gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

History prior to this date belonged to the now deprecated ISAAC project, and was deleted to save space

Browse Source
This commit is contained in:
Philippe Tillet
2021-07-27 12:38:38 -07:00
commit 6d7cf35123
202 changed files with 94034 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

243
docs/pytriton.md Normal file
View File

@@ -0,0 +1,243 @@
# The PyTriton API
## <span style="color:darkred"> Table of Contents </span>
1. [Motivations](#motivations)
2. [Triton Functions](#pytriton-function)
1. [Creation of Triton Kernels](#creation-triton-kernels)
2. [Usage of Triton Kernels](#usage-triton-kernels)
3. [Integration with Automatic Differentiation](#autodiff)
1. [Basics](#autodiff-basics)
2. [Convenience](#autodiff-convenience)
## <span style="color:darkred"> Motivations </span> <a name="motivations"></a>
The purpose of PyTriton is to provide an API for easily executing Triton-C kernels from PyTorch and Tensorflow. One of the main advantages of PyTriton is that it is framework agnostic: any custom op written using this API will be transparently compatible with both Tensorflow and PyTorch without any additional effort required, as will be shown in this tutorial.
Consider for example the following piece of code:
```python
import numpy as np
import triton
def run_tf():
M, N, K = 128, 128, 128
a = tf.placeholder(tf.float32, shape=[M, K])
b = tf.placeholder(tf.float32, shape=[N, K])
c = triton.ops.dot(a, b, transpose_a = False, transpose_b = True)
da, db = tf.gradients(c, [a, b])
# Run
ha = np.random.rand(M, K).astype(np.float32)
hb = np.random.rand(K, N).astype(np.float32)
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
result = sess.run([da], feed_dict = {a: ha, b: hb})
def run_torch():
M, N, K = 128, 128, 128
a = torch.randn(M, K).cuda()
b = torch.randn(K, N).cuda()
a.requires_grad_(True)
b.requires_grad_(True)
c = triton.ops.dot(a, b, False, True)
c.backward()
da = a.grad.clone()
db = b.grad.clone()
## Run on tensorflow
# import tensorflow as tf
# run_tf()
## Run on pytorch
# import torch
# run_torch()
```
PyTriton works by detecting which frameworks are imported and automatically generating and just-in-time compiling C++ binding code for them. Specifically, the following chain of events is triggered when a Triton operation is executed:
1. The imported frameworks are detected
2. C++ binding code for Tensorflow or PyTorch is generated, compiled and cached.
3. The corresponding custom-op is automatically loaded from the generated .so file, and a framework-agnostic wrapper is created.
4. The wrapper is called and a tf.tensor or a torch.tensor is returned. In the case of Tensorflow, the gradient is also registered at this point if applicable
The remainder of this tutorial will show you how to re-implement the above `triton.ops.dot` operation from scratch.
## <span style="color:darkred"> PyTriton Functions </span> <a name="pytriton-function"></a>
The PyTriton API provides a `triton.function` class which automatically handles the interaction with automatic differentiation in whichever framework was detected. Therefore, every differentiable custom operation written with PyTriton should inherit from this class
```python
import triton
# Entry point
class _dot(triton.function):
@staticmethod
# Forward Pass
def forward(ctx, *args):
#...
@staticmethod
# Backward Pass
def backward(ctx, dy):
#...
```
### <span style="color:darkblue">Creation of Triton Kernels </span> <a name="creation-triton-kernel"></a>
PyTriton also provides a `triton.kernel` class which automatically takes care of interaction with the Triton-JIT as well as the generation and compilation of C++ framework bindings code. For our dot operation we create a kernel from the Triton-C code derived at the end of the [previous tutorial](https://github.com/ptillet/triton/blob/master/docs/triton-c.md)
```
src = """
__global__ void dot(TYPE * A, TYPE * B, TYPE * C,
int M, int N, int K,
int lda __multipleof(8), int ldb __multipleof(8), int ldc __multipleof(8)) {
// prologue
int pm = get_program_id(0);
int pn = get_program_id(1);
int rm[TM] = pm * TM + 0 ... TM;
int rn[TN] = pn * TN + 0 ... TN;
int rk[TK] = 0 ... TK;
float c[TM, TN] = 0;
// pointers to operands
TYPE* pa[SHAPE_A] = A + rk[BROADCAST_AK] * STRIDE_AK + rm[BROADCAST_AM] * STRIDE_AM;
TYPE* pb[SHAPE_B] = B + rk[BROADCAST_BK] * STRIDE_BK + rn[BROADCAST_BN] * STRIDE_BN;
// prefetches operands
TYPE a[SHAPE_A] = (*pa);
TYPE b[SHAPE_B] = (*pb);
// reduction loop
for(int k = K; k > 0; k-= TK){
c += USE_A @ USE_B;
pa = pa + TK * STRIDE_AK;
pb = pb + TK * STRIDE_BK;
a = *pa;
b = *pb;
}
// epilogue
int rcm[TM] = pm * TM + 0 ... TM;
int rcn[TN] = pn * TN + 0 ... TN;
TYPE* pc[TM, TN] = C + rcn[newaxis, :] + rcm[:, newaxis] * ldc;
*pc = c;
}
}
"""
kernel = triton.kernel(src, ['C'])
```
Note that the second argument to `triton.kernel` constructors indicates which of the operands our kernel function should return. Here, we only return `C`.
At this point, `kernel` is a callable object which takes the same signature as the `dot` function in our source code, except that pointers are treated as tensors:
```
[tensor, tensor, tensor, int, int, int, int, int, int]
```
### <span style="color:darkblue">Usage of Triton Kernels </span> <a name="usage-triton-kernels"></a>
However, in practice only A, B are provided by the user, and all the other `int` arguments should be derived from these operands only. Hence, we create a helper function that extracts shapes from the `A` and `B` tensors, and then returns the results of a call to `kernel`:
```python
@staticmethod
def _call(a, b, transpose_a, transpose_b):
# extract shapes
shape_a = triton.shape(a)
shape_b = triton.shape(b)
M, Ka = shape_a[0], shape_a[1]
Kb, N = shape_b[0], shape_b[1]
# transpose shapes
if transpose_a:
M, Ka = Ka, M
if transpose_b:
Kb, N = N, Kb
# contiguous dimensions
lda = M if transpose_a else Ka
ldb = Kb if transpose_b else N
ldc = N
# data-type
dtype = a.dtype
# allocate output
c = triton.empty([M, N], dtype = dtype)
# compute
grid = lambda opt: [triton.cdiv(M, opt.d('TM')), triton.cdiv(N, opt.d('TN'))]
# macros -- not necessary but makes kernel source-code simpler
macros = {# handle A transposition
'USE_A' : '^a' if transpose_a else 'a',
'STRIDE_AK' : 'lda' if transpose_a else '1',
'STRIDE_AM' : '1' if transpose_a else 'lda',
'BROADCAST_AK': ':, newaxis' if transpose_a else 'newaxis, :',
'BROADCAST_AM': 'newaxis, :' if transpose_a else ':, newaxis',
'SHAPE_A' : 'TK, TM' if transpose_a else 'TM, TK',
# handle B transposition
'USE_B' : '^b' if transpose_b else 'b',
'STRIDE_BK' : '1' if transpose_b else 'ldb',
'STRIDE_BN' : 'ldb' if transpose_b else '1',
'BROADCAST_BK': 'newaxis, :' if transpose_b else ':, newaxis',
'BROADCAST_BN': ':, newaxis' if transpose_b else 'newaxis, :',
'SHAPE_B' : 'TN, TK' if transpose_b else 'TK, TN'}
return _dot.kernel(a, b, c, M, N, Ka, lda, ldb, ldc, grid,
AT = transpose_a, BT = transpose_b, TYPE = dtype,
TM = [32, 64, 128], TN = [32, 64, 128], TK = [8], **macros)
```
While this code should be mostly self-explanatory, there are a few of noteworthy things worth pointing out
- `triton.shape` provides a framework-agnostic way to retrieve the shape of a tensor
- `triton.empty` creates an empty tensor of the specified dimensions
- `grid` corresponds to the grid with which our Triton kernel will be launched. Because in our case this grid depends on parametric tile variables, it is supplied as a function of compilation options `opt`, whose compile-time definition can be retrieved using `opt.d(name)`. Here, `opt.d('TM')` and `opt.d('TN')` retrieve the first and second tile dimension our kernel was compiled with. We also provide a helper `triton.cdiv` for ceil divisions.
- `macros` provides a list of preprocessor definitions to compile the kernel with. Alternatively, these can also be supplied as named argument to the `_dot.kernel`. We recall that lists can be supplied to the preprocessor, in which case an auto-tuning procedure will be triggered. Here, the value of `TM` and `TN` are both tuned between 32, 64 and 128.
## <span style="color:darkred"> Compatibility with Automatic Differentiation</span> <a name="autodiff"></a>
At this point, our custom operation only takes two tensor arguments and transposition information, which is good. However, it is still not compatible with PyTorch's or TensorFlow's automatic differentiation engine, and a small amount of additional effort is needed.
### <span style="color:darkblue"> Basics </span> <a name="autodiff-basics"></a>
PyTriton binds to Tensorflow's and PyTorch's automatic differentiation framework using a single, common API inspired by PyTorch. It consists of two static methods `forward` and `backward` that take a context as their first input:
```
@staticmethod
def forward(ctx, a, b, transpose_a = False, transpose_b = False):
ctx.save_for_backward(a, b)
ctx.t_a = transpose_a
ctx.t_b = transpose_b
return _dot._call(a, b, transpose_a, transpose_b)
@staticmethod
def backward(ctx, dy):
a, b = ctx.saved_tensors
t_a, t_b = ctx.t_a, ctx.t_b
if not t_a and not t_b:
da = _dot._call(dy, b, False, True)
db = _dot._call(a, dy, True, False)
elif not t_a and t_b:
da = _dot._call(dy, b, False, False)
db = _dot._call(dy, a, True, False)
elif t_a and not t_b:
da = _dot._call(b, dy, False, True)
db = _dot._call(a, dy, False, False)
elif t_a and t_b:
da = _dot._call(b, dy, True, True)
db = _dot._call(dy, a, True, True)
else:
assert False
return da, db, None, None, None, None, None, None, None
```
### <span style="color:darkblue">Convenience </span> <a name="autodiff-convenience"></a>
Still like for PyTorch, a callable operation can be created using the `apply` method of our `triton.function` class. We wrap it as a module variable for convenience:
```python
dot = _dot.apply
```
And that's it! Our custom op is now created and ready to be used with both PyTorch and Tensorflow.

436
docs/triton-c.md Normal file
View File

@@ -0,0 +1,436 @@
# The Triton-C Programming Language
## <span style="color:darkred"> Table of Contents </span>
1. [Motivations](#motivations)
2. [Vector Addition](#vector-addition)
1. [Differences with CUDA](#differences-with-cuda)
2. [Advantages over CUDA](#advantages-over-cuda)
1. [Vectorization](#vectorization)
2. [Parameterization](#parameterization)
3. [Auto-Tuning](#auto-tuning)
3. [Matrix Transposition](#matrix-transposition)
1. [Compute Kernel](#trans-compute-kernel)
2. [The __multipleof Attribute](#trans-multipleof)
3. [Conditional Dereferencing](#conditional-dereferencing)
4. [Matrix Multiplication](#matrix-multiplication)
1. [Compute Kernel](#matmul-compute-kernel)
2. [Optimizations](#optimizations)
1. [Pre-Fetching](#pre-fetching)
1. [Rematerialization](#rematerialization)
3. [Fused Transpositions and Auto-Tuning](#fused-trans-autotuning)
## <span style="color:darkred"> Motivations </span> <a name="motivations"></a>
In C and C++, arrays and pointers have similar semantics. Indeed, there is no native way to manipulate statically shaped multi-dimensional arrays (beyond initialization) as a whole:
```c
// C99
float x[16][8] = {3.14};
float y[16][8] = {5.17};
// z = x + y
float z[16][8];
#pragma unroll
for(int i = 0; i < 16; i++)
#pragma unroll
for(int j = 0; j < 8; j++)
z[i][j] = x[i][j] + y[i][j];
```
While it does not seem like a big deal at first sight, there are two issues with this:
- **Ergonomics**: Of course, it is possible to simplify the above code using functions in C
```
float z[16][8];
add(z, x, y, 16, 8);
```
but this would be semantically different as the loops can no longer be unrolled due to their bounds being now dynamic arguments of the add function. This can be mitigated using templates metaprogramming (and operator overloads) in C++:
```c
// C++
template<typename T, int M, int N>
class matrix;
matrix<float, 16, 8> x = {3.14};
matrix<float, 16, 8> y = {5.17};
matrix<float, 16, 8> z = x + y;
```
While this is better and now equivalent to our initial code snippet, the syntax is not quite as ergonomically satisfying as what native syntactic support could provide:
```c
// Triton-C
float x[16, 8] = 3.14;
float y[16, 8] = 5.17;
// float z[8, 8] = x + y; // doesn't compile -- incompatible shapes!
float z[16, 8] = x + y;
float u[16] = z[:, +]; // sum along the second axis
float v[16, 32] = u[:, newaxis]; // broadcasting along the second axis
```
which is valid _Triton-C_.
- **Portability**: One other issue with our initial C program is that it is not portable. While it will run well on a single CPU thread, the operation `z = x + y` would underutilize a GPU Streaming Processor as it would execute on a single thread only. For this reason, it would have to be rewritten in CUDA as follows:
```
// CUDA
// Launch on a block of 16 x 8 threads
float x = 3.14;
float y = 5.17;
float z = x + y
```
In Triton-C, the same code can be used across many different platforms (only CPUs and GPUs are supported at the moment). Furthermore, Triton-C is single-threaded, hence easier to write than CUDA.
- **Performance**: Another issue with our initial C code snippet is its performance. Although the loops are unrolled, the program does not carry any data-flow information pertaining to array operations. This issue gets more and more problematic as programs get increasingly complex, eventually culminating in matrix multiplication being remarkably hard to optimize.
This can be worked around using heavy metaprogramming techniques (see [CUTLASS](https://github.com/NVIDIA/cutlass)), but even then programmers still have to allocate and synchronize shared memory manually and endure prohibitively long compilation procedures not easily amenable to auto-tuning. For these reasons, most Deep-Learning frameworks still rely heavily on highly optimized subroutines (e.g., BLAS), which makes the development of novel custom primitives time-consuming for experts and almost impossible for others.
Triton addresses this issue by relying on **Triton-IR**, an LLVM-like IR for array operations, and **Triton-JIT**, an optimizing compiler for Triton-IR. These two systems are, however, beyond the scope of this tutorial. More information can be found [here](http://www.eecs.harvard.edu/~htk/publication/2019-mapl-tillet-kung-cox.pdf).
_Note: You might be thinking that this is exactly what [MLIR](https://github.com/tensorflow/mlir) was made for... and you're right! You can conceptually think of Triton-IR as a dialect for MLIR, and Triton-C as a frontend for it. I would like to integrate Triton-IR into MLIR in the future; If you're interested in making this a thing, let me know._
## <span style="color:darkred"> Vector Addition </span> <a name="vector-addition"></a>
### <span style="color:darkblue"> Differences with CUDA </span> <a name="differences-with-cuda"></a>
Let's start it off by looking at a simple example. Vector addition, in its most trivial Triton-C implementation, can be written as follows:
```c
// Triton-C
// launched on a grid of (N / 32) programs of 1 thread each
__global__ void add(int N, float *a, float *b, float* c) {
int id = get_program_id(0);
int off[32] = id * 32 + (0 ... 32)
*(c + off) = *(a + off) + *(b + off);
}
```
For reference, here is an equivalent CUDA kernel (NVCC will generate the same PTX code as Triton-JIT on the above code):
```c
// CUDA
// launched on a grid of (N / 32) programs of 32 threads each
__global__ void add(int N, float *a, float *b, float *c) {
int off = blockIdx.x * 32 + threadIdx.x;
c[off] = a[off] + b[off];
}
```
As you can see, there are three main differences between our Triton-C kernel and the equivalent CUDA:
- **The programming model is different**.
While Triton-C and CUDA both use a Single-Program, Multiple-Data (SPMD) programming model, each Triton-C kernel is single-threaded.
Therefore, `get_program_id({0, 1, 2})` is equivalent to `blockIdx.{x, y, z}`, but there is no such thing as `blockDim` and `threadIdx`.
- **The semantics of arrays is different**
In the above Triton-C kernel, `off` is an array of 32 consecutive integers: `int off[32] = {id * 32 + 0, id * 32 + 1, ..., id * 32 + 31}`.
As a result, the statement: `c + off` implicitly broadcast `c` and creates an array of 32 pointers. This could also be done explicitly as follows:
```
float* c_broadcast[32] = c;
float* c_ptr[32] = c_broadcast + off; // c_ptr = c + off
```
- **The semantics of the subscript operator is different**.
n C/CUDA, subscripting can be used to offset and dereference a pointer, but in Triton-C it can only be used to index and broadcast an array (think NumPy).
### <span style="color:darkblue"> Advantages over CUDA </span> <a name="advantages-over-cuda"></a>
At this point, the advantages of Triton-C over CUDA may not be obvious. But they should become clearer and clearer as this tutorial progresses. First and foremost, the purpose of this subsection is to show how Triton can be used to optimize vector additions by automatically taking care of load/store vectorization, code parameterization and auto-tuning -- all of which require nontrivial implementation efforts in CUDA.
#### <span style="color:purple"> Vectorization </span> <a name="vectorization"></a>
On some hardware architectures, vectorizing load/store operations can lead to better memory utilization and, in turn, noticeable performance gains. In general, 128-bit memory transactions are favored, leading to the following CUDA kernel:
```c
// CUDA
// launched on a grid of (N / 128) programs of 32 threads each
__global__ void add(int N, float4 *a, float4 *b, float4 *c) {
int off = blockIdx.x * 32 + threadIdx.x;
c[off] = a[off] + b[off];
}
```
Or, for half-precision inputs:
```c
// CUDA
// launched on a grid of (N / 256) programs of 32 threads each
__global__ void add(int N, half8 *a, half8 *b, half8 *c) {
int off = blockIdx.x * 32 + threadIdx.x;
c[off] = a[off] + b[off];
}
```
Now this is a bit annoying, because as a programmer you have to keep track of not only the ideal vector size for each data-type (which might change in future GPU architectures), but also of how many elements are processed in each thread-block -- and adjust the grid size of the kernel accordingly! Not to mention that you may want to tune the thread-block size as well.
In Triton-C, this is not a problem as the compiler will figure out automatically when and where vectorization should be used, without any change in the source-code necessary.
#### <span style="color:purple"> Parameterization </span> <a name="parameterization"></a>
Specifically, the Triton compiler would refuse to 4-way vectorize our above compute kernel because it would require the array `int off[32]` to be distributed over 8 threads, which is less than a warp. Fortunately, it turns out that this problem can be easily solved using preprocessor directrives to _parameterize_ our kernel:
```c
// Triton-C
// launched on a grid of (N / SIZE) programs of 1 thread each
__global__ void add(int N, TYPE* a, TYPE* b, TYPE* c) {
int id = get_program_id(0);
int off[SIZE] = id * SIZE + (0 ... SIZE);
*(c + off) = *(a + off) + *(b + off);
}
// Not vectorized when compiled with -DSIZE=32 -DTYPE=float
// 4-Vectorized when compiled with -DSIZE=128 -DTYPE=float
// 8-Vectorized when compiled with -DSIZE=256 -DTYPE=half
```
Now, `TYPE` and `SIZE` are preprocessors macros which can be specified at compile-time, thereby giving the Triton compiler enough information to vectorize when beneficial without requiring any additional code modification.
#### <span style="color:purple"> Auto-Tuning </span> <a name="auto-tuning"></a>
As it turns out, different input vector lengths `N` may require different values of `SIZE` to perform optimally. Fortunately, the Triton preprocessor also accepts lists of possible definitions for macros, in which case an auto-tuning procedure will be launched every-time new input sizes are encountered. For example, compiling the above kernel with the option`-DSIZE=[32, 64, 128, 256] -DTYPE=float`
will result in the parameter `SIZE` being automatically tuned every time a new value of `N` is encountered.
_Note: Tuning our reference CUDA kernel would be much more cumbersome, as template metaprogramming would have to be used to ensure that proper vector types would be used_
## <span style="color:darkred"> Matrix Transposition </span> <a name="matrix-transposition"></a>
Transpositions are (relatively) hard to efficiently write in CUDA because naive implementations typically suffer from _uncoalesced_ memory operations when writing back the transposed matrix to DRAM. Of course, this can be fixed by using shared memory as shown [here](https://devblogs.nvidia.com/efficient-matrix-transpose-cuda-cc/), but this comes at the cost of simplicity and -- more importantly -- interferes with auto-tuning.
### <span style="color:darkblue"> Compute Kernel </span> <a name="trans-compute-kernel"></a>
In Triton, however, kernels are single-threaded and the compiler automatically detects if and when data should be temporarily stashed to shared memory in order to enable shared memory stores/loads. Therefore, an optimal Triton kernel for this operation would look like:
```c
// launched on a grid of (M / TM) x (N / TN) programs of 1 thread each
__global__ void transpose(TYPE * X, TYPE * Y, int M, int N, int ldx, int ldy) {
// extract program ID
int pidm = get_program_id(0); //(1)
int pidn = get_program_id(1); //(2)
// create 1D range along the two matrix's axes
int rm[TM] = pidm * TM + 0 ... TM; //(3)
int rn[TN] = pidn * TN + 0 ... TN; //(4)
// create 2D array of pointers
TYPE* px[TM, TN] = X + rm[:, newaxis] + rn[newaxis, :] * ldx; //(5)
TYPE* py[TN, TM] = Y + rm[newaxis, :] * ldy + rn[:, newaxis]; //(6)
// write back using the transposition operator '^'
*py = ^(*px); //(7)
}
```
At a high level, this kernel loads a `TM x TN` tile from the input matrix `X`, transposes it and writes the resulting `TN x TM` tile to the output matrix `Y`. Eventually, transposition of the full input matrix is achieved by launching a grid of `(M / TM) x (N / TN)` programs decomposed as follows:
- Statements (1) and (2) extract the coordinates the program in the above 2D launch grid. For example, the program producing the output tile `Y[TN:2TN-1, 2TN:3TN-1]` holds the values:
```
pidm = 2
pidn = 1
```
- Statements (3) and (4) construct the ranges of indices:
```
rm = [pidm*TM + 0, pidm*TM + 1, ..., pidm*TM + (TM - 1)]
rn = [pidn*TN + 0, pidn*TN + 1, ..., pidn*TN + (TN - 1)]
```
which will be used in statements (5) and (6) to construct tiles of pointers
- Statements (5) constructs the following array of pointers `px` using numpy-style broadcasting semantics:
```
│ X + (pidm*TM + 0) + (pidn*TN + 0)*ldx, ..., ..., X + (pidm*TM + 0) + (pidn*TN + TN - 1)*ldx) │
│ ⋮ ⋮ │
│ ⋮ ⋮ │
│ X + (pidm*TM + TM - 1) + (pidn*TN + 0)*ldx, ..., ..., X + (pidm*TM + TM - 1) + (pidn*TN + TN - 1)*ldx) │
```
- Statement (6) constructs the following array of pointers `py` using numpy-style broadcasting semantics:
```
│ Y + (pidn*TN + 0) + (pidm*TM + 0)*ldy, ..., ..., Y + (pidn*TN + 0) + (pidm*TM + TM - 1)*ldy) │
│ ⋮ ⋮ │
│ ⋮ ⋮ │
│ Y + (pidn*TN + TN - 1) + (pidn*TN + 0)*ldy, ..., ..., Y + (pidn*TN + TN - 1) + (pidm*TM + TM - 1)*ldy) │
```
- Statement (7) element-wise dereferences the above array of pointers `*px`, transposes it using the unary transposition operator `^`, and writes it back at the location specified by `py`.
### <span style="color:darkblue"> The __multipleof Attribute </span> <a name="trans-multipleof"></a>
The memory loads and store in our transposition kernel are not vectorizable by default, since `X + ldx` (and `Y + ldy`) may be misaligned when `ldx` (and `ldy`) are not multiples of e.g., 4. This is unfortunate because tensor dimensions can be easily made into nice powers of two in Deep Learning, due to batch-sizes and layer width being flexible.
For this reason, Triton provides a __multipleof(N) attributes for variables that are guaranteed to always be multiple of N. In the case of Matrix Transpositions, vector loads can be enabled by modifying the function's signature as follows:
```c
__global__ void transpose(TYPE * X, TYPE * Y, int M, int N, int ldx __multipleof(8), int ldy __multipleof(8)) {
// ...
}
```
### <span style="color:darkblue"> Conditional Dereferencing </span> <a name="conditional-dereferencing"></a>
You might have noticed that the above code will fail when `M` and `N` are not multiples of `TM` and `TN` respectively. Fortunately, the above kernel can be slightly modified to handle thie situation, as shown below:
```c
// launched on a grid of ((M + TM - 1) / TM) x ((N + TN - 1) / TN) programs
__global__ void transpose(TYPE * X, TYPE * Y, int M, int N, int ldx, int ldy) {
// ...
// create bounds-checking mask
bool checkx[TM, TN] = (rm[:, newaxis] < M) && (rn[newaxis, :] < N); //(7a)
bool checky[TN, TM] = (rm[newaxis, :] < M) && (rn[:, newaxis] < N); //(7b)
// conditional write-back using the conditional dereferencing operatior '*?()'
*?(checky)py = ^(*?(checkx)px); //(7)
}
```
Here, statements (7a) creates an array of booleans `checkx[TM, TN]` such that `checkx(i, j) = True` if and only if `px(i, j)` should be dereferenced. Statement (7b) does the same for `py`. Both `px` and `py` are then conditionally dereferenced using Triton-C's conditional dereferencing operator `*?(predicate) pointer`.
## <span style="color:darkred"> Matrix Multiplication </span> <a name="matrix-multiplication"></a>
The purpose of this section is to present a Triton-C implementation of matrix multiplication that achieves performance competitive with the best existing hand-written CUDA kernels (see [CUTLASS](https://github.com/NVIDIA/cutlass)). We will also see how pre-processors macros can be leveraged to fuse transposition operations as well as to provide support for auto-tuning and FP16 Tensor Cores.
_Note: Bounds-checking is ommitted throughout for the sake of clarity. This feature can be easily added into our kernel, but may result in a slight performance hit because LLVM and PTXAS have issues dealing with conditionals and predicates inside loops._
### <span style="color:darkblue"> Compute Kernel </span> <a name="matmul-compute-kernel"></a>
Matrix multiplications of the form `C = A x B` can be implemented in Triton-C fairly concisely, as shown below:
```c
// Triton-C
// launched on a grid of (M / TM) x (N / TN) programs
__global__ void dot(TYPE * A, TYPE * B, TYPE * C, int M, int N, int K,
int lda __multipleof(8), int ldb __multipleof(8), int ldc __multipleof(8)) {
// prologue
int pm = get_program_id(0); //(1)
int pn = get_program_id(1); //(2)
int rm[TM] = pm * TM + 0 ... TM; //(3)
int rn[TN] = pn * TN + 0 ... TN; //(4)
int rk[TK] = 0 ... TK; //(5)
// initialize accumulator
float c[TM, TN] = 0; //(6)
// pointers to operands
TYPE* pa[TM, TK] = A + rk[newaxis, :] * 1 + rm[:, newaxis] * lda; //(7)
TYPE* pb[TK, TN] = B + rk[:, newaxis] * ldb + rn[newaxis, :] * 1; //(8)
// reduction loop
for(int k = K; k > 0; k-= TK){
// fetch operands
TYPE a[TM, TK] = *pa; //(9)
TYPE b[TK, TN] = *pb; //(10)
// matrix-multiply accumulate
c += a @ b; //(11)
// increment pointers
pa = pa + TK * 1; //(12)
pb = pb + TK * ldb; //(13)
}
// epilogue
TYPE* pc[TM, TN] = C + rn[newaxis, :] + rm[:, newaxis] * ldc; //(14)
*pc = c; //(15)
}
```
Here, each kernel instance produces a `TM x TN` tile of the output matrix C as follows:
- Statements (1) - (2) fetch the id of the current program instance.
- Statements (3) - (4) construct ranges of indices to process for the vertical and horizontal axes of the output matrix `C`
- Statement (5) constructs a range of indices along the reduction axis: `rk = [0, 1, ..., TK - 1]`
- Statement (6) initialize a `TM x TN` array of accumulators to hold the result of `A[rm, :] x B[:, rn]`
- Statements (7) - (8) initializes arrays of pointers `pa` and `pb` to the operands `A` and `B` using logic similar to that of the above transposition kernel
- Statements (9) - (10) load tiles of operands by dereferencing `pa` and `pb`
- Statement (11) performs updates the accumulator array using Triton-C's matrix multiplication operator '@'
- Statements (12) - (13) updates `pa` and `pb`
- Statement (14) creates an array of pointers `pc` to the result matrix `C`
- Statement (15) writes back the accumulator to `C`
Internally, the Triton compiler will perform quite a few optimizations that will ensure good performance for this kernel:
- Automatic coalescing of load/store operations
- Automatic vectorization of load/store operations
- Stashing `a` and `b` to shared memory
- Automatic allocation of shared memory
- Automatic synchronization of shared memory
- Automatic padding of shared memory to avoid bank conflicts
- Automatic usage of tensor cores when TYPE = half and TK % 4 = 0
### <span style="color:darkblue"> Optimizations </span> <a name="optimizations"></a>
Nonetheless, there are two important optimizations that the Triton compiler does not do automatically at the moment yet are critical to achieve peak performance: pre-fetching and rematerialization. In this subsection we describe how these optimizations can be done manually by modifying the above source-code.
#### <span style="color:purple"> Pre-Fetching </span> <a name="pre-fetching"></a>
The purpose of pre-fetching is to overlap the update of the accumulator `c` with the memory loads for the next tiles that will need to be multiplied. This can be done by modifying the above reduction loop as follows:
```
// pre-fetch operands
TYPE a[TM, TK] = *pa; //(9)
TYPE b[TK, TN] = *pb; //(10)
for(int k = K; k > 0; k-= TK){
c += a @ b;
pa = pa + TK * 1;
pb = pb + TK * ldb;
// don't prefetch last iteration
bool check = k > TK;
// pre-fetch operands
a = check ? *pa : 0;
b = check ? *pb : 0;
}
```
Note that the Triton-C compiler will now also be able to use double-buffering techniques to make sure that the array `a` can be used and updated at the same time without any memory hazard.
#### <span style="color:purple"> Rematerialization </span> <a name="rematerialization"></a>
[Rematerialization](https://en.wikipedia.org/wiki/Rematerialization) is a compiler optimization which consists in recomputing some values instead of storing and reloading them from (register) memory, so as to decrease register pressure in the compute kernel. Although LLVM does this automatically to some extent, it fails to find good heuristics for the above kernel -- thereby requiring some source code modification to achieve optimal performance. Fortunately, only `rm` and `rn` need to be rematerialized, leading to the following epilogue:
```c
// epilogue
int rcm[TM] = pm * TM + 0 ... TM;
int rcn[TN] = pn * TN + 0 ... TN;
TYPE* pc[TM, TN] = C + rcn[newaxis, :] + rcm[:, newaxis] * ldc;
*pc = c;
```
### <span style="color:darkblue"> Fused Transpositions and Auto-Tuning </span> <a name="fused-trans-autotuning"></a>
It is common for optimized matrix-multiplication implementations (e.g., BLAS) to provide variants in which one or both operands are transposed. This is also what is done in the [PyTriton](https://github.com/ptillet/triton/blob/master/python/triton/ops/dot.py) implementation of matrix-multiplication. Fortunately, this can be done by using pre-processors macros for tile shapes and broadcasting directives, leading to the following kernel:
```c
// Triton-C
// launched on a grid of (M / TM) x (N / TN) programs
void dot(TYPE * A, TYPE * B, TYPE * C,
int M, int N, int K,
int lda __multipleof(8), int ldb __multipleof(8), int ldc __multipleof(8)) {
// prologue
int pm = get_program_id(0);
int pn = get_program_id(1);
int rm[TM] = pm * TM + 0 ... TM;
int rn[TN] = pn * TN + 0 ... TN;
int rk[TK] = 0 ... TK;
float c[TM, TN] = 0;
// pointers to operands
TYPE* pa[SHAPE_A] = A + rk[BROADCAST_AK] * STRIDE_AK + rm[BROADCAST_AM] * STRIDE_AM;
TYPE* pb[SHAPE_B] = B + rk[BROADCAST_BK] * STRIDE_BK + rn[BROADCAST_BN] * STRIDE_BN;
// prefetches operands
TYPE a[SHAPE_A] = (*pa);
TYPE b[SHAPE_B] = (*pb);
// reduction loop
for(int k = K; k > 0; k-= TK){
c += USE_A @ USE_B;
pa = pa + TK * STRIDE_AK;
pb = pb + TK * STRIDE_BK;
a = *pa;
b = *pb;
}
// epilogue
int rcm[TM] = pm * TM + 0 ... TM;
int rcn[TN] = pn * TN + 0 ... TN;
TYPE* pc[TM, TN] = C + rcn[newaxis, :] + rcm[:, newaxis] * ldc;
*pc = c;
}
```
All matrix multiplications variants can then be retrieved using the following compilation option:
```c
// A is not transposed
-DUSE_A=a -DSTRIDE_AK=1-DSTRIDE_AM=lda -DBROADCAST_AK=newaxis,: -DBROADCAST_AN=:,newaxis -DSHAPE_A=TM,TK
// A is transposed
-DUSE_A=^a -DSTRIDE_AK=lda-DSTRIDE_AM=1 -DBROADCAST_AK=:,newaxis -DBROADCAST_AN=newaxis,: -DSHAPE_A=TK,TM
// B is not transpose
-DUSE_B=b -DSTRIDE_BK=ldb-DSTRIDE_BN=1 -DBROADCAST_BK=:,newaxis -DBROADCAST_BN=newaxis,: -DSHAPE_B=TK,TN
// B is transpose
-DUSE_B=^b -DSTRIDE_BK=1-DSTRIDE_BN=ldb -DBROADCAST_BK=newaxis,: -DBROADCAST_BN=:,newaxis -DSHAPE_B=TN,TK
```
Auto-tuning can also be handled using pre-processor macros:
```c
// Auto-tuning TM and TN in {32, 64, 128}; TK in {8, 16}
-DTM=[32, 64, 128] -DTN=[32, 64, 128] -DTK=[8, 16]
```