Use load intrinsic and loop for intrinsic-test programs. Add --release flag back to intrinsic-test programs.

Author: Jmc18134

crates/intrinsic-test/missing_aarch64.txt

@@ -67,20 +67,6 @@ vrnd64xq_f64
 vrnd64z_f64
 vrnd64zq_f64
 
-# Takes too long to compile tests
-vcopyq_laneq_u8
-vcopyq_laneq_s8
-vcopyq_laneq_p8
-vcopyq_lane_u8
-vcopyq_lane_s8
-vcopyq_lane_p8
-vcopy_laneq_u8
-vcopy_laneq_s8
-vcopy_laneq_p8
-vcopy_lane_u8
-vcopy_lane_s8
-vcopy_lane_p8
-
 # QEMU 6.0 doesn't support these instructions
 vmmlaq_s32
 vmmlaq_u32

crates/intrinsic-test/src/argument.rs

@@ -1,6 +1,6 @@
 use std::ops::Range;
 
-use crate::types::IntrinsicType;
+use crate::types::{IntrinsicType, TypeKind};
 use crate::Language;
 
 /// An argument for the intrinsic.
@@ -90,49 +90,87 @@ impl ArgumentList {
             .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 {
+    /// Creates a line for each argument that initializes an array for C from which `loads` argument
+    /// values can be loaded  as a sliding window.
+    /// e.g `const int32x2_t a_vals = {0x3effffff, 0x3effffff, 0x3f7fffff}`, if loads=2.
+    pub fn gen_arglists_c(&self, loads: u32) -> String {
         self.iter()
             .filter_map(|arg| {
                 (!arg.has_constraint()).then(|| {
                     format!(
-                        "{ty} {name} = {{ {values} }};",
-                        ty = arg.to_c_type(),
+                        "const {ty} {name}_vals[] = {{ {values} }};",
+                        ty = arg.ty.c_scalar_type(),
                         name = arg.name,
-                        values = arg.ty.populate_random(pass, &Language::C)
+                        values = arg.ty.populate_random(loads, &Language::C)
                     )
                 })
             })
             .collect::<Vec<_>>()
-            .join("\n    ")
+            .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 {
+    /// Creates a line for each argument that initializes an array for Rust from which `loads` argument
+    /// values can be loaded as a sliding window, e.g `const A_VALS: [u32; 20]  = [...];`
+    pub fn gen_arglists_rust(&self, loads: u32) -> String {
         self.iter()
             .filter_map(|arg| {
                 (!arg.has_constraint()).then(|| {
-                    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)
-                        )
-                    }
+                    format!(
+                        "const {upper_name}_VALS: [{ty}; {load_size}] = unsafe{{ [{values}] }};",
+                        upper_name = arg.name.to_uppercase(),
+                        ty = arg.ty.rust_scalar_type(),
+                        load_size = arg.ty.num_lanes() * arg.ty.num_vectors() + loads - 1,
+                        values = arg.ty.populate_random(loads, &Language::Rust)
+                    )
+                })
+            })
+            .collect::<Vec<_>>()
+            .join("\n")
+    }
+
+    /// Creates a line for each argument that initalizes the argument from an array [arg]_vals at
+    /// an offset i using a load intrinsic, in C.
+    /// e.g `uint8x8_t a = vld1_u8(&a_vals[i]);`
+    pub fn load_values_c(&self, p64_armv7_workaround: bool) -> String {
+        self.iter()
+            .filter_map(|arg| {
+                // The ACLE doesn't support 64-bit polynomial loads on Armv7
+                // This and the cast are a workaround for this
+                let armv7_p64 = if let TypeKind::Poly = arg.ty.kind() { p64_armv7_workaround } else { false };
+
+                (!arg.has_constraint()).then(|| {
+                    format!(
+                        "{ty} {name} = {open_cast}{load}(&{name}_vals[i]){close_cast};",
+                        ty = arg.to_c_type(),
+                        name = arg.name,
+                        load = if arg.is_simd() { arg.ty.get_load_function(p64_armv7_workaround) } else { "*".to_string() },
+                        open_cast = if armv7_p64 { format!("cast<{}>(", arg.to_c_type()) } else { "".to_string() },
+                        close_cast = if armv7_p64{ ")".to_string() } else { "".to_string() }
+                    )
                 })
             })
             .collect::<Vec<_>>()
             .join("\n        ")
     }
 
+    /// Creates a line for each argument that initalizes the argument from array [ARG]_VALS at
+    /// an offset i using a load intrinsic, in Rust.
+    /// e.g `let a = vld1_u8(A_VALS.as_ptr().offset(i));`
+    pub fn load_values_rust(&self) -> String {
+        self.iter()
+            .filter_map(|arg| {
+                (!arg.has_constraint()).then(|| {
+                    format!("let {name} = {load}({upper_name}_VALS.as_ptr().offset(i));",
+                        name = arg.name,
+                        upper_name = arg.name.to_uppercase(),
+                        load = if arg.is_simd() { arg.ty.get_load_function(false) } else { "*".to_string() },
+                    )
+                })
+            })
+            .collect::<Vec<_>>()
+            .join("\n            ")
+    }
+
     pub fn iter(&self) -> std::slice::Iter<'_, Argument> {
         self.args.iter()
     }

crates/intrinsic-test/src/intrinsic.rs

@@ -20,8 +20,9 @@ pub struct Intrinsic {
 
 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, additional: &str) -> String {
+    /// 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.
+    pub fn print_result_c(&self, additional: &str) -> String {
         let lanes = if self.results.num_vectors() > 1 {
             (0..self.results.num_vectors())
                 .map(|vector| {
@@ -72,7 +73,7 @@ impl Intrinsic {
         };
 
         format!(
-            r#"std::cout << "Result {additional}-{idx}: {ty}" << std::fixed << std::setprecision(150) <<  {lanes} << "{close}" << std::endl;"#,
+            r#"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 {
@@ -81,45 +82,48 @@ impl Intrinsic {
             close = if self.results.is_simd() { ")" } else { "" },
             lanes = lanes,
             additional = additional,
-            idx = index,
         )
     }
 
-    pub fn generate_pass_rust(&self, index: usize, additional: &str) -> String {
+    pub fn generate_loop_c(&self, additional: &str, passes: u32, p64_armv7_workaround: bool) -> String {
+        format!(
+            r#"  {{
+    for (int i=0; i<{passes}; i++) {{
+        {loaded_args}
+        auto __return_value = {intrinsic_call}({args});
+        {print_result}
+    }}
+  }}"#,
+            loaded_args = self.arguments.load_values_c(p64_armv7_workaround),
+            intrinsic_call = self.name,
+            args = self.arguments.as_call_param_c(),
+            print_result = self.print_result_c(additional)
+        )
+    }
+
+    pub fn generate_loop_rust(&self, additional: &str, passes: u32) -> String {
         let constraints = self.arguments.as_constraint_parameters_rust();
         let constraints = if !constraints.is_empty() {
             format!("::<{}>", constraints)
         } else {
             constraints
         };
 
-        format!(
-            r#"
-    unsafe {{
-        {initialized_args}
-        let res = {intrinsic_call}{const}({args});
-        println!("Result {additional}-{idx}: {{:.150?}}", res);
-    }}"#,
-            initialized_args = self.arguments.init_random_values_rust(index),
-            intrinsic_call = self.name,
-            args = self.arguments.as_call_param_rust(),
-            additional = additional,
-            idx = index,
-            const = constraints,
-        )
-    }
-
-    pub fn generate_pass_c(&self, index: usize, additional: &str) -> String {
         format!(
             r#"  {{
-    {initialized_args}
-    auto __return_value = {intrinsic_call}({args});
-    {print_result}
+    for i in 0..{passes} {{
+        unsafe {{
+            {loaded_args}
+            let __return_value = {intrinsic_call}{const}({args});
+            println!("Result {additional}-{{}}: {{:.150?}}", i+1, __return_value);
+        }}
+    }}
   }}"#,
-            initialized_args = self.arguments.init_random_values_c(index),
+            loaded_args = self.arguments.load_values_rust(),
             intrinsic_call = self.name,
-            args = self.arguments.as_call_param_c(),
-            print_result = self.print_result_c(index, additional)
+            const = constraints,
+            args = self.arguments.as_call_param_rust(),
+            additional = additional,
         )
     }
 }

crates/intrinsic-test/src/main.rs

@@ -23,13 +23,16 @@ mod intrinsic;
 mod types;
 mod values;
 
+// The number of times each intrinsic will be called.
+const PASSES: u32 = 20;
+
 #[derive(Debug, PartialEq)]
 pub enum Language {
     Rust,
     C,
 }
 
-fn gen_code_c(intrinsic: &Intrinsic, constraints: &[&Argument], name: String) -> String {
+fn gen_code_c(intrinsic: &Intrinsic, constraints: &[&Argument], name: String, p64_armv7_workaround: bool) -> String {
     if let Some((current, constraints)) = constraints.split_last() {
         let range = current
             .constraints
@@ -47,19 +50,16 @@ fn gen_code_c(intrinsic: &Intrinsic, constraints: &[&Argument], name: String) ->
                     name = current.name,
                     ty = current.ty.c_type(),
                     val = i,
-                    pass = gen_code_c(intrinsic, constraints, format!("{}-{}", name, i))
+                    pass = gen_code_c(intrinsic, constraints, format!("{}-{}", name, i), p64_armv7_workaround)
                 )
             })
             .collect()
     } else {
-        (1..20)
-            .map(|idx| intrinsic.generate_pass_c(idx, &name))
-            .collect::<Vec<_>>()
-            .join("\n")
+        intrinsic.generate_loop_c(&name, PASSES, p64_armv7_workaround)
     }
 }
 
-fn generate_c_program(header_files: &[&str], intrinsic: &Intrinsic) -> String {
+fn generate_c_program(header_files: &[&str], intrinsic: &Intrinsic, p64_armv7_workaround: bool) -> String {
     let constraints = intrinsic
         .arguments
         .iter()
@@ -75,7 +75,7 @@ fn generate_c_program(header_files: &[&str], intrinsic: &Intrinsic) -> String {
 
 template<typename T1, typename T2> T1 cast(T2 x) {{
   static_assert(sizeof(T1) == sizeof(T2), "sizeof T1 and T2 must be the same");
-  T1 ret = 0;
+  T1 ret{{}};
   memcpy(&ret, &x, sizeof(T1));
   return ret;
 }}
@@ -95,6 +95,8 @@ std::ostream& operator<<(std::ostream& os, poly128_t value) {{
 }}
 #endif
 
+{arglists}
+
 int main(int argc, char **argv) {{
 {passes}
     return 0;
@@ -104,7 +106,8 @@ int main(int argc, char **argv) {{
             .map(|header| format!("#include <{}>", header))
             .collect::<Vec<_>>()
             .join("\n"),
-        passes = gen_code_c(intrinsic, constraints.as_slice(), Default::default()),
+        arglists = intrinsic.arguments.gen_arglists_c(PASSES),
+        passes = gen_code_c(intrinsic, constraints.as_slice(), Default::default(), p64_armv7_workaround),
     )
 }
 
@@ -131,10 +134,7 @@ fn gen_code_rust(intrinsic: &Intrinsic, constraints: &[&Argument], name: String)
             })
             .collect()
     } else {
-        (1..20)
-            .map(|idx| intrinsic.generate_pass_rust(idx, &name))
-            .collect::<Vec<_>>()
-            .join("\n")
+        intrinsic.generate_loop_rust(&name, PASSES)
     }
 }
 
@@ -153,11 +153,14 @@ fn generate_rust_program(intrinsic: &Intrinsic, a32: bool) -> String {
 #![allow(non_upper_case_globals)]
 use core_arch::arch::{target_arch}::*;
 
+{arglists}
+
 fn main() {{
 {passes}
 }}
 "#,
         target_arch = if a32 { "arm" } else { "aarch64" },
+        arglists = intrinsic.arguments.gen_arglists_rust(PASSES),
         passes = gen_code_rust(intrinsic, &constraints, Default::default())
     )
 }
@@ -203,7 +206,7 @@ fn build_c(intrinsics: &Vec<Intrinsic>, compiler: &str, a32: bool) -> bool {
             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", "arm_acle.h"], &i);
+            let c_code = generate_c_program(&["arm_neon.h", "arm_acle.h"], &i, a32);
             file.write_all(c_code.into_bytes().as_slice()).unwrap();
             compile_c(&c_filename, &i, compiler, a32)
         })
@@ -259,7 +262,7 @@ path = "{intrinsic}/main.rs""#,
         .current_dir("rust_programs")
         .arg("-c")
         .arg(format!(
-            "cargo {toolchain} build --target {target}",
+            "cargo {toolchain} build --target {target} --release",
             toolchain = toolchain,
             target = if a32 {
                 "armv7-unknown-linux-gnueabihf"
@@ -407,7 +410,7 @@ fn compare_outputs(intrinsics: &Vec<Intrinsic>, toolchain: &str, runner: &str, a
                 .current_dir("rust_programs")
                 .arg("-c")
                 .arg(format!(
-                    "cargo {toolchain} run --target {target} --bin {intrinsic}",
+                    "cargo {toolchain} run --target {target} --bin {intrinsic} --release",
                     intrinsic = intrinsic.name,
                     toolchain = toolchain,
                     target = if a32 {