xref: /aosp_15_r20/external/mesa3d/src/intel/vulkan/anv_va.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright © 2023 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  */
23 
24 #include "anv_private.h"
25 
26 #include "util/u_math.h"
27 
28 static uint64_t
va_add(struct anv_va_range * range,uint64_t addr,uint64_t size)29 va_add(struct anv_va_range *range, uint64_t addr, uint64_t size)
30 {
31    range->addr = addr;
32    range->size = size;
33 
34    return addr + size;
35 }
36 
37 static void
va_at(struct anv_va_range * range,uint64_t addr,uint64_t size)38 va_at(struct anv_va_range *range, uint64_t addr, uint64_t size)
39 {
40    range->addr = addr;
41    range->size = size;
42 }
43 
44 static void
anv_device_print_vas(struct anv_physical_device * device)45 anv_device_print_vas(struct anv_physical_device *device)
46 {
47    fprintf(stderr, "Driver heaps:\n");
48 #define PRINT_HEAP(name) \
49    fprintf(stderr, "   0x%016"PRIx64"-0x%016"PRIx64": %s\n", \
50            device->va.name.addr, \
51            device->va.name.addr + device->va.name.size, \
52            #name);
53    PRINT_HEAP(general_state_pool);
54    PRINT_HEAP(low_heap);
55    PRINT_HEAP(binding_table_pool);
56    PRINT_HEAP(internal_surface_state_pool);
57    PRINT_HEAP(scratch_surface_state_pool);
58    PRINT_HEAP(bindless_surface_state_pool);
59    PRINT_HEAP(indirect_descriptor_pool);
60    PRINT_HEAP(indirect_push_descriptor_pool);
61    PRINT_HEAP(instruction_state_pool);
62    PRINT_HEAP(dynamic_state_pool);
63    PRINT_HEAP(dynamic_visible_pool);
64    PRINT_HEAP(push_descriptor_buffer_pool);
65    PRINT_HEAP(high_heap);
66    PRINT_HEAP(trtt);
67 }
68 
69 void
anv_physical_device_init_va_ranges(struct anv_physical_device * device)70 anv_physical_device_init_va_ranges(struct anv_physical_device *device)
71 {
72    /* anv Virtual Memory Layout
73     * =========================
74     *
75     * When the anv driver is determining the virtual graphics addresses of
76     * memory objects itself using the softpin mechanism, the following memory
77     * ranges will be used.
78     *
79     * Three special considerations to notice:
80     *
81     * (1) the dynamic state pool is located within the same 4 GiB as the low
82     * heap. This is to work around a VF cache issue described in a comment in
83     * anv_physical_device_init_heaps.
84     *
85     * (2) the binding table pool is located at lower addresses than the BT
86     * (binding table) surface state pool, within a 4 GiB range which also
87     * contains the bindless surface state pool. This allows surface state base
88     * addresses to cover both binding tables (16 bit offsets), the internal
89     * surface states (32 bit offsets) and the bindless surface states.
90     *
91     * (3) the last 4 GiB of the address space is withheld from the high heap.
92     * Various hardware units will read past the end of an object for various
93     * reasons. This healthy margin prevents reads from wrapping around 48-bit
94     * addresses.
95     */
96    uint64_t _1Mb = 1ull * 1024 * 1024;
97    uint64_t _1Gb = 1ull * 1024 * 1024 * 1024;
98    uint64_t _4Gb = 4ull * 1024 * 1024 * 1024;
99 
100    uint64_t address = 0x000000200000ULL; /* 2MiB */
101 
102    address = va_add(&device->va.general_state_pool, address,
103                     _1Gb - address);
104 
105    address = va_add(&device->va.low_heap, address, _1Gb);
106 
107    /* The binding table pool has to be located directly in front of the
108     * surface states.
109     */
110    address += _1Gb;
111    address = va_add(&device->va.binding_table_pool, address, _1Gb);
112    address = va_add(&device->va.internal_surface_state_pool, address, 1 * _1Gb);
113    assert(device->va.internal_surface_state_pool.addr ==
114           align64(device->va.internal_surface_state_pool.addr, 2 * _1Gb));
115    /* Scratch surface state overlaps with the internal surface state */
116    va_at(&device->va.scratch_surface_state_pool,
117          device->va.internal_surface_state_pool.addr,
118          8 * _1Mb);
119    address = va_add(&device->va.bindless_surface_state_pool, address, 2 * _1Gb);
120 
121    if (device->indirect_descriptors) {
122       /* With indirect descriptors, descriptor buffers can go anywhere, they
123        * just need to be in a 4Gb aligned range, so all shader accesses can
124        * use a relocatable upper dword for the 64bit address.
125        */
126       address = align64(address, _4Gb);
127       address = va_add(&device->va.indirect_descriptor_pool, address, 3 * _1Gb);
128       address = va_add(&device->va.indirect_push_descriptor_pool, address, _1Gb);
129    }
130 
131    /* We use a trick to compute constant data offsets in the shaders to avoid
132     * unnecessary 64bit address computations (see lower_load_constant() in
133     * anv_nir_apply_pipeline_layout.c). This assumes the instruction pool is
134     * located at an address with the lower 32bits at 0.
135     */
136    address = align64(address, _4Gb);
137    address = va_add(&device->va.instruction_state_pool, address, 2 * _1Gb);
138 
139    address += 1 * _1Gb;
140    address = va_add(&device->va.dynamic_state_pool, address, _1Gb);
141    address = va_add(&device->va.dynamic_visible_pool, address,
142                     device->info.verx10 >= 125 ? (2 * _1Gb) : (3 * _1Gb - 4096));
143    assert(device->va.dynamic_visible_pool.addr % _4Gb == 0);
144    if (device->info.verx10 >= 125)
145       address = va_add(&device->va.push_descriptor_buffer_pool, address, _1Gb - 4096);
146 
147    address = align64(address, device->info.mem_alignment);
148    address = va_add(&device->va.aux_tt_pool, address, 2 * _1Gb);
149 
150    /* What's left to do for us is to set va.high_heap and va.trtt without
151     * overlap, but there are a few things to be considered:
152     *
153     * The TR-TT address space is governed by the GFX_TRTT_VA_RANGE register,
154     * which carves out part of the address space for TR-TT and is independent
155     * of device->gtt_size. We use 47:44 for gen9+, the values we set here
156     * should be in sync with what we write to the register.
157     *
158     * If we ever gain the capability to use more than 48 bits of address space
159     * we'll have to adjust where we put the TR-TT space (and how we set
160     * GFX_TRTT_VA_RANGE).
161     *
162     * We have to leave the last 4GiB out of the high vma range, so that no
163     * state base address + size can overflow 48 bits. For more information see
164     * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
165     *
166     * Despite the comment above, before we had TR-TT we were not only avoiding
167     * the last 4GiB of the 48bit address space, but also avoiding the last
168     * 4GiB from gtt_size, so let's be on the safe side and do the 4GiB
169     * avoiding for both the TR-TT space top and the gtt top.
170     */
171    assert(device->gtt_size <= (1uLL << 48));
172    uint64_t trtt_start = 0xFuLL << 44;
173    uint64_t trtt_end = (1uLL << 48) - 4 * _1Gb;
174    uint64_t addressable_top = MIN2(device->gtt_size, trtt_start) - 4 * _1Gb;
175 
176    uint64_t user_heaps_size = addressable_top - address;
177    address = va_add(&device->va.high_heap, address, user_heaps_size);
178    assert(address <= trtt_start);
179    address = va_add(&device->va.trtt, trtt_start, trtt_end - trtt_start);
180 
181    if (INTEL_DEBUG(DEBUG_HEAPS))
182       anv_device_print_vas(device);
183 }
184