[IR] Adding more support for dynamic shape on Task and FlowGraph level#220
Merged
yaoyaoding merged 2 commits intohidet-org:mainfrom May 9, 2023
Merged
[IR] Adding more support for dynamic shape on Task and FlowGraph level#220yaoyaoding merged 2 commits intohidet-org:mainfrom
yaoyaoding merged 2 commits intohidet-org:mainfrom
Conversation
vadiklyutiy
pushed a commit
that referenced
this pull request
Jul 22, 2024
This PR supports the vectorization of epilogue fusion.
Originally, we fused the operators belonging to the epilogue stage on
scalars.
For example, the code before fusion is as follows
```python
for mi in range(mma_m):
for ni in range(mma_n):
global_m = mi * stride_m
global_n = ni * stride_n
if global_m < m and global_n < n:
# the following statement always be converted to a LDG.U16 instruction at SASS level if we assume `regs_c` is a float16 register array.
# the if statement prevents nvcc to aggressively do vectorized writing back to global memory.
gmem_c[global_m, global_n] = regs_c[mi, ni]
```
The code after fusion looks like
```python
for mi in range(mma_m):
for ni in range(mma_n):
global_m = mi * stride_m
global_n = ni * stride_n
if global_m < m and global_n < n:
# x, y, z are fused tensors
gmem_c[inverse_map(global_m, global_n)] = elementwise_op(regs_c[mi, ni], x, y, z...)
```
The fusion still emits some scalar store instructions to perform the
final write-back, which is inefficient. Although fusion can bring us
some performance benefits, the penalty of inefficient write-back can be
higher than the benefit of fusion itself.
This PR is trying to fuse operators while vectorizing the memory
accesses as much as possible. Basically, we developed a technique
similar to the [epilogue visitor
tree](https://dl.acm.org/doi/pdf/10.1145/3620666.3651369) in CUTLASS to
address this problem. For details, you can refer to the attached
[document](https://drive.google.com/drive/folders/1kXlu2k-lPmH0jPlQ_8CfmqZvAPVtQ3Qs).
## Changes
1. restructure the instruction selection as a compiler pass
2. support the vectorized epilogue fusion algorithm in the
`apply_prologue_epilogue` pass.
3. adapt the schedule template of `matmul_f16_pk` to enable this
optimization.
4. currently, we have a very simple shared memory updater, and we should
design a shared memory planner. A potential proposal can be found in
[this document](https://dl.acm.org/doi/pdf/10.5555/314500.315082), which
is the shared memory allocator used in triton.
## Simple epilogue fusion
This experiment benchmarks the following graph. `dims` is a permutation
of the four axes 0, 1, 2, 3. We iterate over all the possible
permutations to show the generality of fusion capability.
```
def graph(a, b, d, bx1xn, bxmx1, mx1, x1xn):
c = a @ b
c = c + d
c = c + bxmx1 + bx1xn + mx1 + x1xn
L, M, N = c.shape
c = c.reshape(L, M, N // head_size, head_size)
c = c.permute(*dims)
c = c.reshape(L * N // head_size, M, head_size)
c = relu(c)
return c
```
Performance comparison
| problem(m,n,k,l) | dims | max-autotune | hidet(after) | hidet(before)
| speedup (after vs before) |
| -------------------- | ------------ | ------------ | ------------- |
------------ | ------- |
| [512, 1536, 512, 16] | (0, 1, 2, 3) | 0.263 | 0.199 | 0.229 | 13.10% |
| [512, 1536, 512, 16] | (0, 1, 3, 2) | 0.25 | 0.218 | 0.27 | 19.26% |
| [512, 1536, 512, 16] | (0, 2, 1, 3) | 0.248 | 0.217 | 0.223 | 2.69% |
| [512, 1536, 512, 16] | (0, 2, 3, 1) | 0.374 | 0.219 | 0.224 | 2.23% |
| [512, 1536, 512, 16] | (0, 3, 1, 2) | 0.509 | 0.269 | 0.35 | 23.14% |
| [512, 1536, 512, 16] | (0, 3, 2, 1) | 0.278 | 0.226 | 0.225 | \-0.44%
|
| [512, 1536, 512, 16] | (1, 0, 2, 3) | 0.248 | 0.202 | 0.237 | 14.77% |
| [512, 1536, 512, 16] | (1, 0, 3, 2) | 0.251 | 0.222 | 0.269 | 17.47% |
| [512, 1536, 512, 16] | (1, 2, 0, 3) | 0.264 | 0.202 | 0.241 | 16.18% |
| [512, 1536, 512, 16] | (1, 2, 3, 0) | 0.247 | 0.275 | 0.274 | \-0.36%
|
| [512, 1536, 512, 16] | (1, 3, 0, 2) | 0.708 | 0.224 | 0.4 | 44.00% |
| [512, 1536, 512, 16] | (1, 3, 2, 0) | 0.368 | 0.352 | 0.435 | 19.08% |
| [512, 1536, 512, 16] | (2, 0, 1, 3) | 0.261 | 0.211 | 0.222 | 4.95% |
| [512, 1536, 512, 16] | (2, 0, 3, 1) | 0.477 | 0.221 | 0.225 | 1.78% |
| [512, 1536, 512, 16] | (2, 1, 0, 3) | 0.262 | 0.211 | 0.236 | 10.59% |
| [512, 1536, 512, 16] | (2, 1, 3, 0) | 0.268 | 0.289 | 0.288 | \-0.35%
|
| [512, 1536, 512, 16] | (2, 3, 0, 1) | 0.257 | 0.22 | 0.227 | 3.08% |
| [512, 1536, 512, 16] | (2, 3, 1, 0) | 0.251 | 0.373 | 0.355 | \-5.07%
|
| [512, 1536, 512, 16] | (3, 0, 1, 2) | 0.559 | 0.27 | 0.372 | 27.42% |
| [512, 1536, 512, 16] | (3, 0, 2, 1) | 0.376 | 0.225 | 0.223 | \-0.90%
|
| [512, 1536, 512, 16] | (3, 1, 0, 2) | 0.648 | 0.281 | 0.399 | 29.57% |
| [512, 1536, 512, 16] | (3, 1, 2, 0) | 0.624 | 0.352 | 0.474 | 25.74% |
| [512, 1536, 512, 16] | (3, 2, 0, 1) | 0.333 | 0.226 | 0.228 | 0.88% |
| [512, 1536, 512, 16] | (3, 2, 1, 0) | 0.35 | 0.354 | 0.355 | 0.28% |
## Model benchmark
Performance comparison
| model | inputs | eager | reduce-overhead | max-autotune | hidet(after)
| hidet(before) | speedup (after vs before) |
| ------------------------ | ------------------ | ------ |
--------------- | ------------ | ------------ | ------------- |
------------------------- |
| model/gpt2 | f16, bs=1, seq=128 | 2.043 | 0.557 | 0.555 | 0.7 | 0.7 |
0.00% |
| model/gpt2 | f16, bs=1, seq=512 | 2.01 | 1.26 | 1.112 | 1.204 | 1.297
| 7.17% |
| model/gpt2 | f16, bs=8, seq=128 | 2.759 | 1.97 | 1.684 | 1.635 | 1.892
| 13.58% |
| model/gpt2 | f16, bs=8, seq=512 | 13.622 | 7.915 | 6.633 | 7.014 |
8.101 | 13.42% |
| model/bert-base-uncased | f16, bs=1, seq=128 | 1.247 | 0.744 | 0.753 |
0.659 | 0.671 | 1.79% |
| model/bert-base-uncased | f16, bs=1, seq=512 | 1.498 | 1.346 | 1.394 |
1.187 | 1.211 | 1.98% |
| model/bert-base-uncased | f16, bs=8, seq=128 | 2.045 | 1.844 | 1.723 |
1.526 | 1.637 | 6.78% |
| model/bert-base-uncased | f16, bs=8, seq=512 | 7.115 | 7.205 | 6.239 |
5.597 | 5.933 | 5.66% |
| model/bert-large-uncased | f16, bs=1, seq=128 | 2.342 | 1.835 | 1.974
| 1.515 | 1.496 | \-1.27% |
| model/bert-large-uncased | f16, bs=1, seq=512 | 3.586 | 3.367 | 3.63 |
2.932 | 3.108 | 5.66% |
| model/bert-large-uncased | f16, bs=8, seq=128 | 5.385 | 4.913 | 4.789
| 4.451 | 4.7 | 5.30% |
| model/bert-large-uncased | f16, bs=8, seq=512 | 19.817 | 19.994 |
19.693 | 16.324 | 17.269 | 5.47% |
## Regression Pipeline
A10
This PR only enables the optimization for `matmul_f16` so models that
don't have this operator won't have performance gains.
## TODO
- [x] epilogue fusion for dynamic shape
- [x] remove rearrange operator
- [x] fallback support
---------
Co-authored-by: xiaocenxiaocen <xiao.zhang@centml.ai>
vadiklyutiy
pushed a commit
that referenced
this pull request
Jul 23, 2024
This PR supports the vectorization of epilogue fusion.
Originally, we fused the operators belonging to the epilogue stage on
scalars.
For example, the code before fusion is as follows
```python
for mi in range(mma_m):
for ni in range(mma_n):
global_m = mi * stride_m
global_n = ni * stride_n
if global_m < m and global_n < n:
# the following statement always be converted to a LDG.U16 instruction at SASS level if we assume `regs_c` is a float16 register array.
# the if statement prevents nvcc to aggressively do vectorized writing back to global memory.
gmem_c[global_m, global_n] = regs_c[mi, ni]
```
The code after fusion looks like
```python
for mi in range(mma_m):
for ni in range(mma_n):
global_m = mi * stride_m
global_n = ni * stride_n
if global_m < m and global_n < n:
# x, y, z are fused tensors
gmem_c[inverse_map(global_m, global_n)] = elementwise_op(regs_c[mi, ni], x, y, z...)
```
The fusion still emits some scalar store instructions to perform the
final write-back, which is inefficient. Although fusion can bring us
some performance benefits, the penalty of inefficient write-back can be
higher than the benefit of fusion itself.
This PR is trying to fuse operators while vectorizing the memory
accesses as much as possible. Basically, we developed a technique
similar to the [epilogue visitor
tree](https://dl.acm.org/doi/pdf/10.1145/3620666.3651369) in CUTLASS to
address this problem. For details, you can refer to the attached
[document](https://drive.google.com/drive/folders/1kXlu2k-lPmH0jPlQ_8CfmqZvAPVtQ3Qs).
## Changes
1. restructure the instruction selection as a compiler pass
2. support the vectorized epilogue fusion algorithm in the
`apply_prologue_epilogue` pass.
3. adapt the schedule template of `matmul_f16_pk` to enable this
optimization.
4. currently, we have a very simple shared memory updater, and we should
design a shared memory planner. A potential proposal can be found in
[this document](https://dl.acm.org/doi/pdf/10.5555/314500.315082), which
is the shared memory allocator used in triton.
## Simple epilogue fusion
This experiment benchmarks the following graph. `dims` is a permutation
of the four axes 0, 1, 2, 3. We iterate over all the possible
permutations to show the generality of fusion capability.
```
def graph(a, b, d, bx1xn, bxmx1, mx1, x1xn):
c = a @ b
c = c + d
c = c + bxmx1 + bx1xn + mx1 + x1xn
L, M, N = c.shape
c = c.reshape(L, M, N // head_size, head_size)
c = c.permute(*dims)
c = c.reshape(L * N // head_size, M, head_size)
c = relu(c)
return c
```
Performance comparison
| problem(m,n,k,l) | dims | max-autotune | hidet(after) | hidet(before)
| speedup (after vs before) |
| -------------------- | ------------ | ------------ | ------------- |
------------ | ------- |
| [512, 1536, 512, 16] | (0, 1, 2, 3) | 0.263 | 0.199 | 0.229 | 13.10% |
| [512, 1536, 512, 16] | (0, 1, 3, 2) | 0.25 | 0.218 | 0.27 | 19.26% |
| [512, 1536, 512, 16] | (0, 2, 1, 3) | 0.248 | 0.217 | 0.223 | 2.69% |
| [512, 1536, 512, 16] | (0, 2, 3, 1) | 0.374 | 0.219 | 0.224 | 2.23% |
| [512, 1536, 512, 16] | (0, 3, 1, 2) | 0.509 | 0.269 | 0.35 | 23.14% |
| [512, 1536, 512, 16] | (0, 3, 2, 1) | 0.278 | 0.226 | 0.225 | \-0.44%
|
| [512, 1536, 512, 16] | (1, 0, 2, 3) | 0.248 | 0.202 | 0.237 | 14.77% |
| [512, 1536, 512, 16] | (1, 0, 3, 2) | 0.251 | 0.222 | 0.269 | 17.47% |
| [512, 1536, 512, 16] | (1, 2, 0, 3) | 0.264 | 0.202 | 0.241 | 16.18% |
| [512, 1536, 512, 16] | (1, 2, 3, 0) | 0.247 | 0.275 | 0.274 | \-0.36%
|
| [512, 1536, 512, 16] | (1, 3, 0, 2) | 0.708 | 0.224 | 0.4 | 44.00% |
| [512, 1536, 512, 16] | (1, 3, 2, 0) | 0.368 | 0.352 | 0.435 | 19.08% |
| [512, 1536, 512, 16] | (2, 0, 1, 3) | 0.261 | 0.211 | 0.222 | 4.95% |
| [512, 1536, 512, 16] | (2, 0, 3, 1) | 0.477 | 0.221 | 0.225 | 1.78% |
| [512, 1536, 512, 16] | (2, 1, 0, 3) | 0.262 | 0.211 | 0.236 | 10.59% |
| [512, 1536, 512, 16] | (2, 1, 3, 0) | 0.268 | 0.289 | 0.288 | \-0.35%
|
| [512, 1536, 512, 16] | (2, 3, 0, 1) | 0.257 | 0.22 | 0.227 | 3.08% |
| [512, 1536, 512, 16] | (2, 3, 1, 0) | 0.251 | 0.373 | 0.355 | \-5.07%
|
| [512, 1536, 512, 16] | (3, 0, 1, 2) | 0.559 | 0.27 | 0.372 | 27.42% |
| [512, 1536, 512, 16] | (3, 0, 2, 1) | 0.376 | 0.225 | 0.223 | \-0.90%
|
| [512, 1536, 512, 16] | (3, 1, 0, 2) | 0.648 | 0.281 | 0.399 | 29.57% |
| [512, 1536, 512, 16] | (3, 1, 2, 0) | 0.624 | 0.352 | 0.474 | 25.74% |
| [512, 1536, 512, 16] | (3, 2, 0, 1) | 0.333 | 0.226 | 0.228 | 0.88% |
| [512, 1536, 512, 16] | (3, 2, 1, 0) | 0.35 | 0.354 | 0.355 | 0.28% |
## Model benchmark
Performance comparison
| model | inputs | eager | reduce-overhead | max-autotune | hidet(after)
| hidet(before) | speedup (after vs before) |
| ------------------------ | ------------------ | ------ |
--------------- | ------------ | ------------ | ------------- |
------------------------- |
| model/gpt2 | f16, bs=1, seq=128 | 2.043 | 0.557 | 0.555 | 0.7 | 0.7 |
0.00% |
| model/gpt2 | f16, bs=1, seq=512 | 2.01 | 1.26 | 1.112 | 1.204 | 1.297
| 7.17% |
| model/gpt2 | f16, bs=8, seq=128 | 2.759 | 1.97 | 1.684 | 1.635 | 1.892
| 13.58% |
| model/gpt2 | f16, bs=8, seq=512 | 13.622 | 7.915 | 6.633 | 7.014 |
8.101 | 13.42% |
| model/bert-base-uncased | f16, bs=1, seq=128 | 1.247 | 0.744 | 0.753 |
0.659 | 0.671 | 1.79% |
| model/bert-base-uncased | f16, bs=1, seq=512 | 1.498 | 1.346 | 1.394 |
1.187 | 1.211 | 1.98% |
| model/bert-base-uncased | f16, bs=8, seq=128 | 2.045 | 1.844 | 1.723 |
1.526 | 1.637 | 6.78% |
| model/bert-base-uncased | f16, bs=8, seq=512 | 7.115 | 7.205 | 6.239 |
5.597 | 5.933 | 5.66% |
| model/bert-large-uncased | f16, bs=1, seq=128 | 2.342 | 1.835 | 1.974
| 1.515 | 1.496 | \-1.27% |
| model/bert-large-uncased | f16, bs=1, seq=512 | 3.586 | 3.367 | 3.63 |
2.932 | 3.108 | 5.66% |
| model/bert-large-uncased | f16, bs=8, seq=128 | 5.385 | 4.913 | 4.789
| 4.451 | 4.7 | 5.30% |
| model/bert-large-uncased | f16, bs=8, seq=512 | 19.817 | 19.994 |
19.693 | 16.324 | 17.269 | 5.47% |
## Regression Pipeline
A10
This PR only enables the optimization for `matmul_f16` so models that
don't have this operator won't have performance gains.
## TODO
- [x] epilogue fusion for dynamic shape
- [x] remove rearrange operator
- [x] fallback support
---------
Co-authored-by: xiaocenxiaocen <xiao.zhang@centml.ai>
vadiklyutiy
pushed a commit
that referenced
this pull request
Dec 26, 2024
This PR supports the vectorization of epilogue fusion.
Originally, we fused the operators belonging to the epilogue stage on
scalars.
For example, the code before fusion is as follows
```python
for mi in range(mma_m):
for ni in range(mma_n):
global_m = mi * stride_m
global_n = ni * stride_n
if global_m < m and global_n < n:
# the following statement always be converted to a LDG.U16 instruction at SASS level if we assume `regs_c` is a float16 register array.
# the if statement prevents nvcc to aggressively do vectorized writing back to global memory.
gmem_c[global_m, global_n] = regs_c[mi, ni]
```
The code after fusion looks like
```python
for mi in range(mma_m):
for ni in range(mma_n):
global_m = mi * stride_m
global_n = ni * stride_n
if global_m < m and global_n < n:
# x, y, z are fused tensors
gmem_c[inverse_map(global_m, global_n)] = elementwise_op(regs_c[mi, ni], x, y, z...)
```
The fusion still emits some scalar store instructions to perform the
final write-back, which is inefficient. Although fusion can bring us
some performance benefits, the penalty of inefficient write-back can be
higher than the benefit of fusion itself.
This PR is trying to fuse operators while vectorizing the memory
accesses as much as possible. Basically, we developed a technique
similar to the [epilogue visitor
tree](https://dl.acm.org/doi/pdf/10.1145/3620666.3651369) in CUTLASS to
address this problem. For details, you can refer to the attached
[document](https://drive.google.com/drive/folders/1kXlu2k-lPmH0jPlQ_8CfmqZvAPVtQ3Qs).
## Changes
1. restructure the instruction selection as a compiler pass
2. support the vectorized epilogue fusion algorithm in the
`apply_prologue_epilogue` pass.
3. adapt the schedule template of `matmul_f16_pk` to enable this
optimization.
4. currently, we have a very simple shared memory updater, and we should
design a shared memory planner. A potential proposal can be found in
[this document](https://dl.acm.org/doi/pdf/10.5555/314500.315082), which
is the shared memory allocator used in triton.
## Simple epilogue fusion
This experiment benchmarks the following graph. `dims` is a permutation
of the four axes 0, 1, 2, 3. We iterate over all the possible
permutations to show the generality of fusion capability.
```
def graph(a, b, d, bx1xn, bxmx1, mx1, x1xn):
c = a @ b
c = c + d
c = c + bxmx1 + bx1xn + mx1 + x1xn
L, M, N = c.shape
c = c.reshape(L, M, N // head_size, head_size)
c = c.permute(*dims)
c = c.reshape(L * N // head_size, M, head_size)
c = relu(c)
return c
```
Performance comparison
| problem(m,n,k,l) | dims | max-autotune | hidet(after) | hidet(before)
| speedup (after vs before) |
| -------------------- | ------------ | ------------ | ------------- |
------------ | ------- |
| [512, 1536, 512, 16] | (0, 1, 2, 3) | 0.263 | 0.199 | 0.229 | 13.10% |
| [512, 1536, 512, 16] | (0, 1, 3, 2) | 0.25 | 0.218 | 0.27 | 19.26% |
| [512, 1536, 512, 16] | (0, 2, 1, 3) | 0.248 | 0.217 | 0.223 | 2.69% |
| [512, 1536, 512, 16] | (0, 2, 3, 1) | 0.374 | 0.219 | 0.224 | 2.23% |
| [512, 1536, 512, 16] | (0, 3, 1, 2) | 0.509 | 0.269 | 0.35 | 23.14% |
| [512, 1536, 512, 16] | (0, 3, 2, 1) | 0.278 | 0.226 | 0.225 | \-0.44%
|
| [512, 1536, 512, 16] | (1, 0, 2, 3) | 0.248 | 0.202 | 0.237 | 14.77% |
| [512, 1536, 512, 16] | (1, 0, 3, 2) | 0.251 | 0.222 | 0.269 | 17.47% |
| [512, 1536, 512, 16] | (1, 2, 0, 3) | 0.264 | 0.202 | 0.241 | 16.18% |
| [512, 1536, 512, 16] | (1, 2, 3, 0) | 0.247 | 0.275 | 0.274 | \-0.36%
|
| [512, 1536, 512, 16] | (1, 3, 0, 2) | 0.708 | 0.224 | 0.4 | 44.00% |
| [512, 1536, 512, 16] | (1, 3, 2, 0) | 0.368 | 0.352 | 0.435 | 19.08% |
| [512, 1536, 512, 16] | (2, 0, 1, 3) | 0.261 | 0.211 | 0.222 | 4.95% |
| [512, 1536, 512, 16] | (2, 0, 3, 1) | 0.477 | 0.221 | 0.225 | 1.78% |
| [512, 1536, 512, 16] | (2, 1, 0, 3) | 0.262 | 0.211 | 0.236 | 10.59% |
| [512, 1536, 512, 16] | (2, 1, 3, 0) | 0.268 | 0.289 | 0.288 | \-0.35%
|
| [512, 1536, 512, 16] | (2, 3, 0, 1) | 0.257 | 0.22 | 0.227 | 3.08% |
| [512, 1536, 512, 16] | (2, 3, 1, 0) | 0.251 | 0.373 | 0.355 | \-5.07%
|
| [512, 1536, 512, 16] | (3, 0, 1, 2) | 0.559 | 0.27 | 0.372 | 27.42% |
| [512, 1536, 512, 16] | (3, 0, 2, 1) | 0.376 | 0.225 | 0.223 | \-0.90%
|
| [512, 1536, 512, 16] | (3, 1, 0, 2) | 0.648 | 0.281 | 0.399 | 29.57% |
| [512, 1536, 512, 16] | (3, 1, 2, 0) | 0.624 | 0.352 | 0.474 | 25.74% |
| [512, 1536, 512, 16] | (3, 2, 0, 1) | 0.333 | 0.226 | 0.228 | 0.88% |
| [512, 1536, 512, 16] | (3, 2, 1, 0) | 0.35 | 0.354 | 0.355 | 0.28% |
## Model benchmark
Performance comparison
| model | inputs | eager | reduce-overhead | max-autotune | hidet(after)
| hidet(before) | speedup (after vs before) |
| ------------------------ | ------------------ | ------ |
--------------- | ------------ | ------------ | ------------- |
------------------------- |
| model/gpt2 | f16, bs=1, seq=128 | 2.043 | 0.557 | 0.555 | 0.7 | 0.7 |
0.00% |
| model/gpt2 | f16, bs=1, seq=512 | 2.01 | 1.26 | 1.112 | 1.204 | 1.297
| 7.17% |
| model/gpt2 | f16, bs=8, seq=128 | 2.759 | 1.97 | 1.684 | 1.635 | 1.892
| 13.58% |
| model/gpt2 | f16, bs=8, seq=512 | 13.622 | 7.915 | 6.633 | 7.014 |
8.101 | 13.42% |
| model/bert-base-uncased | f16, bs=1, seq=128 | 1.247 | 0.744 | 0.753 |
0.659 | 0.671 | 1.79% |
| model/bert-base-uncased | f16, bs=1, seq=512 | 1.498 | 1.346 | 1.394 |
1.187 | 1.211 | 1.98% |
| model/bert-base-uncased | f16, bs=8, seq=128 | 2.045 | 1.844 | 1.723 |
1.526 | 1.637 | 6.78% |
| model/bert-base-uncased | f16, bs=8, seq=512 | 7.115 | 7.205 | 6.239 |
5.597 | 5.933 | 5.66% |
| model/bert-large-uncased | f16, bs=1, seq=128 | 2.342 | 1.835 | 1.974
| 1.515 | 1.496 | \-1.27% |
| model/bert-large-uncased | f16, bs=1, seq=512 | 3.586 | 3.367 | 3.63 |
2.932 | 3.108 | 5.66% |
| model/bert-large-uncased | f16, bs=8, seq=128 | 5.385 | 4.913 | 4.789
| 4.451 | 4.7 | 5.30% |
| model/bert-large-uncased | f16, bs=8, seq=512 | 19.817 | 19.994 |
19.693 | 16.324 | 17.269 | 5.47% |
## Regression Pipeline
A10
This PR only enables the optimization for `matmul_f16` so models that
don't have this operator won't have performance gains.
## TODO
- [x] epilogue fusion for dynamic shape
- [x] remove rearrange operator
- [x] fallback support
---------
Co-authored-by: xiaocenxiaocen <xiao.zhang@centml.ai>
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.