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"
29 #include "util/u_math.h"
34 #include <amdgpu_drm.h>
38 #define CIK_TILE_MODE_COLOR_2D 14
40 #define CIK__GB_TILE_MODE__PIPE_CONFIG(x) (((x) >> 6) & 0x1f)
41 #define CIK__PIPE_CONFIG__ADDR_SURF_P2 0
42 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16 4
43 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16 5
44 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32 6
45 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32 7
46 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16 8
47 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16 9
48 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16 10
49 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16 11
50 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16 12
51 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32 13
52 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32 14
53 #define CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16 16
54 #define CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16 17
56 static unsigned cik_get_num_tile_pipes(struct amdgpu_gpu_info
*info
)
58 unsigned mode2d
= info
->gb_tile_mode
[CIK_TILE_MODE_COLOR_2D
];
60 switch (CIK__GB_TILE_MODE__PIPE_CONFIG(mode2d
)) {
61 case CIK__PIPE_CONFIG__ADDR_SURF_P2
:
63 case CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16
:
64 case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16
:
65 case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32
:
66 case CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32
:
68 case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16
:
69 case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16
:
70 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16
:
71 case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16
:
72 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16
:
73 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32
:
74 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32
:
76 case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16
:
77 case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16
:
80 fprintf(stderr
, "Invalid GFX7 pipe configuration, assuming P2\n");
81 assert(!"this should never occur");
86 static bool has_syncobj(int fd
)
89 if (drmGetCap(fd
, DRM_CAP_SYNCOBJ
, &value
))
91 return value
? true : false;
94 bool ac_query_gpu_info(int fd
, void *dev_p
,
95 struct radeon_info
*info
,
96 struct amdgpu_gpu_info
*amdinfo
)
98 struct drm_amdgpu_info_device device_info
= {};
99 struct amdgpu_buffer_size_alignments alignment_info
= {};
100 struct drm_amdgpu_info_hw_ip dma
= {}, compute
= {}, uvd
= {};
101 struct drm_amdgpu_info_hw_ip uvd_enc
= {}, vce
= {}, vcn_dec
= {}, vcn_jpeg
= {};
102 struct drm_amdgpu_info_hw_ip vcn_enc
= {}, gfx
= {};
103 struct amdgpu_gds_resource_info gds
= {};
104 uint32_t vce_version
= 0, vce_feature
= 0, uvd_version
= 0, uvd_feature
= 0;
106 amdgpu_device_handle dev
= dev_p
;
107 drmDevicePtr devinfo
;
110 r
= drmGetDevice2(fd
, 0, &devinfo
);
112 fprintf(stderr
, "amdgpu: drmGetDevice2 failed.\n");
115 info
->pci_domain
= devinfo
->businfo
.pci
->domain
;
116 info
->pci_bus
= devinfo
->businfo
.pci
->bus
;
117 info
->pci_dev
= devinfo
->businfo
.pci
->dev
;
118 info
->pci_func
= devinfo
->businfo
.pci
->func
;
119 drmFreeDevice(&devinfo
);
121 assert(info
->drm_major
== 3);
122 info
->is_amdgpu
= true;
124 /* Query hardware and driver information. */
125 r
= amdgpu_query_gpu_info(dev
, amdinfo
);
127 fprintf(stderr
, "amdgpu: amdgpu_query_gpu_info failed.\n");
131 r
= amdgpu_query_info(dev
, AMDGPU_INFO_DEV_INFO
, sizeof(device_info
),
134 fprintf(stderr
, "amdgpu: amdgpu_query_info(dev_info) failed.\n");
138 r
= amdgpu_query_buffer_size_alignment(dev
, &alignment_info
);
140 fprintf(stderr
, "amdgpu: amdgpu_query_buffer_size_alignment failed.\n");
144 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_DMA
, 0, &dma
);
146 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(dma) failed.\n");
150 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_GFX
, 0, &gfx
);
152 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(gfx) failed.\n");
156 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_COMPUTE
, 0, &compute
);
158 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(compute) failed.\n");
162 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_UVD
, 0, &uvd
);
164 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(uvd) failed.\n");
168 if (info
->drm_minor
>= 17) {
169 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_UVD_ENC
, 0, &uvd_enc
);
171 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(uvd_enc) failed.\n");
176 if (info
->drm_minor
>= 17) {
177 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_VCN_DEC
, 0, &vcn_dec
);
179 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vcn_dec) failed.\n");
184 if (info
->drm_minor
>= 17) {
185 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_VCN_ENC
, 0, &vcn_enc
);
187 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vcn_enc) failed.\n");
192 if (info
->drm_minor
>= 27) {
193 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_VCN_JPEG
, 0, &vcn_jpeg
);
195 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vcn_jpeg) failed.\n");
200 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_GFX_ME
, 0, 0,
201 &info
->me_fw_version
,
202 &info
->me_fw_feature
);
204 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(me) failed.\n");
208 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_GFX_PFP
, 0, 0,
209 &info
->pfp_fw_version
,
210 &info
->pfp_fw_feature
);
212 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(pfp) failed.\n");
216 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_GFX_CE
, 0, 0,
217 &info
->ce_fw_version
,
218 &info
->ce_fw_feature
);
220 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(ce) failed.\n");
224 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_UVD
, 0, 0,
225 &uvd_version
, &uvd_feature
);
227 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(uvd) failed.\n");
231 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_VCE
, 0, &vce
);
233 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vce) failed.\n");
237 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_VCE
, 0, 0,
238 &vce_version
, &vce_feature
);
240 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(vce) failed.\n");
244 r
= amdgpu_query_sw_info(dev
, amdgpu_sw_info_address32_hi
, &info
->address32_hi
);
246 fprintf(stderr
, "amdgpu: amdgpu_query_sw_info(address32_hi) failed.\n");
250 r
= amdgpu_query_gds_info(dev
, &gds
);
252 fprintf(stderr
, "amdgpu: amdgpu_query_gds_info failed.\n");
256 if (info
->drm_minor
>= 9) {
257 struct drm_amdgpu_memory_info meminfo
= {};
259 r
= amdgpu_query_info(dev
, AMDGPU_INFO_MEMORY
, sizeof(meminfo
), &meminfo
);
261 fprintf(stderr
, "amdgpu: amdgpu_query_info(memory) failed.\n");
265 /* Note: usable_heap_size values can be random and can't be relied on. */
266 info
->gart_size
= meminfo
.gtt
.total_heap_size
;
267 info
->vram_size
= meminfo
.vram
.total_heap_size
;
268 info
->vram_vis_size
= meminfo
.cpu_accessible_vram
.total_heap_size
;
270 /* This is a deprecated interface, which reports usable sizes
271 * (total minus pinned), but the pinned size computation is
272 * buggy, so the values returned from these functions can be
275 struct amdgpu_heap_info vram
, vram_vis
, gtt
;
277 r
= amdgpu_query_heap_info(dev
, AMDGPU_GEM_DOMAIN_VRAM
, 0, &vram
);
279 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(vram) failed.\n");
283 r
= amdgpu_query_heap_info(dev
, AMDGPU_GEM_DOMAIN_VRAM
,
284 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
,
287 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(vram_vis) failed.\n");
291 r
= amdgpu_query_heap_info(dev
, AMDGPU_GEM_DOMAIN_GTT
, 0, >t
);
293 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(gtt) failed.\n");
297 info
->gart_size
= gtt
.heap_size
;
298 info
->vram_size
= vram
.heap_size
;
299 info
->vram_vis_size
= vram_vis
.heap_size
;
302 /* Set chip identification. */
303 info
->pci_id
= amdinfo
->asic_id
; /* TODO: is this correct? */
304 info
->vce_harvest_config
= amdinfo
->vce_harvest_config
;
306 switch (info
->pci_id
) {
307 #define CHIPSET(pci_id, cfamily) \
309 info->family = CHIP_##cfamily; \
310 info->name = #cfamily; \
312 #include "pci_ids/radeonsi_pci_ids.h"
316 fprintf(stderr
, "amdgpu: Invalid PCI ID.\n");
320 /* Raven2 uses the same PCI IDs as Raven1, but different revision IDs. */
321 if (info
->family
== CHIP_RAVEN
&& amdinfo
->chip_rev
>= 0x8) {
322 info
->family
= CHIP_RAVEN2
;
323 info
->name
= "RAVEN2";
326 if (info
->family
>= CHIP_NAVI10
)
327 info
->chip_class
= GFX10
;
328 else if (info
->family
>= CHIP_VEGA10
)
329 info
->chip_class
= GFX9
;
330 else if (info
->family
>= CHIP_TONGA
)
331 info
->chip_class
= GFX8
;
332 else if (info
->family
>= CHIP_BONAIRE
)
333 info
->chip_class
= GFX7
;
334 else if (info
->family
>= CHIP_TAHITI
)
335 info
->chip_class
= GFX6
;
337 fprintf(stderr
, "amdgpu: Unknown family.\n");
341 info
->family_id
= amdinfo
->family_id
;
342 info
->chip_external_rev
= amdinfo
->chip_external_rev
;
343 info
->marketing_name
= amdgpu_get_marketing_name(dev
);
344 info
->is_pro_graphics
= info
->marketing_name
&&
345 (!strcmp(info
->marketing_name
, "Pro") ||
346 !strcmp(info
->marketing_name
, "PRO") ||
347 !strcmp(info
->marketing_name
, "Frontier"));
349 /* Set which chips have dedicated VRAM. */
350 info
->has_dedicated_vram
=
351 !(amdinfo
->ids_flags
& AMDGPU_IDS_FLAGS_FUSION
);
353 /* The kernel can split large buffers in VRAM but not in GTT, so large
354 * allocations can fail or cause buffer movement failures in the kernel.
356 if (info
->has_dedicated_vram
)
357 info
->max_alloc_size
= info
->vram_size
* 0.8;
359 info
->max_alloc_size
= info
->gart_size
* 0.7;
361 /* Set hardware information. */
362 info
->gds_size
= gds
.gds_total_size
;
363 info
->gds_gfx_partition_size
= gds
.gds_gfx_partition_size
;
364 /* convert the shader clock from KHz to MHz */
365 info
->max_shader_clock
= amdinfo
->max_engine_clk
/ 1000;
366 info
->num_tcc_blocks
= device_info
.num_tcc_blocks
;
367 info
->max_se
= amdinfo
->num_shader_engines
;
368 info
->max_sh_per_se
= amdinfo
->num_shader_arrays_per_engine
;
369 info
->has_hw_decode
=
370 (uvd
.available_rings
!= 0) || (vcn_dec
.available_rings
!= 0) ||
371 (vcn_jpeg
.available_rings
!= 0);
372 info
->uvd_fw_version
=
373 uvd
.available_rings
? uvd_version
: 0;
374 info
->vce_fw_version
=
375 vce
.available_rings
? vce_version
: 0;
376 info
->uvd_enc_supported
=
377 uvd_enc
.available_rings
? true : false;
378 info
->has_userptr
= true;
379 info
->has_syncobj
= has_syncobj(fd
);
380 info
->has_syncobj_wait_for_submit
= info
->has_syncobj
&& info
->drm_minor
>= 20;
381 info
->has_fence_to_handle
= info
->has_syncobj
&& info
->drm_minor
>= 21;
382 info
->has_ctx_priority
= info
->drm_minor
>= 22;
383 info
->has_local_buffers
= info
->drm_minor
>= 20;
384 info
->kernel_flushes_hdp_before_ib
= true;
385 info
->htile_cmask_support_1d_tiling
= true;
386 info
->si_TA_CS_BC_BASE_ADDR_allowed
= true;
387 info
->has_bo_metadata
= true;
388 info
->has_gpu_reset_status_query
= true;
389 info
->has_eqaa_surface_allocator
= true;
390 info
->has_format_bc1_through_bc7
= true;
391 /* DRM 3.1.0 doesn't flush TC for GFX8 correctly. */
392 info
->kernel_flushes_tc_l2_after_ib
= info
->chip_class
!= GFX8
||
393 info
->drm_minor
>= 2;
394 info
->has_indirect_compute_dispatch
= true;
395 /* GFX6 doesn't support unaligned loads. */
396 info
->has_unaligned_shader_loads
= info
->chip_class
!= GFX6
;
397 /* Disable sparse mappings on GFX6 due to VM faults in CP DMA. Enable them once
398 * these faults are mitigated in software.
399 * Disable sparse mappings on GFX9 due to hangs.
401 info
->has_sparse_vm_mappings
=
402 info
->chip_class
>= GFX7
&& info
->chip_class
<= GFX8
&&
403 info
->drm_minor
>= 13;
404 info
->has_2d_tiling
= true;
405 info
->has_read_registers_query
= true;
406 info
->has_scheduled_fence_dependency
= info
->drm_minor
>= 28;
408 info
->pa_sc_tile_steering_override
= device_info
.pa_sc_tile_steering_override
;
409 info
->num_render_backends
= amdinfo
->rb_pipes
;
410 /* The value returned by the kernel driver was wrong. */
411 if (info
->family
== CHIP_KAVERI
)
412 info
->num_render_backends
= 2;
414 info
->clock_crystal_freq
= amdinfo
->gpu_counter_freq
;
415 if (!info
->clock_crystal_freq
) {
416 fprintf(stderr
, "amdgpu: clock crystal frequency is 0, timestamps will be wrong\n");
417 info
->clock_crystal_freq
= 1;
419 if (info
->chip_class
>= GFX10
) {
420 info
->tcc_cache_line_size
= 128;
422 info
->tcc_cache_line_size
= 64;
424 info
->gb_addr_config
= amdinfo
->gb_addr_cfg
;
425 if (info
->chip_class
== GFX9
) {
426 info
->num_tile_pipes
= 1 << G_0098F8_NUM_PIPES(amdinfo
->gb_addr_cfg
);
427 info
->pipe_interleave_bytes
=
428 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(amdinfo
->gb_addr_cfg
);
430 info
->num_tile_pipes
= cik_get_num_tile_pipes(amdinfo
);
431 info
->pipe_interleave_bytes
=
432 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX6(amdinfo
->gb_addr_cfg
);
434 info
->r600_has_virtual_memory
= true;
436 assert(util_is_power_of_two_or_zero(dma
.available_rings
+ 1));
437 assert(util_is_power_of_two_or_zero(compute
.available_rings
+ 1));
439 info
->has_graphics
= gfx
.available_rings
> 0;
440 info
->num_sdma_rings
= util_bitcount(dma
.available_rings
);
441 info
->num_compute_rings
= util_bitcount(compute
.available_rings
);
443 /* Get the number of good compute units. */
444 info
->num_good_compute_units
= 0;
445 for (i
= 0; i
< info
->max_se
; i
++)
446 for (j
= 0; j
< info
->max_sh_per_se
; j
++)
447 info
->num_good_compute_units
+=
448 util_bitcount(amdinfo
->cu_bitmap
[i
][j
]);
449 info
->num_good_cu_per_sh
= info
->num_good_compute_units
/
450 (info
->max_se
* info
->max_sh_per_se
);
452 memcpy(info
->si_tile_mode_array
, amdinfo
->gb_tile_mode
,
453 sizeof(amdinfo
->gb_tile_mode
));
454 info
->enabled_rb_mask
= amdinfo
->enabled_rb_pipes_mask
;
456 memcpy(info
->cik_macrotile_mode_array
, amdinfo
->gb_macro_tile_mode
,
457 sizeof(amdinfo
->gb_macro_tile_mode
));
459 info
->pte_fragment_size
= alignment_info
.size_local
;
460 info
->gart_page_size
= alignment_info
.size_remote
;
462 if (info
->chip_class
== GFX6
)
463 info
->gfx_ib_pad_with_type2
= TRUE
;
465 unsigned ib_align
= 0;
466 ib_align
= MAX2(ib_align
, gfx
.ib_start_alignment
);
467 ib_align
= MAX2(ib_align
, compute
.ib_start_alignment
);
468 ib_align
= MAX2(ib_align
, dma
.ib_start_alignment
);
469 ib_align
= MAX2(ib_align
, uvd
.ib_start_alignment
);
470 ib_align
= MAX2(ib_align
, uvd_enc
.ib_start_alignment
);
471 ib_align
= MAX2(ib_align
, vce
.ib_start_alignment
);
472 ib_align
= MAX2(ib_align
, vcn_dec
.ib_start_alignment
);
473 ib_align
= MAX2(ib_align
, vcn_enc
.ib_start_alignment
);
474 ib_align
= MAX2(ib_align
, vcn_jpeg
.ib_start_alignment
);
476 info
->ib_start_alignment
= ib_align
;
478 if (info
->drm_minor
>= 31 &&
479 (info
->family
== CHIP_RAVEN
||
480 info
->family
== CHIP_RAVEN2
)) {
481 if (info
->num_render_backends
== 1)
482 info
->use_display_dcc_unaligned
= true;
484 info
->use_display_dcc_with_retile_blit
= true;
487 info
->has_gds_ordered_append
= info
->chip_class
>= GFX7
&&
488 info
->drm_minor
>= 29 &&
493 void ac_compute_driver_uuid(char *uuid
, size_t size
)
495 char amd_uuid
[] = "AMD-MESA-DRV";
497 assert(size
>= sizeof(amd_uuid
));
499 memset(uuid
, 0, size
);
500 strncpy(uuid
, amd_uuid
, size
);
503 void ac_compute_device_uuid(struct radeon_info
*info
, char *uuid
, size_t size
)
505 uint32_t *uint_uuid
= (uint32_t*)uuid
;
507 assert(size
>= sizeof(uint32_t)*4);
510 * Use the device info directly instead of using a sha1. GL/VK UUIDs
511 * are 16 byte vs 20 byte for sha1, and the truncation that would be
512 * required would get rid of part of the little entropy we have.
514 memset(uuid
, 0, size
);
515 uint_uuid
[0] = info
->pci_domain
;
516 uint_uuid
[1] = info
->pci_bus
;
517 uint_uuid
[2] = info
->pci_dev
;
518 uint_uuid
[3] = info
->pci_func
;
521 void ac_print_gpu_info(struct radeon_info
*info
)
523 printf("Device info:\n");
524 printf(" pci (domain:bus:dev.func): %04x:%02x:%02x.%x\n",
525 info
->pci_domain
, info
->pci_bus
,
526 info
->pci_dev
, info
->pci_func
);
527 printf(" pci_id = 0x%x\n", info
->pci_id
);
528 printf(" family = %i\n", info
->family
);
529 printf(" chip_class = %i\n", info
->chip_class
);
530 printf(" chip_external_rev = %i\n", info
->chip_external_rev
);
531 printf(" num_compute_rings = %u\n", info
->num_compute_rings
);
532 printf(" num_sdma_rings = %i\n", info
->num_sdma_rings
);
533 printf(" clock_crystal_freq = %i\n", info
->clock_crystal_freq
);
534 printf(" tcc_cache_line_size = %u\n", info
->tcc_cache_line_size
);
536 printf(" use_display_dcc_unaligned = %u\n", info
->use_display_dcc_unaligned
);
537 printf(" use_display_dcc_with_retile_blit = %u\n", info
->use_display_dcc_with_retile_blit
);
539 printf("Memory info:\n");
540 printf(" pte_fragment_size = %u\n", info
->pte_fragment_size
);
541 printf(" gart_page_size = %u\n", info
->gart_page_size
);
542 printf(" gart_size = %i MB\n", (int)DIV_ROUND_UP(info
->gart_size
, 1024*1024));
543 printf(" vram_size = %i MB\n", (int)DIV_ROUND_UP(info
->vram_size
, 1024*1024));
544 printf(" vram_vis_size = %i MB\n", (int)DIV_ROUND_UP(info
->vram_vis_size
, 1024*1024));
545 printf(" gds_size = %u kB\n", info
->gds_size
/ 1024);
546 printf(" gds_gfx_partition_size = %u kB\n", info
->gds_gfx_partition_size
/ 1024);
547 printf(" max_alloc_size = %i MB\n",
548 (int)DIV_ROUND_UP(info
->max_alloc_size
, 1024*1024));
549 printf(" min_alloc_size = %u\n", info
->min_alloc_size
);
550 printf(" address32_hi = %u\n", info
->address32_hi
);
551 printf(" has_dedicated_vram = %u\n", info
->has_dedicated_vram
);
553 printf("CP info:\n");
554 printf(" gfx_ib_pad_with_type2 = %i\n", info
->gfx_ib_pad_with_type2
);
555 printf(" ib_start_alignment = %u\n", info
->ib_start_alignment
);
556 printf(" me_fw_version = %i\n", info
->me_fw_version
);
557 printf(" me_fw_feature = %i\n", info
->me_fw_feature
);
558 printf(" pfp_fw_version = %i\n", info
->pfp_fw_version
);
559 printf(" pfp_fw_feature = %i\n", info
->pfp_fw_feature
);
560 printf(" ce_fw_version = %i\n", info
->ce_fw_version
);
561 printf(" ce_fw_feature = %i\n", info
->ce_fw_feature
);
563 printf("Multimedia info:\n");
564 printf(" has_hw_decode = %u\n", info
->has_hw_decode
);
565 printf(" uvd_enc_supported = %u\n", info
->uvd_enc_supported
);
566 printf(" uvd_fw_version = %u\n", info
->uvd_fw_version
);
567 printf(" vce_fw_version = %u\n", info
->vce_fw_version
);
568 printf(" vce_harvest_config = %i\n", info
->vce_harvest_config
);
570 printf("Kernel & winsys capabilities:\n");
571 printf(" drm = %i.%i.%i\n", info
->drm_major
,
572 info
->drm_minor
, info
->drm_patchlevel
);
573 printf(" has_userptr = %i\n", info
->has_userptr
);
574 printf(" has_syncobj = %u\n", info
->has_syncobj
);
575 printf(" has_syncobj_wait_for_submit = %u\n", info
->has_syncobj_wait_for_submit
);
576 printf(" has_fence_to_handle = %u\n", info
->has_fence_to_handle
);
577 printf(" has_ctx_priority = %u\n", info
->has_ctx_priority
);
578 printf(" has_local_buffers = %u\n", info
->has_local_buffers
);
579 printf(" kernel_flushes_hdp_before_ib = %u\n", info
->kernel_flushes_hdp_before_ib
);
580 printf(" htile_cmask_support_1d_tiling = %u\n", info
->htile_cmask_support_1d_tiling
);
581 printf(" si_TA_CS_BC_BASE_ADDR_allowed = %u\n", info
->si_TA_CS_BC_BASE_ADDR_allowed
);
582 printf(" has_bo_metadata = %u\n", info
->has_bo_metadata
);
583 printf(" has_gpu_reset_status_query = %u\n", info
->has_gpu_reset_status_query
);
584 printf(" has_eqaa_surface_allocator = %u\n", info
->has_eqaa_surface_allocator
);
585 printf(" has_format_bc1_through_bc7 = %u\n", info
->has_format_bc1_through_bc7
);
586 printf(" kernel_flushes_tc_l2_after_ib = %u\n", info
->kernel_flushes_tc_l2_after_ib
);
587 printf(" has_indirect_compute_dispatch = %u\n", info
->has_indirect_compute_dispatch
);
588 printf(" has_unaligned_shader_loads = %u\n", info
->has_unaligned_shader_loads
);
589 printf(" has_sparse_vm_mappings = %u\n", info
->has_sparse_vm_mappings
);
590 printf(" has_2d_tiling = %u\n", info
->has_2d_tiling
);
591 printf(" has_read_registers_query = %u\n", info
->has_read_registers_query
);
592 printf(" has_gds_ordered_append = %u\n", info
->has_gds_ordered_append
);
593 printf(" has_scheduled_fence_dependency = %u\n", info
->has_scheduled_fence_dependency
);
595 printf("Shader core info:\n");
596 printf(" max_shader_clock = %i\n", info
->max_shader_clock
);
597 printf(" num_good_compute_units = %i\n", info
->num_good_compute_units
);
598 printf(" num_good_cu_per_sh = %i\n", info
->num_good_cu_per_sh
);
599 printf(" num_tcc_blocks = %i\n", info
->num_tcc_blocks
);
600 printf(" max_se = %i\n", info
->max_se
);
601 printf(" max_sh_per_se = %i\n", info
->max_sh_per_se
);
603 printf("Render backend info:\n");
604 printf(" pa_sc_tile_steering_override = 0x%x\n", info
->pa_sc_tile_steering_override
);
605 printf(" num_render_backends = %i\n", info
->num_render_backends
);
606 printf(" num_tile_pipes = %i\n", info
->num_tile_pipes
);
607 printf(" pipe_interleave_bytes = %i\n", info
->pipe_interleave_bytes
);
608 printf(" enabled_rb_mask = 0x%x\n", info
->enabled_rb_mask
);
609 printf(" max_alignment = %u\n", (unsigned)info
->max_alignment
);
611 printf("GB_ADDR_CONFIG:\n");
612 if (info
->chip_class
>= GFX9
) {
613 printf(" num_pipes = %u\n",
614 1 << G_0098F8_NUM_PIPES(info
->gb_addr_config
));
615 printf(" pipe_interleave_size = %u\n",
616 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(info
->gb_addr_config
));
617 printf(" max_compressed_frags = %u\n",
618 1 << G_0098F8_MAX_COMPRESSED_FRAGS(info
->gb_addr_config
));
619 printf(" bank_interleave_size = %u\n",
620 1 << G_0098F8_BANK_INTERLEAVE_SIZE(info
->gb_addr_config
));
621 printf(" num_banks = %u\n",
622 1 << G_0098F8_NUM_BANKS(info
->gb_addr_config
));
623 printf(" shader_engine_tile_size = %u\n",
624 16 << G_0098F8_SHADER_ENGINE_TILE_SIZE(info
->gb_addr_config
));
625 printf(" num_shader_engines = %u\n",
626 1 << G_0098F8_NUM_SHADER_ENGINES_GFX9(info
->gb_addr_config
));
627 printf(" num_gpus = %u (raw)\n",
628 G_0098F8_NUM_GPUS_GFX9(info
->gb_addr_config
));
629 printf(" multi_gpu_tile_size = %u (raw)\n",
630 G_0098F8_MULTI_GPU_TILE_SIZE(info
->gb_addr_config
));
631 printf(" num_rb_per_se = %u\n",
632 1 << G_0098F8_NUM_RB_PER_SE(info
->gb_addr_config
));
633 printf(" row_size = %u\n",
634 1024 << G_0098F8_ROW_SIZE(info
->gb_addr_config
));
635 printf(" num_lower_pipes = %u (raw)\n",
636 G_0098F8_NUM_LOWER_PIPES(info
->gb_addr_config
));
637 printf(" se_enable = %u (raw)\n",
638 G_0098F8_SE_ENABLE(info
->gb_addr_config
));
640 printf(" num_pipes = %u\n",
641 1 << G_0098F8_NUM_PIPES(info
->gb_addr_config
));
642 printf(" pipe_interleave_size = %u\n",
643 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX6(info
->gb_addr_config
));
644 printf(" bank_interleave_size = %u\n",
645 1 << G_0098F8_BANK_INTERLEAVE_SIZE(info
->gb_addr_config
));
646 printf(" num_shader_engines = %u\n",
647 1 << G_0098F8_NUM_SHADER_ENGINES_GFX6(info
->gb_addr_config
));
648 printf(" shader_engine_tile_size = %u\n",
649 16 << G_0098F8_SHADER_ENGINE_TILE_SIZE(info
->gb_addr_config
));
650 printf(" num_gpus = %u (raw)\n",
651 G_0098F8_NUM_GPUS_GFX6(info
->gb_addr_config
));
652 printf(" multi_gpu_tile_size = %u (raw)\n",
653 G_0098F8_MULTI_GPU_TILE_SIZE(info
->gb_addr_config
));
654 printf(" row_size = %u\n",
655 1024 << G_0098F8_ROW_SIZE(info
->gb_addr_config
));
656 printf(" num_lower_pipes = %u (raw)\n",
657 G_0098F8_NUM_LOWER_PIPES(info
->gb_addr_config
));
662 ac_get_gs_table_depth(enum chip_class chip_class
, enum radeon_family family
)
664 if (chip_class
>= GFX9
)
689 unreachable("Unknown GPU");
694 ac_get_raster_config(struct radeon_info
*info
,
695 uint32_t *raster_config_p
,
696 uint32_t *raster_config_1_p
,
697 uint32_t *se_tile_repeat_p
)
699 unsigned raster_config
, raster_config_1
, se_tile_repeat
;
701 switch (info
->family
) {
706 raster_config
= 0x00000000;
707 raster_config_1
= 0x00000000;
711 raster_config
= 0x0000124a;
712 raster_config_1
= 0x00000000;
714 /* 1 SE / 2 RBs (Oland is special) */
716 raster_config
= 0x00000082;
717 raster_config_1
= 0x00000000;
723 raster_config
= 0x00000002;
724 raster_config_1
= 0x00000000;
730 raster_config
= 0x16000012;
731 raster_config_1
= 0x00000000;
736 raster_config
= 0x2a00126a;
737 raster_config_1
= 0x00000000;
742 raster_config
= 0x16000012;
743 raster_config_1
= 0x0000002a;
749 raster_config
= 0x3a00161a;
750 raster_config_1
= 0x0000002e;
754 "ac: Unknown GPU, using 0 for raster_config\n");
755 raster_config
= 0x00000000;
756 raster_config_1
= 0x00000000;
760 /* drm/radeon on Kaveri is buggy, so disable 1 RB to work around it.
761 * This decreases performance by up to 50% when the RB is the bottleneck.
763 if (info
->family
== CHIP_KAVERI
&& !info
->is_amdgpu
)
764 raster_config
= 0x00000000;
766 /* Fiji: Old kernels have incorrect tiling config. This decreases
767 * RB performance by 25%. (it disables 1 RB in the second packer)
769 if (info
->family
== CHIP_FIJI
&&
770 info
->cik_macrotile_mode_array
[0] == 0x000000e8) {
771 raster_config
= 0x16000012;
772 raster_config_1
= 0x0000002a;
775 unsigned se_width
= 8 << G_028350_SE_XSEL_GFX6(raster_config
);
776 unsigned se_height
= 8 << G_028350_SE_YSEL_GFX6(raster_config
);
778 /* I don't know how to calculate this, though this is probably a good guess. */
779 se_tile_repeat
= MAX2(se_width
, se_height
) * info
->max_se
;
781 *raster_config_p
= raster_config
;
782 *raster_config_1_p
= raster_config_1
;
783 if (se_tile_repeat_p
)
784 *se_tile_repeat_p
= se_tile_repeat
;
788 ac_get_harvested_configs(struct radeon_info
*info
,
789 unsigned raster_config
,
790 unsigned *cik_raster_config_1_p
,
791 unsigned *raster_config_se
)
793 unsigned sh_per_se
= MAX2(info
->max_sh_per_se
, 1);
794 unsigned num_se
= MAX2(info
->max_se
, 1);
795 unsigned rb_mask
= info
->enabled_rb_mask
;
796 unsigned num_rb
= MIN2(info
->num_render_backends
, 16);
797 unsigned rb_per_pkr
= MIN2(num_rb
/ num_se
/ sh_per_se
, 2);
798 unsigned rb_per_se
= num_rb
/ num_se
;
802 se_mask
[0] = ((1 << rb_per_se
) - 1) & rb_mask
;
803 se_mask
[1] = (se_mask
[0] << rb_per_se
) & rb_mask
;
804 se_mask
[2] = (se_mask
[1] << rb_per_se
) & rb_mask
;
805 se_mask
[3] = (se_mask
[2] << rb_per_se
) & rb_mask
;
807 assert(num_se
== 1 || num_se
== 2 || num_se
== 4);
808 assert(sh_per_se
== 1 || sh_per_se
== 2);
809 assert(rb_per_pkr
== 1 || rb_per_pkr
== 2);
812 if (info
->chip_class
>= GFX7
) {
813 unsigned raster_config_1
= *cik_raster_config_1_p
;
814 if ((num_se
> 2) && ((!se_mask
[0] && !se_mask
[1]) ||
815 (!se_mask
[2] && !se_mask
[3]))) {
816 raster_config_1
&= C_028354_SE_PAIR_MAP
;
818 if (!se_mask
[0] && !se_mask
[1]) {
820 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_3
);
823 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_0
);
825 *cik_raster_config_1_p
= raster_config_1
;
829 for (se
= 0; se
< num_se
; se
++) {
830 unsigned pkr0_mask
= ((1 << rb_per_pkr
) - 1) << (se
* rb_per_se
);
831 unsigned pkr1_mask
= pkr0_mask
<< rb_per_pkr
;
832 int idx
= (se
/ 2) * 2;
834 raster_config_se
[se
] = raster_config
;
835 if ((num_se
> 1) && (!se_mask
[idx
] || !se_mask
[idx
+ 1])) {
836 raster_config_se
[se
] &= C_028350_SE_MAP
;
839 raster_config_se
[se
] |=
840 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_3
);
842 raster_config_se
[se
] |=
843 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_0
);
847 pkr0_mask
&= rb_mask
;
848 pkr1_mask
&= rb_mask
;
849 if (rb_per_se
> 2 && (!pkr0_mask
|| !pkr1_mask
)) {
850 raster_config_se
[se
] &= C_028350_PKR_MAP
;
853 raster_config_se
[se
] |=
854 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_3
);
856 raster_config_se
[se
] |=
857 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_0
);
861 if (rb_per_se
>= 2) {
862 unsigned rb0_mask
= 1 << (se
* rb_per_se
);
863 unsigned rb1_mask
= rb0_mask
<< 1;
867 if (!rb0_mask
|| !rb1_mask
) {
868 raster_config_se
[se
] &= C_028350_RB_MAP_PKR0
;
871 raster_config_se
[se
] |=
872 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_3
);
874 raster_config_se
[se
] |=
875 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_0
);
880 rb0_mask
= 1 << (se
* rb_per_se
+ rb_per_pkr
);
881 rb1_mask
= rb0_mask
<< 1;
884 if (!rb0_mask
|| !rb1_mask
) {
885 raster_config_se
[se
] &= C_028350_RB_MAP_PKR1
;
888 raster_config_se
[se
] |=
889 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_3
);
891 raster_config_se
[se
] |=
892 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_0
);
900 unsigned ac_get_compute_resource_limits(struct radeon_info
*info
,
901 unsigned waves_per_threadgroup
,
902 unsigned max_waves_per_sh
,
903 unsigned threadgroups_per_cu
)
905 unsigned compute_resource_limits
=
906 S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup
% 4 == 0);
908 if (info
->chip_class
>= GFX7
) {
909 unsigned num_cu_per_se
= info
->num_good_compute_units
/
912 /* Force even distribution on all SIMDs in CU if the workgroup
913 * size is 64. This has shown some good improvements if # of CUs
914 * per SE is not a multiple of 4.
916 if (num_cu_per_se
% 4 && waves_per_threadgroup
== 1)
917 compute_resource_limits
|= S_00B854_FORCE_SIMD_DIST(1);
919 assert(threadgroups_per_cu
>= 1 && threadgroups_per_cu
<= 8);
920 compute_resource_limits
|= S_00B854_WAVES_PER_SH(max_waves_per_sh
) |
921 S_00B854_CU_GROUP_COUNT(threadgroups_per_cu
- 1);
924 if (max_waves_per_sh
) {
925 unsigned limit_div16
= DIV_ROUND_UP(max_waves_per_sh
, 16);
926 compute_resource_limits
|= S_00B854_WAVES_PER_SH_SI(limit_div16
);
929 return compute_resource_limits
;