008958_DI_PT_LINELIST_ADJ
] = {4, 4},
723 [V_008958_DI_PT_LINESTRIP_ADJ
] = {4, 1},
724 [V_008958_DI_PT_TRILIST_ADJ
] = {6, 6},
725 [V_008958_DI_PT_TRISTRIP_ADJ
] = {6, 2},
726 [V_008958_DI_PT_RECTLIST
] = {3, 3},
727 [V_008958_DI_PT_LINELOOP
] = {2, 1},
728 [V_008958_DI_PT_POLYGON
] = {3, 1},
729 [V_008958_DI_PT_2D_TRI_STRIP
] = {0, 0},
733 si_get_ia_multi_vgt_param(struct radv_cmd_buffer
*cmd_buffer
,
734 bool instanced_draw
, bool indirect_draw
,
735 bool count_from_stream_output
,
736 uint32_t draw_vertex_count
,
739 enum chip_class chip_class
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
;
740 enum radeon_family family
= cmd_buffer
->device
->physical_device
->rad_info
.family
;
741 struct radeon_info
*info
= &cmd_buffer
->device
->physical_device
->rad_info
;
742 const unsigned max_primgroup_in_wave
= 2;
743 /* SWITCH_ON_EOP(0) is always preferable. */
744 bool wd_switch_on_eop
= false;
745 bool ia_switch_on_eop
= false;
746 bool ia_switch_on_eoi
= false;
747 bool partial_vs_wave
= false;
748 bool partial_es_wave
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.partial_es_wave
;
749 bool multi_instances_smaller_than_primgroup
;
750 struct radv_prim_vertex_count prim_vertex_count
= prim_size_table
[topology
];
752 if (radv_pipeline_has_tess(cmd_buffer
->state
.pipeline
)) {
753 if (topology
== V_008958_DI_PT_PATCH
) {
754 prim_vertex_count
.min
= cmd_buffer
->state
.pipeline
->graphics
.tess_patch_control_points
;
755 prim_vertex_count
.incr
= 1;
759 multi_instances_smaller_than_primgroup
= indirect_draw
;
760 if (!multi_instances_smaller_than_primgroup
&& instanced_draw
) {
761 uint32_t num_prims
= radv_prims_for_vertices(&prim_vertex_count
, draw_vertex_count
);
762 if (num_prims
< cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.primgroup_size
)
763 multi_instances_smaller_than_primgroup
= true;
766 ia_switch_on_eoi
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.ia_switch_on_eoi
;
767 partial_vs_wave
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.partial_vs_wave
;
769 if (chip_class
>= GFX7
) {
770 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
771 * 4 shader engines. Set 1 to pass the assertion below.
772 * The other cases are hardware requirements. */
773 if (cmd_buffer
->device
->physical_device
->rad_info
.max_se
< 4 ||
774 topology
== V_008958_DI_PT_POLYGON
||
775 topology
== V_008958_DI_PT_LINELOOP
||
776 topology
== V_008958_DI_PT_TRIFAN
||
777 topology
== V_008958_DI_PT_TRISTRIP_ADJ
||
778 (cmd_buffer
->state
.pipeline
->graphics
.prim_restart_enable
&&
779 (cmd_buffer
->device
->physical_device
->rad_info
.family
< CHIP_POLARIS10
||
780 (topology
!= V_008958_DI_PT_POINTLIST
&&
781 topology
!= V_008958_DI_PT_LINESTRIP
))))
782 wd_switch_on_eop
= true;
784 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
785 * We don't know that for indirect drawing, so treat it as
786 * always problematic. */
787 if (family
== CHIP_HAWAII
&&
788 (instanced_draw
|| indirect_draw
))
789 wd_switch_on_eop
= true;
791 /* Performance recommendation for 4 SE Gfx7-8 parts if
792 * instances are smaller than a primgroup.
793 * Assume indirect draws always use small instances.
794 * This is needed for good VS wave utilization.
796 if (chip_class
<= GFX8
&&
798 multi_instances_smaller_than_primgroup
)
799 wd_switch_on_eop
= true;
801 /* Required on GFX7 and later. */
802 if (info
->max_se
> 2 && !wd_switch_on_eop
)
803 ia_switch_on_eoi
= true;
805 /* Required by Hawaii and, for some special cases, by GFX8. */
806 if (ia_switch_on_eoi
&&
807 (family
== CHIP_HAWAII
||
808 (chip_class
== GFX8
&&
809 /* max primgroup in wave is always 2 - leave this for documentation */
810 (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
) || max_primgroup_in_wave
!= 2))))
811 partial_vs_wave
= true;
813 /* Instancing bug on Bonaire. */
814 if (family
== CHIP_BONAIRE
&& ia_switch_on_eoi
&&
815 (instanced_draw
|| indirect_draw
))
816 partial_vs_wave
= true;
818 /* Hardware requirement when drawing primitives from a stream
821 if (count_from_stream_output
)
822 wd_switch_on_eop
= true;
824 /* If the WD switch is false, the IA switch must be false too. */
825 assert(wd_switch_on_eop
|| !ia_switch_on_eop
);
827 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
828 if (chip_class
<= GFX8
&& ia_switch_on_eoi
)
829 partial_es_wave
= true;
831 if (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
)) {
832 /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
833 * The hw doc says all multi-SE chips are affected, but amdgpu-pro Vulkan
834 * only applies it to Hawaii. Do what amdgpu-pro Vulkan does.
836 if (family
== CHIP_HAWAII
&& ia_switch_on_eoi
) {
837 bool set_vgt_flush
= indirect_draw
;
838 if (!set_vgt_flush
&& instanced_draw
) {
839 uint32_t num_prims
= radv_prims_for_vertices(&prim_vertex_count
, draw_vertex_count
);
841 set_vgt_flush
= true;
844 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_VGT_FLUSH
;
848 /* Workaround for a VGT hang when strip primitive types are used with
851 if (cmd_buffer
->state
.pipeline
->graphics
.prim_restart_enable
&&
852 (topology
== V_008958_DI_PT_LINESTRIP
||
853 topology
== V_008958_DI_PT_TRISTRIP
||
854 topology
== V_008958_DI_PT_LINESTRIP_ADJ
||
855 topology
== V_008958_DI_PT_TRISTRIP_ADJ
)) {
856 partial_vs_wave
= true;
859 return cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.base
|
860 S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop
) |
861 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi
) |
862 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave
) |
863 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave
) |
864 S_028AA8_WD_SWITCH_ON_EOP(chip_class
>= GFX7
? wd_switch_on_eop
: 0);
868 void si_cs_emit_write_event_eop(struct radeon_cmdbuf
*cs
,
869 enum chip_class chip_class
,
871 unsigned event
, unsigned event_flags
,
872 unsigned dst_sel
, unsigned data_sel
,
875 uint64_t gfx9_eop_bug_va
)
877 unsigned op
= EVENT_TYPE(event
) |
878 EVENT_INDEX(event
== V_028A90_CS_DONE
||
879 event
== V_028A90_PS_DONE
? 6 : 5) |
881 unsigned is_gfx8_mec
= is_mec
&& chip_class
< GFX9
;
882 unsigned sel
= EOP_DST_SEL(dst_sel
) |
883 EOP_DATA_SEL(data_sel
);
885 /* Wait for write confirmation before writing data, but don't send
887 if (data_sel
!= EOP_DATA_SEL_DISCARD
)
888 sel
|= EOP_INT_SEL(EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM
);
890 if (chip_class
>= GFX9
|| is_gfx8_mec
) {
891 /* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
892 * counters) must immediately precede every timestamp event to
893 * prevent a GPU hang on GFX9.
895 if (chip_class
== GFX9
&& !is_mec
) {
896 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 2, 0));
897 radeon_emit(cs
, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE
) | EVENT_INDEX(1));
898 radeon_emit(cs
, gfx9_eop_bug_va
);
899 radeon_emit(cs
, gfx9_eop_bug_va
>> 32);
902 radeon_emit(cs
, PKT3(PKT3_RELEASE_MEM
, is_gfx8_mec
? 5 : 6, false));
904 radeon_emit(cs
, sel
);
905 radeon_emit(cs
, va
); /* address lo */
906 radeon_emit(cs
, va
>> 32); /* address hi */
907 radeon_emit(cs
, new_fence
); /* immediate data lo */
908 radeon_emit(cs
, 0); /* immediate data hi */
910 radeon_emit(cs
, 0); /* unused */
912 if (chip_class
== GFX7
||
913 chip_class
== GFX8
) {
914 /* Two EOP events are required to make all engines go idle
915 * (and optional cache flushes executed) before the timestamp
918 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, false));
921 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
922 radeon_emit(cs
, 0); /* immediate data */
923 radeon_emit(cs
, 0); /* unused */
926 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, false));
929 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
930 radeon_emit(cs
, new_fence
); /* immediate data */
931 radeon_emit(cs
, 0); /* unused */
936 radv_cp_wait_mem(struct radeon_cmdbuf
*cs
, uint32_t op
, uint64_t va
,
937 uint32_t ref
, uint32_t mask
)
939 assert(op
== WAIT_REG_MEM_EQUAL
||
940 op
== WAIT_REG_MEM_NOT_EQUAL
||
941 op
== WAIT_REG_MEM_GREATER_OR_EQUAL
);
943 radeon_emit(cs
, PKT3(PKT3_WAIT_REG_MEM
, 5, false));
944 radeon_emit(cs
, op
| WAIT_REG_MEM_MEM_SPACE(1));
946 radeon_emit(cs
, va
>> 32);
947 radeon_emit(cs
, ref
); /* reference value */
948 radeon_emit(cs
, mask
); /* mask */
949 radeon_emit(cs
, 4); /* poll interval */
953 si_emit_acquire_mem(struct radeon_cmdbuf
*cs
,
956 unsigned cp_coher_cntl
)
958 if (is_mec
|| is_gfx9
) {
959 uint32_t hi_val
= is_gfx9
? 0xffffff : 0xff;
960 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 5, false) |
961 PKT3_SHADER_TYPE_S(is_mec
));
962 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
963 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
964 radeon_emit(cs
, hi_val
); /* CP_COHER_SIZE_HI */
965 radeon_emit(cs
, 0); /* CP_COHER_BASE */
966 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
967 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
969 /* ACQUIRE_MEM is only required on a compute ring. */
970 radeon_emit(cs
, PKT3(PKT3_SURFACE_SYNC
, 3, false));
971 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
972 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
973 radeon_emit(cs
, 0); /* CP_COHER_BASE */
974 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
979 gfx10_cs_emit_cache_flush(struct radeon_cmdbuf
*cs
,
980 enum chip_class chip_class
,
984 enum radv_cmd_flush_bits flush_bits
,
985 uint64_t gfx9_eop_bug_va
)
987 uint32_t gcr_cntl
= 0;
988 unsigned cb_db_event
= 0;
990 /* We don't need these. */
991 assert(!(flush_bits
& (RADV_CMD_FLAG_VGT_STREAMOUT_SYNC
)));
993 if (flush_bits
& RADV_CMD_FLAG_INV_ICACHE
)
994 gcr_cntl
|= S_586_GLI_INV(V_586_GLI_ALL
);
995 if (flush_bits
& RADV_CMD_FLAG_INV_SCACHE
) {
996 /* TODO: When writing to the SMEM L1 cache, we need to set SEQ
997 * to FORWARD when both L1 and L2 are written out (WB or INV).
999 gcr_cntl
|= S_586_GL1_INV(1) | S_586_GLK_INV(1);
1001 if (flush_bits
& RADV_CMD_FLAG_INV_VCACHE
)
1002 gcr_cntl
|= S_586_GL1_INV(1) | S_586_GLV_INV(1);
1003 if (flush_bits
& RADV_CMD_FLAG_INV_L2
) {
1004 /* Writeback and invalidate everything in L2. */
1005 gcr_cntl
|= S_586_GL2_INV(1) | S_586_GL2_WB(1) |
1006 S_586_GLM_INV(1) | S_586_GLM_WB(1);
1007 } else if (flush_bits
& RADV_CMD_FLAG_WB_L2
) {
1008 /* Writeback but do not invalidate.
1009 * GLM doesn't support WB alone. If WB is set, INV must be set too.
1011 gcr_cntl
|= S_586_GL2_WB(1) |
1012 S_586_GLM_WB(1) | S_586_GLM_INV(1);
1015 /* TODO: Implement this new flag for GFX9+.
1016 else if (flush_bits & RADV_CMD_FLAG_INV_L2_METADATA)
1017 gcr_cntl |= S_586_GLM_INV(1) | S_586_GLM_WB(1);
1020 if (flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
| RADV_CMD_FLAG_FLUSH_AND_INV_DB
)) {
1021 /* TODO: trigger on RADV_CMD_FLAG_FLUSH_AND_INV_CB_META */
1022 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB
) {
1023 /* Flush CMASK/FMASK/DCC. Will wait for idle later. */
1024 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1025 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) |
1029 /* TODO: trigger on RADV_CMD_FLAG_FLUSH_AND_INV_DB_META ? */
1030 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB
) {
1031 /* Flush HTILE. Will wait for idle later. */
1032 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1033 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) |
1037 /* First flush CB/DB, then L1/L2. */
1038 gcr_cntl
|= S_586_SEQ(V_586_SEQ_FORWARD
);
1040 if ((flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
| RADV_CMD_FLAG_FLUSH_AND_INV_DB
)) ==
1041 (RADV_CMD_FLAG_FLUSH_AND_INV_CB
| RADV_CMD_FLAG_FLUSH_AND_INV_DB
)) {
1042 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
1043 } else if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB
) {
1044 cb_db_event
= V_028A90_FLUSH_AND_INV_CB_DATA_TS
;
1045 } else if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB
) {
1046 cb_db_event
= V_028A90_FLUSH_AND_INV_DB_DATA_TS
;
1051 /* Wait for graphics shaders to go idle if requested. */
1052 if (flush_bits
& RADV_CMD_FLAG_PS_PARTIAL_FLUSH
) {
1053 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1054 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
1055 } else if (flush_bits
& RADV_CMD_FLAG_VS_PARTIAL_FLUSH
) {
1056 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1057 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
1061 if (flush_bits
& RADV_CMD_FLAG_CS_PARTIAL_FLUSH
) {
1062 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1063 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
| EVENT_INDEX(4)));
1067 /* CB/DB flush and invalidate (or possibly just a wait for a
1068 * meta flush) via RELEASE_MEM.
1070 * Combine this with other cache flushes when possible; this
1071 * requires affected shaders to be idle, so do it after the
1072 * CS_PARTIAL_FLUSH before (VS/PS partial flushes are always
1075 /* Get GCR_CNTL fields, because the encoding is different in RELEASE_MEM. */
1076 unsigned glm_wb
= G_586_GLM_WB(gcr_cntl
);
1077 unsigned glm_inv
= G_586_GLM_INV(gcr_cntl
);
1078 unsigned glv_inv
= G_586_GLV_INV(gcr_cntl
);
1079 unsigned gl1_inv
= G_586_GL1_INV(gcr_cntl
);
1080 assert(G_586_GL2_US(gcr_cntl
) == 0);
1081 assert(G_586_GL2_RANGE(gcr_cntl
) == 0);
1082 assert(G_586_GL2_DISCARD(gcr_cntl
) == 0);
1083 unsigned gl2_inv
= G_586_GL2_INV(gcr_cntl
);
1084 unsigned gl2_wb
= G_586_GL2_WB(gcr_cntl
);
1085 unsigned gcr_seq
= G_586_SEQ(gcr_cntl
);
1087 gcr_cntl
&= C_586_GLM_WB
&
1092 C_586_GL2_WB
; /* keep SEQ */
1097 si_cs_emit_write_event_eop(cs
, chip_class
, false, cb_db_event
,
1098 S_490_GLM_WB(glm_wb
) |
1099 S_490_GLM_INV(glm_inv
) |
1100 S_490_GLV_INV(glv_inv
) |
1101 S_490_GL1_INV(gl1_inv
) |
1102 S_490_GL2_INV(gl2_inv
) |
1103 S_490_GL2_WB(gl2_wb
) |
1106 EOP_DATA_SEL_VALUE_32BIT
,
1107 flush_va
, *flush_cnt
,
1110 radv_cp_wait_mem(cs
, WAIT_REG_MEM_EQUAL
, flush_va
,
1111 *flush_cnt
, 0xffffffff);
1114 /* VGT state sync */
1115 if (flush_bits
& RADV_CMD_FLAG_VGT_FLUSH
) {
1116 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1117 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
1120 /* Ignore fields that only modify the behavior of other fields. */
1121 if (gcr_cntl
& C_586_GL1_RANGE
& C_586_GL2_RANGE
& C_586_SEQ
) {
1122 /* Flush caches and wait for the caches to assert idle.
1123 * The cache flush is executed in the ME, but the PFP waits
1126 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 6, 0));
1127 radeon_emit(cs
, 0); /* CP_COHER_CNTL */
1128 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
1129 radeon_emit(cs
, 0xffffff); /* CP_COHER_SIZE_HI */
1130 radeon_emit(cs
, 0); /* CP_COHER_BASE */
1131 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
1132 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
1133 radeon_emit(cs
, gcr_cntl
); /* GCR_CNTL */
1134 } else if ((cb_db_event
||
1135 (flush_bits
& (RADV_CMD_FLAG_VS_PARTIAL_FLUSH
|
1136 RADV_CMD_FLAG_PS_PARTIAL_FLUSH
|
1137 RADV_CMD_FLAG_CS_PARTIAL_FLUSH
)))
1139 /* We need to ensure that PFP waits as well. */
1140 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
1144 if (flush_bits
& RADV_CMD_FLAG_START_PIPELINE_STATS
) {
1145 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1146 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
1148 } else if (flush_bits
& RADV_CMD_FLAG_STOP_PIPELINE_STATS
) {
1149 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1150 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
1156 si_cs_emit_cache_flush(struct radeon_cmdbuf
*cs
,
1157 enum chip_class chip_class
,
1158 uint32_t *flush_cnt
,
1161 enum radv_cmd_flush_bits flush_bits
,
1162 uint64_t gfx9_eop_bug_va
)
1164 unsigned cp_coher_cntl
= 0;
1165 uint32_t flush_cb_db
= flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
1166 RADV_CMD_FLAG_FLUSH_AND_INV_DB
);
1168 if (chip_class
>= GFX10
) {
1169 /* GFX10 cache flush handling is quite different. */
1170 gfx10_cs_emit_cache_flush(cs
, chip_class
, flush_cnt
, flush_va
,
1171 is_mec
, flush_bits
, gfx9_eop_bug_va
);
1175 if (flush_bits
& RADV_CMD_FLAG_INV_ICACHE
)
1176 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
1177 if (flush_bits
& RADV_CMD_FLAG_INV_SCACHE
)
1178 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
1180 if (chip_class
<= GFX8
) {
1181 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB
) {
1182 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
1183 S_0085F0_CB0_DEST_BASE_ENA(1) |
1184 S_0085F0_CB1_DEST_BASE_ENA(1) |
1185 S_0085F0_CB2_DEST_BASE_ENA(1) |
1186 S_0085F0_CB3_DEST_BASE_ENA(1) |
1187 S_0085F0_CB4_DEST_BASE_ENA(1) |
1188 S_0085F0_CB5_DEST_BASE_ENA(1) |
1189 S_0085F0_CB6_DEST_BASE_ENA(1) |
1190 S_0085F0_CB7_DEST_BASE_ENA(1);
1192 /* Necessary for DCC */
1193 if (chip_class
>= GFX8
) {
1194 si_cs_emit_write_event_eop(cs
,
1197 V_028A90_FLUSH_AND_INV_CB_DATA_TS
,
1200 EOP_DATA_SEL_DISCARD
,
1205 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB
) {
1206 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
1207 S_0085F0_DB_DEST_BASE_ENA(1);
1211 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
) {
1212 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1213 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
1216 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
) {
1217 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1218 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
1221 if (flush_bits
& RADV_CMD_FLAG_PS_PARTIAL_FLUSH
) {
1222 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1223 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
1224 } else if (flush_bits
& RADV_CMD_FLAG_VS_PARTIAL_FLUSH
) {
1225 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1226 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
1229 if (flush_bits
& RADV_CMD_FLAG_CS_PARTIAL_FLUSH
) {
1230 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1231 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
1234 if (chip_class
== GFX9
&& flush_cb_db
) {
1235 unsigned cb_db_event
, tc_flags
;
1237 /* Set the CB/DB flush event. */
1238 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
1240 /* These are the only allowed combinations. If you need to
1241 * do multiple operations at once, do them separately.
1242 * All operations that invalidate L2 also seem to invalidate
1243 * metadata. Volatile (VOL) and WC flushes are not listed here.
1245 * TC | TC_WB = writeback & invalidate L2 & L1
1246 * TC | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
1247 * TC_WB | TC_NC = writeback L2 for MTYPE == NC
1248 * TC | TC_NC = invalidate L2 for MTYPE == NC
1249 * TC | TC_MD = writeback & invalidate L2 metadata (DCC, etc.)
1250 * TCL1 = invalidate L1
1252 tc_flags
= EVENT_TC_ACTION_ENA
|
1253 EVENT_TC_MD_ACTION_ENA
;
1255 /* Ideally flush TC together with CB/DB. */
1256 if (flush_bits
& RADV_CMD_FLAG_INV_L2
) {
1257 /* Writeback and invalidate everything in L2 & L1. */
1258 tc_flags
= EVENT_TC_ACTION_ENA
|
1259 EVENT_TC_WB_ACTION_ENA
;
1262 /* Clear the flags. */
1263 flush_bits
&= ~(RADV_CMD_FLAG_INV_L2
|
1264 RADV_CMD_FLAG_WB_L2
|
1265 RADV_CMD_FLAG_INV_VCACHE
);
1270 si_cs_emit_write_event_eop(cs
, chip_class
, false, cb_db_event
, tc_flags
,
1272 EOP_DATA_SEL_VALUE_32BIT
,
1273 flush_va
, *flush_cnt
,
1275 radv_cp_wait_mem(cs
, WAIT_REG_MEM_EQUAL
, flush_va
,
1276 *flush_cnt
, 0xffffffff);
1279 /* VGT state sync */
1280 if (flush_bits
& RADV_CMD_FLAG_VGT_FLUSH
) {
1281 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1282 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
1285 /* VGT streamout state sync */
1286 if (flush_bits
& RADV_CMD_FLAG_VGT_STREAMOUT_SYNC
) {
1287 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1288 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC
) | EVENT_INDEX(0));
1291 /* Make sure ME is idle (it executes most packets) before continuing.
1292 * This prevents read-after-write hazards between PFP and ME.
1294 if ((cp_coher_cntl
||
1295 (flush_bits
& (RADV_CMD_FLAG_CS_PARTIAL_FLUSH
|
1296 RADV_CMD_FLAG_INV_VCACHE
|
1297 RADV_CMD_FLAG_INV_L2
|
1298 RADV_CMD_FLAG_WB_L2
))) &&
1300 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
1304 if ((flush_bits
& RADV_CMD_FLAG_INV_L2
) ||
1305 (chip_class
<= GFX7
&& (flush_bits
& RADV_CMD_FLAG_WB_L2
))) {
1306 si_emit_acquire_mem(cs
, is_mec
, chip_class
== GFX9
,
1308 S_0085F0_TC_ACTION_ENA(1) |
1309 S_0085F0_TCL1_ACTION_ENA(1) |
1310 S_0301F0_TC_WB_ACTION_ENA(chip_class
>= GFX8
));
1313 if(flush_bits
& RADV_CMD_FLAG_WB_L2
) {
1315 * NC = apply to non-coherent MTYPEs
1316 * (i.e. MTYPE <= 1, which is what we use everywhere)
1318 * WB doesn't work without NC.
1320 si_emit_acquire_mem(cs
, is_mec
,
1323 S_0301F0_TC_WB_ACTION_ENA(1) |
1324 S_0301F0_TC_NC_ACTION_ENA(1));
1327 if (flush_bits
& RADV_CMD_FLAG_INV_VCACHE
) {
1328 si_emit_acquire_mem(cs
, is_mec
,
1331 S_0085F0_TCL1_ACTION_ENA(1));
1336 /* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
1337 * Therefore, it should be last. Done in PFP.
1340 si_emit_acquire_mem(cs
, is_mec
, chip_class
== GFX9
, cp_coher_cntl
);
1342 if (flush_bits
& RADV_CMD_FLAG_START_PIPELINE_STATS
) {
1343 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1344 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
1346 } else if (flush_bits
& RADV_CMD_FLAG_STOP_PIPELINE_STATS
) {
1347 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1348 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
1354 si_emit_cache_flush(struct radv_cmd_buffer
*cmd_buffer
)
1356 bool is_compute
= cmd_buffer
->queue_family_index
== RADV_QUEUE_COMPUTE
;
1359 cmd_buffer
->state
.flush_bits
&= ~(RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
1360 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
|
1361 RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
1362 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
|
1363 RADV_CMD_FLAG_PS_PARTIAL_FLUSH
|
1364 RADV_CMD_FLAG_VS_PARTIAL_FLUSH
|
1365 RADV_CMD_FLAG_VGT_FLUSH
|
1366 RADV_CMD_FLAG_START_PIPELINE_STATS
|
1367 RADV_CMD_FLAG_STOP_PIPELINE_STATS
);
1369 if (!cmd_buffer
->state
.flush_bits
)
1372 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 128);
1374 si_cs_emit_cache_flush(cmd_buffer
->cs
,
1375 cmd_buffer
->device
->physical_device
->rad_info
.chip_class
,
1376 &cmd_buffer
->gfx9_fence_idx
,
1377 cmd_buffer
->gfx9_fence_va
,
1378 radv_cmd_buffer_uses_mec(cmd_buffer
),
1379 cmd_buffer
->state
.flush_bits
,
1380 cmd_buffer
->gfx9_eop_bug_va
);
1383 if (unlikely(cmd_buffer
->device
->trace_bo
))
1384 radv_cmd_buffer_trace_emit(cmd_buffer
);
1386 /* Clear the caches that have been flushed to avoid syncing too much
1387 * when there is some pending active queries.
1389 cmd_buffer
->active_query_flush_bits
&= ~cmd_buffer
->state
.flush_bits
;
1391 cmd_buffer
->state
.flush_bits
= 0;
1393 /* If the driver used a compute shader for resetting a query pool, it
1394 * should be finished at this point.
1396 cmd_buffer
->pending_reset_query
= false;
1399 /* sets the CP predication state using a boolean stored at va */
1401 si_emit_set_predication_state(struct radv_cmd_buffer
*cmd_buffer
,
1402 bool draw_visible
, uint64_t va
)
1407 op
= PRED_OP(PREDICATION_OP_BOOL64
);
1409 /* PREDICATION_DRAW_VISIBLE means that if the 32-bit value is
1410 * zero, all rendering commands are discarded. Otherwise, they
1411 * are discarded if the value is non zero.
1413 op
|= draw_visible
? PREDICATION_DRAW_VISIBLE
:
1414 PREDICATION_DRAW_NOT_VISIBLE
;
1416 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1417 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_SET_PREDICATION
, 2, 0));
1418 radeon_emit(cmd_buffer
->cs
, op
);
1419 radeon_emit(cmd_buffer
->cs
, va
);
1420 radeon_emit(cmd_buffer
->cs
, va
>> 32);
1422 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_SET_PREDICATION
, 1, 0));
1423 radeon_emit(cmd_buffer
->cs
, va
);
1424 radeon_emit(cmd_buffer
->cs
, op
| ((va
>> 32) & 0xFF));
1428 /* Set this if you want the 3D engine to wait until CP DMA is done.
1429 * It should be set on the last CP DMA packet. */
1430 #define CP_DMA_SYNC (1 << 0)
1432 /* Set this if the source data was used as a destination in a previous CP DMA
1433 * packet. It's for preventing a read-after-write (RAW) hazard between two
1434 * CP DMA packets. */
1435 #define CP_DMA_RAW_WAIT (1 << 1)
1436 #define CP_DMA_USE_L2 (1 << 2)
1437 #define CP_DMA_CLEAR (1 << 3)
1439 /* Alignment for optimal performance. */
1440 #define SI_CPDMA_ALIGNMENT 32
1442 /* The max number of bytes that can be copied per packet. */
1443 static inline unsigned cp_dma_max_byte_count(struct radv_cmd_buffer
*cmd_buffer
)
1445 unsigned max
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
?
1446 S_414_BYTE_COUNT_GFX9(~0u) :
1447 S_414_BYTE_COUNT_GFX6(~0u);
1449 /* make it aligned for optimal performance */
1450 return max
& ~(SI_CPDMA_ALIGNMENT
- 1);
1453 /* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
1454 * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
1457 static void si_emit_cp_dma(struct radv_cmd_buffer
*cmd_buffer
,
1458 uint64_t dst_va
, uint64_t src_va
,
1459 unsigned size
, unsigned flags
)
1461 struct radeon_cmdbuf
*cs
= cmd_buffer
->cs
;
1462 uint32_t header
= 0, command
= 0;
1464 assert(size
<= cp_dma_max_byte_count(cmd_buffer
));
1466 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 9);
1467 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1468 command
|= S_414_BYTE_COUNT_GFX9(size
);
1470 command
|= S_414_BYTE_COUNT_GFX6(size
);
1473 if (flags
& CP_DMA_SYNC
)
1474 header
|= S_411_CP_SYNC(1);
1476 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1477 command
|= S_414_DISABLE_WR_CONFIRM_GFX9(1);
1479 command
|= S_414_DISABLE_WR_CONFIRM_GFX6(1);
1482 if (flags
& CP_DMA_RAW_WAIT
)
1483 command
|= S_414_RAW_WAIT(1);
1485 /* Src and dst flags. */
1486 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1487 !(flags
& CP_DMA_CLEAR
) &&
1489 header
|= S_411_DST_SEL(V_411_NOWHERE
); /* prefetch only */
1490 else if (flags
& CP_DMA_USE_L2
)
1491 header
|= S_411_DST_SEL(V_411_DST_ADDR_TC_L2
);
1493 if (flags
& CP_DMA_CLEAR
)
1494 header
|= S_411_SRC_SEL(V_411_DATA
);
1495 else if (flags
& CP_DMA_USE_L2
)
1496 header
|= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2
);
1498 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
1499 radeon_emit(cs
, PKT3(PKT3_DMA_DATA
, 5, cmd_buffer
->state
.predicating
));
1500 radeon_emit(cs
, header
);
1501 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
1502 radeon_emit(cs
, src_va
>> 32); /* SRC_ADDR_HI [31:0] */
1503 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
1504 radeon_emit(cs
, dst_va
>> 32); /* DST_ADDR_HI [31:0] */
1505 radeon_emit(cs
, command
);
1507 assert(!(flags
& CP_DMA_USE_L2
));
1508 header
|= S_411_SRC_ADDR_HI(src_va
>> 32);
1509 radeon_emit(cs
, PKT3(PKT3_CP_DMA
, 4, cmd_buffer
->state
.predicating
));
1510 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
1511 radeon_emit(cs
, header
); /* SRC_ADDR_HI [15:0] + flags. */
1512 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
1513 radeon_emit(cs
, (dst_va
>> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
1514 radeon_emit(cs
, command
);
1517 /* CP DMA is executed in ME, but index buffers are read by PFP.
1518 * This ensures that ME (CP DMA) is idle before PFP starts fetching
1519 * indices. If we wanted to execute CP DMA in PFP, this packet
1520 * should precede it.
1522 if (flags
& CP_DMA_SYNC
) {
1523 if (cmd_buffer
->queue_family_index
== RADV_QUEUE_GENERAL
) {
1524 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, cmd_buffer
->state
.predicating
));
1528 /* CP will see the sync flag and wait for all DMAs to complete. */
1529 cmd_buffer
->state
.dma_is_busy
= false;
1532 if (unlikely(cmd_buffer
->device
->trace_bo
))
1533 radv_cmd_buffer_trace_emit(cmd_buffer
);
1536 void si_cp_dma_prefetch(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1539 uint64_t aligned_va
= va
& ~(SI_CPDMA_ALIGNMENT
- 1);
1540 uint64_t aligned_size
= ((va
+ size
+ SI_CPDMA_ALIGNMENT
-1) & ~(SI_CPDMA_ALIGNMENT
- 1)) - aligned_va
;
1542 si_emit_cp_dma(cmd_buffer
, aligned_va
, aligned_va
,
1543 aligned_size
, CP_DMA_USE_L2
);
1546 static void si_cp_dma_prepare(struct radv_cmd_buffer
*cmd_buffer
, uint64_t byte_count
,
1547 uint64_t remaining_size
, unsigned *flags
)
1550 /* Flush the caches for the first copy only.
1551 * Also wait for the previous CP DMA operations.
1553 if (cmd_buffer
->state
.flush_bits
) {
1554 si_emit_cache_flush(cmd_buffer
);
1555 *flags
|= CP_DMA_RAW_WAIT
;
1558 /* Do the synchronization after the last dma, so that all data
1559 * is written to memory.
1561 if (byte_count
== remaining_size
)
1562 *flags
|= CP_DMA_SYNC
;
1565 static void si_cp_dma_realign_engine(struct radv_cmd_buffer
*cmd_buffer
, unsigned size
)
1569 unsigned dma_flags
= 0;
1570 unsigned buf_size
= SI_CPDMA_ALIGNMENT
* 2;
1573 assert(size
< SI_CPDMA_ALIGNMENT
);
1575 radv_cmd_buffer_upload_alloc(cmd_buffer
, buf_size
, SI_CPDMA_ALIGNMENT
, &offset
, &ptr
);
1577 va
= radv_buffer_get_va(cmd_buffer
->upload
.upload_bo
);
1580 si_cp_dma_prepare(cmd_buffer
, size
, size
, &dma_flags
);
1582 si_emit_cp_dma(cmd_buffer
, va
, va
+ SI_CPDMA_ALIGNMENT
, size
,
1586 void si_cp_dma_buffer_copy(struct radv_cmd_buffer
*cmd_buffer
,
1587 uint64_t src_va
, uint64_t dest_va
,
1590 uint64_t main_src_va
, main_dest_va
;
1591 uint64_t skipped_size
= 0, realign_size
= 0;
1593 /* Assume that we are not going to sync after the last DMA operation. */
1594 cmd_buffer
->state
.dma_is_busy
= true;
1596 if (cmd_buffer
->device
->physical_device
->rad_info
.family
<= CHIP_CARRIZO
||
1597 cmd_buffer
->device
->physical_device
->rad_info
.family
== CHIP_STONEY
) {
1598 /* If the size is not aligned, we must add a dummy copy at the end
1599 * just to align the internal counter. Otherwise, the DMA engine
1600 * would slow down by an order of magnitude for following copies.
1602 if (size
% SI_CPDMA_ALIGNMENT
)
1603 realign_size
= SI_CPDMA_ALIGNMENT
- (size
% SI_CPDMA_ALIGNMENT
);
1605 /* If the copy begins unaligned, we must start copying from the next
1606 * aligned block and the skipped part should be copied after everything
1607 * else has been copied. Only the src alignment matters, not dst.
1609 if (src_va
% SI_CPDMA_ALIGNMENT
) {
1610 skipped_size
= SI_CPDMA_ALIGNMENT
- (src_va
% SI_CPDMA_ALIGNMENT
);
1611 /* The main part will be skipped if the size is too small. */
1612 skipped_size
= MIN2(skipped_size
, size
);
1613 size
-= skipped_size
;
1616 main_src_va
= src_va
+ skipped_size
;
1617 main_dest_va
= dest_va
+ skipped_size
;
1620 unsigned dma_flags
= 0;
1621 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(cmd_buffer
));
1623 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
1624 /* DMA operations via L2 are coherent and faster.
1625 * TODO: GFX7-GFX9 should also support this but it
1626 * requires tests/benchmarks.
1628 dma_flags
|= CP_DMA_USE_L2
;
1631 si_cp_dma_prepare(cmd_buffer
, byte_count
,
1632 size
+ skipped_size
+ realign_size
,
1635 dma_flags
&= ~CP_DMA_SYNC
;
1637 si_emit_cp_dma(cmd_buffer
, main_dest_va
, main_src_va
,
1638 byte_count
, dma_flags
);
1641 main_src_va
+= byte_count
;
1642 main_dest_va
+= byte_count
;
1646 unsigned dma_flags
= 0;
1648 si_cp_dma_prepare(cmd_buffer
, skipped_size
,
1649 size
+ skipped_size
+ realign_size
,
1652 si_emit_cp_dma(cmd_buffer
, dest_va
, src_va
,
1653 skipped_size
, dma_flags
);
1656 si_cp_dma_realign_engine(cmd_buffer
, realign_size
);
1659 void si_cp_dma_clear_buffer(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1660 uint64_t size
, unsigned value
)
1666 assert(va
% 4 == 0 && size
% 4 == 0);
1668 /* Assume that we are not going to sync after the last DMA operation. */
1669 cmd_buffer
->state
.dma_is_busy
= true;
1672 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(cmd_buffer
));
1673 unsigned dma_flags
= CP_DMA_CLEAR
;
1675 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
1676 /* DMA operations via L2 are coherent and faster.
1677 * TODO: GFX7-GFX9 should also support this but it
1678 * requires tests/benchmarks.
1680 dma_flags
|= CP_DMA_USE_L2
;
1683 si_cp_dma_prepare(cmd_buffer
, byte_count
, size
, &dma_flags
);
1685 /* Emit the clear packet. */
1686 si_emit_cp_dma(cmd_buffer
, va
, value
, byte_count
,
1694 void si_cp_dma_wait_for_idle(struct radv_cmd_buffer
*cmd_buffer
)
1696 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
< GFX7
)
1699 if (!cmd_buffer
->state
.dma_is_busy
)
1702 /* Issue a dummy DMA that copies zero bytes.
1704 * The DMA engine will see that there's no work to do and skip this
1705 * DMA request, however, the CP will see the sync flag and still wait
1706 * for all DMAs to complete.
1708 si_emit_cp_dma(cmd_buffer
, 0, 0, 0, CP_DMA_SYNC
);
1710 cmd_buffer
->state
.dma_is_busy
= false;
1713 /* For MSAA sample positions. */
1714 #define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y) \
1715 ((((unsigned)(s0x) & 0xf) << 0) | (((unsigned)(s0y) & 0xf) << 4) | \
1716 (((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) | \
1717 (((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) | \
1718 (((unsigned)(s3x) & 0xf) << 24) | (((unsigned)(s3y) & 0xf) << 28))
1720 /* For obtaining location coordinates from registers */
1721 #define SEXT4(x) ((int)((x) | ((x) & 0x8 ? 0xfffffff0 : 0)))
1722 #define GET_SFIELD(reg, index) SEXT4(((reg) >> ((index) * 4)) & 0xf)
1723 #define GET_SX(reg, index) GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2)
1724 #define GET_SY(reg, index) GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2 + 1)
1727 static const uint32_t sample_locs_1x
=
1728 FILL_SREG(0, 0, 0, 0, 0, 0, 0, 0);
1729 static const unsigned max_dist_1x
= 0;
1730 static const uint64_t centroid_priority_1x
= 0x0000000000000000ull
;
1733 static const uint32_t sample_locs_2x
=
1734 FILL_SREG(4,4, -4, -4, 0, 0, 0, 0);
1735 static const unsigned max_dist_2x
= 4;
1736 static const uint64_t centroid_priority_2x
= 0x1010101010101010ull
;
1739 static const uint32_t sample_locs_4x
=
1740 FILL_SREG(-2,-6, 6, -2, -6, 2, 2, 6);
1741 static const unsigned max_dist_4x
= 6;
1742 static const uint64_t centroid_priority_4x
= 0x3210321032103210ull
;
1745 static const uint32_t sample_locs_8x
[] = {
1746 FILL_SREG( 1,-3, -1, 3, 5, 1, -3,-5),
1747 FILL_SREG(-5, 5, -7,-1, 3, 7, 7,-7),
1748 /* The following are unused by hardware, but we emit them to IBs
1749 * instead of multiple SET_CONTEXT_REG packets. */
1753 static const unsigned max_dist_8x
= 7;
1754 static const uint64_t centroid_priority_8x
= 0x7654321076543210ull
;
1756 unsigned radv_get_default_max_sample_dist(int log_samples
)
1758 unsigned max_dist
[] = {
1764 return max_dist
[log_samples
];
1767 void radv_emit_default_sample_locations(struct radeon_cmdbuf
*cs
, int nr_samples
)
1769 switch (nr_samples
) {
1772 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1773 radeon_emit(cs
, (uint32_t)centroid_priority_1x
);
1774 radeon_emit(cs
, centroid_priority_1x
>> 32);
1775 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, sample_locs_1x
);
1776 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, sample_locs_1x
);
1777 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, sample_locs_1x
);
1778 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, sample_locs_1x
);
1781 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1782 radeon_emit(cs
, (uint32_t)centroid_priority_2x
);
1783 radeon_emit(cs
, centroid_priority_2x
>> 32);
1784 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, sample_locs_2x
);
1785 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, sample_locs_2x
);
1786 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, sample_locs_2x
);
1787 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, sample_locs_2x
);
1790 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1791 radeon_emit(cs
, (uint32_t)centroid_priority_4x
);
1792 radeon_emit(cs
, centroid_priority_4x
>> 32);
1793 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, sample_locs_4x
);
1794 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, sample_locs_4x
);
1795 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, sample_locs_4x
);
1796 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, sample_locs_4x
);
1799 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1800 radeon_emit(cs
, (uint32_t)centroid_priority_8x
);
1801 radeon_emit(cs
, centroid_priority_8x
>> 32);
1802 radeon_set_context_reg_seq(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 14);
1803 radeon_emit_array(cs
, sample_locs_8x
, 4);
1804 radeon_emit_array(cs
, sample_locs_8x
, 4);
1805 radeon_emit_array(cs
, sample_locs_8x
, 4);
1806 radeon_emit_array(cs
, sample_locs_8x
, 2);
1811 static void radv_get_sample_position(struct radv_device
*device
,
1812 unsigned sample_count
,
1813 unsigned sample_index
, float *out_value
)
1815 const uint32_t *sample_locs
;
1817 switch (sample_count
) {
1820 sample_locs
= &sample_locs_1x
;
1823 sample_locs
= &sample_locs_2x
;
1826 sample_locs
= &sample_locs_4x
;
1829 sample_locs
= sample_locs_8x
;
1833 out_value
[0] = (GET_SX(sample_locs
, sample_index
) + 8) / 16.0f
;
1834 out_value
[1] = (GET_SY(sample_locs
, sample_index
) + 8) / 16.0f
;
1837 void radv_device_init_msaa(struct radv_device
*device
)
1841 radv_get_sample_position(device
, 1, 0, device
->sample_locations_1x
[0]);
1843 for (i
= 0; i
< 2; i
++)
1844 radv_get_sample_position(device
, 2, i
, device
->sample_locations_2x
[i
]);
1845 for (i
= 0; i
< 4; i
++)
1846 radv_get_sample_position(device
, 4, i
, device
->sample_locations_4x
[i
]);
1847 for (i
= 0; i
< 8; i
++)
1848 radv_get_sample_position(device
, 8, i
, device
->sample_locations_8x
[i
]);