[ET-VK][Ops] torchao.quantize_affine vulkan impl and shader and cleanup#12575
Merged
Conversation
# Changes
* Implement `torchao.quantize_affine` operator in Vulkan backend with comprehensive texture and buffer storage support
* Add block-wise quantization mode in `quantize_texture.glsl` and `quantize_buffer.glsl` shaders for configurable tensor block quantization
* Introduce comprehensive test suite in `affine_test.cpp` with multi-dimensional tensor validation and reference implementation
* Extend quantization infrastructure in `Quantize.cpp` to handle affine transformations with configurable block sizes and quantization parameters
BE: Improved the documentation in the shader logic which is more detailed and clear
NOTE: I delegated the quantize_affine and future affine operators through a new custom test file denoted as `affine_test.cpp` as the other quantization testing framework was getting a little large, and it makes more sense to separate the namespace between torchao and quantized_decomposed. I believe the _decomposed namespace is getting phased out in favor of this affine operator so deprecation will be easier in the future.
# Motivation
The existing Vulkan quantization infrastructure lacked support for the `torchao.quantize_affine` operator, which is essential for enabling dynamic quantization efficiently. The `quantize_affine` operator provides flexible block-wise quantization that allows different scale and zero-point values for tensor blocks, enabling:
* **Block-wise Quantization**: Applies quantization parameters to configurable tensor blocks rather than entire tensors, improving quantization accuracy for heterogeneous data distributions
* **Affine Transformation**: Uses the formula `qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)` for precise floating-point to integer mapping
# Operator Description
The `quantize_affine` operator converts floating-point tensor values to n-bit integer representations using pre-computed quantization parameters (scale and zero_point) applied to configurable tensor blocks. Block-wise quantization divides tensors into blocks and applies separate quantization parameters to each block, allowing fine-grained control over quantization precision.
The quantization formula is: `qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)`
**Storage Requirements**: Scale and zero_point tensors must use buffer storage with width-packed layout. Input/output tensors support both buffer and texture storage with standard axis mapping.
# Block-wise Quantization Implementation
Block-wise quantization enables fine-grained quantization by dividing tensors into blocks and applying separate quantization parameters to each block. The implementation uses several key data structures computed in `Quantize.cpp`:
* **`block_size_vec`**: WHCN-ordered block dimensions converted from PyTorch NCHW layout (e.g., [3,3,2,1] for 3×3×2×1 blocks)
* **`tensor_size_whcn`**: Input tensor dimensions converted to WHCN layout using `utils::make_whcn_ivec4()`
* **`num_blocks_vec`**: Number of blocks per dimension calculated as `tensor_size_whcn / block_size_vec`
* **`block_stride_vec`**: Pre-computed linear strides for block grid indexing `{1, #W, #W*#H, #W*#H*#C}` to enable efficient block ID calculation
The block coordinate calculation uses: `bcoord = tidx / blockSize` where `tidx` is the tensor coordinate in WHCN layout, then the linear block ID is computed as: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`
# Shader Algorithm Overview
## Texture Storage Implementation (`quantize_texture.glsl`)
**Workgroup Configuration**:
- **Global WG Size**: Default sizing based on texture dimensions
- **Local WG Size**: Default with special handling for batch dimension quantization (Z dimension set to 1 for proper workgroup dispatching when `global_workgroup_size[2] > 1`)
**Block-wise Mode Algorithm**:
The shader processes 3D texture positions where each position represents a texel containing 4 width-packed components. For each texel at position `pos`, it calculates a base tensor index `base_tidx = ivec4(pos.x * 4, pos.y, pos.z, 0)` to account for width-packing.
For each of the 4 components in the texel, it computes the actual tensor coordinate: `tidx = ivec4(base_tidx.x + i, base_tidx.y, (foldedZ % C_total), (foldedZ / C_total))` where `foldedZ = pos.z` handles batch-channel folding in 4D tensors and `C_total = numBlocks.z * blockSize.z` represents the total channel dimension.
The block coordinate is calculated using integer division: `bcoord = tidx / blockSize`, then the linear block ID uses pre-computed strides: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`.
Each component is quantized using its corresponding block's parameters: `qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])` and written to the output texel.
## Buffer Storage Implementation (`quantize_buffer.glsl`)
**Workgroup Configuration**:
- **Global WG Size**: Default sizing based on buffer element count
- **Local WG Size**: Default sizing without special constraints
**Block-wise Mode Algorithm**:
The shader processes linear buffer indices using `gl_GlobalInvocationID.x` as the output buffer index. It converts this to tensor coordinates using `bufi_to_tidx(out_bufi, t_out_strides, out_dim_order)` which handles the buffer-to-tensor index mapping with proper stride calculations.
For each element, it computes the block coordinate directly: `bcoord = out_tidx / blockSize` where `out_tidx` is the 4D tensor coordinate in WHCN layout. The linear block ID calculation uses the same pre-computed stride approach: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`.
The element value is loaded using the corresponding input buffer index: `value = t_in[in_bufi]` where `in_bufi = tidx_to_bufi(out_tidx, t_in_strides)`. Quantization applies the block-specific parameters: `qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])`.
**Future Improvements**: Dynamic workgroup sizing based on block dimensions, there is likely a better method to making it better than what it is currently.
Differential Revision: [D78302195](https://our.internmc.facebook.com/intern/diff/D78302195/)
[ghstack-poisoned]
pytorch-bot
Bot
added
the
module: vulkan
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/12575
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 495b62b with merge base b6b7a16 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Jul 17, 2025
meta-cla
Bot
added
the
CLA Signed
This PR needs a
|
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D78302195 |
…r and cleanup"
# Changes
* Implement `torchao.quantize_affine` operator in Vulkan backend with comprehensive texture and buffer storage support
* Add block-wise quantization mode in `quantize_texture.glsl` and `quantize_buffer.glsl` shaders for configurable tensor block quantization
* Introduce comprehensive test suite in `affine_test.cpp` with multi-dimensional tensor validation and reference implementation
* Extend quantization infrastructure in `Quantize.cpp` to handle affine transformations with configurable block sizes and quantization parameters
BE: Improved the documentation in the shader logic which is more detailed and clear
NOTE: I delegated the quantize_affine and future affine operators through a new custom test file denoted as `affine_test.cpp` as the other quantization testing framework was getting a little large, and it makes more sense to separate the namespace between torchao and quantized_decomposed. I believe the _decomposed namespace is getting phased out in favor of this affine operator so deprecation will be easier in the future.
# Motivation
The existing Vulkan quantization infrastructure lacked support for the `torchao.quantize_affine` operator, which is essential for enabling dynamic quantization efficiently. The `quantize_affine` operator provides flexible block-wise quantization that allows different scale and zero-point values for tensor blocks, enabling:
* **Block-wise Quantization**: Applies quantization parameters to configurable tensor blocks rather than entire tensors, improving quantization accuracy for heterogeneous data distributions
* **Affine Transformation**: Uses the formula `qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)` for precise floating-point to integer mapping
# Operator Description
The `quantize_affine` operator converts floating-point tensor values to n-bit integer representations using pre-computed quantization parameters (scale and zero_point) applied to configurable tensor blocks. Block-wise quantization divides tensors into blocks and applies separate quantization parameters to each block, allowing fine-grained control over quantization precision.
The quantization formula is: `qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)`
**Storage Requirements**: Scale and zero_point tensors must use buffer storage with width-packed layout. Input/output tensors support both buffer and texture storage with standard axis mapping.
# Block-wise Quantization Implementation
Block-wise quantization enables fine-grained quantization by dividing tensors into blocks and applying separate quantization parameters to each block. The implementation uses several key data structures computed in `Quantize.cpp`:
* **`block_size_vec`**: WHCN-ordered block dimensions converted from PyTorch NCHW layout (e.g., [3,3,2,1] for 3×3×2×1 blocks)
* **`tensor_size_whcn`**: Input tensor dimensions converted to WHCN layout using `utils::make_whcn_ivec4()`
* **`num_blocks_vec`**: Number of blocks per dimension calculated as `tensor_size_whcn / block_size_vec`
* **`block_stride_vec`**: Pre-computed linear strides for block grid indexing `{1, #W, #W*#H, #W*#H*#C}` to enable efficient block ID calculation
The block coordinate calculation uses: `bcoord = tidx / blockSize` where `tidx` is the tensor coordinate in WHCN layout, then the linear block ID is computed as: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`
# Shader Algorithm Overview
## Texture Storage Implementation (`quantize_texture.glsl`)
**Workgroup Configuration**:
- **Global WG Size**: Default sizing based on texture dimensions
- **Local WG Size**: Default with special handling for batch dimension quantization (Z dimension set to 1 for proper workgroup dispatching when `global_workgroup_size[2] > 1`)
**Block-wise Mode Algorithm**:
The shader processes 3D texture positions where each position represents a texel containing 4 width-packed components. For each texel at position `pos`, it calculates a base tensor index `base_tidx = ivec4(pos.x * 4, pos.y, pos.z, 0)` to account for width-packing.
For each of the 4 components in the texel, it computes the actual tensor coordinate: `tidx = ivec4(base_tidx.x + i, base_tidx.y, (foldedZ % C_total), (foldedZ / C_total))` where `foldedZ = pos.z` handles batch-channel folding in 4D tensors and `C_total = numBlocks.z * blockSize.z` represents the total channel dimension.
The block coordinate is calculated using integer division: `bcoord = tidx / blockSize`, then the linear block ID uses pre-computed strides: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`.
Each component is quantized using its corresponding block's parameters: `qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])` and written to the output texel.
## Buffer Storage Implementation (`quantize_buffer.glsl`)
**Workgroup Configuration**:
- **Global WG Size**: Default sizing based on buffer element count
- **Local WG Size**: Default sizing without special constraints
**Block-wise Mode Algorithm**:
The shader processes linear buffer indices using `gl_GlobalInvocationID.x` as the output buffer index. It converts this to tensor coordinates using `bufi_to_tidx(out_bufi, t_out_strides, out_dim_order)` which handles the buffer-to-tensor index mapping with proper stride calculations.
For each element, it computes the block coordinate directly: `bcoord = out_tidx / blockSize` where `out_tidx` is the 4D tensor coordinate in WHCN layout. The linear block ID calculation uses the same pre-computed stride approach: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`.
The element value is loaded using the corresponding input buffer index: `value = t_in[in_bufi]` where `in_bufi = tidx_to_bufi(out_tidx, t_in_strides)`. Quantization applies the block-specific parameters: `qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])`.
**Future Improvements**: Dynamic workgroup sizing based on block dimensions, there is likely a better method to making it better than what it is currently.
Differential Revision: [D78302195](https://our.internmc.facebook.com/intern/diff/D78302195/)
cc SS-JIA manuelcandales cbilgin
[ghstack-poisoned]
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D78302195 |
…r and cleanup"
# Changes
* Implement `torchao.quantize_affine` operator in Vulkan backend with comprehensive texture and buffer storage support
* Add block-wise quantization mode in `quantize_texture.glsl` and `quantize_buffer.glsl` shaders for configurable tensor block quantization
* Introduce comprehensive test suite in `affine_test.cpp` with multi-dimensional tensor validation and reference implementation
* Extend quantization infrastructure in `Quantize.cpp` to handle affine transformations with configurable block sizes and quantization parameters
BE: Improved the documentation in the shader logic which is more detailed and clear
NOTE: I delegated the quantize_affine and future affine operators through a new custom test file denoted as `affine_test.cpp` as the other quantization testing framework was getting a little large, and it makes more sense to separate the namespace between torchao and quantized_decomposed. I believe the _decomposed namespace is getting phased out in favor of this affine operator so deprecation will be easier in the future.
# Motivation
The existing Vulkan quantization infrastructure lacked support for the `torchao.quantize_affine` operator, which is essential for enabling dynamic quantization efficiently. The `quantize_affine` operator provides flexible block-wise quantization that allows different scale and zero-point values for tensor blocks, enabling:
* **Block-wise Quantization**: Applies quantization parameters to configurable tensor blocks rather than entire tensors, improving quantization accuracy for heterogeneous data distributions
* **Affine Transformation**: Uses the formula `qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)` for precise floating-point to integer mapping
# Operator Description
The `quantize_affine` operator converts floating-point tensor values to n-bit integer representations using pre-computed quantization parameters (scale and zero_point) applied to configurable tensor blocks. Block-wise quantization divides tensors into blocks and applies separate quantization parameters to each block, allowing fine-grained control over quantization precision.
The quantization formula is: `qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)`
**Storage Requirements**: Scale and zero_point tensors must use buffer storage with width-packed layout. Input/output tensors support both buffer and texture storage with standard axis mapping.
# Block-wise Quantization Implementation
Block-wise quantization enables fine-grained quantization by dividing tensors into blocks and applying separate quantization parameters to each block. The implementation uses several key data structures computed in `Quantize.cpp`:
* **`block_size_vec`**: WHCN-ordered block dimensions converted from PyTorch NCHW layout (e.g., [3,3,2,1] for 3×3×2×1 blocks)
* **`tensor_size_whcn`**: Input tensor dimensions converted to WHCN layout using `utils::make_whcn_ivec4()`
* **`num_blocks_vec`**: Number of blocks per dimension calculated as `tensor_size_whcn / block_size_vec`
* **`block_stride_vec`**: Pre-computed linear strides for block grid indexing `{1, #W, #W*#H, #W*#H*#C}` to enable efficient block ID calculation
The block coordinate calculation uses: `bcoord = tidx / blockSize` where `tidx` is the tensor coordinate in WHCN layout, then the linear block ID is computed as: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`
# Shader Algorithm Overview
## Texture Storage Implementation (`quantize_texture.glsl`)
**Workgroup Configuration**:
- **Global WG Size**: Default sizing based on texture dimensions
- **Local WG Size**: Default with special handling for batch dimension quantization (Z dimension set to 1 for proper workgroup dispatching when `global_workgroup_size[2] > 1`)
**Block-wise Mode Algorithm**:
The shader processes 3D texture positions where each position represents a texel containing 4 width-packed components. For each texel at position `pos`, it calculates a base tensor index `base_tidx = ivec4(pos.x * 4, pos.y, pos.z, 0)` to account for width-packing.
For each of the 4 components in the texel, it computes the actual tensor coordinate: `tidx = ivec4(base_tidx.x + i, base_tidx.y, (foldedZ % C_total), (foldedZ / C_total))` where `foldedZ = pos.z` handles batch-channel folding in 4D tensors and `C_total = numBlocks.z * blockSize.z` represents the total channel dimension.
The block coordinate is calculated using integer division: `bcoord = tidx / blockSize`, then the linear block ID uses pre-computed strides: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`.
Each component is quantized using its corresponding block's parameters: `qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])` and written to the output texel.
## Buffer Storage Implementation (`quantize_buffer.glsl`)
**Workgroup Configuration**:
- **Global WG Size**: Default sizing based on buffer element count
- **Local WG Size**: Default sizing without special constraints
**Block-wise Mode Algorithm**:
The shader processes linear buffer indices using `gl_GlobalInvocationID.x` as the output buffer index. It converts this to tensor coordinates using `bufi_to_tidx(out_bufi, t_out_strides, out_dim_order)` which handles the buffer-to-tensor index mapping with proper stride calculations.
For each element, it computes the block coordinate directly: `bcoord = out_tidx / blockSize` where `out_tidx` is the 4D tensor coordinate in WHCN layout. The linear block ID calculation uses the same pre-computed stride approach: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`.
The element value is loaded using the corresponding input buffer index: `value = t_in[in_bufi]` where `in_bufi = tidx_to_bufi(out_tidx, t_in_strides)`. Quantization applies the block-specific parameters: `qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])`.
**Future Improvements**: Dynamic workgroup sizing based on block dimensions, there is likely a better method to making it better than what it is currently.
Differential Revision: [D78302195](https://our.internmc.facebook.com/intern/diff/D78302195/)
cc SS-JIA manuelcandales cbilgin
[ghstack-poisoned]
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D78302195 |
…r and cleanup"
# Changes
* Implement `torchao.quantize_affine` operator in Vulkan backend with comprehensive texture and buffer storage support
* Add block-wise quantization mode in `quantize_texture.glsl` and `quantize_buffer.glsl` shaders for configurable tensor block quantization
* Introduce comprehensive test suite in `affine_test.cpp` with multi-dimensional tensor validation and reference implementation
* Extend quantization infrastructure in `Quantize.cpp` to handle affine transformations with configurable block sizes and quantization parameters
BE: Improved the documentation in the shader logic which is more detailed and clear
NOTE: I delegated the quantize_affine and future affine operators through a new custom test file denoted as `affine_test.cpp` as the other quantization testing framework was getting a little large, and it makes more sense to separate the namespace between torchao and quantized_decomposed. I believe the _decomposed namespace is getting phased out in favor of this affine operator so deprecation will be easier in the future.
# Motivation
The existing Vulkan quantization infrastructure lacked support for the `torchao.quantize_affine` operator, which is essential for enabling dynamic quantization efficiently. The `quantize_affine` operator provides flexible block-wise quantization that allows different scale and zero-point values for tensor blocks, enabling:
* **Block-wise Quantization**: Applies quantization parameters to configurable tensor blocks rather than entire tensors, improving quantization accuracy for heterogeneous data distributions
* **Affine Transformation**: Uses the formula `qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)` for precise floating-point to integer mapping
# Operator Description
The `quantize_affine` operator converts floating-point tensor values to n-bit integer representations using pre-computed quantization parameters (scale and zero_point) applied to configurable tensor blocks. Block-wise quantization divides tensors into blocks and applies separate quantization parameters to each block, allowing fine-grained control over quantization precision.
The quantization formula is: `qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)`
**Storage Requirements**: Scale and zero_point tensors must use buffer storage with width-packed layout. Input/output tensors support both buffer and texture storage with standard axis mapping.
# Block-wise Quantization Implementation
Block-wise quantization enables fine-grained quantization by dividing tensors into blocks and applying separate quantization parameters to each block. The implementation uses several key data structures computed in `Quantize.cpp`:
* **`block_size_vec`**: WHCN-ordered block dimensions converted from PyTorch NCHW layout (e.g., [3,3,2,1] for 3×3×2×1 blocks)
* **`tensor_size_whcn`**: Input tensor dimensions converted to WHCN layout using `utils::make_whcn_ivec4()`
* **`num_blocks_vec`**: Number of blocks per dimension calculated as `tensor_size_whcn / block_size_vec`
* **`block_stride_vec`**: Pre-computed linear strides for block grid indexing `{1, #W, #W*#H, #W*#H*#C}` to enable efficient block ID calculation
The block coordinate calculation uses: `bcoord = tidx / blockSize` where `tidx` is the tensor coordinate in WHCN layout, then the linear block ID is computed as: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`
# Shader Algorithm Overview
## Texture Storage Implementation (`quantize_texture.glsl`)
**Workgroup Configuration**:
- **Global WG Size**: Default sizing based on texture dimensions
- **Local WG Size**: Default with special handling for batch dimension quantization (Z dimension set to 1 for proper workgroup dispatching when `global_workgroup_size[2] > 1`)
**Block-wise Mode Algorithm**:
The shader processes 3D texture positions where each position represents a texel containing 4 width-packed components. For each texel at position `pos`, it calculates a base tensor index `base_tidx = ivec4(pos.x * 4, pos.y, pos.z, 0)` to account for width-packing.
For each of the 4 components in the texel, it computes the actual tensor coordinate: `tidx = ivec4(base_tidx.x + i, base_tidx.y, (foldedZ % C_total), (foldedZ / C_total))` where `foldedZ = pos.z` handles batch-channel folding in 4D tensors and `C_total = numBlocks.z * blockSize.z` represents the total channel dimension.
The block coordinate is calculated using integer division: `bcoord = tidx / blockSize`, then the linear block ID uses pre-computed strides: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`.
Each component is quantized using its corresponding block's parameters: `qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])` and written to the output texel.
## Buffer Storage Implementation (`quantize_buffer.glsl`)
**Workgroup Configuration**:
- **Global WG Size**: Default sizing based on buffer element count
- **Local WG Size**: Default sizing without special constraints
**Block-wise Mode Algorithm**:
The shader processes linear buffer indices using `gl_GlobalInvocationID.x` as the output buffer index. It converts this to tensor coordinates using `bufi_to_tidx(out_bufi, t_out_strides, out_dim_order)` which handles the buffer-to-tensor index mapping with proper stride calculations.
For each element, it computes the block coordinate directly: `bcoord = out_tidx / blockSize` where `out_tidx` is the 4D tensor coordinate in WHCN layout. The linear block ID calculation uses the same pre-computed stride approach: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`.
The element value is loaded using the corresponding input buffer index: `value = t_in[in_bufi]` where `in_bufi = tidx_to_bufi(out_tidx, t_in_strides)`. Quantization applies the block-specific parameters: `qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])`.
**Future Improvements**: Dynamic workgroup sizing based on block dimensions, there is likely a better method to making it better than what it is currently.
Differential Revision: [D78302195](https://our.internmc.facebook.com/intern/diff/D78302195/)
cc SS-JIA manuelcandales cbilgin
[ghstack-poisoned]
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D78302195 |
SS-JIA
approved these changes
Jul 29, 2025
…r and cleanup"
# Changes
* Implement `torchao.quantize_affine` operator in Vulkan backend with comprehensive texture and buffer storage support
* Add block-wise quantization mode in `quantize_texture.glsl` and `quantize_buffer.glsl` shaders for configurable tensor block quantization
* Introduce comprehensive test suite in `affine_test.cpp` with multi-dimensional tensor validation and reference implementation
* Extend quantization infrastructure in `Quantize.cpp` to handle affine transformations with configurable block sizes and quantization parameters
BE: Improved the documentation in the shader logic which is more detailed and clear
NOTE: I delegated the quantize_affine and future affine operators through a new custom test file denoted as `affine_test.cpp` as the other quantization testing framework was getting a little large, and it makes more sense to separate the namespace between torchao and quantized_decomposed. I believe the _decomposed namespace is getting phased out in favor of this affine operator so deprecation will be easier in the future.
# Motivation
The existing Vulkan quantization infrastructure lacked support for the `torchao.quantize_affine` operator, which is essential for enabling dynamic quantization efficiently. The `quantize_affine` operator provides flexible block-wise quantization that allows different scale and zero-point values for tensor blocks, enabling:
* **Block-wise Quantization**: Applies quantization parameters to configurable tensor blocks rather than entire tensors, improving quantization accuracy for heterogeneous data distributions
* **Affine Transformation**: Uses the formula `qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)` for precise floating-point to integer mapping
# Operator Description
The `quantize_affine` operator converts floating-point tensor values to n-bit integer representations using pre-computed quantization parameters (scale and zero_point) applied to configurable tensor blocks. Block-wise quantization divides tensors into blocks and applies separate quantization parameters to each block, allowing fine-grained control over quantization precision.
The quantization formula is: `qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)`
**Storage Requirements**: Scale and zero_point tensors must use buffer storage with width-packed layout. Input/output tensors support both buffer and texture storage with standard axis mapping.
# Block-wise Quantization Implementation
Block-wise quantization enables fine-grained quantization by dividing tensors into blocks and applying separate quantization parameters to each block. The implementation uses several key data structures computed in `Quantize.cpp`:
* **`block_size_vec`**: WHCN-ordered block dimensions converted from PyTorch NCHW layout (e.g., [3,3,2,1] for 3×3×2×1 blocks)
* **`tensor_size_whcn`**: Input tensor dimensions converted to WHCN layout using `utils::make_whcn_ivec4()`
* **`num_blocks_vec`**: Number of blocks per dimension calculated as `tensor_size_whcn / block_size_vec`
* **`block_stride_vec`**: Pre-computed linear strides for block grid indexing `{1, #W, #W*#H, #W*#H*#C}` to enable efficient block ID calculation
The block coordinate calculation uses: `bcoord = tidx / blockSize` where `tidx` is the tensor coordinate in WHCN layout, then the linear block ID is computed as: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`
# Shader Algorithm Overview
## Texture Storage Implementation (`quantize_texture.glsl`)
**Workgroup Configuration**:
- **Global WG Size**: Default sizing based on texture dimensions
- **Local WG Size**: Default with special handling for batch dimension quantization (Z dimension set to 1 for proper workgroup dispatching when `global_workgroup_size[2] > 1`)
**Block-wise Mode Algorithm**:
The shader processes 3D texture positions where each position represents a texel containing 4 width-packed components. For each texel at position `pos`, it calculates a base tensor index `base_tidx = ivec4(pos.x * 4, pos.y, pos.z, 0)` to account for width-packing.
For each of the 4 components in the texel, it computes the actual tensor coordinate: `tidx = ivec4(base_tidx.x + i, base_tidx.y, (foldedZ % C_total), (foldedZ / C_total))` where `foldedZ = pos.z` handles batch-channel folding in 4D tensors and `C_total = numBlocks.z * blockSize.z` represents the total channel dimension.
The block coordinate is calculated using integer division: `bcoord = tidx / blockSize`, then the linear block ID uses pre-computed strides: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`.
Each component is quantized using its corresponding block's parameters: `qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])` and written to the output texel.
## Buffer Storage Implementation (`quantize_buffer.glsl`)
**Workgroup Configuration**:
- **Global WG Size**: Default sizing based on buffer element count
- **Local WG Size**: Default sizing without special constraints
**Block-wise Mode Algorithm**:
The shader processes linear buffer indices using `gl_GlobalInvocationID.x` as the output buffer index. It converts this to tensor coordinates using `bufi_to_tidx(out_bufi, t_out_strides, out_dim_order)` which handles the buffer-to-tensor index mapping with proper stride calculations.
For each element, it computes the block coordinate directly: `bcoord = out_tidx / blockSize` where `out_tidx` is the 4D tensor coordinate in WHCN layout. The linear block ID calculation uses the same pre-computed stride approach: `block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w`.
The element value is loaded using the corresponding input buffer index: `value = t_in[in_bufi]` where `in_bufi = tidx_to_bufi(out_tidx, t_in_strides)`. Quantization applies the block-specific parameters: `qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])`.
**Future Improvements**: Dynamic workgroup sizing based on block dimensions, there is likely a better method to making it better than what it is currently.
Differential Revision: [D78302195](https://our.internmc.facebook.com/intern/diff/D78302195/)
cc SS-JIA manuelcandales cbilgin
[ghstack-poisoned]
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D78302195 |
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
Gasoonjia
pushed a commit
that referenced
this pull request
Jul 30, 2025
…up (#13001) This PR was created by the merge bot to help merge the original PR into the main branch. ghstack PR number: #12575 by @ahmtox ^ Please use this as the source of truth for the PR details, comments, and reviews ghstack PR base: https://github.com/pytorch/executorch/tree/gh/ahmtox/43/base ghstack PR head: https://github.com/pytorch/executorch/tree/gh/ahmtox/43/head Merge bot PR base: https://github.com/pytorch/executorch/tree/gh/ahmtox/42/orig Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/ahmtox/43/orig @diff-train-skip-merge cc @SS-JIA @manuelcandales @cbilgin --------- Co-authored-by: morelos <morelos@devvm4573.ash0.facebook.com> Co-authored-by: ahmtox <69552192+ahmtox@users.noreply.github.com>
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
4 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):
Changes
torchao.quantize_affineoperator in Vulkan backend with comprehensive texture and buffer storage supportquantize_texture.glslandquantize_buffer.glslshaders for configurable tensor block quantizationaffine_test.cppwith multi-dimensional tensor validation and reference implementationQuantize.cppto handle affine transformations with configurable block sizes and quantization parametersBE: Improved the documentation in the shader logic which is more detailed and clear
NOTE: I delegated the quantize_affine and future affine operators through a new custom test file denoted as
affine_test.cppas the other quantization testing framework was getting a little large, and it makes more sense to separate the namespace between torchao and quantized_decomposed. I believe the _decomposed namespace is getting phased out in favor of this affine operator so deprecation will be easier in the future.Motivation
The existing Vulkan quantization infrastructure lacked support for the
torchao.quantize_affineoperator, which is essential for enabling dynamic quantization efficiently. Thequantize_affineoperator provides flexible block-wise quantization that allows different scale and zero-point values for tensor blocks, enabling:qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)for precise floating-point to integer mappingOperator Description
The
quantize_affineoperator converts floating-point tensor values to n-bit integer representations using pre-computed quantization parameters (scale and zero_point) applied to configurable tensor blocks. Block-wise quantization divides tensors into blocks and applies separate quantization parameters to each block, allowing fine-grained control over quantization precision.The quantization formula is:
qvalue = clamp(round(value / scale) + zero_point, quant_min, quant_max)Storage Requirements: Scale and zero_point tensors must use buffer storage with width-packed layout. Input/output tensors support both buffer and texture storage with standard axis mapping.
Block-wise Quantization Implementation
Block-wise quantization enables fine-grained quantization by dividing tensors into blocks and applying separate quantization parameters to each block. The implementation uses several key data structures computed in
Quantize.cpp:block_size_vec: WHCN-ordered block dimensions converted from PyTorch NCHW layout (e.g., [3,3,2,1] for 3×3×2×1 blocks)tensor_size_whcn: Input tensor dimensions converted to WHCN layout usingutils::make_whcn_ivec4()num_blocks_vec: Number of blocks per dimension calculated astensor_size_whcn / block_size_vecblock_stride_vec: Pre-computed linear strides for block grid indexing{1, #W, #W*#H, #W*#H*#C}to enable efficient block ID calculationThe block coordinate calculation uses:
bcoord = tidx / blockSizewheretidxis the tensor coordinate in WHCN layout, then the linear block ID is computed as:block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.wShader Algorithm Overview
Texture Storage Implementation (
quantize_texture.glsl)Workgroup Configuration:
global_workgroup_size[2] > 1)Block-wise Mode Algorithm:
The shader processes 3D texture positions where each position represents a texel containing 4 width-packed components. For each texel at position
pos, it calculates a base tensor indexbase_tidx = ivec4(pos.x * 4, pos.y, pos.z, 0)to account for width-packing.For each of the 4 components in the texel, it computes the actual tensor coordinate:
tidx = ivec4(base_tidx.x + i, base_tidx.y, (foldedZ % C_total), (foldedZ / C_total))wherefoldedZ = pos.zhandles batch-channel folding in 4D tensors andC_total = numBlocks.z * blockSize.zrepresents the total channel dimension.The block coordinate is calculated using integer division:
bcoord = tidx / blockSize, then the linear block ID uses pre-computed strides:block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w.Each component is quantized using its corresponding block's parameters:
qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id])and written to the output texel.Buffer Storage Implementation (
quantize_buffer.glsl)Workgroup Configuration:
Block-wise Mode Algorithm:
The shader processes linear buffer indices using
gl_GlobalInvocationID.xas the output buffer index. It converts this to tensor coordinates usingbufi_to_tidx(out_bufi, t_out_strides, out_dim_order)which handles the buffer-to-tensor index mapping with proper stride calculations.For each element, it computes the block coordinate directly:
bcoord = out_tidx / blockSizewhereout_tidxis the 4D tensor coordinate in WHCN layout. The linear block ID calculation uses the same pre-computed stride approach:block_id = bcoord.x * blockStride.x + bcoord.y * blockStride.y + bcoord.z * blockStride.z + bcoord.w * blockStride.w.The element value is loaded using the corresponding input buffer index:
value = t_in[in_bufi]wherein_bufi = tidx_to_bufi(out_tidx, t_in_strides). Quantization applies the block-specific parameters:qvalue = quantize_val(value, t_scale[block_id], t_zero_point[block_id]).Future Improvements: Dynamic workgroup sizing based on block dimensions, there is likely a better method to making it better than what it is currently.
Differential Revision: D78302195
cc @SS-JIA @manuelcandales @cbilgin