IREE体系结构介绍
//第10章/build_hal.executable.c
module {
hal.executable @executable_module {
hal.interface @abi {
hal.interface.binding @ret, set=0, binding=0, type="StorageBuffer", access="Read"
...
}
hal.executable.binary {data = dense<[...]>: vector<1620×i8>, format = "SPIR-V"}...
func @main(%arg0: !iree.ref<!hal.buffer>, %arg1: !iree.ref<!hal.buffer>) -> !iree.ref<!hal.buffer> {
%dev = hal.ex.shared_device: !iree.ref<!hal.device>
%allocator = hal.device.allocator %dev: !iree.ref<!hal.allocator>
%buffer = hal.allocator.allocate %allocator,..., shape=[...], element_size=4: !iree.ref<!hal.buffer>
%cmd = hal.command_buffer.create %dev, "OneShot", "Transfer|Dispatch": !iree.ref<!hal.command_buffer>
hal.command_buffer.begin %cmd
hal.ex.push_descriptor_set %cmd,...
hal.device.switch(%dev: !hal.device)
#hal.device.match.id<"vulkan*">: !iree.ref<!hal.buffer> {
...
%exe = hal.executable.look_up,...
hal.command_buffer.dispatch %cmd, %exe, entry_point=0, workgroup_xyz=[%c1, %c5, %c1]
}
%memory_barrier = hal.make_memory_barrier "DispatchWrite", "DispatchRead": tuple<i32, i32>
hal.command_buffer.execution_barrier %cmd, "CommandRetire", "CommandIssue",
memory_barriers=[%memory_barrier]
...
hal.command_buffer.end %cmd
hal.ex.submit_and_wait %dev, %cmd
return %buffer: !iree.ref<!hal.buffer>
|
FP32 Model |
GPU/FLOps |
IREE延迟 |
比较延迟 |
移动 |
MobileBERT |
ARM Mali G78 (Pixel 6)/2T |
120ms |
TFLite OpenCL 123ms |
笔记本电脑 |
miniLM |
Apple M1 Max/10.4T |
11.6ms |
TF-Metal 16.99ms |
桌面 |
miniLM |
AMD RX 5700XT/9.7T |
8ms |
|
服务器 |
miniLM |
NVIDIA Tesla V100/15.7T |
6.3ms |
|
|
FP32 Model |
IREE延迟 |
TFLite延迟(缓存) |
TFLite延迟 (结构) |
谷歌Pixel6 |
MobileBERT |
120ms |
172ms |
123ms |
谷歌Pixel6 |
MobileNetV2 |
9ms |
8ms |
6ms |
谷歌Pixel6 |
DeepLabV3 |
12ms |
12.1ms |
9.2ms |
谷歌Pixel6 |
PoseNet |
15ms |
14.4ms |
8ms |