2 * Copyright © 2017 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining
5 * a copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
13 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
14 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
15 * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
16 * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
17 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
18 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
26 #include "ac_gpu_info.h"
27 #include "addrlib/src/amdgpu_asic_addr.h"
30 #include "util/macros.h"
31 #include "util/u_math.h"
36 #include <amdgpu_drm.h>
40 #define CIK_TILE_MODE_COLOR_2D 14
42 #define CIK__GB_TILE_MODE__PIPE_CONFIG(x) (((x) >> 6) & 0x1f)
43 #define CIK__PIPE_CONFIG__ADDR_SURF_P2 0
44 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16 4
45 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16 5
46 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32 6
47 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32 7
48 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16 8
49 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16 9
50 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16 10
51 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16 11
52 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16 12
53 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32 13
54 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32 14
55 #define CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16 16
56 #define CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16 17
58 static unsigned cik_get_num_tile_pipes(struct amdgpu_gpu_info
*info
)
60 unsigned mode2d
= info
->gb_tile_mode
[CIK_TILE_MODE_COLOR_2D
];
62 switch (CIK__GB_TILE_MODE__PIPE_CONFIG(mode2d
)) {
63 case CIK__PIPE_CONFIG__ADDR_SURF_P2
:
65 case CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16
:
66 case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16
:
67 case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32
:
68 case CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32
:
70 case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16
:
71 case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16
:
72 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16
:
73 case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16
:
74 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16
:
75 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32
:
76 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32
:
78 case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16
:
79 case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16
:
82 fprintf(stderr
, "Invalid GFX7 pipe configuration, assuming P2\n");
83 assert(!"this should never occur");
88 static bool has_syncobj(int fd
)
91 if (drmGetCap(fd
, DRM_CAP_SYNCOBJ
, &value
))
93 return value
? true : false;
96 bool ac_query_gpu_info(int fd
, void *dev_p
,
97 struct radeon_info
*info
,
98 struct amdgpu_gpu_info
*amdinfo
)
100 struct drm_amdgpu_info_device device_info
= {};
101 struct amdgpu_buffer_size_alignments alignment_info
= {};
102 struct drm_amdgpu_info_hw_ip dma
= {}, compute
= {}, uvd
= {};
103 struct drm_amdgpu_info_hw_ip uvd_enc
= {}, vce
= {}, vcn_dec
= {}, vcn_jpeg
= {};
104 struct drm_amdgpu_info_hw_ip vcn_enc
= {}, gfx
= {};
105 struct amdgpu_gds_resource_info gds
= {};
106 uint32_t vce_version
= 0, vce_feature
= 0, uvd_version
= 0, uvd_feature
= 0;
108 amdgpu_device_handle dev
= dev_p
;
109 drmDevicePtr devinfo
;
112 r
= drmGetDevice2(fd
, 0, &devinfo
);
114 fprintf(stderr
, "amdgpu: drmGetDevice2 failed.\n");
117 info
->pci_domain
= devinfo
->businfo
.pci
->domain
;
118 info
->pci_bus
= devinfo
->businfo
.pci
->bus
;
119 info
->pci_dev
= devinfo
->businfo
.pci
->dev
;
120 info
->pci_func
= devinfo
->businfo
.pci
->func
;
121 drmFreeDevice(&devinfo
);
123 assert(info
->drm_major
== 3);
124 info
->is_amdgpu
= true;
126 /* Query hardware and driver information. */
127 r
= amdgpu_query_gpu_info(dev
, amdinfo
);
129 fprintf(stderr
, "amdgpu: amdgpu_query_gpu_info failed.\n");
133 r
= amdgpu_query_info(dev
, AMDGPU_INFO_DEV_INFO
, sizeof(device_info
),
136 fprintf(stderr
, "amdgpu: amdgpu_query_info(dev_info) failed.\n");
140 r
= amdgpu_query_buffer_size_alignment(dev
, &alignment_info
);
142 fprintf(stderr
, "amdgpu: amdgpu_query_buffer_size_alignment failed.\n");
146 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_DMA
, 0, &dma
);
148 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(dma) failed.\n");
152 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_GFX
, 0, &gfx
);
154 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(gfx) failed.\n");
158 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_COMPUTE
, 0, &compute
);
160 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(compute) failed.\n");
164 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_UVD
, 0, &uvd
);
166 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(uvd) failed.\n");
170 if (info
->drm_minor
>= 17) {
171 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_UVD_ENC
, 0, &uvd_enc
);
173 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(uvd_enc) failed.\n");
178 if (info
->drm_minor
>= 17) {
179 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_VCN_DEC
, 0, &vcn_dec
);
181 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vcn_dec) failed.\n");
186 if (info
->drm_minor
>= 17) {
187 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_VCN_ENC
, 0, &vcn_enc
);
189 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vcn_enc) failed.\n");
194 if (info
->drm_minor
>= 27) {
195 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_VCN_JPEG
, 0, &vcn_jpeg
);
197 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vcn_jpeg) failed.\n");
202 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_GFX_ME
, 0, 0,
203 &info
->me_fw_version
,
204 &info
->me_fw_feature
);
206 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(me) failed.\n");
210 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_GFX_PFP
, 0, 0,
211 &info
->pfp_fw_version
,
212 &info
->pfp_fw_feature
);
214 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(pfp) failed.\n");
218 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_GFX_CE
, 0, 0,
219 &info
->ce_fw_version
,
220 &info
->ce_fw_feature
);
222 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(ce) failed.\n");
226 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_UVD
, 0, 0,
227 &uvd_version
, &uvd_feature
);
229 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(uvd) failed.\n");
233 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_VCE
, 0, &vce
);
235 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vce) failed.\n");
239 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_VCE
, 0, 0,
240 &vce_version
, &vce_feature
);
242 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(vce) failed.\n");
246 r
= amdgpu_query_sw_info(dev
, amdgpu_sw_info_address32_hi
, &info
->address32_hi
);
248 fprintf(stderr
, "amdgpu: amdgpu_query_sw_info(address32_hi) failed.\n");
252 r
= amdgpu_query_gds_info(dev
, &gds
);
254 fprintf(stderr
, "amdgpu: amdgpu_query_gds_info failed.\n");
258 if (info
->drm_minor
>= 9) {
259 struct drm_amdgpu_memory_info meminfo
= {};
261 r
= amdgpu_query_info(dev
, AMDGPU_INFO_MEMORY
, sizeof(meminfo
), &meminfo
);
263 fprintf(stderr
, "amdgpu: amdgpu_query_info(memory) failed.\n");
267 /* Note: usable_heap_size values can be random and can't be relied on. */
268 info
->gart_size
= meminfo
.gtt
.total_heap_size
;
269 info
->vram_size
= meminfo
.vram
.total_heap_size
;
270 info
->vram_vis_size
= meminfo
.cpu_accessible_vram
.total_heap_size
;
272 /* This is a deprecated interface, which reports usable sizes
273 * (total minus pinned), but the pinned size computation is
274 * buggy, so the values returned from these functions can be
277 struct amdgpu_heap_info vram
, vram_vis
, gtt
;
279 r
= amdgpu_query_heap_info(dev
, AMDGPU_GEM_DOMAIN_VRAM
, 0, &vram
);
281 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(vram) failed.\n");
285 r
= amdgpu_query_heap_info(dev
, AMDGPU_GEM_DOMAIN_VRAM
,
286 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
,
289 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(vram_vis) failed.\n");
293 r
= amdgpu_query_heap_info(dev
, AMDGPU_GEM_DOMAIN_GTT
, 0, >t
);
295 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(gtt) failed.\n");
299 info
->gart_size
= gtt
.heap_size
;
300 info
->vram_size
= vram
.heap_size
;
301 info
->vram_vis_size
= vram_vis
.heap_size
;
304 /* Set chip identification. */
305 info
->pci_id
= amdinfo
->asic_id
; /* TODO: is this correct? */
306 info
->vce_harvest_config
= amdinfo
->vce_harvest_config
;
308 #define identify_chip2(asic, chipname) \
309 if (ASICREV_IS(amdinfo->chip_external_rev, asic)) { \
310 info->family = CHIP_##chipname; \
311 info->name = #chipname; \
313 #define identify_chip(chipname) identify_chip2(chipname, chipname)
315 switch (amdinfo
->family_id
) {
316 case AMDGPU_FAMILY_SI
:
317 identify_chip(TAHITI
);
318 identify_chip(PITCAIRN
);
319 identify_chip2(CAPEVERDE
, VERDE
);
320 identify_chip(OLAND
);
321 identify_chip(HAINAN
);
323 case AMDGPU_FAMILY_CI
:
324 identify_chip(BONAIRE
);
325 identify_chip(HAWAII
);
327 case AMDGPU_FAMILY_KV
:
328 identify_chip2(SPECTRE
, KAVERI
);
329 identify_chip2(SPOOKY
, KAVERI
);
330 identify_chip2(KALINDI
, KABINI
);
331 identify_chip2(GODAVARI
, KABINI
);
333 case AMDGPU_FAMILY_VI
:
334 identify_chip(ICELAND
);
335 identify_chip(TONGA
);
337 identify_chip(POLARIS10
);
338 identify_chip(POLARIS11
);
339 identify_chip(POLARIS12
);
340 identify_chip(VEGAM
);
342 case AMDGPU_FAMILY_CZ
:
343 identify_chip(CARRIZO
);
344 identify_chip(STONEY
);
346 case AMDGPU_FAMILY_AI
:
347 identify_chip(VEGA10
);
348 identify_chip(VEGA12
);
349 identify_chip(VEGA20
);
350 identify_chip(ARCTURUS
);
352 case AMDGPU_FAMILY_RV
:
353 identify_chip(RAVEN
);
354 identify_chip(RAVEN2
);
355 identify_chip(RENOIR
);
357 case AMDGPU_FAMILY_NV
:
358 identify_chip(NAVI10
);
359 identify_chip(NAVI12
);
360 identify_chip(NAVI14
);
365 fprintf(stderr
, "amdgpu: unknown (family_id, chip_external_rev): (%u, %u)\n",
366 amdinfo
->family_id
, amdinfo
->chip_external_rev
);
370 if (info
->family
>= CHIP_NAVI10
)
371 info
->chip_class
= GFX10
;
372 else if (info
->family
>= CHIP_VEGA10
)
373 info
->chip_class
= GFX9
;
374 else if (info
->family
>= CHIP_TONGA
)
375 info
->chip_class
= GFX8
;
376 else if (info
->family
>= CHIP_BONAIRE
)
377 info
->chip_class
= GFX7
;
378 else if (info
->family
>= CHIP_TAHITI
)
379 info
->chip_class
= GFX6
;
381 fprintf(stderr
, "amdgpu: Unknown family.\n");
385 info
->family_id
= amdinfo
->family_id
;
386 info
->chip_external_rev
= amdinfo
->chip_external_rev
;
387 info
->marketing_name
= amdgpu_get_marketing_name(dev
);
388 info
->is_pro_graphics
= info
->marketing_name
&&
389 (!strcmp(info
->marketing_name
, "Pro") ||
390 !strcmp(info
->marketing_name
, "PRO") ||
391 !strcmp(info
->marketing_name
, "Frontier"));
393 /* Set which chips have dedicated VRAM. */
394 info
->has_dedicated_vram
=
395 !(amdinfo
->ids_flags
& AMDGPU_IDS_FLAGS_FUSION
);
397 /* The kernel can split large buffers in VRAM but not in GTT, so large
398 * allocations can fail or cause buffer movement failures in the kernel.
400 if (info
->has_dedicated_vram
)
401 info
->max_alloc_size
= info
->vram_size
* 0.8;
403 info
->max_alloc_size
= info
->gart_size
* 0.7;
405 /* Set hardware information. */
406 info
->gds_size
= gds
.gds_total_size
;
407 info
->gds_gfx_partition_size
= gds
.gds_gfx_partition_size
;
408 /* convert the shader clock from KHz to MHz */
409 info
->max_shader_clock
= amdinfo
->max_engine_clk
/ 1000;
410 info
->num_tcc_blocks
= device_info
.num_tcc_blocks
;
411 info
->max_se
= amdinfo
->num_shader_engines
;
412 info
->max_sh_per_se
= amdinfo
->num_shader_arrays_per_engine
;
413 info
->has_hw_decode
=
414 (uvd
.available_rings
!= 0) || (vcn_dec
.available_rings
!= 0) ||
415 (vcn_jpeg
.available_rings
!= 0);
416 info
->uvd_fw_version
=
417 uvd
.available_rings
? uvd_version
: 0;
418 info
->vce_fw_version
=
419 vce
.available_rings
? vce_version
: 0;
420 info
->uvd_enc_supported
=
421 uvd_enc
.available_rings
? true : false;
422 info
->has_userptr
= true;
423 info
->has_syncobj
= has_syncobj(fd
);
424 info
->has_syncobj_wait_for_submit
= info
->has_syncobj
&& info
->drm_minor
>= 20;
425 info
->has_fence_to_handle
= info
->has_syncobj
&& info
->drm_minor
>= 21;
426 info
->has_ctx_priority
= info
->drm_minor
>= 22;
427 info
->has_local_buffers
= info
->drm_minor
>= 20;
428 info
->kernel_flushes_hdp_before_ib
= true;
429 info
->htile_cmask_support_1d_tiling
= true;
430 info
->si_TA_CS_BC_BASE_ADDR_allowed
= true;
431 info
->has_bo_metadata
= true;
432 info
->has_gpu_reset_status_query
= true;
433 info
->has_eqaa_surface_allocator
= true;
434 info
->has_format_bc1_through_bc7
= true;
435 /* DRM 3.1.0 doesn't flush TC for GFX8 correctly. */
436 info
->kernel_flushes_tc_l2_after_ib
= info
->chip_class
!= GFX8
||
437 info
->drm_minor
>= 2;
438 info
->has_indirect_compute_dispatch
= true;
439 /* GFX6 doesn't support unaligned loads. */
440 info
->has_unaligned_shader_loads
= info
->chip_class
!= GFX6
;
441 /* Disable sparse mappings on GFX6 due to VM faults in CP DMA. Enable them once
442 * these faults are mitigated in software.
443 * Disable sparse mappings on GFX9 due to hangs.
445 info
->has_sparse_vm_mappings
=
446 info
->chip_class
>= GFX7
&& info
->chip_class
<= GFX8
&&
447 info
->drm_minor
>= 13;
448 info
->has_2d_tiling
= true;
449 info
->has_read_registers_query
= true;
450 info
->has_scheduled_fence_dependency
= info
->drm_minor
>= 28;
452 info
->pa_sc_tile_steering_override
= device_info
.pa_sc_tile_steering_override
;
453 info
->num_render_backends
= amdinfo
->rb_pipes
;
454 /* The value returned by the kernel driver was wrong. */
455 if (info
->family
== CHIP_KAVERI
)
456 info
->num_render_backends
= 2;
458 info
->clock_crystal_freq
= amdinfo
->gpu_counter_freq
;
459 if (!info
->clock_crystal_freq
) {
460 fprintf(stderr
, "amdgpu: clock crystal frequency is 0, timestamps will be wrong\n");
461 info
->clock_crystal_freq
= 1;
463 if (info
->chip_class
>= GFX10
) {
464 info
->tcc_cache_line_size
= 128;
466 info
->tcc_cache_line_size
= 64;
468 info
->gb_addr_config
= amdinfo
->gb_addr_cfg
;
469 if (info
->chip_class
== GFX9
) {
470 info
->num_tile_pipes
= 1 << G_0098F8_NUM_PIPES(amdinfo
->gb_addr_cfg
);
471 info
->pipe_interleave_bytes
=
472 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(amdinfo
->gb_addr_cfg
);
474 info
->num_tile_pipes
= cik_get_num_tile_pipes(amdinfo
);
475 info
->pipe_interleave_bytes
=
476 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX6(amdinfo
->gb_addr_cfg
);
478 info
->r600_has_virtual_memory
= true;
480 assert(util_is_power_of_two_or_zero(dma
.available_rings
+ 1));
481 assert(util_is_power_of_two_or_zero(compute
.available_rings
+ 1));
483 info
->has_graphics
= gfx
.available_rings
> 0;
484 info
->num_sdma_rings
= util_bitcount(dma
.available_rings
);
485 info
->num_compute_rings
= util_bitcount(compute
.available_rings
);
487 /* The mere presence of CLEAR_STATE in the IB causes random GPU hangs
488 * on GFX6. Some CLEAR_STATE cause asic hang on radeon kernel, etc.
489 * SPI_VS_OUT_CONFIG. So only enable GFX7 CLEAR_STATE on amdgpu kernel.
491 info
->has_clear_state
= info
->chip_class
>= GFX7
;
493 info
->has_distributed_tess
= info
->chip_class
>= GFX8
&&
496 info
->has_dcc_constant_encode
= info
->family
== CHIP_RAVEN2
||
497 info
->family
== CHIP_RENOIR
||
498 info
->chip_class
>= GFX10
;
500 info
->has_rbplus
= info
->family
== CHIP_STONEY
||
501 info
->chip_class
>= GFX9
;
503 /* Some chips have RB+ registers, but don't support RB+. Those must
506 info
->rbplus_allowed
= info
->has_rbplus
&&
507 (info
->family
== CHIP_STONEY
||
508 info
->family
== CHIP_VEGA12
||
509 info
->family
== CHIP_RAVEN
||
510 info
->family
== CHIP_RAVEN2
||
511 info
->family
== CHIP_RENOIR
);
513 info
->has_out_of_order_rast
= info
->chip_class
>= GFX8
&&
516 /* TODO: Figure out how to use LOAD_CONTEXT_REG on GFX6-GFX7. */
517 info
->has_load_ctx_reg_pkt
= info
->chip_class
>= GFX9
||
518 (info
->chip_class
>= GFX8
&&
519 info
->me_fw_feature
>= 41);
521 info
->cpdma_prefetch_writes_memory
= info
->chip_class
<= GFX8
;
523 info
->has_gfx9_scissor_bug
= info
->family
== CHIP_VEGA10
||
524 info
->family
== CHIP_RAVEN
;
526 info
->has_tc_compat_zrange_bug
= info
->chip_class
>= GFX8
&&
527 info
->chip_class
<= GFX9
;
529 info
->has_msaa_sample_loc_bug
= (info
->family
>= CHIP_POLARIS10
&&
530 info
->family
<= CHIP_POLARIS12
) ||
531 info
->family
== CHIP_VEGA10
||
532 info
->family
== CHIP_RAVEN
;
534 info
->has_ls_vgpr_init_bug
= info
->family
== CHIP_VEGA10
||
535 info
->family
== CHIP_RAVEN
;
537 /* Get the number of good compute units. */
538 info
->num_good_compute_units
= 0;
539 for (i
= 0; i
< info
->max_se
; i
++)
540 for (j
= 0; j
< info
->max_sh_per_se
; j
++)
541 info
->num_good_compute_units
+=
542 util_bitcount(amdinfo
->cu_bitmap
[i
][j
]);
543 info
->num_good_cu_per_sh
= info
->num_good_compute_units
/
544 (info
->max_se
* info
->max_sh_per_se
);
546 memcpy(info
->si_tile_mode_array
, amdinfo
->gb_tile_mode
,
547 sizeof(amdinfo
->gb_tile_mode
));
548 info
->enabled_rb_mask
= amdinfo
->enabled_rb_pipes_mask
;
550 memcpy(info
->cik_macrotile_mode_array
, amdinfo
->gb_macro_tile_mode
,
551 sizeof(amdinfo
->gb_macro_tile_mode
));
553 info
->pte_fragment_size
= alignment_info
.size_local
;
554 info
->gart_page_size
= alignment_info
.size_remote
;
556 if (info
->chip_class
== GFX6
)
557 info
->gfx_ib_pad_with_type2
= true;
559 unsigned ib_align
= 0;
560 ib_align
= MAX2(ib_align
, gfx
.ib_start_alignment
);
561 ib_align
= MAX2(ib_align
, compute
.ib_start_alignment
);
562 ib_align
= MAX2(ib_align
, dma
.ib_start_alignment
);
563 ib_align
= MAX2(ib_align
, uvd
.ib_start_alignment
);
564 ib_align
= MAX2(ib_align
, uvd_enc
.ib_start_alignment
);
565 ib_align
= MAX2(ib_align
, vce
.ib_start_alignment
);
566 ib_align
= MAX2(ib_align
, vcn_dec
.ib_start_alignment
);
567 ib_align
= MAX2(ib_align
, vcn_enc
.ib_start_alignment
);
568 ib_align
= MAX2(ib_align
, vcn_jpeg
.ib_start_alignment
);
570 info
->ib_start_alignment
= ib_align
;
572 if (info
->drm_minor
>= 31 &&
573 (info
->family
== CHIP_RAVEN
||
574 info
->family
== CHIP_RAVEN2
||
575 info
->family
== CHIP_RENOIR
)) {
576 if (info
->num_render_backends
== 1)
577 info
->use_display_dcc_unaligned
= true;
579 info
->use_display_dcc_with_retile_blit
= true;
582 info
->has_gds_ordered_append
= info
->chip_class
>= GFX7
&&
583 info
->drm_minor
>= 29;
585 if (info
->chip_class
>= GFX9
) {
586 unsigned pc_lines
= 0;
588 switch (info
->family
) {
608 if (info
->chip_class
>= GFX10
) {
609 info
->pbb_max_alloc_count
= pc_lines
/ 3;
611 info
->pbb_max_alloc_count
=
612 MIN2(128, pc_lines
/ (4 * info
->max_se
));
616 if (info
->chip_class
>= GFX10
) {
617 switch (info
->family
) {
620 info
->num_sdp_interfaces
= 16;
623 info
->num_sdp_interfaces
= 8;
630 info
->max_wave64_per_simd
= info
->family
>= CHIP_POLARIS10
&&
631 info
->family
<= CHIP_VEGAM
? 8 : 10;
633 /* The number is per SIMD. There is enough SGPRs for the maximum number
634 * of Wave32, which is double the number for Wave64.
636 if (info
->chip_class
>= GFX10
)
637 info
->num_physical_sgprs_per_simd
= 128 * info
->max_wave64_per_simd
* 2;
638 else if (info
->chip_class
>= GFX8
)
639 info
->num_physical_sgprs_per_simd
= 800;
641 info
->num_physical_sgprs_per_simd
= 512;
643 info
->num_physical_wave64_vgprs_per_simd
= info
->chip_class
>= GFX10
? 512 : 256;
647 void ac_compute_driver_uuid(char *uuid
, size_t size
)
649 char amd_uuid
[] = "AMD-MESA-DRV";
651 assert(size
>= sizeof(amd_uuid
));
653 memset(uuid
, 0, size
);
654 strncpy(uuid
, amd_uuid
, size
);
657 void ac_compute_device_uuid(struct radeon_info
*info
, char *uuid
, size_t size
)
659 uint32_t *uint_uuid
= (uint32_t*)uuid
;
661 assert(size
>= sizeof(uint32_t)*4);
664 * Use the device info directly instead of using a sha1. GL/VK UUIDs
665 * are 16 byte vs 20 byte for sha1, and the truncation that would be
666 * required would get rid of part of the little entropy we have.
668 memset(uuid
, 0, size
);
669 uint_uuid
[0] = info
->pci_domain
;
670 uint_uuid
[1] = info
->pci_bus
;
671 uint_uuid
[2] = info
->pci_dev
;
672 uint_uuid
[3] = info
->pci_func
;
675 void ac_print_gpu_info(struct radeon_info
*info
)
677 printf("Device info:\n");
678 printf(" pci (domain:bus:dev.func): %04x:%02x:%02x.%x\n",
679 info
->pci_domain
, info
->pci_bus
,
680 info
->pci_dev
, info
->pci_func
);
681 printf(" pci_id = 0x%x\n", info
->pci_id
);
682 printf(" family = %i\n", info
->family
);
683 printf(" chip_class = %i\n", info
->chip_class
);
684 printf(" chip_external_rev = %i\n", info
->chip_external_rev
);
685 printf(" num_compute_rings = %u\n", info
->num_compute_rings
);
686 printf(" num_sdma_rings = %i\n", info
->num_sdma_rings
);
687 printf(" clock_crystal_freq = %i\n", info
->clock_crystal_freq
);
688 printf(" tcc_cache_line_size = %u\n", info
->tcc_cache_line_size
);
690 printf(" use_display_dcc_unaligned = %u\n", info
->use_display_dcc_unaligned
);
691 printf(" use_display_dcc_with_retile_blit = %u\n", info
->use_display_dcc_with_retile_blit
);
693 printf("Memory info:\n");
694 printf(" pte_fragment_size = %u\n", info
->pte_fragment_size
);
695 printf(" gart_page_size = %u\n", info
->gart_page_size
);
696 printf(" gart_size = %i MB\n", (int)DIV_ROUND_UP(info
->gart_size
, 1024*1024));
697 printf(" vram_size = %i MB\n", (int)DIV_ROUND_UP(info
->vram_size
, 1024*1024));
698 printf(" vram_vis_size = %i MB\n", (int)DIV_ROUND_UP(info
->vram_vis_size
, 1024*1024));
699 printf(" gds_size = %u kB\n", info
->gds_size
/ 1024);
700 printf(" gds_gfx_partition_size = %u kB\n", info
->gds_gfx_partition_size
/ 1024);
701 printf(" max_alloc_size = %i MB\n",
702 (int)DIV_ROUND_UP(info
->max_alloc_size
, 1024*1024));
703 printf(" min_alloc_size = %u\n", info
->min_alloc_size
);
704 printf(" address32_hi = %u\n", info
->address32_hi
);
705 printf(" has_dedicated_vram = %u\n", info
->has_dedicated_vram
);
707 printf("CP info:\n");
708 printf(" gfx_ib_pad_with_type2 = %i\n", info
->gfx_ib_pad_with_type2
);
709 printf(" ib_start_alignment = %u\n", info
->ib_start_alignment
);
710 printf(" me_fw_version = %i\n", info
->me_fw_version
);
711 printf(" me_fw_feature = %i\n", info
->me_fw_feature
);
712 printf(" pfp_fw_version = %i\n", info
->pfp_fw_version
);
713 printf(" pfp_fw_feature = %i\n", info
->pfp_fw_feature
);
714 printf(" ce_fw_version = %i\n", info
->ce_fw_version
);
715 printf(" ce_fw_feature = %i\n", info
->ce_fw_feature
);
717 printf("Multimedia info:\n");
718 printf(" has_hw_decode = %u\n", info
->has_hw_decode
);
719 printf(" uvd_enc_supported = %u\n", info
->uvd_enc_supported
);
720 printf(" uvd_fw_version = %u\n", info
->uvd_fw_version
);
721 printf(" vce_fw_version = %u\n", info
->vce_fw_version
);
722 printf(" vce_harvest_config = %i\n", info
->vce_harvest_config
);
724 printf("Kernel & winsys capabilities:\n");
725 printf(" drm = %i.%i.%i\n", info
->drm_major
,
726 info
->drm_minor
, info
->drm_patchlevel
);
727 printf(" has_userptr = %i\n", info
->has_userptr
);
728 printf(" has_syncobj = %u\n", info
->has_syncobj
);
729 printf(" has_syncobj_wait_for_submit = %u\n", info
->has_syncobj_wait_for_submit
);
730 printf(" has_fence_to_handle = %u\n", info
->has_fence_to_handle
);
731 printf(" has_ctx_priority = %u\n", info
->has_ctx_priority
);
732 printf(" has_local_buffers = %u\n", info
->has_local_buffers
);
733 printf(" kernel_flushes_hdp_before_ib = %u\n", info
->kernel_flushes_hdp_before_ib
);
734 printf(" htile_cmask_support_1d_tiling = %u\n", info
->htile_cmask_support_1d_tiling
);
735 printf(" si_TA_CS_BC_BASE_ADDR_allowed = %u\n", info
->si_TA_CS_BC_BASE_ADDR_allowed
);
736 printf(" has_bo_metadata = %u\n", info
->has_bo_metadata
);
737 printf(" has_gpu_reset_status_query = %u\n", info
->has_gpu_reset_status_query
);
738 printf(" has_eqaa_surface_allocator = %u\n", info
->has_eqaa_surface_allocator
);
739 printf(" has_format_bc1_through_bc7 = %u\n", info
->has_format_bc1_through_bc7
);
740 printf(" kernel_flushes_tc_l2_after_ib = %u\n", info
->kernel_flushes_tc_l2_after_ib
);
741 printf(" has_indirect_compute_dispatch = %u\n", info
->has_indirect_compute_dispatch
);
742 printf(" has_unaligned_shader_loads = %u\n", info
->has_unaligned_shader_loads
);
743 printf(" has_sparse_vm_mappings = %u\n", info
->has_sparse_vm_mappings
);
744 printf(" has_2d_tiling = %u\n", info
->has_2d_tiling
);
745 printf(" has_read_registers_query = %u\n", info
->has_read_registers_query
);
746 printf(" has_gds_ordered_append = %u\n", info
->has_gds_ordered_append
);
747 printf(" has_scheduled_fence_dependency = %u\n", info
->has_scheduled_fence_dependency
);
749 printf("Shader core info:\n");
750 printf(" max_shader_clock = %i\n", info
->max_shader_clock
);
751 printf(" num_good_compute_units = %i\n", info
->num_good_compute_units
);
752 printf(" num_good_cu_per_sh = %i\n", info
->num_good_cu_per_sh
);
753 printf(" num_tcc_blocks = %i\n", info
->num_tcc_blocks
);
754 printf(" max_se = %i\n", info
->max_se
);
755 printf(" max_sh_per_se = %i\n", info
->max_sh_per_se
);
757 printf("Render backend info:\n");
758 printf(" pa_sc_tile_steering_override = 0x%x\n", info
->pa_sc_tile_steering_override
);
759 printf(" num_render_backends = %i\n", info
->num_render_backends
);
760 printf(" num_tile_pipes = %i\n", info
->num_tile_pipes
);
761 printf(" pipe_interleave_bytes = %i\n", info
->pipe_interleave_bytes
);
762 printf(" enabled_rb_mask = 0x%x\n", info
->enabled_rb_mask
);
763 printf(" max_alignment = %u\n", (unsigned)info
->max_alignment
);
765 printf("GB_ADDR_CONFIG:\n");
766 if (info
->chip_class
>= GFX9
) {
767 printf(" num_pipes = %u\n",
768 1 << G_0098F8_NUM_PIPES(info
->gb_addr_config
));
769 printf(" pipe_interleave_size = %u\n",
770 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(info
->gb_addr_config
));
771 printf(" max_compressed_frags = %u\n",
772 1 << G_0098F8_MAX_COMPRESSED_FRAGS(info
->gb_addr_config
));
773 printf(" bank_interleave_size = %u\n",
774 1 << G_0098F8_BANK_INTERLEAVE_SIZE(info
->gb_addr_config
));
775 printf(" num_banks = %u\n",
776 1 << G_0098F8_NUM_BANKS(info
->gb_addr_config
));
777 printf(" shader_engine_tile_size = %u\n",
778 16 << G_0098F8_SHADER_ENGINE_TILE_SIZE(info
->gb_addr_config
));
779 printf(" num_shader_engines = %u\n",
780 1 << G_0098F8_NUM_SHADER_ENGINES_GFX9(info
->gb_addr_config
));
781 printf(" num_gpus = %u (raw)\n",
782 G_0098F8_NUM_GPUS_GFX9(info
->gb_addr_config
));
783 printf(" multi_gpu_tile_size = %u (raw)\n",
784 G_0098F8_MULTI_GPU_TILE_SIZE(info
->gb_addr_config
));
785 printf(" num_rb_per_se = %u\n",
786 1 << G_0098F8_NUM_RB_PER_SE(info
->gb_addr_config
));
787 printf(" row_size = %u\n",
788 1024 << G_0098F8_ROW_SIZE(info
->gb_addr_config
));
789 printf(" num_lower_pipes = %u (raw)\n",
790 G_0098F8_NUM_LOWER_PIPES(info
->gb_addr_config
));
791 printf(" se_enable = %u (raw)\n",
792 G_0098F8_SE_ENABLE(info
->gb_addr_config
));
794 printf(" num_pipes = %u\n",
795 1 << G_0098F8_NUM_PIPES(info
->gb_addr_config
));
796 printf(" pipe_interleave_size = %u\n",
797 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX6(info
->gb_addr_config
));
798 printf(" bank_interleave_size = %u\n",
799 1 << G_0098F8_BANK_INTERLEAVE_SIZE(info
->gb_addr_config
));
800 printf(" num_shader_engines = %u\n",
801 1 << G_0098F8_NUM_SHADER_ENGINES_GFX6(info
->gb_addr_config
));
802 printf(" shader_engine_tile_size = %u\n",
803 16 << G_0098F8_SHADER_ENGINE_TILE_SIZE(info
->gb_addr_config
));
804 printf(" num_gpus = %u (raw)\n",
805 G_0098F8_NUM_GPUS_GFX6(info
->gb_addr_config
));
806 printf(" multi_gpu_tile_size = %u (raw)\n",
807 G_0098F8_MULTI_GPU_TILE_SIZE(info
->gb_addr_config
));
808 printf(" row_size = %u\n",
809 1024 << G_0098F8_ROW_SIZE(info
->gb_addr_config
));
810 printf(" num_lower_pipes = %u (raw)\n",
811 G_0098F8_NUM_LOWER_PIPES(info
->gb_addr_config
));
816 ac_get_gs_table_depth(enum chip_class chip_class
, enum radeon_family family
)
818 if (chip_class
>= GFX9
)
843 unreachable("Unknown GPU");
848 ac_get_raster_config(struct radeon_info
*info
,
849 uint32_t *raster_config_p
,
850 uint32_t *raster_config_1_p
,
851 uint32_t *se_tile_repeat_p
)
853 unsigned raster_config
, raster_config_1
, se_tile_repeat
;
855 switch (info
->family
) {
860 raster_config
= 0x00000000;
861 raster_config_1
= 0x00000000;
865 raster_config
= 0x0000124a;
866 raster_config_1
= 0x00000000;
868 /* 1 SE / 2 RBs (Oland is special) */
870 raster_config
= 0x00000082;
871 raster_config_1
= 0x00000000;
877 raster_config
= 0x00000002;
878 raster_config_1
= 0x00000000;
884 raster_config
= 0x16000012;
885 raster_config_1
= 0x00000000;
890 raster_config
= 0x2a00126a;
891 raster_config_1
= 0x00000000;
896 raster_config
= 0x16000012;
897 raster_config_1
= 0x0000002a;
903 raster_config
= 0x3a00161a;
904 raster_config_1
= 0x0000002e;
908 "ac: Unknown GPU, using 0 for raster_config\n");
909 raster_config
= 0x00000000;
910 raster_config_1
= 0x00000000;
914 /* drm/radeon on Kaveri is buggy, so disable 1 RB to work around it.
915 * This decreases performance by up to 50% when the RB is the bottleneck.
917 if (info
->family
== CHIP_KAVERI
&& !info
->is_amdgpu
)
918 raster_config
= 0x00000000;
920 /* Fiji: Old kernels have incorrect tiling config. This decreases
921 * RB performance by 25%. (it disables 1 RB in the second packer)
923 if (info
->family
== CHIP_FIJI
&&
924 info
->cik_macrotile_mode_array
[0] == 0x000000e8) {
925 raster_config
= 0x16000012;
926 raster_config_1
= 0x0000002a;
929 unsigned se_width
= 8 << G_028350_SE_XSEL_GFX6(raster_config
);
930 unsigned se_height
= 8 << G_028350_SE_YSEL_GFX6(raster_config
);
932 /* I don't know how to calculate this, though this is probably a good guess. */
933 se_tile_repeat
= MAX2(se_width
, se_height
) * info
->max_se
;
935 *raster_config_p
= raster_config
;
936 *raster_config_1_p
= raster_config_1
;
937 if (se_tile_repeat_p
)
938 *se_tile_repeat_p
= se_tile_repeat
;
942 ac_get_harvested_configs(struct radeon_info
*info
,
943 unsigned raster_config
,
944 unsigned *cik_raster_config_1_p
,
945 unsigned *raster_config_se
)
947 unsigned sh_per_se
= MAX2(info
->max_sh_per_se
, 1);
948 unsigned num_se
= MAX2(info
->max_se
, 1);
949 unsigned rb_mask
= info
->enabled_rb_mask
;
950 unsigned num_rb
= MIN2(info
->num_render_backends
, 16);
951 unsigned rb_per_pkr
= MIN2(num_rb
/ num_se
/ sh_per_se
, 2);
952 unsigned rb_per_se
= num_rb
/ num_se
;
956 se_mask
[0] = ((1 << rb_per_se
) - 1) & rb_mask
;
957 se_mask
[1] = (se_mask
[0] << rb_per_se
) & rb_mask
;
958 se_mask
[2] = (se_mask
[1] << rb_per_se
) & rb_mask
;
959 se_mask
[3] = (se_mask
[2] << rb_per_se
) & rb_mask
;
961 assert(num_se
== 1 || num_se
== 2 || num_se
== 4);
962 assert(sh_per_se
== 1 || sh_per_se
== 2);
963 assert(rb_per_pkr
== 1 || rb_per_pkr
== 2);
966 if (info
->chip_class
>= GFX7
) {
967 unsigned raster_config_1
= *cik_raster_config_1_p
;
968 if ((num_se
> 2) && ((!se_mask
[0] && !se_mask
[1]) ||
969 (!se_mask
[2] && !se_mask
[3]))) {
970 raster_config_1
&= C_028354_SE_PAIR_MAP
;
972 if (!se_mask
[0] && !se_mask
[1]) {
974 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_3
);
977 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_0
);
979 *cik_raster_config_1_p
= raster_config_1
;
983 for (se
= 0; se
< num_se
; se
++) {
984 unsigned pkr0_mask
= ((1 << rb_per_pkr
) - 1) << (se
* rb_per_se
);
985 unsigned pkr1_mask
= pkr0_mask
<< rb_per_pkr
;
986 int idx
= (se
/ 2) * 2;
988 raster_config_se
[se
] = raster_config
;
989 if ((num_se
> 1) && (!se_mask
[idx
] || !se_mask
[idx
+ 1])) {
990 raster_config_se
[se
] &= C_028350_SE_MAP
;
993 raster_config_se
[se
] |=
994 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_3
);
996 raster_config_se
[se
] |=
997 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_0
);
1001 pkr0_mask
&= rb_mask
;
1002 pkr1_mask
&= rb_mask
;
1003 if (rb_per_se
> 2 && (!pkr0_mask
|| !pkr1_mask
)) {
1004 raster_config_se
[se
] &= C_028350_PKR_MAP
;
1007 raster_config_se
[se
] |=
1008 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_3
);
1010 raster_config_se
[se
] |=
1011 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_0
);
1015 if (rb_per_se
>= 2) {
1016 unsigned rb0_mask
= 1 << (se
* rb_per_se
);
1017 unsigned rb1_mask
= rb0_mask
<< 1;
1019 rb0_mask
&= rb_mask
;
1020 rb1_mask
&= rb_mask
;
1021 if (!rb0_mask
|| !rb1_mask
) {
1022 raster_config_se
[se
] &= C_028350_RB_MAP_PKR0
;
1025 raster_config_se
[se
] |=
1026 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_3
);
1028 raster_config_se
[se
] |=
1029 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_0
);
1033 if (rb_per_se
> 2) {
1034 rb0_mask
= 1 << (se
* rb_per_se
+ rb_per_pkr
);
1035 rb1_mask
= rb0_mask
<< 1;
1036 rb0_mask
&= rb_mask
;
1037 rb1_mask
&= rb_mask
;
1038 if (!rb0_mask
|| !rb1_mask
) {
1039 raster_config_se
[se
] &= C_028350_RB_MAP_PKR1
;
1042 raster_config_se
[se
] |=
1043 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_3
);
1045 raster_config_se
[se
] |=
1046 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_0
);
1054 unsigned ac_get_compute_resource_limits(struct radeon_info
*info
,
1055 unsigned waves_per_threadgroup
,
1056 unsigned max_waves_per_sh
,
1057 unsigned threadgroups_per_cu
)
1059 unsigned compute_resource_limits
=
1060 S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup
% 4 == 0);
1062 if (info
->chip_class
>= GFX7
) {
1063 unsigned num_cu_per_se
= info
->num_good_compute_units
/
1066 /* Force even distribution on all SIMDs in CU if the workgroup
1067 * size is 64. This has shown some good improvements if # of CUs
1068 * per SE is not a multiple of 4.
1070 if (num_cu_per_se
% 4 && waves_per_threadgroup
== 1)
1071 compute_resource_limits
|= S_00B854_FORCE_SIMD_DIST(1);
1073 assert(threadgroups_per_cu
>= 1 && threadgroups_per_cu
<= 8);
1074 compute_resource_limits
|= S_00B854_WAVES_PER_SH(max_waves_per_sh
) |
1075 S_00B854_CU_GROUP_COUNT(threadgroups_per_cu
- 1);
1078 if (max_waves_per_sh
) {
1079 unsigned limit_div16
= DIV_ROUND_UP(max_waves_per_sh
, 16);
1080 compute_resource_limits
|= S_00B854_WAVES_PER_SH_SI(limit_div16
);
1083 return compute_resource_limits
;