Cross-Vendor GPU Concepts
19 vendor-neutral operations, each mapped to its NVIDIA PTX and AMDGPU/GFX counterparts where one has been identified.
⚠️ Cross-vendor mappings describe semantic relationships and are not guaranteed one-to-one compiler
translations.
| Concept | Category | Mapping Status | Summary |
|---|---|---|---|
| Atomic Add | Atomics | equivalent with restrictions | Atomically add a value to a memory location and (optionally) return the prior value. |
| Ballot | Warp/Wavefront-Level | equivalent sequence | Collect a 1-bit predicate from every lane of a warp/wavefront into a single bitmask. |
| Compare | Comparison | equivalent with restrictions | Compare two operands and produce a boolean or predicate result. |
| Floating-Point Addition | Arithmetic | equivalent with restrictions | Add two IEEE-754 floating-point operands. |
| Floating-Point Multiplication | Arithmetic | equivalent with restrictions | Multiply two IEEE-754 floating-point operands. |
| Fused Multiply-Add | Arithmetic | equivalent with restrictions | Compute (a * b) + c with a single rounding step instead of two. |
| Global-Memory Load | Memory | equivalent with restrictions | Read a value from the device's global (main GPU) memory address space into a register. |
| Global-Memory Store | Memory | equivalent with restrictions | Write a register value to the device's global memory address space. |
| Integer Addition | Arithmetic | equivalent with restrictions | Add two integer operands of matching or implied width, optionally with saturation. |
| Integer Multiplication | Arithmetic | equivalent with restrictions | Multiply two integer operands, taking the low, high, or wide part of the product. |
| Lane Shuffle | Warp/Wavefront-Level | equivalent with restrictions | Exchange a value between lanes of the same warp/wavefront without going through memory. |
| Matrix Multiply (Tensor/Matrix-Core) | Matrix Operations | unverified | Compute a small matrix-multiply-accumulate using dedicated tensor/matrix-core hardware, cooperatively across a warp/wavefront. |
| Memory Fence | Synchronization | equivalent sequence | Order this thread's prior memory accesses relative to later ones, visible to a specified scope. |
| Population Count | Bit Manipulation | equivalent with restrictions | Count the number of set (1) bits in an integer operand. |
| Shared/Local-Memory Load | Memory | equivalent with restrictions | Read a value from on-chip, workgroup/CTA-scoped shared (local) memory. |
| Shared/Local-Memory Store | Memory | equivalent with restrictions | Write a register value to on-chip, workgroup/CTA-scoped shared (local) memory. |
| Special-Function Approximation | Arithmetic | equivalent with restrictions | Fast, reduced-precision hardware approximation of a transcendental function (reciprocal, reciprocal sqrt, sin, log2, exp2, ...). |
| Type Conversion | Data Movement | equivalent with restrictions | Convert a value between integer and/or floating-point types, with a specified rounding and saturation mode. |
| Workgroup Barrier | Synchronization | equivalent with restrictions | Block all threads in a workgroup/CTA at this point until every thread has arrived. |