Previously ExpandDims always inserts 1 as the new divisibility, which
makes writing (x * stride)[:, None] far more slower than (x[:, None] *
stride). A better divisibility can be afforded by computing the GCD of
the old dims. Now the two code above are equally fast. E.g. the conv
inductor in pytorch may be faster.
---------
Co-authored-by: Yuheng XIE <thinelephant@gmail.com>
Change the dot to allow taking an initial accumulator and add a flag
that will allow the compiler to accumulate in a lower precision than the
output type.
On Hopper this flag is on by default which allows accumualting with
lower precision.
This only affect Hopper fp8 dot.
The initial code merge of Nvidia Hopper features support. Please be
aware that the code merge is not finished yet and the trouble-shooting
is still ongoing. The new hardware features (GMMA, TMA, STMATRIX etc.)
and automatic warp-specialization are experimental for now and turned
off by default. It is recommended for a trial when version 3.0 is
released.
The work is contributed by:
ben-zhang-609, bealwang, donproc, qliu93, jsh20, allatit23, LyricZhao,
ivanyinwz, goostavz & yangjunpro
from Nvidia, in cooperation with:
ptillet, Jokeren, ThomasRaoux & zahimoud
from OpenAI.
Co-authored-by: Goostav Zhu <gzhu@nvidia.com>
Also catch out-of-bounds indices at constructio and throw a proper error
in the frontend.
Finally, let's make the IR a bit prettier:
%0 = tt.get_program_id {axis = 0 : i32} : i32
becomes:
%0 = tt.get_program_id x : i32
Fixes#1718
# Introducing the `noinline` Parameter for Triton JIT Decorator
We're excited to introduce a new parameter, `noinline`, that can be
added to the `jit` decorator in Triton. This parameter allows developers
to specify that a particular Triton function should not be inlined into
its callers. In this post, we'll dive into the syntax, purpose, and
implementation details of this new feature.
## Syntax
To use the `noinline` parameter, simply add `noinline=True` to the `jit`
decorator for the function that you don't want to be inlined. Here's an
example:
```python
@triton.jit(noinline=True)
def device_fn(x, y, Z):
z = x + y
tl.store(Z, z)
def test_noinline():
@triton.jit
def kernel(X, Y, Z):
x = tl.load(X)
y = tl.load(Y)
device_fn(x, y, Z)
```
In this example, the `device_fn` function is decorated with
`@triton.jit(noinline=True)`, indicating that it should not be inlined
into its caller, `kernel`.
## Purpose
The `noinline` parameter serves several key purposes:
- Reducing code size: By preventing inlining, we can reduce the size of
the compiled code.
- Facilitating debugging: Keeping functions separate can make it easier
to debug the code.
- Avoiding common subexpression elimination (CSE) in certain cases: CSE
can sometimes be avoided by using the `noinline` parameter to reduce
register pressure.
- Enabling dynamic linking: This parameter makes it possible to
dynamically link Triton functions.
## Implementation
The implementation of the `noinline` parameter involves significant
changes to three analysis modules in Triton: *Allocation*, *Membar*, and
*AxisInfo*. Prior to this update, these modules assumed that all Triton
functions had been inlined into the root kernel function. With the
introduction of non-inlined functions, we've had to rework these
assumptions and make corresponding changes to the analyses.
### Call Graph and Limitations
<div style="text-align: center;">
<img
src="https://user-images.githubusercontent.com/2306281/234663904-12864247-3412-4405-987b-6991cdf053bb.png"
alt="figure 1" width="200" height="auto">
</div>
To address the changes, we build a call graph and perform all the
analyses on the call graph instead of a single function. The call graph
is constructed by traversing the call edges and storing them in an edge
map. Roots are extracted by checking nodes with no incoming edges.
The call graph has certain limitations:
- It does not support recursive function calls, although this could be
implemented in the future.
- It does not support dynamic function calls, where the function name is
unknown at compilation time.
### Allocation
<div style="text-align: center;">
<img
src="https://user-images.githubusercontent.com/2306281/234665110-bf6a2660-06fb-4648-85dc-16429439e72d.png"
alt="figure 2" width="400" height="auto">
</div>
In Triton, shared memory allocation is achieved through two operations:
`triton_gpu.convert_layout` and `triton_gpu.alloc_tensor`. The
`convert_layout` operation allocates an internal tensor, which we refer
to as a *scratch* buffer, while the `alloc_tensor` operation returns an
allocated tensor and is thus known as an *explicit* buffer.
To accommodate the introduction of function calls, we are introducing a
third type of buffer called a *virtual* buffer. Similar to scratch
buffers, virtual buffers are allocated internally within the scope of a
function call, and the buffers allocated by the called functions remain
invisible to subsequent operations in the calling function. However,
virtual buffers are distinct from scratch buffers in that the call
operation itself does not allocate memory—instead, it specifies the
total amount of memory required by all the child functions being called.
The actual allocation of buffers is performed by individual operations
within these child functions. For example, when invoking edge e1, no
memory is allocated, but the total amount of memory needed by function B
is reserved. Notably, the amount of shared memory used by function B
remains fixed across its call sites due to the consideration of dynamic
control flows within each function.
An additional challenge to address is the calculation of shared memory
offsets for functions within a call graph. While we can assume a shared
memory offset starting at 0 for a single root function, this is not the
case with a call graph, where we must determine each function's starting
offset based on the call path. Although each function has a fixed memory
consumption, the starting offset may vary. For instance, in Figure 2,
the starting offset of function C through edges e1->e2 differs from that
through edges e2->e4. To handle this, we accumulate the starting offset
at each call site and pass it as an argument to the called function.
Additionally, we amend both the function declaration and call sites by
appending an offset variable.
### Membar
<div style="text-align: center;">
<img
src="https://user-images.githubusercontent.com/2306281/234665157-844dd66f-5028-4ef3-bca2-4ca74b8f969d.png"
alt="figure 3" width="300" height="auto">
</div>
The membar pass is dependent on the allocation analysis. Once the offset
and size of each buffer are known, we conduct a post-order traversal of
the call graph and analyze each function on an individual basis. Unlike
previous analyses, we now return buffers that remain unsynchronized at
the end of functions, allowing the calling function to perform
synchronization in cases of overlap.
### AxisInfo
<div style="text-align: center;">
<img
src="https://user-images.githubusercontent.com/2306281/234665183-790a11ac-0ba1-47e1-98b1-e356220405a3.png"
alt="figure 4" width="400" height="auto">
</div>
The AxisInfo analysis operates differently from both membar and
allocation, as it traverses the call graph in topological order. This is
necessary because function arguments may contain axis information that
will be utilized by callee functions. As we do not implement
optimizations like function cloning, each function has a single code
base, and the axis information for an argument is determined as a
conservative result of all axis information passed by the calling
functions.
---------
Co-authored-by: Philippe Tillet <phil@openai.com>
MLIR current only supports a custom inlining interface per dialect, so
we cannot change the inlining decision of `func.func`.
https://discourse.llvm.org/t/avoid-inlining-some-functions-using-the-func-dialect/69830/3
Could revert it back once they've designed a better inliner interface.
Inlining attributes will be implemented in the next PR since this PR is
already huge.
* Frontend:
- `int` kernel arguments are always signed
- Loop induction variable is now determine by integer promotion on
lb/ub/step
* Optimizer:
- Added new ExtractSliceOp that enforces 32-bit offsets
* Backend:
- Use 64-bit indices when lowering functions and control flow
- Removed `idx_val` macro and replaced it with `i32_val`
- Cleaned up comments
- Added new ArithToIndex pass to make sure operations on indices are
done with the `index` dialect, that gets converted to LLVM separately
using a 64-bit target
- Rewrite the AxisInfo analysis to handle each op case by case.
- Add bit shift, min max, div/rem, and select ops to AxisInfo.
- Rematerialize across load/store ops in the following two cases:
- A size 1 tensor is considered not expensive since all threads will
load the same
- the targeEncoding may expose more vectorization opportunities (more
elements per thread on the first dim)
**_res2next_** benchmark GPU Kernel time comparison on A100.
- Average kernel sum. Triton 16838630ns vs Triton-MLIR 17105166ns.
**1.016x slowdown**.
- Total kernel sum. Triton 6511735460ns vs Triton-MLIR 6512370620ns.
This PR adds a couple of optimization passes that should substantially
improve the performance of Triton on fused attention kernels:
- DecomposeConversionsPass: This decomposes some instructions of the
form `convert_layout` into
- ReorderInstructions: this reorders instructions in a way that is more
amenable to good code generation from `ptxas`.
