diff options
author | Terje Bergstrom <tbergstrom@nvidia.com> | 2014-10-23 08:35:27 -0400 |
---|---|---|
committer | Dan Willemsen <dwillemsen@nvidia.com> | 2015-03-18 15:11:51 -0400 |
commit | 13ca1676ef9c43b137b0ca55e8143da9fa4b5032 (patch) | |
tree | be9a03e4f4fbcd9d072777b30d99655919734b11 | |
parent | 83f3c09510a70cc046b548eaff006403c986e9f7 (diff) |
gpu: nvgpu: Split L2 size calculation per chip
gk20a and gm20b calculate L2 size with different parameters. Split
the function for calculating size so that it does not query GPU id.
Bug 1567274
Change-Id: I09510c1bf0286c9df125d74e51df322c32bde646
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
-rw-r--r-- | drivers/gpu/nvgpu/gk20a/ltc_common.c | 59 | ||||
-rw-r--r-- | drivers/gpu/nvgpu/gk20a/ltc_gk20a.c | 41 | ||||
-rw-r--r-- | drivers/gpu/nvgpu/gm20b/ltc_gm20b.c | 43 |
3 files changed, 83 insertions, 60 deletions
diff --git a/drivers/gpu/nvgpu/gk20a/ltc_common.c b/drivers/gpu/nvgpu/gk20a/ltc_common.c index 9d534b7b..e0ab3f9b 100644 --- a/drivers/gpu/nvgpu/gk20a/ltc_common.c +++ b/drivers/gpu/nvgpu/gk20a/ltc_common.c | |||
@@ -24,65 +24,6 @@ | |||
24 | #include "gk20a.h" | 24 | #include "gk20a.h" |
25 | #include "gr_gk20a.h" | 25 | #include "gr_gk20a.h" |
26 | 26 | ||
27 | static int gk20a_determine_L2_size_bytes(struct gk20a *g) | ||
28 | { | ||
29 | const u32 gpuid = GK20A_GPUID(g->gpu_characteristics.arch, | ||
30 | g->gpu_characteristics.impl); | ||
31 | u32 lts_per_ltc; | ||
32 | u32 ways; | ||
33 | u32 sets; | ||
34 | u32 bytes_per_line; | ||
35 | u32 active_ltcs; | ||
36 | u32 cache_size; | ||
37 | |||
38 | u32 tmp; | ||
39 | u32 active_sets_value; | ||
40 | |||
41 | tmp = gk20a_readl(g, ltc_ltc0_lts0_tstg_cfg1_r()); | ||
42 | ways = hweight32(ltc_ltc0_lts0_tstg_cfg1_active_ways_v(tmp)); | ||
43 | |||
44 | active_sets_value = ltc_ltc0_lts0_tstg_cfg1_active_sets_v(tmp); | ||
45 | if (active_sets_value == ltc_ltc0_lts0_tstg_cfg1_active_sets_all_v()) { | ||
46 | sets = 64; | ||
47 | } else if (active_sets_value == | ||
48 | ltc_ltc0_lts0_tstg_cfg1_active_sets_half_v()) { | ||
49 | sets = 32; | ||
50 | } else if (active_sets_value == | ||
51 | ltc_ltc0_lts0_tstg_cfg1_active_sets_quarter_v()) { | ||
52 | sets = 16; | ||
53 | } else { | ||
54 | dev_err(dev_from_gk20a(g), | ||
55 | "Unknown constant %u for active sets", | ||
56 | (unsigned)active_sets_value); | ||
57 | sets = 0; | ||
58 | } | ||
59 | |||
60 | active_ltcs = g->gr.num_fbps; | ||
61 | |||
62 | /* chip-specific values */ | ||
63 | switch (gpuid) { | ||
64 | case GK20A_GPUID_GK20A: | ||
65 | lts_per_ltc = 1; | ||
66 | bytes_per_line = 128; | ||
67 | break; | ||
68 | |||
69 | case GK20A_GPUID_GM20B: | ||
70 | lts_per_ltc = 2; | ||
71 | bytes_per_line = 128; | ||
72 | break; | ||
73 | |||
74 | default: | ||
75 | dev_err(dev_from_gk20a(g), "Unknown GPU id 0x%02x\n", | ||
76 | (unsigned)gpuid); | ||
77 | lts_per_ltc = 0; | ||
78 | bytes_per_line = 0; | ||
79 | } | ||
80 | |||
81 | cache_size = active_ltcs * lts_per_ltc * ways * sets * bytes_per_line; | ||
82 | |||
83 | return cache_size; | ||
84 | } | ||
85 | |||
86 | /* | 27 | /* |
87 | * Set the maximum number of ways that can have the "EVIST_LAST" class. | 28 | * Set the maximum number of ways that can have the "EVIST_LAST" class. |
88 | */ | 29 | */ |
diff --git a/drivers/gpu/nvgpu/gk20a/ltc_gk20a.c b/drivers/gpu/nvgpu/gk20a/ltc_gk20a.c index 2794b3db..aa094dc7 100644 --- a/drivers/gpu/nvgpu/gk20a/ltc_gk20a.c +++ b/drivers/gpu/nvgpu/gk20a/ltc_gk20a.c | |||
@@ -224,6 +224,47 @@ static void gk20a_mm_g_elpg_flush_locked(struct gk20a *g) | |||
224 | 224 | ||
225 | } | 225 | } |
226 | 226 | ||
227 | static int gk20a_determine_L2_size_bytes(struct gk20a *g) | ||
228 | { | ||
229 | u32 lts_per_ltc; | ||
230 | u32 ways; | ||
231 | u32 sets; | ||
232 | u32 bytes_per_line; | ||
233 | u32 active_ltcs; | ||
234 | u32 cache_size; | ||
235 | |||
236 | u32 tmp; | ||
237 | u32 active_sets_value; | ||
238 | |||
239 | tmp = gk20a_readl(g, ltc_ltc0_lts0_tstg_cfg1_r()); | ||
240 | ways = hweight32(ltc_ltc0_lts0_tstg_cfg1_active_ways_v(tmp)); | ||
241 | |||
242 | active_sets_value = ltc_ltc0_lts0_tstg_cfg1_active_sets_v(tmp); | ||
243 | if (active_sets_value == ltc_ltc0_lts0_tstg_cfg1_active_sets_all_v()) { | ||
244 | sets = 64; | ||
245 | } else if (active_sets_value == | ||
246 | ltc_ltc0_lts0_tstg_cfg1_active_sets_half_v()) { | ||
247 | sets = 32; | ||
248 | } else if (active_sets_value == | ||
249 | ltc_ltc0_lts0_tstg_cfg1_active_sets_quarter_v()) { | ||
250 | sets = 16; | ||
251 | } else { | ||
252 | dev_err(dev_from_gk20a(g), | ||
253 | "Unknown constant %u for active sets", | ||
254 | (unsigned)active_sets_value); | ||
255 | sets = 0; | ||
256 | } | ||
257 | |||
258 | active_ltcs = g->gr.num_fbps; | ||
259 | |||
260 | /* chip-specific values */ | ||
261 | lts_per_ltc = 1; | ||
262 | bytes_per_line = 128; | ||
263 | cache_size = active_ltcs * lts_per_ltc * ways * sets * bytes_per_line; | ||
264 | |||
265 | return cache_size; | ||
266 | } | ||
267 | |||
227 | void gk20a_init_ltc(struct gpu_ops *gops) | 268 | void gk20a_init_ltc(struct gpu_ops *gops) |
228 | { | 269 | { |
229 | gops->ltc.determine_L2_size_bytes = gk20a_determine_L2_size_bytes; | 270 | gops->ltc.determine_L2_size_bytes = gk20a_determine_L2_size_bytes; |
diff --git a/drivers/gpu/nvgpu/gm20b/ltc_gm20b.c b/drivers/gpu/nvgpu/gm20b/ltc_gm20b.c index 96e5dbde..a089b59c 100644 --- a/drivers/gpu/nvgpu/gm20b/ltc_gm20b.c +++ b/drivers/gpu/nvgpu/gm20b/ltc_gm20b.c | |||
@@ -347,10 +347,51 @@ void gm20b_flush_ltc(struct gk20a *g) | |||
347 | ltc_ltc1_ltss_tstg_cmgmt0_invalidate_pending_f()); | 347 | ltc_ltc1_ltss_tstg_cmgmt0_invalidate_pending_f()); |
348 | } | 348 | } |
349 | 349 | ||
350 | static int gm20b_determine_L2_size_bytes(struct gk20a *g) | ||
351 | { | ||
352 | u32 lts_per_ltc; | ||
353 | u32 ways; | ||
354 | u32 sets; | ||
355 | u32 bytes_per_line; | ||
356 | u32 active_ltcs; | ||
357 | u32 cache_size; | ||
358 | |||
359 | u32 tmp; | ||
360 | u32 active_sets_value; | ||
361 | |||
362 | tmp = gk20a_readl(g, ltc_ltc0_lts0_tstg_cfg1_r()); | ||
363 | ways = hweight32(ltc_ltc0_lts0_tstg_cfg1_active_ways_v(tmp)); | ||
364 | |||
365 | active_sets_value = ltc_ltc0_lts0_tstg_cfg1_active_sets_v(tmp); | ||
366 | if (active_sets_value == ltc_ltc0_lts0_tstg_cfg1_active_sets_all_v()) { | ||
367 | sets = 64; | ||
368 | } else if (active_sets_value == | ||
369 | ltc_ltc0_lts0_tstg_cfg1_active_sets_half_v()) { | ||
370 | sets = 32; | ||
371 | } else if (active_sets_value == | ||
372 | ltc_ltc0_lts0_tstg_cfg1_active_sets_quarter_v()) { | ||
373 | sets = 16; | ||
374 | } else { | ||
375 | dev_err(dev_from_gk20a(g), | ||
376 | "Unknown constant %u for active sets", | ||
377 | (unsigned)active_sets_value); | ||
378 | sets = 0; | ||
379 | } | ||
380 | |||
381 | active_ltcs = g->gr.num_fbps; | ||
382 | |||
383 | /* chip-specific values */ | ||
384 | lts_per_ltc = 2; | ||
385 | bytes_per_line = 128; | ||
386 | cache_size = active_ltcs * lts_per_ltc * ways * sets * bytes_per_line; | ||
387 | |||
388 | return cache_size; | ||
389 | } | ||
390 | |||
350 | void gm20b_init_ltc(struct gpu_ops *gops) | 391 | void gm20b_init_ltc(struct gpu_ops *gops) |
351 | { | 392 | { |
352 | /* Gk20a reused ops. */ | 393 | /* Gk20a reused ops. */ |
353 | gops->ltc.determine_L2_size_bytes = gk20a_determine_L2_size_bytes; | 394 | gops->ltc.determine_L2_size_bytes = gm20b_determine_L2_size_bytes; |
354 | gops->ltc.set_max_ways_evict_last = gk20a_ltc_set_max_ways_evict_last; | 395 | gops->ltc.set_max_ways_evict_last = gk20a_ltc_set_max_ways_evict_last; |
355 | gops->ltc.set_zbc_color_entry = gk20a_ltc_set_zbc_color_entry; | 396 | gops->ltc.set_zbc_color_entry = gk20a_ltc_set_zbc_color_entry; |
356 | gops->ltc.set_zbc_depth_entry = gk20a_ltc_set_zbc_depth_entry; | 397 | gops->ltc.set_zbc_depth_entry = gk20a_ltc_set_zbc_depth_entry; |