2 * Copyright 2012 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 * Christian König <christian.koenig@amd.com>
30 #include "radeon/r600_cs.h"
31 #include "radeon/r600_query.h"
33 #include "util/u_dual_blend.h"
34 #include "util/u_format.h"
35 #include "util/u_format_s3tc.h"
36 #include "util/u_memory.h"
37 #include "util/u_resource.h"
38 #include "util/u_upload_mgr.h"
40 /* Initialize an external atom (owned by ../radeon). */
42 si_init_external_atom(struct si_context
*sctx
, struct r600_atom
*atom
,
43 struct r600_atom
**list_elem
)
45 atom
->id
= list_elem
- sctx
->atoms
.array
;
49 /* Initialize an atom owned by radeonsi. */
50 void si_init_atom(struct si_context
*sctx
, struct r600_atom
*atom
,
51 struct r600_atom
**list_elem
,
52 void (*emit_func
)(struct si_context
*ctx
, struct r600_atom
*state
))
54 atom
->emit
= (void*)emit_func
;
55 atom
->id
= list_elem
- sctx
->atoms
.array
;
59 static unsigned si_map_swizzle(unsigned swizzle
)
63 return V_008F0C_SQ_SEL_Y
;
65 return V_008F0C_SQ_SEL_Z
;
67 return V_008F0C_SQ_SEL_W
;
69 return V_008F0C_SQ_SEL_0
;
71 return V_008F0C_SQ_SEL_1
;
72 default: /* PIPE_SWIZZLE_X */
73 return V_008F0C_SQ_SEL_X
;
77 /* 12.4 fixed-point */
78 static unsigned si_pack_float_12p4(float x
)
81 x
>= 4096 ? 0xffff : x
* 16;
85 * Inferred framebuffer and blender state.
87 * CB_TARGET_MASK is emitted here to avoid a hang with dual source blending
88 * if there is not enough PS outputs.
90 static void si_emit_cb_render_state(struct si_context
*sctx
, struct r600_atom
*atom
)
92 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
93 struct si_state_blend
*blend
= sctx
->queued
.named
.blend
;
94 /* CB_COLORn_INFO.FORMAT=INVALID should disable unbound colorbuffers,
95 * but you never know. */
96 uint32_t cb_target_mask
= sctx
->framebuffer
.colorbuf_enabled_4bit
;
100 cb_target_mask
&= blend
->cb_target_mask
;
102 /* Avoid a hang that happens when dual source blending is enabled
103 * but there is not enough color outputs. This is undefined behavior,
104 * so disable color writes completely.
106 * Reproducible with Unigine Heaven 4.0 and drirc missing.
108 if (blend
&& blend
->dual_src_blend
&&
109 sctx
->ps_shader
.cso
&&
110 (sctx
->ps_shader
.cso
->info
.colors_written
& 0x3) != 0x3)
113 radeon_set_context_reg(cs
, R_028238_CB_TARGET_MASK
, cb_target_mask
);
115 /* GFX9: Flush DFSM when CB_TARGET_MASK changes.
116 * I think we don't have to do anything between IBs.
118 if (sctx
->screen
->dfsm_allowed
&&
119 sctx
->last_cb_target_mask
!= cb_target_mask
) {
120 sctx
->last_cb_target_mask
= cb_target_mask
;
122 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
123 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_DFSM
) | EVENT_INDEX(0));
126 /* RB+ register settings. */
127 if (sctx
->screen
->b
.rbplus_allowed
) {
128 unsigned spi_shader_col_format
=
129 sctx
->ps_shader
.cso
?
130 sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.spi_shader_col_format
: 0;
131 unsigned sx_ps_downconvert
= 0;
132 unsigned sx_blend_opt_epsilon
= 0;
133 unsigned sx_blend_opt_control
= 0;
135 for (i
= 0; i
< sctx
->framebuffer
.state
.nr_cbufs
; i
++) {
136 struct r600_surface
*surf
=
137 (struct r600_surface
*)sctx
->framebuffer
.state
.cbufs
[i
];
138 unsigned format
, swap
, spi_format
, colormask
;
139 bool has_alpha
, has_rgb
;
144 format
= G_028C70_FORMAT(surf
->cb_color_info
);
145 swap
= G_028C70_COMP_SWAP(surf
->cb_color_info
);
146 spi_format
= (spi_shader_col_format
>> (i
* 4)) & 0xf;
147 colormask
= (cb_target_mask
>> (i
* 4)) & 0xf;
149 /* Set if RGB and A are present. */
150 has_alpha
= !G_028C74_FORCE_DST_ALPHA_1(surf
->cb_color_attrib
);
152 if (format
== V_028C70_COLOR_8
||
153 format
== V_028C70_COLOR_16
||
154 format
== V_028C70_COLOR_32
)
155 has_rgb
= !has_alpha
;
159 /* Check the colormask and export format. */
160 if (!(colormask
& (PIPE_MASK_RGBA
& ~PIPE_MASK_A
)))
162 if (!(colormask
& PIPE_MASK_A
))
165 if (spi_format
== V_028714_SPI_SHADER_ZERO
) {
170 /* Disable value checking for disabled channels. */
172 sx_blend_opt_control
|= S_02875C_MRT0_COLOR_OPT_DISABLE(1) << (i
* 4);
174 sx_blend_opt_control
|= S_02875C_MRT0_ALPHA_OPT_DISABLE(1) << (i
* 4);
176 /* Enable down-conversion for 32bpp and smaller formats. */
178 case V_028C70_COLOR_8
:
179 case V_028C70_COLOR_8_8
:
180 case V_028C70_COLOR_8_8_8_8
:
181 /* For 1 and 2-channel formats, use the superset thereof. */
182 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
||
183 spi_format
== V_028714_SPI_SHADER_UINT16_ABGR
||
184 spi_format
== V_028714_SPI_SHADER_SINT16_ABGR
) {
185 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_8_8_8_8
<< (i
* 4);
186 sx_blend_opt_epsilon
|= V_028758_8BIT_FORMAT
<< (i
* 4);
190 case V_028C70_COLOR_5_6_5
:
191 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
192 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_5_6_5
<< (i
* 4);
193 sx_blend_opt_epsilon
|= V_028758_6BIT_FORMAT
<< (i
* 4);
197 case V_028C70_COLOR_1_5_5_5
:
198 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
199 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_1_5_5_5
<< (i
* 4);
200 sx_blend_opt_epsilon
|= V_028758_5BIT_FORMAT
<< (i
* 4);
204 case V_028C70_COLOR_4_4_4_4
:
205 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
206 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_4_4_4_4
<< (i
* 4);
207 sx_blend_opt_epsilon
|= V_028758_4BIT_FORMAT
<< (i
* 4);
211 case V_028C70_COLOR_32
:
212 if (swap
== V_0280A0_SWAP_STD
&&
213 spi_format
== V_028714_SPI_SHADER_32_R
)
214 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_32_R
<< (i
* 4);
215 else if (swap
== V_0280A0_SWAP_ALT_REV
&&
216 spi_format
== V_028714_SPI_SHADER_32_AR
)
217 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_32_A
<< (i
* 4);
220 case V_028C70_COLOR_16
:
221 case V_028C70_COLOR_16_16
:
222 /* For 1-channel formats, use the superset thereof. */
223 if (spi_format
== V_028714_SPI_SHADER_UNORM16_ABGR
||
224 spi_format
== V_028714_SPI_SHADER_SNORM16_ABGR
||
225 spi_format
== V_028714_SPI_SHADER_UINT16_ABGR
||
226 spi_format
== V_028714_SPI_SHADER_SINT16_ABGR
) {
227 if (swap
== V_0280A0_SWAP_STD
||
228 swap
== V_0280A0_SWAP_STD_REV
)
229 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_16_16_GR
<< (i
* 4);
231 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_16_16_AR
<< (i
* 4);
235 case V_028C70_COLOR_10_11_11
:
236 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
237 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_10_11_11
<< (i
* 4);
238 sx_blend_opt_epsilon
|= V_028758_11BIT_FORMAT
<< (i
* 4);
242 case V_028C70_COLOR_2_10_10_10
:
243 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
244 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_2_10_10_10
<< (i
* 4);
245 sx_blend_opt_epsilon
|= V_028758_10BIT_FORMAT
<< (i
* 4);
251 radeon_set_context_reg_seq(cs
, R_028754_SX_PS_DOWNCONVERT
, 3);
252 radeon_emit(cs
, sx_ps_downconvert
); /* R_028754_SX_PS_DOWNCONVERT */
253 radeon_emit(cs
, sx_blend_opt_epsilon
); /* R_028758_SX_BLEND_OPT_EPSILON */
254 radeon_emit(cs
, sx_blend_opt_control
); /* R_02875C_SX_BLEND_OPT_CONTROL */
255 } else if (sctx
->screen
->b
.has_rbplus
) {
256 radeon_set_context_reg_seq(cs
, R_028754_SX_PS_DOWNCONVERT
, 3);
257 radeon_emit(cs
, 0); /* R_028754_SX_PS_DOWNCONVERT */
258 radeon_emit(cs
, 0); /* R_028758_SX_BLEND_OPT_EPSILON */
259 radeon_emit(cs
, 0); /* R_02875C_SX_BLEND_OPT_CONTROL */
267 static uint32_t si_translate_blend_function(int blend_func
)
269 switch (blend_func
) {
271 return V_028780_COMB_DST_PLUS_SRC
;
272 case PIPE_BLEND_SUBTRACT
:
273 return V_028780_COMB_SRC_MINUS_DST
;
274 case PIPE_BLEND_REVERSE_SUBTRACT
:
275 return V_028780_COMB_DST_MINUS_SRC
;
277 return V_028780_COMB_MIN_DST_SRC
;
279 return V_028780_COMB_MAX_DST_SRC
;
281 R600_ERR("Unknown blend function %d\n", blend_func
);
288 static uint32_t si_translate_blend_factor(int blend_fact
)
290 switch (blend_fact
) {
291 case PIPE_BLENDFACTOR_ONE
:
292 return V_028780_BLEND_ONE
;
293 case PIPE_BLENDFACTOR_SRC_COLOR
:
294 return V_028780_BLEND_SRC_COLOR
;
295 case PIPE_BLENDFACTOR_SRC_ALPHA
:
296 return V_028780_BLEND_SRC_ALPHA
;
297 case PIPE_BLENDFACTOR_DST_ALPHA
:
298 return V_028780_BLEND_DST_ALPHA
;
299 case PIPE_BLENDFACTOR_DST_COLOR
:
300 return V_028780_BLEND_DST_COLOR
;
301 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
302 return V_028780_BLEND_SRC_ALPHA_SATURATE
;
303 case PIPE_BLENDFACTOR_CONST_COLOR
:
304 return V_028780_BLEND_CONSTANT_COLOR
;
305 case PIPE_BLENDFACTOR_CONST_ALPHA
:
306 return V_028780_BLEND_CONSTANT_ALPHA
;
307 case PIPE_BLENDFACTOR_ZERO
:
308 return V_028780_BLEND_ZERO
;
309 case PIPE_BLENDFACTOR_INV_SRC_COLOR
:
310 return V_028780_BLEND_ONE_MINUS_SRC_COLOR
;
311 case PIPE_BLENDFACTOR_INV_SRC_ALPHA
:
312 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA
;
313 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
314 return V_028780_BLEND_ONE_MINUS_DST_ALPHA
;
315 case PIPE_BLENDFACTOR_INV_DST_COLOR
:
316 return V_028780_BLEND_ONE_MINUS_DST_COLOR
;
317 case PIPE_BLENDFACTOR_INV_CONST_COLOR
:
318 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR
;
319 case PIPE_BLENDFACTOR_INV_CONST_ALPHA
:
320 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA
;
321 case PIPE_BLENDFACTOR_SRC1_COLOR
:
322 return V_028780_BLEND_SRC1_COLOR
;
323 case PIPE_BLENDFACTOR_SRC1_ALPHA
:
324 return V_028780_BLEND_SRC1_ALPHA
;
325 case PIPE_BLENDFACTOR_INV_SRC1_COLOR
:
326 return V_028780_BLEND_INV_SRC1_COLOR
;
327 case PIPE_BLENDFACTOR_INV_SRC1_ALPHA
:
328 return V_028780_BLEND_INV_SRC1_ALPHA
;
330 R600_ERR("Bad blend factor %d not supported!\n", blend_fact
);
337 static uint32_t si_translate_blend_opt_function(int blend_func
)
339 switch (blend_func
) {
341 return V_028760_OPT_COMB_ADD
;
342 case PIPE_BLEND_SUBTRACT
:
343 return V_028760_OPT_COMB_SUBTRACT
;
344 case PIPE_BLEND_REVERSE_SUBTRACT
:
345 return V_028760_OPT_COMB_REVSUBTRACT
;
347 return V_028760_OPT_COMB_MIN
;
349 return V_028760_OPT_COMB_MAX
;
351 return V_028760_OPT_COMB_BLEND_DISABLED
;
355 static uint32_t si_translate_blend_opt_factor(int blend_fact
, bool is_alpha
)
357 switch (blend_fact
) {
358 case PIPE_BLENDFACTOR_ZERO
:
359 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL
;
360 case PIPE_BLENDFACTOR_ONE
:
361 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
;
362 case PIPE_BLENDFACTOR_SRC_COLOR
:
363 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
364 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0
;
365 case PIPE_BLENDFACTOR_INV_SRC_COLOR
:
366 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
367 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1
;
368 case PIPE_BLENDFACTOR_SRC_ALPHA
:
369 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
;
370 case PIPE_BLENDFACTOR_INV_SRC_ALPHA
:
371 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
;
372 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
373 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
374 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0
;
376 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
380 static void si_blend_check_commutativity(struct si_screen
*sscreen
,
381 struct si_state_blend
*blend
,
382 enum pipe_blend_func func
,
383 enum pipe_blendfactor src
,
384 enum pipe_blendfactor dst
,
387 /* Src factor is allowed when it does not depend on Dst */
388 static const uint32_t src_allowed
=
389 (1u << PIPE_BLENDFACTOR_ONE
) |
390 (1u << PIPE_BLENDFACTOR_SRC_COLOR
) |
391 (1u << PIPE_BLENDFACTOR_SRC_ALPHA
) |
392 (1u << PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
) |
393 (1u << PIPE_BLENDFACTOR_CONST_COLOR
) |
394 (1u << PIPE_BLENDFACTOR_CONST_ALPHA
) |
395 (1u << PIPE_BLENDFACTOR_SRC1_COLOR
) |
396 (1u << PIPE_BLENDFACTOR_SRC1_ALPHA
) |
397 (1u << PIPE_BLENDFACTOR_ZERO
) |
398 (1u << PIPE_BLENDFACTOR_INV_SRC_COLOR
) |
399 (1u << PIPE_BLENDFACTOR_INV_SRC_ALPHA
) |
400 (1u << PIPE_BLENDFACTOR_INV_CONST_COLOR
) |
401 (1u << PIPE_BLENDFACTOR_INV_CONST_ALPHA
) |
402 (1u << PIPE_BLENDFACTOR_INV_SRC1_COLOR
) |
403 (1u << PIPE_BLENDFACTOR_INV_SRC1_ALPHA
);
405 if (dst
== PIPE_BLENDFACTOR_ONE
&&
406 (src_allowed
& (1u << src
))) {
407 /* Addition is commutative, but floating point addition isn't
408 * associative: subtle changes can be introduced via different
411 * Out-of-order is also non-deterministic, which means that
412 * this breaks OpenGL invariance requirements. So only enable
413 * out-of-order additive blending if explicitly allowed by a
416 if (func
== PIPE_BLEND_MAX
|| func
== PIPE_BLEND_MIN
||
417 (func
== PIPE_BLEND_ADD
&& sscreen
->commutative_blend_add
))
418 blend
->commutative_4bit
|= chanmask
;
423 * Get rid of DST in the blend factors by commuting the operands:
424 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
426 static void si_blend_remove_dst(unsigned *func
, unsigned *src_factor
,
427 unsigned *dst_factor
, unsigned expected_dst
,
428 unsigned replacement_src
)
430 if (*src_factor
== expected_dst
&&
431 *dst_factor
== PIPE_BLENDFACTOR_ZERO
) {
432 *src_factor
= PIPE_BLENDFACTOR_ZERO
;
433 *dst_factor
= replacement_src
;
435 /* Commuting the operands requires reversing subtractions. */
436 if (*func
== PIPE_BLEND_SUBTRACT
)
437 *func
= PIPE_BLEND_REVERSE_SUBTRACT
;
438 else if (*func
== PIPE_BLEND_REVERSE_SUBTRACT
)
439 *func
= PIPE_BLEND_SUBTRACT
;
443 static bool si_blend_factor_uses_dst(unsigned factor
)
445 return factor
== PIPE_BLENDFACTOR_DST_COLOR
||
446 factor
== PIPE_BLENDFACTOR_DST_ALPHA
||
447 factor
== PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
||
448 factor
== PIPE_BLENDFACTOR_INV_DST_ALPHA
||
449 factor
== PIPE_BLENDFACTOR_INV_DST_COLOR
;
452 static void *si_create_blend_state_mode(struct pipe_context
*ctx
,
453 const struct pipe_blend_state
*state
,
456 struct si_context
*sctx
= (struct si_context
*)ctx
;
457 struct si_state_blend
*blend
= CALLOC_STRUCT(si_state_blend
);
458 struct si_pm4_state
*pm4
= &blend
->pm4
;
459 uint32_t sx_mrt_blend_opt
[8] = {0};
460 uint32_t color_control
= 0;
465 blend
->alpha_to_coverage
= state
->alpha_to_coverage
;
466 blend
->alpha_to_one
= state
->alpha_to_one
;
467 blend
->dual_src_blend
= util_blend_state_is_dual(state
, 0);
468 blend
->logicop_enable
= state
->logicop_enable
;
470 if (state
->logicop_enable
) {
471 color_control
|= S_028808_ROP3(state
->logicop_func
| (state
->logicop_func
<< 4));
473 color_control
|= S_028808_ROP3(0xcc);
476 si_pm4_set_reg(pm4
, R_028B70_DB_ALPHA_TO_MASK
,
477 S_028B70_ALPHA_TO_MASK_ENABLE(state
->alpha_to_coverage
) |
478 S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
479 S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
480 S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
481 S_028B70_ALPHA_TO_MASK_OFFSET3(2));
483 if (state
->alpha_to_coverage
)
484 blend
->need_src_alpha_4bit
|= 0xf;
486 blend
->cb_target_mask
= 0;
487 blend
->cb_target_enabled_4bit
= 0;
489 for (int i
= 0; i
< 8; i
++) {
490 /* state->rt entries > 0 only written if independent blending */
491 const int j
= state
->independent_blend_enable
? i
: 0;
493 unsigned eqRGB
= state
->rt
[j
].rgb_func
;
494 unsigned srcRGB
= state
->rt
[j
].rgb_src_factor
;
495 unsigned dstRGB
= state
->rt
[j
].rgb_dst_factor
;
496 unsigned eqA
= state
->rt
[j
].alpha_func
;
497 unsigned srcA
= state
->rt
[j
].alpha_src_factor
;
498 unsigned dstA
= state
->rt
[j
].alpha_dst_factor
;
500 unsigned srcRGB_opt
, dstRGB_opt
, srcA_opt
, dstA_opt
;
501 unsigned blend_cntl
= 0;
503 sx_mrt_blend_opt
[i
] =
504 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED
) |
505 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED
);
507 /* Only set dual source blending for MRT0 to avoid a hang. */
508 if (i
>= 1 && blend
->dual_src_blend
) {
509 /* Vulkan does this for dual source blending. */
511 blend_cntl
|= S_028780_ENABLE(1);
513 si_pm4_set_reg(pm4
, R_028780_CB_BLEND0_CONTROL
+ i
* 4, blend_cntl
);
517 /* Only addition and subtraction equations are supported with
518 * dual source blending.
520 if (blend
->dual_src_blend
&&
521 (eqRGB
== PIPE_BLEND_MIN
|| eqRGB
== PIPE_BLEND_MAX
||
522 eqA
== PIPE_BLEND_MIN
|| eqA
== PIPE_BLEND_MAX
)) {
523 assert(!"Unsupported equation for dual source blending");
524 si_pm4_set_reg(pm4
, R_028780_CB_BLEND0_CONTROL
+ i
* 4, blend_cntl
);
528 /* cb_render_state will disable unused ones */
529 blend
->cb_target_mask
|= (unsigned)state
->rt
[j
].colormask
<< (4 * i
);
530 if (state
->rt
[j
].colormask
)
531 blend
->cb_target_enabled_4bit
|= 0xf << (4 * i
);
533 if (!state
->rt
[j
].colormask
|| !state
->rt
[j
].blend_enable
) {
534 si_pm4_set_reg(pm4
, R_028780_CB_BLEND0_CONTROL
+ i
* 4, blend_cntl
);
538 si_blend_check_commutativity(sctx
->screen
, blend
,
539 eqRGB
, srcRGB
, dstRGB
, 0x7 << (4 * i
));
540 si_blend_check_commutativity(sctx
->screen
, blend
,
541 eqA
, srcA
, dstA
, 0x8 << (4 * i
));
543 /* Blending optimizations for RB+.
544 * These transformations don't change the behavior.
546 * First, get rid of DST in the blend factors:
547 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
549 si_blend_remove_dst(&eqRGB
, &srcRGB
, &dstRGB
,
550 PIPE_BLENDFACTOR_DST_COLOR
,
551 PIPE_BLENDFACTOR_SRC_COLOR
);
552 si_blend_remove_dst(&eqA
, &srcA
, &dstA
,
553 PIPE_BLENDFACTOR_DST_COLOR
,
554 PIPE_BLENDFACTOR_SRC_COLOR
);
555 si_blend_remove_dst(&eqA
, &srcA
, &dstA
,
556 PIPE_BLENDFACTOR_DST_ALPHA
,
557 PIPE_BLENDFACTOR_SRC_ALPHA
);
559 /* Look up the ideal settings from tables. */
560 srcRGB_opt
= si_translate_blend_opt_factor(srcRGB
, false);
561 dstRGB_opt
= si_translate_blend_opt_factor(dstRGB
, false);
562 srcA_opt
= si_translate_blend_opt_factor(srcA
, true);
563 dstA_opt
= si_translate_blend_opt_factor(dstA
, true);
565 /* Handle interdependencies. */
566 if (si_blend_factor_uses_dst(srcRGB
))
567 dstRGB_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
568 if (si_blend_factor_uses_dst(srcA
))
569 dstA_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
571 if (srcRGB
== PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
&&
572 (dstRGB
== PIPE_BLENDFACTOR_ZERO
||
573 dstRGB
== PIPE_BLENDFACTOR_SRC_ALPHA
||
574 dstRGB
== PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
))
575 dstRGB_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0
;
577 /* Set the final value. */
578 sx_mrt_blend_opt
[i
] =
579 S_028760_COLOR_SRC_OPT(srcRGB_opt
) |
580 S_028760_COLOR_DST_OPT(dstRGB_opt
) |
581 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB
)) |
582 S_028760_ALPHA_SRC_OPT(srcA_opt
) |
583 S_028760_ALPHA_DST_OPT(dstA_opt
) |
584 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA
));
586 /* Set blend state. */
587 blend_cntl
|= S_028780_ENABLE(1);
588 blend_cntl
|= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB
));
589 blend_cntl
|= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB
));
590 blend_cntl
|= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB
));
592 if (srcA
!= srcRGB
|| dstA
!= dstRGB
|| eqA
!= eqRGB
) {
593 blend_cntl
|= S_028780_SEPARATE_ALPHA_BLEND(1);
594 blend_cntl
|= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA
));
595 blend_cntl
|= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA
));
596 blend_cntl
|= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA
));
598 si_pm4_set_reg(pm4
, R_028780_CB_BLEND0_CONTROL
+ i
* 4, blend_cntl
);
600 blend
->blend_enable_4bit
|= 0xfu
<< (i
* 4);
602 /* This is only important for formats without alpha. */
603 if (srcRGB
== PIPE_BLENDFACTOR_SRC_ALPHA
||
604 dstRGB
== PIPE_BLENDFACTOR_SRC_ALPHA
||
605 srcRGB
== PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
||
606 dstRGB
== PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
||
607 srcRGB
== PIPE_BLENDFACTOR_INV_SRC_ALPHA
||
608 dstRGB
== PIPE_BLENDFACTOR_INV_SRC_ALPHA
)
609 blend
->need_src_alpha_4bit
|= 0xfu
<< (i
* 4);
612 if (blend
->cb_target_mask
) {
613 color_control
|= S_028808_MODE(mode
);
615 color_control
|= S_028808_MODE(V_028808_CB_DISABLE
);
618 if (sctx
->screen
->b
.has_rbplus
) {
619 /* Disable RB+ blend optimizations for dual source blending.
622 if (blend
->dual_src_blend
) {
623 for (int i
= 0; i
< 8; i
++) {
624 sx_mrt_blend_opt
[i
] =
625 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_NONE
) |
626 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_NONE
);
630 for (int i
= 0; i
< 8; i
++)
631 si_pm4_set_reg(pm4
, R_028760_SX_MRT0_BLEND_OPT
+ i
* 4,
632 sx_mrt_blend_opt
[i
]);
634 /* RB+ doesn't work with dual source blending, logic op, and RESOLVE. */
635 if (blend
->dual_src_blend
|| state
->logicop_enable
||
636 mode
== V_028808_CB_RESOLVE
)
637 color_control
|= S_028808_DISABLE_DUAL_QUAD(1);
640 si_pm4_set_reg(pm4
, R_028808_CB_COLOR_CONTROL
, color_control
);
644 static void *si_create_blend_state(struct pipe_context
*ctx
,
645 const struct pipe_blend_state
*state
)
647 return si_create_blend_state_mode(ctx
, state
, V_028808_CB_NORMAL
);
650 static void si_bind_blend_state(struct pipe_context
*ctx
, void *state
)
652 struct si_context
*sctx
= (struct si_context
*)ctx
;
653 struct si_state_blend
*old_blend
= sctx
->queued
.named
.blend
;
654 struct si_state_blend
*blend
= (struct si_state_blend
*)state
;
660 old_blend
->cb_target_mask
!= blend
->cb_target_mask
||
661 old_blend
->dual_src_blend
!= blend
->dual_src_blend
)
662 si_mark_atom_dirty(sctx
, &sctx
->cb_render_state
);
664 si_pm4_bind_state(sctx
, blend
, state
);
667 old_blend
->cb_target_mask
!= blend
->cb_target_mask
||
668 old_blend
->alpha_to_coverage
!= blend
->alpha_to_coverage
||
669 old_blend
->alpha_to_one
!= blend
->alpha_to_one
||
670 old_blend
->dual_src_blend
!= blend
->dual_src_blend
||
671 old_blend
->blend_enable_4bit
!= blend
->blend_enable_4bit
||
672 old_blend
->need_src_alpha_4bit
!= blend
->need_src_alpha_4bit
)
673 sctx
->do_update_shaders
= true;
675 if (sctx
->screen
->dpbb_allowed
&&
677 old_blend
->alpha_to_coverage
!= blend
->alpha_to_coverage
||
678 old_blend
->blend_enable_4bit
!= blend
->blend_enable_4bit
||
679 old_blend
->cb_target_enabled_4bit
!= blend
->cb_target_enabled_4bit
))
680 si_mark_atom_dirty(sctx
, &sctx
->dpbb_state
);
682 if (sctx
->screen
->has_out_of_order_rast
&&
684 (old_blend
->blend_enable_4bit
!= blend
->blend_enable_4bit
||
685 old_blend
->cb_target_enabled_4bit
!= blend
->cb_target_enabled_4bit
||
686 old_blend
->commutative_4bit
!= blend
->commutative_4bit
||
687 old_blend
->logicop_enable
!= blend
->logicop_enable
)))
688 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
691 static void si_delete_blend_state(struct pipe_context
*ctx
, void *state
)
693 struct si_context
*sctx
= (struct si_context
*)ctx
;
694 si_pm4_delete_state(sctx
, blend
, (struct si_state_blend
*)state
);
697 static void si_set_blend_color(struct pipe_context
*ctx
,
698 const struct pipe_blend_color
*state
)
700 struct si_context
*sctx
= (struct si_context
*)ctx
;
701 static const struct pipe_blend_color zeros
;
703 sctx
->blend_color
.state
= *state
;
704 sctx
->blend_color
.any_nonzeros
= memcmp(state
, &zeros
, sizeof(*state
)) != 0;
705 si_mark_atom_dirty(sctx
, &sctx
->blend_color
.atom
);
708 static void si_emit_blend_color(struct si_context
*sctx
, struct r600_atom
*atom
)
710 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
712 radeon_set_context_reg_seq(cs
, R_028414_CB_BLEND_RED
, 4);
713 radeon_emit_array(cs
, (uint32_t*)sctx
->blend_color
.state
.color
, 4);
720 static void si_set_clip_state(struct pipe_context
*ctx
,
721 const struct pipe_clip_state
*state
)
723 struct si_context
*sctx
= (struct si_context
*)ctx
;
724 struct pipe_constant_buffer cb
;
725 static const struct pipe_clip_state zeros
;
727 if (memcmp(&sctx
->clip_state
.state
, state
, sizeof(*state
)) == 0)
730 sctx
->clip_state
.state
= *state
;
731 sctx
->clip_state
.any_nonzeros
= memcmp(state
, &zeros
, sizeof(*state
)) != 0;
732 si_mark_atom_dirty(sctx
, &sctx
->clip_state
.atom
);
735 cb
.user_buffer
= state
->ucp
;
736 cb
.buffer_offset
= 0;
737 cb
.buffer_size
= 4*4*8;
738 si_set_rw_buffer(sctx
, SI_VS_CONST_CLIP_PLANES
, &cb
);
739 pipe_resource_reference(&cb
.buffer
, NULL
);
742 static void si_emit_clip_state(struct si_context
*sctx
, struct r600_atom
*atom
)
744 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
746 radeon_set_context_reg_seq(cs
, R_0285BC_PA_CL_UCP_0_X
, 6*4);
747 radeon_emit_array(cs
, (uint32_t*)sctx
->clip_state
.state
.ucp
, 6*4);
750 static void si_emit_clip_regs(struct si_context
*sctx
, struct r600_atom
*atom
)
752 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
753 struct si_shader
*vs
= si_get_vs_state(sctx
);
754 struct si_shader_selector
*vs_sel
= vs
->selector
;
755 struct tgsi_shader_info
*info
= &vs_sel
->info
;
756 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
757 unsigned window_space
=
758 info
->properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
];
759 unsigned clipdist_mask
= vs_sel
->clipdist_mask
;
760 unsigned ucp_mask
= clipdist_mask
? 0 : rs
->clip_plane_enable
& SIX_BITS
;
761 unsigned culldist_mask
= vs_sel
->culldist_mask
;
764 if (vs
->key
.opt
.clip_disable
) {
765 assert(!info
->culldist_writemask
);
769 total_mask
= clipdist_mask
| culldist_mask
;
771 /* Clip distances on points have no effect, so need to be implemented
772 * as cull distances. This applies for the clipvertex case as well.
774 * Setting this for primitives other than points should have no adverse
777 clipdist_mask
&= rs
->clip_plane_enable
;
778 culldist_mask
|= clipdist_mask
;
780 radeon_set_context_reg(cs
, R_02881C_PA_CL_VS_OUT_CNTL
,
781 vs_sel
->pa_cl_vs_out_cntl
|
782 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask
& 0x0F) != 0) |
783 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask
& 0xF0) != 0) |
784 clipdist_mask
| (culldist_mask
<< 8));
785 radeon_set_context_reg(cs
, R_028810_PA_CL_CLIP_CNTL
,
786 rs
->pa_cl_clip_cntl
|
788 S_028810_CLIP_DISABLE(window_space
));
792 * inferred state between framebuffer and rasterizer
794 static void si_update_poly_offset_state(struct si_context
*sctx
)
796 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
798 if (!rs
|| !rs
->uses_poly_offset
|| !sctx
->framebuffer
.state
.zsbuf
) {
799 si_pm4_bind_state(sctx
, poly_offset
, NULL
);
803 /* Use the user format, not db_render_format, so that the polygon
804 * offset behaves as expected by applications.
806 switch (sctx
->framebuffer
.state
.zsbuf
->texture
->format
) {
807 case PIPE_FORMAT_Z16_UNORM
:
808 si_pm4_bind_state(sctx
, poly_offset
, &rs
->pm4_poly_offset
[0]);
810 default: /* 24-bit */
811 si_pm4_bind_state(sctx
, poly_offset
, &rs
->pm4_poly_offset
[1]);
813 case PIPE_FORMAT_Z32_FLOAT
:
814 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
815 si_pm4_bind_state(sctx
, poly_offset
, &rs
->pm4_poly_offset
[2]);
824 static uint32_t si_translate_fill(uint32_t func
)
827 case PIPE_POLYGON_MODE_FILL
:
828 return V_028814_X_DRAW_TRIANGLES
;
829 case PIPE_POLYGON_MODE_LINE
:
830 return V_028814_X_DRAW_LINES
;
831 case PIPE_POLYGON_MODE_POINT
:
832 return V_028814_X_DRAW_POINTS
;
835 return V_028814_X_DRAW_POINTS
;
839 static void *si_create_rs_state(struct pipe_context
*ctx
,
840 const struct pipe_rasterizer_state
*state
)
842 struct si_screen
*sscreen
= ((struct si_context
*)ctx
)->screen
;
843 struct si_state_rasterizer
*rs
= CALLOC_STRUCT(si_state_rasterizer
);
844 struct si_pm4_state
*pm4
= &rs
->pm4
;
846 float psize_min
, psize_max
;
852 rs
->scissor_enable
= state
->scissor
;
853 rs
->clip_halfz
= state
->clip_halfz
;
854 rs
->two_side
= state
->light_twoside
;
855 rs
->multisample_enable
= state
->multisample
;
856 rs
->force_persample_interp
= state
->force_persample_interp
;
857 rs
->clip_plane_enable
= state
->clip_plane_enable
;
858 rs
->line_stipple_enable
= state
->line_stipple_enable
;
859 rs
->poly_stipple_enable
= state
->poly_stipple_enable
;
860 rs
->line_smooth
= state
->line_smooth
;
861 rs
->line_width
= state
->line_width
;
862 rs
->poly_smooth
= state
->poly_smooth
;
863 rs
->uses_poly_offset
= state
->offset_point
|| state
->offset_line
||
865 rs
->clamp_fragment_color
= state
->clamp_fragment_color
;
866 rs
->clamp_vertex_color
= state
->clamp_vertex_color
;
867 rs
->flatshade
= state
->flatshade
;
868 rs
->sprite_coord_enable
= state
->sprite_coord_enable
;
869 rs
->rasterizer_discard
= state
->rasterizer_discard
;
870 rs
->pa_sc_line_stipple
= state
->line_stipple_enable
?
871 S_028A0C_LINE_PATTERN(state
->line_stipple_pattern
) |
872 S_028A0C_REPEAT_COUNT(state
->line_stipple_factor
) : 0;
873 rs
->pa_cl_clip_cntl
=
874 S_028810_DX_CLIP_SPACE_DEF(state
->clip_halfz
) |
875 S_028810_ZCLIP_NEAR_DISABLE(!state
->depth_clip
) |
876 S_028810_ZCLIP_FAR_DISABLE(!state
->depth_clip
) |
877 S_028810_DX_RASTERIZATION_KILL(state
->rasterizer_discard
) |
878 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
880 si_pm4_set_reg(pm4
, R_0286D4_SPI_INTERP_CONTROL_0
,
881 S_0286D4_FLAT_SHADE_ENA(1) |
882 S_0286D4_PNT_SPRITE_ENA(1) |
883 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S
) |
884 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T
) |
885 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0
) |
886 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1
) |
887 S_0286D4_PNT_SPRITE_TOP_1(state
->sprite_coord_mode
!= PIPE_SPRITE_COORD_UPPER_LEFT
));
889 /* point size 12.4 fixed point */
890 tmp
= (unsigned)(state
->point_size
* 8.0);
891 si_pm4_set_reg(pm4
, R_028A00_PA_SU_POINT_SIZE
, S_028A00_HEIGHT(tmp
) | S_028A00_WIDTH(tmp
));
893 if (state
->point_size_per_vertex
) {
894 psize_min
= util_get_min_point_size(state
);
897 /* Force the point size to be as if the vertex output was disabled. */
898 psize_min
= state
->point_size
;
899 psize_max
= state
->point_size
;
901 rs
->max_point_size
= psize_max
;
903 /* Divide by two, because 0.5 = 1 pixel. */
904 si_pm4_set_reg(pm4
, R_028A04_PA_SU_POINT_MINMAX
,
905 S_028A04_MIN_SIZE(si_pack_float_12p4(psize_min
/2)) |
906 S_028A04_MAX_SIZE(si_pack_float_12p4(psize_max
/2)));
908 si_pm4_set_reg(pm4
, R_028A08_PA_SU_LINE_CNTL
,
909 S_028A08_WIDTH(si_pack_float_12p4(state
->line_width
/2)));
910 si_pm4_set_reg(pm4
, R_028A48_PA_SC_MODE_CNTL_0
,
911 S_028A48_LINE_STIPPLE_ENABLE(state
->line_stipple_enable
) |
912 S_028A48_MSAA_ENABLE(state
->multisample
||
913 state
->poly_smooth
||
914 state
->line_smooth
) |
915 S_028A48_VPORT_SCISSOR_ENABLE(1) |
916 S_028A48_ALTERNATE_RBS_PER_TILE(sscreen
->b
.chip_class
>= GFX9
));
918 si_pm4_set_reg(pm4
, R_028BE4_PA_SU_VTX_CNTL
,
919 S_028BE4_PIX_CENTER(state
->half_pixel_center
) |
920 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH
));
922 si_pm4_set_reg(pm4
, R_028B7C_PA_SU_POLY_OFFSET_CLAMP
, fui(state
->offset_clamp
));
923 si_pm4_set_reg(pm4
, R_028814_PA_SU_SC_MODE_CNTL
,
924 S_028814_PROVOKING_VTX_LAST(!state
->flatshade_first
) |
925 S_028814_CULL_FRONT((state
->cull_face
& PIPE_FACE_FRONT
) ? 1 : 0) |
926 S_028814_CULL_BACK((state
->cull_face
& PIPE_FACE_BACK
) ? 1 : 0) |
927 S_028814_FACE(!state
->front_ccw
) |
928 S_028814_POLY_OFFSET_FRONT_ENABLE(util_get_offset(state
, state
->fill_front
)) |
929 S_028814_POLY_OFFSET_BACK_ENABLE(util_get_offset(state
, state
->fill_back
)) |
930 S_028814_POLY_OFFSET_PARA_ENABLE(state
->offset_point
|| state
->offset_line
) |
931 S_028814_POLY_MODE(state
->fill_front
!= PIPE_POLYGON_MODE_FILL
||
932 state
->fill_back
!= PIPE_POLYGON_MODE_FILL
) |
933 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(state
->fill_front
)) |
934 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(state
->fill_back
)));
936 if (!rs
->uses_poly_offset
)
939 rs
->pm4_poly_offset
= CALLOC(3, sizeof(struct si_pm4_state
));
940 if (!rs
->pm4_poly_offset
) {
945 /* Precalculate polygon offset states for 16-bit, 24-bit, and 32-bit zbuffers. */
946 for (i
= 0; i
< 3; i
++) {
947 struct si_pm4_state
*pm4
= &rs
->pm4_poly_offset
[i
];
948 float offset_units
= state
->offset_units
;
949 float offset_scale
= state
->offset_scale
* 16.0f
;
950 uint32_t pa_su_poly_offset_db_fmt_cntl
= 0;
952 if (!state
->offset_units_unscaled
) {
954 case 0: /* 16-bit zbuffer */
955 offset_units
*= 4.0f
;
956 pa_su_poly_offset_db_fmt_cntl
=
957 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
959 case 1: /* 24-bit zbuffer */
960 offset_units
*= 2.0f
;
961 pa_su_poly_offset_db_fmt_cntl
=
962 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
964 case 2: /* 32-bit zbuffer */
965 offset_units
*= 1.0f
;
966 pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
967 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
972 si_pm4_set_reg(pm4
, R_028B80_PA_SU_POLY_OFFSET_FRONT_SCALE
,
974 si_pm4_set_reg(pm4
, R_028B84_PA_SU_POLY_OFFSET_FRONT_OFFSET
,
976 si_pm4_set_reg(pm4
, R_028B88_PA_SU_POLY_OFFSET_BACK_SCALE
,
978 si_pm4_set_reg(pm4
, R_028B8C_PA_SU_POLY_OFFSET_BACK_OFFSET
,
980 si_pm4_set_reg(pm4
, R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL
,
981 pa_su_poly_offset_db_fmt_cntl
);
987 static void si_bind_rs_state(struct pipe_context
*ctx
, void *state
)
989 struct si_context
*sctx
= (struct si_context
*)ctx
;
990 struct si_state_rasterizer
*old_rs
=
991 (struct si_state_rasterizer
*)sctx
->queued
.named
.rasterizer
;
992 struct si_state_rasterizer
*rs
= (struct si_state_rasterizer
*)state
;
997 if (!old_rs
|| old_rs
->multisample_enable
!= rs
->multisample_enable
) {
998 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
1000 /* Update the small primitive filter workaround if necessary. */
1001 if (sctx
->screen
->has_msaa_sample_loc_bug
&&
1002 sctx
->framebuffer
.nr_samples
> 1)
1003 si_mark_atom_dirty(sctx
, &sctx
->msaa_sample_locs
.atom
);
1006 sctx
->current_vs_state
&= C_VS_STATE_CLAMP_VERTEX_COLOR
;
1007 sctx
->current_vs_state
|= S_VS_STATE_CLAMP_VERTEX_COLOR(rs
->clamp_vertex_color
);
1009 si_pm4_bind_state(sctx
, rasterizer
, rs
);
1010 si_update_poly_offset_state(sctx
);
1013 (old_rs
->scissor_enable
!= rs
->scissor_enable
||
1014 old_rs
->line_width
!= rs
->line_width
||
1015 old_rs
->max_point_size
!= rs
->max_point_size
)) {
1016 sctx
->scissors
.dirty_mask
= (1 << SI_MAX_VIEWPORTS
) - 1;
1017 si_mark_atom_dirty(sctx
, &sctx
->scissors
.atom
);
1021 old_rs
->clip_halfz
!= rs
->clip_halfz
) {
1022 sctx
->viewports
.depth_range_dirty_mask
= (1 << SI_MAX_VIEWPORTS
) - 1;
1023 si_mark_atom_dirty(sctx
, &sctx
->viewports
.atom
);
1027 old_rs
->clip_plane_enable
!= rs
->clip_plane_enable
||
1028 old_rs
->pa_cl_clip_cntl
!= rs
->pa_cl_clip_cntl
)
1029 si_mark_atom_dirty(sctx
, &sctx
->clip_regs
);
1031 sctx
->ia_multi_vgt_param_key
.u
.line_stipple_enabled
=
1032 rs
->line_stipple_enable
;
1035 old_rs
->clip_plane_enable
!= rs
->clip_plane_enable
||
1036 old_rs
->rasterizer_discard
!= rs
->rasterizer_discard
||
1037 old_rs
->sprite_coord_enable
!= rs
->sprite_coord_enable
||
1038 old_rs
->flatshade
!= rs
->flatshade
||
1039 old_rs
->two_side
!= rs
->two_side
||
1040 old_rs
->multisample_enable
!= rs
->multisample_enable
||
1041 old_rs
->poly_stipple_enable
!= rs
->poly_stipple_enable
||
1042 old_rs
->poly_smooth
!= rs
->poly_smooth
||
1043 old_rs
->line_smooth
!= rs
->line_smooth
||
1044 old_rs
->clamp_fragment_color
!= rs
->clamp_fragment_color
||
1045 old_rs
->force_persample_interp
!= rs
->force_persample_interp
)
1046 sctx
->do_update_shaders
= true;
1049 static void si_delete_rs_state(struct pipe_context
*ctx
, void *state
)
1051 struct si_context
*sctx
= (struct si_context
*)ctx
;
1052 struct si_state_rasterizer
*rs
= (struct si_state_rasterizer
*)state
;
1054 if (sctx
->queued
.named
.rasterizer
== state
)
1055 si_pm4_bind_state(sctx
, poly_offset
, NULL
);
1057 FREE(rs
->pm4_poly_offset
);
1058 si_pm4_delete_state(sctx
, rasterizer
, rs
);
1062 * infeered state between dsa and stencil ref
1064 static void si_emit_stencil_ref(struct si_context
*sctx
, struct r600_atom
*atom
)
1066 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
1067 struct pipe_stencil_ref
*ref
= &sctx
->stencil_ref
.state
;
1068 struct si_dsa_stencil_ref_part
*dsa
= &sctx
->stencil_ref
.dsa_part
;
1070 radeon_set_context_reg_seq(cs
, R_028430_DB_STENCILREFMASK
, 2);
1071 radeon_emit(cs
, S_028430_STENCILTESTVAL(ref
->ref_value
[0]) |
1072 S_028430_STENCILMASK(dsa
->valuemask
[0]) |
1073 S_028430_STENCILWRITEMASK(dsa
->writemask
[0]) |
1074 S_028430_STENCILOPVAL(1));
1075 radeon_emit(cs
, S_028434_STENCILTESTVAL_BF(ref
->ref_value
[1]) |
1076 S_028434_STENCILMASK_BF(dsa
->valuemask
[1]) |
1077 S_028434_STENCILWRITEMASK_BF(dsa
->writemask
[1]) |
1078 S_028434_STENCILOPVAL_BF(1));
1081 static void si_set_stencil_ref(struct pipe_context
*ctx
,
1082 const struct pipe_stencil_ref
*state
)
1084 struct si_context
*sctx
= (struct si_context
*)ctx
;
1086 if (memcmp(&sctx
->stencil_ref
.state
, state
, sizeof(*state
)) == 0)
1089 sctx
->stencil_ref
.state
= *state
;
1090 si_mark_atom_dirty(sctx
, &sctx
->stencil_ref
.atom
);
1098 static uint32_t si_translate_stencil_op(int s_op
)
1101 case PIPE_STENCIL_OP_KEEP
:
1102 return V_02842C_STENCIL_KEEP
;
1103 case PIPE_STENCIL_OP_ZERO
:
1104 return V_02842C_STENCIL_ZERO
;
1105 case PIPE_STENCIL_OP_REPLACE
:
1106 return V_02842C_STENCIL_REPLACE_TEST
;
1107 case PIPE_STENCIL_OP_INCR
:
1108 return V_02842C_STENCIL_ADD_CLAMP
;
1109 case PIPE_STENCIL_OP_DECR
:
1110 return V_02842C_STENCIL_SUB_CLAMP
;
1111 case PIPE_STENCIL_OP_INCR_WRAP
:
1112 return V_02842C_STENCIL_ADD_WRAP
;
1113 case PIPE_STENCIL_OP_DECR_WRAP
:
1114 return V_02842C_STENCIL_SUB_WRAP
;
1115 case PIPE_STENCIL_OP_INVERT
:
1116 return V_02842C_STENCIL_INVERT
;
1118 R600_ERR("Unknown stencil op %d", s_op
);
1125 static bool si_dsa_writes_stencil(const struct pipe_stencil_state
*s
)
1127 return s
->enabled
&& s
->writemask
&&
1128 (s
->fail_op
!= PIPE_STENCIL_OP_KEEP
||
1129 s
->zfail_op
!= PIPE_STENCIL_OP_KEEP
||
1130 s
->zpass_op
!= PIPE_STENCIL_OP_KEEP
);
1133 static bool si_order_invariant_stencil_op(enum pipe_stencil_op op
)
1135 /* REPLACE is normally order invariant, except when the stencil
1136 * reference value is written by the fragment shader. Tracking this
1137 * interaction does not seem worth the effort, so be conservative. */
1138 return op
!= PIPE_STENCIL_OP_INCR
&&
1139 op
!= PIPE_STENCIL_OP_DECR
&&
1140 op
!= PIPE_STENCIL_OP_REPLACE
;
1143 /* Compute whether, assuming Z writes are disabled, this stencil state is order
1144 * invariant in the sense that the set of passing fragments as well as the
1145 * final stencil buffer result does not depend on the order of fragments. */
1146 static bool si_order_invariant_stencil_state(const struct pipe_stencil_state
*state
)
1148 return !state
->enabled
|| !state
->writemask
||
1149 /* The following assumes that Z writes are disabled. */
1150 (state
->func
== PIPE_FUNC_ALWAYS
&&
1151 si_order_invariant_stencil_op(state
->zpass_op
) &&
1152 si_order_invariant_stencil_op(state
->zfail_op
)) ||
1153 (state
->func
== PIPE_FUNC_NEVER
&&
1154 si_order_invariant_stencil_op(state
->fail_op
));
1157 static void *si_create_dsa_state(struct pipe_context
*ctx
,
1158 const struct pipe_depth_stencil_alpha_state
*state
)
1160 struct si_context
*sctx
= (struct si_context
*)ctx
;
1161 struct si_state_dsa
*dsa
= CALLOC_STRUCT(si_state_dsa
);
1162 struct si_pm4_state
*pm4
= &dsa
->pm4
;
1163 unsigned db_depth_control
;
1164 uint32_t db_stencil_control
= 0;
1170 dsa
->stencil_ref
.valuemask
[0] = state
->stencil
[0].valuemask
;
1171 dsa
->stencil_ref
.valuemask
[1] = state
->stencil
[1].valuemask
;
1172 dsa
->stencil_ref
.writemask
[0] = state
->stencil
[0].writemask
;
1173 dsa
->stencil_ref
.writemask
[1] = state
->stencil
[1].writemask
;
1175 db_depth_control
= S_028800_Z_ENABLE(state
->depth
.enabled
) |
1176 S_028800_Z_WRITE_ENABLE(state
->depth
.writemask
) |
1177 S_028800_ZFUNC(state
->depth
.func
) |
1178 S_028800_DEPTH_BOUNDS_ENABLE(state
->depth
.bounds_test
);
1181 if (state
->stencil
[0].enabled
) {
1182 db_depth_control
|= S_028800_STENCIL_ENABLE(1);
1183 db_depth_control
|= S_028800_STENCILFUNC(state
->stencil
[0].func
);
1184 db_stencil_control
|= S_02842C_STENCILFAIL(si_translate_stencil_op(state
->stencil
[0].fail_op
));
1185 db_stencil_control
|= S_02842C_STENCILZPASS(si_translate_stencil_op(state
->stencil
[0].zpass_op
));
1186 db_stencil_control
|= S_02842C_STENCILZFAIL(si_translate_stencil_op(state
->stencil
[0].zfail_op
));
1188 if (state
->stencil
[1].enabled
) {
1189 db_depth_control
|= S_028800_BACKFACE_ENABLE(1);
1190 db_depth_control
|= S_028800_STENCILFUNC_BF(state
->stencil
[1].func
);
1191 db_stencil_control
|= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(state
->stencil
[1].fail_op
));
1192 db_stencil_control
|= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(state
->stencil
[1].zpass_op
));
1193 db_stencil_control
|= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(state
->stencil
[1].zfail_op
));
1198 if (state
->alpha
.enabled
) {
1199 dsa
->alpha_func
= state
->alpha
.func
;
1201 si_pm4_set_reg(pm4
, R_00B030_SPI_SHADER_USER_DATA_PS_0
+
1202 SI_SGPR_ALPHA_REF
* 4, fui(state
->alpha
.ref_value
));
1204 dsa
->alpha_func
= PIPE_FUNC_ALWAYS
;
1207 si_pm4_set_reg(pm4
, R_028800_DB_DEPTH_CONTROL
, db_depth_control
);
1208 if (state
->stencil
[0].enabled
)
1209 si_pm4_set_reg(pm4
, R_02842C_DB_STENCIL_CONTROL
, db_stencil_control
);
1210 if (state
->depth
.bounds_test
) {
1211 si_pm4_set_reg(pm4
, R_028020_DB_DEPTH_BOUNDS_MIN
, fui(state
->depth
.bounds_min
));
1212 si_pm4_set_reg(pm4
, R_028024_DB_DEPTH_BOUNDS_MAX
, fui(state
->depth
.bounds_max
));
1215 dsa
->depth_enabled
= state
->depth
.enabled
;
1216 dsa
->depth_write_enabled
= state
->depth
.enabled
&&
1217 state
->depth
.writemask
;
1218 dsa
->stencil_enabled
= state
->stencil
[0].enabled
;
1219 dsa
->stencil_write_enabled
= state
->stencil
[0].enabled
&&
1220 (si_dsa_writes_stencil(&state
->stencil
[0]) ||
1221 si_dsa_writes_stencil(&state
->stencil
[1]));
1222 dsa
->db_can_write
= dsa
->depth_write_enabled
||
1223 dsa
->stencil_write_enabled
;
1225 bool zfunc_is_ordered
=
1226 state
->depth
.func
== PIPE_FUNC_NEVER
||
1227 state
->depth
.func
== PIPE_FUNC_LESS
||
1228 state
->depth
.func
== PIPE_FUNC_LEQUAL
||
1229 state
->depth
.func
== PIPE_FUNC_GREATER
||
1230 state
->depth
.func
== PIPE_FUNC_GEQUAL
;
1232 bool nozwrite_and_order_invariant_stencil
=
1233 !dsa
->db_can_write
||
1234 (!dsa
->depth_write_enabled
&&
1235 si_order_invariant_stencil_state(&state
->stencil
[0]) &&
1236 si_order_invariant_stencil_state(&state
->stencil
[1]));
1238 dsa
->order_invariance
[1].zs
=
1239 nozwrite_and_order_invariant_stencil
||
1240 (!dsa
->stencil_write_enabled
&& zfunc_is_ordered
);
1241 dsa
->order_invariance
[0].zs
= !dsa
->depth_write_enabled
|| zfunc_is_ordered
;
1243 dsa
->order_invariance
[1].pass_set
=
1244 nozwrite_and_order_invariant_stencil
||
1245 (!dsa
->stencil_write_enabled
&&
1246 (state
->depth
.func
== PIPE_FUNC_ALWAYS
||
1247 state
->depth
.func
== PIPE_FUNC_NEVER
));
1248 dsa
->order_invariance
[0].pass_set
=
1249 !dsa
->depth_write_enabled
||
1250 (state
->depth
.func
== PIPE_FUNC_ALWAYS
||
1251 state
->depth
.func
== PIPE_FUNC_NEVER
);
1253 dsa
->order_invariance
[1].pass_last
=
1254 sctx
->screen
->assume_no_z_fights
&&
1255 !dsa
->stencil_write_enabled
&&
1256 dsa
->depth_write_enabled
&& zfunc_is_ordered
;
1257 dsa
->order_invariance
[0].pass_last
=
1258 sctx
->screen
->assume_no_z_fights
&&
1259 dsa
->depth_write_enabled
&& zfunc_is_ordered
;
1264 static void si_bind_dsa_state(struct pipe_context
*ctx
, void *state
)
1266 struct si_context
*sctx
= (struct si_context
*)ctx
;
1267 struct si_state_dsa
*old_dsa
= sctx
->queued
.named
.dsa
;
1268 struct si_state_dsa
*dsa
= state
;
1273 si_pm4_bind_state(sctx
, dsa
, dsa
);
1275 if (memcmp(&dsa
->stencil_ref
, &sctx
->stencil_ref
.dsa_part
,
1276 sizeof(struct si_dsa_stencil_ref_part
)) != 0) {
1277 sctx
->stencil_ref
.dsa_part
= dsa
->stencil_ref
;
1278 si_mark_atom_dirty(sctx
, &sctx
->stencil_ref
.atom
);
1281 if (!old_dsa
|| old_dsa
->alpha_func
!= dsa
->alpha_func
)
1282 sctx
->do_update_shaders
= true;
1284 if (sctx
->screen
->dpbb_allowed
&&
1286 (old_dsa
->depth_enabled
!= dsa
->depth_enabled
||
1287 old_dsa
->stencil_enabled
!= dsa
->stencil_enabled
||
1288 old_dsa
->db_can_write
!= dsa
->db_can_write
)))
1289 si_mark_atom_dirty(sctx
, &sctx
->dpbb_state
);
1291 if (sctx
->screen
->has_out_of_order_rast
&&
1293 memcmp(old_dsa
->order_invariance
, dsa
->order_invariance
,
1294 sizeof(old_dsa
->order_invariance
))))
1295 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
1298 static void si_delete_dsa_state(struct pipe_context
*ctx
, void *state
)
1300 struct si_context
*sctx
= (struct si_context
*)ctx
;
1301 si_pm4_delete_state(sctx
, dsa
, (struct si_state_dsa
*)state
);
1304 static void *si_create_db_flush_dsa(struct si_context
*sctx
)
1306 struct pipe_depth_stencil_alpha_state dsa
= {};
1308 return sctx
->b
.b
.create_depth_stencil_alpha_state(&sctx
->b
.b
, &dsa
);
1311 /* DB RENDER STATE */
1313 static void si_set_active_query_state(struct pipe_context
*ctx
, boolean enable
)
1315 struct si_context
*sctx
= (struct si_context
*)ctx
;
1317 /* Pipeline stat & streamout queries. */
1319 sctx
->b
.flags
&= ~R600_CONTEXT_STOP_PIPELINE_STATS
;
1320 sctx
->b
.flags
|= R600_CONTEXT_START_PIPELINE_STATS
;
1322 sctx
->b
.flags
&= ~R600_CONTEXT_START_PIPELINE_STATS
;
1323 sctx
->b
.flags
|= R600_CONTEXT_STOP_PIPELINE_STATS
;
1326 /* Occlusion queries. */
1327 if (sctx
->occlusion_queries_disabled
!= !enable
) {
1328 sctx
->occlusion_queries_disabled
= !enable
;
1329 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
1333 static void si_set_occlusion_query_state(struct pipe_context
*ctx
,
1335 bool old_perfect_enable
)
1337 struct si_context
*sctx
= (struct si_context
*)ctx
;
1339 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
1341 bool perfect_enable
= sctx
->b
.num_perfect_occlusion_queries
!= 0;
1343 if (perfect_enable
!= old_perfect_enable
)
1344 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
1347 static void si_save_qbo_state(struct pipe_context
*ctx
, struct r600_qbo_state
*st
)
1349 struct si_context
*sctx
= (struct si_context
*)ctx
;
1351 st
->saved_compute
= sctx
->cs_shader_state
.program
;
1353 si_get_pipe_constant_buffer(sctx
, PIPE_SHADER_COMPUTE
, 0, &st
->saved_const0
);
1354 si_get_shader_buffers(sctx
, PIPE_SHADER_COMPUTE
, 0, 3, st
->saved_ssbo
);
1357 static void si_emit_db_render_state(struct si_context
*sctx
, struct r600_atom
*state
)
1359 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
1360 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1361 unsigned db_shader_control
;
1363 radeon_set_context_reg_seq(cs
, R_028000_DB_RENDER_CONTROL
, 2);
1365 /* DB_RENDER_CONTROL */
1366 if (sctx
->dbcb_depth_copy_enabled
||
1367 sctx
->dbcb_stencil_copy_enabled
) {
1369 S_028000_DEPTH_COPY(sctx
->dbcb_depth_copy_enabled
) |
1370 S_028000_STENCIL_COPY(sctx
->dbcb_stencil_copy_enabled
) |
1371 S_028000_COPY_CENTROID(1) |
1372 S_028000_COPY_SAMPLE(sctx
->dbcb_copy_sample
));
1373 } else if (sctx
->db_flush_depth_inplace
|| sctx
->db_flush_stencil_inplace
) {
1375 S_028000_DEPTH_COMPRESS_DISABLE(sctx
->db_flush_depth_inplace
) |
1376 S_028000_STENCIL_COMPRESS_DISABLE(sctx
->db_flush_stencil_inplace
));
1379 S_028000_DEPTH_CLEAR_ENABLE(sctx
->db_depth_clear
) |
1380 S_028000_STENCIL_CLEAR_ENABLE(sctx
->db_stencil_clear
));
1383 /* DB_COUNT_CONTROL (occlusion queries) */
1384 if (sctx
->b
.num_occlusion_queries
> 0 &&
1385 !sctx
->occlusion_queries_disabled
) {
1386 bool perfect
= sctx
->b
.num_perfect_occlusion_queries
> 0;
1388 if (sctx
->b
.chip_class
>= CIK
) {
1390 S_028004_PERFECT_ZPASS_COUNTS(perfect
) |
1391 S_028004_SAMPLE_RATE(sctx
->framebuffer
.log_samples
) |
1392 S_028004_ZPASS_ENABLE(1) |
1393 S_028004_SLICE_EVEN_ENABLE(1) |
1394 S_028004_SLICE_ODD_ENABLE(1));
1397 S_028004_PERFECT_ZPASS_COUNTS(perfect
) |
1398 S_028004_SAMPLE_RATE(sctx
->framebuffer
.log_samples
));
1401 /* Disable occlusion queries. */
1402 if (sctx
->b
.chip_class
>= CIK
) {
1405 radeon_emit(cs
, S_028004_ZPASS_INCREMENT_DISABLE(1));
1409 /* DB_RENDER_OVERRIDE2 */
1410 radeon_set_context_reg(cs
, R_028010_DB_RENDER_OVERRIDE2
,
1411 S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(sctx
->db_depth_disable_expclear
) |
1412 S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(sctx
->db_stencil_disable_expclear
) |
1413 S_028010_DECOMPRESS_Z_ON_FLUSH(sctx
->framebuffer
.nr_samples
>= 4));
1415 db_shader_control
= sctx
->ps_db_shader_control
;
1417 /* Bug workaround for smoothing (overrasterization) on SI. */
1418 if (sctx
->b
.chip_class
== SI
&& sctx
->smoothing_enabled
) {
1419 db_shader_control
&= C_02880C_Z_ORDER
;
1420 db_shader_control
|= S_02880C_Z_ORDER(V_02880C_LATE_Z
);
1423 /* Disable the gl_SampleMask fragment shader output if MSAA is disabled. */
1424 if (!rs
|| !rs
->multisample_enable
)
1425 db_shader_control
&= C_02880C_MASK_EXPORT_ENABLE
;
1427 if (sctx
->screen
->b
.has_rbplus
&&
1428 !sctx
->screen
->b
.rbplus_allowed
)
1429 db_shader_control
|= S_02880C_DUAL_QUAD_DISABLE(1);
1431 radeon_set_context_reg(cs
, R_02880C_DB_SHADER_CONTROL
,
1436 * format translation
1438 static uint32_t si_translate_colorformat(enum pipe_format format
)
1440 const struct util_format_description
*desc
= util_format_description(format
);
1442 return V_028C70_COLOR_INVALID
;
1444 #define HAS_SIZE(x,y,z,w) \
1445 (desc->channel[0].size == (x) && desc->channel[1].size == (y) && \
1446 desc->channel[2].size == (z) && desc->channel[3].size == (w))
1448 if (format
== PIPE_FORMAT_R11G11B10_FLOAT
) /* isn't plain */
1449 return V_028C70_COLOR_10_11_11
;
1451 if (desc
->layout
!= UTIL_FORMAT_LAYOUT_PLAIN
)
1452 return V_028C70_COLOR_INVALID
;
1454 /* hw cannot support mixed formats (except depth/stencil, since
1455 * stencil is not written to). */
1456 if (desc
->is_mixed
&& desc
->colorspace
!= UTIL_FORMAT_COLORSPACE_ZS
)
1457 return V_028C70_COLOR_INVALID
;
1459 switch (desc
->nr_channels
) {
1461 switch (desc
->channel
[0].size
) {
1463 return V_028C70_COLOR_8
;
1465 return V_028C70_COLOR_16
;
1467 return V_028C70_COLOR_32
;
1471 if (desc
->channel
[0].size
== desc
->channel
[1].size
) {
1472 switch (desc
->channel
[0].size
) {
1474 return V_028C70_COLOR_8_8
;
1476 return V_028C70_COLOR_16_16
;
1478 return V_028C70_COLOR_32_32
;
1480 } else if (HAS_SIZE(8,24,0,0)) {
1481 return V_028C70_COLOR_24_8
;
1482 } else if (HAS_SIZE(24,8,0,0)) {
1483 return V_028C70_COLOR_8_24
;
1487 if (HAS_SIZE(5,6,5,0)) {
1488 return V_028C70_COLOR_5_6_5
;
1489 } else if (HAS_SIZE(32,8,24,0)) {
1490 return V_028C70_COLOR_X24_8_32_FLOAT
;
1494 if (desc
->channel
[0].size
== desc
->channel
[1].size
&&
1495 desc
->channel
[0].size
== desc
->channel
[2].size
&&
1496 desc
->channel
[0].size
== desc
->channel
[3].size
) {
1497 switch (desc
->channel
[0].size
) {
1499 return V_028C70_COLOR_4_4_4_4
;
1501 return V_028C70_COLOR_8_8_8_8
;
1503 return V_028C70_COLOR_16_16_16_16
;
1505 return V_028C70_COLOR_32_32_32_32
;
1507 } else if (HAS_SIZE(5,5,5,1)) {
1508 return V_028C70_COLOR_1_5_5_5
;
1509 } else if (HAS_SIZE(10,10,10,2)) {
1510 return V_028C70_COLOR_2_10_10_10
;
1514 return V_028C70_COLOR_INVALID
;
1517 static uint32_t si_colorformat_endian_swap(uint32_t colorformat
)
1519 if (SI_BIG_ENDIAN
) {
1520 switch(colorformat
) {
1521 /* 8-bit buffers. */
1522 case V_028C70_COLOR_8
:
1523 return V_028C70_ENDIAN_NONE
;
1525 /* 16-bit buffers. */
1526 case V_028C70_COLOR_5_6_5
:
1527 case V_028C70_COLOR_1_5_5_5
:
1528 case V_028C70_COLOR_4_4_4_4
:
1529 case V_028C70_COLOR_16
:
1530 case V_028C70_COLOR_8_8
:
1531 return V_028C70_ENDIAN_8IN16
;
1533 /* 32-bit buffers. */
1534 case V_028C70_COLOR_8_8_8_8
:
1535 case V_028C70_COLOR_2_10_10_10
:
1536 case V_028C70_COLOR_8_24
:
1537 case V_028C70_COLOR_24_8
:
1538 case V_028C70_COLOR_16_16
:
1539 return V_028C70_ENDIAN_8IN32
;
1541 /* 64-bit buffers. */
1542 case V_028C70_COLOR_16_16_16_16
:
1543 return V_028C70_ENDIAN_8IN16
;
1545 case V_028C70_COLOR_32_32
:
1546 return V_028C70_ENDIAN_8IN32
;
1548 /* 128-bit buffers. */
1549 case V_028C70_COLOR_32_32_32_32
:
1550 return V_028C70_ENDIAN_8IN32
;
1552 return V_028C70_ENDIAN_NONE
; /* Unsupported. */
1555 return V_028C70_ENDIAN_NONE
;
1559 static uint32_t si_translate_dbformat(enum pipe_format format
)
1562 case PIPE_FORMAT_Z16_UNORM
:
1563 return V_028040_Z_16
;
1564 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
1565 case PIPE_FORMAT_X8Z24_UNORM
:
1566 case PIPE_FORMAT_Z24X8_UNORM
:
1567 case PIPE_FORMAT_Z24_UNORM_S8_UINT
:
1568 return V_028040_Z_24
; /* deprecated on SI */
1569 case PIPE_FORMAT_Z32_FLOAT
:
1570 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
1571 return V_028040_Z_32_FLOAT
;
1573 return V_028040_Z_INVALID
;
1578 * Texture translation
1581 static uint32_t si_translate_texformat(struct pipe_screen
*screen
,
1582 enum pipe_format format
,
1583 const struct util_format_description
*desc
,
1586 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
1587 bool enable_compressed_formats
= (sscreen
->b
.info
.drm_major
== 2 &&
1588 sscreen
->b
.info
.drm_minor
>= 31) ||
1589 sscreen
->b
.info
.drm_major
== 3;
1590 bool uniform
= true;
1593 /* Colorspace (return non-RGB formats directly). */
1594 switch (desc
->colorspace
) {
1595 /* Depth stencil formats */
1596 case UTIL_FORMAT_COLORSPACE_ZS
:
1598 case PIPE_FORMAT_Z16_UNORM
:
1599 return V_008F14_IMG_DATA_FORMAT_16
;
1600 case PIPE_FORMAT_X24S8_UINT
:
1601 case PIPE_FORMAT_S8X24_UINT
:
1603 * Implemented as an 8_8_8_8 data format to fix texture
1604 * gathers in stencil sampling. This affects at least
1605 * GL45-CTS.texture_cube_map_array.sampling on VI.
1607 return V_008F14_IMG_DATA_FORMAT_8_8_8_8
;
1608 case PIPE_FORMAT_Z24X8_UNORM
:
1609 case PIPE_FORMAT_Z24_UNORM_S8_UINT
:
1610 return V_008F14_IMG_DATA_FORMAT_8_24
;
1611 case PIPE_FORMAT_X8Z24_UNORM
:
1612 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
1613 return V_008F14_IMG_DATA_FORMAT_24_8
;
1614 case PIPE_FORMAT_S8_UINT
:
1615 return V_008F14_IMG_DATA_FORMAT_8
;
1616 case PIPE_FORMAT_Z32_FLOAT
:
1617 return V_008F14_IMG_DATA_FORMAT_32
;
1618 case PIPE_FORMAT_X32_S8X24_UINT
:
1619 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
1620 return V_008F14_IMG_DATA_FORMAT_X24_8_32
;
1625 case UTIL_FORMAT_COLORSPACE_YUV
:
1626 goto out_unknown
; /* TODO */
1628 case UTIL_FORMAT_COLORSPACE_SRGB
:
1629 if (desc
->nr_channels
!= 4 && desc
->nr_channels
!= 1)
1637 if (desc
->layout
== UTIL_FORMAT_LAYOUT_RGTC
) {
1638 if (!enable_compressed_formats
)
1642 case PIPE_FORMAT_RGTC1_SNORM
:
1643 case PIPE_FORMAT_LATC1_SNORM
:
1644 case PIPE_FORMAT_RGTC1_UNORM
:
1645 case PIPE_FORMAT_LATC1_UNORM
:
1646 return V_008F14_IMG_DATA_FORMAT_BC4
;
1647 case PIPE_FORMAT_RGTC2_SNORM
:
1648 case PIPE_FORMAT_LATC2_SNORM
:
1649 case PIPE_FORMAT_RGTC2_UNORM
:
1650 case PIPE_FORMAT_LATC2_UNORM
:
1651 return V_008F14_IMG_DATA_FORMAT_BC5
;
1657 if (desc
->layout
== UTIL_FORMAT_LAYOUT_ETC
&&
1658 (sscreen
->b
.family
== CHIP_STONEY
||
1659 sscreen
->b
.chip_class
>= GFX9
)) {
1661 case PIPE_FORMAT_ETC1_RGB8
:
1662 case PIPE_FORMAT_ETC2_RGB8
:
1663 case PIPE_FORMAT_ETC2_SRGB8
:
1664 return V_008F14_IMG_DATA_FORMAT_ETC2_RGB
;
1665 case PIPE_FORMAT_ETC2_RGB8A1
:
1666 case PIPE_FORMAT_ETC2_SRGB8A1
:
1667 return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA1
;
1668 case PIPE_FORMAT_ETC2_RGBA8
:
1669 case PIPE_FORMAT_ETC2_SRGBA8
:
1670 return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA
;
1671 case PIPE_FORMAT_ETC2_R11_UNORM
:
1672 case PIPE_FORMAT_ETC2_R11_SNORM
:
1673 return V_008F14_IMG_DATA_FORMAT_ETC2_R
;
1674 case PIPE_FORMAT_ETC2_RG11_UNORM
:
1675 case PIPE_FORMAT_ETC2_RG11_SNORM
:
1676 return V_008F14_IMG_DATA_FORMAT_ETC2_RG
;
1682 if (desc
->layout
== UTIL_FORMAT_LAYOUT_BPTC
) {
1683 if (!enable_compressed_formats
)
1687 case PIPE_FORMAT_BPTC_RGBA_UNORM
:
1688 case PIPE_FORMAT_BPTC_SRGBA
:
1689 return V_008F14_IMG_DATA_FORMAT_BC7
;
1690 case PIPE_FORMAT_BPTC_RGB_FLOAT
:
1691 case PIPE_FORMAT_BPTC_RGB_UFLOAT
:
1692 return V_008F14_IMG_DATA_FORMAT_BC6
;
1698 if (desc
->layout
== UTIL_FORMAT_LAYOUT_SUBSAMPLED
) {
1700 case PIPE_FORMAT_R8G8_B8G8_UNORM
:
1701 case PIPE_FORMAT_G8R8_B8R8_UNORM
:
1702 return V_008F14_IMG_DATA_FORMAT_GB_GR
;
1703 case PIPE_FORMAT_G8R8_G8B8_UNORM
:
1704 case PIPE_FORMAT_R8G8_R8B8_UNORM
:
1705 return V_008F14_IMG_DATA_FORMAT_BG_RG
;
1711 if (desc
->layout
== UTIL_FORMAT_LAYOUT_S3TC
) {
1712 if (!enable_compressed_formats
)
1715 if (!util_format_s3tc_enabled
) {
1720 case PIPE_FORMAT_DXT1_RGB
:
1721 case PIPE_FORMAT_DXT1_RGBA
:
1722 case PIPE_FORMAT_DXT1_SRGB
:
1723 case PIPE_FORMAT_DXT1_SRGBA
:
1724 return V_008F14_IMG_DATA_FORMAT_BC1
;
1725 case PIPE_FORMAT_DXT3_RGBA
:
1726 case PIPE_FORMAT_DXT3_SRGBA
:
1727 return V_008F14_IMG_DATA_FORMAT_BC2
;
1728 case PIPE_FORMAT_DXT5_RGBA
:
1729 case PIPE_FORMAT_DXT5_SRGBA
:
1730 return V_008F14_IMG_DATA_FORMAT_BC3
;
1736 if (format
== PIPE_FORMAT_R9G9B9E5_FLOAT
) {
1737 return V_008F14_IMG_DATA_FORMAT_5_9_9_9
;
1738 } else if (format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1739 return V_008F14_IMG_DATA_FORMAT_10_11_11
;
1742 /* R8G8Bx_SNORM - TODO CxV8U8 */
1744 /* hw cannot support mixed formats (except depth/stencil, since only
1746 if (desc
->is_mixed
&& desc
->colorspace
!= UTIL_FORMAT_COLORSPACE_ZS
)
1749 /* See whether the components are of the same size. */
1750 for (i
= 1; i
< desc
->nr_channels
; i
++) {
1751 uniform
= uniform
&& desc
->channel
[0].size
== desc
->channel
[i
].size
;
1754 /* Non-uniform formats. */
1756 switch(desc
->nr_channels
) {
1758 if (desc
->channel
[0].size
== 5 &&
1759 desc
->channel
[1].size
== 6 &&
1760 desc
->channel
[2].size
== 5) {
1761 return V_008F14_IMG_DATA_FORMAT_5_6_5
;
1765 if (desc
->channel
[0].size
== 5 &&
1766 desc
->channel
[1].size
== 5 &&
1767 desc
->channel
[2].size
== 5 &&
1768 desc
->channel
[3].size
== 1) {
1769 return V_008F14_IMG_DATA_FORMAT_1_5_5_5
;
1771 if (desc
->channel
[0].size
== 10 &&
1772 desc
->channel
[1].size
== 10 &&
1773 desc
->channel
[2].size
== 10 &&
1774 desc
->channel
[3].size
== 2) {
1775 return V_008F14_IMG_DATA_FORMAT_2_10_10_10
;
1782 if (first_non_void
< 0 || first_non_void
> 3)
1785 /* uniform formats */
1786 switch (desc
->channel
[first_non_void
].size
) {
1788 switch (desc
->nr_channels
) {
1789 #if 0 /* Not supported for render targets */
1791 return V_008F14_IMG_DATA_FORMAT_4_4
;
1794 return V_008F14_IMG_DATA_FORMAT_4_4_4_4
;
1798 switch (desc
->nr_channels
) {
1800 return V_008F14_IMG_DATA_FORMAT_8
;
1802 return V_008F14_IMG_DATA_FORMAT_8_8
;
1804 return V_008F14_IMG_DATA_FORMAT_8_8_8_8
;
1808 switch (desc
->nr_channels
) {
1810 return V_008F14_IMG_DATA_FORMAT_16
;
1812 return V_008F14_IMG_DATA_FORMAT_16_16
;
1814 return V_008F14_IMG_DATA_FORMAT_16_16_16_16
;
1818 switch (desc
->nr_channels
) {
1820 return V_008F14_IMG_DATA_FORMAT_32
;
1822 return V_008F14_IMG_DATA_FORMAT_32_32
;
1823 #if 0 /* Not supported for render targets */
1825 return V_008F14_IMG_DATA_FORMAT_32_32_32
;
1828 return V_008F14_IMG_DATA_FORMAT_32_32_32_32
;
1833 /* R600_ERR("Unable to handle texformat %d %s\n", format, util_format_name(format)); */
1837 static unsigned si_tex_wrap(unsigned wrap
)
1841 case PIPE_TEX_WRAP_REPEAT
:
1842 return V_008F30_SQ_TEX_WRAP
;
1843 case PIPE_TEX_WRAP_CLAMP
:
1844 return V_008F30_SQ_TEX_CLAMP_HALF_BORDER
;
1845 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
1846 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
1847 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
1848 return V_008F30_SQ_TEX_CLAMP_BORDER
;
1849 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
1850 return V_008F30_SQ_TEX_MIRROR
;
1851 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
1852 return V_008F30_SQ_TEX_MIRROR_ONCE_HALF_BORDER
;
1853 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
1854 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
1855 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
1856 return V_008F30_SQ_TEX_MIRROR_ONCE_BORDER
;
1860 static unsigned si_tex_mipfilter(unsigned filter
)
1863 case PIPE_TEX_MIPFILTER_NEAREST
:
1864 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
1865 case PIPE_TEX_MIPFILTER_LINEAR
:
1866 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
1868 case PIPE_TEX_MIPFILTER_NONE
:
1869 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
1873 static unsigned si_tex_compare(unsigned compare
)
1877 case PIPE_FUNC_NEVER
:
1878 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
1879 case PIPE_FUNC_LESS
:
1880 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
1881 case PIPE_FUNC_EQUAL
:
1882 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
1883 case PIPE_FUNC_LEQUAL
:
1884 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
1885 case PIPE_FUNC_GREATER
:
1886 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
1887 case PIPE_FUNC_NOTEQUAL
:
1888 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
1889 case PIPE_FUNC_GEQUAL
:
1890 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
1891 case PIPE_FUNC_ALWAYS
:
1892 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
1896 static unsigned si_tex_dim(struct si_screen
*sscreen
, struct r600_texture
*rtex
,
1897 unsigned view_target
, unsigned nr_samples
)
1899 unsigned res_target
= rtex
->resource
.b
.b
.target
;
1901 if (view_target
== PIPE_TEXTURE_CUBE
||
1902 view_target
== PIPE_TEXTURE_CUBE_ARRAY
)
1903 res_target
= view_target
;
1904 /* If interpreting cubemaps as something else, set 2D_ARRAY. */
1905 else if (res_target
== PIPE_TEXTURE_CUBE
||
1906 res_target
== PIPE_TEXTURE_CUBE_ARRAY
)
1907 res_target
= PIPE_TEXTURE_2D_ARRAY
;
1909 /* GFX9 allocates 1D textures as 2D. */
1910 if ((res_target
== PIPE_TEXTURE_1D
||
1911 res_target
== PIPE_TEXTURE_1D_ARRAY
) &&
1912 sscreen
->b
.chip_class
>= GFX9
&&
1913 rtex
->surface
.u
.gfx9
.resource_type
== RADEON_RESOURCE_2D
) {
1914 if (res_target
== PIPE_TEXTURE_1D
)
1915 res_target
= PIPE_TEXTURE_2D
;
1917 res_target
= PIPE_TEXTURE_2D_ARRAY
;
1920 switch (res_target
) {
1922 case PIPE_TEXTURE_1D
:
1923 return V_008F1C_SQ_RSRC_IMG_1D
;
1924 case PIPE_TEXTURE_1D_ARRAY
:
1925 return V_008F1C_SQ_RSRC_IMG_1D_ARRAY
;
1926 case PIPE_TEXTURE_2D
:
1927 case PIPE_TEXTURE_RECT
:
1928 return nr_samples
> 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA
:
1929 V_008F1C_SQ_RSRC_IMG_2D
;
1930 case PIPE_TEXTURE_2D_ARRAY
:
1931 return nr_samples
> 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
:
1932 V_008F1C_SQ_RSRC_IMG_2D_ARRAY
;
1933 case PIPE_TEXTURE_3D
:
1934 return V_008F1C_SQ_RSRC_IMG_3D
;
1935 case PIPE_TEXTURE_CUBE
:
1936 case PIPE_TEXTURE_CUBE_ARRAY
:
1937 return V_008F1C_SQ_RSRC_IMG_CUBE
;
1942 * Format support testing
1945 static bool si_is_sampler_format_supported(struct pipe_screen
*screen
, enum pipe_format format
)
1947 const struct util_format_description
*desc
= util_format_description(format
);
1951 return si_translate_texformat(screen
, format
, desc
,
1952 util_format_get_first_non_void_channel(format
)) != ~0U;
1955 static uint32_t si_translate_buffer_dataformat(struct pipe_screen
*screen
,
1956 const struct util_format_description
*desc
,
1961 if (desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
)
1962 return V_008F0C_BUF_DATA_FORMAT_10_11_11
;
1964 assert(first_non_void
>= 0);
1966 if (desc
->nr_channels
== 4 &&
1967 desc
->channel
[0].size
== 10 &&
1968 desc
->channel
[1].size
== 10 &&
1969 desc
->channel
[2].size
== 10 &&
1970 desc
->channel
[3].size
== 2)
1971 return V_008F0C_BUF_DATA_FORMAT_2_10_10_10
;
1973 /* See whether the components are of the same size. */
1974 for (i
= 0; i
< desc
->nr_channels
; i
++) {
1975 if (desc
->channel
[first_non_void
].size
!= desc
->channel
[i
].size
)
1976 return V_008F0C_BUF_DATA_FORMAT_INVALID
;
1979 switch (desc
->channel
[first_non_void
].size
) {
1981 switch (desc
->nr_channels
) {
1983 case 3: /* 3 loads */
1984 return V_008F0C_BUF_DATA_FORMAT_8
;
1986 return V_008F0C_BUF_DATA_FORMAT_8_8
;
1988 return V_008F0C_BUF_DATA_FORMAT_8_8_8_8
;
1992 switch (desc
->nr_channels
) {
1994 case 3: /* 3 loads */
1995 return V_008F0C_BUF_DATA_FORMAT_16
;
1997 return V_008F0C_BUF_DATA_FORMAT_16_16
;
1999 return V_008F0C_BUF_DATA_FORMAT_16_16_16_16
;
2003 switch (desc
->nr_channels
) {
2005 return V_008F0C_BUF_DATA_FORMAT_32
;
2007 return V_008F0C_BUF_DATA_FORMAT_32_32
;
2009 return V_008F0C_BUF_DATA_FORMAT_32_32_32
;
2011 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32
;
2015 /* Legacy double formats. */
2016 switch (desc
->nr_channels
) {
2017 case 1: /* 1 load */
2018 return V_008F0C_BUF_DATA_FORMAT_32_32
;
2019 case 2: /* 1 load */
2020 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32
;
2021 case 3: /* 3 loads */
2022 return V_008F0C_BUF_DATA_FORMAT_32_32
;
2023 case 4: /* 2 loads */
2024 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32
;
2029 return V_008F0C_BUF_DATA_FORMAT_INVALID
;
2032 static uint32_t si_translate_buffer_numformat(struct pipe_screen
*screen
,
2033 const struct util_format_description
*desc
,
2036 if (desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
)
2037 return V_008F0C_BUF_NUM_FORMAT_FLOAT
;
2039 assert(first_non_void
>= 0);
2041 switch (desc
->channel
[first_non_void
].type
) {
2042 case UTIL_FORMAT_TYPE_SIGNED
:
2043 case UTIL_FORMAT_TYPE_FIXED
:
2044 if (desc
->channel
[first_non_void
].size
>= 32 ||
2045 desc
->channel
[first_non_void
].pure_integer
)
2046 return V_008F0C_BUF_NUM_FORMAT_SINT
;
2047 else if (desc
->channel
[first_non_void
].normalized
)
2048 return V_008F0C_BUF_NUM_FORMAT_SNORM
;
2050 return V_008F0C_BUF_NUM_FORMAT_SSCALED
;
2052 case UTIL_FORMAT_TYPE_UNSIGNED
:
2053 if (desc
->channel
[first_non_void
].size
>= 32 ||
2054 desc
->channel
[first_non_void
].pure_integer
)
2055 return V_008F0C_BUF_NUM_FORMAT_UINT
;
2056 else if (desc
->channel
[first_non_void
].normalized
)
2057 return V_008F0C_BUF_NUM_FORMAT_UNORM
;
2059 return V_008F0C_BUF_NUM_FORMAT_USCALED
;
2061 case UTIL_FORMAT_TYPE_FLOAT
:
2063 return V_008F0C_BUF_NUM_FORMAT_FLOAT
;
2067 static unsigned si_is_vertex_format_supported(struct pipe_screen
*screen
,
2068 enum pipe_format format
,
2071 const struct util_format_description
*desc
;
2073 unsigned data_format
;
2075 assert((usage
& ~(PIPE_BIND_SHADER_IMAGE
|
2076 PIPE_BIND_SAMPLER_VIEW
|
2077 PIPE_BIND_VERTEX_BUFFER
)) == 0);
2079 desc
= util_format_description(format
);
2083 /* There are no native 8_8_8 or 16_16_16 data formats, and we currently
2084 * select 8_8_8_8 and 16_16_16_16 instead. This works reasonably well
2085 * for read-only access (with caveats surrounding bounds checks), but
2086 * obviously fails for write access which we have to implement for
2087 * shader images. Luckily, OpenGL doesn't expect this to be supported
2088 * anyway, and so the only impact is on PBO uploads / downloads, which
2089 * shouldn't be expected to be fast for GL_RGB anyway.
2091 if (desc
->block
.bits
== 3 * 8 ||
2092 desc
->block
.bits
== 3 * 16) {
2093 if (usage
& (PIPE_BIND_SHADER_IMAGE
| PIPE_BIND_SAMPLER_VIEW
)) {
2094 usage
&= ~(PIPE_BIND_SHADER_IMAGE
| PIPE_BIND_SAMPLER_VIEW
);
2100 first_non_void
= util_format_get_first_non_void_channel(format
);
2101 data_format
= si_translate_buffer_dataformat(screen
, desc
, first_non_void
);
2102 if (data_format
== V_008F0C_BUF_DATA_FORMAT_INVALID
)
2108 static bool si_is_colorbuffer_format_supported(enum pipe_format format
)
2110 return si_translate_colorformat(format
) != V_028C70_COLOR_INVALID
&&
2111 si_translate_colorswap(format
, false) != ~0U;
2114 static bool si_is_zs_format_supported(enum pipe_format format
)
2116 return si_translate_dbformat(format
) != V_028040_Z_INVALID
;
2119 static boolean
si_is_format_supported(struct pipe_screen
*screen
,
2120 enum pipe_format format
,
2121 enum pipe_texture_target target
,
2122 unsigned sample_count
,
2125 unsigned retval
= 0;
2127 if (target
>= PIPE_MAX_TEXTURE_TYPES
) {
2128 R600_ERR("r600: unsupported texture type %d\n", target
);
2132 if (!util_format_is_supported(format
, usage
))
2135 if (sample_count
> 1) {
2136 if (!screen
->get_param(screen
, PIPE_CAP_TEXTURE_MULTISAMPLE
))
2139 if (usage
& PIPE_BIND_SHADER_IMAGE
)
2142 switch (sample_count
) {
2148 if (format
== PIPE_FORMAT_NONE
)
2157 if (usage
& (PIPE_BIND_SAMPLER_VIEW
|
2158 PIPE_BIND_SHADER_IMAGE
)) {
2159 if (target
== PIPE_BUFFER
) {
2160 retval
|= si_is_vertex_format_supported(
2161 screen
, format
, usage
& (PIPE_BIND_SAMPLER_VIEW
|
2162 PIPE_BIND_SHADER_IMAGE
));
2164 if (si_is_sampler_format_supported(screen
, format
))
2165 retval
|= usage
& (PIPE_BIND_SAMPLER_VIEW
|
2166 PIPE_BIND_SHADER_IMAGE
);
2170 if ((usage
& (PIPE_BIND_RENDER_TARGET
|
2171 PIPE_BIND_DISPLAY_TARGET
|
2174 PIPE_BIND_BLENDABLE
)) &&
2175 si_is_colorbuffer_format_supported(format
)) {
2177 (PIPE_BIND_RENDER_TARGET
|
2178 PIPE_BIND_DISPLAY_TARGET
|
2181 if (!util_format_is_pure_integer(format
) &&
2182 !util_format_is_depth_or_stencil(format
))
2183 retval
|= usage
& PIPE_BIND_BLENDABLE
;
2186 if ((usage
& PIPE_BIND_DEPTH_STENCIL
) &&
2187 si_is_zs_format_supported(format
)) {
2188 retval
|= PIPE_BIND_DEPTH_STENCIL
;
2191 if (usage
& PIPE_BIND_VERTEX_BUFFER
) {
2192 retval
|= si_is_vertex_format_supported(screen
, format
,
2193 PIPE_BIND_VERTEX_BUFFER
);
2196 if ((usage
& PIPE_BIND_LINEAR
) &&
2197 !util_format_is_compressed(format
) &&
2198 !(usage
& PIPE_BIND_DEPTH_STENCIL
))
2199 retval
|= PIPE_BIND_LINEAR
;
2201 return retval
== usage
;
2205 * framebuffer handling
2208 static void si_choose_spi_color_formats(struct r600_surface
*surf
,
2209 unsigned format
, unsigned swap
,
2210 unsigned ntype
, bool is_depth
)
2212 /* Alpha is needed for alpha-to-coverage.
2213 * Blending may be with or without alpha.
2215 unsigned normal
= 0; /* most optimal, may not support blending or export alpha */
2216 unsigned alpha
= 0; /* exports alpha, but may not support blending */
2217 unsigned blend
= 0; /* supports blending, but may not export alpha */
2218 unsigned blend_alpha
= 0; /* least optimal, supports blending and exports alpha */
2220 /* Choose the SPI color formats. These are required values for RB+.
2221 * Other chips have multiple choices, though they are not necessarily better.
2224 case V_028C70_COLOR_5_6_5
:
2225 case V_028C70_COLOR_1_5_5_5
:
2226 case V_028C70_COLOR_5_5_5_1
:
2227 case V_028C70_COLOR_4_4_4_4
:
2228 case V_028C70_COLOR_10_11_11
:
2229 case V_028C70_COLOR_11_11_10
:
2230 case V_028C70_COLOR_8
:
2231 case V_028C70_COLOR_8_8
:
2232 case V_028C70_COLOR_8_8_8_8
:
2233 case V_028C70_COLOR_10_10_10_2
:
2234 case V_028C70_COLOR_2_10_10_10
:
2235 if (ntype
== V_028C70_NUMBER_UINT
)
2236 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
2237 else if (ntype
== V_028C70_NUMBER_SINT
)
2238 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
2240 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
2243 case V_028C70_COLOR_16
:
2244 case V_028C70_COLOR_16_16
:
2245 case V_028C70_COLOR_16_16_16_16
:
2246 if (ntype
== V_028C70_NUMBER_UNORM
||
2247 ntype
== V_028C70_NUMBER_SNORM
) {
2248 /* UNORM16 and SNORM16 don't support blending */
2249 if (ntype
== V_028C70_NUMBER_UNORM
)
2250 normal
= alpha
= V_028714_SPI_SHADER_UNORM16_ABGR
;
2252 normal
= alpha
= V_028714_SPI_SHADER_SNORM16_ABGR
;
2254 /* Use 32 bits per channel for blending. */
2255 if (format
== V_028C70_COLOR_16
) {
2256 if (swap
== V_028C70_SWAP_STD
) { /* R */
2257 blend
= V_028714_SPI_SHADER_32_R
;
2258 blend_alpha
= V_028714_SPI_SHADER_32_AR
;
2259 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
2260 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
2263 } else if (format
== V_028C70_COLOR_16_16
) {
2264 if (swap
== V_028C70_SWAP_STD
) { /* RG */
2265 blend
= V_028714_SPI_SHADER_32_GR
;
2266 blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
2267 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
2268 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
2271 } else /* 16_16_16_16 */
2272 blend
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
2273 } else if (ntype
== V_028C70_NUMBER_UINT
)
2274 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
2275 else if (ntype
== V_028C70_NUMBER_SINT
)
2276 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
2277 else if (ntype
== V_028C70_NUMBER_FLOAT
)
2278 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
2283 case V_028C70_COLOR_32
:
2284 if (swap
== V_028C70_SWAP_STD
) { /* R */
2285 blend
= normal
= V_028714_SPI_SHADER_32_R
;
2286 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
2287 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
2288 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
2293 case V_028C70_COLOR_32_32
:
2294 if (swap
== V_028C70_SWAP_STD
) { /* RG */
2295 blend
= normal
= V_028714_SPI_SHADER_32_GR
;
2296 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
2297 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
2298 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
2303 case V_028C70_COLOR_32_32_32_32
:
2304 case V_028C70_COLOR_8_24
:
2305 case V_028C70_COLOR_24_8
:
2306 case V_028C70_COLOR_X24_8_32_FLOAT
:
2307 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_ABGR
;
2315 /* The DB->CB copy needs 32_ABGR. */
2317 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_ABGR
;
2319 surf
->spi_shader_col_format
= normal
;
2320 surf
->spi_shader_col_format_alpha
= alpha
;
2321 surf
->spi_shader_col_format_blend
= blend
;
2322 surf
->spi_shader_col_format_blend_alpha
= blend_alpha
;
2325 static void si_initialize_color_surface(struct si_context
*sctx
,
2326 struct r600_surface
*surf
)
2328 struct r600_texture
*rtex
= (struct r600_texture
*)surf
->base
.texture
;
2329 unsigned color_info
, color_attrib
, color_view
;
2330 unsigned format
, swap
, ntype
, endian
;
2331 const struct util_format_description
*desc
;
2333 unsigned blend_clamp
= 0, blend_bypass
= 0;
2335 color_view
= S_028C6C_SLICE_START(surf
->base
.u
.tex
.first_layer
) |
2336 S_028C6C_SLICE_MAX(surf
->base
.u
.tex
.last_layer
);
2338 desc
= util_format_description(surf
->base
.format
);
2339 for (firstchan
= 0; firstchan
< 4; firstchan
++) {
2340 if (desc
->channel
[firstchan
].type
!= UTIL_FORMAT_TYPE_VOID
) {
2344 if (firstchan
== 4 || desc
->channel
[firstchan
].type
== UTIL_FORMAT_TYPE_FLOAT
) {
2345 ntype
= V_028C70_NUMBER_FLOAT
;
2347 ntype
= V_028C70_NUMBER_UNORM
;
2348 if (desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
)
2349 ntype
= V_028C70_NUMBER_SRGB
;
2350 else if (desc
->channel
[firstchan
].type
== UTIL_FORMAT_TYPE_SIGNED
) {
2351 if (desc
->channel
[firstchan
].pure_integer
) {
2352 ntype
= V_028C70_NUMBER_SINT
;
2354 assert(desc
->channel
[firstchan
].normalized
);
2355 ntype
= V_028C70_NUMBER_SNORM
;
2357 } else if (desc
->channel
[firstchan
].type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
2358 if (desc
->channel
[firstchan
].pure_integer
) {
2359 ntype
= V_028C70_NUMBER_UINT
;
2361 assert(desc
->channel
[firstchan
].normalized
);
2362 ntype
= V_028C70_NUMBER_UNORM
;
2367 format
= si_translate_colorformat(surf
->base
.format
);
2368 if (format
== V_028C70_COLOR_INVALID
) {
2369 R600_ERR("Invalid CB format: %d, disabling CB.\n", surf
->base
.format
);
2371 assert(format
!= V_028C70_COLOR_INVALID
);
2372 swap
= si_translate_colorswap(surf
->base
.format
, false);
2373 endian
= si_colorformat_endian_swap(format
);
2375 /* blend clamp should be set for all NORM/SRGB types */
2376 if (ntype
== V_028C70_NUMBER_UNORM
||
2377 ntype
== V_028C70_NUMBER_SNORM
||
2378 ntype
== V_028C70_NUMBER_SRGB
)
2381 /* set blend bypass according to docs if SINT/UINT or
2382 8/24 COLOR variants */
2383 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
2384 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
2385 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
2390 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) {
2391 if (format
== V_028C70_COLOR_8
||
2392 format
== V_028C70_COLOR_8_8
||
2393 format
== V_028C70_COLOR_8_8_8_8
)
2394 surf
->color_is_int8
= true;
2395 else if (format
== V_028C70_COLOR_10_10_10_2
||
2396 format
== V_028C70_COLOR_2_10_10_10
)
2397 surf
->color_is_int10
= true;
2400 color_info
= S_028C70_FORMAT(format
) |
2401 S_028C70_COMP_SWAP(swap
) |
2402 S_028C70_BLEND_CLAMP(blend_clamp
) |
2403 S_028C70_BLEND_BYPASS(blend_bypass
) |
2404 S_028C70_SIMPLE_FLOAT(1) |
2405 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
2406 ntype
!= V_028C70_NUMBER_SNORM
&&
2407 ntype
!= V_028C70_NUMBER_SRGB
&&
2408 format
!= V_028C70_COLOR_8_24
&&
2409 format
!= V_028C70_COLOR_24_8
) |
2410 S_028C70_NUMBER_TYPE(ntype
) |
2411 S_028C70_ENDIAN(endian
);
2413 /* Intensity is implemented as Red, so treat it that way. */
2414 color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == PIPE_SWIZZLE_1
||
2415 util_format_is_intensity(surf
->base
.format
));
2417 if (rtex
->resource
.b
.b
.nr_samples
> 1) {
2418 unsigned log_samples
= util_logbase2(rtex
->resource
.b
.b
.nr_samples
);
2420 color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
2421 S_028C74_NUM_FRAGMENTS(log_samples
);
2423 if (rtex
->fmask
.size
) {
2424 color_info
|= S_028C70_COMPRESSION(1);
2425 unsigned fmask_bankh
= util_logbase2(rtex
->fmask
.bank_height
);
2427 if (sctx
->b
.chip_class
== SI
) {
2428 /* due to a hw bug, FMASK_BANK_HEIGHT must be set on SI too */
2429 color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
2434 surf
->cb_color_view
= color_view
;
2435 surf
->cb_color_info
= color_info
;
2436 surf
->cb_color_attrib
= color_attrib
;
2438 if (sctx
->b
.chip_class
>= VI
) {
2439 unsigned max_uncompressed_block_size
= 2;
2441 if (rtex
->resource
.b
.b
.nr_samples
> 1) {
2442 if (rtex
->surface
.bpe
== 1)
2443 max_uncompressed_block_size
= 0;
2444 else if (rtex
->surface
.bpe
== 2)
2445 max_uncompressed_block_size
= 1;
2448 surf
->cb_dcc_control
= S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
2449 S_028C78_INDEPENDENT_64B_BLOCKS(1);
2452 /* This must be set for fast clear to work without FMASK. */
2453 if (!rtex
->fmask
.size
&& sctx
->b
.chip_class
== SI
) {
2454 unsigned bankh
= util_logbase2(rtex
->surface
.u
.legacy
.bankh
);
2455 surf
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
2458 if (sctx
->b
.chip_class
>= GFX9
) {
2459 unsigned mip0_depth
= util_max_layer(&rtex
->resource
.b
.b
, 0);
2461 surf
->cb_color_view
|= S_028C6C_MIP_LEVEL(surf
->base
.u
.tex
.level
);
2462 surf
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
2463 S_028C74_RESOURCE_TYPE(rtex
->surface
.u
.gfx9
.resource_type
);
2464 surf
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(surf
->width0
- 1) |
2465 S_028C68_MIP0_HEIGHT(surf
->height0
- 1) |
2466 S_028C68_MAX_MIP(rtex
->resource
.b
.b
.last_level
);
2469 /* Determine pixel shader export format */
2470 si_choose_spi_color_formats(surf
, format
, swap
, ntype
, rtex
->is_depth
);
2472 surf
->color_initialized
= true;
2475 static void si_init_depth_surface(struct si_context
*sctx
,
2476 struct r600_surface
*surf
)
2478 struct r600_texture
*rtex
= (struct r600_texture
*)surf
->base
.texture
;
2479 unsigned level
= surf
->base
.u
.tex
.level
;
2480 unsigned format
, stencil_format
;
2481 uint32_t z_info
, s_info
;
2483 format
= si_translate_dbformat(rtex
->db_render_format
);
2484 stencil_format
= rtex
->surface
.has_stencil
?
2485 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
2487 assert(format
!= V_028040_Z_INVALID
);
2488 if (format
== V_028040_Z_INVALID
)
2489 R600_ERR("Invalid DB format: %d, disabling DB.\n", rtex
->resource
.b
.b
.format
);
2491 surf
->db_depth_view
= S_028008_SLICE_START(surf
->base
.u
.tex
.first_layer
) |
2492 S_028008_SLICE_MAX(surf
->base
.u
.tex
.last_layer
);
2493 surf
->db_htile_data_base
= 0;
2494 surf
->db_htile_surface
= 0;
2496 if (sctx
->b
.chip_class
>= GFX9
) {
2497 assert(rtex
->surface
.u
.gfx9
.surf_offset
== 0);
2498 surf
->db_depth_base
= rtex
->resource
.gpu_address
>> 8;
2499 surf
->db_stencil_base
= (rtex
->resource
.gpu_address
+
2500 rtex
->surface
.u
.gfx9
.stencil_offset
) >> 8;
2501 z_info
= S_028038_FORMAT(format
) |
2502 S_028038_NUM_SAMPLES(util_logbase2(rtex
->resource
.b
.b
.nr_samples
)) |
2503 S_028038_SW_MODE(rtex
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
2504 S_028038_MAXMIP(rtex
->resource
.b
.b
.last_level
);
2505 s_info
= S_02803C_FORMAT(stencil_format
) |
2506 S_02803C_SW_MODE(rtex
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
2507 surf
->db_z_info2
= S_028068_EPITCH(rtex
->surface
.u
.gfx9
.surf
.epitch
);
2508 surf
->db_stencil_info2
= S_02806C_EPITCH(rtex
->surface
.u
.gfx9
.stencil
.epitch
);
2509 surf
->db_depth_view
|= S_028008_MIPID(level
);
2510 surf
->db_depth_size
= S_02801C_X_MAX(rtex
->resource
.b
.b
.width0
- 1) |
2511 S_02801C_Y_MAX(rtex
->resource
.b
.b
.height0
- 1);
2513 if (r600_htile_enabled(rtex
, level
)) {
2514 z_info
|= S_028038_TILE_SURFACE_ENABLE(1) |
2515 S_028038_ALLOW_EXPCLEAR(1);
2517 if (rtex
->tc_compatible_htile
) {
2518 unsigned max_zplanes
= 4;
2520 if (rtex
->db_render_format
== PIPE_FORMAT_Z16_UNORM
&&
2521 rtex
->resource
.b
.b
.nr_samples
> 1)
2524 z_info
|= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes
+ 1) |
2525 S_028038_ITERATE_FLUSH(1);
2526 s_info
|= S_02803C_ITERATE_FLUSH(1);
2529 if (rtex
->surface
.has_stencil
) {
2530 /* Stencil buffer workaround ported from the SI-CI-VI code.
2531 * See that for explanation.
2533 s_info
|= S_02803C_ALLOW_EXPCLEAR(rtex
->resource
.b
.b
.nr_samples
<= 1);
2535 /* Use all HTILE for depth if there's no stencil. */
2536 s_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
2539 surf
->db_htile_data_base
= (rtex
->resource
.gpu_address
+
2540 rtex
->htile_offset
) >> 8;
2541 surf
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
2542 S_028ABC_PIPE_ALIGNED(rtex
->surface
.u
.gfx9
.htile
.pipe_aligned
) |
2543 S_028ABC_RB_ALIGNED(rtex
->surface
.u
.gfx9
.htile
.rb_aligned
);
2547 struct legacy_surf_level
*levelinfo
= &rtex
->surface
.u
.legacy
.level
[level
];
2549 assert(levelinfo
->nblk_x
% 8 == 0 && levelinfo
->nblk_y
% 8 == 0);
2551 surf
->db_depth_base
= (rtex
->resource
.gpu_address
+
2552 rtex
->surface
.u
.legacy
.level
[level
].offset
) >> 8;
2553 surf
->db_stencil_base
= (rtex
->resource
.gpu_address
+
2554 rtex
->surface
.u
.legacy
.stencil_level
[level
].offset
) >> 8;
2556 z_info
= S_028040_FORMAT(format
) |
2557 S_028040_NUM_SAMPLES(util_logbase2(rtex
->resource
.b
.b
.nr_samples
));
2558 s_info
= S_028044_FORMAT(stencil_format
);
2559 surf
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(!rtex
->tc_compatible_htile
);
2561 if (sctx
->b
.chip_class
>= CIK
) {
2562 struct radeon_info
*info
= &sctx
->screen
->b
.info
;
2563 unsigned index
= rtex
->surface
.u
.legacy
.tiling_index
[level
];
2564 unsigned stencil_index
= rtex
->surface
.u
.legacy
.stencil_tiling_index
[level
];
2565 unsigned macro_index
= rtex
->surface
.u
.legacy
.macro_tile_index
;
2566 unsigned tile_mode
= info
->si_tile_mode_array
[index
];
2567 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
2568 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
2570 surf
->db_depth_info
|=
2571 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
2572 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
2573 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
2574 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
2575 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
2576 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
2577 z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
2578 s_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
2580 unsigned tile_mode_index
= si_tile_mode_index(rtex
, level
, false);
2581 z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
2582 tile_mode_index
= si_tile_mode_index(rtex
, level
, true);
2583 s_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
2586 surf
->db_depth_size
= S_028058_PITCH_TILE_MAX((levelinfo
->nblk_x
/ 8) - 1) |
2587 S_028058_HEIGHT_TILE_MAX((levelinfo
->nblk_y
/ 8) - 1);
2588 surf
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((levelinfo
->nblk_x
*
2589 levelinfo
->nblk_y
) / 64 - 1);
2591 if (r600_htile_enabled(rtex
, level
)) {
2592 z_info
|= S_028040_TILE_SURFACE_ENABLE(1) |
2593 S_028040_ALLOW_EXPCLEAR(1);
2595 if (rtex
->surface
.has_stencil
) {
2596 /* Workaround: For a not yet understood reason, the
2597 * combination of MSAA, fast stencil clear and stencil
2598 * decompress messes with subsequent stencil buffer
2599 * uses. Problem was reproduced on Verde, Bonaire,
2600 * Tonga, and Carrizo.
2602 * Disabling EXPCLEAR works around the problem.
2604 * Check piglit's arb_texture_multisample-stencil-clear
2605 * test if you want to try changing this.
2607 if (rtex
->resource
.b
.b
.nr_samples
<= 1)
2608 s_info
|= S_028044_ALLOW_EXPCLEAR(1);
2609 } else if (!rtex
->tc_compatible_htile
) {
2610 /* Use all of the htile_buffer for depth if there's no stencil.
2611 * This must not be set when TC-compatible HTILE is enabled
2614 s_info
|= S_028044_TILE_STENCIL_DISABLE(1);
2617 surf
->db_htile_data_base
= (rtex
->resource
.gpu_address
+
2618 rtex
->htile_offset
) >> 8;
2619 surf
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
2621 if (rtex
->tc_compatible_htile
) {
2622 surf
->db_htile_surface
|= S_028ABC_TC_COMPATIBLE(1);
2624 if (rtex
->resource
.b
.b
.nr_samples
<= 1)
2625 z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(5);
2626 else if (rtex
->resource
.b
.b
.nr_samples
<= 4)
2627 z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(3);
2629 z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(2);
2634 surf
->db_z_info
= z_info
;
2635 surf
->db_stencil_info
= s_info
;
2637 surf
->depth_initialized
= true;
2640 void si_update_fb_dirtiness_after_rendering(struct si_context
*sctx
)
2642 if (sctx
->decompression_enabled
)
2645 if (sctx
->framebuffer
.state
.zsbuf
) {
2646 struct pipe_surface
*surf
= sctx
->framebuffer
.state
.zsbuf
;
2647 struct r600_texture
*rtex
= (struct r600_texture
*)surf
->texture
;
2649 rtex
->dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
2651 if (rtex
->surface
.has_stencil
)
2652 rtex
->stencil_dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
2654 if (sctx
->framebuffer
.compressed_cb_mask
) {
2655 struct pipe_surface
*surf
;
2656 struct r600_texture
*rtex
;
2657 unsigned mask
= sctx
->framebuffer
.compressed_cb_mask
;
2660 unsigned i
= u_bit_scan(&mask
);
2661 surf
= sctx
->framebuffer
.state
.cbufs
[i
];
2662 rtex
= (struct r600_texture
*)surf
->texture
;
2664 if (rtex
->fmask
.size
)
2665 rtex
->dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
2666 if (rtex
->dcc_gather_statistics
)
2667 rtex
->separate_dcc_dirty
= true;
2672 static void si_dec_framebuffer_counters(const struct pipe_framebuffer_state
*state
)
2674 for (int i
= 0; i
< state
->nr_cbufs
; ++i
) {
2675 struct r600_surface
*surf
= NULL
;
2676 struct r600_texture
*rtex
;
2678 if (!state
->cbufs
[i
])
2680 surf
= (struct r600_surface
*)state
->cbufs
[i
];
2681 rtex
= (struct r600_texture
*)surf
->base
.texture
;
2683 p_atomic_dec(&rtex
->framebuffers_bound
);
2687 static void si_set_framebuffer_state(struct pipe_context
*ctx
,
2688 const struct pipe_framebuffer_state
*state
)
2690 struct si_context
*sctx
= (struct si_context
*)ctx
;
2691 struct pipe_constant_buffer constbuf
= {0};
2692 struct r600_surface
*surf
= NULL
;
2693 struct r600_texture
*rtex
;
2694 bool old_any_dst_linear
= sctx
->framebuffer
.any_dst_linear
;
2695 unsigned old_nr_samples
= sctx
->framebuffer
.nr_samples
;
2696 unsigned old_colorbuf_enabled_4bit
= sctx
->framebuffer
.colorbuf_enabled_4bit
;
2697 bool old_has_zsbuf
= !!sctx
->framebuffer
.state
.zsbuf
;
2698 bool old_has_stencil
=
2700 ((struct r600_texture
*)sctx
->framebuffer
.state
.zsbuf
->texture
)->surface
.has_stencil
;
2701 bool unbound
= false;
2704 si_update_fb_dirtiness_after_rendering(sctx
);
2706 for (i
= 0; i
< sctx
->framebuffer
.state
.nr_cbufs
; i
++) {
2707 if (!sctx
->framebuffer
.state
.cbufs
[i
])
2710 rtex
= (struct r600_texture
*)sctx
->framebuffer
.state
.cbufs
[i
]->texture
;
2711 if (rtex
->dcc_gather_statistics
)
2712 vi_separate_dcc_stop_query(ctx
, rtex
);
2715 /* Disable DCC if the formats are incompatible. */
2716 for (i
= 0; i
< state
->nr_cbufs
; i
++) {
2717 if (!state
->cbufs
[i
])
2720 surf
= (struct r600_surface
*)state
->cbufs
[i
];
2721 rtex
= (struct r600_texture
*)surf
->base
.texture
;
2723 if (!surf
->dcc_incompatible
)
2726 /* Since the DCC decompression calls back into set_framebuffer-
2727 * _state, we need to unbind the framebuffer, so that
2728 * vi_separate_dcc_stop_query isn't called twice with the same
2732 util_copy_framebuffer_state(&sctx
->framebuffer
.state
, NULL
);
2736 if (vi_dcc_enabled(rtex
, surf
->base
.u
.tex
.level
))
2737 if (!si_texture_disable_dcc(&sctx
->b
, rtex
))
2738 sctx
->b
.decompress_dcc(ctx
, rtex
);
2740 surf
->dcc_incompatible
= false;
2743 /* Only flush TC when changing the framebuffer state, because
2744 * the only client not using TC that can change textures is
2747 * Wait for compute shaders because of possible transitions:
2748 * - FB write -> shader read
2749 * - shader write -> FB read
2751 * DB caches are flushed on demand (using si_decompress_textures).
2753 * When MSAA is enabled, CB and TC caches are flushed on demand
2754 * (after FMASK decompression). Shader write -> FB read transitions
2755 * cannot happen for MSAA textures, because MSAA shader images are
2758 * Only flush and wait for CB if there is actually a bound color buffer.
2760 if (sctx
->framebuffer
.nr_samples
<= 1 &&
2761 sctx
->framebuffer
.state
.nr_cbufs
)
2762 si_make_CB_shader_coherent(sctx
, sctx
->framebuffer
.nr_samples
,
2763 sctx
->framebuffer
.CB_has_shader_readable_metadata
);
2765 sctx
->b
.flags
|= SI_CONTEXT_CS_PARTIAL_FLUSH
;
2767 /* u_blitter doesn't invoke depth decompression when it does multiple
2768 * blits in a row, but the only case when it matters for DB is when
2769 * doing generate_mipmap. So here we flush DB manually between
2770 * individual generate_mipmap blits.
2771 * Note that lower mipmap levels aren't compressed.
2773 if (sctx
->generate_mipmap_for_depth
) {
2774 si_make_DB_shader_coherent(sctx
, 1, false,
2775 sctx
->framebuffer
.DB_has_shader_readable_metadata
);
2776 } else if (sctx
->b
.chip_class
== GFX9
) {
2777 /* It appears that DB metadata "leaks" in a sequence of:
2779 * - DCC decompress for shader image writes (with DB disabled)
2780 * - render with DEPTH_BEFORE_SHADER=1
2781 * Flushing DB metadata works around the problem.
2783 sctx
->b
.flags
|= SI_CONTEXT_FLUSH_AND_INV_DB_META
;
2786 /* Take the maximum of the old and new count. If the new count is lower,
2787 * dirtying is needed to disable the unbound colorbuffers.
2789 sctx
->framebuffer
.dirty_cbufs
|=
2790 (1 << MAX2(sctx
->framebuffer
.state
.nr_cbufs
, state
->nr_cbufs
)) - 1;
2791 sctx
->framebuffer
.dirty_zsbuf
|= sctx
->framebuffer
.state
.zsbuf
!= state
->zsbuf
;
2793 si_dec_framebuffer_counters(&sctx
->framebuffer
.state
);
2794 util_copy_framebuffer_state(&sctx
->framebuffer
.state
, state
);
2796 sctx
->framebuffer
.colorbuf_enabled_4bit
= 0;
2797 sctx
->framebuffer
.spi_shader_col_format
= 0;
2798 sctx
->framebuffer
.spi_shader_col_format_alpha
= 0;
2799 sctx
->framebuffer
.spi_shader_col_format_blend
= 0;
2800 sctx
->framebuffer
.spi_shader_col_format_blend_alpha
= 0;
2801 sctx
->framebuffer
.color_is_int8
= 0;
2802 sctx
->framebuffer
.color_is_int10
= 0;
2804 sctx
->framebuffer
.compressed_cb_mask
= 0;
2805 sctx
->framebuffer
.nr_samples
= util_framebuffer_get_num_samples(state
);
2806 sctx
->framebuffer
.log_samples
= util_logbase2(sctx
->framebuffer
.nr_samples
);
2807 sctx
->framebuffer
.any_dst_linear
= false;
2808 sctx
->framebuffer
.CB_has_shader_readable_metadata
= false;
2809 sctx
->framebuffer
.DB_has_shader_readable_metadata
= false;
2811 for (i
= 0; i
< state
->nr_cbufs
; i
++) {
2812 if (!state
->cbufs
[i
])
2815 surf
= (struct r600_surface
*)state
->cbufs
[i
];
2816 rtex
= (struct r600_texture
*)surf
->base
.texture
;
2818 if (!surf
->color_initialized
) {
2819 si_initialize_color_surface(sctx
, surf
);
2822 sctx
->framebuffer
.colorbuf_enabled_4bit
|= 0xf << (i
* 4);
2823 sctx
->framebuffer
.spi_shader_col_format
|=
2824 surf
->spi_shader_col_format
<< (i
* 4);
2825 sctx
->framebuffer
.spi_shader_col_format_alpha
|=
2826 surf
->spi_shader_col_format_alpha
<< (i
* 4);
2827 sctx
->framebuffer
.spi_shader_col_format_blend
|=
2828 surf
->spi_shader_col_format_blend
<< (i
* 4);
2829 sctx
->framebuffer
.spi_shader_col_format_blend_alpha
|=
2830 surf
->spi_shader_col_format_blend_alpha
<< (i
* 4);
2832 if (surf
->color_is_int8
)
2833 sctx
->framebuffer
.color_is_int8
|= 1 << i
;
2834 if (surf
->color_is_int10
)
2835 sctx
->framebuffer
.color_is_int10
|= 1 << i
;
2837 if (rtex
->fmask
.size
) {
2838 sctx
->framebuffer
.compressed_cb_mask
|= 1 << i
;
2841 if (rtex
->surface
.is_linear
)
2842 sctx
->framebuffer
.any_dst_linear
= true;
2844 if (vi_dcc_enabled(rtex
, surf
->base
.u
.tex
.level
))
2845 sctx
->framebuffer
.CB_has_shader_readable_metadata
= true;
2847 r600_context_add_resource_size(ctx
, surf
->base
.texture
);
2849 p_atomic_inc(&rtex
->framebuffers_bound
);
2851 if (rtex
->dcc_gather_statistics
) {
2852 /* Dirty tracking must be enabled for DCC usage analysis. */
2853 sctx
->framebuffer
.compressed_cb_mask
|= 1 << i
;
2854 vi_separate_dcc_start_query(ctx
, rtex
);
2858 struct r600_texture
*zstex
= NULL
;
2861 surf
= (struct r600_surface
*)state
->zsbuf
;
2862 zstex
= (struct r600_texture
*)surf
->base
.texture
;
2864 if (!surf
->depth_initialized
) {
2865 si_init_depth_surface(sctx
, surf
);
2868 if (vi_tc_compat_htile_enabled(zstex
, surf
->base
.u
.tex
.level
))
2869 sctx
->framebuffer
.DB_has_shader_readable_metadata
= true;
2871 r600_context_add_resource_size(ctx
, surf
->base
.texture
);
2874 si_update_poly_offset_state(sctx
);
2875 si_mark_atom_dirty(sctx
, &sctx
->cb_render_state
);
2876 si_mark_atom_dirty(sctx
, &sctx
->framebuffer
.atom
);
2878 if (sctx
->screen
->dpbb_allowed
)
2879 si_mark_atom_dirty(sctx
, &sctx
->dpbb_state
);
2881 if (sctx
->framebuffer
.any_dst_linear
!= old_any_dst_linear
)
2882 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
2884 if (sctx
->screen
->has_out_of_order_rast
&&
2885 (sctx
->framebuffer
.colorbuf_enabled_4bit
!= old_colorbuf_enabled_4bit
||
2886 !!sctx
->framebuffer
.state
.zsbuf
!= old_has_zsbuf
||
2887 (zstex
&& zstex
->surface
.has_stencil
!= old_has_stencil
)))
2888 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
2890 if (sctx
->framebuffer
.nr_samples
!= old_nr_samples
) {
2891 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
2892 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
2894 /* Set sample locations as fragment shader constants. */
2895 switch (sctx
->framebuffer
.nr_samples
) {
2897 constbuf
.user_buffer
= sctx
->b
.sample_locations_1x
;
2900 constbuf
.user_buffer
= sctx
->b
.sample_locations_2x
;
2903 constbuf
.user_buffer
= sctx
->b
.sample_locations_4x
;
2906 constbuf
.user_buffer
= sctx
->b
.sample_locations_8x
;
2909 constbuf
.user_buffer
= sctx
->b
.sample_locations_16x
;
2912 R600_ERR("Requested an invalid number of samples %i.\n",
2913 sctx
->framebuffer
.nr_samples
);
2916 constbuf
.buffer_size
= sctx
->framebuffer
.nr_samples
* 2 * 4;
2917 si_set_rw_buffer(sctx
, SI_PS_CONST_SAMPLE_POSITIONS
, &constbuf
);
2919 si_mark_atom_dirty(sctx
, &sctx
->msaa_sample_locs
.atom
);
2922 sctx
->do_update_shaders
= true;
2924 if (!sctx
->decompression_enabled
) {
2925 /* Prevent textures decompression when the framebuffer state
2926 * changes come from the decompression passes themselves.
2928 sctx
->need_check_render_feedback
= true;
2932 static void si_emit_framebuffer_state(struct si_context
*sctx
, struct r600_atom
*atom
)
2934 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
2935 struct pipe_framebuffer_state
*state
= &sctx
->framebuffer
.state
;
2936 unsigned i
, nr_cbufs
= state
->nr_cbufs
;
2937 struct r600_texture
*tex
= NULL
;
2938 struct r600_surface
*cb
= NULL
;
2939 unsigned cb_color_info
= 0;
2942 for (i
= 0; i
< nr_cbufs
; i
++) {
2943 uint64_t cb_color_base
, cb_color_fmask
, cb_dcc_base
;
2944 unsigned cb_color_attrib
;
2946 if (!(sctx
->framebuffer
.dirty_cbufs
& (1 << i
)))
2949 cb
= (struct r600_surface
*)state
->cbufs
[i
];
2951 radeon_set_context_reg(cs
, R_028C70_CB_COLOR0_INFO
+ i
* 0x3C,
2952 S_028C70_FORMAT(V_028C70_COLOR_INVALID
));
2956 tex
= (struct r600_texture
*)cb
->base
.texture
;
2957 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2958 &tex
->resource
, RADEON_USAGE_READWRITE
,
2959 tex
->resource
.b
.b
.nr_samples
> 1 ?
2960 RADEON_PRIO_COLOR_BUFFER_MSAA
:
2961 RADEON_PRIO_COLOR_BUFFER
);
2963 if (tex
->cmask_buffer
&& tex
->cmask_buffer
!= &tex
->resource
) {
2964 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2965 tex
->cmask_buffer
, RADEON_USAGE_READWRITE
,
2969 if (tex
->dcc_separate_buffer
)
2970 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2971 tex
->dcc_separate_buffer
,
2972 RADEON_USAGE_READWRITE
,
2975 /* Compute mutable surface parameters. */
2976 cb_color_base
= tex
->resource
.gpu_address
>> 8;
2979 cb_color_info
= cb
->cb_color_info
| tex
->cb_color_info
;
2980 cb_color_attrib
= cb
->cb_color_attrib
;
2982 if (tex
->fmask
.size
) {
2983 cb_color_fmask
= (tex
->resource
.gpu_address
+ tex
->fmask
.offset
) >> 8;
2984 cb_color_fmask
|= tex
->fmask
.tile_swizzle
;
2988 if (vi_dcc_enabled(tex
, cb
->base
.u
.tex
.level
)) {
2989 bool is_msaa_resolve_dst
= state
->cbufs
[0] &&
2990 state
->cbufs
[0]->texture
->nr_samples
> 1 &&
2991 state
->cbufs
[1] == &cb
->base
&&
2992 state
->cbufs
[1]->texture
->nr_samples
<= 1;
2994 if (!is_msaa_resolve_dst
)
2995 cb_color_info
|= S_028C70_DCC_ENABLE(1);
2997 cb_dcc_base
= ((!tex
->dcc_separate_buffer
? tex
->resource
.gpu_address
: 0) +
2998 tex
->dcc_offset
) >> 8;
2999 cb_dcc_base
|= tex
->surface
.tile_swizzle
;
3002 if (sctx
->b
.chip_class
>= GFX9
) {
3003 struct gfx9_surf_meta_flags meta
;
3005 if (tex
->dcc_offset
)
3006 meta
= tex
->surface
.u
.gfx9
.dcc
;
3008 meta
= tex
->surface
.u
.gfx9
.cmask
;
3010 /* Set mutable surface parameters. */
3011 cb_color_base
+= tex
->surface
.u
.gfx9
.surf_offset
>> 8;
3012 cb_color_base
|= tex
->surface
.tile_swizzle
;
3013 if (!tex
->fmask
.size
)
3014 cb_color_fmask
= cb_color_base
;
3015 cb_color_attrib
|= S_028C74_COLOR_SW_MODE(tex
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
3016 S_028C74_FMASK_SW_MODE(tex
->surface
.u
.gfx9
.fmask
.swizzle_mode
) |
3017 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
3018 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
3020 radeon_set_context_reg_seq(cs
, R_028C60_CB_COLOR0_BASE
+ i
* 0x3C, 15);
3021 radeon_emit(cs
, cb_color_base
); /* CB_COLOR0_BASE */
3022 radeon_emit(cs
, cb_color_base
>> 32); /* CB_COLOR0_BASE_EXT */
3023 radeon_emit(cs
, cb
->cb_color_attrib2
); /* CB_COLOR0_ATTRIB2 */
3024 radeon_emit(cs
, cb
->cb_color_view
); /* CB_COLOR0_VIEW */
3025 radeon_emit(cs
, cb_color_info
); /* CB_COLOR0_INFO */
3026 radeon_emit(cs
, cb_color_attrib
); /* CB_COLOR0_ATTRIB */
3027 radeon_emit(cs
, cb
->cb_dcc_control
); /* CB_COLOR0_DCC_CONTROL */
3028 radeon_emit(cs
, tex
->cmask
.base_address_reg
); /* CB_COLOR0_CMASK */
3029 radeon_emit(cs
, tex
->cmask
.base_address_reg
>> 32); /* CB_COLOR0_CMASK_BASE_EXT */
3030 radeon_emit(cs
, cb_color_fmask
); /* CB_COLOR0_FMASK */
3031 radeon_emit(cs
, cb_color_fmask
>> 32); /* CB_COLOR0_FMASK_BASE_EXT */
3032 radeon_emit(cs
, tex
->color_clear_value
[0]); /* CB_COLOR0_CLEAR_WORD0 */
3033 radeon_emit(cs
, tex
->color_clear_value
[1]); /* CB_COLOR0_CLEAR_WORD1 */
3034 radeon_emit(cs
, cb_dcc_base
); /* CB_COLOR0_DCC_BASE */
3035 radeon_emit(cs
, cb_dcc_base
>> 32); /* CB_COLOR0_DCC_BASE_EXT */
3037 radeon_set_context_reg(cs
, R_0287A0_CB_MRT0_EPITCH
+ i
* 4,
3038 S_0287A0_EPITCH(tex
->surface
.u
.gfx9
.surf
.epitch
));
3040 /* Compute mutable surface parameters (SI-CI-VI). */
3041 const struct legacy_surf_level
*level_info
=
3042 &tex
->surface
.u
.legacy
.level
[cb
->base
.u
.tex
.level
];
3043 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
3044 unsigned cb_color_pitch
, cb_color_slice
, cb_color_fmask_slice
;
3046 cb_color_base
+= level_info
->offset
>> 8;
3047 /* Only macrotiled modes can set tile swizzle. */
3048 if (level_info
->mode
== RADEON_SURF_MODE_2D
)
3049 cb_color_base
|= tex
->surface
.tile_swizzle
;
3051 if (!tex
->fmask
.size
)
3052 cb_color_fmask
= cb_color_base
;
3054 cb_dcc_base
+= level_info
->dcc_offset
>> 8;
3056 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
3057 slice_tile_max
= level_info
->nblk_x
*
3058 level_info
->nblk_y
/ 64 - 1;
3059 tile_mode_index
= si_tile_mode_index(tex
, cb
->base
.u
.tex
.level
, false);
3061 cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
3062 cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
3063 cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
3065 if (tex
->fmask
.size
) {
3066 if (sctx
->b
.chip_class
>= CIK
)
3067 cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(tex
->fmask
.pitch_in_pixels
/ 8 - 1);
3068 cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tex
->fmask
.tile_mode_index
);
3069 cb_color_fmask_slice
= S_028C88_TILE_MAX(tex
->fmask
.slice_tile_max
);
3071 /* This must be set for fast clear to work without FMASK. */
3072 if (sctx
->b
.chip_class
>= CIK
)
3073 cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
3074 cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
3075 cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
3078 radeon_set_context_reg_seq(cs
, R_028C60_CB_COLOR0_BASE
+ i
* 0x3C,
3079 sctx
->b
.chip_class
>= VI
? 14 : 13);
3080 radeon_emit(cs
, cb_color_base
); /* CB_COLOR0_BASE */
3081 radeon_emit(cs
, cb_color_pitch
); /* CB_COLOR0_PITCH */
3082 radeon_emit(cs
, cb_color_slice
); /* CB_COLOR0_SLICE */
3083 radeon_emit(cs
, cb
->cb_color_view
); /* CB_COLOR0_VIEW */
3084 radeon_emit(cs
, cb_color_info
); /* CB_COLOR0_INFO */
3085 radeon_emit(cs
, cb_color_attrib
); /* CB_COLOR0_ATTRIB */
3086 radeon_emit(cs
, cb
->cb_dcc_control
); /* CB_COLOR0_DCC_CONTROL */
3087 radeon_emit(cs
, tex
->cmask
.base_address_reg
); /* CB_COLOR0_CMASK */
3088 radeon_emit(cs
, tex
->cmask
.slice_tile_max
); /* CB_COLOR0_CMASK_SLICE */
3089 radeon_emit(cs
, cb_color_fmask
); /* CB_COLOR0_FMASK */
3090 radeon_emit(cs
, cb_color_fmask_slice
); /* CB_COLOR0_FMASK_SLICE */
3091 radeon_emit(cs
, tex
->color_clear_value
[0]); /* CB_COLOR0_CLEAR_WORD0 */
3092 radeon_emit(cs
, tex
->color_clear_value
[1]); /* CB_COLOR0_CLEAR_WORD1 */
3094 if (sctx
->b
.chip_class
>= VI
) /* R_028C94_CB_COLOR0_DCC_BASE */
3095 radeon_emit(cs
, cb_dcc_base
);
3099 if (sctx
->framebuffer
.dirty_cbufs
& (1 << i
))
3100 radeon_set_context_reg(cs
, R_028C70_CB_COLOR0_INFO
+ i
* 0x3C, 0);
3103 if (state
->zsbuf
&& sctx
->framebuffer
.dirty_zsbuf
) {
3104 struct r600_surface
*zb
= (struct r600_surface
*)state
->zsbuf
;
3105 struct r600_texture
*rtex
= (struct r600_texture
*)zb
->base
.texture
;
3107 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
3108 &rtex
->resource
, RADEON_USAGE_READWRITE
,
3109 zb
->base
.texture
->nr_samples
> 1 ?
3110 RADEON_PRIO_DEPTH_BUFFER_MSAA
:
3111 RADEON_PRIO_DEPTH_BUFFER
);
3113 if (sctx
->b
.chip_class
>= GFX9
) {
3114 radeon_set_context_reg_seq(cs
, R_028014_DB_HTILE_DATA_BASE
, 3);
3115 radeon_emit(cs
, zb
->db_htile_data_base
); /* DB_HTILE_DATA_BASE */
3116 radeon_emit(cs
, zb
->db_htile_data_base
>> 32); /* DB_HTILE_DATA_BASE_HI */
3117 radeon_emit(cs
, zb
->db_depth_size
); /* DB_DEPTH_SIZE */
3119 radeon_set_context_reg_seq(cs
, R_028038_DB_Z_INFO
, 10);
3120 radeon_emit(cs
, zb
->db_z_info
| /* DB_Z_INFO */
3121 S_028038_ZRANGE_PRECISION(rtex
->depth_clear_value
!= 0));
3122 radeon_emit(cs
, zb
->db_stencil_info
); /* DB_STENCIL_INFO */
3123 radeon_emit(cs
, zb
->db_depth_base
); /* DB_Z_READ_BASE */
3124 radeon_emit(cs
, zb
->db_depth_base
>> 32); /* DB_Z_READ_BASE_HI */
3125 radeon_emit(cs
, zb
->db_stencil_base
); /* DB_STENCIL_READ_BASE */
3126 radeon_emit(cs
, zb
->db_stencil_base
>> 32); /* DB_STENCIL_READ_BASE_HI */
3127 radeon_emit(cs
, zb
->db_depth_base
); /* DB_Z_WRITE_BASE */
3128 radeon_emit(cs
, zb
->db_depth_base
>> 32); /* DB_Z_WRITE_BASE_HI */
3129 radeon_emit(cs
, zb
->db_stencil_base
); /* DB_STENCIL_WRITE_BASE */
3130 radeon_emit(cs
, zb
->db_stencil_base
>> 32); /* DB_STENCIL_WRITE_BASE_HI */
3132 radeon_set_context_reg_seq(cs
, R_028068_DB_Z_INFO2
, 2);
3133 radeon_emit(cs
, zb
->db_z_info2
); /* DB_Z_INFO2 */
3134 radeon_emit(cs
, zb
->db_stencil_info2
); /* DB_STENCIL_INFO2 */
3136 radeon_set_context_reg(cs
, R_028014_DB_HTILE_DATA_BASE
, zb
->db_htile_data_base
);
3138 radeon_set_context_reg_seq(cs
, R_02803C_DB_DEPTH_INFO
, 9);
3139 radeon_emit(cs
, zb
->db_depth_info
); /* DB_DEPTH_INFO */
3140 radeon_emit(cs
, zb
->db_z_info
| /* DB_Z_INFO */
3141 S_028040_ZRANGE_PRECISION(rtex
->depth_clear_value
!= 0));
3142 radeon_emit(cs
, zb
->db_stencil_info
); /* DB_STENCIL_INFO */
3143 radeon_emit(cs
, zb
->db_depth_base
); /* DB_Z_READ_BASE */
3144 radeon_emit(cs
, zb
->db_stencil_base
); /* DB_STENCIL_READ_BASE */
3145 radeon_emit(cs
, zb
->db_depth_base
); /* DB_Z_WRITE_BASE */
3146 radeon_emit(cs
, zb
->db_stencil_base
); /* DB_STENCIL_WRITE_BASE */
3147 radeon_emit(cs
, zb
->db_depth_size
); /* DB_DEPTH_SIZE */
3148 radeon_emit(cs
, zb
->db_depth_slice
); /* DB_DEPTH_SLICE */
3151 radeon_set_context_reg_seq(cs
, R_028028_DB_STENCIL_CLEAR
, 2);
3152 radeon_emit(cs
, rtex
->stencil_clear_value
); /* R_028028_DB_STENCIL_CLEAR */
3153 radeon_emit(cs
, fui(rtex
->depth_clear_value
)); /* R_02802C_DB_DEPTH_CLEAR */
3155 radeon_set_context_reg(cs
, R_028008_DB_DEPTH_VIEW
, zb
->db_depth_view
);
3156 radeon_set_context_reg(cs
, R_028ABC_DB_HTILE_SURFACE
, zb
->db_htile_surface
);
3157 } else if (sctx
->framebuffer
.dirty_zsbuf
) {
3158 if (sctx
->b
.chip_class
>= GFX9
)
3159 radeon_set_context_reg_seq(cs
, R_028038_DB_Z_INFO
, 2);
3161 radeon_set_context_reg_seq(cs
, R_028040_DB_Z_INFO
, 2);
3163 radeon_emit(cs
, S_028040_FORMAT(V_028040_Z_INVALID
)); /* DB_Z_INFO */
3164 radeon_emit(cs
, S_028044_FORMAT(V_028044_STENCIL_INVALID
)); /* DB_STENCIL_INFO */
3167 /* Framebuffer dimensions. */
3168 /* PA_SC_WINDOW_SCISSOR_TL is set in si_init_config() */
3169 radeon_set_context_reg(cs
, R_028208_PA_SC_WINDOW_SCISSOR_BR
,
3170 S_028208_BR_X(state
->width
) | S_028208_BR_Y(state
->height
));
3172 if (sctx
->screen
->dfsm_allowed
) {
3173 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
3174 radeon_emit(cs
, EVENT_TYPE(V_028A90_BREAK_BATCH
) | EVENT_INDEX(0));
3177 sctx
->framebuffer
.dirty_cbufs
= 0;
3178 sctx
->framebuffer
.dirty_zsbuf
= false;
3181 static void si_emit_msaa_sample_locs(struct si_context
*sctx
,
3182 struct r600_atom
*atom
)
3184 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
3185 unsigned nr_samples
= sctx
->framebuffer
.nr_samples
;
3186 bool has_msaa_sample_loc_bug
= sctx
->screen
->has_msaa_sample_loc_bug
;
3188 /* Smoothing (only possible with nr_samples == 1) uses the same
3189 * sample locations as the MSAA it simulates.
3191 if (nr_samples
<= 1 && sctx
->smoothing_enabled
)
3192 nr_samples
= SI_NUM_SMOOTH_AA_SAMPLES
;
3194 /* On Polaris, the small primitive filter uses the sample locations
3195 * even when MSAA is off, so we need to make sure they're set to 0.
3197 if (has_msaa_sample_loc_bug
)
3198 nr_samples
= MAX2(nr_samples
, 1);
3200 if (nr_samples
!= sctx
->msaa_sample_locs
.nr_samples
) {
3201 sctx
->msaa_sample_locs
.nr_samples
= nr_samples
;
3202 si_common_emit_msaa_sample_locs(cs
, nr_samples
);
3205 if (sctx
->b
.family
>= CHIP_POLARIS10
) {
3206 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
3207 unsigned small_prim_filter_cntl
=
3208 S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
3210 S_028830_LINE_FILTER_DISABLE(sctx
->b
.family
<= CHIP_POLARIS12
);
3212 /* The alternative of setting sample locations to 0 would
3213 * require a DB flush to avoid Z errors, see
3214 * https://bugs.freedesktop.org/show_bug.cgi?id=96908
3216 if (has_msaa_sample_loc_bug
&&
3217 sctx
->framebuffer
.nr_samples
> 1 &&
3218 rs
&& !rs
->multisample_enable
)
3219 small_prim_filter_cntl
&= C_028830_SMALL_PRIM_FILTER_ENABLE
;
3221 radeon_set_context_reg(cs
, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL
,
3222 small_prim_filter_cntl
);
3226 static bool si_out_of_order_rasterization(struct si_context
*sctx
)
3228 struct si_state_blend
*blend
= sctx
->queued
.named
.blend
;
3229 struct si_state_dsa
*dsa
= sctx
->queued
.named
.dsa
;
3231 if (!sctx
->screen
->has_out_of_order_rast
)
3234 unsigned colormask
= sctx
->framebuffer
.colorbuf_enabled_4bit
;
3237 colormask
&= blend
->cb_target_enabled_4bit
;
3242 /* Conservative: No logic op. */
3243 if (colormask
&& blend
->logicop_enable
)
3246 struct si_dsa_order_invariance dsa_order_invariant
= {
3247 .zs
= true, .pass_set
= true, .pass_last
= false
3250 if (sctx
->framebuffer
.state
.zsbuf
) {
3251 struct r600_texture
*zstex
=
3252 (struct r600_texture
*)sctx
->framebuffer
.state
.zsbuf
->texture
;
3253 bool has_stencil
= zstex
->surface
.has_stencil
;
3254 dsa_order_invariant
= dsa
->order_invariance
[has_stencil
];
3255 if (!dsa_order_invariant
.zs
)
3258 /* The set of PS invocations is always order invariant,
3259 * except when early Z/S tests are requested. */
3260 if (sctx
->ps_shader
.cso
&&
3261 sctx
->ps_shader
.cso
->info
.writes_memory
&&
3262 sctx
->ps_shader
.cso
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
] &&
3263 !dsa_order_invariant
.pass_set
)
3266 if (sctx
->b
.num_perfect_occlusion_queries
!= 0 &&
3267 !dsa_order_invariant
.pass_set
)
3274 unsigned blendmask
= colormask
& blend
->blend_enable_4bit
;
3277 /* Only commutative blending. */
3278 if (blendmask
& ~blend
->commutative_4bit
)
3281 if (!dsa_order_invariant
.pass_set
)
3285 if (colormask
& ~blendmask
) {
3286 if (!dsa_order_invariant
.pass_last
)
3293 static void si_emit_msaa_config(struct si_context
*sctx
, struct r600_atom
*atom
)
3295 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
3296 unsigned num_tile_pipes
= sctx
->screen
->b
.info
.num_tile_pipes
;
3297 /* 33% faster rendering to linear color buffers */
3298 bool dst_is_linear
= sctx
->framebuffer
.any_dst_linear
;
3299 bool out_of_order_rast
= si_out_of_order_rasterization(sctx
);
3300 unsigned sc_mode_cntl_1
=
3301 S_028A4C_WALK_SIZE(dst_is_linear
) |
3302 S_028A4C_WALK_FENCE_ENABLE(!dst_is_linear
) |
3303 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes
== 2 ? 2 : 3) |
3304 S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(out_of_order_rast
) |
3305 S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7) |
3307 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
3308 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
3309 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
3310 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
3311 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
3312 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
3314 si_common_emit_msaa_config(cs
, sctx
->framebuffer
.nr_samples
,
3315 sctx
->ps_iter_samples
,
3316 sctx
->smoothing_enabled
? SI_NUM_SMOOTH_AA_SAMPLES
: 0,
3319 /* GFX9: Flush DFSM when the AA mode changes. */
3320 if (sctx
->screen
->dfsm_allowed
) {
3321 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
3322 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_DFSM
) | EVENT_INDEX(0));
3326 static void si_set_min_samples(struct pipe_context
*ctx
, unsigned min_samples
)
3328 struct si_context
*sctx
= (struct si_context
*)ctx
;
3330 if (sctx
->ps_iter_samples
== min_samples
)
3333 sctx
->ps_iter_samples
= min_samples
;
3334 sctx
->do_update_shaders
= true;
3336 if (sctx
->framebuffer
.nr_samples
> 1)
3337 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
3338 if (sctx
->screen
->dpbb_allowed
)
3339 si_mark_atom_dirty(sctx
, &sctx
->dpbb_state
);
3347 * Build the sampler view descriptor for a buffer texture.
3348 * @param state 256-bit descriptor; only the high 128 bits are filled in
3351 si_make_buffer_descriptor(struct si_screen
*screen
, struct r600_resource
*buf
,
3352 enum pipe_format format
,
3353 unsigned offset
, unsigned size
,
3356 const struct util_format_description
*desc
;
3359 unsigned num_records
;
3360 unsigned num_format
, data_format
;
3362 desc
= util_format_description(format
);
3363 first_non_void
= util_format_get_first_non_void_channel(format
);
3364 stride
= desc
->block
.bits
/ 8;
3365 num_format
= si_translate_buffer_numformat(&screen
->b
.b
, desc
, first_non_void
);
3366 data_format
= si_translate_buffer_dataformat(&screen
->b
.b
, desc
, first_non_void
);
3368 num_records
= size
/ stride
;
3369 num_records
= MIN2(num_records
, (buf
->b
.b
.width0
- offset
) / stride
);
3371 /* The NUM_RECORDS field has a different meaning depending on the chip,
3372 * instruction type, STRIDE, and SWIZZLE_ENABLE.
3375 * - If STRIDE == 0, it's in byte units.
3376 * - If STRIDE != 0, it's in units of STRIDE, used with inst.IDXEN.
3379 * - For SMEM and STRIDE == 0, it's in byte units.
3380 * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3381 * - For VMEM and STRIDE == 0 or SWIZZLE_ENABLE == 0, it's in byte units.
3382 * - For VMEM and STRIDE != 0 and SWIZZLE_ENABLE == 1, it's in units of STRIDE.
3383 * NOTE: There is incompatibility between VMEM and SMEM opcodes due to SWIZZLE_-
3384 * ENABLE. The workaround is to set STRIDE = 0 if SWIZZLE_ENABLE == 0 when
3385 * using SMEM. This can be done in the shader by clearing STRIDE with s_and.
3386 * That way the same descriptor can be used by both SMEM and VMEM.
3389 * - For SMEM and STRIDE == 0, it's in byte units.
3390 * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3391 * - For VMEM and inst.IDXEN == 0 or STRIDE == 0, it's in byte units.
3392 * - For VMEM and inst.IDXEN == 1 and STRIDE != 0, it's in units of STRIDE.
3394 if (screen
->b
.chip_class
>= GFX9
)
3395 /* When vindex == 0, LLVM sets IDXEN = 0, thus changing units
3396 * from STRIDE to bytes. This works around it by setting
3397 * NUM_RECORDS to at least the size of one element, so that
3398 * the first element is readable when IDXEN == 0.
3400 * TODO: Fix this in LLVM, but do we need a new intrinsic where
3401 * IDXEN is enforced?
3403 num_records
= num_records
? MAX2(num_records
, stride
) : 0;
3404 else if (screen
->b
.chip_class
== VI
)
3405 num_records
*= stride
;
3408 state
[5] = S_008F04_STRIDE(stride
);
3409 state
[6] = num_records
;
3410 state
[7] = S_008F0C_DST_SEL_X(si_map_swizzle(desc
->swizzle
[0])) |
3411 S_008F0C_DST_SEL_Y(si_map_swizzle(desc
->swizzle
[1])) |
3412 S_008F0C_DST_SEL_Z(si_map_swizzle(desc
->swizzle
[2])) |
3413 S_008F0C_DST_SEL_W(si_map_swizzle(desc
->swizzle
[3])) |
3414 S_008F0C_NUM_FORMAT(num_format
) |
3415 S_008F0C_DATA_FORMAT(data_format
);
3418 static unsigned gfx9_border_color_swizzle(const unsigned char swizzle
[4])
3420 unsigned bc_swizzle
= V_008F20_BC_SWIZZLE_XYZW
;
3422 if (swizzle
[3] == PIPE_SWIZZLE_X
) {
3423 /* For the pre-defined border color values (white, opaque
3424 * black, transparent black), the only thing that matters is
3425 * that the alpha channel winds up in the correct place
3426 * (because the RGB channels are all the same) so either of
3427 * these enumerations will work.
3429 if (swizzle
[2] == PIPE_SWIZZLE_Y
)
3430 bc_swizzle
= V_008F20_BC_SWIZZLE_WZYX
;
3432 bc_swizzle
= V_008F20_BC_SWIZZLE_WXYZ
;
3433 } else if (swizzle
[0] == PIPE_SWIZZLE_X
) {
3434 if (swizzle
[1] == PIPE_SWIZZLE_Y
)
3435 bc_swizzle
= V_008F20_BC_SWIZZLE_XYZW
;
3437 bc_swizzle
= V_008F20_BC_SWIZZLE_XWYZ
;
3438 } else if (swizzle
[1] == PIPE_SWIZZLE_X
) {
3439 bc_swizzle
= V_008F20_BC_SWIZZLE_YXWZ
;
3440 } else if (swizzle
[2] == PIPE_SWIZZLE_X
) {
3441 bc_swizzle
= V_008F20_BC_SWIZZLE_ZYXW
;
3448 * Build the sampler view descriptor for a texture.
3451 si_make_texture_descriptor(struct si_screen
*screen
,
3452 struct r600_texture
*tex
,
3454 enum pipe_texture_target target
,
3455 enum pipe_format pipe_format
,
3456 const unsigned char state_swizzle
[4],
3457 unsigned first_level
, unsigned last_level
,
3458 unsigned first_layer
, unsigned last_layer
,
3459 unsigned width
, unsigned height
, unsigned depth
,
3461 uint32_t *fmask_state
)
3463 struct pipe_resource
*res
= &tex
->resource
.b
.b
;
3464 const struct util_format_description
*desc
;
3465 unsigned char swizzle
[4];
3467 unsigned num_format
, data_format
, type
;
3470 desc
= util_format_description(pipe_format
);
3472 if (desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
) {
3473 const unsigned char swizzle_xxxx
[4] = {0, 0, 0, 0};
3474 const unsigned char swizzle_yyyy
[4] = {1, 1, 1, 1};
3475 const unsigned char swizzle_wwww
[4] = {3, 3, 3, 3};
3477 switch (pipe_format
) {
3478 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
3479 case PIPE_FORMAT_X32_S8X24_UINT
:
3480 case PIPE_FORMAT_X8Z24_UNORM
:
3481 util_format_compose_swizzles(swizzle_yyyy
, state_swizzle
, swizzle
);
3483 case PIPE_FORMAT_X24S8_UINT
:
3485 * X24S8 is implemented as an 8_8_8_8 data format, to
3486 * fix texture gathers. This affects at least
3487 * GL45-CTS.texture_cube_map_array.sampling on VI.
3489 util_format_compose_swizzles(swizzle_wwww
, state_swizzle
, swizzle
);
3492 util_format_compose_swizzles(swizzle_xxxx
, state_swizzle
, swizzle
);
3495 util_format_compose_swizzles(desc
->swizzle
, state_swizzle
, swizzle
);
3498 first_non_void
= util_format_get_first_non_void_channel(pipe_format
);
3500 switch (pipe_format
) {
3501 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
3502 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3505 if (first_non_void
< 0) {
3506 if (util_format_is_compressed(pipe_format
)) {
3507 switch (pipe_format
) {
3508 case PIPE_FORMAT_DXT1_SRGB
:
3509 case PIPE_FORMAT_DXT1_SRGBA
:
3510 case PIPE_FORMAT_DXT3_SRGBA
:
3511 case PIPE_FORMAT_DXT5_SRGBA
:
3512 case PIPE_FORMAT_BPTC_SRGBA
:
3513 case PIPE_FORMAT_ETC2_SRGB8
:
3514 case PIPE_FORMAT_ETC2_SRGB8A1
:
3515 case PIPE_FORMAT_ETC2_SRGBA8
:
3516 num_format
= V_008F14_IMG_NUM_FORMAT_SRGB
;
3518 case PIPE_FORMAT_RGTC1_SNORM
:
3519 case PIPE_FORMAT_LATC1_SNORM
:
3520 case PIPE_FORMAT_RGTC2_SNORM
:
3521 case PIPE_FORMAT_LATC2_SNORM
:
3522 case PIPE_FORMAT_ETC2_R11_SNORM
:
3523 case PIPE_FORMAT_ETC2_RG11_SNORM
:
3524 /* implies float, so use SNORM/UNORM to determine
3525 whether data is signed or not */
3526 case PIPE_FORMAT_BPTC_RGB_FLOAT
:
3527 num_format
= V_008F14_IMG_NUM_FORMAT_SNORM
;
3530 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3533 } else if (desc
->layout
== UTIL_FORMAT_LAYOUT_SUBSAMPLED
) {
3534 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3536 num_format
= V_008F14_IMG_NUM_FORMAT_FLOAT
;
3538 } else if (desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
3539 num_format
= V_008F14_IMG_NUM_FORMAT_SRGB
;
3541 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3543 switch (desc
->channel
[first_non_void
].type
) {
3544 case UTIL_FORMAT_TYPE_FLOAT
:
3545 num_format
= V_008F14_IMG_NUM_FORMAT_FLOAT
;
3547 case UTIL_FORMAT_TYPE_SIGNED
:
3548 if (desc
->channel
[first_non_void
].normalized
)
3549 num_format
= V_008F14_IMG_NUM_FORMAT_SNORM
;
3550 else if (desc
->channel
[first_non_void
].pure_integer
)
3551 num_format
= V_008F14_IMG_NUM_FORMAT_SINT
;
3553 num_format
= V_008F14_IMG_NUM_FORMAT_SSCALED
;
3555 case UTIL_FORMAT_TYPE_UNSIGNED
:
3556 if (desc
->channel
[first_non_void
].normalized
)
3557 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3558 else if (desc
->channel
[first_non_void
].pure_integer
)
3559 num_format
= V_008F14_IMG_NUM_FORMAT_UINT
;
3561 num_format
= V_008F14_IMG_NUM_FORMAT_USCALED
;
3566 data_format
= si_translate_texformat(&screen
->b
.b
, pipe_format
, desc
, first_non_void
);
3567 if (data_format
== ~0) {
3571 /* S8 with Z32 HTILE needs a special format. */
3572 if (screen
->b
.chip_class
>= GFX9
&&
3573 pipe_format
== PIPE_FORMAT_S8_UINT
&&
3574 tex
->tc_compatible_htile
)
3575 data_format
= V_008F14_IMG_DATA_FORMAT_S8_32
;
3578 (res
->target
== PIPE_TEXTURE_CUBE
||
3579 res
->target
== PIPE_TEXTURE_CUBE_ARRAY
||
3580 (screen
->b
.chip_class
<= VI
&&
3581 res
->target
== PIPE_TEXTURE_3D
))) {
3582 /* For the purpose of shader images, treat cube maps and 3D
3583 * textures as 2D arrays. For 3D textures, the address
3584 * calculations for mipmaps are different, so we rely on the
3585 * caller to effectively disable mipmaps.
3587 type
= V_008F1C_SQ_RSRC_IMG_2D_ARRAY
;
3589 assert(res
->target
!= PIPE_TEXTURE_3D
|| (first_level
== 0 && last_level
== 0));
3591 type
= si_tex_dim(screen
, tex
, target
, res
->nr_samples
);
3594 if (type
== V_008F1C_SQ_RSRC_IMG_1D_ARRAY
) {
3596 depth
= res
->array_size
;
3597 } else if (type
== V_008F1C_SQ_RSRC_IMG_2D_ARRAY
||
3598 type
== V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
) {
3599 if (sampler
|| res
->target
!= PIPE_TEXTURE_3D
)
3600 depth
= res
->array_size
;
3601 } else if (type
== V_008F1C_SQ_RSRC_IMG_CUBE
)
3602 depth
= res
->array_size
/ 6;
3605 state
[1] = (S_008F14_DATA_FORMAT_GFX6(data_format
) |
3606 S_008F14_NUM_FORMAT_GFX6(num_format
));
3607 state
[2] = (S_008F18_WIDTH(width
- 1) |
3608 S_008F18_HEIGHT(height
- 1) |
3609 S_008F18_PERF_MOD(4));
3610 state
[3] = (S_008F1C_DST_SEL_X(si_map_swizzle(swizzle
[0])) |
3611 S_008F1C_DST_SEL_Y(si_map_swizzle(swizzle
[1])) |
3612 S_008F1C_DST_SEL_Z(si_map_swizzle(swizzle
[2])) |
3613 S_008F1C_DST_SEL_W(si_map_swizzle(swizzle
[3])) |
3614 S_008F1C_BASE_LEVEL(res
->nr_samples
> 1 ?
3616 S_008F1C_LAST_LEVEL(res
->nr_samples
> 1 ?
3617 util_logbase2(res
->nr_samples
) :
3619 S_008F1C_TYPE(type
));
3621 state
[5] = S_008F24_BASE_ARRAY(first_layer
);
3625 if (screen
->b
.chip_class
>= GFX9
) {
3626 unsigned bc_swizzle
= gfx9_border_color_swizzle(desc
->swizzle
);
3628 /* Depth is the the last accessible layer on Gfx9.
3629 * The hw doesn't need to know the total number of layers.
3631 if (type
== V_008F1C_SQ_RSRC_IMG_3D
)
3632 state
[4] |= S_008F20_DEPTH(depth
- 1);
3634 state
[4] |= S_008F20_DEPTH(last_layer
);
3636 state
[4] |= S_008F20_BC_SWIZZLE(bc_swizzle
);
3637 state
[5] |= S_008F24_MAX_MIP(res
->nr_samples
> 1 ?
3638 util_logbase2(res
->nr_samples
) :
3639 tex
->resource
.b
.b
.last_level
);
3641 state
[3] |= S_008F1C_POW2_PAD(res
->last_level
> 0);
3642 state
[4] |= S_008F20_DEPTH(depth
- 1);
3643 state
[5] |= S_008F24_LAST_ARRAY(last_layer
);
3646 if (tex
->dcc_offset
) {
3647 unsigned swap
= si_translate_colorswap(pipe_format
, false);
3649 state
[6] = S_008F28_ALPHA_IS_ON_MSB(swap
<= 1);
3651 /* The last dword is unused by hw. The shader uses it to clear
3652 * bits in the first dword of sampler state.
3654 if (screen
->b
.chip_class
<= CIK
&& res
->nr_samples
<= 1) {
3655 if (first_level
== last_level
)
3656 state
[7] = C_008F30_MAX_ANISO_RATIO
;
3658 state
[7] = 0xffffffff;
3662 /* Initialize the sampler view for FMASK. */
3663 if (tex
->fmask
.size
) {
3664 uint32_t data_format
, num_format
;
3666 va
= tex
->resource
.gpu_address
+ tex
->fmask
.offset
;
3668 if (screen
->b
.chip_class
>= GFX9
) {
3669 data_format
= V_008F14_IMG_DATA_FORMAT_FMASK
;
3670 switch (res
->nr_samples
) {
3672 num_format
= V_008F14_IMG_FMASK_8_2_2
;
3675 num_format
= V_008F14_IMG_FMASK_8_4_4
;
3678 num_format
= V_008F14_IMG_FMASK_32_8_8
;
3681 unreachable("invalid nr_samples");
3684 switch (res
->nr_samples
) {
3686 data_format
= V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2
;
3689 data_format
= V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4
;
3692 data_format
= V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8
;
3695 unreachable("invalid nr_samples");
3697 num_format
= V_008F14_IMG_NUM_FORMAT_UINT
;
3700 fmask_state
[0] = (va
>> 8) | tex
->fmask
.tile_swizzle
;
3701 fmask_state
[1] = S_008F14_BASE_ADDRESS_HI(va
>> 40) |
3702 S_008F14_DATA_FORMAT_GFX6(data_format
) |
3703 S_008F14_NUM_FORMAT_GFX6(num_format
);
3704 fmask_state
[2] = S_008F18_WIDTH(width
- 1) |
3705 S_008F18_HEIGHT(height
- 1);
3706 fmask_state
[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X
) |
3707 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X
) |
3708 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X
) |
3709 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X
) |
3710 S_008F1C_TYPE(si_tex_dim(screen
, tex
, target
, 0));
3712 fmask_state
[5] = S_008F24_BASE_ARRAY(first_layer
);
3716 if (screen
->b
.chip_class
>= GFX9
) {
3717 fmask_state
[3] |= S_008F1C_SW_MODE(tex
->surface
.u
.gfx9
.fmask
.swizzle_mode
);
3718 fmask_state
[4] |= S_008F20_DEPTH(last_layer
) |
3719 S_008F20_PITCH_GFX9(tex
->surface
.u
.gfx9
.fmask
.epitch
);
3720 fmask_state
[5] |= S_008F24_META_PIPE_ALIGNED(tex
->surface
.u
.gfx9
.cmask
.pipe_aligned
) |
3721 S_008F24_META_RB_ALIGNED(tex
->surface
.u
.gfx9
.cmask
.rb_aligned
);
3723 fmask_state
[3] |= S_008F1C_TILING_INDEX(tex
->fmask
.tile_mode_index
);
3724 fmask_state
[4] |= S_008F20_DEPTH(depth
- 1) |
3725 S_008F20_PITCH_GFX6(tex
->fmask
.pitch_in_pixels
- 1);
3726 fmask_state
[5] |= S_008F24_LAST_ARRAY(last_layer
);
3732 * Create a sampler view.
3734 * @param ctx context
3735 * @param texture texture
3736 * @param state sampler view template
3737 * @param width0 width0 override (for compressed textures as int)
3738 * @param height0 height0 override (for compressed textures as int)
3739 * @param force_level set the base address to the level (for compressed textures)
3741 struct pipe_sampler_view
*
3742 si_create_sampler_view_custom(struct pipe_context
*ctx
,
3743 struct pipe_resource
*texture
,
3744 const struct pipe_sampler_view
*state
,
3745 unsigned width0
, unsigned height0
,
3746 unsigned force_level
)
3748 struct si_context
*sctx
= (struct si_context
*)ctx
;
3749 struct si_sampler_view
*view
= CALLOC_STRUCT(si_sampler_view
);
3750 struct r600_texture
*tmp
= (struct r600_texture
*)texture
;
3751 unsigned base_level
, first_level
, last_level
;
3752 unsigned char state_swizzle
[4];
3753 unsigned height
, depth
, width
;
3754 unsigned last_layer
= state
->u
.tex
.last_layer
;
3755 enum pipe_format pipe_format
;
3756 const struct legacy_surf_level
*surflevel
;
3761 /* initialize base object */
3762 view
->base
= *state
;
3763 view
->base
.texture
= NULL
;
3764 view
->base
.reference
.count
= 1;
3765 view
->base
.context
= ctx
;
3768 pipe_resource_reference(&view
->base
.texture
, texture
);
3770 if (state
->format
== PIPE_FORMAT_X24S8_UINT
||
3771 state
->format
== PIPE_FORMAT_S8X24_UINT
||
3772 state
->format
== PIPE_FORMAT_X32_S8X24_UINT
||
3773 state
->format
== PIPE_FORMAT_S8_UINT
)
3774 view
->is_stencil_sampler
= true;
3776 /* Buffer resource. */
3777 if (texture
->target
== PIPE_BUFFER
) {
3778 si_make_buffer_descriptor(sctx
->screen
,
3779 (struct r600_resource
*)texture
,
3781 state
->u
.buf
.offset
,
3787 state_swizzle
[0] = state
->swizzle_r
;
3788 state_swizzle
[1] = state
->swizzle_g
;
3789 state_swizzle
[2] = state
->swizzle_b
;
3790 state_swizzle
[3] = state
->swizzle_a
;
3793 first_level
= state
->u
.tex
.first_level
;
3794 last_level
= state
->u
.tex
.last_level
;
3797 depth
= texture
->depth0
;
3799 if (sctx
->b
.chip_class
<= VI
&& force_level
) {
3800 assert(force_level
== first_level
&&
3801 force_level
== last_level
);
3802 base_level
= force_level
;
3805 width
= u_minify(width
, force_level
);
3806 height
= u_minify(height
, force_level
);
3807 depth
= u_minify(depth
, force_level
);
3810 /* This is not needed if state trackers set last_layer correctly. */
3811 if (state
->target
== PIPE_TEXTURE_1D
||
3812 state
->target
== PIPE_TEXTURE_2D
||
3813 state
->target
== PIPE_TEXTURE_RECT
||
3814 state
->target
== PIPE_TEXTURE_CUBE
)
3815 last_layer
= state
->u
.tex
.first_layer
;
3817 /* Texturing with separate depth and stencil. */
3818 pipe_format
= state
->format
;
3820 /* Depth/stencil texturing sometimes needs separate texture. */
3821 if (tmp
->is_depth
&& !r600_can_sample_zs(tmp
, view
->is_stencil_sampler
)) {
3822 if (!tmp
->flushed_depth_texture
&&
3823 !si_init_flushed_depth_texture(ctx
, texture
, NULL
)) {
3824 pipe_resource_reference(&view
->base
.texture
, NULL
);
3829 assert(tmp
->flushed_depth_texture
);
3831 /* Override format for the case where the flushed texture
3832 * contains only Z or only S.
3834 if (tmp
->flushed_depth_texture
->resource
.b
.b
.format
!= tmp
->resource
.b
.b
.format
)
3835 pipe_format
= tmp
->flushed_depth_texture
->resource
.b
.b
.format
;
3837 tmp
= tmp
->flushed_depth_texture
;
3840 surflevel
= tmp
->surface
.u
.legacy
.level
;
3842 if (tmp
->db_compatible
) {
3843 if (!view
->is_stencil_sampler
)
3844 pipe_format
= tmp
->db_render_format
;
3846 switch (pipe_format
) {
3847 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
3848 pipe_format
= PIPE_FORMAT_Z32_FLOAT
;
3850 case PIPE_FORMAT_X8Z24_UNORM
:
3851 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
3852 /* Z24 is always stored like this for DB
3855 pipe_format
= PIPE_FORMAT_Z24X8_UNORM
;
3857 case PIPE_FORMAT_X24S8_UINT
:
3858 case PIPE_FORMAT_S8X24_UINT
:
3859 case PIPE_FORMAT_X32_S8X24_UINT
:
3860 pipe_format
= PIPE_FORMAT_S8_UINT
;
3861 surflevel
= tmp
->surface
.u
.legacy
.stencil_level
;
3867 view
->dcc_incompatible
=
3868 vi_dcc_formats_are_incompatible(texture
,
3869 state
->u
.tex
.first_level
,
3872 si_make_texture_descriptor(sctx
->screen
, tmp
, true,
3873 state
->target
, pipe_format
, state_swizzle
,
3874 first_level
, last_level
,
3875 state
->u
.tex
.first_layer
, last_layer
,
3876 width
, height
, depth
,
3877 view
->state
, view
->fmask_state
);
3879 unsigned num_format
= G_008F14_NUM_FORMAT_GFX6(view
->state
[1]);
3881 num_format
== V_008F14_IMG_NUM_FORMAT_USCALED
||
3882 num_format
== V_008F14_IMG_NUM_FORMAT_SSCALED
||
3883 num_format
== V_008F14_IMG_NUM_FORMAT_UINT
||
3884 num_format
== V_008F14_IMG_NUM_FORMAT_SINT
;
3885 view
->base_level_info
= &surflevel
[base_level
];
3886 view
->base_level
= base_level
;
3887 view
->block_width
= util_format_get_blockwidth(pipe_format
);
3891 static struct pipe_sampler_view
*
3892 si_create_sampler_view(struct pipe_context
*ctx
,
3893 struct pipe_resource
*texture
,
3894 const struct pipe_sampler_view
*state
)
3896 return si_create_sampler_view_custom(ctx
, texture
, state
,
3897 texture
? texture
->width0
: 0,
3898 texture
? texture
->height0
: 0, 0);
3901 static void si_sampler_view_destroy(struct pipe_context
*ctx
,
3902 struct pipe_sampler_view
*state
)
3904 struct si_sampler_view
*view
= (struct si_sampler_view
*)state
;
3906 pipe_resource_reference(&state
->texture
, NULL
);
3910 static bool wrap_mode_uses_border_color(unsigned wrap
, bool linear_filter
)
3912 return wrap
== PIPE_TEX_WRAP_CLAMP_TO_BORDER
||
3913 wrap
== PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
||
3915 (wrap
== PIPE_TEX_WRAP_CLAMP
||
3916 wrap
== PIPE_TEX_WRAP_MIRROR_CLAMP
));
3919 static uint32_t si_translate_border_color(struct si_context
*sctx
,
3920 const struct pipe_sampler_state
*state
,
3921 const union pipe_color_union
*color
,
3924 bool linear_filter
= state
->min_img_filter
!= PIPE_TEX_FILTER_NEAREST
||
3925 state
->mag_img_filter
!= PIPE_TEX_FILTER_NEAREST
;
3927 if (!wrap_mode_uses_border_color(state
->wrap_s
, linear_filter
) &&
3928 !wrap_mode_uses_border_color(state
->wrap_t
, linear_filter
) &&
3929 !wrap_mode_uses_border_color(state
->wrap_r
, linear_filter
))
3930 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
);
3932 #define simple_border_types(elt) \
3934 if (color->elt[0] == 0 && color->elt[1] == 0 && \
3935 color->elt[2] == 0 && color->elt[3] == 0) \
3936 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK); \
3937 if (color->elt[0] == 0 && color->elt[1] == 0 && \
3938 color->elt[2] == 0 && color->elt[3] == 1) \
3939 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK); \
3940 if (color->elt[0] == 1 && color->elt[1] == 1 && \
3941 color->elt[2] == 1 && color->elt[3] == 1) \
3942 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE); \
3946 simple_border_types(ui
);
3948 simple_border_types(f
);
3950 #undef simple_border_types
3954 /* Check if the border has been uploaded already. */
3955 for (i
= 0; i
< sctx
->border_color_count
; i
++)
3956 if (memcmp(&sctx
->border_color_table
[i
], color
,
3957 sizeof(*color
)) == 0)
3960 if (i
>= SI_MAX_BORDER_COLORS
) {
3961 /* Getting 4096 unique border colors is very unlikely. */
3962 fprintf(stderr
, "radeonsi: The border color table is full. "
3963 "Any new border colors will be just black. "
3964 "Please file a bug.\n");
3965 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
);
3968 if (i
== sctx
->border_color_count
) {
3969 /* Upload a new border color. */
3970 memcpy(&sctx
->border_color_table
[i
], color
,
3972 util_memcpy_cpu_to_le32(&sctx
->border_color_map
[i
],
3973 color
, sizeof(*color
));
3974 sctx
->border_color_count
++;
3977 return S_008F3C_BORDER_COLOR_PTR(i
) |
3978 S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_REGISTER
);
3981 static void *si_create_sampler_state(struct pipe_context
*ctx
,
3982 const struct pipe_sampler_state
*state
)
3984 struct si_context
*sctx
= (struct si_context
*)ctx
;
3985 struct r600_common_screen
*rscreen
= sctx
->b
.screen
;
3986 struct si_sampler_state
*rstate
= CALLOC_STRUCT(si_sampler_state
);
3987 unsigned max_aniso
= rscreen
->force_aniso
>= 0 ? rscreen
->force_aniso
3988 : state
->max_anisotropy
;
3989 unsigned max_aniso_ratio
= r600_tex_aniso_filter(max_aniso
);
3990 union pipe_color_union clamped_border_color
;
3997 rstate
->magic
= SI_SAMPLER_STATE_MAGIC
;
3999 rstate
->val
[0] = (S_008F30_CLAMP_X(si_tex_wrap(state
->wrap_s
)) |
4000 S_008F30_CLAMP_Y(si_tex_wrap(state
->wrap_t
)) |
4001 S_008F30_CLAMP_Z(si_tex_wrap(state
->wrap_r
)) |
4002 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
4003 S_008F30_DEPTH_COMPARE_FUNC(si_tex_compare(state
->compare_func
)) |
4004 S_008F30_FORCE_UNNORMALIZED(!state
->normalized_coords
) |
4005 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
4006 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
4007 S_008F30_DISABLE_CUBE_WRAP(!state
->seamless_cube_map
) |
4008 S_008F30_COMPAT_MODE(sctx
->b
.chip_class
>= VI
));
4009 rstate
->val
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(state
->min_lod
, 0, 15), 8)) |
4010 S_008F34_MAX_LOD(S_FIXED(CLAMP(state
->max_lod
, 0, 15), 8)) |
4011 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
4012 rstate
->val
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(state
->lod_bias
, -16, 16), 8)) |
4013 S_008F38_XY_MAG_FILTER(eg_tex_filter(state
->mag_img_filter
, max_aniso
)) |
4014 S_008F38_XY_MIN_FILTER(eg_tex_filter(state
->min_img_filter
, max_aniso
)) |
4015 S_008F38_MIP_FILTER(si_tex_mipfilter(state
->min_mip_filter
)) |
4016 S_008F38_MIP_POINT_PRECLAMP(0) |
4017 S_008F38_DISABLE_LSB_CEIL(sctx
->b
.chip_class
<= VI
) |
4018 S_008F38_FILTER_PREC_FIX(1) |
4019 S_008F38_ANISO_OVERRIDE(sctx
->b
.chip_class
>= VI
));
4020 rstate
->val
[3] = si_translate_border_color(sctx
, state
, &state
->border_color
, false);
4022 /* Create sampler resource for integer textures. */
4023 memcpy(rstate
->integer_val
, rstate
->val
, sizeof(rstate
->val
));
4024 rstate
->integer_val
[3] = si_translate_border_color(sctx
, state
, &state
->border_color
, true);
4026 /* Create sampler resource for upgraded depth textures. */
4027 memcpy(rstate
->upgraded_depth_val
, rstate
->val
, sizeof(rstate
->val
));
4029 for (unsigned i
= 0; i
< 4; ++i
) {
4030 /* Use channel 0 on purpose, so that we can use OPAQUE_WHITE
4031 * when the border color is 1.0. */
4032 clamped_border_color
.f
[i
] = CLAMP(state
->border_color
.f
[0], 0, 1);
4035 if (memcmp(&state
->border_color
, &clamped_border_color
, sizeof(clamped_border_color
)) == 0)
4036 rstate
->upgraded_depth_val
[3] |= S_008F3C_UPGRADED_DEPTH(1);
4038 rstate
->upgraded_depth_val
[3] =
4039 si_translate_border_color(sctx
, state
, &clamped_border_color
, false) |
4040 S_008F3C_UPGRADED_DEPTH(1);
4045 static void si_set_sample_mask(struct pipe_context
*ctx
, unsigned sample_mask
)
4047 struct si_context
*sctx
= (struct si_context
*)ctx
;
4049 if (sctx
->sample_mask
.sample_mask
== (uint16_t)sample_mask
)
4052 sctx
->sample_mask
.sample_mask
= sample_mask
;
4053 si_mark_atom_dirty(sctx
, &sctx
->sample_mask
.atom
);
4056 static void si_emit_sample_mask(struct si_context
*sctx
, struct r600_atom
*atom
)
4058 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
4059 unsigned mask
= sctx
->sample_mask
.sample_mask
;
4061 /* Needed for line and polygon smoothing as well as for the Polaris
4062 * small primitive filter. We expect the state tracker to take care of
4065 assert(mask
== 0xffff || sctx
->framebuffer
.nr_samples
> 1 ||
4066 (mask
& 1 && sctx
->blitter
->running
));
4068 radeon_set_context_reg_seq(cs
, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0
, 2);
4069 radeon_emit(cs
, mask
| (mask
<< 16));
4070 radeon_emit(cs
, mask
| (mask
<< 16));
4073 static void si_delete_sampler_state(struct pipe_context
*ctx
, void *state
)
4076 struct si_sampler_state
*s
= state
;
4078 assert(s
->magic
== SI_SAMPLER_STATE_MAGIC
);
4085 * Vertex elements & buffers
4088 static void *si_create_vertex_elements(struct pipe_context
*ctx
,
4090 const struct pipe_vertex_element
*elements
)
4092 struct si_screen
*sscreen
= (struct si_screen
*)ctx
->screen
;
4093 struct si_vertex_elements
*v
= CALLOC_STRUCT(si_vertex_elements
);
4094 bool used
[SI_NUM_VERTEX_BUFFERS
] = {};
4097 assert(count
<= SI_MAX_ATTRIBS
);
4102 v
->desc_list_byte_size
= align(count
* 16, SI_CPDMA_ALIGNMENT
);
4104 for (i
= 0; i
< count
; ++i
) {
4105 const struct util_format_description
*desc
;
4106 const struct util_format_channel_description
*channel
;
4107 unsigned data_format
, num_format
;
4109 unsigned vbo_index
= elements
[i
].vertex_buffer_index
;
4110 unsigned char swizzle
[4];
4112 if (vbo_index
>= SI_NUM_VERTEX_BUFFERS
) {
4117 if (elements
[i
].instance_divisor
) {
4118 v
->uses_instance_divisors
= true;
4119 v
->instance_divisors
[i
] = elements
[i
].instance_divisor
;
4121 if (v
->instance_divisors
[i
] == 1)
4122 v
->instance_divisor_is_one
|= 1u << i
;
4124 v
->instance_divisor_is_fetched
|= 1u << i
;
4127 if (!used
[vbo_index
]) {
4128 v
->first_vb_use_mask
|= 1 << i
;
4129 used
[vbo_index
] = true;
4132 desc
= util_format_description(elements
[i
].src_format
);
4133 first_non_void
= util_format_get_first_non_void_channel(elements
[i
].src_format
);
4134 data_format
= si_translate_buffer_dataformat(ctx
->screen
, desc
, first_non_void
);
4135 num_format
= si_translate_buffer_numformat(ctx
->screen
, desc
, first_non_void
);
4136 channel
= first_non_void
>= 0 ? &desc
->channel
[first_non_void
] : NULL
;
4137 memcpy(swizzle
, desc
->swizzle
, sizeof(swizzle
));
4139 v
->format_size
[i
] = desc
->block
.bits
/ 8;
4140 v
->src_offset
[i
] = elements
[i
].src_offset
;
4141 v
->vertex_buffer_index
[i
] = vbo_index
;
4143 /* The hardware always treats the 2-bit alpha channel as
4144 * unsigned, so a shader workaround is needed. The affected
4145 * chips are VI and older except Stoney (GFX8.1).
4147 if (data_format
== V_008F0C_BUF_DATA_FORMAT_2_10_10_10
&&
4148 sscreen
->b
.chip_class
<= VI
&&
4149 sscreen
->b
.family
!= CHIP_STONEY
) {
4150 if (num_format
== V_008F0C_BUF_NUM_FORMAT_SNORM
) {
4151 v
->fix_fetch
[i
] = SI_FIX_FETCH_A2_SNORM
;
4152 } else if (num_format
== V_008F0C_BUF_NUM_FORMAT_SSCALED
) {
4153 v
->fix_fetch
[i
] = SI_FIX_FETCH_A2_SSCALED
;
4154 } else if (num_format
== V_008F0C_BUF_NUM_FORMAT_SINT
) {
4155 /* This isn't actually used in OpenGL. */
4156 v
->fix_fetch
[i
] = SI_FIX_FETCH_A2_SINT
;
4158 } else if (channel
&& channel
->type
== UTIL_FORMAT_TYPE_FIXED
) {
4159 if (desc
->swizzle
[3] == PIPE_SWIZZLE_1
)
4160 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBX_32_FIXED
;
4162 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_FIXED
;
4163 } else if (channel
&& channel
->size
== 32 && !channel
->pure_integer
) {
4164 if (channel
->type
== UTIL_FORMAT_TYPE_SIGNED
) {
4165 if (channel
->normalized
) {
4166 if (desc
->swizzle
[3] == PIPE_SWIZZLE_1
)
4167 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBX_32_SNORM
;
4169 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_SNORM
;
4171 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_SSCALED
;
4173 } else if (channel
->type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
4174 if (channel
->normalized
) {
4175 if (desc
->swizzle
[3] == PIPE_SWIZZLE_1
)
4176 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBX_32_UNORM
;
4178 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_UNORM
;
4180 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_USCALED
;
4183 } else if (channel
&& channel
->size
== 64 &&
4184 channel
->type
== UTIL_FORMAT_TYPE_FLOAT
) {
4185 switch (desc
->nr_channels
) {
4188 v
->fix_fetch
[i
] = SI_FIX_FETCH_RG_64_FLOAT
;
4189 swizzle
[0] = PIPE_SWIZZLE_X
;
4190 swizzle
[1] = PIPE_SWIZZLE_Y
;
4191 swizzle
[2] = desc
->nr_channels
== 2 ? PIPE_SWIZZLE_Z
: PIPE_SWIZZLE_0
;
4192 swizzle
[3] = desc
->nr_channels
== 2 ? PIPE_SWIZZLE_W
: PIPE_SWIZZLE_0
;
4195 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_64_FLOAT
;
4196 swizzle
[0] = PIPE_SWIZZLE_X
; /* 3 loads */
4197 swizzle
[1] = PIPE_SWIZZLE_Y
;
4198 swizzle
[2] = PIPE_SWIZZLE_0
;
4199 swizzle
[3] = PIPE_SWIZZLE_0
;
4202 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_64_FLOAT
;
4203 swizzle
[0] = PIPE_SWIZZLE_X
; /* 2 loads */
4204 swizzle
[1] = PIPE_SWIZZLE_Y
;
4205 swizzle
[2] = PIPE_SWIZZLE_Z
;
4206 swizzle
[3] = PIPE_SWIZZLE_W
;
4211 } else if (channel
&& desc
->nr_channels
== 3) {
4212 assert(desc
->swizzle
[0] == PIPE_SWIZZLE_X
);
4214 if (channel
->size
== 8) {
4215 if (channel
->pure_integer
)
4216 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_8_INT
;
4218 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_8
;
4219 } else if (channel
->size
== 16) {
4220 if (channel
->pure_integer
)
4221 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_16_INT
;
4223 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_16
;
4227 v
->rsrc_word3
[i
] = S_008F0C_DST_SEL_X(si_map_swizzle(swizzle
[0])) |
4228 S_008F0C_DST_SEL_Y(si_map_swizzle(swizzle
[1])) |
4229 S_008F0C_DST_SEL_Z(si_map_swizzle(swizzle
[2])) |
4230 S_008F0C_DST_SEL_W(si_map_swizzle(swizzle
[3])) |
4231 S_008F0C_NUM_FORMAT(num_format
) |
4232 S_008F0C_DATA_FORMAT(data_format
);
4237 static void si_bind_vertex_elements(struct pipe_context
*ctx
, void *state
)
4239 struct si_context
*sctx
= (struct si_context
*)ctx
;
4240 struct si_vertex_elements
*old
= sctx
->vertex_elements
;
4241 struct si_vertex_elements
*v
= (struct si_vertex_elements
*)state
;
4243 sctx
->vertex_elements
= v
;
4244 sctx
->vertex_buffers_dirty
= true;
4248 old
->count
!= v
->count
||
4249 old
->uses_instance_divisors
!= v
->uses_instance_divisors
||
4250 v
->uses_instance_divisors
|| /* we don't check which divisors changed */
4251 memcmp(old
->fix_fetch
, v
->fix_fetch
, sizeof(v
->fix_fetch
[0]) * v
->count
)))
4252 sctx
->do_update_shaders
= true;
4254 if (v
&& v
->instance_divisor_is_fetched
) {
4255 struct pipe_constant_buffer cb
;
4258 cb
.user_buffer
= v
->instance_divisors
;
4259 cb
.buffer_offset
= 0;
4260 cb
.buffer_size
= sizeof(uint32_t) * v
->count
;
4261 si_set_rw_buffer(sctx
, SI_VS_CONST_INSTANCE_DIVISORS
, &cb
);
4265 static void si_delete_vertex_element(struct pipe_context
*ctx
, void *state
)
4267 struct si_context
*sctx
= (struct si_context
*)ctx
;
4269 if (sctx
->vertex_elements
== state
)
4270 sctx
->vertex_elements
= NULL
;
4274 static void si_set_vertex_buffers(struct pipe_context
*ctx
,
4275 unsigned start_slot
, unsigned count
,
4276 const struct pipe_vertex_buffer
*buffers
)
4278 struct si_context
*sctx
= (struct si_context
*)ctx
;
4279 struct pipe_vertex_buffer
*dst
= sctx
->vertex_buffer
+ start_slot
;
4282 assert(start_slot
+ count
<= ARRAY_SIZE(sctx
->vertex_buffer
));
4285 for (i
= 0; i
< count
; i
++) {
4286 const struct pipe_vertex_buffer
*src
= buffers
+ i
;
4287 struct pipe_vertex_buffer
*dsti
= dst
+ i
;
4288 struct pipe_resource
*buf
= src
->buffer
.resource
;
4290 pipe_resource_reference(&dsti
->buffer
.resource
, buf
);
4291 dsti
->buffer_offset
= src
->buffer_offset
;
4292 dsti
->stride
= src
->stride
;
4293 r600_context_add_resource_size(ctx
, buf
);
4295 r600_resource(buf
)->bind_history
|= PIPE_BIND_VERTEX_BUFFER
;
4298 for (i
= 0; i
< count
; i
++) {
4299 pipe_resource_reference(&dst
[i
].buffer
.resource
, NULL
);
4302 sctx
->vertex_buffers_dirty
= true;
4309 static void si_set_tess_state(struct pipe_context
*ctx
,
4310 const float default_outer_level
[4],
4311 const float default_inner_level
[2])
4313 struct si_context
*sctx
= (struct si_context
*)ctx
;
4314 struct pipe_constant_buffer cb
;
4317 memcpy(array
, default_outer_level
, sizeof(float) * 4);
4318 memcpy(array
+4, default_inner_level
, sizeof(float) * 2);
4321 cb
.user_buffer
= NULL
;
4322 cb
.buffer_size
= sizeof(array
);
4324 si_upload_const_buffer(sctx
, (struct r600_resource
**)&cb
.buffer
,
4325 (void*)array
, sizeof(array
),
4328 si_set_rw_buffer(sctx
, SI_HS_CONST_DEFAULT_TESS_LEVELS
, &cb
);
4329 pipe_resource_reference(&cb
.buffer
, NULL
);
4332 static void si_texture_barrier(struct pipe_context
*ctx
, unsigned flags
)
4334 struct si_context
*sctx
= (struct si_context
*)ctx
;
4336 si_update_fb_dirtiness_after_rendering(sctx
);
4338 /* Multisample surfaces are flushed in si_decompress_textures. */
4339 if (sctx
->framebuffer
.nr_samples
<= 1 &&
4340 sctx
->framebuffer
.state
.nr_cbufs
)
4341 si_make_CB_shader_coherent(sctx
, sctx
->framebuffer
.nr_samples
,
4342 sctx
->framebuffer
.CB_has_shader_readable_metadata
);
4345 /* This only ensures coherency for shader image/buffer stores. */
4346 static void si_memory_barrier(struct pipe_context
*ctx
, unsigned flags
)
4348 struct si_context
*sctx
= (struct si_context
*)ctx
;
4350 /* Subsequent commands must wait for all shader invocations to
4352 sctx
->b
.flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
|
4353 SI_CONTEXT_CS_PARTIAL_FLUSH
;
4355 if (flags
& PIPE_BARRIER_CONSTANT_BUFFER
)
4356 sctx
->b
.flags
|= SI_CONTEXT_INV_SMEM_L1
|
4357 SI_CONTEXT_INV_VMEM_L1
;
4359 if (flags
& (PIPE_BARRIER_VERTEX_BUFFER
|
4360 PIPE_BARRIER_SHADER_BUFFER
|
4361 PIPE_BARRIER_TEXTURE
|
4362 PIPE_BARRIER_IMAGE
|
4363 PIPE_BARRIER_STREAMOUT_BUFFER
|
4364 PIPE_BARRIER_GLOBAL_BUFFER
)) {
4365 /* As far as I can tell, L1 contents are written back to L2
4366 * automatically at end of shader, but the contents of other
4367 * L1 caches might still be stale. */
4368 sctx
->b
.flags
|= SI_CONTEXT_INV_VMEM_L1
;
4371 if (flags
& PIPE_BARRIER_INDEX_BUFFER
) {
4372 /* Indices are read through TC L2 since VI.
4375 if (sctx
->screen
->b
.chip_class
<= CIK
)
4376 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
4379 /* MSAA color, any depth and any stencil are flushed in
4380 * si_decompress_textures when needed.
4382 if (flags
& PIPE_BARRIER_FRAMEBUFFER
&&
4383 sctx
->framebuffer
.nr_samples
<= 1 &&
4384 sctx
->framebuffer
.state
.nr_cbufs
) {
4385 sctx
->b
.flags
|= SI_CONTEXT_FLUSH_AND_INV_CB
;
4387 if (sctx
->b
.chip_class
<= VI
)
4388 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
4391 /* Indirect buffers use TC L2 on GFX9, but not older hw. */
4392 if (sctx
->screen
->b
.chip_class
<= VI
&&
4393 flags
& PIPE_BARRIER_INDIRECT_BUFFER
)
4394 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
4397 static void *si_create_blend_custom(struct si_context
*sctx
, unsigned mode
)
4399 struct pipe_blend_state blend
;
4401 memset(&blend
, 0, sizeof(blend
));
4402 blend
.independent_blend_enable
= true;
4403 blend
.rt
[0].colormask
= 0xf;
4404 return si_create_blend_state_mode(&sctx
->b
.b
, &blend
, mode
);
4407 static void si_need_gfx_cs_space(struct pipe_context
*ctx
, unsigned num_dw
,
4408 bool include_draw_vbo
)
4410 si_need_cs_space((struct si_context
*)ctx
);
4413 static void si_init_config(struct si_context
*sctx
);
4415 void si_init_state_functions(struct si_context
*sctx
)
4417 si_init_external_atom(sctx
, &sctx
->b
.render_cond_atom
, &sctx
->atoms
.s
.render_cond
);
4418 si_init_external_atom(sctx
, &sctx
->b
.streamout
.begin_atom
, &sctx
->atoms
.s
.streamout_begin
);
4419 si_init_external_atom(sctx
, &sctx
->b
.streamout
.enable_atom
, &sctx
->atoms
.s
.streamout_enable
);
4420 si_init_external_atom(sctx
, &sctx
->scissors
.atom
, &sctx
->atoms
.s
.scissors
);
4421 si_init_external_atom(sctx
, &sctx
->viewports
.atom
, &sctx
->atoms
.s
.viewports
);
4423 si_init_atom(sctx
, &sctx
->framebuffer
.atom
, &sctx
->atoms
.s
.framebuffer
, si_emit_framebuffer_state
);
4424 si_init_atom(sctx
, &sctx
->msaa_sample_locs
.atom
, &sctx
->atoms
.s
.msaa_sample_locs
, si_emit_msaa_sample_locs
);
4425 si_init_atom(sctx
, &sctx
->db_render_state
, &sctx
->atoms
.s
.db_render_state
, si_emit_db_render_state
);
4426 si_init_atom(sctx
, &sctx
->dpbb_state
, &sctx
->atoms
.s
.dpbb_state
, si_emit_dpbb_state
);
4427 si_init_atom(sctx
, &sctx
->msaa_config
, &sctx
->atoms
.s
.msaa_config
, si_emit_msaa_config
);
4428 si_init_atom(sctx
, &sctx
->sample_mask
.atom
, &sctx
->atoms
.s
.sample_mask
, si_emit_sample_mask
);
4429 si_init_atom(sctx
, &sctx
->cb_render_state
, &sctx
->atoms
.s
.cb_render_state
, si_emit_cb_render_state
);
4430 si_init_atom(sctx
, &sctx
->blend_color
.atom
, &sctx
->atoms
.s
.blend_color
, si_emit_blend_color
);
4431 si_init_atom(sctx
, &sctx
->clip_regs
, &sctx
->atoms
.s
.clip_regs
, si_emit_clip_regs
);
4432 si_init_atom(sctx
, &sctx
->clip_state
.atom
, &sctx
->atoms
.s
.clip_state
, si_emit_clip_state
);
4433 si_init_atom(sctx
, &sctx
->stencil_ref
.atom
, &sctx
->atoms
.s
.stencil_ref
, si_emit_stencil_ref
);
4435 sctx
->b
.b
.create_blend_state
= si_create_blend_state
;
4436 sctx
->b
.b
.bind_blend_state
= si_bind_blend_state
;
4437 sctx
->b
.b
.delete_blend_state
= si_delete_blend_state
;
4438 sctx
->b
.b
.set_blend_color
= si_set_blend_color
;
4440 sctx
->b
.b
.create_rasterizer_state
= si_create_rs_state
;
4441 sctx
->b
.b
.bind_rasterizer_state
= si_bind_rs_state
;
4442 sctx
->b
.b
.delete_rasterizer_state
= si_delete_rs_state
;
4444 sctx
->b
.b
.create_depth_stencil_alpha_state
= si_create_dsa_state
;
4445 sctx
->b
.b
.bind_depth_stencil_alpha_state
= si_bind_dsa_state
;
4446 sctx
->b
.b
.delete_depth_stencil_alpha_state
= si_delete_dsa_state
;
4448 sctx
->custom_dsa_flush
= si_create_db_flush_dsa(sctx
);
4449 sctx
->custom_blend_resolve
= si_create_blend_custom(sctx
, V_028808_CB_RESOLVE
);
4450 sctx
->custom_blend_fmask_decompress
= si_create_blend_custom(sctx
, V_028808_CB_FMASK_DECOMPRESS
);
4451 sctx
->custom_blend_eliminate_fastclear
= si_create_blend_custom(sctx
, V_028808_CB_ELIMINATE_FAST_CLEAR
);
4452 sctx
->custom_blend_dcc_decompress
= si_create_blend_custom(sctx
, V_028808_CB_DCC_DECOMPRESS
);
4454 sctx
->b
.b
.set_clip_state
= si_set_clip_state
;
4455 sctx
->b
.b
.set_stencil_ref
= si_set_stencil_ref
;
4457 sctx
->b
.b
.set_framebuffer_state
= si_set_framebuffer_state
;
4458 sctx
->b
.b
.get_sample_position
= si_get_sample_position
;
4460 sctx
->b
.b
.create_sampler_state
= si_create_sampler_state
;
4461 sctx
->b
.b
.delete_sampler_state
= si_delete_sampler_state
;
4463 sctx
->b
.b
.create_sampler_view
= si_create_sampler_view
;
4464 sctx
->b
.b
.sampler_view_destroy
= si_sampler_view_destroy
;
4466 sctx
->b
.b
.set_sample_mask
= si_set_sample_mask
;
4468 sctx
->b
.b
.create_vertex_elements_state
= si_create_vertex_elements
;
4469 sctx
->b
.b
.bind_vertex_elements_state
= si_bind_vertex_elements
;
4470 sctx
->b
.b
.delete_vertex_elements_state
= si_delete_vertex_element
;
4471 sctx
->b
.b
.set_vertex_buffers
= si_set_vertex_buffers
;
4473 sctx
->b
.b
.texture_barrier
= si_texture_barrier
;
4474 sctx
->b
.b
.memory_barrier
= si_memory_barrier
;
4475 sctx
->b
.b
.set_min_samples
= si_set_min_samples
;
4476 sctx
->b
.b
.set_tess_state
= si_set_tess_state
;
4478 sctx
->b
.b
.set_active_query_state
= si_set_active_query_state
;
4479 sctx
->b
.set_occlusion_query_state
= si_set_occlusion_query_state
;
4480 sctx
->b
.save_qbo_state
= si_save_qbo_state
;
4481 sctx
->b
.need_gfx_cs_space
= si_need_gfx_cs_space
;
4483 sctx
->b
.b
.draw_vbo
= si_draw_vbo
;
4485 si_init_config(sctx
);
4488 static uint32_t si_get_bo_metadata_word1(struct r600_common_screen
*rscreen
)
4490 return (ATI_VENDOR_ID
<< 16) | rscreen
->info
.pci_id
;
4493 static void si_query_opaque_metadata(struct r600_common_screen
*rscreen
,
4494 struct r600_texture
*rtex
,
4495 struct radeon_bo_metadata
*md
)
4497 struct si_screen
*sscreen
= (struct si_screen
*)rscreen
;
4498 struct pipe_resource
*res
= &rtex
->resource
.b
.b
;
4499 static const unsigned char swizzle
[] = {
4505 uint32_t desc
[8], i
;
4506 bool is_array
= util_resource_is_array_texture(res
);
4508 /* DRM 2.x.x doesn't support this. */
4509 if (rscreen
->info
.drm_major
!= 3)
4512 assert(rtex
->dcc_separate_buffer
== NULL
);
4513 assert(rtex
->fmask
.size
== 0);
4515 /* Metadata image format format version 1:
4516 * [0] = 1 (metadata format identifier)
4517 * [1] = (VENDOR_ID << 16) | PCI_ID
4518 * [2:9] = image descriptor for the whole resource
4519 * [2] is always 0, because the base address is cleared
4520 * [9] is the DCC offset bits [39:8] from the beginning of
4522 * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
4525 md
->metadata
[0] = 1; /* metadata image format version 1 */
4527 /* TILE_MODE_INDEX is ambiguous without a PCI ID. */
4528 md
->metadata
[1] = si_get_bo_metadata_word1(rscreen
);
4530 si_make_texture_descriptor(sscreen
, rtex
, true,
4531 res
->target
, res
->format
,
4532 swizzle
, 0, res
->last_level
, 0,
4533 is_array
? res
->array_size
- 1 : 0,
4534 res
->width0
, res
->height0
, res
->depth0
,
4537 si_set_mutable_tex_desc_fields(sscreen
, rtex
, &rtex
->surface
.u
.legacy
.level
[0],
4538 0, 0, rtex
->surface
.blk_w
, false, desc
);
4540 /* Clear the base address and set the relative DCC offset. */
4542 desc
[1] &= C_008F14_BASE_ADDRESS_HI
;
4543 desc
[7] = rtex
->dcc_offset
>> 8;
4545 /* Dwords [2:9] contain the image descriptor. */
4546 memcpy(&md
->metadata
[2], desc
, sizeof(desc
));
4547 md
->size_metadata
= 10 * 4;
4549 /* Dwords [10:..] contain the mipmap level offsets. */
4550 if (rscreen
->chip_class
<= VI
) {
4551 for (i
= 0; i
<= res
->last_level
; i
++)
4552 md
->metadata
[10+i
] = rtex
->surface
.u
.legacy
.level
[i
].offset
>> 8;
4554 md
->size_metadata
+= (1 + res
->last_level
) * 4;
4558 static void si_apply_opaque_metadata(struct r600_common_screen
*rscreen
,
4559 struct r600_texture
*rtex
,
4560 struct radeon_bo_metadata
*md
)
4562 uint32_t *desc
= &md
->metadata
[2];
4564 if (rscreen
->chip_class
< VI
)
4567 /* Return if DCC is enabled. The texture should be set up with it
4570 if (md
->size_metadata
>= 11 * 4 &&
4571 md
->metadata
[0] != 0 &&
4572 md
->metadata
[1] == si_get_bo_metadata_word1(rscreen
) &&
4573 G_008F28_COMPRESSION_EN(desc
[6])) {
4574 assert(rtex
->dcc_offset
== ((uint64_t)desc
[7] << 8));
4578 /* Disable DCC. These are always set by texture_from_handle and must
4581 rtex
->dcc_offset
= 0;
4584 void si_init_screen_state_functions(struct si_screen
*sscreen
)
4586 sscreen
->b
.b
.is_format_supported
= si_is_format_supported
;
4587 sscreen
->b
.query_opaque_metadata
= si_query_opaque_metadata
;
4588 sscreen
->b
.apply_opaque_metadata
= si_apply_opaque_metadata
;
4591 static void si_set_grbm_gfx_index(struct si_context
*sctx
,
4592 struct si_pm4_state
*pm4
, unsigned value
)
4594 unsigned reg
= sctx
->b
.chip_class
>= CIK
? R_030800_GRBM_GFX_INDEX
:
4596 si_pm4_set_reg(pm4
, reg
, value
);
4599 static void si_set_grbm_gfx_index_se(struct si_context
*sctx
,
4600 struct si_pm4_state
*pm4
, unsigned se
)
4602 assert(se
== ~0 || se
< sctx
->screen
->b
.info
.max_se
);
4603 si_set_grbm_gfx_index(sctx
, pm4
,
4604 (se
== ~0 ? S_030800_SE_BROADCAST_WRITES(1) :
4605 S_030800_SE_INDEX(se
)) |
4606 S_030800_SH_BROADCAST_WRITES(1) |
4607 S_030800_INSTANCE_BROADCAST_WRITES(1));
4611 si_write_harvested_raster_configs(struct si_context
*sctx
,
4612 struct si_pm4_state
*pm4
,
4613 unsigned raster_config
,
4614 unsigned raster_config_1
)
4616 unsigned sh_per_se
= MAX2(sctx
->screen
->b
.info
.max_sh_per_se
, 1);
4617 unsigned num_se
= MAX2(sctx
->screen
->b
.info
.max_se
, 1);
4618 unsigned rb_mask
= sctx
->screen
->b
.info
.enabled_rb_mask
;
4619 unsigned num_rb
= MIN2(sctx
->screen
->b
.info
.num_render_backends
, 16);
4620 unsigned rb_per_pkr
= MIN2(num_rb
/ num_se
/ sh_per_se
, 2);
4621 unsigned rb_per_se
= num_rb
/ num_se
;
4622 unsigned se_mask
[4];
4625 se_mask
[0] = ((1 << rb_per_se
) - 1);
4626 se_mask
[1] = (se_mask
[0] << rb_per_se
);
4627 se_mask
[2] = (se_mask
[1] << rb_per_se
);
4628 se_mask
[3] = (se_mask
[2] << rb_per_se
);
4630 se_mask
[0] &= rb_mask
;
4631 se_mask
[1] &= rb_mask
;
4632 se_mask
[2] &= rb_mask
;
4633 se_mask
[3] &= rb_mask
;
4635 assert(num_se
== 1 || num_se
== 2 || num_se
== 4);
4636 assert(sh_per_se
== 1 || sh_per_se
== 2);
4637 assert(rb_per_pkr
== 1 || rb_per_pkr
== 2);
4639 /* XXX: I can't figure out what the *_XSEL and *_YSEL
4640 * fields are for, so I'm leaving them as their default
4643 for (se
= 0; se
< num_se
; se
++) {
4644 unsigned raster_config_se
= raster_config
;
4645 unsigned pkr0_mask
= ((1 << rb_per_pkr
) - 1) << (se
* rb_per_se
);
4646 unsigned pkr1_mask
= pkr0_mask
<< rb_per_pkr
;
4647 int idx
= (se
/ 2) * 2;
4649 if ((num_se
> 1) && (!se_mask
[idx
] || !se_mask
[idx
+ 1])) {
4650 raster_config_se
&= C_028350_SE_MAP
;
4652 if (!se_mask
[idx
]) {
4654 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_3
);
4657 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_0
);
4661 pkr0_mask
&= rb_mask
;
4662 pkr1_mask
&= rb_mask
;
4663 if (rb_per_se
> 2 && (!pkr0_mask
|| !pkr1_mask
)) {
4664 raster_config_se
&= C_028350_PKR_MAP
;
4668 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_3
);
4671 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_0
);
4675 if (rb_per_se
>= 2) {
4676 unsigned rb0_mask
= 1 << (se
* rb_per_se
);
4677 unsigned rb1_mask
= rb0_mask
<< 1;
4679 rb0_mask
&= rb_mask
;
4680 rb1_mask
&= rb_mask
;
4681 if (!rb0_mask
|| !rb1_mask
) {
4682 raster_config_se
&= C_028350_RB_MAP_PKR0
;
4686 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_3
);
4689 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_0
);
4693 if (rb_per_se
> 2) {
4694 rb0_mask
= 1 << (se
* rb_per_se
+ rb_per_pkr
);
4695 rb1_mask
= rb0_mask
<< 1;
4696 rb0_mask
&= rb_mask
;
4697 rb1_mask
&= rb_mask
;
4698 if (!rb0_mask
|| !rb1_mask
) {
4699 raster_config_se
&= C_028350_RB_MAP_PKR1
;
4703 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_3
);
4706 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_0
);
4712 si_set_grbm_gfx_index_se(sctx
, pm4
, se
);
4713 si_pm4_set_reg(pm4
, R_028350_PA_SC_RASTER_CONFIG
, raster_config_se
);
4715 si_set_grbm_gfx_index(sctx
, pm4
, ~0);
4717 if (sctx
->b
.chip_class
>= CIK
) {
4718 if ((num_se
> 2) && ((!se_mask
[0] && !se_mask
[1]) ||
4719 (!se_mask
[2] && !se_mask
[3]))) {
4720 raster_config_1
&= C_028354_SE_PAIR_MAP
;
4722 if (!se_mask
[0] && !se_mask
[1]) {
4724 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_3
);
4727 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_0
);
4731 si_pm4_set_reg(pm4
, R_028354_PA_SC_RASTER_CONFIG_1
, raster_config_1
);
4735 static void si_set_raster_config(struct si_context
*sctx
, struct si_pm4_state
*pm4
)
4737 struct si_screen
*sscreen
= sctx
->screen
;
4738 unsigned num_rb
= MIN2(sctx
->screen
->b
.info
.num_render_backends
, 16);
4739 unsigned rb_mask
= sctx
->screen
->b
.info
.enabled_rb_mask
;
4740 unsigned raster_config
, raster_config_1
;
4742 switch (sctx
->b
.family
) {
4745 raster_config
= 0x2a00126a;
4746 raster_config_1
= 0x00000000;
4749 raster_config
= 0x0000124a;
4750 raster_config_1
= 0x00000000;
4753 raster_config
= 0x00000082;
4754 raster_config_1
= 0x00000000;
4757 raster_config
= 0x00000000;
4758 raster_config_1
= 0x00000000;
4761 raster_config
= 0x16000012;
4762 raster_config_1
= 0x00000000;
4765 raster_config
= 0x3a00161a;
4766 raster_config_1
= 0x0000002e;
4769 if (sscreen
->b
.info
.cik_macrotile_mode_array
[0] == 0x000000e8) {
4770 /* old kernels with old tiling config */
4771 raster_config
= 0x16000012;
4772 raster_config_1
= 0x0000002a;
4774 raster_config
= 0x3a00161a;
4775 raster_config_1
= 0x0000002e;
4778 case CHIP_POLARIS10
:
4779 raster_config
= 0x16000012;
4780 raster_config_1
= 0x0000002a;
4782 case CHIP_POLARIS11
:
4783 case CHIP_POLARIS12
:
4784 raster_config
= 0x16000012;
4785 raster_config_1
= 0x00000000;
4788 raster_config
= 0x16000012;
4789 raster_config_1
= 0x0000002a;
4793 raster_config
= 0x00000000;
4795 raster_config
= 0x00000002;
4796 raster_config_1
= 0x00000000;
4799 raster_config
= 0x00000002;
4800 raster_config_1
= 0x00000000;
4803 /* KV should be 0x00000002, but that causes problems with radeon */
4804 raster_config
= 0x00000000; /* 0x00000002 */
4805 raster_config_1
= 0x00000000;
4810 raster_config
= 0x00000000;
4811 raster_config_1
= 0x00000000;
4815 "radeonsi: Unknown GPU, using 0 for raster_config\n");
4816 raster_config
= 0x00000000;
4817 raster_config_1
= 0x00000000;
4820 if (!rb_mask
|| util_bitcount(rb_mask
) >= num_rb
) {
4821 /* Always use the default config when all backends are enabled
4822 * (or when we failed to determine the enabled backends).
4824 si_pm4_set_reg(pm4
, R_028350_PA_SC_RASTER_CONFIG
,
4826 if (sctx
->b
.chip_class
>= CIK
)
4827 si_pm4_set_reg(pm4
, R_028354_PA_SC_RASTER_CONFIG_1
,
4830 si_write_harvested_raster_configs(sctx
, pm4
, raster_config
, raster_config_1
);
4834 static void si_init_config(struct si_context
*sctx
)
4836 struct si_screen
*sscreen
= sctx
->screen
;
4837 uint64_t border_color_va
= sctx
->border_color_buffer
->gpu_address
;
4838 bool has_clear_state
= sscreen
->has_clear_state
;
4839 struct si_pm4_state
*pm4
= CALLOC_STRUCT(si_pm4_state
);
4841 /* Only SI can disable CLEAR_STATE for now. */
4842 assert(has_clear_state
|| sscreen
->b
.chip_class
== SI
);
4847 si_pm4_cmd_begin(pm4
, PKT3_CONTEXT_CONTROL
);
4848 si_pm4_cmd_add(pm4
, CONTEXT_CONTROL_LOAD_ENABLE(1));
4849 si_pm4_cmd_add(pm4
, CONTEXT_CONTROL_SHADOW_ENABLE(1));
4850 si_pm4_cmd_end(pm4
, false);
4852 if (has_clear_state
) {
4853 si_pm4_cmd_begin(pm4
, PKT3_CLEAR_STATE
);
4854 si_pm4_cmd_add(pm4
, 0);
4855 si_pm4_cmd_end(pm4
, false);
4858 if (sctx
->b
.chip_class
<= VI
)
4859 si_set_raster_config(sctx
, pm4
);
4861 si_pm4_set_reg(pm4
, R_028A18_VGT_HOS_MAX_TESS_LEVEL
, fui(64));
4862 if (!has_clear_state
)
4863 si_pm4_set_reg(pm4
, R_028A1C_VGT_HOS_MIN_TESS_LEVEL
, fui(0));
4865 /* FIXME calculate these values somehow ??? */
4866 if (sctx
->b
.chip_class
<= VI
) {
4867 si_pm4_set_reg(pm4
, R_028A54_VGT_GS_PER_ES
, SI_GS_PER_ES
);
4868 si_pm4_set_reg(pm4
, R_028A58_VGT_ES_PER_GS
, 0x40);
4871 if (!has_clear_state
) {
4872 si_pm4_set_reg(pm4
, R_028A5C_VGT_GS_PER_VS
, 0x2);
4873 si_pm4_set_reg(pm4
, R_028A8C_VGT_PRIMITIVEID_RESET
, 0x0);
4874 si_pm4_set_reg(pm4
, R_028B98_VGT_STRMOUT_BUFFER_CONFIG
, 0x0);
4877 si_pm4_set_reg(pm4
, R_028AA0_VGT_INSTANCE_STEP_RATE_0
, 1);
4878 if (!has_clear_state
)
4879 si_pm4_set_reg(pm4
, R_028AB8_VGT_VTX_CNT_EN
, 0x0);
4880 if (sctx
->b
.chip_class
< CIK
)
4881 si_pm4_set_reg(pm4
, R_008A14_PA_CL_ENHANCE
, S_008A14_NUM_CLIP_SEQ(3) |
4882 S_008A14_CLIP_VTX_REORDER_ENA(1));
4884 si_pm4_set_reg(pm4
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 0x76543210);
4885 si_pm4_set_reg(pm4
, R_028BD8_PA_SC_CENTROID_PRIORITY_1
, 0xfedcba98);
4887 if (!has_clear_state
)
4888 si_pm4_set_reg(pm4
, R_02882C_PA_SU_PRIM_FILTER_CNTL
, 0);
4890 /* CLEAR_STATE doesn't clear these correctly on certain generations.
4891 * I don't know why. Deduced by trial and error.
4893 if (sctx
->b
.chip_class
<= CIK
) {
4894 si_pm4_set_reg(pm4
, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET
, 0);
4895 si_pm4_set_reg(pm4
, R_028204_PA_SC_WINDOW_SCISSOR_TL
, S_028204_WINDOW_OFFSET_DISABLE(1));
4896 si_pm4_set_reg(pm4
, R_028240_PA_SC_GENERIC_SCISSOR_TL
, S_028240_WINDOW_OFFSET_DISABLE(1));
4897 si_pm4_set_reg(pm4
, R_028244_PA_SC_GENERIC_SCISSOR_BR
,
4898 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
4899 si_pm4_set_reg(pm4
, R_028030_PA_SC_SCREEN_SCISSOR_TL
, 0);
4900 si_pm4_set_reg(pm4
, R_028034_PA_SC_SCREEN_SCISSOR_BR
,
4901 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
4904 if (!has_clear_state
) {
4905 si_pm4_set_reg(pm4
, R_02820C_PA_SC_CLIPRECT_RULE
, 0xFFFF);
4906 si_pm4_set_reg(pm4
, R_028230_PA_SC_EDGERULE
,
4907 S_028230_ER_TRI(0xA) |
4908 S_028230_ER_POINT(0xA) |
4909 S_028230_ER_RECT(0xA) |
4910 /* Required by DX10_DIAMOND_TEST_ENA: */
4911 S_028230_ER_LINE_LR(0x1A) |
4912 S_028230_ER_LINE_RL(0x26) |
4913 S_028230_ER_LINE_TB(0xA) |
4914 S_028230_ER_LINE_BT(0xA));
4915 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
4916 si_pm4_set_reg(pm4
, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET
, 0);
4917 si_pm4_set_reg(pm4
, R_028820_PA_CL_NANINF_CNTL
, 0);
4918 si_pm4_set_reg(pm4
, R_028AC0_DB_SRESULTS_COMPARE_STATE0
, 0x0);
4919 si_pm4_set_reg(pm4
, R_028AC4_DB_SRESULTS_COMPARE_STATE1
, 0x0);
4920 si_pm4_set_reg(pm4
, R_028AC8_DB_PRELOAD_CONTROL
, 0x0);
4921 si_pm4_set_reg(pm4
, R_02800C_DB_RENDER_OVERRIDE
, 0);
4924 if (sctx
->b
.chip_class
>= GFX9
) {
4925 si_pm4_set_reg(pm4
, R_030920_VGT_MAX_VTX_INDX
, ~0);
4926 si_pm4_set_reg(pm4
, R_030924_VGT_MIN_VTX_INDX
, 0);
4927 si_pm4_set_reg(pm4
, R_030928_VGT_INDX_OFFSET
, 0);
4929 /* These registers, when written, also overwrite the CLEAR_STATE
4930 * context, so we can't rely on CLEAR_STATE setting them.
4931 * It would be an issue if there was another UMD changing them.
4933 si_pm4_set_reg(pm4
, R_028400_VGT_MAX_VTX_INDX
, ~0);
4934 si_pm4_set_reg(pm4
, R_028404_VGT_MIN_VTX_INDX
, 0);
4935 si_pm4_set_reg(pm4
, R_028408_VGT_INDX_OFFSET
, 0);
4938 if (sctx
->b
.chip_class
>= CIK
) {
4939 if (sctx
->b
.chip_class
>= GFX9
) {
4940 si_pm4_set_reg(pm4
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
, S_00B41C_CU_EN(0xffff));
4942 si_pm4_set_reg(pm4
, R_00B51C_SPI_SHADER_PGM_RSRC3_LS
, S_00B51C_CU_EN(0xffff));
4943 si_pm4_set_reg(pm4
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
, 0);
4944 si_pm4_set_reg(pm4
, R_00B31C_SPI_SHADER_PGM_RSRC3_ES
, S_00B31C_CU_EN(0xffff));
4946 /* If this is 0, Bonaire can hang even if GS isn't being used.
4947 * Other chips are unaffected. These are suboptimal values,
4948 * but we don't use on-chip GS.
4950 si_pm4_set_reg(pm4
, R_028A44_VGT_GS_ONCHIP_CNTL
,
4951 S_028A44_ES_VERTS_PER_SUBGRP(64) |
4952 S_028A44_GS_PRIMS_PER_SUBGRP(4));
4954 si_pm4_set_reg(pm4
, R_00B21C_SPI_SHADER_PGM_RSRC3_GS
, S_00B21C_CU_EN(0xffff));
4956 /* Compute LATE_ALLOC_VS.LIMIT. */
4957 unsigned num_cu_per_sh
= sscreen
->b
.info
.num_good_compute_units
/
4958 (sscreen
->b
.info
.max_se
*
4959 sscreen
->b
.info
.max_sh_per_se
);
4960 unsigned late_alloc_limit
; /* The limit is per SH. */
4962 if (sctx
->b
.family
== CHIP_KABINI
) {
4963 late_alloc_limit
= 0; /* Potential hang on Kabini. */
4964 } else if (num_cu_per_sh
<= 4) {
4965 /* Too few available compute units per SH. Disallowing
4966 * VS to run on one CU could hurt us more than late VS
4967 * allocation would help.
4969 * 2 is the highest safe number that allows us to keep
4972 late_alloc_limit
= 2;
4974 /* This is a good initial value, allowing 1 late_alloc
4975 * wave per SIMD on num_cu - 2.
4977 late_alloc_limit
= (num_cu_per_sh
- 2) * 4;
4979 /* The limit is 0-based, so 0 means 1. */
4980 assert(late_alloc_limit
> 0 && late_alloc_limit
<= 64);
4981 late_alloc_limit
-= 1;
4984 /* VS can't execute on one CU if the limit is > 2. */
4985 si_pm4_set_reg(pm4
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
,
4986 S_00B118_CU_EN(late_alloc_limit
> 2 ? 0xfffe : 0xffff));
4987 si_pm4_set_reg(pm4
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
,
4988 S_00B11C_LIMIT(late_alloc_limit
));
4989 si_pm4_set_reg(pm4
, R_00B01C_SPI_SHADER_PGM_RSRC3_PS
, S_00B01C_CU_EN(0xffff));
4992 if (sctx
->b
.chip_class
>= VI
) {
4993 unsigned vgt_tess_distribution
;
4995 si_pm4_set_reg(pm4
, R_028424_CB_DCC_CONTROL
,
4996 S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
4997 S_028424_OVERWRITE_COMBINER_WATERMARK(4));
4999 vgt_tess_distribution
=
5000 S_028B50_ACCUM_ISOLINE(32) |
5001 S_028B50_ACCUM_TRI(11) |
5002 S_028B50_ACCUM_QUAD(11) |
5003 S_028B50_DONUT_SPLIT(16);
5005 /* Testing with Unigine Heaven extreme tesselation yielded best results
5006 * with TRAP_SPLIT = 3.
5008 if (sctx
->b
.family
== CHIP_FIJI
||
5009 sctx
->b
.family
>= CHIP_POLARIS10
)
5010 vgt_tess_distribution
|= S_028B50_TRAP_SPLIT(3);
5012 si_pm4_set_reg(pm4
, R_028B50_VGT_TESS_DISTRIBUTION
, vgt_tess_distribution
);
5013 } else if (!has_clear_state
) {
5014 si_pm4_set_reg(pm4
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 14);
5015 si_pm4_set_reg(pm4
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 16);
5018 si_pm4_set_reg(pm4
, R_028080_TA_BC_BASE_ADDR
, border_color_va
>> 8);
5019 if (sctx
->b
.chip_class
>= CIK
)
5020 si_pm4_set_reg(pm4
, R_028084_TA_BC_BASE_ADDR_HI
, border_color_va
>> 40);
5021 si_pm4_add_bo(pm4
, sctx
->border_color_buffer
, RADEON_USAGE_READ
,
5022 RADEON_PRIO_BORDER_COLORS
);
5024 if (sctx
->b
.chip_class
>= GFX9
) {
5025 unsigned num_se
= sscreen
->b
.info
.max_se
;
5026 unsigned pc_lines
= 0;
5028 switch (sctx
->b
.family
) {
5039 si_pm4_set_reg(pm4
, R_028C48_PA_SC_BINNER_CNTL_1
,
5040 S_028C48_MAX_ALLOC_COUNT(MIN2(128, pc_lines
/ (4 * num_se
))) |
5041 S_028C48_MAX_PRIM_PER_BATCH(1023));
5042 si_pm4_set_reg(pm4
, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL
,
5043 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
5044 si_pm4_set_reg(pm4
, R_030968_VGT_INSTANCE_BASE_ID
, 0);
5047 si_pm4_upload_indirect_buffer(sctx
, pm4
);
5048 sctx
->init_config
= pm4
;