Re-organize intrinsic-test to enable seamless addition of behaviour testing for more architectures

crates/intrinsic-test/src/arm/config.rs

@@ -0,0 +1,34 @@
+pub fn build_notices(line_prefix: &str) -> String {
+    format!(
+        "\
+{line_prefix}This is a transient test file, not intended for distribution. Some aspects of the
+{line_prefix}test are derived from a JSON specification, published under the same license as the
+{line_prefix}`intrinsic-test` crate.\n
+"
+    )
+}
+
+pub const POLY128_OSTREAM_DEF: &str = r#"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;
+}"#;
+
+pub const AARCH_CONFIGURATIONS: &str = r#"
+#![cfg_attr(target_arch = "arm", feature(stdarch_arm_neon_intrinsics))]
+#![cfg_attr(target_arch = "arm", feature(stdarch_aarch32_crc32))]
+#![cfg_attr(any(target_arch = "aarch64", target_arch = "arm64ec"), feature(stdarch_neon_fcma))]
+#![cfg_attr(any(target_arch = "aarch64", target_arch = "arm64ec"), feature(stdarch_neon_dotprod))]
+#![cfg_attr(any(target_arch = "aarch64", target_arch = "arm64ec"), feature(stdarch_neon_i8mm))]
+#![cfg_attr(any(target_arch = "aarch64", target_arch = "arm64ec"), feature(stdarch_neon_sha3))]
+#![cfg_attr(any(target_arch = "aarch64", target_arch = "arm64ec"), feature(stdarch_neon_sm4))]
+#![cfg_attr(any(target_arch = "aarch64", target_arch = "arm64ec"), feature(stdarch_neon_ftts))]
+#![feature(stdarch_neon_f16)]
+"#;

crates/intrinsic-test/src/arm/functions.rs

@@ -0,0 +1,270 @@
+use super::config::{AARCH_CONFIGURATIONS, POLY128_OSTREAM_DEF, build_notices};
+use super::intrinsic::ArmIntrinsicType;
+use crate::common::argument::Argument;
+use crate::common::compile_c::CompilationCommandBuilder;
+use crate::common::gen_c::{compile_c, create_c_filenames, generate_c_program};
+use crate::common::gen_rust::{compile_rust, create_rust_filenames, generate_rust_program};
+use crate::common::indentation::Indentation;
+use crate::common::intrinsic::{Intrinsic, IntrinsicDefinition};
+use crate::common::intrinsic_types::IntrinsicTypeDefinition;
+use crate::common::write_file;
+use itertools::Itertools;
+use rayon::prelude::*;
+
+// The number of times each intrinsic will be called.
+const PASSES: u32 = 20;
+
+fn gen_code_c(
+    indentation: Indentation,
+    intrinsic: &Intrinsic<ArmIntrinsicType>,
+    constraints: &[&Argument<ArmIntrinsicType>],
+    name: String,
+    target: &str,
+) -> String {
+    if let Some((current, constraints)) = constraints.split_last() {
+        let range = current
+            .constraint
+            .iter()
+            .map(|c| c.to_range())
+            .flat_map(|r| r.into_iter());
+
+        let body_indentation = indentation.nested();
+        range
+            .map(|i| {
+                format!(
+                    "{indentation}{{\n\
+                        {body_indentation}{ty} {name} = {val};\n\
+                        {pass}\n\
+                    {indentation}}}",
+                    name = current.name,
+                    ty = current.ty.c_type(),
+                    val = i,
+                    pass = gen_code_c(
+                        body_indentation,
+                        intrinsic,
+                        constraints,
+                        format!("{name}-{i}"),
+                        target,
+                    )
+                )
+            })
+            .join("\n")
+    } else {
+        intrinsic.generate_loop_c(indentation, &name, PASSES, target)
+    }
+}
+
+fn generate_c_program_arm(
+    header_files: &[&str],
+    intrinsic: &Intrinsic<ArmIntrinsicType>,
+    target: &str,
+) -> String {
+    let constraints = intrinsic
+        .arguments
+        .iter()
+        .filter(|&i| i.has_constraint())
+        .collect_vec();
+
+    let indentation = Indentation::default();
+    generate_c_program(
+        build_notices("// ").as_str(),
+        header_files,
+        "aarch64",
+        &[POLY128_OSTREAM_DEF],
+        intrinsic
+            .arguments
+            .gen_arglists_c(indentation, PASSES)
+            .as_str(),
+        gen_code_c(
+            indentation.nested(),
+            intrinsic,
+            constraints.as_slice(),
+            Default::default(),
+            target,
+        )
+        .as_str(),
+    )
+}
+
+fn gen_code_rust(
+    indentation: Indentation,
+    intrinsic: &Intrinsic<ArmIntrinsicType>,
+    constraints: &[&Argument<ArmIntrinsicType>],
+    name: String,
+) -> String {
+    println!("{}", name);
+    if let Some((current, constraints)) = constraints.split_last() {
+        let range = current
+            .constraint
+            .iter()
+            .map(|c| c.to_range())
+            .flat_map(|r| r.into_iter());
+
+        let body_indentation = indentation.nested();
+        range
+            .map(|i| {
+                format!(
+                    "{indentation}{{\n\
+                        {body_indentation}const {name}: {ty} = {val};\n\
+                        {pass}\n\
+                    {indentation}}}",
+                    name = current.name,
+                    ty = current.ty.rust_type(),
+                    val = i,
+                    pass = gen_code_rust(
+                        body_indentation,
+                        intrinsic,
+                        constraints,
+                        format!("{name}-{i}")
+                    )
+                )
+            })
+            .join("\n")
+    } else {
+        intrinsic.generate_loop_rust(indentation, &name, PASSES)
+    }
+}
+
+fn generate_rust_program_arm(intrinsic: &Intrinsic<ArmIntrinsicType>, target: &str) -> String {
+    let constraints = intrinsic
+        .arguments
+        .iter()
+        .filter(|i| i.has_constraint())
+        .collect_vec();
+
+    let indentation = Indentation::default();
+    let final_target = if target.contains("v7") {
+        "arm"
+    } else {
+        "aarch64"
+    };
+    generate_rust_program(
+        build_notices("// ").as_str(),
+        AARCH_CONFIGURATIONS,
+        final_target,
+        intrinsic
+            .arguments
+            .gen_arglists_rust(indentation.nested(), PASSES)
+            .as_str(),
+        gen_code_rust(
+            indentation.nested(),
+            intrinsic,
+            &constraints,
+            Default::default(),
+        )
+        .as_str(),
+    )
+}
+
+fn compile_c_arm(
+    intrinsics_name_list: &Vec<String>,
+    compiler: &str,
+    target: &str,
+    cxx_toolchain_dir: Option<&str>,
+) -> bool {
+    // -ffp-contract=off emulates Rust's approach of not fusing separate mul-add operations
+    let mut command = CompilationCommandBuilder::new()
+        .add_arch_flags(vec!["armv8.6-a", "crypto", "crc", "dotprod", "fp16"])
+        .set_compiler(compiler)
+        .set_target(target)
+        .set_opt_level("2")
+        .set_cxx_toolchain_dir(cxx_toolchain_dir)
+        .set_project_root("c_programs")
+        .add_extra_flags(vec!["-ffp-contract=off", "-Wno-narrowing"]);
+
+    if !target.contains("v7") {
+        command = command.add_arch_flags(vec!["faminmax", "lut", "sha3"]);
+    }
+
+    /*
+     * clang++ cannot link an aarch64_be object file, so we invoke
+     * aarch64_be-unknown-linux-gnu's C++ linker. This ensures that we
+     * are testing the intrinsics against LLVM.
+     *
+     * Note: setting `--sysroot=<...>` which is the obvious thing to do
+     * does not work as it gets caught up with `#include_next <stdlib.h>`
+     * not existing...
+     */
+    if target == "aarch64_be-unknown-linux-gnu" {
+        command = command
+            .set_linker(
+                cxx_toolchain_dir.unwrap_or("").to_string() + "/bin/aarch64_be-none-linux-gnu-g++",
+            )
+            .set_include_paths(vec![
+                "/include",
+                "/aarch64_be-none-linux-gnu/include",
+                "/aarch64_be-none-linux-gnu/include/c++/14.2.1",
+                "/aarch64_be-none-linux-gnu/include/c++/14.2.1/aarch64_be-none-linux-gnu",
+                "/aarch64_be-none-linux-gnu/include/c++/14.2.1/backward",
+                "/aarch64_be-none-linux-gnu/libc/usr/include",
+            ]);
+    }
+
+    if !compiler.contains("clang") {
+        command = command.add_extra_flag("-flax-vector-conversions");
+    }
+
+    let compiler_commands = intrinsics_name_list
+        .iter()
+        .map(|intrinsic_name| {
+            command
+                .clone()
+                .set_input_name(intrinsic_name)
+                .set_output_name(intrinsic_name)
+                .to_string()
+        })
+        .collect::<Vec<_>>();
+
+    compile_c(&compiler_commands)
+}
+
+pub fn build_c(
+    intrinsics: &Vec<Intrinsic<ArmIntrinsicType>>,
+    compiler: Option<&str>,
+    target: &str,
+    cxx_toolchain_dir: Option<&str>,
+) -> bool {
+    let intrinsics_name_list = intrinsics
+        .par_iter()
+        .map(|i| i.name.clone())
+        .collect::<Vec<_>>();
+    let filename_mapping = create_c_filenames(&intrinsics_name_list);
+
+    intrinsics.par_iter().for_each(|i| {
+        let c_code = generate_c_program_arm(&["arm_neon.h", "arm_acle.h", "arm_fp16.h"], i, target);
+        match filename_mapping.get(&i.name) {
+            Some(filename) => write_file(filename, c_code),
+            None => {}
+        };
+    });
+
+    match compiler {
+        None => true,
+        Some(compiler) => compile_c_arm(&intrinsics_name_list, compiler, target, cxx_toolchain_dir),
+    }
+}
+
+pub fn build_rust(
+    intrinsics: &[Intrinsic<ArmIntrinsicType>],
+    toolchain: Option<&str>,
+    target: &str,
+    linker: Option<&str>,
+) -> bool {
+    let intrinsics_name_list = intrinsics
+        .par_iter()
+        .map(|i| i.name.clone())
+        .collect::<Vec<_>>();
+    let filename_mapping = create_rust_filenames(&intrinsics_name_list);
+
+    intrinsics.par_iter().for_each(|i| {
+        let rust_code = generate_rust_program_arm(i, target);
+        match filename_mapping.get(&i.name) {
+            Some(filename) => write_file(filename, rust_code),
+            None => {}
+        }
+    });
+
+    let intrinsics_name_list = intrinsics.iter().map(|i| i.name.as_str()).collect_vec();
+
+    compile_rust(&intrinsics_name_list, toolchain, target, linker)
+}

crates/intrinsic-test/src/arm/intrinsic.rs

@@ -0,0 +1,89 @@
+use crate::base_intrinsictype_trait_def_macro;
+use crate::common::argument::ArgumentList;
+use crate::common::cli::Language;
+use crate::common::indentation::Indentation;
+use crate::common::intrinsic::{Intrinsic, IntrinsicDefinition};
+use crate::common::intrinsic_types::{
+    BaseIntrinsicTypeDefinition, IntrinsicTypeDefinition, TypeKind,
+};
+
+base_intrinsictype_trait_def_macro! {ArmIntrinsicType}
+
+impl IntrinsicDefinition<ArmIntrinsicType> for Intrinsic<ArmIntrinsicType> {
+    fn arguments(&self) -> ArgumentList<ArmIntrinsicType> {
+        self.arguments.clone()
+    }
+
+    fn results(&self) -> ArmIntrinsicType {
+        self.results.clone()
+    }
+
+    fn name(&self) -> String {
+        self.name.clone()
+    }
+
+    /// Generates a std::cout for the intrinsics results that will match the
+    /// rust debug output format for the return type. The generated line assumes
+    /// there is an int i in scope which is the current pass number.
+    fn print_result_c(&self, indentation: Indentation, additional: &str) -> String {
+        let lanes = if self.results().num_vectors() > 1 {
+            (0..self.results().num_vectors())
+                .map(|vector| {
+                    format!(
+                        r#""{ty}(" << {lanes} << ")""#,
+                        ty = self.results().c_single_vector_type(),
+                        lanes = (0..self.results().num_lanes())
+                            .map(move |idx| -> std::string::String {
+                                format!(
+                                    "{cast}{lane_fn}(__return_value.val[{vector}], {lane})",
+                                    cast = self.results().c_promotion(),
+                                    lane_fn = self.results().get_lane_function(),
+                                    lane = idx,
+                                    vector = vector,
+                                )
+                            })
+                            .collect::<Vec<_>>()
+                            .join(r#" << ", " << "#)
+                    )
+                })
+                .collect::<Vec<_>>()
+                .join(r#" << ", " << "#)
+        } else 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::<Vec<_>>()
+                .join(r#" << ", " << "#)
+        } else {
+            format!(
+                "{promote}cast<{cast}>(__return_value)",
+                cast = match self.results.kind() {
+                    TypeKind::Float if self.results().inner_size() == 16 => "float16_t".to_string(),
+                    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#"{indentation}std::cout << "Result {additional}-" << i+1 << ": {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 { "" },
+        )
+    }
+}