Update README.md to align with the Algorithm concept.

This commit is contained in:
John Shumway
2025-10-02 00:51:13 +00:00
parent 816be4c417
commit 22b5da468f

View File

@@ -3,33 +3,28 @@
This is a framework for instantiate kernels and query kernel settings.
The framework is built on a semantic understanding of the convolution kernels.
This is a prototype implemenentation for CK integration in MIOpen.
This is a prototype implemenentation specific to convolutions for CK integration in MIOpen.
Draft of kernel description, to be iterated and aligned with existing CK convolution kernels.
- **Signature**
*The mathematical description of the kernel operation.*
- **Spatial dimensionality**
1D, 2D, or 3D.
- **Direction**
Describes the convolution direction: forward, backward-input, or backward-weight.
Describes the convolution direction: forward, backward-data, or backward-weight.
- **Tensor Layouts**
Channels first or channels last.
Specifies memory layout conventions for input, weight, and output tensors.
- **Groups**
Indicates grouping strategy: standard, grouped, or depthwise.
- **Data Types**
Tenosor input and output types (FP32, FP16, BF16).
- **Operations**
Lists unary and binary operations applied to tensors (e.g., activation, bias add, residual fusion).
- **Spatial Parameters**
Defines padding, stride, and dilation used in the convolution.
- **Data Types**
Specifies precision formats for input, weight, and output tensors.
- **Broadcasting Semantics**
Indicates whether broadcasting is applied for binary operations.
Lists unary and binary operations applied to tensors (e.g., bias or clamp).
- **Algorithm**
*The implementation strategy and hardware-specific optimizations.*