std.simd: return comptime_int from suggestVectorSize

This commit is contained in:
Veikka Tuominen 2023-10-13 16:57:50 +03:00
parent f09313dbc4
commit c919e9a280
4 changed files with 8 additions and 8 deletions

View File

@ -82,7 +82,7 @@ pub const HeadersParser = struct {
/// If the amount returned is less than `bytes.len`, you may assume that the parser is in a content state and the /// If the amount returned is less than `bytes.len`, you may assume that the parser is in a content state and the
/// first byte of content is located at `bytes[result]`. /// first byte of content is located at `bytes[result]`.
pub fn findHeadersEnd(r: *HeadersParser, bytes: []const u8) u32 { pub fn findHeadersEnd(r: *HeadersParser, bytes: []const u8) u32 {
const vector_len: comptime_int = comptime @max(std.simd.suggestVectorSize(u8) orelse 1, 8); const vector_len: comptime_int = @max(std.simd.suggestVectorSize(u8) orelse 1, 8);
const len = @as(u32, @intCast(bytes.len)); const len = @as(u32, @intCast(bytes.len));
var index: u32 = 0; var index: u32 = 0;

View File

@ -974,7 +974,7 @@ pub fn indexOfSentinel(comptime T: type, comptime sentinel: T, p: [*:sentinel]co
// The below branch assumes that reading past the end of the buffer is valid, as long // The below branch assumes that reading past the end of the buffer is valid, as long
// as we don't read into a new page. This should be the case for most architectures // as we don't read into a new page. This should be the case for most architectures
// which use paged memory, however should be confirmed before adding a new arch below. // which use paged memory, however should be confirmed before adding a new arch below.
.aarch64, .x86, .x86_64 => if (comptime std.simd.suggestVectorSize(T)) |block_len| { .aarch64, .x86, .x86_64 => if (std.simd.suggestVectorSize(T)) |block_len| {
comptime std.debug.assert(std.mem.page_size % block_len == 0); comptime std.debug.assert(std.mem.page_size % block_len == 0);
const Block = @Vector(block_len, T); const Block = @Vector(block_len, T);
const mask: Block = @splat(sentinel); const mask: Block = @splat(sentinel);
@ -1027,7 +1027,7 @@ test "indexOfSentinel vector paths" {
const allocator = std.testing.allocator; const allocator = std.testing.allocator;
inline for (Types) |T| { inline for (Types) |T| {
const block_len = comptime std.simd.suggestVectorSize(T) orelse continue; const block_len = std.simd.suggestVectorSize(T) orelse continue;
// Allocate three pages so we guarantee a page-crossing address with a full page after // Allocate three pages so we guarantee a page-crossing address with a full page after
const memory = try allocator.alloc(T, 3 * std.mem.page_size / @sizeOf(T)); const memory = try allocator.alloc(T, 3 * std.mem.page_size / @sizeOf(T));
@ -1118,11 +1118,11 @@ pub fn indexOfScalarPos(comptime T: type, slice: []const T, start_index: usize,
!@inComptime() and !@inComptime() and
(@typeInfo(T) == .Int or @typeInfo(T) == .Float) and std.math.isPowerOfTwo(@bitSizeOf(T))) (@typeInfo(T) == .Int or @typeInfo(T) == .Float) and std.math.isPowerOfTwo(@bitSizeOf(T)))
{ {
if (comptime std.simd.suggestVectorSize(T)) |block_len| { if (std.simd.suggestVectorSize(T)) |block_len| {
// For Intel Nehalem (2009) and AMD Bulldozer (2012) or later, unaligned loads on aligned data result // For Intel Nehalem (2009) and AMD Bulldozer (2012) or later, unaligned loads on aligned data result
// in the same execution as aligned loads. We ignore older arch's here and don't bother pre-aligning. // in the same execution as aligned loads. We ignore older arch's here and don't bother pre-aligning.
// //
// Use `comptime std.simd.suggestVectorSize(T)` to get the same alignment as used in this function // Use `std.simd.suggestVectorSize(T)` to get the same alignment as used in this function
// however this usually isn't necessary unless your arch has a performance penalty due to this. // however this usually isn't necessary unless your arch has a performance penalty due to this.
// //
// This may differ for other arch's. Arm for example costs a cycle when loading across a cache // This may differ for other arch's. Arm for example costs a cycle when loading across a cache

View File

@ -6,7 +6,7 @@
const std = @import("std"); const std = @import("std");
const builtin = @import("builtin"); const builtin = @import("builtin");
pub fn suggestVectorSizeForCpu(comptime T: type, comptime cpu: std.Target.Cpu) ?usize { pub fn suggestVectorSizeForCpu(comptime T: type, comptime cpu: std.Target.Cpu) ?comptime_int {
// This is guesswork, if you have better suggestions can add it or edit the current here // This is guesswork, if you have better suggestions can add it or edit the current here
// This can run in comptime only, but stage 1 fails at it, stage 2 can understand it // This can run in comptime only, but stage 1 fails at it, stage 2 can understand it
const element_bit_size = @max(8, std.math.ceilPowerOfTwo(u16, @bitSizeOf(T)) catch unreachable); const element_bit_size = @max(8, std.math.ceilPowerOfTwo(u16, @bitSizeOf(T)) catch unreachable);
@ -55,7 +55,7 @@ pub fn suggestVectorSizeForCpu(comptime T: type, comptime cpu: std.Target.Cpu) ?
/// Suggests a target-dependant vector size for a given type, or null if scalars are recommended. /// Suggests a target-dependant vector size for a given type, or null if scalars are recommended.
/// Not yet implemented for every CPU architecture. /// Not yet implemented for every CPU architecture.
pub fn suggestVectorSize(comptime T: type) ?usize { pub fn suggestVectorSize(comptime T: type) ?comptime_int {
return suggestVectorSizeForCpu(T, builtin.cpu); return suggestVectorSizeForCpu(T, builtin.cpu);
} }

View File

@ -200,7 +200,7 @@ pub fn utf8CountCodepoints(s: []const u8) !usize {
pub fn utf8ValidateSlice(input: []const u8) bool { pub fn utf8ValidateSlice(input: []const u8) bool {
var remaining = input; var remaining = input;
const V_len = comptime std.simd.suggestVectorSize(usize) orelse 1; const V_len = std.simd.suggestVectorSize(usize) orelse 1;
const V = @Vector(V_len, usize); const V = @Vector(V_len, usize);
const u8s_in_vector = @sizeOf(usize) * V_len; const u8s_in_vector = @sizeOf(usize) * V_len;