• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2017 Advanced Micro Devices, Inc.
3  *
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:
11  *
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.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  */
25 
26 #include "ac_gpu_info.h"
27 #include "ac_shader_util.h"
28 #include "ac_debug.h"
29 
30 #include "addrlib/src/amdgpu_asic_addr.h"
31 #include "sid.h"
32 #include "util/macros.h"
33 #include "util/u_cpu_detect.h"
34 #include "util/u_math.h"
35 #include "util/os_misc.h"
36 #include "util/bitset.h"
37 
38 #include <stdio.h>
39 #include <ctype.h>
40 
41 #define AMDGPU_ARCTURUS_RANGE   0x32, 0x3C
42 #define AMDGPU_ALDEBARAN_RANGE  0x3C, 0xFF
43 
44 #define ASICREV_IS_ARCTURUS(r)         ASICREV_IS(r, ARCTURUS)
45 #define ASICREV_IS_ALDEBARAN(r)        ASICREV_IS(r, ALDEBARAN)
46 
47 #ifdef _WIN32
48 #define DRM_CAP_ADDFB2_MODIFIERS 0x10
49 #define DRM_CAP_SYNCOBJ 0x13
50 #define DRM_CAP_SYNCOBJ_TIMELINE 0x14
51 #define AMDGPU_GEM_DOMAIN_GTT 0x2
52 #define AMDGPU_GEM_DOMAIN_VRAM 0x4
53 #define AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED (1 << 0)
54 #define AMDGPU_GEM_CREATE_ENCRYPTED (1 << 10)
55 #define AMDGPU_HW_IP_GFX 0
56 #define AMDGPU_HW_IP_COMPUTE 1
57 #define AMDGPU_HW_IP_DMA 2
58 #define AMDGPU_HW_IP_UVD 3
59 #define AMDGPU_HW_IP_VCE 4
60 #define AMDGPU_HW_IP_UVD_ENC 5
61 #define AMDGPU_HW_IP_VCN_DEC 6
62 #define AMDGPU_HW_IP_VCN_ENC 7
63 #define AMDGPU_HW_IP_VCN_JPEG 8
64 #define AMDGPU_IDS_FLAGS_FUSION 0x1
65 #define AMDGPU_IDS_FLAGS_PREEMPTION 0x2
66 #define AMDGPU_IDS_FLAGS_TMZ 0x4
67 #define AMDGPU_INFO_FW_VCE 0x1
68 #define AMDGPU_INFO_FW_UVD 0x2
69 #define AMDGPU_INFO_FW_GFX_ME 0x04
70 #define AMDGPU_INFO_FW_GFX_PFP 0x05
71 #define AMDGPU_INFO_FW_GFX_CE 0x06
72 #define AMDGPU_INFO_DEV_INFO 0x16
73 #define AMDGPU_INFO_MEMORY 0x19
74 #define AMDGPU_INFO_VIDEO_CAPS_DECODE 0
75 #define AMDGPU_INFO_VIDEO_CAPS_ENCODE 1
76 #define AMDGPU_INFO_FW_GFX_MEC 0x08
77 
78 #define AMDGPU_VRAM_TYPE_UNKNOWN 0
79 #define AMDGPU_VRAM_TYPE_GDDR1 1
80 #define AMDGPU_VRAM_TYPE_DDR2  2
81 #define AMDGPU_VRAM_TYPE_GDDR3 3
82 #define AMDGPU_VRAM_TYPE_GDDR4 4
83 #define AMDGPU_VRAM_TYPE_GDDR5 5
84 #define AMDGPU_VRAM_TYPE_HBM   6
85 #define AMDGPU_VRAM_TYPE_DDR3  7
86 #define AMDGPU_VRAM_TYPE_DDR4  8
87 #define AMDGPU_VRAM_TYPE_GDDR6 9
88 #define AMDGPU_VRAM_TYPE_DDR5  10
89 
90 struct drm_amdgpu_heap_info {
91    uint64_t total_heap_size;
92 };
93 struct drm_amdgpu_memory_info {
94    struct drm_amdgpu_heap_info vram;
95    struct drm_amdgpu_heap_info cpu_accessible_vram;
96    struct drm_amdgpu_heap_info gtt;
97 };
98 struct drm_amdgpu_info_device {
99    /** PCI Device ID */
100    uint32_t device_id;
101    /** Internal chip revision: A0, A1, etc.) */
102    uint32_t chip_rev;
103    uint32_t external_rev;
104    /** Revision id in PCI Config space */
105    uint32_t pci_rev;
106    uint32_t family;
107    uint32_t num_shader_engines;
108    uint32_t num_shader_arrays_per_engine;
109    /* in KHz */
110    uint32_t gpu_counter_freq;
111    uint64_t max_engine_clock;
112    uint64_t max_memory_clock;
113    /* cu information */
114    uint32_t cu_active_number;
115    /* NOTE: cu_ao_mask is INVALID, DON'T use it */
116    uint32_t cu_ao_mask;
117    uint32_t cu_bitmap[4][4];
118    /** Render backend pipe mask. One render backend is CB+DB. */
119    uint32_t enabled_rb_pipes_mask;
120    uint32_t num_rb_pipes;
121    uint32_t num_hw_gfx_contexts;
122    uint32_t _pad;
123    uint64_t ids_flags;
124    /** Starting virtual address for UMDs. */
125    uint64_t virtual_address_offset;
126    /** The maximum virtual address */
127    uint64_t virtual_address_max;
128    /** Required alignment of virtual addresses. */
129    uint32_t virtual_address_alignment;
130    /** Page table entry - fragment size */
131    uint32_t pte_fragment_size;
132    uint32_t gart_page_size;
133    /** constant engine ram size*/
134    uint32_t ce_ram_size;
135    /** video memory type info*/
136    uint32_t vram_type;
137    /** video memory bit width*/
138    uint32_t vram_bit_width;
139    /* vce harvesting instance */
140    uint32_t vce_harvest_config;
141    /* gfx double offchip LDS buffers */
142    uint32_t gc_double_offchip_lds_buf;
143    /* NGG Primitive Buffer */
144    uint64_t prim_buf_gpu_addr;
145    /* NGG Position Buffer */
146    uint64_t pos_buf_gpu_addr;
147    /* NGG Control Sideband */
148    uint64_t cntl_sb_buf_gpu_addr;
149    /* NGG Parameter Cache */
150    uint64_t param_buf_gpu_addr;
151    uint32_t prim_buf_size;
152    uint32_t pos_buf_size;
153    uint32_t cntl_sb_buf_size;
154    uint32_t param_buf_size;
155    /* wavefront size*/
156    uint32_t wave_front_size;
157    /* shader visible vgprs*/
158    uint32_t num_shader_visible_vgprs;
159    /* CU per shader array*/
160    uint32_t num_cu_per_sh;
161    /* number of tcc blocks*/
162    uint32_t num_tcc_blocks;
163    /* gs vgt table depth*/
164    uint32_t gs_vgt_table_depth;
165    /* gs primitive buffer depth*/
166    uint32_t gs_prim_buffer_depth;
167    /* max gs wavefront per vgt*/
168    uint32_t max_gs_waves_per_vgt;
169    uint32_t _pad1;
170    /* always on cu bitmap */
171    uint32_t cu_ao_bitmap[4][4];
172    /** Starting high virtual address for UMDs. */
173    uint64_t high_va_offset;
174    /** The maximum high virtual address */
175    uint64_t high_va_max;
176    /* gfx10 pa_sc_tile_steering_override */
177    uint32_t pa_sc_tile_steering_override;
178    /* disabled TCCs */
179    uint64_t tcc_disabled_mask;
180 };
181 struct drm_amdgpu_info_hw_ip {
182    uint32_t hw_ip_version_major;
183    uint32_t hw_ip_version_minor;
184    uint32_t ib_start_alignment;
185    uint32_t ib_size_alignment;
186    uint32_t available_rings;
187    uint32_t ip_discovery_version;
188 };
189 typedef struct _drmPciBusInfo {
190    uint16_t domain;
191    uint8_t bus;
192    uint8_t dev;
193    uint8_t func;
194 } drmPciBusInfo, *drmPciBusInfoPtr;
195 typedef struct _drmDevice {
196    union {
197       drmPciBusInfoPtr pci;
198    } businfo;
199 } drmDevice, *drmDevicePtr;
200 enum amdgpu_sw_info {
201    amdgpu_sw_info_address32_hi = 0,
202 };
203 typedef struct amdgpu_device *amdgpu_device_handle;
204 typedef struct amdgpu_bo *amdgpu_bo_handle;
205 struct amdgpu_bo_alloc_request {
206    uint64_t alloc_size;
207    uint64_t phys_alignment;
208    uint32_t preferred_heap;
209    uint64_t flags;
210 };
211 struct amdgpu_gds_resource_info {
212    uint32_t gds_gfx_partition_size;
213    uint32_t gds_total_size;
214 };
215 struct amdgpu_buffer_size_alignments {
216    uint64_t size_local;
217    uint64_t size_remote;
218 };
219 struct amdgpu_heap_info {
220    uint64_t heap_size;
221 };
222 struct amdgpu_gpu_info {
223    uint32_t asic_id;
224    uint32_t chip_external_rev;
225    uint32_t family_id;
226    uint64_t ids_flags;
227    uint64_t max_engine_clk;
228    uint64_t max_memory_clk;
229    uint32_t num_shader_engines;
230    uint32_t num_shader_arrays_per_engine;
231    uint32_t rb_pipes;
232    uint32_t enabled_rb_pipes_mask;
233    uint32_t gpu_counter_freq;
234    uint32_t mc_arb_ramcfg;
235    uint32_t gb_addr_cfg;
236    uint32_t gb_tile_mode[32];
237    uint32_t gb_macro_tile_mode[16];
238    uint32_t cu_bitmap[4][4];
239    uint32_t vram_type;
240    uint32_t vram_bit_width;
241    uint32_t ce_ram_size;
242    uint32_t vce_harvest_config;
243    uint32_t pci_rev_id;
244 };
drmGetCap(int fd,uint64_t capability,uint64_t * value)245 static int drmGetCap(int fd, uint64_t capability, uint64_t *value)
246 {
247    return -EINVAL;
248 }
drmFreeDevice(drmDevicePtr * device)249 static void drmFreeDevice(drmDevicePtr *device)
250 {
251 }
drmGetDevice2(int fd,uint32_t flags,drmDevicePtr * device)252 static int drmGetDevice2(int fd, uint32_t flags, drmDevicePtr *device)
253 {
254    return -ENODEV;
255 }
amdgpu_bo_alloc(amdgpu_device_handle dev,struct amdgpu_bo_alloc_request * alloc_buffer,amdgpu_bo_handle * buf_handle)256 static int amdgpu_bo_alloc(amdgpu_device_handle dev,
257    struct amdgpu_bo_alloc_request *alloc_buffer,
258    amdgpu_bo_handle *buf_handle)
259 {
260    return -EINVAL;
261 }
amdgpu_bo_free(amdgpu_bo_handle buf_handle)262 static int amdgpu_bo_free(amdgpu_bo_handle buf_handle)
263 {
264    return -EINVAL;
265 }
amdgpu_query_buffer_size_alignment(amdgpu_device_handle dev,struct amdgpu_buffer_size_alignments * info)266 static int amdgpu_query_buffer_size_alignment(amdgpu_device_handle dev,
267    struct amdgpu_buffer_size_alignments
268    *info)
269 {
270    return -EINVAL;
271 }
amdgpu_query_firmware_version(amdgpu_device_handle dev,unsigned fw_type,unsigned ip_instance,unsigned index,uint32_t * version,uint32_t * feature)272 static int amdgpu_query_firmware_version(amdgpu_device_handle dev, unsigned fw_type,
273    unsigned ip_instance, unsigned index,
274    uint32_t *version, uint32_t *feature)
275 {
276    return -EINVAL;
277 }
amdgpu_query_hw_ip_info(amdgpu_device_handle dev,unsigned type,unsigned ip_instance,struct drm_amdgpu_info_hw_ip * info)278 static int amdgpu_query_hw_ip_info(amdgpu_device_handle dev, unsigned type,
279    unsigned ip_instance,
280    struct drm_amdgpu_info_hw_ip *info)
281 {
282    return -EINVAL;
283 }
amdgpu_query_heap_info(amdgpu_device_handle dev,uint32_t heap,uint32_t flags,struct amdgpu_heap_info * info)284 static int amdgpu_query_heap_info(amdgpu_device_handle dev, uint32_t heap,
285    uint32_t flags, struct amdgpu_heap_info *info)
286 {
287    return -EINVAL;
288 }
amdgpu_query_gpu_info(amdgpu_device_handle dev,struct amdgpu_gpu_info * info)289 static int amdgpu_query_gpu_info(amdgpu_device_handle dev,
290    struct amdgpu_gpu_info *info)
291 {
292    return -EINVAL;
293 }
amdgpu_query_info(amdgpu_device_handle dev,unsigned info_id,unsigned size,void * value)294 static int amdgpu_query_info(amdgpu_device_handle dev, unsigned info_id,
295    unsigned size, void *value)
296 {
297    return -EINVAL;
298 }
amdgpu_query_sw_info(amdgpu_device_handle dev,enum amdgpu_sw_info info,void * value)299 static int amdgpu_query_sw_info(amdgpu_device_handle dev, enum amdgpu_sw_info info,
300    void *value)
301 {
302    return -EINVAL;
303 }
amdgpu_query_gds_info(amdgpu_device_handle dev,struct amdgpu_gds_resource_info * gds_info)304 static int amdgpu_query_gds_info(amdgpu_device_handle dev,
305    struct amdgpu_gds_resource_info *gds_info)
306 {
307    return -EINVAL;
308 }
amdgpu_query_video_caps_info(amdgpu_device_handle dev,unsigned cap_type,unsigned size,void * value)309 static int amdgpu_query_video_caps_info(amdgpu_device_handle dev, unsigned cap_type,
310                                  unsigned size, void *value)
311 {
312    return -EINVAL;
313 }
amdgpu_get_marketing_name(amdgpu_device_handle dev)314 static const char *amdgpu_get_marketing_name(amdgpu_device_handle dev)
315 {
316    return NULL;
317 }
318 #else
319 #include "drm-uapi/amdgpu_drm.h"
320 #include <amdgpu.h>
321 #include <xf86drm.h>
322 #endif
323 
324 #define CIK_TILE_MODE_COLOR_2D 14
325 
326 #define CIK__GB_TILE_MODE__PIPE_CONFIG(x)           (((x) >> 6) & 0x1f)
327 #define CIK__PIPE_CONFIG__ADDR_SURF_P2              0
328 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16         4
329 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16        5
330 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32        6
331 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32        7
332 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16   8
333 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16   9
334 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16   10
335 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16  11
336 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16  12
337 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32  13
338 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32  14
339 #define CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16  16
340 #define CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16 17
341 
cik_get_num_tile_pipes(struct amdgpu_gpu_info * info)342 static unsigned cik_get_num_tile_pipes(struct amdgpu_gpu_info *info)
343 {
344    unsigned mode2d = info->gb_tile_mode[CIK_TILE_MODE_COLOR_2D];
345 
346    switch (CIK__GB_TILE_MODE__PIPE_CONFIG(mode2d)) {
347    case CIK__PIPE_CONFIG__ADDR_SURF_P2:
348       return 2;
349    case CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16:
350    case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16:
351    case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32:
352    case CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32:
353       return 4;
354    case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16:
355    case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16:
356    case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16:
357    case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16:
358    case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16:
359    case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32:
360    case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32:
361       return 8;
362    case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16:
363    case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16:
364       return 16;
365    default:
366       fprintf(stderr, "Invalid GFX7 pipe configuration, assuming P2\n");
367       assert(!"this should never occur");
368       return 2;
369    }
370 }
371 
has_syncobj(int fd)372 static bool has_syncobj(int fd)
373 {
374    uint64_t value;
375    if (drmGetCap(fd, DRM_CAP_SYNCOBJ, &value))
376       return false;
377    return value ? true : false;
378 }
379 
has_timeline_syncobj(int fd)380 static bool has_timeline_syncobj(int fd)
381 {
382    uint64_t value;
383    if (drmGetCap(fd, DRM_CAP_SYNCOBJ_TIMELINE, &value))
384       return false;
385    return value ? true : false;
386 }
387 
has_modifiers(int fd)388 static bool has_modifiers(int fd)
389 {
390    uint64_t value;
391    if (drmGetCap(fd, DRM_CAP_ADDFB2_MODIFIERS, &value))
392       return false;
393    return value ? true : false;
394 }
395 
fix_vram_size(uint64_t size)396 static uint64_t fix_vram_size(uint64_t size)
397 {
398    /* The VRAM size is underreported, so we need to fix it, because
399     * it's used to compute the number of memory modules for harvesting.
400     */
401    return align64(size, 256 * 1024 * 1024);
402 }
403 
404 static bool
has_tmz_support(amdgpu_device_handle dev,struct radeon_info * info,uint32_t ids_flags)405 has_tmz_support(amdgpu_device_handle dev, struct radeon_info *info, uint32_t ids_flags)
406 {
407    struct amdgpu_bo_alloc_request request = {0};
408    int r;
409    amdgpu_bo_handle bo;
410 
411    if (ids_flags & AMDGPU_IDS_FLAGS_TMZ)
412       return true;
413 
414    /* AMDGPU_IDS_FLAGS_TMZ is supported starting from drm_minor 40 */
415    if (info->drm_minor >= 40)
416       return false;
417 
418    /* Find out ourselves if TMZ is enabled */
419    if (info->gfx_level < GFX9)
420       return false;
421 
422    if (info->drm_minor < 36)
423       return false;
424 
425    request.alloc_size = 256;
426    request.phys_alignment = 1024;
427    request.preferred_heap = AMDGPU_GEM_DOMAIN_VRAM;
428    request.flags = AMDGPU_GEM_CREATE_ENCRYPTED;
429    r = amdgpu_bo_alloc(dev, &request, &bo);
430    if (r)
431       return false;
432    amdgpu_bo_free(bo);
433    return true;
434 }
435 
set_custom_cu_en_mask(struct radeon_info * info)436 static void set_custom_cu_en_mask(struct radeon_info *info)
437 {
438    info->spi_cu_en = ~0;
439 
440    const char *cu_env_var = os_get_option("AMD_CU_MASK");
441    if (!cu_env_var)
442       return;
443 
444    int size = strlen(cu_env_var);
445    char *str = alloca(size + 1);
446    memset(str, 0, size + 1);
447 
448    size = 0;
449 
450    /* Strip whitespace. */
451    for (unsigned src = 0; cu_env_var[src]; src++) {
452       if (cu_env_var[src] != ' ' && cu_env_var[src] != '\t' &&
453           cu_env_var[src] != '\n' && cu_env_var[src] != '\r') {
454          str[size++] = cu_env_var[src];
455       }
456    }
457 
458    /* The following syntax is used, all whitespace is ignored:
459     *   ID = [0-9][0-9]*                         ex. base 10 numbers
460     *   ID_list = (ID | ID-ID)[, (ID | ID-ID)]*  ex. 0,2-4,7
461     *   CU_list = 0x[0-F]* | ID_list             ex. 0x337F OR 0,2-4,7
462     *   AMD_CU_MASK = CU_list
463     *
464     * It's a CU mask within a shader array. It's applied to all shader arrays.
465     */
466    bool is_good_form = true;
467    uint32_t spi_cu_en = 0;
468 
469    if (size > 2 && str[0] == '0' && (str[1] == 'x' || str[1] == 'X')) {
470       str += 2;
471       size -= 2;
472 
473       for (unsigned i = 0; i < size; i++)
474          is_good_form &= isxdigit(str[i]) != 0;
475 
476       if (!is_good_form) {
477          fprintf(stderr, "amd: invalid AMD_CU_MASK: ill-formed hex value\n");
478       } else {
479          spi_cu_en = strtol(str, NULL, 16);
480       }
481    } else {
482       /* Parse ID_list. */
483       long first = 0, last = -1;
484 
485       if (!isdigit(*str)) {
486          is_good_form = false;
487       } else {
488          while (*str) {
489             bool comma = false;
490 
491             if (isdigit(*str)) {
492                first = last = strtol(str, &str, 10);
493             } else if (*str == '-') {
494                str++;
495                /* Parse a digit after a dash. */
496                if (isdigit(*str)) {
497                   last = strtol(str, &str, 10);
498                } else {
499                   fprintf(stderr, "amd: invalid AMD_CU_MASK: expected a digit after -\n");
500                   is_good_form = false;
501                   break;
502                }
503             } else if (*str == ',') {
504                comma = true;
505                str++;
506                if (!isdigit(*str)) {
507                   fprintf(stderr, "amd: invalid AMD_CU_MASK: expected a digit after ,\n");
508                   is_good_form = false;
509                   break;
510                }
511             }
512 
513             if (comma || !*str) {
514                if (first > last) {
515                   fprintf(stderr, "amd: invalid AMD_CU_MASK: range not increasing (%li, %li)\n", first, last);
516                   is_good_form = false;
517                   break;
518                }
519                if (last > 31) {
520                   fprintf(stderr, "amd: invalid AMD_CU_MASK: index too large (%li)\n", last);
521                   is_good_form = false;
522                   break;
523                }
524 
525                spi_cu_en |= BITFIELD_RANGE(first, last - first + 1);
526                last = -1;
527             }
528          }
529       }
530    }
531 
532    /* The mask is parsed. Now assign bits to CUs. */
533    if (is_good_form) {
534       bool error = false;
535 
536       /* Clear bits that have no effect. */
537       spi_cu_en &= BITFIELD_MASK(info->max_good_cu_per_sa);
538 
539       if (!spi_cu_en) {
540          fprintf(stderr, "amd: invalid AMD_CU_MASK: at least 1 CU in each SA must be enabled\n");
541          error = true;
542       }
543 
544       if (info->has_graphics) {
545          uint32_t min_full_cu_mask = BITFIELD_MASK(info->min_good_cu_per_sa);
546 
547          /* The hw ignores all non-compute CU masks if any of them is 0. Disallow that. */
548          if ((spi_cu_en & min_full_cu_mask) == 0) {
549             fprintf(stderr, "amd: invalid AMD_CU_MASK: at least 1 CU from 0x%x per SA must be "
550                             "enabled (SPI limitation)\n", min_full_cu_mask);
551             error = true;
552          }
553 
554          /* We usually disable 1 or 2 CUs for VS and GS, which means at last 1 other CU
555           * must be enabled.
556           */
557          uint32_t cu_mask_ge, unused;
558          ac_compute_late_alloc(info, false, false, false, &unused, &cu_mask_ge);
559          cu_mask_ge &= min_full_cu_mask;
560 
561          if ((spi_cu_en & cu_mask_ge) == 0) {
562             fprintf(stderr, "amd: invalid AMD_CU_MASK: at least 1 CU from 0x%x per SA must be "
563                             "enabled (late alloc constraint for GE)\n", cu_mask_ge);
564             error = true;
565          }
566 
567          if ((min_full_cu_mask & spi_cu_en & ~cu_mask_ge) == 0) {
568             fprintf(stderr, "amd: invalid AMD_CU_MASK: at least 1 CU from 0x%x per SA must be "
569                             "enabled (late alloc constraint for PS)\n",
570                     min_full_cu_mask & ~cu_mask_ge);
571             error = true;
572          }
573       }
574 
575       if (!error) {
576          info->spi_cu_en = spi_cu_en;
577          info->spi_cu_en_has_effect = spi_cu_en & BITFIELD_MASK(info->max_good_cu_per_sa);
578       }
579    }
580 }
581 
ac_query_gpu_info(int fd,void * dev_p,struct radeon_info * info)582 bool ac_query_gpu_info(int fd, void *dev_p, struct radeon_info *info)
583 {
584    struct amdgpu_gpu_info amdinfo;
585    struct drm_amdgpu_info_device device_info = {0};
586    struct amdgpu_buffer_size_alignments alignment_info = {0};
587    uint32_t vce_version = 0, vce_feature = 0, uvd_version = 0, uvd_feature = 0;
588    int r, i, j;
589    amdgpu_device_handle dev = dev_p;
590    drmDevicePtr devinfo;
591 
592    STATIC_ASSERT(AMDGPU_HW_IP_GFX == AMD_IP_GFX);
593    STATIC_ASSERT(AMDGPU_HW_IP_COMPUTE == AMD_IP_COMPUTE);
594    STATIC_ASSERT(AMDGPU_HW_IP_DMA == AMD_IP_SDMA);
595    STATIC_ASSERT(AMDGPU_HW_IP_UVD == AMD_IP_UVD);
596    STATIC_ASSERT(AMDGPU_HW_IP_VCE == AMD_IP_VCE);
597    STATIC_ASSERT(AMDGPU_HW_IP_UVD_ENC == AMD_IP_UVD_ENC);
598    STATIC_ASSERT(AMDGPU_HW_IP_VCN_DEC == AMD_IP_VCN_DEC);
599    STATIC_ASSERT(AMDGPU_HW_IP_VCN_ENC == AMD_IP_VCN_ENC);
600    STATIC_ASSERT(AMDGPU_HW_IP_VCN_JPEG == AMD_IP_VCN_JPEG);
601 
602    /* Get PCI info. */
603    r = drmGetDevice2(fd, 0, &devinfo);
604    if (r) {
605       fprintf(stderr, "amdgpu: drmGetDevice2 failed.\n");
606       return false;
607    }
608    info->pci_domain = devinfo->businfo.pci->domain;
609    info->pci_bus = devinfo->businfo.pci->bus;
610    info->pci_dev = devinfo->businfo.pci->dev;
611    info->pci_func = devinfo->businfo.pci->func;
612    drmFreeDevice(&devinfo);
613 
614    assert(info->drm_major == 3);
615    info->is_amdgpu = true;
616 
617    if (info->drm_minor < 15) {
618       fprintf(stderr, "amdgpu: DRM version is %u.%u.%u, but this driver is "
619                       "only compatible with 3.15.0 (kernel 4.12) or later.\n",
620               info->drm_major, info->drm_minor, info->drm_patchlevel);
621       return false;
622    }
623 
624    /* Query hardware and driver information. */
625    r = amdgpu_query_gpu_info(dev, &amdinfo);
626    if (r) {
627       fprintf(stderr, "amdgpu: amdgpu_query_gpu_info failed.\n");
628       return false;
629    }
630 
631    r = amdgpu_query_info(dev, AMDGPU_INFO_DEV_INFO, sizeof(device_info), &device_info);
632    if (r) {
633       fprintf(stderr, "amdgpu: amdgpu_query_info(dev_info) failed.\n");
634       return false;
635    }
636 
637    r = amdgpu_query_buffer_size_alignment(dev, &alignment_info);
638    if (r) {
639       fprintf(stderr, "amdgpu: amdgpu_query_buffer_size_alignment failed.\n");
640       return false;
641    }
642 
643    for (unsigned ip_type = 0; ip_type < AMD_NUM_IP_TYPES; ip_type++) {
644       struct drm_amdgpu_info_hw_ip ip_info = {0};
645 
646       r = amdgpu_query_hw_ip_info(dev, ip_type, 0, &ip_info);
647       if (r || !ip_info.available_rings)
648          continue;
649 
650       /* Gfx6-8 don't set ip_discovery_version. */
651       if (info->drm_minor >= 48 && ip_info.ip_discovery_version) {
652          info->ip[ip_type].ver_major = (ip_info.ip_discovery_version >> 16) & 0xff;
653          info->ip[ip_type].ver_minor = (ip_info.ip_discovery_version >> 8) & 0xff;
654       } else {
655          info->ip[ip_type].ver_major = ip_info.hw_ip_version_major;
656          info->ip[ip_type].ver_minor = ip_info.hw_ip_version_minor;
657 
658          /* Fix incorrect IP versions reported by the kernel. */
659          if (device_info.family == FAMILY_NV &&
660              (ASICREV_IS(device_info.external_rev, NAVI10) ||
661               ASICREV_IS(device_info.external_rev, NAVI12) ||
662               ASICREV_IS(device_info.external_rev, NAVI14)))
663             info->ip[AMD_IP_GFX].ver_minor = info->ip[AMD_IP_COMPUTE].ver_minor = 1;
664          else if (device_info.family == FAMILY_NV ||
665                   device_info.family == FAMILY_VGH ||
666                   device_info.family == FAMILY_RMB ||
667                   device_info.family == FAMILY_GC_10_3_6 ||
668                   device_info.family == FAMILY_GC_10_3_7)
669             info->ip[AMD_IP_GFX].ver_minor = info->ip[AMD_IP_COMPUTE].ver_minor = 3;
670       }
671       info->ip[ip_type].num_queues = util_bitcount(ip_info.available_rings);
672       info->ib_alignment = MAX3(info->ib_alignment, ip_info.ib_start_alignment,
673                                 ip_info.ib_size_alignment);
674    }
675 
676    /* Only require gfx or compute. */
677    if (!info->ip[AMD_IP_GFX].num_queues && !info->ip[AMD_IP_COMPUTE].num_queues) {
678       fprintf(stderr, "amdgpu: failed to find gfx or compute.\n");
679       return false;
680    }
681 
682    assert(util_is_power_of_two_or_zero(info->ip[AMD_IP_COMPUTE].num_queues));
683    assert(util_is_power_of_two_or_zero(info->ip[AMD_IP_SDMA].num_queues));
684 
685    /* The kernel pads gfx and compute IBs to 256 dwords since:
686     *   66f3b2d527154bd258a57c8815004b5964aa1cf5
687     * Do the same.
688     */
689    info->ib_alignment = MAX2(info->ib_alignment, 1024);
690 
691    r = amdgpu_query_firmware_version(dev, AMDGPU_INFO_FW_GFX_ME, 0, 0, &info->me_fw_version,
692                                      &info->me_fw_feature);
693    if (r) {
694       fprintf(stderr, "amdgpu: amdgpu_query_firmware_version(me) failed.\n");
695       return false;
696    }
697 
698    r = amdgpu_query_firmware_version(dev, AMDGPU_INFO_FW_GFX_MEC, 0, 0, &info->mec_fw_version,
699                                      &info->mec_fw_feature);
700    if (r) {
701       fprintf(stderr, "amdgpu: amdgpu_query_firmware_version(mec) failed.\n");
702       return false;
703    }
704 
705    r = amdgpu_query_firmware_version(dev, AMDGPU_INFO_FW_GFX_PFP, 0, 0, &info->pfp_fw_version,
706                                      &info->pfp_fw_feature);
707    if (r) {
708       fprintf(stderr, "amdgpu: amdgpu_query_firmware_version(pfp) failed.\n");
709       return false;
710    }
711 
712    r = amdgpu_query_firmware_version(dev, AMDGPU_INFO_FW_UVD, 0, 0, &uvd_version, &uvd_feature);
713    if (r) {
714       fprintf(stderr, "amdgpu: amdgpu_query_firmware_version(uvd) failed.\n");
715       return false;
716    }
717 
718    r = amdgpu_query_firmware_version(dev, AMDGPU_INFO_FW_VCE, 0, 0, &vce_version, &vce_feature);
719    if (r) {
720       fprintf(stderr, "amdgpu: amdgpu_query_firmware_version(vce) failed.\n");
721       return false;
722    }
723 
724    r = amdgpu_query_sw_info(dev, amdgpu_sw_info_address32_hi, &info->address32_hi);
725    if (r) {
726       fprintf(stderr, "amdgpu: amdgpu_query_sw_info(address32_hi) failed.\n");
727       return false;
728    }
729 
730    struct drm_amdgpu_memory_info meminfo = {0};
731 
732    r = amdgpu_query_info(dev, AMDGPU_INFO_MEMORY, sizeof(meminfo), &meminfo);
733    if (r) {
734       fprintf(stderr, "amdgpu: amdgpu_query_info(memory) failed.\n");
735       return false;
736    }
737 
738    /* Note: usable_heap_size values can be random and can't be relied on. */
739    info->gart_size_kb = DIV_ROUND_UP(meminfo.gtt.total_heap_size, 1024);
740    info->vram_size_kb = DIV_ROUND_UP(fix_vram_size(meminfo.vram.total_heap_size), 1024);
741    info->vram_vis_size_kb = DIV_ROUND_UP(meminfo.cpu_accessible_vram.total_heap_size, 1024);
742 
743    if (info->drm_minor >= 41) {
744       amdgpu_query_video_caps_info(dev, AMDGPU_INFO_VIDEO_CAPS_DECODE,
745                                    sizeof(info->dec_caps), &(info->dec_caps));
746       amdgpu_query_video_caps_info(dev, AMDGPU_INFO_VIDEO_CAPS_ENCODE,
747                                    sizeof(info->enc_caps), &(info->enc_caps));
748    }
749 
750    /* Add some margin of error, though this shouldn't be needed in theory. */
751    info->all_vram_visible = info->vram_size_kb * 0.9 < info->vram_vis_size_kb;
752 
753    /* Set chip identification. */
754    info->pci_id = device_info.device_id;
755    info->pci_rev_id = device_info.pci_rev;
756    info->vce_harvest_config = device_info.vce_harvest_config;
757 
758 #define identify_chip2(asic, chipname)                                                             \
759    if (ASICREV_IS(device_info.external_rev, asic)) {                                             \
760       info->family = CHIP_##chipname;                                                              \
761       info->name = #chipname;                                                                      \
762    }
763 #define identify_chip(chipname) identify_chip2(chipname, chipname)
764 
765    switch (device_info.family) {
766    case FAMILY_SI:
767       identify_chip(TAHITI);
768       identify_chip(PITCAIRN);
769       identify_chip2(CAPEVERDE, VERDE);
770       identify_chip(OLAND);
771       identify_chip(HAINAN);
772       break;
773    case FAMILY_CI:
774       identify_chip(BONAIRE);
775       identify_chip(HAWAII);
776       break;
777    case FAMILY_KV:
778       identify_chip2(SPECTRE, KAVERI);
779       identify_chip2(SPOOKY, KAVERI);
780       identify_chip2(KALINDI, KABINI);
781       identify_chip2(GODAVARI, KABINI);
782       break;
783    case FAMILY_VI:
784       identify_chip(ICELAND);
785       identify_chip(TONGA);
786       identify_chip(FIJI);
787       identify_chip(POLARIS10);
788       identify_chip(POLARIS11);
789       identify_chip(POLARIS12);
790       identify_chip(VEGAM);
791       break;
792    case FAMILY_CZ:
793       identify_chip(CARRIZO);
794       identify_chip(STONEY);
795       break;
796    case FAMILY_AI:
797       identify_chip(VEGA10);
798       identify_chip(VEGA12);
799       identify_chip(VEGA20);
800       identify_chip(ARCTURUS);
801       identify_chip(ALDEBARAN);
802       break;
803    case FAMILY_RV:
804       identify_chip(RAVEN);
805       identify_chip(RAVEN2);
806       identify_chip(RENOIR);
807       break;
808    case FAMILY_NV:
809       identify_chip(NAVI10);
810       identify_chip(NAVI12);
811       identify_chip(NAVI14);
812       identify_chip(NAVI21);
813       identify_chip(NAVI22);
814       identify_chip(NAVI23);
815       identify_chip(NAVI24);
816       break;
817    case FAMILY_VGH:
818       identify_chip(VANGOGH);
819       break;
820    case FAMILY_RMB:
821       identify_chip(REMBRANDT);
822       break;
823    case FAMILY_GC_10_3_6:
824       identify_chip(GFX1036);
825       break;
826    case FAMILY_GC_10_3_7:
827       identify_chip2(GFX1037, GFX1036);
828       break;
829    case FAMILY_GFX1100:
830       identify_chip(GFX1100);
831       identify_chip(GFX1101);
832       identify_chip(GFX1102);
833       break;
834    case FAMILY_GFX1103:
835       identify_chip(GFX1103);
836       break;
837    }
838 
839    if (!info->name) {
840       fprintf(stderr, "amdgpu: unknown (family_id, chip_external_rev): (%u, %u)\n",
841               device_info.family, device_info.external_rev);
842       return false;
843    }
844 
845    memset(info->lowercase_name, 0, sizeof(info->lowercase_name));
846    for (unsigned i = 0; info->name[i] && i < ARRAY_SIZE(info->lowercase_name) - 1; i++)
847       info->lowercase_name[i] = tolower(info->name[i]);
848 
849    if (info->ip[AMD_IP_GFX].ver_major == 11)
850       info->gfx_level = GFX11;
851    else if (info->ip[AMD_IP_GFX].ver_major == 10 && info->ip[AMD_IP_GFX].ver_minor == 3)
852       info->gfx_level = GFX10_3;
853    else if (info->ip[AMD_IP_GFX].ver_major == 10 && info->ip[AMD_IP_GFX].ver_minor == 1)
854       info->gfx_level = GFX10;
855    else if (info->ip[AMD_IP_GFX].ver_major == 9 || info->ip[AMD_IP_COMPUTE].ver_major == 9)
856       info->gfx_level = GFX9;
857    else if (info->ip[AMD_IP_GFX].ver_major == 8)
858       info->gfx_level = GFX8;
859    else if (info->ip[AMD_IP_GFX].ver_major == 7)
860       info->gfx_level = GFX7;
861    else if (info->ip[AMD_IP_GFX].ver_major == 6)
862       info->gfx_level = GFX6;
863    else {
864       fprintf(stderr, "amdgpu: Unknown gfx version: %u.%u\n",
865               info->ip[AMD_IP_GFX].ver_major, info->ip[AMD_IP_GFX].ver_minor);
866       return false;
867    }
868 
869    info->smart_access_memory = info->all_vram_visible &&
870                                info->gfx_level >= GFX10_3 &&
871                                util_get_cpu_caps()->family >= CPU_AMD_ZEN3 &&
872                                util_get_cpu_caps()->family < CPU_AMD_LAST;
873 
874    info->family_id = device_info.family;
875    info->chip_external_rev = device_info.external_rev;
876    info->chip_rev = device_info.chip_rev;
877    info->marketing_name = amdgpu_get_marketing_name(dev);
878    info->is_pro_graphics = info->marketing_name && (strstr(info->marketing_name, "Pro") ||
879                                                     strstr(info->marketing_name, "PRO") ||
880                                                     strstr(info->marketing_name, "Frontier"));
881 
882    /* Set which chips have dedicated VRAM. */
883    info->has_dedicated_vram = !(device_info.ids_flags & AMDGPU_IDS_FLAGS_FUSION);
884 
885    /* The kernel can split large buffers in VRAM but not in GTT, so large
886     * allocations can fail or cause buffer movement failures in the kernel.
887     */
888    if (info->has_dedicated_vram)
889       info->max_heap_size_kb = info->vram_size_kb;
890    else
891       info->max_heap_size_kb = info->gart_size_kb;
892 
893    info->vram_type = device_info.vram_type;
894    info->memory_bus_width = device_info.vram_bit_width;
895 
896    /* Set which chips have uncached device memory. */
897    info->has_l2_uncached = info->gfx_level >= GFX9;
898 
899    /* Set hardware information. */
900    /* convert the shader/memory clocks from KHz to MHz */
901    info->max_gpu_freq_mhz = device_info.max_engine_clock / 1000;
902    info->memory_freq_mhz_effective = info->memory_freq_mhz = device_info.max_memory_clock / 1000;
903    info->max_tcc_blocks = device_info.num_tcc_blocks;
904    info->max_se = device_info.num_shader_engines;
905    info->max_sa_per_se = device_info.num_shader_arrays_per_engine;
906    info->uvd_fw_version = info->ip[AMD_IP_UVD].num_queues ? uvd_version : 0;
907    info->vce_fw_version = info->ip[AMD_IP_VCE].num_queues ? vce_version : 0;
908 
909    /* Based on MemoryOpsPerClockTable from PAL. */
910    switch (info->vram_type) {
911    case AMDGPU_VRAM_TYPE_DDR2:
912    case AMDGPU_VRAM_TYPE_DDR3:
913    case AMDGPU_VRAM_TYPE_DDR4: /* same for LPDDR4 */
914    case AMDGPU_VRAM_TYPE_HBM: /* same for HBM2 and HBM3 */
915       info->memory_freq_mhz_effective *= 2;
916       break;
917    case AMDGPU_VRAM_TYPE_DDR5: /* same for LPDDR5 */
918    case AMDGPU_VRAM_TYPE_GDDR5:
919       info->memory_freq_mhz_effective *= 4;
920       break;
921    case AMDGPU_VRAM_TYPE_GDDR6:
922       info->memory_freq_mhz_effective *= 16;
923       break;
924    }
925 
926    /* unified ring */
927    info->has_video_hw.vcn_decode
928                   = info->family >= CHIP_GFX1100
929                     ? info->ip[AMD_IP_VCN_UNIFIED].num_queues != 0
930                     : info->ip[AMD_IP_VCN_DEC].num_queues != 0;
931    info->has_userptr = true;
932    info->has_syncobj = has_syncobj(fd);
933    info->has_timeline_syncobj = has_timeline_syncobj(fd);
934    info->has_fence_to_handle = info->has_syncobj && info->drm_minor >= 21;
935    info->has_local_buffers = info->drm_minor >= 20;
936    info->has_bo_metadata = true;
937    info->has_eqaa_surface_allocator = info->gfx_level < GFX11;
938    /* Disable sparse mappings on GFX6 due to VM faults in CP DMA. Enable them once
939     * these faults are mitigated in software.
940     */
941    info->has_sparse_vm_mappings = info->gfx_level >= GFX7;
942    info->has_scheduled_fence_dependency = info->drm_minor >= 28;
943    info->mid_command_buffer_preemption_enabled = device_info.ids_flags & AMDGPU_IDS_FLAGS_PREEMPTION;
944    info->has_tmz_support = has_tmz_support(dev, info, device_info.ids_flags);
945    info->kernel_has_modifiers = has_modifiers(fd);
946    info->has_graphics = info->ip[AMD_IP_GFX].num_queues > 0;
947 
948    info->pa_sc_tile_steering_override = device_info.pa_sc_tile_steering_override;
949    info->max_render_backends = device_info.num_rb_pipes;
950    /* The value returned by the kernel driver was wrong. */
951    if (info->family == CHIP_KAVERI)
952       info->max_render_backends = 2;
953 
954    info->clock_crystal_freq = device_info.gpu_counter_freq;
955    if (!info->clock_crystal_freq) {
956       fprintf(stderr, "amdgpu: clock crystal frequency is 0, timestamps will be wrong\n");
957       info->clock_crystal_freq = 1;
958    }
959    if (info->gfx_level >= GFX10) {
960       info->tcc_cache_line_size = 128;
961 
962       if (info->drm_minor >= 35) {
963          info->num_tcc_blocks = info->max_tcc_blocks - util_bitcount64(device_info.tcc_disabled_mask);
964       } else {
965          /* This is a hack, but it's all we can do without a kernel upgrade. */
966          info->num_tcc_blocks = info->vram_size_kb / (512 * 1024);
967          if (info->num_tcc_blocks > info->max_tcc_blocks)
968             info->num_tcc_blocks /= 2;
969       }
970    } else {
971       if (!info->has_graphics && info->family >= CHIP_ALDEBARAN)
972          info->tcc_cache_line_size = 128;
973       else
974          info->tcc_cache_line_size = 64;
975 
976       info->num_tcc_blocks = info->max_tcc_blocks;
977    }
978 
979    info->tcc_rb_non_coherent = !util_is_power_of_two_or_zero(info->num_tcc_blocks);
980 
981    switch (info->family) {
982    case CHIP_TAHITI:
983    case CHIP_PITCAIRN:
984    case CHIP_OLAND:
985    case CHIP_HAWAII:
986    case CHIP_KABINI:
987    case CHIP_TONGA:
988    case CHIP_STONEY:
989    case CHIP_RAVEN2:
990       info->l2_cache_size = info->num_tcc_blocks * 64 * 1024;
991       break;
992    case CHIP_VERDE:
993    case CHIP_HAINAN:
994    case CHIP_BONAIRE:
995    case CHIP_KAVERI:
996    case CHIP_ICELAND:
997    case CHIP_CARRIZO:
998    case CHIP_FIJI:
999    case CHIP_POLARIS12:
1000    case CHIP_VEGAM:
1001       info->l2_cache_size = info->num_tcc_blocks * 128 * 1024;
1002       break;
1003    default:
1004       info->l2_cache_size = info->num_tcc_blocks * 256 * 1024;
1005       break;
1006    case CHIP_REMBRANDT:
1007       info->l2_cache_size = info->num_tcc_blocks * 512 * 1024;
1008       break;
1009    }
1010 
1011    info->l1_cache_size = 16384;
1012 
1013    info->mc_arb_ramcfg = amdinfo.mc_arb_ramcfg;
1014    info->gb_addr_config = amdinfo.gb_addr_cfg;
1015    if (info->gfx_level >= GFX9) {
1016       info->num_tile_pipes = 1 << G_0098F8_NUM_PIPES(info->gb_addr_config);
1017       info->pipe_interleave_bytes = 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(info->gb_addr_config);
1018    } else {
1019       info->num_tile_pipes = cik_get_num_tile_pipes(&amdinfo);
1020       info->pipe_interleave_bytes = 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX6(info->gb_addr_config);
1021    }
1022    info->r600_has_virtual_memory = true;
1023 
1024    /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
1025     * 16KB makes some SIMDs unoccupied).
1026     *
1027     * LDS is 128KB in WGP mode and 64KB in CU mode. Assume the WGP mode is used.
1028     */
1029    info->lds_size_per_workgroup = info->gfx_level >= GFX10 ? 128 * 1024 : 64 * 1024;
1030    /* lds_encode_granularity is the block size used for encoding registers.
1031     * lds_alloc_granularity is what the hardware will align the LDS size to.
1032     */
1033    info->lds_encode_granularity = info->gfx_level >= GFX7 ? 128 * 4 : 64 * 4;
1034    info->lds_alloc_granularity = info->gfx_level >= GFX10_3 ? 256 * 4 : info->lds_encode_granularity;
1035 
1036    /* This is "align_mask" copied from the kernel, maximums of all IP versions. */
1037    info->ib_pad_dw_mask[AMD_IP_GFX] = 0xff;
1038    info->ib_pad_dw_mask[AMD_IP_COMPUTE] = 0xff;
1039    info->ib_pad_dw_mask[AMD_IP_SDMA] = 0xf;
1040    info->ib_pad_dw_mask[AMD_IP_UVD] = 0xf;
1041    info->ib_pad_dw_mask[AMD_IP_VCE] = 0x3f;
1042    info->ib_pad_dw_mask[AMD_IP_UVD_ENC] = 0x3f;
1043    info->ib_pad_dw_mask[AMD_IP_VCN_DEC] = 0xf;
1044    info->ib_pad_dw_mask[AMD_IP_VCN_ENC] = 0x3f;
1045    info->ib_pad_dw_mask[AMD_IP_VCN_JPEG] = 0xf;
1046 
1047    /* The mere presence of CLEAR_STATE in the IB causes random GPU hangs
1048     * on GFX6. Some CLEAR_STATE cause asic hang on radeon kernel, etc.
1049     * SPI_VS_OUT_CONFIG. So only enable GFX7 CLEAR_STATE on amdgpu kernel.
1050     */
1051    info->has_clear_state = info->gfx_level >= GFX7;
1052 
1053    info->has_distributed_tess =
1054       info->gfx_level >= GFX10 || (info->gfx_level >= GFX8 && info->max_se >= 2);
1055 
1056    info->has_dcc_constant_encode =
1057       info->family == CHIP_RAVEN2 || info->family == CHIP_RENOIR || info->gfx_level >= GFX10;
1058 
1059    info->has_rbplus = info->family == CHIP_STONEY || info->gfx_level >= GFX9;
1060 
1061    /* Some chips have RB+ registers, but don't support RB+. Those must
1062     * always disable it.
1063     */
1064    info->rbplus_allowed =
1065       info->has_rbplus &&
1066       (info->family == CHIP_STONEY || info->family == CHIP_VEGA12 || info->family == CHIP_RAVEN ||
1067        info->family == CHIP_RAVEN2 || info->family == CHIP_RENOIR || info->gfx_level >= GFX10_3);
1068 
1069    info->has_out_of_order_rast =
1070       info->gfx_level >= GFX8 && info->gfx_level <= GFX9 && info->max_se >= 2;
1071 
1072    /* Whether chips support double rate packed math instructions. */
1073    info->has_packed_math_16bit = info->gfx_level >= GFX9;
1074 
1075    /* Whether chips support dot product instructions. A subset of these support a smaller
1076     * instruction encoding which accumulates with the destination.
1077     */
1078    info->has_accelerated_dot_product =
1079       info->family == CHIP_ARCTURUS || info->family == CHIP_ALDEBARAN ||
1080       info->family == CHIP_VEGA20 || info->family >= CHIP_NAVI12;
1081 
1082    /* TODO: Figure out how to use LOAD_CONTEXT_REG on GFX6-GFX7. */
1083    info->has_load_ctx_reg_pkt =
1084       info->gfx_level >= GFX9 || (info->gfx_level >= GFX8 && info->me_fw_feature >= 41);
1085 
1086    info->cpdma_prefetch_writes_memory = info->gfx_level <= GFX8;
1087 
1088    info->has_gfx9_scissor_bug = info->family == CHIP_VEGA10 || info->family == CHIP_RAVEN;
1089 
1090    info->has_tc_compat_zrange_bug = info->gfx_level >= GFX8 && info->gfx_level <= GFX9;
1091 
1092    info->has_msaa_sample_loc_bug =
1093       (info->family >= CHIP_POLARIS10 && info->family <= CHIP_POLARIS12) ||
1094       info->family == CHIP_VEGA10 || info->family == CHIP_RAVEN;
1095 
1096    info->has_ls_vgpr_init_bug = info->family == CHIP_VEGA10 || info->family == CHIP_RAVEN;
1097 
1098    /* Drawing from 0-sized index buffers causes hangs on gfx10. */
1099    info->has_zero_index_buffer_bug = info->gfx_level == GFX10;
1100 
1101    /* Whether chips are affected by the image load/sample/gather hw bug when
1102     * DCC is enabled (ie. WRITE_COMPRESS_ENABLE should be 0).
1103     */
1104    info->has_image_load_dcc_bug = info->family == CHIP_NAVI23 ||
1105                                   info->family == CHIP_VANGOGH ||
1106                                   info->family == CHIP_REMBRANDT;
1107 
1108    /* DB has a bug when ITERATE_256 is set to 1 that can cause a hang. The
1109     * workaround is to set DECOMPRESS_ON_Z_PLANES to 2 for 4X MSAA D/S images.
1110     */
1111    info->has_two_planes_iterate256_bug = info->gfx_level == GFX10;
1112 
1113    /* GFX10+Navi21: NGG->legacy transitions require VGT_FLUSH. */
1114    info->has_vgt_flush_ngg_legacy_bug = info->gfx_level == GFX10 ||
1115                                         info->family == CHIP_NAVI21;
1116 
1117    /* HW bug workaround when CS threadgroups > 256 threads and async compute
1118     * isn't used, i.e. only one compute job can run at a time.  If async
1119     * compute is possible, the threadgroup size must be limited to 256 threads
1120     * on all queues to avoid the bug.
1121     * Only GFX6 and certain GFX7 chips are affected.
1122     *
1123     * FIXME: RADV doesn't limit the number of threads for async compute.
1124     */
1125    info->has_cs_regalloc_hang_bug = info->gfx_level == GFX6 ||
1126                                     info->family == CHIP_BONAIRE ||
1127                                     info->family == CHIP_KABINI;
1128 
1129    /* Support for GFX10.3 was added with F32_ME_FEATURE_VERSION_31 but the
1130     * feature version wasn't bumped.
1131     */
1132    info->has_32bit_predication = (info->gfx_level >= GFX10 &&
1133                                   info->me_fw_feature >= 32) ||
1134                                  (info->gfx_level == GFX9 &&
1135                                   info->me_fw_feature >= 52);
1136 
1137    info->has_export_conflict_bug = info->gfx_level == GFX11;
1138 
1139    /* Get the number of good compute units. */
1140    info->num_cu = 0;
1141    for (i = 0; i < info->max_se; i++) {
1142       for (j = 0; j < info->max_sa_per_se; j++) {
1143          if (info->gfx_level >= GFX11) {
1144             assert(info->max_sa_per_se <= 2);
1145             info->cu_mask[i][j] = device_info.cu_bitmap[i % 4][(i / 4) * 2 + j];
1146          } else if (info->family == CHIP_ARCTURUS) {
1147             /* The CU bitmap in amd gpu info structure is
1148              * 4x4 size array, and it's usually suitable for Vega
1149              * ASICs which has 4*2 SE/SA layout.
1150              * But for Arcturus, SE/SA layout is changed to 8*1.
1151              * To mostly reduce the impact, we make it compatible
1152              * with current bitmap array as below:
1153              *    SE4 --> cu_bitmap[0][1]
1154              *    SE5 --> cu_bitmap[1][1]
1155              *    SE6 --> cu_bitmap[2][1]
1156              *    SE7 --> cu_bitmap[3][1]
1157              */
1158             assert(info->max_sa_per_se == 1);
1159             info->cu_mask[i][0] = device_info.cu_bitmap[i % 4][i / 4];
1160          } else {
1161             info->cu_mask[i][j] = device_info.cu_bitmap[i][j];
1162          }
1163          info->num_cu += util_bitcount(info->cu_mask[i][j]);
1164       }
1165    }
1166 
1167    /* Derive the number of enabled SEs from the CU mask. */
1168    if (info->gfx_level >= GFX10_3 && info->max_se > 1) {
1169       info->num_se = 0;
1170 
1171       for (unsigned se = 0; se < info->max_se; se++) {
1172          for (unsigned sa = 0; sa < info->max_sa_per_se; sa++) {
1173             if (info->cu_mask[se][sa]) {
1174                info->num_se++;
1175                break;
1176             }
1177          }
1178       }
1179    } else {
1180       /* GFX10 and older always enable all SEs because they don't support SE harvesting. */
1181       info->num_se = info->max_se;
1182    }
1183 
1184    /* On GFX10, only whole WGPs (in units of 2 CUs) can be disabled,
1185     * and max - min <= 2.
1186     */
1187    unsigned cu_group = info->gfx_level >= GFX10 ? 2 : 1;
1188    info->max_good_cu_per_sa =
1189       DIV_ROUND_UP(info->num_cu, (info->num_se * info->max_sa_per_se * cu_group)) *
1190       cu_group;
1191    info->min_good_cu_per_sa =
1192       (info->num_cu / (info->num_se * info->max_sa_per_se * cu_group)) * cu_group;
1193 
1194    memcpy(info->si_tile_mode_array, amdinfo.gb_tile_mode, sizeof(amdinfo.gb_tile_mode));
1195    info->enabled_rb_mask = amdinfo.enabled_rb_pipes_mask;
1196 
1197    memcpy(info->cik_macrotile_mode_array, amdinfo.gb_macro_tile_mode,
1198           sizeof(amdinfo.gb_macro_tile_mode));
1199 
1200    info->pte_fragment_size = alignment_info.size_local;
1201    info->gart_page_size = alignment_info.size_remote;
1202 
1203    if (info->gfx_level == GFX6)
1204       info->gfx_ib_pad_with_type2 = true;
1205 
1206    /* GFX10 and maybe GFX9 need this alignment for cache coherency. */
1207    if (info->gfx_level >= GFX9)
1208       info->ib_alignment = MAX2(info->ib_alignment, info->tcc_cache_line_size);
1209 
1210    if ((info->drm_minor >= 31 && (info->family == CHIP_RAVEN || info->family == CHIP_RAVEN2 ||
1211                                   info->family == CHIP_RENOIR)) ||
1212        info->gfx_level >= GFX10_3) {
1213       /* GFX10+ requires retiling in all cases. */
1214       if (info->max_render_backends == 1 && info->gfx_level == GFX9)
1215          info->use_display_dcc_unaligned = true;
1216       else
1217          info->use_display_dcc_with_retile_blit = true;
1218    }
1219 
1220    info->has_stable_pstate = info->drm_minor >= 45;
1221 
1222    if (info->gfx_level >= GFX11) {
1223       info->pc_lines = 1024;
1224       info->pbb_max_alloc_count = 255; /* minimum is 2, maximum is 256 */
1225    } else if (info->gfx_level >= GFX9 && info->has_graphics) {
1226       unsigned pc_lines = 0;
1227 
1228       switch (info->family) {
1229       case CHIP_VEGA10:
1230       case CHIP_VEGA12:
1231       case CHIP_VEGA20:
1232          pc_lines = 2048;
1233          break;
1234       case CHIP_RAVEN:
1235       case CHIP_RAVEN2:
1236       case CHIP_RENOIR:
1237       case CHIP_NAVI10:
1238       case CHIP_NAVI12:
1239       case CHIP_NAVI21:
1240       case CHIP_NAVI22:
1241       case CHIP_NAVI23:
1242          pc_lines = 1024;
1243          break;
1244       case CHIP_NAVI14:
1245       case CHIP_NAVI24:
1246          pc_lines = 512;
1247          break;
1248       case CHIP_VANGOGH:
1249       case CHIP_REMBRANDT:
1250       case CHIP_GFX1036:
1251          pc_lines = 256;
1252          break;
1253       default:
1254          assert(0);
1255       }
1256 
1257       info->pc_lines = pc_lines;
1258 
1259       if (info->gfx_level >= GFX10) {
1260          info->pbb_max_alloc_count = pc_lines / 3;
1261       } else {
1262          info->pbb_max_alloc_count = MIN2(128, pc_lines / (4 * info->max_se));
1263       }
1264    }
1265 
1266    if (info->gfx_level >= GFX10_3)
1267       info->max_wave64_per_simd = 16;
1268    else if (info->gfx_level == GFX10)
1269       info->max_wave64_per_simd = 20;
1270    else if (info->family >= CHIP_POLARIS10 && info->family <= CHIP_VEGAM)
1271       info->max_wave64_per_simd = 8;
1272    else
1273       info->max_wave64_per_simd = 10;
1274 
1275    if (info->gfx_level >= GFX10) {
1276       info->num_physical_sgprs_per_simd = 128 * info->max_wave64_per_simd;
1277       info->min_sgpr_alloc = 128;
1278       info->sgpr_alloc_granularity = 128;
1279    } else if (info->gfx_level >= GFX8) {
1280       info->num_physical_sgprs_per_simd = 800;
1281       info->min_sgpr_alloc = 16;
1282       info->sgpr_alloc_granularity = 16;
1283    } else {
1284       info->num_physical_sgprs_per_simd = 512;
1285       info->min_sgpr_alloc = 8;
1286       info->sgpr_alloc_granularity = 8;
1287    }
1288 
1289    info->has_3d_cube_border_color_mipmap = info->has_graphics || info->family == CHIP_ARCTURUS;
1290    info->never_stop_sq_perf_counters = info->gfx_level == GFX10 ||
1291                                        info->gfx_level == GFX10_3;
1292    info->never_send_perfcounter_stop = info->gfx_level == GFX11;
1293    info->has_sqtt_rb_harvest_bug = (info->family == CHIP_NAVI23 ||
1294                                     info->family == CHIP_NAVI24 ||
1295                                     info->family == CHIP_REMBRANDT ||
1296                                     info->family == CHIP_VANGOGH) &&
1297                                    util_bitcount(info->enabled_rb_mask) !=
1298                                    info->max_render_backends;
1299 
1300    /* On GFX10.3, the polarity of AUTO_FLUSH_MODE is inverted. */
1301    info->has_sqtt_auto_flush_mode_bug = info->gfx_level == GFX10_3;
1302 
1303    info->max_sgpr_alloc = info->family == CHIP_TONGA || info->family == CHIP_ICELAND ? 96 : 104;
1304 
1305    if (!info->has_graphics && info->family >= CHIP_ALDEBARAN) {
1306       info->min_wave64_vgpr_alloc = 8;
1307       info->max_vgpr_alloc = 512;
1308       info->wave64_vgpr_alloc_granularity = 8;
1309    } else {
1310       info->min_wave64_vgpr_alloc = 4;
1311       info->max_vgpr_alloc = 256;
1312       info->wave64_vgpr_alloc_granularity = 4;
1313    }
1314 
1315    info->num_physical_wave64_vgprs_per_simd = info->gfx_level >= GFX10 ? 512 : 256;
1316    info->num_simd_per_compute_unit = info->gfx_level >= GFX10 ? 2 : 4;
1317 
1318    /* BIG_PAGE is supported since gfx10.3 and requires VRAM. VRAM is only guaranteed
1319     * with AMDGPU_GEM_CREATE_DISCARDABLE. DISCARDABLE was added in DRM 3.47.0.
1320     */
1321    info->discardable_allows_big_page = info->gfx_level >= GFX10_3 &&
1322                                        info->has_dedicated_vram &&
1323                                        info->drm_minor >= 47;
1324 
1325    /* The maximum number of scratch waves. The number is only a function of the number of CUs.
1326     * It should be large enough to hold at least 1 threadgroup. Use the minimum per-SA CU count.
1327     *
1328     * We can decrease the number to make it fit into the infinity cache.
1329     */
1330    const unsigned max_waves_per_tg = 32; /* 1024 threads in Wave32 */
1331    info->max_scratch_waves = MAX2(32 * info->min_good_cu_per_sa * info->max_sa_per_se * info->num_se,
1332                                   max_waves_per_tg);
1333    info->num_rb = util_bitcount(info->enabled_rb_mask);
1334    info->max_gflops = info->num_cu * 128 * info->max_gpu_freq_mhz / 1000;
1335    info->memory_bandwidth_gbps = DIV_ROUND_UP(info->memory_freq_mhz_effective * info->memory_bus_width / 8, 1000);
1336 
1337    if (info->gfx_level >= GFX10_3 && info->has_dedicated_vram) {
1338       info->l3_cache_size_mb = info->num_tcc_blocks *
1339                                (info->family == CHIP_NAVI21 ||
1340                                 info->family == CHIP_NAVI22 ? 8 : 4);
1341    }
1342 
1343    set_custom_cu_en_mask(info);
1344 
1345    const char *ib_filename = debug_get_option("AMD_PARSE_IB", NULL);
1346    if (ib_filename) {
1347       FILE *f = fopen(ib_filename, "r");
1348       if (f) {
1349          fseek(f, 0, SEEK_END);
1350          size_t size = ftell(f);
1351          uint32_t *ib = (uint32_t *)malloc(size);
1352          fseek(f, 0, SEEK_SET);
1353          size_t n_read = fread(ib, 1, size, f);
1354          fclose(f);
1355 
1356          if (n_read != size) {
1357             fprintf(stderr, "failed to read %zu bytes from '%s'\n", size, ib_filename);
1358             exit(1);
1359          }
1360 
1361          ac_parse_ib(stdout, ib, size / 4, NULL, 0, "IB", info->gfx_level, NULL, NULL);
1362          free(ib);
1363          exit(0);
1364       }
1365    }
1366    return true;
1367 }
1368 
ac_compute_driver_uuid(char * uuid,size_t size)1369 void ac_compute_driver_uuid(char *uuid, size_t size)
1370 {
1371    char amd_uuid[] = "AMD-MESA-DRV";
1372 
1373    assert(size >= sizeof(amd_uuid));
1374 
1375    memset(uuid, 0, size);
1376    strncpy(uuid, amd_uuid, size);
1377 }
1378 
ac_compute_device_uuid(struct radeon_info * info,char * uuid,size_t size)1379 void ac_compute_device_uuid(struct radeon_info *info, char *uuid, size_t size)
1380 {
1381    uint32_t *uint_uuid = (uint32_t *)uuid;
1382 
1383    assert(size >= sizeof(uint32_t) * 4);
1384 
1385    /**
1386     * Use the device info directly instead of using a sha1. GL/VK UUIDs
1387     * are 16 byte vs 20 byte for sha1, and the truncation that would be
1388     * required would get rid of part of the little entropy we have.
1389     * */
1390    memset(uuid, 0, size);
1391    uint_uuid[0] = info->pci_domain;
1392    uint_uuid[1] = info->pci_bus;
1393    uint_uuid[2] = info->pci_dev;
1394    uint_uuid[3] = info->pci_func;
1395 }
1396 
ac_print_gpu_info(struct radeon_info * info,FILE * f)1397 void ac_print_gpu_info(struct radeon_info *info, FILE *f)
1398 {
1399    fprintf(f, "Device info:\n");
1400    fprintf(f, "    name = %s\n", info->name);
1401    fprintf(f, "    marketing_name = %s\n", info->marketing_name);
1402    fprintf(f, "    num_se = %i\n", info->num_se);
1403    fprintf(f, "    num_rb = %i\n", info->num_rb);
1404    fprintf(f, "    num_cu = %i\n", info->num_cu);
1405    fprintf(f, "    max_gpu_freq = %i MHz\n", info->max_gpu_freq_mhz);
1406    fprintf(f, "    max_gflops = %u GFLOPS\n", info->max_gflops);
1407 
1408    if (info->gfx_level >= GFX10) {
1409       fprintf(f, "    l0_cache_size = %i KB\n", DIV_ROUND_UP(info->l1_cache_size, 1024));
1410       fprintf(f, "    l1_cache_size = %i KB\n", 128);
1411    } else {
1412       fprintf(f, "    l1_cache_size = %i KB\n", DIV_ROUND_UP(info->l1_cache_size, 1024));
1413    }
1414 
1415    fprintf(f, "    l2_cache_size = %i KB\n", DIV_ROUND_UP(info->l2_cache_size, 1024));
1416 
1417    if (info->l3_cache_size_mb)
1418       fprintf(f, "    l3_cache_size = %i MB\n", info->l3_cache_size_mb);
1419 
1420    fprintf(f, "    memory_channels = %u (TCC blocks)\n", info->num_tcc_blocks);
1421    fprintf(f, "    memory_size = %u GB (%u MB)\n",
1422            DIV_ROUND_UP(info->vram_size_kb, (1024 * 1024)),
1423            DIV_ROUND_UP(info->vram_size_kb, 1024));
1424    fprintf(f, "    memory_freq = %u GHz\n", DIV_ROUND_UP(info->memory_freq_mhz_effective, 1000));
1425    fprintf(f, "    memory_bus_width = %u bits\n", info->memory_bus_width);
1426    fprintf(f, "    memory_bandwidth = %u GB/s\n", info->memory_bandwidth_gbps);
1427    fprintf(f, "    clock_crystal_freq = %i KHz\n", info->clock_crystal_freq);
1428 
1429    const char *ip_string[] = {
1430       [AMD_IP_GFX] = "GFX",
1431       [AMD_IP_COMPUTE] = "COMP",
1432       [AMD_IP_SDMA] = "SDMA",
1433       [AMD_IP_UVD] = "UVD",
1434       [AMD_IP_VCE] = "VCE",
1435       [AMD_IP_UVD_ENC] = "UVD_ENC",
1436       [AMD_IP_VCN_DEC] = "VCN_DEC",
1437       [AMD_IP_VCN_ENC] = info->family >= CHIP_GFX1100 ? "VCN" : "VCN_ENC",
1438       [AMD_IP_VCN_JPEG] = "VCN_JPG",
1439    };
1440 
1441    for (unsigned i = 0; i < AMD_NUM_IP_TYPES; i++) {
1442       if (info->ip[i].num_queues) {
1443          fprintf(f, "    IP %-7s %2u.%u \tqueues:%u\n", ip_string[i],
1444                  info->ip[i].ver_major, info->ip[i].ver_minor, info->ip[i].num_queues);
1445       }
1446    }
1447 
1448    fprintf(f, "Identification:\n");
1449    fprintf(f, "    pci (domain:bus:dev.func): %04x:%02x:%02x.%x\n", info->pci_domain, info->pci_bus,
1450            info->pci_dev, info->pci_func);
1451    fprintf(f, "    pci_id = 0x%x\n", info->pci_id);
1452    fprintf(f, "    pci_rev_id = 0x%x\n", info->pci_rev_id);
1453    fprintf(f, "    family = %i\n", info->family);
1454    fprintf(f, "    gfx_level = %i\n", info->gfx_level);
1455    fprintf(f, "    family_id = %i\n", info->family_id);
1456    fprintf(f, "    chip_external_rev = %i\n", info->chip_external_rev);
1457    fprintf(f, "    chip_rev = %i\n", info->chip_rev);
1458 
1459    fprintf(f, "Flags:\n");
1460    fprintf(f, "    is_pro_graphics = %u\n", info->is_pro_graphics);
1461    fprintf(f, "    has_graphics = %i\n", info->has_graphics);
1462    fprintf(f, "    has_clear_state = %u\n", info->has_clear_state);
1463    fprintf(f, "    has_distributed_tess = %u\n", info->has_distributed_tess);
1464    fprintf(f, "    has_dcc_constant_encode = %u\n", info->has_dcc_constant_encode);
1465    fprintf(f, "    has_rbplus = %u\n", info->has_rbplus);
1466    fprintf(f, "    rbplus_allowed = %u\n", info->rbplus_allowed);
1467    fprintf(f, "    has_load_ctx_reg_pkt = %u\n", info->has_load_ctx_reg_pkt);
1468    fprintf(f, "    has_out_of_order_rast = %u\n", info->has_out_of_order_rast);
1469    fprintf(f, "    cpdma_prefetch_writes_memory = %u\n", info->cpdma_prefetch_writes_memory);
1470    fprintf(f, "    has_gfx9_scissor_bug = %i\n", info->has_gfx9_scissor_bug);
1471    fprintf(f, "    has_tc_compat_zrange_bug = %i\n", info->has_tc_compat_zrange_bug);
1472    fprintf(f, "    has_msaa_sample_loc_bug = %i\n", info->has_msaa_sample_loc_bug);
1473    fprintf(f, "    has_ls_vgpr_init_bug = %i\n", info->has_ls_vgpr_init_bug);
1474    fprintf(f, "    has_32bit_predication = %i\n", info->has_32bit_predication);
1475    fprintf(f, "    has_3d_cube_border_color_mipmap = %i\n", info->has_3d_cube_border_color_mipmap);
1476    fprintf(f, "    never_stop_sq_perf_counters = %i\n", info->never_stop_sq_perf_counters);
1477    fprintf(f, "    has_sqtt_rb_harvest_bug = %i\n", info->has_sqtt_rb_harvest_bug);
1478    fprintf(f, "    has_sqtt_auto_flush_mode_bug = %i\n", info->has_sqtt_auto_flush_mode_bug);
1479    fprintf(f, "    never_send_perfcounter_stop = %i\n", info->never_send_perfcounter_stop);
1480    fprintf(f, "    discardable_allows_big_page = %i\n", info->discardable_allows_big_page);
1481 
1482    fprintf(f, "Display features:\n");
1483    fprintf(f, "    use_display_dcc_unaligned = %u\n", info->use_display_dcc_unaligned);
1484    fprintf(f, "    use_display_dcc_with_retile_blit = %u\n", info->use_display_dcc_with_retile_blit);
1485 
1486    fprintf(f, "Memory info:\n");
1487    fprintf(f, "    pte_fragment_size = %u\n", info->pte_fragment_size);
1488    fprintf(f, "    gart_page_size = %u\n", info->gart_page_size);
1489    fprintf(f, "    gart_size = %i MB\n", (int)DIV_ROUND_UP(info->gart_size_kb, 1024));
1490    fprintf(f, "    vram_size = %i MB\n", (int)DIV_ROUND_UP(info->vram_size_kb, 1024));
1491    fprintf(f, "    vram_vis_size = %i MB\n", (int)DIV_ROUND_UP(info->vram_vis_size_kb, 1024));
1492    fprintf(f, "    vram_type = %i\n", info->vram_type);
1493    fprintf(f, "    max_heap_size_kb = %i MB\n", (int)DIV_ROUND_UP(info->max_heap_size_kb, 1024));
1494    fprintf(f, "    min_alloc_size = %u\n", info->min_alloc_size);
1495    fprintf(f, "    address32_hi = 0x%x\n", info->address32_hi);
1496    fprintf(f, "    has_dedicated_vram = %u\n", info->has_dedicated_vram);
1497    fprintf(f, "    all_vram_visible = %u\n", info->all_vram_visible);
1498    fprintf(f, "    smart_access_memory = %u\n", info->smart_access_memory);
1499    fprintf(f, "    max_tcc_blocks = %i\n", info->max_tcc_blocks);
1500    fprintf(f, "    tcc_cache_line_size = %u\n", info->tcc_cache_line_size);
1501    fprintf(f, "    tcc_rb_non_coherent = %u\n", info->tcc_rb_non_coherent);
1502    fprintf(f, "    pc_lines = %u\n", info->pc_lines);
1503    fprintf(f, "    lds_size_per_workgroup = %u\n", info->lds_size_per_workgroup);
1504    fprintf(f, "    lds_alloc_granularity = %i\n", info->lds_alloc_granularity);
1505    fprintf(f, "    lds_encode_granularity = %i\n", info->lds_encode_granularity);
1506    fprintf(f, "    max_memory_clock = %i MHz\n", info->memory_freq_mhz);
1507 
1508    fprintf(f, "CP info:\n");
1509    fprintf(f, "    gfx_ib_pad_with_type2 = %i\n", info->gfx_ib_pad_with_type2);
1510    fprintf(f, "    ib_alignment = %u\n", info->ib_alignment);
1511    fprintf(f, "    me_fw_version = %i\n", info->me_fw_version);
1512    fprintf(f, "    me_fw_feature = %i\n", info->me_fw_feature);
1513    fprintf(f, "    mec_fw_version = %i\n", info->mec_fw_version);
1514    fprintf(f, "    mec_fw_feature = %i\n", info->mec_fw_feature);
1515    fprintf(f, "    pfp_fw_version = %i\n", info->pfp_fw_version);
1516    fprintf(f, "    pfp_fw_feature = %i\n", info->pfp_fw_feature);
1517 
1518    fprintf(f, "Multimedia info:\n");
1519    fprintf(f, "    vce_encode = %u\n", info->ip[AMD_IP_VCE].num_queues);
1520 
1521    if (info->family >= CHIP_GFX1100)
1522       fprintf(f, "    vcn_unified = %u\n", info->has_video_hw.vcn_decode);
1523    else {
1524       fprintf(f, "    vcn_decode = %u\n", info->has_video_hw.vcn_decode);
1525       fprintf(f, "    vcn_encode = %u\n", info->ip[AMD_IP_VCN_ENC].num_queues);
1526    }
1527 
1528    fprintf(f, "    uvd_fw_version = %u\n", info->uvd_fw_version);
1529    fprintf(f, "    vce_fw_version = %u\n", info->vce_fw_version);
1530    fprintf(f, "    vce_harvest_config = %i\n", info->vce_harvest_config);
1531 
1532    fprintf(f, "Kernel & winsys capabilities:\n");
1533    fprintf(f, "    drm = %i.%i.%i\n", info->drm_major, info->drm_minor, info->drm_patchlevel);
1534    fprintf(f, "    has_userptr = %i\n", info->has_userptr);
1535    fprintf(f, "    has_syncobj = %u\n", info->has_syncobj);
1536    fprintf(f, "    has_timeline_syncobj = %u\n", info->has_timeline_syncobj);
1537    fprintf(f, "    has_fence_to_handle = %u\n", info->has_fence_to_handle);
1538    fprintf(f, "    has_local_buffers = %u\n", info->has_local_buffers);
1539    fprintf(f, "    has_bo_metadata = %u\n", info->has_bo_metadata);
1540    fprintf(f, "    has_eqaa_surface_allocator = %u\n", info->has_eqaa_surface_allocator);
1541    fprintf(f, "    has_sparse_vm_mappings = %u\n", info->has_sparse_vm_mappings);
1542    fprintf(f, "    has_stable_pstate = %u\n", info->has_stable_pstate);
1543    fprintf(f, "    has_scheduled_fence_dependency = %u\n", info->has_scheduled_fence_dependency);
1544    fprintf(f, "    mid_command_buffer_preemption_enabled = %u\n",
1545            info->mid_command_buffer_preemption_enabled);
1546    fprintf(f, "    has_tmz_support = %u\n", info->has_tmz_support);
1547 
1548    fprintf(f, "Shader core info:\n");
1549    for (unsigned i = 0; i < info->max_se; i++) {
1550       for (unsigned j = 0; j < info->max_sa_per_se; j++) {
1551          fprintf(f, "    cu_mask[SE%u][SA%u] = 0x%x \t(%u)\tCU_EN = 0x%x\n", i, j,
1552                  info->cu_mask[i][j], util_bitcount(info->cu_mask[i][j]),
1553                  info->spi_cu_en & BITFIELD_MASK(util_bitcount(info->cu_mask[i][j])));
1554       }
1555    }
1556    fprintf(f, "    spi_cu_en_has_effect = %i\n", info->spi_cu_en_has_effect);
1557    fprintf(f, "    max_good_cu_per_sa = %i\n", info->max_good_cu_per_sa);
1558    fprintf(f, "    min_good_cu_per_sa = %i\n", info->min_good_cu_per_sa);
1559    fprintf(f, "    max_se = %i\n", info->max_se);
1560    fprintf(f, "    max_sa_per_se = %i\n", info->max_sa_per_se);
1561    fprintf(f, "    max_wave64_per_simd = %i\n", info->max_wave64_per_simd);
1562    fprintf(f, "    num_physical_sgprs_per_simd = %i\n", info->num_physical_sgprs_per_simd);
1563    fprintf(f, "    num_physical_wave64_vgprs_per_simd = %i\n",
1564            info->num_physical_wave64_vgprs_per_simd);
1565    fprintf(f, "    num_simd_per_compute_unit = %i\n", info->num_simd_per_compute_unit);
1566    fprintf(f, "    min_sgpr_alloc = %i\n", info->min_sgpr_alloc);
1567    fprintf(f, "    max_sgpr_alloc = %i\n", info->max_sgpr_alloc);
1568    fprintf(f, "    sgpr_alloc_granularity = %i\n", info->sgpr_alloc_granularity);
1569    fprintf(f, "    min_wave64_vgpr_alloc = %i\n", info->min_wave64_vgpr_alloc);
1570    fprintf(f, "    max_vgpr_alloc = %i\n", info->max_vgpr_alloc);
1571    fprintf(f, "    wave64_vgpr_alloc_granularity = %i\n", info->wave64_vgpr_alloc_granularity);
1572    fprintf(f, "    max_scratch_waves = %i\n", info->max_scratch_waves);
1573 
1574    fprintf(f, "Render backend info:\n");
1575    fprintf(f, "    pa_sc_tile_steering_override = 0x%x\n", info->pa_sc_tile_steering_override);
1576    fprintf(f, "    max_render_backends = %i\n", info->max_render_backends);
1577    fprintf(f, "    num_tile_pipes = %i\n", info->num_tile_pipes);
1578    fprintf(f, "    pipe_interleave_bytes = %i\n", info->pipe_interleave_bytes);
1579    fprintf(f, "    enabled_rb_mask = 0x%x\n", info->enabled_rb_mask);
1580    fprintf(f, "    max_alignment = %u\n", (unsigned)info->max_alignment);
1581    fprintf(f, "    pbb_max_alloc_count = %u\n", info->pbb_max_alloc_count);
1582 
1583    fprintf(f, "GB_ADDR_CONFIG: 0x%08x\n", info->gb_addr_config);
1584    if (info->gfx_level >= GFX10) {
1585       fprintf(f, "    num_pipes = %u\n", 1 << G_0098F8_NUM_PIPES(info->gb_addr_config));
1586       fprintf(f, "    pipe_interleave_size = %u\n",
1587               256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(info->gb_addr_config));
1588       fprintf(f, "    max_compressed_frags = %u\n",
1589               1 << G_0098F8_MAX_COMPRESSED_FRAGS(info->gb_addr_config));
1590       if (info->gfx_level >= GFX10_3)
1591          fprintf(f, "    num_pkrs = %u\n", 1 << G_0098F8_NUM_PKRS(info->gb_addr_config));
1592    } else if (info->gfx_level == GFX9) {
1593       fprintf(f, "    num_pipes = %u\n", 1 << G_0098F8_NUM_PIPES(info->gb_addr_config));
1594       fprintf(f, "    pipe_interleave_size = %u\n",
1595               256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(info->gb_addr_config));
1596       fprintf(f, "    max_compressed_frags = %u\n",
1597               1 << G_0098F8_MAX_COMPRESSED_FRAGS(info->gb_addr_config));
1598       fprintf(f, "    bank_interleave_size = %u\n",
1599               1 << G_0098F8_BANK_INTERLEAVE_SIZE(info->gb_addr_config));
1600       fprintf(f, "    num_banks = %u\n", 1 << G_0098F8_NUM_BANKS(info->gb_addr_config));
1601       fprintf(f, "    shader_engine_tile_size = %u\n",
1602               16 << G_0098F8_SHADER_ENGINE_TILE_SIZE(info->gb_addr_config));
1603       fprintf(f, "    num_shader_engines = %u\n",
1604               1 << G_0098F8_NUM_SHADER_ENGINES_GFX9(info->gb_addr_config));
1605       fprintf(f, "    num_gpus = %u (raw)\n", G_0098F8_NUM_GPUS_GFX9(info->gb_addr_config));
1606       fprintf(f, "    multi_gpu_tile_size = %u (raw)\n",
1607               G_0098F8_MULTI_GPU_TILE_SIZE(info->gb_addr_config));
1608       fprintf(f, "    num_rb_per_se = %u\n", 1 << G_0098F8_NUM_RB_PER_SE(info->gb_addr_config));
1609       fprintf(f, "    row_size = %u\n", 1024 << G_0098F8_ROW_SIZE(info->gb_addr_config));
1610       fprintf(f, "    num_lower_pipes = %u (raw)\n", G_0098F8_NUM_LOWER_PIPES(info->gb_addr_config));
1611       fprintf(f, "    se_enable = %u (raw)\n", G_0098F8_SE_ENABLE(info->gb_addr_config));
1612    } else {
1613       fprintf(f, "    num_pipes = %u\n", 1 << G_0098F8_NUM_PIPES(info->gb_addr_config));
1614       fprintf(f, "    pipe_interleave_size = %u\n",
1615               256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX6(info->gb_addr_config));
1616       fprintf(f, "    bank_interleave_size = %u\n",
1617               1 << G_0098F8_BANK_INTERLEAVE_SIZE(info->gb_addr_config));
1618       fprintf(f, "    num_shader_engines = %u\n",
1619               1 << G_0098F8_NUM_SHADER_ENGINES_GFX6(info->gb_addr_config));
1620       fprintf(f, "    shader_engine_tile_size = %u\n",
1621               16 << G_0098F8_SHADER_ENGINE_TILE_SIZE(info->gb_addr_config));
1622       fprintf(f, "    num_gpus = %u (raw)\n", G_0098F8_NUM_GPUS_GFX6(info->gb_addr_config));
1623       fprintf(f, "    multi_gpu_tile_size = %u (raw)\n",
1624               G_0098F8_MULTI_GPU_TILE_SIZE(info->gb_addr_config));
1625       fprintf(f, "    row_size = %u\n", 1024 << G_0098F8_ROW_SIZE(info->gb_addr_config));
1626       fprintf(f, "    num_lower_pipes = %u (raw)\n", G_0098F8_NUM_LOWER_PIPES(info->gb_addr_config));
1627    }
1628 }
1629 
ac_get_gs_table_depth(enum amd_gfx_level gfx_level,enum radeon_family family)1630 int ac_get_gs_table_depth(enum amd_gfx_level gfx_level, enum radeon_family family)
1631 {
1632    if (gfx_level >= GFX9)
1633       return -1;
1634 
1635    switch (family) {
1636    case CHIP_OLAND:
1637    case CHIP_HAINAN:
1638    case CHIP_KAVERI:
1639    case CHIP_KABINI:
1640    case CHIP_ICELAND:
1641    case CHIP_CARRIZO:
1642    case CHIP_STONEY:
1643       return 16;
1644    case CHIP_TAHITI:
1645    case CHIP_PITCAIRN:
1646    case CHIP_VERDE:
1647    case CHIP_BONAIRE:
1648    case CHIP_HAWAII:
1649    case CHIP_TONGA:
1650    case CHIP_FIJI:
1651    case CHIP_POLARIS10:
1652    case CHIP_POLARIS11:
1653    case CHIP_POLARIS12:
1654    case CHIP_VEGAM:
1655       return 32;
1656    default:
1657       unreachable("Unknown GPU");
1658    }
1659 }
1660 
ac_get_raster_config(struct radeon_info * info,uint32_t * raster_config_p,uint32_t * raster_config_1_p,uint32_t * se_tile_repeat_p)1661 void ac_get_raster_config(struct radeon_info *info, uint32_t *raster_config_p,
1662                           uint32_t *raster_config_1_p, uint32_t *se_tile_repeat_p)
1663 {
1664    unsigned raster_config, raster_config_1, se_tile_repeat;
1665 
1666    switch (info->family) {
1667    /* 1 SE / 1 RB */
1668    case CHIP_HAINAN:
1669    case CHIP_KABINI:
1670    case CHIP_STONEY:
1671       raster_config = 0x00000000;
1672       raster_config_1 = 0x00000000;
1673       break;
1674    /* 1 SE / 4 RBs */
1675    case CHIP_VERDE:
1676       raster_config = 0x0000124a;
1677       raster_config_1 = 0x00000000;
1678       break;
1679    /* 1 SE / 2 RBs (Oland is special) */
1680    case CHIP_OLAND:
1681       raster_config = 0x00000082;
1682       raster_config_1 = 0x00000000;
1683       break;
1684    /* 1 SE / 2 RBs */
1685    case CHIP_KAVERI:
1686    case CHIP_ICELAND:
1687    case CHIP_CARRIZO:
1688       raster_config = 0x00000002;
1689       raster_config_1 = 0x00000000;
1690       break;
1691    /* 2 SEs / 4 RBs */
1692    case CHIP_BONAIRE:
1693    case CHIP_POLARIS11:
1694    case CHIP_POLARIS12:
1695       raster_config = 0x16000012;
1696       raster_config_1 = 0x00000000;
1697       break;
1698    /* 2 SEs / 8 RBs */
1699    case CHIP_TAHITI:
1700    case CHIP_PITCAIRN:
1701       raster_config = 0x2a00126a;
1702       raster_config_1 = 0x00000000;
1703       break;
1704    /* 4 SEs / 8 RBs */
1705    case CHIP_TONGA:
1706    case CHIP_POLARIS10:
1707       raster_config = 0x16000012;
1708       raster_config_1 = 0x0000002a;
1709       break;
1710    /* 4 SEs / 16 RBs */
1711    case CHIP_HAWAII:
1712    case CHIP_FIJI:
1713    case CHIP_VEGAM:
1714       raster_config = 0x3a00161a;
1715       raster_config_1 = 0x0000002e;
1716       break;
1717    default:
1718       fprintf(stderr, "ac: Unknown GPU, using 0 for raster_config\n");
1719       raster_config = 0x00000000;
1720       raster_config_1 = 0x00000000;
1721       break;
1722    }
1723 
1724    /* drm/radeon on Kaveri is buggy, so disable 1 RB to work around it.
1725     * This decreases performance by up to 50% when the RB is the bottleneck.
1726     */
1727    if (info->family == CHIP_KAVERI && !info->is_amdgpu)
1728       raster_config = 0x00000000;
1729 
1730    /* Fiji: Old kernels have incorrect tiling config. This decreases
1731     * RB performance by 25%. (it disables 1 RB in the second packer)
1732     */
1733    if (info->family == CHIP_FIJI && info->cik_macrotile_mode_array[0] == 0x000000e8) {
1734       raster_config = 0x16000012;
1735       raster_config_1 = 0x0000002a;
1736    }
1737 
1738    unsigned se_width = 8 << G_028350_SE_XSEL_GFX6(raster_config);
1739    unsigned se_height = 8 << G_028350_SE_YSEL_GFX6(raster_config);
1740 
1741    /* I don't know how to calculate this, though this is probably a good guess. */
1742    se_tile_repeat = MAX2(se_width, se_height) * info->max_se;
1743 
1744    *raster_config_p = raster_config;
1745    *raster_config_1_p = raster_config_1;
1746    if (se_tile_repeat_p)
1747       *se_tile_repeat_p = se_tile_repeat;
1748 }
1749 
ac_get_harvested_configs(struct radeon_info * info,unsigned raster_config,unsigned * cik_raster_config_1_p,unsigned * raster_config_se)1750 void ac_get_harvested_configs(struct radeon_info *info, unsigned raster_config,
1751                               unsigned *cik_raster_config_1_p, unsigned *raster_config_se)
1752 {
1753    unsigned sh_per_se = MAX2(info->max_sa_per_se, 1);
1754    unsigned num_se = MAX2(info->max_se, 1);
1755    unsigned rb_mask = info->enabled_rb_mask;
1756    unsigned num_rb = MIN2(info->max_render_backends, 16);
1757    unsigned rb_per_pkr = MIN2(num_rb / num_se / sh_per_se, 2);
1758    unsigned rb_per_se = num_rb / num_se;
1759    unsigned se_mask[4];
1760    unsigned se;
1761 
1762    se_mask[0] = ((1 << rb_per_se) - 1) & rb_mask;
1763    se_mask[1] = (se_mask[0] << rb_per_se) & rb_mask;
1764    se_mask[2] = (se_mask[1] << rb_per_se) & rb_mask;
1765    se_mask[3] = (se_mask[2] << rb_per_se) & rb_mask;
1766 
1767    assert(num_se == 1 || num_se == 2 || num_se == 4);
1768    assert(sh_per_se == 1 || sh_per_se == 2);
1769    assert(rb_per_pkr == 1 || rb_per_pkr == 2);
1770 
1771    if (info->gfx_level >= GFX7) {
1772       unsigned raster_config_1 = *cik_raster_config_1_p;
1773       if ((num_se > 2) && ((!se_mask[0] && !se_mask[1]) || (!se_mask[2] && !se_mask[3]))) {
1774          raster_config_1 &= C_028354_SE_PAIR_MAP;
1775 
1776          if (!se_mask[0] && !se_mask[1]) {
1777             raster_config_1 |= S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_3);
1778          } else {
1779             raster_config_1 |= S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_0);
1780          }
1781          *cik_raster_config_1_p = raster_config_1;
1782       }
1783    }
1784 
1785    for (se = 0; se < num_se; se++) {
1786       unsigned pkr0_mask = ((1 << rb_per_pkr) - 1) << (se * rb_per_se);
1787       unsigned pkr1_mask = pkr0_mask << rb_per_pkr;
1788       int idx = (se / 2) * 2;
1789 
1790       raster_config_se[se] = raster_config;
1791       if ((num_se > 1) && (!se_mask[idx] || !se_mask[idx + 1])) {
1792          raster_config_se[se] &= C_028350_SE_MAP;
1793 
1794          if (!se_mask[idx]) {
1795             raster_config_se[se] |= S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_3);
1796          } else {
1797             raster_config_se[se] |= S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_0);
1798          }
1799       }
1800 
1801       pkr0_mask &= rb_mask;
1802       pkr1_mask &= rb_mask;
1803       if (rb_per_se > 2 && (!pkr0_mask || !pkr1_mask)) {
1804          raster_config_se[se] &= C_028350_PKR_MAP;
1805 
1806          if (!pkr0_mask) {
1807             raster_config_se[se] |= S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_3);
1808          } else {
1809             raster_config_se[se] |= S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_0);
1810          }
1811       }
1812 
1813       if (rb_per_se >= 2) {
1814          unsigned rb0_mask = 1 << (se * rb_per_se);
1815          unsigned rb1_mask = rb0_mask << 1;
1816 
1817          rb0_mask &= rb_mask;
1818          rb1_mask &= rb_mask;
1819          if (!rb0_mask || !rb1_mask) {
1820             raster_config_se[se] &= C_028350_RB_MAP_PKR0;
1821 
1822             if (!rb0_mask) {
1823                raster_config_se[se] |= S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_3);
1824             } else {
1825                raster_config_se[se] |= S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_0);
1826             }
1827          }
1828 
1829          if (rb_per_se > 2) {
1830             rb0_mask = 1 << (se * rb_per_se + rb_per_pkr);
1831             rb1_mask = rb0_mask << 1;
1832             rb0_mask &= rb_mask;
1833             rb1_mask &= rb_mask;
1834             if (!rb0_mask || !rb1_mask) {
1835                raster_config_se[se] &= C_028350_RB_MAP_PKR1;
1836 
1837                if (!rb0_mask) {
1838                   raster_config_se[se] |= S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_3);
1839                } else {
1840                   raster_config_se[se] |= S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_0);
1841                }
1842             }
1843          }
1844       }
1845    }
1846 }
1847 
1848 unsigned
ac_get_compute_resource_limits(const struct radeon_info * info,unsigned waves_per_threadgroup,unsigned max_waves_per_sh,unsigned threadgroups_per_cu)1849 ac_get_compute_resource_limits(const struct radeon_info *info, unsigned waves_per_threadgroup,
1850                                unsigned max_waves_per_sh, unsigned threadgroups_per_cu)
1851 {
1852    unsigned compute_resource_limits = S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup % 4 == 0);
1853 
1854    if (info->gfx_level >= GFX7) {
1855       unsigned num_cu_per_se = info->num_cu / info->num_se;
1856 
1857       /* Gfx9 should set the limit to max instead of 0 to fix high priority compute. */
1858       if (info->gfx_level == GFX9 && !max_waves_per_sh) {
1859          max_waves_per_sh = info->max_good_cu_per_sa * info->num_simd_per_compute_unit *
1860                             info->max_wave64_per_simd;
1861       }
1862 
1863       /* Force even distribution on all SIMDs in CU if the workgroup
1864        * size is 64. This has shown some good improvements if # of CUs
1865        * per SE is not a multiple of 4.
1866        */
1867       if (num_cu_per_se % 4 && waves_per_threadgroup == 1)
1868          compute_resource_limits |= S_00B854_FORCE_SIMD_DIST(1);
1869 
1870       assert(threadgroups_per_cu >= 1 && threadgroups_per_cu <= 8);
1871       compute_resource_limits |=
1872          S_00B854_WAVES_PER_SH(max_waves_per_sh) | S_00B854_CU_GROUP_COUNT(threadgroups_per_cu - 1);
1873    } else {
1874       /* GFX6 */
1875       if (max_waves_per_sh) {
1876          unsigned limit_div16 = DIV_ROUND_UP(max_waves_per_sh, 16);
1877          compute_resource_limits |= S_00B854_WAVES_PER_SH_GFX6(limit_div16);
1878       }
1879    }
1880    return compute_resource_limits;
1881 }
1882 
ac_get_hs_info(struct radeon_info * info,struct ac_hs_info * hs)1883 void ac_get_hs_info(struct radeon_info *info,
1884                     struct ac_hs_info *hs)
1885 {
1886    bool double_offchip_buffers = info->gfx_level >= GFX7 &&
1887                                  info->family != CHIP_CARRIZO &&
1888                                  info->family != CHIP_STONEY;
1889    unsigned max_offchip_buffers_per_se;
1890    unsigned max_offchip_buffers;
1891    unsigned offchip_granularity;
1892    unsigned hs_offchip_param;
1893 
1894    hs->tess_offchip_block_dw_size =
1895       info->family == CHIP_HAWAII ? 4096 : 8192;
1896 
1897    /*
1898     * Per RadeonSI:
1899     * This must be one less than the maximum number due to a hw limitation.
1900     * Various hardware bugs need this.
1901     *
1902     * Per AMDVLK:
1903     * Vega10 should limit max_offchip_buffers to 508 (4 * 127).
1904     * Gfx7 should limit max_offchip_buffers to 508
1905     * Gfx6 should limit max_offchip_buffers to 126 (2 * 63)
1906     *
1907     * Follow AMDVLK here.
1908     */
1909    if (info->gfx_level >= GFX11) {
1910       max_offchip_buffers_per_se = 256; /* TODO: we could decrease this to reduce memory/cache usage */
1911    } else if (info->gfx_level >= GFX10) {
1912       max_offchip_buffers_per_se = 128;
1913    } else if (info->family == CHIP_VEGA12 || info->family == CHIP_VEGA20) {
1914       /* Only certain chips can use the maximum value. */
1915       max_offchip_buffers_per_se = double_offchip_buffers ? 128 : 64;
1916    } else {
1917       max_offchip_buffers_per_se = double_offchip_buffers ? 127 : 63;
1918    }
1919 
1920    max_offchip_buffers = max_offchip_buffers_per_se * info->max_se;
1921 
1922    /* Hawaii has a bug with offchip buffers > 256 that can be worked
1923     * around by setting 4K granularity.
1924     */
1925    if (hs->tess_offchip_block_dw_size == 4096) {
1926       assert(info->family == CHIP_HAWAII);
1927       offchip_granularity = V_03093C_X_4K_DWORDS;
1928    } else {
1929       assert(hs->tess_offchip_block_dw_size == 8192);
1930       offchip_granularity = V_03093C_X_8K_DWORDS;
1931    }
1932 
1933    switch (info->gfx_level) {
1934    case GFX6:
1935       max_offchip_buffers = MIN2(max_offchip_buffers, 126);
1936       break;
1937    case GFX7:
1938    case GFX8:
1939    case GFX9:
1940       max_offchip_buffers = MIN2(max_offchip_buffers, 508);
1941       break;
1942    case GFX10:
1943       break;
1944    default:
1945       break;
1946    }
1947 
1948    hs->max_offchip_buffers = max_offchip_buffers;
1949 
1950    if (info->gfx_level >= GFX11) {
1951       /* OFFCHIP_BUFFERING is per SE. */
1952       hs_offchip_param = S_03093C_OFFCHIP_BUFFERING_GFX103(max_offchip_buffers_per_se - 1) |
1953                          S_03093C_OFFCHIP_GRANULARITY_GFX103(offchip_granularity);
1954    } else if (info->gfx_level >= GFX10_3) {
1955       hs_offchip_param = S_03093C_OFFCHIP_BUFFERING_GFX103(max_offchip_buffers - 1) |
1956                          S_03093C_OFFCHIP_GRANULARITY_GFX103(offchip_granularity);
1957    } else if (info->gfx_level >= GFX7) {
1958       if (info->gfx_level >= GFX8)
1959          --max_offchip_buffers;
1960       hs_offchip_param = S_03093C_OFFCHIP_BUFFERING_GFX7(max_offchip_buffers) |
1961                          S_03093C_OFFCHIP_GRANULARITY_GFX7(offchip_granularity);
1962    } else {
1963       hs_offchip_param = S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers);
1964    }
1965 
1966    hs->hs_offchip_param = hs_offchip_param;
1967 
1968    hs->tess_factor_ring_size = 48 * 1024 * info->max_se;
1969    hs->tess_offchip_ring_offset = align(hs->tess_factor_ring_size, 64 * 1024);
1970    hs->tess_offchip_ring_size = hs->max_offchip_buffers * hs->tess_offchip_block_dw_size * 4;
1971 }
1972 
get_task_num_entries(enum radeon_family fam)1973 static uint16_t get_task_num_entries(enum radeon_family fam)
1974 {
1975    /* Number of task shader ring entries. Needs to be a power of two.
1976     * Use a low number on smaller chips so we don't waste space,
1977     * but keep it high on bigger chips so it doesn't inhibit parallelism.
1978     *
1979     * This number is compiled into task/mesh shaders as a constant.
1980     * In order to ensure this works fine with the shader cache, we must
1981     * base this decision on the chip family, not the number of CUs in
1982     * the current GPU. (So, the cache remains consistent for all
1983     * chips in the same family.)
1984     */
1985    switch (fam) {
1986    case CHIP_VANGOGH:
1987    case CHIP_NAVI24:
1988    case CHIP_REMBRANDT:
1989       return 256;
1990    case CHIP_NAVI21:
1991    case CHIP_NAVI22:
1992    case CHIP_NAVI23:
1993    default:
1994       return 1024;
1995    }
1996 }
1997 
ac_get_task_info(struct radeon_info * info,struct ac_task_info * task_info)1998 void ac_get_task_info(struct radeon_info *info,
1999                       struct ac_task_info *task_info)
2000 {
2001    const uint16_t num_entries = get_task_num_entries(info->family);
2002    const uint32_t draw_ring_bytes = num_entries * AC_TASK_DRAW_ENTRY_BYTES;
2003    const uint32_t payload_ring_bytes = num_entries * AC_TASK_PAYLOAD_ENTRY_BYTES;
2004 
2005    /* Ensure that the addresses of each ring are 256 byte aligned. */
2006    task_info->num_entries = num_entries;
2007    task_info->draw_ring_offset = ALIGN(AC_TASK_CTRLBUF_BYTES, 256);
2008    task_info->payload_ring_offset = ALIGN(task_info->draw_ring_offset + draw_ring_bytes, 256);
2009    task_info->bo_size_bytes = task_info->payload_ring_offset + payload_ring_bytes;
2010 }
2011