This enforces the invariant that every backend implements the same set of ops and removes a layer of indirection for BasicMathOps.
Interestingly this is a small compile time win:
```
...
WIN: benchmark ('add_loop_inductor', 'compile_time_instruction_count') failed, actual result 30151159301 is -6.13% lower than expected 32120000000 ±1.50% please update the expected results.
please update all results that changed significantly, and not only the failed ones
PASS: benchmark ('add_loop_inductor_dynamic_gpu', 'compile_time_instruction_count') pass, actual result 44447549162 -1.69% is within expected 45210000000 ±2.50%
WIN: benchmark ('add_loop_inductor_gpu', 'compile_time_instruction_count') failed, actual result 26743557195 is -2.25% lower than expected 27360000000 ±1.50% please update the expected results.
please update all results that changed significantly, and not only the failed ones
PASS: benchmark ('basic_modules_ListOfLinears_eager', 'compile_time_instruction_count') pass, actual result 945129734 +0.93% is within expected 936400000 ±1.50%
WIN: benchmark ('basic_modules_ListOfLinears_inductor', 'compile_time_instruction_count') failed, actual result 18984384503 is -3.19% lower than expected 19610000000 ±1.50% please update the expected results.
please update all results that changed significantly, and not only the failed ones
WIN: benchmark ('basic_modules_ListOfLinears_inductor_gpu_force_shape_pad', 'compile_time_instruction_count') failed, actual result 17258025389 is -1.94% lower than expected 17600000000 ±1.50% please update the expected results.
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146235
Approved by: https://github.com/shunting314
ghstack dependencies: #146225, #146226
This fixes handling for "1" and "None" args with new Triton versions. TL;DR: triton_meta["constants"] (which is passed to ASTSource) should be a map of {"kwarg_name": constant_value} for values which are tl.constexpr, or have a value of 1 or None (i.e. "specialized" constants). For constant args, triton_meta["signature"][arg_name] should be "constexpr" (even for specialized constants).
Note: This adds support for Triton versions after 5512; but not for versions in between 5220 and 5512 (i.e. `TritonAttrsDescriptorVersion.V3_BACKENDS_TUPLE`). There's a completely different format for constants/signature in the commit range in between.
To test: I ran `test_torchinductor.py` and `test_triton_kernels.py` with the main branch of triton (~jan 27). The only failing tests are aoti-related tests (which need to be fixed as a follow-up), and test_mutable_custom_op_fixed_layout2_cuda (which is failing with or without the new triton version on my machine); additionally, the split-scan/split-reduction kernels rely on https://github.com/triton-lang/triton/pull/5723.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145515
Approved by: https://github.com/SamGinzburg
Prior to this PR, constexprs were appearing in signatures as `{.. "XBLOCK : tl.constexpr": "constexpr"}` when they really should appear as `{.. "XBLOCK": "constexpr"}`.
This PR represents the argument names as ArgName objects, which can optionally be marked as constexpr.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145583
Approved by: https://github.com/jansel
When we attempt prologue or epilogue fusion with a TritonTemplate, we benchmark it at compile time in order to determine profitability. This avoids slowdowns/register spilling, and allows us to pick fusion when a base triton template is slower than cublas but faster when considering an epilogue. However, that fused benchmarking does not do the same async compilation as we do for the base TritonTemplate. The Base TritonTemplate is async compiled during lowering, then later waited on and benchmarked.
This PR extends a similar process to benchmarking fused TritonTemplates in the scheduler. We keep a list of pending fusions which have async compilations. And we resolve any pending fusions a node is in prior to attempting to fuse it with any other node.
Initially, I saw some slowdowns with this because we kick off async compilations of identical fusions in parallel. To address this I added source code caching at the `async_compile` level (we also already cache benchmark runs, but that would not happen in parallel).
Compilation speedups:
<img width="717" alt="image" src="https://github.com/user-attachments/assets/8e8f7d6c-7824-4210-83f9-a2a0f6db5ac9" />
This also should let us be a bit more aggressive with either configs, or benchmarking other fusions which are hard to determine profitability of.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143408
Approved by: https://github.com/jansel, https://github.com/shunting314
Triton commit 5220 adds tuple support in Triton (changing the indexing format in AttrsDescriptor) and commit 5512 replaces AttrsDescriptor with raw tuples. This is an initial PR to add support for Triton versions after commit 5512 landed.
The main changes in 5220 and 5512 that need to be supported:
* AttrsDescriptor() gets replaced with a raw dict. The raw dict has the format `{(TUPLES): [["tt.divisibility", 16]]}`, where `(TUPLES)` is a tuple of indices, e.g. `((0,), (1,), (3,))` to indicate that args 0, 1, and 3 are divisible by 16. These indices are, themselves, represented as tuples to support nested inputs (e.g. an argument that's a tuple), but support for tuples is not implemented right now.
* "signature" changes: the signature now contains _all_ args, including constexpr and constant args.
* ASTSource now takes "constexprs" instead of "constants" - for example, equal-to-1 args are constants but not constexprs so we don't need to pass these args as "constants".
What this PR supports:
* Triton versions before Dec 9, 2024, and (partial support for) Triton versions after Jan 1, 2025
* (triton jan 1+) typical inductor-generated triton: updated AttrsDescriptor, signatures, constexpr/constant handling.
What this PR doesn't support (TODO in follow-up PRs):
* Triton versions between Dec 9, 2024 and before Jan 1, 2025
* (triton jan 1+) user-defined triton kernel support (this is implemented already in @anmyachev's patch)
* (triton jan 1+) triton_helper support (failing in triton codegen - needs investigation)
* (triton jan 1+) AOTI / cpp wrapper
thanks to @anmyachev for patches in https://github.com/intel/intel-xpu-backend-for-triton/blob/main/scripts/pytorch.patch, which contains most of these changes already
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145051
Approved by: https://github.com/jansel
This PR adds a heuristic to potentially fail the block pointer match early. Expressions like below take a long time to match using sympy (e.g. > 100 seconds)
```python
# torch._inductor.config.triton.use_block_ptr = True
# torch._inductor.config.triton.prefer_nd_tiling = True
# Expression from pytest -k test_max_pool2d1_dynamic_shapes_cuda:
((xindex//ps1))*((s2 - 3//2))**2 + 2*((xindex//ps1))*((s2 - 3//2)) + ((xindex//ps1)) + ((s2 - 3//2))*(ModularIndexing(xindex, ps0, ps0)) + (ModularIndexing(xindex, 1, ps0)) + (ModularIndexing(xindex, ps0, ps0))
```
Additionally, the heuristic for the number of dimensions based on the indexing expression is refined to only add dimensions for FloorDiv(index, denom) and ModularIndexing(index, denom, modulo) instead of including FloorDiv/ModularIndexing expressions that don't involve the index.
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144681
Approved by: https://github.com/jansel
### 1. Synopsis
Adds `cache_modifier='.cg'` optional argument into `tl.load` instructions in the inductor-generated triton code for selected buffers.
It makes the `tl.load` instruction to skip the L1 cache for short-lived / non-reused data.
### 2. Using the feature
This feature is experimental and disabled by default. It can be enabled by setting the environmental variable `TORCHINDUCTOR_SKIP_L1` equal to `1`.
### 3. Results
For a simple pointwise addition kernel:
```python
@torch.compile
def add_dummy(x: torch.Tensor, y: torch.Tensor):
return x+y
```
we get (bandwith performance is in GB/s):
(a) feature DISABLED:
(b) feature ENABLED:
### 4. Caveats
The feature boost is only available when using
```python
torch._dynamo.config.cache_size_limit = 64 # or any other sufficiently big number..
torch._dynamo.config.automatic_dynamic_shapes = False # use static shapes
```
When using (the default) dynamic shapes, only 1-2 triton kernels are generated with non-optimal block-sizes for
*all* the cases (vector sizes), hiding any perf benefit from skipping the L1 cache.
In the static case, as an optimal block size is generated for each vector size, the perf benefit of skipping the L1 cache becomes visible.
This block-size optimization issue is a larger problem in pytorch inductor and is outside the scope of this feature.
### 5. References
- [tl.load](https://triton-lang.org/main/python-api/generated/triton.language.load.html#triton.language.load)
- [cache operators](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#cache-operators)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143115
Approved by: https://github.com/jansel
Summary:
Triton compiler does not automatically promote fp16/bf16 reductions to fp32 accumulation. This will result in significant accuracy issue.
This diff will upcast the input to FP32 for all math reductions `["welford_reduce", "welford_combine", "prod", "sum", "xor_sum"]`
Test Plan:
CI
```
python test/inductor/test_torchinductor.py TritonCodeGenTests.test_low_precision_reduction
```
Differential Revision: D65965032
Pull Request resolved: https://github.com/pytorch/pytorch/pull/141052
Approved by: https://github.com/blaine-rister
Fixes#134277 and https://github.com/pytorch/pytorch/issues/142317.
Sub-PRs containing refactors from this one:
- https://github.com/pytorch/pytorch/pull/141733
- https://github.com/pytorch/pytorch/pull/141738
- https://github.com/pytorch/pytorch/pull/141751 (based off the former)
- https://github.com/pytorch/pytorch/pull/142249
- https://github.com/pytorch/pytorch/pull/142020
- https://github.com/pytorch/pytorch/pull/143135
These refactor PRs should land before the main one.
# Feature
*Note: to minimize risk, multi-dimensional reductions are gated by the flag `config.triton.tile_reductions`, which defaults to False.*
Instead of having a single reduction dimension called `"r"`, we can now support 2D reductions with `"r0_"` and `"r1_"` dimensions. 2D reductions generate two nested loops, with different block pointer advancements in each loop body. Most of the implementation is generic to ND reductions, but for now the tiling algorithm sets a hard limit at 2D.
Here's an example of a 2D persistent reduction kernel:
```
@triton.jit
def triton_per_fused_sum_0(in_ptr0, out_ptr0, xnumel, r0_numel, r1_numel, XBLOCK : tl.constexpr):
xnumel = 1
r0_numel = 15
R0_BLOCK: tl.constexpr = 16
r1_numel = 15
R1_BLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None]
xmask = tl.full([XBLOCK, R0_BLOCK, R1_BLOCK], True, tl.int1)
r0_index = tl.arange(0, R0_BLOCK)[None, :, None]
r0_offset = 0
r0_mask = r0_index < r0_numel
r1_index = tl.arange(0, R1_BLOCK)[None, None, :]
r1_offset = 0
r1_mask = r1_index < r1_numel
rnumel = r0_numel * r1_numel
RBLOCK: tl.constexpr = R0_BLOCK*R1_BLOCK
roffset = r1_offset + (r0_offset*r1_numel)
rindex = r1_index + (r0_index*r1_numel)
r0_0 = r0_index
r1_1 = r1_index
tmp0 = tl.load(tl.make_block_ptr(in_ptr0, shape=[15, 15], strides=[30, 1], block_shape=[R0_BLOCK, R1_BLOCK], order=[1, 0], offsets=[r0_offset, r1_offset]), boundary_check=[0, 1], padding_option='zero')[None, :, :]
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK, R1_BLOCK])
tmp3 = tl.where(r0_mask & r1_mask, tmp1, 0)
tmp4 = tl.reshape(tmp3, [XBLOCK, RBLOCK])
tmp5 = tl.sum(tmp4, 1)[:, None, None]
tl.store(out_ptr0 + (tl.full([XBLOCK, 1, 1], 0, tl.int32)), tmp5, None)
''', device_str='cuda')
```
There are a few main differences between this kernel and what Inductor would generate without this PR.
- Instead of an `r`/`RBLOCK` dimension, we have two reduction dimensions: `r0_`/`R0_BLOCK` and `r1_`/`R1_BLOCK`.
- There are special size and indexing variables for reductions, which don't directly correspond to any kernel dimension. (`rindex`, `rnumel`, `RBLOCK`, and `roffset`.) These collapse N-D reduction sizes and indices indices into 1D. This simplifies the codegen for reductions, which sometimes want to access linear indices instead of N-dimensional ones. Doing things this way allows us to generate N-D loads and stores, but access this data as if it were 1D, minimizing the blast radius of this PR. Although this makes the code more verbose, it shouldn't have a perf impact because the triton compiler eliminates dead code.
- We generate the line `tmp4 = tl.reshape(tmp3, [XBLOCK, RBLOCK])` before performing the actual reduction. This reshapes N reduction dimensions into 1D. This allows us to reduce over all N dimensions at once, simplifying the codegen and allowing the Triton complier to decide the order of processing under the hood.
Here's an example of a looped reduction:
```
@triton.jit
def triton_red_fused_sum_0(in_ptr0, out_ptr0, xnumel, r0_numel, r1_numel, XBLOCK : tl.constexpr, R0_BLOCK : tl.constexpr, R1_BLOCK : tl.constexpr):
xnumel = 3
r0_numel = 43
r1_numel = 129
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None]
xmask = xindex < xnumel
r0_base = tl.arange(0, R0_BLOCK)[None, :, None]
r1_base = tl.arange(0, R1_BLOCK)[None, None, :]
rnumel = r0_numel * r1_numel
RBLOCK: tl.constexpr = R0_BLOCK*R1_BLOCK
rbase = r1_base + (r0_base*r1_numel)
x0 = xindex
block_ptr0 = tl.make_block_ptr(in_ptr0, shape=[3, 43, 129], strides=[11094, 258, 1], block_shape=[XBLOCK, R0_BLOCK, R1_BLOCK], order=[2, 1, 0], offsets=[xoffset, 0, 0])
_tmp2 = tl.full([XBLOCK, R0_BLOCK, R1_BLOCK], 0, tl.float32)
for r0_offset in range(0, r0_numel, R0_BLOCK):
r0_index = r0_offset + r0_base
r0_mask = r0_index < r0_numel
for r1_offset in range(0, r1_numel, R1_BLOCK):
r1_index = r1_offset + r1_base
r1_mask = r1_index < r1_numel
roffset = r1_offset + (r0_offset*r1_numel)
rindex = r1_index + (r0_index*r1_numel)
r0_1 = r0_index
r1_2 = r1_index
tmp0 = tl.load(block_ptr0, boundary_check=[0, 1, 2], padding_option='zero', eviction_policy='evict_first')
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK, R1_BLOCK])
tmp3 = _tmp2 + tmp1
_tmp2 = tl.where(r0_mask & r1_mask & xmask, tmp3, _tmp2)
block_ptr0 = tl.advance(block_ptr0, [0, 0, R1_BLOCK])
block_ptr0 = tl.advance(block_ptr0, [0, R0_BLOCK, (-1)*R1_BLOCK*((128 + R1_BLOCK) // R1_BLOCK)])
tmp4 = tl.reshape(_tmp2, [XBLOCK, RBLOCK])
tmp2 = tl.sum(tmp4, 1)[:, None, None]
tl.store(tl.make_block_ptr(out_ptr0, shape=[3], strides=[1], block_shape=[XBLOCK], order=[0], offsets=[xoffset]), tl.reshape(tmp2, [XBLOCK]).to(tl.float32), boundary_check=[0])
''', device_str='cuda')
```
In addition to the aforementioned changes to the persistent reduction, multidimensional looped reductions have a few more lines of code:
- They calculate indices inside the loop using `r0_base` and `r1_base`. For compatibility with existing codegen, these are collapsed to the 1D variant `rbase`.
- Block pointer advancements are more nuanced for multidimensional loops. At the end of each loop body, we emit a `tl.advance` line which not only increments the pointer in its own dimension, but also undoes the cumulative increments of the previous loop level. This is equivalent to the usual practice in nested loops of starting with a fresh iteration variable at each level. Implementing this required refactoring the way we generate pointer advancements into a new `self.pointer_advancements` field of the kernel, which categorizes advancements by dimension.
The biggest difficulty in implementing this feature was that we represented tiling with a tuple like `(5,2)`. In the existing codebase, the compiler can infer that the reduction dimension of `(5,2)` is `2`, since reductions are always the last dimension. This became cumbersome now that we have to support multiple reduction dimensions, so I refactored tiling into a dict like `{"x": 5, "r0_": 2, "r1_": 4}`. This required quite a few code changes, but I don't think it makes the underlying logic much more complex. This will also make it easier to eventually support simultaneous pointwise and reduction tiling, like `{"x": 5, "y": 5, "r0_": 2, "r1_": 4}`. (This is not supported today, but we might want to do it eventually.)
The existing tiling algorithm generalized naturally to support reductions. For pointwise kernels, we tile the pointwise dimensions (`"x"`, `"y"`) as is. For reduction kernels, we never tile the `"x"` dimension, and only tile the reduction dimensions (`"r0_"`, `"r1_"`). Thus we only ever tile pointwise OR reduction dimensions, but not both. In principle it seems possible to support both, but it would likely require changes to the kernel fusion and autotuning logic. I thought it best to keep this PR as minimal as possible since it already touched a lot of different files.
Unfortunately, these changes weren't enough to get block pointers in some seemingly simple test cases. In some tests for `argmax` and `var_mean`, we already collapse reduction dimensions into 1D and generate modular indexing expressions, prior to tiling. So it's not trivial to figure out how to expand the collapsed reduction dimension back to a shape that would simplify the indexing.
To address these cases, this PR adds a new feature to the `config.prefer_nd_tiling` option, which analyzes reads and writes in the kernel, using the same mod-div pattern matching logic that generates block pointers later on. By matching this pattern, we can solve for the tiling splits which *would* simplify the indexing expression, and use then use that tiling to eliminate the modular indexing and emit a block pointer. This tiling mode is still off by default, but it's important for certain applications where we need to get as many block pointers as possible.
# Test plan
This touches pretty much anything that uses the Triton and Halide backends, so the existing CI provides good coverage. However, 2D reductions are gated behind a few feature flags like `config.prefer_nd_tiling` and `config.tile_reductions`, so this really only checks that the PR doesn't break 1D reductions.
In addition to existing CI tests, this PR also adds some new tests that specifically stress 2D reductions:
- `test_2d_reduction_odd_shapes`: test 2D reductions with a variety of ops and sizes. This covers the typical persistent and looped reductions.
- `test_2d_reduce_no_x_dim`: test 2D reductions with no x dimension.
- `test_2d_welford_reduction`: test 2D welford reductions with block pointers.
- `test_welford_non_block_pointer`: test a 2D welford reduction when block pointer analysis fails.
- `test_reduction_multiple_discontiguous_dims`: test reducing over more than one discontiguous dimension. We won't get a block pointer for this case, since that would require 3D tiling, but we're currently limited to 2D.
- `test_2d_reduction_multi_kernel`: test multi kernel autotuning on a 2D softmax kernel.
- `test_enable_tiled_reductions`: test that `config.triton.tile_reductions` enables/disables this feature.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/137243
Approved by: https://github.com/jansel
Co-authored-by: Yueming Hao <yhao@meta.com>
Co-authored-by: Jason Ansel <jansel@meta.com>
This PR aims to add the functionality support of max-autotune for XPU. The current triton templates and configurations are not well optimized for XPU, so the performance is not ready yet. Also the `mm_plus_mm` template have accuracy issues in some cases. We will address these issues in the next PRs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143266
Approved by: https://github.com/EikanWang, https://github.com/jansel
`"compile_id"` had slipped into our generated Triton code (in the
metadata), which will defeat caching because the same kernels generated
in a different order would not cache hit with eachother.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143951
Approved by: https://github.com/oulgen
`"compile_id"` had slipped into our generated Triton code (in the
metadata), which will defeat caching because the same kernels generated
in a different order would not cache hit with eachother.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143951
Approved by: https://github.com/oulgen
This PR aims to add the functionality support of max-autotune for XPU. The current triton templates and configurations are not well optimized for XPU, so the performance is not ready yet. Also the `mm_plus_mm` template have accuracy issues in some cases. We will address these issues in the next PRs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143266
Approved by: https://github.com/EikanWang, https://github.com/jansel
This PR extends our ability to fuse pointwise nodes onto triton templates with the ability to fuse pointwise nodes into triton templates - prologue fusion.
Similar to the store_output api:
`{{store_output(("idx_m", "idx_n"), "acc", "mask")}}`
And the modification api:
```
{{ modification(
subgraph_number=0,
output_name="post_mod_scores",
score="qk",
out="qk"
) | indent_except_first(1) }}
```
We have:
```{{load_input("B", "b", ("idx_m", "idx_n"), mask=None if EVEN_K else "b_mask", indent_width=8)}}```
Because we are now loading the input with explicit indices and mask, I needed to rewrite the mm kernel to no longer update the [pointers by BLOCK_K](bb03ef7aca/torch/_inductor/kernel/mm.py (L110-L111)) on every iteration and instead on each iteration compute indices from the the k_idx of each loop. This did not have any perf difference.
There are a couple main use cases for prologue fusion:
- Fusing dequants into a matmul. particularly for more bandwidth bound scenarios.
- Fusing gather into a matmul. This is useful particularly in MOE. See https://github.com/pytorch/pytorch/issues/134535 for more details.
Prologue fusion is generally much less profitable than epilogue fusion, because it must be applied to an element of an input on each loop of the matmul, compared to only once in the epilogue (gather into matmul is a potential exception). Accordingly, we are much less aggressive in attempting to fuse prologue fusion. We only attempt fusion if it does not increase the number of memory bytes read instead the triton template, multipled by a small factor to allow gathers. This restricts reliably unprofitable fusions like fp32->fp16 inside kernel. In future pr we could potentially have api of being more aggressive if we know we are in a bandwidth bound regime. See: https://github.com/pytorch/pytorch/pull/134532/files#diff-d2539c9c8dc6a3d7e457767a880612e96d3c85752a77ead49a9e4e00a3e4c3c7R3060-R3066
Other notes:
By default we will upcast to fp32 inside every kernel. This matches eager numerics. This is fine enough for epilogue because it is only done once (although it is probably unnecessary for say a relu) but tanks perf for prologue. I am currently using the `codegen_upcast_to_fp32` option to avoid it, but that will not work for libdevice calls that require fp32. We will need https://github.com/pytorch/pytorch/pull/136778/ and dtype-aware codegen to upcast fp16 ops into libdevice calls.
With prologue fusion, we now have essentially separate kernels for each input, and for the output. I had to increase the number of fields that are swapped out in `set_subgraph_body` by a large number :/ I also update the fusion logic because the inputs will have a different group than the outputs. Maybe as part of enabling multiple outputs, this could get cleaned up a bit so..
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134532
Approved by: https://github.com/jansel
Preparatory refactor for https://github.com/pytorch/pytorch/pull/137243.
# Feature
This PR changes the `RINDEX` / `"r"` symbol type to `(R0_INDEX, R1_INDEX)` and `("r0_", "r1_")`, respectively. This allows the relevant code to support 2D (often ND) reductions. Unlike the parent PR, this one does not change the tiling algorithm, so `"r1_"` is never used. However, it prepares other parts of the system to handle `"r1_"` once we start using it. This should significantly reduce the chances of hitting merge conflicts, making the parent PR much easier to land.
The only change to the generated triton code is to rename `"rindex"` -> `"r0_index"`, `"RBLOCK"` -> `"R0_BLOCK"`, etc. To maintain compatibilty with existing codegen, this also generates aliases to the old reduction variables like `rindex = r0_index`. If we generated 2D reductions (which this PR will not do), the aliases would be more complicated and would collapse 2D multi-indices to linear indices. See some example kernels in the parent PR.
These aliases can be eliminated by the Triton compiler, and should not impact the final machine code running on the GPU. See the perf testing in the parent PR which confirms the aliases do not impact perf.
# Test plan
The existing CI provides good coverage. This PR modifies the expected code in a few places, renaming reduction variables from `r.*` to `r0_.*`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/142020
Approved by: https://github.com/jansel
Co-authored-by: Jason Ansel <jansel@meta.com>
This PR extends our ability to fuse pointwise nodes onto triton templates with the ability to fuse pointwise nodes into triton templates - prologue fusion.
Similar to the store_output api:
`{{store_output(("idx_m", "idx_n"), "acc", "mask")}}`
And the modification api:
```
{{ modification(
subgraph_number=0,
output_name="post_mod_scores",
score="qk",
out="qk"
) | indent_except_first(1) }}
```
We have:
```{{load_input("B", "b", ("idx_m", "idx_n"), mask=None if EVEN_K else "b_mask", indent_width=8)}}```
Because we are now loading the input with explicit indices and mask, I needed to rewrite the mm kernel to no longer update the [pointers by BLOCK_K](bb03ef7aca/torch/_inductor/kernel/mm.py (L110-L111)) on every iteration and instead on each iteration compute indices from the the k_idx of each loop. This did not have any perf difference.
There are a couple main use cases for prologue fusion:
- Fusing dequants into a matmul. particularly for more bandwidth bound scenarios.
- Fusing gather into a matmul. This is useful particularly in MOE. See https://github.com/pytorch/pytorch/issues/134535 for more details.
Prologue fusion is generally much less profitable than epilogue fusion, because it must be applied to an element of an input on each loop of the matmul, compared to only once in the epilogue (gather into matmul is a potential exception). Accordingly, we are much less aggressive in attempting to fuse prologue fusion. We only attempt fusion if it does not increase the number of memory bytes read instead the triton template, multipled by a small factor to allow gathers. This restricts reliably unprofitable fusions like fp32->fp16 inside kernel. In future pr we could potentially have api of being more aggressive if we know we are in a bandwidth bound regime. See: https://github.com/pytorch/pytorch/pull/134532/files#diff-d2539c9c8dc6a3d7e457767a880612e96d3c85752a77ead49a9e4e00a3e4c3c7R3060-R3066
Other notes:
By default we will upcast to fp32 inside every kernel. This matches eager numerics. This is fine enough for epilogue because it is only done once (although it is probably unnecessary for say a relu) but tanks perf for prologue. I am currently using the `codegen_upcast_to_fp32` option to avoid it, but that will not work for libdevice calls that require fp32. We will need https://github.com/pytorch/pytorch/pull/136778/ and dtype-aware codegen to upcast fp16 ops into libdevice calls.
With prologue fusion, we now have essentially separate kernels for each input, and for the output. I had to increase the number of fields that are swapped out in `set_subgraph_body` by a large number :/ I also update the fusion logic because the inputs will have a different group than the outputs. Maybe as part of enabling multiple outputs, this could get cleaned up a bit so..
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134532
Approved by: https://github.com/jansel
