2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
6 * Copyright © 2015 Advanced Micro Devices, Inc.
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 /* command buffer handling for SI */
30 #include "radv_private.h"
31 #include "radv_shader.h"
35 #include "radv_util.h"
36 #include "main/macros.h"
39 si_write_harvested_raster_configs(struct radv_physical_device
*physical_device
,
40 struct radeon_cmdbuf
*cs
,
41 unsigned raster_config
,
42 unsigned raster_config_1
)
44 unsigned num_se
= MAX2(physical_device
->rad_info
.max_se
, 1);
45 unsigned raster_config_se
[4];
48 ac_get_harvested_configs(&physical_device
->rad_info
,
53 for (se
= 0; se
< num_se
; se
++) {
54 /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
55 if (physical_device
->rad_info
.chip_class
< CIK
)
56 radeon_set_config_reg(cs
, R_00802C_GRBM_GFX_INDEX
,
57 S_00802C_SE_INDEX(se
) |
58 S_00802C_SH_BROADCAST_WRITES(1) |
59 S_00802C_INSTANCE_BROADCAST_WRITES(1));
61 radeon_set_uconfig_reg(cs
, R_030800_GRBM_GFX_INDEX
,
62 S_030800_SE_INDEX(se
) | S_030800_SH_BROADCAST_WRITES(1) |
63 S_030800_INSTANCE_BROADCAST_WRITES(1));
64 radeon_set_context_reg(cs
, R_028350_PA_SC_RASTER_CONFIG
, raster_config_se
[se
]);
67 /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
68 if (physical_device
->rad_info
.chip_class
< CIK
)
69 radeon_set_config_reg(cs
, R_00802C_GRBM_GFX_INDEX
,
70 S_00802C_SE_BROADCAST_WRITES(1) |
71 S_00802C_SH_BROADCAST_WRITES(1) |
72 S_00802C_INSTANCE_BROADCAST_WRITES(1));
74 radeon_set_uconfig_reg(cs
, R_030800_GRBM_GFX_INDEX
,
75 S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
76 S_030800_INSTANCE_BROADCAST_WRITES(1));
78 if (physical_device
->rad_info
.chip_class
>= CIK
)
79 radeon_set_context_reg(cs
, R_028354_PA_SC_RASTER_CONFIG_1
, raster_config_1
);
83 si_emit_compute(struct radv_physical_device
*physical_device
,
84 struct radeon_cmdbuf
*cs
)
86 radeon_set_sh_reg_seq(cs
, R_00B810_COMPUTE_START_X
, 3);
91 radeon_set_sh_reg_seq(cs
, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0
, 2);
92 /* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1 */
93 radeon_emit(cs
, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
94 radeon_emit(cs
, S_00B85C_SH0_CU_EN(0xffff) | S_00B85C_SH1_CU_EN(0xffff));
96 if (physical_device
->rad_info
.chip_class
>= CIK
) {
97 /* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
98 radeon_set_sh_reg_seq(cs
,
99 R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2
, 2);
100 radeon_emit(cs
, S_00B864_SH0_CU_EN(0xffff) |
101 S_00B864_SH1_CU_EN(0xffff));
102 radeon_emit(cs
, S_00B868_SH0_CU_EN(0xffff) |
103 S_00B868_SH1_CU_EN(0xffff));
106 /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
107 * and is now per pipe, so it should be handled in the
108 * kernel if we want to use something other than the default value,
109 * which is now 0x22f.
111 if (physical_device
->rad_info
.chip_class
<= SI
) {
112 /* XXX: This should be:
113 * (number of compute units) * 4 * (waves per simd) - 1 */
115 radeon_set_sh_reg(cs
, R_00B82C_COMPUTE_MAX_WAVE_ID
,
116 0x190 /* Default value */);
121 si_init_compute(struct radv_cmd_buffer
*cmd_buffer
)
123 struct radv_physical_device
*physical_device
= cmd_buffer
->device
->physical_device
;
124 si_emit_compute(physical_device
, cmd_buffer
->cs
);
127 /* 12.4 fixed-point */
128 static unsigned radv_pack_float_12p4(float x
)
131 x
>= 4096 ? 0xffff : x
* 16;
135 si_set_raster_config(struct radv_physical_device
*physical_device
,
136 struct radeon_cmdbuf
*cs
)
138 unsigned num_rb
= MIN2(physical_device
->rad_info
.num_render_backends
, 16);
139 unsigned rb_mask
= physical_device
->rad_info
.enabled_rb_mask
;
140 unsigned raster_config
, raster_config_1
;
142 ac_get_raster_config(&physical_device
->rad_info
,
146 /* Always use the default config when all backends are enabled
147 * (or when we failed to determine the enabled backends).
149 if (!rb_mask
|| util_bitcount(rb_mask
) >= num_rb
) {
150 radeon_set_context_reg(cs
, R_028350_PA_SC_RASTER_CONFIG
,
152 if (physical_device
->rad_info
.chip_class
>= CIK
)
153 radeon_set_context_reg(cs
, R_028354_PA_SC_RASTER_CONFIG_1
,
156 si_write_harvested_raster_configs(physical_device
, cs
,
163 si_emit_config(struct radv_physical_device
*physical_device
,
164 struct radeon_cmdbuf
*cs
)
168 /* Only SI can disable CLEAR_STATE for now. */
169 assert(physical_device
->has_clear_state
||
170 physical_device
->rad_info
.chip_class
== SI
);
172 radeon_emit(cs
, PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
173 radeon_emit(cs
, CONTEXT_CONTROL_LOAD_ENABLE(1));
174 radeon_emit(cs
, CONTEXT_CONTROL_SHADOW_ENABLE(1));
176 if (physical_device
->has_clear_state
) {
177 radeon_emit(cs
, PKT3(PKT3_CLEAR_STATE
, 0, 0));
181 if (physical_device
->rad_info
.chip_class
<= VI
)
182 si_set_raster_config(physical_device
, cs
);
184 radeon_set_context_reg(cs
, R_028A18_VGT_HOS_MAX_TESS_LEVEL
, fui(64));
185 if (!physical_device
->has_clear_state
)
186 radeon_set_context_reg(cs
, R_028A1C_VGT_HOS_MIN_TESS_LEVEL
, fui(0));
188 /* FIXME calculate these values somehow ??? */
189 if (physical_device
->rad_info
.chip_class
<= VI
) {
190 radeon_set_context_reg(cs
, R_028A54_VGT_GS_PER_ES
, SI_GS_PER_ES
);
191 radeon_set_context_reg(cs
, R_028A58_VGT_ES_PER_GS
, 0x40);
194 if (!physical_device
->has_clear_state
) {
195 radeon_set_context_reg(cs
, R_028A5C_VGT_GS_PER_VS
, 0x2);
196 radeon_set_context_reg(cs
, R_028A8C_VGT_PRIMITIVEID_RESET
, 0x0);
197 radeon_set_context_reg(cs
, R_028B98_VGT_STRMOUT_BUFFER_CONFIG
, 0x0);
200 radeon_set_context_reg(cs
, R_028AA0_VGT_INSTANCE_STEP_RATE_0
, 1);
201 if (!physical_device
->has_clear_state
)
202 radeon_set_context_reg(cs
, R_028AB8_VGT_VTX_CNT_EN
, 0x0);
203 if (physical_device
->rad_info
.chip_class
< CIK
)
204 radeon_set_config_reg(cs
, R_008A14_PA_CL_ENHANCE
, S_008A14_NUM_CLIP_SEQ(3) |
205 S_008A14_CLIP_VTX_REORDER_ENA(1));
207 radeon_set_context_reg(cs
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 0x76543210);
208 radeon_set_context_reg(cs
, R_028BD8_PA_SC_CENTROID_PRIORITY_1
, 0xfedcba98);
210 if (!physical_device
->has_clear_state
)
211 radeon_set_context_reg(cs
, R_02882C_PA_SU_PRIM_FILTER_CNTL
, 0);
213 /* CLEAR_STATE doesn't clear these correctly on certain generations.
214 * I don't know why. Deduced by trial and error.
216 if (physical_device
->rad_info
.chip_class
<= CIK
) {
217 radeon_set_context_reg(cs
, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET
, 0);
218 radeon_set_context_reg(cs
, R_028204_PA_SC_WINDOW_SCISSOR_TL
,
219 S_028204_WINDOW_OFFSET_DISABLE(1));
220 radeon_set_context_reg(cs
, R_028240_PA_SC_GENERIC_SCISSOR_TL
,
221 S_028240_WINDOW_OFFSET_DISABLE(1));
222 radeon_set_context_reg(cs
, R_028244_PA_SC_GENERIC_SCISSOR_BR
,
223 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
224 radeon_set_context_reg(cs
, R_028030_PA_SC_SCREEN_SCISSOR_TL
, 0);
225 radeon_set_context_reg(cs
, R_028034_PA_SC_SCREEN_SCISSOR_BR
,
226 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
229 if (!physical_device
->has_clear_state
) {
230 for (i
= 0; i
< 16; i
++) {
231 radeon_set_context_reg(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+ i
*8, 0);
232 radeon_set_context_reg(cs
, R_0282D4_PA_SC_VPORT_ZMAX_0
+ i
*8, fui(1.0));
236 if (!physical_device
->has_clear_state
) {
237 radeon_set_context_reg(cs
, R_02820C_PA_SC_CLIPRECT_RULE
, 0xFFFF);
238 radeon_set_context_reg(cs
, R_028230_PA_SC_EDGERULE
, 0xAAAAAAAA);
239 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
240 radeon_set_context_reg(cs
, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET
, 0);
241 radeon_set_context_reg(cs
, R_028820_PA_CL_NANINF_CNTL
, 0);
242 radeon_set_context_reg(cs
, R_028AC0_DB_SRESULTS_COMPARE_STATE0
, 0x0);
243 radeon_set_context_reg(cs
, R_028AC4_DB_SRESULTS_COMPARE_STATE1
, 0x0);
244 radeon_set_context_reg(cs
, R_028AC8_DB_PRELOAD_CONTROL
, 0x0);
247 radeon_set_context_reg(cs
, R_02800C_DB_RENDER_OVERRIDE
,
248 S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE
) |
249 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE
));
251 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
252 radeon_set_uconfig_reg(cs
, R_030920_VGT_MAX_VTX_INDX
, ~0);
253 radeon_set_uconfig_reg(cs
, R_030924_VGT_MIN_VTX_INDX
, 0);
254 radeon_set_uconfig_reg(cs
, R_030928_VGT_INDX_OFFSET
, 0);
256 /* These registers, when written, also overwrite the
257 * CLEAR_STATE context, so we can't rely on CLEAR_STATE setting
258 * them. It would be an issue if there was another UMD
261 radeon_set_context_reg(cs
, R_028400_VGT_MAX_VTX_INDX
, ~0);
262 radeon_set_context_reg(cs
, R_028404_VGT_MIN_VTX_INDX
, 0);
263 radeon_set_context_reg(cs
, R_028408_VGT_INDX_OFFSET
, 0);
266 if (physical_device
->rad_info
.chip_class
>= CIK
) {
267 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
268 radeon_set_sh_reg(cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
,
269 S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
271 radeon_set_sh_reg(cs
, R_00B51C_SPI_SHADER_PGM_RSRC3_LS
,
272 S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
273 radeon_set_sh_reg(cs
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
,
274 S_00B41C_WAVE_LIMIT(0x3F));
275 radeon_set_sh_reg(cs
, R_00B31C_SPI_SHADER_PGM_RSRC3_ES
,
276 S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
277 /* If this is 0, Bonaire can hang even if GS isn't being used.
278 * Other chips are unaffected. These are suboptimal values,
279 * but we don't use on-chip GS.
281 radeon_set_context_reg(cs
, R_028A44_VGT_GS_ONCHIP_CNTL
,
282 S_028A44_ES_VERTS_PER_SUBGRP(64) |
283 S_028A44_GS_PRIMS_PER_SUBGRP(4));
285 radeon_set_sh_reg(cs
, R_00B21C_SPI_SHADER_PGM_RSRC3_GS
,
286 S_00B21C_CU_EN(0xffff) | S_00B21C_WAVE_LIMIT(0x3F));
288 if (physical_device
->rad_info
.num_good_compute_units
/
289 (physical_device
->rad_info
.max_se
* physical_device
->rad_info
.max_sh_per_se
) <= 4) {
290 /* Too few available compute units per SH. Disallowing
291 * VS to run on CU0 could hurt us more than late VS
292 * allocation would help.
294 * LATE_ALLOC_VS = 2 is the highest safe number.
296 radeon_set_sh_reg(cs
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
,
297 S_00B118_CU_EN(0xffff) | S_00B118_WAVE_LIMIT(0x3F) );
298 radeon_set_sh_reg(cs
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
, S_00B11C_LIMIT(2));
300 /* Set LATE_ALLOC_VS == 31. It should be less than
301 * the number of scratch waves. Limitations:
302 * - VS can't execute on CU0.
303 * - If HS writes outputs to LDS, LS can't execute on CU0.
305 radeon_set_sh_reg(cs
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
,
306 S_00B118_CU_EN(0xfffe) | S_00B118_WAVE_LIMIT(0x3F));
307 radeon_set_sh_reg(cs
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
, S_00B11C_LIMIT(31));
310 radeon_set_sh_reg(cs
, R_00B01C_SPI_SHADER_PGM_RSRC3_PS
,
311 S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
314 if (physical_device
->rad_info
.chip_class
>= VI
) {
315 uint32_t vgt_tess_distribution
;
316 radeon_set_context_reg(cs
, R_028424_CB_DCC_CONTROL
,
317 S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
318 S_028424_OVERWRITE_COMBINER_WATERMARK(4));
320 vgt_tess_distribution
= S_028B50_ACCUM_ISOLINE(32) |
321 S_028B50_ACCUM_TRI(11) |
322 S_028B50_ACCUM_QUAD(11) |
323 S_028B50_DONUT_SPLIT(16);
325 if (physical_device
->rad_info
.family
== CHIP_FIJI
||
326 physical_device
->rad_info
.family
>= CHIP_POLARIS10
)
327 vgt_tess_distribution
|= S_028B50_TRAP_SPLIT(3);
329 radeon_set_context_reg(cs
, R_028B50_VGT_TESS_DISTRIBUTION
,
330 vgt_tess_distribution
);
331 } else if (!physical_device
->has_clear_state
) {
332 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 14);
333 radeon_set_context_reg(cs
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 16);
336 if (physical_device
->rad_info
.chip_class
>= GFX9
) {
337 unsigned num_se
= physical_device
->rad_info
.max_se
;
338 unsigned pc_lines
= 0;
340 switch (physical_device
->rad_info
.family
) {
352 radeon_set_context_reg(cs
, R_028C48_PA_SC_BINNER_CNTL_1
,
353 S_028C48_MAX_ALLOC_COUNT(MIN2(128, pc_lines
/ (4 * num_se
))) |
354 S_028C48_MAX_PRIM_PER_BATCH(1023));
355 radeon_set_context_reg(cs
, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL
,
356 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
357 radeon_set_uconfig_reg(cs
, R_030968_VGT_INSTANCE_BASE_ID
, 0);
360 unsigned tmp
= (unsigned)(1.0 * 8.0);
361 radeon_set_context_reg_seq(cs
, R_028A00_PA_SU_POINT_SIZE
, 1);
362 radeon_emit(cs
, S_028A00_HEIGHT(tmp
) | S_028A00_WIDTH(tmp
));
363 radeon_set_context_reg_seq(cs
, R_028A04_PA_SU_POINT_MINMAX
, 1);
364 radeon_emit(cs
, S_028A04_MIN_SIZE(radv_pack_float_12p4(0)) |
365 S_028A04_MAX_SIZE(radv_pack_float_12p4(8192/2)));
367 if (!physical_device
->has_clear_state
) {
368 radeon_set_context_reg(cs
, R_028004_DB_COUNT_CONTROL
,
369 S_028004_ZPASS_INCREMENT_DISABLE(1));
372 /* Enable the Polaris small primitive filter control.
373 * XXX: There is possibly an issue when MSAA is off (see RadeonSI
374 * has_msaa_sample_loc_bug). But this doesn't seem to regress anything,
375 * and AMDVLK doesn't have a workaround as well.
377 if (physical_device
->rad_info
.family
>= CHIP_POLARIS10
) {
378 unsigned small_prim_filter_cntl
=
379 S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
380 /* Workaround for a hw line bug. */
381 S_028830_LINE_FILTER_DISABLE(physical_device
->rad_info
.family
<= CHIP_POLARIS12
);
383 radeon_set_context_reg(cs
, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL
,
384 small_prim_filter_cntl
);
387 si_emit_compute(physical_device
, cs
);
390 void si_init_config(struct radv_cmd_buffer
*cmd_buffer
)
392 struct radv_physical_device
*physical_device
= cmd_buffer
->device
->physical_device
;
394 si_emit_config(physical_device
, cmd_buffer
->cs
);
398 cik_create_gfx_config(struct radv_device
*device
)
400 struct radeon_cmdbuf
*cs
= device
->ws
->cs_create(device
->ws
, RING_GFX
);
404 si_emit_config(device
->physical_device
, cs
);
406 while (cs
->cdw
& 7) {
407 if (device
->physical_device
->rad_info
.gfx_ib_pad_with_type2
)
408 radeon_emit(cs
, 0x80000000);
410 radeon_emit(cs
, 0xffff1000);
413 device
->gfx_init
= device
->ws
->buffer_create(device
->ws
,
416 RADEON_FLAG_CPU_ACCESS
|
417 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
418 RADEON_FLAG_READ_ONLY
);
419 if (!device
->gfx_init
)
422 void *map
= device
->ws
->buffer_map(device
->gfx_init
);
424 device
->ws
->buffer_destroy(device
->gfx_init
);
425 device
->gfx_init
= NULL
;
428 memcpy(map
, cs
->buf
, cs
->cdw
* 4);
430 device
->ws
->buffer_unmap(device
->gfx_init
);
431 device
->gfx_init_size_dw
= cs
->cdw
;
433 device
->ws
->cs_destroy(cs
);
437 get_viewport_xform(const VkViewport
*viewport
,
438 float scale
[3], float translate
[3])
440 float x
= viewport
->x
;
441 float y
= viewport
->y
;
442 float half_width
= 0.5f
* viewport
->width
;
443 float half_height
= 0.5f
* viewport
->height
;
444 double n
= viewport
->minDepth
;
445 double f
= viewport
->maxDepth
;
447 scale
[0] = half_width
;
448 translate
[0] = half_width
+ x
;
449 scale
[1] = half_height
;
450 translate
[1] = half_height
+ y
;
457 si_write_viewport(struct radeon_cmdbuf
*cs
, int first_vp
,
458 int count
, const VkViewport
*viewports
)
463 radeon_set_context_reg_seq(cs
, R_02843C_PA_CL_VPORT_XSCALE
+
464 first_vp
* 4 * 6, count
* 6);
466 for (i
= 0; i
< count
; i
++) {
467 float scale
[3], translate
[3];
470 get_viewport_xform(&viewports
[i
], scale
, translate
);
471 radeon_emit(cs
, fui(scale
[0]));
472 radeon_emit(cs
, fui(translate
[0]));
473 radeon_emit(cs
, fui(scale
[1]));
474 radeon_emit(cs
, fui(translate
[1]));
475 radeon_emit(cs
, fui(scale
[2]));
476 radeon_emit(cs
, fui(translate
[2]));
479 radeon_set_context_reg_seq(cs
, R_0282D0_PA_SC_VPORT_ZMIN_0
+
480 first_vp
* 4 * 2, count
* 2);
481 for (i
= 0; i
< count
; i
++) {
482 float zmin
= MIN2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
483 float zmax
= MAX2(viewports
[i
].minDepth
, viewports
[i
].maxDepth
);
484 radeon_emit(cs
, fui(zmin
));
485 radeon_emit(cs
, fui(zmax
));
489 static VkRect2D
si_scissor_from_viewport(const VkViewport
*viewport
)
491 float scale
[3], translate
[3];
494 get_viewport_xform(viewport
, scale
, translate
);
496 rect
.offset
.x
= translate
[0] - fabs(scale
[0]);
497 rect
.offset
.y
= translate
[1] - fabs(scale
[1]);
498 rect
.extent
.width
= ceilf(translate
[0] + fabs(scale
[0])) - rect
.offset
.x
;
499 rect
.extent
.height
= ceilf(translate
[1] + fabs(scale
[1])) - rect
.offset
.y
;
504 static VkRect2D
si_intersect_scissor(const VkRect2D
*a
, const VkRect2D
*b
) {
506 ret
.offset
.x
= MAX2(a
->offset
.x
, b
->offset
.x
);
507 ret
.offset
.y
= MAX2(a
->offset
.y
, b
->offset
.y
);
508 ret
.extent
.width
= MIN2(a
->offset
.x
+ a
->extent
.width
,
509 b
->offset
.x
+ b
->extent
.width
) - ret
.offset
.x
;
510 ret
.extent
.height
= MIN2(a
->offset
.y
+ a
->extent
.height
,
511 b
->offset
.y
+ b
->extent
.height
) - ret
.offset
.y
;
516 si_write_scissors(struct radeon_cmdbuf
*cs
, int first
,
517 int count
, const VkRect2D
*scissors
,
518 const VkViewport
*viewports
, bool can_use_guardband
)
521 float scale
[3], translate
[3], guardband_x
= INFINITY
, guardband_y
= INFINITY
;
522 const float max_range
= 32767.0f
;
526 radeon_set_context_reg_seq(cs
, R_028250_PA_SC_VPORT_SCISSOR_0_TL
+ first
* 4 * 2, count
* 2);
527 for (i
= 0; i
< count
; i
++) {
528 VkRect2D viewport_scissor
= si_scissor_from_viewport(viewports
+ i
);
529 VkRect2D scissor
= si_intersect_scissor(&scissors
[i
], &viewport_scissor
);
531 get_viewport_xform(viewports
+ i
, scale
, translate
);
532 scale
[0] = abs(scale
[0]);
533 scale
[1] = abs(scale
[1]);
540 guardband_x
= MIN2(guardband_x
, (max_range
- abs(translate
[0])) / scale
[0]);
541 guardband_y
= MIN2(guardband_y
, (max_range
- abs(translate
[1])) / scale
[1]);
543 radeon_emit(cs
, S_028250_TL_X(scissor
.offset
.x
) |
544 S_028250_TL_Y(scissor
.offset
.y
) |
545 S_028250_WINDOW_OFFSET_DISABLE(1));
546 radeon_emit(cs
, S_028254_BR_X(scissor
.offset
.x
+ scissor
.extent
.width
) |
547 S_028254_BR_Y(scissor
.offset
.y
+ scissor
.extent
.height
));
549 if (!can_use_guardband
) {
554 radeon_set_context_reg_seq(cs
, R_028BE8_PA_CL_GB_VERT_CLIP_ADJ
, 4);
555 radeon_emit(cs
, fui(guardband_y
));
556 radeon_emit(cs
, fui(1.0));
557 radeon_emit(cs
, fui(guardband_x
));
558 radeon_emit(cs
, fui(1.0));
561 static inline unsigned
562 radv_prims_for_vertices(struct radv_prim_vertex_count
*info
, unsigned num
)
573 return 1 + ((num
- info
->min
) / info
->incr
);
577 si_get_ia_multi_vgt_param(struct radv_cmd_buffer
*cmd_buffer
,
578 bool instanced_draw
, bool indirect_draw
,
579 uint32_t draw_vertex_count
)
581 enum chip_class chip_class
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
;
582 enum radeon_family family
= cmd_buffer
->device
->physical_device
->rad_info
.family
;
583 struct radeon_info
*info
= &cmd_buffer
->device
->physical_device
->rad_info
;
584 const unsigned max_primgroup_in_wave
= 2;
585 /* SWITCH_ON_EOP(0) is always preferable. */
586 bool wd_switch_on_eop
= false;
587 bool ia_switch_on_eop
= false;
588 bool ia_switch_on_eoi
= false;
589 bool partial_vs_wave
= false;
590 bool partial_es_wave
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.partial_es_wave
;
591 bool multi_instances_smaller_than_primgroup
;
593 multi_instances_smaller_than_primgroup
= indirect_draw
;
594 if (!multi_instances_smaller_than_primgroup
&& instanced_draw
) {
595 uint32_t num_prims
= radv_prims_for_vertices(&cmd_buffer
->state
.pipeline
->graphics
.prim_vertex_count
, draw_vertex_count
);
596 if (num_prims
< cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.primgroup_size
)
597 multi_instances_smaller_than_primgroup
= true;
600 ia_switch_on_eoi
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.ia_switch_on_eoi
;
601 partial_vs_wave
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.partial_vs_wave
;
603 if (chip_class
>= CIK
) {
604 wd_switch_on_eop
= cmd_buffer
->state
.pipeline
->graphics
.ia_multi_vgt_param
.wd_switch_on_eop
;
606 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
607 * We don't know that for indirect drawing, so treat it as
608 * always problematic. */
609 if (family
== CHIP_HAWAII
&&
610 (instanced_draw
|| indirect_draw
))
611 wd_switch_on_eop
= true;
613 /* Performance recommendation for 4 SE Gfx7-8 parts if
614 * instances are smaller than a primgroup.
615 * Assume indirect draws always use small instances.
616 * This is needed for good VS wave utilization.
618 if (chip_class
<= VI
&&
620 multi_instances_smaller_than_primgroup
)
621 wd_switch_on_eop
= true;
623 /* Required on CIK and later. */
624 if (info
->max_se
> 2 && !wd_switch_on_eop
)
625 ia_switch_on_eoi
= true;
627 /* Required by Hawaii and, for some special cases, by VI. */
628 if (ia_switch_on_eoi
&&
629 (family
== CHIP_HAWAII
||
631 /* max primgroup in wave is always 2 - leave this for documentation */
632 (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
) || max_primgroup_in_wave
!= 2))))
633 partial_vs_wave
= true;
635 /* Instancing bug on Bonaire. */
636 if (family
== CHIP_BONAIRE
&& ia_switch_on_eoi
&&
637 (instanced_draw
|| indirect_draw
))
638 partial_vs_wave
= 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
<= VI
&& 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
>= CIK
? 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
,
681 uint64_t gfx9_eop_bug_va
)
683 unsigned op
= EVENT_TYPE(event
) |
686 unsigned is_gfx8_mec
= is_mec
&& chip_class
< GFX9
;
687 unsigned sel
= EOP_DATA_SEL(data_sel
);
689 /* Wait for write confirmation before writing data, but don't send
691 if (data_sel
!= EOP_DATA_SEL_DISCARD
)
692 sel
|= EOP_INT_SEL(EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM
);
694 if (chip_class
>= GFX9
|| is_gfx8_mec
) {
695 /* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
696 * counters) must immediately precede every timestamp event to
697 * prevent a GPU hang on GFX9.
699 if (chip_class
== GFX9
) {
700 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 2, 0));
701 radeon_emit(cs
, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE
) | EVENT_INDEX(1));
702 radeon_emit(cs
, gfx9_eop_bug_va
);
703 radeon_emit(cs
, gfx9_eop_bug_va
>> 32);
706 radeon_emit(cs
, PKT3(PKT3_RELEASE_MEM
, is_gfx8_mec
? 5 : 6, false));
708 radeon_emit(cs
, sel
);
709 radeon_emit(cs
, va
); /* address lo */
710 radeon_emit(cs
, va
>> 32); /* address hi */
711 radeon_emit(cs
, new_fence
); /* immediate data lo */
712 radeon_emit(cs
, 0); /* immediate data hi */
714 radeon_emit(cs
, 0); /* unused */
716 if (chip_class
== CIK
||
718 /* Two EOP events are required to make all engines go idle
719 * (and optional cache flushes executed) before the timestamp
722 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, false));
725 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
726 radeon_emit(cs
, old_fence
); /* immediate data */
727 radeon_emit(cs
, 0); /* unused */
730 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, false));
733 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
734 radeon_emit(cs
, new_fence
); /* immediate data */
735 radeon_emit(cs
, 0); /* unused */
740 si_emit_wait_fence(struct radeon_cmdbuf
*cs
,
741 uint64_t va
, uint32_t ref
,
744 radeon_emit(cs
, PKT3(PKT3_WAIT_REG_MEM
, 5, false));
745 radeon_emit(cs
, WAIT_REG_MEM_EQUAL
| WAIT_REG_MEM_MEM_SPACE(1));
747 radeon_emit(cs
, va
>> 32);
748 radeon_emit(cs
, ref
); /* reference value */
749 radeon_emit(cs
, mask
); /* mask */
750 radeon_emit(cs
, 4); /* poll interval */
754 si_emit_acquire_mem(struct radeon_cmdbuf
*cs
,
757 unsigned cp_coher_cntl
)
759 if (is_mec
|| is_gfx9
) {
760 uint32_t hi_val
= is_gfx9
? 0xffffff : 0xff;
761 radeon_emit(cs
, PKT3(PKT3_ACQUIRE_MEM
, 5, false) |
762 PKT3_SHADER_TYPE_S(is_mec
));
763 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
764 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
765 radeon_emit(cs
, hi_val
); /* CP_COHER_SIZE_HI */
766 radeon_emit(cs
, 0); /* CP_COHER_BASE */
767 radeon_emit(cs
, 0); /* CP_COHER_BASE_HI */
768 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
770 /* ACQUIRE_MEM is only required on a compute ring. */
771 radeon_emit(cs
, PKT3(PKT3_SURFACE_SYNC
, 3, false));
772 radeon_emit(cs
, cp_coher_cntl
); /* CP_COHER_CNTL */
773 radeon_emit(cs
, 0xffffffff); /* CP_COHER_SIZE */
774 radeon_emit(cs
, 0); /* CP_COHER_BASE */
775 radeon_emit(cs
, 0x0000000A); /* POLL_INTERVAL */
780 si_cs_emit_cache_flush(struct radeon_cmdbuf
*cs
,
781 enum chip_class chip_class
,
785 enum radv_cmd_flush_bits flush_bits
,
786 uint64_t gfx9_eop_bug_va
)
788 unsigned cp_coher_cntl
= 0;
789 uint32_t flush_cb_db
= flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
790 RADV_CMD_FLAG_FLUSH_AND_INV_DB
);
792 if (flush_bits
& RADV_CMD_FLAG_INV_ICACHE
)
793 cp_coher_cntl
|= S_0085F0_SH_ICACHE_ACTION_ENA(1);
794 if (flush_bits
& RADV_CMD_FLAG_INV_SMEM_L1
)
795 cp_coher_cntl
|= S_0085F0_SH_KCACHE_ACTION_ENA(1);
797 if (chip_class
<= VI
) {
798 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB
) {
799 cp_coher_cntl
|= S_0085F0_CB_ACTION_ENA(1) |
800 S_0085F0_CB0_DEST_BASE_ENA(1) |
801 S_0085F0_CB1_DEST_BASE_ENA(1) |
802 S_0085F0_CB2_DEST_BASE_ENA(1) |
803 S_0085F0_CB3_DEST_BASE_ENA(1) |
804 S_0085F0_CB4_DEST_BASE_ENA(1) |
805 S_0085F0_CB5_DEST_BASE_ENA(1) |
806 S_0085F0_CB6_DEST_BASE_ENA(1) |
807 S_0085F0_CB7_DEST_BASE_ENA(1);
809 /* Necessary for DCC */
810 if (chip_class
>= VI
) {
811 si_cs_emit_write_event_eop(cs
,
814 V_028A90_FLUSH_AND_INV_CB_DATA_TS
,
816 EOP_DATA_SEL_DISCARD
,
821 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB
) {
822 cp_coher_cntl
|= S_0085F0_DB_ACTION_ENA(1) |
823 S_0085F0_DB_DEST_BASE_ENA(1);
827 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
) {
828 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
829 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META
) | EVENT_INDEX(0));
832 if (flush_bits
& RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
) {
833 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
834 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META
) | EVENT_INDEX(0));
837 if (flush_bits
& RADV_CMD_FLAG_PS_PARTIAL_FLUSH
) {
838 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
839 radeon_emit(cs
, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
840 } else if (flush_bits
& RADV_CMD_FLAG_VS_PARTIAL_FLUSH
) {
841 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
842 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
845 if (flush_bits
& RADV_CMD_FLAG_CS_PARTIAL_FLUSH
) {
846 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
847 radeon_emit(cs
, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
850 if (chip_class
>= GFX9
&& flush_cb_db
) {
851 unsigned cb_db_event
, tc_flags
;
853 /* Set the CB/DB flush event. */
854 cb_db_event
= V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
;
856 /* These are the only allowed combinations. If you need to
857 * do multiple operations at once, do them separately.
858 * All operations that invalidate L2 also seem to invalidate
859 * metadata. Volatile (VOL) and WC flushes are not listed here.
861 * TC | TC_WB = writeback & invalidate L2 & L1
862 * TC | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
863 * TC_WB | TC_NC = writeback L2 for MTYPE == NC
864 * TC | TC_NC = invalidate L2 for MTYPE == NC
865 * TC | TC_MD = writeback & invalidate L2 metadata (DCC, etc.)
866 * TCL1 = invalidate L1
868 tc_flags
= EVENT_TC_ACTION_ENA
|
869 EVENT_TC_MD_ACTION_ENA
;
871 /* Ideally flush TC together with CB/DB. */
872 if (flush_bits
& RADV_CMD_FLAG_INV_GLOBAL_L2
) {
873 /* Writeback and invalidate everything in L2 & L1. */
874 tc_flags
= EVENT_TC_ACTION_ENA
|
875 EVENT_TC_WB_ACTION_ENA
;
878 /* Clear the flags. */
879 flush_bits
&= ~(RADV_CMD_FLAG_INV_GLOBAL_L2
|
880 RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2
|
881 RADV_CMD_FLAG_INV_VMEM_L1
);
884 uint32_t old_fence
= (*flush_cnt
)++;
886 si_cs_emit_write_event_eop(cs
, chip_class
, false, cb_db_event
, tc_flags
,
887 EOP_DATA_SEL_VALUE_32BIT
,
888 flush_va
, old_fence
, *flush_cnt
,
890 si_emit_wait_fence(cs
, flush_va
, *flush_cnt
, 0xffffffff);
894 if (flush_bits
& RADV_CMD_FLAG_VGT_FLUSH
) {
895 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
896 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
899 /* Make sure ME is idle (it executes most packets) before continuing.
900 * This prevents read-after-write hazards between PFP and ME.
902 if ((cp_coher_cntl
||
903 (flush_bits
& (RADV_CMD_FLAG_CS_PARTIAL_FLUSH
|
904 RADV_CMD_FLAG_INV_VMEM_L1
|
905 RADV_CMD_FLAG_INV_GLOBAL_L2
|
906 RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2
))) &&
908 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
912 if ((flush_bits
& RADV_CMD_FLAG_INV_GLOBAL_L2
) ||
913 (chip_class
<= CIK
&& (flush_bits
& RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2
))) {
914 si_emit_acquire_mem(cs
, is_mec
, chip_class
>= GFX9
,
916 S_0085F0_TC_ACTION_ENA(1) |
917 S_0085F0_TCL1_ACTION_ENA(1) |
918 S_0301F0_TC_WB_ACTION_ENA(chip_class
>= VI
));
921 if(flush_bits
& RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2
) {
923 * NC = apply to non-coherent MTYPEs
924 * (i.e. MTYPE <= 1, which is what we use everywhere)
926 * WB doesn't work without NC.
928 si_emit_acquire_mem(cs
, is_mec
,
931 S_0301F0_TC_WB_ACTION_ENA(1) |
932 S_0301F0_TC_NC_ACTION_ENA(1));
935 if (flush_bits
& RADV_CMD_FLAG_INV_VMEM_L1
) {
936 si_emit_acquire_mem(cs
, is_mec
,
939 S_0085F0_TCL1_ACTION_ENA(1));
944 /* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
945 * Therefore, it should be last. Done in PFP.
948 si_emit_acquire_mem(cs
, is_mec
, chip_class
>= GFX9
, cp_coher_cntl
);
950 if (flush_bits
& RADV_CMD_FLAG_START_PIPELINE_STATS
) {
951 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
952 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_START
) |
954 } else if (flush_bits
& RADV_CMD_FLAG_STOP_PIPELINE_STATS
) {
955 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
956 radeon_emit(cs
, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP
) |
962 si_emit_cache_flush(struct radv_cmd_buffer
*cmd_buffer
)
964 bool is_compute
= cmd_buffer
->queue_family_index
== RADV_QUEUE_COMPUTE
;
967 cmd_buffer
->state
.flush_bits
&= ~(RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
968 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
|
969 RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
970 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
|
971 RADV_CMD_FLAG_PS_PARTIAL_FLUSH
|
972 RADV_CMD_FLAG_VS_PARTIAL_FLUSH
|
973 RADV_CMD_FLAG_VGT_FLUSH
|
974 RADV_CMD_FLAG_START_PIPELINE_STATS
|
975 RADV_CMD_FLAG_STOP_PIPELINE_STATS
);
977 if (!cmd_buffer
->state
.flush_bits
)
980 enum chip_class chip_class
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
;
981 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 128);
983 uint32_t *ptr
= NULL
;
985 if (chip_class
== GFX9
) {
986 va
= radv_buffer_get_va(cmd_buffer
->gfx9_fence_bo
) + cmd_buffer
->gfx9_fence_offset
;
987 ptr
= &cmd_buffer
->gfx9_fence_idx
;
989 si_cs_emit_cache_flush(cmd_buffer
->cs
,
990 cmd_buffer
->device
->physical_device
->rad_info
.chip_class
,
992 radv_cmd_buffer_uses_mec(cmd_buffer
),
993 cmd_buffer
->state
.flush_bits
,
994 cmd_buffer
->gfx9_eop_bug_va
);
997 if (unlikely(cmd_buffer
->device
->trace_bo
))
998 radv_cmd_buffer_trace_emit(cmd_buffer
);
1000 cmd_buffer
->state
.flush_bits
= 0;
1003 /* sets the CP predication state using a boolean stored at va */
1005 si_emit_set_predication_state(struct radv_cmd_buffer
*cmd_buffer
,
1006 bool inverted
, uint64_t va
)
1011 op
= PRED_OP(PREDICATION_OP_BOOL64
);
1013 /* By default, our internal rendering commands are discarded
1014 * only if the predicate is non-zero (ie. DRAW_VISIBLE). But
1015 * VK_EXT_conditional_rendering also allows to discard commands
1016 * when the predicate is zero, which means we have to use a
1019 op
|= inverted
? PREDICATION_DRAW_VISIBLE
:
1020 PREDICATION_DRAW_NOT_VISIBLE
;
1022 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1023 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_SET_PREDICATION
, 2, 0));
1024 radeon_emit(cmd_buffer
->cs
, op
);
1025 radeon_emit(cmd_buffer
->cs
, va
);
1026 radeon_emit(cmd_buffer
->cs
, va
>> 32);
1028 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_SET_PREDICATION
, 1, 0));
1029 radeon_emit(cmd_buffer
->cs
, va
);
1030 radeon_emit(cmd_buffer
->cs
, op
| ((va
>> 32) & 0xFF));
1034 /* Set this if you want the 3D engine to wait until CP DMA is done.
1035 * It should be set on the last CP DMA packet. */
1036 #define CP_DMA_SYNC (1 << 0)
1038 /* Set this if the source data was used as a destination in a previous CP DMA
1039 * packet. It's for preventing a read-after-write (RAW) hazard between two
1040 * CP DMA packets. */
1041 #define CP_DMA_RAW_WAIT (1 << 1)
1042 #define CP_DMA_USE_L2 (1 << 2)
1043 #define CP_DMA_CLEAR (1 << 3)
1045 /* Alignment for optimal performance. */
1046 #define SI_CPDMA_ALIGNMENT 32
1048 /* The max number of bytes that can be copied per packet. */
1049 static inline unsigned cp_dma_max_byte_count(struct radv_cmd_buffer
*cmd_buffer
)
1051 unsigned max
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
?
1052 S_414_BYTE_COUNT_GFX9(~0u) :
1053 S_414_BYTE_COUNT_GFX6(~0u);
1055 /* make it aligned for optimal performance */
1056 return max
& ~(SI_CPDMA_ALIGNMENT
- 1);
1059 /* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
1060 * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
1063 static void si_emit_cp_dma(struct radv_cmd_buffer
*cmd_buffer
,
1064 uint64_t dst_va
, uint64_t src_va
,
1065 unsigned size
, unsigned flags
)
1067 struct radeon_cmdbuf
*cs
= cmd_buffer
->cs
;
1068 uint32_t header
= 0, command
= 0;
1070 assert(size
<= cp_dma_max_byte_count(cmd_buffer
));
1072 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 9);
1073 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1074 command
|= S_414_BYTE_COUNT_GFX9(size
);
1076 command
|= S_414_BYTE_COUNT_GFX6(size
);
1079 if (flags
& CP_DMA_SYNC
)
1080 header
|= S_411_CP_SYNC(1);
1082 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1083 command
|= S_414_DISABLE_WR_CONFIRM_GFX9(1);
1085 command
|= S_414_DISABLE_WR_CONFIRM_GFX6(1);
1088 if (flags
& CP_DMA_RAW_WAIT
)
1089 command
|= S_414_RAW_WAIT(1);
1091 /* Src and dst flags. */
1092 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1093 !(flags
& CP_DMA_CLEAR
) &&
1095 header
|= S_411_DST_SEL(V_411_NOWHERE
); /* prefetch only */
1096 else if (flags
& CP_DMA_USE_L2
)
1097 header
|= S_411_DST_SEL(V_411_DST_ADDR_TC_L2
);
1099 if (flags
& CP_DMA_CLEAR
)
1100 header
|= S_411_SRC_SEL(V_411_DATA
);
1101 else if (flags
& CP_DMA_USE_L2
)
1102 header
|= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2
);
1104 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1105 radeon_emit(cs
, PKT3(PKT3_DMA_DATA
, 5, cmd_buffer
->state
.predicating
));
1106 radeon_emit(cs
, header
);
1107 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
1108 radeon_emit(cs
, src_va
>> 32); /* SRC_ADDR_HI [31:0] */
1109 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
1110 radeon_emit(cs
, dst_va
>> 32); /* DST_ADDR_HI [31:0] */
1111 radeon_emit(cs
, command
);
1113 assert(!(flags
& CP_DMA_USE_L2
));
1114 header
|= S_411_SRC_ADDR_HI(src_va
>> 32);
1115 radeon_emit(cs
, PKT3(PKT3_CP_DMA
, 4, cmd_buffer
->state
.predicating
));
1116 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
1117 radeon_emit(cs
, header
); /* SRC_ADDR_HI [15:0] + flags. */
1118 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
1119 radeon_emit(cs
, (dst_va
>> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
1120 radeon_emit(cs
, command
);
1123 /* CP DMA is executed in ME, but index buffers are read by PFP.
1124 * This ensures that ME (CP DMA) is idle before PFP starts fetching
1125 * indices. If we wanted to execute CP DMA in PFP, this packet
1126 * should precede it.
1128 if (flags
& CP_DMA_SYNC
) {
1129 if (cmd_buffer
->queue_family_index
== RADV_QUEUE_GENERAL
) {
1130 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, cmd_buffer
->state
.predicating
));
1134 /* CP will see the sync flag and wait for all DMAs to complete. */
1135 cmd_buffer
->state
.dma_is_busy
= false;
1138 if (unlikely(cmd_buffer
->device
->trace_bo
))
1139 radv_cmd_buffer_trace_emit(cmd_buffer
);
1142 void si_cp_dma_prefetch(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1145 uint64_t aligned_va
= va
& ~(SI_CPDMA_ALIGNMENT
- 1);
1146 uint64_t aligned_size
= ((va
+ size
+ SI_CPDMA_ALIGNMENT
-1) & ~(SI_CPDMA_ALIGNMENT
- 1)) - aligned_va
;
1148 si_emit_cp_dma(cmd_buffer
, aligned_va
, aligned_va
,
1149 aligned_size
, CP_DMA_USE_L2
);
1152 static void si_cp_dma_prepare(struct radv_cmd_buffer
*cmd_buffer
, uint64_t byte_count
,
1153 uint64_t remaining_size
, unsigned *flags
)
1156 /* Flush the caches for the first copy only.
1157 * Also wait for the previous CP DMA operations.
1159 if (cmd_buffer
->state
.flush_bits
) {
1160 si_emit_cache_flush(cmd_buffer
);
1161 *flags
|= CP_DMA_RAW_WAIT
;
1164 /* Do the synchronization after the last dma, so that all data
1165 * is written to memory.
1167 if (byte_count
== remaining_size
)
1168 *flags
|= CP_DMA_SYNC
;
1171 static void si_cp_dma_realign_engine(struct radv_cmd_buffer
*cmd_buffer
, unsigned size
)
1175 unsigned dma_flags
= 0;
1176 unsigned buf_size
= SI_CPDMA_ALIGNMENT
* 2;
1179 assert(size
< SI_CPDMA_ALIGNMENT
);
1181 radv_cmd_buffer_upload_alloc(cmd_buffer
, buf_size
, SI_CPDMA_ALIGNMENT
, &offset
, &ptr
);
1183 va
= radv_buffer_get_va(cmd_buffer
->upload
.upload_bo
);
1186 si_cp_dma_prepare(cmd_buffer
, size
, size
, &dma_flags
);
1188 si_emit_cp_dma(cmd_buffer
, va
, va
+ SI_CPDMA_ALIGNMENT
, size
,
1192 void si_cp_dma_buffer_copy(struct radv_cmd_buffer
*cmd_buffer
,
1193 uint64_t src_va
, uint64_t dest_va
,
1196 uint64_t main_src_va
, main_dest_va
;
1197 uint64_t skipped_size
= 0, realign_size
= 0;
1199 /* Assume that we are not going to sync after the last DMA operation. */
1200 cmd_buffer
->state
.dma_is_busy
= true;
1202 if (cmd_buffer
->device
->physical_device
->rad_info
.family
<= CHIP_CARRIZO
||
1203 cmd_buffer
->device
->physical_device
->rad_info
.family
== CHIP_STONEY
) {
1204 /* If the size is not aligned, we must add a dummy copy at the end
1205 * just to align the internal counter. Otherwise, the DMA engine
1206 * would slow down by an order of magnitude for following copies.
1208 if (size
% SI_CPDMA_ALIGNMENT
)
1209 realign_size
= SI_CPDMA_ALIGNMENT
- (size
% SI_CPDMA_ALIGNMENT
);
1211 /* If the copy begins unaligned, we must start copying from the next
1212 * aligned block and the skipped part should be copied after everything
1213 * else has been copied. Only the src alignment matters, not dst.
1215 if (src_va
% SI_CPDMA_ALIGNMENT
) {
1216 skipped_size
= SI_CPDMA_ALIGNMENT
- (src_va
% SI_CPDMA_ALIGNMENT
);
1217 /* The main part will be skipped if the size is too small. */
1218 skipped_size
= MIN2(skipped_size
, size
);
1219 size
-= skipped_size
;
1222 main_src_va
= src_va
+ skipped_size
;
1223 main_dest_va
= dest_va
+ skipped_size
;
1226 unsigned dma_flags
= 0;
1227 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(cmd_buffer
));
1229 si_cp_dma_prepare(cmd_buffer
, byte_count
,
1230 size
+ skipped_size
+ realign_size
,
1233 si_emit_cp_dma(cmd_buffer
, main_dest_va
, main_src_va
,
1234 byte_count
, dma_flags
);
1237 main_src_va
+= byte_count
;
1238 main_dest_va
+= byte_count
;
1242 unsigned dma_flags
= 0;
1244 si_cp_dma_prepare(cmd_buffer
, skipped_size
,
1245 size
+ skipped_size
+ realign_size
,
1248 si_emit_cp_dma(cmd_buffer
, dest_va
, src_va
,
1249 skipped_size
, dma_flags
);
1252 si_cp_dma_realign_engine(cmd_buffer
, realign_size
);
1255 void si_cp_dma_clear_buffer(struct radv_cmd_buffer
*cmd_buffer
, uint64_t va
,
1256 uint64_t size
, unsigned value
)
1262 assert(va
% 4 == 0 && size
% 4 == 0);
1264 /* Assume that we are not going to sync after the last DMA operation. */
1265 cmd_buffer
->state
.dma_is_busy
= true;
1268 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(cmd_buffer
));
1269 unsigned dma_flags
= CP_DMA_CLEAR
;
1271 si_cp_dma_prepare(cmd_buffer
, byte_count
, size
, &dma_flags
);
1273 /* Emit the clear packet. */
1274 si_emit_cp_dma(cmd_buffer
, va
, value
, byte_count
,
1282 void si_cp_dma_wait_for_idle(struct radv_cmd_buffer
*cmd_buffer
)
1284 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
< CIK
)
1287 if (!cmd_buffer
->state
.dma_is_busy
)
1290 /* Issue a dummy DMA that copies zero bytes.
1292 * The DMA engine will see that there's no work to do and skip this
1293 * DMA request, however, the CP will see the sync flag and still wait
1294 * for all DMAs to complete.
1296 si_emit_cp_dma(cmd_buffer
, 0, 0, 0, CP_DMA_SYNC
);
1298 cmd_buffer
->state
.dma_is_busy
= false;
1301 /* For MSAA sample positions. */
1302 #define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y) \
1303 (((s0x) & 0xf) | (((unsigned)(s0y) & 0xf) << 4) | \
1304 (((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) | \
1305 (((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) | \
1306 (((unsigned)(s3x) & 0xf) << 24) | (((unsigned)(s3y) & 0xf) << 28))
1310 * There are two locations (4, 4), (-4, -4). */
1311 const uint32_t eg_sample_locs_2x
[4] = {
1312 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1313 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1314 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1315 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1317 const unsigned eg_max_dist_2x
= 4;
1319 * There are 4 locations: (-2, 6), (6, -2), (-6, 2), (2, 6). */
1320 const uint32_t eg_sample_locs_4x
[4] = {
1321 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1322 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1323 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1324 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1326 const unsigned eg_max_dist_4x
= 6;
1329 static const uint32_t cm_sample_locs_8x
[] = {
1330 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1331 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1332 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1333 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1334 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1335 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1336 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1337 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1339 static const unsigned cm_max_dist_8x
= 8;
1340 /* Cayman 16xMSAA */
1341 static const uint32_t cm_sample_locs_16x
[] = {
1342 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1343 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1344 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1345 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1346 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1347 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1348 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1349 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1350 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1351 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1352 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1353 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1354 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1355 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1356 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1357 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1359 static const unsigned cm_max_dist_16x
= 8;
1361 unsigned radv_cayman_get_maxdist(int log_samples
)
1363 unsigned max_dist
[] = {
1370 return max_dist
[log_samples
];
1373 void radv_cayman_emit_msaa_sample_locs(struct radeon_cmdbuf
*cs
, int nr_samples
)
1375 switch (nr_samples
) {
1378 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 0);
1379 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, 0);
1380 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, 0);
1381 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, 0);
1384 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, eg_sample_locs_2x
[0]);
1385 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, eg_sample_locs_2x
[1]);
1386 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, eg_sample_locs_2x
[2]);
1387 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, eg_sample_locs_2x
[3]);
1390 radeon_set_context_reg(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, eg_sample_locs_4x
[0]);
1391 radeon_set_context_reg(cs
, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0
, eg_sample_locs_4x
[1]);
1392 radeon_set_context_reg(cs
, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0
, eg_sample_locs_4x
[2]);
1393 radeon_set_context_reg(cs
, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0
, eg_sample_locs_4x
[3]);
1396 radeon_set_context_reg_seq(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 14);
1397 radeon_emit(cs
, cm_sample_locs_8x
[0]);
1398 radeon_emit(cs
, cm_sample_locs_8x
[4]);
1401 radeon_emit(cs
, cm_sample_locs_8x
[1]);
1402 radeon_emit(cs
, cm_sample_locs_8x
[5]);
1405 radeon_emit(cs
, cm_sample_locs_8x
[2]);
1406 radeon_emit(cs
, cm_sample_locs_8x
[6]);
1409 radeon_emit(cs
, cm_sample_locs_8x
[3]);
1410 radeon_emit(cs
, cm_sample_locs_8x
[7]);
1413 radeon_set_context_reg_seq(cs
, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0
, 16);
1414 radeon_emit(cs
, cm_sample_locs_16x
[0]);
1415 radeon_emit(cs
, cm_sample_locs_16x
[4]);
1416 radeon_emit(cs
, cm_sample_locs_16x
[8]);
1417 radeon_emit(cs
, cm_sample_locs_16x
[12]);
1418 radeon_emit(cs
, cm_sample_locs_16x
[1]);
1419 radeon_emit(cs
, cm_sample_locs_16x
[5]);
1420 radeon_emit(cs
, cm_sample_locs_16x
[9]);
1421 radeon_emit(cs
, cm_sample_locs_16x
[13]);
1422 radeon_emit(cs
, cm_sample_locs_16x
[2]);
1423 radeon_emit(cs
, cm_sample_locs_16x
[6]);
1424 radeon_emit(cs
, cm_sample_locs_16x
[10]);
1425 radeon_emit(cs
, cm_sample_locs_16x
[14]);
1426 radeon_emit(cs
, cm_sample_locs_16x
[3]);
1427 radeon_emit(cs
, cm_sample_locs_16x
[7]);
1428 radeon_emit(cs
, cm_sample_locs_16x
[11]);
1429 radeon_emit(cs
, cm_sample_locs_16x
[15]);
1434 static void radv_cayman_get_sample_position(struct radv_device
*device
,
1435 unsigned sample_count
,
1436 unsigned sample_index
, float *out_value
)
1442 switch (sample_count
) {
1445 out_value
[0] = out_value
[1] = 0.5;
1448 offset
= 4 * (sample_index
* 2);
1449 val
.idx
= (eg_sample_locs_2x
[0] >> offset
) & 0xf;
1450 out_value
[0] = (float)(val
.idx
+ 8) / 16.0f
;
1451 val
.idx
= (eg_sample_locs_2x
[0] >> (offset
+ 4)) & 0xf;
1452 out_value
[1] = (float)(val
.idx
+ 8) / 16.0f
;
1455 offset
= 4 * (sample_index
* 2);
1456 val
.idx
= (eg_sample_locs_4x
[0] >> offset
) & 0xf;
1457 out_value
[0] = (float)(val
.idx
+ 8) / 16.0f
;
1458 val
.idx
= (eg_sample_locs_4x
[0] >> (offset
+ 4)) & 0xf;
1459 out_value
[1] = (float)(val
.idx
+ 8) / 16.0f
;
1462 offset
= 4 * (sample_index
% 4 * 2);
1463 index
= (sample_index
/ 4) * 4;
1464 val
.idx
= (cm_sample_locs_8x
[index
] >> offset
) & 0xf;
1465 out_value
[0] = (float)(val
.idx
+ 8) / 16.0f
;
1466 val
.idx
= (cm_sample_locs_8x
[index
] >> (offset
+ 4)) & 0xf;
1467 out_value
[1] = (float)(val
.idx
+ 8) / 16.0f
;
1470 offset
= 4 * (sample_index
% 4 * 2);
1471 index
= (sample_index
/ 4) * 4;
1472 val
.idx
= (cm_sample_locs_16x
[index
] >> offset
) & 0xf;
1473 out_value
[0] = (float)(val
.idx
+ 8) / 16.0f
;
1474 val
.idx
= (cm_sample_locs_16x
[index
] >> (offset
+ 4)) & 0xf;
1475 out_value
[1] = (float)(val
.idx
+ 8) / 16.0f
;
1480 void radv_device_init_msaa(struct radv_device
*device
)
1483 radv_cayman_get_sample_position(device
, 1, 0, device
->sample_locations_1x
[0]);
1485 for (i
= 0; i
< 2; i
++)
1486 radv_cayman_get_sample_position(device
, 2, i
, device
->sample_locations_2x
[i
]);
1487 for (i
= 0; i
< 4; i
++)
1488 radv_cayman_get_sample_position(device
, 4, i
, device
->sample_locations_4x
[i
]);
1489 for (i
= 0; i
< 8; i
++)
1490 radv_cayman_get_sample_position(device
, 8, i
, device
->sample_locations_8x
[i
]);
1491 for (i
= 0; i
< 16; i
++)
1492 radv_cayman_get_sample_position(device
, 16, i
, device
->sample_locations_16x
[i
]);