-
Notifications
You must be signed in to change notification settings - Fork 983
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
rfc: graph: support accumulation mode
- Loading branch information
Showing
1 changed file
with
319 additions
and
0 deletions.
There are no files selected for viewing
319 changes: 319 additions & 0 deletions
319
rfcs/20240228-graph-api-support-accumulation-mode/README.md
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,319 @@ | ||
# Graph API: Add Accumulation Mode Support in oneDNN Graph | ||
|
||
## Introduction | ||
|
||
Please refer to the [dcoumentation](https://oneapi-src.github.io/oneDNN/dev_guide_attributes_accumulation_mode.html#doxid-dev-guide-attributes-accumulation-mode) | ||
and [rfc](https://github.com/mgouicem/oneDNN/blob/mgouicem/rfcs/f16-accumulation/rfcs/20230118-f16-accumulation/README.md) for the motivation and design detail of primitive API. | ||
|
||
Currently, oneDNN uses f32 for floating point computation and s32 for integer | ||
computation as the default accumulation data types. However, on some platforms, | ||
using smaller accumulation data types can result in additional speed | ||
improvements. | ||
|
||
By introducing support for the accumulation data type, oneDNN can achieve up to | ||
a 2x speedup while maintaining a high level of accuracy for f16 inference. It is | ||
important to note that f16 accumulation is not suitable for training, which | ||
requires full f32 precision. | ||
|
||
This document proposes the corresponding accumulation mode API for oneDNN Graph. | ||
To provide control granularity and ease of use, the API offers four options: | ||
global setting, graph-level attribute, partition-level attribute, and op-level | ||
attribute. | ||
|
||
## Accumulation mode definition | ||
|
||
oneDNN graph will reuse the C and C++ API for accumulation mode enumerations | ||
defined by primitive API: | ||
|
||
C API: | ||
|
||
```c | ||
typedef enum { | ||
dnnl_accumulation_mode_strict, | ||
dnnl_accumulation_mode_relaxed, | ||
dnnl_accumulation_mode_any, | ||
dnnl_accumulation_mode_s32, | ||
dnnl_accumulation_mode_f32, | ||
dnnl_accumulation_mode_f16, | ||
} dnnl_accumulation_mode_t; | ||
``` | ||
|
||
C++ API: | ||
|
||
```cpp | ||
enum class accumulation_mode { | ||
strict = dnnl_accumulation_mode_strict, | ||
relaxed = dnnl_accumulation_mode_relaxed, | ||
any = dnnl_accumulation_mode_any, | ||
s32 = dnnl_accumulation_mode_s32, | ||
f32 = dnnl_accumulation_mode_f32, | ||
f16 = dnnl_accumulation_mode_f16, | ||
} | ||
``` | ||
The accumulation mode attribute accepts: | ||
- `strict` (default): For floating-point datatypes, the default accumulation | ||
datatype is `f32`(or f64 for f64 primitives). For integral datatypes, the | ||
default accumulation datatype is `s32`. | ||
- `relaxed`: Same as strict except some partial accumulators can be | ||
rounded to the src/dst datatype in memory. | ||
- `any`: Uses the fastest implementation available with one of the | ||
src/dst datatypes or a higher precision accumulation datatype. | ||
- `f32`, `f16` and `s32`: Uses the specified accumulation datatype. | ||
Framework users may need map their own accumulation mode definitions to the | ||
above enumerations. | ||
## Proposal | ||
### Option 1: Specify accumulation mode thorugh global function | ||
Use a global function to set the accumulation mode, which will affect all the | ||
created graphs. Users should set the accumulation mode at the graph build stage | ||
before compilation. | ||
The C API will be like: | ||
```c | ||
status_t DNNL_GRAPH_API | ||
dnnl_graph_set_accumulation_mode(dnnl_accumulation_mode_t mode); | ||
dnnl_accumulation_mode_t DNNL_GRAPH_API | ||
dnnl_graph_get_accumulation_mode(); | ||
``` | ||
|
||
The correspodning C++ API will be like: | ||
|
||
```cpp | ||
namespace dnnl { | ||
namespace graph { | ||
void set_accumulation_mode(accumulation_mode mode); | ||
} | ||
} | ||
``` | ||
#### Pros and Cons | ||
Pros: | ||
- No change on all existing APIs | ||
Cons: | ||
- Only provide coarse-grained control on all graphs | ||
- Need efforts for managing global status. | ||
- Require users to provide essential information at the early graph build stage | ||
### Option 2: Support accumulation mode on graph level( Recommended ) | ||
All the computations or operations in a graph will be specified with the same | ||
accumulation data type. Whether the given accumulation mode can be utilized is | ||
decided by the backend capbility. | ||
This option refers to the current solution for floating-point match mode | ||
attribute support in oneDNN Graph. These attributes share some similarities to | ||
a certain extent: both provide a certain degree of speedup by modifying the | ||
datatype based on backend capabilities. | ||
Since we already have the `fpmath-mode` attribute, in order to make the API | ||
simple and scalable, we will need to introduce a new API, `graph config`, to | ||
wrap all the attributes for graph the constructor. Users can now create a | ||
`config` with specific attributes or set attribute values later. This new | ||
design maintains compatibility with current APIs while offers improved | ||
flexibility. | ||
The API will be like: | ||
```cpp | ||
dnnl::graph::graph::config cfg; | ||
cfg.set_engine_kind(dnnl::engine::kind::cpu); | ||
cfg.set_fpmath_mode(dnnl::fpmath_mode::strict); | ||
cfg.set_accumulation_mode(dnnl::accumulation_mode::strict); | ||
graph g( dnnl::engine::kind::cpu, cfg); | ||
op foo(id, kind, "foo"); g.add_op(foo); | ||
op bar(id, kind, "bar"); g.add_op(bar); | ||
partitions p = g.get_partitions(); | ||
// compile and generate kernel according to the accumulation mode setting on | ||
// graph all the partitions from the same graph should share the same math mode | ||
compiled_partition cp0 = p[0].compile(inputs, outputs, engine); | ||
compiled_partition cp1 = p[1].compile(inputs, outputs, engine); | ||
cp0.execute(…); | ||
cp1.execute(…); | ||
``` | ||
|
||
The current graph ctor API will remain unchanged but might be deprecated in the | ||
future, hence it's recommended to use `graph config` to create a graph object | ||
if the attributes are needed. | ||
|
||
#### Pros and Cons | ||
|
||
Pros: | ||
|
||
1. No changes are needed for op-related and partition-related operations, such | ||
as op creation and compilation. | ||
2. Graph may provide different fusion strategies based on the given accumulation | ||
mode according to backend capabilities. | ||
3. Aligned with the existing similar math mode attribute, it might be easier for | ||
users to handle them together. | ||
|
||
Cons: | ||
|
||
1. Relatively coarse-grained control, no support for partition-level and | ||
op-level attribute setting. | ||
2. All ops in the graph will be set the same accumulation mode, yet in practice | ||
it's possible that only a subset of the subgraphs will use f16 accumulation | ||
( e.g. the first layer in the NN might not use this attribute ). Users should | ||
handle such case by themselves such as creating multiple graphs. | ||
3. Require users to provide essential information at the early graph build stage | ||
4. For multi-thread compilation, cannot specify accumulation mode for each | ||
thread. | ||
|
||
### Option 3: Support accumulation mode on partition level | ||
|
||
Users will be asked to provide the accumulation mode during compilation stage. | ||
In this case, users may need to set the accumulation mode for different | ||
partitions separately. | ||
|
||
The API will be like: | ||
|
||
```cpp | ||
graph g(kind); | ||
op foo(id, kind, "foo"); g.add_op(foo); | ||
op bar(id, kind, "bar"); g.add_op(bar); | ||
partitions p = g.get_partitions(); | ||
// compile and generate kernel according to the accumulation mode setting | ||
// here different partition may be compiled with different accumulation | ||
// mode | ||
compiled_partition cp0 = p[0].compile(inputs, outputs, engine, accumulation_mode0); | ||
cp0.execute(…); | ||
|
||
compiled_partition cp1 = p[1].compile(inputs, outputs, engine, accumulation_mode1); | ||
cp1.execute(…); | ||
``` | ||
#### Pros and Cons | ||
Pros: | ||
1. Not require users to provide the accumulation mode information until | ||
compilation stage. | ||
2. Users may specify different accumulation mode for different backends. | ||
3. Users may compile the same partition with different accumulation mode. | ||
4. Scalable for multi-threading cases. | ||
Cons: | ||
1. If a subset of ops in the subgraph are not required to use f16 accumulation | ||
mode( e.g. imagine only the first matmul in the sdp fusion uses f16 | ||
accumulation), users need to handle such case by themselves. | ||
2. The library need to ensure the partition is acceptable on the given | ||
accumulation mode | ||
### Option 4: Support accumulation mode on op level | ||
Users will be asked to set the accumulation mode as an attribute when creating | ||
an op. The API will be like: | ||
```cpp | ||
graph g(kind); | ||
std::string acc_mode1{"strict"}; | ||
std::string acc_mode2{"relax"}; | ||
// set accumulation_mode as an attribute | ||
op foo(id, kind, "foo"); foo.set_attr<std::string>(op_attr::accumulation_mode, acc_mode1); | ||
op bar(id, kind, "bar"); bar.set_attr<std::string>(op_attr::accumulation_mode, acc_mode2); | ||
g.add_op(foo); | ||
g.add_op(bar); | ||
// partitioning algorithm needs to respect accumulation mode on the | ||
// operators. may not fuse if two operators have different accumulation | ||
// modes | ||
partitions p = g.get_partitions(); | ||
// compile and generate kernel according to the accumulation mode | ||
// setting on op. different partition may have different accumulation mode | ||
compiled_partition cp0 = p[0].compile(inputs, outputs, engine); | ||
cp0.execute(…); | ||
compiled_partition cp1 = p[1].compile(inputs, outputs, engine); | ||
cp1.execute(…); | ||
``` | ||
|
||
#### Pros and Cons | ||
|
||
Pros: | ||
|
||
1. Most fine-grained control over the operations. | ||
2. If part of the ops in the subgraph are not required to set f16 accumulation | ||
mode, the library will be able to do pattern matching accordingly. | ||
|
||
Cons: | ||
|
||
1. Huge code change on operation API and integration. | ||
2. Might introduce more complex logic for fusion. | ||
|
||
## Verbose Support | ||
|
||
The specified accumulation mode will be printed in the verbose in the manner | ||
like `accm:f16`. | ||
|
||
## Validation | ||
|
||
Currently benchdnn has already supported setting accumulation mode for | ||
validation, such as: | ||
|
||
```cpp | ||
--attr-acc-mode=MODE | ||
``` | ||
|
||
benchdnn graph should be able to specify the accumulation mode in the JSON file | ||
based on the implementation. For instance, if the accumulation mode is added as | ||
a graph attribute, it can be validated like: | ||
|
||
```JSON | ||
{ | ||
"version": "3.0.0", | ||
"engine_kind": "cpu", | ||
"fpmath_mode": "strict", | ||
"accumulation_mode": "strict", | ||
"graph": [ | ||
//... | ||
] | ||
} | ||
``` | ||
|
||
If the accumulation mode is enabled through a global API or partition-level | ||
API, benchdnn graph needs to be updated to accommodate it. Additionally, the | ||
validation input should be similar with the requirements of the graph-level | ||
API. This will enable benchdnn graph to parse the JSON file and configure the | ||
accumulation mode accordingly. | ||
|
||
If the accumulation mode is added as an operation attribute, it might be | ||
validated like: | ||
|
||
```JSON | ||
{ | ||
"id": 0, | ||
"name": "Convolution", | ||
"kind": "Convolution", | ||
"attrs": { | ||
"strides": { | ||
"type": "s64[]", | ||
"value": [ | ||
2, | ||
2 | ||
] | ||
}, | ||
"accumulation_mode": { | ||
"type": "string", | ||
"value": "strict" | ||
}, | ||
//... | ||
} | ||
//... | ||
} | ||
``` | ||
|
||
END |