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
)
1716 case PIPE_FORMAT_DXT1_RGB
:
1717 case PIPE_FORMAT_DXT1_RGBA
:
1718 case PIPE_FORMAT_DXT1_SRGB
:
1719 case PIPE_FORMAT_DXT1_SRGBA
:
1720 return V_008F14_IMG_DATA_FORMAT_BC1
;
1721 case PIPE_FORMAT_DXT3_RGBA
:
1722 case PIPE_FORMAT_DXT3_SRGBA
:
1723 return V_008F14_IMG_DATA_FORMAT_BC2
;
1724 case PIPE_FORMAT_DXT5_RGBA
:
1725 case PIPE_FORMAT_DXT5_SRGBA
:
1726 return V_008F14_IMG_DATA_FORMAT_BC3
;
1732 if (format
== PIPE_FORMAT_R9G9B9E5_FLOAT
) {
1733 return V_008F14_IMG_DATA_FORMAT_5_9_9_9
;
1734 } else if (format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1735 return V_008F14_IMG_DATA_FORMAT_10_11_11
;
1738 /* R8G8Bx_SNORM - TODO CxV8U8 */
1740 /* hw cannot support mixed formats (except depth/stencil, since only
1742 if (desc
->is_mixed
&& desc
->colorspace
!= UTIL_FORMAT_COLORSPACE_ZS
)
1745 /* See whether the components are of the same size. */
1746 for (i
= 1; i
< desc
->nr_channels
; i
++) {
1747 uniform
= uniform
&& desc
->channel
[0].size
== desc
->channel
[i
].size
;
1750 /* Non-uniform formats. */
1752 switch(desc
->nr_channels
) {
1754 if (desc
->channel
[0].size
== 5 &&
1755 desc
->channel
[1].size
== 6 &&
1756 desc
->channel
[2].size
== 5) {
1757 return V_008F14_IMG_DATA_FORMAT_5_6_5
;
1761 if (desc
->channel
[0].size
== 5 &&
1762 desc
->channel
[1].size
== 5 &&
1763 desc
->channel
[2].size
== 5 &&
1764 desc
->channel
[3].size
== 1) {
1765 return V_008F14_IMG_DATA_FORMAT_1_5_5_5
;
1767 if (desc
->channel
[0].size
== 10 &&
1768 desc
->channel
[1].size
== 10 &&
1769 desc
->channel
[2].size
== 10 &&
1770 desc
->channel
[3].size
== 2) {
1771 return V_008F14_IMG_DATA_FORMAT_2_10_10_10
;
1778 if (first_non_void
< 0 || first_non_void
> 3)
1781 /* uniform formats */
1782 switch (desc
->channel
[first_non_void
].size
) {
1784 switch (desc
->nr_channels
) {
1785 #if 0 /* Not supported for render targets */
1787 return V_008F14_IMG_DATA_FORMAT_4_4
;
1790 return V_008F14_IMG_DATA_FORMAT_4_4_4_4
;
1794 switch (desc
->nr_channels
) {
1796 return V_008F14_IMG_DATA_FORMAT_8
;
1798 return V_008F14_IMG_DATA_FORMAT_8_8
;
1800 return V_008F14_IMG_DATA_FORMAT_8_8_8_8
;
1804 switch (desc
->nr_channels
) {
1806 return V_008F14_IMG_DATA_FORMAT_16
;
1808 return V_008F14_IMG_DATA_FORMAT_16_16
;
1810 return V_008F14_IMG_DATA_FORMAT_16_16_16_16
;
1814 switch (desc
->nr_channels
) {
1816 return V_008F14_IMG_DATA_FORMAT_32
;
1818 return V_008F14_IMG_DATA_FORMAT_32_32
;
1819 #if 0 /* Not supported for render targets */
1821 return V_008F14_IMG_DATA_FORMAT_32_32_32
;
1824 return V_008F14_IMG_DATA_FORMAT_32_32_32_32
;
1829 /* R600_ERR("Unable to handle texformat %d %s\n", format, util_format_name(format)); */
1833 static unsigned si_tex_wrap(unsigned wrap
)
1837 case PIPE_TEX_WRAP_REPEAT
:
1838 return V_008F30_SQ_TEX_WRAP
;
1839 case PIPE_TEX_WRAP_CLAMP
:
1840 return V_008F30_SQ_TEX_CLAMP_HALF_BORDER
;
1841 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
1842 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
1843 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
1844 return V_008F30_SQ_TEX_CLAMP_BORDER
;
1845 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
1846 return V_008F30_SQ_TEX_MIRROR
;
1847 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
1848 return V_008F30_SQ_TEX_MIRROR_ONCE_HALF_BORDER
;
1849 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
1850 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
1851 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
1852 return V_008F30_SQ_TEX_MIRROR_ONCE_BORDER
;
1856 static unsigned si_tex_mipfilter(unsigned filter
)
1859 case PIPE_TEX_MIPFILTER_NEAREST
:
1860 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
1861 case PIPE_TEX_MIPFILTER_LINEAR
:
1862 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
1864 case PIPE_TEX_MIPFILTER_NONE
:
1865 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
1869 static unsigned si_tex_compare(unsigned compare
)
1873 case PIPE_FUNC_NEVER
:
1874 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
1875 case PIPE_FUNC_LESS
:
1876 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
1877 case PIPE_FUNC_EQUAL
:
1878 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
1879 case PIPE_FUNC_LEQUAL
:
1880 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
1881 case PIPE_FUNC_GREATER
:
1882 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
1883 case PIPE_FUNC_NOTEQUAL
:
1884 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
1885 case PIPE_FUNC_GEQUAL
:
1886 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
1887 case PIPE_FUNC_ALWAYS
:
1888 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
1892 static unsigned si_tex_dim(struct si_screen
*sscreen
, struct r600_texture
*rtex
,
1893 unsigned view_target
, unsigned nr_samples
)
1895 unsigned res_target
= rtex
->resource
.b
.b
.target
;
1897 if (view_target
== PIPE_TEXTURE_CUBE
||
1898 view_target
== PIPE_TEXTURE_CUBE_ARRAY
)
1899 res_target
= view_target
;
1900 /* If interpreting cubemaps as something else, set 2D_ARRAY. */
1901 else if (res_target
== PIPE_TEXTURE_CUBE
||
1902 res_target
== PIPE_TEXTURE_CUBE_ARRAY
)
1903 res_target
= PIPE_TEXTURE_2D_ARRAY
;
1905 /* GFX9 allocates 1D textures as 2D. */
1906 if ((res_target
== PIPE_TEXTURE_1D
||
1907 res_target
== PIPE_TEXTURE_1D_ARRAY
) &&
1908 sscreen
->b
.chip_class
>= GFX9
&&
1909 rtex
->surface
.u
.gfx9
.resource_type
== RADEON_RESOURCE_2D
) {
1910 if (res_target
== PIPE_TEXTURE_1D
)
1911 res_target
= PIPE_TEXTURE_2D
;
1913 res_target
= PIPE_TEXTURE_2D_ARRAY
;
1916 switch (res_target
) {
1918 case PIPE_TEXTURE_1D
:
1919 return V_008F1C_SQ_RSRC_IMG_1D
;
1920 case PIPE_TEXTURE_1D_ARRAY
:
1921 return V_008F1C_SQ_RSRC_IMG_1D_ARRAY
;
1922 case PIPE_TEXTURE_2D
:
1923 case PIPE_TEXTURE_RECT
:
1924 return nr_samples
> 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA
:
1925 V_008F1C_SQ_RSRC_IMG_2D
;
1926 case PIPE_TEXTURE_2D_ARRAY
:
1927 return nr_samples
> 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
:
1928 V_008F1C_SQ_RSRC_IMG_2D_ARRAY
;
1929 case PIPE_TEXTURE_3D
:
1930 return V_008F1C_SQ_RSRC_IMG_3D
;
1931 case PIPE_TEXTURE_CUBE
:
1932 case PIPE_TEXTURE_CUBE_ARRAY
:
1933 return V_008F1C_SQ_RSRC_IMG_CUBE
;
1938 * Format support testing
1941 static bool si_is_sampler_format_supported(struct pipe_screen
*screen
, enum pipe_format format
)
1943 const struct util_format_description
*desc
= util_format_description(format
);
1947 return si_translate_texformat(screen
, format
, desc
,
1948 util_format_get_first_non_void_channel(format
)) != ~0U;
1951 static uint32_t si_translate_buffer_dataformat(struct pipe_screen
*screen
,
1952 const struct util_format_description
*desc
,
1957 if (desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
)
1958 return V_008F0C_BUF_DATA_FORMAT_10_11_11
;
1960 assert(first_non_void
>= 0);
1962 if (desc
->nr_channels
== 4 &&
1963 desc
->channel
[0].size
== 10 &&
1964 desc
->channel
[1].size
== 10 &&
1965 desc
->channel
[2].size
== 10 &&
1966 desc
->channel
[3].size
== 2)
1967 return V_008F0C_BUF_DATA_FORMAT_2_10_10_10
;
1969 /* See whether the components are of the same size. */
1970 for (i
= 0; i
< desc
->nr_channels
; i
++) {
1971 if (desc
->channel
[first_non_void
].size
!= desc
->channel
[i
].size
)
1972 return V_008F0C_BUF_DATA_FORMAT_INVALID
;
1975 switch (desc
->channel
[first_non_void
].size
) {
1977 switch (desc
->nr_channels
) {
1979 case 3: /* 3 loads */
1980 return V_008F0C_BUF_DATA_FORMAT_8
;
1982 return V_008F0C_BUF_DATA_FORMAT_8_8
;
1984 return V_008F0C_BUF_DATA_FORMAT_8_8_8_8
;
1988 switch (desc
->nr_channels
) {
1990 case 3: /* 3 loads */
1991 return V_008F0C_BUF_DATA_FORMAT_16
;
1993 return V_008F0C_BUF_DATA_FORMAT_16_16
;
1995 return V_008F0C_BUF_DATA_FORMAT_16_16_16_16
;
1999 switch (desc
->nr_channels
) {
2001 return V_008F0C_BUF_DATA_FORMAT_32
;
2003 return V_008F0C_BUF_DATA_FORMAT_32_32
;
2005 return V_008F0C_BUF_DATA_FORMAT_32_32_32
;
2007 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32
;
2011 /* Legacy double formats. */
2012 switch (desc
->nr_channels
) {
2013 case 1: /* 1 load */
2014 return V_008F0C_BUF_DATA_FORMAT_32_32
;
2015 case 2: /* 1 load */
2016 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32
;
2017 case 3: /* 3 loads */
2018 return V_008F0C_BUF_DATA_FORMAT_32_32
;
2019 case 4: /* 2 loads */
2020 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32
;
2025 return V_008F0C_BUF_DATA_FORMAT_INVALID
;
2028 static uint32_t si_translate_buffer_numformat(struct pipe_screen
*screen
,
2029 const struct util_format_description
*desc
,
2032 if (desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
)
2033 return V_008F0C_BUF_NUM_FORMAT_FLOAT
;
2035 assert(first_non_void
>= 0);
2037 switch (desc
->channel
[first_non_void
].type
) {
2038 case UTIL_FORMAT_TYPE_SIGNED
:
2039 case UTIL_FORMAT_TYPE_FIXED
:
2040 if (desc
->channel
[first_non_void
].size
>= 32 ||
2041 desc
->channel
[first_non_void
].pure_integer
)
2042 return V_008F0C_BUF_NUM_FORMAT_SINT
;
2043 else if (desc
->channel
[first_non_void
].normalized
)
2044 return V_008F0C_BUF_NUM_FORMAT_SNORM
;
2046 return V_008F0C_BUF_NUM_FORMAT_SSCALED
;
2048 case UTIL_FORMAT_TYPE_UNSIGNED
:
2049 if (desc
->channel
[first_non_void
].size
>= 32 ||
2050 desc
->channel
[first_non_void
].pure_integer
)
2051 return V_008F0C_BUF_NUM_FORMAT_UINT
;
2052 else if (desc
->channel
[first_non_void
].normalized
)
2053 return V_008F0C_BUF_NUM_FORMAT_UNORM
;
2055 return V_008F0C_BUF_NUM_FORMAT_USCALED
;
2057 case UTIL_FORMAT_TYPE_FLOAT
:
2059 return V_008F0C_BUF_NUM_FORMAT_FLOAT
;
2063 static unsigned si_is_vertex_format_supported(struct pipe_screen
*screen
,
2064 enum pipe_format format
,
2067 const struct util_format_description
*desc
;
2069 unsigned data_format
;
2071 assert((usage
& ~(PIPE_BIND_SHADER_IMAGE
|
2072 PIPE_BIND_SAMPLER_VIEW
|
2073 PIPE_BIND_VERTEX_BUFFER
)) == 0);
2075 desc
= util_format_description(format
);
2079 /* There are no native 8_8_8 or 16_16_16 data formats, and we currently
2080 * select 8_8_8_8 and 16_16_16_16 instead. This works reasonably well
2081 * for read-only access (with caveats surrounding bounds checks), but
2082 * obviously fails for write access which we have to implement for
2083 * shader images. Luckily, OpenGL doesn't expect this to be supported
2084 * anyway, and so the only impact is on PBO uploads / downloads, which
2085 * shouldn't be expected to be fast for GL_RGB anyway.
2087 if (desc
->block
.bits
== 3 * 8 ||
2088 desc
->block
.bits
== 3 * 16) {
2089 if (usage
& (PIPE_BIND_SHADER_IMAGE
| PIPE_BIND_SAMPLER_VIEW
)) {
2090 usage
&= ~(PIPE_BIND_SHADER_IMAGE
| PIPE_BIND_SAMPLER_VIEW
);
2096 first_non_void
= util_format_get_first_non_void_channel(format
);
2097 data_format
= si_translate_buffer_dataformat(screen
, desc
, first_non_void
);
2098 if (data_format
== V_008F0C_BUF_DATA_FORMAT_INVALID
)
2104 static bool si_is_colorbuffer_format_supported(enum pipe_format format
)
2106 return si_translate_colorformat(format
) != V_028C70_COLOR_INVALID
&&
2107 si_translate_colorswap(format
, false) != ~0U;
2110 static bool si_is_zs_format_supported(enum pipe_format format
)
2112 return si_translate_dbformat(format
) != V_028040_Z_INVALID
;
2115 static boolean
si_is_format_supported(struct pipe_screen
*screen
,
2116 enum pipe_format format
,
2117 enum pipe_texture_target target
,
2118 unsigned sample_count
,
2121 unsigned retval
= 0;
2123 if (target
>= PIPE_MAX_TEXTURE_TYPES
) {
2124 R600_ERR("r600: unsupported texture type %d\n", target
);
2128 if (!util_format_is_supported(format
, usage
))
2131 if (sample_count
> 1) {
2132 if (!screen
->get_param(screen
, PIPE_CAP_TEXTURE_MULTISAMPLE
))
2135 if (usage
& PIPE_BIND_SHADER_IMAGE
)
2138 switch (sample_count
) {
2144 if (format
== PIPE_FORMAT_NONE
)
2153 if (usage
& (PIPE_BIND_SAMPLER_VIEW
|
2154 PIPE_BIND_SHADER_IMAGE
)) {
2155 if (target
== PIPE_BUFFER
) {
2156 retval
|= si_is_vertex_format_supported(
2157 screen
, format
, usage
& (PIPE_BIND_SAMPLER_VIEW
|
2158 PIPE_BIND_SHADER_IMAGE
));
2160 if (si_is_sampler_format_supported(screen
, format
))
2161 retval
|= usage
& (PIPE_BIND_SAMPLER_VIEW
|
2162 PIPE_BIND_SHADER_IMAGE
);
2166 if ((usage
& (PIPE_BIND_RENDER_TARGET
|
2167 PIPE_BIND_DISPLAY_TARGET
|
2170 PIPE_BIND_BLENDABLE
)) &&
2171 si_is_colorbuffer_format_supported(format
)) {
2173 (PIPE_BIND_RENDER_TARGET
|
2174 PIPE_BIND_DISPLAY_TARGET
|
2177 if (!util_format_is_pure_integer(format
) &&
2178 !util_format_is_depth_or_stencil(format
))
2179 retval
|= usage
& PIPE_BIND_BLENDABLE
;
2182 if ((usage
& PIPE_BIND_DEPTH_STENCIL
) &&
2183 si_is_zs_format_supported(format
)) {
2184 retval
|= PIPE_BIND_DEPTH_STENCIL
;
2187 if (usage
& PIPE_BIND_VERTEX_BUFFER
) {
2188 retval
|= si_is_vertex_format_supported(screen
, format
,
2189 PIPE_BIND_VERTEX_BUFFER
);
2192 if ((usage
& PIPE_BIND_LINEAR
) &&
2193 !util_format_is_compressed(format
) &&
2194 !(usage
& PIPE_BIND_DEPTH_STENCIL
))
2195 retval
|= PIPE_BIND_LINEAR
;
2197 return retval
== usage
;
2201 * framebuffer handling
2204 static void si_choose_spi_color_formats(struct r600_surface
*surf
,
2205 unsigned format
, unsigned swap
,
2206 unsigned ntype
, bool is_depth
)
2208 /* Alpha is needed for alpha-to-coverage.
2209 * Blending may be with or without alpha.
2211 unsigned normal
= 0; /* most optimal, may not support blending or export alpha */
2212 unsigned alpha
= 0; /* exports alpha, but may not support blending */
2213 unsigned blend
= 0; /* supports blending, but may not export alpha */
2214 unsigned blend_alpha
= 0; /* least optimal, supports blending and exports alpha */
2216 /* Choose the SPI color formats. These are required values for RB+.
2217 * Other chips have multiple choices, though they are not necessarily better.
2220 case V_028C70_COLOR_5_6_5
:
2221 case V_028C70_COLOR_1_5_5_5
:
2222 case V_028C70_COLOR_5_5_5_1
:
2223 case V_028C70_COLOR_4_4_4_4
:
2224 case V_028C70_COLOR_10_11_11
:
2225 case V_028C70_COLOR_11_11_10
:
2226 case V_028C70_COLOR_8
:
2227 case V_028C70_COLOR_8_8
:
2228 case V_028C70_COLOR_8_8_8_8
:
2229 case V_028C70_COLOR_10_10_10_2
:
2230 case V_028C70_COLOR_2_10_10_10
:
2231 if (ntype
== V_028C70_NUMBER_UINT
)
2232 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
2233 else if (ntype
== V_028C70_NUMBER_SINT
)
2234 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
2236 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
2239 case V_028C70_COLOR_16
:
2240 case V_028C70_COLOR_16_16
:
2241 case V_028C70_COLOR_16_16_16_16
:
2242 if (ntype
== V_028C70_NUMBER_UNORM
||
2243 ntype
== V_028C70_NUMBER_SNORM
) {
2244 /* UNORM16 and SNORM16 don't support blending */
2245 if (ntype
== V_028C70_NUMBER_UNORM
)
2246 normal
= alpha
= V_028714_SPI_SHADER_UNORM16_ABGR
;
2248 normal
= alpha
= V_028714_SPI_SHADER_SNORM16_ABGR
;
2250 /* Use 32 bits per channel for blending. */
2251 if (format
== V_028C70_COLOR_16
) {
2252 if (swap
== V_028C70_SWAP_STD
) { /* R */
2253 blend
= V_028714_SPI_SHADER_32_R
;
2254 blend_alpha
= V_028714_SPI_SHADER_32_AR
;
2255 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
2256 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
2259 } else if (format
== V_028C70_COLOR_16_16
) {
2260 if (swap
== V_028C70_SWAP_STD
) { /* RG */
2261 blend
= V_028714_SPI_SHADER_32_GR
;
2262 blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
2263 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
2264 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
2267 } else /* 16_16_16_16 */
2268 blend
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
2269 } else if (ntype
== V_028C70_NUMBER_UINT
)
2270 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
2271 else if (ntype
== V_028C70_NUMBER_SINT
)
2272 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
2273 else if (ntype
== V_028C70_NUMBER_FLOAT
)
2274 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
2279 case V_028C70_COLOR_32
:
2280 if (swap
== V_028C70_SWAP_STD
) { /* R */
2281 blend
= normal
= V_028714_SPI_SHADER_32_R
;
2282 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
2283 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
2284 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
2289 case V_028C70_COLOR_32_32
:
2290 if (swap
== V_028C70_SWAP_STD
) { /* RG */
2291 blend
= normal
= V_028714_SPI_SHADER_32_GR
;
2292 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
2293 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
2294 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
2299 case V_028C70_COLOR_32_32_32_32
:
2300 case V_028C70_COLOR_8_24
:
2301 case V_028C70_COLOR_24_8
:
2302 case V_028C70_COLOR_X24_8_32_FLOAT
:
2303 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_ABGR
;
2311 /* The DB->CB copy needs 32_ABGR. */
2313 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_ABGR
;
2315 surf
->spi_shader_col_format
= normal
;
2316 surf
->spi_shader_col_format_alpha
= alpha
;
2317 surf
->spi_shader_col_format_blend
= blend
;
2318 surf
->spi_shader_col_format_blend_alpha
= blend_alpha
;
2321 static void si_initialize_color_surface(struct si_context
*sctx
,
2322 struct r600_surface
*surf
)
2324 struct r600_texture
*rtex
= (struct r600_texture
*)surf
->base
.texture
;
2325 unsigned color_info
, color_attrib
, color_view
;
2326 unsigned format
, swap
, ntype
, endian
;
2327 const struct util_format_description
*desc
;
2329 unsigned blend_clamp
= 0, blend_bypass
= 0;
2331 color_view
= S_028C6C_SLICE_START(surf
->base
.u
.tex
.first_layer
) |
2332 S_028C6C_SLICE_MAX(surf
->base
.u
.tex
.last_layer
);
2334 desc
= util_format_description(surf
->base
.format
);
2335 for (firstchan
= 0; firstchan
< 4; firstchan
++) {
2336 if (desc
->channel
[firstchan
].type
!= UTIL_FORMAT_TYPE_VOID
) {
2340 if (firstchan
== 4 || desc
->channel
[firstchan
].type
== UTIL_FORMAT_TYPE_FLOAT
) {
2341 ntype
= V_028C70_NUMBER_FLOAT
;
2343 ntype
= V_028C70_NUMBER_UNORM
;
2344 if (desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
)
2345 ntype
= V_028C70_NUMBER_SRGB
;
2346 else if (desc
->channel
[firstchan
].type
== UTIL_FORMAT_TYPE_SIGNED
) {
2347 if (desc
->channel
[firstchan
].pure_integer
) {
2348 ntype
= V_028C70_NUMBER_SINT
;
2350 assert(desc
->channel
[firstchan
].normalized
);
2351 ntype
= V_028C70_NUMBER_SNORM
;
2353 } else if (desc
->channel
[firstchan
].type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
2354 if (desc
->channel
[firstchan
].pure_integer
) {
2355 ntype
= V_028C70_NUMBER_UINT
;
2357 assert(desc
->channel
[firstchan
].normalized
);
2358 ntype
= V_028C70_NUMBER_UNORM
;
2363 format
= si_translate_colorformat(surf
->base
.format
);
2364 if (format
== V_028C70_COLOR_INVALID
) {
2365 R600_ERR("Invalid CB format: %d, disabling CB.\n", surf
->base
.format
);
2367 assert(format
!= V_028C70_COLOR_INVALID
);
2368 swap
= si_translate_colorswap(surf
->base
.format
, false);
2369 endian
= si_colorformat_endian_swap(format
);
2371 /* blend clamp should be set for all NORM/SRGB types */
2372 if (ntype
== V_028C70_NUMBER_UNORM
||
2373 ntype
== V_028C70_NUMBER_SNORM
||
2374 ntype
== V_028C70_NUMBER_SRGB
)
2377 /* set blend bypass according to docs if SINT/UINT or
2378 8/24 COLOR variants */
2379 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
2380 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
2381 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
2386 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) {
2387 if (format
== V_028C70_COLOR_8
||
2388 format
== V_028C70_COLOR_8_8
||
2389 format
== V_028C70_COLOR_8_8_8_8
)
2390 surf
->color_is_int8
= true;
2391 else if (format
== V_028C70_COLOR_10_10_10_2
||
2392 format
== V_028C70_COLOR_2_10_10_10
)
2393 surf
->color_is_int10
= true;
2396 color_info
= S_028C70_FORMAT(format
) |
2397 S_028C70_COMP_SWAP(swap
) |
2398 S_028C70_BLEND_CLAMP(blend_clamp
) |
2399 S_028C70_BLEND_BYPASS(blend_bypass
) |
2400 S_028C70_SIMPLE_FLOAT(1) |
2401 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
2402 ntype
!= V_028C70_NUMBER_SNORM
&&
2403 ntype
!= V_028C70_NUMBER_SRGB
&&
2404 format
!= V_028C70_COLOR_8_24
&&
2405 format
!= V_028C70_COLOR_24_8
) |
2406 S_028C70_NUMBER_TYPE(ntype
) |
2407 S_028C70_ENDIAN(endian
);
2409 /* Intensity is implemented as Red, so treat it that way. */
2410 color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == PIPE_SWIZZLE_1
||
2411 util_format_is_intensity(surf
->base
.format
));
2413 if (rtex
->resource
.b
.b
.nr_samples
> 1) {
2414 unsigned log_samples
= util_logbase2(rtex
->resource
.b
.b
.nr_samples
);
2416 color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
2417 S_028C74_NUM_FRAGMENTS(log_samples
);
2419 if (rtex
->fmask
.size
) {
2420 color_info
|= S_028C70_COMPRESSION(1);
2421 unsigned fmask_bankh
= util_logbase2(rtex
->fmask
.bank_height
);
2423 if (sctx
->b
.chip_class
== SI
) {
2424 /* due to a hw bug, FMASK_BANK_HEIGHT must be set on SI too */
2425 color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
2430 surf
->cb_color_view
= color_view
;
2431 surf
->cb_color_info
= color_info
;
2432 surf
->cb_color_attrib
= color_attrib
;
2434 if (sctx
->b
.chip_class
>= VI
) {
2435 unsigned max_uncompressed_block_size
= 2;
2437 if (rtex
->resource
.b
.b
.nr_samples
> 1) {
2438 if (rtex
->surface
.bpe
== 1)
2439 max_uncompressed_block_size
= 0;
2440 else if (rtex
->surface
.bpe
== 2)
2441 max_uncompressed_block_size
= 1;
2444 surf
->cb_dcc_control
= S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
2445 S_028C78_INDEPENDENT_64B_BLOCKS(1);
2448 /* This must be set for fast clear to work without FMASK. */
2449 if (!rtex
->fmask
.size
&& sctx
->b
.chip_class
== SI
) {
2450 unsigned bankh
= util_logbase2(rtex
->surface
.u
.legacy
.bankh
);
2451 surf
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
2454 if (sctx
->b
.chip_class
>= GFX9
) {
2455 unsigned mip0_depth
= util_max_layer(&rtex
->resource
.b
.b
, 0);
2457 surf
->cb_color_view
|= S_028C6C_MIP_LEVEL(surf
->base
.u
.tex
.level
);
2458 surf
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
2459 S_028C74_RESOURCE_TYPE(rtex
->surface
.u
.gfx9
.resource_type
);
2460 surf
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(surf
->width0
- 1) |
2461 S_028C68_MIP0_HEIGHT(surf
->height0
- 1) |
2462 S_028C68_MAX_MIP(rtex
->resource
.b
.b
.last_level
);
2465 /* Determine pixel shader export format */
2466 si_choose_spi_color_formats(surf
, format
, swap
, ntype
, rtex
->is_depth
);
2468 surf
->color_initialized
= true;
2471 static void si_init_depth_surface(struct si_context
*sctx
,
2472 struct r600_surface
*surf
)
2474 struct r600_texture
*rtex
= (struct r600_texture
*)surf
->base
.texture
;
2475 unsigned level
= surf
->base
.u
.tex
.level
;
2476 unsigned format
, stencil_format
;
2477 uint32_t z_info
, s_info
;
2479 format
= si_translate_dbformat(rtex
->db_render_format
);
2480 stencil_format
= rtex
->surface
.has_stencil
?
2481 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
2483 assert(format
!= V_028040_Z_INVALID
);
2484 if (format
== V_028040_Z_INVALID
)
2485 R600_ERR("Invalid DB format: %d, disabling DB.\n", rtex
->resource
.b
.b
.format
);
2487 surf
->db_depth_view
= S_028008_SLICE_START(surf
->base
.u
.tex
.first_layer
) |
2488 S_028008_SLICE_MAX(surf
->base
.u
.tex
.last_layer
);
2489 surf
->db_htile_data_base
= 0;
2490 surf
->db_htile_surface
= 0;
2492 if (sctx
->b
.chip_class
>= GFX9
) {
2493 assert(rtex
->surface
.u
.gfx9
.surf_offset
== 0);
2494 surf
->db_depth_base
= rtex
->resource
.gpu_address
>> 8;
2495 surf
->db_stencil_base
= (rtex
->resource
.gpu_address
+
2496 rtex
->surface
.u
.gfx9
.stencil_offset
) >> 8;
2497 z_info
= S_028038_FORMAT(format
) |
2498 S_028038_NUM_SAMPLES(util_logbase2(rtex
->resource
.b
.b
.nr_samples
)) |
2499 S_028038_SW_MODE(rtex
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
2500 S_028038_MAXMIP(rtex
->resource
.b
.b
.last_level
);
2501 s_info
= S_02803C_FORMAT(stencil_format
) |
2502 S_02803C_SW_MODE(rtex
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
2503 surf
->db_z_info2
= S_028068_EPITCH(rtex
->surface
.u
.gfx9
.surf
.epitch
);
2504 surf
->db_stencil_info2
= S_02806C_EPITCH(rtex
->surface
.u
.gfx9
.stencil
.epitch
);
2505 surf
->db_depth_view
|= S_028008_MIPID(level
);
2506 surf
->db_depth_size
= S_02801C_X_MAX(rtex
->resource
.b
.b
.width0
- 1) |
2507 S_02801C_Y_MAX(rtex
->resource
.b
.b
.height0
- 1);
2509 if (r600_htile_enabled(rtex
, level
)) {
2510 z_info
|= S_028038_TILE_SURFACE_ENABLE(1) |
2511 S_028038_ALLOW_EXPCLEAR(1);
2513 if (rtex
->tc_compatible_htile
) {
2514 unsigned max_zplanes
= 4;
2516 if (rtex
->db_render_format
== PIPE_FORMAT_Z16_UNORM
&&
2517 rtex
->resource
.b
.b
.nr_samples
> 1)
2520 z_info
|= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes
+ 1) |
2521 S_028038_ITERATE_FLUSH(1);
2522 s_info
|= S_02803C_ITERATE_FLUSH(1);
2525 if (rtex
->surface
.has_stencil
) {
2526 /* Stencil buffer workaround ported from the SI-CI-VI code.
2527 * See that for explanation.
2529 s_info
|= S_02803C_ALLOW_EXPCLEAR(rtex
->resource
.b
.b
.nr_samples
<= 1);
2531 /* Use all HTILE for depth if there's no stencil. */
2532 s_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
2535 surf
->db_htile_data_base
= (rtex
->resource
.gpu_address
+
2536 rtex
->htile_offset
) >> 8;
2537 surf
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
2538 S_028ABC_PIPE_ALIGNED(rtex
->surface
.u
.gfx9
.htile
.pipe_aligned
) |
2539 S_028ABC_RB_ALIGNED(rtex
->surface
.u
.gfx9
.htile
.rb_aligned
);
2543 struct legacy_surf_level
*levelinfo
= &rtex
->surface
.u
.legacy
.level
[level
];
2545 assert(levelinfo
->nblk_x
% 8 == 0 && levelinfo
->nblk_y
% 8 == 0);
2547 surf
->db_depth_base
= (rtex
->resource
.gpu_address
+
2548 rtex
->surface
.u
.legacy
.level
[level
].offset
) >> 8;
2549 surf
->db_stencil_base
= (rtex
->resource
.gpu_address
+
2550 rtex
->surface
.u
.legacy
.stencil_level
[level
].offset
) >> 8;
2552 z_info
= S_028040_FORMAT(format
) |
2553 S_028040_NUM_SAMPLES(util_logbase2(rtex
->resource
.b
.b
.nr_samples
));
2554 s_info
= S_028044_FORMAT(stencil_format
);
2555 surf
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(!rtex
->tc_compatible_htile
);
2557 if (sctx
->b
.chip_class
>= CIK
) {
2558 struct radeon_info
*info
= &sctx
->screen
->b
.info
;
2559 unsigned index
= rtex
->surface
.u
.legacy
.tiling_index
[level
];
2560 unsigned stencil_index
= rtex
->surface
.u
.legacy
.stencil_tiling_index
[level
];
2561 unsigned macro_index
= rtex
->surface
.u
.legacy
.macro_tile_index
;
2562 unsigned tile_mode
= info
->si_tile_mode_array
[index
];
2563 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
2564 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
2566 surf
->db_depth_info
|=
2567 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
2568 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
2569 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
2570 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
2571 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
2572 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
2573 z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
2574 s_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
2576 unsigned tile_mode_index
= si_tile_mode_index(rtex
, level
, false);
2577 z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
2578 tile_mode_index
= si_tile_mode_index(rtex
, level
, true);
2579 s_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
2582 surf
->db_depth_size
= S_028058_PITCH_TILE_MAX((levelinfo
->nblk_x
/ 8) - 1) |
2583 S_028058_HEIGHT_TILE_MAX((levelinfo
->nblk_y
/ 8) - 1);
2584 surf
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((levelinfo
->nblk_x
*
2585 levelinfo
->nblk_y
) / 64 - 1);
2587 if (r600_htile_enabled(rtex
, level
)) {
2588 z_info
|= S_028040_TILE_SURFACE_ENABLE(1) |
2589 S_028040_ALLOW_EXPCLEAR(1);
2591 if (rtex
->surface
.has_stencil
) {
2592 /* Workaround: For a not yet understood reason, the
2593 * combination of MSAA, fast stencil clear and stencil
2594 * decompress messes with subsequent stencil buffer
2595 * uses. Problem was reproduced on Verde, Bonaire,
2596 * Tonga, and Carrizo.
2598 * Disabling EXPCLEAR works around the problem.
2600 * Check piglit's arb_texture_multisample-stencil-clear
2601 * test if you want to try changing this.
2603 if (rtex
->resource
.b
.b
.nr_samples
<= 1)
2604 s_info
|= S_028044_ALLOW_EXPCLEAR(1);
2605 } else if (!rtex
->tc_compatible_htile
) {
2606 /* Use all of the htile_buffer for depth if there's no stencil.
2607 * This must not be set when TC-compatible HTILE is enabled
2610 s_info
|= S_028044_TILE_STENCIL_DISABLE(1);
2613 surf
->db_htile_data_base
= (rtex
->resource
.gpu_address
+
2614 rtex
->htile_offset
) >> 8;
2615 surf
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
2617 if (rtex
->tc_compatible_htile
) {
2618 surf
->db_htile_surface
|= S_028ABC_TC_COMPATIBLE(1);
2620 if (rtex
->resource
.b
.b
.nr_samples
<= 1)
2621 z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(5);
2622 else if (rtex
->resource
.b
.b
.nr_samples
<= 4)
2623 z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(3);
2625 z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(2);
2630 surf
->db_z_info
= z_info
;
2631 surf
->db_stencil_info
= s_info
;
2633 surf
->depth_initialized
= true;
2636 void si_update_fb_dirtiness_after_rendering(struct si_context
*sctx
)
2638 if (sctx
->decompression_enabled
)
2641 if (sctx
->framebuffer
.state
.zsbuf
) {
2642 struct pipe_surface
*surf
= sctx
->framebuffer
.state
.zsbuf
;
2643 struct r600_texture
*rtex
= (struct r600_texture
*)surf
->texture
;
2645 rtex
->dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
2647 if (rtex
->surface
.has_stencil
)
2648 rtex
->stencil_dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
2651 unsigned compressed_cb_mask
= sctx
->framebuffer
.compressed_cb_mask
;
2652 while (compressed_cb_mask
) {
2653 unsigned i
= u_bit_scan(&compressed_cb_mask
);
2654 struct pipe_surface
*surf
= sctx
->framebuffer
.state
.cbufs
[i
];
2655 struct r600_texture
*rtex
= (struct r600_texture
*)surf
->texture
;
2657 if (rtex
->fmask
.size
)
2658 rtex
->dirty_level_mask
|= 1 << surf
->u
.tex
.level
;
2659 if (rtex
->dcc_gather_statistics
)
2660 rtex
->separate_dcc_dirty
= true;
2664 static void si_dec_framebuffer_counters(const struct pipe_framebuffer_state
*state
)
2666 for (int i
= 0; i
< state
->nr_cbufs
; ++i
) {
2667 struct r600_surface
*surf
= NULL
;
2668 struct r600_texture
*rtex
;
2670 if (!state
->cbufs
[i
])
2672 surf
= (struct r600_surface
*)state
->cbufs
[i
];
2673 rtex
= (struct r600_texture
*)surf
->base
.texture
;
2675 p_atomic_dec(&rtex
->framebuffers_bound
);
2679 static void si_set_framebuffer_state(struct pipe_context
*ctx
,
2680 const struct pipe_framebuffer_state
*state
)
2682 struct si_context
*sctx
= (struct si_context
*)ctx
;
2683 struct pipe_constant_buffer constbuf
= {0};
2684 struct r600_surface
*surf
= NULL
;
2685 struct r600_texture
*rtex
;
2686 bool old_any_dst_linear
= sctx
->framebuffer
.any_dst_linear
;
2687 unsigned old_nr_samples
= sctx
->framebuffer
.nr_samples
;
2688 unsigned old_colorbuf_enabled_4bit
= sctx
->framebuffer
.colorbuf_enabled_4bit
;
2689 bool old_has_zsbuf
= !!sctx
->framebuffer
.state
.zsbuf
;
2690 bool old_has_stencil
=
2692 ((struct r600_texture
*)sctx
->framebuffer
.state
.zsbuf
->texture
)->surface
.has_stencil
;
2693 bool unbound
= false;
2696 si_update_fb_dirtiness_after_rendering(sctx
);
2698 for (i
= 0; i
< sctx
->framebuffer
.state
.nr_cbufs
; i
++) {
2699 if (!sctx
->framebuffer
.state
.cbufs
[i
])
2702 rtex
= (struct r600_texture
*)sctx
->framebuffer
.state
.cbufs
[i
]->texture
;
2703 if (rtex
->dcc_gather_statistics
)
2704 vi_separate_dcc_stop_query(ctx
, rtex
);
2707 /* Disable DCC if the formats are incompatible. */
2708 for (i
= 0; i
< state
->nr_cbufs
; i
++) {
2709 if (!state
->cbufs
[i
])
2712 surf
= (struct r600_surface
*)state
->cbufs
[i
];
2713 rtex
= (struct r600_texture
*)surf
->base
.texture
;
2715 if (!surf
->dcc_incompatible
)
2718 /* Since the DCC decompression calls back into set_framebuffer-
2719 * _state, we need to unbind the framebuffer, so that
2720 * vi_separate_dcc_stop_query isn't called twice with the same
2724 util_copy_framebuffer_state(&sctx
->framebuffer
.state
, NULL
);
2728 if (vi_dcc_enabled(rtex
, surf
->base
.u
.tex
.level
))
2729 if (!si_texture_disable_dcc(&sctx
->b
, rtex
))
2730 sctx
->b
.decompress_dcc(ctx
, rtex
);
2732 surf
->dcc_incompatible
= false;
2735 /* Only flush TC when changing the framebuffer state, because
2736 * the only client not using TC that can change textures is
2739 * Wait for compute shaders because of possible transitions:
2740 * - FB write -> shader read
2741 * - shader write -> FB read
2743 * DB caches are flushed on demand (using si_decompress_textures).
2745 * When MSAA is enabled, CB and TC caches are flushed on demand
2746 * (after FMASK decompression). Shader write -> FB read transitions
2747 * cannot happen for MSAA textures, because MSAA shader images are
2750 * Only flush and wait for CB if there is actually a bound color buffer.
2752 if (sctx
->framebuffer
.nr_samples
<= 1 &&
2753 sctx
->framebuffer
.state
.nr_cbufs
)
2754 si_make_CB_shader_coherent(sctx
, sctx
->framebuffer
.nr_samples
,
2755 sctx
->framebuffer
.CB_has_shader_readable_metadata
);
2757 sctx
->b
.flags
|= SI_CONTEXT_CS_PARTIAL_FLUSH
;
2759 /* u_blitter doesn't invoke depth decompression when it does multiple
2760 * blits in a row, but the only case when it matters for DB is when
2761 * doing generate_mipmap. So here we flush DB manually between
2762 * individual generate_mipmap blits.
2763 * Note that lower mipmap levels aren't compressed.
2765 if (sctx
->generate_mipmap_for_depth
) {
2766 si_make_DB_shader_coherent(sctx
, 1, false,
2767 sctx
->framebuffer
.DB_has_shader_readable_metadata
);
2768 } else if (sctx
->b
.chip_class
== GFX9
) {
2769 /* It appears that DB metadata "leaks" in a sequence of:
2771 * - DCC decompress for shader image writes (with DB disabled)
2772 * - render with DEPTH_BEFORE_SHADER=1
2773 * Flushing DB metadata works around the problem.
2775 sctx
->b
.flags
|= SI_CONTEXT_FLUSH_AND_INV_DB_META
;
2778 /* Take the maximum of the old and new count. If the new count is lower,
2779 * dirtying is needed to disable the unbound colorbuffers.
2781 sctx
->framebuffer
.dirty_cbufs
|=
2782 (1 << MAX2(sctx
->framebuffer
.state
.nr_cbufs
, state
->nr_cbufs
)) - 1;
2783 sctx
->framebuffer
.dirty_zsbuf
|= sctx
->framebuffer
.state
.zsbuf
!= state
->zsbuf
;
2785 si_dec_framebuffer_counters(&sctx
->framebuffer
.state
);
2786 util_copy_framebuffer_state(&sctx
->framebuffer
.state
, state
);
2788 sctx
->framebuffer
.colorbuf_enabled_4bit
= 0;
2789 sctx
->framebuffer
.spi_shader_col_format
= 0;
2790 sctx
->framebuffer
.spi_shader_col_format_alpha
= 0;
2791 sctx
->framebuffer
.spi_shader_col_format_blend
= 0;
2792 sctx
->framebuffer
.spi_shader_col_format_blend_alpha
= 0;
2793 sctx
->framebuffer
.color_is_int8
= 0;
2794 sctx
->framebuffer
.color_is_int10
= 0;
2796 sctx
->framebuffer
.compressed_cb_mask
= 0;
2797 sctx
->framebuffer
.nr_samples
= util_framebuffer_get_num_samples(state
);
2798 sctx
->framebuffer
.log_samples
= util_logbase2(sctx
->framebuffer
.nr_samples
);
2799 sctx
->framebuffer
.any_dst_linear
= false;
2800 sctx
->framebuffer
.CB_has_shader_readable_metadata
= false;
2801 sctx
->framebuffer
.DB_has_shader_readable_metadata
= false;
2803 for (i
= 0; i
< state
->nr_cbufs
; i
++) {
2804 if (!state
->cbufs
[i
])
2807 surf
= (struct r600_surface
*)state
->cbufs
[i
];
2808 rtex
= (struct r600_texture
*)surf
->base
.texture
;
2810 if (!surf
->color_initialized
) {
2811 si_initialize_color_surface(sctx
, surf
);
2814 sctx
->framebuffer
.colorbuf_enabled_4bit
|= 0xf << (i
* 4);
2815 sctx
->framebuffer
.spi_shader_col_format
|=
2816 surf
->spi_shader_col_format
<< (i
* 4);
2817 sctx
->framebuffer
.spi_shader_col_format_alpha
|=
2818 surf
->spi_shader_col_format_alpha
<< (i
* 4);
2819 sctx
->framebuffer
.spi_shader_col_format_blend
|=
2820 surf
->spi_shader_col_format_blend
<< (i
* 4);
2821 sctx
->framebuffer
.spi_shader_col_format_blend_alpha
|=
2822 surf
->spi_shader_col_format_blend_alpha
<< (i
* 4);
2824 if (surf
->color_is_int8
)
2825 sctx
->framebuffer
.color_is_int8
|= 1 << i
;
2826 if (surf
->color_is_int10
)
2827 sctx
->framebuffer
.color_is_int10
|= 1 << i
;
2829 if (rtex
->fmask
.size
) {
2830 sctx
->framebuffer
.compressed_cb_mask
|= 1 << i
;
2833 if (rtex
->surface
.is_linear
)
2834 sctx
->framebuffer
.any_dst_linear
= true;
2836 if (vi_dcc_enabled(rtex
, surf
->base
.u
.tex
.level
))
2837 sctx
->framebuffer
.CB_has_shader_readable_metadata
= true;
2839 r600_context_add_resource_size(ctx
, surf
->base
.texture
);
2841 p_atomic_inc(&rtex
->framebuffers_bound
);
2843 if (rtex
->dcc_gather_statistics
) {
2844 /* Dirty tracking must be enabled for DCC usage analysis. */
2845 sctx
->framebuffer
.compressed_cb_mask
|= 1 << i
;
2846 vi_separate_dcc_start_query(ctx
, rtex
);
2850 struct r600_texture
*zstex
= NULL
;
2853 surf
= (struct r600_surface
*)state
->zsbuf
;
2854 zstex
= (struct r600_texture
*)surf
->base
.texture
;
2856 if (!surf
->depth_initialized
) {
2857 si_init_depth_surface(sctx
, surf
);
2860 if (vi_tc_compat_htile_enabled(zstex
, surf
->base
.u
.tex
.level
))
2861 sctx
->framebuffer
.DB_has_shader_readable_metadata
= true;
2863 r600_context_add_resource_size(ctx
, surf
->base
.texture
);
2866 si_update_poly_offset_state(sctx
);
2867 si_mark_atom_dirty(sctx
, &sctx
->cb_render_state
);
2868 si_mark_atom_dirty(sctx
, &sctx
->framebuffer
.atom
);
2870 if (sctx
->screen
->dpbb_allowed
)
2871 si_mark_atom_dirty(sctx
, &sctx
->dpbb_state
);
2873 if (sctx
->framebuffer
.any_dst_linear
!= old_any_dst_linear
)
2874 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
2876 if (sctx
->screen
->has_out_of_order_rast
&&
2877 (sctx
->framebuffer
.colorbuf_enabled_4bit
!= old_colorbuf_enabled_4bit
||
2878 !!sctx
->framebuffer
.state
.zsbuf
!= old_has_zsbuf
||
2879 (zstex
&& zstex
->surface
.has_stencil
!= old_has_stencil
)))
2880 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
2882 if (sctx
->framebuffer
.nr_samples
!= old_nr_samples
) {
2883 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
2884 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
2886 /* Set sample locations as fragment shader constants. */
2887 switch (sctx
->framebuffer
.nr_samples
) {
2889 constbuf
.user_buffer
= sctx
->b
.sample_locations_1x
;
2892 constbuf
.user_buffer
= sctx
->b
.sample_locations_2x
;
2895 constbuf
.user_buffer
= sctx
->b
.sample_locations_4x
;
2898 constbuf
.user_buffer
= sctx
->b
.sample_locations_8x
;
2901 constbuf
.user_buffer
= sctx
->b
.sample_locations_16x
;
2904 R600_ERR("Requested an invalid number of samples %i.\n",
2905 sctx
->framebuffer
.nr_samples
);
2908 constbuf
.buffer_size
= sctx
->framebuffer
.nr_samples
* 2 * 4;
2909 si_set_rw_buffer(sctx
, SI_PS_CONST_SAMPLE_POSITIONS
, &constbuf
);
2911 si_mark_atom_dirty(sctx
, &sctx
->msaa_sample_locs
.atom
);
2914 sctx
->do_update_shaders
= true;
2916 if (!sctx
->decompression_enabled
) {
2917 /* Prevent textures decompression when the framebuffer state
2918 * changes come from the decompression passes themselves.
2920 sctx
->need_check_render_feedback
= true;
2924 static void si_emit_framebuffer_state(struct si_context
*sctx
, struct r600_atom
*atom
)
2926 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
2927 struct pipe_framebuffer_state
*state
= &sctx
->framebuffer
.state
;
2928 unsigned i
, nr_cbufs
= state
->nr_cbufs
;
2929 struct r600_texture
*tex
= NULL
;
2930 struct r600_surface
*cb
= NULL
;
2931 unsigned cb_color_info
= 0;
2934 for (i
= 0; i
< nr_cbufs
; i
++) {
2935 uint64_t cb_color_base
, cb_color_fmask
, cb_dcc_base
;
2936 unsigned cb_color_attrib
;
2938 if (!(sctx
->framebuffer
.dirty_cbufs
& (1 << i
)))
2941 cb
= (struct r600_surface
*)state
->cbufs
[i
];
2943 radeon_set_context_reg(cs
, R_028C70_CB_COLOR0_INFO
+ i
* 0x3C,
2944 S_028C70_FORMAT(V_028C70_COLOR_INVALID
));
2948 tex
= (struct r600_texture
*)cb
->base
.texture
;
2949 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2950 &tex
->resource
, RADEON_USAGE_READWRITE
,
2951 tex
->resource
.b
.b
.nr_samples
> 1 ?
2952 RADEON_PRIO_COLOR_BUFFER_MSAA
:
2953 RADEON_PRIO_COLOR_BUFFER
);
2955 if (tex
->cmask_buffer
&& tex
->cmask_buffer
!= &tex
->resource
) {
2956 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2957 tex
->cmask_buffer
, RADEON_USAGE_READWRITE
,
2961 if (tex
->dcc_separate_buffer
)
2962 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2963 tex
->dcc_separate_buffer
,
2964 RADEON_USAGE_READWRITE
,
2967 /* Compute mutable surface parameters. */
2968 cb_color_base
= tex
->resource
.gpu_address
>> 8;
2971 cb_color_info
= cb
->cb_color_info
| tex
->cb_color_info
;
2972 cb_color_attrib
= cb
->cb_color_attrib
;
2974 if (tex
->fmask
.size
) {
2975 cb_color_fmask
= (tex
->resource
.gpu_address
+ tex
->fmask
.offset
) >> 8;
2976 cb_color_fmask
|= tex
->fmask
.tile_swizzle
;
2980 if (vi_dcc_enabled(tex
, cb
->base
.u
.tex
.level
)) {
2981 bool is_msaa_resolve_dst
= state
->cbufs
[0] &&
2982 state
->cbufs
[0]->texture
->nr_samples
> 1 &&
2983 state
->cbufs
[1] == &cb
->base
&&
2984 state
->cbufs
[1]->texture
->nr_samples
<= 1;
2986 if (!is_msaa_resolve_dst
)
2987 cb_color_info
|= S_028C70_DCC_ENABLE(1);
2989 cb_dcc_base
= ((!tex
->dcc_separate_buffer
? tex
->resource
.gpu_address
: 0) +
2990 tex
->dcc_offset
) >> 8;
2991 cb_dcc_base
|= tex
->surface
.tile_swizzle
;
2994 if (sctx
->b
.chip_class
>= GFX9
) {
2995 struct gfx9_surf_meta_flags meta
;
2997 if (tex
->dcc_offset
)
2998 meta
= tex
->surface
.u
.gfx9
.dcc
;
3000 meta
= tex
->surface
.u
.gfx9
.cmask
;
3002 /* Set mutable surface parameters. */
3003 cb_color_base
+= tex
->surface
.u
.gfx9
.surf_offset
>> 8;
3004 cb_color_base
|= tex
->surface
.tile_swizzle
;
3005 if (!tex
->fmask
.size
)
3006 cb_color_fmask
= cb_color_base
;
3007 cb_color_attrib
|= S_028C74_COLOR_SW_MODE(tex
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
3008 S_028C74_FMASK_SW_MODE(tex
->surface
.u
.gfx9
.fmask
.swizzle_mode
) |
3009 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
3010 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
3012 radeon_set_context_reg_seq(cs
, R_028C60_CB_COLOR0_BASE
+ i
* 0x3C, 15);
3013 radeon_emit(cs
, cb_color_base
); /* CB_COLOR0_BASE */
3014 radeon_emit(cs
, cb_color_base
>> 32); /* CB_COLOR0_BASE_EXT */
3015 radeon_emit(cs
, cb
->cb_color_attrib2
); /* CB_COLOR0_ATTRIB2 */
3016 radeon_emit(cs
, cb
->cb_color_view
); /* CB_COLOR0_VIEW */
3017 radeon_emit(cs
, cb_color_info
); /* CB_COLOR0_INFO */
3018 radeon_emit(cs
, cb_color_attrib
); /* CB_COLOR0_ATTRIB */
3019 radeon_emit(cs
, cb
->cb_dcc_control
); /* CB_COLOR0_DCC_CONTROL */
3020 radeon_emit(cs
, tex
->cmask
.base_address_reg
); /* CB_COLOR0_CMASK */
3021 radeon_emit(cs
, tex
->cmask
.base_address_reg
>> 32); /* CB_COLOR0_CMASK_BASE_EXT */
3022 radeon_emit(cs
, cb_color_fmask
); /* CB_COLOR0_FMASK */
3023 radeon_emit(cs
, cb_color_fmask
>> 32); /* CB_COLOR0_FMASK_BASE_EXT */
3024 radeon_emit(cs
, tex
->color_clear_value
[0]); /* CB_COLOR0_CLEAR_WORD0 */
3025 radeon_emit(cs
, tex
->color_clear_value
[1]); /* CB_COLOR0_CLEAR_WORD1 */
3026 radeon_emit(cs
, cb_dcc_base
); /* CB_COLOR0_DCC_BASE */
3027 radeon_emit(cs
, cb_dcc_base
>> 32); /* CB_COLOR0_DCC_BASE_EXT */
3029 radeon_set_context_reg(cs
, R_0287A0_CB_MRT0_EPITCH
+ i
* 4,
3030 S_0287A0_EPITCH(tex
->surface
.u
.gfx9
.surf
.epitch
));
3032 /* Compute mutable surface parameters (SI-CI-VI). */
3033 const struct legacy_surf_level
*level_info
=
3034 &tex
->surface
.u
.legacy
.level
[cb
->base
.u
.tex
.level
];
3035 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
3036 unsigned cb_color_pitch
, cb_color_slice
, cb_color_fmask_slice
;
3038 cb_color_base
+= level_info
->offset
>> 8;
3039 /* Only macrotiled modes can set tile swizzle. */
3040 if (level_info
->mode
== RADEON_SURF_MODE_2D
)
3041 cb_color_base
|= tex
->surface
.tile_swizzle
;
3043 if (!tex
->fmask
.size
)
3044 cb_color_fmask
= cb_color_base
;
3046 cb_dcc_base
+= level_info
->dcc_offset
>> 8;
3048 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
3049 slice_tile_max
= level_info
->nblk_x
*
3050 level_info
->nblk_y
/ 64 - 1;
3051 tile_mode_index
= si_tile_mode_index(tex
, cb
->base
.u
.tex
.level
, false);
3053 cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
3054 cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
3055 cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
3057 if (tex
->fmask
.size
) {
3058 if (sctx
->b
.chip_class
>= CIK
)
3059 cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(tex
->fmask
.pitch_in_pixels
/ 8 - 1);
3060 cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tex
->fmask
.tile_mode_index
);
3061 cb_color_fmask_slice
= S_028C88_TILE_MAX(tex
->fmask
.slice_tile_max
);
3063 /* This must be set for fast clear to work without FMASK. */
3064 if (sctx
->b
.chip_class
>= CIK
)
3065 cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
3066 cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
3067 cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
3070 radeon_set_context_reg_seq(cs
, R_028C60_CB_COLOR0_BASE
+ i
* 0x3C,
3071 sctx
->b
.chip_class
>= VI
? 14 : 13);
3072 radeon_emit(cs
, cb_color_base
); /* CB_COLOR0_BASE */
3073 radeon_emit(cs
, cb_color_pitch
); /* CB_COLOR0_PITCH */
3074 radeon_emit(cs
, cb_color_slice
); /* CB_COLOR0_SLICE */
3075 radeon_emit(cs
, cb
->cb_color_view
); /* CB_COLOR0_VIEW */
3076 radeon_emit(cs
, cb_color_info
); /* CB_COLOR0_INFO */
3077 radeon_emit(cs
, cb_color_attrib
); /* CB_COLOR0_ATTRIB */
3078 radeon_emit(cs
, cb
->cb_dcc_control
); /* CB_COLOR0_DCC_CONTROL */
3079 radeon_emit(cs
, tex
->cmask
.base_address_reg
); /* CB_COLOR0_CMASK */
3080 radeon_emit(cs
, tex
->cmask
.slice_tile_max
); /* CB_COLOR0_CMASK_SLICE */
3081 radeon_emit(cs
, cb_color_fmask
); /* CB_COLOR0_FMASK */
3082 radeon_emit(cs
, cb_color_fmask_slice
); /* CB_COLOR0_FMASK_SLICE */
3083 radeon_emit(cs
, tex
->color_clear_value
[0]); /* CB_COLOR0_CLEAR_WORD0 */
3084 radeon_emit(cs
, tex
->color_clear_value
[1]); /* CB_COLOR0_CLEAR_WORD1 */
3086 if (sctx
->b
.chip_class
>= VI
) /* R_028C94_CB_COLOR0_DCC_BASE */
3087 radeon_emit(cs
, cb_dcc_base
);
3091 if (sctx
->framebuffer
.dirty_cbufs
& (1 << i
))
3092 radeon_set_context_reg(cs
, R_028C70_CB_COLOR0_INFO
+ i
* 0x3C, 0);
3095 if (state
->zsbuf
&& sctx
->framebuffer
.dirty_zsbuf
) {
3096 struct r600_surface
*zb
= (struct r600_surface
*)state
->zsbuf
;
3097 struct r600_texture
*rtex
= (struct r600_texture
*)zb
->base
.texture
;
3099 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
3100 &rtex
->resource
, RADEON_USAGE_READWRITE
,
3101 zb
->base
.texture
->nr_samples
> 1 ?
3102 RADEON_PRIO_DEPTH_BUFFER_MSAA
:
3103 RADEON_PRIO_DEPTH_BUFFER
);
3105 if (sctx
->b
.chip_class
>= GFX9
) {
3106 radeon_set_context_reg_seq(cs
, R_028014_DB_HTILE_DATA_BASE
, 3);
3107 radeon_emit(cs
, zb
->db_htile_data_base
); /* DB_HTILE_DATA_BASE */
3108 radeon_emit(cs
, zb
->db_htile_data_base
>> 32); /* DB_HTILE_DATA_BASE_HI */
3109 radeon_emit(cs
, zb
->db_depth_size
); /* DB_DEPTH_SIZE */
3111 radeon_set_context_reg_seq(cs
, R_028038_DB_Z_INFO
, 10);
3112 radeon_emit(cs
, zb
->db_z_info
| /* DB_Z_INFO */
3113 S_028038_ZRANGE_PRECISION(rtex
->depth_clear_value
!= 0));
3114 radeon_emit(cs
, zb
->db_stencil_info
); /* DB_STENCIL_INFO */
3115 radeon_emit(cs
, zb
->db_depth_base
); /* DB_Z_READ_BASE */
3116 radeon_emit(cs
, zb
->db_depth_base
>> 32); /* DB_Z_READ_BASE_HI */
3117 radeon_emit(cs
, zb
->db_stencil_base
); /* DB_STENCIL_READ_BASE */
3118 radeon_emit(cs
, zb
->db_stencil_base
>> 32); /* DB_STENCIL_READ_BASE_HI */
3119 radeon_emit(cs
, zb
->db_depth_base
); /* DB_Z_WRITE_BASE */
3120 radeon_emit(cs
, zb
->db_depth_base
>> 32); /* DB_Z_WRITE_BASE_HI */
3121 radeon_emit(cs
, zb
->db_stencil_base
); /* DB_STENCIL_WRITE_BASE */
3122 radeon_emit(cs
, zb
->db_stencil_base
>> 32); /* DB_STENCIL_WRITE_BASE_HI */
3124 radeon_set_context_reg_seq(cs
, R_028068_DB_Z_INFO2
, 2);
3125 radeon_emit(cs
, zb
->db_z_info2
); /* DB_Z_INFO2 */
3126 radeon_emit(cs
, zb
->db_stencil_info2
); /* DB_STENCIL_INFO2 */
3128 radeon_set_context_reg(cs
, R_028014_DB_HTILE_DATA_BASE
, zb
->db_htile_data_base
);
3130 radeon_set_context_reg_seq(cs
, R_02803C_DB_DEPTH_INFO
, 9);
3131 radeon_emit(cs
, zb
->db_depth_info
); /* DB_DEPTH_INFO */
3132 radeon_emit(cs
, zb
->db_z_info
| /* DB_Z_INFO */
3133 S_028040_ZRANGE_PRECISION(rtex
->depth_clear_value
!= 0));
3134 radeon_emit(cs
, zb
->db_stencil_info
); /* DB_STENCIL_INFO */
3135 radeon_emit(cs
, zb
->db_depth_base
); /* DB_Z_READ_BASE */
3136 radeon_emit(cs
, zb
->db_stencil_base
); /* DB_STENCIL_READ_BASE */
3137 radeon_emit(cs
, zb
->db_depth_base
); /* DB_Z_WRITE_BASE */
3138 radeon_emit(cs
, zb
->db_stencil_base
); /* DB_STENCIL_WRITE_BASE */
3139 radeon_emit(cs
, zb
->db_depth_size
); /* DB_DEPTH_SIZE */
3140 radeon_emit(cs
, zb
->db_depth_slice
); /* DB_DEPTH_SLICE */
3143 radeon_set_context_reg_seq(cs
, R_028028_DB_STENCIL_CLEAR
, 2);
3144 radeon_emit(cs
, rtex
->stencil_clear_value
); /* R_028028_DB_STENCIL_CLEAR */
3145 radeon_emit(cs
, fui(rtex
->depth_clear_value
)); /* R_02802C_DB_DEPTH_CLEAR */
3147 radeon_set_context_reg(cs
, R_028008_DB_DEPTH_VIEW
, zb
->db_depth_view
);
3148 radeon_set_context_reg(cs
, R_028ABC_DB_HTILE_SURFACE
, zb
->db_htile_surface
);
3149 } else if (sctx
->framebuffer
.dirty_zsbuf
) {
3150 if (sctx
->b
.chip_class
>= GFX9
)
3151 radeon_set_context_reg_seq(cs
, R_028038_DB_Z_INFO
, 2);
3153 radeon_set_context_reg_seq(cs
, R_028040_DB_Z_INFO
, 2);
3155 radeon_emit(cs
, S_028040_FORMAT(V_028040_Z_INVALID
)); /* DB_Z_INFO */
3156 radeon_emit(cs
, S_028044_FORMAT(V_028044_STENCIL_INVALID
)); /* DB_STENCIL_INFO */
3159 /* Framebuffer dimensions. */
3160 /* PA_SC_WINDOW_SCISSOR_TL is set in si_init_config() */
3161 radeon_set_context_reg(cs
, R_028208_PA_SC_WINDOW_SCISSOR_BR
,
3162 S_028208_BR_X(state
->width
) | S_028208_BR_Y(state
->height
));
3164 if (sctx
->screen
->dfsm_allowed
) {
3165 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
3166 radeon_emit(cs
, EVENT_TYPE(V_028A90_BREAK_BATCH
) | EVENT_INDEX(0));
3169 sctx
->framebuffer
.dirty_cbufs
= 0;
3170 sctx
->framebuffer
.dirty_zsbuf
= false;
3173 static void si_emit_msaa_sample_locs(struct si_context
*sctx
,
3174 struct r600_atom
*atom
)
3176 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
3177 unsigned nr_samples
= sctx
->framebuffer
.nr_samples
;
3178 bool has_msaa_sample_loc_bug
= sctx
->screen
->has_msaa_sample_loc_bug
;
3180 /* Smoothing (only possible with nr_samples == 1) uses the same
3181 * sample locations as the MSAA it simulates.
3183 if (nr_samples
<= 1 && sctx
->smoothing_enabled
)
3184 nr_samples
= SI_NUM_SMOOTH_AA_SAMPLES
;
3186 /* On Polaris, the small primitive filter uses the sample locations
3187 * even when MSAA is off, so we need to make sure they're set to 0.
3189 if (has_msaa_sample_loc_bug
)
3190 nr_samples
= MAX2(nr_samples
, 1);
3192 if (nr_samples
!= sctx
->msaa_sample_locs
.nr_samples
) {
3193 sctx
->msaa_sample_locs
.nr_samples
= nr_samples
;
3194 si_common_emit_msaa_sample_locs(cs
, nr_samples
);
3197 if (sctx
->b
.family
>= CHIP_POLARIS10
) {
3198 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
3199 unsigned small_prim_filter_cntl
=
3200 S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
3202 S_028830_LINE_FILTER_DISABLE(sctx
->b
.family
<= CHIP_POLARIS12
);
3204 /* The alternative of setting sample locations to 0 would
3205 * require a DB flush to avoid Z errors, see
3206 * https://bugs.freedesktop.org/show_bug.cgi?id=96908
3208 if (has_msaa_sample_loc_bug
&&
3209 sctx
->framebuffer
.nr_samples
> 1 &&
3210 rs
&& !rs
->multisample_enable
)
3211 small_prim_filter_cntl
&= C_028830_SMALL_PRIM_FILTER_ENABLE
;
3213 radeon_set_context_reg(cs
, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL
,
3214 small_prim_filter_cntl
);
3218 static bool si_out_of_order_rasterization(struct si_context
*sctx
)
3220 struct si_state_blend
*blend
= sctx
->queued
.named
.blend
;
3221 struct si_state_dsa
*dsa
= sctx
->queued
.named
.dsa
;
3223 if (!sctx
->screen
->has_out_of_order_rast
)
3226 unsigned colormask
= sctx
->framebuffer
.colorbuf_enabled_4bit
;
3229 colormask
&= blend
->cb_target_enabled_4bit
;
3234 /* Conservative: No logic op. */
3235 if (colormask
&& blend
->logicop_enable
)
3238 struct si_dsa_order_invariance dsa_order_invariant
= {
3239 .zs
= true, .pass_set
= true, .pass_last
= false
3242 if (sctx
->framebuffer
.state
.zsbuf
) {
3243 struct r600_texture
*zstex
=
3244 (struct r600_texture
*)sctx
->framebuffer
.state
.zsbuf
->texture
;
3245 bool has_stencil
= zstex
->surface
.has_stencil
;
3246 dsa_order_invariant
= dsa
->order_invariance
[has_stencil
];
3247 if (!dsa_order_invariant
.zs
)
3250 /* The set of PS invocations is always order invariant,
3251 * except when early Z/S tests are requested. */
3252 if (sctx
->ps_shader
.cso
&&
3253 sctx
->ps_shader
.cso
->info
.writes_memory
&&
3254 sctx
->ps_shader
.cso
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
] &&
3255 !dsa_order_invariant
.pass_set
)
3258 if (sctx
->b
.num_perfect_occlusion_queries
!= 0 &&
3259 !dsa_order_invariant
.pass_set
)
3266 unsigned blendmask
= colormask
& blend
->blend_enable_4bit
;
3269 /* Only commutative blending. */
3270 if (blendmask
& ~blend
->commutative_4bit
)
3273 if (!dsa_order_invariant
.pass_set
)
3277 if (colormask
& ~blendmask
) {
3278 if (!dsa_order_invariant
.pass_last
)
3285 static void si_emit_msaa_config(struct si_context
*sctx
, struct r600_atom
*atom
)
3287 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
3288 unsigned num_tile_pipes
= sctx
->screen
->b
.info
.num_tile_pipes
;
3289 /* 33% faster rendering to linear color buffers */
3290 bool dst_is_linear
= sctx
->framebuffer
.any_dst_linear
;
3291 bool out_of_order_rast
= si_out_of_order_rasterization(sctx
);
3292 unsigned sc_mode_cntl_1
=
3293 S_028A4C_WALK_SIZE(dst_is_linear
) |
3294 S_028A4C_WALK_FENCE_ENABLE(!dst_is_linear
) |
3295 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes
== 2 ? 2 : 3) |
3296 S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(out_of_order_rast
) |
3297 S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7) |
3299 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
3300 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
3301 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
3302 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
3303 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
3304 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
3306 si_common_emit_msaa_config(cs
, sctx
->framebuffer
.nr_samples
,
3307 sctx
->ps_iter_samples
,
3308 sctx
->smoothing_enabled
? SI_NUM_SMOOTH_AA_SAMPLES
: 0,
3311 /* GFX9: Flush DFSM when the AA mode changes. */
3312 if (sctx
->screen
->dfsm_allowed
) {
3313 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
3314 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_DFSM
) | EVENT_INDEX(0));
3318 static void si_set_min_samples(struct pipe_context
*ctx
, unsigned min_samples
)
3320 struct si_context
*sctx
= (struct si_context
*)ctx
;
3322 if (sctx
->ps_iter_samples
== min_samples
)
3325 sctx
->ps_iter_samples
= min_samples
;
3326 sctx
->do_update_shaders
= true;
3328 if (sctx
->framebuffer
.nr_samples
> 1)
3329 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
3330 if (sctx
->screen
->dpbb_allowed
)
3331 si_mark_atom_dirty(sctx
, &sctx
->dpbb_state
);
3339 * Build the sampler view descriptor for a buffer texture.
3340 * @param state 256-bit descriptor; only the high 128 bits are filled in
3343 si_make_buffer_descriptor(struct si_screen
*screen
, struct r600_resource
*buf
,
3344 enum pipe_format format
,
3345 unsigned offset
, unsigned size
,
3348 const struct util_format_description
*desc
;
3351 unsigned num_records
;
3352 unsigned num_format
, data_format
;
3354 desc
= util_format_description(format
);
3355 first_non_void
= util_format_get_first_non_void_channel(format
);
3356 stride
= desc
->block
.bits
/ 8;
3357 num_format
= si_translate_buffer_numformat(&screen
->b
.b
, desc
, first_non_void
);
3358 data_format
= si_translate_buffer_dataformat(&screen
->b
.b
, desc
, first_non_void
);
3360 num_records
= size
/ stride
;
3361 num_records
= MIN2(num_records
, (buf
->b
.b
.width0
- offset
) / stride
);
3363 /* The NUM_RECORDS field has a different meaning depending on the chip,
3364 * instruction type, STRIDE, and SWIZZLE_ENABLE.
3367 * - If STRIDE == 0, it's in byte units.
3368 * - If STRIDE != 0, it's in units of STRIDE, used with inst.IDXEN.
3371 * - For SMEM and STRIDE == 0, it's in byte units.
3372 * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3373 * - For VMEM and STRIDE == 0 or SWIZZLE_ENABLE == 0, it's in byte units.
3374 * - For VMEM and STRIDE != 0 and SWIZZLE_ENABLE == 1, it's in units of STRIDE.
3375 * NOTE: There is incompatibility between VMEM and SMEM opcodes due to SWIZZLE_-
3376 * ENABLE. The workaround is to set STRIDE = 0 if SWIZZLE_ENABLE == 0 when
3377 * using SMEM. This can be done in the shader by clearing STRIDE with s_and.
3378 * That way the same descriptor can be used by both SMEM and VMEM.
3381 * - For SMEM and STRIDE == 0, it's in byte units.
3382 * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3383 * - For VMEM and inst.IDXEN == 0 or STRIDE == 0, it's in byte units.
3384 * - For VMEM and inst.IDXEN == 1 and STRIDE != 0, it's in units of STRIDE.
3386 if (screen
->b
.chip_class
>= GFX9
)
3387 /* When vindex == 0, LLVM sets IDXEN = 0, thus changing units
3388 * from STRIDE to bytes. This works around it by setting
3389 * NUM_RECORDS to at least the size of one element, so that
3390 * the first element is readable when IDXEN == 0.
3392 * TODO: Fix this in LLVM, but do we need a new intrinsic where
3393 * IDXEN is enforced?
3395 num_records
= num_records
? MAX2(num_records
, stride
) : 0;
3396 else if (screen
->b
.chip_class
== VI
)
3397 num_records
*= stride
;
3400 state
[5] = S_008F04_STRIDE(stride
);
3401 state
[6] = num_records
;
3402 state
[7] = S_008F0C_DST_SEL_X(si_map_swizzle(desc
->swizzle
[0])) |
3403 S_008F0C_DST_SEL_Y(si_map_swizzle(desc
->swizzle
[1])) |
3404 S_008F0C_DST_SEL_Z(si_map_swizzle(desc
->swizzle
[2])) |
3405 S_008F0C_DST_SEL_W(si_map_swizzle(desc
->swizzle
[3])) |
3406 S_008F0C_NUM_FORMAT(num_format
) |
3407 S_008F0C_DATA_FORMAT(data_format
);
3410 static unsigned gfx9_border_color_swizzle(const unsigned char swizzle
[4])
3412 unsigned bc_swizzle
= V_008F20_BC_SWIZZLE_XYZW
;
3414 if (swizzle
[3] == PIPE_SWIZZLE_X
) {
3415 /* For the pre-defined border color values (white, opaque
3416 * black, transparent black), the only thing that matters is
3417 * that the alpha channel winds up in the correct place
3418 * (because the RGB channels are all the same) so either of
3419 * these enumerations will work.
3421 if (swizzle
[2] == PIPE_SWIZZLE_Y
)
3422 bc_swizzle
= V_008F20_BC_SWIZZLE_WZYX
;
3424 bc_swizzle
= V_008F20_BC_SWIZZLE_WXYZ
;
3425 } else if (swizzle
[0] == PIPE_SWIZZLE_X
) {
3426 if (swizzle
[1] == PIPE_SWIZZLE_Y
)
3427 bc_swizzle
= V_008F20_BC_SWIZZLE_XYZW
;
3429 bc_swizzle
= V_008F20_BC_SWIZZLE_XWYZ
;
3430 } else if (swizzle
[1] == PIPE_SWIZZLE_X
) {
3431 bc_swizzle
= V_008F20_BC_SWIZZLE_YXWZ
;
3432 } else if (swizzle
[2] == PIPE_SWIZZLE_X
) {
3433 bc_swizzle
= V_008F20_BC_SWIZZLE_ZYXW
;
3440 * Build the sampler view descriptor for a texture.
3443 si_make_texture_descriptor(struct si_screen
*screen
,
3444 struct r600_texture
*tex
,
3446 enum pipe_texture_target target
,
3447 enum pipe_format pipe_format
,
3448 const unsigned char state_swizzle
[4],
3449 unsigned first_level
, unsigned last_level
,
3450 unsigned first_layer
, unsigned last_layer
,
3451 unsigned width
, unsigned height
, unsigned depth
,
3453 uint32_t *fmask_state
)
3455 struct pipe_resource
*res
= &tex
->resource
.b
.b
;
3456 const struct util_format_description
*desc
;
3457 unsigned char swizzle
[4];
3459 unsigned num_format
, data_format
, type
;
3462 desc
= util_format_description(pipe_format
);
3464 if (desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
) {
3465 const unsigned char swizzle_xxxx
[4] = {0, 0, 0, 0};
3466 const unsigned char swizzle_yyyy
[4] = {1, 1, 1, 1};
3467 const unsigned char swizzle_wwww
[4] = {3, 3, 3, 3};
3469 switch (pipe_format
) {
3470 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
3471 case PIPE_FORMAT_X32_S8X24_UINT
:
3472 case PIPE_FORMAT_X8Z24_UNORM
:
3473 util_format_compose_swizzles(swizzle_yyyy
, state_swizzle
, swizzle
);
3475 case PIPE_FORMAT_X24S8_UINT
:
3477 * X24S8 is implemented as an 8_8_8_8 data format, to
3478 * fix texture gathers. This affects at least
3479 * GL45-CTS.texture_cube_map_array.sampling on VI.
3481 util_format_compose_swizzles(swizzle_wwww
, state_swizzle
, swizzle
);
3484 util_format_compose_swizzles(swizzle_xxxx
, state_swizzle
, swizzle
);
3487 util_format_compose_swizzles(desc
->swizzle
, state_swizzle
, swizzle
);
3490 first_non_void
= util_format_get_first_non_void_channel(pipe_format
);
3492 switch (pipe_format
) {
3493 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
3494 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3497 if (first_non_void
< 0) {
3498 if (util_format_is_compressed(pipe_format
)) {
3499 switch (pipe_format
) {
3500 case PIPE_FORMAT_DXT1_SRGB
:
3501 case PIPE_FORMAT_DXT1_SRGBA
:
3502 case PIPE_FORMAT_DXT3_SRGBA
:
3503 case PIPE_FORMAT_DXT5_SRGBA
:
3504 case PIPE_FORMAT_BPTC_SRGBA
:
3505 case PIPE_FORMAT_ETC2_SRGB8
:
3506 case PIPE_FORMAT_ETC2_SRGB8A1
:
3507 case PIPE_FORMAT_ETC2_SRGBA8
:
3508 num_format
= V_008F14_IMG_NUM_FORMAT_SRGB
;
3510 case PIPE_FORMAT_RGTC1_SNORM
:
3511 case PIPE_FORMAT_LATC1_SNORM
:
3512 case PIPE_FORMAT_RGTC2_SNORM
:
3513 case PIPE_FORMAT_LATC2_SNORM
:
3514 case PIPE_FORMAT_ETC2_R11_SNORM
:
3515 case PIPE_FORMAT_ETC2_RG11_SNORM
:
3516 /* implies float, so use SNORM/UNORM to determine
3517 whether data is signed or not */
3518 case PIPE_FORMAT_BPTC_RGB_FLOAT
:
3519 num_format
= V_008F14_IMG_NUM_FORMAT_SNORM
;
3522 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3525 } else if (desc
->layout
== UTIL_FORMAT_LAYOUT_SUBSAMPLED
) {
3526 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3528 num_format
= V_008F14_IMG_NUM_FORMAT_FLOAT
;
3530 } else if (desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
3531 num_format
= V_008F14_IMG_NUM_FORMAT_SRGB
;
3533 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3535 switch (desc
->channel
[first_non_void
].type
) {
3536 case UTIL_FORMAT_TYPE_FLOAT
:
3537 num_format
= V_008F14_IMG_NUM_FORMAT_FLOAT
;
3539 case UTIL_FORMAT_TYPE_SIGNED
:
3540 if (desc
->channel
[first_non_void
].normalized
)
3541 num_format
= V_008F14_IMG_NUM_FORMAT_SNORM
;
3542 else if (desc
->channel
[first_non_void
].pure_integer
)
3543 num_format
= V_008F14_IMG_NUM_FORMAT_SINT
;
3545 num_format
= V_008F14_IMG_NUM_FORMAT_SSCALED
;
3547 case UTIL_FORMAT_TYPE_UNSIGNED
:
3548 if (desc
->channel
[first_non_void
].normalized
)
3549 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3550 else if (desc
->channel
[first_non_void
].pure_integer
)
3551 num_format
= V_008F14_IMG_NUM_FORMAT_UINT
;
3553 num_format
= V_008F14_IMG_NUM_FORMAT_USCALED
;
3558 data_format
= si_translate_texformat(&screen
->b
.b
, pipe_format
, desc
, first_non_void
);
3559 if (data_format
== ~0) {
3563 /* S8 with Z32 HTILE needs a special format. */
3564 if (screen
->b
.chip_class
>= GFX9
&&
3565 pipe_format
== PIPE_FORMAT_S8_UINT
&&
3566 tex
->tc_compatible_htile
)
3567 data_format
= V_008F14_IMG_DATA_FORMAT_S8_32
;
3570 (res
->target
== PIPE_TEXTURE_CUBE
||
3571 res
->target
== PIPE_TEXTURE_CUBE_ARRAY
||
3572 (screen
->b
.chip_class
<= VI
&&
3573 res
->target
== PIPE_TEXTURE_3D
))) {
3574 /* For the purpose of shader images, treat cube maps and 3D
3575 * textures as 2D arrays. For 3D textures, the address
3576 * calculations for mipmaps are different, so we rely on the
3577 * caller to effectively disable mipmaps.
3579 type
= V_008F1C_SQ_RSRC_IMG_2D_ARRAY
;
3581 assert(res
->target
!= PIPE_TEXTURE_3D
|| (first_level
== 0 && last_level
== 0));
3583 type
= si_tex_dim(screen
, tex
, target
, res
->nr_samples
);
3586 if (type
== V_008F1C_SQ_RSRC_IMG_1D_ARRAY
) {
3588 depth
= res
->array_size
;
3589 } else if (type
== V_008F1C_SQ_RSRC_IMG_2D_ARRAY
||
3590 type
== V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
) {
3591 if (sampler
|| res
->target
!= PIPE_TEXTURE_3D
)
3592 depth
= res
->array_size
;
3593 } else if (type
== V_008F1C_SQ_RSRC_IMG_CUBE
)
3594 depth
= res
->array_size
/ 6;
3597 state
[1] = (S_008F14_DATA_FORMAT_GFX6(data_format
) |
3598 S_008F14_NUM_FORMAT_GFX6(num_format
));
3599 state
[2] = (S_008F18_WIDTH(width
- 1) |
3600 S_008F18_HEIGHT(height
- 1) |
3601 S_008F18_PERF_MOD(4));
3602 state
[3] = (S_008F1C_DST_SEL_X(si_map_swizzle(swizzle
[0])) |
3603 S_008F1C_DST_SEL_Y(si_map_swizzle(swizzle
[1])) |
3604 S_008F1C_DST_SEL_Z(si_map_swizzle(swizzle
[2])) |
3605 S_008F1C_DST_SEL_W(si_map_swizzle(swizzle
[3])) |
3606 S_008F1C_BASE_LEVEL(res
->nr_samples
> 1 ?
3608 S_008F1C_LAST_LEVEL(res
->nr_samples
> 1 ?
3609 util_logbase2(res
->nr_samples
) :
3611 S_008F1C_TYPE(type
));
3613 state
[5] = S_008F24_BASE_ARRAY(first_layer
);
3617 if (screen
->b
.chip_class
>= GFX9
) {
3618 unsigned bc_swizzle
= gfx9_border_color_swizzle(desc
->swizzle
);
3620 /* Depth is the the last accessible layer on Gfx9.
3621 * The hw doesn't need to know the total number of layers.
3623 if (type
== V_008F1C_SQ_RSRC_IMG_3D
)
3624 state
[4] |= S_008F20_DEPTH(depth
- 1);
3626 state
[4] |= S_008F20_DEPTH(last_layer
);
3628 state
[4] |= S_008F20_BC_SWIZZLE(bc_swizzle
);
3629 state
[5] |= S_008F24_MAX_MIP(res
->nr_samples
> 1 ?
3630 util_logbase2(res
->nr_samples
) :
3631 tex
->resource
.b
.b
.last_level
);
3633 state
[3] |= S_008F1C_POW2_PAD(res
->last_level
> 0);
3634 state
[4] |= S_008F20_DEPTH(depth
- 1);
3635 state
[5] |= S_008F24_LAST_ARRAY(last_layer
);
3638 if (tex
->dcc_offset
) {
3639 unsigned swap
= si_translate_colorswap(pipe_format
, false);
3641 state
[6] = S_008F28_ALPHA_IS_ON_MSB(swap
<= 1);
3643 /* The last dword is unused by hw. The shader uses it to clear
3644 * bits in the first dword of sampler state.
3646 if (screen
->b
.chip_class
<= CIK
&& res
->nr_samples
<= 1) {
3647 if (first_level
== last_level
)
3648 state
[7] = C_008F30_MAX_ANISO_RATIO
;
3650 state
[7] = 0xffffffff;
3654 /* Initialize the sampler view for FMASK. */
3655 if (tex
->fmask
.size
) {
3656 uint32_t data_format
, num_format
;
3658 va
= tex
->resource
.gpu_address
+ tex
->fmask
.offset
;
3660 if (screen
->b
.chip_class
>= GFX9
) {
3661 data_format
= V_008F14_IMG_DATA_FORMAT_FMASK
;
3662 switch (res
->nr_samples
) {
3664 num_format
= V_008F14_IMG_FMASK_8_2_2
;
3667 num_format
= V_008F14_IMG_FMASK_8_4_4
;
3670 num_format
= V_008F14_IMG_FMASK_32_8_8
;
3673 unreachable("invalid nr_samples");
3676 switch (res
->nr_samples
) {
3678 data_format
= V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2
;
3681 data_format
= V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4
;
3684 data_format
= V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8
;
3687 unreachable("invalid nr_samples");
3689 num_format
= V_008F14_IMG_NUM_FORMAT_UINT
;
3692 fmask_state
[0] = (va
>> 8) | tex
->fmask
.tile_swizzle
;
3693 fmask_state
[1] = S_008F14_BASE_ADDRESS_HI(va
>> 40) |
3694 S_008F14_DATA_FORMAT_GFX6(data_format
) |
3695 S_008F14_NUM_FORMAT_GFX6(num_format
);
3696 fmask_state
[2] = S_008F18_WIDTH(width
- 1) |
3697 S_008F18_HEIGHT(height
- 1);
3698 fmask_state
[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X
) |
3699 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X
) |
3700 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X
) |
3701 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X
) |
3702 S_008F1C_TYPE(si_tex_dim(screen
, tex
, target
, 0));
3704 fmask_state
[5] = S_008F24_BASE_ARRAY(first_layer
);
3708 if (screen
->b
.chip_class
>= GFX9
) {
3709 fmask_state
[3] |= S_008F1C_SW_MODE(tex
->surface
.u
.gfx9
.fmask
.swizzle_mode
);
3710 fmask_state
[4] |= S_008F20_DEPTH(last_layer
) |
3711 S_008F20_PITCH_GFX9(tex
->surface
.u
.gfx9
.fmask
.epitch
);
3712 fmask_state
[5] |= S_008F24_META_PIPE_ALIGNED(tex
->surface
.u
.gfx9
.cmask
.pipe_aligned
) |
3713 S_008F24_META_RB_ALIGNED(tex
->surface
.u
.gfx9
.cmask
.rb_aligned
);
3715 fmask_state
[3] |= S_008F1C_TILING_INDEX(tex
->fmask
.tile_mode_index
);
3716 fmask_state
[4] |= S_008F20_DEPTH(depth
- 1) |
3717 S_008F20_PITCH_GFX6(tex
->fmask
.pitch_in_pixels
- 1);
3718 fmask_state
[5] |= S_008F24_LAST_ARRAY(last_layer
);
3724 * Create a sampler view.
3726 * @param ctx context
3727 * @param texture texture
3728 * @param state sampler view template
3729 * @param width0 width0 override (for compressed textures as int)
3730 * @param height0 height0 override (for compressed textures as int)
3731 * @param force_level set the base address to the level (for compressed textures)
3733 struct pipe_sampler_view
*
3734 si_create_sampler_view_custom(struct pipe_context
*ctx
,
3735 struct pipe_resource
*texture
,
3736 const struct pipe_sampler_view
*state
,
3737 unsigned width0
, unsigned height0
,
3738 unsigned force_level
)
3740 struct si_context
*sctx
= (struct si_context
*)ctx
;
3741 struct si_sampler_view
*view
= CALLOC_STRUCT(si_sampler_view
);
3742 struct r600_texture
*tmp
= (struct r600_texture
*)texture
;
3743 unsigned base_level
, first_level
, last_level
;
3744 unsigned char state_swizzle
[4];
3745 unsigned height
, depth
, width
;
3746 unsigned last_layer
= state
->u
.tex
.last_layer
;
3747 enum pipe_format pipe_format
;
3748 const struct legacy_surf_level
*surflevel
;
3753 /* initialize base object */
3754 view
->base
= *state
;
3755 view
->base
.texture
= NULL
;
3756 view
->base
.reference
.count
= 1;
3757 view
->base
.context
= ctx
;
3760 pipe_resource_reference(&view
->base
.texture
, texture
);
3762 if (state
->format
== PIPE_FORMAT_X24S8_UINT
||
3763 state
->format
== PIPE_FORMAT_S8X24_UINT
||
3764 state
->format
== PIPE_FORMAT_X32_S8X24_UINT
||
3765 state
->format
== PIPE_FORMAT_S8_UINT
)
3766 view
->is_stencil_sampler
= true;
3768 /* Buffer resource. */
3769 if (texture
->target
== PIPE_BUFFER
) {
3770 si_make_buffer_descriptor(sctx
->screen
,
3771 (struct r600_resource
*)texture
,
3773 state
->u
.buf
.offset
,
3779 state_swizzle
[0] = state
->swizzle_r
;
3780 state_swizzle
[1] = state
->swizzle_g
;
3781 state_swizzle
[2] = state
->swizzle_b
;
3782 state_swizzle
[3] = state
->swizzle_a
;
3785 first_level
= state
->u
.tex
.first_level
;
3786 last_level
= state
->u
.tex
.last_level
;
3789 depth
= texture
->depth0
;
3791 if (sctx
->b
.chip_class
<= VI
&& force_level
) {
3792 assert(force_level
== first_level
&&
3793 force_level
== last_level
);
3794 base_level
= force_level
;
3797 width
= u_minify(width
, force_level
);
3798 height
= u_minify(height
, force_level
);
3799 depth
= u_minify(depth
, force_level
);
3802 /* This is not needed if state trackers set last_layer correctly. */
3803 if (state
->target
== PIPE_TEXTURE_1D
||
3804 state
->target
== PIPE_TEXTURE_2D
||
3805 state
->target
== PIPE_TEXTURE_RECT
||
3806 state
->target
== PIPE_TEXTURE_CUBE
)
3807 last_layer
= state
->u
.tex
.first_layer
;
3809 /* Texturing with separate depth and stencil. */
3810 pipe_format
= state
->format
;
3812 /* Depth/stencil texturing sometimes needs separate texture. */
3813 if (tmp
->is_depth
&& !r600_can_sample_zs(tmp
, view
->is_stencil_sampler
)) {
3814 if (!tmp
->flushed_depth_texture
&&
3815 !si_init_flushed_depth_texture(ctx
, texture
, NULL
)) {
3816 pipe_resource_reference(&view
->base
.texture
, NULL
);
3821 assert(tmp
->flushed_depth_texture
);
3823 /* Override format for the case where the flushed texture
3824 * contains only Z or only S.
3826 if (tmp
->flushed_depth_texture
->resource
.b
.b
.format
!= tmp
->resource
.b
.b
.format
)
3827 pipe_format
= tmp
->flushed_depth_texture
->resource
.b
.b
.format
;
3829 tmp
= tmp
->flushed_depth_texture
;
3832 surflevel
= tmp
->surface
.u
.legacy
.level
;
3834 if (tmp
->db_compatible
) {
3835 if (!view
->is_stencil_sampler
)
3836 pipe_format
= tmp
->db_render_format
;
3838 switch (pipe_format
) {
3839 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
3840 pipe_format
= PIPE_FORMAT_Z32_FLOAT
;
3842 case PIPE_FORMAT_X8Z24_UNORM
:
3843 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
3844 /* Z24 is always stored like this for DB
3847 pipe_format
= PIPE_FORMAT_Z24X8_UNORM
;
3849 case PIPE_FORMAT_X24S8_UINT
:
3850 case PIPE_FORMAT_S8X24_UINT
:
3851 case PIPE_FORMAT_X32_S8X24_UINT
:
3852 pipe_format
= PIPE_FORMAT_S8_UINT
;
3853 surflevel
= tmp
->surface
.u
.legacy
.stencil_level
;
3859 view
->dcc_incompatible
=
3860 vi_dcc_formats_are_incompatible(texture
,
3861 state
->u
.tex
.first_level
,
3864 si_make_texture_descriptor(sctx
->screen
, tmp
, true,
3865 state
->target
, pipe_format
, state_swizzle
,
3866 first_level
, last_level
,
3867 state
->u
.tex
.first_layer
, last_layer
,
3868 width
, height
, depth
,
3869 view
->state
, view
->fmask_state
);
3871 unsigned num_format
= G_008F14_NUM_FORMAT_GFX6(view
->state
[1]);
3873 num_format
== V_008F14_IMG_NUM_FORMAT_USCALED
||
3874 num_format
== V_008F14_IMG_NUM_FORMAT_SSCALED
||
3875 num_format
== V_008F14_IMG_NUM_FORMAT_UINT
||
3876 num_format
== V_008F14_IMG_NUM_FORMAT_SINT
;
3877 view
->base_level_info
= &surflevel
[base_level
];
3878 view
->base_level
= base_level
;
3879 view
->block_width
= util_format_get_blockwidth(pipe_format
);
3883 static struct pipe_sampler_view
*
3884 si_create_sampler_view(struct pipe_context
*ctx
,
3885 struct pipe_resource
*texture
,
3886 const struct pipe_sampler_view
*state
)
3888 return si_create_sampler_view_custom(ctx
, texture
, state
,
3889 texture
? texture
->width0
: 0,
3890 texture
? texture
->height0
: 0, 0);
3893 static void si_sampler_view_destroy(struct pipe_context
*ctx
,
3894 struct pipe_sampler_view
*state
)
3896 struct si_sampler_view
*view
= (struct si_sampler_view
*)state
;
3898 pipe_resource_reference(&state
->texture
, NULL
);
3902 static bool wrap_mode_uses_border_color(unsigned wrap
, bool linear_filter
)
3904 return wrap
== PIPE_TEX_WRAP_CLAMP_TO_BORDER
||
3905 wrap
== PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
||
3907 (wrap
== PIPE_TEX_WRAP_CLAMP
||
3908 wrap
== PIPE_TEX_WRAP_MIRROR_CLAMP
));
3911 static uint32_t si_translate_border_color(struct si_context
*sctx
,
3912 const struct pipe_sampler_state
*state
,
3913 const union pipe_color_union
*color
,
3916 bool linear_filter
= state
->min_img_filter
!= PIPE_TEX_FILTER_NEAREST
||
3917 state
->mag_img_filter
!= PIPE_TEX_FILTER_NEAREST
;
3919 if (!wrap_mode_uses_border_color(state
->wrap_s
, linear_filter
) &&
3920 !wrap_mode_uses_border_color(state
->wrap_t
, linear_filter
) &&
3921 !wrap_mode_uses_border_color(state
->wrap_r
, linear_filter
))
3922 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
);
3924 #define simple_border_types(elt) \
3926 if (color->elt[0] == 0 && color->elt[1] == 0 && \
3927 color->elt[2] == 0 && color->elt[3] == 0) \
3928 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK); \
3929 if (color->elt[0] == 0 && color->elt[1] == 0 && \
3930 color->elt[2] == 0 && color->elt[3] == 1) \
3931 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK); \
3932 if (color->elt[0] == 1 && color->elt[1] == 1 && \
3933 color->elt[2] == 1 && color->elt[3] == 1) \
3934 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE); \
3938 simple_border_types(ui
);
3940 simple_border_types(f
);
3942 #undef simple_border_types
3946 /* Check if the border has been uploaded already. */
3947 for (i
= 0; i
< sctx
->border_color_count
; i
++)
3948 if (memcmp(&sctx
->border_color_table
[i
], color
,
3949 sizeof(*color
)) == 0)
3952 if (i
>= SI_MAX_BORDER_COLORS
) {
3953 /* Getting 4096 unique border colors is very unlikely. */
3954 fprintf(stderr
, "radeonsi: The border color table is full. "
3955 "Any new border colors will be just black. "
3956 "Please file a bug.\n");
3957 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
);
3960 if (i
== sctx
->border_color_count
) {
3961 /* Upload a new border color. */
3962 memcpy(&sctx
->border_color_table
[i
], color
,
3964 util_memcpy_cpu_to_le32(&sctx
->border_color_map
[i
],
3965 color
, sizeof(*color
));
3966 sctx
->border_color_count
++;
3969 return S_008F3C_BORDER_COLOR_PTR(i
) |
3970 S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_REGISTER
);
3973 static void *si_create_sampler_state(struct pipe_context
*ctx
,
3974 const struct pipe_sampler_state
*state
)
3976 struct si_context
*sctx
= (struct si_context
*)ctx
;
3977 struct r600_common_screen
*rscreen
= sctx
->b
.screen
;
3978 struct si_sampler_state
*rstate
= CALLOC_STRUCT(si_sampler_state
);
3979 unsigned max_aniso
= rscreen
->force_aniso
>= 0 ? rscreen
->force_aniso
3980 : state
->max_anisotropy
;
3981 unsigned max_aniso_ratio
= r600_tex_aniso_filter(max_aniso
);
3982 union pipe_color_union clamped_border_color
;
3989 rstate
->magic
= SI_SAMPLER_STATE_MAGIC
;
3991 rstate
->val
[0] = (S_008F30_CLAMP_X(si_tex_wrap(state
->wrap_s
)) |
3992 S_008F30_CLAMP_Y(si_tex_wrap(state
->wrap_t
)) |
3993 S_008F30_CLAMP_Z(si_tex_wrap(state
->wrap_r
)) |
3994 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
3995 S_008F30_DEPTH_COMPARE_FUNC(si_tex_compare(state
->compare_func
)) |
3996 S_008F30_FORCE_UNNORMALIZED(!state
->normalized_coords
) |
3997 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
3998 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
3999 S_008F30_DISABLE_CUBE_WRAP(!state
->seamless_cube_map
) |
4000 S_008F30_COMPAT_MODE(sctx
->b
.chip_class
>= VI
));
4001 rstate
->val
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(state
->min_lod
, 0, 15), 8)) |
4002 S_008F34_MAX_LOD(S_FIXED(CLAMP(state
->max_lod
, 0, 15), 8)) |
4003 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
4004 rstate
->val
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(state
->lod_bias
, -16, 16), 8)) |
4005 S_008F38_XY_MAG_FILTER(eg_tex_filter(state
->mag_img_filter
, max_aniso
)) |
4006 S_008F38_XY_MIN_FILTER(eg_tex_filter(state
->min_img_filter
, max_aniso
)) |
4007 S_008F38_MIP_FILTER(si_tex_mipfilter(state
->min_mip_filter
)) |
4008 S_008F38_MIP_POINT_PRECLAMP(0) |
4009 S_008F38_DISABLE_LSB_CEIL(sctx
->b
.chip_class
<= VI
) |
4010 S_008F38_FILTER_PREC_FIX(1) |
4011 S_008F38_ANISO_OVERRIDE(sctx
->b
.chip_class
>= VI
));
4012 rstate
->val
[3] = si_translate_border_color(sctx
, state
, &state
->border_color
, false);
4014 /* Create sampler resource for integer textures. */
4015 memcpy(rstate
->integer_val
, rstate
->val
, sizeof(rstate
->val
));
4016 rstate
->integer_val
[3] = si_translate_border_color(sctx
, state
, &state
->border_color
, true);
4018 /* Create sampler resource for upgraded depth textures. */
4019 memcpy(rstate
->upgraded_depth_val
, rstate
->val
, sizeof(rstate
->val
));
4021 for (unsigned i
= 0; i
< 4; ++i
) {
4022 /* Use channel 0 on purpose, so that we can use OPAQUE_WHITE
4023 * when the border color is 1.0. */
4024 clamped_border_color
.f
[i
] = CLAMP(state
->border_color
.f
[0], 0, 1);
4027 if (memcmp(&state
->border_color
, &clamped_border_color
, sizeof(clamped_border_color
)) == 0)
4028 rstate
->upgraded_depth_val
[3] |= S_008F3C_UPGRADED_DEPTH(1);
4030 rstate
->upgraded_depth_val
[3] =
4031 si_translate_border_color(sctx
, state
, &clamped_border_color
, false) |
4032 S_008F3C_UPGRADED_DEPTH(1);
4037 static void si_set_sample_mask(struct pipe_context
*ctx
, unsigned sample_mask
)
4039 struct si_context
*sctx
= (struct si_context
*)ctx
;
4041 if (sctx
->sample_mask
.sample_mask
== (uint16_t)sample_mask
)
4044 sctx
->sample_mask
.sample_mask
= sample_mask
;
4045 si_mark_atom_dirty(sctx
, &sctx
->sample_mask
.atom
);
4048 static void si_emit_sample_mask(struct si_context
*sctx
, struct r600_atom
*atom
)
4050 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
4051 unsigned mask
= sctx
->sample_mask
.sample_mask
;
4053 /* Needed for line and polygon smoothing as well as for the Polaris
4054 * small primitive filter. We expect the state tracker to take care of
4057 assert(mask
== 0xffff || sctx
->framebuffer
.nr_samples
> 1 ||
4058 (mask
& 1 && sctx
->blitter
->running
));
4060 radeon_set_context_reg_seq(cs
, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0
, 2);
4061 radeon_emit(cs
, mask
| (mask
<< 16));
4062 radeon_emit(cs
, mask
| (mask
<< 16));
4065 static void si_delete_sampler_state(struct pipe_context
*ctx
, void *state
)
4068 struct si_sampler_state
*s
= state
;
4070 assert(s
->magic
== SI_SAMPLER_STATE_MAGIC
);
4077 * Vertex elements & buffers
4080 static void *si_create_vertex_elements(struct pipe_context
*ctx
,
4082 const struct pipe_vertex_element
*elements
)
4084 struct si_screen
*sscreen
= (struct si_screen
*)ctx
->screen
;
4085 struct si_vertex_elements
*v
= CALLOC_STRUCT(si_vertex_elements
);
4086 bool used
[SI_NUM_VERTEX_BUFFERS
] = {};
4089 assert(count
<= SI_MAX_ATTRIBS
);
4094 v
->desc_list_byte_size
= align(count
* 16, SI_CPDMA_ALIGNMENT
);
4096 for (i
= 0; i
< count
; ++i
) {
4097 const struct util_format_description
*desc
;
4098 const struct util_format_channel_description
*channel
;
4099 unsigned data_format
, num_format
;
4101 unsigned vbo_index
= elements
[i
].vertex_buffer_index
;
4102 unsigned char swizzle
[4];
4104 if (vbo_index
>= SI_NUM_VERTEX_BUFFERS
) {
4109 if (elements
[i
].instance_divisor
) {
4110 v
->uses_instance_divisors
= true;
4111 v
->instance_divisors
[i
] = elements
[i
].instance_divisor
;
4113 if (v
->instance_divisors
[i
] == 1)
4114 v
->instance_divisor_is_one
|= 1u << i
;
4116 v
->instance_divisor_is_fetched
|= 1u << i
;
4119 if (!used
[vbo_index
]) {
4120 v
->first_vb_use_mask
|= 1 << i
;
4121 used
[vbo_index
] = true;
4124 desc
= util_format_description(elements
[i
].src_format
);
4125 first_non_void
= util_format_get_first_non_void_channel(elements
[i
].src_format
);
4126 data_format
= si_translate_buffer_dataformat(ctx
->screen
, desc
, first_non_void
);
4127 num_format
= si_translate_buffer_numformat(ctx
->screen
, desc
, first_non_void
);
4128 channel
= first_non_void
>= 0 ? &desc
->channel
[first_non_void
] : NULL
;
4129 memcpy(swizzle
, desc
->swizzle
, sizeof(swizzle
));
4131 v
->format_size
[i
] = desc
->block
.bits
/ 8;
4132 v
->src_offset
[i
] = elements
[i
].src_offset
;
4133 v
->vertex_buffer_index
[i
] = vbo_index
;
4135 /* The hardware always treats the 2-bit alpha channel as
4136 * unsigned, so a shader workaround is needed. The affected
4137 * chips are VI and older except Stoney (GFX8.1).
4139 if (data_format
== V_008F0C_BUF_DATA_FORMAT_2_10_10_10
&&
4140 sscreen
->b
.chip_class
<= VI
&&
4141 sscreen
->b
.family
!= CHIP_STONEY
) {
4142 if (num_format
== V_008F0C_BUF_NUM_FORMAT_SNORM
) {
4143 v
->fix_fetch
[i
] = SI_FIX_FETCH_A2_SNORM
;
4144 } else if (num_format
== V_008F0C_BUF_NUM_FORMAT_SSCALED
) {
4145 v
->fix_fetch
[i
] = SI_FIX_FETCH_A2_SSCALED
;
4146 } else if (num_format
== V_008F0C_BUF_NUM_FORMAT_SINT
) {
4147 /* This isn't actually used in OpenGL. */
4148 v
->fix_fetch
[i
] = SI_FIX_FETCH_A2_SINT
;
4150 } else if (channel
&& channel
->type
== UTIL_FORMAT_TYPE_FIXED
) {
4151 if (desc
->swizzle
[3] == PIPE_SWIZZLE_1
)
4152 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBX_32_FIXED
;
4154 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_FIXED
;
4155 } else if (channel
&& channel
->size
== 32 && !channel
->pure_integer
) {
4156 if (channel
->type
== UTIL_FORMAT_TYPE_SIGNED
) {
4157 if (channel
->normalized
) {
4158 if (desc
->swizzle
[3] == PIPE_SWIZZLE_1
)
4159 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBX_32_SNORM
;
4161 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_SNORM
;
4163 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_SSCALED
;
4165 } else if (channel
->type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
4166 if (channel
->normalized
) {
4167 if (desc
->swizzle
[3] == PIPE_SWIZZLE_1
)
4168 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBX_32_UNORM
;
4170 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_UNORM
;
4172 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_USCALED
;
4175 } else if (channel
&& channel
->size
== 64 &&
4176 channel
->type
== UTIL_FORMAT_TYPE_FLOAT
) {
4177 switch (desc
->nr_channels
) {
4180 v
->fix_fetch
[i
] = SI_FIX_FETCH_RG_64_FLOAT
;
4181 swizzle
[0] = PIPE_SWIZZLE_X
;
4182 swizzle
[1] = PIPE_SWIZZLE_Y
;
4183 swizzle
[2] = desc
->nr_channels
== 2 ? PIPE_SWIZZLE_Z
: PIPE_SWIZZLE_0
;
4184 swizzle
[3] = desc
->nr_channels
== 2 ? PIPE_SWIZZLE_W
: PIPE_SWIZZLE_0
;
4187 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_64_FLOAT
;
4188 swizzle
[0] = PIPE_SWIZZLE_X
; /* 3 loads */
4189 swizzle
[1] = PIPE_SWIZZLE_Y
;
4190 swizzle
[2] = PIPE_SWIZZLE_0
;
4191 swizzle
[3] = PIPE_SWIZZLE_0
;
4194 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_64_FLOAT
;
4195 swizzle
[0] = PIPE_SWIZZLE_X
; /* 2 loads */
4196 swizzle
[1] = PIPE_SWIZZLE_Y
;
4197 swizzle
[2] = PIPE_SWIZZLE_Z
;
4198 swizzle
[3] = PIPE_SWIZZLE_W
;
4203 } else if (channel
&& desc
->nr_channels
== 3) {
4204 assert(desc
->swizzle
[0] == PIPE_SWIZZLE_X
);
4206 if (channel
->size
== 8) {
4207 if (channel
->pure_integer
)
4208 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_8_INT
;
4210 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_8
;
4211 } else if (channel
->size
== 16) {
4212 if (channel
->pure_integer
)
4213 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_16_INT
;
4215 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_16
;
4219 v
->rsrc_word3
[i
] = S_008F0C_DST_SEL_X(si_map_swizzle(swizzle
[0])) |
4220 S_008F0C_DST_SEL_Y(si_map_swizzle(swizzle
[1])) |
4221 S_008F0C_DST_SEL_Z(si_map_swizzle(swizzle
[2])) |
4222 S_008F0C_DST_SEL_W(si_map_swizzle(swizzle
[3])) |
4223 S_008F0C_NUM_FORMAT(num_format
) |
4224 S_008F0C_DATA_FORMAT(data_format
);
4229 static void si_bind_vertex_elements(struct pipe_context
*ctx
, void *state
)
4231 struct si_context
*sctx
= (struct si_context
*)ctx
;
4232 struct si_vertex_elements
*old
= sctx
->vertex_elements
;
4233 struct si_vertex_elements
*v
= (struct si_vertex_elements
*)state
;
4235 sctx
->vertex_elements
= v
;
4236 sctx
->vertex_buffers_dirty
= true;
4240 old
->count
!= v
->count
||
4241 old
->uses_instance_divisors
!= v
->uses_instance_divisors
||
4242 v
->uses_instance_divisors
|| /* we don't check which divisors changed */
4243 memcmp(old
->fix_fetch
, v
->fix_fetch
, sizeof(v
->fix_fetch
[0]) * v
->count
)))
4244 sctx
->do_update_shaders
= true;
4246 if (v
&& v
->instance_divisor_is_fetched
) {
4247 struct pipe_constant_buffer cb
;
4250 cb
.user_buffer
= v
->instance_divisors
;
4251 cb
.buffer_offset
= 0;
4252 cb
.buffer_size
= sizeof(uint32_t) * v
->count
;
4253 si_set_rw_buffer(sctx
, SI_VS_CONST_INSTANCE_DIVISORS
, &cb
);
4257 static void si_delete_vertex_element(struct pipe_context
*ctx
, void *state
)
4259 struct si_context
*sctx
= (struct si_context
*)ctx
;
4261 if (sctx
->vertex_elements
== state
)
4262 sctx
->vertex_elements
= NULL
;
4266 static void si_set_vertex_buffers(struct pipe_context
*ctx
,
4267 unsigned start_slot
, unsigned count
,
4268 const struct pipe_vertex_buffer
*buffers
)
4270 struct si_context
*sctx
= (struct si_context
*)ctx
;
4271 struct pipe_vertex_buffer
*dst
= sctx
->vertex_buffer
+ start_slot
;
4274 assert(start_slot
+ count
<= ARRAY_SIZE(sctx
->vertex_buffer
));
4277 for (i
= 0; i
< count
; i
++) {
4278 const struct pipe_vertex_buffer
*src
= buffers
+ i
;
4279 struct pipe_vertex_buffer
*dsti
= dst
+ i
;
4280 struct pipe_resource
*buf
= src
->buffer
.resource
;
4282 pipe_resource_reference(&dsti
->buffer
.resource
, buf
);
4283 dsti
->buffer_offset
= src
->buffer_offset
;
4284 dsti
->stride
= src
->stride
;
4285 r600_context_add_resource_size(ctx
, buf
);
4287 r600_resource(buf
)->bind_history
|= PIPE_BIND_VERTEX_BUFFER
;
4290 for (i
= 0; i
< count
; i
++) {
4291 pipe_resource_reference(&dst
[i
].buffer
.resource
, NULL
);
4294 sctx
->vertex_buffers_dirty
= true;
4301 static void si_set_tess_state(struct pipe_context
*ctx
,
4302 const float default_outer_level
[4],
4303 const float default_inner_level
[2])
4305 struct si_context
*sctx
= (struct si_context
*)ctx
;
4306 struct pipe_constant_buffer cb
;
4309 memcpy(array
, default_outer_level
, sizeof(float) * 4);
4310 memcpy(array
+4, default_inner_level
, sizeof(float) * 2);
4313 cb
.user_buffer
= NULL
;
4314 cb
.buffer_size
= sizeof(array
);
4316 si_upload_const_buffer(sctx
, (struct r600_resource
**)&cb
.buffer
,
4317 (void*)array
, sizeof(array
),
4320 si_set_rw_buffer(sctx
, SI_HS_CONST_DEFAULT_TESS_LEVELS
, &cb
);
4321 pipe_resource_reference(&cb
.buffer
, NULL
);
4324 static void si_texture_barrier(struct pipe_context
*ctx
, unsigned flags
)
4326 struct si_context
*sctx
= (struct si_context
*)ctx
;
4328 si_update_fb_dirtiness_after_rendering(sctx
);
4330 /* Multisample surfaces are flushed in si_decompress_textures. */
4331 if (sctx
->framebuffer
.nr_samples
<= 1 &&
4332 sctx
->framebuffer
.state
.nr_cbufs
)
4333 si_make_CB_shader_coherent(sctx
, sctx
->framebuffer
.nr_samples
,
4334 sctx
->framebuffer
.CB_has_shader_readable_metadata
);
4337 /* This only ensures coherency for shader image/buffer stores. */
4338 static void si_memory_barrier(struct pipe_context
*ctx
, unsigned flags
)
4340 struct si_context
*sctx
= (struct si_context
*)ctx
;
4342 /* Subsequent commands must wait for all shader invocations to
4344 sctx
->b
.flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
|
4345 SI_CONTEXT_CS_PARTIAL_FLUSH
;
4347 if (flags
& PIPE_BARRIER_CONSTANT_BUFFER
)
4348 sctx
->b
.flags
|= SI_CONTEXT_INV_SMEM_L1
|
4349 SI_CONTEXT_INV_VMEM_L1
;
4351 if (flags
& (PIPE_BARRIER_VERTEX_BUFFER
|
4352 PIPE_BARRIER_SHADER_BUFFER
|
4353 PIPE_BARRIER_TEXTURE
|
4354 PIPE_BARRIER_IMAGE
|
4355 PIPE_BARRIER_STREAMOUT_BUFFER
|
4356 PIPE_BARRIER_GLOBAL_BUFFER
)) {
4357 /* As far as I can tell, L1 contents are written back to L2
4358 * automatically at end of shader, but the contents of other
4359 * L1 caches might still be stale. */
4360 sctx
->b
.flags
|= SI_CONTEXT_INV_VMEM_L1
;
4363 if (flags
& PIPE_BARRIER_INDEX_BUFFER
) {
4364 /* Indices are read through TC L2 since VI.
4367 if (sctx
->screen
->b
.chip_class
<= CIK
)
4368 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
4371 /* MSAA color, any depth and any stencil are flushed in
4372 * si_decompress_textures when needed.
4374 if (flags
& PIPE_BARRIER_FRAMEBUFFER
&&
4375 sctx
->framebuffer
.nr_samples
<= 1 &&
4376 sctx
->framebuffer
.state
.nr_cbufs
) {
4377 sctx
->b
.flags
|= SI_CONTEXT_FLUSH_AND_INV_CB
;
4379 if (sctx
->b
.chip_class
<= VI
)
4380 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
4383 /* Indirect buffers use TC L2 on GFX9, but not older hw. */
4384 if (sctx
->screen
->b
.chip_class
<= VI
&&
4385 flags
& PIPE_BARRIER_INDIRECT_BUFFER
)
4386 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
4389 static void *si_create_blend_custom(struct si_context
*sctx
, unsigned mode
)
4391 struct pipe_blend_state blend
;
4393 memset(&blend
, 0, sizeof(blend
));
4394 blend
.independent_blend_enable
= true;
4395 blend
.rt
[0].colormask
= 0xf;
4396 return si_create_blend_state_mode(&sctx
->b
.b
, &blend
, mode
);
4399 static void si_need_gfx_cs_space(struct pipe_context
*ctx
, unsigned num_dw
,
4400 bool include_draw_vbo
)
4402 si_need_cs_space((struct si_context
*)ctx
);
4405 static void si_init_config(struct si_context
*sctx
);
4407 void si_init_state_functions(struct si_context
*sctx
)
4409 si_init_external_atom(sctx
, &sctx
->b
.render_cond_atom
, &sctx
->atoms
.s
.render_cond
);
4410 si_init_external_atom(sctx
, &sctx
->b
.streamout
.begin_atom
, &sctx
->atoms
.s
.streamout_begin
);
4411 si_init_external_atom(sctx
, &sctx
->b
.streamout
.enable_atom
, &sctx
->atoms
.s
.streamout_enable
);
4412 si_init_external_atom(sctx
, &sctx
->scissors
.atom
, &sctx
->atoms
.s
.scissors
);
4413 si_init_external_atom(sctx
, &sctx
->viewports
.atom
, &sctx
->atoms
.s
.viewports
);
4415 si_init_atom(sctx
, &sctx
->framebuffer
.atom
, &sctx
->atoms
.s
.framebuffer
, si_emit_framebuffer_state
);
4416 si_init_atom(sctx
, &sctx
->msaa_sample_locs
.atom
, &sctx
->atoms
.s
.msaa_sample_locs
, si_emit_msaa_sample_locs
);
4417 si_init_atom(sctx
, &sctx
->db_render_state
, &sctx
->atoms
.s
.db_render_state
, si_emit_db_render_state
);
4418 si_init_atom(sctx
, &sctx
->dpbb_state
, &sctx
->atoms
.s
.dpbb_state
, si_emit_dpbb_state
);
4419 si_init_atom(sctx
, &sctx
->msaa_config
, &sctx
->atoms
.s
.msaa_config
, si_emit_msaa_config
);
4420 si_init_atom(sctx
, &sctx
->sample_mask
.atom
, &sctx
->atoms
.s
.sample_mask
, si_emit_sample_mask
);
4421 si_init_atom(sctx
, &sctx
->cb_render_state
, &sctx
->atoms
.s
.cb_render_state
, si_emit_cb_render_state
);
4422 si_init_atom(sctx
, &sctx
->blend_color
.atom
, &sctx
->atoms
.s
.blend_color
, si_emit_blend_color
);
4423 si_init_atom(sctx
, &sctx
->clip_regs
, &sctx
->atoms
.s
.clip_regs
, si_emit_clip_regs
);
4424 si_init_atom(sctx
, &sctx
->clip_state
.atom
, &sctx
->atoms
.s
.clip_state
, si_emit_clip_state
);
4425 si_init_atom(sctx
, &sctx
->stencil_ref
.atom
, &sctx
->atoms
.s
.stencil_ref
, si_emit_stencil_ref
);
4427 sctx
->b
.b
.create_blend_state
= si_create_blend_state
;
4428 sctx
->b
.b
.bind_blend_state
= si_bind_blend_state
;
4429 sctx
->b
.b
.delete_blend_state
= si_delete_blend_state
;
4430 sctx
->b
.b
.set_blend_color
= si_set_blend_color
;
4432 sctx
->b
.b
.create_rasterizer_state
= si_create_rs_state
;
4433 sctx
->b
.b
.bind_rasterizer_state
= si_bind_rs_state
;
4434 sctx
->b
.b
.delete_rasterizer_state
= si_delete_rs_state
;
4436 sctx
->b
.b
.create_depth_stencil_alpha_state
= si_create_dsa_state
;
4437 sctx
->b
.b
.bind_depth_stencil_alpha_state
= si_bind_dsa_state
;
4438 sctx
->b
.b
.delete_depth_stencil_alpha_state
= si_delete_dsa_state
;
4440 sctx
->custom_dsa_flush
= si_create_db_flush_dsa(sctx
);
4441 sctx
->custom_blend_resolve
= si_create_blend_custom(sctx
, V_028808_CB_RESOLVE
);
4442 sctx
->custom_blend_fmask_decompress
= si_create_blend_custom(sctx
, V_028808_CB_FMASK_DECOMPRESS
);
4443 sctx
->custom_blend_eliminate_fastclear
= si_create_blend_custom(sctx
, V_028808_CB_ELIMINATE_FAST_CLEAR
);
4444 sctx
->custom_blend_dcc_decompress
= si_create_blend_custom(sctx
, V_028808_CB_DCC_DECOMPRESS
);
4446 sctx
->b
.b
.set_clip_state
= si_set_clip_state
;
4447 sctx
->b
.b
.set_stencil_ref
= si_set_stencil_ref
;
4449 sctx
->b
.b
.set_framebuffer_state
= si_set_framebuffer_state
;
4450 sctx
->b
.b
.get_sample_position
= si_get_sample_position
;
4452 sctx
->b
.b
.create_sampler_state
= si_create_sampler_state
;
4453 sctx
->b
.b
.delete_sampler_state
= si_delete_sampler_state
;
4455 sctx
->b
.b
.create_sampler_view
= si_create_sampler_view
;
4456 sctx
->b
.b
.sampler_view_destroy
= si_sampler_view_destroy
;
4458 sctx
->b
.b
.set_sample_mask
= si_set_sample_mask
;
4460 sctx
->b
.b
.create_vertex_elements_state
= si_create_vertex_elements
;
4461 sctx
->b
.b
.bind_vertex_elements_state
= si_bind_vertex_elements
;
4462 sctx
->b
.b
.delete_vertex_elements_state
= si_delete_vertex_element
;
4463 sctx
->b
.b
.set_vertex_buffers
= si_set_vertex_buffers
;
4465 sctx
->b
.b
.texture_barrier
= si_texture_barrier
;
4466 sctx
->b
.b
.memory_barrier
= si_memory_barrier
;
4467 sctx
->b
.b
.set_min_samples
= si_set_min_samples
;
4468 sctx
->b
.b
.set_tess_state
= si_set_tess_state
;
4470 sctx
->b
.b
.set_active_query_state
= si_set_active_query_state
;
4471 sctx
->b
.set_occlusion_query_state
= si_set_occlusion_query_state
;
4472 sctx
->b
.save_qbo_state
= si_save_qbo_state
;
4473 sctx
->b
.need_gfx_cs_space
= si_need_gfx_cs_space
;
4475 sctx
->b
.b
.draw_vbo
= si_draw_vbo
;
4477 si_init_config(sctx
);
4480 static uint32_t si_get_bo_metadata_word1(struct r600_common_screen
*rscreen
)
4482 return (ATI_VENDOR_ID
<< 16) | rscreen
->info
.pci_id
;
4485 static void si_query_opaque_metadata(struct r600_common_screen
*rscreen
,
4486 struct r600_texture
*rtex
,
4487 struct radeon_bo_metadata
*md
)
4489 struct si_screen
*sscreen
= (struct si_screen
*)rscreen
;
4490 struct pipe_resource
*res
= &rtex
->resource
.b
.b
;
4491 static const unsigned char swizzle
[] = {
4497 uint32_t desc
[8], i
;
4498 bool is_array
= util_resource_is_array_texture(res
);
4500 /* DRM 2.x.x doesn't support this. */
4501 if (rscreen
->info
.drm_major
!= 3)
4504 assert(rtex
->dcc_separate_buffer
== NULL
);
4505 assert(rtex
->fmask
.size
== 0);
4507 /* Metadata image format format version 1:
4508 * [0] = 1 (metadata format identifier)
4509 * [1] = (VENDOR_ID << 16) | PCI_ID
4510 * [2:9] = image descriptor for the whole resource
4511 * [2] is always 0, because the base address is cleared
4512 * [9] is the DCC offset bits [39:8] from the beginning of
4514 * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
4517 md
->metadata
[0] = 1; /* metadata image format version 1 */
4519 /* TILE_MODE_INDEX is ambiguous without a PCI ID. */
4520 md
->metadata
[1] = si_get_bo_metadata_word1(rscreen
);
4522 si_make_texture_descriptor(sscreen
, rtex
, true,
4523 res
->target
, res
->format
,
4524 swizzle
, 0, res
->last_level
, 0,
4525 is_array
? res
->array_size
- 1 : 0,
4526 res
->width0
, res
->height0
, res
->depth0
,
4529 si_set_mutable_tex_desc_fields(sscreen
, rtex
, &rtex
->surface
.u
.legacy
.level
[0],
4530 0, 0, rtex
->surface
.blk_w
, false, desc
);
4532 /* Clear the base address and set the relative DCC offset. */
4534 desc
[1] &= C_008F14_BASE_ADDRESS_HI
;
4535 desc
[7] = rtex
->dcc_offset
>> 8;
4537 /* Dwords [2:9] contain the image descriptor. */
4538 memcpy(&md
->metadata
[2], desc
, sizeof(desc
));
4539 md
->size_metadata
= 10 * 4;
4541 /* Dwords [10:..] contain the mipmap level offsets. */
4542 if (rscreen
->chip_class
<= VI
) {
4543 for (i
= 0; i
<= res
->last_level
; i
++)
4544 md
->metadata
[10+i
] = rtex
->surface
.u
.legacy
.level
[i
].offset
>> 8;
4546 md
->size_metadata
+= (1 + res
->last_level
) * 4;
4550 static void si_apply_opaque_metadata(struct r600_common_screen
*rscreen
,
4551 struct r600_texture
*rtex
,
4552 struct radeon_bo_metadata
*md
)
4554 uint32_t *desc
= &md
->metadata
[2];
4556 if (rscreen
->chip_class
< VI
)
4559 /* Return if DCC is enabled. The texture should be set up with it
4562 if (md
->size_metadata
>= 11 * 4 &&
4563 md
->metadata
[0] != 0 &&
4564 md
->metadata
[1] == si_get_bo_metadata_word1(rscreen
) &&
4565 G_008F28_COMPRESSION_EN(desc
[6])) {
4566 assert(rtex
->dcc_offset
== ((uint64_t)desc
[7] << 8));
4570 /* Disable DCC. These are always set by texture_from_handle and must
4573 rtex
->dcc_offset
= 0;
4576 void si_init_screen_state_functions(struct si_screen
*sscreen
)
4578 sscreen
->b
.b
.is_format_supported
= si_is_format_supported
;
4579 sscreen
->b
.query_opaque_metadata
= si_query_opaque_metadata
;
4580 sscreen
->b
.apply_opaque_metadata
= si_apply_opaque_metadata
;
4583 static void si_set_grbm_gfx_index(struct si_context
*sctx
,
4584 struct si_pm4_state
*pm4
, unsigned value
)
4586 unsigned reg
= sctx
->b
.chip_class
>= CIK
? R_030800_GRBM_GFX_INDEX
:
4588 si_pm4_set_reg(pm4
, reg
, value
);
4591 static void si_set_grbm_gfx_index_se(struct si_context
*sctx
,
4592 struct si_pm4_state
*pm4
, unsigned se
)
4594 assert(se
== ~0 || se
< sctx
->screen
->b
.info
.max_se
);
4595 si_set_grbm_gfx_index(sctx
, pm4
,
4596 (se
== ~0 ? S_030800_SE_BROADCAST_WRITES(1) :
4597 S_030800_SE_INDEX(se
)) |
4598 S_030800_SH_BROADCAST_WRITES(1) |
4599 S_030800_INSTANCE_BROADCAST_WRITES(1));
4603 si_write_harvested_raster_configs(struct si_context
*sctx
,
4604 struct si_pm4_state
*pm4
,
4605 unsigned raster_config
,
4606 unsigned raster_config_1
)
4608 unsigned sh_per_se
= MAX2(sctx
->screen
->b
.info
.max_sh_per_se
, 1);
4609 unsigned num_se
= MAX2(sctx
->screen
->b
.info
.max_se
, 1);
4610 unsigned rb_mask
= sctx
->screen
->b
.info
.enabled_rb_mask
;
4611 unsigned num_rb
= MIN2(sctx
->screen
->b
.info
.num_render_backends
, 16);
4612 unsigned rb_per_pkr
= MIN2(num_rb
/ num_se
/ sh_per_se
, 2);
4613 unsigned rb_per_se
= num_rb
/ num_se
;
4614 unsigned se_mask
[4];
4617 se_mask
[0] = ((1 << rb_per_se
) - 1);
4618 se_mask
[1] = (se_mask
[0] << rb_per_se
);
4619 se_mask
[2] = (se_mask
[1] << rb_per_se
);
4620 se_mask
[3] = (se_mask
[2] << rb_per_se
);
4622 se_mask
[0] &= rb_mask
;
4623 se_mask
[1] &= rb_mask
;
4624 se_mask
[2] &= rb_mask
;
4625 se_mask
[3] &= rb_mask
;
4627 assert(num_se
== 1 || num_se
== 2 || num_se
== 4);
4628 assert(sh_per_se
== 1 || sh_per_se
== 2);
4629 assert(rb_per_pkr
== 1 || rb_per_pkr
== 2);
4631 /* XXX: I can't figure out what the *_XSEL and *_YSEL
4632 * fields are for, so I'm leaving them as their default
4635 for (se
= 0; se
< num_se
; se
++) {
4636 unsigned raster_config_se
= raster_config
;
4637 unsigned pkr0_mask
= ((1 << rb_per_pkr
) - 1) << (se
* rb_per_se
);
4638 unsigned pkr1_mask
= pkr0_mask
<< rb_per_pkr
;
4639 int idx
= (se
/ 2) * 2;
4641 if ((num_se
> 1) && (!se_mask
[idx
] || !se_mask
[idx
+ 1])) {
4642 raster_config_se
&= C_028350_SE_MAP
;
4644 if (!se_mask
[idx
]) {
4646 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_3
);
4649 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_0
);
4653 pkr0_mask
&= rb_mask
;
4654 pkr1_mask
&= rb_mask
;
4655 if (rb_per_se
> 2 && (!pkr0_mask
|| !pkr1_mask
)) {
4656 raster_config_se
&= C_028350_PKR_MAP
;
4660 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_3
);
4663 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_0
);
4667 if (rb_per_se
>= 2) {
4668 unsigned rb0_mask
= 1 << (se
* rb_per_se
);
4669 unsigned rb1_mask
= rb0_mask
<< 1;
4671 rb0_mask
&= rb_mask
;
4672 rb1_mask
&= rb_mask
;
4673 if (!rb0_mask
|| !rb1_mask
) {
4674 raster_config_se
&= C_028350_RB_MAP_PKR0
;
4678 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_3
);
4681 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_0
);
4685 if (rb_per_se
> 2) {
4686 rb0_mask
= 1 << (se
* rb_per_se
+ rb_per_pkr
);
4687 rb1_mask
= rb0_mask
<< 1;
4688 rb0_mask
&= rb_mask
;
4689 rb1_mask
&= rb_mask
;
4690 if (!rb0_mask
|| !rb1_mask
) {
4691 raster_config_se
&= C_028350_RB_MAP_PKR1
;
4695 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_3
);
4698 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_0
);
4704 si_set_grbm_gfx_index_se(sctx
, pm4
, se
);
4705 si_pm4_set_reg(pm4
, R_028350_PA_SC_RASTER_CONFIG
, raster_config_se
);
4707 si_set_grbm_gfx_index(sctx
, pm4
, ~0);
4709 if (sctx
->b
.chip_class
>= CIK
) {
4710 if ((num_se
> 2) && ((!se_mask
[0] && !se_mask
[1]) ||
4711 (!se_mask
[2] && !se_mask
[3]))) {
4712 raster_config_1
&= C_028354_SE_PAIR_MAP
;
4714 if (!se_mask
[0] && !se_mask
[1]) {
4716 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_3
);
4719 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_0
);
4723 si_pm4_set_reg(pm4
, R_028354_PA_SC_RASTER_CONFIG_1
, raster_config_1
);
4727 static void si_set_raster_config(struct si_context
*sctx
, struct si_pm4_state
*pm4
)
4729 struct si_screen
*sscreen
= sctx
->screen
;
4730 unsigned num_rb
= MIN2(sctx
->screen
->b
.info
.num_render_backends
, 16);
4731 unsigned rb_mask
= sctx
->screen
->b
.info
.enabled_rb_mask
;
4732 unsigned raster_config
, raster_config_1
;
4734 switch (sctx
->b
.family
) {
4737 raster_config
= 0x2a00126a;
4738 raster_config_1
= 0x00000000;
4741 raster_config
= 0x0000124a;
4742 raster_config_1
= 0x00000000;
4745 raster_config
= 0x00000082;
4746 raster_config_1
= 0x00000000;
4749 raster_config
= 0x00000000;
4750 raster_config_1
= 0x00000000;
4753 raster_config
= 0x16000012;
4754 raster_config_1
= 0x00000000;
4757 raster_config
= 0x3a00161a;
4758 raster_config_1
= 0x0000002e;
4761 if (sscreen
->b
.info
.cik_macrotile_mode_array
[0] == 0x000000e8) {
4762 /* old kernels with old tiling config */
4763 raster_config
= 0x16000012;
4764 raster_config_1
= 0x0000002a;
4766 raster_config
= 0x3a00161a;
4767 raster_config_1
= 0x0000002e;
4770 case CHIP_POLARIS10
:
4771 raster_config
= 0x16000012;
4772 raster_config_1
= 0x0000002a;
4774 case CHIP_POLARIS11
:
4775 case CHIP_POLARIS12
:
4776 raster_config
= 0x16000012;
4777 raster_config_1
= 0x00000000;
4780 raster_config
= 0x16000012;
4781 raster_config_1
= 0x0000002a;
4785 raster_config
= 0x00000000;
4787 raster_config
= 0x00000002;
4788 raster_config_1
= 0x00000000;
4791 raster_config
= 0x00000002;
4792 raster_config_1
= 0x00000000;
4795 /* KV should be 0x00000002, but that causes problems with radeon */
4796 raster_config
= 0x00000000; /* 0x00000002 */
4797 raster_config_1
= 0x00000000;
4802 raster_config
= 0x00000000;
4803 raster_config_1
= 0x00000000;
4807 "radeonsi: Unknown GPU, using 0 for raster_config\n");
4808 raster_config
= 0x00000000;
4809 raster_config_1
= 0x00000000;
4812 if (!rb_mask
|| util_bitcount(rb_mask
) >= num_rb
) {
4813 /* Always use the default config when all backends are enabled
4814 * (or when we failed to determine the enabled backends).
4816 si_pm4_set_reg(pm4
, R_028350_PA_SC_RASTER_CONFIG
,
4818 if (sctx
->b
.chip_class
>= CIK
)
4819 si_pm4_set_reg(pm4
, R_028354_PA_SC_RASTER_CONFIG_1
,
4822 si_write_harvested_raster_configs(sctx
, pm4
, raster_config
, raster_config_1
);
4826 static void si_init_config(struct si_context
*sctx
)
4828 struct si_screen
*sscreen
= sctx
->screen
;
4829 uint64_t border_color_va
= sctx
->border_color_buffer
->gpu_address
;
4830 bool has_clear_state
= sscreen
->has_clear_state
;
4831 struct si_pm4_state
*pm4
= CALLOC_STRUCT(si_pm4_state
);
4833 /* Only SI can disable CLEAR_STATE for now. */
4834 assert(has_clear_state
|| sscreen
->b
.chip_class
== SI
);
4839 si_pm4_cmd_begin(pm4
, PKT3_CONTEXT_CONTROL
);
4840 si_pm4_cmd_add(pm4
, CONTEXT_CONTROL_LOAD_ENABLE(1));
4841 si_pm4_cmd_add(pm4
, CONTEXT_CONTROL_SHADOW_ENABLE(1));
4842 si_pm4_cmd_end(pm4
, false);
4844 if (has_clear_state
) {
4845 si_pm4_cmd_begin(pm4
, PKT3_CLEAR_STATE
);
4846 si_pm4_cmd_add(pm4
, 0);
4847 si_pm4_cmd_end(pm4
, false);
4850 if (sctx
->b
.chip_class
<= VI
)
4851 si_set_raster_config(sctx
, pm4
);
4853 si_pm4_set_reg(pm4
, R_028A18_VGT_HOS_MAX_TESS_LEVEL
, fui(64));
4854 if (!has_clear_state
)
4855 si_pm4_set_reg(pm4
, R_028A1C_VGT_HOS_MIN_TESS_LEVEL
, fui(0));
4857 /* FIXME calculate these values somehow ??? */
4858 if (sctx
->b
.chip_class
<= VI
) {
4859 si_pm4_set_reg(pm4
, R_028A54_VGT_GS_PER_ES
, SI_GS_PER_ES
);
4860 si_pm4_set_reg(pm4
, R_028A58_VGT_ES_PER_GS
, 0x40);
4863 if (!has_clear_state
) {
4864 si_pm4_set_reg(pm4
, R_028A5C_VGT_GS_PER_VS
, 0x2);
4865 si_pm4_set_reg(pm4
, R_028A8C_VGT_PRIMITIVEID_RESET
, 0x0);
4866 si_pm4_set_reg(pm4
, R_028B98_VGT_STRMOUT_BUFFER_CONFIG
, 0x0);
4869 si_pm4_set_reg(pm4
, R_028AA0_VGT_INSTANCE_STEP_RATE_0
, 1);
4870 if (!has_clear_state
)
4871 si_pm4_set_reg(pm4
, R_028AB8_VGT_VTX_CNT_EN
, 0x0);
4872 if (sctx
->b
.chip_class
< CIK
)
4873 si_pm4_set_reg(pm4
, R_008A14_PA_CL_ENHANCE
, S_008A14_NUM_CLIP_SEQ(3) |
4874 S_008A14_CLIP_VTX_REORDER_ENA(1));
4876 si_pm4_set_reg(pm4
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 0x76543210);
4877 si_pm4_set_reg(pm4
, R_028BD8_PA_SC_CENTROID_PRIORITY_1
, 0xfedcba98);
4879 if (!has_clear_state
)
4880 si_pm4_set_reg(pm4
, R_02882C_PA_SU_PRIM_FILTER_CNTL
, 0);
4882 /* CLEAR_STATE doesn't clear these correctly on certain generations.
4883 * I don't know why. Deduced by trial and error.
4885 if (sctx
->b
.chip_class
<= CIK
) {
4886 si_pm4_set_reg(pm4
, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET
, 0);
4887 si_pm4_set_reg(pm4
, R_028204_PA_SC_WINDOW_SCISSOR_TL
, S_028204_WINDOW_OFFSET_DISABLE(1));
4888 si_pm4_set_reg(pm4
, R_028240_PA_SC_GENERIC_SCISSOR_TL
, S_028240_WINDOW_OFFSET_DISABLE(1));
4889 si_pm4_set_reg(pm4
, R_028244_PA_SC_GENERIC_SCISSOR_BR
,
4890 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
4891 si_pm4_set_reg(pm4
, R_028030_PA_SC_SCREEN_SCISSOR_TL
, 0);
4892 si_pm4_set_reg(pm4
, R_028034_PA_SC_SCREEN_SCISSOR_BR
,
4893 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
4896 if (!has_clear_state
) {
4897 si_pm4_set_reg(pm4
, R_02820C_PA_SC_CLIPRECT_RULE
, 0xFFFF);
4898 si_pm4_set_reg(pm4
, R_028230_PA_SC_EDGERULE
,
4899 S_028230_ER_TRI(0xA) |
4900 S_028230_ER_POINT(0xA) |
4901 S_028230_ER_RECT(0xA) |
4902 /* Required by DX10_DIAMOND_TEST_ENA: */
4903 S_028230_ER_LINE_LR(0x1A) |
4904 S_028230_ER_LINE_RL(0x26) |
4905 S_028230_ER_LINE_TB(0xA) |
4906 S_028230_ER_LINE_BT(0xA));
4907 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
4908 si_pm4_set_reg(pm4
, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET
, 0);
4909 si_pm4_set_reg(pm4
, R_028820_PA_CL_NANINF_CNTL
, 0);
4910 si_pm4_set_reg(pm4
, R_028AC0_DB_SRESULTS_COMPARE_STATE0
, 0x0);
4911 si_pm4_set_reg(pm4
, R_028AC4_DB_SRESULTS_COMPARE_STATE1
, 0x0);
4912 si_pm4_set_reg(pm4
, R_028AC8_DB_PRELOAD_CONTROL
, 0x0);
4913 si_pm4_set_reg(pm4
, R_02800C_DB_RENDER_OVERRIDE
, 0);
4916 if (sctx
->b
.chip_class
>= GFX9
) {
4917 si_pm4_set_reg(pm4
, R_030920_VGT_MAX_VTX_INDX
, ~0);
4918 si_pm4_set_reg(pm4
, R_030924_VGT_MIN_VTX_INDX
, 0);
4919 si_pm4_set_reg(pm4
, R_030928_VGT_INDX_OFFSET
, 0);
4921 /* These registers, when written, also overwrite the CLEAR_STATE
4922 * context, so we can't rely on CLEAR_STATE setting them.
4923 * It would be an issue if there was another UMD changing them.
4925 si_pm4_set_reg(pm4
, R_028400_VGT_MAX_VTX_INDX
, ~0);
4926 si_pm4_set_reg(pm4
, R_028404_VGT_MIN_VTX_INDX
, 0);
4927 si_pm4_set_reg(pm4
, R_028408_VGT_INDX_OFFSET
, 0);
4930 if (sctx
->b
.chip_class
>= CIK
) {
4931 if (sctx
->b
.chip_class
>= GFX9
) {
4932 si_pm4_set_reg(pm4
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
, S_00B41C_CU_EN(0xffff));
4934 si_pm4_set_reg(pm4
, R_00B51C_SPI_SHADER_PGM_RSRC3_LS
, S_00B51C_CU_EN(0xffff));
4935 si_pm4_set_reg(pm4
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
, 0);
4936 si_pm4_set_reg(pm4
, R_00B31C_SPI_SHADER_PGM_RSRC3_ES
, S_00B31C_CU_EN(0xffff));
4938 /* If this is 0, Bonaire can hang even if GS isn't being used.
4939 * Other chips are unaffected. These are suboptimal values,
4940 * but we don't use on-chip GS.
4942 si_pm4_set_reg(pm4
, R_028A44_VGT_GS_ONCHIP_CNTL
,
4943 S_028A44_ES_VERTS_PER_SUBGRP(64) |
4944 S_028A44_GS_PRIMS_PER_SUBGRP(4));
4946 si_pm4_set_reg(pm4
, R_00B21C_SPI_SHADER_PGM_RSRC3_GS
, S_00B21C_CU_EN(0xffff));
4948 /* Compute LATE_ALLOC_VS.LIMIT. */
4949 unsigned num_cu_per_sh
= sscreen
->b
.info
.num_good_compute_units
/
4950 (sscreen
->b
.info
.max_se
*
4951 sscreen
->b
.info
.max_sh_per_se
);
4952 unsigned late_alloc_limit
; /* The limit is per SH. */
4954 if (sctx
->b
.family
== CHIP_KABINI
) {
4955 late_alloc_limit
= 0; /* Potential hang on Kabini. */
4956 } else if (num_cu_per_sh
<= 4) {
4957 /* Too few available compute units per SH. Disallowing
4958 * VS to run on one CU could hurt us more than late VS
4959 * allocation would help.
4961 * 2 is the highest safe number that allows us to keep
4964 late_alloc_limit
= 2;
4966 /* This is a good initial value, allowing 1 late_alloc
4967 * wave per SIMD on num_cu - 2.
4969 late_alloc_limit
= (num_cu_per_sh
- 2) * 4;
4971 /* The limit is 0-based, so 0 means 1. */
4972 assert(late_alloc_limit
> 0 && late_alloc_limit
<= 64);
4973 late_alloc_limit
-= 1;
4976 /* VS can't execute on one CU if the limit is > 2. */
4977 si_pm4_set_reg(pm4
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
,
4978 S_00B118_CU_EN(late_alloc_limit
> 2 ? 0xfffe : 0xffff));
4979 si_pm4_set_reg(pm4
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
,
4980 S_00B11C_LIMIT(late_alloc_limit
));
4981 si_pm4_set_reg(pm4
, R_00B01C_SPI_SHADER_PGM_RSRC3_PS
, S_00B01C_CU_EN(0xffff));
4984 if (sctx
->b
.chip_class
>= VI
) {
4985 unsigned vgt_tess_distribution
;
4987 si_pm4_set_reg(pm4
, R_028424_CB_DCC_CONTROL
,
4988 S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
4989 S_028424_OVERWRITE_COMBINER_WATERMARK(4));
4991 vgt_tess_distribution
=
4992 S_028B50_ACCUM_ISOLINE(32) |
4993 S_028B50_ACCUM_TRI(11) |
4994 S_028B50_ACCUM_QUAD(11) |
4995 S_028B50_DONUT_SPLIT(16);
4997 /* Testing with Unigine Heaven extreme tesselation yielded best results
4998 * with TRAP_SPLIT = 3.
5000 if (sctx
->b
.family
== CHIP_FIJI
||
5001 sctx
->b
.family
>= CHIP_POLARIS10
)
5002 vgt_tess_distribution
|= S_028B50_TRAP_SPLIT(3);
5004 si_pm4_set_reg(pm4
, R_028B50_VGT_TESS_DISTRIBUTION
, vgt_tess_distribution
);
5005 } else if (!has_clear_state
) {
5006 si_pm4_set_reg(pm4
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 14);
5007 si_pm4_set_reg(pm4
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 16);
5010 si_pm4_set_reg(pm4
, R_028080_TA_BC_BASE_ADDR
, border_color_va
>> 8);
5011 if (sctx
->b
.chip_class
>= CIK
)
5012 si_pm4_set_reg(pm4
, R_028084_TA_BC_BASE_ADDR_HI
, border_color_va
>> 40);
5013 si_pm4_add_bo(pm4
, sctx
->border_color_buffer
, RADEON_USAGE_READ
,
5014 RADEON_PRIO_BORDER_COLORS
);
5016 if (sctx
->b
.chip_class
>= GFX9
) {
5017 unsigned num_se
= sscreen
->b
.info
.max_se
;
5018 unsigned pc_lines
= 0;
5020 switch (sctx
->b
.family
) {
5031 si_pm4_set_reg(pm4
, R_028C48_PA_SC_BINNER_CNTL_1
,
5032 S_028C48_MAX_ALLOC_COUNT(MIN2(128, pc_lines
/ (4 * num_se
))) |
5033 S_028C48_MAX_PRIM_PER_BATCH(1023));
5034 si_pm4_set_reg(pm4
, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL
,
5035 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
5036 si_pm4_set_reg(pm4
, R_030968_VGT_INSTANCE_BASE_ID
, 0);
5039 si_pm4_upload_indirect_buffer(sctx
, pm4
);
5040 sctx
->init_config
= pm4
;