Adding `hal.device.memoize` op. (#17938)
This is an inline nested region that returns HAL resources for the given
device and affinity pair. The region may only implicitly capture
constants and global values in order for it to be outlined for
initialization, though in the future we may extend this to support
dynamic values by adding an LRU cache.
To start all memoize regions are inlined after conversion. This optional
inlining will also serve as a useful debug feature for isolating
memoization related issues. The initial outlining and proper memoization
pass is present but disabled by default until we can verify all targets
properly support reusable command buffers.
Progress on #17875. Nothing is generating these yet.
Example:
```mlir
util.global private @device = #hal.device.target<"local"> : !hal.device
util.global private @executable : !hal.executable
util.global private @buffer : !hal.buffer
util.func public @memoize_command_buffer() -> !hal.command_buffer {
%device = util.global.load immutable @device : !hal.device
%affinity = arith.constant -1 : i64
%executable = util.global.load immutable @executable : !hal.executable
%buffer = util.global.load immutable @buffer : !hal.buffer
%result = hal.device.memoize<%device : !hal.device> affinity(%affinity) -> !hal.command_buffer {
%cmd = hal.command_buffer.create device(%device : !hal.device) mode(None) categories("Transfer|Dispatch") : !hal.command_buffer
%dispatch_ordinal = arith.constant 123 : index
%offset = arith.constant 100 : index
hal.command_buffer.dispatch.indirect<%cmd : !hal.command_buffer>
target(%executable : !hal.executable)[%dispatch_ordinal]
workgroups(%buffer : !hal.buffer)[%offset]
hal.command_buffer.execution_barrier<%cmd : !hal.command_buffer>
source(CommandIssue) target(CommandProcess) flags(None)
hal.command_buffer.finalize<%cmd : !hal.command_buffer>
hal.return %cmd : !hal.command_buffer
}
util.return %result : !hal.command_buffer
}
```
->
```mlir
util.global private @device = #hal.device.target<"local"> : !hal.device
util.global private @executable : !hal.executable
util.global private @global_resource : !hal.buffer
util.func private @__memoize_command_buffer_memoize_apply(%arg0: !hal.device, %arg1: i64, %arg2: !hal.executable, %arg3: !hal.buffer) -> !hal.command_buffer {
%cmd = hal.command_buffer.create device(%arg0 : !hal.device) mode(OneShot) categories("Transfer|Dispatch") : !hal.command_buffer
%c123 = arith.constant 123 : index
%c100 = arith.constant 100 : index
hal.command_buffer.dispatch.indirect<%cmd : !hal.command_buffer> target(%arg2 : !hal.executable)[%c123] workgroups(%arg3 : !hal.buffer)[%c100]
hal.command_buffer.execution_barrier<%cmd : !hal.command_buffer> source(CommandIssue) target(CommandProcess) flags("None")
hal.command_buffer.finalize<%cmd : !hal.command_buffer>
util.return %cmd : !hal.command_buffer
}
util.global private @__memoize_command_buffer_memoize_result_0_device : !hal.command_buffer
util.initializer {
%device = util.global.load immutable @device : !hal.device
%c-1_i64 = arith.constant -1 : i64
%executable = util.global.load immutable @executable : !hal.executable
%global_resource = util.global.load immutable @global_resource : !hal.buffer
%0 = util.call @__memoize_command_buffer_memoize_apply(%device, %c-1_i64, %executable, %global_resource) : (!hal.device, i64, !hal.executable, !hal.buffer) -> !hal.command_buffer
util.global.store %0, @__memoize_command_buffer_memoize_result_0_device : !hal.command_buffer
util.return
}
util.func private @__memoize_command_buffer_memoize_lookup(%arg0: !hal.device, %arg1: i64) -> !hal.command_buffer {
%device = util.global.load @device : !hal.device
%0 = util.cmp.eq %arg0, %device : !hal.device
%1 = scf.if %0 -> (!hal.command_buffer) {
%__memoize_command_buffer_memoize_result_0_device = util.global.load @__memoize_command_buffer_memoize_result_0_device : !hal.command_buffer
scf.yield %__memoize_command_buffer_memoize_result_0_device : !hal.command_buffer
} else {
%2 = util.null : !hal.command_buffer
scf.yield %2 : !hal.command_buffer
}
util.return %1 : !hal.command_buffer
}
util.func public @memoize_command_buffer() -> !hal.command_buffer {
%device = util.global.load immutable @device : !hal.device
%c-1_i64 = arith.constant -1 : i64
%executable = util.global.load immutable @executable : !hal.executable
%global_resource = util.global.load immutable @global_resource : !hal.buffer
%0 = util.call @__memoize_command_buffer_memoize_lookup(%device, %c-1_i64) : (!hal.device, i64) -> !hal.command_buffer
util.return %0 : !hal.command_buffer
}
```IREE (Intermediate Representation Execution Environment, pronounced as “eerie”) is an MLIR-based end-to-end compiler and runtime that lowers Machine Learning (ML) models to a unified IR that scales up to meet the needs of the datacenter and down to satisfy the constraints and special considerations of mobile and edge deployments.
See our website for project details, user guides, and instructions on building from source.
IREE is still in its early phase. We have settled down on the overarching infrastructure and are actively improving various software components as well as project logistics. It is still quite far from ready for everyday use and is made available without any support at the moment. With that said, we welcome any kind of feedback on any communication channels!
See our website for more information.
Community meeting recordings: IREE YouTube channel
IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.