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
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
- Set the dtype of "acc" appropriately so that epilogue fusion will have args with dtype
- Update dtype propagation to use `type_to_dtype` instead of instantiating tensor
- Throw if we have a string arg where we should have a proper CSEVariable, unless we're doing the Modification Subgraph thing which is nyi. everything else is appropriately typed (cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov @drisspg ).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/141991
Approved by: https://github.com/drisspg
ghstack dependencies: #139945, #140057, #141495, #141882
- Add in upcast_compute_type on creation of new tensors (loads, constants)
- Fixes index_expr - right now we are sort of inconsistent in dtype and dont always respect the dtype specified. would be nice to fix but not doing in this pr.
- bug fix in view dtype where we were always upcasting back to fp32 when input was in bf16/fp16. we should only be doing that if the output is also in bf16/fp16.
- for masked, avoid calling dtype propagation and just use output dtype.
Turns on the runtime dtype verification for opinfo tests. The separate test file is still useful because we can use it for testing turning off codegen_upcast_to_fp32.
Follow ups:
- We could consider requiring less explicit upcast_compute_types calls and do it automatically. That would potentially make things easier but be less flexible in the future. Maybe I should have done it this pr.
- Be more consistent on our index expr dtype printing.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/141495
Approved by: https://github.com/blaine-rister, https://github.com/arui-meta, https://github.com/ezyang
ghstack dependencies: #139945, #140057
A couple changes.
- Tries to reuse dtype propagation rules that were already registered in inductor. These were present both with `pointwise_overrides_data` and the `boolean_ops` list. Additionally, the registration of pointwise ops already specified dtype propagation rules. Saves those registrations and reuses them later.
- Factors out `get_promoted_dtype` which uses functools.lru_cache to take in non - CSEVariable args because those will not work with the functools cache.
Tests get added later in the stack when everything is implemented.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/139945
Approved by: https://github.com/blaine-rister, https://github.com/arui-meta, https://github.com/ezyang
Here's a markdown summary for the PR:
# Add workspace buffer support for Triton templates
## Summary
Adds support for templates to allocate and use temporary workspace buffers
## Key Changes
- Add `WorkspaceArg` support in Triton template system
- Automatic workspace allocation/deallocation around kernel execution
- Zero-initialization support for workspace buffers
- Seamless integration with existing tensor management
## Example Usage
```python
def generate(self, ...):
workspace_arg = WorkspaceArg(
count=1024*1024, # 1MB workspace
zero_fill=True # Zero-initialized
)
return TritonTemplateCaller(..., workspace_arg=workspace_arg)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138050
Approved by: https://github.com/Chillee, https://github.com/eellison
This adds host-side Triton TMA support to AOTInductor. Notes:
- Two helper functions, `init1DTMADescriptor` and `init2DTMADescriptor` are added to the C++ wrapper codegen on GPU, conditioned on the model having user-defined Triton kernels with host-side TMA (CUDA-specific).
- C++ wrapper codegen on GPU emits TMA descriptor initialization via the aforementioned helper functions.
- Special handling added for the TMA descriptors (in the Python wrapper codegen) during the compile-time autotuning, as the underlying tensor can't be passed directly to the user-defined Triton kernel. TMA descriptors are generated in-between the source tensor's buffer and the kernel call, like in the full Python wrapper codegen.
- This PR concludes the host-side Triton TMA support in PT2.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138878
Approved by: https://github.com/desertfire, https://github.com/chenyang78
ghstack dependencies: #138759, #138877
Summary:
- This diff introduces `dtype` attribute to `TritonCSEVariable` and a dtype propagation helper function to infer dtype from input to output for each op.
- There will be a follow-up diff that uses this `dtype` information in `TritonCSEVariable` to perform dtype-aware codegen.
Test Plan: CI
Differential Revision: D61815079
Pull Request resolved: https://github.com/pytorch/pytorch/pull/136778
Approved by: https://github.com/eellison, https://github.com/blaine-rister
Summary: while reading through inductor template code I found a few places where else statements were driving me crazy. Fixing them as I read
Test Plan: CI
Differential Revision: D64882385
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138789
Approved by: https://github.com/aakhundov
Summary: The problem happened after splitting CppWrapperCpu and CppWrapperCpuArrayRef, because CppWrapperCpuArrayRef.generate_index_put_fallback missed a statement. Running test_aot_inductor.py as a whole didn't reveal the problem, but running test_index_put_with_none_index_cpu_with_stack_allocation individually did. Digging deeper, the root cause is init_backend_registration has incorrectly cached CPU CppWrapperCodegen class, which means CppWrapperCpuArrayRef was never picked when running test_aot_inductor.py as a whole.
Differential Revision: [D64598714](https://our.internmc.facebook.com/intern/diff/D64598714)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138303
Approved by: https://github.com/hl475
This adds Dynamo tracing support for the host-side Triton TMA API (see `create_2d_tma_descriptor` calls on the host in the [Triton tutorial](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#sphx-glr-getting-started-tutorials-09-persistent-matmul-py)). A few notes:
- Here we assume the availability of the host-side TMA API added to upstream Triton in https://github.com/triton-lang/triton/pull/4498. As of time of writing, this is not a part of the PT2 OSS Triton pin (although back-ported internally). OSS Triton pin update should be done in December 2024.
- Due to Dynamo support implemented in the previous PR, the `tma_descriptor_metadata` dict is delivered to the `triton_kerenl_wrap_` lowering and passed to the `ir.UserDefinedTritonKernel` as additional argument.
- Looking into the `tma_descriptor_metadata`, `ir.UserDefinedTritonKernel` substitutes the corresponding `TensorBox` arguments of the kernel (swapped upstream in Dynamo) by the new `ir.TMADescriptor` nodes implementing TMA descriptors in Inductor IR.
- `ir.TMADescriptor.__init__` provides the wiring between the upstream underlying `ir.TensorBox` and the downstream `ir.UserDefinedTritonKernel` kernel. In particular, we use `ir.NonOwnedLayout` wrapping `ir.ReinterpretView` to avoid the upstream tensor's buffer being deleted prematurely (before the TMA descriptor is used in the Triton kernel).
- Via `ir.TMADescriptor.codegen`, the Triton's `create_{1d,2d}_tma_descriptor` function call is codegened in the wrapper (in the host code).
- New `TMADescriptorArg` dataclass is added to handle the Triton kernel metadata pertinent to host-side TMA.
- AOT Inductor support will be implemented in a follow-up PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/137950
Approved by: https://github.com/eellison
ghstack dependencies: #137677
Adds lowering for `aten.searchsorted`. This entails:
1. Adding support for multi-dimensional bucket tensors to `ops.bucketize`.
2. Adding support for striding to `ops.bucketize`.
3. Adding support for sorting tensors to `ops.bucketize`.
4. Adding a lowering for `aten.searchsorted.Tensor`.
5. Adding a basic decomposition for `aten.searchsorted.Scalar` that calls into the lowering for tensors.
6. Updating the meta-function for `aten.searchsorted` to properly check some of the sizing conditions.
Closes#135873
Differential Revision: [D63766514](https://our.internmc.facebook.com/intern/diff/D63766514)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/135701
Approved by: https://github.com/amjames, https://github.com/eellison, https://github.com/davidberard98
