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"
30 #include "util/u_math.h"
35 #include <amdgpu_drm.h>
39 #define CIK_TILE_MODE_COLOR_2D 14
41 #define CIK__GB_TILE_MODE__PIPE_CONFIG(x) (((x) >> 6) & 0x1f)
42 #define CIK__PIPE_CONFIG__ADDR_SURF_P2 0
43 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16 4
44 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16 5
45 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32 6
46 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32 7
47 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16 8
48 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16 9
49 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16 10
50 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16 11
51 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16 12
52 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32 13
53 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32 14
54 #define CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16 16
55 #define CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16 17
57 static unsigned cik_get_num_tile_pipes(struct amdgpu_gpu_info
*info
)
59 unsigned mode2d
= info
->gb_tile_mode
[CIK_TILE_MODE_COLOR_2D
];
61 switch (CIK__GB_TILE_MODE__PIPE_CONFIG(mode2d
)) {
62 case CIK__PIPE_CONFIG__ADDR_SURF_P2
:
64 case CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16
:
65 case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16
:
66 case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32
:
67 case CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32
:
69 case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16
:
70 case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16
:
71 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16
:
72 case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16
:
73 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16
:
74 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32
:
75 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32
:
77 case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16
:
78 case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16
:
81 fprintf(stderr
, "Invalid CIK pipe configuration, assuming P2\n");
82 assert(!"this should never occur");
87 static bool has_syncobj(int fd
)
90 if (drmGetCap(fd
, DRM_CAP_SYNCOBJ
, &value
))
92 return value
? true : false;
95 bool ac_query_gpu_info(int fd
, amdgpu_device_handle dev
,
96 struct radeon_info
*info
,
97 struct amdgpu_gpu_info
*amdinfo
)
99 struct drm_amdgpu_info_device device_info
= {};
100 struct amdgpu_buffer_size_alignments alignment_info
= {};
101 struct drm_amdgpu_info_hw_ip dma
= {}, compute
= {}, uvd
= {};
102 struct drm_amdgpu_info_hw_ip uvd_enc
= {}, vce
= {}, vcn_dec
= {}, vcn_jpeg
= {};
103 struct drm_amdgpu_info_hw_ip vcn_enc
= {}, gfx
= {};
104 struct amdgpu_gds_resource_info gds
= {};
105 uint32_t vce_version
= 0, vce_feature
= 0, uvd_version
= 0, uvd_feature
= 0;
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 /* Query hardware and driver information. */
122 r
= amdgpu_query_gpu_info(dev
, amdinfo
);
124 fprintf(stderr
, "amdgpu: amdgpu_query_gpu_info failed.\n");
128 r
= amdgpu_query_info(dev
, AMDGPU_INFO_DEV_INFO
, sizeof(device_info
),
131 fprintf(stderr
, "amdgpu: amdgpu_query_info(dev_info) failed.\n");
135 r
= amdgpu_query_buffer_size_alignment(dev
, &alignment_info
);
137 fprintf(stderr
, "amdgpu: amdgpu_query_buffer_size_alignment failed.\n");
141 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_DMA
, 0, &dma
);
143 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(dma) failed.\n");
147 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_GFX
, 0, &gfx
);
149 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(gfx) failed.\n");
153 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_COMPUTE
, 0, &compute
);
155 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(compute) failed.\n");
159 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_UVD
, 0, &uvd
);
161 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(uvd) failed.\n");
165 if (info
->drm_major
== 3 && info
->drm_minor
>= 17) {
166 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_UVD_ENC
, 0, &uvd_enc
);
168 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(uvd_enc) failed.\n");
173 if (info
->drm_major
== 3 && info
->drm_minor
>= 17) {
174 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_VCN_DEC
, 0, &vcn_dec
);
176 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vcn_dec) failed.\n");
181 if (info
->drm_major
== 3 && info
->drm_minor
>= 17) {
182 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_VCN_ENC
, 0, &vcn_enc
);
184 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vcn_enc) failed.\n");
189 if (info
->drm_major
== 3 && info
->drm_minor
>= 27) {
190 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_VCN_JPEG
, 0, &vcn_jpeg
);
192 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vcn_jpeg) failed.\n");
197 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_GFX_ME
, 0, 0,
198 &info
->me_fw_version
,
199 &info
->me_fw_feature
);
201 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(me) failed.\n");
205 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_GFX_PFP
, 0, 0,
206 &info
->pfp_fw_version
,
207 &info
->pfp_fw_feature
);
209 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(pfp) failed.\n");
213 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_GFX_CE
, 0, 0,
214 &info
->ce_fw_version
,
215 &info
->ce_fw_feature
);
217 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(ce) failed.\n");
221 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_UVD
, 0, 0,
222 &uvd_version
, &uvd_feature
);
224 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(uvd) failed.\n");
228 r
= amdgpu_query_hw_ip_info(dev
, AMDGPU_HW_IP_VCE
, 0, &vce
);
230 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vce) failed.\n");
234 r
= amdgpu_query_firmware_version(dev
, AMDGPU_INFO_FW_VCE
, 0, 0,
235 &vce_version
, &vce_feature
);
237 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(vce) failed.\n");
241 r
= amdgpu_query_sw_info(dev
, amdgpu_sw_info_address32_hi
, &info
->address32_hi
);
243 fprintf(stderr
, "amdgpu: amdgpu_query_sw_info(address32_hi) failed.\n");
247 r
= amdgpu_query_gds_info(dev
, &gds
);
249 fprintf(stderr
, "amdgpu: amdgpu_query_gds_info failed.\n");
253 if (info
->drm_minor
>= 9) {
254 struct drm_amdgpu_memory_info meminfo
= {};
256 r
= amdgpu_query_info(dev
, AMDGPU_INFO_MEMORY
, sizeof(meminfo
), &meminfo
);
258 fprintf(stderr
, "amdgpu: amdgpu_query_info(memory) failed.\n");
262 /* Note: usable_heap_size values can be random and can't be relied on. */
263 info
->gart_size
= meminfo
.gtt
.total_heap_size
;
264 info
->vram_size
= meminfo
.vram
.total_heap_size
;
265 info
->vram_vis_size
= meminfo
.cpu_accessible_vram
.total_heap_size
;
267 /* This is a deprecated interface, which reports usable sizes
268 * (total minus pinned), but the pinned size computation is
269 * buggy, so the values returned from these functions can be
272 struct amdgpu_heap_info vram
, vram_vis
, gtt
;
274 r
= amdgpu_query_heap_info(dev
, AMDGPU_GEM_DOMAIN_VRAM
, 0, &vram
);
276 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(vram) failed.\n");
280 r
= amdgpu_query_heap_info(dev
, AMDGPU_GEM_DOMAIN_VRAM
,
281 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
,
284 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(vram_vis) failed.\n");
288 r
= amdgpu_query_heap_info(dev
, AMDGPU_GEM_DOMAIN_GTT
, 0, >t
);
290 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(gtt) failed.\n");
294 info
->gart_size
= gtt
.heap_size
;
295 info
->vram_size
= vram
.heap_size
;
296 info
->vram_vis_size
= vram_vis
.heap_size
;
299 /* Set chip identification. */
300 info
->pci_id
= amdinfo
->asic_id
; /* TODO: is this correct? */
301 info
->vce_harvest_config
= amdinfo
->vce_harvest_config
;
303 switch (info
->pci_id
) {
304 #define CHIPSET(pci_id, cfamily) \
306 info->family = CHIP_##cfamily; \
307 info->name = #cfamily; \
309 #include "pci_ids/radeonsi_pci_ids.h"
313 fprintf(stderr
, "amdgpu: Invalid PCI ID.\n");
317 /* Raven2 uses the same PCI IDs as Raven1, but different revision IDs. */
318 if (info
->family
== CHIP_RAVEN
&& amdinfo
->chip_rev
>= 0x8) {
319 info
->family
= CHIP_RAVEN2
;
320 info
->name
= "RAVEN2";
323 if (info
->family
>= CHIP_VEGA10
)
324 info
->chip_class
= GFX9
;
325 else if (info
->family
>= CHIP_TONGA
)
326 info
->chip_class
= VI
;
327 else if (info
->family
>= CHIP_BONAIRE
)
328 info
->chip_class
= CIK
;
329 else if (info
->family
>= CHIP_TAHITI
)
330 info
->chip_class
= SI
;
332 fprintf(stderr
, "amdgpu: Unknown family.\n");
336 /* Set which chips have dedicated VRAM. */
337 info
->has_dedicated_vram
=
338 !(amdinfo
->ids_flags
& AMDGPU_IDS_FLAGS_FUSION
);
340 /* The kernel can split large buffers in VRAM but not in GTT, so large
341 * allocations can fail or cause buffer movement failures in the kernel.
343 if (info
->has_dedicated_vram
)
344 info
->max_alloc_size
= info
->vram_size
* 0.8;
346 info
->max_alloc_size
= info
->gart_size
* 0.7;
348 /* Set hardware information. */
349 info
->gds_size
= gds
.gds_total_size
;
350 info
->gds_gfx_partition_size
= gds
.gds_gfx_partition_size
;
351 /* convert the shader clock from KHz to MHz */
352 info
->max_shader_clock
= amdinfo
->max_engine_clk
/ 1000;
353 info
->num_tcc_blocks
= device_info
.num_tcc_blocks
;
354 info
->max_se
= amdinfo
->num_shader_engines
;
355 info
->max_sh_per_se
= amdinfo
->num_shader_arrays_per_engine
;
356 info
->has_hw_decode
=
357 (uvd
.available_rings
!= 0) || (vcn_dec
.available_rings
!= 0) ||
358 (vcn_jpeg
.available_rings
!= 0);
359 info
->uvd_fw_version
=
360 uvd
.available_rings
? uvd_version
: 0;
361 info
->vce_fw_version
=
362 vce
.available_rings
? vce_version
: 0;
363 info
->uvd_enc_supported
=
364 uvd_enc
.available_rings
? true : false;
365 info
->has_userptr
= true;
366 info
->has_syncobj
= has_syncobj(fd
);
367 info
->has_syncobj_wait_for_submit
= info
->has_syncobj
&& info
->drm_minor
>= 20;
368 info
->has_fence_to_handle
= info
->has_syncobj
&& info
->drm_minor
>= 21;
369 info
->has_ctx_priority
= info
->drm_minor
>= 22;
370 /* TODO: Enable this once the kernel handles it efficiently. */
371 info
->has_local_buffers
= info
->drm_minor
>= 20 &&
372 !info
->has_dedicated_vram
;
373 info
->kernel_flushes_hdp_before_ib
= true;
374 info
->htile_cmask_support_1d_tiling
= true;
375 info
->si_TA_CS_BC_BASE_ADDR_allowed
= true;
376 info
->has_bo_metadata
= true;
377 info
->has_gpu_reset_status_query
= true;
378 info
->has_gpu_reset_counter_query
= false;
379 info
->has_eqaa_surface_allocator
= true;
380 info
->has_format_bc1_through_bc7
= true;
381 /* DRM 3.1.0 doesn't flush TC for VI correctly. */
382 info
->kernel_flushes_tc_l2_after_ib
= info
->chip_class
!= VI
||
383 info
->drm_minor
>= 2;
384 info
->has_indirect_compute_dispatch
= true;
385 /* SI doesn't support unaligned loads. */
386 info
->has_unaligned_shader_loads
= info
->chip_class
!= SI
;
387 /* Disable sparse mappings on SI due to VM faults in CP DMA. Enable them once
388 * these faults are mitigated in software.
389 * Disable sparse mappings on GFX9 due to hangs.
391 info
->has_sparse_vm_mappings
=
392 info
->chip_class
>= CIK
&& info
->chip_class
<= VI
&&
393 info
->drm_minor
>= 13;
394 info
->has_2d_tiling
= true;
395 info
->has_read_registers_query
= true;
397 info
->num_render_backends
= amdinfo
->rb_pipes
;
398 /* The value returned by the kernel driver was wrong. */
399 if (info
->family
== CHIP_KAVERI
)
400 info
->num_render_backends
= 2;
402 info
->clock_crystal_freq
= amdinfo
->gpu_counter_freq
;
403 if (!info
->clock_crystal_freq
) {
404 fprintf(stderr
, "amdgpu: clock crystal frequency is 0, timestamps will be wrong\n");
405 info
->clock_crystal_freq
= 1;
407 info
->tcc_cache_line_size
= 64; /* TC L2 line size on GCN */
408 info
->gb_addr_config
= amdinfo
->gb_addr_cfg
;
409 if (info
->chip_class
== GFX9
) {
410 info
->num_tile_pipes
= 1 << G_0098F8_NUM_PIPES(amdinfo
->gb_addr_cfg
);
411 info
->pipe_interleave_bytes
=
412 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(amdinfo
->gb_addr_cfg
);
414 info
->num_tile_pipes
= cik_get_num_tile_pipes(amdinfo
);
415 info
->pipe_interleave_bytes
=
416 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX6(amdinfo
->gb_addr_cfg
);
418 info
->r600_has_virtual_memory
= true;
420 assert(util_is_power_of_two_or_zero(dma
.available_rings
+ 1));
421 assert(util_is_power_of_two_or_zero(compute
.available_rings
+ 1));
423 info
->num_sdma_rings
= util_bitcount(dma
.available_rings
);
424 info
->num_compute_rings
= util_bitcount(compute
.available_rings
);
426 /* Get the number of good compute units. */
427 info
->num_good_compute_units
= 0;
428 for (i
= 0; i
< info
->max_se
; i
++)
429 for (j
= 0; j
< info
->max_sh_per_se
; j
++)
430 info
->num_good_compute_units
+=
431 util_bitcount(amdinfo
->cu_bitmap
[i
][j
]);
432 info
->num_good_cu_per_sh
= info
->num_good_compute_units
/
433 (info
->max_se
* info
->max_sh_per_se
);
435 memcpy(info
->si_tile_mode_array
, amdinfo
->gb_tile_mode
,
436 sizeof(amdinfo
->gb_tile_mode
));
437 info
->enabled_rb_mask
= amdinfo
->enabled_rb_pipes_mask
;
439 memcpy(info
->cik_macrotile_mode_array
, amdinfo
->gb_macro_tile_mode
,
440 sizeof(amdinfo
->gb_macro_tile_mode
));
442 info
->pte_fragment_size
= alignment_info
.size_local
;
443 info
->gart_page_size
= alignment_info
.size_remote
;
445 if (info
->chip_class
== SI
)
446 info
->gfx_ib_pad_with_type2
= TRUE
;
448 unsigned ib_align
= 0;
449 ib_align
= MAX2(ib_align
, gfx
.ib_start_alignment
);
450 ib_align
= MAX2(ib_align
, compute
.ib_start_alignment
);
451 ib_align
= MAX2(ib_align
, dma
.ib_start_alignment
);
452 ib_align
= MAX2(ib_align
, uvd
.ib_start_alignment
);
453 ib_align
= MAX2(ib_align
, uvd_enc
.ib_start_alignment
);
454 ib_align
= MAX2(ib_align
, vce
.ib_start_alignment
);
455 ib_align
= MAX2(ib_align
, vcn_dec
.ib_start_alignment
);
456 ib_align
= MAX2(ib_align
, vcn_enc
.ib_start_alignment
);
457 ib_align
= MAX2(ib_align
, vcn_jpeg
.ib_start_alignment
);
459 info
->ib_start_alignment
= ib_align
;
464 void ac_compute_driver_uuid(char *uuid
, size_t size
)
466 char amd_uuid
[] = "AMD-MESA-DRV";
468 assert(size
>= sizeof(amd_uuid
));
470 memset(uuid
, 0, size
);
471 strncpy(uuid
, amd_uuid
, size
);
474 void ac_compute_device_uuid(struct radeon_info
*info
, char *uuid
, size_t size
)
476 uint32_t *uint_uuid
= (uint32_t*)uuid
;
478 assert(size
>= sizeof(uint32_t)*4);
481 * Use the device info directly instead of using a sha1. GL/VK UUIDs
482 * are 16 byte vs 20 byte for sha1, and the truncation that would be
483 * required would get rid of part of the little entropy we have.
485 memset(uuid
, 0, size
);
486 uint_uuid
[0] = info
->pci_domain
;
487 uint_uuid
[1] = info
->pci_bus
;
488 uint_uuid
[2] = info
->pci_dev
;
489 uint_uuid
[3] = info
->pci_func
;
492 void ac_print_gpu_info(struct radeon_info
*info
)
494 printf("Device info:\n");
495 printf(" pci (domain:bus:dev.func): %04x:%02x:%02x.%x\n",
496 info
->pci_domain
, info
->pci_bus
,
497 info
->pci_dev
, info
->pci_func
);
498 printf(" pci_id = 0x%x\n", info
->pci_id
);
499 printf(" family = %i\n", info
->family
);
500 printf(" chip_class = %i\n", info
->chip_class
);
501 printf(" num_compute_rings = %u\n", info
->num_compute_rings
);
502 printf(" num_sdma_rings = %i\n", info
->num_sdma_rings
);
503 printf(" clock_crystal_freq = %i\n", info
->clock_crystal_freq
);
504 printf(" tcc_cache_line_size = %u\n", info
->tcc_cache_line_size
);
506 printf(" use_display_dcc_unaligned = %u\n", info
->use_display_dcc_unaligned
);
507 printf(" use_display_dcc_with_retile_blit = %u\n", info
->use_display_dcc_with_retile_blit
);
509 printf("Memory info:\n");
510 printf(" pte_fragment_size = %u\n", info
->pte_fragment_size
);
511 printf(" gart_page_size = %u\n", info
->gart_page_size
);
512 printf(" gart_size = %i MB\n", (int)DIV_ROUND_UP(info
->gart_size
, 1024*1024));
513 printf(" vram_size = %i MB\n", (int)DIV_ROUND_UP(info
->vram_size
, 1024*1024));
514 printf(" vram_vis_size = %i MB\n", (int)DIV_ROUND_UP(info
->vram_vis_size
, 1024*1024));
515 printf(" gds_size = %u kB\n", info
->gds_size
/ 1024);
516 printf(" gds_gfx_partition_size = %u kB\n", info
->gds_gfx_partition_size
/ 1024);
517 printf(" max_alloc_size = %i MB\n",
518 (int)DIV_ROUND_UP(info
->max_alloc_size
, 1024*1024));
519 printf(" min_alloc_size = %u\n", info
->min_alloc_size
);
520 printf(" address32_hi = %u\n", info
->address32_hi
);
521 printf(" has_dedicated_vram = %u\n", info
->has_dedicated_vram
);
523 printf("CP info:\n");
524 printf(" gfx_ib_pad_with_type2 = %i\n", info
->gfx_ib_pad_with_type2
);
525 printf(" ib_start_alignment = %u\n", info
->ib_start_alignment
);
526 printf(" me_fw_version = %i\n", info
->me_fw_version
);
527 printf(" me_fw_feature = %i\n", info
->me_fw_feature
);
528 printf(" pfp_fw_version = %i\n", info
->pfp_fw_version
);
529 printf(" pfp_fw_feature = %i\n", info
->pfp_fw_feature
);
530 printf(" ce_fw_version = %i\n", info
->ce_fw_version
);
531 printf(" ce_fw_feature = %i\n", info
->ce_fw_feature
);
533 printf("Multimedia info:\n");
534 printf(" has_hw_decode = %u\n", info
->has_hw_decode
);
535 printf(" uvd_enc_supported = %u\n", info
->uvd_enc_supported
);
536 printf(" uvd_fw_version = %u\n", info
->uvd_fw_version
);
537 printf(" vce_fw_version = %u\n", info
->vce_fw_version
);
538 printf(" vce_harvest_config = %i\n", info
->vce_harvest_config
);
540 printf("Kernel & winsys capabilities:\n");
541 printf(" drm = %i.%i.%i\n", info
->drm_major
,
542 info
->drm_minor
, info
->drm_patchlevel
);
543 printf(" has_userptr = %i\n", info
->has_userptr
);
544 printf(" has_syncobj = %u\n", info
->has_syncobj
);
545 printf(" has_syncobj_wait_for_submit = %u\n", info
->has_syncobj_wait_for_submit
);
546 printf(" has_fence_to_handle = %u\n", info
->has_fence_to_handle
);
547 printf(" has_ctx_priority = %u\n", info
->has_ctx_priority
);
548 printf(" has_local_buffers = %u\n", info
->has_local_buffers
);
549 printf(" kernel_flushes_hdp_before_ib = %u\n", info
->kernel_flushes_hdp_before_ib
);
550 printf(" htile_cmask_support_1d_tiling = %u\n", info
->htile_cmask_support_1d_tiling
);
551 printf(" si_TA_CS_BC_BASE_ADDR_allowed = %u\n", info
->si_TA_CS_BC_BASE_ADDR_allowed
);
552 printf(" has_bo_metadata = %u\n", info
->has_bo_metadata
);
553 printf(" has_gpu_reset_status_query = %u\n", info
->has_gpu_reset_status_query
);
554 printf(" has_gpu_reset_counter_query = %u\n", info
->has_gpu_reset_counter_query
);
555 printf(" has_eqaa_surface_allocator = %u\n", info
->has_eqaa_surface_allocator
);
556 printf(" has_format_bc1_through_bc7 = %u\n", info
->has_format_bc1_through_bc7
);
557 printf(" kernel_flushes_tc_l2_after_ib = %u\n", info
->kernel_flushes_tc_l2_after_ib
);
558 printf(" has_indirect_compute_dispatch = %u\n", info
->has_indirect_compute_dispatch
);
559 printf(" has_unaligned_shader_loads = %u\n", info
->has_unaligned_shader_loads
);
560 printf(" has_sparse_vm_mappings = %u\n", info
->has_sparse_vm_mappings
);
561 printf(" has_2d_tiling = %u\n", info
->has_2d_tiling
);
562 printf(" has_read_registers_query = %u\n", info
->has_read_registers_query
);
564 printf("Shader core info:\n");
565 printf(" max_shader_clock = %i\n", info
->max_shader_clock
);
566 printf(" num_good_compute_units = %i\n", info
->num_good_compute_units
);
567 printf(" num_good_cu_per_sh = %i\n", info
->num_good_cu_per_sh
);
568 printf(" num_tcc_blocks = %i\n", info
->num_tcc_blocks
);
569 printf(" max_se = %i\n", info
->max_se
);
570 printf(" max_sh_per_se = %i\n", info
->max_sh_per_se
);
572 printf("Render backend info:\n");
573 printf(" num_render_backends = %i\n", info
->num_render_backends
);
574 printf(" num_tile_pipes = %i\n", info
->num_tile_pipes
);
575 printf(" pipe_interleave_bytes = %i\n", info
->pipe_interleave_bytes
);
576 printf(" enabled_rb_mask = 0x%x\n", info
->enabled_rb_mask
);
577 printf(" max_alignment = %u\n", (unsigned)info
->max_alignment
);
579 printf("GB_ADDR_CONFIG:\n");
580 if (info
->chip_class
>= GFX9
) {
581 printf(" num_pipes = %u\n",
582 1 << G_0098F8_NUM_PIPES(info
->gb_addr_config
));
583 printf(" pipe_interleave_size = %u\n",
584 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(info
->gb_addr_config
));
585 printf(" max_compressed_frags = %u\n",
586 1 << G_0098F8_MAX_COMPRESSED_FRAGS(info
->gb_addr_config
));
587 printf(" bank_interleave_size = %u\n",
588 1 << G_0098F8_BANK_INTERLEAVE_SIZE(info
->gb_addr_config
));
589 printf(" num_banks = %u\n",
590 1 << G_0098F8_NUM_BANKS(info
->gb_addr_config
));
591 printf(" shader_engine_tile_size = %u\n",
592 16 << G_0098F8_SHADER_ENGINE_TILE_SIZE(info
->gb_addr_config
));
593 printf(" num_shader_engines = %u\n",
594 1 << G_0098F8_NUM_SHADER_ENGINES_GFX9(info
->gb_addr_config
));
595 printf(" num_gpus = %u (raw)\n",
596 G_0098F8_NUM_GPUS_GFX9(info
->gb_addr_config
));
597 printf(" multi_gpu_tile_size = %u (raw)\n",
598 G_0098F8_MULTI_GPU_TILE_SIZE(info
->gb_addr_config
));
599 printf(" num_rb_per_se = %u\n",
600 1 << G_0098F8_NUM_RB_PER_SE(info
->gb_addr_config
));
601 printf(" row_size = %u\n",
602 1024 << G_0098F8_ROW_SIZE(info
->gb_addr_config
));
603 printf(" num_lower_pipes = %u (raw)\n",
604 G_0098F8_NUM_LOWER_PIPES(info
->gb_addr_config
));
605 printf(" se_enable = %u (raw)\n",
606 G_0098F8_SE_ENABLE(info
->gb_addr_config
));
608 printf(" num_pipes = %u\n",
609 1 << G_0098F8_NUM_PIPES(info
->gb_addr_config
));
610 printf(" pipe_interleave_size = %u\n",
611 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX6(info
->gb_addr_config
));
612 printf(" bank_interleave_size = %u\n",
613 1 << G_0098F8_BANK_INTERLEAVE_SIZE(info
->gb_addr_config
));
614 printf(" num_shader_engines = %u\n",
615 1 << G_0098F8_NUM_SHADER_ENGINES_GFX6(info
->gb_addr_config
));
616 printf(" shader_engine_tile_size = %u\n",
617 16 << G_0098F8_SHADER_ENGINE_TILE_SIZE(info
->gb_addr_config
));
618 printf(" num_gpus = %u (raw)\n",
619 G_0098F8_NUM_GPUS_GFX6(info
->gb_addr_config
));
620 printf(" multi_gpu_tile_size = %u (raw)\n",
621 G_0098F8_MULTI_GPU_TILE_SIZE(info
->gb_addr_config
));
622 printf(" row_size = %u\n",
623 1024 << G_0098F8_ROW_SIZE(info
->gb_addr_config
));
624 printf(" num_lower_pipes = %u (raw)\n",
625 G_0098F8_NUM_LOWER_PIPES(info
->gb_addr_config
));
630 ac_get_gs_table_depth(enum chip_class chip_class
, enum radeon_family family
)
632 if (chip_class
>= GFX9
)
658 unreachable("Unknown GPU");
663 ac_get_raster_config(struct radeon_info
*info
,
664 uint32_t *raster_config_p
,
665 uint32_t *raster_config_1_p
,
666 uint32_t *se_tile_repeat_p
)
668 unsigned raster_config
, raster_config_1
, se_tile_repeat
;
670 switch (info
->family
) {
676 raster_config
= 0x00000000;
677 raster_config_1
= 0x00000000;
681 raster_config
= 0x0000124a;
682 raster_config_1
= 0x00000000;
684 /* 1 SE / 2 RBs (Oland is special) */
686 raster_config
= 0x00000082;
687 raster_config_1
= 0x00000000;
693 raster_config
= 0x00000002;
694 raster_config_1
= 0x00000000;
700 raster_config
= 0x16000012;
701 raster_config_1
= 0x00000000;
706 raster_config
= 0x2a00126a;
707 raster_config_1
= 0x00000000;
712 raster_config
= 0x16000012;
713 raster_config_1
= 0x0000002a;
719 raster_config
= 0x3a00161a;
720 raster_config_1
= 0x0000002e;
724 "ac: Unknown GPU, using 0 for raster_config\n");
725 raster_config
= 0x00000000;
726 raster_config_1
= 0x00000000;
730 /* drm/radeon on Kaveri is buggy, so disable 1 RB to work around it.
731 * This decreases performance by up to 50% when the RB is the bottleneck.
733 if (info
->family
== CHIP_KAVERI
&& info
->drm_major
== 2)
734 raster_config
= 0x00000000;
736 /* Fiji: Old kernels have incorrect tiling config. This decreases
737 * RB performance by 25%. (it disables 1 RB in the second packer)
739 if (info
->family
== CHIP_FIJI
&&
740 info
->cik_macrotile_mode_array
[0] == 0x000000e8) {
741 raster_config
= 0x16000012;
742 raster_config_1
= 0x0000002a;
745 unsigned se_width
= 8 << G_028350_SE_XSEL_GFX6(raster_config
);
746 unsigned se_height
= 8 << G_028350_SE_YSEL_GFX6(raster_config
);
748 /* I don't know how to calculate this, though this is probably a good guess. */
749 se_tile_repeat
= MAX2(se_width
, se_height
) * info
->max_se
;
751 *raster_config_p
= raster_config
;
752 *raster_config_1_p
= raster_config_1
;
753 if (se_tile_repeat_p
)
754 *se_tile_repeat_p
= se_tile_repeat
;
758 ac_get_harvested_configs(struct radeon_info
*info
,
759 unsigned raster_config
,
760 unsigned *cik_raster_config_1_p
,
761 unsigned *raster_config_se
)
763 unsigned sh_per_se
= MAX2(info
->max_sh_per_se
, 1);
764 unsigned num_se
= MAX2(info
->max_se
, 1);
765 unsigned rb_mask
= info
->enabled_rb_mask
;
766 unsigned num_rb
= MIN2(info
->num_render_backends
, 16);
767 unsigned rb_per_pkr
= MIN2(num_rb
/ num_se
/ sh_per_se
, 2);
768 unsigned rb_per_se
= num_rb
/ num_se
;
772 se_mask
[0] = ((1 << rb_per_se
) - 1) & rb_mask
;
773 se_mask
[1] = (se_mask
[0] << rb_per_se
) & rb_mask
;
774 se_mask
[2] = (se_mask
[1] << rb_per_se
) & rb_mask
;
775 se_mask
[3] = (se_mask
[2] << rb_per_se
) & rb_mask
;
777 assert(num_se
== 1 || num_se
== 2 || num_se
== 4);
778 assert(sh_per_se
== 1 || sh_per_se
== 2);
779 assert(rb_per_pkr
== 1 || rb_per_pkr
== 2);
782 if (info
->chip_class
>= CIK
) {
783 unsigned raster_config_1
= *cik_raster_config_1_p
;
784 if ((num_se
> 2) && ((!se_mask
[0] && !se_mask
[1]) ||
785 (!se_mask
[2] && !se_mask
[3]))) {
786 raster_config_1
&= C_028354_SE_PAIR_MAP
;
788 if (!se_mask
[0] && !se_mask
[1]) {
790 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_3
);
793 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_0
);
795 *cik_raster_config_1_p
= raster_config_1
;
799 for (se
= 0; se
< num_se
; se
++) {
800 unsigned pkr0_mask
= ((1 << rb_per_pkr
) - 1) << (se
* rb_per_se
);
801 unsigned pkr1_mask
= pkr0_mask
<< rb_per_pkr
;
802 int idx
= (se
/ 2) * 2;
804 raster_config_se
[se
] = raster_config
;
805 if ((num_se
> 1) && (!se_mask
[idx
] || !se_mask
[idx
+ 1])) {
806 raster_config_se
[se
] &= C_028350_SE_MAP
;
809 raster_config_se
[se
] |=
810 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_3
);
812 raster_config_se
[se
] |=
813 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_0
);
817 pkr0_mask
&= rb_mask
;
818 pkr1_mask
&= rb_mask
;
819 if (rb_per_se
> 2 && (!pkr0_mask
|| !pkr1_mask
)) {
820 raster_config_se
[se
] &= C_028350_PKR_MAP
;
823 raster_config_se
[se
] |=
824 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_3
);
826 raster_config_se
[se
] |=
827 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_0
);
831 if (rb_per_se
>= 2) {
832 unsigned rb0_mask
= 1 << (se
* rb_per_se
);
833 unsigned rb1_mask
= rb0_mask
<< 1;
837 if (!rb0_mask
|| !rb1_mask
) {
838 raster_config_se
[se
] &= C_028350_RB_MAP_PKR0
;
841 raster_config_se
[se
] |=
842 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_3
);
844 raster_config_se
[se
] |=
845 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_0
);
850 rb0_mask
= 1 << (se
* rb_per_se
+ rb_per_pkr
);
851 rb1_mask
= rb0_mask
<< 1;
854 if (!rb0_mask
|| !rb1_mask
) {
855 raster_config_se
[se
] &= C_028350_RB_MAP_PKR1
;
858 raster_config_se
[se
] |=
859 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_3
);
861 raster_config_se
[se
] |=
862 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_0
);