From 452010810493743f70c25b7ecd821668b942fc85 Mon Sep 17 00:00:00 2001 From: Jamie Cunliffe Date: Wed, 5 May 2021 14:39:09 +0100 Subject: [PATCH] Intrinsic test tool to compare neon intrinsics with C This tool generates rust and C programs that will call intrinsics and print the results with multiple different inputs. It will build all the programs and then run each of them and diff the output printing any differences that are found. It uses the tracking spreadsheet (with the % column renamed to enabled) to determine which intrinsics to test. It will filter out any intrinsics that have an argument with the name "n" or "lane" as those have constraints on them as to what numbers can be used. --- .gitignore | 2 + Cargo.toml | 1 + crates/intrinsic-test/Cargo.toml | 16 + crates/intrinsic-test/README.md | 23 ++ crates/intrinsic-test/src/argument.rs | 137 +++++++ crates/intrinsic-test/src/intrinsic.rs | 112 ++++++ crates/intrinsic-test/src/main.rs | 380 +++++++++++++++++++ crates/intrinsic-test/src/types.rs | 483 +++++++++++++++++++++++++ crates/intrinsic-test/src/values.rs | 126 +++++++ 9 files changed, 1280 insertions(+) create mode 100644 crates/intrinsic-test/Cargo.toml create mode 100644 crates/intrinsic-test/README.md create mode 100644 crates/intrinsic-test/src/argument.rs create mode 100644 crates/intrinsic-test/src/intrinsic.rs create mode 100644 crates/intrinsic-test/src/main.rs create mode 100644 crates/intrinsic-test/src/types.rs create mode 100644 crates/intrinsic-test/src/values.rs diff --git a/.gitignore b/.gitignore index 97647e1e70..3a1fbc6423 100644 --- a/.gitignore +++ b/.gitignore @@ -4,3 +4,5 @@ target tags crates/stdarch-gen/aarch64.rs crates/stdarch-gen/arm.rs +c_programs/* +rust_programs/* \ No newline at end of file diff --git a/Cargo.toml b/Cargo.toml index 73f69ca46f..6efd6b189a 100644 --- a/Cargo.toml +++ b/Cargo.toml @@ -4,6 +4,7 @@ members = [ "crates/core_arch", "crates/std_detect", "crates/stdarch-gen", + "crates/intrinsic-test", "examples/" ] exclude = [ diff --git a/crates/intrinsic-test/Cargo.toml b/crates/intrinsic-test/Cargo.toml new file mode 100644 index 0000000000..4cdf811e5c --- /dev/null +++ b/crates/intrinsic-test/Cargo.toml @@ -0,0 +1,16 @@ +[package] +name = "intrinsic-test" +version = "0.1.0" +authors = ["Jamie Cunliffe "] +edition = "2018" + +[dependencies] +lazy_static = "1.4.0" +serde = { version = "1", features = ["derive"] } +csv = "1.1" +clap = "2.33.3" +regex = "1.4.2" +log = "0.4.11" +pretty_env_logger = "0.4.0" +rayon = "1.5.0" +diff = "0.1.12" \ No newline at end of file diff --git a/crates/intrinsic-test/README.md b/crates/intrinsic-test/README.md new file mode 100644 index 0000000000..0f60aaa040 --- /dev/null +++ b/crates/intrinsic-test/README.md @@ -0,0 +1,23 @@ +Generate and run programs using equivalent C and Rust intrinsics, checking that +each produces the same result from random inputs. + +# Usage +``` +USAGE: + intrinsic-test [OPTIONS] + +FLAGS: + -h, --help Prints help information + -V, --version Prints version information + +OPTIONS: + --cppcompiler The C++ compiler to use for compiling the c++ code [default: clang++] + --toolchain The rust toolchain to use for building the rust code [default: nightly] + +ARGS: + The input file containing the intrinsics +``` + +The intrinsic.csv is the arm neon tracking google sheet (https://docs.google.com/spreadsheets/d/1MqW1g8c7tlhdRWQixgdWvR4uJHNZzCYAf4V0oHjZkwA/edit#gid=0) +that contains the intrinsic list. The done percentage column should be renamed to "enabled". + diff --git a/crates/intrinsic-test/src/argument.rs b/crates/intrinsic-test/src/argument.rs new file mode 100644 index 0000000000..96ed440ee7 --- /dev/null +++ b/crates/intrinsic-test/src/argument.rs @@ -0,0 +1,137 @@ +use serde::{Deserialize, Deserializer}; + +use crate::types::IntrinsicType; +use crate::Language; + +/// An argument for the intrinsic. +#[derive(Debug, PartialEq, Clone)] +pub struct Argument { + /// The argument's index in the intrinsic function call. + pub pos: usize, + /// The argument name. + pub name: String, + /// The type of the argument. + pub ty: IntrinsicType, +} + +impl Argument { + /// Creates an argument from a Rust style signature i.e. `name: type` + fn from_rust(pos: usize, arg: &str) -> Result { + let mut parts = arg.split(':'); + let name = parts.next().unwrap().trim().to_string(); + let ty = IntrinsicType::from_rust(parts.next().unwrap().trim())?; + + Ok(Self { pos, name, ty }) + } + + fn to_c_type(&self) -> String { + self.ty.c_type() + } + + fn is_simd(&self) -> bool { + self.ty.is_simd() + } + + pub fn is_ptr(&self) -> bool { + self.ty.is_ptr() + } +} + +#[derive(Debug, PartialEq, Clone)] +pub struct ArgumentList { + pub args: Vec, +} + +impl ArgumentList { + /// Creates an argument list from a Rust function signature, the data for + /// this function should only be the arguments. + /// e.g. for `fn test(a: u32, b: u32) -> u32` data should just be `a: u32, b: u32` + fn from_rust_arguments(data: &str) -> Result { + let args = data + .split(',') + .enumerate() + .map(|(idx, arg)| Argument::from_rust(idx, arg)) + .collect::>()?; + + Ok(Self { args }) + } + + /// Converts the argument list into the call paramters for a C function call. + /// e.g. this would generate something like `a, &b, c` + pub fn as_call_param_c(&self) -> String { + self.args + .iter() + .map(|arg| match arg.ty { + IntrinsicType::Ptr { .. } => { + format!("&{}", arg.name) + } + IntrinsicType::Type { .. } => arg.name.clone(), + }) + .collect::>() + .join(", ") + } + + /// Converts the argument list into the call paramters for a Rust function. + /// e.g. this would generate something like `a, b, c` + pub fn as_call_param_rust(&self) -> String { + self.args + .iter() + .map(|arg| arg.name.clone()) + .collect::>() + .join(", ") + } + + /// Creates a line that initializes this argument for C code. + /// e.g. `int32x2_t a = { 0x1, 0x2 };` + pub fn init_random_values_c(&self, pass: usize) -> String { + self.iter() + .map(|arg| { + format!( + "{ty} {name} = {{ {values} }};", + ty = arg.to_c_type(), + name = arg.name, + values = arg.ty.populate_random(pass, &Language::C) + ) + }) + .collect::>() + .join("\n ") + } + + /// Creates a line that initializes this argument for Rust code. + /// e.g. `let a = transmute([0x1, 0x2]);` + pub fn init_random_values_rust(&self, pass: usize) -> String { + self.iter() + .map(|arg| { + if arg.is_simd() { + format!( + "let {name} = ::std::mem::transmute([{values}]);", + name = arg.name, + values = arg.ty.populate_random(pass, &Language::Rust), + ) + } else { + format!( + "let {name} = {value};", + name = arg.name, + value = arg.ty.populate_random(pass, &Language::Rust) + ) + } + }) + .collect::>() + .join("\n ") + } + + pub fn iter(&self) -> std::slice::Iter<'_, Argument> { + self.args.iter() + } +} + +impl<'de> Deserialize<'de> for ArgumentList { + fn deserialize(deserializer: D) -> Result + where + D: Deserializer<'de>, + { + use serde::de::Error; + let s = String::deserialize(deserializer)?; + Self::from_rust_arguments(&s).map_err(Error::custom) + } +} diff --git a/crates/intrinsic-test/src/intrinsic.rs b/crates/intrinsic-test/src/intrinsic.rs new file mode 100644 index 0000000000..499cf7612c --- /dev/null +++ b/crates/intrinsic-test/src/intrinsic.rs @@ -0,0 +1,112 @@ +use crate::types::{IntrinsicType, TypeKind}; + +use super::argument::ArgumentList; +use serde::de::Unexpected; +use serde::{de, Deserialize, Deserializer}; + +/// An intrinsic +#[derive(Deserialize, Debug, PartialEq, Clone)] +pub struct Intrinsic { + /// If the intrinsic should be tested. + #[serde(deserialize_with = "bool_from_string")] + pub enabled: bool, + + /// The function name of this intrinsic. + pub name: String, + + /// Any arguments for this intrinsinc. + #[serde(rename = "args")] + pub arguments: ArgumentList, + + /// The return type of this intrinsic. + #[serde(rename = "return")] + pub results: IntrinsicType, +} + +impl Intrinsic { + /// Generates a std::cout for the intrinsics results that will match the + /// rust debug output format for the return type. + pub fn print_result_c(&self, index: usize) -> String { + let lanes = if self.results.num_lanes() > 1 { + (0..self.results.num_lanes()) + .map(|idx| -> std::string::String { + format!( + "{cast}{lane_fn}(__return_value, {lane})", + cast = self.results.c_promotion(), + lane_fn = self.results.get_lane_function(), + lane = idx + ) + }) + .collect::>() + .join(r#" << ", " << "#) + } else { + format!( + "{promote}cast<{cast}>(__return_value)", + cast = match self.results.kind() { + TypeKind::Float if self.results.inner_size() == 32 => "float".to_string(), + TypeKind::Float if self.results.inner_size() == 64 => "double".to_string(), + TypeKind::Int => format!("int{}_t", self.results.inner_size()), + TypeKind::UInt => format!("uint{}_t", self.results.inner_size()), + TypeKind::Poly => format!("poly{}_t", self.results.inner_size()), + ty => todo!("print_result_c - Unknown type: {:#?}", ty), + }, + promote = self.results.c_promotion(), + ) + }; + + format!( + r#"std::cout << "Result {idx}: {ty}" << std::fixed << std::setprecision(150) << {lanes} << "{close}" << std::endl;"#, + ty = if self.results.is_simd() { + format!("{}(", self.results.c_type()) + } else { + String::from("") + }, + close = if self.results.is_simd() { ")" } else { "" }, + lanes = lanes, + idx = index, + ) + } + + pub fn generate_pass_rust(&self, index: usize) -> String { + format!( + r#" + unsafe {{ + {initialized_args} + let res = {intrinsic_call}({args}); + println!("Result {idx}: {{:.150?}}", res); + }}"#, + initialized_args = self.arguments.init_random_values_rust(index), + intrinsic_call = self.name, + args = self.arguments.as_call_param_rust(), + idx = index, + ) + } + + pub fn generate_pass_c(&self, index: usize) -> String { + format!( + r#" {{ + {initialized_args} + auto __return_value = {intrinsic_call}({args}); + {print_result} + }}"#, + initialized_args = self.arguments.init_random_values_c(index), + intrinsic_call = self.name, + args = self.arguments.as_call_param_c(), + print_result = self.print_result_c(index) + ) + } +} + +fn bool_from_string<'de, D>(deserializer: D) -> Result +where + D: Deserializer<'de>, +{ + match String::deserialize(deserializer)?.to_uppercase().as_ref() { + "TRUE" => Ok(true), + "FALSE" => Ok(false), + other => Err(de::Error::invalid_value( + Unexpected::Str(other), + &"TRUE or FALSE", + )), + } +} diff --git a/crates/intrinsic-test/src/main.rs b/crates/intrinsic-test/src/main.rs new file mode 100644 index 0000000000..22e1be5e4f --- /dev/null +++ b/crates/intrinsic-test/src/main.rs @@ -0,0 +1,380 @@ +#[macro_use] +extern crate lazy_static; +#[macro_use] +extern crate log; + +use std::fs::File; +use std::io::Write; +use std::process::Command; + +use clap::{App, Arg}; +use intrinsic::Intrinsic; +use rayon::prelude::*; +use types::TypeKind; + +mod argument; +mod intrinsic; +mod types; +mod values; + +#[derive(Debug, PartialEq)] +pub enum Language { + Rust, + C, +} + +fn generate_c_program(header_file: &str, intrinsic: &Intrinsic) -> String { + format!( + r#"#include <{header_file}> +#include +#include +#include +#include +template T1 cast(T2 x) {{ + static_assert(sizeof(T1) == sizeof(T2), "sizeof T1 and T2 must be the same"); + T1 ret = 0; + memcpy(&ret, &x, sizeof(T1)); + return ret; +}} +std::ostream& operator<<(std::ostream& os, poly128_t value) {{ + std::stringstream temp; + do {{ + int n = value % 10; + value /= 10; + temp << n; + }} while (value != 0); + std::string tempstr(temp.str()); + std::string res(tempstr.rbegin(), tempstr.rend()); + os << res; + return os; +}} +int main(int argc, char **argv) {{ +{passes} + return 0; +}}"#, + header_file = header_file, + passes = (1..20) + .map(|idx| intrinsic.generate_pass_c(idx)) + .collect::>() + .join("\n"), + ) +} + +fn generate_rust_program(intrinsic: &Intrinsic) -> String { + format!( + r#"#![feature(simd_ffi)] +#![feature(link_llvm_intrinsics)] +#![feature(stdsimd)] +#![allow(overflowing_literals)] +use core::arch::aarch64::*; +#[allow(unused_macros)] +macro_rules! bits_to_float( + ($t:ty, $bits:expr) => ( + {{ let x: $t = ::std::mem::transmute($bits); x }} + ) +); + +fn main() {{ +{passes} +}} +"#, + passes = (1..20) + .map(|idx| intrinsic.generate_pass_rust(idx)) + .collect::>() + .join("\n"), + ) +} + +fn compile_c(c_filename: &str, intrinsic: &Intrinsic, compiler: &str) -> bool { + let output = Command::new("sh") + .arg("-c") + .arg(format!( + "{cpp} -Wno-narrowing -O2 -target {target} -o c_programs/{intrinsic} {filename}", + target = "aarch64-unknown-linux-gnu", + filename = c_filename, + intrinsic = intrinsic.name, + cpp = compiler, + )) + .output(); + if let Ok(output) = output { + if output.status.success() { + true + } else { + let stderr = std::str::from_utf8(&output.stderr).unwrap_or(""); + if stderr.contains("error: use of undeclared identifier") { + warn!("Skipping intrinsic due to no support: {}", intrinsic.name); + true + } else { + error!( + "Failed to compile code for intrinsic: {}\n\nstdout:\n{}\n\nstderr:\n{}", + intrinsic.name, + std::str::from_utf8(&output.stdout).unwrap_or(""), + std::str::from_utf8(&output.stderr).unwrap_or("") + ); + false + } + } + } else { + error!("Command failed: {:#?}", output); + false + } +} + +fn build_c(intrinsics: &Vec, compiler: &str) -> bool { + let _ = std::fs::create_dir("c_programs"); + intrinsics + .par_iter() + .map(|i| { + let c_filename = format!(r#"c_programs/{}.cpp"#, i.name); + let mut file = File::create(&c_filename).unwrap(); + + let c_code = generate_c_program("arm_neon.h", &i); + file.write_all(c_code.into_bytes().as_slice()).unwrap(); + compile_c(&c_filename, &i, compiler) + }) + .find_any(|x| !x) + .is_none() +} + +fn build_rust(intrinsics: &Vec, toolchain: &str) -> bool { + intrinsics.iter().for_each(|i| { + let rust_dir = format!(r#"rust_programs/{}"#, i.name); + let _ = std::fs::create_dir_all(&rust_dir); + let rust_filename = format!(r#"{}/main.rs"#, rust_dir); + let mut file = File::create(&rust_filename).unwrap(); + + let c_code = generate_rust_program(&i); + file.write_all(c_code.into_bytes().as_slice()).unwrap(); + }); + + let mut cargo = File::create("rust_programs/Cargo.toml").unwrap(); + cargo + .write_all( + format!( + r#"[package] +name = "intrinsic-test" +version = "{version}" +authors = ["{authors}"] +edition = "2018" +[workspace] +{binaries}"#, + version = env!("CARGO_PKG_VERSION"), + authors = env!("CARGO_PKG_AUTHORS"), + binaries = intrinsics + .iter() + .map(|i| { + format!( + r#"[[bin]] +name = "{intrinsic}" +path = "{intrinsic}/main.rs""#, + intrinsic = i.name + ) + }) + .collect::>() + .join("\n") + ) + .into_bytes() + .as_slice(), + ) + .unwrap(); + + let output = Command::new("sh") + .current_dir("rust_programs") + .arg("-c") + .arg(format!( + "cargo +{toolchain} build --release", + toolchain = toolchain, + )) + .output(); + if let Ok(output) = output { + if output.status.success() { + true + } else { + error!( + "Failed to compile code for intrinsics\n\nstdout:\n{}\n\nstderr:\n{}", + std::str::from_utf8(&output.stdout).unwrap_or(""), + std::str::from_utf8(&output.stderr).unwrap_or("") + ); + false + } + } else { + error!("Command failed: {:#?}", output); + false + } +} + +fn main() { + pretty_env_logger::init(); + + let matches = App::new("Intrinsic test tool") + .about("Generates Rust and C programs for intrinsics and compares the output") + .arg( + Arg::with_name("INPUT") + .help("The input file containing the intrinsics") + .required(true) + .index(1), + ) + .arg( + Arg::with_name("TOOLCHAIN") + .takes_value(true) + .default_value("nightly") + .long("toolchain") + .help("The rust toolchain to use for building the rust code"), + ) + .arg( + Arg::with_name("CPPCOMPILER") + .takes_value(true) + .default_value("clang++") + .long("cppcompiler") + .help("The C++ compiler to use for compiling the c++ code"), + ) + .get_matches(); + + let filename = matches.value_of("INPUT").unwrap(); + let toolchain = matches.value_of("TOOLCHAIN").unwrap(); + let cpp_compiler = matches.value_of("CPPCOMPILER").unwrap(); + + let mut csv_reader = csv::Reader::from_reader(std::fs::File::open(filename).unwrap()); + + let mut intrinsics = csv_reader + .deserialize() + .filter_map(|x| -> Option { + debug!("Processing {:#?}", x); + match x { + Ok(a) => Some(a), + e => { + error!("{:#?}", e); + None + } + } + }) + // Only perform the test for intrinsics that are enabled... + .filter(|i| i.enabled) + // Not sure how we would compare intrinsic that returns void. + .filter(|i| i.results.kind() != TypeKind::Void) + .filter(|i| i.results.kind() != TypeKind::BFloat) + .filter(|i| !(i.results.kind() == TypeKind::Float && i.results.inner_size() == 16)) + .filter(|i| { + i.arguments + .iter() + .find(|a| a.ty.kind() == TypeKind::BFloat) + .is_none() + }) + .filter(|i| { + i.arguments + .iter() + .find(|a| a.ty.kind() == TypeKind::Float && a.ty.inner_size() == 16) + .is_none() + }) + // Skip pointers for now, we would probably need to look at the return + // type to work out how many elements we need to point to. + .filter(|i| i.arguments.iter().find(|a| a.is_ptr()).is_none()) + // intrinsics with a lane parameter have constraints, deal with them later. + .filter(|i| { + i.arguments + .iter() + .find(|a| a.name.starts_with("lane")) + .is_none() + }) + .filter(|i| i.arguments.iter().find(|a| a.name == "n").is_none()) + .collect::>(); + intrinsics.dedup(); + + if !build_c(&intrinsics, cpp_compiler) { + std::process::exit(2); + } + + if !build_rust(&intrinsics, &toolchain) { + std::process::exit(3); + } + + if !compare_outputs(&intrinsics, &toolchain) { + std::process::exit(1) + } +} + +enum FailureReason { + RunC(String), + RunRust(String), + Difference(String, String, String), +} + +fn compare_outputs(intrinsics: &Vec, toolchain: &str) -> bool { + let intrinsics = intrinsics + .par_iter() + .filter_map(|intrinsic| { + let c = Command::new("sh") + .arg("-c") + .arg(format!( + "./c_programs/{intrinsic}", + intrinsic = intrinsic.name, + )) + .output(); + let rust = Command::new("sh") + .current_dir("rust_programs") + .arg("-c") + .arg(format!( + "cargo +{toolchain} run --release --bin {intrinsic}", + intrinsic = intrinsic.name, + toolchain = toolchain, + )) + .output(); + + let (c, rust) = match (c, rust) { + (Ok(c), Ok(rust)) => (c, rust), + a => panic!("{:#?}", a), + }; + + if !c.status.success() { + error!("Failed to run C program for intrinsic {}", intrinsic.name); + return Some(FailureReason::RunC(intrinsic.name.clone())); + } + + if !rust.status.success() { + error!( + "Failed to run rust program for intrinsic {}", + intrinsic.name + ); + return Some(FailureReason::RunRust(intrinsic.name.clone())); + } + + info!("Comparing intrinsic: {}", intrinsic.name); + + let c = std::str::from_utf8(&c.stdout) + .unwrap() + .to_lowercase() + .replace("-nan", "nan"); + let rust = std::str::from_utf8(&rust.stdout) + .unwrap() + .to_lowercase() + .replace("-nan", "nan"); + + if c == rust { + None + } else { + Some(FailureReason::Difference(intrinsic.name.clone(), c, rust)) + } + }) + .collect::>(); + + intrinsics.iter().for_each(|reason| match reason { + FailureReason::Difference(intrinsic, c, rust) => { + println!("Difference for intrinsic: {}", intrinsic); + let diff = diff::lines(c, rust); + diff.iter().for_each(|diff| match diff { + diff::Result::Left(c) => println!("C: {}", c), + diff::Result::Right(rust) => println!("Rust: {}", rust), + diff::Result::Both(_, _) => (), + }); + println!("****************************************************************"); + } + FailureReason::RunC(intrinsic) => { + println!("Failed to run C program for intrinsic {}", intrinsic) + } + FailureReason::RunRust(intrinsic) => { + println!("Failed to run rust program for intrinsic {}", intrinsic) + } + }); + println!("{} differences found", intrinsics.len()); + intrinsics.is_empty() +} diff --git a/crates/intrinsic-test/src/types.rs b/crates/intrinsic-test/src/types.rs new file mode 100644 index 0000000000..c037ed3ae5 --- /dev/null +++ b/crates/intrinsic-test/src/types.rs @@ -0,0 +1,483 @@ +use regex::Regex; +use serde::{Deserialize, Deserializer}; +use std::fmt; +use std::str::FromStr; + +use crate::values::values_for_pass; +use crate::Language; + +#[derive(Debug, PartialEq, Copy, Clone)] +pub enum TypeKind { + BFloat, + Float, + Int, + UInt, + Poly, + Void, +} + +impl FromStr for TypeKind { + type Err = String; + + fn from_str(s: &str) -> Result { + match s { + "bfloat" => Ok(Self::BFloat), + "float" => Ok(Self::Float), + "int" => Ok(Self::Int), + "poly" => Ok(Self::Poly), + "uint" | "unsigned" => Ok(Self::UInt), + "void" => Ok(Self::Void), + _ => Err(format!("Impossible to parse argument kind {}", s)), + } + } +} + +impl fmt::Display for TypeKind { + fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { + write!( + f, + "{}", + match self { + Self::BFloat => "bfloat", + Self::Float => "float", + Self::Int => "int", + Self::UInt => "uint", + Self::Poly => "poly", + Self::Void => "void", + } + ) + } +} + +impl TypeKind { + /// Gets the type part of a c typedef for a type that's in the form of {type}{size}_t. + pub fn c_prefix(&self) -> &str { + match self { + Self::Float => "float", + Self::Int => "int", + Self::UInt => "uint", + Self::Poly => "poly", + _ => unreachable!("Not used: {:#?}", self), + } + } + + /// Gets the rust prefix for the type kind i.e. i, u, f. + pub fn rust_prefix(&self) -> &str { + match self { + Self::Float => "f", + Self::Int => "i", + Self::UInt => "u", + Self::Poly => "u", + _ => unreachable!("Unused type kind: {:#?}", self), + } + } +} + +#[derive(Debug, PartialEq, Clone)] +pub enum IntrinsicType { + Ptr { + constant: bool, + child: Box, + }, + Type { + constant: bool, + kind: TypeKind, + /// The bit length of this type (e.g. 32 for u32). + bit_len: Option, + + /// Length of the SIMD vector (i.e. 4 for uint32x4_t), A value of `None` + /// means this is not a simd type. A `None` can be assumed to be 1, + /// although in some places a distinction is needed between `u64` and + /// `uint64x1_t` this signals that. + simd_len: Option, + + /// The number of rows for SIMD matrices (i.e. 2 for uint8x8x2_t). + /// A value of `None` represents a type that does not contain any + /// rows encoded in the type (e.g. uint8x8_t). + /// A value of `None` can be assumed to be 1 though. + vec_len: Option, + }, +} + +impl IntrinsicType { + /// Get the TypeKind for this type, recursing into pointers. + pub fn kind(&self) -> TypeKind { + match *self { + IntrinsicType::Ptr { ref child, .. } => child.kind(), + IntrinsicType::Type { kind, .. } => kind, + } + } + + /// Get the size of a single element inside this type, recursing into + /// pointers, i.e. a pointer to a u16 would be 16 rather than the size + /// of a pointer. + pub fn inner_size(&self) -> u32 { + match *self { + IntrinsicType::Ptr { ref child, .. } => child.inner_size(), + IntrinsicType::Type { + bit_len: Some(bl), .. + } => bl, + _ => unreachable!(""), + } + } + + pub fn num_lanes(&self) -> u32 { + match *self { + IntrinsicType::Ptr { ref child, .. } => child.num_lanes(), + IntrinsicType::Type { + simd_len: Some(sl), .. + } => sl, + _ => 1, + } + } + + /// Determine if the type is a simd type, this will treat a type such as + /// `uint64x1` as simd. + pub fn is_simd(&self) -> bool { + match *self { + IntrinsicType::Ptr { ref child, .. } => child.is_simd(), + IntrinsicType::Type { + simd_len: None, + vec_len: None, + .. + } => false, + _ => true, + } + } + + pub fn is_ptr(&self) -> bool { + match *self { + IntrinsicType::Ptr { .. } => true, + IntrinsicType::Type { .. } => false, + } + } + + pub fn from_rust(ty: &str) -> Result { + lazy_static! { + static ref SIMD_TYPE: Regex = Regex::new(r#"([a-z]*)([0-9]*)x([0-9]*)_t"#).unwrap(); + static ref MULTI_SIMD_TYPE: Regex = + Regex::new(r#"([a-z]*)([0-9]*)x([0-9]*)x([0-9]*)_t"#).unwrap(); + static ref RUST_TYPE: Regex = Regex::new(r#"([iuf]|float|poly)([0-9]+)"#).unwrap(); + } + + debug!("Parsing type: {}", ty); + + if let Some(ty) = ty.strip_prefix('*') { + let (constant, ty) = if let Some(ty) = ty.strip_prefix("const") { + (true, ty.trim()) + } else if let Some(ty) = ty.strip_prefix("mut") { + (false, ty.trim()) + } else { + (false, ty) + }; + return Ok(Self::Ptr { + constant, + child: Box::new(Self::from_rust(ty)?), + }); + } + + let (constant, ty) = if let Some(ty) = ty.strip_prefix("const") { + (true, ty.trim()) + } else { + (false, ty) + }; + + if let Some(captures) = MULTI_SIMD_TYPE.captures(ty) { + let kind = captures + .get(1) + .map(|s| s.as_str().parse::().unwrap()) + .unwrap(); + let bit_len = captures.get(2).map(|s| s.as_str().parse::().unwrap()); + let simd_len = captures.get(3).map(|s| s.as_str().parse::().unwrap()); + let vec_len = captures.get(4).map(|s| s.as_str().parse::().unwrap()); + Ok(Self::Type { + constant, + kind, + bit_len, + simd_len, + vec_len, + }) + } else if let Some(captures) = SIMD_TYPE.captures(ty) { + let kind = captures + .get(1) + .map(|s| s.as_str().parse::().unwrap()) + .unwrap(); + let bit_len = captures.get(2).map(|s| s.as_str().parse::().unwrap()); + let simd_len = captures.get(3).map(|s| s.as_str().parse::().unwrap()); + + Ok(Self::Type { + constant, + kind, + bit_len, + simd_len, + vec_len: None, + }) + } else if let Some(captures) = RUST_TYPE.captures(ty) { + let kind = captures + .get(1) + .map(|s| match s.as_str() { + "i" => TypeKind::Int, + "u" => TypeKind::UInt, + "f" => TypeKind::Float, + "float" => TypeKind::Float, + "poly" => TypeKind::Poly, + a => panic!("Unexpected type: {} found", a), + }) + .unwrap(); + let bit_len = captures.get(2).map(|s| s.as_str().parse::().unwrap()); + Ok(Self::Type { + constant, + kind, + bit_len, + simd_len: None, + vec_len: None, + }) + } else { + match ty { + "int" => Ok(Self::Type { + constant, + kind: TypeKind::Int, + bit_len: Some(32), + simd_len: None, + vec_len: None, + }), + "void" => Ok(Self::Type { + constant: false, + kind: TypeKind::Void, + bit_len: None, + simd_len: None, + vec_len: None, + }), + _ => Err(format!("Failed to parse type: {}", ty)), + } + } + } + + #[allow(unused)] + fn c_scalar_type(&self) -> String { + format!( + "{prefix}{bits}_t", + prefix = self.kind().c_prefix(), + bits = self.inner_size() + ) + } + + fn rust_scalar_type(&self) -> String { + format!( + "{prefix}{bits}", + prefix = self.kind().rust_prefix(), + bits = self.inner_size() + ) + } + + /// Gets a string containing the typename for this type in C format. + pub fn c_type(&self) -> String { + match self { + IntrinsicType::Ptr { child, .. } => child.c_type(), + IntrinsicType::Type { + constant, + kind, + bit_len: Some(bit_len), + simd_len: None, + vec_len: None, + .. + } => format!( + "{}{}{}_t", + if *constant { "const " } else { "" }, + kind.c_prefix(), + bit_len + ), + IntrinsicType::Type { + kind, + bit_len: Some(bit_len), + simd_len: Some(simd_len), + vec_len: None, + .. + } => format!("{}{}x{}_t", kind.c_prefix(), bit_len, simd_len), + IntrinsicType::Type { + kind, + bit_len: Some(bit_len), + simd_len: Some(simd_len), + vec_len: Some(vec_len), + .. + } => format!("{}{}x{}x{}_t", kind.c_prefix(), bit_len, simd_len, vec_len), + _ => todo!("{:#?}", self), + } + } + + /// Gets a cast for this type if needs promotion. + /// This is required for 8 bit types due to printing as the 8 bit types use + /// a char and when using that in `std::cout` it will print as a character, + /// which means value of 0 will be printed as a null byte. + pub fn c_promotion(&self) -> &str { + match *self { + IntrinsicType::Type { + kind, + bit_len: Some(bit_len), + .. + } if bit_len == 8 => match kind { + TypeKind::Int => "(int)", + TypeKind::UInt => "(unsigned int)", + TypeKind::Poly => "(unsigned int)", + _ => "", + }, + _ => "", + } + } + + /// Generates a comma list of values that can be used to initialize an + /// argument for the intrinsic call. + /// This is determistic based on the pass number. + /// + /// * `pass`: The pass index, i.e. the iteration index for the call to an intrinsic + /// + /// Returns a string such as + /// * `0x1, 0x7F, 0xFF` if `language` is `Language::C` + /// * `0x1 as _, 0x7F as _, 0xFF as _` if `language` is `Language::Rust` + pub fn populate_random(&self, pass: usize, language: &Language) -> String { + match self { + IntrinsicType::Ptr { child, .. } => child.populate_random(pass, language), + IntrinsicType::Type { + bit_len: Some(bit_len), + kind, + simd_len, + vec_len, + .. + } if kind == &TypeKind::Int || kind == &TypeKind::UInt || kind == &TypeKind::Poly => (0 + ..(simd_len.unwrap_or(1) * vec_len.unwrap_or(1))) + .map(|i| { + format!( + "{}{}", + values_for_pass(*bit_len, i, pass), + match language { + &Language::Rust => format!(" as {ty} ", ty = self.rust_scalar_type()), + &Language::C => String::from(""), + } + ) + }) + .collect::>() + .join(","), + IntrinsicType::Type { + kind: TypeKind::Float, + bit_len: Some(32), + simd_len, + vec_len, + .. + } => (0..(simd_len.unwrap_or(1) * vec_len.unwrap_or(1))) + .map(|i| { + format!( + "{}{})", + match language { + &Language::Rust => "bits_to_float!(f32, ", + &Language::C => "cast(", + }, + values_for_pass(32, i, pass), + ) + }) + .collect::>() + .join(","), + IntrinsicType::Type { + kind: TypeKind::Float, + bit_len: Some(64), + simd_len, + vec_len, + .. + } => (0..(simd_len.unwrap_or(1) * vec_len.unwrap_or(1))) + .map(|i| { + format!( + "{}{}{})", + match language { + &Language::Rust => "bits_to_float!(f64,", + &Language::C => "cast(", + }, + values_for_pass(64, i, pass), + match language { + &Language::Rust => " as u64", + &Language::C => "", + } + ) + }) + .collect::>() + .join(","), + _ => unreachable!("populate random: {:#?}", self), + } + } + + /// Determines the load function for this type. + #[allow(unused)] + pub fn get_load_function(&self) -> String { + match self { + IntrinsicType::Ptr { child, .. } => child.get_load_function(), + IntrinsicType::Type { + kind: k, + bit_len: Some(bl), + simd_len, + vec_len, + .. + } => { + let quad = if (simd_len.unwrap_or(1) * bl) > 64 { + "q" + } else { + "" + }; + format!( + "vld{len}{quad}_{type}{size}", + type = match k { + TypeKind::UInt => "u", + TypeKind::Int => "s", + TypeKind::Float => "f", + TypeKind::Poly => "p", + x => todo!("get_load_function TypeKind: {:#?}", x), + }, + size = bl, + quad = quad, + len = vec_len.unwrap_or(1), + ) + } + _ => todo!("get_load_function IntrinsicType: {:#?}", self), + } + } + + /// Determines the get lane function for this type. + pub fn get_lane_function(&self) -> String { + match self { + IntrinsicType::Ptr { child, .. } => child.get_lane_function(), + IntrinsicType::Type { + kind: k, + bit_len: Some(bl), + simd_len, + .. + } => { + let quad = if (simd_len.unwrap_or(1) * bl) > 64 { + "q" + } else { + "" + }; + format!( + "vget{quad}_lane_{type}{size}", + type = match k { + TypeKind::UInt => "u", + TypeKind::Int => "s", + TypeKind::Float => "f", + TypeKind::Poly => "p", + x => todo!("get_load_function TypeKind: {:#?}", x), + }, + size = bl, + quad = quad, + ) + } + _ => todo!("get_lane_function IntrinsicType: {:#?}", self), + } + } +} + +impl<'de> Deserialize<'de> for IntrinsicType { + fn deserialize(deserializer: D) -> Result + where + D: Deserializer<'de>, + { + use serde::de::Error; + let s = String::deserialize(deserializer)?; + Self::from_rust(&s).map_err(Error::custom) + } +} diff --git a/crates/intrinsic-test/src/values.rs b/crates/intrinsic-test/src/values.rs new file mode 100644 index 0000000000..4565edca09 --- /dev/null +++ b/crates/intrinsic-test/src/values.rs @@ -0,0 +1,126 @@ +/// Gets a hex constant value for a single lane in in a determistic way +/// * `bits`: The number of bits for the type, only 8, 16, 32, 64 are valid values +/// * `simd`: The index of the simd lane we are generating for +/// * `pass`: The index of the pass we are generating the values for +pub fn values_for_pass(bits: u32, simd: u32, pass: usize) -> String { + let index = pass + (simd as usize); + + if bits == 8 { + format!("{:#X}", VALUES_8[index % VALUES_8.len()]) + } else if bits == 16 { + format!("{:#X}", VALUES_16[index % VALUES_16.len()]) + } else if bits == 32 { + format!("{:#X}", VALUES_32[index % VALUES_32.len()]) + } else if bits == 64 { + format!("{:#X}", VALUES_64[index % VALUES_64.len()]) + } else { + panic!("Unknown size: {}", bits); + } +} + +pub const VALUES_8: &[u8] = &[ + 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, + 0xf0, 0x80, 0x3b, 0xff, +]; + +pub const VALUES_16: &[u16] = &[ + 0x0000, // 0.0 + 0x0400, // The smallest normal value. + 0x37ff, // The value just below 0.5. + 0x3800, // 0.5 + 0x3801, // The value just above 0.5. + 0x3bff, // The value just below 1.0. + 0x3c00, // 1.0 + 0x3c01, // The value just above 1.0. + 0x3e00, // 1.5 + 0x4900, // 10 + 0x7bff, // The largest finite value. + 0x7c00, // Infinity. + // NaNs. + // - Quiet NaNs + 0x7f23, 0x7e00, // - Signalling NaNs + 0x7d23, 0x7c01, // Subnormals. + // - A recognisable bit pattern. + 0x0012, // - The largest subnormal value. + 0x03ff, // - The smallest subnormal value. + 0x0001, // The same values again, but negated. + 0x8000, 0x8400, 0xb7ff, 0xb800, 0xb801, 0xbbff, 0xbc00, 0xbc01, 0xbe00, 0xc900, 0xfbff, 0xfc00, + 0xff23, 0xfe00, 0xfd23, 0xfc01, 0x8012, 0x83ff, 0x8001, +]; + +pub const VALUES_32: &[u32] = &[ + // Simple values. + 0x00000000, // 0.0 + 0x00800000, // The smallest normal value. + 0x3effffff, // The value just below 0.5. + 0x3f000000, // 0.5 + 0x3f000001, // The value just above 0.5. + 0x3f7fffff, // The value just below 1.0. + 0x3f800000, // 1.0 + 0x3f800001, // The value just above 1.0. + 0x3fc00000, // 1.5 + 0x41200000, // 10 + 0x7f8fffff, // The largest finite value. + 0x7f800000, // Infinity. + // NaNs. + // - Quiet NaNs + 0x7fd23456, 0x7fc00000, // - Signalling NaNs + 0x7f923456, 0x7f800001, // Subnormals. + // - A recognisable bit pattern. + 0x00123456, // - The largest subnormal value. + 0x007fffff, // - The smallest subnormal value. + 0x00000001, // The same values again, but negated. + 0x80000000, 0x80800000, 0xbeffffff, 0xbf000000, 0xbf000001, 0xbf7fffff, 0xbf800000, 0xbf800001, + 0xbfc00000, 0xc1200000, 0xff8fffff, 0xff800000, 0xffd23456, 0xffc00000, 0xff923456, 0xff800001, + 0x80123456, 0x807fffff, 0x80000001, +]; + +pub const VALUES_64: &[u64] = &[ + // Simple values. + 0x0000000000000000, // 0.0 + 0x0010000000000000, // The smallest normal value. + 0x3fdfffffffffffff, // The value just below 0.5. + 0x3fe0000000000000, // 0.5 + 0x3fe0000000000001, // The value just above 0.5. + 0x3fefffffffffffff, // The value just below 1.0. + 0x3ff0000000000000, // 1.0 + 0x3ff0000000000001, // The value just above 1.0. + 0x3ff8000000000000, // 1.5 + 0x4024000000000000, // 10 + 0x7fefffffffffffff, // The largest finite value. + 0x7ff0000000000000, // Infinity. + // NaNs. + // - Quiet NaNs + 0x7ff923456789abcd, + 0x7ff8000000000000, + // - Signalling NaNs + 0x7ff123456789abcd, + 0x7ff0000000000000, + // Subnormals. + // - A recognisable bit pattern. + 0x000123456789abcd, + // - The largest subnormal value. + 0x000fffffffffffff, + // - The smallest subnormal value. + 0x0000000000000001, + // The same values again, but negated. + 0x8000000000000000, + 0x8010000000000000, + 0xbfdfffffffffffff, + 0xbfe0000000000000, + 0xbfe0000000000001, + 0xbfefffffffffffff, + 0xbff0000000000000, + 0xbff0000000000001, + 0xbff8000000000000, + 0xc024000000000000, + 0xffefffffffffffff, + 0xfff0000000000000, + 0xfff923456789abcd, + 0xfff8000000000000, + 0xfff123456789abcd, + 0xfff0000000000000, + 0x800123456789abcd, + 0x800fffffffffffff, + 0x8000000000000001, +];