Rename SIMD_IS_EMULATED to capability-affirmative HAS_HARDWARE_SIMD

This commit is contained in:
Feoramund
2025-05-29 17:17:51 -04:00
parent edbd247972
commit 45219f240e
4 changed files with 16 additions and 15 deletions

View File

@@ -16,11 +16,12 @@ RUNTIME_REQUIRE :: false // !ODIN_TILDE
@(private)
__float16 :: f16 when __ODIN_LLVM_F16_SUPPORTED else u16
SIMD_IS_EMULATED :: true when (ODIN_ARCH == .amd64 || ODIN_ARCH == .i386) && !intrinsics.has_target_feature("sse2") else
true when (ODIN_ARCH == .arm64 || ODIN_ARCH == .arm32) && !intrinsics.has_target_feature("neon") else
true when (ODIN_ARCH == .wasm64p32 || ODIN_ARCH == .wasm32) && !intrinsics.has_target_feature("simd128") else
true when (ODIN_ARCH == .riscv64) && !intrinsics.has_target_feature("v") else
false
HAS_HARDWARE_SIMD :: false when (ODIN_ARCH == .amd64 || ODIN_ARCH == .i386) && !intrinsics.has_target_feature("sse2") else
false when (ODIN_ARCH == .arm64 || ODIN_ARCH == .arm32) && !intrinsics.has_target_feature("neon") else
false when (ODIN_ARCH == .wasm64p32 || ODIN_ARCH == .wasm32) && !intrinsics.has_target_feature("simd128") else
false when (ODIN_ARCH == .riscv64) && !intrinsics.has_target_feature("v") else
true
@(private)
byte_slice :: #force_inline proc "contextless" (data: rawptr, len: int) -> []byte #no_bounds_check {
@@ -241,7 +242,7 @@ memory_equal :: proc "contextless" (x, y: rawptr, n: int) -> bool {
m := uint(0)
if n >= 8 {
when !SIMD_IS_EMULATED {
when HAS_HARDWARE_SIMD {
// Avoid using 256-bit SIMD on platforms where its emulation is
// likely to be less than ideal.
when ODIN_ARCH == .amd64 && intrinsics.has_target_feature("avx2") {
@@ -295,7 +296,7 @@ memory_compare :: proc "contextless" (x, y: rawptr, n: int) -> int #no_bounds_ch
i := uint(0)
m := uint(0)
when !SIMD_IS_EMULATED {
when HAS_HARDWARE_SIMD {
when ODIN_ARCH == .amd64 && intrinsics.has_target_feature("avx2") {
m = n / 32 * 32
for /**/; i < m; i += 32 {
@@ -364,7 +365,7 @@ memory_compare_zero :: proc "contextless" (a: rawptr, n: int) -> int #no_bounds_
bytes := ([^]u8)(a)
if n >= 8 {
when !SIMD_IS_EMULATED {
when HAS_HARDWARE_SIMD {
when ODIN_ARCH == .amd64 && intrinsics.has_target_feature("avx2") {
scanner32: #simd[32]u8
m = n / 32 * 32

View File

@@ -350,7 +350,7 @@ index_byte :: proc "contextless" (s: []byte, c: byte) -> (index: int) #no_bounds
}
c_vec: simd.u8x16 = c
when !simd.IS_EMULATED {
when simd.HAS_HARDWARE_SIMD {
// Note: While this is something that could also logically take
// advantage of AVX512, the various downclocking and power
// consumption related woes make premature to have a dedicated
@@ -485,7 +485,7 @@ last_index_byte :: proc "contextless" (s: []byte, c: byte) -> int #no_bounds_che
}
c_vec: simd.u8x16 = c
when !simd.IS_EMULATED {
when simd.HAS_HARDWARE_SIMD {
// Note: While this is something that could also logically take
// advantage of AVX512, the various downclocking and power
// consumption related woes make premature to have a dedicated

View File

@@ -39,7 +39,7 @@ when ODIN_ARCH == .arm64 || ODIN_ARCH == .arm32 {
// Some targets lack runtime feature detection, and will flat out refuse
// to load binaries that have unknown instructions. This is distinct from
// `simd.IS_EMULATED` as actually good designs support runtime feature
// `simd.HAS_HARDWARE_SIMD` as actually good designs support runtime feature
// detection and that constant establishes a baseline.
//
// See:

View File

@@ -26,12 +26,12 @@ import "base:runtime"
/*
Check if SIMD is software-emulated on a target platform.
This value is `false`, when the compile-time target has the hardware support for
at 128-bit (or wider) SIMD. If the compile-time target lacks the hardware support
for 128-bit SIMD, this value is `true`, and all SIMD operations will likely be
This value is `true`, when the compile-time target has the hardware support for
at least 128-bit (or wider) SIMD. If the compile-time target lacks the hardware support
for 128-bit SIMD, this value is `false`, and all SIMD operations will likely be
emulated.
*/
IS_EMULATED :: runtime.SIMD_IS_EMULATED
HAS_HARDWARE_SIMD :: runtime.HAS_HARDWARE_SIMD
/*
Vector of 16 `u8` lanes (128 bits).