Files
compiler-explorer/lib/compilers/scale-nvcc-nvidia.ts
moletteremi ef9a1ce7c3 [CUDA] add scale's nvcc compiler (nvidia & amd backends) (#8849)
Hi !

This PR adds the [scale nvcc
compiler](https://docs.scale-lang.com/stable/) to the CE live site for
both AMD and Nvidia backends

It works locally and shows host asm, device LLVMIR and:

 - PTX ans SASS for Nvidia backend,
 - AMDGPU code for AMD backend.
 
There is a bug in scale 1.7.1 that prevents the compilation with both
`-S, -o` flags. As a result I had to make some temp workarounds.

For Nvidia backend, PTX is written in the device `.s` file. I run it
through `ptax` and `nvidasm` to get the SASS. To get the LLVMIR, I
decode the device `.bc` file with `llvm-dis`.

For AMD backend, the `.s` file is the AMDGPU code, and the LLVMIR is
obtained the same way.

I plan to make the `.ts` script closer to say `nvcc.ts` once the flag
bugs are fixed in scale.

Let me know if there is anything I missed for the live site integration.

infra PR: https://github.com/compiler-explorer/infra/pull/2191
Issue:
https://github.com/compiler-explorer/compiler-explorer/issues/8865

I acknowledge the use of generative AI to help drafting the code of this
PR.

---------

Co-authored-by: mattgodbolt-molty <mattgodbolt-molty@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-02 22:14:19 +01:00

414 lines
16 KiB
TypeScript

// Copyright (c) 2026, Compiler Explorer Authors
// All rights reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are met:
//
// * Redistributions of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
// * Redistributions in binary form must reproduce the above copyright
// notice, this list of conditions and the following disclaimer in the
// documentation and/or other materials provided with the distribution.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
// ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
// LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
// CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
// CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
// ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
// POSSIBILITY OF SUCH DAMAGE.
import * as fs from 'node:fs/promises';
import Path from 'node:path';
import _ from 'underscore';
import type {CompilationInfo, CompilationResult} from '../../types/compilation/compilation.interfaces.js';
import type {PreliminaryCompilerInfo} from '../../types/compiler.interfaces.js';
import type {ParseFiltersAndOutputOptions} from '../../types/features/filters.interfaces.js';
import {unwrap} from '../assert.js';
import {BaseCompiler} from '../base-compiler.js';
import {CompilationEnvironment} from '../compilation-env.js';
import {logger} from '../logger.js';
import {PTXAsmParser} from '../parsers/asm-parser-ptx.js';
import {SassAsmParser} from '../parsers/asm-parser-sass.js';
import {ClangParser} from './argument-parsers.js';
interface ExecResult {
code: number;
stdout: string;
stderr: string;
}
export class ScaleNvccNvidiaCompiler extends BaseCompiler {
static get key() {
return 'scale-nvcc-nvidia';
}
deviceAsmParser: SassAsmParser;
ptxParser: PTXAsmParser;
constructor(info: PreliminaryCompilerInfo, env: CompilationEnvironment) {
super(info, env);
this.compiler.supportsOptOutput = true;
this.compiler.supportsDeviceAsmView = true;
this.deviceAsmParser = new SassAsmParser(this.compilerProps);
this.ptxParser = new PTXAsmParser(this.compilerProps);
}
// TEMP: -o commented out because scale can't combine an explicit `-o`
// with `-Xcompiler=-S`. This means scale (and now nvcc too, while this
// is shared) falls back to basename-derived default naming instead of
// a predictable `output.s`. See findHostAsmFile()/extractDeviceCode()
// below, which auto-detect either naming convention.
override optionsForFilter(filters: ParseFiltersAndOutputOptions, outputFilename: string) {
const opts = ['-g', '-lineinfo', '--keep-device-functions'];
if (!filters.execute) {
opts.push('-keep', '-keep-dir', Path.dirname(outputFilename));
if (!filters.binary) {
opts.push('-Xcompiler=-S');
}
}
return opts;
}
override getArgumentParserClass() {
return ClangParser;
}
override optOutputRequested(options: string[]) {
return (
super.optOutputRequested(options) ||
options.includes('--optimization-info') ||
options.includes('-opt-info')
);
}
//Helpers to get de device ptx (.s) and bitcode files
private static readonly scaleDeviceFileRe = /-cuda-nvptx64-nvidia-cuda-([^./]+)\.s$/;
private static readonly scaleDeviceBcFileRe = /-cuda-nvptx64-nvidia-cuda-([^./]+)\.bc$/;
// Temp workaround since -o -S does not work in scale-1.7.1
// Side effect: it works only if one device file is present.
// will be cleaned when scale's bug is fixed
private async findHostAsmFile(dirPath: string): Promise<string | null> {
try {
const files = await fs.readdir(dirPath);
const hostFiles = files.filter(f => f.endsWith('.s') && !ScaleNvccNvidiaCompiler.scaleDeviceFileRe.test(f));
if (hostFiles.length !== 1) {
return null;
}
return Path.join(dirPath, hostFiles[0]);
} catch {
return null;
}
}
override async postProcess(result, outputFilename: string, filters: ParseFiltersAndOutputOptions) {
// TEMP (scale support): outputFilename as originally computed may not
// exist since we no longer pass `-o`. Try to recover the real file.
if (!filters.binary && result.dirPath) {
try {
await fs.stat(outputFilename);
} catch {
const hostAsm = await this.findHostAsmFile(result.dirPath);
if (hostAsm) {
try {
result.asmSize = (await fs.stat(hostAsm)).size;
} catch {
// leave asmSize as-is; base behaviour reports "no output" below
}
outputFilename = hostAsm;
}
}
}
const maxSize = this.env.ceProps('max-asm-size', 64 * 1024 * 1024);
const optPromise = result.optPath ? this.processOptOutput(result.optPath) : Promise.resolve([]);
const postProcess = _.compact(this.compiler.postProcess);
const asmPromise = (
filters.binary
? this.objdump(outputFilename, {}, maxSize, !!filters.intel, !!filters.demangle, false, false, filters)
: (async () => {
if (result.asmSize === undefined) {
result.asm = '<No output file>';
return result;
}
if (result.asmSize >= maxSize) {
result.asm =
'<No output: generated assembly was too large' +
` (${result.asmSize} > ${maxSize} bytes)>`;
return result;
}
if (postProcess.length > 0) {
return await this.execPostProcess(result, postProcess, outputFilename, maxSize);
}
const contents = await fs.readFile(outputFilename, {encoding: 'utf8'});
result.asm = contents.toString();
return result;
})()
).then(asm => {
result.asm = typeof asm === 'string' ? asm : asm.asm;
return result;
});
return Promise.all([asmPromise, optPromise, []]);
}
// Matches the start/end of a GAS inline-assembly block emitted by the host compiler.
private static readonly appBlockStartRe = /^#APP\b/;
private static readonly appBlockEndRe = /^#NO_APP\b/;
// Matches the .nv_fatbin section directive that NVCC injects to hold the fat binary blob.
private static readonly nvFatBinSectionRe = /^\s*\.section\s+\.nv_fatbin\b/;
/**
* Strip `#APP`/`#NO_APP` inline-assembly blocks that contain a `.nv_fatbin`
* section from the host-side x86 assembly. These blocks hold the raw CUDA
* fat binary blob (the `fatbinData` label followed by hundreds of `.quad`
* hex lines) which is never useful to inspect in the asm view.
*
* Only blocks that contain `.nv_fatbin` are removed; any `#APP`/`#NO_APP`
* blocks originating from genuine user inline-assembly are left intact.
*/
protected removeNvccFatbinaryBlob(asm: string): string {
const lines = asm.split('\n');
const result: string[] = [];
let inAppBlock = false;
let hasFatBin = false;
let appBuffer: string[] = [];
for (const line of lines) {
const trimmed = line.trim();
if (ScaleNvccNvidiaCompiler.appBlockStartRe.test(trimmed)) {
inAppBlock = true;
hasFatBin = false;
appBuffer = [line];
} else if (ScaleNvccNvidiaCompiler.appBlockEndRe.test(trimmed)) {
inAppBlock = false;
if (!hasFatBin) {
// Not a fat-binary block — keep it
appBuffer.push(line);
result.push(...appBuffer);
}
appBuffer = [];
} else if (inAppBlock) {
if (ScaleNvccNvidiaCompiler.nvFatBinSectionRe.test(line)) {
hasFatBin = true;
}
appBuffer.push(line);
} else {
result.push(line);
}
}
// Handle (malformed) unclosed #APP block: keep it
if (appBuffer.length > 0) {
result.push(...appBuffer);
}
return result.join('\n');
}
override async processAsm(result, filters: ParseFiltersAndOutputOptions, options: string[]) {
if (filters.labels && typeof result.asm === 'string') {
result = {...result, asm: this.removeNvccFatbinaryBlob(result.asm)};
}
return super.processAsm(result, filters, options);
}
// ---- device-code extraction helpers -------------------------------------------------
/** Pulls the captured arch (e.g. `sm_75`) out of a scale per-target device filename. */
private static archFromDeviceFileName(name: string, re: RegExp): string {
return unwrap(name.match(re))[1];
}
private async runPtxas(ptxFileName: string, cubinPath: string, arch: string, dirPath: string): Promise<ExecResult> {
const {ptxas} = this.compiler;
const args = ['-arch', arch, ptxFileName, '-o', cubinPath];
return this.exec(unwrap(ptxas), args, {customCwd: dirPath});
}
private async runNvdisasm(cubinPath: string, dirPath: string): Promise<ExecResult> {
const {nvdisasm} = this.compiler;
const args = [cubinPath, '-c', '-g', '-hex'];
return this.exec(unwrap(nvdisasm), args, {customCwd: dirPath});
}
private async runLlvmDis(bcPath: string, dirPath: string): Promise<ExecResult> {
const result = await this.exec(unwrap(this.compiler.llvmDisassembler), [bcPath], {customCwd: dirPath});
return result;
}
//Tune the display of llvmir here !
private static llvmIrOptions(demangle: boolean) {
return {
demangle,
filterDebugInfo: false,
filterIRMetadata: false,
filterAttributes: false,
filterComments: false,
};
}
/**
* Handles a single scale/nvcc device `.s` (PTX) file: registers it under
* `PTX (<arch>)`, then pipes it through ptxas -> nvdisasm to also
* register the corresponding `SASS (<arch>)` entry.
*/
private async processPtxFile(
name: string,
dirPath: string,
filters: ParseFiltersAndOutputOptions,
demangle: boolean,
devices: Record<string, CompilationResult>,
): Promise<void> {
const archAndCode = ScaleNvccNvidiaCompiler.archFromDeviceFileName(
name,
ScaleNvccNvidiaCompiler.scaleDeviceFileRe,
);
const asm = await fs.readFile(Path.join(dirPath, name), 'utf8');
const ptxNameAndArch = `PTX (${archAndCode.toLowerCase()})`;
Object.assign(devices, {
[ptxNameAndArch]: await this.postProcessAsm(
{
okToCache: demangle,
...this.ptxParser.process(asm, {...filters, binary: false}),
},
{...filters, binary: false},
),
});
//
// PTX -> CUBIN -> SASS
//
const cubinPath = Path.join(dirPath, `${Path.basename(name, '.s')}.cubin`);
let ptxasResult: ExecResult;
try {
ptxasResult = await this.runPtxas(name, cubinPath, archAndCode, dirPath);
} catch (err) {
logger.error('[processPtxFile]: exception running ptxas for', name, err);
return;
}
if (ptxasResult.code !== 0) {
return;
}
let nvdisasmResult: ExecResult;
try {
nvdisasmResult = await this.runNvdisasm(cubinPath, dirPath);
} catch (err) {
logger.error('[processPtxFile]: exception running nvdisasm for', name, err);
return;
}
const sassAsm =
nvdisasmResult.code === 0
? this.postProcessObjdumpOutput(nvdisasmResult.stdout)
: `<nvdisasm failed with code ${nvdisasmResult.code}>`;
const sassNameAndArch = `SASS (${archAndCode.toLowerCase()})`;
Object.assign(devices, {
[sassNameAndArch]: await this.postProcessAsm(
{
okToCache: demangle,
...this.deviceAsmParser.process(sassAsm, {...filters, binary: true}),
},
{...filters, binary: true},
),
});
}
/**
* Handles a single scale/nvcc device `.bc` (LLVM bitcode) file: runs
* `llvm-dis` to get readable LLVM IR text
*/
private async processBcDeviceFile(
name: string,
dirPath: string,
filters: ParseFiltersAndOutputOptions,
demangle: boolean,
devices: Record<string, CompilationResult>,
): Promise<void> {
const archAndCode = ScaleNvccNvidiaCompiler.archFromDeviceFileName(
name,
ScaleNvccNvidiaCompiler.scaleDeviceBcFileRe,
);
const irNameAndArch = `Device LLVM IR (${archAndCode.toLowerCase()})`;
if (!this.compiler.llvmDisassembler) {
Object.assign(devices, {
[irNameAndArch]: await this.postProcessAsm(
{
okToCache: false,
...(await this.llvmIr.process(
'<error: no llvm-dis found to disassemble bitcode>',
ScaleNvccNvidiaCompiler.llvmIrOptions(demangle),
)),
},
{...filters, binary: false},
),
});
return;
}
const bcPath = Path.join(dirPath, name);
const llPath = Path.join(dirPath, `${Path.basename(name, '.bc')}.ll`);
let irText: string;
try {
const disResult = await this.runLlvmDis(bcPath, dirPath);
irText =
disResult.code === 0
? await fs.readFile(llPath, 'utf8')
: `<llvm-dis failed with code ${disResult.code}>`;
} catch (err) {
irText = `<llvm-dis failed: ${err}>`;
}
Object.assign(devices, {
[irNameAndArch]: await this.postProcessAsm(
{
okToCache: demangle,
...(await this.llvmIr.process(irText, ScaleNvccNvidiaCompiler.llvmIrOptions(demangle))),
},
{...filters, binary: false},
),
});
}
override async extractDeviceCode(
result: CompilationResult,
filters: ParseFiltersAndOutputOptions,
compilationInfo: CompilationInfo,
) {
const {dirPath} = result;
const {demangle} = filters;
const devices: Record<string, CompilationResult> = {...result.devices};
if (dirPath) {
const files = await fs.readdir(dirPath);
const ptxFile = files.filter(f => ScaleNvccNvidiaCompiler.scaleDeviceFileRe.test(f)); //Ends with .s, not the host
const bcDeviceFile = files.filter(f => ScaleNvccNvidiaCompiler.scaleDeviceBcFileRe.test(f)); //ends with .bc, not the host
await Promise.all([
...ptxFile.map(name => this.processPtxFile(name, dirPath, filters, !!demangle, devices)),
...bcDeviceFile.map(name => this.processBcDeviceFile(name, dirPath, filters, !!demangle, devices)),
]);
result.devices = devices;
}
return result;
}
}