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_physical_device
*physical_device
,
159 struct radeon_cmdbuf
*cs
)
161 bool has_clear_state
= physical_device
->rad_info
.has_clear_state
;
164 radeon_emit(cs
, PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
165 radeon_emit(cs
, CONTEXT_CONTROL_LOAD_ENABLE(1));
166 radeon_emit(cs
, CONTEXT_CONTROL_SHADOW_ENABLE(1));
168 if (has_clear_state
) {
169 radeon_emit(cs
, PKT3(PKT3_CLEAR_STATE
, 0, 0));
173 if (physical_device
->rad_info
.chip_class
<= GFX8
)
174 si_set_raster_config(physical_device
, cs
);
176 radeon_set_context_reg(cs
, R_028A18_VGT_HOS_MAX_TESS_LEVEL
, fui(64));
177 if (!has_clear_state
)
178 radeon_set_context_reg(cs
, R_028A1C_VGT_HOS_MIN_TESS_LEVEL
, fui(0));
180 /* FIXME calculate these values somehow ??? */
181 if (physical_device
->rad_info
.chip_class
<= GFX8
) {
182 radeon_set_context_reg(cs
, R_028A54_VGT_GS_PER_ES
, SI_GS_PER_ES
);
183 radeon_set_context_reg(cs
, R_028A58_VGT_ES_PER_GS
, 0x40);
186 if (!has_clear_state
) {
187 radeon_set_context_reg(cs
, R_028A5C_VGT_GS_PER_VS
, 0x2);
188 radeon_set_context_reg(cs
, R_028A8C_VGT_PRIMITIVEID_RESET
, 0x0);
189 radeon_set_context_reg(cs
, R_028B98_VGT_STRMOUT_BUFFER_CONFIG
, 0x0);
192 if (physical_device
->rad_info
.chip_class
<= GFX9
)
193 radeon_set_context_reg(cs
, R_028AA0_VGT_INSTANCE_STEP_RATE_0
, 1);
194 if (!has_clear_state
)
195 radeon_set_context_reg(cs
, R_028AB8_VGT_VTX_CNT_EN
, 0x0);
196 if (physical_device
->rad_info
.chip_class
< GFX7
)
197 radeon_set_config_reg(cs
, R_008A14_PA_CL_ENHANCE
, S_008A14_NUM_CLIP_SEQ(3) |
198 S_008A14_CLIP_VTX_REORDER_ENA(1));
200 if (!has_clear_state
)
201 radeon_set_context_reg(cs
, R_02882C_PA_SU_PRIM_FILTER_CNTL
, 0);
203 /* CLEAR_STATE doesn't clear these correctly on certain generations.
204 * I don't know why. Deduced by trial and error.
206 if (physical_device
->rad_info
.chip_class
<= GFX7
|| !has_clear_state
) {
207 radeon_set_context_reg(cs
, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET
, 0);
208 radeon_set_context_reg(cs
, R_028204_PA_SC_WINDOW_SCISSOR_TL
,
209 S_028204_WINDOW_OFFSET_DISABLE(1));
210 radeon_set_context_reg(cs
, R_028240_PA_SC_GENERIC_SCISSOR_TL
,
211 S_028240_WINDOW_OFFSET_DISABLE(1));
212 radeon_set_context_reg(cs
, R_028244_PA_SC_GENERIC_SCISSOR_BR
,
213 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
214 radeon_set_context_reg(cs
, R_028030_PA_SC_SCREEN_SCISSOR_TL
, 0);
215 radeon_set_context_reg(cs
, R_028034_PA_SC_SCREEN_SCISSOR_BR
,
216 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
219 if (!has_clear_state
) {
220 for (i
= 0; i
< 16; i
++) {
221 radeon_set_context_reg(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+ i
*8, 0);
222 radeon_set_context_reg(cs
, R_0282D4_PA_SC_VPORT_ZMAX_0
+ i
*8, fui(1.0));
226 if (!has_clear_state
) {
227 radeon_set_context_reg(cs
, R_02820C_PA_SC_CLIPRECT_RULE
, 0xFFFF);
228 radeon_set_context_reg(cs
, R_028230_PA_SC_EDGERULE
, 0xAAAAAAAA);
229 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on GFX6 */
230 radeon_set_context_reg(cs
, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET
, 0);
231 radeon_set_context_reg(cs
, R_028820_PA_CL_NANINF_CNTL
, 0);
232 radeon_set_context_reg(cs
, R_028AC0_DB_SRESULTS_COMPARE_STATE0
, 0x0);
233 radeon_set_context_reg(cs
, R_028AC4_DB_SRESULTS_COMPARE_STATE1
, 0x0);
234 radeon_set_context_reg(cs
, R_028AC8_DB_PRELOAD_CONTROL
, 0x0);
237 radeon_set_context_reg(cs
, R_02800C_DB_RENDER_OVERRIDE
,
238 S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE
) |
239 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE
));
241 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
242 radeon_set_context_reg(cs
, R_028A98_VGT_DRAW_PAYLOAD_CNTL
, 0);
243 radeon_set_uconfig_reg(cs
, R_030964_GE_MAX_VTX_INDX
, ~0);
244 radeon_set_uconfig_reg(cs
, R_030924_GE_MIN_VTX_INDX
, 0);
245 radeon_set_uconfig_reg(cs
, R_030928_GE_INDX_OFFSET
, 0);
246 radeon_set_uconfig_reg(cs
, R_03097C_GE_STEREO_CNTL
, 0);
247 radeon_set_uconfig_reg(cs
, R_030988_GE_USER_VGPR_EN
, 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_idx(physical_device
, cs
, R_00B404_SPI_SHADER_PGM_RSRC4_HS
,
267 3, S_00B404_CU_EN(0xffff));
268 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B104_SPI_SHADER_PGM_RSRC4_VS
,
269 3, S_00B104_CU_EN(0xffff));
270 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B004_SPI_SHADER_PGM_RSRC4_PS
,
271 3, S_00B004_CU_EN(0xffff));
274 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
275 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
,
276 3, S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
278 radeon_set_sh_reg(cs
, R_00B51C_SPI_SHADER_PGM_RSRC3_LS
,
279 S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
280 radeon_set_sh_reg(cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
,
281 S_00B41C_WAVE_LIMIT(0x3F));
282 radeon_set_sh_reg(cs
, R_00B31C_SPI_SHADER_PGM_RSRC3_ES
,
283 S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
284 /* If this is 0, Bonaire can hang even if GS isn't being used.
285 * Other chips are unaffected. These are suboptimal values,
286 * but we don't use on-chip GS.
288 radeon_set_context_reg(cs
, R_028A44_VGT_GS_ONCHIP_CNTL
,
289 S_028A44_ES_VERTS_PER_SUBGRP(64) |
290 S_028A44_GS_PRIMS_PER_SUBGRP(4));
293 /* Compute LATE_ALLOC_VS.LIMIT. */
294 unsigned num_cu_per_sh
= physical_device
->rad_info
.num_good_cu_per_sh
;
295 unsigned late_alloc_wave64
= 0; /* The limit is per SH. */
296 unsigned late_alloc_wave64_gs
= 0;
297 unsigned cu_mask_vs
= 0xffff;
298 unsigned cu_mask_gs
= 0xffff;
300 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
301 /* For Wave32, the hw will launch twice the number of late
302 * alloc waves, so 1 == 2x wave32.
304 if (!physical_device
->rad_info
.use_late_alloc
) {
305 late_alloc_wave64
= 0;
306 } else if (num_cu_per_sh
<= 6) {
307 late_alloc_wave64
= num_cu_per_sh
- 2;
309 late_alloc_wave64
= (num_cu_per_sh
- 2) * 4;
311 /* CU2 & CU3 disabled because of the dual CU design */
313 cu_mask_gs
= 0xfff3; /* NGG only */
316 late_alloc_wave64_gs
= late_alloc_wave64
;
318 /* Don't use late alloc for NGG on Navi14 due to a hw
319 * bug. If NGG is never used, enable all CUs.
321 if (!physical_device
->use_ngg
||
322 physical_device
->rad_info
.family
== CHIP_NAVI14
) {
323 late_alloc_wave64_gs
= 0;
327 if (!physical_device
->rad_info
.use_late_alloc
) {
328 late_alloc_wave64
= 0;
329 } else if (num_cu_per_sh
<= 4) {
330 /* Too few available compute units per SH.
331 * Disallowing VS to run on one CU could hurt
332 * us more than late VS allocation would help.
334 * 2 is the highest safe number that allows us
335 * to keep all CUs enabled.
337 late_alloc_wave64
= 2;
339 /* This is a good initial value, allowing 1
340 * late_alloc wave per SIMD on num_cu - 2.
342 late_alloc_wave64
= (num_cu_per_sh
- 2) * 4;
345 if (late_alloc_wave64
> 2)
346 cu_mask_vs
= 0xfffe; /* 1 CU disabled */
349 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
,
350 3, S_00B118_CU_EN(cu_mask_vs
) |
351 S_00B118_WAVE_LIMIT(0x3F));
352 radeon_set_sh_reg(cs
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
,
353 S_00B11C_LIMIT(late_alloc_wave64
));
355 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B21C_SPI_SHADER_PGM_RSRC3_GS
,
356 3, S_00B21C_CU_EN(cu_mask_gs
) | S_00B21C_WAVE_LIMIT(0x3F));
358 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
359 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B204_SPI_SHADER_PGM_RSRC4_GS
,
360 3, S_00B204_CU_EN(0xffff) |
361 S_00B204_SPI_SHADER_LATE_ALLOC_GS_GFX10(late_alloc_wave64_gs
));
364 radeon_set_sh_reg_idx(physical_device
, cs
, R_00B01C_SPI_SHADER_PGM_RSRC3_PS
,
365 3, S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
368 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
369 /* Break up a pixel wave if it contains deallocs for more than
370 * half the parameter cache.
372 * To avoid a deadlock where pixel waves aren't launched
373 * because they're waiting for more pixels while the frontend
374 * is stuck waiting for PC space, the maximum allowed value is
375 * the size of the PC minus the largest possible allocation for
376 * a single primitive shader subgroup.
378 radeon_set_context_reg(cs
, R_028C50_PA_SC_NGG_MODE_CNTL
,
379 S_028C50_MAX_DEALLOCS_IN_WAVE(512));
380 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 14);
382 /* Enable CMASK/FMASK/HTILE/DCC caching in L2 for small chips. */
383 unsigned meta_write_policy
, meta_read_policy
;
385 /* TODO: investigate whether LRU improves performance on other chips too */
386 if (physical_device
->rad_info
.num_render_backends
<= 4) {
387 meta_write_policy
= V_02807C_CACHE_LRU_WR
; /* cache writes */
388 meta_read_policy
= V_02807C_CACHE_LRU_RD
; /* cache reads */
390 meta_write_policy
= V_02807C_CACHE_STREAM_WR
; /* write combine */
391 meta_read_policy
= V_02807C_CACHE_NOA_RD
; /* don't cache reads */
394 radeon_set_context_reg(cs
, R_02807C_DB_RMI_L2_CACHE_CONTROL
,
395 S_02807C_Z_WR_POLICY(V_02807C_CACHE_STREAM_WR
) |
396 S_02807C_S_WR_POLICY(V_02807C_CACHE_STREAM_WR
) |
397 S_02807C_HTILE_WR_POLICY(meta_write_policy
) |
398 S_02807C_ZPCPSD_WR_POLICY(V_02807C_CACHE_STREAM_WR
) |
399 S_02807C_Z_RD_POLICY(V_02807C_CACHE_NOA_RD
) |
400 S_02807C_S_RD_POLICY(V_02807C_CACHE_NOA_RD
) |
401 S_02807C_HTILE_RD_POLICY(meta_read_policy
));
403 radeon_set_context_reg(cs
, R_028410_CB_RMI_GL2_CACHE_CONTROL
,
404 S_028410_CMASK_WR_POLICY(meta_write_policy
) |
405 S_028410_FMASK_WR_POLICY(meta_write_policy
) |
406 S_028410_DCC_WR_POLICY(meta_write_policy
) |
407 S_028410_COLOR_WR_POLICY(V_028410_CACHE_STREAM_WR
) |
408 S_028410_CMASK_RD_POLICY(meta_read_policy
) |
409 S_028410_FMASK_RD_POLICY(meta_read_policy
) |
410 S_028410_DCC_RD_POLICY(meta_read_policy
) |
411 S_028410_COLOR_RD_POLICY(V_028410_CACHE_NOA_RD
));
412 radeon_set_context_reg(cs
, R_028428_CB_COVERAGE_OUT_CONTROL
, 0);
414 radeon_set_sh_reg(cs
, R_00B0C0_SPI_SHADER_REQ_CTRL_PS
,
415 S_00B0C0_SOFT_GROUPING_EN(1) |
416 S_00B0C0_NUMBER_OF_REQUESTS_PER_CU(4 - 1));
417 radeon_set_sh_reg(cs
, R_00B1C0_SPI_SHADER_REQ_CTRL_VS
, 0);
419 if (physical_device
->rad_info
.chip_class
== GFX10
) {
420 /* SQ_NON_EVENT must be emitted before GE_PC_ALLOC is written. */
421 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
422 radeon_emit(cs
, EVENT_TYPE(V_028A90_SQ_NON_EVENT
) | EVENT_INDEX(0));
425 /* TODO: For culling, replace 128 with 256. */
426 radeon_set_uconfig_reg(cs
, R_030980_GE_PC_ALLOC
,
427 S_030980_OVERSUB_EN(physical_device
->rad_info
.use_late_alloc
) |
428 S_030980_NUM_PC_LINES(128 * physical_device
->rad_info
.max_se
- 1));
431 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
432 radeon_set_context_reg(cs
, R_028B50_VGT_TESS_DISTRIBUTION
,
433 S_028B50_ACCUM_ISOLINE(40) |
434 S_028B50_ACCUM_TRI(30) |
435 S_028B50_ACCUM_QUAD(24) |
436 S_028B50_DONUT_SPLIT(24) |
437 S_028B50_TRAP_SPLIT(6));
438 } else if (physical_device
->rad_info
.chip_class
>= GFX8
) {
439 uint32_t vgt_tess_distribution
;
441 vgt_tess_distribution
= S_028B50_ACCUM_ISOLINE(32) |
442 S_028B50_ACCUM_TRI(11) |
443 S_028B50_ACCUM_QUAD(11) |
444 S_028B50_DONUT_SPLIT(16);
446 if (physical_device
->rad_info
.family
== CHIP_FIJI
||
447 physical_device
->rad_info
.family
>= CHIP_POLARIS10
)
448 vgt_tess_distribution
|= S_028B50_TRAP_SPLIT(3);
450 radeon_set_context_reg(cs
, R_028B50_VGT_TESS_DISTRIBUTION
,
451 vgt_tess_distribution
);
452 } else if (!has_clear_state
) {
453 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 14);
454 radeon_set_context_reg(cs
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 16);
457 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
458 radeon_set_context_reg(cs
, R_028C48_PA_SC_BINNER_CNTL_1
,
459 S_028C48_MAX_ALLOC_COUNT(physical_device
->rad_info
.pbb_max_alloc_count
- 1) |
460 S_028C48_MAX_PRIM_PER_BATCH(1023));
461 radeon_set_context_reg(cs
, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL
,
462 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
463 radeon_set_uconfig_reg(cs
, R_030968_VGT_INSTANCE_BASE_ID
, 0);
466 unsigned tmp
= (unsigned)(1.0 * 8.0);
467 radeon_set_context_reg_seq(cs
, R_028A00_PA_SU_POINT_SIZE
, 1);
468 radeon_emit(cs
, S_028A00_HEIGHT(tmp
) | S_028A00_WIDTH(tmp
));
469 radeon_set_context_reg_seq(cs
, R_028A04_PA_SU_POINT_MINMAX
, 1);
470 radeon_emit(cs
, S_028A04_MIN_SIZE(radv_pack_float_12p4(0)) |
471 S_028A04_MAX_SIZE(radv_pack_float_12p4(8192/2)));
473 if (!has_clear_state
) {
474 radeon_set_context_reg(cs
, R_028004_DB_COUNT_CONTROL
,
475 S_028004_ZPASS_INCREMENT_DISABLE(1));
478 /* Enable the Polaris small primitive filter control.
479 * XXX: There is possibly an issue when MSAA is off (see RadeonSI
480 * has_msaa_sample_loc_bug). But this doesn't seem to regress anything,
481 * and AMDVLK doesn't have a workaround as well.
483 if (physical_device
->rad_info
.family
>= CHIP_POLARIS10
) {
484 unsigned small_prim_filter_cntl
=
485 S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
486 /* Workaround for a hw line bug. */
487 S_028830_LINE_FILTER_DISABLE(physical_device
->rad_info
.family
<= CHIP_POLARIS12
);
489 radeon_set_context_reg(cs
, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL
,
490 small_prim_filter_cntl
);
493 si_emit_compute(physical_device
, cs
);
497 cik_create_gfx_config(struct radv_device
*device
)
499 struct radeon_cmdbuf
*cs
= device
->ws
->cs_create(device
->ws
, RING_GFX
);
503 si_emit_graphics(device
->physical_device
, cs
);
505 while (cs
->cdw
& 7) {
506 if (device
->physical_device
->rad_info
.gfx_ib_pad_with_type2
)
507 radeon_emit(cs
, 0x80000000);
509 radeon_emit(cs
, 0xffff1000);
512 device
->gfx_init
= device
->ws
->buffer_create(device
->ws
,
515 RADEON_FLAG_CPU_ACCESS
|
516 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
517 RADEON_FLAG_READ_ONLY
,
518 RADV_BO_PRIORITY_CS
);
519 if (!device
->gfx_init
)
522 void *map
= device
->ws
->buffer_map(device
->gfx_init
);
524 device
->ws
->buffer_destroy(device
->gfx_init
);
525 device
->gfx_init
= NULL
;
528 memcpy(map
, cs
->buf
, cs
->cdw
* 4);
530 device
->ws
->buffer_unmap(device
->gfx_init
);
531 device
->gfx_init_size_dw
= cs
->cdw
;
533 device
->ws
->cs_destroy(cs
);
537 get_viewport_xform(const VkViewport
*viewport
,
538 float scale
[3], float translate
[3])
540 float x
= viewport
->x
;
541 float y
= viewport
->y
;
542 float half_width
= 0.5f
* viewport
->width
;
543 float half_height
= 0.5f
* viewport
->height
;
544 double n
= viewport
->minDepth
;
545 double f
= viewport
->maxDepth
;
547 scale
[0] = half_width
;
548 translate
[0] = half_width
+ x
;
549 scale
[1] = half_height
;
550 translate
[1] = half_height
+ y
;
557 si_write_viewport(struct radeon_cmdbuf
*cs
, int first_vp
,
558 int count
, const VkViewport
*viewports
)
563 radeon_set_context_reg_seq(cs
, R_02843C_PA_CL_VPORT_XSCALE
+
564 first_vp
* 4 * 6, count
* 6);
566 for (i
= 0; i
< count
; i
++) {
567 float scale
[3], translate
[3];
570 get_viewport_xform(&viewports
[i
], scale
, translate
);
571 radeon_emit(cs
, fui(scale
[0]));
572 radeon_emit(cs
, fui(translate
[0]));
573 radeon_emit(cs
, fui(scale
[1]));
574 radeon_emit(cs
, fui(translate
[1]));
575 radeon_emit(cs
, fui(scale
[2]));
576 radeon_emit(cs
, fui(translate
[2]));
579 radeon_set_context_reg_seq(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+
580 first_vp
* 4 * 2, count
* 2);
581 for (i
= 0; i
< count
; i
++) {
582 float zmin
= MIN2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
583 float zmax
= MAX2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
584 radeon_emit(cs
, fui(zmin
));
585 radeon_emit(cs
, fui(zmax
));
589 static VkRect2D
si_scissor_from_viewport(const VkViewport
*viewport
)
591 float scale
[3], translate
[3];
594 get_viewport_xform(viewport
, scale
, translate
);
596 rect
.offset
.x
= translate
[0] - fabsf(scale
[0]);
597 rect
.offset
.y
= translate
[1] - fabsf(scale
[1]);
598 rect
.extent
.width
= ceilf(translate
[0] + fabsf(scale
[0])) - rect
.offset
.x
;
599 rect
.extent
.height
= ceilf(translate
[1] + fabsf(scale
[1])) - rect
.offset
.y
;
604 static VkRect2D
si_intersect_scissor(const VkRect2D
*a
, const VkRect2D
*b
) {
606 ret
.offset
.x
= MAX2(a
->offset
.x
, b
->offset
.x
);
607 ret
.offset
.y
= MAX2(a
->offset
.y
, b
->offset
.y
);
608 ret
.extent
.width
= MIN2(a
->offset
.x
+ a
->extent
.width
,
609 b
->offset
.x
+ b
->extent
.width
) - ret
.offset
.x
;
610 ret
.extent
.height
= MIN2(a
->offset
.y
+ a
->extent
.height
,
611 b
->offset
.y
+ b
->extent
.height
) - ret
.offset
.y
;
616 si_write_scissors(struct radeon_cmdbuf
*cs
, int first
,
617 int count
, const VkRect2D
*scissors
,
618 const VkViewport
*viewports
, bool can_use_guardband
)
621 float scale
[3], translate
[3], guardband_x
= INFINITY
, guardband_y
= INFINITY
;
622 const float max_range
= 32767.0f
;
626 radeon_set_context_reg_seq(cs
, R_028250_PA_SC_VPORT_SCISSOR_0_TL
+ first
* 4 * 2, count
* 2);
627 for (i
= 0; i
< count
; i
++) {
628 VkRect2D viewport_scissor
= si_scissor_from_viewport(viewports
+ i
);
629 VkRect2D scissor
= si_intersect_scissor(&scissors
[i
], &viewport_scissor
);
631 get_viewport_xform(viewports
+ i
, scale
, translate
);
632 scale
[0] = fabsf(scale
[0]);
633 scale
[1] = fabsf(scale
[1]);
640 guardband_x
= MIN2(guardband_x
, (max_range
- fabsf(translate
[0])) / scale
[0]);
641 guardband_y
= MIN2(guardband_y
, (max_range
- fabsf(translate
[1])) / scale
[1]);
643 radeon_emit(cs
, S_028250_TL_X(scissor
.offset
.x
) |
644 S_028250_TL_Y(scissor
.offset
.y
) |
645 S_028250_WINDOW_OFFSET_DISABLE(1));
646 radeon_emit(cs
, S_028254_BR_X(scissor
.offset
.x
+ scissor
.extent
.width
) |
647 S_028254_BR_Y(scissor
.offset
.y
+ scissor
.extent
.height
));
649 if (!can_use_guardband
) {
654 radeon_set_context_reg_seq(cs
, R_028BE8_PA_CL_GB_VERT_CLIP_ADJ
, 4);
655 radeon_emit(cs
, fui(guardband_y
));
656 radeon_emit(cs
, fui(1.0));
657 radeon_emit(cs
, fui(guardband_x
));
658 radeon_emit(cs
, fui(1.0));
661 static inline unsigned
662 radv_prims_for_vertices(struct radv_prim_vertex_count
*info
, unsigned num
)
673 return 1 + ((num
- info
->min
) / info
->incr
);
677 si_get_ia_multi_vgt_param(struct radv_cmd_buffer
*cmd_buffer
,
678 bool instanced_draw
, bool indirect_draw
,
679 bool count_from_stream_output
,
680 uint32_t draw_vertex_count
)
682 enum chip_class chip_class
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
;
683 enum radeon_family family
= cmd_buffer
->device
->physical_device
->rad_info
.family
;
684 struct radeon_info
*info
= &cmd_buffer
->device
->physical_device
->rad_info
;
685 const unsigned max_primgroup_in_wave
= 2;
686 /* SWITCH_ON_EOP(0) is always preferable. */
687 bool wd_switch_on_eop
= false;
688 bool ia_switch_on_eop
= false;
689 bool ia_switch_on_eoi
= false;
690 bool partial_vs_wave
= false;
691 bool partial_es_wave
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.partial_es_wave
;
692 bool multi_instances_smaller_than_primgroup
;
694 multi_instances_smaller_than_primgroup
= indirect_draw
;
695 if (!multi_instances_smaller_than_primgroup
&& instanced_draw
) {
696 uint32_t num_prims
= radv_prims_for_vertices(&cmd_buffer
->state
.pipeline
->graphics
.prim_vertex_count
, draw_vertex_count
);
697 if (num_prims
< cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.primgroup_size
)
698 multi_instances_smaller_than_primgroup
= true;
701 ia_switch_on_eoi
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.ia_switch_on_eoi
;
702 partial_vs_wave
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.partial_vs_wave
;
704 if (chip_class
>= GFX7
) {
705 wd_switch_on_eop
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.wd_switch_on_eop
;
707 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
708 * We don't know that for indirect drawing, so treat it as
709 * always problematic. */
710 if (family
== CHIP_HAWAII
&&
711 (instanced_draw
|| indirect_draw
))
712 wd_switch_on_eop
= true;
714 /* Performance recommendation for 4 SE Gfx7-8 parts if
715 * instances are smaller than a primgroup.
716 * Assume indirect draws always use small instances.
717 * This is needed for good VS wave utilization.
719 if (chip_class
<= GFX8
&&
721 multi_instances_smaller_than_primgroup
)
722 wd_switch_on_eop
= true;
724 /* Required on GFX7 and later. */
725 if (info
->max_se
> 2 && !wd_switch_on_eop
)
726 ia_switch_on_eoi
= true;
728 /* Required by Hawaii and, for some special cases, by GFX8. */
729 if (ia_switch_on_eoi
&&
730 (family
== CHIP_HAWAII
||
731 (chip_class
== GFX8
&&
732 /* max primgroup in wave is always 2 - leave this for documentation */
733 (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
) || max_primgroup_in_wave
!= 2))))
734 partial_vs_wave
= true;
736 /* Instancing bug on Bonaire. */
737 if (family
== CHIP_BONAIRE
&& ia_switch_on_eoi
&&
738 (instanced_draw
|| indirect_draw
))
739 partial_vs_wave
= true;
741 /* Hardware requirement when drawing primitives from a stream
744 if (count_from_stream_output
)
745 wd_switch_on_eop
= true;
747 /* If the WD switch is false, the IA switch must be false too. */
748 assert(wd_switch_on_eop
|| !ia_switch_on_eop
);
750 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
751 if (chip_class
<= GFX8
&& ia_switch_on_eoi
)
752 partial_es_wave
= true;
754 if (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
)) {
755 /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
756 * The hw doc says all multi-SE chips are affected, but amdgpu-pro Vulkan
757 * only applies it to Hawaii. Do what amdgpu-pro Vulkan does.
759 if (family
== CHIP_HAWAII
&& ia_switch_on_eoi
) {
760 bool set_vgt_flush
= indirect_draw
;
761 if (!set_vgt_flush
&& instanced_draw
) {
762 uint32_t num_prims
= radv_prims_for_vertices(&cmd_buffer
->state
.pipeline
->graphics
.prim_vertex_count
, draw_vertex_count
);
764 set_vgt_flush
= true;
767 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_VGT_FLUSH
;
771 return cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.base
|
772 S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop
) |
773 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi
) |
774 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave
) |
775 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave
) |
776 S_028AA8_WD_SWITCH_ON_EOP(chip_class
>= GFX7
? wd_switch_on_eop
: 0);
780 void si_cs_emit_write_event_eop(struct radeon_cmdbuf
*cs
,
781 enum chip_class chip_class
,
783 unsigned event
, unsigned event_flags
,
784 unsigned dst_sel
, unsigned data_sel
,
787 uint64_t gfx9_eop_bug_va
)
789 unsigned op
= EVENT_TYPE(event
) |
790 EVENT_INDEX(event
== V_028A90_CS_DONE
||
791 event
== V_028A90_PS_DONE
? 6 : 5) |
793 unsigned is_gfx8_mec
= is_mec
&& chip_class
< GFX9
;
794 unsigned sel
= EOP_DST_SEL(dst_sel
) |
795 EOP_DATA_SEL(data_sel
);
797 /* Wait for write confirmation before writing data, but don't send
799 if (data_sel
!= EOP_DATA_SEL_DISCARD
)
800 sel
|= EOP_INT_SEL(EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM
);
802 if (chip_class
>= GFX9
|| is_gfx8_mec
) {
803 /* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
804 * counters) must immediately precede every timestamp event to
805 * prevent a GPU hang on GFX9.
807 if (chip_class
== GFX9
&& !is_mec
) {
808 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 2, 0));
809 radeon_emit(cs
, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE
) | EVENT_INDEX(1));
810 radeon_emit(cs
, gfx9_eop_bug_va
);
811 radeon_emit(cs
, gfx9_eop_bug_va
>> 32);
814 radeon_emit(cs
, PKT3(PKT3_RELEASE_MEM
, is_gfx8_mec
? 5 : 6, false));
816 radeon_emit(cs
, sel
);
817 radeon_emit(cs
, va
); /* address lo */
818 radeon_emit(cs
, va
>> 32); /* address hi */
819 radeon_emit(cs
, new_fence
); /* immediate data lo */
820 radeon_emit(cs
, 0); /* immediate data hi */
822 radeon_emit(cs
, 0); /* unused */
824 if (chip_class
== GFX7
||
825 chip_class
== GFX8
) {
826 /* Two EOP events are required to make all engines go idle
827 * (and optional cache flushes executed) before the timestamp
830 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, false));
833 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
834 radeon_emit(cs
, 0); /* immediate data */
835 radeon_emit(cs
, 0); /* unused */
838 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, false));
841 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
842 radeon_emit(cs
, new_fence
); /* immediate data */
843 radeon_emit(cs
, 0); /* unused */
848 radv_cp_wait_mem(struct radeon_cmdbuf
*cs
, uint32_t op
, uint64_t va
,
849 uint32_t ref
, uint32_t mask
)
851 assert(op
== WAIT_REG_MEM_EQUAL
||
852 op
== WAIT_REG_MEM_NOT_EQUAL
||
853 op
== WAIT_REG_MEM_GREATER_OR_EQUAL
);
855 radeon_emit(cs
, PKT3(PKT3_WAIT_REG_MEM
, 5, false));
856 radeon_emit(cs
, op
| WAIT_REG_MEM_MEM_SPACE(1));
858 radeon_emit(cs
, va
>> 32);
859 radeon_emit(cs
, ref
); /* reference value */
860 radeon_emit(cs
, mask
); /* mask */
861 radeon_emit(cs
, 4); /* poll interval */
865 si_emit_acquire_mem(struct radeon_cmdbuf
*cs
,
868 unsigned cp_coher_cntl
)
870 if (is_mec
|| is_gfx9
) {
871 uint32_t hi_val
= is_gfx9
? 0xffffff : 0xff;
872 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 5, false) |
873 PKT3_SHADER_TYPE_S(is_mec
));
874 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
875 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
876 radeon_emit(cs
, hi_val
); /* CP_COHER_SIZE_HI */
877 radeon_emit(cs
, 0); /* CP_COHER_BASE */
878 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
879 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
881 /* ACQUIRE_MEM is only required on a compute ring. */
882 radeon_emit(cs
, PKT3(PKT3_SURFACE_SYNC
, 3, false));
883 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
884 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
885 radeon_emit(cs
, 0); /* CP_COHER_BASE */
886 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
891 gfx10_cs_emit_cache_flush(struct radeon_cmdbuf
*cs
,
892 enum chip_class chip_class
,
896 enum radv_cmd_flush_bits flush_bits
,
897 uint64_t gfx9_eop_bug_va
)
899 uint32_t gcr_cntl
= 0;
900 unsigned cb_db_event
= 0;
902 /* We don't need these. */
903 assert(!(flush_bits
& (RADV_CMD_FLAG_VGT_STREAMOUT_SYNC
)));
905 if (flush_bits
& RADV_CMD_FLAG_INV_ICACHE
)
906 gcr_cntl
|= S_586_GLI_INV(V_586_GLI_ALL
);
907 if (flush_bits
& RADV_CMD_FLAG_INV_SCACHE
) {
908 /* TODO: When writing to the SMEM L1 cache, we need to set SEQ
909 * to FORWARD when both L1 and L2 are written out (WB or INV).
911 gcr_cntl
|= S_586_GL1_INV(1) | S_586_GLK_INV(1);
913 if (flush_bits
& RADV_CMD_FLAG_INV_VCACHE
)
914 gcr_cntl
|= S_586_GL1_INV(1) | S_586_GLV_INV(1);
915 if (flush_bits
& RADV_CMD_FLAG_INV_L2
) {
916 /* Writeback and invalidate everything in L2. */
917 gcr_cntl
|= S_586_GL2_INV(1) | S_586_GL2_WB(1) |
918 S_586_GLM_INV(1) | S_586_GLM_WB(1);
919 } else if (flush_bits
& RADV_CMD_FLAG_WB_L2
) {
920 /* Writeback but do not invalidate.
921 * GLM doesn't support WB alone. If WB is set, INV must be set too.
923 gcr_cntl
|= S_586_GL2_WB(1) |
924 S_586_GLM_WB(1) | S_586_GLM_INV(1);
927 /* TODO: Implement this new flag for GFX9+.
928 else if (flush_bits & RADV_CMD_FLAG_INV_L2_METADATA)
929 gcr_cntl |= S_586_GLM_INV(1) | S_586_GLM_WB(1);
932 if (flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
| RADV_CMD_FLAG_FLUSH_AND_INV_DB
)) {
933 /* TODO: trigger on RADV_CMD_FLAG_FLUSH_AND_INV_CB_META */
934 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB
) {
935 /* Flush CMASK/FMASK/DCC. Will wait for idle later. */
936 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
937 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) |
941 /* TODO: trigger on RADV_CMD_FLAG_FLUSH_AND_INV_DB_META ? */
942 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB
) {
943 /* Flush HTILE. Will wait for idle later. */
944 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
945 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) |
949 /* First flush CB/DB, then L1/L2. */
950 gcr_cntl
|= S_586_SEQ(V_586_SEQ_FORWARD
);
952 if ((flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
| RADV_CMD_FLAG_FLUSH_AND_INV_DB
)) ==
953 (RADV_CMD_FLAG_FLUSH_AND_INV_CB
| RADV_CMD_FLAG_FLUSH_AND_INV_DB
)) {
954 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
955 } else if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB
) {
956 cb_db_event
= V_028A90_FLUSH_AND_INV_CB_DATA_TS
;
957 } else if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB
) {
958 cb_db_event
= V_028A90_FLUSH_AND_INV_DB_DATA_TS
;
963 /* Wait for graphics shaders to go idle if requested. */
964 if (flush_bits
& RADV_CMD_FLAG_PS_PARTIAL_FLUSH
) {
965 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
966 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
967 } else if (flush_bits
& RADV_CMD_FLAG_VS_PARTIAL_FLUSH
) {
968 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
969 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
973 if (flush_bits
& RADV_CMD_FLAG_CS_PARTIAL_FLUSH
) {
974 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
975 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
| EVENT_INDEX(4)));
979 /* CB/DB flush and invalidate (or possibly just a wait for a
980 * meta flush) via RELEASE_MEM.
982 * Combine this with other cache flushes when possible; this
983 * requires affected shaders to be idle, so do it after the
984 * CS_PARTIAL_FLUSH before (VS/PS partial flushes are always
987 /* Get GCR_CNTL fields, because the encoding is different in RELEASE_MEM. */
988 unsigned glm_wb
= G_586_GLM_WB(gcr_cntl
);
989 unsigned glm_inv
= G_586_GLM_INV(gcr_cntl
);
990 unsigned glv_inv
= G_586_GLV_INV(gcr_cntl
);
991 unsigned gl1_inv
= G_586_GL1_INV(gcr_cntl
);
992 assert(G_586_GL2_US(gcr_cntl
) == 0);
993 assert(G_586_GL2_RANGE(gcr_cntl
) == 0);
994 assert(G_586_GL2_DISCARD(gcr_cntl
) == 0);
995 unsigned gl2_inv
= G_586_GL2_INV(gcr_cntl
);
996 unsigned gl2_wb
= G_586_GL2_WB(gcr_cntl
);
997 unsigned gcr_seq
= G_586_SEQ(gcr_cntl
);
999 gcr_cntl
&= C_586_GLM_WB
&
1004 C_586_GL2_WB
; /* keep SEQ */
1009 si_cs_emit_write_event_eop(cs
, chip_class
, false, cb_db_event
,
1010 S_490_GLM_WB(glm_wb
) |
1011 S_490_GLM_INV(glm_inv
) |
1012 S_490_GLV_INV(glv_inv
) |
1013 S_490_GL1_INV(gl1_inv
) |
1014 S_490_GL2_INV(gl2_inv
) |
1015 S_490_GL2_WB(gl2_wb
) |
1018 EOP_DATA_SEL_VALUE_32BIT
,
1019 flush_va
, *flush_cnt
,
1022 radv_cp_wait_mem(cs
, WAIT_REG_MEM_EQUAL
, flush_va
,
1023 *flush_cnt
, 0xffffffff);
1026 /* VGT state sync */
1027 if (flush_bits
& RADV_CMD_FLAG_VGT_FLUSH
) {
1028 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1029 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
1032 /* Ignore fields that only modify the behavior of other fields. */
1033 if (gcr_cntl
& C_586_GL1_RANGE
& C_586_GL2_RANGE
& C_586_SEQ
) {
1034 /* Flush caches and wait for the caches to assert idle.
1035 * The cache flush is executed in the ME, but the PFP waits
1038 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 6, 0));
1039 radeon_emit(cs
, 0); /* CP_COHER_CNTL */
1040 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
1041 radeon_emit(cs
, 0xffffff); /* CP_COHER_SIZE_HI */
1042 radeon_emit(cs
, 0); /* CP_COHER_BASE */
1043 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
1044 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
1045 radeon_emit(cs
, gcr_cntl
); /* GCR_CNTL */
1046 } else if ((cb_db_event
||
1047 (flush_bits
& (RADV_CMD_FLAG_VS_PARTIAL_FLUSH
|
1048 RADV_CMD_FLAG_PS_PARTIAL_FLUSH
|
1049 RADV_CMD_FLAG_CS_PARTIAL_FLUSH
)))
1051 /* We need to ensure that PFP waits as well. */
1052 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
1056 if (flush_bits
& RADV_CMD_FLAG_START_PIPELINE_STATS
) {
1057 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1058 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
1060 } else if (flush_bits
& RADV_CMD_FLAG_STOP_PIPELINE_STATS
) {
1061 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1062 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
1068 si_cs_emit_cache_flush(struct radeon_cmdbuf
*cs
,
1069 enum chip_class chip_class
,
1070 uint32_t *flush_cnt
,
1073 enum radv_cmd_flush_bits flush_bits
,
1074 uint64_t gfx9_eop_bug_va
)
1076 unsigned cp_coher_cntl
= 0;
1077 uint32_t flush_cb_db
= flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
1078 RADV_CMD_FLAG_FLUSH_AND_INV_DB
);
1080 if (chip_class
>= GFX10
) {
1081 /* GFX10 cache flush handling is quite different. */
1082 gfx10_cs_emit_cache_flush(cs
, chip_class
, flush_cnt
, flush_va
,
1083 is_mec
, flush_bits
, gfx9_eop_bug_va
);
1087 if (flush_bits
& RADV_CMD_FLAG_INV_ICACHE
)
1088 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
1089 if (flush_bits
& RADV_CMD_FLAG_INV_SCACHE
)
1090 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
1092 if (chip_class
<= GFX8
) {
1093 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB
) {
1094 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
1095 S_0085F0_CB0_DEST_BASE_ENA(1) |
1096 S_0085F0_CB1_DEST_BASE_ENA(1) |
1097 S_0085F0_CB2_DEST_BASE_ENA(1) |
1098 S_0085F0_CB3_DEST_BASE_ENA(1) |
1099 S_0085F0_CB4_DEST_BASE_ENA(1) |
1100 S_0085F0_CB5_DEST_BASE_ENA(1) |
1101 S_0085F0_CB6_DEST_BASE_ENA(1) |
1102 S_0085F0_CB7_DEST_BASE_ENA(1);
1104 /* Necessary for DCC */
1105 if (chip_class
>= GFX8
) {
1106 si_cs_emit_write_event_eop(cs
,
1109 V_028A90_FLUSH_AND_INV_CB_DATA_TS
,
1112 EOP_DATA_SEL_DISCARD
,
1117 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB
) {
1118 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
1119 S_0085F0_DB_DEST_BASE_ENA(1);
1123 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
) {
1124 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1125 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
1128 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
) {
1129 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1130 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
1133 if (flush_bits
& RADV_CMD_FLAG_PS_PARTIAL_FLUSH
) {
1134 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1135 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
1136 } else if (flush_bits
& RADV_CMD_FLAG_VS_PARTIAL_FLUSH
) {
1137 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1138 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
1141 if (flush_bits
& RADV_CMD_FLAG_CS_PARTIAL_FLUSH
) {
1142 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1143 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
1146 if (chip_class
== GFX9
&& flush_cb_db
) {
1147 unsigned cb_db_event
, tc_flags
;
1149 /* Set the CB/DB flush event. */
1150 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
1152 /* These are the only allowed combinations. If you need to
1153 * do multiple operations at once, do them separately.
1154 * All operations that invalidate L2 also seem to invalidate
1155 * metadata. Volatile (VOL) and WC flushes are not listed here.
1157 * TC | TC_WB = writeback & invalidate L2 & L1
1158 * TC | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
1159 * TC_WB | TC_NC = writeback L2 for MTYPE == NC
1160 * TC | TC_NC = invalidate L2 for MTYPE == NC
1161 * TC | TC_MD = writeback & invalidate L2 metadata (DCC, etc.)
1162 * TCL1 = invalidate L1
1164 tc_flags
= EVENT_TC_ACTION_ENA
|
1165 EVENT_TC_MD_ACTION_ENA
;
1167 /* Ideally flush TC together with CB/DB. */
1168 if (flush_bits
& RADV_CMD_FLAG_INV_L2
) {
1169 /* Writeback and invalidate everything in L2 & L1. */
1170 tc_flags
= EVENT_TC_ACTION_ENA
|
1171 EVENT_TC_WB_ACTION_ENA
;
1174 /* Clear the flags. */
1175 flush_bits
&= ~(RADV_CMD_FLAG_INV_L2
|
1176 RADV_CMD_FLAG_WB_L2
|
1177 RADV_CMD_FLAG_INV_VCACHE
);
1182 si_cs_emit_write_event_eop(cs
, chip_class
, false, cb_db_event
, tc_flags
,
1184 EOP_DATA_SEL_VALUE_32BIT
,
1185 flush_va
, *flush_cnt
,
1187 radv_cp_wait_mem(cs
, WAIT_REG_MEM_EQUAL
, flush_va
,
1188 *flush_cnt
, 0xffffffff);
1191 /* VGT state sync */
1192 if (flush_bits
& RADV_CMD_FLAG_VGT_FLUSH
) {
1193 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1194 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
1197 /* VGT streamout state sync */
1198 if (flush_bits
& RADV_CMD_FLAG_VGT_STREAMOUT_SYNC
) {
1199 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1200 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC
) | EVENT_INDEX(0));
1203 /* Make sure ME is idle (it executes most packets) before continuing.
1204 * This prevents read-after-write hazards between PFP and ME.
1206 if ((cp_coher_cntl
||
1207 (flush_bits
& (RADV_CMD_FLAG_CS_PARTIAL_FLUSH
|
1208 RADV_CMD_FLAG_INV_VCACHE
|
1209 RADV_CMD_FLAG_INV_L2
|
1210 RADV_CMD_FLAG_WB_L2
))) &&
1212 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
1216 if ((flush_bits
& RADV_CMD_FLAG_INV_L2
) ||
1217 (chip_class
<= GFX7
&& (flush_bits
& RADV_CMD_FLAG_WB_L2
))) {
1218 si_emit_acquire_mem(cs
, is_mec
, chip_class
== GFX9
,
1220 S_0085F0_TC_ACTION_ENA(1) |
1221 S_0085F0_TCL1_ACTION_ENA(1) |
1222 S_0301F0_TC_WB_ACTION_ENA(chip_class
>= GFX8
));
1225 if(flush_bits
& RADV_CMD_FLAG_WB_L2
) {
1227 * NC = apply to non-coherent MTYPEs
1228 * (i.e. MTYPE <= 1, which is what we use everywhere)
1230 * WB doesn't work without NC.
1232 si_emit_acquire_mem(cs
, is_mec
,
1235 S_0301F0_TC_WB_ACTION_ENA(1) |
1236 S_0301F0_TC_NC_ACTION_ENA(1));
1239 if (flush_bits
& RADV_CMD_FLAG_INV_VCACHE
) {
1240 si_emit_acquire_mem(cs
, is_mec
,
1243 S_0085F0_TCL1_ACTION_ENA(1));
1248 /* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
1249 * Therefore, it should be last. Done in PFP.
1252 si_emit_acquire_mem(cs
, is_mec
, chip_class
== GFX9
, cp_coher_cntl
);
1254 if (flush_bits
& RADV_CMD_FLAG_START_PIPELINE_STATS
) {
1255 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1256 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
1258 } else if (flush_bits
& RADV_CMD_FLAG_STOP_PIPELINE_STATS
) {
1259 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1260 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
1266 si_emit_cache_flush(struct radv_cmd_buffer
*cmd_buffer
)
1268 bool is_compute
= cmd_buffer
->queue_family_index
== RADV_QUEUE_COMPUTE
;
1271 cmd_buffer
->state
.flush_bits
&= ~(RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
1272 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
|
1273 RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
1274 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
|
1275 RADV_CMD_FLAG_PS_PARTIAL_FLUSH
|
1276 RADV_CMD_FLAG_VS_PARTIAL_FLUSH
|
1277 RADV_CMD_FLAG_VGT_FLUSH
|
1278 RADV_CMD_FLAG_START_PIPELINE_STATS
|
1279 RADV_CMD_FLAG_STOP_PIPELINE_STATS
);
1281 if (!cmd_buffer
->state
.flush_bits
)
1284 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 128);
1286 si_cs_emit_cache_flush(cmd_buffer
->cs
,
1287 cmd_buffer
->device
->physical_device
->rad_info
.chip_class
,
1288 &cmd_buffer
->gfx9_fence_idx
,
1289 cmd_buffer
->gfx9_fence_va
,
1290 radv_cmd_buffer_uses_mec(cmd_buffer
),
1291 cmd_buffer
->state
.flush_bits
,
1292 cmd_buffer
->gfx9_eop_bug_va
);
1295 if (unlikely(cmd_buffer
->device
->trace_bo
))
1296 radv_cmd_buffer_trace_emit(cmd_buffer
);
1298 /* Clear the caches that have been flushed to avoid syncing too much
1299 * when there is some pending active queries.
1301 cmd_buffer
->active_query_flush_bits
&= ~cmd_buffer
->state
.flush_bits
;
1303 cmd_buffer
->state
.flush_bits
= 0;
1305 /* If the driver used a compute shader for resetting a query pool, it
1306 * should be finished at this point.
1308 cmd_buffer
->pending_reset_query
= false;
1311 /* sets the CP predication state using a boolean stored at va */
1313 si_emit_set_predication_state(struct radv_cmd_buffer
*cmd_buffer
,
1314 bool draw_visible
, uint64_t va
)
1319 op
= PRED_OP(PREDICATION_OP_BOOL64
);
1321 /* PREDICATION_DRAW_VISIBLE means that if the 32-bit value is
1322 * zero, all rendering commands are discarded. Otherwise, they
1323 * are discarded if the value is non zero.
1325 op
|= draw_visible
? PREDICATION_DRAW_VISIBLE
:
1326 PREDICATION_DRAW_NOT_VISIBLE
;
1328 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1329 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_SET_PREDICATION
, 2, 0));
1330 radeon_emit(cmd_buffer
->cs
, op
);
1331 radeon_emit(cmd_buffer
->cs
, va
);
1332 radeon_emit(cmd_buffer
->cs
, va
>> 32);
1334 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_SET_PREDICATION
, 1, 0));
1335 radeon_emit(cmd_buffer
->cs
, va
);
1336 radeon_emit(cmd_buffer
->cs
, op
| ((va
>> 32) & 0xFF));
1340 /* Set this if you want the 3D engine to wait until CP DMA is done.
1341 * It should be set on the last CP DMA packet. */
1342 #define CP_DMA_SYNC (1 << 0)
1344 /* Set this if the source data was used as a destination in a previous CP DMA
1345 * packet. It's for preventing a read-after-write (RAW) hazard between two
1346 * CP DMA packets. */
1347 #define CP_DMA_RAW_WAIT (1 << 1)
1348 #define CP_DMA_USE_L2 (1 << 2)
1349 #define CP_DMA_CLEAR (1 << 3)
1351 /* Alignment for optimal performance. */
1352 #define SI_CPDMA_ALIGNMENT 32
1354 /* The max number of bytes that can be copied per packet. */
1355 static inline unsigned cp_dma_max_byte_count(struct radv_cmd_buffer
*cmd_buffer
)
1357 unsigned max
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
?
1358 S_414_BYTE_COUNT_GFX9(~0u) :
1359 S_414_BYTE_COUNT_GFX6(~0u);
1361 /* make it aligned for optimal performance */
1362 return max
& ~(SI_CPDMA_ALIGNMENT
- 1);
1365 /* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
1366 * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
1369 static void si_emit_cp_dma(struct radv_cmd_buffer
*cmd_buffer
,
1370 uint64_t dst_va
, uint64_t src_va
,
1371 unsigned size
, unsigned flags
)
1373 struct radeon_cmdbuf
*cs
= cmd_buffer
->cs
;
1374 uint32_t header
= 0, command
= 0;
1376 assert(size
<= cp_dma_max_byte_count(cmd_buffer
));
1378 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 9);
1379 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1380 command
|= S_414_BYTE_COUNT_GFX9(size
);
1382 command
|= S_414_BYTE_COUNT_GFX6(size
);
1385 if (flags
& CP_DMA_SYNC
)
1386 header
|= S_411_CP_SYNC(1);
1388 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1389 command
|= S_414_DISABLE_WR_CONFIRM_GFX9(1);
1391 command
|= S_414_DISABLE_WR_CONFIRM_GFX6(1);
1394 if (flags
& CP_DMA_RAW_WAIT
)
1395 command
|= S_414_RAW_WAIT(1);
1397 /* Src and dst flags. */
1398 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1399 !(flags
& CP_DMA_CLEAR
) &&
1401 header
|= S_411_DST_SEL(V_411_NOWHERE
); /* prefetch only */
1402 else if (flags
& CP_DMA_USE_L2
)
1403 header
|= S_411_DST_SEL(V_411_DST_ADDR_TC_L2
);
1405 if (flags
& CP_DMA_CLEAR
)
1406 header
|= S_411_SRC_SEL(V_411_DATA
);
1407 else if (flags
& CP_DMA_USE_L2
)
1408 header
|= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2
);
1410 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
1411 radeon_emit(cs
, PKT3(PKT3_DMA_DATA
, 5, cmd_buffer
->state
.predicating
));
1412 radeon_emit(cs
, header
);
1413 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
1414 radeon_emit(cs
, src_va
>> 32); /* SRC_ADDR_HI [31:0] */
1415 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
1416 radeon_emit(cs
, dst_va
>> 32); /* DST_ADDR_HI [31:0] */
1417 radeon_emit(cs
, command
);
1419 assert(!(flags
& CP_DMA_USE_L2
));
1420 header
|= S_411_SRC_ADDR_HI(src_va
>> 32);
1421 radeon_emit(cs
, PKT3(PKT3_CP_DMA
, 4, cmd_buffer
->state
.predicating
));
1422 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
1423 radeon_emit(cs
, header
); /* SRC_ADDR_HI [15:0] + flags. */
1424 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
1425 radeon_emit(cs
, (dst_va
>> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
1426 radeon_emit(cs
, command
);
1429 /* CP DMA is executed in ME, but index buffers are read by PFP.
1430 * This ensures that ME (CP DMA) is idle before PFP starts fetching
1431 * indices. If we wanted to execute CP DMA in PFP, this packet
1432 * should precede it.
1434 if (flags
& CP_DMA_SYNC
) {
1435 if (cmd_buffer
->queue_family_index
== RADV_QUEUE_GENERAL
) {
1436 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, cmd_buffer
->state
.predicating
));
1440 /* CP will see the sync flag and wait for all DMAs to complete. */
1441 cmd_buffer
->state
.dma_is_busy
= false;
1444 if (unlikely(cmd_buffer
->device
->trace_bo
))
1445 radv_cmd_buffer_trace_emit(cmd_buffer
);
1448 void si_cp_dma_prefetch(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1451 uint64_t aligned_va
= va
& ~(SI_CPDMA_ALIGNMENT
- 1);
1452 uint64_t aligned_size
= ((va
+ size
+ SI_CPDMA_ALIGNMENT
-1) & ~(SI_CPDMA_ALIGNMENT
- 1)) - aligned_va
;
1454 si_emit_cp_dma(cmd_buffer
, aligned_va
, aligned_va
,
1455 aligned_size
, CP_DMA_USE_L2
);
1458 static void si_cp_dma_prepare(struct radv_cmd_buffer
*cmd_buffer
, uint64_t byte_count
,
1459 uint64_t remaining_size
, unsigned *flags
)
1462 /* Flush the caches for the first copy only.
1463 * Also wait for the previous CP DMA operations.
1465 if (cmd_buffer
->state
.flush_bits
) {
1466 si_emit_cache_flush(cmd_buffer
);
1467 *flags
|= CP_DMA_RAW_WAIT
;
1470 /* Do the synchronization after the last dma, so that all data
1471 * is written to memory.
1473 if (byte_count
== remaining_size
)
1474 *flags
|= CP_DMA_SYNC
;
1477 static void si_cp_dma_realign_engine(struct radv_cmd_buffer
*cmd_buffer
, unsigned size
)
1481 unsigned dma_flags
= 0;
1482 unsigned buf_size
= SI_CPDMA_ALIGNMENT
* 2;
1485 assert(size
< SI_CPDMA_ALIGNMENT
);
1487 radv_cmd_buffer_upload_alloc(cmd_buffer
, buf_size
, SI_CPDMA_ALIGNMENT
, &offset
, &ptr
);
1489 va
= radv_buffer_get_va(cmd_buffer
->upload
.upload_bo
);
1492 si_cp_dma_prepare(cmd_buffer
, size
, size
, &dma_flags
);
1494 si_emit_cp_dma(cmd_buffer
, va
, va
+ SI_CPDMA_ALIGNMENT
, size
,
1498 void si_cp_dma_buffer_copy(struct radv_cmd_buffer
*cmd_buffer
,
1499 uint64_t src_va
, uint64_t dest_va
,
1502 uint64_t main_src_va
, main_dest_va
;
1503 uint64_t skipped_size
= 0, realign_size
= 0;
1505 /* Assume that we are not going to sync after the last DMA operation. */
1506 cmd_buffer
->state
.dma_is_busy
= true;
1508 if (cmd_buffer
->device
->physical_device
->rad_info
.family
<= CHIP_CARRIZO
||
1509 cmd_buffer
->device
->physical_device
->rad_info
.family
== CHIP_STONEY
) {
1510 /* If the size is not aligned, we must add a dummy copy at the end
1511 * just to align the internal counter. Otherwise, the DMA engine
1512 * would slow down by an order of magnitude for following copies.
1514 if (size
% SI_CPDMA_ALIGNMENT
)
1515 realign_size
= SI_CPDMA_ALIGNMENT
- (size
% SI_CPDMA_ALIGNMENT
);
1517 /* If the copy begins unaligned, we must start copying from the next
1518 * aligned block and the skipped part should be copied after everything
1519 * else has been copied. Only the src alignment matters, not dst.
1521 if (src_va
% SI_CPDMA_ALIGNMENT
) {
1522 skipped_size
= SI_CPDMA_ALIGNMENT
- (src_va
% SI_CPDMA_ALIGNMENT
);
1523 /* The main part will be skipped if the size is too small. */
1524 skipped_size
= MIN2(skipped_size
, size
);
1525 size
-= skipped_size
;
1528 main_src_va
= src_va
+ skipped_size
;
1529 main_dest_va
= dest_va
+ skipped_size
;
1532 unsigned dma_flags
= 0;
1533 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(cmd_buffer
));
1535 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
1536 /* DMA operations via L2 are coherent and faster.
1537 * TODO: GFX7-GFX9 should also support this but it
1538 * requires tests/benchmarks.
1540 dma_flags
|= CP_DMA_USE_L2
;
1543 si_cp_dma_prepare(cmd_buffer
, byte_count
,
1544 size
+ skipped_size
+ realign_size
,
1547 dma_flags
&= ~CP_DMA_SYNC
;
1549 si_emit_cp_dma(cmd_buffer
, main_dest_va
, main_src_va
,
1550 byte_count
, dma_flags
);
1553 main_src_va
+= byte_count
;
1554 main_dest_va
+= byte_count
;
1558 unsigned dma_flags
= 0;
1560 si_cp_dma_prepare(cmd_buffer
, skipped_size
,
1561 size
+ skipped_size
+ realign_size
,
1564 si_emit_cp_dma(cmd_buffer
, dest_va
, src_va
,
1565 skipped_size
, dma_flags
);
1568 si_cp_dma_realign_engine(cmd_buffer
, realign_size
);
1571 void si_cp_dma_clear_buffer(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1572 uint64_t size
, unsigned value
)
1578 assert(va
% 4 == 0 && size
% 4 == 0);
1580 /* Assume that we are not going to sync after the last DMA operation. */
1581 cmd_buffer
->state
.dma_is_busy
= true;
1584 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(cmd_buffer
));
1585 unsigned dma_flags
= CP_DMA_CLEAR
;
1587 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
1588 /* DMA operations via L2 are coherent and faster.
1589 * TODO: GFX7-GFX9 should also support this but it
1590 * requires tests/benchmarks.
1592 dma_flags
|= CP_DMA_USE_L2
;
1595 si_cp_dma_prepare(cmd_buffer
, byte_count
, size
, &dma_flags
);
1597 /* Emit the clear packet. */
1598 si_emit_cp_dma(cmd_buffer
, va
, value
, byte_count
,
1606 void si_cp_dma_wait_for_idle(struct radv_cmd_buffer
*cmd_buffer
)
1608 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
< GFX7
)
1611 if (!cmd_buffer
->state
.dma_is_busy
)
1614 /* Issue a dummy DMA that copies zero bytes.
1616 * The DMA engine will see that there's no work to do and skip this
1617 * DMA request, however, the CP will see the sync flag and still wait
1618 * for all DMAs to complete.
1620 si_emit_cp_dma(cmd_buffer
, 0, 0, 0, CP_DMA_SYNC
);
1622 cmd_buffer
->state
.dma_is_busy
= false;
1625 /* For MSAA sample positions. */
1626 #define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y) \
1627 ((((unsigned)(s0x) & 0xf) << 0) | (((unsigned)(s0y) & 0xf) << 4) | \
1628 (((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) | \
1629 (((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) | \
1630 (((unsigned)(s3x) & 0xf) << 24) | (((unsigned)(s3y) & 0xf) << 28))
1632 /* For obtaining location coordinates from registers */
1633 #define SEXT4(x) ((int)((x) | ((x) & 0x8 ? 0xfffffff0 : 0)))
1634 #define GET_SFIELD(reg, index) SEXT4(((reg) >> ((index) * 4)) & 0xf)
1635 #define GET_SX(reg, index) GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2)
1636 #define GET_SY(reg, index) GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2 + 1)
1639 static const uint32_t sample_locs_1x
=
1640 FILL_SREG(0, 0, 0, 0, 0, 0, 0, 0);
1641 static const unsigned max_dist_1x
= 0;
1642 static const uint64_t centroid_priority_1x
= 0x0000000000000000ull
;
1645 static const uint32_t sample_locs_2x
=
1646 FILL_SREG(4,4, -4, -4, 0, 0, 0, 0);
1647 static const unsigned max_dist_2x
= 4;
1648 static const uint64_t centroid_priority_2x
= 0x1010101010101010ull
;
1651 static const uint32_t sample_locs_4x
=
1652 FILL_SREG(-2,-6, 6, -2, -6, 2, 2, 6);
1653 static const unsigned max_dist_4x
= 6;
1654 static const uint64_t centroid_priority_4x
= 0x3210321032103210ull
;
1657 static const uint32_t sample_locs_8x
[] = {
1658 FILL_SREG( 1,-3, -1, 3, 5, 1, -3,-5),
1659 FILL_SREG(-5, 5, -7,-1, 3, 7, 7,-7),
1660 /* The following are unused by hardware, but we emit them to IBs
1661 * instead of multiple SET_CONTEXT_REG packets. */
1665 static const unsigned max_dist_8x
= 7;
1666 static const uint64_t centroid_priority_8x
= 0x7654321076543210ull
;
1668 unsigned radv_get_default_max_sample_dist(int log_samples
)
1670 unsigned max_dist
[] = {
1676 return max_dist
[log_samples
];
1679 void radv_emit_default_sample_locations(struct radeon_cmdbuf
*cs
, int nr_samples
)
1681 switch (nr_samples
) {
1684 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1685 radeon_emit(cs
, (uint32_t)centroid_priority_1x
);
1686 radeon_emit(cs
, centroid_priority_1x
>> 32);
1687 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, sample_locs_1x
);
1688 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, sample_locs_1x
);
1689 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, sample_locs_1x
);
1690 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, sample_locs_1x
);
1693 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1694 radeon_emit(cs
, (uint32_t)centroid_priority_2x
);
1695 radeon_emit(cs
, centroid_priority_2x
>> 32);
1696 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, sample_locs_2x
);
1697 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, sample_locs_2x
);
1698 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, sample_locs_2x
);
1699 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, sample_locs_2x
);
1702 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1703 radeon_emit(cs
, (uint32_t)centroid_priority_4x
);
1704 radeon_emit(cs
, centroid_priority_4x
>> 32);
1705 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, sample_locs_4x
);
1706 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, sample_locs_4x
);
1707 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, sample_locs_4x
);
1708 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, sample_locs_4x
);
1711 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1712 radeon_emit(cs
, (uint32_t)centroid_priority_8x
);
1713 radeon_emit(cs
, centroid_priority_8x
>> 32);
1714 radeon_set_context_reg_seq(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 14);
1715 radeon_emit_array(cs
, sample_locs_8x
, 4);
1716 radeon_emit_array(cs
, sample_locs_8x
, 4);
1717 radeon_emit_array(cs
, sample_locs_8x
, 4);
1718 radeon_emit_array(cs
, sample_locs_8x
, 2);
1723 static void radv_get_sample_position(struct radv_device
*device
,
1724 unsigned sample_count
,
1725 unsigned sample_index
, float *out_value
)
1727 const uint32_t *sample_locs
;
1729 switch (sample_count
) {
1732 sample_locs
= &sample_locs_1x
;
1735 sample_locs
= &sample_locs_2x
;
1738 sample_locs
= &sample_locs_4x
;
1741 sample_locs
= sample_locs_8x
;
1745 out_value
[0] = (GET_SX(sample_locs
, sample_index
) + 8) / 16.0f
;
1746 out_value
[1] = (GET_SY(sample_locs
, sample_index
) + 8) / 16.0f
;
1749 void radv_device_init_msaa(struct radv_device
*device
)
1753 radv_get_sample_position(device
, 1, 0, device
->sample_locations_1x
[0]);
1755 for (i
= 0; i
< 2; i
++)
1756 radv_get_sample_position(device
, 2, i
, device
->sample_locations_2x
[i
]);
1757 for (i
= 0; i
< 4; i
++)
1758 radv_get_sample_position(device
, 4, i
, device
->sample_locations_4x
[i
]);
1759 for (i
= 0; i
< 8; i
++)
1760 radv_get_sample_position(device
, 8, i
, device
->sample_locations_8x
[i
]);