diff --git a/src/singeli/src/avx.singeli b/src/singeli/src/avx.singeli index 5879747d..e4401b98 100644 --- a/src/singeli/src/avx.singeli +++ b/src/singeli/src/avx.singeli @@ -19,31 +19,27 @@ def w256u = wgen256{{T} => isint{T} & isunsigned{T}} def w256f = wgen256{{T} => isfloat{T}} -def cast_vp{T, x & w256{T}} = emit{*T, '(void*)', x} -def cast_v{R, x:S & w256i{R } & w256{S}} = emit{R, '(__m256i)', x} -def cast_v{R, x:S & w256f{R,32} & w256{S}} = emit{R, '(__m256)', x} -def cast_v{R, x:S & w256f{R,64} & w256{S}} = emit{R, '(__m256d)', x} -def v2i{x:T & w256{T}} = cast_v{[32]u8, x} # for compact casting for the annoying intrinsic type system -def v2f{x:T & w256{T}} = cast_v{[8]f32, x} -def v2d{x:T & w256{T}} = cast_v{[4]f64, x} +def v2i{x:T & w256{T}} = [32]u8 ~~ x # for compact casting for the annoying intrinsic type system +def v2f{x:T & w256{T}} = [8]f32 ~~ x +def v2d{x:T & w256{T}} = [4]f64 ~~ x def to_el{E, T & isvec{T}} = [width{T}/width{E}]E -def to_el{E, x:T} = cast_v{to_el{E,T}, x} -def forv{T & w256{T}} = forc{{v}=>cast_vp{T,v}} +def to_el{E, x:T} = to_el{E,T} ~~ x +def forv{T & w256{T}} = forc{{v} => *T ~~ v} # load & store -def load {a:T, n & w256i{eltype{T} }} = emit{eltype{T}, '_mm256_loadu_si256', a+n} -def loada{a:T, n & w256i{eltype{T} }} = emit{eltype{T}, '_mm256_load_si256', a+n} -def load {a:T, n & w256f{eltype{T},64}} = emit{eltype{T}, '_mm256_loadu_pd', cast_p{f64, a+n}} -def loada{a:T, n & w256f{eltype{T},64}} = emit{eltype{T}, '_mm256_load_pd', cast_p{f64, a+n}} -def load {a:T, n & w256f{eltype{T},32}} = emit{eltype{T}, '_mm256_loadu_ps', cast_p{f32, a+n}} -def loada{a:T, n & w256f{eltype{T},32}} = emit{eltype{T}, '_mm256_load_ps', cast_p{f32, a+n}} -def store {a:T, n, v & w256i{eltype{T} }} = emit{void, '_mm256_storeu_si256', a+n, v} -def storea{a:T, n, v & w256i{eltype{T} }} = emit{void, '_mm256_store_si256', a+n, v} -def store {a:T, n, v & w256f{eltype{T},64}} = emit{void, '_mm256_storeu_pd', cast_p{f64, a+n}, v} -def storea{a:T, n, v & w256f{eltype{T},64}} = emit{void, '_mm256_store_pd', cast_p{f64, a+n}, v} -def store {a:T, n, v & w256f{eltype{T},32}} = emit{void, '_mm256_storeu_ps', cast_p{f32, a+n}, v} -def storea{a:T, n, v & w256f{eltype{T},32}} = emit{void, '_mm256_store_ps', cast_p{f32, a+n}, v} +def load {a:T, n & w256i{eltype{T} }} = emit{eltype{T}, '_mm256_loadu_si256', a+n} +def loada{a:T, n & w256i{eltype{T} }} = emit{eltype{T}, '_mm256_load_si256', a+n} +def load {a:T, n & w256f{eltype{T},64}} = emit{eltype{T}, '_mm256_loadu_pd', *f64 ~~ (a+n)} +def loada{a:T, n & w256f{eltype{T},64}} = emit{eltype{T}, '_mm256_load_pd', *f64 ~~ (a+n)} +def load {a:T, n & w256f{eltype{T},32}} = emit{eltype{T}, '_mm256_loadu_ps', *f32 ~~ (a+n)} +def loada{a:T, n & w256f{eltype{T},32}} = emit{eltype{T}, '_mm256_load_ps', *f32 ~~ (a+n)} +def store {a:T, n, v & w256i{eltype{T} }} = emit{void, '_mm256_storeu_si256', a+n, v} +def storea{a:T, n, v & w256i{eltype{T} }} = emit{void, '_mm256_store_si256', a+n, v} +def store {a:T, n, v & w256f{eltype{T},64}} = emit{void, '_mm256_storeu_pd', *f64 ~~ (a+n), v} +def storea{a:T, n, v & w256f{eltype{T},64}} = emit{void, '_mm256_store_pd', *f64 ~~ (a+n), v} +def store {a:T, n, v & w256f{eltype{T},32}} = emit{void, '_mm256_storeu_ps', *f32 ~~ (a+n), v} +def storea{a:T, n, v & w256f{eltype{T},32}} = emit{void, '_mm256_store_ps', *f32 ~~ (a+n), v} # broadcast def broadcast{T, v & w256i{T, 8}} = emit{T, '_mm256_set1_epi8', ext{eltype{T},v}} @@ -68,14 +64,14 @@ def iota{T & w256{T,32}} = make{T,0,1,2,3,4,5,6,7} def iota{T & w256{T,16}} = make{T,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15} def iota{T & w256{T,8}} = make{T,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31} -def __xor{a:T, b:T & w256i{T}} = cast_v{T, emit{[8]f32, '_mm256_xor_ps', v2f{a}, v2f{b}}} -def __and{a:T, b:T & w256i{T}} = cast_v{T, emit{[8]f32, '_mm256_and_ps', v2f{a}, v2f{b}}} -def __or {a:T, b:T & w256i{T}} = cast_v{T, emit{[8]f32, '_mm256_or_ps', v2f{a}, v2f{b}}} +def __xor{a:T, b:T & w256i{T}} = T ~~ emit{[8]f32, '_mm256_xor_ps', v2f{a}, v2f{b}} +def __and{a:T, b:T & w256i{T}} = T ~~ emit{[8]f32, '_mm256_and_ps', v2f{a}, v2f{b}} +def __or {a:T, b:T & w256i{T}} = T ~~ emit{[8]f32, '_mm256_or_ps', v2f{a}, v2f{b}} def __not{a:T & w256u{T}} = a ^ broadcast{T, ~cast{eltype{T},0}} # f64 comparison -def f64cmpAVX{a,b,n} = cast_v{[4]u64, emit{[4]f64, '_mm256_cmp_pd', a, b, n}} +def f64cmpAVX{a,b,n} = [4]u64 ~~ emit{[4]f64, '_mm256_cmp_pd', a, b, n} def __eq{a:T,b:T & T==[4]f64} = f64cmpAVX{a,b, 0} def __ne{a:T,b:T & T==[4]f64} = f64cmpAVX{a,b, 4} def __gt{a:T,b:T & T==[4]f64} = f64cmpAVX{a,b,30} @@ -126,12 +122,12 @@ def insert{x:T, i, v & w256i{T,32} & knum{i}} = emit{T, '_mm256_insert_epi32', x def insert{x:T, i, v & w256i{T,64} & knum{i}} = emit{T, '_mm256_insert_epi64', x, v, i} # blend by sign bit -def blend{f:T, t:T, m:M & w256{T} & w256i{M,32}} = cast_v{T, emit{[8]f32, '_mm256_blendv_ps', v2f{f}, v2f{t}, v2f{m}}} -def blend{f:T, t:T, m:M & w256{T} & w256i{M,64}} = cast_v{T, emit{[4]f64, '_mm256_blendv_pd', v2d{f}, v2d{t}, v2d{m}}} +def blend{f:T, t:T, m:M & w256{T} & w256i{M,32}} = T ~~ emit{[8]f32, '_mm256_blendv_ps', v2f{f}, v2f{t}, v2f{m}} +def blend{f:T, t:T, m:M & w256{T} & w256i{M,64}} = T ~~ emit{[4]f64, '_mm256_blendv_pd', v2d{f}, v2d{t}, v2d{m}} # mixed-width operations -def half{x:T, i & w256{T} & knum{i}} = cast_v{[vcount{T}/2](eltype{T}), emit{[8]i16, '_mm256_extracti128_si256', v2i{x}, i}} -def pair{a:T,b:T & width{T}==128} = cast_v{[vcount{T}*2](eltype{T}), emit{[8]i32, '_mm256_setr_m128i', a, b}} +def half{x:T, i & w256{T} & knum{i}} = [vcount{T}/2](eltype{T}) ~~ emit{[8]i16, '_mm256_extracti128_si256', v2i{x}, i} +def pair{a:T,b:T & width{T}==128} = [vcount{T}*2](eltype{T}) ~~ emit{[8]i32, '_mm256_setr_m128i', a, b} def pair{x} = pair{tupsel{0,x},tupsel{1,x}} # mask stuff diff --git a/src/singeli/src/avx2.singeli b/src/singeli/src/avx2.singeli index bfd749ea..799cef3c 100644 --- a/src/singeli/src/avx2.singeli +++ b/src/singeli/src/avx2.singeli @@ -1,6 +1,6 @@ # maskstore -def maskstore{a:T, m:M, n, v & w256{eltype{T}, 32} & w256i{M, 32}} = emit{void, '_mm256_maskstore_epi32', cast_p{i32, a+n}, m, v} -def maskstore{a:T, m:M, n, v & w256{eltype{T}, 64} & w256i{M, 64}} = emit{void, '_mm256_maskstore_pd', cast_p{f64, a+n}, m, v} +def maskstore{a:T, m:M, n, v & w256{eltype{T}, 32} & w256i{M, 32}} = emit{void, '_mm256_maskstore_epi32', *i32 ~~ (a+n), m, v} +def maskstore{a:T, m:M, n, v & w256{eltype{T}, 64} & w256i{M, 64}} = emit{void, '_mm256_maskstore_pd', *f64 ~~ (a+n), m, v} # maskstore with all cases defined, at the cost of not being a single instruction def maskstoreF{p, m, n, x:T} = store{p, n, blendF{load{p,n}, x, m}} @@ -108,26 +108,26 @@ def __subs{a:T,b:T & T==[32]u8 } = emit{T, '_mm256_subs_epu8', a, b} # structural operations -def shl{S==[16]u8, x:T, n & w256{T}} = cast_v{T, emit{T, '_mm256_bslli_epi128', x, n}} -def shr{S==[16]u8, x:T, n & w256{T}} = cast_v{T, emit{T, '_mm256_bsrli_epi128', x, n}} +def shl{S==[16]u8, x:T, n & w256{T}} = T ~~ emit{T, '_mm256_bslli_epi128', x, n} +def shr{S==[16]u8, x:T, n & w256{T}} = T ~~ emit{T, '_mm256_bsrli_epi128', x, n} -def blend{I==[8]u16, a:T, b:T, m & w256{T} & knum{m}} = cast_v{T, emit{[16]i16, '_mm256_blend_epi16', v2i{a}, v2i{b}, m}} -def blend{I==[8]u32, a:T, b:T, m & w256{T} & knum{m}} = cast_v{T, emit{[ 8]i32, '_mm256_blend_epi32', v2i{a}, v2i{b}, m}} -def blend{I==[4]u64, a:T, b:T, m & w256{T} & knum{m}} = cast_v{T, emit{[ 4]f64, '_mm256_blend_pd', v2d{a}, v2d{b}, m}} +def blend{I==[8]u16, a:T, b:T, m & w256{T} & knum{m}} = T ~~ emit{[16]i16, '_mm256_blend_epi16', v2i{a}, v2i{b}, m} +def blend{I==[8]u32, a:T, b:T, m & w256{T} & knum{m}} = T ~~ emit{[ 8]i32, '_mm256_blend_epi32', v2i{a}, v2i{b}, m} +def blend{I==[4]u64, a:T, b:T, m & w256{T} & knum{m}} = T ~~ emit{[ 4]f64, '_mm256_blend_pd', v2d{a}, v2d{b}, m} # blend by sign bit; still no 16-bit case -def blend{f:T, t:T, m:M & w256{T} & w256i{M, 8}} = cast_v{T, emit{[32]i8, '_mm256_blendv_epi8', v2i{f}, v2i{t}, v2i{m}}} +def blend{f:T, t:T, m:M & w256{T} & w256i{M, 8}} = T ~~ emit{[32]i8, '_mm256_blendv_epi8', v2i{f}, v2i{t}, v2i{m}} # assumes all bits are the same in each mask item def blendF{f:T, t:T, m:M & w256{T} & w256{M} & width{eltype{M}}!=16} = blend{f, t, m} -def blendF{f:T, t:T, m:M & w256{T} & w256{M,16}} = blend{f, t, cast_v{[32]i8,m}} +def blendF{f:T, t:T, m:M & w256{T} & w256{M,16}} = blend{f, t, [32]i8 ~~ m} -def shuf{I==[4]u32, x:T, n & w256{T} & knum{n}} = cast_v{T, emit{[8]i32, '_mm256_shuffle_epi32', v2i{x}, n}} -def shuf{I==[4]u64, x:T, n & w256{T} & knum{n}} = cast_v{T, emit{[4]f64, '_mm256_permute4x64_pd', v2d{x}, n}} -def shufHalves{x:T, y:T, n & w256{T} & knum{n}} = T~~emit{[4]i64, '_mm256_permute2x128_si256', v2i{x}, v2i{y}, n} +def shuf{I==[4]u32, x:T, n & w256{T} & knum{n}} = T ~~ emit{[8]i32, '_mm256_shuffle_epi32', v2i{x}, n} +def shuf{I==[4]u64, x:T, n & w256{T} & knum{n}} = T ~~ emit{[4]f64, '_mm256_permute4x64_pd', v2d{x}, n} +def shufHalves{x:T, y:T, n & w256{T} & knum{n}} = T ~~ emit{[4]i64, '_mm256_permute2x128_si256', v2i{x}, v2i{y}, n} -def sel{I, x:T, i:[8]i32 & w256{I,32}} = cast_v{T, emit{[32]u8, '_mm256_permutevar8x32_epi32', v2i{x}, i}} -def sel{I, x:T, i:[32]i8 & w128{I, 8}} = cast_v{T, emit{[32]u8, '_mm256_shuffle_epi8', v2i{x}, i}} +def sel{I, x:T, i:[8]i32 & w256{I,32}} = T ~~ emit{[32]u8, '_mm256_permutevar8x32_epi32', v2i{x}, i} +def sel{I, x:T, i:[32]i8 & w128{I, 8}} = T ~~ emit{[32]u8, '_mm256_shuffle_epi8', v2i{x}, i} def extract{x:T, i & w256i{T,8 } & knum{i}} = emit{eltype{T}, '_mm256_extract_epi8', x, i} def extract{x:T, i & w256i{T,16} & knum{i}} = emit{eltype{T}, '_mm256_extract_epi16', x, i} @@ -138,9 +138,9 @@ def getmask{x:T & w256{T, 16}} = { msk:u32 = getmask{emit{[32]u8, '_mm256_packs_epi16', x, broadcast{[16]u16, 0}}} (msk&255) | (msk>>8) } -def any{x:T & w256i{T, 16}} = getmask{cast_v{[32]u8,x}}!=0 -def all{x:T & w256i{T, 16}} = getmask{cast_v{[32]u8,x}} == (1<<32)-1 -def anyneg{x:T & w256s{T, 16}} = getmask{cast_v{[32]u8, cast_v{[16]i16,x} < broadcast{[16]i16, 0}}}!=0 +def any{x:T & w256i{T, 16}} = getmask{[32]u8~~x}!=0 +def all{x:T & w256i{T, 16}} = getmask{[32]u8~~x} == (1<<32)-1 +def anyneg{x:T & w256s{T, 16}} = getmask{[32]u8 ~~ ([16]i16~~x < broadcast{[16]i16, 0})}!=0 # conversion # convert packed elements of type F to a result T diff --git a/src/singeli/src/base.singeli b/src/singeli/src/base.singeli index 14ade689..d1840013 100644 --- a/src/singeli/src/base.singeli +++ b/src/singeli/src/base.singeli @@ -3,6 +3,7 @@ include 'arch/c' def Size = u64 oper infix right ~~ reinterpret 55 +def load{x} = *x # TODO move these to a more base file def knum{x} = match{'number',kind{x}} def trunc{T, x:U & isint{T} & isint{U} & T<=U} = emit{T, '', x} diff --git a/src/singeli/src/bitops.singeli b/src/singeli/src/bitops.singeli index 7d780208..9f63f55f 100644 --- a/src/singeli/src/bitops.singeli +++ b/src/singeli/src/bitops.singeli @@ -24,7 +24,7 @@ def b_setBatch{sz, x:*u64, n:(Size), v} = { vc:u64 = ext{u64,v} } def b_setBatch{sz, x:*u64, n:(Size), v & sz==4} = { - x8:= cast_p{u8, x} + x8:= *u8 ~~ x #w:u64 = cast_i{u64, load{x8,n/2}} #sh:u64 = (n&1) * 4 @@ -42,7 +42,7 @@ def b_setBatch{sz, x:*u64, n:(Size), v & sz==4} = { store{x8, n/2, cast_i{u8,w}} } -def b_setBatch{sz, x:*u64, n:(Size), v & sz== 8} = store{cast_p{u8, x}, n, cast_i{u8, v}} -def b_setBatch{sz, x:*u64, n:(Size), v & sz==16} = store{cast_p{u16, x}, n, cast_i{u16,v}} -def b_setBatch{sz, x:*u64, n:(Size), v & sz==32} = store{cast_p{u32, x}, n, cast_i{u32,v}} -def b_setBatch{sz, x:*u64, n:(Size), v & sz==64} = store{x, n, cast_i{u64,v}} +def b_setBatch{sz, x:*u64, n:(Size), v & sz== 8} = store{*u8 ~~ x, n, cast_i{u8, v}} +def b_setBatch{sz, x:*u64, n:(Size), v & sz==16} = store{*u16 ~~ x, n, cast_i{u16,v}} +def b_setBatch{sz, x:*u64, n:(Size), v & sz==32} = store{*u32 ~~ x, n, cast_i{u32,v}} +def b_setBatch{sz, x:*u64, n:(Size), v & sz==64} = store{ x, n, cast_i{u64,v}} diff --git a/src/singeli/src/cmp.singeli b/src/singeli/src/cmp.singeli index b4f2cdf7..8bb7fcdb 100644 --- a/src/singeli/src/cmp.singeli +++ b/src/singeli/src/cmp.singeli @@ -83,20 +83,20 @@ def any2bit{VT, unr, op, wS, wV, xS, xV, dst:*u64, len:(Size)} = { } } aa2bit{VT, unr, op}(dst:*u64, wr:*u8, xr:*u8, len:Size) : void = { - wv:= cast_vp{VT, wr}; ws:= cast_p{eltype{VT}, wr} - xv:= cast_vp{VT, xr}; xs:= cast_p{eltype{VT}, xr} + wv:= *VT ~~ wr; ws:= *eltype{VT} ~~ wr + xv:= *VT ~~ xr; xs:= *eltype{VT} ~~ xr any2bit{VT, unr, op, {i}=>load{ws,i}, {i}=>load{wv,i}, {i}=>load{xs,i}, {i}=>load{xv,i}, dst, len} } as2bit{VT, unr, op}(dst:*u64, wr:*u8, x:u64, len:Size) : void = { # show{VT,unr,fmt{op}} - wv:= cast_vp{VT, wr}; ws:= cast_p{eltype{VT}, wr} + wv:= *VT~~wr; ws:= *eltype{VT}~~wr xv:= broadcast{VT, pathAS{dst, len, eltype{VT}, op, x}} any2bit{VT, unr, op, {i}=>load{ws,i}, {i}=>load{wv,i}, {i}=>x, {i}=>xv, dst, len} } bitAA{bitop}(dst:*u64, wr:*u8, xr:*u8, len:Size) : void = { - ws:= cast_p{u64, wr} - xs:= cast_p{u64, xr} + ws:= *u64~~wr + xs:= *u64~~xr @for (dst,ws,xs over _ from 0 to cdiv{len,64}) dst = bitop{ws,xs} } @@ -115,8 +115,8 @@ bitAS{op}(dst:*u64, wr:*u8, x:u64, len:Size) : void = { # show{'bitAS'} fillbits{dst, len, r0} return{} } - if (r0) call{not, dst, cast_p{u64,wr}, len} - else call{cpy, dst, cast_p{u64,wr}, len} + if (r0) call{not, dst, *u64~~wr, len} + else call{cpy, dst, *u64~~wr, len} } diff --git a/src/singeli/src/dyarith.singeli b/src/singeli/src/dyarith.singeli index 374aad78..8cee68b9 100644 --- a/src/singeli/src/dyarith.singeli +++ b/src/singeli/src/dyarith.singeli @@ -41,10 +41,10 @@ def arithChk2{F, M, w:T, x:T, i & match{F,__mul} & isvec{T} & i16==eltype{T}} = tup{rl, anyne{rh, rl>>15, M}} } def arithChk2{F, M, w:T, x:T, i & match{F,__mul} & isvec{T} & i8==eltype{T}} = { - def wp = unpackQ{w, cast_v{T,broadcast{T,0}>w}} - def xp = unpackQ{x, cast_v{T,broadcast{T,0}>x}} + def wp = unpackQ{w, T ~~ (broadcast{T,0}>w)} + def xp = unpackQ{x, T ~~ (broadcast{T,0}>x)} def rp = each{__mul, wp, xp} - def bad = each{{v}=>cast_v{[16]i16,(v<<8)>>8 != v}, rp} + def bad = each{{v} => [16]i16 ~~ ((v<<8)>>8 != v), rp} if (M{0}) { # masked check tup{packQ{rp}, any{M{packQ{bad}}}} } else { # unmasked check; can do check in a simpler way @@ -52,10 +52,10 @@ def arithChk2{F, M, w:T, x:T, i & match{F,__mul} & isvec{T} & i8==eltype{T}} = { } } def arithChk2{F, M, w:T, x:T, i & match{F,__mul} & isvec{T} & i32==eltype{T}} = { - max:= cast_v{[8]f32, broadcast{[8]u32, 0x4efffffe}} + max:= [8]f32 ~~ broadcast{[8]u32, 0x4efffffe} def cf32{x} = emit{[8]f32, '_mm256_cvtepi32_ps', x} f32mul:= cf32{w} * cf32{x} - tup{w*x, any{M{cast_v{[8]u32, emit{[8]f32, '_mm256_cmp_ps', abs{f32mul}, max, 29}}}}} + tup{w*x, any{M{[8]u32 ~~ emit{[8]f32, '_mm256_cmp_ps', abs{f32mul}, max, 29}}}} # TODO fallback to the below if the above fails # TODO don't do this, but instead shuffle one half, do math, unshuffle that half # def wp = unpackQ{w, broadcast{T, 0}} @@ -63,7 +63,7 @@ def arithChk2{F, M, w:T, x:T, i & match{F,__mul} & isvec{T} & i32==eltype{T}} = # def rp = each{__mul32, wp, xp} # def T2 = ty_dbl{T} # def bad = each{{v}=>{ - # ((cast_v{T2,v} + broadcast{T2,0x80000000}) ^ broadcast{T2, cast{i64,1}<<63}) > broadcast{T2, cast_i{i64, (cast{u64,1}<<63) | 0xFFFFFFFF}} + # (((T2~~v) + broadcast{T2,0x80000000}) ^ broadcast{T2, cast{i64,1}<<63}) > broadcast{T2, cast_i{i64, (cast{u64,1}<<63) | 0xFFFFFFFF}} # }, rp} # tup{packQQ{each{{v} => v&broadcast{T2, 0xFFFFFFFF}, rp}}, any{tupsel{0,bad}|tupsel{1,bad}}} TODO use M } @@ -88,15 +88,15 @@ def arithAny{VT, F, W, X, r, len} = { } def arithAA{VT, F, w, x, r, len} = { - arithAny{VT, F, {i}=>load{cast_p{VT,w}, i}, {i}=>load{cast_p{VT,x}, i}, r, len} + arithAny{VT, F, {i}=>load{*VT~~w, i}, {i}=>load{*VT~~x, i}, r, len} } def arithAS{VT, F, w, x, r, len} = { xv:= broadcast{VT, x} - arithAny{VT, F, {i}=>load{cast_p{VT,w}, i}, {i}=>xv, r, len} + arithAny{VT, F, {i}=>load{*VT~~w, i}, {i}=>xv, r, len} } def arithSA{VT, F, w, x, r, len} = { wv:= broadcast{VT, w} - arithAny{VT, F, {i}=>wv, {i}=>load{cast_p{VT,x}, i}, r, len} + arithAny{VT, F, {i}=>wv, {i}=>load{*VT~~x, i}, r, len} } @@ -109,9 +109,9 @@ def cast_fB{T, x:(u64) & issigned{T} & T { - cw:= load{cast_p{[32]u8, w}, i} - cx:= load{cast_p{[32]u8, x}, i} + cw:= load{*[32]u8 ~~ w, i} + cx:= load{*[32]u8 ~~ x, i} if (anyneBit{cw,cx,M}) return{0} }} } else if (X==f64) { # bitarr ≡ f64arr @@ -28,21 +28,21 @@ equal{W, X}(w:*u8, x:*u8, l:u64, d:u64) : u1 = { f0:= broadcast{T, 0.0} f1:= broadcast{T, 1.0} maskedLoop{bulk, l, {i, M} => { - cw:= load{cast_p{u8,w}, i>>1} >> cast_i{u8, 4*(i&1)} - cx:= load{cast_p{T, x}, i} + cw:= load{*u8 ~~ w, i>>1} >> cast_i{u8, 4*(i&1)} + cx:= load{*T ~~ x, i} wu:= blend{f0, f1, broadcast{[4]u64, cw} << make{[4]u64,63,62,61,60}} if (anyne{wu, cx, M}) return{0} }} } else { # bitarr ≡ i8/i16/i32arr def T = [256/width{X}]X def sh{c} = c << (width{X}-1) - def sh{c & X==u8} = cast_v{T, to_el{u16,c}<<7} + def sh{c & X==u8} = T ~~ (to_el{u16,c}<<7) # TODO compare with doing the comparison in vector registers badBits:= broadcast{T, ~cast{X,1}} maskedLoop{bulk, l, {i, M} => { - cw:= load{cast_p{ty_u{bulk}, w}, i} - cx:= load{cast_p{T,x}, i} + cw:= load{*ty_u{bulk} ~~ w, i} + cx:= load{*T ~~ x, i} if (~andIsZero{M{cx}, badBits}) return{0} if (anyne{promote{u64,getmask{sh{cx}}}, promote{u64,cw}, M}) return{0} }} @@ -54,8 +54,8 @@ equal{W, X}(w:*u8, x:*u8, l:u64, d:u64) : u1 = { def fac = width{X}/width{W} maskedLoop{bulk, l, {i, M} => { - cw:= load{cast_p{ww{tern{fac==1, 256, 128}, u8}, w + i*32/fac}, 0} - cx:= load{cast_p{ww{256, X}, x}, i} + cw:= load{*ww{tern{fac==1, 256, 128}, u8} ~~ (w + i*32/fac)} + cx:= load{*ww{256, X} ~~ x, i} cwc:= cvt{W, ww{256, X}, cw} if (anyne{cwc,cx,M}) return{0} }} diff --git a/src/singeli/src/mask.singeli b/src/singeli/src/mask.singeli index 6d58b1ce..92f816b6 100644 --- a/src/singeli/src/mask.singeli +++ b/src/singeli/src/mask.singeli @@ -16,17 +16,17 @@ mask256_64:*i64 = maskInit{256,i64}; mask128_64:*i64 = maskInit{128,i64} mask256_1 :*u8 = maskInit1{256}; mask128_1 :*u8 = maskInit1{128} -def maskOf{T,n & w256{T, 8}} = load{cast_p{[32]u8, mask256_8 + (n^31)}, 0} -def maskOf{T,n & w256{T,16}} = load{cast_p{[16]u16, mask256_16 + (n^15)}, 0} -def maskOf{T,n & w256{T,32}} = load{cast_p{[ 8]u32, mask256_32 + (n^7)}, 0} -def maskOf{T,n & w256{T,64}} = load{cast_p{[ 4]u64, mask256_64 + (n^3)}, 0} -def maskOfBit{T,n & w256{T}} = load{cast_p{[32]u8, mask256_1 + (n>>3)^31 + 64*(n&7)}, 0} +def maskOf{T,n & w256{T, 8}} = load{*[32]u8 ~~ (mask256_8 + (n^31))} +def maskOf{T,n & w256{T,16}} = load{*[16]u16 ~~ (mask256_16 + (n^15))} +def maskOf{T,n & w256{T,32}} = load{*[ 8]u32 ~~ (mask256_32 + (n^7))} +def maskOf{T,n & w256{T,64}} = load{*[ 4]u64 ~~ (mask256_64 + (n^3))} +def maskOfBit{T,n & w256{T}} = load{*[32]u8 ~~ (mask256_1 + (n>>3)^31 + 64*(n&7))} -def maskOf{T,n & w128{T, 8}} = load{cast_p{[16]u8, mask128_8 + (n^15)}, 0} -def maskOf{T,n & w128{T,16}} = load{cast_p{[ 8]u16, mask128_16 + (n^7)}, 0} -def maskOf{T,n & w128{T,32}} = load{cast_p{[ 4]u32, mask128_32 + (n^3)}, 0} -def maskOf{T,n & w128{T,64}} = load{cast_p{[ 2]u64, mask128_64 + (n^1)}, 0} -def maskOfBit{T,n & w128{T}} = load{cast_p{[16]u8, mask128_1 + (n>>3)^15 + 32*(n&7)}, 0} +def maskOf{T,n & w128{T, 8}} = load{*[16]u8 ~~ (mask128_8 + (n^15))} +def maskOf{T,n & w128{T,16}} = load{*[ 8]u16 ~~ (mask128_16 + (n^7))} +def maskOf{T,n & w128{T,32}} = load{*[ 4]u32 ~~ (mask128_32 + (n^3))} +def maskOf{T,n & w128{T,64}} = load{*[ 2]u64 ~~ (mask128_64 + (n^1))} +def maskOfBit{T,n & w128{T}} = load{*[16]u8 ~~ (mask128_1 + (n>>3)^15 + 32*(n&7))} def anyne{x:T, y:T, M & M{0}==0 & isvec{T}} = ~all{x==y} def anyne{x:T, y:T, M & M{0}==1 & isvec{T}} = any{M{x!=y}} @@ -37,9 +37,9 @@ def anyneBit{x:T, y:T, M} = ~M{x^y, 'all bits zeroes'} def maskNone{x} = x def maskNone{x, mode=='all bits zeroes'} = andIsZero{x, x} def maskAfter{n} = { - def mask{x:X & isvec{X}} = x&cast_v{X,maskOf{X,n}} + def mask{x:X & isvec{X}} = x & (X~~maskOf{X,n}) def mask{x:X & anyInt{x}} = x & ((1<>k<>k}<=width{B}} = w/width{A} # b:B - pointer to data to index; if width{B} { cw0:= loadBatch{w, i, TIF} diff --git a/src/singeli/src/slash.singeli b/src/singeli/src/slash.singeli index 5b09a7c8..5de2d2a4 100644 --- a/src/singeli/src/slash.singeli +++ b/src/singeli/src/slash.singeli @@ -32,7 +32,7 @@ def comp16{w:*u64, X, r:*i16, l:u64} = { slash2{F, T}(w:*u64, x:*T, r:*T, l:u64) : void = { xv:= reinterpret{*u64, x} - F{w, {} => {c:=load{xv,0}; xv+=1; c}, r, l} + F{w, {} => {c:= *xv; xv+=1; c}, r, l} } slash1{F, T, iota, add}(w:*u64, r:*T, l:u64) : void = { diff --git a/src/singeli/src/sse3.singeli b/src/singeli/src/sse3.singeli index 09cdb62f..2f207252 100644 --- a/src/singeli/src/sse3.singeli +++ b/src/singeli/src/sse3.singeli @@ -16,27 +16,23 @@ def w128s = wgen128{{T} => isint{T} & issigned{T}} def w128u = wgen128{{T} => isint{T} & isunsigned{T}} def w128f = wgen128{{T} => isfloat{T}} -def cast_v{T, x:S & match{show{T},show{type{x}}}} = x -def cast_v{R, x:S & w128i{R } & w128{S}} = emit{R, '(__m128i)', x} -def cast_v{R, x:S & w128f{R,32} & w128{S}} = emit{R, '(__m128)', x} -def cast_v{R, x:S & w128f{R,64} & w128{S}} = emit{R, '(__m128d)', x} -def v2i{x:T & w128{T}} = cast_v{[16]u8, x} # for compact casting for the annoying intrinsic type system -def v2f{x:T & w128{T}} = cast_v{[4]f32, x} -def v2d{x:T & w128{T}} = cast_v{[2]f64, x} +def v2i{x:T & w128{T}} = [16]u8 ~~ x # for compact casting for the annoying intrinsic type system +def v2f{x:T & w128{T}} = [4]f32 ~~ x +def v2d{x:T & w128{T}} = [2]f64 ~~ x # load & store -def load {a:T, n & w128i{eltype{T} }} = emit{eltype{T}, '_mm_loadu_si128', a+n} -def loada{a:T, n & w128i{eltype{T} }} = emit{eltype{T}, '_mm_load_si128', a+n} -def load {a:T, n & w128f{eltype{T},64}} = emit{eltype{T}, '_mm_loadu_pd', cast_p{f64, a+n}} -def loada{a:T, n & w128f{eltype{T},64}} = emit{eltype{T}, '_mm_load_pd', cast_p{f64, a+n}} -def load {a:T, n & w128f{eltype{T},32}} = emit{eltype{T}, '_mm_loadu_ps', cast_p{f32, a+n}} -def loada{a:T, n & w128f{eltype{T},32}} = emit{eltype{T}, '_mm_load_ps', cast_p{f32, a+n}} -def store {a:T, n, v & w128i{eltype{T} }} = emit{void, '_mm_storeu_si128', a+n, v} -def storea{a:T, n, v & w128i{eltype{T} }} = emit{void, '_mm_store_si128', a+n, v} -def store {a:T, n, v & w128f{eltype{T},64}} = emit{void, '_mm_storeu_pd', cast_p{f64, a+n}, v} -def storea{a:T, n, v & w128f{eltype{T},64}} = emit{void, '_mm_store_pd', cast_p{f64, a+n}, v} -def store {a:T, n, v & w128f{eltype{T},32}} = emit{void, '_mm_storeu_ps', cast_p{f32, a+n}, v} -def storea{a:T, n, v & w128f{eltype{T},32}} = emit{void, '_mm_store_ps', cast_p{f32, a+n}, v} +def load {a:T, n & w128i{eltype{T} }} = emit{eltype{T}, '_mm_loadu_si128', a+n} +def loada{a:T, n & w128i{eltype{T} }} = emit{eltype{T}, '_mm_load_si128', a+n} +def load {a:T, n & w128f{eltype{T},64}} = emit{eltype{T}, '_mm_loadu_pd', *f64 ~~ (a+n)} +def loada{a:T, n & w128f{eltype{T},64}} = emit{eltype{T}, '_mm_load_pd', *f64 ~~ (a+n)} +def load {a:T, n & w128f{eltype{T},32}} = emit{eltype{T}, '_mm_loadu_ps', *f32 ~~ (a+n)} +def loada{a:T, n & w128f{eltype{T},32}} = emit{eltype{T}, '_mm_load_ps', *f32 ~~ (a+n)} +def store {a:T, n, v & w128i{eltype{T} }} = emit{void, '_mm_storeu_si128', a+n, v} +def storea{a:T, n, v & w128i{eltype{T} }} = emit{void, '_mm_store_si128', a+n, v} +def store {a:T, n, v & w128f{eltype{T},64}} = emit{void, '_mm_storeu_pd', *f64 ~~ (a+n), v} +def storea{a:T, n, v & w128f{eltype{T},64}} = emit{void, '_mm_store_pd', *f64 ~~ (a+n), v} +def store {a:T, n, v & w128f{eltype{T},32}} = emit{void, '_mm_storeu_ps', *f32 ~~ (a+n), v} +def storea{a:T, n, v & w128f{eltype{T},32}} = emit{void, '_mm_store_ps', *f32 ~~ (a+n), v} # broadcast def broadcast{T, v & w128i{T, 8}} = emit{T, '_mm_set1_epi8', ext{eltype{T},v}} @@ -82,9 +78,9 @@ def getmask{x:T & w128{T, 64}} = emit{u8, '_mm_movemask_pd', v2d{x}} def any{x:T & w128i{T}} = getmask{x} != 0 # assumes elements of x all have equal bits (avx2 utilizes this for 16 bits) def all{x:T & w128i{T}} = getmask{x} == (1<