diff --git a/src/singeli/src/slash.singeli b/src/singeli/src/slash.singeli index 07fbdace..27f00f37 100644 --- a/src/singeli/src/slash.singeli +++ b/src/singeli/src/slash.singeli @@ -145,18 +145,10 @@ fn slash2{T==i8 & hasarch{'AVX2'}}(w:*u64, x:*T, r:*T, l:u64, sum:u64) : void = } } -# 8-bit writes ~8 bytes of garbage past end, 16-bit writes ~16 bytes -if (hasarch{'BMI2'}) { -export{'bmipopc_2slash8', slash2{i8}} -export{'bmipopc_2slash16', slash2{i16}} -export{'bmipopc_1slash8', slash1{i8}} -export{'bmipopc_1slash16', slash1{i16}} -} - itab :*u64 = fold{{t,k} => join{each{tup,t,k+(t<<8)%(1<<64)}}, tup{0x8080808080808080}, reverse{iota{8}}} i64tab:*u32 = fold{{t,k} => join{each{tup,t,k+(t<<8)%(1<<32)}}, tup{0x80808080}, reverse{2*iota{4}}} -fn avx2_compress{T & width{T}>=32}(wp:*u64, x:*T, r:*T, l:u64, sum:u64) : void = { +fn slash2{T & hasarch{'AVX2'} & width{T}>=32}(wp:*u64, x:*T, r:*T, l:u64, sum:u64) : void = { def tw = width{T} def V = [8]u32 expander := make{[32]u8, merge{...each{{i}=>tup{i, ... 3**128}, iota{8}>>lb{tw/32}}}} @@ -179,9 +171,39 @@ fn avx2_compress{T & width{T}>=32}(wp:*u64, x:*T, r:*T, l:u64, sum:u64) : void = } } +fn slash2{T & hasarch{if (width{T}>=32) 'AVX512F' else 'AVX512VBMI2'}}(w:*u64, x:*T, r:*T, l:u64, sum:u64) : void = { + def f = match { {_==8}=>'8'; {_==16}=>'16'; {_==32}=>'32'; {_==64}=>'64' } + def wt = width{T} + def vl = 512/wt + def V = [vl]T + def wu = max{32,vl} + def load {a:T, n & 512==width{eltype{T}}} = emit{eltype{T}, '_mm512_loadu_si512', a+n} + @for (w in *(ty_u{vl})~~w, x in *V~~x over cdiv{l,vl}) { + def I = ty_u{wu} + def emitT{O, name, ...a} = emit{O, merge{'_mm512_',name,'_epi',f{wt}}, ...a} + def to_mask{a} = emit{[vl]u1, merge{'_cvtu',f{wu},'_mask',f{vl}}, a} + m := to_mask{promote{I,w}} + c := popc{w} + # The compress-store instruction performs very poorly on Zen4, + # and is also a lot worse than the following on Tiger Lake + # emitT{void, 'mask_compressstoreu', r, m, x} + cs := cast_i{I,promote{i64,1}<>6 + v := emitT{V, 'mask_compress', x, m, x} + emitT{void, 'mask_storeu', r, to_mask{cs}, v} + r += c + } +} + +if (hasarch{'BMI2'}) { +export{'bmipopc_2slash8', slash2{i8}} +export{'bmipopc_2slash16', slash2{i16}} +export{'bmipopc_1slash8', slash1{i8}} +export{'bmipopc_1slash16', slash1{i16}} +} if (hasarch{'AVX2'}) { -export{'avx2_2slash32', avx2_compress{i32}} -export{'avx2_2slash64', avx2_compress{i64}} +export{'avx2_2slash32', slash2{i32}} +export{'avx2_2slash64', slash2{i64}} } # pext, or boolean compress