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 static uint32_t S_FIXED(float value
, uint32_t frac_bits
)
79 return value
* (1 << frac_bits
);
82 /* 12.4 fixed-point */
83 static unsigned si_pack_float_12p4(float x
)
86 x
>= 4096 ? 0xffff : x
* 16;
90 * Inferred framebuffer and blender state.
92 * CB_TARGET_MASK is emitted here to avoid a hang with dual source blending
93 * if there is not enough PS outputs.
95 static void si_emit_cb_render_state(struct si_context
*sctx
, struct r600_atom
*atom
)
97 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
98 struct si_state_blend
*blend
= sctx
->queued
.named
.blend
;
99 /* CB_COLORn_INFO.FORMAT=INVALID should disable unbound colorbuffers,
100 * but you never know. */
101 uint32_t cb_target_mask
= sctx
->framebuffer
.colorbuf_enabled_4bit
;
105 cb_target_mask
&= blend
->cb_target_mask
;
107 /* Avoid a hang that happens when dual source blending is enabled
108 * but there is not enough color outputs. This is undefined behavior,
109 * so disable color writes completely.
111 * Reproducible with Unigine Heaven 4.0 and drirc missing.
113 if (blend
&& blend
->dual_src_blend
&&
114 sctx
->ps_shader
.cso
&&
115 (sctx
->ps_shader
.cso
->info
.colors_written
& 0x3) != 0x3)
118 radeon_set_context_reg(cs
, R_028238_CB_TARGET_MASK
, cb_target_mask
);
120 /* GFX9: Flush DFSM when CB_TARGET_MASK changes.
121 * I think we don't have to do anything between IBs.
123 if (sctx
->b
.chip_class
>= GFX9
&&
124 sctx
->last_cb_target_mask
!= cb_target_mask
) {
125 sctx
->last_cb_target_mask
= cb_target_mask
;
127 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
128 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_DFSM
) | EVENT_INDEX(0));
131 /* RB+ register settings. */
132 if (sctx
->screen
->b
.rbplus_allowed
) {
133 unsigned spi_shader_col_format
=
134 sctx
->ps_shader
.cso
?
135 sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.spi_shader_col_format
: 0;
136 unsigned sx_ps_downconvert
= 0;
137 unsigned sx_blend_opt_epsilon
= 0;
138 unsigned sx_blend_opt_control
= 0;
140 for (i
= 0; i
< sctx
->framebuffer
.state
.nr_cbufs
; i
++) {
141 struct r600_surface
*surf
=
142 (struct r600_surface
*)sctx
->framebuffer
.state
.cbufs
[i
];
143 unsigned format
, swap
, spi_format
, colormask
;
144 bool has_alpha
, has_rgb
;
149 format
= G_028C70_FORMAT(surf
->cb_color_info
);
150 swap
= G_028C70_COMP_SWAP(surf
->cb_color_info
);
151 spi_format
= (spi_shader_col_format
>> (i
* 4)) & 0xf;
152 colormask
= (cb_target_mask
>> (i
* 4)) & 0xf;
154 /* Set if RGB and A are present. */
155 has_alpha
= !G_028C74_FORCE_DST_ALPHA_1(surf
->cb_color_attrib
);
157 if (format
== V_028C70_COLOR_8
||
158 format
== V_028C70_COLOR_16
||
159 format
== V_028C70_COLOR_32
)
160 has_rgb
= !has_alpha
;
164 /* Check the colormask and export format. */
165 if (!(colormask
& (PIPE_MASK_RGBA
& ~PIPE_MASK_A
)))
167 if (!(colormask
& PIPE_MASK_A
))
170 if (spi_format
== V_028714_SPI_SHADER_ZERO
) {
175 /* Disable value checking for disabled channels. */
177 sx_blend_opt_control
|= S_02875C_MRT0_COLOR_OPT_DISABLE(1) << (i
* 4);
179 sx_blend_opt_control
|= S_02875C_MRT0_ALPHA_OPT_DISABLE(1) << (i
* 4);
181 /* Enable down-conversion for 32bpp and smaller formats. */
183 case V_028C70_COLOR_8
:
184 case V_028C70_COLOR_8_8
:
185 case V_028C70_COLOR_8_8_8_8
:
186 /* For 1 and 2-channel formats, use the superset thereof. */
187 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
||
188 spi_format
== V_028714_SPI_SHADER_UINT16_ABGR
||
189 spi_format
== V_028714_SPI_SHADER_SINT16_ABGR
) {
190 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_8_8_8_8
<< (i
* 4);
191 sx_blend_opt_epsilon
|= V_028758_8BIT_FORMAT
<< (i
* 4);
195 case V_028C70_COLOR_5_6_5
:
196 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
197 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_5_6_5
<< (i
* 4);
198 sx_blend_opt_epsilon
|= V_028758_6BIT_FORMAT
<< (i
* 4);
202 case V_028C70_COLOR_1_5_5_5
:
203 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
204 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_1_5_5_5
<< (i
* 4);
205 sx_blend_opt_epsilon
|= V_028758_5BIT_FORMAT
<< (i
* 4);
209 case V_028C70_COLOR_4_4_4_4
:
210 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
211 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_4_4_4_4
<< (i
* 4);
212 sx_blend_opt_epsilon
|= V_028758_4BIT_FORMAT
<< (i
* 4);
216 case V_028C70_COLOR_32
:
217 if (swap
== V_0280A0_SWAP_STD
&&
218 spi_format
== V_028714_SPI_SHADER_32_R
)
219 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_32_R
<< (i
* 4);
220 else if (swap
== V_0280A0_SWAP_ALT_REV
&&
221 spi_format
== V_028714_SPI_SHADER_32_AR
)
222 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_32_A
<< (i
* 4);
225 case V_028C70_COLOR_16
:
226 case V_028C70_COLOR_16_16
:
227 /* For 1-channel formats, use the superset thereof. */
228 if (spi_format
== V_028714_SPI_SHADER_UNORM16_ABGR
||
229 spi_format
== V_028714_SPI_SHADER_SNORM16_ABGR
||
230 spi_format
== V_028714_SPI_SHADER_UINT16_ABGR
||
231 spi_format
== V_028714_SPI_SHADER_SINT16_ABGR
) {
232 if (swap
== V_0280A0_SWAP_STD
||
233 swap
== V_0280A0_SWAP_STD_REV
)
234 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_16_16_GR
<< (i
* 4);
236 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_16_16_AR
<< (i
* 4);
240 case V_028C70_COLOR_10_11_11
:
241 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
242 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_10_11_11
<< (i
* 4);
243 sx_blend_opt_epsilon
|= V_028758_11BIT_FORMAT
<< (i
* 4);
247 case V_028C70_COLOR_2_10_10_10
:
248 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
249 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_2_10_10_10
<< (i
* 4);
250 sx_blend_opt_epsilon
|= V_028758_10BIT_FORMAT
<< (i
* 4);
256 radeon_set_context_reg_seq(cs
, R_028754_SX_PS_DOWNCONVERT
, 3);
257 radeon_emit(cs
, sx_ps_downconvert
); /* R_028754_SX_PS_DOWNCONVERT */
258 radeon_emit(cs
, sx_blend_opt_epsilon
); /* R_028758_SX_BLEND_OPT_EPSILON */
259 radeon_emit(cs
, sx_blend_opt_control
); /* R_02875C_SX_BLEND_OPT_CONTROL */
260 } else if (sctx
->screen
->b
.has_rbplus
) {
261 radeon_set_context_reg_seq(cs
, R_028754_SX_PS_DOWNCONVERT
, 3);
262 radeon_emit(cs
, 0); /* R_028754_SX_PS_DOWNCONVERT */
263 radeon_emit(cs
, 0); /* R_028758_SX_BLEND_OPT_EPSILON */
264 radeon_emit(cs
, 0); /* R_02875C_SX_BLEND_OPT_CONTROL */
272 static uint32_t si_translate_blend_function(int blend_func
)
274 switch (blend_func
) {
276 return V_028780_COMB_DST_PLUS_SRC
;
277 case PIPE_BLEND_SUBTRACT
:
278 return V_028780_COMB_SRC_MINUS_DST
;
279 case PIPE_BLEND_REVERSE_SUBTRACT
:
280 return V_028780_COMB_DST_MINUS_SRC
;
282 return V_028780_COMB_MIN_DST_SRC
;
284 return V_028780_COMB_MAX_DST_SRC
;
286 R600_ERR("Unknown blend function %d\n", blend_func
);
293 static uint32_t si_translate_blend_factor(int blend_fact
)
295 switch (blend_fact
) {
296 case PIPE_BLENDFACTOR_ONE
:
297 return V_028780_BLEND_ONE
;
298 case PIPE_BLENDFACTOR_SRC_COLOR
:
299 return V_028780_BLEND_SRC_COLOR
;
300 case PIPE_BLENDFACTOR_SRC_ALPHA
:
301 return V_028780_BLEND_SRC_ALPHA
;
302 case PIPE_BLENDFACTOR_DST_ALPHA
:
303 return V_028780_BLEND_DST_ALPHA
;
304 case PIPE_BLENDFACTOR_DST_COLOR
:
305 return V_028780_BLEND_DST_COLOR
;
306 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
307 return V_028780_BLEND_SRC_ALPHA_SATURATE
;
308 case PIPE_BLENDFACTOR_CONST_COLOR
:
309 return V_028780_BLEND_CONSTANT_COLOR
;
310 case PIPE_BLENDFACTOR_CONST_ALPHA
:
311 return V_028780_BLEND_CONSTANT_ALPHA
;
312 case PIPE_BLENDFACTOR_ZERO
:
313 return V_028780_BLEND_ZERO
;
314 case PIPE_BLENDFACTOR_INV_SRC_COLOR
:
315 return V_028780_BLEND_ONE_MINUS_SRC_COLOR
;
316 case PIPE_BLENDFACTOR_INV_SRC_ALPHA
:
317 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA
;
318 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
319 return V_028780_BLEND_ONE_MINUS_DST_ALPHA
;
320 case PIPE_BLENDFACTOR_INV_DST_COLOR
:
321 return V_028780_BLEND_ONE_MINUS_DST_COLOR
;
322 case PIPE_BLENDFACTOR_INV_CONST_COLOR
:
323 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR
;
324 case PIPE_BLENDFACTOR_INV_CONST_ALPHA
:
325 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA
;
326 case PIPE_BLENDFACTOR_SRC1_COLOR
:
327 return V_028780_BLEND_SRC1_COLOR
;
328 case PIPE_BLENDFACTOR_SRC1_ALPHA
:
329 return V_028780_BLEND_SRC1_ALPHA
;
330 case PIPE_BLENDFACTOR_INV_SRC1_COLOR
:
331 return V_028780_BLEND_INV_SRC1_COLOR
;
332 case PIPE_BLENDFACTOR_INV_SRC1_ALPHA
:
333 return V_028780_BLEND_INV_SRC1_ALPHA
;
335 R600_ERR("Bad blend factor %d not supported!\n", blend_fact
);
342 static uint32_t si_translate_blend_opt_function(int blend_func
)
344 switch (blend_func
) {
346 return V_028760_OPT_COMB_ADD
;
347 case PIPE_BLEND_SUBTRACT
:
348 return V_028760_OPT_COMB_SUBTRACT
;
349 case PIPE_BLEND_REVERSE_SUBTRACT
:
350 return V_028760_OPT_COMB_REVSUBTRACT
;
352 return V_028760_OPT_COMB_MIN
;
354 return V_028760_OPT_COMB_MAX
;
356 return V_028760_OPT_COMB_BLEND_DISABLED
;
360 static uint32_t si_translate_blend_opt_factor(int blend_fact
, bool is_alpha
)
362 switch (blend_fact
) {
363 case PIPE_BLENDFACTOR_ZERO
:
364 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL
;
365 case PIPE_BLENDFACTOR_ONE
:
366 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
;
367 case PIPE_BLENDFACTOR_SRC_COLOR
:
368 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
369 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0
;
370 case PIPE_BLENDFACTOR_INV_SRC_COLOR
:
371 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
372 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1
;
373 case PIPE_BLENDFACTOR_SRC_ALPHA
:
374 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
;
375 case PIPE_BLENDFACTOR_INV_SRC_ALPHA
:
376 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
;
377 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
378 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
379 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0
;
381 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
386 * Get rid of DST in the blend factors by commuting the operands:
387 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
389 static void si_blend_remove_dst(unsigned *func
, unsigned *src_factor
,
390 unsigned *dst_factor
, unsigned expected_dst
,
391 unsigned replacement_src
)
393 if (*src_factor
== expected_dst
&&
394 *dst_factor
== PIPE_BLENDFACTOR_ZERO
) {
395 *src_factor
= PIPE_BLENDFACTOR_ZERO
;
396 *dst_factor
= replacement_src
;
398 /* Commuting the operands requires reversing subtractions. */
399 if (*func
== PIPE_BLEND_SUBTRACT
)
400 *func
= PIPE_BLEND_REVERSE_SUBTRACT
;
401 else if (*func
== PIPE_BLEND_REVERSE_SUBTRACT
)
402 *func
= PIPE_BLEND_SUBTRACT
;
406 static bool si_blend_factor_uses_dst(unsigned factor
)
408 return factor
== PIPE_BLENDFACTOR_DST_COLOR
||
409 factor
== PIPE_BLENDFACTOR_DST_ALPHA
||
410 factor
== PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
||
411 factor
== PIPE_BLENDFACTOR_INV_DST_ALPHA
||
412 factor
== PIPE_BLENDFACTOR_INV_DST_COLOR
;
415 static void *si_create_blend_state_mode(struct pipe_context
*ctx
,
416 const struct pipe_blend_state
*state
,
419 struct si_context
*sctx
= (struct si_context
*)ctx
;
420 struct si_state_blend
*blend
= CALLOC_STRUCT(si_state_blend
);
421 struct si_pm4_state
*pm4
= &blend
->pm4
;
422 uint32_t sx_mrt_blend_opt
[8] = {0};
423 uint32_t color_control
= 0;
428 blend
->alpha_to_coverage
= state
->alpha_to_coverage
;
429 blend
->alpha_to_one
= state
->alpha_to_one
;
430 blend
->dual_src_blend
= util_blend_state_is_dual(state
, 0);
432 if (state
->logicop_enable
) {
433 color_control
|= S_028808_ROP3(state
->logicop_func
| (state
->logicop_func
<< 4));
435 color_control
|= S_028808_ROP3(0xcc);
438 si_pm4_set_reg(pm4
, R_028B70_DB_ALPHA_TO_MASK
,
439 S_028B70_ALPHA_TO_MASK_ENABLE(state
->alpha_to_coverage
) |
440 S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
441 S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
442 S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
443 S_028B70_ALPHA_TO_MASK_OFFSET3(2));
445 if (state
->alpha_to_coverage
)
446 blend
->need_src_alpha_4bit
|= 0xf;
448 blend
->cb_target_mask
= 0;
449 for (int i
= 0; i
< 8; i
++) {
450 /* state->rt entries > 0 only written if independent blending */
451 const int j
= state
->independent_blend_enable
? i
: 0;
453 unsigned eqRGB
= state
->rt
[j
].rgb_func
;
454 unsigned srcRGB
= state
->rt
[j
].rgb_src_factor
;
455 unsigned dstRGB
= state
->rt
[j
].rgb_dst_factor
;
456 unsigned eqA
= state
->rt
[j
].alpha_func
;
457 unsigned srcA
= state
->rt
[j
].alpha_src_factor
;
458 unsigned dstA
= state
->rt
[j
].alpha_dst_factor
;
460 unsigned srcRGB_opt
, dstRGB_opt
, srcA_opt
, dstA_opt
;
461 unsigned blend_cntl
= 0;
463 sx_mrt_blend_opt
[i
] =
464 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED
) |
465 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED
);
467 /* Only set dual source blending for MRT0 to avoid a hang. */
468 if (i
>= 1 && blend
->dual_src_blend
) {
469 /* Vulkan does this for dual source blending. */
471 blend_cntl
|= S_028780_ENABLE(1);
473 si_pm4_set_reg(pm4
, R_028780_CB_BLEND0_CONTROL
+ i
* 4, blend_cntl
);
477 /* Only addition and subtraction equations are supported with
478 * dual source blending.
480 if (blend
->dual_src_blend
&&
481 (eqRGB
== PIPE_BLEND_MIN
|| eqRGB
== PIPE_BLEND_MAX
||
482 eqA
== PIPE_BLEND_MIN
|| eqA
== PIPE_BLEND_MAX
)) {
483 assert(!"Unsupported equation for dual source blending");
484 si_pm4_set_reg(pm4
, R_028780_CB_BLEND0_CONTROL
+ i
* 4, blend_cntl
);
488 /* cb_render_state will disable unused ones */
489 blend
->cb_target_mask
|= (unsigned)state
->rt
[j
].colormask
<< (4 * i
);
491 if (!state
->rt
[j
].colormask
|| !state
->rt
[j
].blend_enable
) {
492 si_pm4_set_reg(pm4
, R_028780_CB_BLEND0_CONTROL
+ i
* 4, blend_cntl
);
496 /* Blending optimizations for RB+.
497 * These transformations don't change the behavior.
499 * First, get rid of DST in the blend factors:
500 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
502 si_blend_remove_dst(&eqRGB
, &srcRGB
, &dstRGB
,
503 PIPE_BLENDFACTOR_DST_COLOR
,
504 PIPE_BLENDFACTOR_SRC_COLOR
);
505 si_blend_remove_dst(&eqA
, &srcA
, &dstA
,
506 PIPE_BLENDFACTOR_DST_COLOR
,
507 PIPE_BLENDFACTOR_SRC_COLOR
);
508 si_blend_remove_dst(&eqA
, &srcA
, &dstA
,
509 PIPE_BLENDFACTOR_DST_ALPHA
,
510 PIPE_BLENDFACTOR_SRC_ALPHA
);
512 /* Look up the ideal settings from tables. */
513 srcRGB_opt
= si_translate_blend_opt_factor(srcRGB
, false);
514 dstRGB_opt
= si_translate_blend_opt_factor(dstRGB
, false);
515 srcA_opt
= si_translate_blend_opt_factor(srcA
, true);
516 dstA_opt
= si_translate_blend_opt_factor(dstA
, true);
518 /* Handle interdependencies. */
519 if (si_blend_factor_uses_dst(srcRGB
))
520 dstRGB_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
521 if (si_blend_factor_uses_dst(srcA
))
522 dstA_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
524 if (srcRGB
== PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
&&
525 (dstRGB
== PIPE_BLENDFACTOR_ZERO
||
526 dstRGB
== PIPE_BLENDFACTOR_SRC_ALPHA
||
527 dstRGB
== PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
))
528 dstRGB_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0
;
530 /* Set the final value. */
531 sx_mrt_blend_opt
[i
] =
532 S_028760_COLOR_SRC_OPT(srcRGB_opt
) |
533 S_028760_COLOR_DST_OPT(dstRGB_opt
) |
534 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB
)) |
535 S_028760_ALPHA_SRC_OPT(srcA_opt
) |
536 S_028760_ALPHA_DST_OPT(dstA_opt
) |
537 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA
));
539 /* Set blend state. */
540 blend_cntl
|= S_028780_ENABLE(1);
541 blend_cntl
|= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB
));
542 blend_cntl
|= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB
));
543 blend_cntl
|= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB
));
545 if (srcA
!= srcRGB
|| dstA
!= dstRGB
|| eqA
!= eqRGB
) {
546 blend_cntl
|= S_028780_SEPARATE_ALPHA_BLEND(1);
547 blend_cntl
|= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA
));
548 blend_cntl
|= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA
));
549 blend_cntl
|= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA
));
551 si_pm4_set_reg(pm4
, R_028780_CB_BLEND0_CONTROL
+ i
* 4, blend_cntl
);
553 blend
->blend_enable_4bit
|= 0xfu
<< (i
* 4);
555 /* This is only important for formats without alpha. */
556 if (srcRGB
== PIPE_BLENDFACTOR_SRC_ALPHA
||
557 dstRGB
== PIPE_BLENDFACTOR_SRC_ALPHA
||
558 srcRGB
== PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
||
559 dstRGB
== PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
||
560 srcRGB
== PIPE_BLENDFACTOR_INV_SRC_ALPHA
||
561 dstRGB
== PIPE_BLENDFACTOR_INV_SRC_ALPHA
)
562 blend
->need_src_alpha_4bit
|= 0xfu
<< (i
* 4);
565 if (blend
->cb_target_mask
) {
566 color_control
|= S_028808_MODE(mode
);
568 color_control
|= S_028808_MODE(V_028808_CB_DISABLE
);
571 if (sctx
->screen
->b
.has_rbplus
) {
572 /* Disable RB+ blend optimizations for dual source blending.
575 if (blend
->dual_src_blend
) {
576 for (int i
= 0; i
< 8; i
++) {
577 sx_mrt_blend_opt
[i
] =
578 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_NONE
) |
579 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_NONE
);
583 for (int i
= 0; i
< 8; i
++)
584 si_pm4_set_reg(pm4
, R_028760_SX_MRT0_BLEND_OPT
+ i
* 4,
585 sx_mrt_blend_opt
[i
]);
587 /* RB+ doesn't work with dual source blending, logic op, and RESOLVE. */
588 if (blend
->dual_src_blend
|| state
->logicop_enable
||
589 mode
== V_028808_CB_RESOLVE
)
590 color_control
|= S_028808_DISABLE_DUAL_QUAD(1);
593 si_pm4_set_reg(pm4
, R_028808_CB_COLOR_CONTROL
, color_control
);
597 static void *si_create_blend_state(struct pipe_context
*ctx
,
598 const struct pipe_blend_state
*state
)
600 return si_create_blend_state_mode(ctx
, state
, V_028808_CB_NORMAL
);
603 static void si_bind_blend_state(struct pipe_context
*ctx
, void *state
)
605 struct si_context
*sctx
= (struct si_context
*)ctx
;
606 si_pm4_bind_state(sctx
, blend
, (struct si_state_blend
*)state
);
607 si_mark_atom_dirty(sctx
, &sctx
->cb_render_state
);
608 sctx
->do_update_shaders
= true;
611 static void si_delete_blend_state(struct pipe_context
*ctx
, void *state
)
613 struct si_context
*sctx
= (struct si_context
*)ctx
;
614 si_pm4_delete_state(sctx
, blend
, (struct si_state_blend
*)state
);
617 static void si_set_blend_color(struct pipe_context
*ctx
,
618 const struct pipe_blend_color
*state
)
620 struct si_context
*sctx
= (struct si_context
*)ctx
;
622 if (memcmp(&sctx
->blend_color
.state
, state
, sizeof(*state
)) == 0)
625 sctx
->blend_color
.state
= *state
;
626 si_mark_atom_dirty(sctx
, &sctx
->blend_color
.atom
);
629 static void si_emit_blend_color(struct si_context
*sctx
, struct r600_atom
*atom
)
631 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
633 radeon_set_context_reg_seq(cs
, R_028414_CB_BLEND_RED
, 4);
634 radeon_emit_array(cs
, (uint32_t*)sctx
->blend_color
.state
.color
, 4);
641 static void si_set_clip_state(struct pipe_context
*ctx
,
642 const struct pipe_clip_state
*state
)
644 struct si_context
*sctx
= (struct si_context
*)ctx
;
645 struct pipe_constant_buffer cb
;
647 if (memcmp(&sctx
->clip_state
.state
, state
, sizeof(*state
)) == 0)
650 sctx
->clip_state
.state
= *state
;
651 si_mark_atom_dirty(sctx
, &sctx
->clip_state
.atom
);
654 cb
.user_buffer
= state
->ucp
;
655 cb
.buffer_offset
= 0;
656 cb
.buffer_size
= 4*4*8;
657 si_set_rw_buffer(sctx
, SI_VS_CONST_CLIP_PLANES
, &cb
);
658 pipe_resource_reference(&cb
.buffer
, NULL
);
661 static void si_emit_clip_state(struct si_context
*sctx
, struct r600_atom
*atom
)
663 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
665 radeon_set_context_reg_seq(cs
, R_0285BC_PA_CL_UCP_0_X
, 6*4);
666 radeon_emit_array(cs
, (uint32_t*)sctx
->clip_state
.state
.ucp
, 6*4);
669 #define SIX_BITS 0x3F
671 static void si_emit_clip_regs(struct si_context
*sctx
, struct r600_atom
*atom
)
673 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
674 struct si_shader
*vs
= si_get_vs_state(sctx
);
675 struct tgsi_shader_info
*info
= si_get_vs_info(sctx
);
676 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
677 unsigned window_space
=
678 info
->properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
];
679 unsigned clipdist_mask
=
680 info
->writes_clipvertex
? SIX_BITS
: info
->clipdist_writemask
;
681 unsigned ucp_mask
= clipdist_mask
? 0 : rs
->clip_plane_enable
& SIX_BITS
;
682 unsigned culldist_mask
= info
->culldist_writemask
<< info
->num_written_clipdistance
;
686 if (vs
->key
.opt
.hw_vs
.clip_disable
) {
687 assert(!info
->culldist_writemask
);
691 total_mask
= clipdist_mask
| culldist_mask
;
693 /* Clip distances on points have no effect, so need to be implemented
694 * as cull distances. This applies for the clipvertex case as well.
696 * Setting this for primitives other than points should have no adverse
699 clipdist_mask
&= rs
->clip_plane_enable
;
700 culldist_mask
|= clipdist_mask
;
702 misc_vec_ena
= info
->writes_psize
|| info
->writes_edgeflag
||
703 info
->writes_layer
|| info
->writes_viewport_index
;
705 radeon_set_context_reg(cs
, R_02881C_PA_CL_VS_OUT_CNTL
,
706 S_02881C_USE_VTX_POINT_SIZE(info
->writes_psize
) |
707 S_02881C_USE_VTX_EDGE_FLAG(info
->writes_edgeflag
) |
708 S_02881C_USE_VTX_RENDER_TARGET_INDX(info
->writes_layer
) |
709 S_02881C_USE_VTX_VIEWPORT_INDX(info
->writes_viewport_index
) |
710 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask
& 0x0F) != 0) |
711 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask
& 0xF0) != 0) |
712 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena
) |
713 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena
) |
714 clipdist_mask
| (culldist_mask
<< 8));
715 radeon_set_context_reg(cs
, R_028810_PA_CL_CLIP_CNTL
,
716 rs
->pa_cl_clip_cntl
|
718 S_028810_CLIP_DISABLE(window_space
));
720 /* reuse needs to be set off if we write oViewport */
721 radeon_set_context_reg(cs
, R_028AB4_VGT_REUSE_OFF
,
722 S_028AB4_REUSE_OFF(info
->writes_viewport_index
));
726 * inferred state between framebuffer and rasterizer
728 static void si_update_poly_offset_state(struct si_context
*sctx
)
730 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
732 if (!rs
|| !rs
->uses_poly_offset
|| !sctx
->framebuffer
.state
.zsbuf
) {
733 si_pm4_bind_state(sctx
, poly_offset
, NULL
);
737 /* Use the user format, not db_render_format, so that the polygon
738 * offset behaves as expected by applications.
740 switch (sctx
->framebuffer
.state
.zsbuf
->texture
->format
) {
741 case PIPE_FORMAT_Z16_UNORM
:
742 si_pm4_bind_state(sctx
, poly_offset
, &rs
->pm4_poly_offset
[0]);
744 default: /* 24-bit */
745 si_pm4_bind_state(sctx
, poly_offset
, &rs
->pm4_poly_offset
[1]);
747 case PIPE_FORMAT_Z32_FLOAT
:
748 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
749 si_pm4_bind_state(sctx
, poly_offset
, &rs
->pm4_poly_offset
[2]);
758 static uint32_t si_translate_fill(uint32_t func
)
761 case PIPE_POLYGON_MODE_FILL
:
762 return V_028814_X_DRAW_TRIANGLES
;
763 case PIPE_POLYGON_MODE_LINE
:
764 return V_028814_X_DRAW_LINES
;
765 case PIPE_POLYGON_MODE_POINT
:
766 return V_028814_X_DRAW_POINTS
;
769 return V_028814_X_DRAW_POINTS
;
773 static void *si_create_rs_state(struct pipe_context
*ctx
,
774 const struct pipe_rasterizer_state
*state
)
776 struct si_screen
*sscreen
= ((struct si_context
*)ctx
)->screen
;
777 struct si_state_rasterizer
*rs
= CALLOC_STRUCT(si_state_rasterizer
);
778 struct si_pm4_state
*pm4
= &rs
->pm4
;
780 float psize_min
, psize_max
;
786 rs
->scissor_enable
= state
->scissor
;
787 rs
->clip_halfz
= state
->clip_halfz
;
788 rs
->two_side
= state
->light_twoside
;
789 rs
->multisample_enable
= state
->multisample
;
790 rs
->force_persample_interp
= state
->force_persample_interp
;
791 rs
->clip_plane_enable
= state
->clip_plane_enable
;
792 rs
->line_stipple_enable
= state
->line_stipple_enable
;
793 rs
->poly_stipple_enable
= state
->poly_stipple_enable
;
794 rs
->line_smooth
= state
->line_smooth
;
795 rs
->poly_smooth
= state
->poly_smooth
;
796 rs
->uses_poly_offset
= state
->offset_point
|| state
->offset_line
||
798 rs
->clamp_fragment_color
= state
->clamp_fragment_color
;
799 rs
->flatshade
= state
->flatshade
;
800 rs
->sprite_coord_enable
= state
->sprite_coord_enable
;
801 rs
->rasterizer_discard
= state
->rasterizer_discard
;
802 rs
->pa_sc_line_stipple
= state
->line_stipple_enable
?
803 S_028A0C_LINE_PATTERN(state
->line_stipple_pattern
) |
804 S_028A0C_REPEAT_COUNT(state
->line_stipple_factor
) : 0;
805 rs
->pa_cl_clip_cntl
=
806 S_028810_DX_CLIP_SPACE_DEF(state
->clip_halfz
) |
807 S_028810_ZCLIP_NEAR_DISABLE(!state
->depth_clip
) |
808 S_028810_ZCLIP_FAR_DISABLE(!state
->depth_clip
) |
809 S_028810_DX_RASTERIZATION_KILL(state
->rasterizer_discard
) |
810 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
812 si_pm4_set_reg(pm4
, R_0286D4_SPI_INTERP_CONTROL_0
,
813 S_0286D4_FLAT_SHADE_ENA(1) |
814 S_0286D4_PNT_SPRITE_ENA(1) |
815 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S
) |
816 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T
) |
817 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0
) |
818 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1
) |
819 S_0286D4_PNT_SPRITE_TOP_1(state
->sprite_coord_mode
!= PIPE_SPRITE_COORD_UPPER_LEFT
));
821 /* point size 12.4 fixed point */
822 tmp
= (unsigned)(state
->point_size
* 8.0);
823 si_pm4_set_reg(pm4
, R_028A00_PA_SU_POINT_SIZE
, S_028A00_HEIGHT(tmp
) | S_028A00_WIDTH(tmp
));
825 if (state
->point_size_per_vertex
) {
826 psize_min
= util_get_min_point_size(state
);
829 /* Force the point size to be as if the vertex output was disabled. */
830 psize_min
= state
->point_size
;
831 psize_max
= state
->point_size
;
833 /* Divide by two, because 0.5 = 1 pixel. */
834 si_pm4_set_reg(pm4
, R_028A04_PA_SU_POINT_MINMAX
,
835 S_028A04_MIN_SIZE(si_pack_float_12p4(psize_min
/2)) |
836 S_028A04_MAX_SIZE(si_pack_float_12p4(psize_max
/2)));
838 tmp
= (unsigned)state
->line_width
* 8;
839 si_pm4_set_reg(pm4
, R_028A08_PA_SU_LINE_CNTL
, S_028A08_WIDTH(tmp
));
840 si_pm4_set_reg(pm4
, R_028A48_PA_SC_MODE_CNTL_0
,
841 S_028A48_LINE_STIPPLE_ENABLE(state
->line_stipple_enable
) |
842 S_028A48_MSAA_ENABLE(state
->multisample
||
843 state
->poly_smooth
||
844 state
->line_smooth
) |
845 S_028A48_VPORT_SCISSOR_ENABLE(1) |
846 S_028A48_ALTERNATE_RBS_PER_TILE(sscreen
->b
.chip_class
>= GFX9
));
848 si_pm4_set_reg(pm4
, R_028BE4_PA_SU_VTX_CNTL
,
849 S_028BE4_PIX_CENTER(state
->half_pixel_center
) |
850 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH
));
852 si_pm4_set_reg(pm4
, R_028B7C_PA_SU_POLY_OFFSET_CLAMP
, fui(state
->offset_clamp
));
853 si_pm4_set_reg(pm4
, R_028814_PA_SU_SC_MODE_CNTL
,
854 S_028814_PROVOKING_VTX_LAST(!state
->flatshade_first
) |
855 S_028814_CULL_FRONT((state
->cull_face
& PIPE_FACE_FRONT
) ? 1 : 0) |
856 S_028814_CULL_BACK((state
->cull_face
& PIPE_FACE_BACK
) ? 1 : 0) |
857 S_028814_FACE(!state
->front_ccw
) |
858 S_028814_POLY_OFFSET_FRONT_ENABLE(util_get_offset(state
, state
->fill_front
)) |
859 S_028814_POLY_OFFSET_BACK_ENABLE(util_get_offset(state
, state
->fill_back
)) |
860 S_028814_POLY_OFFSET_PARA_ENABLE(state
->offset_point
|| state
->offset_line
) |
861 S_028814_POLY_MODE(state
->fill_front
!= PIPE_POLYGON_MODE_FILL
||
862 state
->fill_back
!= PIPE_POLYGON_MODE_FILL
) |
863 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(state
->fill_front
)) |
864 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(state
->fill_back
)));
865 si_pm4_set_reg(pm4
, R_00B130_SPI_SHADER_USER_DATA_VS_0
+
866 SI_SGPR_VS_STATE_BITS
* 4, state
->clamp_vertex_color
);
868 /* Precalculate polygon offset states for 16-bit, 24-bit, and 32-bit zbuffers. */
869 for (i
= 0; i
< 3; i
++) {
870 struct si_pm4_state
*pm4
= &rs
->pm4_poly_offset
[i
];
871 float offset_units
= state
->offset_units
;
872 float offset_scale
= state
->offset_scale
* 16.0f
;
873 uint32_t pa_su_poly_offset_db_fmt_cntl
= 0;
875 if (!state
->offset_units_unscaled
) {
877 case 0: /* 16-bit zbuffer */
878 offset_units
*= 4.0f
;
879 pa_su_poly_offset_db_fmt_cntl
=
880 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
882 case 1: /* 24-bit zbuffer */
883 offset_units
*= 2.0f
;
884 pa_su_poly_offset_db_fmt_cntl
=
885 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
887 case 2: /* 32-bit zbuffer */
888 offset_units
*= 1.0f
;
889 pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
890 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
895 si_pm4_set_reg(pm4
, R_028B80_PA_SU_POLY_OFFSET_FRONT_SCALE
,
897 si_pm4_set_reg(pm4
, R_028B84_PA_SU_POLY_OFFSET_FRONT_OFFSET
,
899 si_pm4_set_reg(pm4
, R_028B88_PA_SU_POLY_OFFSET_BACK_SCALE
,
901 si_pm4_set_reg(pm4
, R_028B8C_PA_SU_POLY_OFFSET_BACK_OFFSET
,
903 si_pm4_set_reg(pm4
, R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL
,
904 pa_su_poly_offset_db_fmt_cntl
);
910 static void si_bind_rs_state(struct pipe_context
*ctx
, void *state
)
912 struct si_context
*sctx
= (struct si_context
*)ctx
;
913 struct si_state_rasterizer
*old_rs
=
914 (struct si_state_rasterizer
*)sctx
->queued
.named
.rasterizer
;
915 struct si_state_rasterizer
*rs
= (struct si_state_rasterizer
*)state
;
920 if (!old_rs
|| old_rs
->multisample_enable
!= rs
->multisample_enable
) {
921 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
923 /* Update the small primitive filter workaround if necessary. */
924 if (sctx
->b
.family
>= CHIP_POLARIS10
&&
925 sctx
->framebuffer
.nr_samples
> 1)
926 si_mark_atom_dirty(sctx
, &sctx
->msaa_sample_locs
.atom
);
929 r600_viewport_set_rast_deps(&sctx
->b
, rs
->scissor_enable
, rs
->clip_halfz
);
931 si_pm4_bind_state(sctx
, rasterizer
, rs
);
932 si_update_poly_offset_state(sctx
);
934 si_mark_atom_dirty(sctx
, &sctx
->clip_regs
);
935 sctx
->ia_multi_vgt_param_key
.u
.line_stipple_enabled
=
936 rs
->line_stipple_enable
;
937 sctx
->do_update_shaders
= true;
940 static void si_delete_rs_state(struct pipe_context
*ctx
, void *state
)
942 struct si_context
*sctx
= (struct si_context
*)ctx
;
944 if (sctx
->queued
.named
.rasterizer
== state
)
945 si_pm4_bind_state(sctx
, poly_offset
, NULL
);
946 si_pm4_delete_state(sctx
, rasterizer
, (struct si_state_rasterizer
*)state
);
950 * infeered state between dsa and stencil ref
952 static void si_emit_stencil_ref(struct si_context
*sctx
, struct r600_atom
*atom
)
954 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
955 struct pipe_stencil_ref
*ref
= &sctx
->stencil_ref
.state
;
956 struct si_dsa_stencil_ref_part
*dsa
= &sctx
->stencil_ref
.dsa_part
;
958 radeon_set_context_reg_seq(cs
, R_028430_DB_STENCILREFMASK
, 2);
959 radeon_emit(cs
, S_028430_STENCILTESTVAL(ref
->ref_value
[0]) |
960 S_028430_STENCILMASK(dsa
->valuemask
[0]) |
961 S_028430_STENCILWRITEMASK(dsa
->writemask
[0]) |
962 S_028430_STENCILOPVAL(1));
963 radeon_emit(cs
, S_028434_STENCILTESTVAL_BF(ref
->ref_value
[1]) |
964 S_028434_STENCILMASK_BF(dsa
->valuemask
[1]) |
965 S_028434_STENCILWRITEMASK_BF(dsa
->writemask
[1]) |
966 S_028434_STENCILOPVAL_BF(1));
969 static void si_set_stencil_ref(struct pipe_context
*ctx
,
970 const struct pipe_stencil_ref
*state
)
972 struct si_context
*sctx
= (struct si_context
*)ctx
;
974 if (memcmp(&sctx
->stencil_ref
.state
, state
, sizeof(*state
)) == 0)
977 sctx
->stencil_ref
.state
= *state
;
978 si_mark_atom_dirty(sctx
, &sctx
->stencil_ref
.atom
);
986 static uint32_t si_translate_stencil_op(int s_op
)
989 case PIPE_STENCIL_OP_KEEP
:
990 return V_02842C_STENCIL_KEEP
;
991 case PIPE_STENCIL_OP_ZERO
:
992 return V_02842C_STENCIL_ZERO
;
993 case PIPE_STENCIL_OP_REPLACE
:
994 return V_02842C_STENCIL_REPLACE_TEST
;
995 case PIPE_STENCIL_OP_INCR
:
996 return V_02842C_STENCIL_ADD_CLAMP
;
997 case PIPE_STENCIL_OP_DECR
:
998 return V_02842C_STENCIL_SUB_CLAMP
;
999 case PIPE_STENCIL_OP_INCR_WRAP
:
1000 return V_02842C_STENCIL_ADD_WRAP
;
1001 case PIPE_STENCIL_OP_DECR_WRAP
:
1002 return V_02842C_STENCIL_SUB_WRAP
;
1003 case PIPE_STENCIL_OP_INVERT
:
1004 return V_02842C_STENCIL_INVERT
;
1006 R600_ERR("Unknown stencil op %d", s_op
);
1013 static void *si_create_dsa_state(struct pipe_context
*ctx
,
1014 const struct pipe_depth_stencil_alpha_state
*state
)
1016 struct si_state_dsa
*dsa
= CALLOC_STRUCT(si_state_dsa
);
1017 struct si_pm4_state
*pm4
= &dsa
->pm4
;
1018 unsigned db_depth_control
;
1019 uint32_t db_stencil_control
= 0;
1025 dsa
->stencil_ref
.valuemask
[0] = state
->stencil
[0].valuemask
;
1026 dsa
->stencil_ref
.valuemask
[1] = state
->stencil
[1].valuemask
;
1027 dsa
->stencil_ref
.writemask
[0] = state
->stencil
[0].writemask
;
1028 dsa
->stencil_ref
.writemask
[1] = state
->stencil
[1].writemask
;
1030 db_depth_control
= S_028800_Z_ENABLE(state
->depth
.enabled
) |
1031 S_028800_Z_WRITE_ENABLE(state
->depth
.writemask
) |
1032 S_028800_ZFUNC(state
->depth
.func
) |
1033 S_028800_DEPTH_BOUNDS_ENABLE(state
->depth
.bounds_test
);
1036 if (state
->stencil
[0].enabled
) {
1037 db_depth_control
|= S_028800_STENCIL_ENABLE(1);
1038 db_depth_control
|= S_028800_STENCILFUNC(state
->stencil
[0].func
);
1039 db_stencil_control
|= S_02842C_STENCILFAIL(si_translate_stencil_op(state
->stencil
[0].fail_op
));
1040 db_stencil_control
|= S_02842C_STENCILZPASS(si_translate_stencil_op(state
->stencil
[0].zpass_op
));
1041 db_stencil_control
|= S_02842C_STENCILZFAIL(si_translate_stencil_op(state
->stencil
[0].zfail_op
));
1043 if (state
->stencil
[1].enabled
) {
1044 db_depth_control
|= S_028800_BACKFACE_ENABLE(1);
1045 db_depth_control
|= S_028800_STENCILFUNC_BF(state
->stencil
[1].func
);
1046 db_stencil_control
|= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(state
->stencil
[1].fail_op
));
1047 db_stencil_control
|= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(state
->stencil
[1].zpass_op
));
1048 db_stencil_control
|= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(state
->stencil
[1].zfail_op
));
1053 if (state
->alpha
.enabled
) {
1054 dsa
->alpha_func
= state
->alpha
.func
;
1056 si_pm4_set_reg(pm4
, R_00B030_SPI_SHADER_USER_DATA_PS_0
+
1057 SI_SGPR_ALPHA_REF
* 4, fui(state
->alpha
.ref_value
));
1059 dsa
->alpha_func
= PIPE_FUNC_ALWAYS
;
1062 si_pm4_set_reg(pm4
, R_028800_DB_DEPTH_CONTROL
, db_depth_control
);
1063 si_pm4_set_reg(pm4
, R_02842C_DB_STENCIL_CONTROL
, db_stencil_control
);
1064 if (state
->depth
.bounds_test
) {
1065 si_pm4_set_reg(pm4
, R_028020_DB_DEPTH_BOUNDS_MIN
, fui(state
->depth
.bounds_min
));
1066 si_pm4_set_reg(pm4
, R_028024_DB_DEPTH_BOUNDS_MAX
, fui(state
->depth
.bounds_max
));
1072 static void si_bind_dsa_state(struct pipe_context
*ctx
, void *state
)
1074 struct si_context
*sctx
= (struct si_context
*)ctx
;
1075 struct si_state_dsa
*dsa
= state
;
1080 si_pm4_bind_state(sctx
, dsa
, dsa
);
1082 if (memcmp(&dsa
->stencil_ref
, &sctx
->stencil_ref
.dsa_part
,
1083 sizeof(struct si_dsa_stencil_ref_part
)) != 0) {
1084 sctx
->stencil_ref
.dsa_part
= dsa
->stencil_ref
;
1085 si_mark_atom_dirty(sctx
, &sctx
->stencil_ref
.atom
);
1087 sctx
->do_update_shaders
= true;
1090 static void si_delete_dsa_state(struct pipe_context
*ctx
, void *state
)
1092 struct si_context
*sctx
= (struct si_context
*)ctx
;
1093 si_pm4_delete_state(sctx
, dsa
, (struct si_state_dsa
*)state
);
1096 static void *si_create_db_flush_dsa(struct si_context
*sctx
)
1098 struct pipe_depth_stencil_alpha_state dsa
= {};
1100 return sctx
->b
.b
.create_depth_stencil_alpha_state(&sctx
->b
.b
, &dsa
);
1103 /* DB RENDER STATE */
1105 static void si_set_active_query_state(struct pipe_context
*ctx
, boolean enable
)
1107 struct si_context
*sctx
= (struct si_context
*)ctx
;
1109 /* Pipeline stat & streamout queries. */
1111 sctx
->b
.flags
&= ~R600_CONTEXT_STOP_PIPELINE_STATS
;
1112 sctx
->b
.flags
|= R600_CONTEXT_START_PIPELINE_STATS
;
1114 sctx
->b
.flags
&= ~R600_CONTEXT_START_PIPELINE_STATS
;
1115 sctx
->b
.flags
|= R600_CONTEXT_STOP_PIPELINE_STATS
;
1118 /* Occlusion queries. */
1119 if (sctx
->occlusion_queries_disabled
!= !enable
) {
1120 sctx
->occlusion_queries_disabled
= !enable
;
1121 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
1125 static void si_set_occlusion_query_state(struct pipe_context
*ctx
, bool enable
)
1127 struct si_context
*sctx
= (struct si_context
*)ctx
;
1129 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
1132 static void si_save_qbo_state(struct pipe_context
*ctx
, struct r600_qbo_state
*st
)
1134 struct si_context
*sctx
= (struct si_context
*)ctx
;
1136 st
->saved_compute
= sctx
->cs_shader_state
.program
;
1138 si_get_pipe_constant_buffer(sctx
, PIPE_SHADER_COMPUTE
, 0, &st
->saved_const0
);
1139 si_get_shader_buffers(sctx
, PIPE_SHADER_COMPUTE
, 0, 3, st
->saved_ssbo
);
1142 static void si_emit_db_render_state(struct si_context
*sctx
, struct r600_atom
*state
)
1144 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
1145 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1146 unsigned db_shader_control
;
1148 radeon_set_context_reg_seq(cs
, R_028000_DB_RENDER_CONTROL
, 2);
1150 /* DB_RENDER_CONTROL */
1151 if (sctx
->dbcb_depth_copy_enabled
||
1152 sctx
->dbcb_stencil_copy_enabled
) {
1154 S_028000_DEPTH_COPY(sctx
->dbcb_depth_copy_enabled
) |
1155 S_028000_STENCIL_COPY(sctx
->dbcb_stencil_copy_enabled
) |
1156 S_028000_COPY_CENTROID(1) |
1157 S_028000_COPY_SAMPLE(sctx
->dbcb_copy_sample
));
1158 } else if (sctx
->db_flush_depth_inplace
|| sctx
->db_flush_stencil_inplace
) {
1160 S_028000_DEPTH_COMPRESS_DISABLE(sctx
->db_flush_depth_inplace
) |
1161 S_028000_STENCIL_COMPRESS_DISABLE(sctx
->db_flush_stencil_inplace
));
1164 S_028000_DEPTH_CLEAR_ENABLE(sctx
->db_depth_clear
) |
1165 S_028000_STENCIL_CLEAR_ENABLE(sctx
->db_stencil_clear
));
1168 /* DB_COUNT_CONTROL (occlusion queries) */
1169 if (sctx
->b
.num_occlusion_queries
> 0 &&
1170 !sctx
->occlusion_queries_disabled
) {
1171 bool perfect
= sctx
->b
.num_perfect_occlusion_queries
> 0;
1173 if (sctx
->b
.chip_class
>= CIK
) {
1175 S_028004_PERFECT_ZPASS_COUNTS(perfect
) |
1176 S_028004_SAMPLE_RATE(sctx
->framebuffer
.log_samples
) |
1177 S_028004_ZPASS_ENABLE(1) |
1178 S_028004_SLICE_EVEN_ENABLE(1) |
1179 S_028004_SLICE_ODD_ENABLE(1));
1182 S_028004_PERFECT_ZPASS_COUNTS(perfect
) |
1183 S_028004_SAMPLE_RATE(sctx
->framebuffer
.log_samples
));
1186 /* Disable occlusion queries. */
1187 if (sctx
->b
.chip_class
>= CIK
) {
1190 radeon_emit(cs
, S_028004_ZPASS_INCREMENT_DISABLE(1));
1194 /* DB_RENDER_OVERRIDE2 */
1195 radeon_set_context_reg(cs
, R_028010_DB_RENDER_OVERRIDE2
,
1196 S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(sctx
->db_depth_disable_expclear
) |
1197 S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(sctx
->db_stencil_disable_expclear
) |
1198 S_028010_DECOMPRESS_Z_ON_FLUSH(sctx
->framebuffer
.nr_samples
>= 4));
1200 db_shader_control
= sctx
->ps_db_shader_control
;
1202 /* Bug workaround for smoothing (overrasterization) on SI. */
1203 if (sctx
->b
.chip_class
== SI
&& sctx
->smoothing_enabled
) {
1204 db_shader_control
&= C_02880C_Z_ORDER
;
1205 db_shader_control
|= S_02880C_Z_ORDER(V_02880C_LATE_Z
);
1208 /* Disable the gl_SampleMask fragment shader output if MSAA is disabled. */
1209 if (!rs
|| !rs
->multisample_enable
)
1210 db_shader_control
&= C_02880C_MASK_EXPORT_ENABLE
;
1212 if (sctx
->screen
->b
.has_rbplus
&&
1213 !sctx
->screen
->b
.rbplus_allowed
)
1214 db_shader_control
|= S_02880C_DUAL_QUAD_DISABLE(1);
1216 radeon_set_context_reg(cs
, R_02880C_DB_SHADER_CONTROL
,
1221 * format translation
1223 static uint32_t si_translate_colorformat(enum pipe_format format
)
1225 const struct util_format_description
*desc
= util_format_description(format
);
1227 #define HAS_SIZE(x,y,z,w) \
1228 (desc->channel[0].size == (x) && desc->channel[1].size == (y) && \
1229 desc->channel[2].size == (z) && desc->channel[3].size == (w))
1231 if (format
== PIPE_FORMAT_R11G11B10_FLOAT
) /* isn't plain */
1232 return V_028C70_COLOR_10_11_11
;
1234 if (desc
->layout
!= UTIL_FORMAT_LAYOUT_PLAIN
)
1235 return V_028C70_COLOR_INVALID
;
1237 /* hw cannot support mixed formats (except depth/stencil, since
1238 * stencil is not written to). */
1239 if (desc
->is_mixed
&& desc
->colorspace
!= UTIL_FORMAT_COLORSPACE_ZS
)
1240 return V_028C70_COLOR_INVALID
;
1242 switch (desc
->nr_channels
) {
1244 switch (desc
->channel
[0].size
) {
1246 return V_028C70_COLOR_8
;
1248 return V_028C70_COLOR_16
;
1250 return V_028C70_COLOR_32
;
1254 if (desc
->channel
[0].size
== desc
->channel
[1].size
) {
1255 switch (desc
->channel
[0].size
) {
1257 return V_028C70_COLOR_8_8
;
1259 return V_028C70_COLOR_16_16
;
1261 return V_028C70_COLOR_32_32
;
1263 } else if (HAS_SIZE(8,24,0,0)) {
1264 return V_028C70_COLOR_24_8
;
1265 } else if (HAS_SIZE(24,8,0,0)) {
1266 return V_028C70_COLOR_8_24
;
1270 if (HAS_SIZE(5,6,5,0)) {
1271 return V_028C70_COLOR_5_6_5
;
1272 } else if (HAS_SIZE(32,8,24,0)) {
1273 return V_028C70_COLOR_X24_8_32_FLOAT
;
1277 if (desc
->channel
[0].size
== desc
->channel
[1].size
&&
1278 desc
->channel
[0].size
== desc
->channel
[2].size
&&
1279 desc
->channel
[0].size
== desc
->channel
[3].size
) {
1280 switch (desc
->channel
[0].size
) {
1282 return V_028C70_COLOR_4_4_4_4
;
1284 return V_028C70_COLOR_8_8_8_8
;
1286 return V_028C70_COLOR_16_16_16_16
;
1288 return V_028C70_COLOR_32_32_32_32
;
1290 } else if (HAS_SIZE(5,5,5,1)) {
1291 return V_028C70_COLOR_1_5_5_5
;
1292 } else if (HAS_SIZE(10,10,10,2)) {
1293 return V_028C70_COLOR_2_10_10_10
;
1297 return V_028C70_COLOR_INVALID
;
1300 static uint32_t si_colorformat_endian_swap(uint32_t colorformat
)
1302 if (SI_BIG_ENDIAN
) {
1303 switch(colorformat
) {
1304 /* 8-bit buffers. */
1305 case V_028C70_COLOR_8
:
1306 return V_028C70_ENDIAN_NONE
;
1308 /* 16-bit buffers. */
1309 case V_028C70_COLOR_5_6_5
:
1310 case V_028C70_COLOR_1_5_5_5
:
1311 case V_028C70_COLOR_4_4_4_4
:
1312 case V_028C70_COLOR_16
:
1313 case V_028C70_COLOR_8_8
:
1314 return V_028C70_ENDIAN_8IN16
;
1316 /* 32-bit buffers. */
1317 case V_028C70_COLOR_8_8_8_8
:
1318 case V_028C70_COLOR_2_10_10_10
:
1319 case V_028C70_COLOR_8_24
:
1320 case V_028C70_COLOR_24_8
:
1321 case V_028C70_COLOR_16_16
:
1322 return V_028C70_ENDIAN_8IN32
;
1324 /* 64-bit buffers. */
1325 case V_028C70_COLOR_16_16_16_16
:
1326 return V_028C70_ENDIAN_8IN16
;
1328 case V_028C70_COLOR_32_32
:
1329 return V_028C70_ENDIAN_8IN32
;
1331 /* 128-bit buffers. */
1332 case V_028C70_COLOR_32_32_32_32
:
1333 return V_028C70_ENDIAN_8IN32
;
1335 return V_028C70_ENDIAN_NONE
; /* Unsupported. */
1338 return V_028C70_ENDIAN_NONE
;
1342 static uint32_t si_translate_dbformat(enum pipe_format format
)
1345 case PIPE_FORMAT_Z16_UNORM
:
1346 return V_028040_Z_16
;
1347 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
1348 case PIPE_FORMAT_X8Z24_UNORM
:
1349 case PIPE_FORMAT_Z24X8_UNORM
:
1350 case PIPE_FORMAT_Z24_UNORM_S8_UINT
:
1351 return V_028040_Z_24
; /* deprecated on SI */
1352 case PIPE_FORMAT_Z32_FLOAT
:
1353 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
1354 return V_028040_Z_32_FLOAT
;
1356 return V_028040_Z_INVALID
;
1361 * Texture translation
1364 static uint32_t si_translate_texformat(struct pipe_screen
*screen
,
1365 enum pipe_format format
,
1366 const struct util_format_description
*desc
,
1369 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
1370 bool enable_compressed_formats
= (sscreen
->b
.info
.drm_major
== 2 &&
1371 sscreen
->b
.info
.drm_minor
>= 31) ||
1372 sscreen
->b
.info
.drm_major
== 3;
1373 bool uniform
= true;
1376 /* Colorspace (return non-RGB formats directly). */
1377 switch (desc
->colorspace
) {
1378 /* Depth stencil formats */
1379 case UTIL_FORMAT_COLORSPACE_ZS
:
1381 case PIPE_FORMAT_Z16_UNORM
:
1382 return V_008F14_IMG_DATA_FORMAT_16
;
1383 case PIPE_FORMAT_X24S8_UINT
:
1384 case PIPE_FORMAT_S8X24_UINT
:
1386 * Implemented as an 8_8_8_8 data format to fix texture
1387 * gathers in stencil sampling. This affects at least
1388 * GL45-CTS.texture_cube_map_array.sampling on VI.
1390 return V_008F14_IMG_DATA_FORMAT_8_8_8_8
;
1391 case PIPE_FORMAT_Z24X8_UNORM
:
1392 case PIPE_FORMAT_Z24_UNORM_S8_UINT
:
1393 return V_008F14_IMG_DATA_FORMAT_8_24
;
1394 case PIPE_FORMAT_X8Z24_UNORM
:
1395 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
1396 return V_008F14_IMG_DATA_FORMAT_24_8
;
1397 case PIPE_FORMAT_S8_UINT
:
1398 return V_008F14_IMG_DATA_FORMAT_8
;
1399 case PIPE_FORMAT_Z32_FLOAT
:
1400 return V_008F14_IMG_DATA_FORMAT_32
;
1401 case PIPE_FORMAT_X32_S8X24_UINT
:
1402 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
1403 return V_008F14_IMG_DATA_FORMAT_X24_8_32
;
1408 case UTIL_FORMAT_COLORSPACE_YUV
:
1409 goto out_unknown
; /* TODO */
1411 case UTIL_FORMAT_COLORSPACE_SRGB
:
1412 if (desc
->nr_channels
!= 4 && desc
->nr_channels
!= 1)
1420 if (desc
->layout
== UTIL_FORMAT_LAYOUT_RGTC
) {
1421 if (!enable_compressed_formats
)
1425 case PIPE_FORMAT_RGTC1_SNORM
:
1426 case PIPE_FORMAT_LATC1_SNORM
:
1427 case PIPE_FORMAT_RGTC1_UNORM
:
1428 case PIPE_FORMAT_LATC1_UNORM
:
1429 return V_008F14_IMG_DATA_FORMAT_BC4
;
1430 case PIPE_FORMAT_RGTC2_SNORM
:
1431 case PIPE_FORMAT_LATC2_SNORM
:
1432 case PIPE_FORMAT_RGTC2_UNORM
:
1433 case PIPE_FORMAT_LATC2_UNORM
:
1434 return V_008F14_IMG_DATA_FORMAT_BC5
;
1440 if (desc
->layout
== UTIL_FORMAT_LAYOUT_ETC
&&
1441 (sscreen
->b
.family
== CHIP_STONEY
||
1442 sscreen
->b
.chip_class
>= GFX9
)) {
1444 case PIPE_FORMAT_ETC1_RGB8
:
1445 case PIPE_FORMAT_ETC2_RGB8
:
1446 case PIPE_FORMAT_ETC2_SRGB8
:
1447 return V_008F14_IMG_DATA_FORMAT_ETC2_RGB
;
1448 case PIPE_FORMAT_ETC2_RGB8A1
:
1449 case PIPE_FORMAT_ETC2_SRGB8A1
:
1450 return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA1
;
1451 case PIPE_FORMAT_ETC2_RGBA8
:
1452 case PIPE_FORMAT_ETC2_SRGBA8
:
1453 return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA
;
1454 case PIPE_FORMAT_ETC2_R11_UNORM
:
1455 case PIPE_FORMAT_ETC2_R11_SNORM
:
1456 return V_008F14_IMG_DATA_FORMAT_ETC2_R
;
1457 case PIPE_FORMAT_ETC2_RG11_UNORM
:
1458 case PIPE_FORMAT_ETC2_RG11_SNORM
:
1459 return V_008F14_IMG_DATA_FORMAT_ETC2_RG
;
1465 if (desc
->layout
== UTIL_FORMAT_LAYOUT_BPTC
) {
1466 if (!enable_compressed_formats
)
1470 case PIPE_FORMAT_BPTC_RGBA_UNORM
:
1471 case PIPE_FORMAT_BPTC_SRGBA
:
1472 return V_008F14_IMG_DATA_FORMAT_BC7
;
1473 case PIPE_FORMAT_BPTC_RGB_FLOAT
:
1474 case PIPE_FORMAT_BPTC_RGB_UFLOAT
:
1475 return V_008F14_IMG_DATA_FORMAT_BC6
;
1481 if (desc
->layout
== UTIL_FORMAT_LAYOUT_SUBSAMPLED
) {
1483 case PIPE_FORMAT_R8G8_B8G8_UNORM
:
1484 case PIPE_FORMAT_G8R8_B8R8_UNORM
:
1485 return V_008F14_IMG_DATA_FORMAT_GB_GR
;
1486 case PIPE_FORMAT_G8R8_G8B8_UNORM
:
1487 case PIPE_FORMAT_R8G8_R8B8_UNORM
:
1488 return V_008F14_IMG_DATA_FORMAT_BG_RG
;
1494 if (desc
->layout
== UTIL_FORMAT_LAYOUT_S3TC
) {
1495 if (!enable_compressed_formats
)
1498 if (!util_format_s3tc_enabled
) {
1503 case PIPE_FORMAT_DXT1_RGB
:
1504 case PIPE_FORMAT_DXT1_RGBA
:
1505 case PIPE_FORMAT_DXT1_SRGB
:
1506 case PIPE_FORMAT_DXT1_SRGBA
:
1507 return V_008F14_IMG_DATA_FORMAT_BC1
;
1508 case PIPE_FORMAT_DXT3_RGBA
:
1509 case PIPE_FORMAT_DXT3_SRGBA
:
1510 return V_008F14_IMG_DATA_FORMAT_BC2
;
1511 case PIPE_FORMAT_DXT5_RGBA
:
1512 case PIPE_FORMAT_DXT5_SRGBA
:
1513 return V_008F14_IMG_DATA_FORMAT_BC3
;
1519 if (format
== PIPE_FORMAT_R9G9B9E5_FLOAT
) {
1520 return V_008F14_IMG_DATA_FORMAT_5_9_9_9
;
1521 } else if (format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1522 return V_008F14_IMG_DATA_FORMAT_10_11_11
;
1525 /* R8G8Bx_SNORM - TODO CxV8U8 */
1527 /* hw cannot support mixed formats (except depth/stencil, since only
1529 if (desc
->is_mixed
&& desc
->colorspace
!= UTIL_FORMAT_COLORSPACE_ZS
)
1532 /* See whether the components are of the same size. */
1533 for (i
= 1; i
< desc
->nr_channels
; i
++) {
1534 uniform
= uniform
&& desc
->channel
[0].size
== desc
->channel
[i
].size
;
1537 /* Non-uniform formats. */
1539 switch(desc
->nr_channels
) {
1541 if (desc
->channel
[0].size
== 5 &&
1542 desc
->channel
[1].size
== 6 &&
1543 desc
->channel
[2].size
== 5) {
1544 return V_008F14_IMG_DATA_FORMAT_5_6_5
;
1548 if (desc
->channel
[0].size
== 5 &&
1549 desc
->channel
[1].size
== 5 &&
1550 desc
->channel
[2].size
== 5 &&
1551 desc
->channel
[3].size
== 1) {
1552 return V_008F14_IMG_DATA_FORMAT_1_5_5_5
;
1554 if (desc
->channel
[0].size
== 10 &&
1555 desc
->channel
[1].size
== 10 &&
1556 desc
->channel
[2].size
== 10 &&
1557 desc
->channel
[3].size
== 2) {
1558 return V_008F14_IMG_DATA_FORMAT_2_10_10_10
;
1565 if (first_non_void
< 0 || first_non_void
> 3)
1568 /* uniform formats */
1569 switch (desc
->channel
[first_non_void
].size
) {
1571 switch (desc
->nr_channels
) {
1572 #if 0 /* Not supported for render targets */
1574 return V_008F14_IMG_DATA_FORMAT_4_4
;
1577 return V_008F14_IMG_DATA_FORMAT_4_4_4_4
;
1581 switch (desc
->nr_channels
) {
1583 return V_008F14_IMG_DATA_FORMAT_8
;
1585 return V_008F14_IMG_DATA_FORMAT_8_8
;
1587 return V_008F14_IMG_DATA_FORMAT_8_8_8_8
;
1591 switch (desc
->nr_channels
) {
1593 return V_008F14_IMG_DATA_FORMAT_16
;
1595 return V_008F14_IMG_DATA_FORMAT_16_16
;
1597 return V_008F14_IMG_DATA_FORMAT_16_16_16_16
;
1601 switch (desc
->nr_channels
) {
1603 return V_008F14_IMG_DATA_FORMAT_32
;
1605 return V_008F14_IMG_DATA_FORMAT_32_32
;
1606 #if 0 /* Not supported for render targets */
1608 return V_008F14_IMG_DATA_FORMAT_32_32_32
;
1611 return V_008F14_IMG_DATA_FORMAT_32_32_32_32
;
1616 /* R600_ERR("Unable to handle texformat %d %s\n", format, util_format_name(format)); */
1620 static unsigned si_tex_wrap(unsigned wrap
)
1624 case PIPE_TEX_WRAP_REPEAT
:
1625 return V_008F30_SQ_TEX_WRAP
;
1626 case PIPE_TEX_WRAP_CLAMP
:
1627 return V_008F30_SQ_TEX_CLAMP_HALF_BORDER
;
1628 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
1629 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
1630 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
1631 return V_008F30_SQ_TEX_CLAMP_BORDER
;
1632 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
1633 return V_008F30_SQ_TEX_MIRROR
;
1634 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
1635 return V_008F30_SQ_TEX_MIRROR_ONCE_HALF_BORDER
;
1636 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
1637 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
1638 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
1639 return V_008F30_SQ_TEX_MIRROR_ONCE_BORDER
;
1643 static unsigned si_tex_mipfilter(unsigned filter
)
1646 case PIPE_TEX_MIPFILTER_NEAREST
:
1647 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
1648 case PIPE_TEX_MIPFILTER_LINEAR
:
1649 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
1651 case PIPE_TEX_MIPFILTER_NONE
:
1652 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
1656 static unsigned si_tex_compare(unsigned compare
)
1660 case PIPE_FUNC_NEVER
:
1661 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
1662 case PIPE_FUNC_LESS
:
1663 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
1664 case PIPE_FUNC_EQUAL
:
1665 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
1666 case PIPE_FUNC_LEQUAL
:
1667 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
1668 case PIPE_FUNC_GREATER
:
1669 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
1670 case PIPE_FUNC_NOTEQUAL
:
1671 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
1672 case PIPE_FUNC_GEQUAL
:
1673 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
1674 case PIPE_FUNC_ALWAYS
:
1675 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
1679 static unsigned si_tex_dim(struct si_screen
*sscreen
, struct r600_texture
*rtex
,
1680 unsigned view_target
, unsigned nr_samples
)
1682 unsigned res_target
= rtex
->resource
.b
.b
.target
;
1684 if (view_target
== PIPE_TEXTURE_CUBE
||
1685 view_target
== PIPE_TEXTURE_CUBE_ARRAY
)
1686 res_target
= view_target
;
1687 /* If interpreting cubemaps as something else, set 2D_ARRAY. */
1688 else if (res_target
== PIPE_TEXTURE_CUBE
||
1689 res_target
== PIPE_TEXTURE_CUBE_ARRAY
)
1690 res_target
= PIPE_TEXTURE_2D_ARRAY
;
1692 /* GFX9 allocates 1D textures as 2D. */
1693 if ((res_target
== PIPE_TEXTURE_1D
||
1694 res_target
== PIPE_TEXTURE_1D_ARRAY
) &&
1695 sscreen
->b
.chip_class
>= GFX9
&&
1696 rtex
->surface
.u
.gfx9
.resource_type
== RADEON_RESOURCE_2D
) {
1697 if (res_target
== PIPE_TEXTURE_1D
)
1698 res_target
= PIPE_TEXTURE_2D
;
1700 res_target
= PIPE_TEXTURE_2D_ARRAY
;
1703 switch (res_target
) {
1705 case PIPE_TEXTURE_1D
:
1706 return V_008F1C_SQ_RSRC_IMG_1D
;
1707 case PIPE_TEXTURE_1D_ARRAY
:
1708 return V_008F1C_SQ_RSRC_IMG_1D_ARRAY
;
1709 case PIPE_TEXTURE_2D
:
1710 case PIPE_TEXTURE_RECT
:
1711 return nr_samples
> 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA
:
1712 V_008F1C_SQ_RSRC_IMG_2D
;
1713 case PIPE_TEXTURE_2D_ARRAY
:
1714 return nr_samples
> 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
:
1715 V_008F1C_SQ_RSRC_IMG_2D_ARRAY
;
1716 case PIPE_TEXTURE_3D
:
1717 return V_008F1C_SQ_RSRC_IMG_3D
;
1718 case PIPE_TEXTURE_CUBE
:
1719 case PIPE_TEXTURE_CUBE_ARRAY
:
1720 return V_008F1C_SQ_RSRC_IMG_CUBE
;
1725 * Format support testing
1728 static bool si_is_sampler_format_supported(struct pipe_screen
*screen
, enum pipe_format format
)
1730 return si_translate_texformat(screen
, format
, util_format_description(format
),
1731 util_format_get_first_non_void_channel(format
)) != ~0U;
1734 static uint32_t si_translate_buffer_dataformat(struct pipe_screen
*screen
,
1735 const struct util_format_description
*desc
,
1740 if (desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
)
1741 return V_008F0C_BUF_DATA_FORMAT_10_11_11
;
1743 assert(first_non_void
>= 0);
1745 if (desc
->nr_channels
== 4 &&
1746 desc
->channel
[0].size
== 10 &&
1747 desc
->channel
[1].size
== 10 &&
1748 desc
->channel
[2].size
== 10 &&
1749 desc
->channel
[3].size
== 2)
1750 return V_008F0C_BUF_DATA_FORMAT_2_10_10_10
;
1752 /* See whether the components are of the same size. */
1753 for (i
= 0; i
< desc
->nr_channels
; i
++) {
1754 if (desc
->channel
[first_non_void
].size
!= desc
->channel
[i
].size
)
1755 return V_008F0C_BUF_DATA_FORMAT_INVALID
;
1758 switch (desc
->channel
[first_non_void
].size
) {
1760 switch (desc
->nr_channels
) {
1762 case 3: /* 3 loads */
1763 return V_008F0C_BUF_DATA_FORMAT_8
;
1765 return V_008F0C_BUF_DATA_FORMAT_8_8
;
1767 return V_008F0C_BUF_DATA_FORMAT_8_8_8_8
;
1771 switch (desc
->nr_channels
) {
1773 case 3: /* 3 loads */
1774 return V_008F0C_BUF_DATA_FORMAT_16
;
1776 return V_008F0C_BUF_DATA_FORMAT_16_16
;
1778 return V_008F0C_BUF_DATA_FORMAT_16_16_16_16
;
1782 switch (desc
->nr_channels
) {
1784 return V_008F0C_BUF_DATA_FORMAT_32
;
1786 return V_008F0C_BUF_DATA_FORMAT_32_32
;
1788 return V_008F0C_BUF_DATA_FORMAT_32_32_32
;
1790 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32
;
1794 /* Legacy double formats. */
1795 switch (desc
->nr_channels
) {
1796 case 1: /* 1 load */
1797 return V_008F0C_BUF_DATA_FORMAT_32_32
;
1798 case 2: /* 1 load */
1799 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32
;
1800 case 3: /* 3 loads */
1801 return V_008F0C_BUF_DATA_FORMAT_32_32
;
1802 case 4: /* 2 loads */
1803 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32
;
1808 return V_008F0C_BUF_DATA_FORMAT_INVALID
;
1811 static uint32_t si_translate_buffer_numformat(struct pipe_screen
*screen
,
1812 const struct util_format_description
*desc
,
1815 if (desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
)
1816 return V_008F0C_BUF_NUM_FORMAT_FLOAT
;
1818 assert(first_non_void
>= 0);
1820 switch (desc
->channel
[first_non_void
].type
) {
1821 case UTIL_FORMAT_TYPE_SIGNED
:
1822 case UTIL_FORMAT_TYPE_FIXED
:
1823 if (desc
->channel
[first_non_void
].size
>= 32 ||
1824 desc
->channel
[first_non_void
].pure_integer
)
1825 return V_008F0C_BUF_NUM_FORMAT_SINT
;
1826 else if (desc
->channel
[first_non_void
].normalized
)
1827 return V_008F0C_BUF_NUM_FORMAT_SNORM
;
1829 return V_008F0C_BUF_NUM_FORMAT_SSCALED
;
1831 case UTIL_FORMAT_TYPE_UNSIGNED
:
1832 if (desc
->channel
[first_non_void
].size
>= 32 ||
1833 desc
->channel
[first_non_void
].pure_integer
)
1834 return V_008F0C_BUF_NUM_FORMAT_UINT
;
1835 else if (desc
->channel
[first_non_void
].normalized
)
1836 return V_008F0C_BUF_NUM_FORMAT_UNORM
;
1838 return V_008F0C_BUF_NUM_FORMAT_USCALED
;
1840 case UTIL_FORMAT_TYPE_FLOAT
:
1842 return V_008F0C_BUF_NUM_FORMAT_FLOAT
;
1846 static unsigned si_is_vertex_format_supported(struct pipe_screen
*screen
,
1847 enum pipe_format format
,
1850 const struct util_format_description
*desc
;
1852 unsigned data_format
;
1854 assert((usage
& ~(PIPE_BIND_SHADER_IMAGE
|
1855 PIPE_BIND_SAMPLER_VIEW
|
1856 PIPE_BIND_VERTEX_BUFFER
)) == 0);
1858 desc
= util_format_description(format
);
1860 /* There are no native 8_8_8 or 16_16_16 data formats, and we currently
1861 * select 8_8_8_8 and 16_16_16_16 instead. This works reasonably well
1862 * for read-only access (with caveats surrounding bounds checks), but
1863 * obviously fails for write access which we have to implement for
1864 * shader images. Luckily, OpenGL doesn't expect this to be supported
1865 * anyway, and so the only impact is on PBO uploads / downloads, which
1866 * shouldn't be expected to be fast for GL_RGB anyway.
1868 if (desc
->block
.bits
== 3 * 8 ||
1869 desc
->block
.bits
== 3 * 16) {
1870 if (usage
& (PIPE_BIND_SHADER_IMAGE
| PIPE_BIND_SAMPLER_VIEW
)) {
1871 usage
&= ~(PIPE_BIND_SHADER_IMAGE
| PIPE_BIND_SAMPLER_VIEW
);
1877 first_non_void
= util_format_get_first_non_void_channel(format
);
1878 data_format
= si_translate_buffer_dataformat(screen
, desc
, first_non_void
);
1879 if (data_format
== V_008F0C_BUF_DATA_FORMAT_INVALID
)
1885 static bool si_is_colorbuffer_format_supported(enum pipe_format format
)
1887 return si_translate_colorformat(format
) != V_028C70_COLOR_INVALID
&&
1888 r600_translate_colorswap(format
, false) != ~0U;
1891 static bool si_is_zs_format_supported(enum pipe_format format
)
1893 return si_translate_dbformat(format
) != V_028040_Z_INVALID
;
1896 static boolean
si_is_format_supported(struct pipe_screen
*screen
,
1897 enum pipe_format format
,
1898 enum pipe_texture_target target
,
1899 unsigned sample_count
,
1902 unsigned retval
= 0;
1904 if (target
>= PIPE_MAX_TEXTURE_TYPES
) {
1905 R600_ERR("r600: unsupported texture type %d\n", target
);
1909 if (!util_format_is_supported(format
, usage
))
1912 if (sample_count
> 1) {
1913 if (!screen
->get_param(screen
, PIPE_CAP_TEXTURE_MULTISAMPLE
))
1916 if (usage
& PIPE_BIND_SHADER_IMAGE
)
1919 switch (sample_count
) {
1925 if (format
== PIPE_FORMAT_NONE
)
1934 if (usage
& (PIPE_BIND_SAMPLER_VIEW
|
1935 PIPE_BIND_SHADER_IMAGE
)) {
1936 if (target
== PIPE_BUFFER
) {
1937 retval
|= si_is_vertex_format_supported(
1938 screen
, format
, usage
& (PIPE_BIND_SAMPLER_VIEW
|
1939 PIPE_BIND_SHADER_IMAGE
));
1941 if (si_is_sampler_format_supported(screen
, format
))
1942 retval
|= usage
& (PIPE_BIND_SAMPLER_VIEW
|
1943 PIPE_BIND_SHADER_IMAGE
);
1947 if ((usage
& (PIPE_BIND_RENDER_TARGET
|
1948 PIPE_BIND_DISPLAY_TARGET
|
1951 PIPE_BIND_BLENDABLE
)) &&
1952 si_is_colorbuffer_format_supported(format
)) {
1954 (PIPE_BIND_RENDER_TARGET
|
1955 PIPE_BIND_DISPLAY_TARGET
|
1958 if (!util_format_is_pure_integer(format
) &&
1959 !util_format_is_depth_or_stencil(format
))
1960 retval
|= usage
& PIPE_BIND_BLENDABLE
;
1963 if ((usage
& PIPE_BIND_DEPTH_STENCIL
) &&
1964 si_is_zs_format_supported(format
)) {
1965 retval
|= PIPE_BIND_DEPTH_STENCIL
;
1968 if (usage
& PIPE_BIND_VERTEX_BUFFER
) {
1969 retval
|= si_is_vertex_format_supported(screen
, format
,
1970 PIPE_BIND_VERTEX_BUFFER
);
1973 if ((usage
& PIPE_BIND_LINEAR
) &&
1974 !util_format_is_compressed(format
) &&
1975 !(usage
& PIPE_BIND_DEPTH_STENCIL
))
1976 retval
|= PIPE_BIND_LINEAR
;
1978 return retval
== usage
;
1982 * framebuffer handling
1985 static void si_choose_spi_color_formats(struct r600_surface
*surf
,
1986 unsigned format
, unsigned swap
,
1987 unsigned ntype
, bool is_depth
)
1989 /* Alpha is needed for alpha-to-coverage.
1990 * Blending may be with or without alpha.
1992 unsigned normal
= 0; /* most optimal, may not support blending or export alpha */
1993 unsigned alpha
= 0; /* exports alpha, but may not support blending */
1994 unsigned blend
= 0; /* supports blending, but may not export alpha */
1995 unsigned blend_alpha
= 0; /* least optimal, supports blending and exports alpha */
1997 /* Choose the SPI color formats. These are required values for RB+.
1998 * Other chips have multiple choices, though they are not necessarily better.
2001 case V_028C70_COLOR_5_6_5
:
2002 case V_028C70_COLOR_1_5_5_5
:
2003 case V_028C70_COLOR_5_5_5_1
:
2004 case V_028C70_COLOR_4_4_4_4
:
2005 case V_028C70_COLOR_10_11_11
:
2006 case V_028C70_COLOR_11_11_10
:
2007 case V_028C70_COLOR_8
:
2008 case V_028C70_COLOR_8_8
:
2009 case V_028C70_COLOR_8_8_8_8
:
2010 case V_028C70_COLOR_10_10_10_2
:
2011 case V_028C70_COLOR_2_10_10_10
:
2012 if (ntype
== V_028C70_NUMBER_UINT
)
2013 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
2014 else if (ntype
== V_028C70_NUMBER_SINT
)
2015 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
2017 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
2020 case V_028C70_COLOR_16
:
2021 case V_028C70_COLOR_16_16
:
2022 case V_028C70_COLOR_16_16_16_16
:
2023 if (ntype
== V_028C70_NUMBER_UNORM
||
2024 ntype
== V_028C70_NUMBER_SNORM
) {
2025 /* UNORM16 and SNORM16 don't support blending */
2026 if (ntype
== V_028C70_NUMBER_UNORM
)
2027 normal
= alpha
= V_028714_SPI_SHADER_UNORM16_ABGR
;
2029 normal
= alpha
= V_028714_SPI_SHADER_SNORM16_ABGR
;
2031 /* Use 32 bits per channel for blending. */
2032 if (format
== V_028C70_COLOR_16
) {
2033 if (swap
== V_028C70_SWAP_STD
) { /* R */
2034 blend
= V_028714_SPI_SHADER_32_R
;
2035 blend_alpha
= V_028714_SPI_SHADER_32_AR
;
2036 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
2037 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
2040 } else if (format
== V_028C70_COLOR_16_16
) {
2041 if (swap
== V_028C70_SWAP_STD
) { /* RG */
2042 blend
= V_028714_SPI_SHADER_32_GR
;
2043 blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
2044 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
2045 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
2048 } else /* 16_16_16_16 */
2049 blend
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
2050 } else if (ntype
== V_028C70_NUMBER_UINT
)
2051 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
2052 else if (ntype
== V_028C70_NUMBER_SINT
)
2053 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
2054 else if (ntype
== V_028C70_NUMBER_FLOAT
)
2055 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
2060 case V_028C70_COLOR_32
:
2061 if (swap
== V_028C70_SWAP_STD
) { /* R */
2062 blend
= normal
= V_028714_SPI_SHADER_32_R
;
2063 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
2064 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
2065 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
2070 case V_028C70_COLOR_32_32
:
2071 if (swap
== V_028C70_SWAP_STD
) { /* RG */
2072 blend
= normal
= V_028714_SPI_SHADER_32_GR
;
2073 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
2074 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
2075 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
2080 case V_028C70_COLOR_32_32_32_32
:
2081 case V_028C70_COLOR_8_24
:
2082 case V_028C70_COLOR_24_8
:
2083 case V_028C70_COLOR_X24_8_32_FLOAT
:
2084 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_ABGR
;
2092 /* The DB->CB copy needs 32_ABGR. */
2094 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_ABGR
;
2096 surf
->spi_shader_col_format
= normal
;
2097 surf
->spi_shader_col_format_alpha
= alpha
;
2098 surf
->spi_shader_col_format_blend
= blend
;
2099 surf
->spi_shader_col_format_blend_alpha
= blend_alpha
;
2102 static void si_initialize_color_surface(struct si_context
*sctx
,
2103 struct r600_surface
*surf
)
2105 struct r600_texture
*rtex
= (struct r600_texture
*)surf
->base
.texture
;
2106 unsigned color_info
, color_attrib
, color_view
;
2107 unsigned format
, swap
, ntype
, endian
;
2108 const struct util_format_description
*desc
;
2110 unsigned blend_clamp
= 0, blend_bypass
= 0;
2112 color_view
= S_028C6C_SLICE_START(surf
->base
.u
.tex
.first_layer
) |
2113 S_028C6C_SLICE_MAX(surf
->base
.u
.tex
.last_layer
);
2115 desc
= util_format_description(surf
->base
.format
);
2116 for (i
= 0; i
< 4; i
++) {
2117 if (desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
) {
2121 if (i
== 4 || desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_FLOAT
) {
2122 ntype
= V_028C70_NUMBER_FLOAT
;
2124 ntype
= V_028C70_NUMBER_UNORM
;
2125 if (desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
)
2126 ntype
= V_028C70_NUMBER_SRGB
;
2127 else if (desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_SIGNED
) {
2128 if (desc
->channel
[i
].pure_integer
) {
2129 ntype
= V_028C70_NUMBER_SINT
;
2131 assert(desc
->channel
[i
].normalized
);
2132 ntype
= V_028C70_NUMBER_SNORM
;
2134 } else if (desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
2135 if (desc
->channel
[i
].pure_integer
) {
2136 ntype
= V_028C70_NUMBER_UINT
;
2138 assert(desc
->channel
[i
].normalized
);
2139 ntype
= V_028C70_NUMBER_UNORM
;
2144 format
= si_translate_colorformat(surf
->base
.format
);
2145 if (format
== V_028C70_COLOR_INVALID
) {
2146 R600_ERR("Invalid CB format: %d, disabling CB.\n", surf
->base
.format
);
2148 assert(format
!= V_028C70_COLOR_INVALID
);
2149 swap
= r600_translate_colorswap(surf
->base
.format
, false);
2150 endian
= si_colorformat_endian_swap(format
);
2152 /* blend clamp should be set for all NORM/SRGB types */
2153 if (ntype
== V_028C70_NUMBER_UNORM
||
2154 ntype
== V_028C70_NUMBER_SNORM
||
2155 ntype
== V_028C70_NUMBER_SRGB
)
2158 /* set blend bypass according to docs if SINT/UINT or
2159 8/24 COLOR variants */
2160 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
2161 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
2162 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
2167 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) {
2168 if (format
== V_028C70_COLOR_8
||
2169 format
== V_028C70_COLOR_8_8
||
2170 format
== V_028C70_COLOR_8_8_8_8
)
2171 surf
->color_is_int8
= true;
2172 else if (format
== V_028C70_COLOR_10_10_10_2
||
2173 format
== V_028C70_COLOR_2_10_10_10
)
2174 surf
->color_is_int10
= true;
2177 color_info
= S_028C70_FORMAT(format
) |
2178 S_028C70_COMP_SWAP(swap
) |
2179 S_028C70_BLEND_CLAMP(blend_clamp
) |
2180 S_028C70_BLEND_BYPASS(blend_bypass
) |
2181 S_028C70_SIMPLE_FLOAT(1) |
2182 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
2183 ntype
!= V_028C70_NUMBER_SNORM
&&
2184 ntype
!= V_028C70_NUMBER_SRGB
&&
2185 format
!= V_028C70_COLOR_8_24
&&
2186 format
!= V_028C70_COLOR_24_8
) |
2187 S_028C70_NUMBER_TYPE(ntype
) |
2188 S_028C70_ENDIAN(endian
);
2190 /* Intensity is implemented as Red, so treat it that way. */
2191 color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == PIPE_SWIZZLE_1
||
2192 util_format_is_intensity(surf
->base
.format
));
2194 if (rtex
->resource
.b
.b
.nr_samples
> 1) {
2195 unsigned log_samples
= util_logbase2(rtex
->resource
.b
.b
.nr_samples
);
2197 color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
2198 S_028C74_NUM_FRAGMENTS(log_samples
);
2200 if (rtex
->fmask
.size
) {
2201 /* TODO: fix FMASK on GFX9: */
2202 color_info
|= S_028C70_COMPRESSION(sctx
->b
.chip_class
<= VI
);
2203 unsigned fmask_bankh
= util_logbase2(rtex
->fmask
.bank_height
);
2205 if (sctx
->b
.chip_class
== SI
) {
2206 /* due to a hw bug, FMASK_BANK_HEIGHT must be set on SI too */
2207 color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
2212 surf
->cb_color_view
= color_view
;
2213 surf
->cb_color_info
= color_info
;
2214 surf
->cb_color_attrib
= color_attrib
;
2216 if (sctx
->b
.chip_class
>= VI
) {
2217 unsigned max_uncompressed_block_size
= 2;
2219 if (rtex
->resource
.b
.b
.nr_samples
> 1) {
2220 if (rtex
->surface
.bpe
== 1)
2221 max_uncompressed_block_size
= 0;
2222 else if (rtex
->surface
.bpe
== 2)
2223 max_uncompressed_block_size
= 1;
2226 surf
->cb_dcc_control
= S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
2227 S_028C78_INDEPENDENT_64B_BLOCKS(1);
2230 /* This must be set for fast clear to work without FMASK. */
2231 if (!rtex
->fmask
.size
&& sctx
->b
.chip_class
== SI
) {
2232 unsigned bankh
= util_logbase2(rtex
->surface
.u
.legacy
.bankh
);
2233 surf
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
2236 if (sctx
->b
.chip_class
>= GFX9
) {
2237 unsigned mip0_depth
= util_max_layer(&rtex
->resource
.b
.b
, 0);
2239 surf
->cb_color_view
|= S_028C6C_MIP_LEVEL(surf
->base
.u
.tex
.level
);
2240 surf
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
2241 S_028C74_RESOURCE_TYPE(rtex
->surface
.u
.gfx9
.resource_type
);
2242 surf
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(surf
->width0
- 1) |
2243 S_028C68_MIP0_HEIGHT(surf
->height0
- 1) |
2244 S_028C68_MAX_MIP(rtex
->resource
.b
.b
.last_level
);
2247 /* Determine pixel shader export format */
2248 si_choose_spi_color_formats(surf
, format
, swap
, ntype
, rtex
->is_depth
);
2250 surf
->color_initialized
= true;
2253 static void si_init_depth_surface(struct si_context
*sctx
,
2254 struct r600_surface
*surf
)
2256 struct r600_texture
*rtex
= (struct r600_texture
*)surf
->base
.texture
;
2257 unsigned level
= surf
->base
.u
.tex
.level
;
2258 unsigned format
, stencil_format
;
2259 uint32_t z_info
, s_info
;
2261 format
= si_translate_dbformat(rtex
->db_render_format
);
2262 stencil_format
= rtex
->surface
.flags
& RADEON_SURF_SBUFFER
?
2263 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
2265 assert(format
!= V_028040_Z_INVALID
);
2266 if (format
== V_028040_Z_INVALID
)
2267 R600_ERR("Invalid DB format: %d, disabling DB.\n", rtex
->resource
.b
.b
.format
);
2269 surf
->db_depth_view
= S_028008_SLICE_START(surf
->base
.u
.tex
.first_layer
) |
2270 S_028008_SLICE_MAX(surf
->base
.u
.tex
.last_layer
);
2271 surf
->db_htile_data_base
= 0;
2272 surf
->db_htile_surface
= 0;
2274 if (sctx
->b
.chip_class
>= GFX9
) {
2275 assert(rtex
->surface
.u
.gfx9
.surf_offset
== 0);
2276 surf
->db_depth_base
= rtex
->resource
.gpu_address
>> 8;
2277 surf
->db_stencil_base
= (rtex
->resource
.gpu_address
+
2278 rtex
->surface
.u
.gfx9
.stencil_offset
) >> 8;
2279 z_info
= S_028038_FORMAT(format
) |
2280 S_028038_NUM_SAMPLES(util_logbase2(rtex
->resource
.b
.b
.nr_samples
)) |
2281 S_028038_SW_MODE(rtex
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
2282 S_028038_MAXMIP(rtex
->resource
.b
.b
.last_level
);
2283 s_info
= S_02803C_FORMAT(stencil_format
) |
2284 S_02803C_SW_MODE(rtex
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
2285 surf
->db_z_info2
= S_028068_EPITCH(rtex
->surface
.u
.gfx9
.surf
.epitch
);
2286 surf
->db_stencil_info2
= S_02806C_EPITCH(rtex
->surface
.u
.gfx9
.stencil
.epitch
);
2287 surf
->db_depth_view
|= S_028008_MIPID(level
);
2288 surf
->db_depth_size
= S_02801C_X_MAX(rtex
->resource
.b
.b
.width0
- 1) |
2289 S_02801C_Y_MAX(rtex
->resource
.b
.b
.height0
- 1);
2291 /* Only use HTILE for the first level. */
2292 if (rtex
->htile_buffer
&& !level
) {
2293 z_info
|= S_028038_TILE_SURFACE_ENABLE(1) |
2294 S_028038_ALLOW_EXPCLEAR(1);
2296 if (rtex
->tc_compatible_htile
) {
2297 unsigned max_zplanes
= 4;
2299 if (rtex
->db_render_format
== PIPE_FORMAT_Z16_UNORM
&&
2300 rtex
->resource
.b
.b
.nr_samples
> 1)
2303 z_info
|= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes
+ 1) |
2304 S_028038_ITERATE_FLUSH(1);
2305 s_info
|= S_02803C_ITERATE_FLUSH(1);
2308 if (rtex
->surface
.flags
& RADEON_SURF_SBUFFER
) {
2309 /* Stencil buffer workaround ported from the SI-CI-VI code.
2310 * See that for explanation.
2312 s_info
|= S_02803C_ALLOW_EXPCLEAR(rtex
->resource
.b
.b
.nr_samples
<= 1);
2314 /* Use all HTILE for depth if there's no stencil. */
2315 s_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
2318 surf
->db_htile_data_base
= rtex
->htile_buffer
->gpu_address
>> 8;
2319 surf
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
2320 S_028ABC_PIPE_ALIGNED(rtex
->surface
.u
.gfx9
.htile
.pipe_aligned
) |
2321 S_028ABC_RB_ALIGNED(rtex
->surface
.u
.gfx9
.htile
.rb_aligned
);
2325 struct legacy_surf_level
*levelinfo
= &rtex
->surface
.u
.legacy
.level
[level
];
2327 assert(levelinfo
->nblk_x
% 8 == 0 && levelinfo
->nblk_y
% 8 == 0);
2329 surf
->db_depth_base
= (rtex
->resource
.gpu_address
+
2330 rtex
->surface
.u
.legacy
.level
[level
].offset
) >> 8;
2331 surf
->db_stencil_base
= (rtex
->resource
.gpu_address
+
2332 rtex
->surface
.u
.legacy
.stencil_level
[level
].offset
) >> 8;
2334 z_info
= S_028040_FORMAT(format
) |
2335 S_028040_NUM_SAMPLES(util_logbase2(rtex
->resource
.b
.b
.nr_samples
));
2336 s_info
= S_028044_FORMAT(stencil_format
);
2337 surf
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(!rtex
->tc_compatible_htile
);
2339 if (sctx
->b
.chip_class
>= CIK
) {
2340 struct radeon_info
*info
= &sctx
->screen
->b
.info
;
2341 unsigned index
= rtex
->surface
.u
.legacy
.tiling_index
[level
];
2342 unsigned stencil_index
= rtex
->surface
.u
.legacy
.stencil_tiling_index
[level
];
2343 unsigned macro_index
= rtex
->surface
.u
.legacy
.macro_tile_index
;
2344 unsigned tile_mode
= info
->si_tile_mode_array
[index
];
2345 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
2346 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
2348 surf
->db_depth_info
|=
2349 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
2350 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
2351 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
2352 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
2353 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
2354 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
2355 z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
2356 s_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
2358 unsigned tile_mode_index
= si_tile_mode_index(rtex
, level
, false);
2359 z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
2360 tile_mode_index
= si_tile_mode_index(rtex
, level
, true);
2361 s_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
2364 surf
->db_depth_size
= S_028058_PITCH_TILE_MAX((levelinfo
->nblk_x
/ 8) - 1) |
2365 S_028058_HEIGHT_TILE_MAX((levelinfo
->nblk_y
/ 8) - 1);
2366 surf
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((levelinfo
->nblk_x
*
2367 levelinfo
->nblk_y
) / 64 - 1);
2369 /* Only use HTILE for the first level. */
2370 if (rtex
->htile_buffer
&& !level
) {
2371 z_info
|= S_028040_TILE_SURFACE_ENABLE(1) |
2372 S_028040_ALLOW_EXPCLEAR(1);
2374 if (rtex
->surface
.flags
& RADEON_SURF_SBUFFER
) {
2375 /* Workaround: For a not yet understood reason, the
2376 * combination of MSAA, fast stencil clear and stencil
2377 * decompress messes with subsequent stencil buffer
2378 * uses. Problem was reproduced on Verde, Bonaire,
2379 * Tonga, and Carrizo.
2381 * Disabling EXPCLEAR works around the problem.
2383 * Check piglit's arb_texture_multisample-stencil-clear
2384 * test if you want to try changing this.
2386 if (rtex
->resource
.b
.b
.nr_samples
<= 1)
2387 s_info
|= S_028044_ALLOW_EXPCLEAR(1);
2388 } else if (!rtex
->tc_compatible_htile
) {
2389 /* Use all of the htile_buffer for depth if there's no stencil.
2390 * This must not be set when TC-compatible HTILE is enabled
2393 s_info
|= S_028044_TILE_STENCIL_DISABLE(1);
2396 surf
->db_htile_data_base
= rtex
->htile_buffer
->gpu_address
>> 8;
2397 surf
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
2399 if (rtex
->tc_compatible_htile
) {
2400 surf
->db_htile_surface
|= S_028ABC_TC_COMPATIBLE(1);
2402 if (rtex
->resource
.b
.b
.nr_samples
<= 1)
2403 z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(5);
2404 else if (rtex
->resource
.b
.b
.nr_samples
<= 4)
2405 z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(3);
2407 z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(2);
2412 surf
->db_z_info
= z_info
;
2413 surf
->db_stencil_info
= s_info
;
2415 surf
->depth_initialized
= true;
2418 static void si_dec_framebuffer_counters(const struct pipe_framebuffer_state
*state
)
2420 for (int i
= 0; i
< state
->nr_cbufs
; ++i
) {
2421 struct r600_surface
*surf
= NULL
;
2422 struct r600_texture
*rtex
;
2424 if (!state
->cbufs
[i
])
2426 surf
= (struct r600_surface
*)state
->cbufs
[i
];
2427 rtex
= (struct r600_texture
*)surf
->base
.texture
;
2429 p_atomic_dec(&rtex
->framebuffers_bound
);
2433 static void si_set_framebuffer_state(struct pipe_context
*ctx
,
2434 const struct pipe_framebuffer_state
*state
)
2436 struct si_context
*sctx
= (struct si_context
*)ctx
;
2437 struct pipe_constant_buffer constbuf
= {0};
2438 struct r600_surface
*surf
= NULL
;
2439 struct r600_texture
*rtex
;
2440 bool old_any_dst_linear
= sctx
->framebuffer
.any_dst_linear
;
2441 unsigned old_nr_samples
= sctx
->framebuffer
.nr_samples
;
2444 for (i
= 0; i
< sctx
->framebuffer
.state
.nr_cbufs
; i
++) {
2445 if (!sctx
->framebuffer
.state
.cbufs
[i
])
2448 rtex
= (struct r600_texture
*)sctx
->framebuffer
.state
.cbufs
[i
]->texture
;
2449 if (rtex
->dcc_gather_statistics
)
2450 vi_separate_dcc_stop_query(ctx
, rtex
);
2453 /* Only flush TC when changing the framebuffer state, because
2454 * the only client not using TC that can change textures is
2457 * Flush all CB and DB caches here because all buffers can be used
2458 * for write by both TC (with shader image stores) and CB/DB.
2460 sctx
->b
.flags
|= SI_CONTEXT_INV_VMEM_L1
|
2461 SI_CONTEXT_INV_GLOBAL_L2
|
2462 SI_CONTEXT_FLUSH_AND_INV_CB
|
2463 SI_CONTEXT_FLUSH_AND_INV_DB
|
2464 SI_CONTEXT_CS_PARTIAL_FLUSH
;
2466 /* Take the maximum of the old and new count. If the new count is lower,
2467 * dirtying is needed to disable the unbound colorbuffers.
2469 sctx
->framebuffer
.dirty_cbufs
|=
2470 (1 << MAX2(sctx
->framebuffer
.state
.nr_cbufs
, state
->nr_cbufs
)) - 1;
2471 sctx
->framebuffer
.dirty_zsbuf
|= sctx
->framebuffer
.state
.zsbuf
!= state
->zsbuf
;
2473 si_dec_framebuffer_counters(&sctx
->framebuffer
.state
);
2474 util_copy_framebuffer_state(&sctx
->framebuffer
.state
, state
);
2476 sctx
->framebuffer
.colorbuf_enabled_4bit
= 0;
2477 sctx
->framebuffer
.spi_shader_col_format
= 0;
2478 sctx
->framebuffer
.spi_shader_col_format_alpha
= 0;
2479 sctx
->framebuffer
.spi_shader_col_format_blend
= 0;
2480 sctx
->framebuffer
.spi_shader_col_format_blend_alpha
= 0;
2481 sctx
->framebuffer
.color_is_int8
= 0;
2482 sctx
->framebuffer
.color_is_int10
= 0;
2484 sctx
->framebuffer
.compressed_cb_mask
= 0;
2485 sctx
->framebuffer
.nr_samples
= util_framebuffer_get_num_samples(state
);
2486 sctx
->framebuffer
.log_samples
= util_logbase2(sctx
->framebuffer
.nr_samples
);
2487 sctx
->framebuffer
.any_dst_linear
= false;
2489 for (i
= 0; i
< state
->nr_cbufs
; i
++) {
2490 if (!state
->cbufs
[i
])
2493 surf
= (struct r600_surface
*)state
->cbufs
[i
];
2494 rtex
= (struct r600_texture
*)surf
->base
.texture
;
2496 if (!surf
->color_initialized
) {
2497 si_initialize_color_surface(sctx
, surf
);
2500 sctx
->framebuffer
.colorbuf_enabled_4bit
|= 0xf << (i
* 4);
2501 sctx
->framebuffer
.spi_shader_col_format
|=
2502 surf
->spi_shader_col_format
<< (i
* 4);
2503 sctx
->framebuffer
.spi_shader_col_format_alpha
|=
2504 surf
->spi_shader_col_format_alpha
<< (i
* 4);
2505 sctx
->framebuffer
.spi_shader_col_format_blend
|=
2506 surf
->spi_shader_col_format_blend
<< (i
* 4);
2507 sctx
->framebuffer
.spi_shader_col_format_blend_alpha
|=
2508 surf
->spi_shader_col_format_blend_alpha
<< (i
* 4);
2510 if (surf
->color_is_int8
)
2511 sctx
->framebuffer
.color_is_int8
|= 1 << i
;
2512 if (surf
->color_is_int10
)
2513 sctx
->framebuffer
.color_is_int10
|= 1 << i
;
2515 if (rtex
->fmask
.size
) {
2516 sctx
->framebuffer
.compressed_cb_mask
|= 1 << i
;
2519 if (rtex
->surface
.is_linear
)
2520 sctx
->framebuffer
.any_dst_linear
= true;
2522 r600_context_add_resource_size(ctx
, surf
->base
.texture
);
2524 p_atomic_inc(&rtex
->framebuffers_bound
);
2526 if (rtex
->dcc_gather_statistics
) {
2527 /* Dirty tracking must be enabled for DCC usage analysis. */
2528 sctx
->framebuffer
.compressed_cb_mask
|= 1 << i
;
2529 vi_separate_dcc_start_query(ctx
, rtex
);
2534 surf
= (struct r600_surface
*)state
->zsbuf
;
2535 rtex
= (struct r600_texture
*)surf
->base
.texture
;
2537 if (!surf
->depth_initialized
) {
2538 si_init_depth_surface(sctx
, surf
);
2540 r600_context_add_resource_size(ctx
, surf
->base
.texture
);
2543 si_update_poly_offset_state(sctx
);
2544 si_mark_atom_dirty(sctx
, &sctx
->cb_render_state
);
2545 si_mark_atom_dirty(sctx
, &sctx
->framebuffer
.atom
);
2547 if (sctx
->framebuffer
.any_dst_linear
!= old_any_dst_linear
)
2548 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
2550 if (sctx
->framebuffer
.nr_samples
!= old_nr_samples
) {
2551 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
2552 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
2554 /* Set sample locations as fragment shader constants. */
2555 switch (sctx
->framebuffer
.nr_samples
) {
2557 constbuf
.user_buffer
= sctx
->b
.sample_locations_1x
;
2560 constbuf
.user_buffer
= sctx
->b
.sample_locations_2x
;
2563 constbuf
.user_buffer
= sctx
->b
.sample_locations_4x
;
2566 constbuf
.user_buffer
= sctx
->b
.sample_locations_8x
;
2569 constbuf
.user_buffer
= sctx
->b
.sample_locations_16x
;
2572 R600_ERR("Requested an invalid number of samples %i.\n",
2573 sctx
->framebuffer
.nr_samples
);
2576 constbuf
.buffer_size
= sctx
->framebuffer
.nr_samples
* 2 * 4;
2577 si_set_rw_buffer(sctx
, SI_PS_CONST_SAMPLE_POSITIONS
, &constbuf
);
2579 si_mark_atom_dirty(sctx
, &sctx
->msaa_sample_locs
.atom
);
2582 sctx
->need_check_render_feedback
= true;
2583 sctx
->do_update_shaders
= true;
2584 sctx
->framebuffer
.do_update_surf_dirtiness
= true;
2587 static void si_emit_framebuffer_state(struct si_context
*sctx
, struct r600_atom
*atom
)
2589 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
2590 struct pipe_framebuffer_state
*state
= &sctx
->framebuffer
.state
;
2591 unsigned i
, nr_cbufs
= state
->nr_cbufs
;
2592 struct r600_texture
*tex
= NULL
;
2593 struct r600_surface
*cb
= NULL
;
2594 unsigned cb_color_info
= 0;
2597 for (i
= 0; i
< nr_cbufs
; i
++) {
2598 uint64_t cb_color_base
, cb_color_fmask
, cb_dcc_base
;
2599 unsigned cb_color_attrib
;
2601 if (!(sctx
->framebuffer
.dirty_cbufs
& (1 << i
)))
2604 cb
= (struct r600_surface
*)state
->cbufs
[i
];
2606 radeon_set_context_reg(cs
, R_028C70_CB_COLOR0_INFO
+ i
* 0x3C,
2607 S_028C70_FORMAT(V_028C70_COLOR_INVALID
));
2611 tex
= (struct r600_texture
*)cb
->base
.texture
;
2612 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2613 &tex
->resource
, RADEON_USAGE_READWRITE
,
2614 tex
->resource
.b
.b
.nr_samples
> 1 ?
2615 RADEON_PRIO_COLOR_BUFFER_MSAA
:
2616 RADEON_PRIO_COLOR_BUFFER
);
2618 if (tex
->cmask_buffer
&& tex
->cmask_buffer
!= &tex
->resource
) {
2619 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2620 tex
->cmask_buffer
, RADEON_USAGE_READWRITE
,
2624 if (tex
->dcc_separate_buffer
)
2625 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2626 tex
->dcc_separate_buffer
,
2627 RADEON_USAGE_READWRITE
,
2630 /* Compute mutable surface parameters. */
2631 cb_color_base
= tex
->resource
.gpu_address
>> 8;
2632 cb_color_fmask
= cb_color_base
;
2634 cb_color_info
= cb
->cb_color_info
| tex
->cb_color_info
;
2635 cb_color_attrib
= cb
->cb_color_attrib
;
2637 if (tex
->fmask
.size
)
2638 cb_color_fmask
= (tex
->resource
.gpu_address
+ tex
->fmask
.offset
) >> 8;
2641 if (vi_dcc_enabled(tex
, cb
->base
.u
.tex
.level
)) {
2642 bool is_msaa_resolve_dst
= state
->cbufs
[0] &&
2643 state
->cbufs
[0]->texture
->nr_samples
> 1 &&
2644 state
->cbufs
[1] == &cb
->base
&&
2645 state
->cbufs
[1]->texture
->nr_samples
<= 1;
2647 if (!is_msaa_resolve_dst
)
2648 cb_color_info
|= S_028C70_DCC_ENABLE(1);
2650 cb_dcc_base
= ((!tex
->dcc_separate_buffer
? tex
->resource
.gpu_address
: 0) +
2651 tex
->dcc_offset
) >> 8;
2654 if (sctx
->b
.chip_class
>= GFX9
) {
2655 struct gfx9_surf_meta_flags meta
;
2657 if (tex
->dcc_offset
)
2658 meta
= tex
->surface
.u
.gfx9
.dcc
;
2660 meta
= tex
->surface
.u
.gfx9
.cmask
;
2662 /* Set mutable surface parameters. */
2663 cb_color_base
+= tex
->surface
.u
.gfx9
.surf_offset
>> 8;
2664 cb_color_attrib
|= S_028C74_COLOR_SW_MODE(tex
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
2665 S_028C74_FMASK_SW_MODE(tex
->surface
.u
.gfx9
.fmask
.swizzle_mode
) |
2666 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
2667 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
2669 radeon_set_context_reg_seq(cs
, R_028C60_CB_COLOR0_BASE
+ i
* 0x3C, 15);
2670 radeon_emit(cs
, cb_color_base
); /* CB_COLOR0_BASE */
2671 radeon_emit(cs
, cb_color_base
>> 32); /* CB_COLOR0_BASE_EXT */
2672 radeon_emit(cs
, cb
->cb_color_attrib2
); /* CB_COLOR0_ATTRIB2 */
2673 radeon_emit(cs
, cb
->cb_color_view
); /* CB_COLOR0_VIEW */
2674 radeon_emit(cs
, cb_color_info
); /* CB_COLOR0_INFO */
2675 radeon_emit(cs
, cb_color_attrib
); /* CB_COLOR0_ATTRIB */
2676 radeon_emit(cs
, cb
->cb_dcc_control
); /* CB_COLOR0_DCC_CONTROL */
2677 radeon_emit(cs
, tex
->cmask
.base_address_reg
); /* CB_COLOR0_CMASK */
2678 radeon_emit(cs
, tex
->cmask
.base_address_reg
>> 32); /* CB_COLOR0_CMASK_BASE_EXT */
2679 radeon_emit(cs
, cb_color_fmask
); /* CB_COLOR0_FMASK */
2680 radeon_emit(cs
, cb_color_fmask
>> 32); /* CB_COLOR0_FMASK_BASE_EXT */
2681 radeon_emit(cs
, tex
->color_clear_value
[0]); /* CB_COLOR0_CLEAR_WORD0 */
2682 radeon_emit(cs
, tex
->color_clear_value
[1]); /* CB_COLOR0_CLEAR_WORD1 */
2683 radeon_emit(cs
, cb_dcc_base
); /* CB_COLOR0_DCC_BASE */
2684 radeon_emit(cs
, cb_dcc_base
>> 32); /* CB_COLOR0_DCC_BASE_EXT */
2686 radeon_set_context_reg(cs
, R_0287A0_CB_MRT0_EPITCH
+ i
* 4,
2687 S_0287A0_EPITCH(tex
->surface
.u
.gfx9
.surf
.epitch
));
2689 /* Compute mutable surface parameters (SI-CI-VI). */
2690 const struct legacy_surf_level
*level_info
=
2691 &tex
->surface
.u
.legacy
.level
[cb
->base
.u
.tex
.level
];
2692 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
2693 unsigned cb_color_pitch
, cb_color_slice
, cb_color_fmask_slice
;
2695 cb_color_base
+= level_info
->offset
>> 8;
2697 cb_dcc_base
+= level_info
->dcc_offset
>> 8;
2699 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
2700 slice_tile_max
= level_info
->nblk_x
*
2701 level_info
->nblk_y
/ 64 - 1;
2702 tile_mode_index
= si_tile_mode_index(tex
, cb
->base
.u
.tex
.level
, false);
2704 cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
2705 cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
2706 cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
2708 if (tex
->fmask
.size
) {
2709 if (sctx
->b
.chip_class
>= CIK
)
2710 cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(tex
->fmask
.pitch_in_pixels
/ 8 - 1);
2711 cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tex
->fmask
.tile_mode_index
);
2712 cb_color_fmask_slice
= S_028C88_TILE_MAX(tex
->fmask
.slice_tile_max
);
2714 /* This must be set for fast clear to work without FMASK. */
2715 if (sctx
->b
.chip_class
>= CIK
)
2716 cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
2717 cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
2718 cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
2721 radeon_set_context_reg_seq(cs
, R_028C60_CB_COLOR0_BASE
+ i
* 0x3C,
2722 sctx
->b
.chip_class
>= VI
? 14 : 13);
2723 radeon_emit(cs
, cb_color_base
); /* CB_COLOR0_BASE */
2724 radeon_emit(cs
, cb_color_pitch
); /* CB_COLOR0_PITCH */
2725 radeon_emit(cs
, cb_color_slice
); /* CB_COLOR0_SLICE */
2726 radeon_emit(cs
, cb
->cb_color_view
); /* CB_COLOR0_VIEW */
2727 radeon_emit(cs
, cb_color_info
); /* CB_COLOR0_INFO */
2728 radeon_emit(cs
, cb_color_attrib
); /* CB_COLOR0_ATTRIB */
2729 radeon_emit(cs
, cb
->cb_dcc_control
); /* CB_COLOR0_DCC_CONTROL */
2730 radeon_emit(cs
, tex
->cmask
.base_address_reg
); /* CB_COLOR0_CMASK */
2731 radeon_emit(cs
, tex
->cmask
.slice_tile_max
); /* CB_COLOR0_CMASK_SLICE */
2732 radeon_emit(cs
, cb_color_fmask
); /* CB_COLOR0_FMASK */
2733 radeon_emit(cs
, cb_color_fmask_slice
); /* CB_COLOR0_FMASK_SLICE */
2734 radeon_emit(cs
, tex
->color_clear_value
[0]); /* CB_COLOR0_CLEAR_WORD0 */
2735 radeon_emit(cs
, tex
->color_clear_value
[1]); /* CB_COLOR0_CLEAR_WORD1 */
2737 if (sctx
->b
.chip_class
>= VI
) /* R_028C94_CB_COLOR0_DCC_BASE */
2738 radeon_emit(cs
, cb_dcc_base
);
2742 if (sctx
->framebuffer
.dirty_cbufs
& (1 << i
))
2743 radeon_set_context_reg(cs
, R_028C70_CB_COLOR0_INFO
+ i
* 0x3C, 0);
2746 if (state
->zsbuf
&& sctx
->framebuffer
.dirty_zsbuf
) {
2747 struct r600_surface
*zb
= (struct r600_surface
*)state
->zsbuf
;
2748 struct r600_texture
*rtex
= (struct r600_texture
*)zb
->base
.texture
;
2750 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2751 &rtex
->resource
, RADEON_USAGE_READWRITE
,
2752 zb
->base
.texture
->nr_samples
> 1 ?
2753 RADEON_PRIO_DEPTH_BUFFER_MSAA
:
2754 RADEON_PRIO_DEPTH_BUFFER
);
2756 if (zb
->db_htile_data_base
) {
2757 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2758 rtex
->htile_buffer
, RADEON_USAGE_READWRITE
,
2762 if (sctx
->b
.chip_class
>= GFX9
) {
2763 radeon_set_context_reg_seq(cs
, R_028014_DB_HTILE_DATA_BASE
, 3);
2764 radeon_emit(cs
, zb
->db_htile_data_base
); /* DB_HTILE_DATA_BASE */
2765 radeon_emit(cs
, zb
->db_htile_data_base
>> 32); /* DB_HTILE_DATA_BASE_HI */
2766 radeon_emit(cs
, zb
->db_depth_size
); /* DB_DEPTH_SIZE */
2768 radeon_set_context_reg_seq(cs
, R_028038_DB_Z_INFO
, 10);
2769 radeon_emit(cs
, zb
->db_z_info
| /* DB_Z_INFO */
2770 S_028038_ZRANGE_PRECISION(rtex
->depth_clear_value
!= 0));
2771 radeon_emit(cs
, zb
->db_stencil_info
); /* DB_STENCIL_INFO */
2772 radeon_emit(cs
, zb
->db_depth_base
); /* DB_Z_READ_BASE */
2773 radeon_emit(cs
, zb
->db_depth_base
>> 32); /* DB_Z_READ_BASE_HI */
2774 radeon_emit(cs
, zb
->db_stencil_base
); /* DB_STENCIL_READ_BASE */
2775 radeon_emit(cs
, zb
->db_stencil_base
>> 32); /* DB_STENCIL_READ_BASE_HI */
2776 radeon_emit(cs
, zb
->db_depth_base
); /* DB_Z_WRITE_BASE */
2777 radeon_emit(cs
, zb
->db_depth_base
>> 32); /* DB_Z_WRITE_BASE_HI */
2778 radeon_emit(cs
, zb
->db_stencil_base
); /* DB_STENCIL_WRITE_BASE */
2779 radeon_emit(cs
, zb
->db_stencil_base
>> 32); /* DB_STENCIL_WRITE_BASE_HI */
2781 radeon_set_context_reg_seq(cs
, R_028068_DB_Z_INFO2
, 2);
2782 radeon_emit(cs
, zb
->db_z_info2
); /* DB_Z_INFO2 */
2783 radeon_emit(cs
, zb
->db_stencil_info2
); /* DB_STENCIL_INFO2 */
2785 radeon_set_context_reg(cs
, R_028014_DB_HTILE_DATA_BASE
, zb
->db_htile_data_base
);
2787 radeon_set_context_reg_seq(cs
, R_02803C_DB_DEPTH_INFO
, 9);
2788 radeon_emit(cs
, zb
->db_depth_info
); /* DB_DEPTH_INFO */
2789 radeon_emit(cs
, zb
->db_z_info
| /* DB_Z_INFO */
2790 S_028040_ZRANGE_PRECISION(rtex
->depth_clear_value
!= 0));
2791 radeon_emit(cs
, zb
->db_stencil_info
); /* DB_STENCIL_INFO */
2792 radeon_emit(cs
, zb
->db_depth_base
); /* DB_Z_READ_BASE */
2793 radeon_emit(cs
, zb
->db_stencil_base
); /* DB_STENCIL_READ_BASE */
2794 radeon_emit(cs
, zb
->db_depth_base
); /* DB_Z_WRITE_BASE */
2795 radeon_emit(cs
, zb
->db_stencil_base
); /* DB_STENCIL_WRITE_BASE */
2796 radeon_emit(cs
, zb
->db_depth_size
); /* DB_DEPTH_SIZE */
2797 radeon_emit(cs
, zb
->db_depth_slice
); /* DB_DEPTH_SLICE */
2800 radeon_set_context_reg_seq(cs
, R_028028_DB_STENCIL_CLEAR
, 2);
2801 radeon_emit(cs
, rtex
->stencil_clear_value
); /* R_028028_DB_STENCIL_CLEAR */
2802 radeon_emit(cs
, fui(rtex
->depth_clear_value
)); /* R_02802C_DB_DEPTH_CLEAR */
2804 radeon_set_context_reg(cs
, R_028008_DB_DEPTH_VIEW
, zb
->db_depth_view
);
2805 radeon_set_context_reg(cs
, R_028ABC_DB_HTILE_SURFACE
, zb
->db_htile_surface
);
2806 } else if (sctx
->framebuffer
.dirty_zsbuf
) {
2807 if (sctx
->b
.chip_class
>= GFX9
)
2808 radeon_set_context_reg_seq(cs
, R_028038_DB_Z_INFO
, 2);
2810 radeon_set_context_reg_seq(cs
, R_028040_DB_Z_INFO
, 2);
2812 radeon_emit(cs
, S_028040_FORMAT(V_028040_Z_INVALID
)); /* DB_Z_INFO */
2813 radeon_emit(cs
, S_028044_FORMAT(V_028044_STENCIL_INVALID
)); /* DB_STENCIL_INFO */
2816 /* Framebuffer dimensions. */
2817 /* PA_SC_WINDOW_SCISSOR_TL is set in si_init_config() */
2818 radeon_set_context_reg(cs
, R_028208_PA_SC_WINDOW_SCISSOR_BR
,
2819 S_028208_BR_X(state
->width
) | S_028208_BR_Y(state
->height
));
2821 if (sctx
->b
.chip_class
>= GFX9
) {
2822 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2823 radeon_emit(cs
, EVENT_TYPE(V_028A90_BREAK_BATCH
) | EVENT_INDEX(0));
2826 sctx
->framebuffer
.dirty_cbufs
= 0;
2827 sctx
->framebuffer
.dirty_zsbuf
= false;
2830 static void si_emit_msaa_sample_locs(struct si_context
*sctx
,
2831 struct r600_atom
*atom
)
2833 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
2834 unsigned nr_samples
= sctx
->framebuffer
.nr_samples
;
2836 /* Smoothing (only possible with nr_samples == 1) uses the same
2837 * sample locations as the MSAA it simulates.
2839 if (nr_samples
<= 1 && sctx
->smoothing_enabled
)
2840 nr_samples
= SI_NUM_SMOOTH_AA_SAMPLES
;
2842 /* On Polaris, the small primitive filter uses the sample locations
2843 * even when MSAA is off, so we need to make sure they're set to 0.
2845 if (sctx
->b
.family
>= CHIP_POLARIS10
)
2846 nr_samples
= MAX2(nr_samples
, 1);
2848 if (nr_samples
>= 1 &&
2849 (nr_samples
!= sctx
->msaa_sample_locs
.nr_samples
)) {
2850 sctx
->msaa_sample_locs
.nr_samples
= nr_samples
;
2851 cayman_emit_msaa_sample_locs(cs
, nr_samples
);
2854 if (sctx
->b
.family
>= CHIP_POLARIS10
) {
2855 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
2856 unsigned small_prim_filter_cntl
=
2857 S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
2858 S_028830_LINE_FILTER_DISABLE(sctx
->b
.chip_class
== VI
); /* line bug */
2860 /* The alternative of setting sample locations to 0 would
2861 * require a DB flush to avoid Z errors, see
2862 * https://bugs.freedesktop.org/show_bug.cgi?id=96908
2864 if (sctx
->framebuffer
.nr_samples
> 1 && rs
&& !rs
->multisample_enable
)
2865 small_prim_filter_cntl
&= C_028830_SMALL_PRIM_FILTER_ENABLE
;
2867 radeon_set_context_reg(cs
, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL
,
2868 small_prim_filter_cntl
);
2872 static void si_emit_msaa_config(struct si_context
*sctx
, struct r600_atom
*atom
)
2874 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
2875 unsigned num_tile_pipes
= sctx
->screen
->b
.info
.num_tile_pipes
;
2876 /* 33% faster rendering to linear color buffers */
2877 bool dst_is_linear
= sctx
->framebuffer
.any_dst_linear
;
2878 unsigned sc_mode_cntl_1
=
2879 S_028A4C_WALK_SIZE(dst_is_linear
) |
2880 S_028A4C_WALK_FENCE_ENABLE(!dst_is_linear
) |
2881 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes
== 2 ? 2 : 3) |
2883 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
2884 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
2885 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
2886 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
2887 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
2888 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
2890 cayman_emit_msaa_config(cs
, sctx
->framebuffer
.nr_samples
,
2891 sctx
->ps_iter_samples
,
2892 sctx
->smoothing_enabled
? SI_NUM_SMOOTH_AA_SAMPLES
: 0,
2895 /* GFX9: Flush DFSM when the AA mode changes. */
2896 if (sctx
->b
.chip_class
>= GFX9
) {
2897 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2898 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_DFSM
) | EVENT_INDEX(0));
2902 static void si_set_min_samples(struct pipe_context
*ctx
, unsigned min_samples
)
2904 struct si_context
*sctx
= (struct si_context
*)ctx
;
2906 if (sctx
->ps_iter_samples
== min_samples
)
2909 sctx
->ps_iter_samples
= min_samples
;
2910 sctx
->do_update_shaders
= true;
2912 if (sctx
->framebuffer
.nr_samples
> 1)
2913 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
2921 * Build the sampler view descriptor for a buffer texture.
2922 * @param state 256-bit descriptor; only the high 128 bits are filled in
2925 si_make_buffer_descriptor(struct si_screen
*screen
, struct r600_resource
*buf
,
2926 enum pipe_format format
,
2927 unsigned offset
, unsigned size
,
2930 const struct util_format_description
*desc
;
2933 unsigned num_records
;
2934 unsigned num_format
, data_format
;
2936 desc
= util_format_description(format
);
2937 first_non_void
= util_format_get_first_non_void_channel(format
);
2938 stride
= desc
->block
.bits
/ 8;
2939 num_format
= si_translate_buffer_numformat(&screen
->b
.b
, desc
, first_non_void
);
2940 data_format
= si_translate_buffer_dataformat(&screen
->b
.b
, desc
, first_non_void
);
2942 num_records
= size
/ stride
;
2943 num_records
= MIN2(num_records
, (buf
->b
.b
.width0
- offset
) / stride
);
2945 if (screen
->b
.chip_class
== VI
)
2946 num_records
*= stride
;
2949 state
[5] = S_008F04_STRIDE(stride
);
2950 state
[6] = num_records
;
2951 state
[7] = S_008F0C_DST_SEL_X(si_map_swizzle(desc
->swizzle
[0])) |
2952 S_008F0C_DST_SEL_Y(si_map_swizzle(desc
->swizzle
[1])) |
2953 S_008F0C_DST_SEL_Z(si_map_swizzle(desc
->swizzle
[2])) |
2954 S_008F0C_DST_SEL_W(si_map_swizzle(desc
->swizzle
[3])) |
2955 S_008F0C_NUM_FORMAT(num_format
) |
2956 S_008F0C_DATA_FORMAT(data_format
);
2959 static unsigned gfx9_border_color_swizzle(const unsigned char swizzle
[4])
2961 unsigned bc_swizzle
= V_008F20_BC_SWIZZLE_XYZW
;
2963 if (swizzle
[3] == PIPE_SWIZZLE_X
) {
2964 /* For the pre-defined border color values (white, opaque
2965 * black, transparent black), the only thing that matters is
2966 * that the alpha channel winds up in the correct place
2967 * (because the RGB channels are all the same) so either of
2968 * these enumerations will work.
2970 if (swizzle
[2] == PIPE_SWIZZLE_Y
)
2971 bc_swizzle
= V_008F20_BC_SWIZZLE_WZYX
;
2973 bc_swizzle
= V_008F20_BC_SWIZZLE_WXYZ
;
2974 } else if (swizzle
[0] == PIPE_SWIZZLE_X
) {
2975 if (swizzle
[1] == PIPE_SWIZZLE_Y
)
2976 bc_swizzle
= V_008F20_BC_SWIZZLE_XYZW
;
2978 bc_swizzle
= V_008F20_BC_SWIZZLE_XWYZ
;
2979 } else if (swizzle
[1] == PIPE_SWIZZLE_X
) {
2980 bc_swizzle
= V_008F20_BC_SWIZZLE_YXWZ
;
2981 } else if (swizzle
[2] == PIPE_SWIZZLE_X
) {
2982 bc_swizzle
= V_008F20_BC_SWIZZLE_ZYXW
;
2989 * Build the sampler view descriptor for a texture.
2992 si_make_texture_descriptor(struct si_screen
*screen
,
2993 struct r600_texture
*tex
,
2995 enum pipe_texture_target target
,
2996 enum pipe_format pipe_format
,
2997 const unsigned char state_swizzle
[4],
2998 unsigned first_level
, unsigned last_level
,
2999 unsigned first_layer
, unsigned last_layer
,
3000 unsigned width
, unsigned height
, unsigned depth
,
3002 uint32_t *fmask_state
)
3004 struct pipe_resource
*res
= &tex
->resource
.b
.b
;
3005 const struct util_format_description
*base_desc
, *desc
;
3006 unsigned char swizzle
[4];
3008 unsigned num_format
, data_format
, type
;
3011 desc
= util_format_description(pipe_format
);
3012 base_desc
= util_format_description(res
->format
);
3014 if (desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
) {
3015 const unsigned char swizzle_xxxx
[4] = {0, 0, 0, 0};
3016 const unsigned char swizzle_yyyy
[4] = {1, 1, 1, 1};
3017 const unsigned char swizzle_wwww
[4] = {3, 3, 3, 3};
3019 switch (pipe_format
) {
3020 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
3021 case PIPE_FORMAT_X32_S8X24_UINT
:
3022 case PIPE_FORMAT_X8Z24_UNORM
:
3023 util_format_compose_swizzles(swizzle_yyyy
, state_swizzle
, swizzle
);
3025 case PIPE_FORMAT_X24S8_UINT
:
3027 * X24S8 is implemented as an 8_8_8_8 data format, to
3028 * fix texture gathers. This affects at least
3029 * GL45-CTS.texture_cube_map_array.sampling on VI.
3031 util_format_compose_swizzles(swizzle_wwww
, state_swizzle
, swizzle
);
3034 util_format_compose_swizzles(swizzle_xxxx
, state_swizzle
, swizzle
);
3037 util_format_compose_swizzles(desc
->swizzle
, state_swizzle
, swizzle
);
3040 first_non_void
= util_format_get_first_non_void_channel(pipe_format
);
3042 switch (pipe_format
) {
3043 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
3044 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3047 if (first_non_void
< 0) {
3048 if (util_format_is_compressed(pipe_format
)) {
3049 switch (pipe_format
) {
3050 case PIPE_FORMAT_DXT1_SRGB
:
3051 case PIPE_FORMAT_DXT1_SRGBA
:
3052 case PIPE_FORMAT_DXT3_SRGBA
:
3053 case PIPE_FORMAT_DXT5_SRGBA
:
3054 case PIPE_FORMAT_BPTC_SRGBA
:
3055 case PIPE_FORMAT_ETC2_SRGB8
:
3056 case PIPE_FORMAT_ETC2_SRGB8A1
:
3057 case PIPE_FORMAT_ETC2_SRGBA8
:
3058 num_format
= V_008F14_IMG_NUM_FORMAT_SRGB
;
3060 case PIPE_FORMAT_RGTC1_SNORM
:
3061 case PIPE_FORMAT_LATC1_SNORM
:
3062 case PIPE_FORMAT_RGTC2_SNORM
:
3063 case PIPE_FORMAT_LATC2_SNORM
:
3064 case PIPE_FORMAT_ETC2_R11_SNORM
:
3065 case PIPE_FORMAT_ETC2_RG11_SNORM
:
3066 /* implies float, so use SNORM/UNORM to determine
3067 whether data is signed or not */
3068 case PIPE_FORMAT_BPTC_RGB_FLOAT
:
3069 num_format
= V_008F14_IMG_NUM_FORMAT_SNORM
;
3072 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3075 } else if (desc
->layout
== UTIL_FORMAT_LAYOUT_SUBSAMPLED
) {
3076 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3078 num_format
= V_008F14_IMG_NUM_FORMAT_FLOAT
;
3080 } else if (desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
3081 num_format
= V_008F14_IMG_NUM_FORMAT_SRGB
;
3083 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3085 switch (desc
->channel
[first_non_void
].type
) {
3086 case UTIL_FORMAT_TYPE_FLOAT
:
3087 num_format
= V_008F14_IMG_NUM_FORMAT_FLOAT
;
3089 case UTIL_FORMAT_TYPE_SIGNED
:
3090 if (desc
->channel
[first_non_void
].normalized
)
3091 num_format
= V_008F14_IMG_NUM_FORMAT_SNORM
;
3092 else if (desc
->channel
[first_non_void
].pure_integer
)
3093 num_format
= V_008F14_IMG_NUM_FORMAT_SINT
;
3095 num_format
= V_008F14_IMG_NUM_FORMAT_SSCALED
;
3097 case UTIL_FORMAT_TYPE_UNSIGNED
:
3098 if (desc
->channel
[first_non_void
].normalized
)
3099 num_format
= V_008F14_IMG_NUM_FORMAT_UNORM
;
3100 else if (desc
->channel
[first_non_void
].pure_integer
)
3101 num_format
= V_008F14_IMG_NUM_FORMAT_UINT
;
3103 num_format
= V_008F14_IMG_NUM_FORMAT_USCALED
;
3108 data_format
= si_translate_texformat(&screen
->b
.b
, pipe_format
, desc
, first_non_void
);
3109 if (data_format
== ~0) {
3113 /* Enable clamping for UNORM depth formats promoted to Z32F. */
3114 if (screen
->b
.chip_class
>= GFX9
&&
3115 util_format_has_depth(desc
) &&
3116 num_format
== V_008F14_IMG_NUM_FORMAT_FLOAT
&&
3117 util_get_depth_format_type(base_desc
) != UTIL_FORMAT_TYPE_FLOAT
) {
3118 /* NUM_FORMAT=FLOAT and DATA_FORMAT=24_8 means "clamp to [0,1]". */
3119 data_format
= V_008F14_IMG_DATA_FORMAT_24_8
;
3123 (res
->target
== PIPE_TEXTURE_CUBE
||
3124 res
->target
== PIPE_TEXTURE_CUBE_ARRAY
||
3125 res
->target
== PIPE_TEXTURE_3D
)) {
3126 /* For the purpose of shader images, treat cube maps and 3D
3127 * textures as 2D arrays. For 3D textures, the address
3128 * calculations for mipmaps are different, so we rely on the
3129 * caller to effectively disable mipmaps.
3131 type
= V_008F1C_SQ_RSRC_IMG_2D_ARRAY
;
3133 assert(res
->target
!= PIPE_TEXTURE_3D
|| (first_level
== 0 && last_level
== 0));
3135 type
= si_tex_dim(screen
, tex
, target
, res
->nr_samples
);
3138 if (type
== V_008F1C_SQ_RSRC_IMG_1D_ARRAY
) {
3140 depth
= res
->array_size
;
3141 } else if (type
== V_008F1C_SQ_RSRC_IMG_2D_ARRAY
||
3142 type
== V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
) {
3143 if (sampler
|| res
->target
!= PIPE_TEXTURE_3D
)
3144 depth
= res
->array_size
;
3145 } else if (type
== V_008F1C_SQ_RSRC_IMG_CUBE
)
3146 depth
= res
->array_size
/ 6;
3149 state
[1] = (S_008F14_DATA_FORMAT_GFX6(data_format
) |
3150 S_008F14_NUM_FORMAT_GFX6(num_format
));
3151 state
[2] = (S_008F18_WIDTH(width
- 1) |
3152 S_008F18_HEIGHT(height
- 1) |
3153 S_008F18_PERF_MOD(4));
3154 state
[3] = (S_008F1C_DST_SEL_X(si_map_swizzle(swizzle
[0])) |
3155 S_008F1C_DST_SEL_Y(si_map_swizzle(swizzle
[1])) |
3156 S_008F1C_DST_SEL_Z(si_map_swizzle(swizzle
[2])) |
3157 S_008F1C_DST_SEL_W(si_map_swizzle(swizzle
[3])) |
3158 S_008F1C_BASE_LEVEL(res
->nr_samples
> 1 ?
3160 S_008F1C_LAST_LEVEL(res
->nr_samples
> 1 ?
3161 util_logbase2(res
->nr_samples
) :
3163 S_008F1C_TYPE(type
));
3165 state
[5] = S_008F24_BASE_ARRAY(first_layer
);
3169 if (screen
->b
.chip_class
>= GFX9
) {
3170 unsigned bc_swizzle
= gfx9_border_color_swizzle(desc
->swizzle
);
3172 /* Depth is the the last accessible layer on Gfx9.
3173 * The hw doesn't need to know the total number of layers.
3175 if (type
== V_008F1C_SQ_RSRC_IMG_3D
)
3176 state
[4] |= S_008F20_DEPTH(depth
- 1);
3178 state
[4] |= S_008F20_DEPTH(last_layer
);
3180 state
[4] |= S_008F20_BC_SWIZZLE(bc_swizzle
);
3181 state
[5] |= S_008F24_MAX_MIP(res
->nr_samples
> 1 ?
3182 util_logbase2(res
->nr_samples
) :
3183 tex
->resource
.b
.b
.last_level
);
3185 state
[3] |= S_008F1C_POW2_PAD(res
->last_level
> 0);
3186 state
[4] |= S_008F20_DEPTH(depth
- 1);
3187 state
[5] |= S_008F24_LAST_ARRAY(last_layer
);
3190 if (tex
->dcc_offset
) {
3191 unsigned swap
= r600_translate_colorswap(pipe_format
, false);
3193 state
[6] = S_008F28_ALPHA_IS_ON_MSB(swap
<= 1);
3195 /* The last dword is unused by hw. The shader uses it to clear
3196 * bits in the first dword of sampler state.
3198 if (screen
->b
.chip_class
<= CIK
&& res
->nr_samples
<= 1) {
3199 if (first_level
== last_level
)
3200 state
[7] = C_008F30_MAX_ANISO_RATIO
;
3202 state
[7] = 0xffffffff;
3206 /* Initialize the sampler view for FMASK. */
3207 if (tex
->fmask
.size
) {
3208 uint32_t data_format
, num_format
;
3210 va
= tex
->resource
.gpu_address
+ tex
->fmask
.offset
;
3212 if (screen
->b
.chip_class
>= GFX9
) {
3213 data_format
= V_008F14_IMG_DATA_FORMAT_FMASK
;
3214 switch (res
->nr_samples
) {
3216 num_format
= V_008F14_IMG_FMASK_8_2_2
;
3219 num_format
= V_008F14_IMG_FMASK_8_4_4
;
3222 num_format
= V_008F14_IMG_FMASK_32_8_8
;
3228 switch (res
->nr_samples
) {
3230 data_format
= V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2
;
3233 data_format
= V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4
;
3236 data_format
= V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8
;
3241 num_format
= V_008F14_IMG_NUM_FORMAT_UINT
;
3244 fmask_state
[0] = va
>> 8;
3245 fmask_state
[1] = S_008F14_BASE_ADDRESS_HI(va
>> 40) |
3246 S_008F14_DATA_FORMAT_GFX6(data_format
) |
3247 S_008F14_NUM_FORMAT_GFX6(num_format
);
3248 fmask_state
[2] = S_008F18_WIDTH(width
- 1) |
3249 S_008F18_HEIGHT(height
- 1);
3250 fmask_state
[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X
) |
3251 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X
) |
3252 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X
) |
3253 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X
) |
3254 S_008F1C_TYPE(si_tex_dim(screen
, tex
, target
, 0));
3256 fmask_state
[5] = S_008F24_BASE_ARRAY(first_layer
);
3260 if (screen
->b
.chip_class
>= GFX9
) {
3261 fmask_state
[3] |= S_008F1C_SW_MODE(tex
->surface
.u
.gfx9
.fmask
.swizzle_mode
);
3262 fmask_state
[4] |= S_008F20_DEPTH(last_layer
) |
3263 S_008F20_PITCH_GFX9(tex
->surface
.u
.gfx9
.fmask
.epitch
);
3264 fmask_state
[5] |= S_008F24_META_PIPE_ALIGNED(tex
->surface
.u
.gfx9
.cmask
.pipe_aligned
) |
3265 S_008F24_META_RB_ALIGNED(tex
->surface
.u
.gfx9
.cmask
.rb_aligned
);
3267 fmask_state
[3] |= S_008F1C_TILING_INDEX(tex
->fmask
.tile_mode_index
);
3268 fmask_state
[4] |= S_008F20_DEPTH(depth
- 1) |
3269 S_008F20_PITCH_GFX6(tex
->fmask
.pitch_in_pixels
- 1);
3270 fmask_state
[5] |= S_008F24_LAST_ARRAY(last_layer
);
3276 * Create a sampler view.
3278 * @param ctx context
3279 * @param texture texture
3280 * @param state sampler view template
3281 * @param width0 width0 override (for compressed textures as int)
3282 * @param height0 height0 override (for compressed textures as int)
3283 * @param force_level set the base address to the level (for compressed textures)
3285 struct pipe_sampler_view
*
3286 si_create_sampler_view_custom(struct pipe_context
*ctx
,
3287 struct pipe_resource
*texture
,
3288 const struct pipe_sampler_view
*state
,
3289 unsigned width0
, unsigned height0
,
3290 unsigned force_level
)
3292 struct si_context
*sctx
= (struct si_context
*)ctx
;
3293 struct si_sampler_view
*view
= CALLOC_STRUCT(si_sampler_view
);
3294 struct r600_texture
*tmp
= (struct r600_texture
*)texture
;
3295 unsigned base_level
, first_level
, last_level
;
3296 unsigned char state_swizzle
[4];
3297 unsigned height
, depth
, width
;
3298 unsigned last_layer
= state
->u
.tex
.last_layer
;
3299 enum pipe_format pipe_format
;
3300 const struct legacy_surf_level
*surflevel
;
3305 /* initialize base object */
3306 view
->base
= *state
;
3307 view
->base
.texture
= NULL
;
3308 view
->base
.reference
.count
= 1;
3309 view
->base
.context
= ctx
;
3312 pipe_resource_reference(&view
->base
.texture
, texture
);
3314 if (state
->format
== PIPE_FORMAT_X24S8_UINT
||
3315 state
->format
== PIPE_FORMAT_S8X24_UINT
||
3316 state
->format
== PIPE_FORMAT_X32_S8X24_UINT
||
3317 state
->format
== PIPE_FORMAT_S8_UINT
)
3318 view
->is_stencil_sampler
= true;
3320 /* Buffer resource. */
3321 if (texture
->target
== PIPE_BUFFER
) {
3322 si_make_buffer_descriptor(sctx
->screen
,
3323 (struct r600_resource
*)texture
,
3325 state
->u
.buf
.offset
,
3331 state_swizzle
[0] = state
->swizzle_r
;
3332 state_swizzle
[1] = state
->swizzle_g
;
3333 state_swizzle
[2] = state
->swizzle_b
;
3334 state_swizzle
[3] = state
->swizzle_a
;
3337 first_level
= state
->u
.tex
.first_level
;
3338 last_level
= state
->u
.tex
.last_level
;
3341 depth
= texture
->depth0
;
3343 if (sctx
->b
.chip_class
<= VI
&& force_level
) {
3344 assert(force_level
== first_level
&&
3345 force_level
== last_level
);
3346 base_level
= force_level
;
3349 width
= u_minify(width
, force_level
);
3350 height
= u_minify(height
, force_level
);
3351 depth
= u_minify(depth
, force_level
);
3354 /* This is not needed if state trackers set last_layer correctly. */
3355 if (state
->target
== PIPE_TEXTURE_1D
||
3356 state
->target
== PIPE_TEXTURE_2D
||
3357 state
->target
== PIPE_TEXTURE_RECT
||
3358 state
->target
== PIPE_TEXTURE_CUBE
)
3359 last_layer
= state
->u
.tex
.first_layer
;
3361 /* Texturing with separate depth and stencil. */
3362 pipe_format
= state
->format
;
3364 /* Depth/stencil texturing sometimes needs separate texture. */
3365 if (tmp
->is_depth
&& !r600_can_sample_zs(tmp
, view
->is_stencil_sampler
)) {
3366 if (!tmp
->flushed_depth_texture
&&
3367 !r600_init_flushed_depth_texture(ctx
, texture
, NULL
)) {
3368 pipe_resource_reference(&view
->base
.texture
, NULL
);
3373 assert(tmp
->flushed_depth_texture
);
3375 /* Override format for the case where the flushed texture
3376 * contains only Z or only S.
3378 if (tmp
->flushed_depth_texture
->resource
.b
.b
.format
!= tmp
->resource
.b
.b
.format
)
3379 pipe_format
= tmp
->flushed_depth_texture
->resource
.b
.b
.format
;
3381 tmp
= tmp
->flushed_depth_texture
;
3384 surflevel
= tmp
->surface
.u
.legacy
.level
;
3386 if (tmp
->db_compatible
) {
3387 if (!view
->is_stencil_sampler
)
3388 pipe_format
= tmp
->db_render_format
;
3390 switch (pipe_format
) {
3391 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
3392 pipe_format
= PIPE_FORMAT_Z32_FLOAT
;
3394 case PIPE_FORMAT_X8Z24_UNORM
:
3395 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
3396 /* Z24 is always stored like this for DB
3399 pipe_format
= PIPE_FORMAT_Z24X8_UNORM
;
3401 case PIPE_FORMAT_X24S8_UINT
:
3402 case PIPE_FORMAT_S8X24_UINT
:
3403 case PIPE_FORMAT_X32_S8X24_UINT
:
3404 pipe_format
= PIPE_FORMAT_S8_UINT
;
3405 surflevel
= tmp
->surface
.u
.legacy
.stencil_level
;
3411 vi_disable_dcc_if_incompatible_format(&sctx
->b
, texture
,
3412 state
->u
.tex
.first_level
,
3415 si_make_texture_descriptor(sctx
->screen
, tmp
, true,
3416 state
->target
, pipe_format
, state_swizzle
,
3417 first_level
, last_level
,
3418 state
->u
.tex
.first_layer
, last_layer
,
3419 width
, height
, depth
,
3420 view
->state
, view
->fmask_state
);
3422 view
->base_level_info
= &surflevel
[base_level
];
3423 view
->base_level
= base_level
;
3424 view
->block_width
= util_format_get_blockwidth(pipe_format
);
3428 static struct pipe_sampler_view
*
3429 si_create_sampler_view(struct pipe_context
*ctx
,
3430 struct pipe_resource
*texture
,
3431 const struct pipe_sampler_view
*state
)
3433 return si_create_sampler_view_custom(ctx
, texture
, state
,
3434 texture
? texture
->width0
: 0,
3435 texture
? texture
->height0
: 0, 0);
3438 static void si_sampler_view_destroy(struct pipe_context
*ctx
,
3439 struct pipe_sampler_view
*state
)
3441 struct si_sampler_view
*view
= (struct si_sampler_view
*)state
;
3443 pipe_resource_reference(&state
->texture
, NULL
);
3447 static bool wrap_mode_uses_border_color(unsigned wrap
, bool linear_filter
)
3449 return wrap
== PIPE_TEX_WRAP_CLAMP_TO_BORDER
||
3450 wrap
== PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
||
3452 (wrap
== PIPE_TEX_WRAP_CLAMP
||
3453 wrap
== PIPE_TEX_WRAP_MIRROR_CLAMP
));
3456 static bool sampler_state_needs_border_color(const struct pipe_sampler_state
*state
)
3458 bool linear_filter
= state
->min_img_filter
!= PIPE_TEX_FILTER_NEAREST
||
3459 state
->mag_img_filter
!= PIPE_TEX_FILTER_NEAREST
;
3461 return (state
->border_color
.ui
[0] || state
->border_color
.ui
[1] ||
3462 state
->border_color
.ui
[2] || state
->border_color
.ui
[3]) &&
3463 (wrap_mode_uses_border_color(state
->wrap_s
, linear_filter
) ||
3464 wrap_mode_uses_border_color(state
->wrap_t
, linear_filter
) ||
3465 wrap_mode_uses_border_color(state
->wrap_r
, linear_filter
));
3468 static void *si_create_sampler_state(struct pipe_context
*ctx
,
3469 const struct pipe_sampler_state
*state
)
3471 struct si_context
*sctx
= (struct si_context
*)ctx
;
3472 struct r600_common_screen
*rscreen
= sctx
->b
.screen
;
3473 struct si_sampler_state
*rstate
= CALLOC_STRUCT(si_sampler_state
);
3474 unsigned border_color_type
, border_color_index
= 0;
3475 unsigned max_aniso
= rscreen
->force_aniso
>= 0 ? rscreen
->force_aniso
3476 : state
->max_anisotropy
;
3477 unsigned max_aniso_ratio
= r600_tex_aniso_filter(max_aniso
);
3483 if (!sampler_state_needs_border_color(state
))
3484 border_color_type
= V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
3485 else if (state
->border_color
.f
[0] == 0 &&
3486 state
->border_color
.f
[1] == 0 &&
3487 state
->border_color
.f
[2] == 0 &&
3488 state
->border_color
.f
[3] == 0)
3489 border_color_type
= V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
3490 else if (state
->border_color
.f
[0] == 0 &&
3491 state
->border_color
.f
[1] == 0 &&
3492 state
->border_color
.f
[2] == 0 &&
3493 state
->border_color
.f
[3] == 1)
3494 border_color_type
= V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
3495 else if (state
->border_color
.f
[0] == 1 &&
3496 state
->border_color
.f
[1] == 1 &&
3497 state
->border_color
.f
[2] == 1 &&
3498 state
->border_color
.f
[3] == 1)
3499 border_color_type
= V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
3503 border_color_type
= V_008F3C_SQ_TEX_BORDER_COLOR_REGISTER
;
3505 /* Check if the border has been uploaded already. */
3506 for (i
= 0; i
< sctx
->border_color_count
; i
++)
3507 if (memcmp(&sctx
->border_color_table
[i
], &state
->border_color
,
3508 sizeof(state
->border_color
)) == 0)
3511 if (i
>= SI_MAX_BORDER_COLORS
) {
3512 /* Getting 4096 unique border colors is very unlikely. */
3513 fprintf(stderr
, "radeonsi: The border color table is full. "
3514 "Any new border colors will be just black. "
3515 "Please file a bug.\n");
3516 border_color_type
= V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
3518 if (i
== sctx
->border_color_count
) {
3519 /* Upload a new border color. */
3520 memcpy(&sctx
->border_color_table
[i
], &state
->border_color
,
3521 sizeof(state
->border_color
));
3522 util_memcpy_cpu_to_le32(&sctx
->border_color_map
[i
],
3523 &state
->border_color
,
3524 sizeof(state
->border_color
));
3525 sctx
->border_color_count
++;
3528 border_color_index
= i
;
3533 rstate
->magic
= SI_SAMPLER_STATE_MAGIC
;
3535 rstate
->val
[0] = (S_008F30_CLAMP_X(si_tex_wrap(state
->wrap_s
)) |
3536 S_008F30_CLAMP_Y(si_tex_wrap(state
->wrap_t
)) |
3537 S_008F30_CLAMP_Z(si_tex_wrap(state
->wrap_r
)) |
3538 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
3539 S_008F30_DEPTH_COMPARE_FUNC(si_tex_compare(state
->compare_func
)) |
3540 S_008F30_FORCE_UNNORMALIZED(!state
->normalized_coords
) |
3541 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
3542 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
3543 S_008F30_DISABLE_CUBE_WRAP(!state
->seamless_cube_map
) |
3544 S_008F30_COMPAT_MODE(sctx
->b
.chip_class
>= VI
));
3545 rstate
->val
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(state
->min_lod
, 0, 15), 8)) |
3546 S_008F34_MAX_LOD(S_FIXED(CLAMP(state
->max_lod
, 0, 15), 8)) |
3547 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
3548 rstate
->val
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(state
->lod_bias
, -16, 16), 8)) |
3549 S_008F38_XY_MAG_FILTER(eg_tex_filter(state
->mag_img_filter
, max_aniso
)) |
3550 S_008F38_XY_MIN_FILTER(eg_tex_filter(state
->min_img_filter
, max_aniso
)) |
3551 S_008F38_MIP_FILTER(si_tex_mipfilter(state
->min_mip_filter
)) |
3552 S_008F38_MIP_POINT_PRECLAMP(1) |
3553 S_008F38_DISABLE_LSB_CEIL(sctx
->b
.chip_class
<= VI
) |
3554 S_008F38_FILTER_PREC_FIX(1) |
3555 S_008F38_ANISO_OVERRIDE(sctx
->b
.chip_class
>= VI
));
3556 rstate
->val
[3] = S_008F3C_BORDER_COLOR_PTR(border_color_index
) |
3557 S_008F3C_BORDER_COLOR_TYPE(border_color_type
);
3561 static void si_set_sample_mask(struct pipe_context
*ctx
, unsigned sample_mask
)
3563 struct si_context
*sctx
= (struct si_context
*)ctx
;
3565 if (sctx
->sample_mask
.sample_mask
== (uint16_t)sample_mask
)
3568 sctx
->sample_mask
.sample_mask
= sample_mask
;
3569 si_mark_atom_dirty(sctx
, &sctx
->sample_mask
.atom
);
3572 static void si_emit_sample_mask(struct si_context
*sctx
, struct r600_atom
*atom
)
3574 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
3575 unsigned mask
= sctx
->sample_mask
.sample_mask
;
3577 /* Needed for line and polygon smoothing as well as for the Polaris
3578 * small primitive filter. We expect the state tracker to take care of
3581 assert(mask
== 0xffff || sctx
->framebuffer
.nr_samples
> 1 ||
3582 (mask
& 1 && sctx
->blitter
->running
));
3584 radeon_set_context_reg_seq(cs
, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0
, 2);
3585 radeon_emit(cs
, mask
| (mask
<< 16));
3586 radeon_emit(cs
, mask
| (mask
<< 16));
3589 static void si_delete_sampler_state(struct pipe_context
*ctx
, void *state
)
3592 struct si_sampler_state
*s
= state
;
3594 assert(s
->magic
== SI_SAMPLER_STATE_MAGIC
);
3601 * Vertex elements & buffers
3604 static void *si_create_vertex_elements(struct pipe_context
*ctx
,
3606 const struct pipe_vertex_element
*elements
)
3608 struct si_screen
*sscreen
= (struct si_screen
*)ctx
->screen
;
3609 struct si_vertex_element
*v
= CALLOC_STRUCT(si_vertex_element
);
3610 bool used
[SI_NUM_VERTEX_BUFFERS
] = {};
3613 assert(count
<= SI_MAX_ATTRIBS
);
3618 v
->desc_list_byte_size
= align(count
* 16, SI_CPDMA_ALIGNMENT
);
3620 for (i
= 0; i
< count
; ++i
) {
3621 const struct util_format_description
*desc
;
3622 const struct util_format_channel_description
*channel
;
3623 unsigned data_format
, num_format
;
3625 unsigned vbo_index
= elements
[i
].vertex_buffer_index
;
3626 unsigned char swizzle
[4];
3628 if (vbo_index
>= SI_NUM_VERTEX_BUFFERS
) {
3633 if (!used
[vbo_index
]) {
3634 v
->first_vb_use_mask
|= 1 << i
;
3635 used
[vbo_index
] = true;
3638 desc
= util_format_description(elements
[i
].src_format
);
3639 first_non_void
= util_format_get_first_non_void_channel(elements
[i
].src_format
);
3640 data_format
= si_translate_buffer_dataformat(ctx
->screen
, desc
, first_non_void
);
3641 num_format
= si_translate_buffer_numformat(ctx
->screen
, desc
, first_non_void
);
3642 channel
= first_non_void
>= 0 ? &desc
->channel
[first_non_void
] : NULL
;
3643 memcpy(swizzle
, desc
->swizzle
, sizeof(swizzle
));
3645 v
->format_size
[i
] = desc
->block
.bits
/ 8;
3647 /* The hardware always treats the 2-bit alpha channel as
3648 * unsigned, so a shader workaround is needed. The affected
3649 * chips are VI and older except Stoney (GFX8.1).
3651 if (data_format
== V_008F0C_BUF_DATA_FORMAT_2_10_10_10
&&
3652 sscreen
->b
.chip_class
<= VI
&&
3653 sscreen
->b
.family
!= CHIP_STONEY
) {
3654 if (num_format
== V_008F0C_BUF_NUM_FORMAT_SNORM
) {
3655 v
->fix_fetch
[i
] = SI_FIX_FETCH_A2_SNORM
;
3656 } else if (num_format
== V_008F0C_BUF_NUM_FORMAT_SSCALED
) {
3657 v
->fix_fetch
[i
] = SI_FIX_FETCH_A2_SSCALED
;
3658 } else if (num_format
== V_008F0C_BUF_NUM_FORMAT_SINT
) {
3659 /* This isn't actually used in OpenGL. */
3660 v
->fix_fetch
[i
] = SI_FIX_FETCH_A2_SINT
;
3662 } else if (channel
&& channel
->type
== UTIL_FORMAT_TYPE_FIXED
) {
3663 if (desc
->swizzle
[3] == PIPE_SWIZZLE_1
)
3664 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBX_32_FIXED
;
3666 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_FIXED
;
3667 } else if (channel
&& channel
->size
== 32 && !channel
->pure_integer
) {
3668 if (channel
->type
== UTIL_FORMAT_TYPE_SIGNED
) {
3669 if (channel
->normalized
) {
3670 if (desc
->swizzle
[3] == PIPE_SWIZZLE_1
)
3671 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBX_32_SNORM
;
3673 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_SNORM
;
3675 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_SSCALED
;
3677 } else if (channel
->type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
3678 if (channel
->normalized
) {
3679 if (desc
->swizzle
[3] == PIPE_SWIZZLE_1
)
3680 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBX_32_UNORM
;
3682 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_UNORM
;
3684 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_32_USCALED
;
3687 } else if (channel
&& channel
->size
== 64 &&
3688 channel
->type
== UTIL_FORMAT_TYPE_FLOAT
) {
3689 switch (desc
->nr_channels
) {
3692 v
->fix_fetch
[i
] = SI_FIX_FETCH_RG_64_FLOAT
;
3693 swizzle
[0] = PIPE_SWIZZLE_X
;
3694 swizzle
[1] = PIPE_SWIZZLE_Y
;
3695 swizzle
[2] = desc
->nr_channels
== 2 ? PIPE_SWIZZLE_Z
: PIPE_SWIZZLE_0
;
3696 swizzle
[3] = desc
->nr_channels
== 2 ? PIPE_SWIZZLE_W
: PIPE_SWIZZLE_0
;
3699 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_64_FLOAT
;
3700 swizzle
[0] = PIPE_SWIZZLE_X
; /* 3 loads */
3701 swizzle
[1] = PIPE_SWIZZLE_Y
;
3702 swizzle
[2] = PIPE_SWIZZLE_0
;
3703 swizzle
[3] = PIPE_SWIZZLE_0
;
3706 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGBA_64_FLOAT
;
3707 swizzle
[0] = PIPE_SWIZZLE_X
; /* 2 loads */
3708 swizzle
[1] = PIPE_SWIZZLE_Y
;
3709 swizzle
[2] = PIPE_SWIZZLE_Z
;
3710 swizzle
[3] = PIPE_SWIZZLE_W
;
3715 } else if (channel
&& desc
->nr_channels
== 3) {
3716 assert(desc
->swizzle
[0] == PIPE_SWIZZLE_X
);
3718 if (channel
->size
== 8) {
3719 if (channel
->pure_integer
)
3720 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_8_INT
;
3722 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_8
;
3723 } else if (channel
->size
== 16) {
3724 if (channel
->pure_integer
)
3725 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_16_INT
;
3727 v
->fix_fetch
[i
] = SI_FIX_FETCH_RGB_16
;
3731 v
->rsrc_word3
[i
] = S_008F0C_DST_SEL_X(si_map_swizzle(swizzle
[0])) |
3732 S_008F0C_DST_SEL_Y(si_map_swizzle(swizzle
[1])) |
3733 S_008F0C_DST_SEL_Z(si_map_swizzle(swizzle
[2])) |
3734 S_008F0C_DST_SEL_W(si_map_swizzle(swizzle
[3])) |
3735 S_008F0C_NUM_FORMAT(num_format
) |
3736 S_008F0C_DATA_FORMAT(data_format
);
3738 memcpy(v
->elements
, elements
, sizeof(struct pipe_vertex_element
) * count
);
3743 static void si_bind_vertex_elements(struct pipe_context
*ctx
, void *state
)
3745 struct si_context
*sctx
= (struct si_context
*)ctx
;
3746 struct si_vertex_element
*v
= (struct si_vertex_element
*)state
;
3748 sctx
->vertex_elements
= v
;
3749 sctx
->vertex_buffers_dirty
= true;
3750 sctx
->do_update_shaders
= true;
3753 static void si_delete_vertex_element(struct pipe_context
*ctx
, void *state
)
3755 struct si_context
*sctx
= (struct si_context
*)ctx
;
3757 if (sctx
->vertex_elements
== state
)
3758 sctx
->vertex_elements
= NULL
;
3762 static void si_set_vertex_buffers(struct pipe_context
*ctx
,
3763 unsigned start_slot
, unsigned count
,
3764 const struct pipe_vertex_buffer
*buffers
)
3766 struct si_context
*sctx
= (struct si_context
*)ctx
;
3767 struct pipe_vertex_buffer
*dst
= sctx
->vertex_buffer
+ start_slot
;
3770 assert(start_slot
+ count
<= ARRAY_SIZE(sctx
->vertex_buffer
));
3773 for (i
= 0; i
< count
; i
++) {
3774 const struct pipe_vertex_buffer
*src
= buffers
+ i
;
3775 struct pipe_vertex_buffer
*dsti
= dst
+ i
;
3777 if (unlikely(src
->user_buffer
)) {
3778 /* Zero-stride attribs only. */
3779 assert(src
->stride
== 0);
3781 /* Assume that the user_buffer comes from
3782 * gl_current_attrib, which implies it has
3783 * 4 * 8 bytes (for dvec4 attributes).
3785 * Use const_uploader to upload into VRAM directly.
3787 u_upload_data(sctx
->b
.b
.const_uploader
, 0, 32, 32,
3789 &dsti
->buffer_offset
,
3793 struct pipe_resource
*buf
= src
->buffer
;
3795 pipe_resource_reference(&dsti
->buffer
, buf
);
3796 dsti
->buffer_offset
= src
->buffer_offset
;
3797 dsti
->stride
= src
->stride
;
3798 r600_context_add_resource_size(ctx
, buf
);
3800 r600_resource(buf
)->bind_history
|= PIPE_BIND_VERTEX_BUFFER
;
3804 for (i
= 0; i
< count
; i
++) {
3805 pipe_resource_reference(&dst
[i
].buffer
, NULL
);
3808 sctx
->vertex_buffers_dirty
= true;
3811 static void si_set_index_buffer(struct pipe_context
*ctx
,
3812 const struct pipe_index_buffer
*ib
)
3814 struct si_context
*sctx
= (struct si_context
*)ctx
;
3817 struct pipe_resource
*buf
= ib
->buffer
;
3819 pipe_resource_reference(&sctx
->index_buffer
.buffer
, buf
);
3820 memcpy(&sctx
->index_buffer
, ib
, sizeof(*ib
));
3821 r600_context_add_resource_size(ctx
, buf
);
3823 r600_resource(buf
)->bind_history
|= PIPE_BIND_INDEX_BUFFER
;
3825 pipe_resource_reference(&sctx
->index_buffer
.buffer
, NULL
);
3833 static void si_set_tess_state(struct pipe_context
*ctx
,
3834 const float default_outer_level
[4],
3835 const float default_inner_level
[2])
3837 struct si_context
*sctx
= (struct si_context
*)ctx
;
3838 struct pipe_constant_buffer cb
;
3841 memcpy(array
, default_outer_level
, sizeof(float) * 4);
3842 memcpy(array
+4, default_inner_level
, sizeof(float) * 2);
3845 cb
.user_buffer
= NULL
;
3846 cb
.buffer_size
= sizeof(array
);
3848 si_upload_const_buffer(sctx
, (struct r600_resource
**)&cb
.buffer
,
3849 (void*)array
, sizeof(array
),
3852 si_set_rw_buffer(sctx
, SI_HS_CONST_DEFAULT_TESS_LEVELS
, &cb
);
3853 pipe_resource_reference(&cb
.buffer
, NULL
);
3856 static void si_texture_barrier(struct pipe_context
*ctx
, unsigned flags
)
3858 struct si_context
*sctx
= (struct si_context
*)ctx
;
3860 sctx
->b
.flags
|= SI_CONTEXT_INV_VMEM_L1
|
3861 SI_CONTEXT_INV_GLOBAL_L2
|
3862 SI_CONTEXT_FLUSH_AND_INV_CB
;
3863 sctx
->framebuffer
.do_update_surf_dirtiness
= true;
3866 /* This only ensures coherency for shader image/buffer stores. */
3867 static void si_memory_barrier(struct pipe_context
*ctx
, unsigned flags
)
3869 struct si_context
*sctx
= (struct si_context
*)ctx
;
3871 /* Subsequent commands must wait for all shader invocations to
3873 sctx
->b
.flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
|
3874 SI_CONTEXT_CS_PARTIAL_FLUSH
;
3876 if (flags
& PIPE_BARRIER_CONSTANT_BUFFER
)
3877 sctx
->b
.flags
|= SI_CONTEXT_INV_SMEM_L1
|
3878 SI_CONTEXT_INV_VMEM_L1
;
3880 if (flags
& (PIPE_BARRIER_VERTEX_BUFFER
|
3881 PIPE_BARRIER_SHADER_BUFFER
|
3882 PIPE_BARRIER_TEXTURE
|
3883 PIPE_BARRIER_IMAGE
|
3884 PIPE_BARRIER_STREAMOUT_BUFFER
|
3885 PIPE_BARRIER_GLOBAL_BUFFER
)) {
3886 /* As far as I can tell, L1 contents are written back to L2
3887 * automatically at end of shader, but the contents of other
3888 * L1 caches might still be stale. */
3889 sctx
->b
.flags
|= SI_CONTEXT_INV_VMEM_L1
;
3892 if (flags
& PIPE_BARRIER_INDEX_BUFFER
) {
3893 /* Indices are read through TC L2 since VI.
3896 if (sctx
->screen
->b
.chip_class
<= CIK
)
3897 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
3900 if (flags
& PIPE_BARRIER_FRAMEBUFFER
)
3901 sctx
->b
.flags
|= SI_CONTEXT_FLUSH_AND_INV_CB
|
3902 SI_CONTEXT_FLUSH_AND_INV_DB
;
3904 if (flags
& (PIPE_BARRIER_FRAMEBUFFER
|
3905 PIPE_BARRIER_INDIRECT_BUFFER
))
3906 sctx
->b
.flags
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
3909 static void *si_create_blend_custom(struct si_context
*sctx
, unsigned mode
)
3911 struct pipe_blend_state blend
;
3913 memset(&blend
, 0, sizeof(blend
));
3914 blend
.independent_blend_enable
= true;
3915 blend
.rt
[0].colormask
= 0xf;
3916 return si_create_blend_state_mode(&sctx
->b
.b
, &blend
, mode
);
3919 static void si_need_gfx_cs_space(struct pipe_context
*ctx
, unsigned num_dw
,
3920 bool include_draw_vbo
)
3922 si_need_cs_space((struct si_context
*)ctx
);
3925 static void si_init_config(struct si_context
*sctx
);
3927 void si_init_state_functions(struct si_context
*sctx
)
3929 si_init_external_atom(sctx
, &sctx
->b
.render_cond_atom
, &sctx
->atoms
.s
.render_cond
);
3930 si_init_external_atom(sctx
, &sctx
->b
.streamout
.begin_atom
, &sctx
->atoms
.s
.streamout_begin
);
3931 si_init_external_atom(sctx
, &sctx
->b
.streamout
.enable_atom
, &sctx
->atoms
.s
.streamout_enable
);
3932 si_init_external_atom(sctx
, &sctx
->b
.scissors
.atom
, &sctx
->atoms
.s
.scissors
);
3933 si_init_external_atom(sctx
, &sctx
->b
.viewports
.atom
, &sctx
->atoms
.s
.viewports
);
3935 si_init_atom(sctx
, &sctx
->framebuffer
.atom
, &sctx
->atoms
.s
.framebuffer
, si_emit_framebuffer_state
);
3936 si_init_atom(sctx
, &sctx
->msaa_sample_locs
.atom
, &sctx
->atoms
.s
.msaa_sample_locs
, si_emit_msaa_sample_locs
);
3937 si_init_atom(sctx
, &sctx
->db_render_state
, &sctx
->atoms
.s
.db_render_state
, si_emit_db_render_state
);
3938 si_init_atom(sctx
, &sctx
->msaa_config
, &sctx
->atoms
.s
.msaa_config
, si_emit_msaa_config
);
3939 si_init_atom(sctx
, &sctx
->sample_mask
.atom
, &sctx
->atoms
.s
.sample_mask
, si_emit_sample_mask
);
3940 si_init_atom(sctx
, &sctx
->cb_render_state
, &sctx
->atoms
.s
.cb_render_state
, si_emit_cb_render_state
);
3941 si_init_atom(sctx
, &sctx
->blend_color
.atom
, &sctx
->atoms
.s
.blend_color
, si_emit_blend_color
);
3942 si_init_atom(sctx
, &sctx
->clip_regs
, &sctx
->atoms
.s
.clip_regs
, si_emit_clip_regs
);
3943 si_init_atom(sctx
, &sctx
->clip_state
.atom
, &sctx
->atoms
.s
.clip_state
, si_emit_clip_state
);
3944 si_init_atom(sctx
, &sctx
->stencil_ref
.atom
, &sctx
->atoms
.s
.stencil_ref
, si_emit_stencil_ref
);
3946 sctx
->b
.b
.create_blend_state
= si_create_blend_state
;
3947 sctx
->b
.b
.bind_blend_state
= si_bind_blend_state
;
3948 sctx
->b
.b
.delete_blend_state
= si_delete_blend_state
;
3949 sctx
->b
.b
.set_blend_color
= si_set_blend_color
;
3951 sctx
->b
.b
.create_rasterizer_state
= si_create_rs_state
;
3952 sctx
->b
.b
.bind_rasterizer_state
= si_bind_rs_state
;
3953 sctx
->b
.b
.delete_rasterizer_state
= si_delete_rs_state
;
3955 sctx
->b
.b
.create_depth_stencil_alpha_state
= si_create_dsa_state
;
3956 sctx
->b
.b
.bind_depth_stencil_alpha_state
= si_bind_dsa_state
;
3957 sctx
->b
.b
.delete_depth_stencil_alpha_state
= si_delete_dsa_state
;
3959 sctx
->custom_dsa_flush
= si_create_db_flush_dsa(sctx
);
3960 sctx
->custom_blend_resolve
= si_create_blend_custom(sctx
, V_028808_CB_RESOLVE
);
3961 sctx
->custom_blend_decompress
= si_create_blend_custom(sctx
, V_028808_CB_FMASK_DECOMPRESS
);
3962 sctx
->custom_blend_fastclear
= si_create_blend_custom(sctx
, V_028808_CB_ELIMINATE_FAST_CLEAR
);
3963 sctx
->custom_blend_dcc_decompress
= si_create_blend_custom(sctx
, V_028808_CB_DCC_DECOMPRESS
);
3965 sctx
->b
.b
.set_clip_state
= si_set_clip_state
;
3966 sctx
->b
.b
.set_stencil_ref
= si_set_stencil_ref
;
3968 sctx
->b
.b
.set_framebuffer_state
= si_set_framebuffer_state
;
3969 sctx
->b
.b
.get_sample_position
= cayman_get_sample_position
;
3971 sctx
->b
.b
.create_sampler_state
= si_create_sampler_state
;
3972 sctx
->b
.b
.delete_sampler_state
= si_delete_sampler_state
;
3974 sctx
->b
.b
.create_sampler_view
= si_create_sampler_view
;
3975 sctx
->b
.b
.sampler_view_destroy
= si_sampler_view_destroy
;
3977 sctx
->b
.b
.set_sample_mask
= si_set_sample_mask
;
3979 sctx
->b
.b
.create_vertex_elements_state
= si_create_vertex_elements
;
3980 sctx
->b
.b
.bind_vertex_elements_state
= si_bind_vertex_elements
;
3981 sctx
->b
.b
.delete_vertex_elements_state
= si_delete_vertex_element
;
3982 sctx
->b
.b
.set_vertex_buffers
= si_set_vertex_buffers
;
3983 sctx
->b
.b
.set_index_buffer
= si_set_index_buffer
;
3985 sctx
->b
.b
.texture_barrier
= si_texture_barrier
;
3986 sctx
->b
.b
.memory_barrier
= si_memory_barrier
;
3987 sctx
->b
.b
.set_min_samples
= si_set_min_samples
;
3988 sctx
->b
.b
.set_tess_state
= si_set_tess_state
;
3990 sctx
->b
.b
.set_active_query_state
= si_set_active_query_state
;
3991 sctx
->b
.set_occlusion_query_state
= si_set_occlusion_query_state
;
3992 sctx
->b
.save_qbo_state
= si_save_qbo_state
;
3993 sctx
->b
.need_gfx_cs_space
= si_need_gfx_cs_space
;
3995 sctx
->b
.b
.draw_vbo
= si_draw_vbo
;
3997 si_init_config(sctx
);
4000 static uint32_t si_get_bo_metadata_word1(struct r600_common_screen
*rscreen
)
4002 return (ATI_VENDOR_ID
<< 16) | rscreen
->info
.pci_id
;
4005 static void si_query_opaque_metadata(struct r600_common_screen
*rscreen
,
4006 struct r600_texture
*rtex
,
4007 struct radeon_bo_metadata
*md
)
4009 struct si_screen
*sscreen
= (struct si_screen
*)rscreen
;
4010 struct pipe_resource
*res
= &rtex
->resource
.b
.b
;
4011 static const unsigned char swizzle
[] = {
4017 uint32_t desc
[8], i
;
4018 bool is_array
= util_resource_is_array_texture(res
);
4020 /* DRM 2.x.x doesn't support this. */
4021 if (rscreen
->info
.drm_major
!= 3)
4024 assert(rtex
->dcc_separate_buffer
== NULL
);
4025 assert(rtex
->fmask
.size
== 0);
4027 /* Metadata image format format version 1:
4028 * [0] = 1 (metadata format identifier)
4029 * [1] = (VENDOR_ID << 16) | PCI_ID
4030 * [2:9] = image descriptor for the whole resource
4031 * [2] is always 0, because the base address is cleared
4032 * [9] is the DCC offset bits [39:8] from the beginning of
4034 * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
4037 md
->metadata
[0] = 1; /* metadata image format version 1 */
4039 /* TILE_MODE_INDEX is ambiguous without a PCI ID. */
4040 md
->metadata
[1] = si_get_bo_metadata_word1(rscreen
);
4042 si_make_texture_descriptor(sscreen
, rtex
, true,
4043 res
->target
, res
->format
,
4044 swizzle
, 0, res
->last_level
, 0,
4045 is_array
? res
->array_size
- 1 : 0,
4046 res
->width0
, res
->height0
, res
->depth0
,
4049 si_set_mutable_tex_desc_fields(sscreen
, rtex
, &rtex
->surface
.u
.legacy
.level
[0],
4050 0, 0, rtex
->surface
.blk_w
, false, desc
);
4052 /* Clear the base address and set the relative DCC offset. */
4054 desc
[1] &= C_008F14_BASE_ADDRESS_HI
;
4055 desc
[7] = rtex
->dcc_offset
>> 8;
4057 /* Dwords [2:9] contain the image descriptor. */
4058 memcpy(&md
->metadata
[2], desc
, sizeof(desc
));
4059 md
->size_metadata
= 10 * 4;
4061 /* Dwords [10:..] contain the mipmap level offsets. */
4062 if (rscreen
->chip_class
<= VI
) {
4063 for (i
= 0; i
<= res
->last_level
; i
++)
4064 md
->metadata
[10+i
] = rtex
->surface
.u
.legacy
.level
[i
].offset
>> 8;
4066 md
->size_metadata
+= (1 + res
->last_level
) * 4;
4070 static void si_apply_opaque_metadata(struct r600_common_screen
*rscreen
,
4071 struct r600_texture
*rtex
,
4072 struct radeon_bo_metadata
*md
)
4074 uint32_t *desc
= &md
->metadata
[2];
4076 if (rscreen
->chip_class
< VI
)
4079 /* Return if DCC is enabled. The texture should be set up with it
4082 if (md
->size_metadata
>= 11 * 4 &&
4083 md
->metadata
[0] != 0 &&
4084 md
->metadata
[1] == si_get_bo_metadata_word1(rscreen
) &&
4085 G_008F28_COMPRESSION_EN(desc
[6])) {
4086 assert(rtex
->dcc_offset
== ((uint64_t)desc
[7] << 8));
4090 /* Disable DCC. These are always set by texture_from_handle and must
4093 rtex
->dcc_offset
= 0;
4096 void si_init_screen_state_functions(struct si_screen
*sscreen
)
4098 sscreen
->b
.b
.is_format_supported
= si_is_format_supported
;
4099 sscreen
->b
.query_opaque_metadata
= si_query_opaque_metadata
;
4100 sscreen
->b
.apply_opaque_metadata
= si_apply_opaque_metadata
;
4104 si_write_harvested_raster_configs(struct si_context
*sctx
,
4105 struct si_pm4_state
*pm4
,
4106 unsigned raster_config
,
4107 unsigned raster_config_1
)
4109 unsigned sh_per_se
= MAX2(sctx
->screen
->b
.info
.max_sh_per_se
, 1);
4110 unsigned num_se
= MAX2(sctx
->screen
->b
.info
.max_se
, 1);
4111 unsigned rb_mask
= sctx
->screen
->b
.info
.enabled_rb_mask
;
4112 unsigned num_rb
= MIN2(sctx
->screen
->b
.info
.num_render_backends
, 16);
4113 unsigned rb_per_pkr
= MIN2(num_rb
/ num_se
/ sh_per_se
, 2);
4114 unsigned rb_per_se
= num_rb
/ num_se
;
4115 unsigned se_mask
[4];
4118 se_mask
[0] = ((1 << rb_per_se
) - 1);
4119 se_mask
[1] = (se_mask
[0] << rb_per_se
);
4120 se_mask
[2] = (se_mask
[1] << rb_per_se
);
4121 se_mask
[3] = (se_mask
[2] << rb_per_se
);
4123 se_mask
[0] &= rb_mask
;
4124 se_mask
[1] &= rb_mask
;
4125 se_mask
[2] &= rb_mask
;
4126 se_mask
[3] &= rb_mask
;
4128 assert(num_se
== 1 || num_se
== 2 || num_se
== 4);
4129 assert(sh_per_se
== 1 || sh_per_se
== 2);
4130 assert(rb_per_pkr
== 1 || rb_per_pkr
== 2);
4132 /* XXX: I can't figure out what the *_XSEL and *_YSEL
4133 * fields are for, so I'm leaving them as their default
4136 for (se
= 0; se
< num_se
; se
++) {
4137 unsigned raster_config_se
= raster_config
;
4138 unsigned pkr0_mask
= ((1 << rb_per_pkr
) - 1) << (se
* rb_per_se
);
4139 unsigned pkr1_mask
= pkr0_mask
<< rb_per_pkr
;
4140 int idx
= (se
/ 2) * 2;
4142 if ((num_se
> 1) && (!se_mask
[idx
] || !se_mask
[idx
+ 1])) {
4143 raster_config_se
&= C_028350_SE_MAP
;
4145 if (!se_mask
[idx
]) {
4147 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_3
);
4150 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_0
);
4154 pkr0_mask
&= rb_mask
;
4155 pkr1_mask
&= rb_mask
;
4156 if (rb_per_se
> 2 && (!pkr0_mask
|| !pkr1_mask
)) {
4157 raster_config_se
&= C_028350_PKR_MAP
;
4161 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_3
);
4164 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_0
);
4168 if (rb_per_se
>= 2) {
4169 unsigned rb0_mask
= 1 << (se
* rb_per_se
);
4170 unsigned rb1_mask
= rb0_mask
<< 1;
4172 rb0_mask
&= rb_mask
;
4173 rb1_mask
&= rb_mask
;
4174 if (!rb0_mask
|| !rb1_mask
) {
4175 raster_config_se
&= C_028350_RB_MAP_PKR0
;
4179 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_3
);
4182 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_0
);
4186 if (rb_per_se
> 2) {
4187 rb0_mask
= 1 << (se
* rb_per_se
+ rb_per_pkr
);
4188 rb1_mask
= rb0_mask
<< 1;
4189 rb0_mask
&= rb_mask
;
4190 rb1_mask
&= rb_mask
;
4191 if (!rb0_mask
|| !rb1_mask
) {
4192 raster_config_se
&= C_028350_RB_MAP_PKR1
;
4196 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_3
);
4199 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_0
);
4205 /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
4206 if (sctx
->b
.chip_class
< CIK
)
4207 si_pm4_set_reg(pm4
, GRBM_GFX_INDEX
,
4208 SE_INDEX(se
) | SH_BROADCAST_WRITES
|
4209 INSTANCE_BROADCAST_WRITES
);
4211 si_pm4_set_reg(pm4
, R_030800_GRBM_GFX_INDEX
,
4212 S_030800_SE_INDEX(se
) | S_030800_SH_BROADCAST_WRITES(1) |
4213 S_030800_INSTANCE_BROADCAST_WRITES(1));
4214 si_pm4_set_reg(pm4
, R_028350_PA_SC_RASTER_CONFIG
, raster_config_se
);
4217 /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
4218 if (sctx
->b
.chip_class
< CIK
)
4219 si_pm4_set_reg(pm4
, GRBM_GFX_INDEX
,
4220 SE_BROADCAST_WRITES
| SH_BROADCAST_WRITES
|
4221 INSTANCE_BROADCAST_WRITES
);
4223 si_pm4_set_reg(pm4
, R_030800_GRBM_GFX_INDEX
,
4224 S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
4225 S_030800_INSTANCE_BROADCAST_WRITES(1));
4227 if ((num_se
> 2) && ((!se_mask
[0] && !se_mask
[1]) ||
4228 (!se_mask
[2] && !se_mask
[3]))) {
4229 raster_config_1
&= C_028354_SE_PAIR_MAP
;
4231 if (!se_mask
[0] && !se_mask
[1]) {
4233 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_3
);
4236 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_0
);
4240 si_pm4_set_reg(pm4
, R_028354_PA_SC_RASTER_CONFIG_1
, raster_config_1
);
4244 static void si_init_config(struct si_context
*sctx
)
4246 struct si_screen
*sscreen
= sctx
->screen
;
4247 unsigned num_rb
= MIN2(sctx
->screen
->b
.info
.num_render_backends
, 16);
4248 unsigned rb_mask
= sctx
->screen
->b
.info
.enabled_rb_mask
;
4249 unsigned raster_config
, raster_config_1
;
4250 uint64_t border_color_va
= sctx
->border_color_buffer
->gpu_address
;
4251 struct si_pm4_state
*pm4
= CALLOC_STRUCT(si_pm4_state
);
4256 si_pm4_cmd_begin(pm4
, PKT3_CONTEXT_CONTROL
);
4257 si_pm4_cmd_add(pm4
, CONTEXT_CONTROL_LOAD_ENABLE(1));
4258 si_pm4_cmd_add(pm4
, CONTEXT_CONTROL_SHADOW_ENABLE(1));
4259 si_pm4_cmd_end(pm4
, false);
4261 si_pm4_set_reg(pm4
, R_028A18_VGT_HOS_MAX_TESS_LEVEL
, fui(64));
4262 si_pm4_set_reg(pm4
, R_028A1C_VGT_HOS_MIN_TESS_LEVEL
, fui(0));
4264 /* FIXME calculate these values somehow ??? */
4265 si_pm4_set_reg(pm4
, R_028A54_VGT_GS_PER_ES
, SI_GS_PER_ES
);
4266 si_pm4_set_reg(pm4
, R_028A58_VGT_ES_PER_GS
, 0x40);
4267 si_pm4_set_reg(pm4
, R_028A5C_VGT_GS_PER_VS
, 0x2);
4269 si_pm4_set_reg(pm4
, R_028A8C_VGT_PRIMITIVEID_RESET
, 0x0);
4270 si_pm4_set_reg(pm4
, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET
, 0);
4272 si_pm4_set_reg(pm4
, R_028B98_VGT_STRMOUT_BUFFER_CONFIG
, 0x0);
4273 si_pm4_set_reg(pm4
, R_028AB8_VGT_VTX_CNT_EN
, 0x0);
4274 if (sctx
->b
.chip_class
< CIK
)
4275 si_pm4_set_reg(pm4
, R_008A14_PA_CL_ENHANCE
, S_008A14_NUM_CLIP_SEQ(3) |
4276 S_008A14_CLIP_VTX_REORDER_ENA(1));
4278 si_pm4_set_reg(pm4
, R_028BD4_PA_SC_CENTROID_PRIORITY_0
, 0x76543210);
4279 si_pm4_set_reg(pm4
, R_028BD8_PA_SC_CENTROID_PRIORITY_1
, 0xfedcba98);
4281 si_pm4_set_reg(pm4
, R_02882C_PA_SU_PRIM_FILTER_CNTL
, 0);
4283 switch (sctx
->screen
->b
.family
) {
4286 raster_config
= 0x2a00126a;
4287 raster_config_1
= 0x00000000;
4290 raster_config
= 0x0000124a;
4291 raster_config_1
= 0x00000000;
4294 raster_config
= 0x00000082;
4295 raster_config_1
= 0x00000000;
4298 raster_config
= 0x00000000;
4299 raster_config_1
= 0x00000000;
4302 raster_config
= 0x16000012;
4303 raster_config_1
= 0x00000000;
4306 raster_config
= 0x3a00161a;
4307 raster_config_1
= 0x0000002e;
4310 if (sscreen
->b
.info
.cik_macrotile_mode_array
[0] == 0x000000e8) {
4311 /* old kernels with old tiling config */
4312 raster_config
= 0x16000012;
4313 raster_config_1
= 0x0000002a;
4315 raster_config
= 0x3a00161a;
4316 raster_config_1
= 0x0000002e;
4319 case CHIP_POLARIS10
:
4320 raster_config
= 0x16000012;
4321 raster_config_1
= 0x0000002a;
4323 case CHIP_POLARIS11
:
4324 case CHIP_POLARIS12
:
4325 raster_config
= 0x16000012;
4326 raster_config_1
= 0x00000000;
4329 raster_config
= 0x16000012;
4330 raster_config_1
= 0x0000002a;
4334 raster_config
= 0x00000000;
4336 raster_config
= 0x00000002;
4337 raster_config_1
= 0x00000000;
4340 raster_config
= 0x00000002;
4341 raster_config_1
= 0x00000000;
4344 /* KV should be 0x00000002, but that causes problems with radeon */
4345 raster_config
= 0x00000000; /* 0x00000002 */
4346 raster_config_1
= 0x00000000;
4351 raster_config
= 0x00000000;
4352 raster_config_1
= 0x00000000;
4355 if (sctx
->b
.chip_class
<= VI
) {
4357 "radeonsi: Unknown GPU, using 0 for raster_config\n");
4358 raster_config
= 0x00000000;
4359 raster_config_1
= 0x00000000;
4364 if (sctx
->b
.chip_class
<= VI
) {
4365 if (!rb_mask
|| util_bitcount(rb_mask
) >= num_rb
) {
4366 /* Always use the default config when all backends are enabled
4367 * (or when we failed to determine the enabled backends).
4369 si_pm4_set_reg(pm4
, R_028350_PA_SC_RASTER_CONFIG
,
4371 if (sctx
->b
.chip_class
>= CIK
)
4372 si_pm4_set_reg(pm4
, R_028354_PA_SC_RASTER_CONFIG_1
,
4375 si_write_harvested_raster_configs(sctx
, pm4
, raster_config
, raster_config_1
);
4379 si_pm4_set_reg(pm4
, R_028204_PA_SC_WINDOW_SCISSOR_TL
, S_028204_WINDOW_OFFSET_DISABLE(1));
4380 si_pm4_set_reg(pm4
, R_028240_PA_SC_GENERIC_SCISSOR_TL
, S_028240_WINDOW_OFFSET_DISABLE(1));
4381 si_pm4_set_reg(pm4
, R_028244_PA_SC_GENERIC_SCISSOR_BR
,
4382 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
4383 si_pm4_set_reg(pm4
, R_028030_PA_SC_SCREEN_SCISSOR_TL
, 0);
4384 si_pm4_set_reg(pm4
, R_028034_PA_SC_SCREEN_SCISSOR_BR
,
4385 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
4387 si_pm4_set_reg(pm4
, R_02820C_PA_SC_CLIPRECT_RULE
, 0xFFFF);
4388 si_pm4_set_reg(pm4
, R_028230_PA_SC_EDGERULE
,
4389 S_028230_ER_TRI(0xA) |
4390 S_028230_ER_POINT(0xA) |
4391 S_028230_ER_RECT(0xA) |
4392 /* Required by DX10_DIAMOND_TEST_ENA: */
4393 S_028230_ER_LINE_LR(0x1A) |
4394 S_028230_ER_LINE_RL(0x26) |
4395 S_028230_ER_LINE_TB(0xA) |
4396 S_028230_ER_LINE_BT(0xA));
4397 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
4398 si_pm4_set_reg(pm4
, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET
, 0);
4399 si_pm4_set_reg(pm4
, R_028820_PA_CL_NANINF_CNTL
, 0);
4400 si_pm4_set_reg(pm4
, R_028AC0_DB_SRESULTS_COMPARE_STATE0
, 0x0);
4401 si_pm4_set_reg(pm4
, R_028AC4_DB_SRESULTS_COMPARE_STATE1
, 0x0);
4402 si_pm4_set_reg(pm4
, R_028AC8_DB_PRELOAD_CONTROL
, 0x0);
4403 si_pm4_set_reg(pm4
, R_02800C_DB_RENDER_OVERRIDE
, 0);
4405 if (sctx
->b
.chip_class
>= GFX9
) {
4406 si_pm4_set_reg(pm4
, R_030920_VGT_MAX_VTX_INDX
, ~0);
4407 si_pm4_set_reg(pm4
, R_030924_VGT_MIN_VTX_INDX
, 0);
4408 si_pm4_set_reg(pm4
, R_030928_VGT_INDX_OFFSET
, 0);
4410 si_pm4_set_reg(pm4
, R_028400_VGT_MAX_VTX_INDX
, ~0);
4411 si_pm4_set_reg(pm4
, R_028404_VGT_MIN_VTX_INDX
, 0);
4412 si_pm4_set_reg(pm4
, R_028408_VGT_INDX_OFFSET
, 0);
4415 if (sctx
->b
.chip_class
>= CIK
) {
4416 /* If this is 0, Bonaire can hang even if GS isn't being used.
4417 * Other chips are unaffected. These are suboptimal values,
4418 * but we don't use on-chip GS.
4420 si_pm4_set_reg(pm4
, R_028A44_VGT_GS_ONCHIP_CNTL
,
4421 S_028A44_ES_VERTS_PER_SUBGRP(64) |
4422 S_028A44_GS_PRIMS_PER_SUBGRP(4));
4424 if (sctx
->b
.chip_class
>= GFX9
) {
4425 si_pm4_set_reg(pm4
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
, S_00B41C_CU_EN(0xffff));
4427 si_pm4_set_reg(pm4
, R_00B51C_SPI_SHADER_PGM_RSRC3_LS
, S_00B51C_CU_EN(0xffff));
4428 si_pm4_set_reg(pm4
, R_00B41C_SPI_SHADER_PGM_RSRC3_HS
, 0);
4429 si_pm4_set_reg(pm4
, R_00B31C_SPI_SHADER_PGM_RSRC3_ES
, S_00B31C_CU_EN(0xffff));
4431 si_pm4_set_reg(pm4
, R_00B21C_SPI_SHADER_PGM_RSRC3_GS
, S_00B21C_CU_EN(0xffff));
4433 if (sscreen
->b
.info
.num_good_compute_units
/
4434 (sscreen
->b
.info
.max_se
* sscreen
->b
.info
.max_sh_per_se
) <= 4) {
4435 /* Too few available compute units per SH. Disallowing
4436 * VS to run on CU0 could hurt us more than late VS
4437 * allocation would help.
4439 * LATE_ALLOC_VS = 2 is the highest safe number.
4441 si_pm4_set_reg(pm4
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
, S_00B118_CU_EN(0xffff));
4442 si_pm4_set_reg(pm4
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
, S_00B11C_LIMIT(2));
4444 /* Set LATE_ALLOC_VS == 31. It should be less than
4445 * the number of scratch waves. Limitations:
4446 * - VS can't execute on CU0.
4447 * - If HS writes outputs to LDS, LS can't execute on CU0.
4449 si_pm4_set_reg(pm4
, R_00B118_SPI_SHADER_PGM_RSRC3_VS
, S_00B118_CU_EN(0xfffe));
4450 si_pm4_set_reg(pm4
, R_00B11C_SPI_SHADER_LATE_ALLOC_VS
, S_00B11C_LIMIT(31));
4453 si_pm4_set_reg(pm4
, R_00B01C_SPI_SHADER_PGM_RSRC3_PS
, S_00B01C_CU_EN(0xffff));
4456 if (sctx
->b
.chip_class
>= VI
) {
4457 unsigned vgt_tess_distribution
;
4459 si_pm4_set_reg(pm4
, R_028424_CB_DCC_CONTROL
,
4460 S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
4461 S_028424_OVERWRITE_COMBINER_WATERMARK(4));
4462 if (sctx
->b
.family
< CHIP_POLARIS10
)
4463 si_pm4_set_reg(pm4
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 30);
4464 si_pm4_set_reg(pm4
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 32);
4466 vgt_tess_distribution
=
4467 S_028B50_ACCUM_ISOLINE(32) |
4468 S_028B50_ACCUM_TRI(11) |
4469 S_028B50_ACCUM_QUAD(11) |
4470 S_028B50_DONUT_SPLIT(16);
4472 /* Testing with Unigine Heaven extreme tesselation yielded best results
4473 * with TRAP_SPLIT = 3.
4475 if (sctx
->b
.family
== CHIP_FIJI
||
4476 sctx
->b
.family
>= CHIP_POLARIS10
)
4477 vgt_tess_distribution
|= S_028B50_TRAP_SPLIT(3);
4479 si_pm4_set_reg(pm4
, R_028B50_VGT_TESS_DISTRIBUTION
, vgt_tess_distribution
);
4481 si_pm4_set_reg(pm4
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
, 14);
4482 si_pm4_set_reg(pm4
, R_028C5C_VGT_OUT_DEALLOC_CNTL
, 16);
4485 if (sctx
->screen
->b
.has_rbplus
)
4486 si_pm4_set_reg(pm4
, R_028C40_PA_SC_SHADER_CONTROL
, 0);
4488 si_pm4_set_reg(pm4
, R_028080_TA_BC_BASE_ADDR
, border_color_va
>> 8);
4489 if (sctx
->b
.chip_class
>= CIK
)
4490 si_pm4_set_reg(pm4
, R_028084_TA_BC_BASE_ADDR_HI
, border_color_va
>> 40);
4491 si_pm4_add_bo(pm4
, sctx
->border_color_buffer
, RADEON_USAGE_READ
,
4492 RADEON_PRIO_BORDER_COLORS
);
4494 if (sctx
->b
.chip_class
>= GFX9
) {
4495 si_pm4_set_reg(pm4
, R_028060_DB_DFSM_CONTROL
, 0);
4496 si_pm4_set_reg(pm4
, R_028064_DB_RENDER_FILTER
, 0);
4497 /* TODO: We can use this to disable RBs for rendering to GART: */
4498 si_pm4_set_reg(pm4
, R_02835C_PA_SC_TILE_STEERING_OVERRIDE
, 0);
4499 si_pm4_set_reg(pm4
, R_02883C_PA_SU_OVER_RASTERIZATION_CNTL
, 0);
4500 /* TODO: Enable the binner: */
4501 si_pm4_set_reg(pm4
, R_028C44_PA_SC_BINNER_CNTL_0
,
4502 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC
));
4503 si_pm4_set_reg(pm4
, R_028C48_PA_SC_BINNER_CNTL_1
, 0);
4504 si_pm4_set_reg(pm4
, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL
,
4505 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
4506 si_pm4_set_reg(pm4
, R_030968_VGT_INSTANCE_BASE_ID
, 0);
4509 si_pm4_upload_indirect_buffer(sctx
, pm4
);
4510 sctx
->init_config
= pm4
;