|
| 1 | +// Copyright 2024 The Khronos Group, Inc. |
| 2 | +// SPDX-License-Identifier: CC-BY-4.0 |
| 3 | + |
| 4 | +// Required for both single-page and combined guide xrefs to work |
| 5 | +ifndef::chapters[:chapters:] |
| 6 | +ifndef::images[:images: images/] |
| 7 | + |
| 8 | +[[buffer-device-address-alignment]] |
| 9 | += Buffer Device Address Alignment |
| 10 | + |
| 11 | +All variables accessed with `PhysicalStorageBuffer` must have an `Aligned` memory operand to it. |
| 12 | + |
| 13 | +[source,swift] |
| 14 | +---- |
| 15 | +%x = OpLoad %type %ptr Aligned 16 |
| 16 | +OpStore %ptr %obj Aligned 16 |
| 17 | +---- |
| 18 | + |
| 19 | +Shading languages will have a default, but can allow you to align it explicitly (ex `buffer_reference_alignment`). |
| 20 | + |
| 21 | +The goal of this alignment is this is a promise for how aligned this specific pointer is. |
| 22 | +The compiler has no idea what the address will be when the shader is compiled. |
| 23 | +By providing an alignment it can generate valid code to match the requirement. |
| 24 | +The user is responsible to confirm the address they use is aligned to it. |
| 25 | + |
| 26 | +[source,glsl] |
| 27 | +---- |
| 28 | +layout(buffer_reference, buffer_reference_align = 64) buffer MyBDA { |
| 29 | + uint data; |
| 30 | +}; |
| 31 | +
|
| 32 | +MyBDA ptr_a; // at 0x1000 |
| 33 | +MyBDA ptr_b; // at 0x1010 |
| 34 | +MyBDA ptr_c; // at 0x1040 |
| 35 | +
|
| 36 | +ptr_a.data = 0; // (Aligned 64) valid! |
| 37 | +ptr_b.data = 0; // (Aligned 64) invalid! |
| 38 | +ptr_c.data = 0; // (Aligned 64) valid! |
| 39 | +---- |
| 40 | + |
| 41 | +When deciding on an alignment, the minimum value will always be the size greater than or equal to the largest scalar/component type in the block. |
| 42 | + |
| 43 | +[source,glsl] |
| 44 | +---- |
| 45 | +// alignment must be at least 4 |
| 46 | +layout(buffer_reference) buffer MyBDA { |
| 47 | + vec4 a; // scalar is float |
| 48 | +}; |
| 49 | +
|
| 50 | +// alignment must be at least 1 |
| 51 | +layout(buffer_reference) buffer MyBDA { |
| 52 | + uint8_t a; // scalar is 8-bit int |
| 53 | +}; |
| 54 | +
|
| 55 | +// alignment must be at least 8 |
| 56 | +layout(buffer_reference) buffer MyBDA { |
| 57 | + uint a; // 32-bit |
| 58 | + double b; // 64-bit |
| 59 | +}; |
| 60 | +---- |
| 61 | + |
| 62 | +== Setting Alignment Example |
| 63 | + |
| 64 | +To help explain alignment, lets take an example of loading an array of vectors |
| 65 | + |
| 66 | +[source,glsl] |
| 67 | +---- |
| 68 | +layout(buffer_reference, buffer_reference_align = ???) buffer MyBDA { |
| 69 | + uvec4 data[]; |
| 70 | +}; |
| 71 | +
|
| 72 | +MyBDA ptr; // at 0x1000 |
| 73 | +ptr.data[i] = uvec4(0); |
| 74 | +---- |
| 75 | + |
| 76 | +Here we have 2 options, we could set the `Aligned` to be `4` or `16`. |
| 77 | + |
| 78 | +If we set alignment to `16` we are letting the compiler know it can load 16 bytes at a time, so it will hopefully do a vector load/store on the memory. |
| 79 | + |
| 80 | +If we set alignment to `4` the compiler will likely have no way to infer the real alignment and will now do 4 scalar int load/store on the memory. |
| 81 | + |
| 82 | +[NOTE] |
| 83 | +==== |
| 84 | +Some GPUs can do vector load/store even on unaligned addresses. |
| 85 | +==== |
| 86 | + |
| 87 | +For the next case, if we had `uvec3` instead of `uvec4` such as |
| 88 | + |
| 89 | +[source,glsl] |
| 90 | +---- |
| 91 | +layout(buffer_reference, buffer_reference_align = 4, scalar) buffer MyBDA { |
| 92 | + uvec3 data[]; |
| 93 | +}; |
| 94 | +
|
| 95 | +data[0]; // 0x1000 |
| 96 | +data[1]; // 0x100C |
| 97 | +data[2]; // 0x1018 |
| 98 | +data[3]; // 0x1024 |
| 99 | +---- |
| 100 | + |
| 101 | +We know that setting the alignment to `16` would be violated at `data[1]` and therefore we need to use an alignment of `4` in this case. |
| 102 | +Luckily shading languages will help do this for you as seen in both link:https://godbolt.org/z/jWGKax1ed[glslang] and link:https://godbolt.org/z/Y7xW3Mfd4[slang]. |
| 103 | + |
| 104 | +== Matching Alignment From The Host |
| 105 | + |
| 106 | +When dealing with buffer device address, you are able to do a simple `memcpy` to that memory on the host, which can easily lead to bugs if you aren't careful about things being aligned. |
| 107 | + |
| 108 | +[NOTE] |
| 109 | +==== |
| 110 | +The following issues are not directly tied to Buffer Device Address, and still can occur with any uniform or storage buffer. |
| 111 | +==== |
| 112 | + |
| 113 | +Take the following GLSL code as an example (link:https://godbolt.org/z/G4P8GdG9q[view online]) |
| 114 | + |
| 115 | +[source,glsl] |
| 116 | +---- |
| 117 | +// ArrayStride is 16 |
| 118 | +struct Metadata { |
| 119 | + uint64_t address; |
| 120 | + uint status; |
| 121 | +}; |
| 122 | +
|
| 123 | +layout(buffer_reference, buffer_reference_align = 8, scalar) readonly buffer Payload { |
| 124 | + uint count; // offset 0 |
| 125 | + Metadata meta[]; // offset 8 |
| 126 | +}; |
| 127 | +
|
| 128 | +layout(set = 0, binding = 0) buffer SSBO_0 { |
| 129 | + Payload data; |
| 130 | +}; |
| 131 | +---- |
| 132 | + |
| 133 | +Because the `uint64_t` needs be accessed at an 8-byte alignment, `glslang` (and any other compiler) will be smart and pack things as tightly as possible for you. |
| 134 | + |
| 135 | +The first thing you might notice is `Metadata` needs to have an array stride of 16 instead of 12. This is because otherwise `uint64_t address` will land on a non 8-byte alignment every other instance of the array. |
| 136 | + |
| 137 | +The next thing happening is because `struct Metedata` **largest scalar** is an 8-byte value, it knows to have the offset at `8` instead of `4`. This is why trying to change the struct to |
| 138 | + |
| 139 | +[source,glsl] |
| 140 | +---- |
| 141 | +struct Metadata { |
| 142 | + uint status; |
| 143 | + uint64_t address; |
| 144 | +}; |
| 145 | +---- |
| 146 | + |
| 147 | +or |
| 148 | + |
| 149 | +[source,glsl] |
| 150 | +---- |
| 151 | +struct Metadata { |
| 152 | + uint64_t address; |
| 153 | + uint status; |
| 154 | + uint pad; |
| 155 | +}; |
| 156 | +---- |
| 157 | + |
| 158 | +won't change the offset from `8`. |
| 159 | + |
| 160 | +Here is how the memory is laid out in memory: |
| 161 | + |
| 162 | +image::{images}buffer_device_address_alignment_1.svg[buffer_device_address_alignment_1.svg] |
| 163 | + |
| 164 | +So the issue here becomes when we try to map our host memory. When you call `vkMapMemory` and get a `void*` you need to cautious that memory needs to be laid out the same as the diagram above. One way to ensure this is use a struct on host as it will match the shader code. |
| 165 | + |
| 166 | +[source,c++] |
| 167 | +---- |
| 168 | +struct Metadata { |
| 169 | + uint64_t address; |
| 170 | + uint32_t status; |
| 171 | +}; |
| 172 | +
|
| 173 | +struct Payload { |
| 174 | + uint32_t count; |
| 175 | + Metadata meta[2]; |
| 176 | +} payload; |
| 177 | +
|
| 178 | +payload.count = 2; |
| 179 | +payload.meta[0].address = 0xDEADBEEF; |
| 180 | +payload.meta[0].status = 20; |
| 181 | +payload.meta[1].address = 0xDEADBEEF; |
| 182 | +payload.meta[1].status = 5; |
| 183 | +
|
| 184 | +void* data; |
| 185 | +vkMapMemory(device, device_memory, 0, VK_WHOLE_SIZE, 0, &data); |
| 186 | +
|
| 187 | +// You can also just memcpy here as well! |
| 188 | +Payload *payload_ptr = (Payload*)data; |
| 189 | +*payload_ptr = payload; |
| 190 | +---- |
| 191 | + |
| 192 | +If we examine the C++ code here (https://godbolt.org/z/Gq75qq1x6) we can see the assembly also automatically maps the offsets the same as the GLSL code above! |
0 commit comments