[ET-VK] Add kInt8x4 dtype and GPUMemoryLayouts for packed quantized tensors
#14329
Conversation
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
…ed tensors ## Motivation Lay the foundations for being able to execute statically quantized CNNs with ET-VK. Unlike with dynamic quantization, static quantization allows the output of quantized operators to stay in integer representation and be fed directly to the next quantized operator. ## Context Typically, int8 quantized tensors can be represented by simply having the tensor use the int8 data type. While this is possible in ET-VK, in practice quantized operators expect int8 quantized tensors to be packed so that 16 8-bit values are packed into each `ivec4`, such that quantized int8 tensors will load/store with a granularity of 16 elements. The reason for this is twofold: * Support for shader int8 / storage buffer int8 extension is not guaranteed, meaning some devices do not allow using int8 types in shaders * We have found that load/store from storage buffers/textures that use int8 data types sometimes results in worse memory load performance, due to vectorized load/store instructions not being used. Therefore, in ET-VK we need a way to mark that a quantized tensor should 1. Use int32 as the underlying data type for the storage buffer/texture 2. Account for the block-packing that may be used ## Changes First, introduce the `Int8x4` dtype that can be used for packed int8 tensors. This dtype is functionally the same as `Int`, but denotes that each int32 actually contains 4 packed 8-bit values. Second, introduce new memory layouts: `kPackedInt8_4W4C` and `kPackedInt8_4H4W`. The former will be used for convolution, whil the latter will be used for matrix multiplication. See the inline comments for more details about these memory layouts. Then, update `QuantizedConvolution.cpp` and `QuantizedLinear.cpp` to use the new data type and memory layouts for the packed int8 input tensor. Differential Revision: [D82542336](https://our.internmc.facebook.com/intern/diff/D82542336/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14329
Note: Links to docs will display an error until the docs builds have been completed. ❌ 4 New Failures, 5 Unrelated FailuresAs of commit 6a4879f with merge base c18abc8 ( NEW FAILURES - The following jobs have failed:
FLAKY - The following job failed but was likely due to flakiness present on trunk:
BROKEN TRUNK - The following jobs failed but was present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
This PR needs a
|
…ked quantized tensors" ## Motivation Lay the foundations for being able to execute statically quantized CNNs with ET-VK. Unlike with dynamic quantization, static quantization allows the output of quantized operators to stay in integer representation and be fed directly to the next quantized operator. ## Context Typically, int8 quantized tensors can be represented by simply having the tensor use the int8 data type. While this is possible in ET-VK, in practice quantized operators expect int8 quantized tensors to be packed so that 16 8-bit values are packed into each `ivec4`, such that quantized int8 tensors will load/store with a granularity of 16 elements. The reason for this is twofold: * Support for shader int8 / storage buffer int8 extension is not guaranteed, meaning some devices do not allow using int8 types in shaders * We have found that load/store from storage buffers/textures that use int8 data types sometimes results in worse memory load performance, due to vectorized load/store instructions not being used. Therefore, in ET-VK we need a way to mark that a quantized tensor should 1. Use int32 as the underlying data type for the storage buffer/texture 2. Account for the block-packing that may be used ## Changes First, introduce the `Int8x4` dtype that can be used for packed int8 tensors. This dtype is functionally the same as `Int`, but denotes that each int32 actually contains 4 packed 8-bit values. Second, introduce new memory layouts: `kPackedInt8_4W4C` and `kPackedInt8_4H4W`. The former will be used for convolution, whil the latter will be used for matrix multiplication. See the inline comments for more details about these memory layouts. Then, update `QuantizedConvolution.cpp` and `QuantizedLinear.cpp` to use the new data type and memory layouts for the packed int8 input tensor. Differential Revision: [D82542336](https://our.internmc.facebook.com/intern/diff/D82542336/) [ghstack-poisoned]
…ked quantized tensors" ## Motivation Lay the foundations for being able to execute statically quantized CNNs with ET-VK. Unlike with dynamic quantization, static quantization allows the output of quantized operators to stay in integer representation and be fed directly to the next quantized operator. ## Context Typically, int8 quantized tensors can be represented by simply having the tensor use the int8 data type. While this is possible in ET-VK, in practice quantized operators expect int8 quantized tensors to be packed so that 16 8-bit values are packed into each `ivec4`, such that quantized int8 tensors will load/store with a granularity of 16 elements. The reason for this is twofold: * Support for shader int8 / storage buffer int8 extension is not guaranteed, meaning some devices do not allow using int8 types in shaders * We have found that load/store from storage buffers/textures that use int8 data types sometimes results in worse memory load performance, due to vectorized load/store instructions not being used. Therefore, in ET-VK we need a way to mark that a quantized tensor should 1. Use int32 as the underlying data type for the storage buffer/texture 2. Account for the block-packing that may be used ## Changes First, introduce the `Int8x4` dtype that can be used for packed int8 tensors. This dtype is functionally the same as `Int`, but denotes that each int32 actually contains 4 packed 8-bit values. Second, introduce new memory layouts: `kPackedInt8_4W4C` and `kPackedInt8_4H4W`. The former will be used for convolution, whil the latter will be used for matrix multiplication. See the inline comments for more details about these memory layouts. Then, update `QuantizedConvolution.cpp` and `QuantizedLinear.cpp` to use the new data type and memory layouts for the packed int8 input tensor. Differential Revision: [D82542336](https://our.internmc.facebook.com/intern/diff/D82542336/) [ghstack-poisoned]
…ked quantized tensors" ## Motivation Lay the foundations for being able to execute statically quantized CNNs with ET-VK. Unlike with dynamic quantization, static quantization allows the output of quantized operators to stay in integer representation and be fed directly to the next quantized operator. ## Context Typically, int8 quantized tensors can be represented by simply having the tensor use the int8 data type. While this is possible in ET-VK, in practice quantized operators expect int8 quantized tensors to be packed so that 16 8-bit values are packed into each `ivec4`, such that quantized int8 tensors will load/store with a granularity of 16 elements. The reason for this is twofold: * Support for shader int8 / storage buffer int8 extension is not guaranteed, meaning some devices do not allow using int8 types in shaders * We have found that load/store from storage buffers/textures that use int8 data types sometimes results in worse memory load performance, due to vectorized load/store instructions not being used. Therefore, in ET-VK we need a way to mark that a quantized tensor should 1. Use int32 as the underlying data type for the storage buffer/texture 2. Account for the block-packing that may be used ## Changes First, introduce the `Int8x4` dtype that can be used for packed int8 tensors. This dtype is functionally the same as `Int`, but denotes that each int32 actually contains 4 packed 8-bit values. Second, introduce new memory layouts: `kPackedInt8_4W4C` and `kPackedInt8_4H4W`. The former will be used for convolution, whil the latter will be used for matrix multiplication. See the inline comments for more details about these memory layouts. Then, update `QuantizedConvolution.cpp` and `QuantizedLinear.cpp` to use the new data type and memory layouts for the packed int8 input tensor. Differential Revision: [D82542336](https://our.internmc.facebook.com/intern/diff/D82542336/) [ghstack-poisoned]
…ked quantized tensors" ## Motivation Lay the foundations for being able to execute statically quantized CNNs with ET-VK. Unlike with dynamic quantization, static quantization allows the output of quantized operators to stay in integer representation and be fed directly to the next quantized operator. ## Context Typically, int8 quantized tensors can be represented by simply having the tensor use the int8 data type. While this is possible in ET-VK, in practice quantized operators expect int8 quantized tensors to be packed so that 16 8-bit values are packed into each `ivec4`, such that quantized int8 tensors will load/store with a granularity of 16 elements. The reason for this is twofold: * Support for shader int8 / storage buffer int8 extension is not guaranteed, meaning some devices do not allow using int8 types in shaders * We have found that load/store from storage buffers/textures that use int8 data types sometimes results in worse memory load performance, due to vectorized load/store instructions not being used. Therefore, in ET-VK we need a way to mark that a quantized tensor should 1. Use int32 as the underlying data type for the storage buffer/texture 2. Account for the block-packing that may be used ## Changes First, introduce the `Int8x4` dtype that can be used for packed int8 tensors. This dtype is functionally the same as `Int`, but denotes that each int32 actually contains 4 packed 8-bit values. Second, introduce new memory layouts: `kPackedInt8_4W4C` and `kPackedInt8_4H4W`. The former will be used for convolution, whil the latter will be used for matrix multiplication. See the inline comments for more details about these memory layouts. Then, update `QuantizedConvolution.cpp` and `QuantizedLinear.cpp` to use the new data type and memory layouts for the packed int8 input tensor. Differential Revision: [D82542336](https://our.internmc.facebook.com/intern/diff/D82542336/) [ghstack-poisoned]
manuelcandales
approved these changes
Sep 25, 2025
facebook-github-bot
merged commit 307a138
into
gh/SS-JIA/329/base
121 of 132 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Stack from ghstack (oldest at bottom):
kInt8x4dtype andGPUMemoryLayouts for packed quantized tensors #14329Motivation
Lay the foundations for being able to execute statically quantized CNNs with ET-VK. Unlike with dynamic quantization, static quantization allows the output of quantized operators to stay in integer representation and be fed directly to the next quantized operator.
Context
Typically, int8 quantized tensors can be represented by simply having the tensor use the int8 data type. While this is possible in ET-VK, in practice quantized operators expect int8 quantized tensors to be packed so that 16 8-bit values are packed into each
ivec4, such that quantized int8 tensors will load/store with a granularity of 16 elements.The reason for this is twofold:
Therefore, in ET-VK we need a way to mark that a quantized tensor should
Changes
First, introduce the
Int8x4dtype that can be used for packed int8 tensors. This dtype is functionally the same asInt, but denotes that each int32 actually contains 4 packed 8-bit values.Second, introduce new memory layouts:
kPackedInt8_4W4CandkPackedInt8_4H4W. The former will be used for convolution, whil the latter will be used for matrix multiplication. See the inline comments for more details about these memory layouts.Then, update
QuantizedConvolution.cppandQuantizedLinear.cppto use the new data type and memory layouts for the packed int8 input tensor.Differential Revision: D82542336