Skip to content

NNNN-8-bit-integers.md

Latest commit

 

History

History
67 lines (49 loc) · 3.48 KB

NNNN-8-bit-integers.md

File metadata and controls

67 lines (49 loc) · 3.48 KB
title params
NNNN - 8-bit integer support
authors sponsors status
anteru
Matthäus Chajdas
llvm-beanz
Chris Bieneman
Under Consideration
  • Planned Version: SM 6.x
  • Issues: HLSL 337

Introduction

This proposal introduces 8-bit integer types to HLSL, to align it more closely with C/C++ and other languages, and allow tight interop with the 8-bit types used in the cooperative vector proposal.

Motivation

8-bit types are very special in HLSL as they only exist for some arithmetic instructions, but aren't generally accessible. For example, cooperative vectors allow inputs to be specified as 8-bit quantities, but only by packing them into 32-bit. This is cumbersome if for example an application needs to modify the values before passing in, for example, to add a bias -- the application is now forced to unpack the value, modify it, and pack it back in.

Proposed solution

Introduce two new, native types:

  • uint8_t: 8-bit unsigned integer
  • int8_t: 8-bit signed integer

For support and conversion rules, those new types would match the 16-bit scalar types, that is, new uint8_t.

DXIL changes

DXIL mentions under Primitive types that i8 is supported only for limited operations. With this proposal: i8 gets supported for computation by shader. For memory access, we'd only support loading multiples of 4 i8.

DXIL/LLVM IR doesn't discern between i8 and u8, so during lowering, we need to restrict/select between the I and U instructions.

For the elementwise overloads defined in DXIL vectors, the following subset would be supported:

Opcode Name Class Note
30 Bfrev Unary See note
31 Countbits UnaryBits
32 FirstBitLo UnaryBits
33 FirstBitHi UnaryBits
34 FirstBitSHi UnaryBits
37 IMax Binary i8 only
38 IMin Binary i8 only
39 UMax Binary u8 only
40 UMin Binary u8 only
48 IMad Tertiary i8 only
49 UMad Tertiary u8 only
115 WaveActiveAllEqual WaveActiveAllEqual
117 WaveReadLaneAt WaveReadLaneAt
118 WaveReadLaneFirst WaveReadLaneFirst
119 WaveActiveOp WaveActiveOp
120 WaveActiveBit WaveActiveBit
121 WavePrefixOp WavePrefixOp
122 QuadReadLaneAt QuadReadLaneAt
123 QuadOp QuadOp
165 WaveMatch WaveMatch

TODO: Do we really need Bfrev?

SPIR-V changes

SPIR-V already supports 8-bit integers.