1 /*
2 * Copyright (c) 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <stdlib.h>
25 #include <math.h>
26
27 #include "util/macros.h"
28
29 #include "intel_l3_config.h"
30
31 struct intel_l3_list {
32 const struct intel_l3_config *configs;
33 int length;
34 };
35
36 #define DECLARE_L3_LIST(hw) \
37 struct intel_l3_list hw##_l3_list = \
38 { .configs = hw##_l3_configs, .length = ARRAY_SIZE(hw##_l3_configs) }
39
40 /**
41 * IVB/HSW validated L3 configurations. The first entry will be used as
42 * default by gfx7_restore_default_l3_config(), otherwise the ordering is
43 * unimportant.
44 */
45 static const struct intel_l3_config ivb_l3_configs[] = {
46 /* SLM URB ALL DC RO IS C T */
47 {{ 0, 32, 0, 0, 32, 0, 0, 0 }},
48 {{ 0, 32, 0, 16, 16, 0, 0, 0 }},
49 {{ 0, 32, 0, 4, 0, 8, 4, 16 }},
50 {{ 0, 28, 0, 8, 0, 8, 4, 16 }},
51 {{ 0, 28, 0, 16, 0, 8, 4, 8 }},
52 {{ 0, 28, 0, 8, 0, 16, 4, 8 }},
53 {{ 0, 28, 0, 0, 0, 16, 4, 16 }},
54 {{ 0, 32, 0, 0, 0, 16, 0, 16 }},
55 {{ 0, 28, 0, 4, 32, 0, 0, 0 }},
56 {{ 16, 16, 0, 16, 16, 0, 0, 0 }},
57 {{ 16, 16, 0, 8, 0, 8, 8, 8 }},
58 {{ 16, 16, 0, 4, 0, 8, 4, 16 }},
59 {{ 16, 16, 0, 4, 0, 16, 4, 8 }},
60 {{ 16, 16, 0, 0, 32, 0, 0, 0 }},
61 };
62 DECLARE_L3_LIST(ivb);
63
64 /**
65 * VLV validated L3 configurations. \sa ivb_l3_configs.
66 */
67 static const struct intel_l3_config vlv_l3_configs[] = {
68 /* SLM URB ALL DC RO IS C T */
69 {{ 0, 64, 0, 0, 32, 0, 0, 0 }},
70 {{ 0, 80, 0, 0, 16, 0, 0, 0 }},
71 {{ 0, 80, 0, 8, 8, 0, 0, 0 }},
72 {{ 0, 64, 0, 16, 16, 0, 0, 0 }},
73 {{ 0, 60, 0, 4, 32, 0, 0, 0 }},
74 {{ 32, 32, 0, 16, 16, 0, 0, 0 }},
75 {{ 32, 40, 0, 8, 16, 0, 0, 0 }},
76 {{ 32, 40, 0, 16, 8, 0, 0, 0 }},
77 };
78 DECLARE_L3_LIST(vlv);
79
80 /**
81 * BDW validated L3 configurations. \sa ivb_l3_configs.
82 */
83 static const struct intel_l3_config bdw_l3_configs[] = {
84 /* SLM URB ALL DC RO IS C T */
85 {{ 0, 48, 48, 0, 0, 0, 0, 0 }},
86 {{ 0, 48, 0, 16, 32, 0, 0, 0 }},
87 {{ 0, 32, 0, 16, 48, 0, 0, 0 }},
88 {{ 0, 32, 0, 0, 64, 0, 0, 0 }},
89 {{ 0, 32, 64, 0, 0, 0, 0, 0 }},
90 {{ 24, 16, 48, 0, 0, 0, 0, 0 }},
91 {{ 24, 16, 0, 16, 32, 0, 0, 0 }},
92 {{ 24, 16, 0, 32, 16, 0, 0, 0 }},
93 };
94 DECLARE_L3_LIST(bdw);
95
96 /**
97 * CHV/SKL validated L3 configurations. \sa ivb_l3_configs.
98 */
99 static const struct intel_l3_config chv_l3_configs[] = {
100 /* SLM URB ALL DC RO IS C T */
101 {{ 0, 48, 48, 0, 0, 0, 0, 0 }},
102 {{ 0, 48, 0, 16, 32, 0, 0, 0 }},
103 {{ 0, 32, 0, 16, 48, 0, 0, 0 }},
104 {{ 0, 32, 0, 0, 64, 0, 0, 0 }},
105 {{ 0, 32, 64, 0, 0, 0, 0, 0 }},
106 {{ 32, 16, 48, 0, 0, 0, 0, 0 }},
107 {{ 32, 16, 0, 16, 32, 0, 0, 0 }},
108 {{ 32, 16, 0, 32, 16, 0, 0, 0 }},
109 };
110 DECLARE_L3_LIST(chv);
111
112 /**
113 * BXT 2x6 validated L3 configurations. \sa ivb_l3_configs.
114 */
115 static const struct intel_l3_config bxt_2x6_l3_configs[] = {
116 /* SLM URB ALL DC RO IS C T */
117 {{ 0, 32, 48, 0, 0, 0, 0, 0 }},
118 {{ 0, 32, 0, 8, 40, 0, 0, 0 }},
119 {{ 0, 32, 0, 32, 16, 0, 0, 0 }},
120 {{ 16, 16, 48, 0, 0, 0, 0, 0 }},
121 {{ 16, 16, 0, 40, 8, 0, 0, 0 }},
122 {{ 16, 16, 0, 16, 32, 0, 0, 0 }},
123 };
124 DECLARE_L3_LIST(bxt_2x6);
125
126 /**
127 * ICL validated L3 configurations. \sa icl_l3_configs.
128 * Zeroth entry in below table has been commented out intentionally
129 * due to known issues with this configuration. Many other entries
130 * suggested by h/w specification aren't added here because they
131 * do under allocation of L3 cache with below partitioning.
132 */
133 static const struct intel_l3_config icl_l3_configs[] = {
134 /* SLM URB ALL DC RO IS C T */
135 /*{{ 0, 16, 80, 0, 0, 0, 0, 0 }},*/
136 {{ 0, 32, 64, 0, 0, 0, 0, 0 }},
137 };
138 DECLARE_L3_LIST(icl);
139
140 /**
141 * TGL validated L3 configurations. \sa tgl_l3_configs.
142 */
143 static const struct intel_l3_config tgl_l3_configs[] = {
144 /* SLM URB ALL DC RO IS C T */
145 {{ 0, 32, 88, 0, 0, 0, 0, 0 }},
146 {{ 0, 16, 104, 0, 0, 0, 0, 0 }},
147 };
148 DECLARE_L3_LIST(tgl);
149
150 /**
151 * Empty L3 configurations. \sa empty_l3_configs.
152 */
153 static const struct intel_l3_config empty_l3_configs[] = {
154 /* No configurations. L3FullWayAllocationEnable is always set. */
155 };
156 DECLARE_L3_LIST(empty);
157
158 /**
159 * Return a zero-terminated array of validated L3 configurations for the
160 * specified device.
161 */
162 static const struct intel_l3_list *
get_l3_list(const struct intel_device_info * devinfo)163 get_l3_list(const struct intel_device_info *devinfo)
164 {
165 switch (devinfo->ver) {
166 case 7:
167 return (devinfo->platform == INTEL_PLATFORM_BYT ? &vlv_l3_list : &ivb_l3_list);
168
169 case 8:
170 return (devinfo->platform == INTEL_PLATFORM_CHV ? &chv_l3_list : &bdw_l3_list);
171
172 case 9:
173 if (devinfo->l3_banks == 1)
174 return &bxt_2x6_l3_list;
175 return &chv_l3_list;
176
177 case 11:
178 return &icl_l3_list;
179
180 case 12:
181 if (devinfo->platform == INTEL_PLATFORM_DG1 ||
182 intel_device_info_is_dg2(devinfo))
183 return &empty_l3_list;
184 else
185 return &tgl_l3_list;
186
187 default:
188 unreachable("Not implemented");
189 }
190 }
191
192 /**
193 * L1-normalize a vector of L3 partition weights.
194 */
195 static struct intel_l3_weights
norm_l3_weights(struct intel_l3_weights w)196 norm_l3_weights(struct intel_l3_weights w)
197 {
198 float sz = 0;
199
200 for (unsigned i = 0; i < INTEL_NUM_L3P; i++)
201 sz += w.w[i];
202
203 for (unsigned i = 0; i < INTEL_NUM_L3P; i++)
204 w.w[i] /= sz;
205
206 return w;
207 }
208
209 /**
210 * Get the relative partition weights of the specified L3 configuration.
211 */
212 struct intel_l3_weights
intel_get_l3_config_weights(const struct intel_l3_config * cfg)213 intel_get_l3_config_weights(const struct intel_l3_config *cfg)
214 {
215 if (cfg) {
216 struct intel_l3_weights w;
217
218 for (unsigned i = 0; i < INTEL_NUM_L3P; i++)
219 w.w[i] = cfg->n[i];
220
221 return norm_l3_weights(w);
222 } else {
223 const struct intel_l3_weights w = { { 0 } };
224 return w;
225 }
226 }
227
228 /**
229 * Distance between two L3 configurations represented as vectors of weights.
230 * Usually just the L1 metric except when the two configurations are
231 * considered incompatible in which case the distance will be infinite. Note
232 * that the compatibility condition is asymmetric -- They will be considered
233 * incompatible whenever the reference configuration \p w0 requires SLM, DC,
234 * or URB but \p w1 doesn't provide it.
235 */
236 float
intel_diff_l3_weights(struct intel_l3_weights w0,struct intel_l3_weights w1)237 intel_diff_l3_weights(struct intel_l3_weights w0, struct intel_l3_weights w1)
238 {
239 if ((w0.w[INTEL_L3P_SLM] && !w1.w[INTEL_L3P_SLM]) ||
240 (w0.w[INTEL_L3P_DC] && !w1.w[INTEL_L3P_DC] && !w1.w[INTEL_L3P_ALL]) ||
241 (w0.w[INTEL_L3P_URB] && !w1.w[INTEL_L3P_URB])) {
242 return HUGE_VALF;
243
244 } else {
245 float dw = 0;
246
247 for (unsigned i = 0; i < INTEL_NUM_L3P; i++)
248 dw += fabsf(w0.w[i] - w1.w[i]);
249
250 return dw;
251 }
252 }
253
254 /**
255 * Return a reasonable default L3 configuration for the specified device based
256 * on whether SLM and DC are required. In the non-SLM non-DC case the result
257 * is intended to approximately resemble the hardware defaults.
258 */
259 struct intel_l3_weights
intel_get_default_l3_weights(const struct intel_device_info * devinfo,bool needs_dc,bool needs_slm)260 intel_get_default_l3_weights(const struct intel_device_info *devinfo,
261 bool needs_dc, bool needs_slm)
262 {
263 struct intel_l3_weights w = {{ 0 }};
264
265 w.w[INTEL_L3P_SLM] = devinfo->ver < 11 && needs_slm;
266 w.w[INTEL_L3P_URB] = 1.0;
267
268 if (devinfo->ver >= 8) {
269 w.w[INTEL_L3P_ALL] = 1.0;
270 } else {
271 w.w[INTEL_L3P_DC] = needs_dc ? 0.1 : 0;
272 w.w[INTEL_L3P_RO] = devinfo->platform == INTEL_PLATFORM_BYT ? 0.5 : 1.0;
273 }
274
275 return norm_l3_weights(w);
276 }
277
278 /**
279 * Get the default L3 configuration
280 */
281 const struct intel_l3_config *
intel_get_default_l3_config(const struct intel_device_info * devinfo)282 intel_get_default_l3_config(const struct intel_device_info *devinfo)
283 {
284 /* For efficiency assume that the first entry of the array matches the
285 * default configuration.
286 */
287 const struct intel_l3_list *const list = get_l3_list(devinfo);
288 assert(list->length > 0 || devinfo->ver >= 12);
289 if (list->length > 0) {
290 const struct intel_l3_config *const cfg = &list->configs[0];
291 assert(cfg == intel_get_l3_config(devinfo,
292 intel_get_default_l3_weights(devinfo, false, false)));
293 return cfg;
294 } else {
295 return NULL;
296 }
297 }
298
299 /**
300 * Return the closest validated L3 configuration for the specified device and
301 * weight vector.
302 */
303 const struct intel_l3_config *
intel_get_l3_config(const struct intel_device_info * devinfo,struct intel_l3_weights w0)304 intel_get_l3_config(const struct intel_device_info *devinfo,
305 struct intel_l3_weights w0)
306 {
307 const struct intel_l3_list *const list = get_l3_list(devinfo);
308 const struct intel_l3_config *const cfgs = list->configs;
309 const struct intel_l3_config *cfg_best = NULL;
310 float dw_best = HUGE_VALF;
311
312 for (int i = 0; i < list->length; i++) {
313 const struct intel_l3_config *cfg = &cfgs[i];
314 const float dw = intel_diff_l3_weights(w0, intel_get_l3_config_weights(cfg));
315
316 if (dw < dw_best) {
317 cfg_best = cfg;
318 dw_best = dw;
319 }
320 }
321
322 assert(cfg_best || devinfo->ver >= 12);
323 return cfg_best;
324 }
325
326 /**
327 * Return the size of an L3 way in KB.
328 */
329 static unsigned
get_l3_way_size(const struct intel_device_info * devinfo)330 get_l3_way_size(const struct intel_device_info *devinfo)
331 {
332 const unsigned way_size_per_bank =
333 (devinfo->ver >= 9 && devinfo->l3_banks == 1) || devinfo->ver >= 11 ?
334 4 : 2;
335
336 assert(devinfo->l3_banks);
337 return way_size_per_bank * devinfo->l3_banks;
338 }
339
340 /**
341 * Return the unit brw_context::urb::size is expressed in, in KB. \sa
342 * intel_device_info::urb::size.
343 */
344 static unsigned
get_urb_size_scale(const struct intel_device_info * devinfo)345 get_urb_size_scale(const struct intel_device_info *devinfo)
346 {
347 return (devinfo->ver >= 8 ? devinfo->num_slices : 1);
348 }
349
350 unsigned
intel_get_l3_config_urb_size(const struct intel_device_info * devinfo,const struct intel_l3_config * cfg)351 intel_get_l3_config_urb_size(const struct intel_device_info *devinfo,
352 const struct intel_l3_config *cfg)
353 {
354 /* We don't have to program the URB size for some platforms. It's a fixed
355 * value.
356 */
357 if (cfg == NULL) {
358 ASSERTED const struct intel_l3_list *const list = get_l3_list(devinfo);
359 assert(list->length == 0);
360 return devinfo->urb.size;
361 }
362
363 /* From the SKL "L3 Allocation and Programming" documentation:
364 *
365 * "URB is limited to 1008KB due to programming restrictions. This is not
366 * a restriction of the L3 implementation, but of the FF and other clients.
367 * Therefore, in a GT4 implementation it is possible for the programmed
368 * allocation of the L3 data array to provide 3*384KB=1152KB for URB, but
369 * only 1008KB of this will be used."
370 */
371 const unsigned max = (devinfo->ver == 9 ? 1008 : ~0);
372 return MIN2(max, cfg->n[INTEL_L3P_URB] * get_l3_way_size(devinfo)) /
373 get_urb_size_scale(devinfo);
374 }
375
376 /**
377 * Print out the specified L3 configuration.
378 */
379 void
intel_dump_l3_config(const struct intel_l3_config * cfg,FILE * fp)380 intel_dump_l3_config(const struct intel_l3_config *cfg, FILE *fp)
381 {
382 fprintf(stderr, "SLM=%d URB=%d ALL=%d DC=%d RO=%d IS=%d C=%d T=%d\n",
383 cfg->n[INTEL_L3P_SLM], cfg->n[INTEL_L3P_URB], cfg->n[INTEL_L3P_ALL],
384 cfg->n[INTEL_L3P_DC], cfg->n[INTEL_L3P_RO],
385 cfg->n[INTEL_L3P_IS], cfg->n[INTEL_L3P_C], cfg->n[INTEL_L3P_T]);
386 }
387