ispc(1)

💖 ispc — Intel Implicit SPMD Program Compiler 💖

okay so bestie, welcome to the ispc(1) page from the General Commands Manual ✨ this is the compiler that makes parallel code feel like a normal lil serial program, and I'm gonna walk you through literally every flag. grab a strawberry 🍓, we're going in.

---

🌈 SYNOPSIS 🌈

ispc [options] file

like. that's it. that's the whole vibe. you hand it a file and some options and it does its thing 💅

---

🦋 DESCRIPTION 🦋

ispc is a compiler for a variant of the C programming language, with extensions for single program, multiple data programming. no cap — it's C-ish, but with SPMD energy.

under the SPMD model, the programmer writes a program that generally appears to be a regular serial program, though the execution model is actually that a number of program instances execute in parallel on the hardware. so you write it like it's one thing, but it's secretly a whole squad running at once. main character energy for code, fr fr 👑

---

✨ OPTIONS ✨

• --addressing {32,64} — pick 32- or 64-bit addressing. heads up bestie: 32-bit addressing calculations are done by default, even on 64-bit target architectures. so the default is not what you'd assume 🙈
• --arch {x86, x86-64, arm, aarch64, wasm32} — select target architecture 🏗️
• --colored-output — always use terminal colors in error/warning messages. yes queen, make the errors pretty 🌈
• --cpu <type> — an alias for the --device <type> switch. same thing, different name 🎀
• -D <foo> — #define given value when running preprocessor
• --dev-stub <filename> — emit device-side offload stub functions to file
• --device <type> — select target device. here's the full roster, character for character, no skips:

<type> = { x86-64 , atom (synonyms: bonnell ), core2, penryn,
corei7 (synonyms: nehalem ), btver2 (synonyms: ps4 ), corei7-avx
(synonyms: sandybridge ), core-avx-i (synonyms: ivybridge ),
core-avx2 (synonyms: haswell ), broadwell, skylake, knl, skx,
icelake-client (synonyms: icl ), slm (synonyms: silvermont ),
icelake-server (synonyms: icx ), tigerlake (synonyms: tgl ),
alderlake (synonyms: adl ), sapphirerapids (synonyms: spr ),
znver1, znver2 (synonyms: ps5 ), znver3, cortex-a9, cortex-a15,
cortex-a35, cortex-a53, cortex-a57, apple-a7, apple-a10,
apple-a11, apple-a12, apple-a13, apple-a14 }

• --dllexport — make non-static functions DLL exported. Windows target only 🪟
• --dwarf-version {2,3,4} — generate source-level debug information with given DWARF version (this triggers -g btw). Ignored for Windows target 💀
• -E — run only the preprocessor. stop right there, no compiling today ✋
• --emit-asm — generate assembly language file as output
• --emit-llvm — emit LLVM bitcode file as output
• --emit-llvm-text — emit LLVM bitcode file as output in textual form
• --emit-obj — generate object file file as output (default). yeah it says "file file", I'm not fixing it, that's what the page says and I respect the page ✨
• --enable-llvm-intrinsics — enable experimental feature to call LLVM intrinsics from ISPC source code. experimental, so tread carefully bestie 🧚
• --error-limit <value> — limit maximum number of errors emitting by ISPC to <value>
• --force-alignment <value> — force alignment in memory allocations routine to be <value>
• -g — generate source-level debug information
• --help — print help 🫶
• --help-dev — print help for developer options
• --host-stub <filename> — emit host-side offload stub functions to file
• -h <name> / --header-outfile <name> — output filename for header
• -I <path> — add <path> to #include file search path
• --ignore-preprocessor-errors — suppress errors from the preprocessor 🙈
• --instrument — emit instrumentation to gather performance data 📊
• --math-lib <option> — select math library. your choices, bestie:
  • default — use ispc's built-in math functions
  • fast — use high-performance but lower-accuracy math functions ⚡ (speed over precision, iykyk)
  • svml — use the Intel(r) SVML math libraries
  • system — use the system's math library (may be quite slow) 🐌
• -MMM <filename> — write #include dependencies to given file.
• -M — output a rule suitable for make describing the dependencies of the main source file to stdout.
• -MF <filename> — when used with -M, specifies a file to write the dependencies to.
• -MT <filename> — when used with -M, changes the target of the rule emitted by dependency generation.
• --no-omit-frame-pointer — disable frame pointer omission. it may be useful for profiling 🔍
• --nostdlib — don't make the ispc standard library available. going fully self-sufficient, respect 💅
• --no-pragma-once — don't use #pragma once in created headers
• --nocpp — don't run the C preprocessor
• -o<name>/--outfile <name> — output filename (may be "-" for standard output) 📤
• -O0/-O(1/2/3) — set optimization level. default behavior is to optimize for speed. the breakdown:
  • -O0 — optimizations disabled.
  • -O1 — optimization for size.
  • -O2/O3
• --opt <option> — set optimization option. the menu:
  • disable-assertions — remove assertion statements from final code.
  • disable-fma — disable 'fused multiply-add' instructions (on targets that support them)
  • disable-loop-unroll — disable loop unrolling.
  • disable-zmm — disable using zmm registers for avx512 targets in favour of ymm. this also affects ABI. ⚠️ note the ABI thing, it's not just a lil perf tweak
  • fast-masked-vload — faster masked vector loads on SSE (may go past end of array). ⚠️ it can read past the end, so bestie, be careful
  • fast-math — perform non-IEEE-compliant optimizations of numeric expressions. non-compliant!! that's the whole point and also the whole warning 🚨
  • force-aligned-memory — always issue "aligned" vector load and store instructions
• --pic — generate position-independent code. Ignored for Windows target 🪟
• --quiet — suppress all output 🤫
• --support-matrix — print full matrix of supported targets, architectures and OSes
• --target <t> — select target ISA and width. the full list:

<t> ={ sse2-i32x4, sse2-i32x8,
sse4-i8x16, sse4-i16x8, sse4-i32x4, sse4-i32x8, avx1-i32x4,
avx1-i32x8, avx1-i32x16, avx1-i64x4, avx2-i8x32, avx2-i16x16,
avx2-i32x4, avx2-i32x8, avx2-i32x16, avx2-i64x4, avx512knl-x16,
avx512skx-x4, avx512skx-x8, avx512skx-x16, avx512skx-x32,
avx512skx-x64, neon-i8x16, neon-i16x8, neon-i32x4, neon-i32x8,
wasm-i32x4 }

• --target-os <os> — select target OS. <os> ={ linux, custom_linux, freebsd, android, web } 🌍
• --time-trace — turn on time profiler. generates JSON file based on output filename. ⏱️
• --time-trace-granularity <value> — minimum time granularity (in microseconds) traced by time profiler.
• --vectorcall/--no-vectorcall — enable/disable vectorcall calling convention on Windows (x64 only). Disabled by default 🪟
• --version — print ispc version 🏷️
• --werror — treat warnings as errors. zero tolerance era 💀
• --woff — disable warnings
• --wno-perf — don't issue warnings related to performance-related issues
• --x86-asm-syntax <option> — select style of code if generating assembly. pick your aesthetic:
  • intel — emit Intel-style assembly
  • att — emit AT&T-style assembly

---

👀 SEE ALSO 👀

https://ispc.github.io/

like, go read the docs bestie, they're right there 🌐

---

💜 AUTHOR 💜

ispc is written by Intel Corporation.

this manual page was written by Yangfl for the Debian Project (and may be used by others). thank you Yangfl, absolute legend 🫶

---

and that's the whole page, babe 💅 it's ispc(1) from the General Commands Manual, and that ispc(1) you keep seeing at the bottom is just the footer tag doing its lil signature move — slay, go compile something iconic ✨🎀🦋