2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
6 * Copyright © 2015 Advanced Micro Devices, Inc.
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 /* command buffer handling for AMD GCN */
30 #include "radv_private.h"
31 #include "radv_shader.h"
34 #include "radv_util.h"
35 #include "main/macros.h"
38 si_write_harvested_raster_configs(struct radv_physical_device
*physical_device
,
39 struct radeon_cmdbuf
*cs
,
40 unsigned raster_config
,
41 unsigned raster_config_1
)
43 unsigned num_se
= MAX2(physical_device
->rad_info
.max_se
, 1);
44 unsigned raster_config_se
[4];
47 ac_get_harvested_configs(&physical_device
->rad_info
,
52 for (se
= 0; se
< num_se
; se
++) {
53 /* GRBM_GFX_INDEX has a different offset on GFX6 and GFX7+ */
54 if (physical_device
->rad_info
.chip_class
< GFX7
)
55 radeon_set_config_reg(cs
, R_00802C_GRBM_GFX_INDEX
,
56 S_00802C_SE_INDEX(se
) |
57 S_00802C_SH_BROADCAST_WRITES(1) |
58 S_00802C_INSTANCE_BROADCAST_WRITES(1));
60 radeon_set_uconfig_reg(cs
, R_030800_GRBM_GFX_INDEX
,
61 S_030800_SE_INDEX(se
) | S_030800_SH_BROADCAST_WRITES(1) |
62 S_030800_INSTANCE_BROADCAST_WRITES(1));
63 radeon_set_context_reg(cs
, R_028350_PA_SC_RASTER_CONFIG
, raster_config_se
[se
]);
66 /* GRBM_GFX_INDEX has a different offset on GFX6 and GFX7+ */
67 if (physical_device
->rad_info
.chip_class
< GFX7
)
68 radeon_set_config_reg(cs
, R_00802C_GRBM_GFX_INDEX
,
69 S_00802C_SE_BROADCAST_WRITES(1) |
70 S_00802C_SH_BROADCAST_WRITES(1) |
71 S_00802C_INSTANCE_BROADCAST_WRITES(1));
73 radeon_set_uconfig_reg(cs
, R_030800_GRBM_GFX_INDEX
,
74 S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
75 S_030800_INSTANCE_BROADCAST_WRITES(1));
77 if (physical_device
->rad_info
.chip_class
>= GFX7
)
78 radeon_set_context_reg(cs
, R_028354_PA_SC_RASTER_CONFIG_1
, raster_config_1
);
82 si_emit_compute(struct radv_physical_device
*physical_device
,
83 struct radeon_cmdbuf
*cs
)
85 radeon_set_sh_reg_seq(cs
, R_00B810_COMPUTE_START_X
, 3);
90 radeon_set_sh_reg_seq(cs
, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0
, 2);
91 /* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1 */
92 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 /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
106 * and is now per pipe, so it should be handled in the
107 * kernel if we want to use something other than the default value,
108 * which is now 0x22f.
110 if (physical_device
->rad_info
.chip_class
<= GFX6
) {
111 /* XXX: This should be:
112 * (number of compute units) * 4 * (waves per simd) - 1 */
114 radeon_set_sh_reg(cs
, R_00B82C_COMPUTE_MAX_WAVE_ID
,
115 0x190 /* Default value */);
119 /* 12.4 fixed-point */
120 static unsigned radv_pack_float_12p4(float x
)
123 x
>= 4096 ? 0xffff : x
* 16;
127 si_set_raster_config(struct radv_physical_device
*physical_device
,
128 struct radeon_cmdbuf
*cs
)
130 unsigned num_rb
= MIN2(physical_device
->rad_info
.num_render_backends
, 16);
131 unsigned rb_mask
= physical_device
->rad_info
.enabled_rb_mask
;
132 unsigned raster_config
, raster_config_1
;
134 ac_get_raster_config(&physical_device
->rad_info
,
136 &raster_config_1
, NULL
);
138 /* Always use the default config when all backends are enabled
139 * (or when we failed to determine the enabled backends).
141 if (!rb_mask
|| util_bitcount(rb_mask
) >= num_rb
) {
142 radeon_set_context_reg(cs
, R_028350_PA_SC_RASTER_CONFIG
,
144 if (physical_device
->rad_info
.chip_class
>= GFX7
)
145 radeon_set_context_reg(cs
, R_028354_PA_SC_RASTER_CONFIG_1
,
148 si_write_harvested_raster_configs(physical_device
, cs
,
155 si_emit_graphics(struct radv_physical_device
*physical_device
,
156 struct radeon_cmdbuf
*cs
)
160 /* Only GFX6 can disable CLEAR_STATE for now. */
161 assert(physical_device
->has_clear_state
||
162 physical_device
->rad_info
.chip_class
== GFX6
);
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 (physical_device
->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 (!physical_device
->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 (!physical_device
->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 radeon_set_context_reg(cs
, R_028AA0_VGT_INSTANCE_STEP_RATE_0
, 1);
193 if (!physical_device
->has_clear_state
)
194 radeon_set_context_reg(cs
, R_028AB8_VGT_VTX_CNT_EN
, 0x0);
195 if (physical_device
->rad_info
.chip_class
< GFX7
)
196 radeon_set_config_reg(cs
, R_008A14_PA_CL_ENHANCE
, S_008A14_NUM_CLIP_SEQ(3) |
197 S_008A14_CLIP_VTX_REORDER_ENA(1));
199 if (!physical_device
->has_clear_state
)
200 radeon_set_context_reg(cs
, R_02882C_PA_SU_PRIM_FILTER_CNTL
, 0);
202 /* CLEAR_STATE doesn't clear these correctly on certain generations.
203 * I don't know why. Deduced by trial and error.
205 if (physical_device
->rad_info
.chip_class
<= GFX7
) {
206 radeon_set_context_reg(cs
, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET
, 0);
207 radeon_set_context_reg(cs
, R_028204_PA_SC_WINDOW_SCISSOR_TL
,
208 S_028204_WINDOW_OFFSET_DISABLE(1));
209 radeon_set_context_reg(cs
, R_028240_PA_SC_GENERIC_SCISSOR_TL
,
210 S_028240_WINDOW_OFFSET_DISABLE(1));
211 radeon_set_context_reg(cs
, R_028244_PA_SC_GENERIC_SCISSOR_BR
,
212 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
213 radeon_set_context_reg(cs
, R_028030_PA_SC_SCREEN_SCISSOR_TL
, 0);
214 radeon_set_context_reg(cs
, R_028034_PA_SC_SCREEN_SCISSOR_BR
,
215 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
218 if (!physical_device
->has_clear_state
) {
219 for (i
= 0; i
< 16; i
++) {
220 radeon_set_context_reg(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+ i
*8, 0);
221 radeon_set_context_reg(cs
, R_0282D4_PA_SC_VPORT_ZMAX_0
+ i
*8, fui(1.0));
225 if (!physical_device
->has_clear_state
) {
226 radeon_set_context_reg(cs
, R_02820C_PA_SC_CLIPRECT_RULE
, 0xFFFF);
227 radeon_set_context_reg(cs
, R_028230_PA_SC_EDGERULE
, 0xAAAAAAAA);
228 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on GFX6 */
229 radeon_set_context_reg(cs
, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET
, 0);
230 radeon_set_context_reg(cs
, R_028820_PA_CL_NANINF_CNTL
, 0);
231 radeon_set_context_reg(cs
, R_028AC0_DB_SRESULTS_COMPARE_STATE0
, 0x0);
232 radeon_set_context_reg(cs
, R_028AC4_DB_SRESULTS_COMPARE_STATE1
, 0x0);
233 radeon_set_context_reg(cs
, R_028AC8_DB_PRELOAD_CONTROL
, 0x0);
236 radeon_set_context_reg(cs
, R_02800C_DB_RENDER_OVERRIDE
,
237 S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE
) |
238 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE
));
240 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
241 radeon_set_uconfig_reg(cs
, R_030920_VGT_MAX_VTX_INDX
, ~0);
242 radeon_set_uconfig_reg(cs
, R_030924_VGT_MIN_VTX_INDX
, 0);
243 radeon_set_uconfig_reg(cs
, R_030928_VGT_INDX_OFFSET
, 0);
245 /* These registers, when written, also overwrite the
246 * CLEAR_STATE context, so we can't rely on CLEAR_STATE setting
247 * them. It would be an issue if there was another UMD
250 radeon_set_context_reg(cs
, R_028400_VGT_MAX_VTX_INDX
, ~0);
251 radeon_set_context_reg(cs
, R_028404_VGT_MIN_VTX_INDX
, 0);
252 radeon_set_context_reg(cs
, R_028408_VGT_INDX_OFFSET
, 0);
255 if (physical_device
->rad_info
.chip_class
>= GFX7
) {
256 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
257 radeon_set_sh_reg(cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
,
258 S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
260 radeon_set_sh_reg(cs
, R_00B51C_SPI_SHADER_PGM_RSRC3_LS
,
261 S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
262 radeon_set_sh_reg(cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
,
263 S_00B41C_WAVE_LIMIT(0x3F));
264 radeon_set_sh_reg(cs
, R_00B31C_SPI_SHADER_PGM_RSRC3_ES
,
265 S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
266 /* If this is 0, Bonaire can hang even if GS isn't being used.
267 * Other chips are unaffected. These are suboptimal values,
268 * but we don't use on-chip GS.
270 radeon_set_context_reg(cs
, R_028A44_VGT_GS_ONCHIP_CNTL
,
271 S_028A44_ES_VERTS_PER_SUBGRP(64) |
272 S_028A44_GS_PRIMS_PER_SUBGRP(4));
274 radeon_set_sh_reg(cs
, R_00B21C_SPI_SHADER_PGM_RSRC3_GS
,
275 S_00B21C_CU_EN(0xffff) | S_00B21C_WAVE_LIMIT(0x3F));
277 if (physical_device
->rad_info
.num_good_cu_per_sh
<= 4) {
278 /* Too few available compute units per SH. Disallowing
279 * VS to run on CU0 could hurt us more than late VS
280 * allocation would help.
282 * LATE_ALLOC_VS = 2 is the highest safe number.
284 radeon_set_sh_reg(cs
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
,
285 S_00B118_CU_EN(0xffff) | S_00B118_WAVE_LIMIT(0x3F) );
286 radeon_set_sh_reg(cs
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
, S_00B11C_LIMIT(2));
288 /* Set LATE_ALLOC_VS == 31. It should be less than
289 * the number of scratch waves. Limitations:
290 * - VS can't execute on CU0.
291 * - If HS writes outputs to LDS, LS can't execute on CU0.
293 radeon_set_sh_reg(cs
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
,
294 S_00B118_CU_EN(0xfffe) | S_00B118_WAVE_LIMIT(0x3F));
295 radeon_set_sh_reg(cs
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
, S_00B11C_LIMIT(31));
298 radeon_set_sh_reg(cs
, R_00B01C_SPI_SHADER_PGM_RSRC3_PS
,
299 S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
302 if (physical_device
->rad_info
.chip_class
>= GFX8
) {
303 uint32_t vgt_tess_distribution
;
305 vgt_tess_distribution
= S_028B50_ACCUM_ISOLINE(32) |
306 S_028B50_ACCUM_TRI(11) |
307 S_028B50_ACCUM_QUAD(11) |
308 S_028B50_DONUT_SPLIT(16);
310 if (physical_device
->rad_info
.family
== CHIP_FIJI
||
311 physical_device
->rad_info
.family
>= CHIP_POLARIS10
)
312 vgt_tess_distribution
|= S_028B50_TRAP_SPLIT(3);
314 radeon_set_context_reg(cs
, R_028B50_VGT_TESS_DISTRIBUTION
,
315 vgt_tess_distribution
);
316 } else if (!physical_device
->has_clear_state
) {
317 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 14);
318 radeon_set_context_reg(cs
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 16);
321 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
322 unsigned num_se
= physical_device
->rad_info
.max_se
;
323 unsigned pc_lines
= 0;
324 unsigned max_alloc_count
= 0;
326 switch (physical_device
->rad_info
.family
) {
345 if (physical_device
->rad_info
.chip_class
>= GFX10
) {
346 max_alloc_count
= pc_lines
/ 3;
348 max_alloc_count
= MIN2(128, pc_lines
/ (4 * num_se
));
351 radeon_set_context_reg(cs
, R_028C48_PA_SC_BINNER_CNTL_1
,
352 S_028C48_MAX_ALLOC_COUNT(max_alloc_count
) |
353 S_028C48_MAX_PRIM_PER_BATCH(1023));
354 radeon_set_context_reg(cs
, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL
,
355 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
356 radeon_set_uconfig_reg(cs
, R_030968_VGT_INSTANCE_BASE_ID
, 0);
359 unsigned tmp
= (unsigned)(1.0 * 8.0);
360 radeon_set_context_reg_seq(cs
, R_028A00_PA_SU_POINT_SIZE
, 1);
361 radeon_emit(cs
, S_028A00_HEIGHT(tmp
) | S_028A00_WIDTH(tmp
));
362 radeon_set_context_reg_seq(cs
, R_028A04_PA_SU_POINT_MINMAX
, 1);
363 radeon_emit(cs
, S_028A04_MIN_SIZE(radv_pack_float_12p4(0)) |
364 S_028A04_MAX_SIZE(radv_pack_float_12p4(8192/2)));
366 if (!physical_device
->has_clear_state
) {
367 radeon_set_context_reg(cs
, R_028004_DB_COUNT_CONTROL
,
368 S_028004_ZPASS_INCREMENT_DISABLE(1));
371 /* Enable the Polaris small primitive filter control.
372 * XXX: There is possibly an issue when MSAA is off (see RadeonSI
373 * has_msaa_sample_loc_bug). But this doesn't seem to regress anything,
374 * and AMDVLK doesn't have a workaround as well.
376 if (physical_device
->rad_info
.family
>= CHIP_POLARIS10
) {
377 unsigned small_prim_filter_cntl
=
378 S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
379 /* Workaround for a hw line bug. */
380 S_028830_LINE_FILTER_DISABLE(physical_device
->rad_info
.family
<= CHIP_POLARIS12
);
382 radeon_set_context_reg(cs
, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL
,
383 small_prim_filter_cntl
);
386 si_emit_compute(physical_device
, cs
);
390 cik_create_gfx_config(struct radv_device
*device
)
392 struct radeon_cmdbuf
*cs
= device
->ws
->cs_create(device
->ws
, RING_GFX
);
396 si_emit_graphics(device
->physical_device
, cs
);
398 while (cs
->cdw
& 7) {
399 if (device
->physical_device
->rad_info
.gfx_ib_pad_with_type2
)
400 radeon_emit(cs
, 0x80000000);
402 radeon_emit(cs
, 0xffff1000);
405 device
->gfx_init
= device
->ws
->buffer_create(device
->ws
,
408 RADEON_FLAG_CPU_ACCESS
|
409 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
410 RADEON_FLAG_READ_ONLY
,
411 RADV_BO_PRIORITY_CS
);
412 if (!device
->gfx_init
)
415 void *map
= device
->ws
->buffer_map(device
->gfx_init
);
417 device
->ws
->buffer_destroy(device
->gfx_init
);
418 device
->gfx_init
= NULL
;
421 memcpy(map
, cs
->buf
, cs
->cdw
* 4);
423 device
->ws
->buffer_unmap(device
->gfx_init
);
424 device
->gfx_init_size_dw
= cs
->cdw
;
426 device
->ws
->cs_destroy(cs
);
430 get_viewport_xform(const VkViewport
*viewport
,
431 float scale
[3], float translate
[3])
433 float x
= viewport
->x
;
434 float y
= viewport
->y
;
435 float half_width
= 0.5f
* viewport
->width
;
436 float half_height
= 0.5f
* viewport
->height
;
437 double n
= viewport
->minDepth
;
438 double f
= viewport
->maxDepth
;
440 scale
[0] = half_width
;
441 translate
[0] = half_width
+ x
;
442 scale
[1] = half_height
;
443 translate
[1] = half_height
+ y
;
450 si_write_viewport(struct radeon_cmdbuf
*cs
, int first_vp
,
451 int count
, const VkViewport
*viewports
)
456 radeon_set_context_reg_seq(cs
, R_02843C_PA_CL_VPORT_XSCALE
+
457 first_vp
* 4 * 6, count
* 6);
459 for (i
= 0; i
< count
; i
++) {
460 float scale
[3], translate
[3];
463 get_viewport_xform(&viewports
[i
], scale
, translate
);
464 radeon_emit(cs
, fui(scale
[0]));
465 radeon_emit(cs
, fui(translate
[0]));
466 radeon_emit(cs
, fui(scale
[1]));
467 radeon_emit(cs
, fui(translate
[1]));
468 radeon_emit(cs
, fui(scale
[2]));
469 radeon_emit(cs
, fui(translate
[2]));
472 radeon_set_context_reg_seq(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+
473 first_vp
* 4 * 2, count
* 2);
474 for (i
= 0; i
< count
; i
++) {
475 float zmin
= MIN2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
476 float zmax
= MAX2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
477 radeon_emit(cs
, fui(zmin
));
478 radeon_emit(cs
, fui(zmax
));
482 static VkRect2D
si_scissor_from_viewport(const VkViewport
*viewport
)
484 float scale
[3], translate
[3];
487 get_viewport_xform(viewport
, scale
, translate
);
489 rect
.offset
.x
= translate
[0] - fabs(scale
[0]);
490 rect
.offset
.y
= translate
[1] - fabs(scale
[1]);
491 rect
.extent
.width
= ceilf(translate
[0] + fabs(scale
[0])) - rect
.offset
.x
;
492 rect
.extent
.height
= ceilf(translate
[1] + fabs(scale
[1])) - rect
.offset
.y
;
497 static VkRect2D
si_intersect_scissor(const VkRect2D
*a
, const VkRect2D
*b
) {
499 ret
.offset
.x
= MAX2(a
->offset
.x
, b
->offset
.x
);
500 ret
.offset
.y
= MAX2(a
->offset
.y
, b
->offset
.y
);
501 ret
.extent
.width
= MIN2(a
->offset
.x
+ a
->extent
.width
,
502 b
->offset
.x
+ b
->extent
.width
) - ret
.offset
.x
;
503 ret
.extent
.height
= MIN2(a
->offset
.y
+ a
->extent
.height
,
504 b
->offset
.y
+ b
->extent
.height
) - ret
.offset
.y
;
509 si_write_scissors(struct radeon_cmdbuf
*cs
, int first
,
510 int count
, const VkRect2D
*scissors
,
511 const VkViewport
*viewports
, bool can_use_guardband
)
514 float scale
[3], translate
[3], guardband_x
= INFINITY
, guardband_y
= INFINITY
;
515 const float max_range
= 32767.0f
;
519 radeon_set_context_reg_seq(cs
, R_028250_PA_SC_VPORT_SCISSOR_0_TL
+ first
* 4 * 2, count
* 2);
520 for (i
= 0; i
< count
; i
++) {
521 VkRect2D viewport_scissor
= si_scissor_from_viewport(viewports
+ i
);
522 VkRect2D scissor
= si_intersect_scissor(&scissors
[i
], &viewport_scissor
);
524 get_viewport_xform(viewports
+ i
, scale
, translate
);
525 scale
[0] = fabsf(scale
[0]);
526 scale
[1] = fabsf(scale
[1]);
533 guardband_x
= MIN2(guardband_x
, (max_range
- fabsf(translate
[0])) / scale
[0]);
534 guardband_y
= MIN2(guardband_y
, (max_range
- fabsf(translate
[1])) / scale
[1]);
536 radeon_emit(cs
, S_028250_TL_X(scissor
.offset
.x
) |
537 S_028250_TL_Y(scissor
.offset
.y
) |
538 S_028250_WINDOW_OFFSET_DISABLE(1));
539 radeon_emit(cs
, S_028254_BR_X(scissor
.offset
.x
+ scissor
.extent
.width
) |
540 S_028254_BR_Y(scissor
.offset
.y
+ scissor
.extent
.height
));
542 if (!can_use_guardband
) {
547 radeon_set_context_reg_seq(cs
, R_028BE8_PA_CL_GB_VERT_CLIP_ADJ
, 4);
548 radeon_emit(cs
, fui(guardband_y
));
549 radeon_emit(cs
, fui(1.0));
550 radeon_emit(cs
, fui(guardband_x
));
551 radeon_emit(cs
, fui(1.0));
554 static inline unsigned
555 radv_prims_for_vertices(struct radv_prim_vertex_count
*info
, unsigned num
)
566 return 1 + ((num
- info
->min
) / info
->incr
);
570 si_get_ia_multi_vgt_param(struct radv_cmd_buffer
*cmd_buffer
,
571 bool instanced_draw
, bool indirect_draw
,
572 bool count_from_stream_output
,
573 uint32_t draw_vertex_count
)
575 enum chip_class chip_class
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
;
576 enum radeon_family family
= cmd_buffer
->device
->physical_device
->rad_info
.family
;
577 struct radeon_info
*info
= &cmd_buffer
->device
->physical_device
->rad_info
;
578 const unsigned max_primgroup_in_wave
= 2;
579 /* SWITCH_ON_EOP(0) is always preferable. */
580 bool wd_switch_on_eop
= false;
581 bool ia_switch_on_eop
= false;
582 bool ia_switch_on_eoi
= false;
583 bool partial_vs_wave
= false;
584 bool partial_es_wave
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.partial_es_wave
;
585 bool multi_instances_smaller_than_primgroup
;
587 multi_instances_smaller_than_primgroup
= indirect_draw
;
588 if (!multi_instances_smaller_than_primgroup
&& instanced_draw
) {
589 uint32_t num_prims
= radv_prims_for_vertices(&cmd_buffer
->state
.pipeline
->graphics
.prim_vertex_count
, draw_vertex_count
);
590 if (num_prims
< cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.primgroup_size
)
591 multi_instances_smaller_than_primgroup
= true;
594 ia_switch_on_eoi
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.ia_switch_on_eoi
;
595 partial_vs_wave
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.partial_vs_wave
;
597 if (chip_class
>= GFX7
) {
598 wd_switch_on_eop
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.wd_switch_on_eop
;
600 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
601 * We don't know that for indirect drawing, so treat it as
602 * always problematic. */
603 if (family
== CHIP_HAWAII
&&
604 (instanced_draw
|| indirect_draw
))
605 wd_switch_on_eop
= true;
607 /* Performance recommendation for 4 SE Gfx7-8 parts if
608 * instances are smaller than a primgroup.
609 * Assume indirect draws always use small instances.
610 * This is needed for good VS wave utilization.
612 if (chip_class
<= GFX8
&&
614 multi_instances_smaller_than_primgroup
)
615 wd_switch_on_eop
= true;
617 /* Required on GFX7 and later. */
618 if (info
->max_se
> 2 && !wd_switch_on_eop
)
619 ia_switch_on_eoi
= true;
621 /* Required by Hawaii and, for some special cases, by GFX8. */
622 if (ia_switch_on_eoi
&&
623 (family
== CHIP_HAWAII
||
624 (chip_class
== GFX8
&&
625 /* max primgroup in wave is always 2 - leave this for documentation */
626 (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
) || max_primgroup_in_wave
!= 2))))
627 partial_vs_wave
= true;
629 /* Instancing bug on Bonaire. */
630 if (family
== CHIP_BONAIRE
&& ia_switch_on_eoi
&&
631 (instanced_draw
|| indirect_draw
))
632 partial_vs_wave
= true;
634 /* Hardware requirement when drawing primitives from a stream
637 if (count_from_stream_output
)
638 wd_switch_on_eop
= true;
640 /* If the WD switch is false, the IA switch must be false too. */
641 assert(wd_switch_on_eop
|| !ia_switch_on_eop
);
643 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
644 if (chip_class
<= GFX8
&& ia_switch_on_eoi
)
645 partial_es_wave
= true;
647 if (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
)) {
648 /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
649 * The hw doc says all multi-SE chips are affected, but amdgpu-pro Vulkan
650 * only applies it to Hawaii. Do what amdgpu-pro Vulkan does.
652 if (family
== CHIP_HAWAII
&& ia_switch_on_eoi
) {
653 bool set_vgt_flush
= indirect_draw
;
654 if (!set_vgt_flush
&& instanced_draw
) {
655 uint32_t num_prims
= radv_prims_for_vertices(&cmd_buffer
->state
.pipeline
->graphics
.prim_vertex_count
, draw_vertex_count
);
657 set_vgt_flush
= true;
660 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_VGT_FLUSH
;
664 return cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.base
|
665 S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop
) |
666 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi
) |
667 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave
) |
668 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave
) |
669 S_028AA8_WD_SWITCH_ON_EOP(chip_class
>= GFX7
? wd_switch_on_eop
: 0);
673 void si_cs_emit_write_event_eop(struct radeon_cmdbuf
*cs
,
674 enum chip_class chip_class
,
676 unsigned event
, unsigned event_flags
,
680 uint64_t gfx9_eop_bug_va
)
682 unsigned op
= EVENT_TYPE(event
) |
685 unsigned is_gfx8_mec
= is_mec
&& chip_class
< GFX9
;
686 unsigned sel
= EOP_DATA_SEL(data_sel
);
688 /* Wait for write confirmation before writing data, but don't send
690 if (data_sel
!= EOP_DATA_SEL_DISCARD
)
691 sel
|= EOP_INT_SEL(EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM
);
693 if (chip_class
>= GFX9
|| is_gfx8_mec
) {
694 /* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
695 * counters) must immediately precede every timestamp event to
696 * prevent a GPU hang on GFX9.
698 if (chip_class
== GFX9
&& !is_mec
) {
699 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 2, 0));
700 radeon_emit(cs
, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE
) | EVENT_INDEX(1));
701 radeon_emit(cs
, gfx9_eop_bug_va
);
702 radeon_emit(cs
, gfx9_eop_bug_va
>> 32);
705 radeon_emit(cs
, PKT3(PKT3_RELEASE_MEM
, is_gfx8_mec
? 5 : 6, false));
707 radeon_emit(cs
, sel
);
708 radeon_emit(cs
, va
); /* address lo */
709 radeon_emit(cs
, va
>> 32); /* address hi */
710 radeon_emit(cs
, new_fence
); /* immediate data lo */
711 radeon_emit(cs
, 0); /* immediate data hi */
713 radeon_emit(cs
, 0); /* unused */
715 if (chip_class
== GFX7
||
716 chip_class
== GFX8
) {
717 /* Two EOP events are required to make all engines go idle
718 * (and optional cache flushes executed) before the timestamp
721 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, false));
724 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
725 radeon_emit(cs
, 0); /* immediate data */
726 radeon_emit(cs
, 0); /* unused */
729 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, false));
732 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
733 radeon_emit(cs
, new_fence
); /* immediate data */
734 radeon_emit(cs
, 0); /* unused */
739 radv_cp_wait_mem(struct radeon_cmdbuf
*cs
, uint32_t op
, uint64_t va
,
740 uint32_t ref
, uint32_t mask
)
742 assert(op
== WAIT_REG_MEM_EQUAL
||
743 op
== WAIT_REG_MEM_NOT_EQUAL
||
744 op
== WAIT_REG_MEM_GREATER_OR_EQUAL
);
746 radeon_emit(cs
, PKT3(PKT3_WAIT_REG_MEM
, 5, false));
747 radeon_emit(cs
, op
| WAIT_REG_MEM_MEM_SPACE(1));
749 radeon_emit(cs
, va
>> 32);
750 radeon_emit(cs
, ref
); /* reference value */
751 radeon_emit(cs
, mask
); /* mask */
752 radeon_emit(cs
, 4); /* poll interval */
756 si_emit_acquire_mem(struct radeon_cmdbuf
*cs
,
759 unsigned cp_coher_cntl
)
761 if (is_mec
|| is_gfx9
) {
762 uint32_t hi_val
= is_gfx9
? 0xffffff : 0xff;
763 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 5, false) |
764 PKT3_SHADER_TYPE_S(is_mec
));
765 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
766 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
767 radeon_emit(cs
, hi_val
); /* CP_COHER_SIZE_HI */
768 radeon_emit(cs
, 0); /* CP_COHER_BASE */
769 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
770 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
772 /* ACQUIRE_MEM is only required on a compute ring. */
773 radeon_emit(cs
, PKT3(PKT3_SURFACE_SYNC
, 3, false));
774 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
775 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
776 radeon_emit(cs
, 0); /* CP_COHER_BASE */
777 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
782 si_cs_emit_cache_flush(struct radeon_cmdbuf
*cs
,
783 enum chip_class chip_class
,
787 enum radv_cmd_flush_bits flush_bits
,
788 uint64_t gfx9_eop_bug_va
)
790 unsigned cp_coher_cntl
= 0;
791 uint32_t flush_cb_db
= flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
792 RADV_CMD_FLAG_FLUSH_AND_INV_DB
);
794 if (flush_bits
& RADV_CMD_FLAG_INV_ICACHE
)
795 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
796 if (flush_bits
& RADV_CMD_FLAG_INV_SCACHE
)
797 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
799 if (chip_class
<= GFX8
) {
800 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB
) {
801 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
802 S_0085F0_CB0_DEST_BASE_ENA(1) |
803 S_0085F0_CB1_DEST_BASE_ENA(1) |
804 S_0085F0_CB2_DEST_BASE_ENA(1) |
805 S_0085F0_CB3_DEST_BASE_ENA(1) |
806 S_0085F0_CB4_DEST_BASE_ENA(1) |
807 S_0085F0_CB5_DEST_BASE_ENA(1) |
808 S_0085F0_CB6_DEST_BASE_ENA(1) |
809 S_0085F0_CB7_DEST_BASE_ENA(1);
811 /* Necessary for DCC */
812 if (chip_class
>= GFX8
) {
813 si_cs_emit_write_event_eop(cs
,
816 V_028A90_FLUSH_AND_INV_CB_DATA_TS
,
818 EOP_DATA_SEL_DISCARD
,
823 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB
) {
824 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
825 S_0085F0_DB_DEST_BASE_ENA(1);
829 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
) {
830 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
831 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
834 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
) {
835 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
836 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
839 if (flush_bits
& RADV_CMD_FLAG_PS_PARTIAL_FLUSH
) {
840 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
841 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
842 } else if (flush_bits
& RADV_CMD_FLAG_VS_PARTIAL_FLUSH
) {
843 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
844 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
847 if (flush_bits
& RADV_CMD_FLAG_CS_PARTIAL_FLUSH
) {
848 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
849 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
852 if (chip_class
>= GFX9
&& flush_cb_db
) {
853 unsigned cb_db_event
, tc_flags
;
855 /* Set the CB/DB flush event. */
856 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
858 /* These are the only allowed combinations. If you need to
859 * do multiple operations at once, do them separately.
860 * All operations that invalidate L2 also seem to invalidate
861 * metadata. Volatile (VOL) and WC flushes are not listed here.
863 * TC | TC_WB = writeback & invalidate L2 & L1
864 * TC | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
865 * TC_WB | TC_NC = writeback L2 for MTYPE == NC
866 * TC | TC_NC = invalidate L2 for MTYPE == NC
867 * TC | TC_MD = writeback & invalidate L2 metadata (DCC, etc.)
868 * TCL1 = invalidate L1
870 tc_flags
= EVENT_TC_ACTION_ENA
|
871 EVENT_TC_MD_ACTION_ENA
;
873 /* Ideally flush TC together with CB/DB. */
874 if (flush_bits
& RADV_CMD_FLAG_INV_L2
) {
875 /* Writeback and invalidate everything in L2 & L1. */
876 tc_flags
= EVENT_TC_ACTION_ENA
|
877 EVENT_TC_WB_ACTION_ENA
;
880 /* Clear the flags. */
881 flush_bits
&= ~(RADV_CMD_FLAG_INV_L2
|
882 RADV_CMD_FLAG_WB_L2
|
883 RADV_CMD_FLAG_INV_VCACHE
);
888 si_cs_emit_write_event_eop(cs
, chip_class
, false, cb_db_event
, tc_flags
,
889 EOP_DATA_SEL_VALUE_32BIT
,
890 flush_va
, *flush_cnt
,
892 radv_cp_wait_mem(cs
, WAIT_REG_MEM_EQUAL
, flush_va
,
893 *flush_cnt
, 0xffffffff);
897 if (flush_bits
& RADV_CMD_FLAG_VGT_FLUSH
) {
898 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
899 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
902 /* VGT streamout state sync */
903 if (flush_bits
& RADV_CMD_FLAG_VGT_STREAMOUT_SYNC
) {
904 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
905 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC
) | EVENT_INDEX(0));
908 /* Make sure ME is idle (it executes most packets) before continuing.
909 * This prevents read-after-write hazards between PFP and ME.
911 if ((cp_coher_cntl
||
912 (flush_bits
& (RADV_CMD_FLAG_CS_PARTIAL_FLUSH
|
913 RADV_CMD_FLAG_INV_VCACHE
|
914 RADV_CMD_FLAG_INV_L2
|
915 RADV_CMD_FLAG_WB_L2
))) &&
917 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
921 if ((flush_bits
& RADV_CMD_FLAG_INV_L2
) ||
922 (chip_class
<= GFX7
&& (flush_bits
& RADV_CMD_FLAG_WB_L2
))) {
923 si_emit_acquire_mem(cs
, is_mec
, chip_class
>= GFX9
,
925 S_0085F0_TC_ACTION_ENA(1) |
926 S_0085F0_TCL1_ACTION_ENA(1) |
927 S_0301F0_TC_WB_ACTION_ENA(chip_class
>= GFX8
));
930 if(flush_bits
& RADV_CMD_FLAG_WB_L2
) {
932 * NC = apply to non-coherent MTYPEs
933 * (i.e. MTYPE <= 1, which is what we use everywhere)
935 * WB doesn't work without NC.
937 si_emit_acquire_mem(cs
, is_mec
,
940 S_0301F0_TC_WB_ACTION_ENA(1) |
941 S_0301F0_TC_NC_ACTION_ENA(1));
944 if (flush_bits
& RADV_CMD_FLAG_INV_VCACHE
) {
945 si_emit_acquire_mem(cs
, is_mec
,
948 S_0085F0_TCL1_ACTION_ENA(1));
953 /* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
954 * Therefore, it should be last. Done in PFP.
957 si_emit_acquire_mem(cs
, is_mec
, chip_class
>= GFX9
, cp_coher_cntl
);
959 if (flush_bits
& RADV_CMD_FLAG_START_PIPELINE_STATS
) {
960 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
961 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
963 } else if (flush_bits
& RADV_CMD_FLAG_STOP_PIPELINE_STATS
) {
964 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
965 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
971 si_emit_cache_flush(struct radv_cmd_buffer
*cmd_buffer
)
973 bool is_compute
= cmd_buffer
->queue_family_index
== RADV_QUEUE_COMPUTE
;
976 cmd_buffer
->state
.flush_bits
&= ~(RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
977 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
|
978 RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
979 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
|
980 RADV_CMD_FLAG_PS_PARTIAL_FLUSH
|
981 RADV_CMD_FLAG_VS_PARTIAL_FLUSH
|
982 RADV_CMD_FLAG_VGT_FLUSH
|
983 RADV_CMD_FLAG_START_PIPELINE_STATS
|
984 RADV_CMD_FLAG_STOP_PIPELINE_STATS
);
986 if (!cmd_buffer
->state
.flush_bits
)
989 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 128);
991 si_cs_emit_cache_flush(cmd_buffer
->cs
,
992 cmd_buffer
->device
->physical_device
->rad_info
.chip_class
,
993 &cmd_buffer
->gfx9_fence_idx
,
994 cmd_buffer
->gfx9_fence_va
,
995 radv_cmd_buffer_uses_mec(cmd_buffer
),
996 cmd_buffer
->state
.flush_bits
,
997 cmd_buffer
->gfx9_eop_bug_va
);
1000 if (unlikely(cmd_buffer
->device
->trace_bo
))
1001 radv_cmd_buffer_trace_emit(cmd_buffer
);
1003 /* Clear the caches that have been flushed to avoid syncing too much
1004 * when there is some pending active queries.
1006 cmd_buffer
->active_query_flush_bits
&= ~cmd_buffer
->state
.flush_bits
;
1008 cmd_buffer
->state
.flush_bits
= 0;
1010 /* If the driver used a compute shader for resetting a query pool, it
1011 * should be finished at this point.
1013 cmd_buffer
->pending_reset_query
= false;
1016 /* sets the CP predication state using a boolean stored at va */
1018 si_emit_set_predication_state(struct radv_cmd_buffer
*cmd_buffer
,
1019 bool draw_visible
, uint64_t va
)
1024 op
= PRED_OP(PREDICATION_OP_BOOL64
);
1026 /* PREDICATION_DRAW_VISIBLE means that if the 32-bit value is
1027 * zero, all rendering commands are discarded. Otherwise, they
1028 * are discarded if the value is non zero.
1030 op
|= draw_visible
? PREDICATION_DRAW_VISIBLE
:
1031 PREDICATION_DRAW_NOT_VISIBLE
;
1033 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1034 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_SET_PREDICATION
, 2, 0));
1035 radeon_emit(cmd_buffer
->cs
, op
);
1036 radeon_emit(cmd_buffer
->cs
, va
);
1037 radeon_emit(cmd_buffer
->cs
, va
>> 32);
1039 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_SET_PREDICATION
, 1, 0));
1040 radeon_emit(cmd_buffer
->cs
, va
);
1041 radeon_emit(cmd_buffer
->cs
, op
| ((va
>> 32) & 0xFF));
1045 /* Set this if you want the 3D engine to wait until CP DMA is done.
1046 * It should be set on the last CP DMA packet. */
1047 #define CP_DMA_SYNC (1 << 0)
1049 /* Set this if the source data was used as a destination in a previous CP DMA
1050 * packet. It's for preventing a read-after-write (RAW) hazard between two
1051 * CP DMA packets. */
1052 #define CP_DMA_RAW_WAIT (1 << 1)
1053 #define CP_DMA_USE_L2 (1 << 2)
1054 #define CP_DMA_CLEAR (1 << 3)
1056 /* Alignment for optimal performance. */
1057 #define SI_CPDMA_ALIGNMENT 32
1059 /* The max number of bytes that can be copied per packet. */
1060 static inline unsigned cp_dma_max_byte_count(struct radv_cmd_buffer
*cmd_buffer
)
1062 unsigned max
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
?
1063 S_414_BYTE_COUNT_GFX9(~0u) :
1064 S_414_BYTE_COUNT_GFX6(~0u);
1066 /* make it aligned for optimal performance */
1067 return max
& ~(SI_CPDMA_ALIGNMENT
- 1);
1070 /* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
1071 * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
1074 static void si_emit_cp_dma(struct radv_cmd_buffer
*cmd_buffer
,
1075 uint64_t dst_va
, uint64_t src_va
,
1076 unsigned size
, unsigned flags
)
1078 struct radeon_cmdbuf
*cs
= cmd_buffer
->cs
;
1079 uint32_t header
= 0, command
= 0;
1081 assert(size
<= cp_dma_max_byte_count(cmd_buffer
));
1083 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 9);
1084 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1085 command
|= S_414_BYTE_COUNT_GFX9(size
);
1087 command
|= S_414_BYTE_COUNT_GFX6(size
);
1090 if (flags
& CP_DMA_SYNC
)
1091 header
|= S_411_CP_SYNC(1);
1093 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1094 command
|= S_414_DISABLE_WR_CONFIRM_GFX9(1);
1096 command
|= S_414_DISABLE_WR_CONFIRM_GFX6(1);
1099 if (flags
& CP_DMA_RAW_WAIT
)
1100 command
|= S_414_RAW_WAIT(1);
1102 /* Src and dst flags. */
1103 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1104 !(flags
& CP_DMA_CLEAR
) &&
1106 header
|= S_411_DST_SEL(V_411_NOWHERE
); /* prefetch only */
1107 else if (flags
& CP_DMA_USE_L2
)
1108 header
|= S_411_DST_SEL(V_411_DST_ADDR_TC_L2
);
1110 if (flags
& CP_DMA_CLEAR
)
1111 header
|= S_411_SRC_SEL(V_411_DATA
);
1112 else if (flags
& CP_DMA_USE_L2
)
1113 header
|= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2
);
1115 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
1116 radeon_emit(cs
, PKT3(PKT3_DMA_DATA
, 5, cmd_buffer
->state
.predicating
));
1117 radeon_emit(cs
, header
);
1118 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
1119 radeon_emit(cs
, src_va
>> 32); /* SRC_ADDR_HI [31:0] */
1120 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
1121 radeon_emit(cs
, dst_va
>> 32); /* DST_ADDR_HI [31:0] */
1122 radeon_emit(cs
, command
);
1124 assert(!(flags
& CP_DMA_USE_L2
));
1125 header
|= S_411_SRC_ADDR_HI(src_va
>> 32);
1126 radeon_emit(cs
, PKT3(PKT3_CP_DMA
, 4, cmd_buffer
->state
.predicating
));
1127 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
1128 radeon_emit(cs
, header
); /* SRC_ADDR_HI [15:0] + flags. */
1129 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
1130 radeon_emit(cs
, (dst_va
>> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
1131 radeon_emit(cs
, command
);
1134 /* CP DMA is executed in ME, but index buffers are read by PFP.
1135 * This ensures that ME (CP DMA) is idle before PFP starts fetching
1136 * indices. If we wanted to execute CP DMA in PFP, this packet
1137 * should precede it.
1139 if (flags
& CP_DMA_SYNC
) {
1140 if (cmd_buffer
->queue_family_index
== RADV_QUEUE_GENERAL
) {
1141 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, cmd_buffer
->state
.predicating
));
1145 /* CP will see the sync flag and wait for all DMAs to complete. */
1146 cmd_buffer
->state
.dma_is_busy
= false;
1149 if (unlikely(cmd_buffer
->device
->trace_bo
))
1150 radv_cmd_buffer_trace_emit(cmd_buffer
);
1153 void si_cp_dma_prefetch(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1156 uint64_t aligned_va
= va
& ~(SI_CPDMA_ALIGNMENT
- 1);
1157 uint64_t aligned_size
= ((va
+ size
+ SI_CPDMA_ALIGNMENT
-1) & ~(SI_CPDMA_ALIGNMENT
- 1)) - aligned_va
;
1159 si_emit_cp_dma(cmd_buffer
, aligned_va
, aligned_va
,
1160 aligned_size
, CP_DMA_USE_L2
);
1163 static void si_cp_dma_prepare(struct radv_cmd_buffer
*cmd_buffer
, uint64_t byte_count
,
1164 uint64_t remaining_size
, unsigned *flags
)
1167 /* Flush the caches for the first copy only.
1168 * Also wait for the previous CP DMA operations.
1170 if (cmd_buffer
->state
.flush_bits
) {
1171 si_emit_cache_flush(cmd_buffer
);
1172 *flags
|= CP_DMA_RAW_WAIT
;
1175 /* Do the synchronization after the last dma, so that all data
1176 * is written to memory.
1178 if (byte_count
== remaining_size
)
1179 *flags
|= CP_DMA_SYNC
;
1182 static void si_cp_dma_realign_engine(struct radv_cmd_buffer
*cmd_buffer
, unsigned size
)
1186 unsigned dma_flags
= 0;
1187 unsigned buf_size
= SI_CPDMA_ALIGNMENT
* 2;
1190 assert(size
< SI_CPDMA_ALIGNMENT
);
1192 radv_cmd_buffer_upload_alloc(cmd_buffer
, buf_size
, SI_CPDMA_ALIGNMENT
, &offset
, &ptr
);
1194 va
= radv_buffer_get_va(cmd_buffer
->upload
.upload_bo
);
1197 si_cp_dma_prepare(cmd_buffer
, size
, size
, &dma_flags
);
1199 si_emit_cp_dma(cmd_buffer
, va
, va
+ SI_CPDMA_ALIGNMENT
, size
,
1203 void si_cp_dma_buffer_copy(struct radv_cmd_buffer
*cmd_buffer
,
1204 uint64_t src_va
, uint64_t dest_va
,
1207 uint64_t main_src_va
, main_dest_va
;
1208 uint64_t skipped_size
= 0, realign_size
= 0;
1210 /* Assume that we are not going to sync after the last DMA operation. */
1211 cmd_buffer
->state
.dma_is_busy
= true;
1213 if (cmd_buffer
->device
->physical_device
->rad_info
.family
<= CHIP_CARRIZO
||
1214 cmd_buffer
->device
->physical_device
->rad_info
.family
== CHIP_STONEY
) {
1215 /* If the size is not aligned, we must add a dummy copy at the end
1216 * just to align the internal counter. Otherwise, the DMA engine
1217 * would slow down by an order of magnitude for following copies.
1219 if (size
% SI_CPDMA_ALIGNMENT
)
1220 realign_size
= SI_CPDMA_ALIGNMENT
- (size
% SI_CPDMA_ALIGNMENT
);
1222 /* If the copy begins unaligned, we must start copying from the next
1223 * aligned block and the skipped part should be copied after everything
1224 * else has been copied. Only the src alignment matters, not dst.
1226 if (src_va
% SI_CPDMA_ALIGNMENT
) {
1227 skipped_size
= SI_CPDMA_ALIGNMENT
- (src_va
% SI_CPDMA_ALIGNMENT
);
1228 /* The main part will be skipped if the size is too small. */
1229 skipped_size
= MIN2(skipped_size
, size
);
1230 size
-= skipped_size
;
1233 main_src_va
= src_va
+ skipped_size
;
1234 main_dest_va
= dest_va
+ skipped_size
;
1237 unsigned dma_flags
= 0;
1238 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(cmd_buffer
));
1240 si_cp_dma_prepare(cmd_buffer
, byte_count
,
1241 size
+ skipped_size
+ realign_size
,
1244 dma_flags
&= ~CP_DMA_SYNC
;
1246 si_emit_cp_dma(cmd_buffer
, main_dest_va
, main_src_va
,
1247 byte_count
, dma_flags
);
1250 main_src_va
+= byte_count
;
1251 main_dest_va
+= byte_count
;
1255 unsigned dma_flags
= 0;
1257 si_cp_dma_prepare(cmd_buffer
, skipped_size
,
1258 size
+ skipped_size
+ realign_size
,
1261 si_emit_cp_dma(cmd_buffer
, dest_va
, src_va
,
1262 skipped_size
, dma_flags
);
1265 si_cp_dma_realign_engine(cmd_buffer
, realign_size
);
1268 void si_cp_dma_clear_buffer(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1269 uint64_t size
, unsigned value
)
1275 assert(va
% 4 == 0 && size
% 4 == 0);
1277 /* Assume that we are not going to sync after the last DMA operation. */
1278 cmd_buffer
->state
.dma_is_busy
= true;
1281 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(cmd_buffer
));
1282 unsigned dma_flags
= CP_DMA_CLEAR
;
1284 si_cp_dma_prepare(cmd_buffer
, byte_count
, size
, &dma_flags
);
1286 /* Emit the clear packet. */
1287 si_emit_cp_dma(cmd_buffer
, va
, value
, byte_count
,
1295 void si_cp_dma_wait_for_idle(struct radv_cmd_buffer
*cmd_buffer
)
1297 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
< GFX7
)
1300 if (!cmd_buffer
->state
.dma_is_busy
)
1303 /* Issue a dummy DMA that copies zero bytes.
1305 * The DMA engine will see that there's no work to do and skip this
1306 * DMA request, however, the CP will see the sync flag and still wait
1307 * for all DMAs to complete.
1309 si_emit_cp_dma(cmd_buffer
, 0, 0, 0, CP_DMA_SYNC
);
1311 cmd_buffer
->state
.dma_is_busy
= false;
1314 /* For MSAA sample positions. */
1315 #define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y) \
1316 ((((unsigned)(s0x) & 0xf) << 0) | (((unsigned)(s0y) & 0xf) << 4) | \
1317 (((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) | \
1318 (((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) | \
1319 (((unsigned)(s3x) & 0xf) << 24) | (((unsigned)(s3y) & 0xf) << 28))
1321 /* For obtaining location coordinates from registers */
1322 #define SEXT4(x) ((int)((x) | ((x) & 0x8 ? 0xfffffff0 : 0)))
1323 #define GET_SFIELD(reg, index) SEXT4(((reg) >> ((index) * 4)) & 0xf)
1324 #define GET_SX(reg, index) GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2)
1325 #define GET_SY(reg, index) GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2 + 1)
1328 static const uint32_t sample_locs_1x
=
1329 FILL_SREG(0, 0, 0, 0, 0, 0, 0, 0);
1330 static const unsigned max_dist_1x
= 0;
1331 static const uint64_t centroid_priority_1x
= 0x0000000000000000ull
;
1334 static const uint32_t sample_locs_2x
=
1335 FILL_SREG(4,4, -4, -4, 0, 0, 0, 0);
1336 static const unsigned max_dist_2x
= 4;
1337 static const uint64_t centroid_priority_2x
= 0x1010101010101010ull
;
1340 static const uint32_t sample_locs_4x
=
1341 FILL_SREG(-2,-6, 6, -2, -6, 2, 2, 6);
1342 static const unsigned max_dist_4x
= 6;
1343 static const uint64_t centroid_priority_4x
= 0x3210321032103210ull
;
1346 static const uint32_t sample_locs_8x
[] = {
1347 FILL_SREG( 1,-3, -1, 3, 5, 1, -3,-5),
1348 FILL_SREG(-5, 5, -7,-1, 3, 7, 7,-7),
1349 /* The following are unused by hardware, but we emit them to IBs
1350 * instead of multiple SET_CONTEXT_REG packets. */
1354 static const unsigned max_dist_8x
= 7;
1355 static const uint64_t centroid_priority_8x
= 0x7654321076543210ull
;
1357 unsigned radv_get_default_max_sample_dist(int log_samples
)
1359 unsigned max_dist
[] = {
1365 return max_dist
[log_samples
];
1368 void radv_emit_default_sample_locations(struct radeon_cmdbuf
*cs
, int nr_samples
)
1370 switch (nr_samples
) {
1373 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1374 radeon_emit(cs
, (uint32_t)centroid_priority_1x
);
1375 radeon_emit(cs
, centroid_priority_1x
>> 32);
1376 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, sample_locs_1x
);
1377 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, sample_locs_1x
);
1378 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, sample_locs_1x
);
1379 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, sample_locs_1x
);
1382 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1383 radeon_emit(cs
, (uint32_t)centroid_priority_2x
);
1384 radeon_emit(cs
, centroid_priority_2x
>> 32);
1385 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, sample_locs_2x
);
1386 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, sample_locs_2x
);
1387 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, sample_locs_2x
);
1388 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, sample_locs_2x
);
1391 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1392 radeon_emit(cs
, (uint32_t)centroid_priority_4x
);
1393 radeon_emit(cs
, centroid_priority_4x
>> 32);
1394 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, sample_locs_4x
);
1395 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, sample_locs_4x
);
1396 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, sample_locs_4x
);
1397 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, sample_locs_4x
);
1400 radeon_set_context_reg_seq(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 2);
1401 radeon_emit(cs
, (uint32_t)centroid_priority_8x
);
1402 radeon_emit(cs
, centroid_priority_8x
>> 32);
1403 radeon_set_context_reg_seq(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 14);
1404 radeon_emit_array(cs
, sample_locs_8x
, 4);
1405 radeon_emit_array(cs
, sample_locs_8x
, 4);
1406 radeon_emit_array(cs
, sample_locs_8x
, 4);
1407 radeon_emit_array(cs
, sample_locs_8x
, 2);
1412 static void radv_get_sample_position(struct radv_device
*device
,
1413 unsigned sample_count
,
1414 unsigned sample_index
, float *out_value
)
1416 const uint32_t *sample_locs
;
1418 switch (sample_count
) {
1421 sample_locs
= &sample_locs_1x
;
1424 sample_locs
= &sample_locs_2x
;
1427 sample_locs
= &sample_locs_4x
;
1430 sample_locs
= sample_locs_8x
;
1434 out_value
[0] = (GET_SX(sample_locs
, sample_index
) + 8) / 16.0f
;
1435 out_value
[1] = (GET_SY(sample_locs
, sample_index
) + 8) / 16.0f
;
1438 void radv_device_init_msaa(struct radv_device
*device
)
1442 radv_get_sample_position(device
, 1, 0, device
->sample_locations_1x
[0]);
1444 for (i
= 0; i
< 2; i
++)
1445 radv_get_sample_position(device
, 2, i
, device
->sample_locations_2x
[i
]);
1446 for (i
= 0; i
< 4; i
++)
1447 radv_get_sample_position(device
, 4, i
, device
->sample_locations_4x
[i
]);
1448 for (i
= 0; i
< 8; i
++)
1449 radv_get_sample_position(device
, 8, i
, device
->sample_locations_8x
[i
]);