add some file-level doc for the MFMA classes that already existed (prior to this PR), classes defined in include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma_impl.hpp

2026-07-07 15:47:05 +00:00 · 2025-07-15 06:47:11 -04:00
parent e84eb0f284
commit 8e671e4018
1 changed files with 123 additions and 0 deletions
--- a/docs/reference/modules/warp_gemm_attribute_mfma_impl.rst
+++ b/docs/reference/modules/warp_gemm_attribute_mfma_impl.rst
@@ -0,0 +1,123 @@
+.. |rst_start_tag| raw:: html
+
+   <div class="section" id="file-doc">
+
+.. |rst_end_tag| raw:: html
+
+   </div>
+
+|rst_start_tag|
+
+.. highlight:: cpp
+
+.. include_file:: include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma_impl.hpp
+
+.. _file_ck_tile_ops_gemm_warp_warp_gemm_attribute_mfma_impl.hpp:
+
+include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma_impl.hpp
+================================================================
+
+This file defines the `WGAttrCtlEnum` enumeration and various `WarpGemmAttributeMfmaImpl` struct definitions. These structs specify attributes for warp-level GEMM (General Matrix Multiply) operations, particularly those utilizing MFMA (Matrix Fused Multiply-Add) instructions, with different data types and matrix dimensions.
+
+**Contents:**
+
+.. contents::
+   :local:
+   :depth: 2
+
+.. cpp:enum-class:: WGAttrCtlEnum
+   :project: ck_tile
+
+   .. _ex_WGAttrCtlEnum:
+
+   `WGAttrCtlEnum`
+   --------------------
+
+   Enumeration controlling the attributes for warp-level GEMM operations, primarily related to register allocation (e.g., vgpr for vector general-purpose register, agpr for accumulator general-purpose register) for C, A, and B matrices in the context of MFMA instructions.
+
+   .. container:: custom-attributes
+
+      .. cpp:enumerator:: Default_ = 0
+
+         Default attribute control.
+
+      .. cpp:enumerator:: Raw_vvv = 1
+
+         C-vgpr, A-vgpr, B-vgpr.
+
+      .. cpp:enumerator:: Raw_vaa = 2
+
+         C-vgpr, A-agpr, B-agpr.
+
+      .. cpp:enumerator:: Raw_vav = 3
+
+         C-vgpr, A-agpr, B-vgpr.
+
+      .. cpp:enumerator:: Raw_vva = 4
+
+         C-vgpr, A-vgpr, B-agpr.
+
+      .. cpp:enumerator:: Raw_avv = 5
+
+         C-agpr, A-vgpr, B-vgpr.
+
+.. _ex_WarpGemmAttributeMfmaImpl_Structs:
+
+WarpGemmAttributeMfmaImpl Struct Definitions
+----------------------------------------------
+
+These structs define specific configurations for warp-level GEMM, parameterized by input/output data types (e.g., BF16, FP16, FP8, INT8) and matrix dimensions (M, N, K). The naming convention generally follows `WarpGemmAttributeMfmaImpl<InputTypeA><InputTypeB><OutputType>M<M>N<N>K<K>`.
+
+.. _WarpGemmAttributeMfmaImpl_BF16:
+
+Bf16Bf16F32 Combinations
+^^^^^^^^^^^^^^^^^^^^^^^
+
+These structs specify MFMA attributes for operations with BF16 inputs and FP32 accumulation/output.
+
+* `WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K32`
+* `WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8`
+* `WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16`
+* `WarpGemmAttributeMfmaImplBf16Bf16F32M4N64K4`
+* `WarpGemmAttributeMfmaImplBf16Bf16F32M64N4K4`
+* `WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K16`
+
+.. _WarpGemmAttributeMfmaImpl_FP16:
+
+F16F16F32 Combinations
+^^^^^^^^^^^^^^^^^^^^^^
+
+These structs specify MFMA attributes for operations with FP16 inputs and FP32 accumulation/output.
+
+* `WarpGemmAttributeMfmaImplF16F16F32M32N32K8`
+* `WarpGemmAttributeMfmaImplF16F16F32M16N16K16`
+* `WarpGemmAttributeMfmaImplF16F16F32M16N16K32`
+* `WarpGemmAttributeMfmaImplF16F16F32M4N64K4`
+* `WarpGemmAttributeMfmaImplF16F16F32M64N4K4`
+* `WarpGemmAttributeMfmaImplF16F16F32M32N32K16`
+
+.. _WarpGemmAttributeMfmaImpl_FP8:
+
+FP8 Related Base Structs
+^^^^^^^^^^^^^^^^^^^^^^^
+
+These are base template structs for MFMA attributes involving FP8 or BF8 input types, typically accumulating to FP32.
+
+* `WarpGemmAttributeMfmaImpl_f32_16x16x32_f8_base`
+* `WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base`
+* `WarpGemmAttributeMfmaImpl_f32_16x16x128_f8_bf8_base`
+* `WarpGemmAttributeMfmaImpl_f32_32x32x64_f8_bf8_base`
+
+.. _WarpGemmAttributeMfmaImpl_INT8:
+
+Int8 Related Structs (`_i32_`)
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+These structs specify MFMA attributes for operations with INT8 inputs and INT32 accumulation/output.
+
+* `WarpGemmAttributeMfmaImpl_i32_32x32x16_i8`
+* `WarpGemmAttributeMfmaImpl_i32_16x16x32_i8`
+* `WarpGemmAttributeMfmaImpl_i32_16x16x64_i8`
+* `WarpGemmAttributeMfmaImpl_i32_32x32x32_i8`
+
+|rst_end_tag|