projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
amd: replace SH -> SA (shader array) in comments
[mesa.git] / src / amd / vulkan / si_cmd_buffer.c
1 /*
2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
4 *
5 * based on si_state.c
6 * Copyright © 2015 Advanced Micro Devices, Inc.
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
17 * Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 * IN THE SOFTWARE.
26 */
27
28 /* command buffer handling for AMD GCN */
29
30 #include "radv_private.h"
31 #include "radv_shader.h"
32 #include "radv_cs.h"
33 #include "sid.h"
34 #include "radv_util.h"
35
36 static void
37 si_write_harvested_raster_configs(struct radv_physical_device *physical_device,
38 struct radeon_cmdbuf *cs,
39 unsigned raster_config,
40 unsigned raster_config_1)
41 {
42 unsigned num_se = MAX2(physical_device->rad_info.max_se, 1);
43 unsigned raster_config_se[4];
44 unsigned se;
45
46 ac_get_harvested_configs(&physical_device->rad_info,
47 raster_config,
48 &raster_config_1,
49 raster_config_se);
50
51 for (se = 0; se < num_se; se++) {
52 /* GRBM_GFX_INDEX has a different offset on GFX6 and GFX7+ */
53 if (physical_device->rad_info.chip_class < GFX7)
54 radeon_set_config_reg(cs, R_00802C_GRBM_GFX_INDEX,
55 S_00802C_SE_INDEX(se) |
56 S_00802C_SH_BROADCAST_WRITES(1) |
57 S_00802C_INSTANCE_BROADCAST_WRITES(1));
58 else
59 radeon_set_uconfig_reg(cs, R_030800_GRBM_GFX_INDEX,
60 S_030800_SE_INDEX(se) | S_030800_SH_BROADCAST_WRITES(1) |
61 S_030800_INSTANCE_BROADCAST_WRITES(1));
62 radeon_set_context_reg(cs, R_028350_PA_SC_RASTER_CONFIG, raster_config_se[se]);
63 }
64
65 /* GRBM_GFX_INDEX has a different offset on GFX6 and GFX7+ */
66 if (physical_device->rad_info.chip_class < GFX7)
67 radeon_set_config_reg(cs, R_00802C_GRBM_GFX_INDEX,
68 S_00802C_SE_BROADCAST_WRITES(1) |
69 S_00802C_SH_BROADCAST_WRITES(1) |
70 S_00802C_INSTANCE_BROADCAST_WRITES(1));
71 else
72 radeon_set_uconfig_reg(cs, R_030800_GRBM_GFX_INDEX,
73 S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
74 S_030800_INSTANCE_BROADCAST_WRITES(1));
75
76 if (physical_device->rad_info.chip_class >= GFX7)
77 radeon_set_context_reg(cs, R_028354_PA_SC_RASTER_CONFIG_1, raster_config_1);
78 }
79
80 void
81 si_emit_compute(struct radv_physical_device *physical_device,
82 struct radeon_cmdbuf *cs)
83 {
84 radeon_set_sh_reg_seq(cs, R_00B810_COMPUTE_START_X, 3);
85 radeon_emit(cs, 0);
86 radeon_emit(cs, 0);
87 radeon_emit(cs, 0);
88
89 radeon_set_sh_reg_seq(cs, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0, 2);
90 /* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1,
91 * renamed COMPUTE_DESTINATION_EN_SEn on gfx10. */
92 radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
93 radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
94
95 if (physical_device->rad_info.chip_class >= GFX7) {
96 /* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
97 radeon_set_sh_reg_seq(cs,
98 R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2, 2);
99 radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) |
100 S_00B858_SH1_CU_EN(0xffff));
101 radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) |
102 S_00B858_SH1_CU_EN(0xffff));
103 }
104
105 if (physical_device->rad_info.chip_class >= GFX10)
106 radeon_set_sh_reg(cs, R_00B8A0_COMPUTE_PGM_RSRC3, 0);
107
108 /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
109 * and is now per pipe, so it should be handled in the
110 * kernel if we want to use something other than the default value,
111 * which is now 0x22f.
112 */
113 if (physical_device->rad_info.chip_class <= GFX6) {
114 /* XXX: This should be:
115 * (number of compute units) * 4 * (waves per simd) - 1 */
116
117 radeon_set_sh_reg(cs, R_00B82C_COMPUTE_MAX_WAVE_ID,
118 0x190 /* Default value */);
119 }
120 }
121
122 /* 12.4 fixed-point */
123 static unsigned radv_pack_float_12p4(float x)
124 {
125 return x <= 0 ? 0 :
126 x >= 4096 ? 0xffff : x * 16;
127 }
128
129 static void
130 si_set_raster_config(struct radv_physical_device *physical_device,
131 struct radeon_cmdbuf *cs)
132 {
133 unsigned num_rb = MIN2(physical_device->rad_info.num_render_backends, 16);
134 unsigned rb_mask = physical_device->rad_info.enabled_rb_mask;
135 unsigned raster_config, raster_config_1;
136
137 ac_get_raster_config(&physical_device->rad_info,
138 &raster_config,
139 &raster_config_1, NULL);
140
141 /* Always use the default config when all backends are enabled
142 * (or when we failed to determine the enabled backends).
143 */
144 if (!rb_mask || util_bitcount(rb_mask) >= num_rb) {
145 radeon_set_context_reg(cs, R_028350_PA_SC_RASTER_CONFIG,
146 raster_config);
147 if (physical_device->rad_info.chip_class >= GFX7)
148 radeon_set_context_reg(cs, R_028354_PA_SC_RASTER_CONFIG_1,
149 raster_config_1);
150 } else {
151 si_write_harvested_raster_configs(physical_device, cs,
152 raster_config,
153 raster_config_1);
154 }
155 }
156
157 void
158 si_emit_graphics(struct radv_device *device,
159 struct radeon_cmdbuf *cs)
160 {
161 struct radv_physical_device *physical_device = device->physical_device;
162
163 bool has_clear_state = physical_device->rad_info.has_clear_state;
164 int i;
165
166 radeon_emit(cs, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
167 radeon_emit(cs, CC0_UPDATE_LOAD_ENABLES(1));
168 radeon_emit(cs, CC1_UPDATE_SHADOW_ENABLES(1));
169
170 if (has_clear_state) {
171 radeon_emit(cs, PKT3(PKT3_CLEAR_STATE, 0, 0));
172 radeon_emit(cs, 0);
173 }
174
175 if (physical_device->rad_info.chip_class <= GFX8)
176 si_set_raster_config(physical_device, cs);
177
178 radeon_set_context_reg(cs, R_028A18_VGT_HOS_MAX_TESS_LEVEL, fui(64));
179 if (!has_clear_state)
180 radeon_set_context_reg(cs, R_028A1C_VGT_HOS_MIN_TESS_LEVEL, fui(0));
181
182 /* FIXME calculate these values somehow ??? */
183 if (physical_device->rad_info.chip_class <= GFX8) {
184 radeon_set_context_reg(cs, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
185 radeon_set_context_reg(cs, R_028A58_VGT_ES_PER_GS, 0x40);
186 }
187
188 if (!has_clear_state) {
189 radeon_set_context_reg(cs, R_028A5C_VGT_GS_PER_VS, 0x2);
190 radeon_set_context_reg(cs, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
191 radeon_set_context_reg(cs, R_028B98_VGT_STRMOUT_BUFFER_CONFIG, 0x0);
192 }
193
194 if (physical_device->rad_info.chip_class <= GFX9)
195 radeon_set_context_reg(cs, R_028AA0_VGT_INSTANCE_STEP_RATE_0, 1);
196 if (!has_clear_state)
197 radeon_set_context_reg(cs, R_028AB8_VGT_VTX_CNT_EN, 0x0);
198 if (physical_device->rad_info.chip_class < GFX7)
199 radeon_set_config_reg(cs, R_008A14_PA_CL_ENHANCE, S_008A14_NUM_CLIP_SEQ(3) |
200 S_008A14_CLIP_VTX_REORDER_ENA(1));
201
202 if (!has_clear_state)
203 radeon_set_context_reg(cs, R_02882C_PA_SU_PRIM_FILTER_CNTL, 0);
204
205 /* CLEAR_STATE doesn't clear these correctly on certain generations.
206 * I don't know why. Deduced by trial and error.
207 */
208 if (physical_device->rad_info.chip_class <= GFX7 || !has_clear_state) {
209 radeon_set_context_reg(cs, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET, 0);
210 radeon_set_context_reg(cs, R_028204_PA_SC_WINDOW_SCISSOR_TL,
211 S_028204_WINDOW_OFFSET_DISABLE(1));
212 radeon_set_context_reg(cs, R_028240_PA_SC_GENERIC_SCISSOR_TL,
213 S_028240_WINDOW_OFFSET_DISABLE(1));
214 radeon_set_context_reg(cs, R_028244_PA_SC_GENERIC_SCISSOR_BR,
215 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
216 radeon_set_context_reg(cs, R_028030_PA_SC_SCREEN_SCISSOR_TL, 0);
217 radeon_set_context_reg(cs, R_028034_PA_SC_SCREEN_SCISSOR_BR,
218 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
219 }
220
221 if (!has_clear_state) {
222 for (i = 0; i < 16; i++) {
223 radeon_set_context_reg(cs, R_0282D0_PA_SC_VPORT_ZMIN_0 + i*8, 0);
224 radeon_set_context_reg(cs, R_0282D4_PA_SC_VPORT_ZMAX_0 + i*8, fui(1.0));
225 }
226 }
227
228 if (!has_clear_state) {
229 radeon_set_context_reg(cs, R_02820C_PA_SC_CLIPRECT_RULE, 0xFFFF);
230 radeon_set_context_reg(cs, R_028230_PA_SC_EDGERULE, 0xAAAAAAAA);
231 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on GFX6 */
232 radeon_set_context_reg(cs, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET, 0);
233 radeon_set_context_reg(cs, R_028820_PA_CL_NANINF_CNTL, 0);
234 radeon_set_context_reg(cs, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 0x0);
235 radeon_set_context_reg(cs, R_028AC4_DB_SRESULTS_COMPARE_STATE1, 0x0);
236 radeon_set_context_reg(cs, R_028AC8_DB_PRELOAD_CONTROL, 0x0);
237 }
238
239 radeon_set_context_reg(cs, R_02800C_DB_RENDER_OVERRIDE,
240 S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
241 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE));
242
243 if (physical_device->rad_info.chip_class >= GFX10) {
244 radeon_set_context_reg(cs, R_028A98_VGT_DRAW_PAYLOAD_CNTL, 0);
245 radeon_set_uconfig_reg(cs, R_030964_GE_MAX_VTX_INDX, ~0);
246 radeon_set_uconfig_reg(cs, R_030924_GE_MIN_VTX_INDX, 0);
247 radeon_set_uconfig_reg(cs, R_030928_GE_INDX_OFFSET, 0);
248 radeon_set_uconfig_reg(cs, R_03097C_GE_STEREO_CNTL, 0);
249 radeon_set_uconfig_reg(cs, R_030988_GE_USER_VGPR_EN, 0);
250 } else if (physical_device->rad_info.chip_class == GFX9) {
251 radeon_set_uconfig_reg(cs, R_030920_VGT_MAX_VTX_INDX, ~0);
252 radeon_set_uconfig_reg(cs, R_030924_VGT_MIN_VTX_INDX, 0);
253 radeon_set_uconfig_reg(cs, R_030928_VGT_INDX_OFFSET, 0);
254 } else {
255 /* These registers, when written, also overwrite the
256 * CLEAR_STATE context, so we can't rely on CLEAR_STATE setting
257 * them. It would be an issue if there was another UMD
258 * changing them.
259 */
260 radeon_set_context_reg(cs, R_028400_VGT_MAX_VTX_INDX, ~0);
261 radeon_set_context_reg(cs, R_028404_VGT_MIN_VTX_INDX, 0);
262 radeon_set_context_reg(cs, R_028408_VGT_INDX_OFFSET, 0);
263 }
264
265 if (physical_device->rad_info.chip_class >= GFX7) {
266 if (physical_device->rad_info.chip_class >= GFX10) {
267 /* Logical CUs 16 - 31 */
268 radeon_set_sh_reg_idx(physical_device, cs, R_00B404_SPI_SHADER_PGM_RSRC4_HS,
269 3, S_00B404_CU_EN(0xffff));
270 radeon_set_sh_reg_idx(physical_device, cs, R_00B104_SPI_SHADER_PGM_RSRC4_VS,
271 3, S_00B104_CU_EN(0xffff));
272 radeon_set_sh_reg_idx(physical_device, cs, R_00B004_SPI_SHADER_PGM_RSRC4_PS,
273 3, S_00B004_CU_EN(0xffff));
274 }
275
276 if (physical_device->rad_info.chip_class >= GFX9) {
277 radeon_set_sh_reg_idx(physical_device, cs, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
278 3, S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
279 } else {
280 radeon_set_sh_reg(cs, R_00B51C_SPI_SHADER_PGM_RSRC3_LS,
281 S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
282 radeon_set_sh_reg(cs, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
283 S_00B41C_WAVE_LIMIT(0x3F));
284 radeon_set_sh_reg(cs, R_00B31C_SPI_SHADER_PGM_RSRC3_ES,
285 S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
286 /* If this is 0, Bonaire can hang even if GS isn't being used.
287 * Other chips are unaffected. These are suboptimal values,
288 * but we don't use on-chip GS.
289 */
290 radeon_set_context_reg(cs, R_028A44_VGT_GS_ONCHIP_CNTL,
291 S_028A44_ES_VERTS_PER_SUBGRP(64) |
292 S_028A44_GS_PRIMS_PER_SUBGRP(4));
293 }
294
295 /* Compute LATE_ALLOC_VS.LIMIT. */
296 unsigned num_cu_per_sh = physical_device->rad_info.min_good_cu_per_sa;
297 unsigned late_alloc_wave64 = 0; /* The limit is per SA. */
298 unsigned late_alloc_wave64_gs = 0;
299 unsigned cu_mask_vs = 0xffff;
300 unsigned cu_mask_gs = 0xffff;
301
302 if (physical_device->rad_info.chip_class >= GFX10) {
303 /* For Wave32, the hw will launch twice the number of late
304 * alloc waves, so 1 == 2x wave32.
305 */
306 if (!physical_device->rad_info.use_late_alloc) {
307 late_alloc_wave64 = 0;
308 } else if (num_cu_per_sh <= 6) {
309 late_alloc_wave64 = num_cu_per_sh - 2;
310 } else {
311 late_alloc_wave64 = (num_cu_per_sh - 2) * 4;
312
313 /* CU2 & CU3 disabled because of the dual CU design */
314 cu_mask_vs = 0xfff3;
315 cu_mask_gs = 0xfff3; /* NGG only */
316 }
317
318 late_alloc_wave64_gs = late_alloc_wave64;
319
320 /* Don't use late alloc for NGG on Navi14 due to a hw
321 * bug. If NGG is never used, enable all CUs.
322 */
323 if (!physical_device->use_ngg ||
324 physical_device->rad_info.family == CHIP_NAVI14) {
325 late_alloc_wave64_gs = 0;
326 cu_mask_gs = 0xffff;
327 }
328 } else {
329 if (!physical_device->rad_info.use_late_alloc) {
330 late_alloc_wave64 = 0;
331 } else if (num_cu_per_sh <= 4) {
332 /* Too few available compute units per SA.
333 * Disallowing VS to run on one CU could hurt
334 * us more than late VS allocation would help.
335 *
336 * 2 is the highest safe number that allows us
337 * to keep all CUs enabled.
338 */
339 late_alloc_wave64 = 2;
340 } else {
341 /* This is a good initial value, allowing 1
342 * late_alloc wave per SIMD on num_cu - 2.
343 */
344 late_alloc_wave64 = (num_cu_per_sh - 2) * 4;
345 }
346
347 if (late_alloc_wave64 > 2)
348 cu_mask_vs = 0xfffe; /* 1 CU disabled */
349 }
350
351 radeon_set_sh_reg_idx(physical_device, cs, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
352 3, S_00B118_CU_EN(cu_mask_vs) |
353 S_00B118_WAVE_LIMIT(0x3F));
354 radeon_set_sh_reg(cs, R_00B11C_SPI_SHADER_LATE_ALLOC_VS,
355 S_00B11C_LIMIT(late_alloc_wave64));
356
357 radeon_set_sh_reg_idx(physical_device, cs, R_00B21C_SPI_SHADER_PGM_RSRC3_GS,
358 3, S_00B21C_CU_EN(cu_mask_gs) | S_00B21C_WAVE_LIMIT(0x3F));
359
360 if (physical_device->rad_info.chip_class >= GFX10) {
361 radeon_set_sh_reg_idx(physical_device, cs, R_00B204_SPI_SHADER_PGM_RSRC4_GS,
362 3, S_00B204_CU_EN(0xffff) |
363 S_00B204_SPI_SHADER_LATE_ALLOC_GS_GFX10(late_alloc_wave64_gs));
364 }
365
366 radeon_set_sh_reg_idx(physical_device, cs, R_00B01C_SPI_SHADER_PGM_RSRC3_PS,
367 3, S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
368 }
369
370 if (physical_device->rad_info.chip_class >= GFX10) {
371 /* Break up a pixel wave if it contains deallocs for more than
372 * half the parameter cache.
373 *
374 * To avoid a deadlock where pixel waves aren't launched
375 * because they're waiting for more pixels while the frontend
376 * is stuck waiting for PC space, the maximum allowed value is
377 * the size of the PC minus the largest possible allocation for
378 * a single primitive shader subgroup.
379 */
380 radeon_set_context_reg(cs, R_028C50_PA_SC_NGG_MODE_CNTL,
381 S_028C50_MAX_DEALLOCS_IN_WAVE(512));
382 radeon_set_context_reg(cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
383
384 /* Enable CMASK/FMASK/HTILE/DCC caching in L2 for small chips. */
385 unsigned meta_write_policy, meta_read_policy;
386
387 /* TODO: investigate whether LRU improves performance on other chips too */
388 if (physical_device->rad_info.num_render_backends <= 4) {
389 meta_write_policy = V_02807C_CACHE_LRU_WR; /* cache writes */
390 meta_read_policy = V_02807C_CACHE_LRU_RD; /* cache reads */
391 } else {
392 meta_write_policy = V_02807C_CACHE_STREAM_WR; /* write combine */
393 meta_read_policy = V_02807C_CACHE_NOA_RD; /* don't cache reads */
394 }
395
396 radeon_set_context_reg(cs, R_02807C_DB_RMI_L2_CACHE_CONTROL,
397 S_02807C_Z_WR_POLICY(V_02807C_CACHE_STREAM_WR) |
398 S_02807C_S_WR_POLICY(V_02807C_CACHE_STREAM_WR) |
399 S_02807C_HTILE_WR_POLICY(meta_write_policy) |
400 S_02807C_ZPCPSD_WR_POLICY(V_02807C_CACHE_STREAM_WR) |
401 S_02807C_Z_RD_POLICY(V_02807C_CACHE_NOA_RD) |
402 S_02807C_S_RD_POLICY(V_02807C_CACHE_NOA_RD) |
403 S_02807C_HTILE_RD_POLICY(meta_read_policy));
404
405 radeon_set_context_reg(cs, R_028410_CB_RMI_GL2_CACHE_CONTROL,
406 S_028410_CMASK_WR_POLICY(meta_write_policy) |
407 S_028410_FMASK_WR_POLICY(meta_write_policy) |
408 S_028410_DCC_WR_POLICY(meta_write_policy) |
409 S_028410_COLOR_WR_POLICY(V_028410_CACHE_STREAM_WR) |
410 S_028410_CMASK_RD_POLICY(meta_read_policy) |
411 S_028410_FMASK_RD_POLICY(meta_read_policy) |
412 S_028410_DCC_RD_POLICY(meta_read_policy) |
413 S_028410_COLOR_RD_POLICY(V_028410_CACHE_NOA_RD));
414 radeon_set_context_reg(cs, R_028428_CB_COVERAGE_OUT_CONTROL, 0);
415
416 radeon_set_sh_reg(cs, R_00B0C0_SPI_SHADER_REQ_CTRL_PS,
417 S_00B0C0_SOFT_GROUPING_EN(1) |
418 S_00B0C0_NUMBER_OF_REQUESTS_PER_CU(4 - 1));
419 radeon_set_sh_reg(cs, R_00B1C0_SPI_SHADER_REQ_CTRL_VS, 0);
420
421 if (physical_device->rad_info.chip_class == GFX10) {
422 /* SQ_NON_EVENT must be emitted before GE_PC_ALLOC is written. */
423 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
424 radeon_emit(cs, EVENT_TYPE(V_028A90_SQ_NON_EVENT) | EVENT_INDEX(0));
425 }
426
427 /* TODO: For culling, replace 128 with 256. */
428 radeon_set_uconfig_reg(cs, R_030980_GE_PC_ALLOC,
429 S_030980_OVERSUB_EN(physical_device->rad_info.use_late_alloc) |
430 S_030980_NUM_PC_LINES(128 * physical_device->rad_info.max_se - 1));
431 }
432
433 if (physical_device->rad_info.chip_class >= GFX9) {
434 radeon_set_context_reg(cs, R_028B50_VGT_TESS_DISTRIBUTION,
435 S_028B50_ACCUM_ISOLINE(40) |
436 S_028B50_ACCUM_TRI(30) |
437 S_028B50_ACCUM_QUAD(24) |
438 S_028B50_DONUT_SPLIT(24) |
439 S_028B50_TRAP_SPLIT(6));
440 } else if (physical_device->rad_info.chip_class >= GFX8) {
441 uint32_t vgt_tess_distribution;
442
443 vgt_tess_distribution = S_028B50_ACCUM_ISOLINE(32) |
444 S_028B50_ACCUM_TRI(11) |
445 S_028B50_ACCUM_QUAD(11) |
446 S_028B50_DONUT_SPLIT(16);
447
448 if (physical_device->rad_info.family == CHIP_FIJI ||
449 physical_device->rad_info.family >= CHIP_POLARIS10)
450 vgt_tess_distribution |= S_028B50_TRAP_SPLIT(3);
451
452 radeon_set_context_reg(cs, R_028B50_VGT_TESS_DISTRIBUTION,
453 vgt_tess_distribution);
454 } else if (!has_clear_state) {
455 radeon_set_context_reg(cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
456 radeon_set_context_reg(cs, R_028C5C_VGT_OUT_DEALLOC_CNTL, 16);
457 }
458
459 if (device->border_color_data.bo) {
460 uint64_t border_color_va = radv_buffer_get_va(device->border_color_data.bo);
461
462 radeon_set_context_reg(cs, R_028080_TA_BC_BASE_ADDR, border_color_va >> 8);
463 if (physical_device->rad_info.chip_class >= GFX7) {
464 radeon_set_context_reg(cs, R_028084_TA_BC_BASE_ADDR_HI,
465 S_028084_ADDRESS(border_color_va >> 40));
466 }
467 }
468
469 if (physical_device->rad_info.chip_class >= GFX9) {
470 radeon_set_context_reg(cs, R_028C48_PA_SC_BINNER_CNTL_1,
471 S_028C48_MAX_ALLOC_COUNT(physical_device->rad_info.pbb_max_alloc_count - 1) |
472 S_028C48_MAX_PRIM_PER_BATCH(1023));
473 radeon_set_context_reg(cs, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
474 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
475 radeon_set_uconfig_reg(cs, R_030968_VGT_INSTANCE_BASE_ID, 0);
476 }
477
478 unsigned tmp = (unsigned)(1.0 * 8.0);
479 radeon_set_context_reg_seq(cs, R_028A00_PA_SU_POINT_SIZE, 1);
480 radeon_emit(cs, S_028A00_HEIGHT(tmp) | S_028A00_WIDTH(tmp));
481 radeon_set_context_reg_seq(cs, R_028A04_PA_SU_POINT_MINMAX, 1);
482 radeon_emit(cs, S_028A04_MIN_SIZE(radv_pack_float_12p4(0)) |
483 S_028A04_MAX_SIZE(radv_pack_float_12p4(8192/2)));
484
485 if (!has_clear_state) {
486 radeon_set_context_reg(cs, R_028004_DB_COUNT_CONTROL,
487 S_028004_ZPASS_INCREMENT_DISABLE(1));
488 }
489
490 /* Enable the Polaris small primitive filter control.
491 * XXX: There is possibly an issue when MSAA is off (see RadeonSI
492 * has_msaa_sample_loc_bug). But this doesn't seem to regress anything,
493 * and AMDVLK doesn't have a workaround as well.
494 */
495 if (physical_device->rad_info.family >= CHIP_POLARIS10) {
496 unsigned small_prim_filter_cntl =
497 S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
498 /* Workaround for a hw line bug. */
499 S_028830_LINE_FILTER_DISABLE(physical_device->rad_info.family <= CHIP_POLARIS12);
500
501 radeon_set_context_reg(cs, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL,
502 small_prim_filter_cntl);
503 }
504
505 si_emit_compute(physical_device, cs);
506 }
507
508 void
509 cik_create_gfx_config(struct radv_device *device)
510 {
511 struct radeon_cmdbuf *cs = device->ws->cs_create(device->ws, RING_GFX);
512 if (!cs)
513 return;
514
515 si_emit_graphics(device, cs);
516
517 while (cs->cdw & 7) {
518 if (device->physical_device->rad_info.gfx_ib_pad_with_type2)
519 radeon_emit(cs, 0x80000000);
520 else
521 radeon_emit(cs, 0xffff1000);
522 }
523
524 device->gfx_init = device->ws->buffer_create(device->ws,
525 cs->cdw * 4, 4096,
526 RADEON_DOMAIN_GTT,
527 RADEON_FLAG_CPU_ACCESS|
528 RADEON_FLAG_NO_INTERPROCESS_SHARING |
529 RADEON_FLAG_READ_ONLY,
530 RADV_BO_PRIORITY_CS);
531 if (!device->gfx_init)
532 goto fail;
533
534 void *map = device->ws->buffer_map(device->gfx_init);
535 if (!map) {
536 device->ws->buffer_destroy(device->gfx_init);
537 device->gfx_init = NULL;
538 goto fail;
539 }
540 memcpy(map, cs->buf, cs->cdw * 4);
541
542 device->ws->buffer_unmap(device->gfx_init);
543 device->gfx_init_size_dw = cs->cdw;
544 fail:
545 device->ws->cs_destroy(cs);
546 }
547
548 static void
549 get_viewport_xform(const VkViewport *viewport,
550 float scale[3], float translate[3])
551 {
552 float x = viewport->x;
553 float y = viewport->y;
554 float half_width = 0.5f * viewport->width;
555 float half_height = 0.5f * viewport->height;
556 double n = viewport->minDepth;
557 double f = viewport->maxDepth;
558
559 scale[0] = half_width;
560 translate[0] = half_width + x;
561 scale[1] = half_height;
562 translate[1] = half_height + y;
563
564 scale[2] = (f - n);
565 translate[2] = n;
566 }
567
568 void
569 si_write_viewport(struct radeon_cmdbuf *cs, int first_vp,
570 int count, const VkViewport *viewports)
571 {
572 int i;
573
574 assert(count);
575 radeon_set_context_reg_seq(cs, R_02843C_PA_CL_VPORT_XSCALE +
576 first_vp * 4 * 6, count * 6);
577
578 for (i = 0; i < count; i++) {
579 float scale[3], translate[3];
580
581
582 get_viewport_xform(&viewports[i], scale, translate);
583 radeon_emit(cs, fui(scale[0]));
584 radeon_emit(cs, fui(translate[0]));
585 radeon_emit(cs, fui(scale[1]));
586 radeon_emit(cs, fui(translate[1]));
587 radeon_emit(cs, fui(scale[2]));
588 radeon_emit(cs, fui(translate[2]));
589 }
590
591 radeon_set_context_reg_seq(cs, R_0282D0_PA_SC_VPORT_ZMIN_0 +
592 first_vp * 4 * 2, count * 2);
593 for (i = 0; i < count; i++) {
594 float zmin = MIN2(viewports[i].minDepth, viewports[i].maxDepth);
595 float zmax = MAX2(viewports[i].minDepth, viewports[i].maxDepth);
596 radeon_emit(cs, fui(zmin));
597 radeon_emit(cs, fui(zmax));
598 }
599 }
600
601 static VkRect2D si_scissor_from_viewport(const VkViewport *viewport)
602 {
603 float scale[3], translate[3];
604 VkRect2D rect;
605
606 get_viewport_xform(viewport, scale, translate);
607
608 rect.offset.x = translate[0] - fabsf(scale[0]);
609 rect.offset.y = translate[1] - fabsf(scale[1]);
610 rect.extent.width = ceilf(translate[0] + fabsf(scale[0])) - rect.offset.x;
611 rect.extent.height = ceilf(translate[1] + fabsf(scale[1])) - rect.offset.y;
612
613 return rect;
614 }
615
616 static VkRect2D si_intersect_scissor(const VkRect2D *a, const VkRect2D *b) {
617 VkRect2D ret;
618 ret.offset.x = MAX2(a->offset.x, b->offset.x);
619 ret.offset.y = MAX2(a->offset.y, b->offset.y);
620 ret.extent.width = MIN2(a->offset.x + a->extent.width,
621 b->offset.x + b->extent.width) - ret.offset.x;
622 ret.extent.height = MIN2(a->offset.y + a->extent.height,
623 b->offset.y + b->extent.height) - ret.offset.y;
624 return ret;
625 }
626
627 void
628 si_write_scissors(struct radeon_cmdbuf *cs, int first,
629 int count, const VkRect2D *scissors,
630 const VkViewport *viewports, bool can_use_guardband)
631 {
632 int i;
633 float scale[3], translate[3], guardband_x = INFINITY, guardband_y = INFINITY;
634 const float max_range = 32767.0f;
635 if (!count)
636 return;
637
638 radeon_set_context_reg_seq(cs, R_028250_PA_SC_VPORT_SCISSOR_0_TL + first * 4 * 2, count * 2);
639 for (i = 0; i < count; i++) {
640 VkRect2D viewport_scissor = si_scissor_from_viewport(viewports + i);
641 VkRect2D scissor = si_intersect_scissor(&scissors[i], &viewport_scissor);
642
643 get_viewport_xform(viewports + i, scale, translate);
644 scale[0] = fabsf(scale[0]);
645 scale[1] = fabsf(scale[1]);
646
647 if (scale[0] < 0.5)
648 scale[0] = 0.5;
649 if (scale[1] < 0.5)
650 scale[1] = 0.5;
651
652 guardband_x = MIN2(guardband_x, (max_range - fabsf(translate[0])) / scale[0]);
653 guardband_y = MIN2(guardband_y, (max_range - fabsf(translate[1])) / scale[1]);
654
655 radeon_emit(cs, S_028250_TL_X(scissor.offset.x) |
656 S_028250_TL_Y(scissor.offset.y) |
657 S_028250_WINDOW_OFFSET_DISABLE(1));
658 radeon_emit(cs, S_028254_BR_X(scissor.offset.x + scissor.extent.width) |
659 S_028254_BR_Y(scissor.offset.y + scissor.extent.height));
660 }
661 if (!can_use_guardband) {
662 guardband_x = 1.0;
663 guardband_y = 1.0;
664 }
665
666 radeon_set_context_reg_seq(cs, R_028BE8_PA_CL_GB_VERT_CLIP_ADJ, 4);
667 radeon_emit(cs, fui(guardband_y));
668 radeon_emit(cs, fui(1.0));
669 radeon_emit(cs, fui(guardband_x));
670 radeon_emit(cs, fui(1.0));
671 }
672
673 static inline unsigned
674 radv_prims_for_vertices(struct radv_prim_vertex_count *info, unsigned num)
675 {
676 if (num == 0)
677 return 0;
678
679 if (info->incr == 0)
680 return 0;
681
682 if (num < info->min)
683 return 0;
684
685 return 1 + ((num - info->min) / info->incr);
686 }
687
688 uint32_t
689 si_get_ia_multi_vgt_param(struct radv_cmd_buffer *cmd_buffer,
690 bool instanced_draw, bool indirect_draw,
691 bool count_from_stream_output,
692 uint32_t draw_vertex_count)
693 {
694 enum chip_class chip_class = cmd_buffer->device->physical_device->rad_info.chip_class;
695 enum radeon_family family = cmd_buffer->device->physical_device->rad_info.family;
696 struct radeon_info *info = &cmd_buffer->device->physical_device->rad_info;
697 const unsigned max_primgroup_in_wave = 2;
698 /* SWITCH_ON_EOP(0) is always preferable. */
699 bool wd_switch_on_eop = false;
700 bool ia_switch_on_eop = false;
701 bool ia_switch_on_eoi = false;
702 bool partial_vs_wave = false;
703 bool partial_es_wave = cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.partial_es_wave;
704 bool multi_instances_smaller_than_primgroup;
705
706 multi_instances_smaller_than_primgroup = indirect_draw;
707 if (!multi_instances_smaller_than_primgroup && instanced_draw) {
708 uint32_t num_prims = radv_prims_for_vertices(&cmd_buffer->state.pipeline->graphics.prim_vertex_count, draw_vertex_count);
709 if (num_prims < cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.primgroup_size)
710 multi_instances_smaller_than_primgroup = true;
711 }
712
713 ia_switch_on_eoi = cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.ia_switch_on_eoi;
714 partial_vs_wave = cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.partial_vs_wave;
715
716 if (chip_class >= GFX7) {
717 wd_switch_on_eop = cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.wd_switch_on_eop;
718
719 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
720 * We don't know that for indirect drawing, so treat it as
721 * always problematic. */
722 if (family == CHIP_HAWAII &&
723 (instanced_draw || indirect_draw))
724 wd_switch_on_eop = true;
725
726 /* Performance recommendation for 4 SE Gfx7-8 parts if
727 * instances are smaller than a primgroup.
728 * Assume indirect draws always use small instances.
729 * This is needed for good VS wave utilization.
730 */
731 if (chip_class <= GFX8 &&
732 info->max_se == 4 &&
733 multi_instances_smaller_than_primgroup)
734 wd_switch_on_eop = true;
735
736 /* Required on GFX7 and later. */
737 if (info->max_se > 2 && !wd_switch_on_eop)
738 ia_switch_on_eoi = true;
739
740 /* Required by Hawaii and, for some special cases, by GFX8. */
741 if (ia_switch_on_eoi &&
742 (family == CHIP_HAWAII ||
743 (chip_class == GFX8 &&
744 /* max primgroup in wave is always 2 - leave this for documentation */
745 (radv_pipeline_has_gs(cmd_buffer->state.pipeline) || max_primgroup_in_wave != 2))))
746 partial_vs_wave = true;
747
748 /* Instancing bug on Bonaire. */
749 if (family == CHIP_BONAIRE && ia_switch_on_eoi &&
750 (instanced_draw || indirect_draw))
751 partial_vs_wave = true;
752
753 /* Hardware requirement when drawing primitives from a stream
754 * output buffer.
755 */
756 if (count_from_stream_output)
757 wd_switch_on_eop = true;
758
759 /* If the WD switch is false, the IA switch must be false too. */
760 assert(wd_switch_on_eop || !ia_switch_on_eop);
761 }
762 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
763 if (chip_class <= GFX8 && ia_switch_on_eoi)
764 partial_es_wave = true;
765
766 if (radv_pipeline_has_gs(cmd_buffer->state.pipeline)) {
767 /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
768 * The hw doc says all multi-SE chips are affected, but amdgpu-pro Vulkan
769 * only applies it to Hawaii. Do what amdgpu-pro Vulkan does.
770 */
771 if (family == CHIP_HAWAII && ia_switch_on_eoi) {
772 bool set_vgt_flush = indirect_draw;
773 if (!set_vgt_flush && instanced_draw) {
774 uint32_t num_prims = radv_prims_for_vertices(&cmd_buffer->state.pipeline->graphics.prim_vertex_count, draw_vertex_count);
775 if (num_prims <= 1)
776 set_vgt_flush = true;
777 }
778 if (set_vgt_flush)
779 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_VGT_FLUSH;
780 }
781 }
782
783 return cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.base |
784 S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop) |
785 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi) |
786 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave) |
787 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave) |
788 S_028AA8_WD_SWITCH_ON_EOP(chip_class >= GFX7 ? wd_switch_on_eop : 0);
789
790 }
791
792 void si_cs_emit_write_event_eop(struct radeon_cmdbuf *cs,
793 enum chip_class chip_class,
794 bool is_mec,
795 unsigned event, unsigned event_flags,
796 unsigned dst_sel, unsigned data_sel,
797 uint64_t va,
798 uint32_t new_fence,
799 uint64_t gfx9_eop_bug_va)
800 {
801 unsigned op = EVENT_TYPE(event) |
802 EVENT_INDEX(event == V_028A90_CS_DONE ||
803 event == V_028A90_PS_DONE ? 6 : 5) |
804 event_flags;
805 unsigned is_gfx8_mec = is_mec && chip_class < GFX9;
806 unsigned sel = EOP_DST_SEL(dst_sel) |
807 EOP_DATA_SEL(data_sel);
808
809 /* Wait for write confirmation before writing data, but don't send
810 * an interrupt. */
811 if (data_sel != EOP_DATA_SEL_DISCARD)
812 sel |= EOP_INT_SEL(EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM);
813
814 if (chip_class >= GFX9 || is_gfx8_mec) {
815 /* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
816 * counters) must immediately precede every timestamp event to
817 * prevent a GPU hang on GFX9.
818 */
819 if (chip_class == GFX9 && !is_mec) {
820 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
821 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1));
822 radeon_emit(cs, gfx9_eop_bug_va);
823 radeon_emit(cs, gfx9_eop_bug_va >> 32);
824 }
825
826 radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, is_gfx8_mec ? 5 : 6, false));
827 radeon_emit(cs, op);
828 radeon_emit(cs, sel);
829 radeon_emit(cs, va); /* address lo */
830 radeon_emit(cs, va >> 32); /* address hi */
831 radeon_emit(cs, new_fence); /* immediate data lo */
832 radeon_emit(cs, 0); /* immediate data hi */
833 if (!is_gfx8_mec)
834 radeon_emit(cs, 0); /* unused */
835 } else {
836 if (chip_class == GFX7 ||
837 chip_class == GFX8) {
838 /* Two EOP events are required to make all engines go idle
839 * (and optional cache flushes executed) before the timestamp
840 * is written.
841 */
842 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, false));
843 radeon_emit(cs, op);
844 radeon_emit(cs, va);
845 radeon_emit(cs, ((va >> 32) & 0xffff) | sel);
846 radeon_emit(cs, 0); /* immediate data */
847 radeon_emit(cs, 0); /* unused */
848 }
849
850 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, false));
851 radeon_emit(cs, op);
852 radeon_emit(cs, va);
853 radeon_emit(cs, ((va >> 32) & 0xffff) | sel);
854 radeon_emit(cs, new_fence); /* immediate data */
855 radeon_emit(cs, 0); /* unused */
856 }
857 }
858
859 void
860 radv_cp_wait_mem(struct radeon_cmdbuf *cs, uint32_t op, uint64_t va,
861 uint32_t ref, uint32_t mask)
862 {
863 assert(op == WAIT_REG_MEM_EQUAL ||
864 op == WAIT_REG_MEM_NOT_EQUAL ||
865 op == WAIT_REG_MEM_GREATER_OR_EQUAL);
866
867 radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, false));
868 radeon_emit(cs, op | WAIT_REG_MEM_MEM_SPACE(1));
869 radeon_emit(cs, va);
870 radeon_emit(cs, va >> 32);
871 radeon_emit(cs, ref); /* reference value */
872 radeon_emit(cs, mask); /* mask */
873 radeon_emit(cs, 4); /* poll interval */
874 }
875
876 static void
877 si_emit_acquire_mem(struct radeon_cmdbuf *cs,
878 bool is_mec,
879 bool is_gfx9,
880 unsigned cp_coher_cntl)
881 {
882 if (is_mec || is_gfx9) {
883 uint32_t hi_val = is_gfx9 ? 0xffffff : 0xff;
884 radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, false) |
885 PKT3_SHADER_TYPE_S(is_mec));
886 radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
887 radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
888 radeon_emit(cs, hi_val); /* CP_COHER_SIZE_HI */
889 radeon_emit(cs, 0); /* CP_COHER_BASE */
890 radeon_emit(cs, 0); /* CP_COHER_BASE_HI */
891 radeon_emit(cs, 0x0000000A); /* POLL_INTERVAL */
892 } else {
893 /* ACQUIRE_MEM is only required on a compute ring. */
894 radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, false));
895 radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
896 radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
897 radeon_emit(cs, 0); /* CP_COHER_BASE */
898 radeon_emit(cs, 0x0000000A); /* POLL_INTERVAL */
899 }
900 }
901
902 static void
903 gfx10_cs_emit_cache_flush(struct radeon_cmdbuf *cs,
904 enum chip_class chip_class,
905 uint32_t *flush_cnt,
906 uint64_t flush_va,
907 bool is_mec,
908 enum radv_cmd_flush_bits flush_bits,
909 uint64_t gfx9_eop_bug_va)
910 {
911 uint32_t gcr_cntl = 0;
912 unsigned cb_db_event = 0;
913
914 /* We don't need these. */
915 assert(!(flush_bits & (RADV_CMD_FLAG_VGT_STREAMOUT_SYNC)));
916
917 if (flush_bits & RADV_CMD_FLAG_INV_ICACHE)
918 gcr_cntl |= S_586_GLI_INV(V_586_GLI_ALL);
919 if (flush_bits & RADV_CMD_FLAG_INV_SCACHE) {
920 /* TODO: When writing to the SMEM L1 cache, we need to set SEQ
921 * to FORWARD when both L1 and L2 are written out (WB or INV).
922 */
923 gcr_cntl |= S_586_GL1_INV(1) | S_586_GLK_INV(1);
924 }
925 if (flush_bits & RADV_CMD_FLAG_INV_VCACHE)
926 gcr_cntl |= S_586_GL1_INV(1) | S_586_GLV_INV(1);
927 if (flush_bits & RADV_CMD_FLAG_INV_L2) {
928 /* Writeback and invalidate everything in L2. */
929 gcr_cntl |= S_586_GL2_INV(1) | S_586_GL2_WB(1) |
930 S_586_GLM_INV(1) | S_586_GLM_WB(1);
931 } else if (flush_bits & RADV_CMD_FLAG_WB_L2) {
932 /* Writeback but do not invalidate.
933 * GLM doesn't support WB alone. If WB is set, INV must be set too.
934 */
935 gcr_cntl |= S_586_GL2_WB(1) |
936 S_586_GLM_WB(1) | S_586_GLM_INV(1);
937 }
938
939 /* TODO: Implement this new flag for GFX9+.
940 else if (flush_bits & RADV_CMD_FLAG_INV_L2_METADATA)
941 gcr_cntl |= S_586_GLM_INV(1) | S_586_GLM_WB(1);
942 */
943
944 if (flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB | RADV_CMD_FLAG_FLUSH_AND_INV_DB)) {
945 /* TODO: trigger on RADV_CMD_FLAG_FLUSH_AND_INV_CB_META */
946 if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB) {
947 /* Flush CMASK/FMASK/DCC. Will wait for idle later. */
948 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
949 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) |
950 EVENT_INDEX(0));
951 }
952
953 /* TODO: trigger on RADV_CMD_FLAG_FLUSH_AND_INV_DB_META ? */
954 if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB) {
955 /* Flush HTILE. Will wait for idle later. */
956 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
957 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) |
958 EVENT_INDEX(0));
959 }
960
961 /* First flush CB/DB, then L1/L2. */
962 gcr_cntl |= S_586_SEQ(V_586_SEQ_FORWARD);
963
964 if ((flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB | RADV_CMD_FLAG_FLUSH_AND_INV_DB)) ==
965 (RADV_CMD_FLAG_FLUSH_AND_INV_CB | RADV_CMD_FLAG_FLUSH_AND_INV_DB)) {
966 cb_db_event = V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT;
967 } else if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB) {
968 cb_db_event = V_028A90_FLUSH_AND_INV_CB_DATA_TS;
969 } else if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB) {
970 cb_db_event = V_028A90_FLUSH_AND_INV_DB_DATA_TS;
971 } else {
972 assert(0);
973 }
974 } else {
975 /* Wait for graphics shaders to go idle if requested. */
976 if (flush_bits & RADV_CMD_FLAG_PS_PARTIAL_FLUSH) {
977 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
978 radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
979 } else if (flush_bits & RADV_CMD_FLAG_VS_PARTIAL_FLUSH) {
980 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
981 radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
982 }
983 }
984
985 if (flush_bits & RADV_CMD_FLAG_CS_PARTIAL_FLUSH) {
986 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
987 radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH | EVENT_INDEX(4)));
988 }
989
990 if (cb_db_event) {
991 /* CB/DB flush and invalidate (or possibly just a wait for a
992 * meta flush) via RELEASE_MEM.
993 *
994 * Combine this with other cache flushes when possible; this
995 * requires affected shaders to be idle, so do it after the
996 * CS_PARTIAL_FLUSH before (VS/PS partial flushes are always
997 * implied).
998 */
999 /* Get GCR_CNTL fields, because the encoding is different in RELEASE_MEM. */
1000 unsigned glm_wb = G_586_GLM_WB(gcr_cntl);
1001 unsigned glm_inv = G_586_GLM_INV(gcr_cntl);
1002 unsigned glv_inv = G_586_GLV_INV(gcr_cntl);
1003 unsigned gl1_inv = G_586_GL1_INV(gcr_cntl);
1004 assert(G_586_GL2_US(gcr_cntl) == 0);
1005 assert(G_586_GL2_RANGE(gcr_cntl) == 0);
1006 assert(G_586_GL2_DISCARD(gcr_cntl) == 0);
1007 unsigned gl2_inv = G_586_GL2_INV(gcr_cntl);
1008 unsigned gl2_wb = G_586_GL2_WB(gcr_cntl);
1009 unsigned gcr_seq = G_586_SEQ(gcr_cntl);
1010
1011 gcr_cntl &= C_586_GLM_WB &
1012 C_586_GLM_INV &
1013 C_586_GLV_INV &
1014 C_586_GL1_INV &
1015 C_586_GL2_INV &
1016 C_586_GL2_WB; /* keep SEQ */
1017
1018 assert(flush_cnt);
1019 (*flush_cnt)++;
1020
1021 si_cs_emit_write_event_eop(cs, chip_class, false, cb_db_event,
1022 S_490_GLM_WB(glm_wb) |
1023 S_490_GLM_INV(glm_inv) |
1024 S_490_GLV_INV(glv_inv) |
1025 S_490_GL1_INV(gl1_inv) |
1026 S_490_GL2_INV(gl2_inv) |
1027 S_490_GL2_WB(gl2_wb) |
1028 S_490_SEQ(gcr_seq),
1029 EOP_DST_SEL_MEM,
1030 EOP_DATA_SEL_VALUE_32BIT,
1031 flush_va, *flush_cnt,
1032 gfx9_eop_bug_va);
1033
1034 radv_cp_wait_mem(cs, WAIT_REG_MEM_EQUAL, flush_va,
1035 *flush_cnt, 0xffffffff);
1036 }
1037
1038 /* VGT state sync */
1039 if (flush_bits & RADV_CMD_FLAG_VGT_FLUSH) {
1040 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1041 radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
1042 }
1043
1044 /* Ignore fields that only modify the behavior of other fields. */
1045 if (gcr_cntl & C_586_GL1_RANGE & C_586_GL2_RANGE & C_586_SEQ) {
1046 /* Flush caches and wait for the caches to assert idle.
1047 * The cache flush is executed in the ME, but the PFP waits
1048 * for completion.
1049 */
1050 radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 6, 0));
1051 radeon_emit(cs, 0); /* CP_COHER_CNTL */
1052 radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
1053 radeon_emit(cs, 0xffffff); /* CP_COHER_SIZE_HI */
1054 radeon_emit(cs, 0); /* CP_COHER_BASE */
1055 radeon_emit(cs, 0); /* CP_COHER_BASE_HI */
1056 radeon_emit(cs, 0x0000000A); /* POLL_INTERVAL */
1057 radeon_emit(cs, gcr_cntl); /* GCR_CNTL */
1058 } else if ((cb_db_event ||
1059 (flush_bits & (RADV_CMD_FLAG_VS_PARTIAL_FLUSH |
1060 RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
1061 RADV_CMD_FLAG_CS_PARTIAL_FLUSH)))
1062 && !is_mec) {
1063 /* We need to ensure that PFP waits as well. */
1064 radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
1065 radeon_emit(cs, 0);
1066 }
1067
1068 if (flush_bits & RADV_CMD_FLAG_START_PIPELINE_STATS) {
1069 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1070 radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_START) |
1071 EVENT_INDEX(0));
1072 } else if (flush_bits & RADV_CMD_FLAG_STOP_PIPELINE_STATS) {
1073 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1074 radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP) |
1075 EVENT_INDEX(0));
1076 }
1077 }
1078
1079 void
1080 si_cs_emit_cache_flush(struct radeon_cmdbuf *cs,
1081 enum chip_class chip_class,
1082 uint32_t *flush_cnt,
1083 uint64_t flush_va,
1084 bool is_mec,
1085 enum radv_cmd_flush_bits flush_bits,
1086 uint64_t gfx9_eop_bug_va)
1087 {
1088 unsigned cp_coher_cntl = 0;
1089 uint32_t flush_cb_db = flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
1090 RADV_CMD_FLAG_FLUSH_AND_INV_DB);
1091
1092 if (chip_class >= GFX10) {
1093 /* GFX10 cache flush handling is quite different. */
1094 gfx10_cs_emit_cache_flush(cs, chip_class, flush_cnt, flush_va,
1095 is_mec, flush_bits, gfx9_eop_bug_va);
1096 return;
1097 }
1098
1099 if (flush_bits & RADV_CMD_FLAG_INV_ICACHE)
1100 cp_coher_cntl |= S_0085F0_SH_ICACHE_ACTION_ENA(1);
1101 if (flush_bits & RADV_CMD_FLAG_INV_SCACHE)
1102 cp_coher_cntl |= S_0085F0_SH_KCACHE_ACTION_ENA(1);
1103
1104 if (chip_class <= GFX8) {
1105 if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB) {
1106 cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) |
1107 S_0085F0_CB0_DEST_BASE_ENA(1) |
1108 S_0085F0_CB1_DEST_BASE_ENA(1) |
1109 S_0085F0_CB2_DEST_BASE_ENA(1) |
1110 S_0085F0_CB3_DEST_BASE_ENA(1) |
1111 S_0085F0_CB4_DEST_BASE_ENA(1) |
1112 S_0085F0_CB5_DEST_BASE_ENA(1) |
1113 S_0085F0_CB6_DEST_BASE_ENA(1) |
1114 S_0085F0_CB7_DEST_BASE_ENA(1);
1115
1116 /* Necessary for DCC */
1117 if (chip_class >= GFX8) {
1118 si_cs_emit_write_event_eop(cs,
1119 chip_class,
1120 is_mec,
1121 V_028A90_FLUSH_AND_INV_CB_DATA_TS,
1122 0,
1123 EOP_DST_SEL_MEM,
1124 EOP_DATA_SEL_DISCARD,
1125 0, 0,
1126 gfx9_eop_bug_va);
1127 }
1128 }
1129 if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB) {
1130 cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) |
1131 S_0085F0_DB_DEST_BASE_ENA(1);
1132 }
1133 }
1134
1135 if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB_META) {
1136 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1137 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
1138 }
1139
1140 if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB_META) {
1141 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1142 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
1143 }
1144
1145 if (flush_bits & RADV_CMD_FLAG_PS_PARTIAL_FLUSH) {
1146 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1147 radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
1148 } else if (flush_bits & RADV_CMD_FLAG_VS_PARTIAL_FLUSH) {
1149 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1150 radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
1151 }
1152
1153 if (flush_bits & RADV_CMD_FLAG_CS_PARTIAL_FLUSH) {
1154 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1155 radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH) | EVENT_INDEX(4));
1156 }
1157
1158 if (chip_class == GFX9 && flush_cb_db) {
1159 unsigned cb_db_event, tc_flags;
1160
1161 /* Set the CB/DB flush event. */
1162 cb_db_event = V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT;
1163
1164 /* These are the only allowed combinations. If you need to
1165 * do multiple operations at once, do them separately.
1166 * All operations that invalidate L2 also seem to invalidate
1167 * metadata. Volatile (VOL) and WC flushes are not listed here.
1168 *
1169 * TC | TC_WB = writeback & invalidate L2 & L1
1170 * TC | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
1171 * TC_WB | TC_NC = writeback L2 for MTYPE == NC
1172 * TC | TC_NC = invalidate L2 for MTYPE == NC
1173 * TC | TC_MD = writeback & invalidate L2 metadata (DCC, etc.)
1174 * TCL1 = invalidate L1
1175 */
1176 tc_flags = EVENT_TC_ACTION_ENA |
1177 EVENT_TC_MD_ACTION_ENA;
1178
1179 /* Ideally flush TC together with CB/DB. */
1180 if (flush_bits & RADV_CMD_FLAG_INV_L2) {
1181 /* Writeback and invalidate everything in L2 & L1. */
1182 tc_flags = EVENT_TC_ACTION_ENA |
1183 EVENT_TC_WB_ACTION_ENA;
1184
1185
1186 /* Clear the flags. */
1187 flush_bits &= ~(RADV_CMD_FLAG_INV_L2 |
1188 RADV_CMD_FLAG_WB_L2 |
1189 RADV_CMD_FLAG_INV_VCACHE);
1190 }
1191 assert(flush_cnt);
1192 (*flush_cnt)++;
1193
1194 si_cs_emit_write_event_eop(cs, chip_class, false, cb_db_event, tc_flags,
1195 EOP_DST_SEL_MEM,
1196 EOP_DATA_SEL_VALUE_32BIT,
1197 flush_va, *flush_cnt,
1198 gfx9_eop_bug_va);
1199 radv_cp_wait_mem(cs, WAIT_REG_MEM_EQUAL, flush_va,
1200 *flush_cnt, 0xffffffff);
1201 }
1202
1203 /* VGT state sync */
1204 if (flush_bits & RADV_CMD_FLAG_VGT_FLUSH) {
1205 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1206 radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
1207 }
1208
1209 /* VGT streamout state sync */
1210 if (flush_bits & RADV_CMD_FLAG_VGT_STREAMOUT_SYNC) {
1211 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1212 radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC) | EVENT_INDEX(0));
1213 }
1214
1215 /* Make sure ME is idle (it executes most packets) before continuing.
1216 * This prevents read-after-write hazards between PFP and ME.
1217 */
1218 if ((cp_coher_cntl ||
1219 (flush_bits & (RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
1220 RADV_CMD_FLAG_INV_VCACHE |
1221 RADV_CMD_FLAG_INV_L2 |
1222 RADV_CMD_FLAG_WB_L2))) &&
1223 !is_mec) {
1224 radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
1225 radeon_emit(cs, 0);
1226 }
1227
1228 if ((flush_bits & RADV_CMD_FLAG_INV_L2) ||
1229 (chip_class <= GFX7 && (flush_bits & RADV_CMD_FLAG_WB_L2))) {
1230 si_emit_acquire_mem(cs, is_mec, chip_class == GFX9,
1231 cp_coher_cntl |
1232 S_0085F0_TC_ACTION_ENA(1) |
1233 S_0085F0_TCL1_ACTION_ENA(1) |
1234 S_0301F0_TC_WB_ACTION_ENA(chip_class >= GFX8));
1235 cp_coher_cntl = 0;
1236 } else {
1237 if(flush_bits & RADV_CMD_FLAG_WB_L2) {
1238 /* WB = write-back
1239 * NC = apply to non-coherent MTYPEs
1240 * (i.e. MTYPE <= 1, which is what we use everywhere)
1241 *
1242 * WB doesn't work without NC.
1243 */
1244 si_emit_acquire_mem(cs, is_mec,
1245 chip_class == GFX9,
1246 cp_coher_cntl |
1247 S_0301F0_TC_WB_ACTION_ENA(1) |
1248 S_0301F0_TC_NC_ACTION_ENA(1));
1249 cp_coher_cntl = 0;
1250 }
1251 if (flush_bits & RADV_CMD_FLAG_INV_VCACHE) {
1252 si_emit_acquire_mem(cs, is_mec,
1253 chip_class == GFX9,
1254 cp_coher_cntl |
1255 S_0085F0_TCL1_ACTION_ENA(1));
1256 cp_coher_cntl = 0;
1257 }
1258 }
1259
1260 /* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
1261 * Therefore, it should be last. Done in PFP.
1262 */
1263 if (cp_coher_cntl)
1264 si_emit_acquire_mem(cs, is_mec, chip_class == GFX9, cp_coher_cntl);
1265
1266 if (flush_bits & RADV_CMD_FLAG_START_PIPELINE_STATS) {
1267 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1268 radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_START) |
1269 EVENT_INDEX(0));
1270 } else if (flush_bits & RADV_CMD_FLAG_STOP_PIPELINE_STATS) {
1271 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1272 radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP) |
1273 EVENT_INDEX(0));
1274 }
1275 }
1276
1277 void
1278 si_emit_cache_flush(struct radv_cmd_buffer *cmd_buffer)
1279 {
1280 bool is_compute = cmd_buffer->queue_family_index == RADV_QUEUE_COMPUTE;
1281
1282 if (is_compute)
1283 cmd_buffer->state.flush_bits &= ~(RADV_CMD_FLAG_FLUSH_AND_INV_CB |
1284 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
1285 RADV_CMD_FLAG_FLUSH_AND_INV_DB |
1286 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META |
1287 RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
1288 RADV_CMD_FLAG_VS_PARTIAL_FLUSH |
1289 RADV_CMD_FLAG_VGT_FLUSH |
1290 RADV_CMD_FLAG_START_PIPELINE_STATS |
1291 RADV_CMD_FLAG_STOP_PIPELINE_STATS);
1292
1293 if (!cmd_buffer->state.flush_bits)
1294 return;
1295
1296 radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 128);
1297
1298 si_cs_emit_cache_flush(cmd_buffer->cs,
1299 cmd_buffer->device->physical_device->rad_info.chip_class,
1300 &cmd_buffer->gfx9_fence_idx,
1301 cmd_buffer->gfx9_fence_va,
1302 radv_cmd_buffer_uses_mec(cmd_buffer),
1303 cmd_buffer->state.flush_bits,
1304 cmd_buffer->gfx9_eop_bug_va);
1305
1306
1307 if (unlikely(cmd_buffer->device->trace_bo))
1308 radv_cmd_buffer_trace_emit(cmd_buffer);
1309
1310 /* Clear the caches that have been flushed to avoid syncing too much
1311 * when there is some pending active queries.
1312 */
1313 cmd_buffer->active_query_flush_bits &= ~cmd_buffer->state.flush_bits;
1314
1315 cmd_buffer->state.flush_bits = 0;
1316
1317 /* If the driver used a compute shader for resetting a query pool, it
1318 * should be finished at this point.
1319 */
1320 cmd_buffer->pending_reset_query = false;
1321 }
1322
1323 /* sets the CP predication state using a boolean stored at va */
1324 void
1325 si_emit_set_predication_state(struct radv_cmd_buffer *cmd_buffer,
1326 bool draw_visible, uint64_t va)
1327 {
1328 uint32_t op = 0;
1329
1330 if (va) {
1331 op = PRED_OP(PREDICATION_OP_BOOL64);
1332
1333 /* PREDICATION_DRAW_VISIBLE means that if the 32-bit value is
1334 * zero, all rendering commands are discarded. Otherwise, they
1335 * are discarded if the value is non zero.
1336 */
1337 op |= draw_visible ? PREDICATION_DRAW_VISIBLE :
1338 PREDICATION_DRAW_NOT_VISIBLE;
1339 }
1340 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
1341 radeon_emit(cmd_buffer->cs, PKT3(PKT3_SET_PREDICATION, 2, 0));
1342 radeon_emit(cmd_buffer->cs, op);
1343 radeon_emit(cmd_buffer->cs, va);
1344 radeon_emit(cmd_buffer->cs, va >> 32);
1345 } else {
1346 radeon_emit(cmd_buffer->cs, PKT3(PKT3_SET_PREDICATION, 1, 0));
1347 radeon_emit(cmd_buffer->cs, va);
1348 radeon_emit(cmd_buffer->cs, op | ((va >> 32) & 0xFF));
1349 }
1350 }
1351
1352 /* Set this if you want the 3D engine to wait until CP DMA is done.
1353 * It should be set on the last CP DMA packet. */
1354 #define CP_DMA_SYNC (1 << 0)
1355
1356 /* Set this if the source data was used as a destination in a previous CP DMA
1357 * packet. It's for preventing a read-after-write (RAW) hazard between two
1358 * CP DMA packets. */
1359 #define CP_DMA_RAW_WAIT (1 << 1)
1360 #define CP_DMA_USE_L2 (1 << 2)
1361 #define CP_DMA_CLEAR (1 << 3)
1362
1363 /* Alignment for optimal performance. */
1364 #define SI_CPDMA_ALIGNMENT 32
1365
1366 /* The max number of bytes that can be copied per packet. */
1367 static inline unsigned cp_dma_max_byte_count(struct radv_cmd_buffer *cmd_buffer)
1368 {
1369 unsigned max = cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9 ?
1370 S_414_BYTE_COUNT_GFX9(~0u) :
1371 S_414_BYTE_COUNT_GFX6(~0u);
1372
1373 /* make it aligned for optimal performance */
1374 return max & ~(SI_CPDMA_ALIGNMENT - 1);
1375 }
1376
1377 /* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
1378 * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
1379 * clear value.
1380 */
1381 static void si_emit_cp_dma(struct radv_cmd_buffer *cmd_buffer,
1382 uint64_t dst_va, uint64_t src_va,
1383 unsigned size, unsigned flags)
1384 {
1385 struct radeon_cmdbuf *cs = cmd_buffer->cs;
1386 uint32_t header = 0, command = 0;
1387
1388 assert(size <= cp_dma_max_byte_count(cmd_buffer));
1389
1390 radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 9);
1391 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9)
1392 command |= S_414_BYTE_COUNT_GFX9(size);
1393 else
1394 command |= S_414_BYTE_COUNT_GFX6(size);
1395
1396 /* Sync flags. */
1397 if (flags & CP_DMA_SYNC)
1398 header |= S_411_CP_SYNC(1);
1399 else {
1400 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9)
1401 command |= S_414_DISABLE_WR_CONFIRM_GFX9(1);
1402 else
1403 command |= S_414_DISABLE_WR_CONFIRM_GFX6(1);
1404 }
1405
1406 if (flags & CP_DMA_RAW_WAIT)
1407 command |= S_414_RAW_WAIT(1);
1408
1409 /* Src and dst flags. */
1410 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9 &&
1411 !(flags & CP_DMA_CLEAR) &&
1412 src_va == dst_va)
1413 header |= S_411_DST_SEL(V_411_NOWHERE); /* prefetch only */
1414 else if (flags & CP_DMA_USE_L2)
1415 header |= S_411_DST_SEL(V_411_DST_ADDR_TC_L2);
1416
1417 if (flags & CP_DMA_CLEAR)
1418 header |= S_411_SRC_SEL(V_411_DATA);
1419 else if (flags & CP_DMA_USE_L2)
1420 header |= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2);
1421
1422 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX7) {
1423 radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, cmd_buffer->state.predicating));
1424 radeon_emit(cs, header);
1425 radeon_emit(cs, src_va); /* SRC_ADDR_LO [31:0] */
1426 radeon_emit(cs, src_va >> 32); /* SRC_ADDR_HI [31:0] */
1427 radeon_emit(cs, dst_va); /* DST_ADDR_LO [31:0] */
1428 radeon_emit(cs, dst_va >> 32); /* DST_ADDR_HI [31:0] */
1429 radeon_emit(cs, command);
1430 } else {
1431 assert(!(flags & CP_DMA_USE_L2));
1432 header |= S_411_SRC_ADDR_HI(src_va >> 32);
1433 radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, cmd_buffer->state.predicating));
1434 radeon_emit(cs, src_va); /* SRC_ADDR_LO [31:0] */
1435 radeon_emit(cs, header); /* SRC_ADDR_HI [15:0] + flags. */
1436 radeon_emit(cs, dst_va); /* DST_ADDR_LO [31:0] */
1437 radeon_emit(cs, (dst_va >> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
1438 radeon_emit(cs, command);
1439 }
1440
1441 /* CP DMA is executed in ME, but index buffers are read by PFP.
1442 * This ensures that ME (CP DMA) is idle before PFP starts fetching
1443 * indices. If we wanted to execute CP DMA in PFP, this packet
1444 * should precede it.
1445 */
1446 if (flags & CP_DMA_SYNC) {
1447 if (cmd_buffer->queue_family_index == RADV_QUEUE_GENERAL) {
1448 radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, cmd_buffer->state.predicating));
1449 radeon_emit(cs, 0);
1450 }
1451
1452 /* CP will see the sync flag and wait for all DMAs to complete. */
1453 cmd_buffer->state.dma_is_busy = false;
1454 }
1455
1456 if (unlikely(cmd_buffer->device->trace_bo))
1457 radv_cmd_buffer_trace_emit(cmd_buffer);
1458 }
1459
1460 void si_cp_dma_prefetch(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
1461 unsigned size)
1462 {
1463 uint64_t aligned_va = va & ~(SI_CPDMA_ALIGNMENT - 1);
1464 uint64_t aligned_size = ((va + size + SI_CPDMA_ALIGNMENT -1) & ~(SI_CPDMA_ALIGNMENT - 1)) - aligned_va;
1465
1466 si_emit_cp_dma(cmd_buffer, aligned_va, aligned_va,
1467 aligned_size, CP_DMA_USE_L2);
1468 }
1469
1470 static void si_cp_dma_prepare(struct radv_cmd_buffer *cmd_buffer, uint64_t byte_count,
1471 uint64_t remaining_size, unsigned *flags)
1472 {
1473
1474 /* Flush the caches for the first copy only.
1475 * Also wait for the previous CP DMA operations.
1476 */
1477 if (cmd_buffer->state.flush_bits) {
1478 si_emit_cache_flush(cmd_buffer);
1479 *flags |= CP_DMA_RAW_WAIT;
1480 }
1481
1482 /* Do the synchronization after the last dma, so that all data
1483 * is written to memory.
1484 */
1485 if (byte_count == remaining_size)
1486 *flags |= CP_DMA_SYNC;
1487 }
1488
1489 static void si_cp_dma_realign_engine(struct radv_cmd_buffer *cmd_buffer, unsigned size)
1490 {
1491 uint64_t va;
1492 uint32_t offset;
1493 unsigned dma_flags = 0;
1494 unsigned buf_size = SI_CPDMA_ALIGNMENT * 2;
1495 void *ptr;
1496
1497 assert(size < SI_CPDMA_ALIGNMENT);
1498
1499 radv_cmd_buffer_upload_alloc(cmd_buffer, buf_size, SI_CPDMA_ALIGNMENT, &offset, &ptr);
1500
1501 va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
1502 va += offset;
1503
1504 si_cp_dma_prepare(cmd_buffer, size, size, &dma_flags);
1505
1506 si_emit_cp_dma(cmd_buffer, va, va + SI_CPDMA_ALIGNMENT, size,
1507 dma_flags);
1508 }
1509
1510 void si_cp_dma_buffer_copy(struct radv_cmd_buffer *cmd_buffer,
1511 uint64_t src_va, uint64_t dest_va,
1512 uint64_t size)
1513 {
1514 uint64_t main_src_va, main_dest_va;
1515 uint64_t skipped_size = 0, realign_size = 0;
1516
1517 /* Assume that we are not going to sync after the last DMA operation. */
1518 cmd_buffer->state.dma_is_busy = true;
1519
1520 if (cmd_buffer->device->physical_device->rad_info.family <= CHIP_CARRIZO ||
1521 cmd_buffer->device->physical_device->rad_info.family == CHIP_STONEY) {
1522 /* If the size is not aligned, we must add a dummy copy at the end
1523 * just to align the internal counter. Otherwise, the DMA engine
1524 * would slow down by an order of magnitude for following copies.
1525 */
1526 if (size % SI_CPDMA_ALIGNMENT)
1527 realign_size = SI_CPDMA_ALIGNMENT - (size % SI_CPDMA_ALIGNMENT);
1528
1529 /* If the copy begins unaligned, we must start copying from the next
1530 * aligned block and the skipped part should be copied after everything
1531 * else has been copied. Only the src alignment matters, not dst.
1532 */
1533 if (src_va % SI_CPDMA_ALIGNMENT) {
1534 skipped_size = SI_CPDMA_ALIGNMENT - (src_va % SI_CPDMA_ALIGNMENT);
1535 /* The main part will be skipped if the size is too small. */
1536 skipped_size = MIN2(skipped_size, size);
1537 size -= skipped_size;
1538 }
1539 }
1540 main_src_va = src_va + skipped_size;
1541 main_dest_va = dest_va + skipped_size;
1542
1543 while (size) {
1544 unsigned dma_flags = 0;
1545 unsigned byte_count = MIN2(size, cp_dma_max_byte_count(cmd_buffer));
1546
1547 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX10) {
1548 /* DMA operations via L2 are coherent and faster.
1549 * TODO: GFX7-GFX9 should also support this but it
1550 * requires tests/benchmarks.
1551 */
1552 dma_flags |= CP_DMA_USE_L2;
1553 }
1554
1555 si_cp_dma_prepare(cmd_buffer, byte_count,
1556 size + skipped_size + realign_size,
1557 &dma_flags);
1558
1559 dma_flags &= ~CP_DMA_SYNC;
1560
1561 si_emit_cp_dma(cmd_buffer, main_dest_va, main_src_va,
1562 byte_count, dma_flags);
1563
1564 size -= byte_count;
1565 main_src_va += byte_count;
1566 main_dest_va += byte_count;
1567 }
1568
1569 if (skipped_size) {
1570 unsigned dma_flags = 0;
1571
1572 si_cp_dma_prepare(cmd_buffer, skipped_size,
1573 size + skipped_size + realign_size,
1574 &dma_flags);
1575
1576 si_emit_cp_dma(cmd_buffer, dest_va, src_va,
1577 skipped_size, dma_flags);
1578 }
1579 if (realign_size)
1580 si_cp_dma_realign_engine(cmd_buffer, realign_size);
1581 }
1582
1583 void si_cp_dma_clear_buffer(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
1584 uint64_t size, unsigned value)
1585 {
1586
1587 if (!size)
1588 return;
1589
1590 assert(va % 4 == 0 && size % 4 == 0);
1591
1592 /* Assume that we are not going to sync after the last DMA operation. */
1593 cmd_buffer->state.dma_is_busy = true;
1594
1595 while (size) {
1596 unsigned byte_count = MIN2(size, cp_dma_max_byte_count(cmd_buffer));
1597 unsigned dma_flags = CP_DMA_CLEAR;
1598
1599 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX10) {
1600 /* DMA operations via L2 are coherent and faster.
1601 * TODO: GFX7-GFX9 should also support this but it
1602 * requires tests/benchmarks.
1603 */
1604 dma_flags |= CP_DMA_USE_L2;
1605 }
1606
1607 si_cp_dma_prepare(cmd_buffer, byte_count, size, &dma_flags);
1608
1609 /* Emit the clear packet. */
1610 si_emit_cp_dma(cmd_buffer, va, value, byte_count,
1611 dma_flags);
1612
1613 size -= byte_count;
1614 va += byte_count;
1615 }
1616 }
1617
1618 void si_cp_dma_wait_for_idle(struct radv_cmd_buffer *cmd_buffer)
1619 {
1620 if (cmd_buffer->device->physical_device->rad_info.chip_class < GFX7)
1621 return;
1622
1623 if (!cmd_buffer->state.dma_is_busy)
1624 return;
1625
1626 /* Issue a dummy DMA that copies zero bytes.
1627 *
1628 * The DMA engine will see that there's no work to do and skip this
1629 * DMA request, however, the CP will see the sync flag and still wait
1630 * for all DMAs to complete.
1631 */
1632 si_emit_cp_dma(cmd_buffer, 0, 0, 0, CP_DMA_SYNC);
1633
1634 cmd_buffer->state.dma_is_busy = false;
1635 }
1636
1637 /* For MSAA sample positions. */
1638 #define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y) \
1639 ((((unsigned)(s0x) & 0xf) << 0) | (((unsigned)(s0y) & 0xf) << 4) | \
1640 (((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) | \
1641 (((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) | \
1642 (((unsigned)(s3x) & 0xf) << 24) | (((unsigned)(s3y) & 0xf) << 28))
1643
1644 /* For obtaining location coordinates from registers */
1645 #define SEXT4(x) ((int)((x) | ((x) & 0x8 ? 0xfffffff0 : 0)))
1646 #define GET_SFIELD(reg, index) SEXT4(((reg) >> ((index) * 4)) & 0xf)
1647 #define GET_SX(reg, index) GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2)
1648 #define GET_SY(reg, index) GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2 + 1)
1649
1650 /* 1x MSAA */
1651 static const uint32_t sample_locs_1x =
1652 FILL_SREG(0, 0, 0, 0, 0, 0, 0, 0);
1653 static const unsigned max_dist_1x = 0;
1654 static const uint64_t centroid_priority_1x = 0x0000000000000000ull;
1655
1656 /* 2xMSAA */
1657 static const uint32_t sample_locs_2x =
1658 FILL_SREG(4,4, -4, -4, 0, 0, 0, 0);
1659 static const unsigned max_dist_2x = 4;
1660 static const uint64_t centroid_priority_2x = 0x1010101010101010ull;
1661
1662 /* 4xMSAA */
1663 static const uint32_t sample_locs_4x =
1664 FILL_SREG(-2,-6, 6, -2, -6, 2, 2, 6);
1665 static const unsigned max_dist_4x = 6;
1666 static const uint64_t centroid_priority_4x = 0x3210321032103210ull;
1667
1668 /* 8xMSAA */
1669 static const uint32_t sample_locs_8x[] = {
1670 FILL_SREG( 1,-3, -1, 3, 5, 1, -3,-5),
1671 FILL_SREG(-5, 5, -7,-1, 3, 7, 7,-7),
1672 /* The following are unused by hardware, but we emit them to IBs
1673 * instead of multiple SET_CONTEXT_REG packets. */
1674 0,
1675 0,
1676 };
1677 static const unsigned max_dist_8x = 7;
1678 static const uint64_t centroid_priority_8x = 0x7654321076543210ull;
1679
1680 unsigned radv_get_default_max_sample_dist(int log_samples)
1681 {
1682 unsigned max_dist[] = {
1683 max_dist_1x,
1684 max_dist_2x,
1685 max_dist_4x,
1686 max_dist_8x,
1687 };
1688 return max_dist[log_samples];
1689 }
1690
1691 void radv_emit_default_sample_locations(struct radeon_cmdbuf *cs, int nr_samples)
1692 {
1693 switch (nr_samples) {
1694 default:
1695 case 1:
1696 radeon_set_context_reg_seq(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 2);
1697 radeon_emit(cs, (uint32_t)centroid_priority_1x);
1698 radeon_emit(cs, centroid_priority_1x >> 32);
1699 radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, sample_locs_1x);
1700 radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, sample_locs_1x);
1701 radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, sample_locs_1x);
1702 radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, sample_locs_1x);
1703 break;
1704 case 2:
1705 radeon_set_context_reg_seq(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 2);
1706 radeon_emit(cs, (uint32_t)centroid_priority_2x);
1707 radeon_emit(cs, centroid_priority_2x >> 32);
1708 radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, sample_locs_2x);
1709 radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, sample_locs_2x);
1710 radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, sample_locs_2x);
1711 radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, sample_locs_2x);
1712 break;
1713 case 4:
1714 radeon_set_context_reg_seq(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 2);
1715 radeon_emit(cs, (uint32_t)centroid_priority_4x);
1716 radeon_emit(cs, centroid_priority_4x >> 32);
1717 radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, sample_locs_4x);
1718 radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, sample_locs_4x);
1719 radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, sample_locs_4x);
1720 radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, sample_locs_4x);
1721 break;
1722 case 8:
1723 radeon_set_context_reg_seq(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 2);
1724 radeon_emit(cs, (uint32_t)centroid_priority_8x);
1725 radeon_emit(cs, centroid_priority_8x >> 32);
1726 radeon_set_context_reg_seq(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 14);
1727 radeon_emit_array(cs, sample_locs_8x, 4);
1728 radeon_emit_array(cs, sample_locs_8x, 4);
1729 radeon_emit_array(cs, sample_locs_8x, 4);
1730 radeon_emit_array(cs, sample_locs_8x, 2);
1731 break;
1732 }
1733 }
1734
1735 static void radv_get_sample_position(struct radv_device *device,
1736 unsigned sample_count,
1737 unsigned sample_index, float *out_value)
1738 {
1739 const uint32_t *sample_locs;
1740
1741 switch (sample_count) {
1742 case 1:
1743 default:
1744 sample_locs = &sample_locs_1x;
1745 break;
1746 case 2:
1747 sample_locs = &sample_locs_2x;
1748 break;
1749 case 4:
1750 sample_locs = &sample_locs_4x;
1751 break;
1752 case 8:
1753 sample_locs = sample_locs_8x;
1754 break;
1755 }
1756
1757 out_value[0] = (GET_SX(sample_locs, sample_index) + 8) / 16.0f;
1758 out_value[1] = (GET_SY(sample_locs, sample_index) + 8) / 16.0f;
1759 }
1760
1761 void radv_device_init_msaa(struct radv_device *device)
1762 {
1763 int i;
1764
1765 radv_get_sample_position(device, 1, 0, device->sample_locations_1x[0]);
1766
1767 for (i = 0; i < 2; i++)
1768 radv_get_sample_position(device, 2, i, device->sample_locations_2x[i]);
1769 for (i = 0; i < 4; i++)
1770 radv_get_sample_position(device, 4, i, device->sample_locations_4x[i]);
1771 for (i = 0; i < 8; i++)
1772 radv_get_sample_position(device, 8, i, device->sample_locations_8x[i]);
1773 }