[llvm] [DirectX][DXIL] Design document for TableGen Spec of DXIL Operations (PR #85170)
S. Bharadwaj Yadavalli via llvm-commits
llvm-commits at lists.llvm.org
Wed May 29 05:06:09 PDT 2024
================
@@ -0,0 +1,179 @@
+==============================================================
+Specification of DXIL Operations using TableGen Representation
+==============================================================
+.. contents::
+ :local:
+
+.. toctree
+ :hidden
+
+Introduction
+============
+
+`DirectXShaderCompiler <https://github.com/microsoft/DirectXShaderCompiler>`_
+encapsulates, among other information, various DXIL Operations in
+`hctdb.py <https://github.com/microsoft/DirectXShaderCompiler/blob/main/utils/hct/hctdb.py>`_.
+DXIL Operations are represented in one of the following `two ways
+<https://github.com/microsoft/DirectXShaderCompiler/blob/130877392c263888ef06bab768856d3dab1f1c9a/docs/DXIL.rst#L1978>`_:
+
+#. Using LLVM instructions
+#. Using LLVM External functions. These are represented in LLVM IR as follows:
+ * "Standard" LLVM intrinsics (e.g., ``llvm.sin.*``) and
+ * HLSL intrinsics (defined as LLVM intrinsics in ``llvm/include/llvm/IR/IntrinsicsDirectX.td``, e.g., ``llvm.dx.*``)
+
+ These are collectively referred to as `LLVM Intrinsics` in this note.
+
+Following is the complete list of properties of DXIL Ops with the corresponding field name
+as used in ``hctdb.py``, if one exists. A DXIL Op is represented by a set of associated properties
+
+1. Name of operation (``dxil_op``)
+2. The generic or HLSL-specific intrinsic that maps to the operation (``llvm_name``).
+3. Unique Integer ID (``dxil_opid``)
+4. Operation Class signifying the name and function signature of the operation (``dxil_class``).
+ This string is an integral part of the DXIL Op function name and is constructed in
+ the format ``dx.op.<class-name>.<overload-type>``. The DXIL validator checks for any
+ deviation from this for each of the DXIL Op call.
+
+5. List of valid overload types for the operation (``oload_types``).
+6. Required minimum Shader Model version with support for the operation. (``shader_model``).
+7. Required minimum DXIL version with support for the operation.
+8. Minimum shader model required with translation by linker (``shader_model_translated``)
+9. List of shader stages the operation is applicable to (``shader_stages``); empty if applicable to all stages.
+10. Memory access attributes of the operation (``fn_attr``).
+11. Boolean attributes of operation to indicate if it
+ * is some kind of a derivative (``is_derivative``)
+ * requires gradient calculation (``is_gradient``)
+ * is a sampler feedback (``is_feedback``)
+ * requires in-wave, cross-lane functionality (``is_wave``)
+ * requires that all of its inputs are uniform across the wave (``requires_uniform_inputs``).
+ * is a barrier operation (``is_barrier``).
+12. A string that documents the operation (``doc``)
+
+Motivation
+==========
+
+DXIL backend passes depend on the knowledge of various properties of DXIL Operations.
+For example, ``DXILLowering`` pass will need information such as the DXIL operation an
+LLVM intrinsic is to be lowered to, along with valid overload and parameter types etc.
+TableGen file - ``llvm/lib/Target/DirectX/DXIL.td`` - is used to represent DXIL Operations
+by specifying their properties listed above. ``DXIL.td`` is designed to be the single source
+of reference of DXIL Operations for DXIL backend implementation in ``llvm-project`` repo -
+analogous to ``hctdb.py`` for ``DirectXShadeCompiler`` repo. It needs to have a rich
+representation capabilities that TableGen backends (such as ``DXILEmitter``) can rely on.
+Additionally, the DXIL Op specification should be easy to read and comprehend.
+
+Design
+======
+
+1. Each DXIL Operation is represented as a TableGen record. The name of each of the records
+ signifies operation name.
+2. The LLVM Intrinsic that maps to the operation is represented using ``Intrinsic::*``.
+3. The unique operation id is represented by an integer.
+4. DXIL Operation Class is represented as follows
+
+ .. code-block::
+
+ // Abstraction of DXIL Operation class.
+ // It encapsulates an associated function signature viz.,
+ // returnTy(param1Ty, param2Ty, ...) represented as a list of LLVMTypes.
+ // DXIL Ops that belong to a DXILOpClass record the signature of that DXILOpClass
+
+ class DXILOpClass<list<LLVMType> OpSig> {
+ list<LLVMType> OpSignature = OpSig;
+ }
+
+ Concrete operation classes, such as ``unary`` are defined inheriting from ``DXILOpClass``.
+5. Valid overload types are represented as a list of ``LLVMType`` s.
+6. Concrete records of Shader Model version and DXIL versions and are defined
+ by inheriting from the class
+
+ .. code-block::
+
+ // Abstract class to represent major and minor version values
+ class Version<int major, int minor> {
+ int Major = major;
+ int Minor = minor;
+ }
+
+7. Shader stages for which the operation is applicable are represented as a list of
+ concrete records that inherit from the Class
+
+ .. code-block::
+
+ class ShaderStage;
+
+
+8. All remaining properties of the operation are represented as a list as of concrete records
+ that inherit from the class
+
+ .. code-block::
+
+ class OpAttributes;
+
+ - memory access - ``ReadNone``, ``ReadNone``
+ - ``IsDerivative``, ``IsGradient``, ``IsFeedback``, ``IsWave``, ``NeedsUniformInputs``, ``IsBarrier``
+
+9. A documentation string for the operation.
+
+
+A DXIL Operation is represented by the following TableGne class by encapsulating the various
+TableGen representations of its properties described above.
+
+.. code-block::
+
+ // Abstraction DXIL Operation
+ class DXILOpPropertiesBase {
+ int OpCode = 0; // Opcode of DXIL Operation
+ DXILOpClass OpClass = UnknownOpClass; // Class of DXIL Operation.
+ Intrinsic LLVMIntrinsic = ?; // LLVM Intrinsic DXIL Operation maps to
+ list<LLVMType> OpOverloadTypes = ?; // Valid overload type
+ // of DXIL Operation
+ Version SMVer = ?; // Min Shader Model version
+ Version SMVerLinker = ?; // Min Shader Model required for linking
+ Version DXILVer = ?; // Min DXIL version
+ list<ShaderStage> ShaderStages = ?; // List of applicable shader stages
+ list<OpAttributes> OpAttribs = ?; // Operation attributes
+ string Doc = ""; // A short description of the operation
+ }
+
+
+The following convenience class is used to demonstrate the definitions of a couple of
+operations:
+
+ .. code-block::
+
+ class DXILOpProperties<int opCode,
+ Intrinsic intrinsic,
+ list<LLVMType> overloadTypes,
+ string doc> : DXILOpPropertiesBase {
+ int OpCode = opCode;
+ Intrinsic LLVMIntrinsic = intrinsic;
+ list<LLVMType> OpOverloadTypes = overloadTypes;
+ string Doc = doc;
+ }
+
+Additionally, following definition of ``unary`` class is also used:
+
+.. code-block::
+
+ def unary : DXILOpClass<[llvm_any_ty, LLVMMatchType<0>]>;
+
+Following is the definition of ``Sin`` and ``Cos``
+
+.. code-block::
+
+ let OpClass = unary in {
+ def Cos : DXILOpProperties<12, int_cos, [llvm_half_ty, llvm_float_ty],
+ "Returns cosine(theta) for theta in radians.">;
+ def Sin : DXILOpProperties<13, int_sin, [llvm_half_ty, llvm_float_ty],
+ "Returns sine(theta) for theta in radians.">;
----------------
bharadwajy wrote:
> I'm curious what the representation looks like for a more exotic DXIL Op. Like maybe `ddx`? How do we represent that `ddx` is available in SM 2.0+ for Pixel shaders and SM 6.6+ for Compute, Mesh and Amplification?
>
> How do we handle `RawBufferLoad` where the `double` and `i64` overloads are only valid for SM 6.3+?
Thanks for flagging these. I've noted this and plan to update the design spec in subsequent PRs, if needed, as I make progress with implementation and learn about the specific requirements to represent DXIL ops that may not be sufficiently represented using the current specification mechanism.
https://github.com/llvm/llvm-project/pull/85170
More information about the llvm-commits
mailing list