diff --git a/build/singeliSubmodule b/build/singeliSubmodule index 528faaf9..5f9cbd46 160000 --- a/build/singeliSubmodule +++ b/build/singeliSubmodule @@ -1 +1 @@ -Subproject commit 528faaf9e2a7f4f3434365bcd91d6c18c87c4f08 +Subproject commit 5f9cbd46c265491ff167a5d9377d1462539dbdd8 diff --git a/src/singeli/src/avx512.singeli b/src/singeli/src/avx512.singeli new file mode 100644 index 00000000..d73ab10f --- /dev/null +++ b/src/singeli/src/avx512.singeli @@ -0,0 +1,11 @@ +local def re_mask{M, sub} = { + def l = vcount{M}; def w = max{32,l} + sub{fmtnat{l}, fmtnat{w}, ty_u{w}} +} +local def ismask{M} = if (isvec{M}) u1==eltype{M} else 0 +def reinterpret{M, a:T & ismask{M} & width{T}==width{M}} = { + re_mask{M, {l,w,W} => emit{M, merge{'_cvtu',w,'_mask',l}, promote{W, a}}} +} +def reinterpret{T, a:M & ismask{M} & width{T}==width{M}} = { + re_mask{M, {l,w,W} => cast_i{T, emit{W, merge{'_cvtmask',l,'_u',w}, a}}} +} diff --git a/src/singeli/src/scan.singeli b/src/singeli/src/scan.singeli index 92e16230..ace6b2d3 100644 --- a/src/singeli/src/scan.singeli +++ b/src/singeli/src/scan.singeli @@ -1,5 +1,8 @@ include './base' -include './clmul' +if (hasarch{'X86_64'}) { + if (hasarch{'PCLMUL'}) include './clmul' + if (hasarch{'AVX512BW', 'VPCLMULQDQ', 'GFNI'}) include './avx512' +} include './mask' include './f64' include './scan_common' @@ -77,7 +80,7 @@ fn scan_neq{}(p:u64, x:*u64, r:*u64, nw:u64) : void = { p = -(r>>63) # repeat sign bit } } -fn clmul_scan_ne_any{..._ & hasarch{'PCLMUL'}}(x:*void, r:*void, init:u64, words:u64, mark:u64) : void = { +fn clmul_scan_ne_any{& hasarch{'PCLMUL'}}(x:*void, r:*void, init:u64, words:u64, mark:u64) : void = { def V = [2]u64 m := V**mark def xor64{a, i, carry} = { # carry is 64-bit broadcasted current total @@ -98,9 +101,34 @@ fn clmul_scan_ne_any{..._ & hasarch{'PCLMUL'}}(x:*void, r:*void, init:u64, words storeLow{rv+e, 64, clmul{loadLow{xv+e, 64}, m, 0} ^ c} } } -fn scan_neq{..._ & hasarch{'PCLMUL'}}(init:u64, x:*u64, r:*u64, nw:u64) : void = { +fn scan_neq{& hasarch{'PCLMUL'}}(init:u64, x:*u64, r:*u64, nw:u64) : void = { clmul_scan_ne_any{}(*void~~x, *void~~r, init, nw, -(u64~~1)) } +fn scan_neq{& hasarch{'AVX512BW', 'VPCLMULQDQ', 'GFNI'}}(init:u64, x:*u64, r:*u64, nw:u64) : void = { + def emitM{T,s,...a} = emit{T, merge{'_mm512_',s}, ...a} + def V = [8]u64 + def sse{a} = make{[2]u64, a, 0} + carry := sse{init} + # xor-scan on bytes + xmat := V**base{256, 1<<(8-iota{8}) - 1} + def xor8 = emitM{V, 'gf2p8affine_epi64_epi8', ., xmat, 0} + # Exclusive xor-scan on one word + def exor64 = clmul{., sse{1<<64 - 2}, 0} + @for (xv in *V~~x, rv in *V~~r over i to cdiv{nw,vcount{V}}) { + x8 := xor8{xv} + hb := sse{u64~~emitM{[64]u1, 'movepi8_mask', x8}} + xh := exor64{hb} # Exclusive xor of high bits + xc := xh ^ carry + v := x8 ^ emitM{V, 'movm_epi8', [64]u1~~extract{xc,0}} + carry = (xc ^ hb) ^ shuf{[4]u32, xh, 4b3232} + rem:= nw - 8*i + if (rem < 8) { + emitM{void, 'mask_storeu_epi64', *V~~r+i, [8]u1~~(~(u8~~0xff<