xref: /aosp_15_r20/external/mesa3d/src/gallium/drivers/radeonsi/si_state.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright 2012 Advanced Micro Devices, Inc.
3  *
4  * SPDX-License-Identifier: MIT
5  */
6 
7 #include "si_build_pm4.h"
8 #include "si_query.h"
9 #include "si_shader_internal.h"
10 #include "sid.h"
11 #include "util/fast_idiv_by_const.h"
12 #include "util/format/u_format.h"
13 #include "util/format/u_format_s3tc.h"
14 #include "util/u_dual_blend.h"
15 #include "util/u_helpers.h"
16 #include "util/u_memory.h"
17 #include "util/u_resource.h"
18 #include "util/u_upload_mgr.h"
19 #include "util/u_blend.h"
20 
21 #include "ac_cmdbuf.h"
22 #include "ac_descriptors.h"
23 #include "ac_formats.h"
24 #include "gfx10_format_table.h"
25 
26 /* 12.4 fixed-point */
si_pack_float_12p4(float x)27 static unsigned si_pack_float_12p4(float x)
28 {
29    return x <= 0 ? 0 : x >= 4096 ? 0xffff : x * 16;
30 }
31 
32 /*
33  * Inferred framebuffer and blender state.
34  *
35  * CB_TARGET_MASK is emitted here to avoid a hang with dual source blending
36  * if there is not enough PS outputs.
37  */
si_emit_cb_render_state(struct si_context * sctx,unsigned index)38 static void si_emit_cb_render_state(struct si_context *sctx, unsigned index)
39 {
40    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
41    struct si_state_blend *blend = sctx->queued.named.blend;
42    /* CB_COLORn_INFO.FORMAT=INVALID should disable unbound colorbuffers,
43     * but you never know. */
44    uint32_t cb_target_mask = sctx->framebuffer.colorbuf_enabled_4bit & blend->cb_target_mask;
45    unsigned i;
46 
47    /* Avoid a hang that happens when dual source blending is enabled
48     * but there is not enough color outputs. This is undefined behavior,
49     * so disable color writes completely.
50     *
51     * Reproducible with Unigine Heaven 4.0 and drirc missing.
52     */
53    if (blend->dual_src_blend && sctx->shader.ps.cso &&
54        (sctx->shader.ps.cso->info.colors_written & 0x3) != 0x3)
55       cb_target_mask = 0;
56 
57    /* GFX9: Flush DFSM when CB_TARGET_MASK changes.
58     * I think we don't have to do anything between IBs.
59     */
60    if (sctx->screen->dpbb_allowed && sctx->last_cb_target_mask != cb_target_mask &&
61        sctx->screen->pbb_context_states_per_bin > 1) {
62       sctx->last_cb_target_mask = cb_target_mask;
63 
64       radeon_begin(cs);
65       radeon_event_write(V_028A90_BREAK_BATCH);
66       radeon_end();
67    }
68 
69    uint32_t cb_dcc_control = 0;
70 
71    if (sctx->gfx_level >= GFX8 && sctx->gfx_level < GFX12) {
72       /* DCC MSAA workaround.
73        * Alternatively, we can set CB_COLORi_DCC_CONTROL.OVERWRITE_-
74        * COMBINER_DISABLE, but that would be more complicated.
75        */
76       bool oc_disable =
77          blend->dcc_msaa_corruption_4bit & cb_target_mask && sctx->framebuffer.nr_samples >= 2;
78 
79       if (sctx->gfx_level >= GFX11) {
80          cb_dcc_control =
81             S_028424_SAMPLE_MASK_TRACKER_DISABLE(oc_disable) |
82             S_028424_SAMPLE_MASK_TRACKER_WATERMARK(sctx->screen->info.has_dedicated_vram ? 0 : 15);
83       } else {
84          cb_dcc_control =
85             S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(sctx->gfx_level <= GFX9) |
86             S_028424_OVERWRITE_COMBINER_WATERMARK(sctx->gfx_level >= GFX10 ? 6 : 4) |
87             S_028424_OVERWRITE_COMBINER_DISABLE(oc_disable) |
88             S_028424_DISABLE_CONSTANT_ENCODE_REG(sctx->gfx_level < GFX11 &&
89                                                  sctx->screen->info.has_dcc_constant_encode);
90       }
91    }
92 
93    uint32_t sx_ps_downconvert = 0;
94    uint32_t sx_blend_opt_epsilon = 0;
95    uint32_t sx_blend_opt_control = 0;
96 
97    /* RB+ register settings. */
98    if (sctx->screen->info.rbplus_allowed) {
99       unsigned spi_shader_col_format =
100          sctx->shader.ps.cso ? sctx->shader.ps.current->key.ps.part.epilog.spi_shader_col_format
101                              : 0;
102       unsigned num_cbufs = util_last_bit(sctx->framebuffer.colorbuf_enabled_4bit &
103                                          blend->cb_target_enabled_4bit) / 4;
104 
105       for (i = 0; i < num_cbufs; i++) {
106          struct si_surface *surf = (struct si_surface *)sctx->framebuffer.state.cbufs[i];
107          unsigned format, swap, spi_format, colormask;
108          bool has_alpha, has_rgb;
109 
110          if (!surf) {
111             /* If the color buffer is not set, the driver sets 32_R
112              * as the SPI color format, because the hw doesn't allow
113              * holes between color outputs, so also set this to
114              * enable RB+.
115              */
116             sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_R << (i * 4);
117             continue;
118          }
119 
120          format = sctx->gfx_level >= GFX11 ? G_028C70_FORMAT_GFX11(surf->cb.cb_color_info):
121                                              G_028C70_FORMAT_GFX6(surf->cb.cb_color_info);
122          swap = G_028C70_COMP_SWAP(surf->cb.cb_color_info);
123          spi_format = (spi_shader_col_format >> (i * 4)) & 0xf;
124          colormask = (cb_target_mask >> (i * 4)) & 0xf;
125 
126          /* Set if RGB and A are present. */
127          has_alpha = !(sctx->gfx_level >= GFX11 ? G_028C74_FORCE_DST_ALPHA_1_GFX11(surf->cb.cb_color_attrib):
128                                                   G_028C74_FORCE_DST_ALPHA_1_GFX6(surf->cb.cb_color_attrib));
129 
130          if (format == V_028C70_COLOR_8 || format == V_028C70_COLOR_16 ||
131              format == V_028C70_COLOR_32)
132             has_rgb = !has_alpha;
133          else
134             has_rgb = true;
135 
136          /* Check the colormask and export format. */
137          if (!(colormask & (PIPE_MASK_RGBA & ~PIPE_MASK_A)))
138             has_rgb = false;
139          if (!(colormask & PIPE_MASK_A))
140             has_alpha = false;
141 
142          if (spi_format == V_028714_SPI_SHADER_ZERO) {
143             has_rgb = false;
144             has_alpha = false;
145          }
146 
147          /* Disable value checking for disabled channels. */
148          if (!has_rgb)
149             sx_blend_opt_control |= S_02875C_MRT0_COLOR_OPT_DISABLE(1) << (i * 4);
150          if (!has_alpha)
151             sx_blend_opt_control |= S_02875C_MRT0_ALPHA_OPT_DISABLE(1) << (i * 4);
152 
153          /* Enable down-conversion for 32bpp and smaller formats. */
154          switch (format) {
155          case V_028C70_COLOR_8:
156          case V_028C70_COLOR_8_8:
157          case V_028C70_COLOR_8_8_8_8:
158             /* For 1 and 2-channel formats, use the superset thereof. */
159             if (spi_format == V_028714_SPI_SHADER_FP16_ABGR ||
160                 spi_format == V_028714_SPI_SHADER_UINT16_ABGR ||
161                 spi_format == V_028714_SPI_SHADER_SINT16_ABGR) {
162                sx_ps_downconvert |= V_028754_SX_RT_EXPORT_8_8_8_8 << (i * 4);
163                if (G_028C70_NUMBER_TYPE(surf->cb.cb_color_info) != V_028C70_NUMBER_SRGB)
164                   sx_blend_opt_epsilon |= V_028758_8BIT_FORMAT_0_5 << (i * 4);
165             }
166             break;
167 
168          case V_028C70_COLOR_5_6_5:
169             if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
170                sx_ps_downconvert |= V_028754_SX_RT_EXPORT_5_6_5 << (i * 4);
171                sx_blend_opt_epsilon |= V_028758_6BIT_FORMAT_0_5 << (i * 4);
172             }
173             break;
174 
175          case V_028C70_COLOR_1_5_5_5:
176             if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
177                sx_ps_downconvert |= V_028754_SX_RT_EXPORT_1_5_5_5 << (i * 4);
178                sx_blend_opt_epsilon |= V_028758_5BIT_FORMAT_0_5 << (i * 4);
179             }
180             break;
181 
182          case V_028C70_COLOR_4_4_4_4:
183             if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
184                sx_ps_downconvert |= V_028754_SX_RT_EXPORT_4_4_4_4 << (i * 4);
185                sx_blend_opt_epsilon |= V_028758_4BIT_FORMAT_0_5 << (i * 4);
186             }
187             break;
188 
189          case V_028C70_COLOR_32:
190             if (swap == V_028C70_SWAP_STD && spi_format == V_028714_SPI_SHADER_32_R)
191                sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_R << (i * 4);
192             else if (swap == V_028C70_SWAP_ALT_REV && spi_format == V_028714_SPI_SHADER_32_AR)
193                sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_A << (i * 4);
194             break;
195 
196          case V_028C70_COLOR_16:
197          case V_028C70_COLOR_16_16:
198             /* For 1-channel formats, use the superset thereof. */
199             if (spi_format == V_028714_SPI_SHADER_UNORM16_ABGR ||
200                 spi_format == V_028714_SPI_SHADER_SNORM16_ABGR ||
201                 spi_format == V_028714_SPI_SHADER_UINT16_ABGR ||
202                 spi_format == V_028714_SPI_SHADER_SINT16_ABGR) {
203                if (swap == V_028C70_SWAP_STD || swap == V_028C70_SWAP_STD_REV)
204                   sx_ps_downconvert |= V_028754_SX_RT_EXPORT_16_16_GR << (i * 4);
205                else
206                   sx_ps_downconvert |= V_028754_SX_RT_EXPORT_16_16_AR << (i * 4);
207             }
208             break;
209 
210          case V_028C70_COLOR_10_11_11:
211             if (spi_format == V_028714_SPI_SHADER_FP16_ABGR)
212                sx_ps_downconvert |= V_028754_SX_RT_EXPORT_10_11_11 << (i * 4);
213             break;
214 
215          case V_028C70_COLOR_2_10_10_10:
216          case V_028C70_COLOR_10_10_10_2:
217             if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
218                sx_ps_downconvert |= V_028754_SX_RT_EXPORT_2_10_10_10 << (i * 4);
219                sx_blend_opt_epsilon |= V_028758_10BIT_FORMAT_0_5 << (i * 4);
220             }
221             break;
222 
223          case V_028C70_COLOR_5_9_9_9:
224             if (spi_format == V_028714_SPI_SHADER_FP16_ABGR)
225                sx_ps_downconvert |= V_028754_SX_RT_EXPORT_9_9_9_E5 << (i * 4);
226             break;
227          }
228       }
229 
230       /* If there are no color outputs, the first color export is
231        * always enabled as 32_R, so also set this to enable RB+.
232        */
233       if (!sx_ps_downconvert)
234          sx_ps_downconvert = V_028754_SX_RT_EXPORT_32_R;
235    }
236 
237    if (sctx->gfx_level >= GFX12) {
238       /* GFX12 doesn't have CB_FDCC_CONTROL. */
239       assert(cb_dcc_control == 0);
240 
241       radeon_begin(cs);
242       gfx12_begin_context_regs();
243       gfx12_opt_set_context_reg(R_028850_CB_TARGET_MASK, SI_TRACKED_CB_TARGET_MASK,
244                                 cb_target_mask);
245       gfx12_opt_set_context_reg(R_028754_SX_PS_DOWNCONVERT, SI_TRACKED_SX_PS_DOWNCONVERT,
246                                 sx_ps_downconvert);
247       gfx12_opt_set_context_reg(R_028758_SX_BLEND_OPT_EPSILON, SI_TRACKED_SX_BLEND_OPT_EPSILON,
248                                 sx_blend_opt_epsilon);
249       gfx12_opt_set_context_reg(R_02875C_SX_BLEND_OPT_CONTROL, SI_TRACKED_SX_BLEND_OPT_CONTROL,
250                                 sx_blend_opt_control);
251       gfx12_end_context_regs();
252       radeon_end(); /* don't track context rolls on GFX12 */
253    } else if (sctx->screen->info.has_set_context_pairs_packed) {
254       radeon_begin(cs);
255       gfx11_begin_packed_context_regs();
256       gfx11_opt_set_context_reg(R_028238_CB_TARGET_MASK, SI_TRACKED_CB_TARGET_MASK,
257                                 cb_target_mask);
258       gfx11_opt_set_context_reg(R_028424_CB_DCC_CONTROL, SI_TRACKED_CB_DCC_CONTROL,
259                                 cb_dcc_control);
260       gfx11_opt_set_context_reg(R_028754_SX_PS_DOWNCONVERT, SI_TRACKED_SX_PS_DOWNCONVERT,
261                                 sx_ps_downconvert);
262       gfx11_opt_set_context_reg(R_028758_SX_BLEND_OPT_EPSILON, SI_TRACKED_SX_BLEND_OPT_EPSILON,
263                                 sx_blend_opt_epsilon);
264       gfx11_opt_set_context_reg(R_02875C_SX_BLEND_OPT_CONTROL, SI_TRACKED_SX_BLEND_OPT_CONTROL,
265                                 sx_blend_opt_control);
266       gfx11_end_packed_context_regs();
267       radeon_end(); /* don't track context rolls on GFX11 */
268    } else {
269       radeon_begin(cs);
270       radeon_opt_set_context_reg(R_028238_CB_TARGET_MASK, SI_TRACKED_CB_TARGET_MASK,
271                                  cb_target_mask);
272       if (sctx->gfx_level >= GFX8) {
273          radeon_opt_set_context_reg(R_028424_CB_DCC_CONTROL, SI_TRACKED_CB_DCC_CONTROL,
274                                     cb_dcc_control);
275       }
276       if (sctx->screen->info.rbplus_allowed) {
277          radeon_opt_set_context_reg3(R_028754_SX_PS_DOWNCONVERT, SI_TRACKED_SX_PS_DOWNCONVERT,
278                                      sx_ps_downconvert, sx_blend_opt_epsilon, sx_blend_opt_control);
279       }
280       radeon_end_update_context_roll();
281    }
282 }
283 
284 /*
285  * Blender functions
286  */
287 
si_translate_blend_function(int blend_func)288 static uint32_t si_translate_blend_function(int blend_func)
289 {
290    switch (blend_func) {
291    case PIPE_BLEND_ADD:
292       return V_028780_COMB_DST_PLUS_SRC;
293    case PIPE_BLEND_SUBTRACT:
294       return V_028780_COMB_SRC_MINUS_DST;
295    case PIPE_BLEND_REVERSE_SUBTRACT:
296       return V_028780_COMB_DST_MINUS_SRC;
297    case PIPE_BLEND_MIN:
298       return V_028780_COMB_MIN_DST_SRC;
299    case PIPE_BLEND_MAX:
300       return V_028780_COMB_MAX_DST_SRC;
301    default:
302       PRINT_ERR("Unknown blend function %d\n", blend_func);
303       assert(0);
304       break;
305    }
306    return 0;
307 }
308 
si_translate_blend_factor(enum amd_gfx_level gfx_level,int blend_fact)309 static uint32_t si_translate_blend_factor(enum amd_gfx_level gfx_level, int blend_fact)
310 {
311    switch (blend_fact) {
312    case PIPE_BLENDFACTOR_ONE:
313       return V_028780_BLEND_ONE;
314    case PIPE_BLENDFACTOR_SRC_COLOR:
315       return V_028780_BLEND_SRC_COLOR;
316    case PIPE_BLENDFACTOR_SRC_ALPHA:
317       return V_028780_BLEND_SRC_ALPHA;
318    case PIPE_BLENDFACTOR_DST_ALPHA:
319       return V_028780_BLEND_DST_ALPHA;
320    case PIPE_BLENDFACTOR_DST_COLOR:
321       return V_028780_BLEND_DST_COLOR;
322    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
323       return V_028780_BLEND_SRC_ALPHA_SATURATE;
324    case PIPE_BLENDFACTOR_CONST_COLOR:
325       return gfx_level >= GFX11 ? V_028780_BLEND_CONSTANT_COLOR_GFX11:
326                                    V_028780_BLEND_CONSTANT_COLOR_GFX6;
327    case PIPE_BLENDFACTOR_CONST_ALPHA:
328       return gfx_level >= GFX11 ? V_028780_BLEND_CONSTANT_ALPHA_GFX11 :
329                                    V_028780_BLEND_CONSTANT_ALPHA_GFX6;
330    case PIPE_BLENDFACTOR_ZERO:
331       return V_028780_BLEND_ZERO;
332    case PIPE_BLENDFACTOR_INV_SRC_COLOR:
333       return V_028780_BLEND_ONE_MINUS_SRC_COLOR;
334    case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
335       return V_028780_BLEND_ONE_MINUS_SRC_ALPHA;
336    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
337       return V_028780_BLEND_ONE_MINUS_DST_ALPHA;
338    case PIPE_BLENDFACTOR_INV_DST_COLOR:
339       return V_028780_BLEND_ONE_MINUS_DST_COLOR;
340    case PIPE_BLENDFACTOR_INV_CONST_COLOR:
341       return gfx_level >= GFX11 ? V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR_GFX11:
342                                    V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR_GFX6;
343    case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
344       return gfx_level >= GFX11 ? V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA_GFX11:
345                                    V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA_GFX6;
346    case PIPE_BLENDFACTOR_SRC1_COLOR:
347       return gfx_level >= GFX11 ? V_028780_BLEND_SRC1_COLOR_GFX11:
348                                    V_028780_BLEND_SRC1_COLOR_GFX6;
349    case PIPE_BLENDFACTOR_SRC1_ALPHA:
350       return gfx_level >= GFX11 ? V_028780_BLEND_SRC1_ALPHA_GFX11:
351                                    V_028780_BLEND_SRC1_ALPHA_GFX6;
352    case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
353       return gfx_level >= GFX11 ? V_028780_BLEND_INV_SRC1_COLOR_GFX11:
354                                    V_028780_BLEND_INV_SRC1_COLOR_GFX6;
355    case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
356       return gfx_level >= GFX11 ? V_028780_BLEND_INV_SRC1_ALPHA_GFX11:
357                                    V_028780_BLEND_INV_SRC1_ALPHA_GFX6;
358    default:
359       PRINT_ERR("Bad blend factor %d not supported!\n", blend_fact);
360       assert(0);
361       break;
362    }
363    return 0;
364 }
365 
si_translate_blend_opt_function(int blend_func)366 static uint32_t si_translate_blend_opt_function(int blend_func)
367 {
368    switch (blend_func) {
369    case PIPE_BLEND_ADD:
370       return V_028760_OPT_COMB_ADD;
371    case PIPE_BLEND_SUBTRACT:
372       return V_028760_OPT_COMB_SUBTRACT;
373    case PIPE_BLEND_REVERSE_SUBTRACT:
374       return V_028760_OPT_COMB_REVSUBTRACT;
375    case PIPE_BLEND_MIN:
376       return V_028760_OPT_COMB_MIN;
377    case PIPE_BLEND_MAX:
378       return V_028760_OPT_COMB_MAX;
379    default:
380       return V_028760_OPT_COMB_BLEND_DISABLED;
381    }
382 }
383 
si_translate_blend_opt_factor(int blend_fact,bool is_alpha)384 static uint32_t si_translate_blend_opt_factor(int blend_fact, bool is_alpha)
385 {
386    switch (blend_fact) {
387    case PIPE_BLENDFACTOR_ZERO:
388       return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL;
389    case PIPE_BLENDFACTOR_ONE:
390       return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE;
391    case PIPE_BLENDFACTOR_SRC_COLOR:
392       return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
393                       : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0;
394    case PIPE_BLENDFACTOR_INV_SRC_COLOR:
395       return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
396                       : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1;
397    case PIPE_BLENDFACTOR_SRC_ALPHA:
398       return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0;
399    case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
400       return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1;
401    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
402       return is_alpha ? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
403                       : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
404    default:
405       return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
406    }
407 }
408 
si_blend_check_commutativity(struct si_screen * sscreen,struct si_state_blend * blend,enum pipe_blend_func func,enum pipe_blendfactor src,enum pipe_blendfactor dst,unsigned chanmask)409 static void si_blend_check_commutativity(struct si_screen *sscreen, struct si_state_blend *blend,
410                                          enum pipe_blend_func func, enum pipe_blendfactor src,
411                                          enum pipe_blendfactor dst, unsigned chanmask)
412 {
413    /* Src factor is allowed when it does not depend on Dst */
414    static const uint32_t src_allowed =
415       (1u << PIPE_BLENDFACTOR_ONE) | (1u << PIPE_BLENDFACTOR_SRC_COLOR) |
416       (1u << PIPE_BLENDFACTOR_SRC_ALPHA) | (1u << PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE) |
417       (1u << PIPE_BLENDFACTOR_CONST_COLOR) | (1u << PIPE_BLENDFACTOR_CONST_ALPHA) |
418       (1u << PIPE_BLENDFACTOR_SRC1_COLOR) | (1u << PIPE_BLENDFACTOR_SRC1_ALPHA) |
419       (1u << PIPE_BLENDFACTOR_ZERO) | (1u << PIPE_BLENDFACTOR_INV_SRC_COLOR) |
420       (1u << PIPE_BLENDFACTOR_INV_SRC_ALPHA) | (1u << PIPE_BLENDFACTOR_INV_CONST_COLOR) |
421       (1u << PIPE_BLENDFACTOR_INV_CONST_ALPHA) | (1u << PIPE_BLENDFACTOR_INV_SRC1_COLOR) |
422       (1u << PIPE_BLENDFACTOR_INV_SRC1_ALPHA);
423 
424    if (dst == PIPE_BLENDFACTOR_ONE && (src_allowed & (1u << src)) &&
425        (func == PIPE_BLEND_MAX || func == PIPE_BLEND_MIN))
426       blend->commutative_4bit |= chanmask;
427 }
428 
429 /**
430  * Get rid of DST in the blend factors by commuting the operands:
431  *    func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
432  */
si_blend_remove_dst(unsigned * func,unsigned * src_factor,unsigned * dst_factor,unsigned expected_dst,unsigned replacement_src)433 static void si_blend_remove_dst(unsigned *func, unsigned *src_factor, unsigned *dst_factor,
434                                 unsigned expected_dst, unsigned replacement_src)
435 {
436    if (*src_factor == expected_dst && *dst_factor == PIPE_BLENDFACTOR_ZERO) {
437       *src_factor = PIPE_BLENDFACTOR_ZERO;
438       *dst_factor = replacement_src;
439 
440       /* Commuting the operands requires reversing subtractions. */
441       if (*func == PIPE_BLEND_SUBTRACT)
442          *func = PIPE_BLEND_REVERSE_SUBTRACT;
443       else if (*func == PIPE_BLEND_REVERSE_SUBTRACT)
444          *func = PIPE_BLEND_SUBTRACT;
445    }
446 }
447 
si_create_blend_state_mode(struct pipe_context * ctx,const struct pipe_blend_state * state,unsigned mode)448 static void *si_create_blend_state_mode(struct pipe_context *ctx,
449                                         const struct pipe_blend_state *state, unsigned mode)
450 {
451    struct si_context *sctx = (struct si_context *)ctx;
452    struct si_state_blend *blend = CALLOC_STRUCT(si_state_blend);
453    struct si_pm4_state *pm4 = &blend->pm4;
454    uint32_t sx_mrt_blend_opt[8] = {0};
455    uint32_t color_control = 0;
456    bool logicop_enable = state->logicop_enable && state->logicop_func != PIPE_LOGICOP_COPY;
457 
458    if (!blend)
459       return NULL;
460 
461    si_pm4_clear_state(pm4, sctx->screen, false);
462 
463    blend->alpha_to_coverage = state->alpha_to_coverage;
464    blend->alpha_to_one = state->alpha_to_one;
465    blend->dual_src_blend = util_blend_state_is_dual(state, 0);
466    blend->logicop_enable = logicop_enable;
467    blend->allows_noop_optimization =
468       state->rt[0].rgb_func == PIPE_BLEND_ADD &&
469       state->rt[0].alpha_func == PIPE_BLEND_ADD &&
470       state->rt[0].rgb_src_factor == PIPE_BLENDFACTOR_DST_COLOR &&
471       state->rt[0].alpha_src_factor == PIPE_BLENDFACTOR_DST_COLOR &&
472       state->rt[0].rgb_dst_factor == PIPE_BLENDFACTOR_ZERO &&
473       state->rt[0].alpha_dst_factor == PIPE_BLENDFACTOR_ZERO &&
474       mode == V_028808_CB_NORMAL;
475 
476    unsigned num_shader_outputs = state->max_rt + 1; /* estimate */
477    if (blend->dual_src_blend)
478       num_shader_outputs = MAX2(num_shader_outputs, 2);
479 
480    if (logicop_enable) {
481       color_control |= S_028808_ROP3(state->logicop_func | (state->logicop_func << 4));
482    } else {
483       color_control |= S_028808_ROP3(0xcc);
484    }
485 
486    unsigned db_alpha_to_mask;
487    if (state->alpha_to_coverage && state->alpha_to_coverage_dither) {
488       db_alpha_to_mask = S_028B70_ALPHA_TO_MASK_ENABLE(state->alpha_to_coverage) |
489                          S_028B70_ALPHA_TO_MASK_OFFSET0(3) | S_028B70_ALPHA_TO_MASK_OFFSET1(1) |
490                          S_028B70_ALPHA_TO_MASK_OFFSET2(0) | S_028B70_ALPHA_TO_MASK_OFFSET3(2) |
491                          S_028B70_OFFSET_ROUND(1);
492    } else {
493       db_alpha_to_mask = S_028B70_ALPHA_TO_MASK_ENABLE(state->alpha_to_coverage) |
494                          S_028B70_ALPHA_TO_MASK_OFFSET0(2) | S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
495                          S_028B70_ALPHA_TO_MASK_OFFSET2(2) | S_028B70_ALPHA_TO_MASK_OFFSET3(2) |
496                          S_028B70_OFFSET_ROUND(0);
497    }
498 
499    if (sctx->gfx_level >= GFX12)
500       ac_pm4_set_reg(&pm4->base, R_02807C_DB_ALPHA_TO_MASK, db_alpha_to_mask);
501    else
502       ac_pm4_set_reg(&pm4->base, R_028B70_DB_ALPHA_TO_MASK, db_alpha_to_mask);
503 
504    blend->cb_target_mask = 0;
505    blend->cb_target_enabled_4bit = 0;
506 
507    unsigned last_blend_cntl;
508 
509    for (int i = 0; i < num_shader_outputs; i++) {
510       /* state->rt entries > 0 only written if independent blending */
511       const int j = state->independent_blend_enable ? i : 0;
512 
513       unsigned eqRGB = state->rt[j].rgb_func;
514       unsigned srcRGB = state->rt[j].rgb_src_factor;
515       unsigned dstRGB = state->rt[j].rgb_dst_factor;
516       unsigned eqA = state->rt[j].alpha_func;
517       unsigned srcA = state->rt[j].alpha_src_factor;
518       unsigned dstA = state->rt[j].alpha_dst_factor;
519 
520       unsigned srcRGB_opt, dstRGB_opt, srcA_opt, dstA_opt;
521       unsigned blend_cntl = 0;
522 
523       sx_mrt_blend_opt[i] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) |
524                             S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED);
525 
526       /* Only set dual source blending for MRT0 to avoid a hang. */
527       if (i >= 1 && blend->dual_src_blend) {
528          if (i == 1) {
529             if (sctx->gfx_level >= GFX11)
530                blend_cntl = last_blend_cntl;
531             else
532                blend_cntl = S_028780_ENABLE(1);
533          }
534 
535          ac_pm4_set_reg(&pm4->base, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
536          continue;
537       }
538 
539       /* Only addition and subtraction equations are supported with
540        * dual source blending.
541        */
542       if (blend->dual_src_blend && (eqRGB == PIPE_BLEND_MIN || eqRGB == PIPE_BLEND_MAX ||
543                                     eqA == PIPE_BLEND_MIN || eqA == PIPE_BLEND_MAX)) {
544          assert(!"Unsupported equation for dual source blending");
545          ac_pm4_set_reg(&pm4->base, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
546          continue;
547       }
548 
549       /* cb_render_state will disable unused ones */
550       blend->cb_target_mask |= (unsigned)state->rt[j].colormask << (4 * i);
551       if (state->rt[j].colormask)
552          blend->cb_target_enabled_4bit |= 0xf << (4 * i);
553 
554       if (!state->rt[j].colormask || !state->rt[j].blend_enable) {
555          ac_pm4_set_reg(&pm4->base, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
556          continue;
557       }
558 
559       si_blend_check_commutativity(sctx->screen, blend, eqRGB, srcRGB, dstRGB, 0x7 << (4 * i));
560       si_blend_check_commutativity(sctx->screen, blend, eqA, srcA, dstA, 0x8 << (4 * i));
561 
562       /* Blending optimizations for RB+.
563        * These transformations don't change the behavior.
564        *
565        * First, get rid of DST in the blend factors:
566        *    func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
567        */
568       si_blend_remove_dst(&eqRGB, &srcRGB, &dstRGB, PIPE_BLENDFACTOR_DST_COLOR,
569                           PIPE_BLENDFACTOR_SRC_COLOR);
570       si_blend_remove_dst(&eqA, &srcA, &dstA, PIPE_BLENDFACTOR_DST_COLOR,
571                           PIPE_BLENDFACTOR_SRC_COLOR);
572       si_blend_remove_dst(&eqA, &srcA, &dstA, PIPE_BLENDFACTOR_DST_ALPHA,
573                           PIPE_BLENDFACTOR_SRC_ALPHA);
574 
575       /* Look up the ideal settings from tables. */
576       srcRGB_opt = si_translate_blend_opt_factor(srcRGB, false);
577       dstRGB_opt = si_translate_blend_opt_factor(dstRGB, false);
578       srcA_opt = si_translate_blend_opt_factor(srcA, true);
579       dstA_opt = si_translate_blend_opt_factor(dstA, true);
580 
581       /* Handle interdependencies. */
582       if (util_blend_factor_uses_dest(srcRGB, false))
583          dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
584       if (util_blend_factor_uses_dest(srcA, false))
585          dstA_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
586 
587       if (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE &&
588           (dstRGB == PIPE_BLENDFACTOR_ZERO || dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
589            dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE))
590          dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
591 
592       /* Set the final value. */
593       sx_mrt_blend_opt[i] = S_028760_COLOR_SRC_OPT(srcRGB_opt) |
594                             S_028760_COLOR_DST_OPT(dstRGB_opt) |
595                             S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB)) |
596                             S_028760_ALPHA_SRC_OPT(srcA_opt) | S_028760_ALPHA_DST_OPT(dstA_opt) |
597                             S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA));
598 
599       /* Alpha-to-coverage with blending enabled, depth writes enabled, and having no MRTZ export
600        * should disable SX blend optimizations.
601        *
602        * TODO: Add a piglit test for this. It should fail on gfx11 without this.
603        */
604       if (sctx->gfx_level >= GFX11 && state->alpha_to_coverage && i == 0) {
605          sx_mrt_blend_opt[0] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_NONE) |
606                                S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_NONE);
607       }
608 
609       /* Set blend state. */
610       blend_cntl |= S_028780_ENABLE(1);
611       blend_cntl |= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB));
612       blend_cntl |= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(sctx->gfx_level, srcRGB));
613       blend_cntl |= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(sctx->gfx_level, dstRGB));
614 
615       if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) {
616          blend_cntl |= S_028780_SEPARATE_ALPHA_BLEND(1);
617          blend_cntl |= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA));
618          blend_cntl |= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(sctx->gfx_level, srcA));
619          blend_cntl |= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(sctx->gfx_level, dstA));
620       }
621       ac_pm4_set_reg(&pm4->base, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
622       last_blend_cntl = blend_cntl;
623 
624       blend->blend_enable_4bit |= 0xfu << (i * 4);
625 
626       if (sctx->gfx_level >= GFX8 && sctx->gfx_level <= GFX10)
627          blend->dcc_msaa_corruption_4bit |= 0xfu << (i * 4);
628 
629       /* This is only important for formats without alpha. */
630       if (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA || dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
631           srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
632           dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
633           srcRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA || dstRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA)
634          blend->need_src_alpha_4bit |= 0xfu << (i * 4);
635    }
636 
637    if (sctx->gfx_level >= GFX8 && sctx->gfx_level <= GFX10 && logicop_enable)
638       blend->dcc_msaa_corruption_4bit |= blend->cb_target_enabled_4bit;
639 
640    if (blend->cb_target_mask) {
641       color_control |= S_028808_MODE(mode);
642    } else {
643       color_control |= S_028808_MODE(V_028808_CB_DISABLE);
644    }
645 
646    if (sctx->screen->info.rbplus_allowed) {
647       /* Disable RB+ blend optimizations for dual source blending.
648        * Vulkan does this.
649        */
650       if (blend->dual_src_blend) {
651          for (int i = 0; i < num_shader_outputs; i++) {
652             sx_mrt_blend_opt[i] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_NONE) |
653                                   S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_NONE);
654          }
655       }
656 
657       for (int i = 0; i < num_shader_outputs; i++)
658          ac_pm4_set_reg(&pm4->base, R_028760_SX_MRT0_BLEND_OPT + i * 4, sx_mrt_blend_opt[i]);
659 
660       /* RB+ doesn't work with dual source blending, logic op, and RESOLVE. */
661       if (blend->dual_src_blend || logicop_enable || mode == V_028808_CB_RESOLVE ||
662           /* Disabling RB+ improves blending performance in synthetic tests on GFX11. */
663           (sctx->gfx_level == GFX11 && blend->blend_enable_4bit))
664          color_control |= S_028808_DISABLE_DUAL_QUAD(1);
665    }
666 
667    if (sctx->gfx_level >= GFX12)
668       ac_pm4_set_reg(&pm4->base, R_028858_CB_COLOR_CONTROL, color_control);
669    else
670       ac_pm4_set_reg(&pm4->base, R_028808_CB_COLOR_CONTROL, color_control);
671 
672    ac_pm4_finalize(&pm4->base);
673    return blend;
674 }
675 
si_create_blend_state(struct pipe_context * ctx,const struct pipe_blend_state * state)676 static void *si_create_blend_state(struct pipe_context *ctx, const struct pipe_blend_state *state)
677 {
678    return si_create_blend_state_mode(ctx, state, V_028808_CB_NORMAL);
679 }
680 
si_check_blend_dst_sampler_noop(struct si_context * sctx)681 static bool si_check_blend_dst_sampler_noop(struct si_context *sctx)
682 {
683    if (sctx->framebuffer.state.nr_cbufs == 1) {
684       struct si_shader_selector *sel = sctx->shader.ps.cso;
685 
686       if (unlikely(sel->info.writes_1_if_tex_is_1 == 0xff)) {
687          /* Wait for the shader to be ready. */
688          util_queue_fence_wait(&sel->ready);
689          assert(sel->nir_binary);
690 
691          struct nir_shader *nir = si_deserialize_shader(sel);
692 
693          /* Determine if this fragment shader always writes vec4(1) if a specific texture
694           * is all 1s.
695           */
696          float in[4] = { 1.0, 1.0, 1.0, 1.0 };
697          float out[4];
698          int texunit;
699          if (si_nir_is_output_const_if_tex_is_const(nir, in, out, &texunit) &&
700              !memcmp(in, out, 4 * sizeof(float))) {
701             sel->info.writes_1_if_tex_is_1 = 1 + texunit;
702          } else {
703             sel->info.writes_1_if_tex_is_1 = 0;
704          }
705 
706          ralloc_free(nir);
707       }
708 
709       if (sel->info.writes_1_if_tex_is_1 &&
710           sel->info.writes_1_if_tex_is_1 != 0xff) {
711          /* Now check if the texture is cleared to 1 */
712          int unit = sctx->shader.ps.cso->info.writes_1_if_tex_is_1 - 1;
713          struct si_samplers *samp = &sctx->samplers[PIPE_SHADER_FRAGMENT];
714          if ((1u << unit) & samp->enabled_mask) {
715             struct si_texture* tex = (struct si_texture*) samp->views[unit]->texture;
716             if (tex->is_depth &&
717                 tex->depth_cleared_level_mask & BITFIELD_BIT(samp->views[unit]->u.tex.first_level) &&
718                 tex->depth_clear_value[0] == 1) {
719                return false;
720             }
721             /* TODO: handle color textures */
722          }
723       }
724    }
725 
726    return true;
727 }
728 
si_draw_blend_dst_sampler_noop(struct pipe_context * ctx,const struct pipe_draw_info * info,unsigned drawid_offset,const struct pipe_draw_indirect_info * indirect,const struct pipe_draw_start_count_bias * draws,unsigned num_draws)729 static void si_draw_blend_dst_sampler_noop(struct pipe_context *ctx,
730                                            const struct pipe_draw_info *info,
731                                            unsigned drawid_offset,
732                                            const struct pipe_draw_indirect_info *indirect,
733                                            const struct pipe_draw_start_count_bias *draws,
734                                            unsigned num_draws) {
735    struct si_context *sctx = (struct si_context *)ctx;
736 
737    if (!si_check_blend_dst_sampler_noop(sctx))
738       return;
739 
740    sctx->real_draw_vbo(ctx, info, drawid_offset, indirect, draws, num_draws);
741 }
742 
si_draw_vstate_blend_dst_sampler_noop(struct pipe_context * ctx,struct pipe_vertex_state * state,uint32_t partial_velem_mask,struct pipe_draw_vertex_state_info info,const struct pipe_draw_start_count_bias * draws,unsigned num_draws)743 static void si_draw_vstate_blend_dst_sampler_noop(struct pipe_context *ctx,
744                                                   struct pipe_vertex_state *state,
745                                                   uint32_t partial_velem_mask,
746                                                   struct pipe_draw_vertex_state_info info,
747                                                   const struct pipe_draw_start_count_bias *draws,
748                                                   unsigned num_draws) {
749    struct si_context *sctx = (struct si_context *)ctx;
750 
751    if (!si_check_blend_dst_sampler_noop(sctx))
752       return;
753 
754    sctx->real_draw_vertex_state(ctx, state, partial_velem_mask, info, draws, num_draws);
755 }
756 
si_bind_blend_state(struct pipe_context * ctx,void * state)757 static void si_bind_blend_state(struct pipe_context *ctx, void *state)
758 {
759    struct si_context *sctx = (struct si_context *)ctx;
760    struct si_state_blend *old_blend = sctx->queued.named.blend;
761    struct si_state_blend *blend = (struct si_state_blend *)state;
762 
763    if (!blend)
764       blend = (struct si_state_blend *)sctx->noop_blend;
765 
766    si_pm4_bind_state(sctx, blend, blend);
767 
768    if (old_blend->cb_target_mask != blend->cb_target_mask ||
769        old_blend->dual_src_blend != blend->dual_src_blend ||
770        (old_blend->dcc_msaa_corruption_4bit != blend->dcc_msaa_corruption_4bit &&
771         sctx->framebuffer.has_dcc_msaa))
772       si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
773 
774    if ((sctx->screen->info.has_export_conflict_bug &&
775         old_blend->blend_enable_4bit != blend->blend_enable_4bit) ||
776        (sctx->occlusion_query_mode == SI_OCCLUSION_QUERY_MODE_PRECISE_BOOLEAN &&
777         !!old_blend->cb_target_mask != !!blend->cb_target_enabled_4bit))
778       si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
779 
780    if (old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit ||
781        old_blend->alpha_to_coverage != blend->alpha_to_coverage ||
782        old_blend->alpha_to_one != blend->alpha_to_one ||
783        old_blend->dual_src_blend != blend->dual_src_blend ||
784        old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
785        old_blend->need_src_alpha_4bit != blend->need_src_alpha_4bit)
786       si_ps_key_update_framebuffer_blend_rasterizer(sctx);
787 
788    if (old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit ||
789        old_blend->alpha_to_coverage != blend->alpha_to_coverage)
790       si_update_ps_inputs_read_or_disabled(sctx);
791 
792    if (sctx->screen->dpbb_allowed &&
793        (old_blend->alpha_to_coverage != blend->alpha_to_coverage ||
794         old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
795         old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit))
796       si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
797 
798    if (sctx->screen->info.has_out_of_order_rast &&
799        ((old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
800          old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit ||
801          old_blend->commutative_4bit != blend->commutative_4bit ||
802          old_blend->logicop_enable != blend->logicop_enable)))
803       si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
804 
805    /* RB+ depth-only rendering. See the comment where we set rbplus_depth_only_opt for more
806     * information.
807     */
808    if (sctx->screen->info.rbplus_allowed &&
809        !!old_blend->cb_target_mask != !!blend->cb_target_mask) {
810       sctx->framebuffer.dirty_cbufs |= BITFIELD_BIT(0);
811       si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
812    }
813 
814    if (likely(!radeon_uses_secure_bos(sctx->ws))) {
815       if (unlikely(blend->allows_noop_optimization)) {
816          si_install_draw_wrapper(sctx, si_draw_blend_dst_sampler_noop,
817                                  si_draw_vstate_blend_dst_sampler_noop);
818       } else {
819          si_install_draw_wrapper(sctx, NULL, NULL);
820       }
821    }
822 }
823 
si_delete_blend_state(struct pipe_context * ctx,void * state)824 static void si_delete_blend_state(struct pipe_context *ctx, void *state)
825 {
826    struct si_context *sctx = (struct si_context *)ctx;
827 
828    if (sctx->queued.named.blend == state)
829       si_bind_blend_state(ctx, sctx->noop_blend);
830 
831    si_pm4_free_state(sctx, (struct si_pm4_state*)state, SI_STATE_IDX(blend));
832 }
833 
si_set_blend_color(struct pipe_context * ctx,const struct pipe_blend_color * state)834 static void si_set_blend_color(struct pipe_context *ctx, const struct pipe_blend_color *state)
835 {
836    struct si_context *sctx = (struct si_context *)ctx;
837    static const struct pipe_blend_color zeros;
838 
839    sctx->blend_color = *state;
840    sctx->blend_color_any_nonzeros = memcmp(state, &zeros, sizeof(*state)) != 0;
841    si_mark_atom_dirty(sctx, &sctx->atoms.s.blend_color);
842 }
843 
si_emit_blend_color(struct si_context * sctx,unsigned index)844 static void si_emit_blend_color(struct si_context *sctx, unsigned index)
845 {
846    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
847 
848    radeon_begin(cs);
849    radeon_set_context_reg_seq(R_028414_CB_BLEND_RED, 4);
850    radeon_emit_array((uint32_t *)sctx->blend_color.color, 4);
851    radeon_end();
852 }
853 
854 /*
855  * Clipping
856  */
857 
si_set_clip_state(struct pipe_context * ctx,const struct pipe_clip_state * state)858 static void si_set_clip_state(struct pipe_context *ctx, const struct pipe_clip_state *state)
859 {
860    struct si_context *sctx = (struct si_context *)ctx;
861    struct pipe_constant_buffer cb;
862    static const struct pipe_clip_state zeros;
863 
864    if (memcmp(&sctx->clip_state, state, sizeof(*state)) == 0)
865       return;
866 
867    sctx->clip_state = *state;
868    sctx->clip_state_any_nonzeros = memcmp(state, &zeros, sizeof(*state)) != 0;
869    si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_state);
870 
871    cb.buffer = NULL;
872    cb.user_buffer = state->ucp;
873    cb.buffer_offset = 0;
874    cb.buffer_size = 4 * 4 * 8;
875    si_set_internal_const_buffer(sctx, SI_VS_CONST_CLIP_PLANES, &cb);
876 }
877 
si_emit_clip_state(struct si_context * sctx,unsigned index)878 static void si_emit_clip_state(struct si_context *sctx, unsigned index)
879 {
880    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
881 
882    radeon_begin(cs);
883    if (sctx->gfx_level >= GFX12)
884       radeon_set_context_reg_seq(R_0282D0_PA_CL_UCP_0_X, 6 * 4);
885    else
886       radeon_set_context_reg_seq(R_0285BC_PA_CL_UCP_0_X, 6 * 4);
887    radeon_emit_array((uint32_t *)sctx->clip_state.ucp, 6 * 4);
888    radeon_end();
889 }
890 
si_emit_clip_regs(struct si_context * sctx,unsigned index)891 static void si_emit_clip_regs(struct si_context *sctx, unsigned index)
892 {
893    struct si_shader *vs = si_get_vs(sctx)->current;
894    struct si_shader_selector *vs_sel = vs->selector;
895    struct si_shader_info *info = &vs_sel->info;
896    struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
897    bool window_space = vs_sel->stage == MESA_SHADER_VERTEX ?
898                           info->base.vs.window_space_position : 0;
899    unsigned clipdist_mask = vs_sel->info.clipdist_mask;
900    unsigned ucp_mask = clipdist_mask ? 0 : rs->clip_plane_enable & SI_USER_CLIP_PLANE_MASK;
901    unsigned culldist_mask = vs_sel->info.culldist_mask;
902 
903    /* Clip distances on points have no effect, so need to be implemented
904     * as cull distances. This applies for the clipvertex case as well.
905     *
906     * Setting this for primitives other than points should have no adverse
907     * effects.
908     */
909    clipdist_mask &= rs->clip_plane_enable;
910    culldist_mask |= clipdist_mask;
911 
912    unsigned pa_cl_cntl = S_02881C_BYPASS_VTX_RATE_COMBINER(sctx->gfx_level >= GFX10_3 &&
913                                                            !sctx->screen->options.vrs2x2) |
914                          S_02881C_BYPASS_PRIM_RATE_COMBINER(sctx->gfx_level >= GFX10_3) |
915                          clipdist_mask | (culldist_mask << 8);
916 
917    unsigned pa_cl_clip_cntl = rs->pa_cl_clip_cntl | ucp_mask |
918                               S_028810_CLIP_DISABLE(window_space);
919    unsigned pa_cl_vs_out_cntl = pa_cl_cntl | vs->pa_cl_vs_out_cntl;
920 
921    if (sctx->gfx_level >= GFX12) {
922       radeon_begin(&sctx->gfx_cs);
923       gfx12_begin_context_regs();
924       gfx12_opt_set_context_reg(R_028810_PA_CL_CLIP_CNTL, SI_TRACKED_PA_CL_CLIP_CNTL,
925                                 pa_cl_clip_cntl);
926       gfx12_opt_set_context_reg(R_028818_PA_CL_VS_OUT_CNTL, SI_TRACKED_PA_CL_VS_OUT_CNTL,
927                                 pa_cl_vs_out_cntl);
928       gfx12_end_context_regs();
929       radeon_end(); /* don't track context rolls on GFX12 */
930    } else if (sctx->screen->info.has_set_context_pairs_packed) {
931       radeon_begin(&sctx->gfx_cs);
932       gfx11_begin_packed_context_regs();
933       gfx11_opt_set_context_reg(R_028810_PA_CL_CLIP_CNTL, SI_TRACKED_PA_CL_CLIP_CNTL,
934                                 pa_cl_clip_cntl);
935       gfx11_opt_set_context_reg(R_02881C_PA_CL_VS_OUT_CNTL, SI_TRACKED_PA_CL_VS_OUT_CNTL,
936                                 pa_cl_vs_out_cntl);
937       gfx11_end_packed_context_regs();
938       radeon_end(); /* don't track context rolls on GFX11 */
939    } else {
940       radeon_begin(&sctx->gfx_cs);
941       radeon_opt_set_context_reg(R_028810_PA_CL_CLIP_CNTL, SI_TRACKED_PA_CL_CLIP_CNTL,
942                                  pa_cl_clip_cntl);
943       radeon_opt_set_context_reg(R_02881C_PA_CL_VS_OUT_CNTL, SI_TRACKED_PA_CL_VS_OUT_CNTL,
944                                  pa_cl_vs_out_cntl);
945       radeon_end_update_context_roll();
946    }
947 }
948 
949 /*
950  * Rasterizer
951  */
952 
si_translate_fill(uint32_t func)953 static uint32_t si_translate_fill(uint32_t func)
954 {
955    switch (func) {
956    case PIPE_POLYGON_MODE_FILL:
957       return V_028814_X_DRAW_TRIANGLES;
958    case PIPE_POLYGON_MODE_LINE:
959       return V_028814_X_DRAW_LINES;
960    case PIPE_POLYGON_MODE_POINT:
961       return V_028814_X_DRAW_POINTS;
962    default:
963       assert(0);
964       return V_028814_X_DRAW_POINTS;
965    }
966 }
967 
si_create_rs_state(struct pipe_context * ctx,const struct pipe_rasterizer_state * state)968 static void *si_create_rs_state(struct pipe_context *ctx, const struct pipe_rasterizer_state *state)
969 {
970    struct si_screen *sscreen = ((struct si_context *)ctx)->screen;
971    struct si_state_rasterizer *rs = CALLOC_STRUCT(si_state_rasterizer);
972 
973    if (!rs) {
974       return NULL;
975    }
976 
977    rs->scissor_enable = state->scissor;
978    rs->clip_halfz = state->clip_halfz;
979    rs->two_side = state->light_twoside;
980    rs->multisample_enable = state->multisample;
981    rs->force_persample_interp = state->force_persample_interp;
982    rs->clip_plane_enable = state->clip_plane_enable;
983    rs->half_pixel_center = state->half_pixel_center;
984    rs->line_stipple_enable = state->line_stipple_enable;
985    rs->poly_stipple_enable = state->poly_stipple_enable;
986    rs->line_smooth = state->line_smooth;
987    rs->line_width = state->line_width;
988    rs->poly_smooth = state->poly_smooth;
989    rs->point_smooth = state->point_smooth;
990    rs->uses_poly_offset = state->offset_point || state->offset_line || state->offset_tri;
991    rs->clamp_fragment_color = state->clamp_fragment_color;
992    rs->clamp_vertex_color = state->clamp_vertex_color;
993    rs->flatshade = state->flatshade;
994    rs->flatshade_first = state->flatshade_first;
995    rs->sprite_coord_enable = state->sprite_coord_enable;
996    rs->rasterizer_discard = state->rasterizer_discard;
997    rs->bottom_edge_rule = state->bottom_edge_rule;
998    rs->polygon_mode_is_lines =
999       (state->fill_front == PIPE_POLYGON_MODE_LINE && !(state->cull_face & PIPE_FACE_FRONT)) ||
1000       (state->fill_back == PIPE_POLYGON_MODE_LINE && !(state->cull_face & PIPE_FACE_BACK));
1001    rs->polygon_mode_is_points =
1002       (state->fill_front == PIPE_POLYGON_MODE_POINT && !(state->cull_face & PIPE_FACE_FRONT)) ||
1003       (state->fill_back == PIPE_POLYGON_MODE_POINT && !(state->cull_face & PIPE_FACE_BACK));
1004    rs->pa_sc_line_stipple = state->line_stipple_enable ?
1005                                S_028A0C_LINE_PATTERN(state->line_stipple_pattern) |
1006                                S_028A0C_REPEAT_COUNT(state->line_stipple_factor) : 0;
1007    /* TODO: implement line stippling with perpendicular end caps. */
1008    /* Line width > 2 is an internal recommendation. */
1009    rs->perpendicular_end_caps = state->multisample &&
1010                                 state->line_width > 2 && !state->line_stipple_enable;
1011 
1012    rs->pa_cl_clip_cntl = S_028810_DX_CLIP_SPACE_DEF(state->clip_halfz) |
1013                          S_028810_ZCLIP_NEAR_DISABLE(!state->depth_clip_near) |
1014                          S_028810_ZCLIP_FAR_DISABLE(!state->depth_clip_far) |
1015                          S_028810_DX_RASTERIZATION_KILL(state->rasterizer_discard) |
1016                          S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
1017 
1018    rs->ngg_cull_flags_tris = SI_NGG_CULL_TRIANGLES |
1019                              SI_NGG_CULL_CLIP_PLANE_ENABLE(state->clip_plane_enable);
1020    rs->ngg_cull_flags_tris_y_inverted = rs->ngg_cull_flags_tris;
1021 
1022    rs->ngg_cull_flags_lines = SI_NGG_CULL_LINES |
1023                               (!rs->perpendicular_end_caps ? SI_NGG_CULL_SMALL_LINES_DIAMOND_EXIT : 0) |
1024                               SI_NGG_CULL_CLIP_PLANE_ENABLE(state->clip_plane_enable);
1025 
1026    if (rs->rasterizer_discard) {
1027       rs->ngg_cull_flags_tris |= SI_NGG_CULL_FRONT_FACE |
1028                                  SI_NGG_CULL_BACK_FACE;
1029       rs->ngg_cull_flags_tris_y_inverted = rs->ngg_cull_flags_tris;
1030    } else {
1031       bool cull_front, cull_back;
1032 
1033       if (!state->front_ccw) {
1034          cull_front = !!(state->cull_face & PIPE_FACE_FRONT);
1035          cull_back = !!(state->cull_face & PIPE_FACE_BACK);
1036       } else {
1037          cull_back = !!(state->cull_face & PIPE_FACE_FRONT);
1038          cull_front = !!(state->cull_face & PIPE_FACE_BACK);
1039       }
1040 
1041       if (cull_front) {
1042          rs->ngg_cull_flags_tris |= SI_NGG_CULL_FRONT_FACE;
1043          rs->ngg_cull_flags_tris_y_inverted |= SI_NGG_CULL_BACK_FACE;
1044       }
1045 
1046       if (cull_back) {
1047          rs->ngg_cull_flags_tris |= SI_NGG_CULL_BACK_FACE;
1048          rs->ngg_cull_flags_tris_y_inverted |= SI_NGG_CULL_FRONT_FACE;
1049       }
1050    }
1051 
1052    /* Force gl_FrontFacing to true or false if the other face is culled. */
1053    if (util_bitcount(state->cull_face) == 1) {
1054       if (state->cull_face & PIPE_FACE_FRONT)
1055          rs->force_front_face_input = -1;
1056       else
1057          rs->force_front_face_input = 1;
1058    }
1059 
1060    rs->spi_interp_control_0 = S_0286D4_FLAT_SHADE_ENA(1) |
1061                               S_0286D4_PNT_SPRITE_ENA(state->point_quad_rasterization) |
1062                               S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) |
1063                               S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T) |
1064                               S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0) |
1065                               S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) |
1066                               S_0286D4_PNT_SPRITE_TOP_1(state->sprite_coord_mode !=
1067                                                         PIPE_SPRITE_COORD_UPPER_LEFT);
1068 
1069    /* point size 12.4 fixed point */
1070    float psize_min, psize_max;
1071    unsigned tmp = (unsigned)(state->point_size * 8.0);
1072    rs->pa_su_point_size = S_028A00_HEIGHT(tmp) | S_028A00_WIDTH(tmp);
1073 
1074    if (state->point_size_per_vertex) {
1075       psize_min = util_get_min_point_size(state);
1076       psize_max = SI_MAX_POINT_SIZE;
1077    } else {
1078       /* Force the point size to be as if the vertex output was disabled. */
1079       psize_min = state->point_size;
1080       psize_max = state->point_size;
1081    }
1082    rs->max_point_size = psize_max;
1083 
1084    /* Divide by two, because 0.5 = 1 pixel. */
1085    rs->pa_su_point_minmax = S_028A04_MIN_SIZE(si_pack_float_12p4(psize_min / 2)) |
1086                             S_028A04_MAX_SIZE(si_pack_float_12p4(psize_max / 2));
1087    rs->pa_su_line_cntl = S_028A08_WIDTH(si_pack_float_12p4(state->line_width / 2));
1088 
1089    rs->pa_sc_mode_cntl_0 = S_028A48_LINE_STIPPLE_ENABLE(state->line_stipple_enable) |
1090                            S_028A48_MSAA_ENABLE(state->multisample || state->poly_smooth ||
1091                                                 state->line_smooth) |
1092                            S_028A48_VPORT_SCISSOR_ENABLE(1) |
1093                            S_028A48_ALTERNATE_RBS_PER_TILE(sscreen->info.gfx_level >= GFX9);
1094 
1095    bool polygon_mode_enabled =
1096       (state->fill_front != PIPE_POLYGON_MODE_FILL && !(state->cull_face & PIPE_FACE_FRONT)) ||
1097       (state->fill_back != PIPE_POLYGON_MODE_FILL && !(state->cull_face & PIPE_FACE_BACK));
1098 
1099    rs->pa_su_sc_mode_cntl = S_028814_PROVOKING_VTX_LAST(!state->flatshade_first) |
1100                             S_028814_CULL_FRONT((state->cull_face & PIPE_FACE_FRONT) ? 1 : 0) |
1101                             S_028814_CULL_BACK((state->cull_face & PIPE_FACE_BACK) ? 1 : 0) |
1102                             S_028814_FACE(!state->front_ccw) |
1103                             S_028814_POLY_OFFSET_FRONT_ENABLE(util_get_offset(state, state->fill_front)) |
1104                             S_028814_POLY_OFFSET_BACK_ENABLE(util_get_offset(state, state->fill_back)) |
1105                             S_028814_POLY_OFFSET_PARA_ENABLE(state->offset_point || state->offset_line) |
1106                             S_028814_POLY_MODE(polygon_mode_enabled) |
1107                             S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(state->fill_front)) |
1108                             S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(state->fill_back)) |
1109                             /* this must be set if POLY_MODE or PERPENDICULAR_ENDCAP_ENA is set */
1110                             S_028814_KEEP_TOGETHER_ENABLE(sscreen->info.gfx_level >= GFX10 &&
1111                                                           sscreen->info.gfx_level < GFX12 ?
1112                                                              polygon_mode_enabled ||
1113                                                              rs->perpendicular_end_caps : 0);
1114    if (sscreen->info.gfx_level >= GFX10) {
1115       rs->pa_cl_ngg_cntl = S_028838_INDEX_BUF_EDGE_FLAG_ENA(rs->polygon_mode_is_points ||
1116                                                             rs->polygon_mode_is_lines) |
1117                            S_028838_VERTEX_REUSE_DEPTH(sscreen->info.gfx_level >= GFX10_3 ? 30 : 0);
1118    }
1119 
1120    if (state->bottom_edge_rule) {
1121       /* OpenGL windows should set this. */
1122       rs->pa_sc_edgerule = S_028230_ER_TRI(0xA) |
1123                            S_028230_ER_POINT(0x5) |
1124                            S_028230_ER_RECT(0x9) |
1125                            S_028230_ER_LINE_LR(0x2A) |
1126                            S_028230_ER_LINE_RL(0x2A) |
1127                            S_028230_ER_LINE_TB(0xA) |
1128                            S_028230_ER_LINE_BT(0xA);
1129    } else {
1130       /* OpenGL FBOs and Direct3D should set this. */
1131       rs->pa_sc_edgerule = S_028230_ER_TRI(0xA) |
1132                            S_028230_ER_POINT(0x6) |
1133                            S_028230_ER_RECT(0xA) |
1134                            S_028230_ER_LINE_LR(0x19) |
1135                            S_028230_ER_LINE_RL(0x25) |
1136                            S_028230_ER_LINE_TB(0xA) |
1137                            S_028230_ER_LINE_BT(0xA);
1138    }
1139 
1140    if (rs->uses_poly_offset) {
1141       /* Calculate polygon offset states for 16-bit, 24-bit, and 32-bit zbuffers. */
1142       rs->pa_su_poly_offset_clamp = fui(state->offset_clamp);
1143       rs->pa_su_poly_offset_frontback_scale = fui(state->offset_scale * 16);
1144 
1145       if (!state->offset_units_unscaled) {
1146          /* 16-bit zbuffer */
1147          rs->pa_su_poly_offset_db_fmt_cntl[0] = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
1148          rs->pa_su_poly_offset_frontback_offset[0] = fui(state->offset_units * 4);
1149 
1150          /* 24-bit zbuffer */
1151          rs->pa_su_poly_offset_db_fmt_cntl[1] = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
1152          rs->pa_su_poly_offset_frontback_offset[1] = fui(state->offset_units * 2);
1153 
1154          /* 32-bit zbuffer */
1155          rs->pa_su_poly_offset_db_fmt_cntl[2] = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
1156                                                 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
1157          rs->pa_su_poly_offset_frontback_offset[2] = fui(state->offset_units);
1158       } else {
1159          rs->pa_su_poly_offset_frontback_offset[0] = fui(state->offset_units);
1160          rs->pa_su_poly_offset_frontback_offset[1] = fui(state->offset_units);
1161          rs->pa_su_poly_offset_frontback_offset[2] = fui(state->offset_units);
1162       }
1163    }
1164 
1165    return rs;
1166 }
1167 
si_pm4_emit_rasterizer(struct si_context * sctx,unsigned index)1168 static void si_pm4_emit_rasterizer(struct si_context *sctx, unsigned index)
1169 {
1170    struct si_state_rasterizer *state = sctx->queued.named.rasterizer;
1171 
1172    if (sctx->screen->info.gfx_level >= GFX12) {
1173       radeon_begin(&sctx->gfx_cs);
1174       gfx12_begin_context_regs();
1175       if (state->line_stipple_enable) {
1176          gfx12_opt_set_context_reg(R_028A0C_PA_SC_LINE_STIPPLE, SI_TRACKED_PA_SC_LINE_STIPPLE,
1177                                    state->pa_sc_line_stipple);
1178       }
1179 
1180       gfx12_opt_set_context_reg(R_028644_SPI_INTERP_CONTROL_0, SI_TRACKED_SPI_INTERP_CONTROL_0,
1181                                 state->spi_interp_control_0);
1182       gfx12_opt_set_context_reg(R_028A00_PA_SU_POINT_SIZE, SI_TRACKED_PA_SU_POINT_SIZE,
1183                                 state->pa_su_point_size);
1184       gfx12_opt_set_context_reg(R_028A04_PA_SU_POINT_MINMAX, SI_TRACKED_PA_SU_POINT_MINMAX,
1185                                 state->pa_su_point_minmax);
1186       gfx12_opt_set_context_reg(R_028A08_PA_SU_LINE_CNTL, SI_TRACKED_PA_SU_LINE_CNTL,
1187                                 state->pa_su_line_cntl);
1188       gfx12_opt_set_context_reg(R_028A48_PA_SC_MODE_CNTL_0, SI_TRACKED_PA_SC_MODE_CNTL_0,
1189                                 state->pa_sc_mode_cntl_0);
1190       gfx12_opt_set_context_reg(R_02881C_PA_SU_SC_MODE_CNTL, SI_TRACKED_PA_SU_SC_MODE_CNTL,
1191                                 state->pa_su_sc_mode_cntl);
1192       gfx12_opt_set_context_reg(R_028838_PA_CL_NGG_CNTL, SI_TRACKED_PA_CL_NGG_CNTL,
1193                                 state->pa_cl_ngg_cntl);
1194       gfx12_opt_set_context_reg(R_028230_PA_SC_EDGERULE, SI_TRACKED_PA_SC_EDGERULE,
1195                                 state->pa_sc_edgerule);
1196 
1197       if (state->uses_poly_offset && sctx->framebuffer.state.zsbuf) {
1198          unsigned db_format_index =
1199             ((struct si_surface *)sctx->framebuffer.state.zsbuf)->db_format_index;
1200 
1201          gfx12_opt_set_context_reg(R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
1202                                    SI_TRACKED_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
1203                                    state->pa_su_poly_offset_db_fmt_cntl[db_format_index]);
1204          gfx12_opt_set_context_reg(R_028B7C_PA_SU_POLY_OFFSET_CLAMP,
1205                                    SI_TRACKED_PA_SU_POLY_OFFSET_CLAMP,
1206                                    state->pa_su_poly_offset_clamp);
1207          gfx12_opt_set_context_reg(R_028B80_PA_SU_POLY_OFFSET_FRONT_SCALE,
1208                                    SI_TRACKED_PA_SU_POLY_OFFSET_FRONT_SCALE,
1209                                    state->pa_su_poly_offset_frontback_scale);
1210          gfx12_opt_set_context_reg(R_028B84_PA_SU_POLY_OFFSET_FRONT_OFFSET,
1211                                    SI_TRACKED_PA_SU_POLY_OFFSET_FRONT_OFFSET,
1212                                    state->pa_su_poly_offset_frontback_offset[db_format_index]);
1213          gfx12_opt_set_context_reg(R_028B88_PA_SU_POLY_OFFSET_BACK_SCALE,
1214                                    SI_TRACKED_PA_SU_POLY_OFFSET_BACK_SCALE,
1215                                    state->pa_su_poly_offset_frontback_scale);
1216          gfx12_opt_set_context_reg(R_028B8C_PA_SU_POLY_OFFSET_BACK_OFFSET,
1217                                    SI_TRACKED_PA_SU_POLY_OFFSET_BACK_OFFSET,
1218                                    state->pa_su_poly_offset_frontback_offset[db_format_index]);
1219       }
1220       gfx12_end_context_regs();
1221       radeon_end(); /* don't track context rolls on GFX12 */
1222    } else if (sctx->screen->info.has_set_context_pairs_packed) {
1223       radeon_begin(&sctx->gfx_cs);
1224       gfx11_begin_packed_context_regs();
1225       gfx11_opt_set_context_reg(R_0286D4_SPI_INTERP_CONTROL_0, SI_TRACKED_SPI_INTERP_CONTROL_0,
1226                                 state->spi_interp_control_0);
1227       gfx11_opt_set_context_reg(R_028A00_PA_SU_POINT_SIZE, SI_TRACKED_PA_SU_POINT_SIZE,
1228                                 state->pa_su_point_size);
1229       gfx11_opt_set_context_reg(R_028A04_PA_SU_POINT_MINMAX, SI_TRACKED_PA_SU_POINT_MINMAX,
1230                                 state->pa_su_point_minmax);
1231       gfx11_opt_set_context_reg(R_028A08_PA_SU_LINE_CNTL, SI_TRACKED_PA_SU_LINE_CNTL,
1232                                 state->pa_su_line_cntl);
1233       gfx11_opt_set_context_reg(R_028A48_PA_SC_MODE_CNTL_0, SI_TRACKED_PA_SC_MODE_CNTL_0,
1234                                 state->pa_sc_mode_cntl_0);
1235       gfx11_opt_set_context_reg(R_028814_PA_SU_SC_MODE_CNTL, SI_TRACKED_PA_SU_SC_MODE_CNTL,
1236                                 state->pa_su_sc_mode_cntl);
1237       gfx11_opt_set_context_reg(R_028838_PA_CL_NGG_CNTL, SI_TRACKED_PA_CL_NGG_CNTL,
1238                                 state->pa_cl_ngg_cntl);
1239       gfx11_opt_set_context_reg(R_028230_PA_SC_EDGERULE, SI_TRACKED_PA_SC_EDGERULE,
1240                                 state->pa_sc_edgerule);
1241 
1242       if (state->uses_poly_offset && sctx->framebuffer.state.zsbuf) {
1243          unsigned db_format_index =
1244             ((struct si_surface *)sctx->framebuffer.state.zsbuf)->db_format_index;
1245 
1246          gfx11_opt_set_context_reg(R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
1247                                    SI_TRACKED_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
1248                                    state->pa_su_poly_offset_db_fmt_cntl[db_format_index]);
1249          gfx11_opt_set_context_reg(R_028B7C_PA_SU_POLY_OFFSET_CLAMP,
1250                                    SI_TRACKED_PA_SU_POLY_OFFSET_CLAMP,
1251                                    state->pa_su_poly_offset_clamp);
1252          gfx11_opt_set_context_reg(R_028B80_PA_SU_POLY_OFFSET_FRONT_SCALE,
1253                                    SI_TRACKED_PA_SU_POLY_OFFSET_FRONT_SCALE,
1254                                    state->pa_su_poly_offset_frontback_scale);
1255          gfx11_opt_set_context_reg(R_028B84_PA_SU_POLY_OFFSET_FRONT_OFFSET,
1256                                    SI_TRACKED_PA_SU_POLY_OFFSET_FRONT_OFFSET,
1257                                    state->pa_su_poly_offset_frontback_offset[db_format_index]);
1258          gfx11_opt_set_context_reg(R_028B88_PA_SU_POLY_OFFSET_BACK_SCALE,
1259                                    SI_TRACKED_PA_SU_POLY_OFFSET_BACK_SCALE,
1260                                    state->pa_su_poly_offset_frontback_scale);
1261          gfx11_opt_set_context_reg(R_028B8C_PA_SU_POLY_OFFSET_BACK_OFFSET,
1262                                    SI_TRACKED_PA_SU_POLY_OFFSET_BACK_OFFSET,
1263                                    state->pa_su_poly_offset_frontback_offset[db_format_index]);
1264       }
1265       gfx11_end_packed_context_regs();
1266       radeon_end(); /* don't track context rolls on GFX11 */
1267    } else {
1268       radeon_begin(&sctx->gfx_cs);
1269       radeon_opt_set_context_reg(R_0286D4_SPI_INTERP_CONTROL_0,
1270                                  SI_TRACKED_SPI_INTERP_CONTROL_0,
1271                                  state->spi_interp_control_0);
1272       radeon_opt_set_context_reg(R_028A00_PA_SU_POINT_SIZE, SI_TRACKED_PA_SU_POINT_SIZE,
1273                                  state->pa_su_point_size);
1274       radeon_opt_set_context_reg(R_028A04_PA_SU_POINT_MINMAX, SI_TRACKED_PA_SU_POINT_MINMAX,
1275                                  state->pa_su_point_minmax);
1276       radeon_opt_set_context_reg(R_028A08_PA_SU_LINE_CNTL, SI_TRACKED_PA_SU_LINE_CNTL,
1277                                  state->pa_su_line_cntl);
1278       radeon_opt_set_context_reg(R_028A48_PA_SC_MODE_CNTL_0, SI_TRACKED_PA_SC_MODE_CNTL_0,
1279                                  state->pa_sc_mode_cntl_0);
1280       radeon_opt_set_context_reg(R_028814_PA_SU_SC_MODE_CNTL,
1281                                  SI_TRACKED_PA_SU_SC_MODE_CNTL, state->pa_su_sc_mode_cntl);
1282       if (sctx->gfx_level >= GFX10) {
1283          radeon_opt_set_context_reg(R_028838_PA_CL_NGG_CNTL, SI_TRACKED_PA_CL_NGG_CNTL,
1284                                     state->pa_cl_ngg_cntl);
1285       }
1286       radeon_opt_set_context_reg(R_028230_PA_SC_EDGERULE, SI_TRACKED_PA_SC_EDGERULE,
1287                                  state->pa_sc_edgerule);
1288 
1289       if (state->uses_poly_offset && sctx->framebuffer.state.zsbuf) {
1290          unsigned db_format_index =
1291             ((struct si_surface *)sctx->framebuffer.state.zsbuf)->db_format_index;
1292 
1293          radeon_opt_set_context_reg6(R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
1294                                      SI_TRACKED_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
1295                                      state->pa_su_poly_offset_db_fmt_cntl[db_format_index],
1296                                      state->pa_su_poly_offset_clamp,
1297                                      state->pa_su_poly_offset_frontback_scale,
1298                                      state->pa_su_poly_offset_frontback_offset[db_format_index],
1299                                      state->pa_su_poly_offset_frontback_scale,
1300                                      state->pa_su_poly_offset_frontback_offset[db_format_index]);
1301       }
1302       radeon_end_update_context_roll();
1303    }
1304 
1305    sctx->emitted.named.rasterizer = state;
1306 }
1307 
si_bind_rs_state(struct pipe_context * ctx,void * state)1308 static void si_bind_rs_state(struct pipe_context *ctx, void *state)
1309 {
1310    struct si_context *sctx = (struct si_context *)ctx;
1311    struct si_state_rasterizer *old_rs = (struct si_state_rasterizer *)sctx->queued.named.rasterizer;
1312    struct si_state_rasterizer *rs = (struct si_state_rasterizer *)state;
1313 
1314    if (!rs)
1315       rs = (struct si_state_rasterizer *)sctx->discard_rasterizer_state;
1316 
1317    if (old_rs->multisample_enable != rs->multisample_enable) {
1318       si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
1319 
1320       /* Update the small primitive filter workaround if necessary. */
1321       if (sctx->screen->info.has_small_prim_filter_sample_loc_bug && sctx->framebuffer.nr_samples > 1)
1322          si_mark_atom_dirty(sctx, &sctx->atoms.s.sample_locations);
1323 
1324       /* NGG cull state uses multisample_enable. */
1325       if (sctx->screen->use_ngg_culling)
1326          si_mark_atom_dirty(sctx, &sctx->atoms.s.ngg_cull_state);
1327    }
1328 
1329    if (old_rs->perpendicular_end_caps != rs->perpendicular_end_caps)
1330       si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
1331 
1332    if (sctx->screen->use_ngg_culling &&
1333        (old_rs->half_pixel_center != rs->half_pixel_center ||
1334         old_rs->line_width != rs->line_width))
1335       si_mark_atom_dirty(sctx, &sctx->atoms.s.ngg_cull_state);
1336 
1337    SET_FIELD(sctx->current_vs_state, VS_STATE_CLAMP_VERTEX_COLOR, rs->clamp_vertex_color);
1338 
1339    si_pm4_bind_state(sctx, rasterizer, rs);
1340 
1341    if (old_rs->scissor_enable != rs->scissor_enable)
1342       si_mark_atom_dirty(sctx, &sctx->atoms.s.scissors);
1343 
1344    /* This never changes for OpenGL. */
1345    if (old_rs->half_pixel_center != rs->half_pixel_center)
1346       si_mark_atom_dirty(sctx, &sctx->atoms.s.guardband);
1347 
1348    if (util_prim_is_lines(sctx->current_rast_prim))
1349       si_set_clip_discard_distance(sctx, rs->line_width);
1350    else if (sctx->current_rast_prim == MESA_PRIM_POINTS)
1351       si_set_clip_discard_distance(sctx, rs->max_point_size);
1352 
1353    if (old_rs->clip_halfz != rs->clip_halfz)
1354       si_mark_atom_dirty(sctx, &sctx->atoms.s.viewports);
1355 
1356    if (old_rs->clip_plane_enable != rs->clip_plane_enable ||
1357        old_rs->pa_cl_clip_cntl != rs->pa_cl_clip_cntl)
1358       si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_regs);
1359 
1360    if (old_rs->sprite_coord_enable != rs->sprite_coord_enable ||
1361        old_rs->flatshade != rs->flatshade)
1362       si_mark_atom_dirty(sctx, &sctx->atoms.s.spi_map);
1363 
1364    if (sctx->screen->dpbb_allowed && (old_rs->bottom_edge_rule != rs->bottom_edge_rule))
1365       si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
1366 
1367    if (old_rs->multisample_enable != rs->multisample_enable)
1368       si_ps_key_update_framebuffer_blend_rasterizer(sctx);
1369 
1370    if (old_rs->flatshade != rs->flatshade ||
1371        old_rs->clamp_fragment_color != rs->clamp_fragment_color)
1372       si_ps_key_update_rasterizer(sctx);
1373 
1374    if (old_rs->flatshade != rs->flatshade ||
1375        old_rs->force_persample_interp != rs->force_persample_interp ||
1376        old_rs->multisample_enable != rs->multisample_enable)
1377       si_ps_key_update_framebuffer_rasterizer_sample_shading(sctx);
1378 
1379    if (old_rs->rasterizer_discard != rs->rasterizer_discard ||
1380        old_rs->two_side != rs->two_side ||
1381        old_rs->poly_stipple_enable != rs->poly_stipple_enable ||
1382        old_rs->point_smooth != rs->point_smooth)
1383       si_update_ps_inputs_read_or_disabled(sctx);
1384 
1385    if (old_rs->point_smooth != rs->point_smooth ||
1386        old_rs->line_smooth != rs->line_smooth ||
1387        old_rs->poly_smooth != rs->poly_smooth ||
1388        old_rs->polygon_mode_is_points != rs->polygon_mode_is_points ||
1389        old_rs->poly_stipple_enable != rs->poly_stipple_enable ||
1390        old_rs->two_side != rs->two_side ||
1391        old_rs->force_front_face_input != rs->force_front_face_input)
1392       si_vs_ps_key_update_rast_prim_smooth_stipple(sctx);
1393 
1394    /* Used by si_get_vs_key_outputs in si_update_shaders: */
1395    if (old_rs->clip_plane_enable != rs->clip_plane_enable)
1396       sctx->do_update_shaders = true;
1397 
1398    if (old_rs->line_smooth != rs->line_smooth ||
1399        old_rs->poly_smooth != rs->poly_smooth ||
1400        old_rs->point_smooth != rs->point_smooth ||
1401        old_rs->poly_stipple_enable != rs->poly_stipple_enable ||
1402        old_rs->flatshade != rs->flatshade)
1403       si_update_vrs_flat_shading(sctx);
1404 
1405    if (old_rs->flatshade_first != rs->flatshade_first)
1406       si_update_ngg_sgpr_state_provoking_vtx(sctx, si_get_vs(sctx)->current, sctx->ngg);
1407 }
1408 
si_delete_rs_state(struct pipe_context * ctx,void * state)1409 static void si_delete_rs_state(struct pipe_context *ctx, void *state)
1410 {
1411    struct si_context *sctx = (struct si_context *)ctx;
1412    struct si_state_rasterizer *rs = (struct si_state_rasterizer *)state;
1413 
1414    if (sctx->queued.named.rasterizer == state)
1415       si_bind_rs_state(ctx, sctx->discard_rasterizer_state);
1416 
1417    si_pm4_free_state(sctx, &rs->pm4, SI_STATE_IDX(rasterizer));
1418 }
1419 
1420 /*
1421  * inferred state between dsa and stencil ref
1422  */
si_emit_stencil_ref(struct si_context * sctx,unsigned index)1423 static void si_emit_stencil_ref(struct si_context *sctx, unsigned index)
1424 {
1425    struct pipe_stencil_ref *ref = &sctx->stencil_ref.state;
1426 
1427    if (sctx->gfx_level >= GFX12) {
1428       radeon_begin(&sctx->gfx_cs);
1429       radeon_set_context_reg(R_028088_DB_STENCIL_REF,
1430                              S_028088_TESTVAL(ref->ref_value[0]) |
1431                              S_028088_TESTVAL_BF(ref->ref_value[1]));
1432       radeon_end();
1433    } else {
1434       struct si_dsa_stencil_ref_part *dsa = &sctx->stencil_ref.dsa_part;
1435 
1436       radeon_begin(&sctx->gfx_cs);
1437       radeon_set_context_reg_seq(R_028430_DB_STENCILREFMASK, 2);
1438       radeon_emit(S_028430_STENCILTESTVAL(ref->ref_value[0]) |
1439                   S_028430_STENCILMASK(dsa->valuemask[0]) |
1440                   S_028430_STENCILWRITEMASK(dsa->writemask[0]) |
1441                   S_028430_STENCILOPVAL(1));
1442       radeon_emit(S_028434_STENCILTESTVAL_BF(ref->ref_value[1]) |
1443                   S_028434_STENCILMASK_BF(dsa->valuemask[1]) |
1444                   S_028434_STENCILWRITEMASK_BF(dsa->writemask[1]) |
1445                   S_028434_STENCILOPVAL_BF(1));
1446       radeon_end();
1447    }
1448 }
1449 
si_set_stencil_ref(struct pipe_context * ctx,const struct pipe_stencil_ref state)1450 static void si_set_stencil_ref(struct pipe_context *ctx, const struct pipe_stencil_ref state)
1451 {
1452    struct si_context *sctx = (struct si_context *)ctx;
1453 
1454    if (memcmp(&sctx->stencil_ref.state, &state, sizeof(state)) == 0)
1455       return;
1456 
1457    sctx->stencil_ref.state = state;
1458    si_mark_atom_dirty(sctx, &sctx->atoms.s.stencil_ref);
1459 }
1460 
1461 /*
1462  * DSA
1463  */
1464 
si_translate_stencil_op(int s_op)1465 static uint32_t si_translate_stencil_op(int s_op)
1466 {
1467    switch (s_op) {
1468    case PIPE_STENCIL_OP_KEEP:
1469       return V_02842C_STENCIL_KEEP;
1470    case PIPE_STENCIL_OP_ZERO:
1471       return V_02842C_STENCIL_ZERO;
1472    case PIPE_STENCIL_OP_REPLACE:
1473       return V_02842C_STENCIL_REPLACE_TEST;
1474    case PIPE_STENCIL_OP_INCR:
1475       return V_02842C_STENCIL_ADD_CLAMP;
1476    case PIPE_STENCIL_OP_DECR:
1477       return V_02842C_STENCIL_SUB_CLAMP;
1478    case PIPE_STENCIL_OP_INCR_WRAP:
1479       return V_02842C_STENCIL_ADD_WRAP;
1480    case PIPE_STENCIL_OP_DECR_WRAP:
1481       return V_02842C_STENCIL_SUB_WRAP;
1482    case PIPE_STENCIL_OP_INVERT:
1483       return V_02842C_STENCIL_INVERT;
1484    default:
1485       PRINT_ERR("Unknown stencil op %d", s_op);
1486       assert(0);
1487       break;
1488    }
1489    return 0;
1490 }
1491 
si_order_invariant_stencil_op(enum pipe_stencil_op op)1492 static bool si_order_invariant_stencil_op(enum pipe_stencil_op op)
1493 {
1494    /* REPLACE is normally order invariant, except when the stencil
1495     * reference value is written by the fragment shader. Tracking this
1496     * interaction does not seem worth the effort, so be conservative. */
1497    return op != PIPE_STENCIL_OP_INCR && op != PIPE_STENCIL_OP_DECR && op != PIPE_STENCIL_OP_REPLACE;
1498 }
1499 
1500 /* Compute whether, assuming Z writes are disabled, this stencil state is order
1501  * invariant in the sense that the set of passing fragments as well as the
1502  * final stencil buffer result does not depend on the order of fragments. */
si_order_invariant_stencil_state(const struct pipe_stencil_state * state)1503 static bool si_order_invariant_stencil_state(const struct pipe_stencil_state *state)
1504 {
1505    return !state->enabled || !state->writemask ||
1506           /* The following assumes that Z writes are disabled. */
1507           (state->func == PIPE_FUNC_ALWAYS && si_order_invariant_stencil_op(state->zpass_op) &&
1508            si_order_invariant_stencil_op(state->zfail_op)) ||
1509           (state->func == PIPE_FUNC_NEVER && si_order_invariant_stencil_op(state->fail_op));
1510 }
1511 
si_create_dsa_state(struct pipe_context * ctx,const struct pipe_depth_stencil_alpha_state * state)1512 static void *si_create_dsa_state(struct pipe_context *ctx,
1513                                  const struct pipe_depth_stencil_alpha_state *state)
1514 {
1515    struct si_context *sctx = (struct si_context *)ctx;
1516    struct si_state_dsa *dsa = CALLOC_STRUCT(si_state_dsa);
1517    if (!dsa) {
1518       return NULL;
1519    }
1520 
1521    dsa->stencil_ref.valuemask[0] = state->stencil[0].valuemask;
1522    dsa->stencil_ref.valuemask[1] = state->stencil[1].valuemask;
1523    dsa->stencil_ref.writemask[0] = state->stencil[0].writemask;
1524    dsa->stencil_ref.writemask[1] = state->stencil[1].writemask;
1525 
1526    dsa->db_depth_control =
1527       S_028800_Z_ENABLE(state->depth_enabled) | S_028800_Z_WRITE_ENABLE(state->depth_writemask) |
1528       S_028800_ZFUNC(state->depth_func) | S_028800_DEPTH_BOUNDS_ENABLE(state->depth_bounds_test);
1529 
1530    /* stencil */
1531    if (state->stencil[0].enabled) {
1532       dsa->db_depth_control |= S_028800_STENCIL_ENABLE(1);
1533       dsa->db_depth_control |= S_028800_STENCILFUNC(state->stencil[0].func);
1534       dsa->db_stencil_control |=
1535          S_02842C_STENCILFAIL(si_translate_stencil_op(state->stencil[0].fail_op));
1536       dsa->db_stencil_control |=
1537          S_02842C_STENCILZPASS(si_translate_stencil_op(state->stencil[0].zpass_op));
1538       dsa->db_stencil_control |=
1539          S_02842C_STENCILZFAIL(si_translate_stencil_op(state->stencil[0].zfail_op));
1540 
1541       if (state->stencil[1].enabled) {
1542          dsa->db_depth_control |= S_028800_BACKFACE_ENABLE(1);
1543          dsa->db_depth_control |= S_028800_STENCILFUNC_BF(state->stencil[1].func);
1544          dsa->db_stencil_control |=
1545             S_02842C_STENCILFAIL_BF(si_translate_stencil_op(state->stencil[1].fail_op));
1546          dsa->db_stencil_control |=
1547             S_02842C_STENCILZPASS_BF(si_translate_stencil_op(state->stencil[1].zpass_op));
1548          dsa->db_stencil_control |=
1549             S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(state->stencil[1].zfail_op));
1550       }
1551    }
1552 
1553    dsa->db_depth_bounds_min = fui(state->depth_bounds_min);
1554    dsa->db_depth_bounds_max = fui(state->depth_bounds_max);
1555 
1556    /* alpha */
1557    if (state->alpha_enabled) {
1558       dsa->alpha_func = state->alpha_func;
1559       dsa->spi_shader_user_data_ps_alpha_ref = fui(state->alpha_ref_value);
1560    } else {
1561       dsa->alpha_func = PIPE_FUNC_ALWAYS;
1562    }
1563 
1564    dsa->depth_enabled = state->depth_enabled;
1565    dsa->depth_write_enabled = state->depth_enabled && state->depth_writemask;
1566    dsa->stencil_enabled = state->stencil[0].enabled;
1567    dsa->stencil_write_enabled =
1568       (util_writes_stencil(&state->stencil[0]) || util_writes_stencil(&state->stencil[1]));
1569    dsa->db_can_write = dsa->depth_write_enabled || dsa->stencil_write_enabled;
1570    dsa->depth_bounds_enabled = state->depth_bounds_test;
1571 
1572    if (sctx->gfx_level >= GFX12) {
1573       dsa->db_stencil_read_mask = S_028090_TESTMASK(state->stencil[0].valuemask) |
1574                                   S_028090_TESTMASK_BF(state->stencil[1].valuemask);
1575       dsa->db_stencil_write_mask = S_028094_WRITEMASK(state->stencil[0].writemask) |
1576                                    S_028094_WRITEMASK_BF(state->stencil[1].writemask);
1577    }
1578 
1579    bool zfunc_is_ordered =
1580       state->depth_func == PIPE_FUNC_NEVER || state->depth_func == PIPE_FUNC_LESS ||
1581       state->depth_func == PIPE_FUNC_LEQUAL || state->depth_func == PIPE_FUNC_GREATER ||
1582       state->depth_func == PIPE_FUNC_GEQUAL;
1583 
1584    bool nozwrite_and_order_invariant_stencil =
1585       !dsa->db_can_write ||
1586       (!dsa->depth_write_enabled && si_order_invariant_stencil_state(&state->stencil[0]) &&
1587        si_order_invariant_stencil_state(&state->stencil[1]));
1588 
1589    dsa->order_invariance[1].zs =
1590       nozwrite_and_order_invariant_stencil || (!dsa->stencil_write_enabled && zfunc_is_ordered);
1591    dsa->order_invariance[0].zs = !dsa->depth_write_enabled || zfunc_is_ordered;
1592 
1593    dsa->order_invariance[1].pass_set =
1594       nozwrite_and_order_invariant_stencil ||
1595       (!dsa->stencil_write_enabled &&
1596        (state->depth_func == PIPE_FUNC_ALWAYS || state->depth_func == PIPE_FUNC_NEVER));
1597    dsa->order_invariance[0].pass_set =
1598       !dsa->depth_write_enabled ||
1599       (state->depth_func == PIPE_FUNC_ALWAYS || state->depth_func == PIPE_FUNC_NEVER);
1600 
1601    return dsa;
1602 }
1603 
si_pm4_emit_dsa(struct si_context * sctx,unsigned index)1604 static void si_pm4_emit_dsa(struct si_context *sctx, unsigned index)
1605 {
1606    struct si_state_dsa *state = sctx->queued.named.dsa;
1607    assert(state && state != sctx->emitted.named.dsa);
1608 
1609    if (sctx->gfx_level >= GFX12) {
1610       radeon_begin(&sctx->gfx_cs);
1611       gfx12_begin_context_regs();
1612       gfx12_opt_set_context_reg(R_028070_DB_DEPTH_CONTROL, SI_TRACKED_DB_DEPTH_CONTROL,
1613                                 state->db_depth_control);
1614       if (state->stencil_enabled) {
1615          gfx12_opt_set_context_reg(R_028074_DB_STENCIL_CONTROL, SI_TRACKED_DB_STENCIL_CONTROL,
1616                                    state->db_stencil_control);
1617          gfx12_opt_set_context_reg(R_028090_DB_STENCIL_READ_MASK, SI_TRACKED_DB_STENCIL_READ_MASK,
1618                                    state->db_stencil_read_mask);
1619          gfx12_opt_set_context_reg(R_028094_DB_STENCIL_WRITE_MASK, SI_TRACKED_DB_STENCIL_WRITE_MASK,
1620                                    state->db_stencil_write_mask);
1621       }
1622       if (state->depth_bounds_enabled) {
1623          gfx12_opt_set_context_reg(R_028050_DB_DEPTH_BOUNDS_MIN, SI_TRACKED_DB_DEPTH_BOUNDS_MIN,
1624                                    state->db_depth_bounds_min);
1625          gfx12_opt_set_context_reg(R_028054_DB_DEPTH_BOUNDS_MAX, SI_TRACKED_DB_DEPTH_BOUNDS_MAX,
1626                                    state->db_depth_bounds_max);
1627       }
1628       gfx12_end_context_regs();
1629       radeon_end(); /* don't track context rolls on GFX12 */
1630 
1631       gfx12_opt_push_gfx_sh_reg(R_00B030_SPI_SHADER_USER_DATA_PS_0 + SI_SGPR_ALPHA_REF * 4,
1632                                 SI_TRACKED_SPI_SHADER_USER_DATA_PS__ALPHA_REF,
1633                                 state->spi_shader_user_data_ps_alpha_ref);
1634    } else if (sctx->screen->info.has_set_context_pairs_packed) {
1635       radeon_begin(&sctx->gfx_cs);
1636       gfx11_begin_packed_context_regs();
1637       gfx11_opt_set_context_reg(R_028800_DB_DEPTH_CONTROL, SI_TRACKED_DB_DEPTH_CONTROL,
1638                                 state->db_depth_control);
1639       if (state->stencil_enabled) {
1640          gfx11_opt_set_context_reg(R_02842C_DB_STENCIL_CONTROL, SI_TRACKED_DB_STENCIL_CONTROL,
1641                                    state->db_stencil_control);
1642       }
1643       if (state->depth_bounds_enabled) {
1644          gfx11_opt_set_context_reg(R_028020_DB_DEPTH_BOUNDS_MIN, SI_TRACKED_DB_DEPTH_BOUNDS_MIN,
1645                                    state->db_depth_bounds_min);
1646          gfx11_opt_set_context_reg(R_028024_DB_DEPTH_BOUNDS_MAX, SI_TRACKED_DB_DEPTH_BOUNDS_MAX,
1647                                    state->db_depth_bounds_max);
1648       }
1649       gfx11_end_packed_context_regs();
1650 
1651       if (state->alpha_func != PIPE_FUNC_ALWAYS) {
1652          if (sctx->screen->info.has_set_sh_pairs_packed) {
1653             gfx11_opt_push_gfx_sh_reg(R_00B030_SPI_SHADER_USER_DATA_PS_0 + SI_SGPR_ALPHA_REF * 4,
1654                                       SI_TRACKED_SPI_SHADER_USER_DATA_PS__ALPHA_REF,
1655                                       state->spi_shader_user_data_ps_alpha_ref);
1656          } else {
1657             radeon_opt_set_sh_reg(R_00B030_SPI_SHADER_USER_DATA_PS_0 + SI_SGPR_ALPHA_REF * 4,
1658                                   SI_TRACKED_SPI_SHADER_USER_DATA_PS__ALPHA_REF,
1659                                   state->spi_shader_user_data_ps_alpha_ref);
1660          }
1661       }
1662       radeon_end(); /* don't track context rolls on GFX11 */
1663    } else {
1664       radeon_begin(&sctx->gfx_cs);
1665       radeon_opt_set_context_reg(R_028800_DB_DEPTH_CONTROL, SI_TRACKED_DB_DEPTH_CONTROL,
1666                                  state->db_depth_control);
1667       if (state->stencil_enabled) {
1668          radeon_opt_set_context_reg(R_02842C_DB_STENCIL_CONTROL, SI_TRACKED_DB_STENCIL_CONTROL,
1669                                     state->db_stencil_control);
1670       }
1671       if (state->depth_bounds_enabled) {
1672          radeon_opt_set_context_reg2(R_028020_DB_DEPTH_BOUNDS_MIN,
1673                                      SI_TRACKED_DB_DEPTH_BOUNDS_MIN,
1674                                      state->db_depth_bounds_min,
1675                                      state->db_depth_bounds_max);
1676       }
1677       radeon_end_update_context_roll();
1678 
1679       if (state->alpha_func != PIPE_FUNC_ALWAYS) {
1680          radeon_begin(&sctx->gfx_cs);
1681          radeon_opt_set_sh_reg(R_00B030_SPI_SHADER_USER_DATA_PS_0 + SI_SGPR_ALPHA_REF * 4,
1682                                SI_TRACKED_SPI_SHADER_USER_DATA_PS__ALPHA_REF,
1683                                state->spi_shader_user_data_ps_alpha_ref);
1684          radeon_end();
1685       }
1686    }
1687 
1688    sctx->emitted.named.dsa = state;
1689 }
1690 
si_bind_dsa_state(struct pipe_context * ctx,void * state)1691 static void si_bind_dsa_state(struct pipe_context *ctx, void *state)
1692 {
1693    struct si_context *sctx = (struct si_context *)ctx;
1694    struct si_state_dsa *old_dsa = sctx->queued.named.dsa;
1695    struct si_state_dsa *dsa = state;
1696 
1697    if (!dsa)
1698       dsa = (struct si_state_dsa *)sctx->noop_dsa;
1699 
1700    si_pm4_bind_state(sctx, dsa, dsa);
1701 
1702    /* Gfx12 doesn't need to combine a DSA state with a stencil ref state. */
1703    if (sctx->gfx_level < GFX12 &&
1704        memcmp(&dsa->stencil_ref, &sctx->stencil_ref.dsa_part,
1705               sizeof(struct si_dsa_stencil_ref_part)) != 0) {
1706       sctx->stencil_ref.dsa_part = dsa->stencil_ref;
1707       si_mark_atom_dirty(sctx, &sctx->atoms.s.stencil_ref);
1708    }
1709 
1710    if (old_dsa->alpha_func != dsa->alpha_func) {
1711       si_ps_key_update_dsa(sctx);
1712       si_update_ps_inputs_read_or_disabled(sctx);
1713       sctx->do_update_shaders = true;
1714    }
1715 
1716    if (sctx->occlusion_query_mode == SI_OCCLUSION_QUERY_MODE_PRECISE_BOOLEAN &&
1717        (old_dsa->depth_enabled != dsa->depth_enabled ||
1718         old_dsa->depth_write_enabled != dsa->depth_write_enabled))
1719       si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1720 
1721    if (sctx->screen->dpbb_allowed && ((old_dsa->depth_enabled != dsa->depth_enabled ||
1722                                        old_dsa->stencil_enabled != dsa->stencil_enabled ||
1723                                        old_dsa->db_can_write != dsa->db_can_write)))
1724       si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
1725 
1726    if (sctx->screen->info.has_out_of_order_rast &&
1727        (memcmp(old_dsa->order_invariance, dsa->order_invariance,
1728                sizeof(old_dsa->order_invariance))))
1729       si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
1730 }
1731 
si_delete_dsa_state(struct pipe_context * ctx,void * state)1732 static void si_delete_dsa_state(struct pipe_context *ctx, void *state)
1733 {
1734    struct si_context *sctx = (struct si_context *)ctx;
1735 
1736    if (sctx->queued.named.dsa == state)
1737       si_bind_dsa_state(ctx, sctx->noop_dsa);
1738 
1739    si_pm4_free_state(sctx, (struct si_pm4_state*)state, SI_STATE_IDX(dsa));
1740 }
1741 
si_create_db_flush_dsa(struct si_context * sctx)1742 static void *si_create_db_flush_dsa(struct si_context *sctx)
1743 {
1744    struct pipe_depth_stencil_alpha_state dsa = {};
1745 
1746    return sctx->b.create_depth_stencil_alpha_state(&sctx->b, &dsa);
1747 }
1748 
1749 /* DB RENDER STATE */
1750 
si_set_active_query_state(struct pipe_context * ctx,bool enable)1751 static void si_set_active_query_state(struct pipe_context *ctx, bool enable)
1752 {
1753    struct si_context *sctx = (struct si_context *)ctx;
1754 
1755    /* Pipeline stat & streamout queries. */
1756    if (enable) {
1757       /* Disable pipeline stats if there are no active queries. */
1758       if (sctx->num_hw_pipestat_streamout_queries) {
1759          sctx->barrier_flags &= ~SI_BARRIER_EVENT_PIPELINESTAT_STOP;
1760          sctx->barrier_flags |= SI_BARRIER_EVENT_PIPELINESTAT_START;
1761          si_mark_atom_dirty(sctx, &sctx->atoms.s.barrier);
1762       }
1763    } else {
1764       if (sctx->num_hw_pipestat_streamout_queries) {
1765          sctx->barrier_flags &= ~SI_BARRIER_EVENT_PIPELINESTAT_START;
1766          sctx->barrier_flags |= SI_BARRIER_EVENT_PIPELINESTAT_STOP;
1767          si_mark_atom_dirty(sctx, &sctx->atoms.s.barrier);
1768       }
1769    }
1770 
1771    /* Occlusion queries. */
1772    if (sctx->occlusion_queries_disabled != !enable) {
1773       sctx->occlusion_queries_disabled = !enable;
1774       si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1775    }
1776 }
1777 
si_save_qbo_state(struct si_context * sctx,struct si_qbo_state * st)1778 void si_save_qbo_state(struct si_context *sctx, struct si_qbo_state *st)
1779 {
1780    si_get_pipe_constant_buffer(sctx, PIPE_SHADER_COMPUTE, 0, &st->saved_const0);
1781 }
1782 
si_restore_qbo_state(struct si_context * sctx,struct si_qbo_state * st)1783 void si_restore_qbo_state(struct si_context *sctx, struct si_qbo_state *st)
1784 {
1785    sctx->b.set_constant_buffer(&sctx->b, PIPE_SHADER_COMPUTE, 0, true, &st->saved_const0);
1786 }
1787 
si_emit_db_render_state(struct si_context * sctx,unsigned index)1788 static void si_emit_db_render_state(struct si_context *sctx, unsigned index)
1789 {
1790    unsigned db_shader_control = 0, db_render_control = 0, db_count_control = 0, vrs_override_cntl = 0;
1791 
1792    /* DB_RENDER_CONTROL */
1793    /* Program OREO_MODE optimally for GFX11+. */
1794    if (sctx->gfx_level >= GFX11) {
1795       bool z_export = G_02880C_Z_EXPORT_ENABLE(sctx->ps_db_shader_control);
1796       db_render_control |= S_028000_OREO_MODE(z_export ? V_028000_OMODE_BLEND : V_028000_OMODE_O_THEN_B);
1797    }
1798 
1799    if (sctx->gfx_level >= GFX12) {
1800       assert(!sctx->dbcb_depth_copy_enabled && !sctx->dbcb_stencil_copy_enabled);
1801       assert(!sctx->db_flush_depth_inplace && !sctx->db_flush_stencil_inplace);
1802       assert(!sctx->db_depth_clear && !sctx->db_stencil_clear);
1803    } else {
1804       if (sctx->dbcb_depth_copy_enabled || sctx->dbcb_stencil_copy_enabled) {
1805          db_render_control |= S_028000_DEPTH_COPY(sctx->dbcb_depth_copy_enabled) |
1806                               S_028000_STENCIL_COPY(sctx->dbcb_stencil_copy_enabled) |
1807                               S_028000_COPY_CENTROID(1) | S_028000_COPY_SAMPLE(sctx->dbcb_copy_sample);
1808       } else if (sctx->db_flush_depth_inplace || sctx->db_flush_stencil_inplace) {
1809          db_render_control |= S_028000_DEPTH_COMPRESS_DISABLE(sctx->db_flush_depth_inplace) |
1810                               S_028000_STENCIL_COMPRESS_DISABLE(sctx->db_flush_stencil_inplace);
1811       } else {
1812          db_render_control |= S_028000_DEPTH_CLEAR_ENABLE(sctx->db_depth_clear) |
1813                               S_028000_STENCIL_CLEAR_ENABLE(sctx->db_stencil_clear);
1814       }
1815 
1816       if (sctx->gfx_level >= GFX11) {
1817          unsigned max_allowed_tiles_in_wave;
1818 
1819          if (sctx->screen->info.has_dedicated_vram) {
1820             if (sctx->framebuffer.nr_samples == 8)
1821                max_allowed_tiles_in_wave = 6;
1822             else if (sctx->framebuffer.nr_samples == 4)
1823                max_allowed_tiles_in_wave = 13;
1824             else
1825                max_allowed_tiles_in_wave = 0;
1826          } else {
1827             if (sctx->framebuffer.nr_samples == 8)
1828                max_allowed_tiles_in_wave = 7;
1829             else if (sctx->framebuffer.nr_samples == 4)
1830                max_allowed_tiles_in_wave = 15;
1831             else
1832                max_allowed_tiles_in_wave = 0;
1833          }
1834 
1835          db_render_control |= S_028000_MAX_ALLOWED_TILES_IN_WAVE(max_allowed_tiles_in_wave);
1836       }
1837    }
1838 
1839    /* DB_COUNT_CONTROL (occlusion queries) */
1840    if (sctx->occlusion_query_mode == SI_OCCLUSION_QUERY_MODE_DISABLE ||
1841        sctx->occlusion_queries_disabled) {
1842       /* Occlusion queries disabled. */
1843       if (sctx->gfx_level >= GFX7)
1844          db_count_control |= S_028004_ZPASS_ENABLE(0);
1845       else
1846          db_count_control |= S_028004_ZPASS_INCREMENT_DISABLE(1);
1847    } else {
1848       /* Occlusion queries enabled. */
1849       if (sctx->gfx_level < GFX12)
1850          db_count_control |= S_028004_SAMPLE_RATE(sctx->framebuffer.log_samples);
1851 
1852       if (sctx->gfx_level >= GFX7) {
1853          db_count_control |= S_028004_ZPASS_ENABLE(1) |
1854                              S_028004_SLICE_EVEN_ENABLE(1) |
1855                              S_028004_SLICE_ODD_ENABLE(1);
1856       }
1857 
1858       if (sctx->occlusion_query_mode == SI_OCCLUSION_QUERY_MODE_PRECISE_INTEGER ||
1859           /* Boolean occlusion queries must set PERFECT_ZPASS_COUNTS for depth-only rendering
1860            * without depth writes or when depth testing is disabled. */
1861           (sctx->occlusion_query_mode == SI_OCCLUSION_QUERY_MODE_PRECISE_BOOLEAN &&
1862            (!sctx->queued.named.dsa->depth_enabled ||
1863             (!sctx->queued.named.blend->cb_target_mask &&
1864              !sctx->queued.named.dsa->depth_write_enabled))))
1865          db_count_control |= S_028004_PERFECT_ZPASS_COUNTS(1);
1866 
1867       if (sctx->gfx_level >= GFX10 &&
1868           sctx->occlusion_query_mode != SI_OCCLUSION_QUERY_MODE_CONSERVATIVE_BOOLEAN)
1869          db_count_control |= S_028004_DISABLE_CONSERVATIVE_ZPASS_COUNTS(1);
1870    }
1871 
1872    /* This should always be set on GFX11. */
1873    if (sctx->gfx_level >= GFX11)
1874       db_count_control |= S_028004_DISABLE_CONSERVATIVE_ZPASS_COUNTS(1);
1875 
1876    db_shader_control |= sctx->ps_db_shader_control;
1877 
1878    if (sctx->screen->info.has_export_conflict_bug &&
1879        sctx->queued.named.blend->blend_enable_4bit &&
1880        si_get_num_coverage_samples(sctx) == 1) {
1881       db_shader_control |= S_02880C_OVERRIDE_INTRINSIC_RATE_ENABLE(1) |
1882                            S_02880C_OVERRIDE_INTRINSIC_RATE(2);
1883    }
1884 
1885    if (sctx->gfx_level >= GFX10_3) {
1886       /* Variable rate shading. */
1887       unsigned mode, log_rate_x, log_rate_y;
1888 
1889       if (sctx->allow_flat_shading) {
1890          mode = V_028064_SC_VRS_COMB_MODE_OVERRIDE;
1891          log_rate_x = log_rate_y = 1; /* 2x2 VRS (log2(2) == 1) */
1892       } else {
1893          /* If the shader is using discard, turn off coarse shading because discarding at 2x2 pixel
1894           * granularity degrades quality too much.
1895           *
1896           * The shader writes the VRS rate and we either pass it through or do MIN(shader, 1x1)
1897           * to disable coarse shading.
1898           */
1899          mode = sctx->screen->options.vrs2x2 && G_02880C_KILL_ENABLE(db_shader_control) ?
1900                    V_028064_SC_VRS_COMB_MODE_MIN : V_028064_SC_VRS_COMB_MODE_PASSTHRU;
1901          log_rate_x = log_rate_y = 0; /* 1x1 VRS (log2(1) == 0) */
1902       }
1903 
1904       if (sctx->gfx_level >= GFX11) {
1905          vrs_override_cntl = S_0283D0_VRS_OVERRIDE_RATE_COMBINER_MODE(mode) |
1906                              S_0283D0_VRS_RATE(log_rate_x * 4 + log_rate_y);
1907       } else {
1908          vrs_override_cntl = S_028064_VRS_OVERRIDE_RATE_COMBINER_MODE(mode) |
1909                              S_028064_VRS_OVERRIDE_RATE_X(log_rate_x) |
1910                              S_028064_VRS_OVERRIDE_RATE_Y(log_rate_y);
1911       }
1912    }
1913 
1914    unsigned db_render_override2 =
1915          S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(sctx->db_depth_disable_expclear) |
1916          S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(sctx->db_stencil_disable_expclear) |
1917          S_028010_DECOMPRESS_Z_ON_FLUSH(sctx->framebuffer.nr_samples >= 4) |
1918          S_028010_CENTROID_COMPUTATION_MODE(sctx->gfx_level >= GFX10_3 ? 1 : 0);
1919 
1920    if (sctx->gfx_level >= GFX12) {
1921       radeon_begin(&sctx->gfx_cs);
1922       gfx12_begin_context_regs();
1923       gfx12_opt_set_context_reg(R_028000_DB_RENDER_CONTROL, SI_TRACKED_DB_RENDER_CONTROL,
1924                                 db_render_control);
1925       gfx12_opt_set_context_reg(R_028010_DB_RENDER_OVERRIDE2, SI_TRACKED_DB_RENDER_OVERRIDE2,
1926                                 S_028010_DECOMPRESS_Z_ON_FLUSH(sctx->framebuffer.nr_samples >= 4) |
1927                                 S_028010_CENTROID_COMPUTATION_MODE(1));
1928       gfx12_opt_set_context_reg(R_028060_DB_COUNT_CONTROL, SI_TRACKED_DB_COUNT_CONTROL,
1929                                 db_count_control);
1930       gfx12_opt_set_context_reg(R_02806C_DB_SHADER_CONTROL, SI_TRACKED_DB_SHADER_CONTROL,
1931                                 db_shader_control);
1932       gfx12_opt_set_context_reg(R_0283D0_PA_SC_VRS_OVERRIDE_CNTL,
1933                                 SI_TRACKED_DB_PA_SC_VRS_OVERRIDE_CNTL, vrs_override_cntl);
1934       gfx12_end_context_regs();
1935       radeon_end(); /* don't track context rolls on GFX12 */
1936    } else if (sctx->screen->info.has_set_context_pairs_packed) {
1937       radeon_begin(&sctx->gfx_cs);
1938       gfx11_begin_packed_context_regs();
1939       gfx11_opt_set_context_reg(R_028000_DB_RENDER_CONTROL, SI_TRACKED_DB_RENDER_CONTROL,
1940                                 db_render_control);
1941       gfx11_opt_set_context_reg(R_028004_DB_COUNT_CONTROL, SI_TRACKED_DB_COUNT_CONTROL,
1942                                 db_count_control);
1943       gfx11_opt_set_context_reg(R_028010_DB_RENDER_OVERRIDE2, SI_TRACKED_DB_RENDER_OVERRIDE2,
1944                                 db_render_override2);
1945       gfx11_opt_set_context_reg(R_02880C_DB_SHADER_CONTROL, SI_TRACKED_DB_SHADER_CONTROL,
1946                                 db_shader_control);
1947       gfx11_opt_set_context_reg(R_0283D0_PA_SC_VRS_OVERRIDE_CNTL,
1948                                 SI_TRACKED_DB_PA_SC_VRS_OVERRIDE_CNTL, vrs_override_cntl);
1949       gfx11_end_packed_context_regs();
1950       radeon_end(); /* don't track context rolls on GFX11 */
1951    } else {
1952       radeon_begin(&sctx->gfx_cs);
1953       radeon_opt_set_context_reg2(R_028000_DB_RENDER_CONTROL, SI_TRACKED_DB_RENDER_CONTROL,
1954                                   db_render_control, db_count_control);
1955       radeon_opt_set_context_reg(R_028010_DB_RENDER_OVERRIDE2,
1956                                  SI_TRACKED_DB_RENDER_OVERRIDE2, db_render_override2);
1957       radeon_opt_set_context_reg(R_02880C_DB_SHADER_CONTROL, SI_TRACKED_DB_SHADER_CONTROL,
1958                                  db_shader_control);
1959 
1960       if (sctx->gfx_level >= GFX11) {
1961          radeon_opt_set_context_reg(R_0283D0_PA_SC_VRS_OVERRIDE_CNTL,
1962                                     SI_TRACKED_DB_PA_SC_VRS_OVERRIDE_CNTL, vrs_override_cntl);
1963       } else if (sctx->gfx_level >= GFX10_3) {
1964          radeon_opt_set_context_reg(R_028064_DB_VRS_OVERRIDE_CNTL,
1965                                     SI_TRACKED_DB_PA_SC_VRS_OVERRIDE_CNTL, vrs_override_cntl);
1966       }
1967       radeon_end_update_context_roll();
1968    }
1969 }
1970 
1971 /*
1972  * Texture translation
1973  */
1974 
si_translate_texformat(struct pipe_screen * screen,enum pipe_format format,const struct util_format_description * desc,int first_non_void)1975 static uint32_t si_translate_texformat(struct pipe_screen *screen, enum pipe_format format,
1976                                        const struct util_format_description *desc,
1977                                        int first_non_void)
1978 {
1979    struct si_screen *sscreen = (struct si_screen *)screen;
1980 
1981    assert(sscreen->info.gfx_level <= GFX9);
1982 
1983    return ac_translate_tex_dataformat(&sscreen->info, desc, first_non_void);
1984 }
1985 
is_wrap_mode_legal(struct si_screen * screen,unsigned wrap)1986 static unsigned is_wrap_mode_legal(struct si_screen *screen, unsigned wrap)
1987 {
1988    if (!screen->info.has_3d_cube_border_color_mipmap) {
1989       switch (wrap) {
1990       case PIPE_TEX_WRAP_CLAMP:
1991       case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
1992       case PIPE_TEX_WRAP_MIRROR_CLAMP:
1993       case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
1994          return false;
1995       }
1996    }
1997    return true;
1998 }
1999 
si_tex_wrap(unsigned wrap)2000 static unsigned si_tex_wrap(unsigned wrap)
2001 {
2002    switch (wrap) {
2003    default:
2004    case PIPE_TEX_WRAP_REPEAT:
2005       return V_008F30_SQ_TEX_WRAP;
2006    case PIPE_TEX_WRAP_CLAMP:
2007       return V_008F30_SQ_TEX_CLAMP_HALF_BORDER;
2008    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
2009       return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL;
2010    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
2011       return V_008F30_SQ_TEX_CLAMP_BORDER;
2012    case PIPE_TEX_WRAP_MIRROR_REPEAT:
2013       return V_008F30_SQ_TEX_MIRROR;
2014    case PIPE_TEX_WRAP_MIRROR_CLAMP:
2015       return V_008F30_SQ_TEX_MIRROR_ONCE_HALF_BORDER;
2016    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
2017       return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL;
2018    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
2019       return V_008F30_SQ_TEX_MIRROR_ONCE_BORDER;
2020    }
2021 }
2022 
si_tex_mipfilter(unsigned filter)2023 static unsigned si_tex_mipfilter(unsigned filter)
2024 {
2025    switch (filter) {
2026    case PIPE_TEX_MIPFILTER_NEAREST:
2027       return V_008F38_SQ_TEX_Z_FILTER_POINT;
2028    case PIPE_TEX_MIPFILTER_LINEAR:
2029       return V_008F38_SQ_TEX_Z_FILTER_LINEAR;
2030    default:
2031    case PIPE_TEX_MIPFILTER_NONE:
2032       return V_008F38_SQ_TEX_Z_FILTER_NONE;
2033    }
2034 }
2035 
si_tex_compare(unsigned mode,unsigned compare)2036 static unsigned si_tex_compare(unsigned mode, unsigned compare)
2037 {
2038    if (mode == PIPE_TEX_COMPARE_NONE)
2039       return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER;
2040 
2041    switch (compare) {
2042    default:
2043    case PIPE_FUNC_NEVER:
2044       return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER;
2045    case PIPE_FUNC_LESS:
2046       return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS;
2047    case PIPE_FUNC_EQUAL:
2048       return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL;
2049    case PIPE_FUNC_LEQUAL:
2050       return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL;
2051    case PIPE_FUNC_GREATER:
2052       return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER;
2053    case PIPE_FUNC_NOTEQUAL:
2054       return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL;
2055    case PIPE_FUNC_GEQUAL:
2056       return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL;
2057    case PIPE_FUNC_ALWAYS:
2058       return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS;
2059    }
2060 }
2061 
si_tex_dim(struct si_screen * sscreen,struct si_texture * tex,unsigned view_target,unsigned nr_samples)2062 static unsigned si_tex_dim(struct si_screen *sscreen, struct si_texture *tex, unsigned view_target,
2063                            unsigned nr_samples)
2064 {
2065    unsigned res_target = tex->buffer.b.b.target;
2066 
2067    if (view_target == PIPE_TEXTURE_CUBE || view_target == PIPE_TEXTURE_CUBE_ARRAY)
2068       res_target = view_target;
2069    /* If interpreting cubemaps as something else, set 2D_ARRAY. */
2070    else if (res_target == PIPE_TEXTURE_CUBE || res_target == PIPE_TEXTURE_CUBE_ARRAY)
2071       res_target = PIPE_TEXTURE_2D_ARRAY;
2072 
2073    /* GFX9 allocates 1D textures as 2D. */
2074    if ((res_target == PIPE_TEXTURE_1D || res_target == PIPE_TEXTURE_1D_ARRAY) &&
2075        sscreen->info.gfx_level == GFX9 &&
2076        tex->surface.u.gfx9.resource_type == RADEON_RESOURCE_2D) {
2077       if (res_target == PIPE_TEXTURE_1D)
2078          res_target = PIPE_TEXTURE_2D;
2079       else
2080          res_target = PIPE_TEXTURE_2D_ARRAY;
2081    }
2082 
2083    switch (res_target) {
2084    default:
2085    case PIPE_TEXTURE_1D:
2086       return V_008F1C_SQ_RSRC_IMG_1D;
2087    case PIPE_TEXTURE_1D_ARRAY:
2088       return V_008F1C_SQ_RSRC_IMG_1D_ARRAY;
2089    case PIPE_TEXTURE_2D:
2090    case PIPE_TEXTURE_RECT:
2091       return nr_samples > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA : V_008F1C_SQ_RSRC_IMG_2D;
2092    case PIPE_TEXTURE_2D_ARRAY:
2093       return nr_samples > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY : V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
2094    case PIPE_TEXTURE_3D:
2095       return V_008F1C_SQ_RSRC_IMG_3D;
2096    case PIPE_TEXTURE_CUBE:
2097    case PIPE_TEXTURE_CUBE_ARRAY:
2098       return V_008F1C_SQ_RSRC_IMG_CUBE;
2099    }
2100 }
2101 
2102 /*
2103  * Format support testing
2104  */
2105 
si_is_sampler_format_supported(struct pipe_screen * screen,enum pipe_format format)2106 static bool si_is_sampler_format_supported(struct pipe_screen *screen, enum pipe_format format)
2107 {
2108    struct si_screen *sscreen = (struct si_screen *)screen;
2109    const struct util_format_description *desc = util_format_description(format);
2110 
2111    /* Samplers don't support 64 bits per channel. */
2112    if (desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
2113        desc->channel[0].size == 64)
2114       return false;
2115 
2116    if (sscreen->info.gfx_level >= GFX10) {
2117       const struct gfx10_format *fmt = &ac_get_gfx10_format_table(sscreen->info.gfx_level)[format];
2118       if (!fmt->img_format || fmt->buffers_only)
2119          return false;
2120       return true;
2121    }
2122 
2123    const int first_non_void =  util_format_get_first_non_void_channel(format);
2124 
2125    if (si_translate_texformat(screen, format, desc, first_non_void) == ~0U)
2126       return false;
2127 
2128    if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB &&
2129        desc->nr_channels != 4 && desc->nr_channels != 1)
2130       return false;
2131 
2132    if (desc->layout == UTIL_FORMAT_LAYOUT_ETC && !sscreen->info.has_etc_support)
2133       return false;
2134 
2135    if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED &&
2136        (desc->format == PIPE_FORMAT_G8B8_G8R8_UNORM ||
2137         desc->format == PIPE_FORMAT_B8G8_R8G8_UNORM))
2138       return false;
2139 
2140    /* Other "OTHER" layouts are unsupported. */
2141    if (desc->layout == UTIL_FORMAT_LAYOUT_OTHER &&
2142        desc->format != PIPE_FORMAT_R11G11B10_FLOAT &&
2143        desc->format != PIPE_FORMAT_R9G9B9E5_FLOAT)
2144       return false;
2145 
2146    /* This must be before using first_non_void. */
2147    if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
2148       return true;
2149 
2150    if (first_non_void < 0 || first_non_void > 3)
2151       return false;
2152 
2153    /* Reject SCALED formats because we don't implement them for CB and do the same for texturing. */
2154    if ((desc->channel[first_non_void].type == UTIL_FORMAT_TYPE_UNSIGNED ||
2155         desc->channel[first_non_void].type == UTIL_FORMAT_TYPE_SIGNED) &&
2156        !desc->channel[first_non_void].normalized &&
2157        !desc->channel[first_non_void].pure_integer)
2158       return false;
2159 
2160    /* Reject unsupported 32_*NORM and FIXED formats. */
2161    if (desc->channel[first_non_void].size == 32 &&
2162        (desc->channel[first_non_void].normalized ||
2163         desc->channel[first_non_void].type == UTIL_FORMAT_TYPE_FIXED))
2164       return false;
2165 
2166    /* This format fails on Gfx8/Carrizo´. */
2167    if (sscreen->info.family == CHIP_CARRIZO && format == PIPE_FORMAT_A8R8_UNORM)
2168       return false;
2169 
2170    /* Reject unsupported 3x 32-bit formats for CB. */
2171    if (desc->nr_channels == 3 && desc->channel[0].size == 32 && desc->channel[1].size == 32 &&
2172        desc->channel[2].size == 32)
2173       return false;
2174 
2175    /* Reject all 64-bit formats. */
2176    if (desc->channel[first_non_void].size == 64)
2177       return false;
2178 
2179    return true;
2180 }
2181 
si_translate_buffer_dataformat(struct pipe_screen * screen,const struct util_format_description * desc,int first_non_void)2182 static uint32_t si_translate_buffer_dataformat(struct pipe_screen *screen,
2183                                                const struct util_format_description *desc,
2184                                                int first_non_void)
2185 {
2186    assert(((struct si_screen *)screen)->info.gfx_level <= GFX9);
2187 
2188    return ac_translate_buffer_dataformat(desc, first_non_void);
2189 }
2190 
si_translate_buffer_numformat(struct pipe_screen * screen,const struct util_format_description * desc,int first_non_void)2191 static uint32_t si_translate_buffer_numformat(struct pipe_screen *screen,
2192                                               const struct util_format_description *desc,
2193                                               int first_non_void)
2194 {
2195    assert(((struct si_screen *)screen)->info.gfx_level <= GFX9);
2196 
2197    return ac_translate_buffer_numformat(desc, first_non_void);
2198 }
2199 
si_is_vertex_format_supported(struct pipe_screen * screen,enum pipe_format format,unsigned usage)2200 static unsigned si_is_vertex_format_supported(struct pipe_screen *screen, enum pipe_format format,
2201                                               unsigned usage)
2202 {
2203    struct si_screen *sscreen = (struct si_screen *)screen;
2204    const struct util_format_description *desc;
2205    int first_non_void;
2206    unsigned data_format;
2207 
2208    assert((usage & ~(PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_VERTEX_BUFFER)) ==
2209           0);
2210 
2211    desc = util_format_description(format);
2212 
2213    /* There are no native 8_8_8 or 16_16_16 data formats, and we currently
2214     * select 8_8_8_8 and 16_16_16_16 instead. This works reasonably well
2215     * for read-only access (with caveats surrounding bounds checks), but
2216     * obviously fails for write access which we have to implement for
2217     * shader images. Luckily, OpenGL doesn't expect this to be supported
2218     * anyway, and so the only impact is on PBO uploads / downloads, which
2219     * shouldn't be expected to be fast for GL_RGB anyway.
2220     */
2221    if (desc->block.bits == 3 * 8 || desc->block.bits == 3 * 16) {
2222       if (usage & (PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SAMPLER_VIEW)) {
2223          usage &= ~(PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SAMPLER_VIEW);
2224          if (!usage)
2225             return 0;
2226       }
2227    }
2228 
2229    if (sscreen->info.gfx_level >= GFX10) {
2230       const struct gfx10_format *fmt = &ac_get_gfx10_format_table(sscreen->info.gfx_level)[format];
2231       unsigned first_image_only_format = sscreen->info.gfx_level >= GFX11 ? 64 : 128;
2232 
2233       if (!fmt->img_format || fmt->img_format >= first_image_only_format)
2234          return 0;
2235       return usage;
2236    }
2237 
2238    first_non_void = util_format_get_first_non_void_channel(format);
2239    data_format = si_translate_buffer_dataformat(screen, desc, first_non_void);
2240    if (data_format == V_008F0C_BUF_DATA_FORMAT_INVALID)
2241       return 0;
2242 
2243    return usage;
2244 }
2245 
si_is_zs_format_supported(enum pipe_format format)2246 static bool si_is_zs_format_supported(enum pipe_format format)
2247 {
2248    if (format == PIPE_FORMAT_Z16_UNORM_S8_UINT)
2249       return false;
2250 
2251    return ac_is_zs_format_supported(format);
2252 }
2253 
si_is_reduction_mode_supported(struct pipe_screen * screen,enum pipe_format format)2254 static bool si_is_reduction_mode_supported(struct pipe_screen *screen, enum pipe_format format)
2255 {
2256    struct si_screen *sscreen = (struct si_screen *)screen;
2257 
2258    return ac_is_reduction_mode_supported(&sscreen->info, format, true);
2259 }
2260 
si_is_format_supported(struct pipe_screen * screen,enum pipe_format format,enum pipe_texture_target target,unsigned sample_count,unsigned storage_sample_count,unsigned usage)2261 static bool si_is_format_supported(struct pipe_screen *screen, enum pipe_format format,
2262                                    enum pipe_texture_target target, unsigned sample_count,
2263                                    unsigned storage_sample_count, unsigned usage)
2264 {
2265    struct si_screen *sscreen = (struct si_screen *)screen;
2266    unsigned retval = 0;
2267 
2268    if (target >= PIPE_MAX_TEXTURE_TYPES) {
2269       PRINT_ERR("radeonsi: unsupported texture type %d\n", target);
2270       return false;
2271    }
2272 
2273    /* Require PIPE_BIND_SAMPLER_VIEW support when PIPE_BIND_RENDER_TARGET
2274     * is requested.
2275     */
2276    if (usage & PIPE_BIND_RENDER_TARGET)
2277       usage |= PIPE_BIND_SAMPLER_VIEW;
2278 
2279    if ((target == PIPE_TEXTURE_3D || target == PIPE_TEXTURE_CUBE) &&
2280         !sscreen->info.has_3d_cube_border_color_mipmap)
2281       return false;
2282 
2283    if (util_format_get_num_planes(format) >= 2)
2284       return false;
2285 
2286    if (MAX2(1, sample_count) < MAX2(1, storage_sample_count))
2287       return false;
2288 
2289    if (sample_count > 1) {
2290       if (!screen->get_param(screen, PIPE_CAP_TEXTURE_MULTISAMPLE))
2291          return false;
2292 
2293       /* Only power-of-two sample counts are supported. */
2294       if (!util_is_power_of_two_or_zero(sample_count) ||
2295           !util_is_power_of_two_or_zero(storage_sample_count))
2296          return false;
2297 
2298       /* Chips with 1 RB don't increment occlusion queries at 16x MSAA sample rate,
2299        * so don't expose 16 samples there.
2300        *
2301        * EQAA also uses max 8 samples because our FMASK fetches only load 32 bits and
2302        * would need to be changed to 64 bits for 16 samples.
2303        */
2304       const unsigned max_samples = 8;
2305 
2306       /* MSAA support without framebuffer attachments. */
2307       if (format == PIPE_FORMAT_NONE && sample_count <= max_samples)
2308          return true;
2309 
2310       if (!sscreen->info.has_eqaa_surface_allocator || util_format_is_depth_or_stencil(format)) {
2311          /* Color without EQAA or depth/stencil. */
2312          if (sample_count > max_samples || sample_count != storage_sample_count)
2313             return false;
2314       } else {
2315          /* Color with EQAA. */
2316          if (sample_count > max_samples || storage_sample_count > max_samples)
2317             return false;
2318       }
2319    }
2320 
2321    if (usage & (PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_SHADER_IMAGE)) {
2322       if (target == PIPE_BUFFER) {
2323          retval |= si_is_vertex_format_supported(
2324             screen, format, usage & (PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_SHADER_IMAGE));
2325       } else {
2326          if (si_is_sampler_format_supported(screen, format))
2327             retval |= usage & (PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_SHADER_IMAGE);
2328       }
2329    }
2330 
2331    if ((usage & (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT |
2332                  PIPE_BIND_SHARED | PIPE_BIND_BLENDABLE)) &&
2333        ac_is_colorbuffer_format_supported(sscreen->info.gfx_level, format)) {
2334       retval |= usage & (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT |
2335                          PIPE_BIND_SHARED);
2336       if (!util_format_is_pure_integer(format) && !util_format_is_depth_or_stencil(format))
2337          retval |= usage & PIPE_BIND_BLENDABLE;
2338    }
2339 
2340    if ((usage & PIPE_BIND_DEPTH_STENCIL) && si_is_zs_format_supported(format)) {
2341       retval |= PIPE_BIND_DEPTH_STENCIL;
2342    }
2343 
2344    if (usage & PIPE_BIND_VERTEX_BUFFER) {
2345       retval |= si_is_vertex_format_supported(screen, format, PIPE_BIND_VERTEX_BUFFER);
2346    }
2347 
2348    if (usage & PIPE_BIND_INDEX_BUFFER) {
2349       if (format == PIPE_FORMAT_R8_UINT ||
2350           format == PIPE_FORMAT_R16_UINT ||
2351           format == PIPE_FORMAT_R32_UINT)
2352          retval |= PIPE_BIND_INDEX_BUFFER;
2353    }
2354 
2355    if ((usage & PIPE_BIND_LINEAR) && !util_format_is_compressed(format) &&
2356        !(usage & PIPE_BIND_DEPTH_STENCIL))
2357       retval |= PIPE_BIND_LINEAR;
2358 
2359    if ((usage & PIPE_BIND_SAMPLER_REDUCTION_MINMAX) &&
2360        screen->get_param(screen, PIPE_CAP_SAMPLER_REDUCTION_MINMAX) &&
2361        si_is_reduction_mode_supported(screen, format))
2362       retval |= PIPE_BIND_SAMPLER_REDUCTION_MINMAX;
2363 
2364    return retval == usage;
2365 }
2366 
2367 /*
2368  * framebuffer handling
2369  */
2370 
si_choose_spi_color_formats(struct si_surface * surf,unsigned format,unsigned swap,unsigned ntype,bool is_depth)2371 static void si_choose_spi_color_formats(struct si_surface *surf, unsigned format, unsigned swap,
2372                                         unsigned ntype, bool is_depth)
2373 {
2374    struct ac_spi_color_formats formats = {};
2375 
2376    ac_choose_spi_color_formats(format, swap, ntype, is_depth, true, &formats);
2377 
2378    surf->spi_shader_col_format = formats.normal;
2379    surf->spi_shader_col_format_alpha = formats.alpha;
2380    surf->spi_shader_col_format_blend = formats.blend;
2381    surf->spi_shader_col_format_blend_alpha = formats.blend_alpha;
2382 }
2383 
si_initialize_color_surface(struct si_context * sctx,struct si_surface * surf)2384 static void si_initialize_color_surface(struct si_context *sctx, struct si_surface *surf)
2385 {
2386    struct si_texture *tex = (struct si_texture *)surf->base.texture;
2387    unsigned format, swap, ntype;//, endian;
2388 
2389    ntype = ac_get_cb_number_type(surf->base.format);
2390    format = ac_get_cb_format(sctx->gfx_level, surf->base.format);
2391 
2392    if (format == V_028C70_COLOR_INVALID) {
2393       PRINT_ERR("Invalid CB format: %d, disabling CB.\n", surf->base.format);
2394    }
2395    assert(format != V_028C70_COLOR_INVALID);
2396    swap = ac_translate_colorswap(sctx->gfx_level, surf->base.format, false);
2397 
2398    if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT) {
2399       if (format == V_028C70_COLOR_8 || format == V_028C70_COLOR_8_8 ||
2400           format == V_028C70_COLOR_8_8_8_8)
2401          surf->color_is_int8 = true;
2402       else if (format == V_028C70_COLOR_10_10_10_2 || format == V_028C70_COLOR_2_10_10_10)
2403          surf->color_is_int10 = true;
2404    }
2405 
2406    const struct ac_cb_state cb_state = {
2407       .surf = &tex->surface,
2408       .format = surf->base.format,
2409       .width = surf->width0,
2410       .height = surf->height0,
2411       .first_layer = surf->base.u.tex.first_layer,
2412       .last_layer = surf->base.u.tex.last_layer,
2413       .num_layers = util_max_layer(&tex->buffer.b.b, 0),
2414       .num_samples = tex->buffer.b.b.nr_samples,
2415       .num_storage_samples = tex->buffer.b.b.nr_storage_samples,
2416       .base_level = surf->base.u.tex.level,
2417       .num_levels = tex->buffer.b.b.last_level + 1,
2418    };
2419 
2420    ac_init_cb_surface(&sctx->screen->info, &cb_state, &surf->cb);
2421 
2422    /* Determine pixel shader export format */
2423    si_choose_spi_color_formats(surf, format, swap, ntype, tex->is_depth);
2424 
2425    surf->color_initialized = true;
2426 }
2427 
si_init_depth_surface(struct si_context * sctx,struct si_surface * surf)2428 static void si_init_depth_surface(struct si_context *sctx, struct si_surface *surf)
2429 {
2430    struct si_texture *tex = (struct si_texture *)surf->base.texture;
2431    unsigned level = surf->base.u.tex.level;
2432    unsigned format;
2433 
2434    format = ac_translate_dbformat(tex->db_render_format);
2435 
2436    assert(format != V_028040_Z_24 || sctx->gfx_level < GFX12);
2437    assert(format != V_028040_Z_INVALID);
2438 
2439    if (format == V_028040_Z_INVALID)
2440       PRINT_ERR("Invalid DB format: %d, disabling DB.\n", tex->buffer.b.b.format);
2441 
2442    /* Use the original Z format, not db_render_format, so that the polygon offset behaves as
2443     * expected by applications.
2444     */
2445    switch (tex->buffer.b.b.format) {
2446    case PIPE_FORMAT_Z16_UNORM:
2447       surf->db_format_index = 0;
2448       break;
2449    default: /* 24-bit */
2450       surf->db_format_index = 1;
2451       break;
2452    case PIPE_FORMAT_Z32_FLOAT:
2453    case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
2454       surf->db_format_index = 2;
2455       break;
2456    }
2457 
2458    const struct ac_ds_state ds_state = {
2459       .surf = &tex->surface,
2460       .va = tex->buffer.gpu_address,
2461       .format = tex->db_render_format,
2462       .width = tex->buffer.b.b.width0,
2463       .height = tex->buffer.b.b.height0,
2464       .level = level,
2465       .num_levels = tex->buffer.b.b.last_level + 1,
2466       .num_samples = tex->buffer.b.b.nr_samples,
2467       .first_layer = surf->base.u.tex.first_layer,
2468       .last_layer = surf->base.u.tex.last_layer,
2469       .allow_expclear = true,
2470       .htile_enabled = sctx->gfx_level < GFX12 && si_htile_enabled(tex, level, PIPE_MASK_ZS),
2471       .htile_stencil_disabled = tex->htile_stencil_disabled,
2472    };
2473 
2474    ac_init_ds_surface(&sctx->screen->info, &ds_state, &surf->ds);
2475 
2476    surf->depth_initialized = true;
2477 }
2478 
si_dec_framebuffer_counters(const struct pipe_framebuffer_state * state)2479 static void si_dec_framebuffer_counters(const struct pipe_framebuffer_state *state)
2480 {
2481    for (int i = 0; i < state->nr_cbufs; ++i) {
2482       struct si_surface *surf = NULL;
2483       struct si_texture *tex;
2484 
2485       if (!state->cbufs[i])
2486          continue;
2487       surf = (struct si_surface *)state->cbufs[i];
2488       tex = (struct si_texture *)surf->base.texture;
2489 
2490       p_atomic_dec(&tex->framebuffers_bound);
2491    }
2492 }
2493 
si_mark_display_dcc_dirty(struct si_context * sctx,struct si_texture * tex)2494 void si_mark_display_dcc_dirty(struct si_context *sctx, struct si_texture *tex)
2495 {
2496    assert(sctx->gfx_level < GFX12);
2497 
2498    if (!tex->surface.display_dcc_offset || tex->displayable_dcc_dirty)
2499       return;
2500 
2501    if (!(tex->buffer.external_usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH)) {
2502       struct hash_entry *entry = _mesa_hash_table_search(sctx->dirty_implicit_resources, tex);
2503       if (!entry) {
2504          struct pipe_resource *dummy = NULL;
2505          pipe_resource_reference(&dummy, &tex->buffer.b.b);
2506          _mesa_hash_table_insert(sctx->dirty_implicit_resources, tex, tex);
2507       }
2508    }
2509    tex->displayable_dcc_dirty = true;
2510 }
2511 
si_update_display_dcc_dirty(struct si_context * sctx)2512 static void si_update_display_dcc_dirty(struct si_context *sctx)
2513 {
2514    const struct pipe_framebuffer_state *state = &sctx->framebuffer.state;
2515 
2516    for (unsigned i = 0; i < state->nr_cbufs; i++) {
2517       if (state->cbufs[i])
2518          si_mark_display_dcc_dirty(sctx, (struct si_texture *)state->cbufs[i]->texture);
2519    }
2520 }
2521 
si_set_framebuffer_state(struct pipe_context * ctx,const struct pipe_framebuffer_state * state)2522 static void si_set_framebuffer_state(struct pipe_context *ctx,
2523                                      const struct pipe_framebuffer_state *state)
2524 {
2525    struct si_context *sctx = (struct si_context *)ctx;
2526    struct si_surface *surf = NULL;
2527    struct si_texture *tex;
2528    bool old_any_dst_linear = sctx->framebuffer.any_dst_linear;
2529    unsigned old_nr_samples = sctx->framebuffer.nr_samples;
2530    unsigned old_colorbuf_enabled_4bit = sctx->framebuffer.colorbuf_enabled_4bit;
2531    bool old_has_zsbuf = !!sctx->framebuffer.state.zsbuf;
2532    bool old_has_stencil =
2533       old_has_zsbuf &&
2534       ((struct si_texture *)sctx->framebuffer.state.zsbuf->texture)->surface.has_stencil;
2535    uint8_t old_db_format_index =
2536       old_has_zsbuf ?
2537       ((struct si_surface *)sctx->framebuffer.state.zsbuf)->db_format_index : -1;
2538    bool old_has_hiz_his = sctx->framebuffer.has_hiz_his;
2539    int i;
2540 
2541    /* Reject zero-sized framebuffers due to a hw bug on GFX6 that occurs
2542     * when PA_SU_HARDWARE_SCREEN_OFFSET != 0 and any_scissor.BR_X/Y <= 0.
2543     * We could implement the full workaround here, but it's a useless case.
2544     */
2545    if ((!state->width || !state->height) && (state->nr_cbufs || state->zsbuf)) {
2546       unreachable("the framebuffer shouldn't have zero area");
2547       return;
2548    }
2549 
2550    si_fb_barrier_after_rendering(sctx, SI_FB_BARRIER_SYNC_ALL);
2551 
2552    /* Disable DCC if the formats are incompatible. */
2553    if (sctx->gfx_level >= GFX8 && sctx->gfx_level < GFX11) {
2554       for (i = 0; i < state->nr_cbufs; i++) {
2555          if (!state->cbufs[i])
2556             continue;
2557 
2558          surf = (struct si_surface *)state->cbufs[i];
2559          tex = (struct si_texture *)surf->base.texture;
2560 
2561          if (!surf->dcc_incompatible)
2562             continue;
2563 
2564          if (vi_dcc_enabled(tex, surf->base.u.tex.level))
2565             if (!si_texture_disable_dcc(sctx, tex))
2566                si_decompress_dcc(sctx, tex);
2567 
2568          surf->dcc_incompatible = false;
2569       }
2570    }
2571 
2572    /* Take the maximum of the old and new count. If the new count is lower,
2573     * dirtying is needed to disable the unbound colorbuffers.
2574     */
2575    sctx->framebuffer.dirty_cbufs |=
2576       (1 << MAX2(sctx->framebuffer.state.nr_cbufs, state->nr_cbufs)) - 1;
2577    sctx->framebuffer.dirty_zsbuf |= sctx->framebuffer.state.zsbuf != state->zsbuf;
2578 
2579    si_dec_framebuffer_counters(&sctx->framebuffer.state);
2580    util_copy_framebuffer_state(&sctx->framebuffer.state, state);
2581 
2582    /* The framebuffer state must be set before the barrier. */
2583    si_fb_barrier_before_rendering(sctx);
2584 
2585    /* Recompute layers because frontends and utils might not set it. */
2586    sctx->framebuffer.state.layers = util_framebuffer_get_num_layers(state);
2587 
2588    sctx->framebuffer.colorbuf_enabled_4bit = 0;
2589    sctx->framebuffer.spi_shader_col_format = 0;
2590    sctx->framebuffer.spi_shader_col_format_alpha = 0;
2591    sctx->framebuffer.spi_shader_col_format_blend = 0;
2592    sctx->framebuffer.spi_shader_col_format_blend_alpha = 0;
2593    sctx->framebuffer.color_is_int8 = 0;
2594    sctx->framebuffer.color_is_int10 = 0;
2595 
2596    sctx->framebuffer.compressed_cb_mask = 0;
2597    sctx->framebuffer.uncompressed_cb_mask = 0;
2598    sctx->framebuffer.nr_samples = util_framebuffer_get_num_samples(state);
2599    sctx->framebuffer.nr_color_samples = sctx->framebuffer.nr_samples;
2600    sctx->framebuffer.log_samples = util_logbase2(sctx->framebuffer.nr_samples);
2601    sctx->framebuffer.any_dst_linear = false;
2602    sctx->framebuffer.CB_has_shader_readable_metadata = false;
2603    sctx->framebuffer.DB_has_shader_readable_metadata = false;
2604    sctx->framebuffer.all_DCC_pipe_aligned = true;
2605    sctx->framebuffer.has_dcc_msaa = false;
2606    sctx->framebuffer.min_bytes_per_pixel = 0;
2607    sctx->framebuffer.disable_vrs_flat_shading = false;
2608    sctx->framebuffer.has_stencil = false;
2609    sctx->framebuffer.has_hiz_his = false;
2610 
2611    for (i = 0; i < state->nr_cbufs; i++) {
2612       if (!state->cbufs[i])
2613          continue;
2614 
2615       surf = (struct si_surface *)state->cbufs[i];
2616       tex = (struct si_texture *)surf->base.texture;
2617 
2618       if (!surf->color_initialized) {
2619          si_initialize_color_surface(sctx, surf);
2620       }
2621 
2622       sctx->framebuffer.colorbuf_enabled_4bit |= 0xf << (i * 4);
2623       sctx->framebuffer.spi_shader_col_format |= surf->spi_shader_col_format << (i * 4);
2624       sctx->framebuffer.spi_shader_col_format_alpha |= surf->spi_shader_col_format_alpha << (i * 4);
2625       sctx->framebuffer.spi_shader_col_format_blend |= surf->spi_shader_col_format_blend << (i * 4);
2626       sctx->framebuffer.spi_shader_col_format_blend_alpha |= surf->spi_shader_col_format_blend_alpha
2627                                                              << (i * 4);
2628 
2629       if (surf->color_is_int8)
2630          sctx->framebuffer.color_is_int8 |= 1 << i;
2631       if (surf->color_is_int10)
2632          sctx->framebuffer.color_is_int10 |= 1 << i;
2633 
2634       if (tex->surface.fmask_offset)
2635          sctx->framebuffer.compressed_cb_mask |= 1 << i;
2636       else
2637          sctx->framebuffer.uncompressed_cb_mask |= 1 << i;
2638 
2639       /* Don't update nr_color_samples for non-AA buffers.
2640        * (e.g. destination of MSAA resolve)
2641        */
2642       if (tex->buffer.b.b.nr_samples >= 2 &&
2643           tex->buffer.b.b.nr_storage_samples < tex->buffer.b.b.nr_samples) {
2644          sctx->framebuffer.nr_color_samples =
2645             MIN2(sctx->framebuffer.nr_color_samples, tex->buffer.b.b.nr_storage_samples);
2646          sctx->framebuffer.nr_color_samples = MAX2(1, sctx->framebuffer.nr_color_samples);
2647       }
2648 
2649       if (tex->surface.is_linear)
2650          sctx->framebuffer.any_dst_linear = true;
2651 
2652       if (vi_dcc_enabled(tex, surf->base.u.tex.level)) {
2653          sctx->framebuffer.CB_has_shader_readable_metadata = true;
2654 
2655          if (sctx->gfx_level >= GFX9 && sctx->gfx_level < GFX12 &&
2656              !tex->surface.u.gfx9.color.dcc.pipe_aligned)
2657             sctx->framebuffer.all_DCC_pipe_aligned = false;
2658 
2659          if (tex->buffer.b.b.nr_storage_samples >= 2)
2660             sctx->framebuffer.has_dcc_msaa = true;
2661       }
2662 
2663       p_atomic_inc(&tex->framebuffers_bound);
2664 
2665       /* Update the minimum but don't keep 0. */
2666       if (!sctx->framebuffer.min_bytes_per_pixel ||
2667           tex->surface.bpe < sctx->framebuffer.min_bytes_per_pixel)
2668          sctx->framebuffer.min_bytes_per_pixel = tex->surface.bpe;
2669 
2670       /* Disable VRS flat shading where it decreases performance.
2671        * This gives the best results for slow clears for AMD_TEST=blitperf on Navi31.
2672        */
2673       if ((sctx->framebuffer.nr_samples == 8 && tex->surface.bpe != 2) ||
2674           (tex->surface.thick_tiling && tex->surface.bpe == 4 &&
2675            util_format_get_nr_components(surf->base.format) == 4))
2676          sctx->framebuffer.disable_vrs_flat_shading = true;
2677    }
2678 
2679    struct si_texture *zstex = NULL;
2680 
2681    if (state->zsbuf) {
2682       surf = (struct si_surface *)state->zsbuf;
2683       zstex = (struct si_texture *)surf->base.texture;
2684 
2685       if (!surf->depth_initialized) {
2686          si_init_depth_surface(sctx, surf);
2687       }
2688 
2689       if (sctx->gfx_level < GFX12 &&
2690           vi_tc_compat_htile_enabled(zstex, surf->base.u.tex.level, PIPE_MASK_ZS))
2691          sctx->framebuffer.DB_has_shader_readable_metadata = true;
2692 
2693       /* Update the minimum but don't keep 0. */
2694       if (!sctx->framebuffer.min_bytes_per_pixel ||
2695           zstex->surface.bpe < sctx->framebuffer.min_bytes_per_pixel)
2696          sctx->framebuffer.min_bytes_per_pixel = zstex->surface.bpe;
2697 
2698       /* Update polygon offset based on the Z format. */
2699       if (sctx->queued.named.rasterizer->uses_poly_offset &&
2700           surf->db_format_index != old_db_format_index)
2701          sctx->dirty_atoms |= SI_STATE_BIT(rasterizer);
2702 
2703       if (util_format_has_stencil(util_format_description(zstex->buffer.b.b.format)))
2704          sctx->framebuffer.has_stencil = true;
2705 
2706       if (sctx->gfx_level >= GFX12) {
2707          sctx->framebuffer.has_hiz_his = zstex->surface.u.gfx9.zs.hiz.offset ||
2708                                          zstex->surface.u.gfx9.zs.his.offset;
2709       }
2710    }
2711 
2712    si_update_ps_colorbuf0_slot(sctx);
2713    si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
2714    si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
2715 
2716    /* NGG cull state uses the sample count. */
2717    if (sctx->screen->use_ngg_culling)
2718       si_mark_atom_dirty(sctx, &sctx->atoms.s.ngg_cull_state);
2719 
2720    if (sctx->screen->dpbb_allowed)
2721       si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
2722 
2723    if (sctx->framebuffer.any_dst_linear != old_any_dst_linear ||
2724        sctx->framebuffer.has_hiz_his != old_has_hiz_his)
2725       si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2726 
2727    if (sctx->screen->info.has_out_of_order_rast &&
2728        (sctx->framebuffer.colorbuf_enabled_4bit != old_colorbuf_enabled_4bit ||
2729         !!sctx->framebuffer.state.zsbuf != old_has_zsbuf ||
2730         (zstex && zstex->surface.has_stencil != old_has_stencil)))
2731       si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2732 
2733    if (sctx->framebuffer.nr_samples != old_nr_samples) {
2734       struct pipe_constant_buffer constbuf = {0};
2735 
2736       si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2737       si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
2738 
2739       if (!sctx->sample_pos_buffer) {
2740          sctx->sample_pos_buffer = pipe_buffer_create_with_data(&sctx->b, 0, PIPE_USAGE_DEFAULT,
2741                                                       sizeof(sctx->sample_positions),
2742                                                       &sctx->sample_positions);
2743       }
2744       constbuf.buffer = sctx->sample_pos_buffer;
2745 
2746       /* Set sample locations as fragment shader constants. */
2747       switch (sctx->framebuffer.nr_samples) {
2748       case 1:
2749          constbuf.buffer_offset = 0;
2750          break;
2751       case 2:
2752          constbuf.buffer_offset =
2753             (uint8_t *)sctx->sample_positions.x2 - (uint8_t *)sctx->sample_positions.x1;
2754          break;
2755       case 4:
2756          constbuf.buffer_offset =
2757             (uint8_t *)sctx->sample_positions.x4 - (uint8_t *)sctx->sample_positions.x1;
2758          break;
2759       case 8:
2760          constbuf.buffer_offset =
2761             (uint8_t *)sctx->sample_positions.x8 - (uint8_t *)sctx->sample_positions.x1;
2762          break;
2763       case 16:
2764          constbuf.buffer_offset =
2765             (uint8_t *)sctx->sample_positions.x16 - (uint8_t *)sctx->sample_positions.x1;
2766          break;
2767       default:
2768          PRINT_ERR("Requested an invalid number of samples %i.\n", sctx->framebuffer.nr_samples);
2769          assert(0);
2770       }
2771       constbuf.buffer_size = sctx->framebuffer.nr_samples * 2 * 4;
2772       si_set_internal_const_buffer(sctx, SI_PS_CONST_SAMPLE_POSITIONS, &constbuf);
2773 
2774       si_mark_atom_dirty(sctx, &sctx->atoms.s.sample_locations);
2775    }
2776 
2777    si_ps_key_update_framebuffer(sctx);
2778    si_ps_key_update_framebuffer_blend_rasterizer(sctx);
2779    si_ps_key_update_framebuffer_rasterizer_sample_shading(sctx);
2780    si_vs_ps_key_update_rast_prim_smooth_stipple(sctx);
2781    si_update_ps_inputs_read_or_disabled(sctx);
2782    si_update_vrs_flat_shading(sctx);
2783    sctx->do_update_shaders = true;
2784 
2785    if (sctx->gfx_level < GFX12 && !sctx->decompression_enabled) {
2786       /* Prevent textures decompression when the framebuffer state
2787        * changes come from the decompression passes themselves.
2788        */
2789       sctx->need_check_render_feedback = true;
2790    }
2791 }
2792 
gfx6_emit_framebuffer_state(struct si_context * sctx,unsigned index)2793 static void gfx6_emit_framebuffer_state(struct si_context *sctx, unsigned index)
2794 {
2795    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
2796    struct pipe_framebuffer_state *state = &sctx->framebuffer.state;
2797    unsigned i, nr_cbufs = state->nr_cbufs;
2798    struct si_texture *tex = NULL;
2799    struct si_surface *cb = NULL;
2800    bool is_msaa_resolve = state->nr_cbufs == 2 &&
2801                           state->cbufs[0] && state->cbufs[0]->texture->nr_samples > 1 &&
2802                           state->cbufs[1] && state->cbufs[1]->texture->nr_samples <= 1;
2803 
2804    /* CB can't do MSAA resolve on gfx11. */
2805    assert(!is_msaa_resolve || sctx->gfx_level < GFX11);
2806 
2807    radeon_begin(cs);
2808 
2809    /* Colorbuffers. */
2810    for (i = 0; i < nr_cbufs; i++) {
2811       if (!(sctx->framebuffer.dirty_cbufs & (1 << i)))
2812          continue;
2813 
2814       /* RB+ depth-only rendering. See the comment where we set rbplus_depth_only_opt for more
2815        * information.
2816        */
2817       if (i == 0 &&
2818           sctx->screen->info.rbplus_allowed &&
2819           !sctx->queued.named.blend->cb_target_mask) {
2820          radeon_set_context_reg(R_028C70_CB_COLOR0_INFO + i * 0x3C,
2821                                 (sctx->gfx_level >= GFX11 ?
2822                                    S_028C70_FORMAT_GFX11(V_028C70_COLOR_32) :
2823                                    S_028C70_FORMAT_GFX6(V_028C70_COLOR_32)) |
2824                                 S_028C70_NUMBER_TYPE(V_028C70_NUMBER_FLOAT));
2825          continue;
2826       }
2827 
2828       cb = (struct si_surface *)state->cbufs[i];
2829       if (!cb) {
2830          radeon_set_context_reg(R_028C70_CB_COLOR0_INFO + i * 0x3C,
2831                                 sctx->gfx_level >= GFX11 ?
2832                                    S_028C70_FORMAT_GFX11(V_028C70_COLOR_INVALID) :
2833                                    S_028C70_FORMAT_GFX6(V_028C70_COLOR_INVALID));
2834          continue;
2835       }
2836 
2837       tex = (struct si_texture *)cb->base.texture;
2838       radeon_add_to_buffer_list(
2839          sctx, &sctx->gfx_cs, &tex->buffer, RADEON_USAGE_READWRITE | RADEON_USAGE_CB_NEEDS_IMPLICIT_SYNC |
2840          (tex->buffer.b.b.nr_samples > 1 ? RADEON_PRIO_COLOR_BUFFER_MSAA : RADEON_PRIO_COLOR_BUFFER));
2841 
2842       if (tex->cmask_buffer && tex->cmask_buffer != &tex->buffer) {
2843          radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, tex->cmask_buffer,
2844                                    RADEON_USAGE_READWRITE | RADEON_USAGE_CB_NEEDS_IMPLICIT_SYNC |
2845                                    RADEON_PRIO_SEPARATE_META);
2846       }
2847 
2848       /* Compute mutable surface parameters. */
2849       const struct ac_mutable_cb_state mutable_cb_state = {
2850          .surf = &tex->surface,
2851          .cb = &cb->cb,
2852          .va = tex->buffer.gpu_address,
2853          .base_level = cb->base.u.tex.level,
2854          .num_samples = cb->base.texture->nr_samples,
2855          .fmask_enabled = !!tex->surface.fmask_offset,
2856          /* CMASK and fast clears are configured elsewhere. */
2857          .cmask_enabled = false,
2858          .fast_clear_enabled = false,
2859          .dcc_enabled = vi_dcc_enabled(tex, cb->base.u.tex.level) &&
2860                         (i != 1 || !is_msaa_resolve),
2861       };
2862       struct ac_cb_surface cb_surf;
2863 
2864       ac_set_mutable_cb_surface_fields(&sctx->screen->info, &mutable_cb_state, &cb_surf);
2865 
2866       cb_surf.cb_color_info |= tex->cb_color_info;
2867 
2868       if (sctx->gfx_level < GFX11) {
2869          if (tex->swap_rgb_to_bgr) {
2870             /* Swap R and B channels. */
2871             static unsigned rgb_to_bgr[4] = {
2872                [V_028C70_SWAP_STD] = V_028C70_SWAP_ALT,
2873                [V_028C70_SWAP_ALT] = V_028C70_SWAP_STD,
2874                [V_028C70_SWAP_STD_REV] = V_028C70_SWAP_ALT_REV,
2875                [V_028C70_SWAP_ALT_REV] = V_028C70_SWAP_STD_REV,
2876             };
2877             unsigned swap = rgb_to_bgr[G_028C70_COMP_SWAP(cb_surf.cb_color_info)];
2878 
2879             cb_surf.cb_color_info &= C_028C70_COMP_SWAP;
2880             cb_surf.cb_color_info |= S_028C70_COMP_SWAP(swap);
2881          }
2882 
2883          if (cb->base.u.tex.level > 0)
2884             cb_surf.cb_color_info &= C_028C70_FAST_CLEAR;
2885          else
2886             cb_surf.cb_color_cmask = tex->cmask_base_address_reg;
2887       }
2888 
2889       if (sctx->gfx_level >= GFX11) {
2890          radeon_set_context_reg(R_028C60_CB_COLOR0_BASE + i * 0x3C, cb_surf.cb_color_base);
2891 
2892          radeon_set_context_reg_seq(R_028C6C_CB_COLOR0_VIEW + i * 0x3C, 4);
2893          radeon_emit(cb_surf.cb_color_view);                      /* CB_COLOR0_VIEW */
2894          radeon_emit(cb_surf.cb_color_info);                          /* CB_COLOR0_INFO */
2895          radeon_emit(cb_surf.cb_color_attrib);                    /* CB_COLOR0_ATTRIB */
2896          radeon_emit(cb_surf.cb_dcc_control);                        /* CB_COLOR0_FDCC_CONTROL */
2897 
2898          radeon_set_context_reg(R_028C94_CB_COLOR0_DCC_BASE + i * 0x3C, cb_surf.cb_dcc_base);
2899          radeon_set_context_reg(R_028E40_CB_COLOR0_BASE_EXT + i * 4, cb_surf.cb_color_base >> 32);
2900          radeon_set_context_reg(R_028EA0_CB_COLOR0_DCC_BASE_EXT + i * 4, cb_surf.cb_dcc_base >> 32);
2901          radeon_set_context_reg(R_028EC0_CB_COLOR0_ATTRIB2 + i * 4, cb_surf.cb_color_attrib2);
2902          radeon_set_context_reg(R_028EE0_CB_COLOR0_ATTRIB3 + i * 4, cb_surf.cb_color_attrib3);
2903       } else if (sctx->gfx_level >= GFX10) {
2904          radeon_set_context_reg_seq(R_028C60_CB_COLOR0_BASE + i * 0x3C, 14);
2905          radeon_emit(cb_surf.cb_color_base);             /* CB_COLOR0_BASE */
2906          radeon_emit(0);                         /* hole */
2907          radeon_emit(0);                         /* hole */
2908          radeon_emit(cb_surf.cb_color_view);         /* CB_COLOR0_VIEW */
2909          radeon_emit(cb_surf.cb_color_info);             /* CB_COLOR0_INFO */
2910          radeon_emit(cb_surf.cb_color_attrib);       /* CB_COLOR0_ATTRIB */
2911          radeon_emit(cb_surf.cb_dcc_control);        /* CB_COLOR0_DCC_CONTROL */
2912          radeon_emit(cb_surf.cb_color_cmask);            /* CB_COLOR0_CMASK */
2913          radeon_emit(0);                         /* hole */
2914          radeon_emit(cb_surf.cb_color_fmask);            /* CB_COLOR0_FMASK */
2915          radeon_emit(0);                         /* hole */
2916          radeon_emit(tex->color_clear_value[0]); /* CB_COLOR0_CLEAR_WORD0 */
2917          radeon_emit(tex->color_clear_value[1]); /* CB_COLOR0_CLEAR_WORD1 */
2918          radeon_emit(cb_surf.cb_dcc_base);               /* CB_COLOR0_DCC_BASE */
2919 
2920          radeon_set_context_reg(R_028E40_CB_COLOR0_BASE_EXT + i * 4, cb_surf.cb_color_base >> 32);
2921          radeon_set_context_reg(R_028E60_CB_COLOR0_CMASK_BASE_EXT + i * 4,
2922                                 cb_surf.cb_color_cmask >> 32);
2923          radeon_set_context_reg(R_028E80_CB_COLOR0_FMASK_BASE_EXT + i * 4,
2924                                 cb_surf.cb_color_fmask >> 32);
2925          radeon_set_context_reg(R_028EA0_CB_COLOR0_DCC_BASE_EXT + i * 4, cb_surf.cb_dcc_base >> 32);
2926          radeon_set_context_reg(R_028EC0_CB_COLOR0_ATTRIB2 + i * 4, cb_surf.cb_color_attrib2);
2927          radeon_set_context_reg(R_028EE0_CB_COLOR0_ATTRIB3 + i * 4, cb_surf.cb_color_attrib3);
2928       } else if (sctx->gfx_level == GFX9) {
2929          radeon_set_context_reg_seq(R_028C60_CB_COLOR0_BASE + i * 0x3C, 15);
2930          radeon_emit(cb_surf.cb_color_base);                            /* CB_COLOR0_BASE */
2931          radeon_emit(S_028C64_BASE_256B(cb_surf.cb_color_base >> 32));  /* CB_COLOR0_BASE_EXT */
2932          radeon_emit(cb_surf.cb_color_attrib2);                     /* CB_COLOR0_ATTRIB2 */
2933          radeon_emit(cb_surf.cb_color_view);                        /* CB_COLOR0_VIEW */
2934          radeon_emit(cb_surf.cb_color_info);                            /* CB_COLOR0_INFO */
2935          radeon_emit(cb_surf.cb_color_attrib);                          /* CB_COLOR0_ATTRIB */
2936          radeon_emit(cb_surf.cb_dcc_control);                       /* CB_COLOR0_DCC_CONTROL */
2937          radeon_emit(cb_surf.cb_color_cmask);                           /* CB_COLOR0_CMASK */
2938          radeon_emit(S_028C80_BASE_256B(cb_surf.cb_color_cmask >> 32)); /* CB_COLOR0_CMASK_BASE_EXT */
2939          radeon_emit(cb_surf.cb_color_fmask);                           /* CB_COLOR0_FMASK */
2940          radeon_emit(S_028C88_BASE_256B(cb_surf.cb_color_fmask >> 32)); /* CB_COLOR0_FMASK_BASE_EXT */
2941          radeon_emit(tex->color_clear_value[0]);                /* CB_COLOR0_CLEAR_WORD0 */
2942          radeon_emit(tex->color_clear_value[1]);                /* CB_COLOR0_CLEAR_WORD1 */
2943          radeon_emit(cb_surf.cb_dcc_base);                              /* CB_COLOR0_DCC_BASE */
2944          radeon_emit(S_028C98_BASE_256B(cb_surf.cb_dcc_base >> 32));    /* CB_COLOR0_DCC_BASE_EXT */
2945 
2946          radeon_set_context_reg(R_0287A0_CB_MRT0_EPITCH + i * 4, cb_surf.cb_mrt_epitch);
2947       } else {
2948          /* GFX6-8 */
2949          radeon_set_context_reg_seq(R_028C60_CB_COLOR0_BASE + i * 0x3C,
2950                                     sctx->gfx_level >= GFX8 ? 14 : 13);
2951          radeon_emit(cb_surf.cb_color_base);                              /* CB_COLOR0_BASE */
2952          radeon_emit(cb_surf.cb_color_pitch);                             /* CB_COLOR0_PITCH */
2953          radeon_emit(cb_surf.cb_color_slice);                             /* CB_COLOR0_SLICE */
2954          radeon_emit(cb_surf.cb_color_view);                          /* CB_COLOR0_VIEW */
2955          radeon_emit(cb_surf.cb_color_info);                              /* CB_COLOR0_INFO */
2956          radeon_emit(cb_surf.cb_color_attrib);                            /* CB_COLOR0_ATTRIB */
2957          radeon_emit(cb_surf.cb_dcc_control);                         /* CB_COLOR0_DCC_CONTROL */
2958          radeon_emit(cb_surf.cb_color_cmask);                             /* CB_COLOR0_CMASK */
2959          radeon_emit(tex->surface.u.legacy.color.cmask_slice_tile_max); /* CB_COLOR0_CMASK_SLICE */
2960          radeon_emit(cb_surf.cb_color_fmask);                             /* CB_COLOR0_FMASK */
2961          radeon_emit(cb_surf.cb_color_fmask_slice);                       /* CB_COLOR0_FMASK_SLICE */
2962          radeon_emit(tex->color_clear_value[0]);                  /* CB_COLOR0_CLEAR_WORD0 */
2963          radeon_emit(tex->color_clear_value[1]);                  /* CB_COLOR0_CLEAR_WORD1 */
2964 
2965          if (sctx->gfx_level >= GFX8) /* R_028C94_CB_COLOR0_DCC_BASE */
2966             radeon_emit(cb_surf.cb_dcc_base);
2967       }
2968    }
2969    for (; i < 8; i++)
2970       if (sctx->framebuffer.dirty_cbufs & (1 << i))
2971          radeon_set_context_reg(R_028C70_CB_COLOR0_INFO + i * 0x3C, 0);
2972 
2973    /* ZS buffer. */
2974    if (state->zsbuf && sctx->framebuffer.dirty_zsbuf) {
2975       struct si_surface *zb = (struct si_surface *)state->zsbuf;
2976       struct si_texture *tex = (struct si_texture *)zb->base.texture;
2977 
2978       radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, &tex->buffer, RADEON_USAGE_READWRITE |
2979                                 (zb->base.texture->nr_samples > 1 ? RADEON_PRIO_DEPTH_BUFFER_MSAA
2980                                                                   : RADEON_PRIO_DEPTH_BUFFER));
2981 
2982       const unsigned level = zb->base.u.tex.level;
2983 
2984       /* Set mutable fields. */
2985       const struct ac_mutable_ds_state mutable_ds_state = {
2986          .ds = &zb->ds,
2987          .format = tex->db_render_format,
2988          .tc_compat_htile_enabled = vi_tc_compat_htile_enabled(tex, level, PIPE_MASK_ZS),
2989          .zrange_precision = tex->depth_clear_value[level] != 0,
2990       };
2991       struct ac_ds_surface ds;
2992 
2993       ac_set_mutable_ds_surface_fields(&sctx->screen->info, &mutable_ds_state, &ds);
2994 
2995       if (sctx->gfx_level >= GFX10) {
2996          radeon_set_context_reg(R_028014_DB_HTILE_DATA_BASE, ds.u.gfx6.db_htile_data_base);
2997          radeon_set_context_reg(R_02801C_DB_DEPTH_SIZE_XY, ds.db_depth_size);
2998 
2999          if (sctx->gfx_level >= GFX11) {
3000             radeon_set_context_reg_seq(R_028040_DB_Z_INFO, 6);
3001          } else {
3002             radeon_set_context_reg_seq(R_02803C_DB_DEPTH_INFO, 7);
3003             radeon_emit(S_02803C_RESOURCE_LEVEL(1)); /* DB_DEPTH_INFO */
3004          }
3005          radeon_emit(ds.db_z_info);                  /* DB_Z_INFO */
3006          radeon_emit(ds.db_stencil_info);     /* DB_STENCIL_INFO */
3007          radeon_emit(ds.db_depth_base);   /* DB_Z_READ_BASE */
3008          radeon_emit(ds.db_stencil_base); /* DB_STENCIL_READ_BASE */
3009          radeon_emit(ds.db_depth_base);   /* DB_Z_WRITE_BASE */
3010          radeon_emit(ds.db_stencil_base); /* DB_STENCIL_WRITE_BASE */
3011 
3012          radeon_set_context_reg_seq(R_028068_DB_Z_READ_BASE_HI, 5);
3013          radeon_emit(ds.db_depth_base >> 32);      /* DB_Z_READ_BASE_HI */
3014          radeon_emit(ds.db_stencil_base >> 32);    /* DB_STENCIL_READ_BASE_HI */
3015          radeon_emit(ds.db_depth_base >> 32);      /* DB_Z_WRITE_BASE_HI */
3016          radeon_emit(ds.db_stencil_base >> 32);    /* DB_STENCIL_WRITE_BASE_HI */
3017          radeon_emit(ds.u.gfx6.db_htile_data_base >> 32); /* DB_HTILE_DATA_BASE_HI */
3018       } else if (sctx->gfx_level == GFX9) {
3019          radeon_set_context_reg_seq(R_028014_DB_HTILE_DATA_BASE, 3);
3020          radeon_emit(ds.u.gfx6.db_htile_data_base); /* DB_HTILE_DATA_BASE */
3021          radeon_emit(S_028018_BASE_HI(ds.u.gfx6.db_htile_data_base >> 32)); /* DB_HTILE_DATA_BASE_HI */
3022          radeon_emit(ds.db_depth_size);                          /* DB_DEPTH_SIZE */
3023 
3024          radeon_set_context_reg_seq(R_028038_DB_Z_INFO, 10);
3025          radeon_emit(ds.db_z_info);                                   /* DB_Z_INFO */
3026          radeon_emit(ds.db_stencil_info);                             /* DB_STENCIL_INFO */
3027          radeon_emit(ds.db_depth_base);                           /* DB_Z_READ_BASE */
3028          radeon_emit(S_028044_BASE_HI(ds.db_depth_base >> 32));   /* DB_Z_READ_BASE_HI */
3029          radeon_emit(ds.db_stencil_base);                         /* DB_STENCIL_READ_BASE */
3030          radeon_emit(S_02804C_BASE_HI(ds.db_stencil_base >> 32)); /* DB_STENCIL_READ_BASE_HI */
3031          radeon_emit(ds.db_depth_base);                           /* DB_Z_WRITE_BASE */
3032          radeon_emit(S_028054_BASE_HI(ds.db_depth_base >> 32));   /* DB_Z_WRITE_BASE_HI */
3033          radeon_emit(ds.db_stencil_base);                         /* DB_STENCIL_WRITE_BASE */
3034          radeon_emit(S_02805C_BASE_HI(ds.db_stencil_base >> 32)); /* DB_STENCIL_WRITE_BASE_HI */
3035 
3036          radeon_set_context_reg_seq(R_028068_DB_Z_INFO2, 2);
3037          radeon_emit(ds.u.gfx6.db_z_info2);       /* DB_Z_INFO2 */
3038          radeon_emit(ds.u.gfx6.db_stencil_info2); /* DB_STENCIL_INFO2 */
3039       } else {
3040          /* GFX6-GFX8 */
3041          radeon_set_context_reg(R_028014_DB_HTILE_DATA_BASE, ds.u.gfx6.db_htile_data_base);
3042 
3043          radeon_set_context_reg_seq(R_02803C_DB_DEPTH_INFO, 9);
3044          radeon_emit(ds.u.gfx6.db_depth_info);   /* DB_DEPTH_INFO */
3045          radeon_emit(ds.db_z_info);           /* DB_Z_INFO */
3046          radeon_emit(ds.db_stencil_info);     /* DB_STENCIL_INFO */
3047          radeon_emit(ds.db_depth_base);   /* DB_Z_READ_BASE */
3048          radeon_emit(ds.db_stencil_base); /* DB_STENCIL_READ_BASE */
3049          radeon_emit(ds.db_depth_base);   /* DB_Z_WRITE_BASE */
3050          radeon_emit(ds.db_stencil_base); /* DB_STENCIL_WRITE_BASE */
3051          radeon_emit(ds.db_depth_size);   /* DB_DEPTH_SIZE */
3052          radeon_emit(ds.u.gfx6.db_depth_slice);  /* DB_DEPTH_SLICE */
3053       }
3054 
3055       radeon_set_context_reg_seq(R_028028_DB_STENCIL_CLEAR, 2);
3056       radeon_emit(tex->stencil_clear_value[level]);    /* R_028028_DB_STENCIL_CLEAR */
3057       radeon_emit(fui(tex->depth_clear_value[level])); /* R_02802C_DB_DEPTH_CLEAR */
3058 
3059       radeon_set_context_reg(R_028008_DB_DEPTH_VIEW, ds.db_depth_view);
3060       radeon_set_context_reg(R_028ABC_DB_HTILE_SURFACE, ds.u.gfx6.db_htile_surface);
3061    } else if (sctx->framebuffer.dirty_zsbuf) {
3062       if (sctx->gfx_level == GFX9)
3063          radeon_set_context_reg_seq(R_028038_DB_Z_INFO, 2);
3064       else
3065          radeon_set_context_reg_seq(R_028040_DB_Z_INFO, 2);
3066 
3067       /* Gfx11+: DB_Z_INFO.NUM_SAMPLES should match the framebuffer samples if no Z/S is bound.
3068        * It determines the sample count for VRS, primitive-ordered pixel shading, and occlusion
3069        * queries.
3070        */
3071       radeon_emit(S_028040_FORMAT(V_028040_Z_INVALID) |       /* DB_Z_INFO */
3072                   S_028040_NUM_SAMPLES(sctx->gfx_level >= GFX11 ? sctx->framebuffer.log_samples : 0));
3073       radeon_emit(S_028044_FORMAT(V_028044_STENCIL_INVALID)); /* DB_STENCIL_INFO */
3074    }
3075 
3076    /* Framebuffer dimensions. */
3077    /* PA_SC_WINDOW_SCISSOR_TL is set to 0,0 in gfx*_init_gfx_preamble_state */
3078    radeon_set_context_reg(R_028208_PA_SC_WINDOW_SCISSOR_BR,
3079                           S_028208_BR_X(state->width) | S_028208_BR_Y(state->height));
3080 
3081    if (sctx->screen->dpbb_allowed &&
3082        sctx->screen->pbb_context_states_per_bin > 1)
3083       radeon_event_write(V_028A90_BREAK_BATCH);
3084 
3085    radeon_end();
3086 
3087    si_update_display_dcc_dirty(sctx);
3088 
3089    sctx->framebuffer.dirty_cbufs = 0;
3090    sctx->framebuffer.dirty_zsbuf = false;
3091 }
3092 
gfx11_dgpu_emit_framebuffer_state(struct si_context * sctx,unsigned index)3093 static void gfx11_dgpu_emit_framebuffer_state(struct si_context *sctx, unsigned index)
3094 {
3095    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
3096    struct pipe_framebuffer_state *state = &sctx->framebuffer.state;
3097    unsigned i, nr_cbufs = state->nr_cbufs;
3098    struct si_texture *tex = NULL;
3099    struct si_surface *cb = NULL;
3100    bool is_msaa_resolve = state->nr_cbufs == 2 &&
3101                           state->cbufs[0] && state->cbufs[0]->texture->nr_samples > 1 &&
3102                           state->cbufs[1] && state->cbufs[1]->texture->nr_samples <= 1;
3103 
3104    /* CB can't do MSAA resolve on gfx11. */
3105    assert(!is_msaa_resolve);
3106 
3107    radeon_begin(cs);
3108    gfx11_begin_packed_context_regs();
3109 
3110    /* Colorbuffers. */
3111    for (i = 0; i < nr_cbufs; i++) {
3112       if (!(sctx->framebuffer.dirty_cbufs & (1 << i)))
3113          continue;
3114 
3115       /* RB+ depth-only rendering. See the comment where we set rbplus_depth_only_opt for more
3116        * information.
3117        */
3118       if (i == 0 &&
3119           sctx->screen->info.rbplus_allowed &&
3120           !sctx->queued.named.blend->cb_target_mask) {
3121          gfx11_set_context_reg(R_028C70_CB_COLOR0_INFO + i * 0x3C,
3122                                S_028C70_FORMAT_GFX11(V_028C70_COLOR_32) |
3123                                S_028C70_NUMBER_TYPE(V_028C70_NUMBER_FLOAT));
3124          continue;
3125       }
3126 
3127       cb = (struct si_surface *)state->cbufs[i];
3128       if (!cb) {
3129          gfx11_set_context_reg(R_028C70_CB_COLOR0_INFO + i * 0x3C,
3130                                S_028C70_FORMAT_GFX11(V_028C70_COLOR_INVALID));
3131          continue;
3132       }
3133 
3134       tex = (struct si_texture *)cb->base.texture;
3135       radeon_add_to_buffer_list(
3136          sctx, &sctx->gfx_cs, &tex->buffer, RADEON_USAGE_READWRITE | RADEON_USAGE_CB_NEEDS_IMPLICIT_SYNC |
3137          (tex->buffer.b.b.nr_samples > 1 ? RADEON_PRIO_COLOR_BUFFER_MSAA : RADEON_PRIO_COLOR_BUFFER));
3138 
3139       if (tex->cmask_buffer && tex->cmask_buffer != &tex->buffer) {
3140          radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, tex->cmask_buffer,
3141                                    RADEON_USAGE_READWRITE | RADEON_USAGE_CB_NEEDS_IMPLICIT_SYNC |
3142                                    RADEON_PRIO_SEPARATE_META);
3143       }
3144 
3145       /* Compute mutable surface parameters. */
3146       const struct ac_mutable_cb_state mutable_cb_state = {
3147          .surf = &tex->surface,
3148          .cb = &cb->cb,
3149          .va = tex->buffer.gpu_address,
3150          .num_samples = cb->base.texture->nr_samples,
3151          .dcc_enabled = vi_dcc_enabled(tex, cb->base.u.tex.level),
3152       };
3153       struct ac_cb_surface cb_surf;
3154 
3155       ac_set_mutable_cb_surface_fields(&sctx->screen->info, &mutable_cb_state, &cb_surf);
3156 
3157       cb_surf.cb_color_info |= tex->cb_color_info;
3158 
3159       gfx11_set_context_reg(R_028C60_CB_COLOR0_BASE + i * 0x3C, cb_surf.cb_color_base);
3160       gfx11_set_context_reg(R_028C6C_CB_COLOR0_VIEW + i * 0x3C, cb_surf.cb_color_view);
3161       gfx11_set_context_reg(R_028C70_CB_COLOR0_INFO + i * 0x3C, cb_surf.cb_color_info);
3162       gfx11_set_context_reg(R_028C74_CB_COLOR0_ATTRIB + i * 0x3C, cb_surf.cb_color_attrib);
3163       gfx11_set_context_reg(R_028C78_CB_COLOR0_DCC_CONTROL + i * 0x3C, cb_surf.cb_dcc_control);
3164       gfx11_set_context_reg(R_028C94_CB_COLOR0_DCC_BASE + i * 0x3C, cb_surf.cb_dcc_base);
3165       gfx11_set_context_reg(R_028E40_CB_COLOR0_BASE_EXT + i * 4, cb_surf.cb_color_base >> 32);
3166       gfx11_set_context_reg(R_028EA0_CB_COLOR0_DCC_BASE_EXT + i * 4, cb_surf.cb_dcc_base >> 32);
3167       gfx11_set_context_reg(R_028EC0_CB_COLOR0_ATTRIB2 + i * 4, cb_surf.cb_color_attrib2);
3168       gfx11_set_context_reg(R_028EE0_CB_COLOR0_ATTRIB3 + i * 4, cb_surf.cb_color_attrib3);
3169    }
3170    for (; i < 8; i++)
3171       if (sctx->framebuffer.dirty_cbufs & (1 << i))
3172          gfx11_set_context_reg(R_028C70_CB_COLOR0_INFO + i * 0x3C, 0);
3173 
3174    /* ZS buffer. */
3175    if (state->zsbuf && sctx->framebuffer.dirty_zsbuf) {
3176       struct si_surface *zb = (struct si_surface *)state->zsbuf;
3177       struct si_texture *tex = (struct si_texture *)zb->base.texture;
3178 
3179       radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, &tex->buffer, RADEON_USAGE_READWRITE |
3180                                 (zb->base.texture->nr_samples > 1 ? RADEON_PRIO_DEPTH_BUFFER_MSAA
3181                                                                   : RADEON_PRIO_DEPTH_BUFFER));
3182 
3183       const unsigned level = zb->base.u.tex.level;
3184 
3185       /* Set mutable fields. */
3186       const struct ac_mutable_ds_state mutable_ds_state = {
3187          .ds = &zb->ds,
3188          .format = tex->db_render_format,
3189          .tc_compat_htile_enabled = vi_tc_compat_htile_enabled(tex, level, PIPE_MASK_ZS),
3190          .zrange_precision = tex->depth_clear_value[level] != 0,
3191       };
3192       struct ac_ds_surface ds;
3193 
3194       ac_set_mutable_ds_surface_fields(&sctx->screen->info, &mutable_ds_state, &ds);
3195 
3196       gfx11_set_context_reg(R_028014_DB_HTILE_DATA_BASE, ds.u.gfx6.db_htile_data_base);
3197       gfx11_set_context_reg(R_02801C_DB_DEPTH_SIZE_XY, ds.db_depth_size);
3198       gfx11_set_context_reg(R_028040_DB_Z_INFO, ds.db_z_info);
3199       gfx11_set_context_reg(R_028044_DB_STENCIL_INFO, ds.db_stencil_info);
3200       gfx11_set_context_reg(R_028048_DB_Z_READ_BASE, ds.db_depth_base);
3201       gfx11_set_context_reg(R_02804C_DB_STENCIL_READ_BASE, ds.db_stencil_base);
3202       gfx11_set_context_reg(R_028050_DB_Z_WRITE_BASE, ds.db_depth_base);
3203       gfx11_set_context_reg(R_028054_DB_STENCIL_WRITE_BASE, ds.db_stencil_base);
3204       gfx11_set_context_reg(R_028068_DB_Z_READ_BASE_HI, ds.db_depth_base >> 32);
3205       gfx11_set_context_reg(R_02806C_DB_STENCIL_READ_BASE_HI, ds.db_stencil_base >> 32);
3206       gfx11_set_context_reg(R_028070_DB_Z_WRITE_BASE_HI, ds.db_depth_base >> 32);
3207       gfx11_set_context_reg(R_028074_DB_STENCIL_WRITE_BASE_HI, ds.db_stencil_base >> 32);
3208       gfx11_set_context_reg(R_028078_DB_HTILE_DATA_BASE_HI, ds.u.gfx6.db_htile_data_base >> 32);
3209       gfx11_set_context_reg(R_028028_DB_STENCIL_CLEAR, tex->stencil_clear_value[level]);
3210       gfx11_set_context_reg(R_02802C_DB_DEPTH_CLEAR, fui(tex->depth_clear_value[level]));
3211       gfx11_set_context_reg(R_028008_DB_DEPTH_VIEW, ds.db_depth_view);
3212       gfx11_set_context_reg(R_028ABC_DB_HTILE_SURFACE, ds.u.gfx6.db_htile_surface);
3213    } else if (sctx->framebuffer.dirty_zsbuf) {
3214       /* Gfx11+: DB_Z_INFO.NUM_SAMPLES should match the framebuffer samples if no Z/S is bound.
3215        * It determines the sample count for VRS, primitive-ordered pixel shading, and occlusion
3216        * queries.
3217        */
3218       gfx11_set_context_reg(R_028040_DB_Z_INFO,
3219                             S_028040_FORMAT(V_028040_Z_INVALID) |
3220                             S_028040_NUM_SAMPLES(sctx->framebuffer.log_samples));
3221       gfx11_set_context_reg(R_028044_DB_STENCIL_INFO, S_028044_FORMAT(V_028044_STENCIL_INVALID));
3222    }
3223 
3224    /* Framebuffer dimensions. */
3225    /* PA_SC_WINDOW_SCISSOR_TL is set to 0,0 in gfx*_init_gfx_preamble_state */
3226    gfx11_set_context_reg(R_028208_PA_SC_WINDOW_SCISSOR_BR,
3227                          S_028208_BR_X(state->width) | S_028208_BR_Y(state->height));
3228    gfx11_end_packed_context_regs();
3229 
3230    if (sctx->screen->dpbb_allowed &&
3231        sctx->screen->pbb_context_states_per_bin > 1)
3232       radeon_event_write(V_028A90_BREAK_BATCH);
3233 
3234    radeon_end();
3235 
3236    si_update_display_dcc_dirty(sctx);
3237 
3238    sctx->framebuffer.dirty_cbufs = 0;
3239    sctx->framebuffer.dirty_zsbuf = false;
3240 }
3241 
gfx12_emit_framebuffer_state(struct si_context * sctx,unsigned index)3242 static void gfx12_emit_framebuffer_state(struct si_context *sctx, unsigned index)
3243 {
3244    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
3245    struct pipe_framebuffer_state *state = &sctx->framebuffer.state;
3246    unsigned i, nr_cbufs = state->nr_cbufs;
3247    struct si_texture *tex = NULL;
3248    struct si_surface *cb = NULL;
3249    bool is_msaa_resolve = state->nr_cbufs == 2 &&
3250                           state->cbufs[0] && state->cbufs[0]->texture->nr_samples > 1 &&
3251                           state->cbufs[1] && state->cbufs[1]->texture->nr_samples <= 1;
3252 
3253    /* CB can't do MSAA resolve. */
3254    assert(!is_msaa_resolve);
3255 
3256    radeon_begin(cs);
3257    gfx12_begin_context_regs();
3258 
3259    /* Colorbuffers. */
3260    for (i = 0; i < nr_cbufs; i++) {
3261       if (!(sctx->framebuffer.dirty_cbufs & (1 << i)))
3262          continue;
3263 
3264       /* RB+ depth-only rendering. See the comment where we set rbplus_depth_only_opt for more
3265        * information.
3266        */
3267       if (i == 0 &&
3268           sctx->screen->info.rbplus_allowed &&
3269           !sctx->queued.named.blend->cb_target_mask) {
3270          gfx12_set_context_reg(R_028EC0_CB_COLOR0_INFO + i * 4,
3271                                S_028EC0_FORMAT(V_028C70_COLOR_32) |
3272                                S_028EC0_NUMBER_TYPE(V_028C70_NUMBER_FLOAT));
3273          continue;
3274       }
3275 
3276       cb = (struct si_surface *)state->cbufs[i];
3277       if (!cb) {
3278          gfx12_set_context_reg(R_028EC0_CB_COLOR0_INFO + i * 4,
3279                                S_028EC0_FORMAT(V_028C70_COLOR_INVALID));
3280          continue;
3281       }
3282 
3283       tex = (struct si_texture *)cb->base.texture;
3284       radeon_add_to_buffer_list(
3285          sctx, &sctx->gfx_cs, &tex->buffer, RADEON_USAGE_READWRITE | RADEON_USAGE_CB_NEEDS_IMPLICIT_SYNC |
3286          (tex->buffer.b.b.nr_samples > 1 ? RADEON_PRIO_COLOR_BUFFER_MSAA : RADEON_PRIO_COLOR_BUFFER));
3287 
3288       /* Compute mutable surface parameters. */
3289       const struct ac_mutable_cb_state mutable_cb_state = {
3290          .surf = &tex->surface,
3291          .cb = &cb->cb,
3292          .va = tex->buffer.gpu_address,
3293       };
3294       struct ac_cb_surface cb_surf;
3295 
3296       ac_set_mutable_cb_surface_fields(&sctx->screen->info, &mutable_cb_state, &cb_surf);
3297 
3298       gfx12_set_context_reg(R_028C60_CB_COLOR0_BASE + i * 0x24, cb_surf.cb_color_base);
3299       gfx12_set_context_reg(R_028C64_CB_COLOR0_VIEW + i * 0x24, cb_surf.cb_color_view);
3300       gfx12_set_context_reg(R_028C68_CB_COLOR0_VIEW2 + i * 0x24, cb_surf.cb_color_view2);
3301       gfx12_set_context_reg(R_028C6C_CB_COLOR0_ATTRIB + i * 0x24, cb_surf.cb_color_attrib);
3302       gfx12_set_context_reg(R_028C70_CB_COLOR0_FDCC_CONTROL + i * 0x24, cb_surf.cb_dcc_control);
3303       gfx12_set_context_reg(R_028C78_CB_COLOR0_ATTRIB2 + i * 0x24, cb_surf.cb_color_attrib2);
3304       gfx12_set_context_reg(R_028C7C_CB_COLOR0_ATTRIB3 + i * 0x24, cb_surf.cb_color_attrib3);
3305       gfx12_set_context_reg(R_028E40_CB_COLOR0_BASE_EXT + i * 4, cb_surf.cb_color_base >> 32);
3306       gfx12_set_context_reg(R_028EC0_CB_COLOR0_INFO + i * 4, cb_surf.cb_color_info);
3307    }
3308    /* Set unbound colorbuffers. */
3309    for (; i < 8; i++)
3310       if (sctx->framebuffer.dirty_cbufs & (1 << i))
3311          gfx12_set_context_reg(R_028EC0_CB_COLOR0_INFO + i * 4, 0);
3312 
3313    /* ZS buffer. */
3314    if (state->zsbuf && sctx->framebuffer.dirty_zsbuf) {
3315       struct si_surface *zb = (struct si_surface *)state->zsbuf;
3316       struct si_texture *tex = (struct si_texture *)zb->base.texture;
3317 
3318       radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, &tex->buffer,
3319                                 RADEON_USAGE_READWRITE | RADEON_USAGE_DB_NEEDS_IMPLICIT_SYNC |
3320                                 (zb->base.texture->nr_samples > 1 ? RADEON_PRIO_DEPTH_BUFFER_MSAA
3321                                                                   : RADEON_PRIO_DEPTH_BUFFER));
3322       gfx12_set_context_reg(R_028004_DB_DEPTH_VIEW, zb->ds.db_depth_view);
3323       gfx12_set_context_reg(R_028008_DB_DEPTH_VIEW1, zb->ds.u.gfx12.db_depth_view1);
3324       gfx12_set_context_reg(R_028014_DB_DEPTH_SIZE_XY, zb->ds.db_depth_size);
3325       gfx12_set_context_reg(R_028018_DB_Z_INFO, zb->ds.db_z_info);
3326       gfx12_set_context_reg(R_02801C_DB_STENCIL_INFO, zb->ds.db_stencil_info);
3327       gfx12_set_context_reg(R_028020_DB_Z_READ_BASE, zb->ds.db_depth_base);
3328       gfx12_set_context_reg(R_028024_DB_Z_READ_BASE_HI, zb->ds.db_depth_base >> 32);
3329       gfx12_set_context_reg(R_028028_DB_Z_WRITE_BASE, zb->ds.db_depth_base);
3330       gfx12_set_context_reg(R_02802C_DB_Z_WRITE_BASE_HI, zb->ds.db_depth_base >> 32);
3331       gfx12_set_context_reg(R_028030_DB_STENCIL_READ_BASE, zb->ds.db_stencil_base);
3332       gfx12_set_context_reg(R_028034_DB_STENCIL_READ_BASE_HI, zb->ds.db_stencil_base >> 32);
3333       gfx12_set_context_reg(R_028038_DB_STENCIL_WRITE_BASE, zb->ds.db_stencil_base);
3334       gfx12_set_context_reg(R_02803C_DB_STENCIL_WRITE_BASE_HI, zb->ds.db_stencil_base >> 32);
3335       gfx12_set_context_reg(R_028B94_PA_SC_HIZ_INFO, zb->ds.u.gfx12.hiz_info);
3336       gfx12_set_context_reg(R_028B98_PA_SC_HIS_INFO, zb->ds.u.gfx12.his_info);
3337 
3338       if (zb->ds.u.gfx12.hiz_info) {
3339          gfx12_set_context_reg(R_028B9C_PA_SC_HIZ_BASE, zb->ds.u.gfx12.hiz_base);
3340          gfx12_set_context_reg(R_028BA0_PA_SC_HIZ_BASE_EXT, zb->ds.u.gfx12.hiz_base >> 32);
3341          gfx12_set_context_reg(R_028BA4_PA_SC_HIZ_SIZE_XY, zb->ds.u.gfx12.hiz_size_xy);
3342       }
3343       if (zb->ds.u.gfx12.his_info) {
3344          gfx12_set_context_reg(R_028BA8_PA_SC_HIS_BASE, zb->ds.u.gfx12.his_base);
3345          gfx12_set_context_reg(R_028BAC_PA_SC_HIS_BASE_EXT, zb->ds.u.gfx12.his_base >> 32);
3346          gfx12_set_context_reg(R_028BB0_PA_SC_HIS_SIZE_XY, zb->ds.u.gfx12.his_size_xy);
3347       }
3348    } else if (sctx->framebuffer.dirty_zsbuf) {
3349       gfx12_set_context_reg(R_028018_DB_Z_INFO,
3350                             S_028040_FORMAT(V_028040_Z_INVALID) |
3351                             S_028040_NUM_SAMPLES(sctx->framebuffer.log_samples));
3352       gfx12_set_context_reg(R_02801C_DB_STENCIL_INFO,
3353                             S_028044_FORMAT(V_028044_STENCIL_INVALID)|
3354                             S_028044_TILE_STENCIL_DISABLE(1));
3355       gfx12_set_context_reg(R_028B94_PA_SC_HIZ_INFO, S_028B94_SURFACE_ENABLE(0));
3356       gfx12_set_context_reg(R_028B98_PA_SC_HIS_INFO, S_028B98_SURFACE_ENABLE(0));
3357    }
3358 
3359    /* Framebuffer dimensions. */
3360    /* PA_SC_WINDOW_SCISSOR_TL is set in gfx12_init_gfx_preamble_state */
3361    gfx12_set_context_reg(R_028208_PA_SC_WINDOW_SCISSOR_BR,
3362                          S_028208_BR_X(state->width - 1) |    /* inclusive */
3363                          S_028208_BR_Y(state->height - 1));   /* inclusive */
3364    gfx12_end_context_regs();
3365 
3366    if (sctx->screen->dpbb_allowed &&
3367        sctx->screen->pbb_context_states_per_bin > 1)
3368       radeon_event_write(V_028A90_BREAK_BATCH);
3369 
3370    radeon_end();
3371 
3372    sctx->framebuffer.dirty_cbufs = 0;
3373    sctx->framebuffer.dirty_zsbuf = false;
3374 }
3375 
si_out_of_order_rasterization(struct si_context * sctx)3376 static bool si_out_of_order_rasterization(struct si_context *sctx)
3377 {
3378    struct si_state_blend *blend = sctx->queued.named.blend;
3379    struct si_state_dsa *dsa = sctx->queued.named.dsa;
3380 
3381    if (!sctx->screen->info.has_out_of_order_rast)
3382       return false;
3383 
3384    unsigned colormask = sctx->framebuffer.colorbuf_enabled_4bit;
3385 
3386    colormask &= blend->cb_target_enabled_4bit;
3387 
3388    /* Conservative: No logic op. */
3389    if (colormask && blend->logicop_enable)
3390       return false;
3391 
3392    struct si_dsa_order_invariance dsa_order_invariant = {.zs = true,
3393                                                          .pass_set = true};
3394 
3395    if (sctx->framebuffer.state.zsbuf) {
3396       struct si_texture *zstex = (struct si_texture *)sctx->framebuffer.state.zsbuf->texture;
3397       bool has_stencil = zstex->surface.has_stencil;
3398       dsa_order_invariant = dsa->order_invariance[has_stencil];
3399       if (!dsa_order_invariant.zs)
3400          return false;
3401 
3402       /* The set of PS invocations is always order invariant,
3403        * except when early Z/S tests are requested. */
3404       if (sctx->shader.ps.cso && sctx->shader.ps.cso->info.base.writes_memory &&
3405           sctx->shader.ps.cso->info.base.fs.early_fragment_tests &&
3406           !dsa_order_invariant.pass_set)
3407          return false;
3408 
3409       if (sctx->occlusion_query_mode == SI_OCCLUSION_QUERY_MODE_PRECISE_INTEGER &&
3410           !dsa_order_invariant.pass_set)
3411          return false;
3412    }
3413 
3414    if (!colormask)
3415       return true;
3416 
3417    unsigned blendmask = colormask & blend->blend_enable_4bit;
3418 
3419    if (blendmask) {
3420       /* Only commutative blending. */
3421       if (blendmask & ~blend->commutative_4bit)
3422          return false;
3423 
3424       if (!dsa_order_invariant.pass_set)
3425          return false;
3426    }
3427 
3428    if (colormask & ~blendmask)
3429       return false;
3430 
3431    return true;
3432 }
3433 
si_emit_msaa_config(struct si_context * sctx,unsigned index)3434 static void si_emit_msaa_config(struct si_context *sctx, unsigned index)
3435 {
3436    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
3437    unsigned num_tile_pipes = sctx->screen->info.num_tile_pipes;
3438    /* 33% faster rendering to linear color buffers */
3439    bool dst_is_linear = sctx->framebuffer.any_dst_linear;
3440    bool out_of_order_rast = si_out_of_order_rasterization(sctx);
3441    unsigned sc_mode_cntl_1 =
3442       S_028A4C_WALK_SIZE(dst_is_linear) | S_028A4C_WALK_FENCE_ENABLE(!dst_is_linear) |
3443       S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) |
3444       S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(out_of_order_rast) |
3445       S_028A4C_OUT_OF_ORDER_WATER_MARK(sctx->gfx_level >= GFX12 ? 0 : 0x7) |
3446       /* This should also be 0 when the VRS image is enabled. */
3447       S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(!sctx->framebuffer.has_hiz_his) |
3448       /* always 1: */
3449       S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
3450       S_028A4C_TILE_WALK_ORDER_ENABLE(1) | S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
3451       S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) | S_028A4C_FORCE_EOV_REZ_ENABLE(1);
3452    unsigned db_eqaa = S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
3453                       S_028804_INCOHERENT_EQAA_READS(sctx->gfx_level < GFX12) |
3454                       S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
3455    unsigned coverage_samples, z_samples;
3456    struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
3457 
3458    /* S: Coverage samples (up to 16x):
3459     * - Scan conversion samples (PA_SC_AA_CONFIG.MSAA_NUM_SAMPLES)
3460     * - CB FMASK samples (CB_COLORi_ATTRIB.NUM_SAMPLES)
3461     *
3462     * Z: Z/S samples (up to 8x, must be <= coverage samples and >= color samples):
3463     * - Value seen by DB (DB_Z_INFO.NUM_SAMPLES)
3464     * - Value seen by CB, must be correct even if Z/S is unbound (DB_EQAA.MAX_ANCHOR_SAMPLES)
3465     * # Missing samples are derived from Z planes if Z is compressed (up to 16x quality), or
3466     * # from the closest defined sample if Z is uncompressed (same quality as the number of
3467     * # Z samples).
3468     *
3469     * F: Color samples (up to 8x, must be <= coverage samples):
3470     * - CB color samples (CB_COLORi_ATTRIB.NUM_FRAGMENTS)
3471     * - PS iter samples (DB_EQAA.PS_ITER_SAMPLES)
3472     *
3473     * Can be anything between coverage and color samples:
3474     * - SampleMaskIn samples (PA_SC_AA_CONFIG.MSAA_EXPOSED_SAMPLES)
3475     * - SampleMaskOut samples (DB_EQAA.MASK_EXPORT_NUM_SAMPLES)
3476     * - Alpha-to-coverage samples (DB_EQAA.ALPHA_TO_MASK_NUM_SAMPLES)
3477     * - Occlusion query samples (DB_COUNT_CONTROL.SAMPLE_RATE)
3478     * # All are currently set the same as coverage samples.
3479     *
3480     * If color samples < coverage samples, FMASK has a higher bpp to store an "unknown"
3481     * flag for undefined color samples. A shader-based resolve must handle unknowns
3482     * or mask them out with AND. Unknowns can also be guessed from neighbors via
3483     * an edge-detect shader-based resolve, which is required to make "color samples = 1"
3484     * useful. The CB resolve always drops unknowns.
3485     *
3486     * Sensible AA configurations:
3487     *   EQAA 16s 8z 8f - might look the same as 16x MSAA if Z is compressed
3488     *   EQAA 16s 8z 4f - might look the same as 16x MSAA if Z is compressed
3489     *   EQAA 16s 4z 4f - might look the same as 16x MSAA if Z is compressed
3490     *   EQAA  8s 8z 8f = 8x MSAA
3491     *   EQAA  8s 8z 4f - might look the same as 8x MSAA
3492     *   EQAA  8s 8z 2f - might look the same as 8x MSAA with low-density geometry
3493     *   EQAA  8s 4z 4f - might look the same as 8x MSAA if Z is compressed
3494     *   EQAA  8s 4z 2f - might look the same as 8x MSAA with low-density geometry if Z is compressed
3495     *   EQAA  4s 4z 4f = 4x MSAA
3496     *   EQAA  4s 4z 2f - might look the same as 4x MSAA with low-density geometry
3497     *   EQAA  2s 2z 2f = 2x MSAA
3498     */
3499    coverage_samples = si_get_num_coverage_samples(sctx);
3500 
3501    /* DCC_DECOMPRESS and ELIMINATE_FAST_CLEAR require MSAA_NUM_SAMPLES=0. */
3502    if (sctx->gfx_level >= GFX11 && sctx->gfx11_force_msaa_num_samples_zero)
3503       coverage_samples = 1;
3504 
3505    /* The DX10 diamond test is not required by GL and decreases line rasterization
3506     * performance, so don't use it.
3507     */
3508    unsigned sc_line_cntl = 0;
3509    unsigned sc_aa_config = 0;
3510 
3511    if (coverage_samples > 1 && (rs->multisample_enable ||
3512                                 sctx->smoothing_enabled)) {
3513       unsigned log_samples = util_logbase2(coverage_samples);
3514 
3515       sc_line_cntl |= S_028BDC_EXPAND_LINE_WIDTH(1) |
3516                       S_028BDC_PERPENDICULAR_ENDCAP_ENA(rs->perpendicular_end_caps) |
3517                       S_028BDC_EXTRA_DX_DY_PRECISION(rs->perpendicular_end_caps &&
3518                                                      (sctx->family == CHIP_VEGA20 ||
3519                                                       sctx->gfx_level >= GFX10));
3520       sc_aa_config = S_028BE0_MSAA_NUM_SAMPLES(log_samples) |
3521                      S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples);
3522 
3523       if (sctx->gfx_level < GFX12) {
3524          sc_aa_config |= S_028BE0_MAX_SAMPLE_DIST(si_msaa_max_distance[log_samples]) |
3525                          S_028BE0_COVERED_CENTROID_IS_CENTER(sctx->gfx_level >= GFX10_3);
3526       }
3527    }
3528 
3529    if (sctx->framebuffer.nr_samples > 1 ||
3530        sctx->smoothing_enabled) {
3531       if (sctx->framebuffer.state.zsbuf) {
3532          z_samples = sctx->framebuffer.state.zsbuf->texture->nr_samples;
3533          z_samples = MAX2(1, z_samples);
3534       } else {
3535          z_samples = coverage_samples;
3536       }
3537       unsigned log_samples = util_logbase2(coverage_samples);
3538       unsigned log_z_samples = util_logbase2(z_samples);
3539       unsigned ps_iter_samples = si_get_ps_iter_samples(sctx);
3540       unsigned log_ps_iter_samples = util_logbase2(ps_iter_samples);
3541       if (sctx->framebuffer.nr_samples > 1) {
3542          if (sctx->gfx_level >= GFX12) {
3543             sc_aa_config |= S_028BE0_PS_ITER_SAMPLES(log_ps_iter_samples);
3544             db_eqaa |= S_028078_MASK_EXPORT_NUM_SAMPLES(log_samples) |
3545                        S_028078_ALPHA_TO_MASK_NUM_SAMPLES(log_samples);
3546          } else {
3547             db_eqaa |= S_028804_MAX_ANCHOR_SAMPLES(log_z_samples) |
3548                        S_028804_PS_ITER_SAMPLES(log_ps_iter_samples) |
3549                        S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples) |
3550                        S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples);
3551          }
3552          sc_mode_cntl_1 |= S_028A4C_PS_ITER_SAMPLE(ps_iter_samples > 1);
3553       } else if (sctx->smoothing_enabled) {
3554          db_eqaa |= S_028804_OVERRASTERIZATION_AMOUNT(log_samples);
3555       }
3556    }
3557 
3558    if (sctx->gfx_level >= GFX12) {
3559       radeon_begin(cs);
3560       gfx12_begin_context_regs();
3561       gfx12_opt_set_context_reg(R_028BDC_PA_SC_LINE_CNTL, SI_TRACKED_PA_SC_LINE_CNTL,
3562                                 sc_line_cntl);
3563       gfx12_opt_set_context_reg(R_028BE0_PA_SC_AA_CONFIG, SI_TRACKED_PA_SC_AA_CONFIG,
3564                                 sc_aa_config);
3565       gfx12_opt_set_context_reg(R_028078_DB_EQAA, SI_TRACKED_DB_EQAA, db_eqaa);
3566       gfx12_opt_set_context_reg(R_028A4C_PA_SC_MODE_CNTL_1, SI_TRACKED_PA_SC_MODE_CNTL_1,
3567                                 sc_mode_cntl_1);
3568       gfx12_end_context_regs();
3569       radeon_end(); /* don't track context rolls on GFX12 */
3570    } else if (sctx->screen->info.has_set_context_pairs_packed) {
3571       radeon_begin(cs);
3572       gfx11_begin_packed_context_regs();
3573       gfx11_opt_set_context_reg(R_028BDC_PA_SC_LINE_CNTL, SI_TRACKED_PA_SC_LINE_CNTL,
3574                                 sc_line_cntl);
3575       gfx11_opt_set_context_reg(R_028BE0_PA_SC_AA_CONFIG, SI_TRACKED_PA_SC_AA_CONFIG,
3576                                 sc_aa_config);
3577       gfx11_opt_set_context_reg(R_028804_DB_EQAA, SI_TRACKED_DB_EQAA, db_eqaa);
3578       gfx11_opt_set_context_reg(R_028A4C_PA_SC_MODE_CNTL_1, SI_TRACKED_PA_SC_MODE_CNTL_1,
3579                                 sc_mode_cntl_1);
3580       gfx11_end_packed_context_regs();
3581       radeon_end(); /* don't track context rolls on GFX11 */
3582    } else {
3583       radeon_begin(cs);
3584       radeon_opt_set_context_reg2(R_028BDC_PA_SC_LINE_CNTL, SI_TRACKED_PA_SC_LINE_CNTL,
3585                                   sc_line_cntl, sc_aa_config);
3586       radeon_opt_set_context_reg(R_028804_DB_EQAA, SI_TRACKED_DB_EQAA, db_eqaa);
3587       radeon_opt_set_context_reg(R_028A4C_PA_SC_MODE_CNTL_1, SI_TRACKED_PA_SC_MODE_CNTL_1,
3588                                  sc_mode_cntl_1);
3589       radeon_end_update_context_roll();
3590    }
3591 }
3592 
si_update_ps_iter_samples(struct si_context * sctx)3593 void si_update_ps_iter_samples(struct si_context *sctx)
3594 {
3595    if (sctx->framebuffer.nr_samples > 1)
3596       si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
3597    if (sctx->screen->dpbb_allowed)
3598       si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
3599 }
3600 
si_set_min_samples(struct pipe_context * ctx,unsigned min_samples)3601 static void si_set_min_samples(struct pipe_context *ctx, unsigned min_samples)
3602 {
3603    struct si_context *sctx = (struct si_context *)ctx;
3604 
3605    /* The hardware can only do sample shading with 2^n samples. */
3606    min_samples = util_next_power_of_two(min_samples);
3607 
3608    if (sctx->ps_iter_samples == min_samples)
3609       return;
3610 
3611    sctx->ps_iter_samples = min_samples;
3612 
3613    si_ps_key_update_sample_shading(sctx);
3614    si_ps_key_update_framebuffer_rasterizer_sample_shading(sctx);
3615    sctx->do_update_shaders = true;
3616 
3617    si_update_ps_iter_samples(sctx);
3618 }
3619 
3620 /*
3621  * Samplers
3622  */
3623 
3624 /**
3625  * Build the sampler view descriptor for a buffer texture.
3626  * @param state 256-bit descriptor; only the high 128 bits are filled in
3627  */
si_make_buffer_descriptor(struct si_screen * screen,struct si_resource * buf,enum pipe_format format,unsigned offset,unsigned num_elements,uint32_t * state)3628 void si_make_buffer_descriptor(struct si_screen *screen, struct si_resource *buf,
3629                                enum pipe_format format, unsigned offset, unsigned num_elements,
3630                                uint32_t *state)
3631 {
3632    const struct util_format_description *desc;
3633    unsigned stride;
3634    unsigned num_records;
3635 
3636    desc = util_format_description(format);
3637    stride = desc->block.bits / 8;
3638 
3639    num_records = num_elements;
3640    num_records = MIN2(num_records, (buf->b.b.width0 - offset) / stride);
3641 
3642    /* The NUM_RECORDS field has a different meaning depending on the chip,
3643     * instruction type, STRIDE, and SWIZZLE_ENABLE.
3644     *
3645     * GFX6-7,10:
3646     * - If STRIDE == 0, it's in byte units.
3647     * - If STRIDE != 0, it's in units of STRIDE, used with inst.IDXEN.
3648     *
3649     * GFX8:
3650     * - For SMEM and STRIDE == 0, it's in byte units.
3651     * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3652     * - For VMEM and STRIDE == 0 or SWIZZLE_ENABLE == 0, it's in byte units.
3653     * - For VMEM and STRIDE != 0 and SWIZZLE_ENABLE == 1, it's in units of STRIDE.
3654     * NOTE: There is incompatibility between VMEM and SMEM opcodes due to SWIZZLE_-
3655     *       ENABLE. The workaround is to set STRIDE = 0 if SWIZZLE_ENABLE == 0 when
3656     *       using SMEM. This can be done in the shader by clearing STRIDE with s_and.
3657     *       That way the same descriptor can be used by both SMEM and VMEM.
3658     *
3659     * GFX9:
3660     * - For SMEM and STRIDE == 0, it's in byte units.
3661     * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3662     * - For VMEM and inst.IDXEN == 0 or STRIDE == 0, it's in byte units.
3663     * - For VMEM and inst.IDXEN == 1 and STRIDE != 0, it's in units of STRIDE.
3664     */
3665    if (screen->info.gfx_level == GFX8)
3666       num_records *= stride;
3667 
3668    const struct ac_buffer_state buffer_state = {
3669       .size = num_records,
3670       .format = format,
3671       .swizzle =
3672          {
3673             desc->swizzle[0],
3674             desc->swizzle[1],
3675             desc->swizzle[2],
3676             desc->swizzle[3],
3677          },
3678       .stride = stride,
3679       .gfx10_oob_select = V_008F0C_OOB_SELECT_STRUCTURED_WITH_OFFSET,
3680    };
3681 
3682    ac_build_buffer_descriptor(screen->info.gfx_level, &buffer_state, &state[4]);
3683 }
3684 
3685 /**
3686  * Translate the parameters to an image descriptor for CDNA image emulation.
3687  * In this function, we choose our own image descriptor format because we emulate image opcodes
3688  * using buffer opcodes.
3689  */
cdna_emu_make_image_descriptor(struct si_screen * screen,struct si_texture * tex,bool sampler,enum pipe_texture_target target,enum pipe_format pipe_format,const unsigned char state_swizzle[4],unsigned first_level,unsigned last_level,unsigned first_layer,unsigned last_layer,unsigned width,unsigned height,unsigned depth,uint32_t * state,uint32_t * fmask_state)3690 static void cdna_emu_make_image_descriptor(struct si_screen *screen, struct si_texture *tex,
3691                                            bool sampler, enum pipe_texture_target target,
3692                                            enum pipe_format pipe_format,
3693                                            const unsigned char state_swizzle[4], unsigned first_level,
3694                                            unsigned last_level, unsigned first_layer,
3695                                            unsigned last_layer, unsigned width, unsigned height,
3696                                            unsigned depth, uint32_t *state, uint32_t *fmask_state)
3697 {
3698    const struct util_format_description *desc = util_format_description(pipe_format);
3699 
3700    /* We don't need support these. We only need enough to support VAAPI and OpenMAX. */
3701    if (target == PIPE_TEXTURE_CUBE ||
3702        target == PIPE_TEXTURE_CUBE_ARRAY ||
3703        tex->buffer.b.b.last_level > 0 ||
3704        tex->buffer.b.b.nr_samples >= 2 ||
3705        desc->colorspace != UTIL_FORMAT_COLORSPACE_RGB ||
3706        desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED ||
3707        util_format_is_compressed(pipe_format)) {
3708       assert(!"unexpected texture type");
3709       memset(state, 0, 8 * 4);
3710       return;
3711    }
3712 
3713    /* Adjust the image parameters according to the texture type. */
3714    switch (target) {
3715    case PIPE_TEXTURE_1D:
3716       height = 1;
3717       FALLTHROUGH;
3718    case PIPE_TEXTURE_2D:
3719    case PIPE_TEXTURE_RECT:
3720       depth = 1;
3721       break;
3722 
3723    case PIPE_TEXTURE_1D_ARRAY:
3724       height = 1;
3725       FALLTHROUGH;
3726    case PIPE_TEXTURE_2D_ARRAY:
3727       first_layer = MIN2(first_layer, tex->buffer.b.b.array_size - 1);
3728       last_layer = MIN2(last_layer, tex->buffer.b.b.array_size - 1);
3729       last_layer = MAX2(last_layer, first_layer);
3730       depth = last_layer - first_layer + 1;
3731       break;
3732 
3733    case PIPE_TEXTURE_3D:
3734       first_layer = 0;
3735       break;
3736 
3737    default:
3738       unreachable("invalid texture target");
3739    }
3740 
3741    unsigned stride = desc->block.bits / 8;
3742    uint64_t num_records = tex->surface.surf_size / stride;
3743    assert(num_records <= UINT32_MAX);
3744 
3745    /* Prepare the format fields. */
3746    unsigned char swizzle[4];
3747    util_format_compose_swizzles(desc->swizzle, state_swizzle, swizzle);
3748 
3749    /* Buffer descriptor */
3750    const struct ac_buffer_state buffer_state = {
3751       .size = num_records,
3752       .format = pipe_format,
3753       .swizzle =
3754          {
3755             desc->swizzle[0],
3756             desc->swizzle[1],
3757             desc->swizzle[2],
3758             desc->swizzle[3],
3759          },
3760       .stride = stride,
3761       .gfx10_oob_select = V_008F0C_OOB_SELECT_STRUCTURED_WITH_OFFSET,
3762    };
3763 
3764    ac_build_buffer_descriptor(screen->info.gfx_level, &buffer_state, &state[0]);
3765 
3766    /* Additional fields used by image opcode emulation. */
3767    state[4] = width | (height << 16);
3768    state[5] = depth | (first_layer << 16);
3769    state[6] = tex->surface.u.gfx9.surf_pitch;
3770    state[7] = (uint32_t)tex->surface.u.gfx9.surf_pitch * tex->surface.u.gfx9.surf_height;
3771 }
3772 
3773 /**
3774  * Build the sampler view descriptor for a texture.
3775  */
gfx10_make_texture_descriptor(struct si_screen * screen,struct si_texture * tex,bool sampler,enum pipe_texture_target target,enum pipe_format pipe_format,const unsigned char state_swizzle[4],unsigned first_level,unsigned last_level,unsigned first_layer,unsigned last_layer,unsigned width,unsigned height,unsigned depth,bool get_bo_metadata,uint32_t * state,uint32_t * fmask_state)3776 static void gfx10_make_texture_descriptor(
3777    struct si_screen *screen, struct si_texture *tex, bool sampler, enum pipe_texture_target target,
3778    enum pipe_format pipe_format, const unsigned char state_swizzle[4], unsigned first_level,
3779    unsigned last_level, unsigned first_layer, unsigned last_layer, unsigned width, unsigned height,
3780    unsigned depth, bool get_bo_metadata, uint32_t *state, uint32_t *fmask_state)
3781 {
3782    if (!screen->info.has_image_opcodes && !get_bo_metadata) {
3783       cdna_emu_make_image_descriptor(screen, tex, sampler, target, pipe_format, state_swizzle,
3784                                      first_level, last_level, first_layer, last_layer, width,
3785                                      height, depth, state, fmask_state);
3786       return;
3787    }
3788 
3789    struct pipe_resource *res = &tex->buffer.b.b;
3790    const struct util_format_description *desc;
3791    unsigned char swizzle[4];
3792    unsigned type;
3793 
3794    desc = util_format_description(pipe_format);
3795 
3796    if (desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
3797       const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0};
3798       const unsigned char swizzle_yyyy[4] = {1, 1, 1, 1};
3799       const unsigned char swizzle_wwww[4] = {3, 3, 3, 3};
3800 
3801       switch (pipe_format) {
3802       case PIPE_FORMAT_S8_UINT_Z24_UNORM:
3803       case PIPE_FORMAT_X32_S8X24_UINT:
3804       case PIPE_FORMAT_X8Z24_UNORM:
3805          util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
3806          break;
3807       case PIPE_FORMAT_X24S8_UINT:
3808          /*
3809           * X24S8 is implemented as an 8_8_8_8 data format, to
3810           * fix texture gathers. This affects at least
3811           * GL45-CTS.texture_cube_map_array.sampling on GFX8.
3812           */
3813          util_format_compose_swizzles(swizzle_wwww, state_swizzle, swizzle);
3814          break;
3815       default:
3816          util_format_compose_swizzles(swizzle_xxxx, state_swizzle, swizzle);
3817       }
3818    } else {
3819       util_format_compose_swizzles(desc->swizzle, state_swizzle, swizzle);
3820    }
3821 
3822    if (!sampler && (res->target == PIPE_TEXTURE_CUBE || res->target == PIPE_TEXTURE_CUBE_ARRAY)) {
3823       /* For the purpose of shader images, treat cube maps as 2D
3824        * arrays.
3825        */
3826       type = V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
3827    } else {
3828       type = si_tex_dim(screen, tex, target, res->nr_samples);
3829    }
3830 
3831    if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
3832       height = 1;
3833       depth = res->array_size;
3834    } else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY || type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) {
3835       if (sampler || res->target != PIPE_TEXTURE_3D)
3836          depth = res->array_size;
3837    } else if (type == V_008F1C_SQ_RSRC_IMG_CUBE)
3838       depth = res->array_size / 6;
3839 
3840    const struct ac_texture_state tex_state = {
3841       .surf = &tex->surface,
3842       .format = pipe_format,
3843       .img_format = res->format,
3844       .width = width,
3845       .height = height,
3846       .depth =  (type == V_008F1C_SQ_RSRC_IMG_3D && sampler) ? depth - 1 : last_layer,
3847       .type = type,
3848       .swizzle =
3849          {
3850             swizzle[0],
3851             swizzle[1],
3852             swizzle[2],
3853             swizzle[3],
3854          },
3855       .num_samples = res->nr_samples,
3856       .num_storage_samples = res->nr_storage_samples,
3857       .first_level = first_level,
3858       .last_level = last_level,
3859       .num_levels = res->last_level + 1,
3860       .first_layer = first_layer,
3861       .last_layer = last_layer,
3862       .gfx10 = {
3863          .uav3d = !!(type == V_008F1C_SQ_RSRC_IMG_3D && !sampler),
3864          .upgraded_depth = tex->upgraded_depth,
3865       },
3866       .dcc_enabled = vi_dcc_enabled(tex, first_level),
3867    };
3868 
3869    ac_build_texture_descriptor(&screen->info, &tex_state, &state[0]);
3870 
3871    /* Initialize the sampler view for FMASK. */
3872    if (tex->surface.fmask_offset) {
3873       const struct ac_fmask_state ac_state = {
3874          .surf = &tex->surface,
3875          .va = tex->buffer.gpu_address,
3876          .width = width,
3877          .height = height,
3878          .depth = depth,
3879          .type = si_tex_dim(screen, tex, target, 0),
3880          .first_layer = first_layer,
3881          .last_layer = last_layer,
3882          .num_samples = res->nr_samples,
3883          .num_storage_samples = res->nr_storage_samples,
3884       };
3885 
3886       ac_build_fmask_descriptor(screen->info.gfx_level, &ac_state, &fmask_state[0]);
3887    }
3888 }
3889 
3890 /**
3891  * Build the sampler view descriptor for a texture (SI-GFX9).
3892  */
si_make_texture_descriptor(struct si_screen * screen,struct si_texture * tex,bool sampler,enum pipe_texture_target target,enum pipe_format pipe_format,const unsigned char state_swizzle[4],unsigned first_level,unsigned last_level,unsigned first_layer,unsigned last_layer,unsigned width,unsigned height,unsigned depth,bool get_bo_metadata,uint32_t * state,uint32_t * fmask_state)3893 static void si_make_texture_descriptor(struct si_screen *screen, struct si_texture *tex,
3894                                        bool sampler, enum pipe_texture_target target,
3895                                        enum pipe_format pipe_format,
3896                                        const unsigned char state_swizzle[4], unsigned first_level,
3897                                        unsigned last_level, unsigned first_layer,
3898                                        unsigned last_layer, unsigned width, unsigned height,
3899                                        unsigned depth, bool get_bo_metadata,
3900                                        uint32_t *state, uint32_t *fmask_state)
3901 {
3902    if (!screen->info.has_image_opcodes && !get_bo_metadata) {
3903       cdna_emu_make_image_descriptor(screen, tex, sampler, target, pipe_format, state_swizzle,
3904                                      first_level, last_level, first_layer, last_layer, width,
3905                                      height, depth, state, fmask_state);
3906       return;
3907    }
3908 
3909    struct pipe_resource *res = &tex->buffer.b.b;
3910    const struct util_format_description *desc;
3911    unsigned char swizzle[4];
3912    unsigned type, num_samples;
3913 
3914    desc = util_format_description(pipe_format);
3915 
3916    num_samples = desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS ? MAX2(1, res->nr_samples)
3917                                                                : MAX2(1, res->nr_storage_samples);
3918 
3919    if (desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
3920       const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0};
3921       const unsigned char swizzle_yyyy[4] = {1, 1, 1, 1};
3922       const unsigned char swizzle_wwww[4] = {3, 3, 3, 3};
3923 
3924       switch (pipe_format) {
3925       case PIPE_FORMAT_S8_UINT_Z24_UNORM:
3926       case PIPE_FORMAT_X32_S8X24_UINT:
3927       case PIPE_FORMAT_X8Z24_UNORM:
3928          util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
3929          break;
3930       case PIPE_FORMAT_X24S8_UINT:
3931          /*
3932           * X24S8 is implemented as an 8_8_8_8 data format, to
3933           * fix texture gathers. This affects at least
3934           * GL45-CTS.texture_cube_map_array.sampling on GFX8.
3935           */
3936          if (screen->info.gfx_level <= GFX8)
3937             util_format_compose_swizzles(swizzle_wwww, state_swizzle, swizzle);
3938          else
3939             util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
3940          break;
3941       default:
3942          util_format_compose_swizzles(swizzle_xxxx, state_swizzle, swizzle);
3943       }
3944    } else {
3945       util_format_compose_swizzles(desc->swizzle, state_swizzle, swizzle);
3946    }
3947 
3948    if (!sampler && (res->target == PIPE_TEXTURE_CUBE || res->target == PIPE_TEXTURE_CUBE_ARRAY ||
3949                     (screen->info.gfx_level <= GFX8 && res->target == PIPE_TEXTURE_3D))) {
3950       /* For the purpose of shader images, treat cube maps and 3D
3951        * textures as 2D arrays. For 3D textures, the address
3952        * calculations for mipmaps are different, so we rely on the
3953        * caller to effectively disable mipmaps.
3954        */
3955       type = V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
3956 
3957       assert(res->target != PIPE_TEXTURE_3D || (first_level == 0 && last_level == 0));
3958    } else {
3959       type = si_tex_dim(screen, tex, target, num_samples);
3960    }
3961 
3962    if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
3963       height = 1;
3964       depth = res->array_size;
3965    } else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY || type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) {
3966       if (sampler || res->target != PIPE_TEXTURE_3D)
3967          depth = res->array_size;
3968    } else if (type == V_008F1C_SQ_RSRC_IMG_CUBE)
3969       depth = res->array_size / 6;
3970 
3971    const struct ac_texture_state tex_state = {
3972       .surf = &tex->surface,
3973       .format = pipe_format,
3974       .img_format = res->format,
3975       .width = width,
3976       .height = height,
3977       .depth = depth,
3978       .type = type,
3979       .swizzle =
3980          {
3981             swizzle[0],
3982             swizzle[1],
3983             swizzle[2],
3984             swizzle[3],
3985          },
3986       .num_samples = res->nr_samples,
3987       .num_storage_samples = res->nr_storage_samples,
3988       .first_level = first_level,
3989       .last_level = last_level,
3990       .num_levels = res->last_level + 1,
3991       .first_layer = first_layer,
3992       .last_layer = last_layer,
3993       .dcc_enabled = vi_dcc_enabled(tex, first_level),
3994       .tc_compat_htile_enabled = true,
3995    };
3996 
3997    ac_build_texture_descriptor(&screen->info, &tex_state, &state[0]);
3998 
3999    /* Initialize the sampler view for FMASK. */
4000    if (tex->surface.fmask_offset) {
4001       const struct ac_fmask_state ac_state = {
4002          .surf = &tex->surface,
4003          .va = tex->buffer.gpu_address,
4004          .width = width,
4005          .height = height,
4006          .depth = depth,
4007          .type = si_tex_dim(screen, tex, target, 0),
4008          .first_layer = first_layer,
4009          .last_layer = last_layer,
4010          .num_samples = res->nr_samples,
4011          .num_storage_samples = res->nr_storage_samples,
4012       };
4013 
4014       ac_build_fmask_descriptor(screen->info.gfx_level, &ac_state, &fmask_state[0]);
4015    }
4016 }
4017 
4018 /**
4019  * Create a sampler view.
4020  *
4021  * @param ctx      context
4022  * @param texture  texture
4023  * @param state    sampler view template
4024  */
si_create_sampler_view(struct pipe_context * ctx,struct pipe_resource * texture,const struct pipe_sampler_view * state)4025 static struct pipe_sampler_view *si_create_sampler_view(struct pipe_context *ctx,
4026                                                         struct pipe_resource *texture,
4027                                                         const struct pipe_sampler_view *state)
4028 {
4029    struct si_context *sctx = (struct si_context *)ctx;
4030    struct si_sampler_view *view = CALLOC_STRUCT_CL(si_sampler_view);
4031    struct si_texture *tex = (struct si_texture *)texture;
4032    unsigned char state_swizzle[4];
4033    unsigned last_layer = state->u.tex.last_layer;
4034    enum pipe_format pipe_format;
4035    const struct legacy_surf_level *surflevel;
4036 
4037    if (!view)
4038       return NULL;
4039 
4040    /* initialize base object */
4041    view->base = *state;
4042    view->base.texture = NULL;
4043    view->base.reference.count = 1;
4044    view->base.context = ctx;
4045 
4046    assert(texture);
4047    pipe_resource_reference(&view->base.texture, texture);
4048 
4049    if (state->format == PIPE_FORMAT_X24S8_UINT || state->format == PIPE_FORMAT_S8X24_UINT ||
4050        state->format == PIPE_FORMAT_X32_S8X24_UINT || state->format == PIPE_FORMAT_S8_UINT)
4051       view->is_stencil_sampler = true;
4052 
4053    /* Buffer resource. */
4054    if (texture->target == PIPE_BUFFER) {
4055       uint32_t elements = si_clamp_texture_texel_count(sctx->screen->max_texel_buffer_elements,
4056                                                        state->format, state->u.buf.size);
4057 
4058       si_make_buffer_descriptor(sctx->screen, si_resource(texture), state->format,
4059                                 state->u.buf.offset, elements, view->state);
4060       return &view->base;
4061    }
4062 
4063    state_swizzle[0] = state->swizzle_r;
4064    state_swizzle[1] = state->swizzle_g;
4065    state_swizzle[2] = state->swizzle_b;
4066    state_swizzle[3] = state->swizzle_a;
4067 
4068    /* This is not needed if gallium frontends set last_layer correctly. */
4069    if (state->target == PIPE_TEXTURE_1D || state->target == PIPE_TEXTURE_2D ||
4070        state->target == PIPE_TEXTURE_RECT || state->target == PIPE_TEXTURE_CUBE)
4071       last_layer = state->u.tex.first_layer;
4072 
4073    /* Texturing with separate depth and stencil. */
4074    pipe_format = state->format;
4075 
4076    /* Depth/stencil texturing sometimes needs separate texture. */
4077    if (tex->is_depth && !si_can_sample_zs(tex, view->is_stencil_sampler)) {
4078       if (!tex->flushed_depth_texture && !si_init_flushed_depth_texture(ctx, texture)) {
4079          pipe_resource_reference(&view->base.texture, NULL);
4080          FREE(view);
4081          return NULL;
4082       }
4083 
4084       assert(tex->flushed_depth_texture);
4085 
4086       /* Override format for the case where the flushed texture
4087        * contains only Z or only S.
4088        */
4089       if (tex->flushed_depth_texture->buffer.b.b.format != tex->buffer.b.b.format)
4090          pipe_format = tex->flushed_depth_texture->buffer.b.b.format;
4091 
4092       tex = tex->flushed_depth_texture;
4093    }
4094 
4095    surflevel = tex->surface.u.legacy.level;
4096 
4097    if (tex->db_compatible) {
4098       if (!view->is_stencil_sampler)
4099          pipe_format = tex->db_render_format;
4100 
4101       switch (pipe_format) {
4102       case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
4103          pipe_format = PIPE_FORMAT_Z32_FLOAT;
4104          break;
4105       case PIPE_FORMAT_X8Z24_UNORM:
4106       case PIPE_FORMAT_S8_UINT_Z24_UNORM:
4107          /* Z24 is always stored like this for DB
4108           * compatibility.
4109           */
4110          pipe_format = PIPE_FORMAT_Z24X8_UNORM;
4111          break;
4112       case PIPE_FORMAT_X24S8_UINT:
4113       case PIPE_FORMAT_S8X24_UINT:
4114       case PIPE_FORMAT_X32_S8X24_UINT:
4115          pipe_format = PIPE_FORMAT_S8_UINT;
4116          surflevel = tex->surface.u.legacy.zs.stencil_level;
4117          break;
4118       default:;
4119       }
4120    }
4121 
4122    view->dcc_incompatible =
4123       vi_dcc_formats_are_incompatible(texture, state->u.tex.first_level, state->format);
4124 
4125    sctx->screen->make_texture_descriptor(
4126       sctx->screen, tex, true, state->target, pipe_format, state_swizzle,
4127       state->u.tex.first_level, state->u.tex.last_level,
4128       state->u.tex.first_layer, last_layer, texture->width0, texture->height0, texture->depth0,
4129       false, view->state, view->fmask_state);
4130 
4131    view->base_level_info = &surflevel[0];
4132    view->block_width = util_format_get_blockwidth(pipe_format);
4133    return &view->base;
4134 }
4135 
si_sampler_view_destroy(struct pipe_context * ctx,struct pipe_sampler_view * state)4136 static void si_sampler_view_destroy(struct pipe_context *ctx, struct pipe_sampler_view *state)
4137 {
4138    struct si_sampler_view *view = (struct si_sampler_view *)state;
4139 
4140    pipe_resource_reference(&state->texture, NULL);
4141    FREE_CL(view);
4142 }
4143 
wrap_mode_uses_border_color(unsigned wrap,bool linear_filter)4144 static bool wrap_mode_uses_border_color(unsigned wrap, bool linear_filter)
4145 {
4146    return wrap == PIPE_TEX_WRAP_CLAMP_TO_BORDER || wrap == PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER ||
4147           (linear_filter && (wrap == PIPE_TEX_WRAP_CLAMP || wrap == PIPE_TEX_WRAP_MIRROR_CLAMP));
4148 }
4149 
si_translate_border_color(struct si_context * sctx,const struct pipe_sampler_state * state,const union pipe_color_union * color,bool is_integer,uint32_t * border_color_ptr)4150 static uint32_t si_translate_border_color(struct si_context *sctx,
4151                                           const struct pipe_sampler_state *state,
4152                                           const union pipe_color_union *color, bool is_integer,
4153                                           uint32_t *border_color_ptr)
4154 {
4155    bool linear_filter = state->min_img_filter != PIPE_TEX_FILTER_NEAREST ||
4156                         state->mag_img_filter != PIPE_TEX_FILTER_NEAREST;
4157 
4158    if (!wrap_mode_uses_border_color(state->wrap_s, linear_filter) &&
4159        !wrap_mode_uses_border_color(state->wrap_t, linear_filter) &&
4160        !wrap_mode_uses_border_color(state->wrap_r, linear_filter))
4161       return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK;
4162 
4163 #define simple_border_types(elt)                                                                   \
4164    do {                                                                                            \
4165       if (color->elt[0] == 0 && color->elt[1] == 0 && color->elt[2] == 0 && color->elt[3] == 0)    \
4166          return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK;                                          \
4167       if (color->elt[0] == 0 && color->elt[1] == 0 && color->elt[2] == 0 && color->elt[3] == 1)    \
4168          return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK;                                         \
4169       if (color->elt[0] == 1 && color->elt[1] == 1 && color->elt[2] == 1 && color->elt[3] == 1)    \
4170          return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE;                                         \
4171    } while (false)
4172 
4173    if (is_integer)
4174       simple_border_types(ui);
4175    else
4176       simple_border_types(f);
4177 
4178 #undef simple_border_types
4179 
4180    int i;
4181 
4182    /* Check if the border has been uploaded already. */
4183    for (i = 0; i < sctx->border_color_count; i++)
4184       if (memcmp(&sctx->border_color_table[i], color, sizeof(*color)) == 0)
4185          break;
4186 
4187    if (i >= SI_MAX_BORDER_COLORS) {
4188       /* Getting 4096 unique border colors is very unlikely. */
4189       static bool printed;
4190       if (!printed) {
4191          fprintf(stderr, "radeonsi: The border color table is full. "
4192                          "Any new border colors will be just black. "
4193                          "This is a hardware limitation.\n");
4194          printed = true;
4195       }
4196       return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK;
4197    }
4198 
4199    if (i == sctx->border_color_count) {
4200       /* Upload a new border color. */
4201       memcpy(&sctx->border_color_table[i], color, sizeof(*color));
4202       util_memcpy_cpu_to_le32(&sctx->border_color_map[i], color, sizeof(*color));
4203       sctx->border_color_count++;
4204    }
4205 
4206    *border_color_ptr = i;
4207 
4208    return V_008F3C_SQ_TEX_BORDER_COLOR_REGISTER;
4209 }
4210 
S_FIXED(float value,unsigned frac_bits)4211 static inline int S_FIXED(float value, unsigned frac_bits)
4212 {
4213    return value * (1 << frac_bits);
4214 }
4215 
si_tex_filter(unsigned filter,unsigned max_aniso)4216 static inline unsigned si_tex_filter(unsigned filter, unsigned max_aniso)
4217 {
4218    if (filter == PIPE_TEX_FILTER_LINEAR)
4219       return max_aniso > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
4220                            : V_008F38_SQ_TEX_XY_FILTER_BILINEAR;
4221    else
4222       return max_aniso > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
4223                            : V_008F38_SQ_TEX_XY_FILTER_POINT;
4224 }
4225 
si_tex_aniso_filter(unsigned filter)4226 static inline unsigned si_tex_aniso_filter(unsigned filter)
4227 {
4228    if (filter < 2)
4229       return 0;
4230    if (filter < 4)
4231       return 1;
4232    if (filter < 8)
4233       return 2;
4234    if (filter < 16)
4235       return 3;
4236    return 4;
4237 }
4238 
si_tex_filter_mode(unsigned mode)4239 static unsigned si_tex_filter_mode(unsigned mode)
4240 {
4241    switch (mode) {
4242    case PIPE_TEX_REDUCTION_WEIGHTED_AVERAGE:
4243       return V_008F30_SQ_IMG_FILTER_MODE_BLEND;
4244    case PIPE_TEX_REDUCTION_MIN:
4245       return V_008F30_SQ_IMG_FILTER_MODE_MIN;
4246    case PIPE_TEX_REDUCTION_MAX:
4247       return V_008F30_SQ_IMG_FILTER_MODE_MAX;
4248    default:
4249       break;
4250    }
4251    return 0;
4252 }
4253 
si_create_sampler_state(struct pipe_context * ctx,const struct pipe_sampler_state * state)4254 static void *si_create_sampler_state(struct pipe_context *ctx,
4255                                      const struct pipe_sampler_state *state)
4256 {
4257    struct si_context *sctx = (struct si_context *)ctx;
4258    struct si_screen *sscreen = sctx->screen;
4259    struct si_sampler_state *rstate = CALLOC_STRUCT(si_sampler_state);
4260    unsigned max_aniso = sscreen->force_aniso >= 0 ? sscreen->force_aniso : state->max_anisotropy;
4261    unsigned max_aniso_ratio = si_tex_aniso_filter(max_aniso);
4262    unsigned filter_mode = si_tex_filter_mode(state->reduction_mode);
4263    bool trunc_coord = (state->min_img_filter == PIPE_TEX_FILTER_NEAREST &&
4264                        state->mag_img_filter == PIPE_TEX_FILTER_NEAREST &&
4265                        state->compare_mode == PIPE_TEX_COMPARE_NONE) ||
4266                       sscreen->info.conformant_trunc_coord;
4267    union pipe_color_union clamped_border_color;
4268 
4269    if (!rstate) {
4270       return NULL;
4271    }
4272 
4273    /* Validate inputs. */
4274    if (!is_wrap_mode_legal(sscreen, state->wrap_s) ||
4275        !is_wrap_mode_legal(sscreen, state->wrap_t) ||
4276        !is_wrap_mode_legal(sscreen, state->wrap_r) ||
4277        (!sscreen->info.has_3d_cube_border_color_mipmap &&
4278         (state->min_mip_filter != PIPE_TEX_MIPFILTER_NONE ||
4279          state->max_anisotropy > 0))) {
4280       assert(0);
4281       return NULL;
4282    }
4283 
4284 #ifndef NDEBUG
4285    rstate->magic = SI_SAMPLER_STATE_MAGIC;
4286 #endif
4287 
4288    unsigned border_color_ptr = 0;
4289    unsigned border_color_type =
4290       si_translate_border_color(sctx, state, &state->border_color,
4291                                 state->border_color_is_integer,
4292                                 &border_color_ptr);
4293 
4294    struct ac_sampler_state ac_state = {
4295       .address_mode_u = si_tex_wrap(state->wrap_s),
4296       .address_mode_v = si_tex_wrap(state->wrap_t),
4297       .address_mode_w = si_tex_wrap(state->wrap_r),
4298       .max_aniso_ratio = max_aniso_ratio,
4299       .depth_compare_func = si_tex_compare(state->compare_mode, state->compare_func),
4300       .unnormalized_coords = state->unnormalized_coords,
4301       .cube_wrap = state->seamless_cube_map,
4302       .trunc_coord = trunc_coord,
4303       .filter_mode = filter_mode,
4304       .mag_filter = si_tex_filter(state->mag_img_filter, max_aniso),
4305       .min_filter = si_tex_filter(state->min_img_filter, max_aniso),
4306       .mip_filter = si_tex_mipfilter(state->min_mip_filter),
4307       .min_lod = state->min_lod,
4308       .max_lod = state->max_lod,
4309       .lod_bias = state->lod_bias,
4310       .border_color_type = border_color_type,
4311       .border_color_ptr = border_color_ptr,
4312    };
4313 
4314    ac_build_sampler_descriptor(sscreen->info.gfx_level, &ac_state, rstate->val);
4315 
4316    /* Create sampler resource for upgraded depth textures. */
4317    memcpy(rstate->upgraded_depth_val, rstate->val, sizeof(rstate->val));
4318 
4319    for (unsigned i = 0; i < 4; ++i) {
4320       /* Use channel 0 on purpose, so that we can use OPAQUE_WHITE
4321        * when the border color is 1.0. */
4322       clamped_border_color.f[i] = CLAMP(state->border_color.f[0], 0, 1);
4323    }
4324 
4325    if (memcmp(&state->border_color, &clamped_border_color, sizeof(clamped_border_color)) == 0) {
4326       if (sscreen->info.gfx_level <= GFX9)
4327          rstate->upgraded_depth_val[3] |= S_008F3C_UPGRADED_DEPTH(1);
4328    } else {
4329       border_color_ptr = 0;
4330       border_color_type = si_translate_border_color(sctx, state, &clamped_border_color, false, &border_color_ptr);
4331 
4332       rstate->upgraded_depth_val[3] = S_008F3C_BORDER_COLOR_TYPE(border_color_type);
4333 
4334       if (sscreen->info.gfx_level >= GFX11) {
4335          rstate->upgraded_depth_val[3] |= S_008F3C_BORDER_COLOR_PTR_GFX11(border_color_ptr);
4336       } else {
4337          rstate->upgraded_depth_val[3] |= S_008F3C_BORDER_COLOR_PTR_GFX6(border_color_ptr);
4338       }
4339    }
4340 
4341    return rstate;
4342 }
4343 
si_set_sample_mask(struct pipe_context * ctx,unsigned sample_mask)4344 static void si_set_sample_mask(struct pipe_context *ctx, unsigned sample_mask)
4345 {
4346    struct si_context *sctx = (struct si_context *)ctx;
4347 
4348    if (sctx->sample_mask == (uint16_t)sample_mask)
4349       return;
4350 
4351    sctx->sample_mask = sample_mask;
4352    si_mark_atom_dirty(sctx, &sctx->atoms.s.sample_mask);
4353 }
4354 
si_emit_sample_mask(struct si_context * sctx,unsigned index)4355 static void si_emit_sample_mask(struct si_context *sctx, unsigned index)
4356 {
4357    struct radeon_cmdbuf *cs = &sctx->gfx_cs;
4358    unsigned mask = sctx->sample_mask;
4359 
4360    /* Needed for line and polygon smoothing as well as for the Polaris
4361     * small primitive filter. We expect the gallium frontend to take care of
4362     * this for us.
4363     */
4364    assert(mask == 0xffff || sctx->framebuffer.nr_samples > 1 ||
4365           (mask & 1 && sctx->blitter_running));
4366 
4367    radeon_begin(cs);
4368    radeon_set_context_reg_seq(R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
4369    radeon_emit(mask | (mask << 16));
4370    radeon_emit(mask | (mask << 16));
4371    radeon_end();
4372 }
4373 
si_delete_sampler_state(struct pipe_context * ctx,void * state)4374 static void si_delete_sampler_state(struct pipe_context *ctx, void *state)
4375 {
4376 #ifndef NDEBUG
4377    struct si_sampler_state *s = state;
4378 
4379    assert(s->magic == SI_SAMPLER_STATE_MAGIC);
4380    s->magic = 0;
4381 #endif
4382    free(state);
4383 }
4384 
4385 /*
4386  * Vertex elements & buffers
4387  */
4388 
si_compute_fast_udiv_info32(uint32_t D,unsigned num_bits)4389 struct si_fast_udiv_info32 si_compute_fast_udiv_info32(uint32_t D, unsigned num_bits)
4390 {
4391    struct util_fast_udiv_info info = util_compute_fast_udiv_info(D, num_bits, 32);
4392 
4393    struct si_fast_udiv_info32 result = {
4394       info.multiplier,
4395       info.pre_shift,
4396       info.post_shift,
4397       info.increment,
4398    };
4399    return result;
4400 }
4401 
si_create_vertex_elements(struct pipe_context * ctx,unsigned count,const struct pipe_vertex_element * elements)4402 static void *si_create_vertex_elements(struct pipe_context *ctx, unsigned count,
4403                                        const struct pipe_vertex_element *elements)
4404 {
4405    struct si_screen *sscreen = (struct si_screen *)ctx->screen;
4406 
4407    if (sscreen->debug_flags & DBG(VERTEX_ELEMENTS)) {
4408       for (int i = 0; i < count; ++i) {
4409          const struct pipe_vertex_element *e = elements + i;
4410          fprintf(stderr, "elements[%d]: offset %2d, buffer_index %d, dual_slot %d, format %3d, divisor %u\n",
4411                 i, e->src_offset, e->vertex_buffer_index, e->dual_slot, e->src_format, e->instance_divisor);
4412       }
4413    }
4414 
4415    struct si_vertex_elements *v = CALLOC_STRUCT(si_vertex_elements);
4416    struct si_fast_udiv_info32 divisor_factors[SI_MAX_ATTRIBS] = {};
4417    STATIC_ASSERT(sizeof(struct si_fast_udiv_info32) == 16);
4418    STATIC_ASSERT(sizeof(divisor_factors[0].multiplier) == 4);
4419    STATIC_ASSERT(sizeof(divisor_factors[0].pre_shift) == 4);
4420    STATIC_ASSERT(sizeof(divisor_factors[0].post_shift) == 4);
4421    STATIC_ASSERT(sizeof(divisor_factors[0].increment) == 4);
4422    int i;
4423 
4424    assert(count <= SI_MAX_ATTRIBS);
4425    if (!v)
4426       return NULL;
4427 
4428    v->count = count;
4429 
4430    unsigned num_vbos_in_user_sgprs = si_num_vbos_in_user_sgprs(sscreen);
4431    unsigned alloc_count =
4432       count > num_vbos_in_user_sgprs ? count - num_vbos_in_user_sgprs : 0;
4433    v->vb_desc_list_alloc_size = align(alloc_count * 16, SI_CPDMA_ALIGNMENT);
4434 
4435    for (i = 0; i < count; ++i) {
4436       const struct util_format_description *desc;
4437       const struct util_format_channel_description *channel;
4438       int first_non_void;
4439       unsigned vbo_index = elements[i].vertex_buffer_index;
4440 
4441       if (vbo_index >= SI_NUM_VERTEX_BUFFERS) {
4442          FREE(v);
4443          return NULL;
4444       }
4445 
4446       unsigned instance_divisor = elements[i].instance_divisor;
4447       if (instance_divisor) {
4448          if (instance_divisor == 1) {
4449             v->instance_divisor_is_one |= 1u << i;
4450          } else {
4451             v->instance_divisor_is_fetched |= 1u << i;
4452             divisor_factors[i] = si_compute_fast_udiv_info32(instance_divisor, 32);
4453          }
4454       }
4455 
4456       desc = util_format_description(elements[i].src_format);
4457       first_non_void = util_format_get_first_non_void_channel(elements[i].src_format);
4458       channel = first_non_void >= 0 ? &desc->channel[first_non_void] : NULL;
4459 
4460       v->elem[i].format_size = desc->block.bits / 8;
4461       v->elem[i].src_offset = elements[i].src_offset;
4462       v->elem[i].stride = elements[i].src_stride;
4463       v->vertex_buffer_index[i] = vbo_index;
4464 
4465       bool always_fix = false;
4466       union si_vs_fix_fetch fix_fetch;
4467       unsigned log_hw_load_size; /* the load element size as seen by the hardware */
4468 
4469       fix_fetch.bits = 0;
4470       log_hw_load_size = MIN2(2, util_logbase2(desc->block.bits) - 3);
4471 
4472       if (channel) {
4473          switch (channel->type) {
4474          case UTIL_FORMAT_TYPE_FLOAT:
4475             fix_fetch.u.format = AC_FETCH_FORMAT_FLOAT;
4476             break;
4477          case UTIL_FORMAT_TYPE_FIXED:
4478             fix_fetch.u.format = AC_FETCH_FORMAT_FIXED;
4479             break;
4480          case UTIL_FORMAT_TYPE_SIGNED: {
4481             if (channel->pure_integer)
4482                fix_fetch.u.format = AC_FETCH_FORMAT_SINT;
4483             else if (channel->normalized)
4484                fix_fetch.u.format = AC_FETCH_FORMAT_SNORM;
4485             else
4486                fix_fetch.u.format = AC_FETCH_FORMAT_SSCALED;
4487             break;
4488          }
4489          case UTIL_FORMAT_TYPE_UNSIGNED: {
4490             if (channel->pure_integer)
4491                fix_fetch.u.format = AC_FETCH_FORMAT_UINT;
4492             else if (channel->normalized)
4493                fix_fetch.u.format = AC_FETCH_FORMAT_UNORM;
4494             else
4495                fix_fetch.u.format = AC_FETCH_FORMAT_USCALED;
4496             break;
4497          }
4498          default:
4499             unreachable("bad format type");
4500          }
4501       } else {
4502          switch (elements[i].src_format) {
4503          case PIPE_FORMAT_R11G11B10_FLOAT:
4504             fix_fetch.u.format = AC_FETCH_FORMAT_FLOAT;
4505             break;
4506          default:
4507             unreachable("bad other format");
4508          }
4509       }
4510 
4511       if (desc->channel[0].size == 10) {
4512          fix_fetch.u.log_size = 3; /* special encoding for 2_10_10_10 */
4513          log_hw_load_size = 2;
4514 
4515          /* The hardware always treats the 2-bit alpha channel as
4516           * unsigned, so a shader workaround is needed. The affected
4517           * chips are GFX8 and older except Stoney (GFX8.1).
4518           */
4519          always_fix = sscreen->info.gfx_level <= GFX8 && sscreen->info.family != CHIP_STONEY &&
4520                       channel->type == UTIL_FORMAT_TYPE_SIGNED;
4521       } else if (elements[i].src_format == PIPE_FORMAT_R11G11B10_FLOAT) {
4522          fix_fetch.u.log_size = 3; /* special encoding */
4523          fix_fetch.u.format = AC_FETCH_FORMAT_FIXED;
4524          log_hw_load_size = 2;
4525       } else {
4526          fix_fetch.u.log_size = util_logbase2(channel->size) - 3;
4527          fix_fetch.u.num_channels_m1 = desc->nr_channels - 1;
4528 
4529          /* Always fix up:
4530           * - doubles (multiple loads + truncate to float)
4531           * - 32-bit requiring a conversion
4532           */
4533          always_fix = (fix_fetch.u.log_size == 3) ||
4534                       (fix_fetch.u.log_size == 2 && fix_fetch.u.format != AC_FETCH_FORMAT_FLOAT &&
4535                        fix_fetch.u.format != AC_FETCH_FORMAT_UINT &&
4536                        fix_fetch.u.format != AC_FETCH_FORMAT_SINT);
4537 
4538          /* Also fixup 8_8_8 and 16_16_16. */
4539          if (desc->nr_channels == 3 && fix_fetch.u.log_size <= 1) {
4540             always_fix = true;
4541             log_hw_load_size = fix_fetch.u.log_size;
4542          }
4543       }
4544 
4545       if (desc->swizzle[0] != PIPE_SWIZZLE_X) {
4546          assert(desc->swizzle[0] == PIPE_SWIZZLE_Z &&
4547                 (desc->swizzle[2] == PIPE_SWIZZLE_X || desc->swizzle[2] == PIPE_SWIZZLE_0));
4548          fix_fetch.u.reverse = 1;
4549       }
4550 
4551       /* Force the workaround for unaligned access here already if the
4552        * offset relative to the vertex buffer base is unaligned.
4553        *
4554        * There is a theoretical case in which this is too conservative:
4555        * if the vertex buffer's offset is also unaligned in just the
4556        * right way, we end up with an aligned address after all.
4557        * However, this case should be extremely rare in practice (it
4558        * won't happen in well-behaved applications), and taking it
4559        * into account would complicate the fast path (where everything
4560        * is nicely aligned).
4561        */
4562       bool check_alignment =
4563             log_hw_load_size >= 1 &&
4564             (sscreen->info.gfx_level == GFX6 || sscreen->info.gfx_level >= GFX10);
4565       bool opencode = sscreen->options.vs_fetch_always_opencode;
4566 
4567       if (check_alignment && ((elements[i].src_offset & ((1 << log_hw_load_size) - 1)) != 0 ||
4568                               elements[i].src_stride & 3))
4569          opencode = true;
4570 
4571       if (always_fix || check_alignment || opencode)
4572          v->fix_fetch[i] = fix_fetch.bits;
4573 
4574       if (opencode)
4575          v->fix_fetch_opencode |= 1 << i;
4576       if (opencode || always_fix)
4577          v->fix_fetch_always |= 1 << i;
4578 
4579       if (check_alignment && !opencode) {
4580          assert(log_hw_load_size == 1 || log_hw_load_size == 2);
4581 
4582          v->fix_fetch_unaligned |= 1 << i;
4583          v->hw_load_is_dword |= (log_hw_load_size - 1) << i;
4584          v->vb_alignment_check_mask |= 1 << vbo_index;
4585       }
4586 
4587       const struct ac_buffer_state buffer_state = {
4588          .format = elements[i].src_format,
4589          .swizzle =
4590             {
4591                desc->swizzle[0],
4592                desc->swizzle[1],
4593                desc->swizzle[2],
4594                desc->swizzle[3],
4595             },
4596          /* OOB_SELECT chooses the out-of-bounds check:
4597           *  - 1: index >= NUM_RECORDS (Structured)
4598           *  - 3: offset >= NUM_RECORDS (Raw)
4599           */
4600          .gfx10_oob_select = v->elem[i].stride ? V_008F0C_OOB_SELECT_STRUCTURED
4601                                                : V_008F0C_OOB_SELECT_RAW,
4602       };
4603 
4604       ac_set_buf_desc_word3(sscreen->info.gfx_level, &buffer_state, &v->elem[i].rsrc_word3);
4605    }
4606 
4607    if (v->instance_divisor_is_fetched) {
4608       unsigned num_divisors = util_last_bit(v->instance_divisor_is_fetched);
4609 
4610       v->instance_divisor_factor_buffer = (struct si_resource *)pipe_buffer_create(
4611          &sscreen->b, 0, PIPE_USAGE_DEFAULT, num_divisors * sizeof(divisor_factors[0]));
4612       if (!v->instance_divisor_factor_buffer) {
4613          FREE(v);
4614          return NULL;
4615       }
4616       void *map =
4617          sscreen->ws->buffer_map(sscreen->ws, v->instance_divisor_factor_buffer->buf, NULL, PIPE_MAP_WRITE);
4618       memcpy(map, divisor_factors, num_divisors * sizeof(divisor_factors[0]));
4619    }
4620    return v;
4621 }
4622 
si_bind_vertex_elements(struct pipe_context * ctx,void * state)4623 static void si_bind_vertex_elements(struct pipe_context *ctx, void *state)
4624 {
4625    struct si_context *sctx = (struct si_context *)ctx;
4626    struct si_vertex_elements *old = sctx->vertex_elements;
4627    struct si_vertex_elements *v = (struct si_vertex_elements *)state;
4628 
4629    if (!v)
4630       v = sctx->no_velems_state;
4631 
4632    sctx->vertex_elements = v;
4633    sctx->num_vertex_elements = v->count;
4634    sctx->vertex_buffers_dirty = sctx->num_vertex_elements > 0;
4635 
4636    if (old->instance_divisor_is_one != v->instance_divisor_is_one ||
4637        old->instance_divisor_is_fetched != v->instance_divisor_is_fetched ||
4638        (old->vb_alignment_check_mask ^ v->vb_alignment_check_mask) &
4639        sctx->vertex_buffer_unaligned ||
4640        ((v->vb_alignment_check_mask & sctx->vertex_buffer_unaligned) &&
4641         memcmp(old->vertex_buffer_index, v->vertex_buffer_index,
4642                sizeof(v->vertex_buffer_index[0]) * MAX2(old->count, v->count))) ||
4643        /* fix_fetch_{always,opencode,unaligned} and hw_load_is_dword are
4644         * functions of fix_fetch and the src_offset alignment.
4645         * If they change and fix_fetch doesn't, it must be due to different
4646         * src_offset alignment, which is reflected in fix_fetch_opencode. */
4647        old->fix_fetch_opencode != v->fix_fetch_opencode ||
4648        memcmp(old->fix_fetch, v->fix_fetch, sizeof(v->fix_fetch[0]) *
4649               MAX2(old->count, v->count))) {
4650       si_vs_key_update_inputs(sctx);
4651       sctx->do_update_shaders = true;
4652    }
4653 
4654    if (v->instance_divisor_is_fetched) {
4655       struct pipe_constant_buffer cb;
4656 
4657       cb.buffer = &v->instance_divisor_factor_buffer->b.b;
4658       cb.user_buffer = NULL;
4659       cb.buffer_offset = 0;
4660       cb.buffer_size = 0xffffffff;
4661       si_set_internal_const_buffer(sctx, SI_VS_CONST_INSTANCE_DIVISORS, &cb);
4662    }
4663 }
4664 
si_delete_vertex_element(struct pipe_context * ctx,void * state)4665 static void si_delete_vertex_element(struct pipe_context *ctx, void *state)
4666 {
4667    struct si_context *sctx = (struct si_context *)ctx;
4668    struct si_vertex_elements *v = (struct si_vertex_elements *)state;
4669 
4670    if (sctx->vertex_elements == state)
4671       si_bind_vertex_elements(ctx, sctx->no_velems_state);
4672 
4673    si_resource_reference(&v->instance_divisor_factor_buffer, NULL);
4674    FREE(state);
4675 }
4676 
si_set_vertex_buffers(struct pipe_context * ctx,unsigned count,const struct pipe_vertex_buffer * buffers)4677 static void si_set_vertex_buffers(struct pipe_context *ctx, unsigned count,
4678                                   const struct pipe_vertex_buffer *buffers)
4679 {
4680    struct si_context *sctx = (struct si_context *)ctx;
4681    uint32_t unaligned = 0;
4682    unsigned i;
4683 
4684    assert(count <= ARRAY_SIZE(sctx->vertex_buffer));
4685    assert(!count || buffers);
4686 
4687    for (i = 0; i < count; i++) {
4688       const struct pipe_vertex_buffer *src = buffers + i;
4689       struct pipe_vertex_buffer *dst = sctx->vertex_buffer + i;
4690       struct pipe_resource *buf = src->buffer.resource;
4691 
4692       dst->buffer_offset = src->buffer_offset;
4693 
4694       /* Only unreference bound vertex buffers. */
4695       pipe_resource_reference(&dst->buffer.resource, NULL);
4696       dst->buffer.resource = src->buffer.resource;
4697 
4698       if (src->buffer_offset & 3)
4699          unaligned |= BITFIELD_BIT(i);
4700 
4701       if (buf) {
4702          si_resource(buf)->bind_history |= SI_BIND_VERTEX_BUFFER;
4703          radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, si_resource(buf),
4704                                    RADEON_USAGE_READ | RADEON_PRIO_VERTEX_BUFFER);
4705       }
4706    }
4707 
4708    unsigned last_count = sctx->num_vertex_buffers;
4709    for (; i < last_count; i++)
4710       pipe_resource_reference(&sctx->vertex_buffer[i].buffer.resource, NULL);
4711 
4712    sctx->num_vertex_buffers = count;
4713    sctx->vertex_buffers_dirty = sctx->num_vertex_elements > 0;
4714    sctx->vertex_buffer_unaligned = unaligned;
4715 
4716    /* Check whether alignment may have changed in a way that requires
4717     * shader changes. This check is conservative: a vertex buffer can only
4718     * trigger a shader change if the misalignment amount changes (e.g.
4719     * from byte-aligned to short-aligned), but we only keep track of
4720     * whether buffers are at least dword-aligned, since that should always
4721     * be the case in well-behaved applications anyway.
4722     */
4723    if (sctx->vertex_elements->vb_alignment_check_mask & unaligned) {
4724       si_vs_key_update_inputs(sctx);
4725       sctx->do_update_shaders = true;
4726    }
4727 }
4728 
4729 static struct pipe_vertex_state *
si_create_vertex_state(struct pipe_screen * screen,struct pipe_vertex_buffer * buffer,const struct pipe_vertex_element * elements,unsigned num_elements,struct pipe_resource * indexbuf,uint32_t full_velem_mask)4730 si_create_vertex_state(struct pipe_screen *screen,
4731                        struct pipe_vertex_buffer *buffer,
4732                        const struct pipe_vertex_element *elements,
4733                        unsigned num_elements,
4734                        struct pipe_resource *indexbuf,
4735                        uint32_t full_velem_mask)
4736 {
4737    struct si_screen *sscreen = (struct si_screen *)screen;
4738    struct si_vertex_state *state = CALLOC_STRUCT(si_vertex_state);
4739 
4740    util_init_pipe_vertex_state(screen, buffer, elements, num_elements, indexbuf, full_velem_mask,
4741                                &state->b);
4742 
4743    /* Initialize the vertex element state in state->element.
4744     * Do it by creating a vertex element state object and copying it there.
4745     */
4746    struct si_context ctx = {};
4747    ctx.b.screen = screen;
4748    struct si_vertex_elements *velems = si_create_vertex_elements(&ctx.b, num_elements, elements);
4749    state->velems = *velems;
4750    si_delete_vertex_element(&ctx.b, velems);
4751 
4752    assert(!state->velems.instance_divisor_is_one);
4753    assert(!state->velems.instance_divisor_is_fetched);
4754    assert(!state->velems.fix_fetch_always);
4755    assert(buffer->buffer_offset % 4 == 0);
4756    assert(!buffer->is_user_buffer);
4757    for (unsigned i = 0; i < num_elements; i++) {
4758       assert(elements[i].src_offset % 4 == 0);
4759       assert(!elements[i].dual_slot);
4760       assert(elements[i].src_stride % 4 == 0);
4761    }
4762 
4763    for (unsigned i = 0; i < num_elements; i++) {
4764       si_set_vertex_buffer_descriptor(sscreen, &state->velems, &state->b.input.vbuffer, i,
4765                                       &state->descriptors[i * 4]);
4766    }
4767 
4768    return &state->b;
4769 }
4770 
si_vertex_state_destroy(struct pipe_screen * screen,struct pipe_vertex_state * state)4771 static void si_vertex_state_destroy(struct pipe_screen *screen,
4772                                     struct pipe_vertex_state *state)
4773 {
4774    pipe_vertex_buffer_unreference(&state->input.vbuffer);
4775    pipe_resource_reference(&state->input.indexbuf, NULL);
4776    FREE(state);
4777 }
4778 
4779 static struct pipe_vertex_state *
si_pipe_create_vertex_state(struct pipe_screen * screen,struct pipe_vertex_buffer * buffer,const struct pipe_vertex_element * elements,unsigned num_elements,struct pipe_resource * indexbuf,uint32_t full_velem_mask)4780 si_pipe_create_vertex_state(struct pipe_screen *screen,
4781                             struct pipe_vertex_buffer *buffer,
4782                             const struct pipe_vertex_element *elements,
4783                             unsigned num_elements,
4784                             struct pipe_resource *indexbuf,
4785                             uint32_t full_velem_mask)
4786 {
4787    struct si_screen *sscreen = (struct si_screen *)screen;
4788 
4789    return util_vertex_state_cache_get(screen, buffer, elements, num_elements, indexbuf,
4790                                       full_velem_mask, &sscreen->vertex_state_cache);
4791 }
4792 
si_pipe_vertex_state_destroy(struct pipe_screen * screen,struct pipe_vertex_state * state)4793 static void si_pipe_vertex_state_destroy(struct pipe_screen *screen,
4794                                          struct pipe_vertex_state *state)
4795 {
4796    struct si_screen *sscreen = (struct si_screen *)screen;
4797 
4798    util_vertex_state_destroy(screen, &sscreen->vertex_state_cache, state);
4799 }
4800 
4801 /*
4802  * Misc
4803  */
4804 
si_set_tess_state(struct pipe_context * ctx,const float default_outer_level[4],const float default_inner_level[2])4805 static void si_set_tess_state(struct pipe_context *ctx, const float default_outer_level[4],
4806                               const float default_inner_level[2])
4807 {
4808    struct si_context *sctx = (struct si_context *)ctx;
4809    struct pipe_constant_buffer cb;
4810    float array[8];
4811 
4812    memcpy(array, default_outer_level, sizeof(float) * 4);
4813    memcpy(array + 4, default_inner_level, sizeof(float) * 2);
4814 
4815    cb.buffer = NULL;
4816    cb.user_buffer = array;
4817    cb.buffer_offset = 0;
4818    cb.buffer_size = sizeof(array);
4819 
4820    si_set_internal_const_buffer(sctx, SI_HS_CONST_DEFAULT_TESS_LEVELS, &cb);
4821 }
4822 
si_create_blend_custom(struct si_context * sctx,unsigned mode)4823 static void *si_create_blend_custom(struct si_context *sctx, unsigned mode)
4824 {
4825    struct pipe_blend_state blend;
4826 
4827    memset(&blend, 0, sizeof(blend));
4828    blend.independent_blend_enable = true;
4829    blend.rt[0].colormask = 0xf;
4830    return si_create_blend_state_mode(&sctx->b, &blend, mode);
4831 }
4832 
si_pm4_emit_sqtt_pipeline(struct si_context * sctx,unsigned index)4833 static void si_pm4_emit_sqtt_pipeline(struct si_context *sctx, unsigned index)
4834 {
4835    struct si_pm4_state *state = sctx->queued.array[index];
4836 
4837    si_pm4_emit_state(sctx, index);
4838 
4839    radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, ((struct si_sqtt_fake_pipeline*)state)->bo,
4840                              RADEON_USAGE_READ | RADEON_PRIO_SHADER_BINARY);
4841 }
4842 
si_init_state_compute_functions(struct si_context * sctx)4843 void si_init_state_compute_functions(struct si_context *sctx)
4844 {
4845    sctx->b.create_sampler_state = si_create_sampler_state;
4846    sctx->b.delete_sampler_state = si_delete_sampler_state;
4847    sctx->b.create_sampler_view = si_create_sampler_view;
4848    sctx->b.sampler_view_destroy = si_sampler_view_destroy;
4849 }
4850 
si_init_state_functions(struct si_context * sctx)4851 void si_init_state_functions(struct si_context *sctx)
4852 {
4853    sctx->atoms.s.pm4_states[SI_STATE_IDX(blend)].emit = si_pm4_emit_state;
4854    sctx->atoms.s.pm4_states[SI_STATE_IDX(rasterizer)].emit = si_pm4_emit_rasterizer;
4855    sctx->atoms.s.pm4_states[SI_STATE_IDX(dsa)].emit = si_pm4_emit_dsa;
4856    sctx->atoms.s.pm4_states[SI_STATE_IDX(sqtt_pipeline)].emit = si_pm4_emit_sqtt_pipeline;
4857    sctx->atoms.s.pm4_states[SI_STATE_IDX(ls)].emit = si_pm4_emit_shader;
4858    sctx->atoms.s.pm4_states[SI_STATE_IDX(hs)].emit = si_pm4_emit_shader;
4859    sctx->atoms.s.pm4_states[SI_STATE_IDX(es)].emit = si_pm4_emit_shader;
4860    sctx->atoms.s.pm4_states[SI_STATE_IDX(gs)].emit = si_pm4_emit_shader;
4861    sctx->atoms.s.pm4_states[SI_STATE_IDX(vs)].emit = si_pm4_emit_shader;
4862    sctx->atoms.s.pm4_states[SI_STATE_IDX(ps)].emit = si_pm4_emit_shader;
4863 
4864    if (sctx->gfx_level >= GFX12)
4865       sctx->atoms.s.framebuffer.emit = gfx12_emit_framebuffer_state;
4866    else if (sctx->screen->info.has_set_context_pairs_packed)
4867       sctx->atoms.s.framebuffer.emit = gfx11_dgpu_emit_framebuffer_state;
4868    else
4869       sctx->atoms.s.framebuffer.emit = gfx6_emit_framebuffer_state;
4870 
4871    sctx->atoms.s.db_render_state.emit = si_emit_db_render_state;
4872    sctx->atoms.s.dpbb_state.emit = si_emit_dpbb_state;
4873    sctx->atoms.s.msaa_config.emit = si_emit_msaa_config;
4874    sctx->atoms.s.sample_mask.emit = si_emit_sample_mask;
4875    sctx->atoms.s.cb_render_state.emit = si_emit_cb_render_state;
4876    sctx->atoms.s.blend_color.emit = si_emit_blend_color;
4877    sctx->atoms.s.clip_regs.emit = si_emit_clip_regs;
4878    sctx->atoms.s.clip_state.emit = si_emit_clip_state;
4879    sctx->atoms.s.stencil_ref.emit = si_emit_stencil_ref;
4880 
4881    sctx->b.create_blend_state = si_create_blend_state;
4882    sctx->b.bind_blend_state = si_bind_blend_state;
4883    sctx->b.delete_blend_state = si_delete_blend_state;
4884    sctx->b.set_blend_color = si_set_blend_color;
4885 
4886    sctx->b.create_rasterizer_state = si_create_rs_state;
4887    sctx->b.bind_rasterizer_state = si_bind_rs_state;
4888    sctx->b.delete_rasterizer_state = si_delete_rs_state;
4889 
4890    sctx->b.create_depth_stencil_alpha_state = si_create_dsa_state;
4891    sctx->b.bind_depth_stencil_alpha_state = si_bind_dsa_state;
4892    sctx->b.delete_depth_stencil_alpha_state = si_delete_dsa_state;
4893 
4894    sctx->custom_dsa_flush = si_create_db_flush_dsa(sctx);
4895 
4896    if (sctx->gfx_level < GFX11) {
4897       sctx->custom_blend_resolve = si_create_blend_custom(sctx, V_028808_CB_RESOLVE);
4898       sctx->custom_blend_fmask_decompress = si_create_blend_custom(sctx, V_028808_CB_FMASK_DECOMPRESS);
4899       sctx->custom_blend_eliminate_fastclear =
4900          si_create_blend_custom(sctx, V_028808_CB_ELIMINATE_FAST_CLEAR);
4901    }
4902 
4903    sctx->custom_blend_dcc_decompress =
4904       si_create_blend_custom(sctx,
4905                              sctx->gfx_level >= GFX12 ? V_028858_CB_DCC_DECOMPRESS :
4906                              sctx->gfx_level >= GFX11 ? V_028808_CB_DCC_DECOMPRESS_GFX11 :
4907                                                         V_028808_CB_DCC_DECOMPRESS_GFX8);
4908 
4909    sctx->b.set_clip_state = si_set_clip_state;
4910    sctx->b.set_stencil_ref = si_set_stencil_ref;
4911 
4912    sctx->b.set_framebuffer_state = si_set_framebuffer_state;
4913 
4914    sctx->b.set_sample_mask = si_set_sample_mask;
4915 
4916    sctx->b.create_vertex_elements_state = si_create_vertex_elements;
4917    sctx->b.bind_vertex_elements_state = si_bind_vertex_elements;
4918    sctx->b.delete_vertex_elements_state = si_delete_vertex_element;
4919    sctx->b.set_vertex_buffers = si_set_vertex_buffers;
4920 
4921    sctx->b.set_min_samples = si_set_min_samples;
4922    sctx->b.set_tess_state = si_set_tess_state;
4923 
4924    sctx->b.set_active_query_state = si_set_active_query_state;
4925 }
4926 
si_init_screen_state_functions(struct si_screen * sscreen)4927 void si_init_screen_state_functions(struct si_screen *sscreen)
4928 {
4929    sscreen->b.is_format_supported = si_is_format_supported;
4930    sscreen->b.create_vertex_state = si_pipe_create_vertex_state;
4931    sscreen->b.vertex_state_destroy = si_pipe_vertex_state_destroy;
4932 
4933    if (sscreen->info.gfx_level >= GFX10)
4934       sscreen->make_texture_descriptor = gfx10_make_texture_descriptor;
4935    else
4936       sscreen->make_texture_descriptor = si_make_texture_descriptor;
4937 
4938    util_vertex_state_cache_init(&sscreen->vertex_state_cache,
4939                                 si_create_vertex_state, si_vertex_state_destroy);
4940 }
4941 
si_set_grbm_gfx_index(struct si_context * sctx,struct si_pm4_state * pm4,unsigned value)4942 static void si_set_grbm_gfx_index(struct si_context *sctx, struct si_pm4_state *pm4, unsigned value)
4943 {
4944    unsigned reg = sctx->gfx_level >= GFX7 ? R_030800_GRBM_GFX_INDEX : R_00802C_GRBM_GFX_INDEX;
4945    ac_pm4_set_reg(&pm4->base, reg, value);
4946 }
4947 
si_set_grbm_gfx_index_se(struct si_context * sctx,struct si_pm4_state * pm4,unsigned se)4948 static void si_set_grbm_gfx_index_se(struct si_context *sctx, struct si_pm4_state *pm4, unsigned se)
4949 {
4950    assert(se == ~0 || se < sctx->screen->info.max_se);
4951    si_set_grbm_gfx_index(sctx, pm4,
4952                          (se == ~0 ? S_030800_SE_BROADCAST_WRITES(1) : S_030800_SE_INDEX(se)) |
4953                             S_030800_SH_BROADCAST_WRITES(1) |
4954                             S_030800_INSTANCE_BROADCAST_WRITES(1));
4955 }
4956 
si_write_harvested_raster_configs(struct si_context * sctx,struct si_pm4_state * pm4,unsigned raster_config,unsigned raster_config_1)4957 static void si_write_harvested_raster_configs(struct si_context *sctx, struct si_pm4_state *pm4,
4958                                               unsigned raster_config, unsigned raster_config_1)
4959 {
4960    unsigned num_se = MAX2(sctx->screen->info.max_se, 1);
4961    unsigned raster_config_se[4];
4962    unsigned se;
4963 
4964    ac_get_harvested_configs(&sctx->screen->info, raster_config, &raster_config_1, raster_config_se);
4965 
4966    for (se = 0; se < num_se; se++) {
4967       si_set_grbm_gfx_index_se(sctx, pm4, se);
4968       ac_pm4_set_reg(&pm4->base, R_028350_PA_SC_RASTER_CONFIG, raster_config_se[se]);
4969    }
4970    si_set_grbm_gfx_index(sctx, pm4, ~0);
4971 
4972    if (sctx->gfx_level >= GFX7) {
4973       ac_pm4_set_reg(&pm4->base, R_028354_PA_SC_RASTER_CONFIG_1, raster_config_1);
4974    }
4975 }
4976 
si_set_raster_config(struct si_context * sctx,struct si_pm4_state * pm4)4977 static void si_set_raster_config(struct si_context *sctx, struct si_pm4_state *pm4)
4978 {
4979    struct si_screen *sscreen = sctx->screen;
4980    unsigned num_rb = MIN2(sscreen->info.max_render_backends, 16);
4981    uint64_t rb_mask = sscreen->info.enabled_rb_mask;
4982    unsigned raster_config = sscreen->pa_sc_raster_config;
4983    unsigned raster_config_1 = sscreen->pa_sc_raster_config_1;
4984 
4985    if (!rb_mask || util_bitcount64(rb_mask) >= num_rb) {
4986       /* Always use the default config when all backends are enabled
4987        * (or when we failed to determine the enabled backends).
4988        */
4989       ac_pm4_set_reg(&pm4->base, R_028350_PA_SC_RASTER_CONFIG, raster_config);
4990       if (sctx->gfx_level >= GFX7)
4991          ac_pm4_set_reg(&pm4->base, R_028354_PA_SC_RASTER_CONFIG_1, raster_config_1);
4992    } else {
4993       si_write_harvested_raster_configs(sctx, pm4, raster_config, raster_config_1);
4994    }
4995 }
4996 
si_init_compute_preamble_state(struct si_context * sctx,struct si_pm4_state * pm4)4997 static void si_init_compute_preamble_state(struct si_context *sctx,
4998                                            struct si_pm4_state *pm4)
4999 {
5000    uint64_t border_color_va =
5001       sctx->border_color_buffer ? sctx->border_color_buffer->gpu_address : 0;
5002 
5003    const struct ac_preamble_state preamble_state = {
5004       .border_color_va = border_color_va,
5005       .gfx11 = {
5006          .compute_dispatch_interleave = 256,
5007       },
5008    };
5009 
5010    ac_init_compute_preamble_state(&preamble_state, &pm4->base);
5011 
5012    if (sctx->gfx_level == GFX10 || sctx->gfx_level == GFX10_3)
5013       ac_pm4_set_reg(&pm4->base, R_00B8A0_COMPUTE_PGM_RSRC3, 0);
5014 }
5015 
5016 
si_init_graphics_preamble_state(struct si_context * sctx,struct si_pm4_state * pm4)5017 static void si_init_graphics_preamble_state(struct si_context *sctx,
5018                                             struct si_pm4_state *pm4)
5019 {
5020    struct si_screen *sscreen = sctx->screen;
5021    uint64_t border_color_va =
5022       sctx->border_color_buffer ? sctx->border_color_buffer->gpu_address : 0;
5023 
5024    const struct ac_preamble_state preamble_state = {
5025       .border_color_va = border_color_va,
5026       .gfx10.cache_rb_gl2 = sctx->gfx_level >= GFX10 && sscreen->options.cache_rb_gl2,
5027    };
5028 
5029    ac_init_graphics_preamble_state(&preamble_state, &pm4->base);
5030 }
5031 
gfx6_init_gfx_preamble_state(struct si_context * sctx)5032 static void gfx6_init_gfx_preamble_state(struct si_context *sctx)
5033 {
5034    struct si_screen *sscreen = sctx->screen;
5035    bool has_clear_state = sscreen->info.has_clear_state;
5036 
5037    /* We need more space because the preamble is large. */
5038    struct si_pm4_state *pm4 = si_pm4_create_sized(sscreen, 214, sctx->has_graphics);
5039    if (!pm4)
5040       return;
5041 
5042    if (sctx->has_graphics && !sctx->shadowing.registers) {
5043       ac_pm4_cmd_add(&pm4->base, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
5044       ac_pm4_cmd_add(&pm4->base, CC0_UPDATE_LOAD_ENABLES(1));
5045       ac_pm4_cmd_add(&pm4->base, CC1_UPDATE_SHADOW_ENABLES(1));
5046 
5047       if (sscreen->dpbb_allowed) {
5048          ac_pm4_cmd_add(&pm4->base, PKT3(PKT3_EVENT_WRITE, 0, 0));
5049          ac_pm4_cmd_add(&pm4->base, EVENT_TYPE(V_028A90_BREAK_BATCH) | EVENT_INDEX(0));
5050       }
5051 
5052       if (has_clear_state) {
5053          ac_pm4_cmd_add(&pm4->base, PKT3(PKT3_CLEAR_STATE, 0, 0));
5054          ac_pm4_cmd_add(&pm4->base, 0);
5055       }
5056    }
5057 
5058    si_init_compute_preamble_state(sctx, pm4);
5059 
5060    if (!sctx->has_graphics)
5061       goto done;
5062 
5063    /* Graphics registers. */
5064    si_init_graphics_preamble_state(sctx, pm4);
5065 
5066    if (!has_clear_state)
5067       ac_pm4_set_reg(&pm4->base, R_02800C_DB_RENDER_OVERRIDE, 0);
5068 
5069    if (sctx->family >= CHIP_POLARIS10 && !sctx->screen->info.has_small_prim_filter_sample_loc_bug) {
5070       /* Polaris10-12 should disable small line culling, but those also have the sample loc bug,
5071        * so they never enter this branch.
5072        */
5073       assert(sctx->family > CHIP_POLARIS12);
5074       ac_pm4_set_reg(&pm4->base, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL,
5075                      S_028830_SMALL_PRIM_FILTER_ENABLE(1));
5076    }
5077 
5078    if (sctx->gfx_level <= GFX7 || !has_clear_state) {
5079       ac_pm4_set_reg(&pm4->base, R_028034_PA_SC_SCREEN_SCISSOR_BR,
5080                      S_028034_BR_X(16384) | S_028034_BR_Y(16384));
5081    }
5082 
5083    if (sctx->gfx_level == GFX9) {
5084       ac_pm4_set_reg(&pm4->base, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
5085                      S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
5086    }
5087 
5088 done:
5089    ac_pm4_finalize(&pm4->base);
5090    sctx->cs_preamble_state = pm4;
5091    sctx->cs_preamble_state_tmz = si_pm4_clone(sscreen, pm4); /* Make a copy of the preamble for TMZ. */
5092 }
5093 
cdna_init_compute_preamble_state(struct si_context * sctx)5094 static void cdna_init_compute_preamble_state(struct si_context *sctx)
5095 {
5096    struct si_screen *sscreen = sctx->screen;
5097    uint64_t border_color_va =
5098       sctx->border_color_buffer ? sctx->border_color_buffer->gpu_address : 0;
5099    uint32_t compute_cu_en = S_00B858_SH0_CU_EN(sscreen->info.spi_cu_en) |
5100                             S_00B858_SH1_CU_EN(sscreen->info.spi_cu_en);
5101 
5102    struct si_pm4_state *pm4 = si_pm4_create_sized(sscreen, 48, true);
5103    if (!pm4)
5104       return;
5105 
5106    /* Compute registers. */
5107    /* Disable profiling on compute chips. */
5108    ac_pm4_set_reg(&pm4->base, R_00B82C_COMPUTE_PERFCOUNT_ENABLE, 0);
5109    ac_pm4_set_reg(&pm4->base, R_00B834_COMPUTE_PGM_HI, S_00B834_DATA(sctx->screen->info.address32_hi >> 8));
5110    ac_pm4_set_reg(&pm4->base, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0, compute_cu_en);
5111    ac_pm4_set_reg(&pm4->base, R_00B85C_COMPUTE_STATIC_THREAD_MGMT_SE1, compute_cu_en);
5112    ac_pm4_set_reg(&pm4->base, R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2, compute_cu_en);
5113    ac_pm4_set_reg(&pm4->base, R_00B868_COMPUTE_STATIC_THREAD_MGMT_SE3, compute_cu_en);
5114    ac_pm4_set_reg(&pm4->base, R_00B878_COMPUTE_THREAD_TRACE_ENABLE, 0);
5115 
5116    if (sscreen->info.family >= CHIP_GFX940) {
5117       ac_pm4_set_reg(&pm4->base, R_00B89C_COMPUTE_TG_CHUNK_SIZE, 0);
5118       ac_pm4_set_reg(&pm4->base, R_00B8B4_COMPUTE_PGM_RSRC3, 0);
5119    } else {
5120       ac_pm4_set_reg(&pm4->base, R_00B894_COMPUTE_STATIC_THREAD_MGMT_SE4, compute_cu_en);
5121       ac_pm4_set_reg(&pm4->base, R_00B898_COMPUTE_STATIC_THREAD_MGMT_SE5, compute_cu_en);
5122       ac_pm4_set_reg(&pm4->base, R_00B89C_COMPUTE_STATIC_THREAD_MGMT_SE6, compute_cu_en);
5123       ac_pm4_set_reg(&pm4->base, R_00B8A0_COMPUTE_STATIC_THREAD_MGMT_SE7, compute_cu_en);
5124    }
5125 
5126    ac_pm4_set_reg(&pm4->base, R_0301EC_CP_COHER_START_DELAY, 0);
5127 
5128    /* Set the pointer to border colors. Only MI100 supports border colors. */
5129    if (sscreen->info.family == CHIP_MI100) {
5130       ac_pm4_set_reg(&pm4->base, R_030E00_TA_CS_BC_BASE_ADDR, border_color_va >> 8);
5131       ac_pm4_set_reg(&pm4->base, R_030E04_TA_CS_BC_BASE_ADDR_HI,
5132                      S_030E04_ADDRESS(border_color_va >> 40));
5133    }
5134 
5135    ac_pm4_finalize(&pm4->base);
5136    sctx->cs_preamble_state = pm4;
5137    sctx->cs_preamble_state_tmz = si_pm4_clone(sscreen, pm4); /* Make a copy of the preamble for TMZ. */
5138 }
5139 
gfx10_init_gfx_preamble_state(struct si_context * sctx)5140 static void gfx10_init_gfx_preamble_state(struct si_context *sctx)
5141 {
5142    struct si_screen *sscreen = sctx->screen;
5143 
5144    /* We need more space because the preamble is large. */
5145    struct si_pm4_state *pm4 = si_pm4_create_sized(sscreen, 214, sctx->has_graphics);
5146    if (!pm4)
5147       return;
5148 
5149    if (sctx->has_graphics && !sctx->shadowing.registers) {
5150       ac_pm4_cmd_add(&pm4->base, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
5151       ac_pm4_cmd_add(&pm4->base, CC0_UPDATE_LOAD_ENABLES(1));
5152       ac_pm4_cmd_add(&pm4->base, CC1_UPDATE_SHADOW_ENABLES(1));
5153 
5154       if (sscreen->dpbb_allowed) {
5155          ac_pm4_cmd_add(&pm4->base, PKT3(PKT3_EVENT_WRITE, 0, 0));
5156          ac_pm4_cmd_add(&pm4->base, EVENT_TYPE(V_028A90_BREAK_BATCH) | EVENT_INDEX(0));
5157       }
5158 
5159       ac_pm4_cmd_add(&pm4->base, PKT3(PKT3_CLEAR_STATE, 0, 0));
5160       ac_pm4_cmd_add(&pm4->base, 0);
5161    }
5162 
5163    si_init_compute_preamble_state(sctx, pm4);
5164 
5165    if (!sctx->has_graphics)
5166       goto done;
5167 
5168    /* Graphics registers. */
5169    si_init_graphics_preamble_state(sctx, pm4);
5170 
5171    ac_pm4_set_reg(&pm4->base, R_028708_SPI_SHADER_IDX_FORMAT,
5172                   S_028708_IDX0_EXPORT_FORMAT(V_028708_SPI_SHADER_1COMP));
5173 
5174    if (sctx->gfx_level >= GFX10_3) {
5175       /* The rate combiners have no effect if they are disabled like this:
5176        *   VERTEX_RATE:    BYPASS_VTX_RATE_COMBINER = 1
5177        *   PRIMITIVE_RATE: BYPASS_PRIM_RATE_COMBINER = 1
5178        *   HTILE_RATE:     VRS_HTILE_ENCODING = 0
5179        *   SAMPLE_ITER:    PS_ITER_SAMPLE = 0
5180        *
5181        * Use OVERRIDE, which will ignore results from previous combiners.
5182        * (e.g. enabled sample shading overrides the vertex rate)
5183        */
5184       ac_pm4_set_reg(&pm4->base, R_028848_PA_CL_VRS_CNTL,
5185                      S_028848_VERTEX_RATE_COMBINER_MODE(V_028848_SC_VRS_COMB_MODE_OVERRIDE) |
5186                      S_028848_SAMPLE_ITER_COMBINER_MODE(V_028848_SC_VRS_COMB_MODE_OVERRIDE));
5187    }
5188 
5189 done:
5190    ac_pm4_finalize(&pm4->base);
5191    sctx->cs_preamble_state = pm4;
5192    sctx->cs_preamble_state_tmz = si_pm4_clone(sscreen, pm4); /* Make a copy of the preamble for TMZ. */
5193 }
5194 
gfx12_init_gfx_preamble_state(struct si_context * sctx)5195 static void gfx12_init_gfx_preamble_state(struct si_context *sctx)
5196 {
5197    struct si_screen *sscreen = sctx->screen;
5198 
5199    struct si_pm4_state *pm4 = si_pm4_create_sized(sscreen, 300, sctx->has_graphics);
5200    if (!pm4)
5201       return;
5202 
5203    if (sctx->has_graphics && !sctx->shadowing.registers) {
5204       ac_pm4_cmd_add(&pm4->base, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
5205       ac_pm4_cmd_add(&pm4->base, CC0_UPDATE_LOAD_ENABLES(1));
5206       ac_pm4_cmd_add(&pm4->base, CC1_UPDATE_SHADOW_ENABLES(1));
5207    }
5208 
5209    if (sctx->has_graphics && sscreen->dpbb_allowed) {
5210       ac_pm4_cmd_add(&pm4->base, PKT3(PKT3_EVENT_WRITE, 0, 0));
5211       ac_pm4_cmd_add(&pm4->base, EVENT_TYPE(V_028A90_BREAK_BATCH) | EVENT_INDEX(0));
5212    }
5213 
5214    si_init_compute_preamble_state(sctx, pm4);
5215 
5216    if (!sctx->has_graphics)
5217       goto done;
5218 
5219    /* Graphics registers. */
5220    si_init_graphics_preamble_state(sctx, pm4);
5221 
5222    ac_pm4_set_reg(&pm4->base, R_028648_SPI_SHADER_IDX_FORMAT,
5223                   S_028648_IDX0_EXPORT_FORMAT(V_028648_SPI_SHADER_1COMP));
5224 
5225    /* The rate combiners have no effect if they are disabled like this:
5226     *   VERTEX_RATE:    BYPASS_VTX_RATE_COMBINER = 1
5227     *   PRIMITIVE_RATE: BYPASS_PRIM_RATE_COMBINER = 1
5228     *   HTILE_RATE:     VRS_HTILE_ENCODING = 0
5229     *   SAMPLE_ITER:    PS_ITER_SAMPLE = 0
5230     *
5231     * Use OVERRIDE, which will ignore results from previous combiners.
5232     * (e.g. enabled sample shading overrides the vertex rate)
5233     */
5234    ac_pm4_set_reg(&pm4->base, R_028848_PA_CL_VRS_CNTL,
5235                   S_028848_VERTEX_RATE_COMBINER_MODE(V_028848_SC_VRS_COMB_MODE_OVERRIDE) |
5236                   S_028848_SAMPLE_ITER_COMBINER_MODE(V_028848_SC_VRS_COMB_MODE_OVERRIDE));
5237 
5238    ac_pm4_set_reg(&pm4->base, R_028C54_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
5239                   S_028C54_NULL_SQUAD_AA_MASK_ENABLE(1));
5240 
5241 done:
5242    sctx->cs_preamble_state = pm4;
5243    sctx->cs_preamble_state_tmz = si_pm4_clone(sscreen, pm4); /* Make a copy of the preamble for TMZ. */
5244 }
5245 
si_init_gfx_preamble_state(struct si_context * sctx)5246 void si_init_gfx_preamble_state(struct si_context *sctx)
5247 {
5248    if (!sctx->screen->info.has_graphics)
5249       cdna_init_compute_preamble_state(sctx);
5250    else if (sctx->gfx_level >= GFX12)
5251       gfx12_init_gfx_preamble_state(sctx);
5252    else if (sctx->gfx_level >= GFX10)
5253       gfx10_init_gfx_preamble_state(sctx);
5254    else
5255       gfx6_init_gfx_preamble_state(sctx);
5256 }
5257