tensors: optimize matrix operations.

tensors: Add benchmarking file

tensors: Add addition and multiplication tests for benchmarking

tensors: inlined slicing to improve metrics.

tensors: fix help-lint warnings.

tensors: restore newer matmul

tensors: add fixnum declaration.

tensors: away with you, unsafe!

tensors: transpose added to benchmarks

tensors: optimize matmul to be within an order of magnitude of np.

tensors: remove type declaration.

tensors: optimize matmul.

extra/tensors/benchmark/benchmark.factor

@@ -0,0 +1,50 @@
+! Copyright (C) 2019 HMC Clinic.
+! See http://factorcode.org/license.txt for BSD license.
+USING: arrays io kernel locals math prettyprint tensors tools.time ;
+IN: tensors.benchmark
+
+<PRIVATE
+
+:: add-tensors ( trials elems -- time )
+    ! Create the arrays to be added
+    elems 1array naturals dup
+    ! Benchmark!
+    [ trials [ 2dup t+ drop ] times ] benchmark
+    ! Normalize
+    trials / >float
+    nip nip ;
+
+:: matmul-tensors ( trials elems -- time )
+    ! Create the arrays to be multiplied
+    elems elems 2array naturals dup
+    ! Benchmark!
+    [ trials [ 2dup matmul drop ] times ] benchmark
+    ! Normalize
+    trials / >float
+    nip nip ;
+
+:: transpose-tensor ( trials elems -- time )
+    ! Create the array to be transposed
+    elems elems 2array naturals
+    ! benchmark
+    [ trials [ dup transpose drop ] times ] benchmark
+    ! Normalize
+    trials / >float
+    nip ;
+
+PRIVATE>
+
+: run-benchmarks ( -- )
+    "Benchmarking the tensors vocabulary" print
+    "Add two 100 element tensors" print
+    1000000 100 add-tensors .
+    "Add two 100,000 element tensors" print
+    10000 100000 add-tensors .
+    "Multiply two 10x10 matrices" print
+    100000 10 matmul-tensors .
+    "Multiply two 100x100 matrices" print
+    1000 100 matmul-tensors .
+    "Transpose a 10x10 matrix" print
+    10000 10 transpose-tensor .
+    "Transpose a 100x100 matrix" print
+    10 100 transpose-tensor . ;

extra/tensors/tensors.factor

@@ -1,9 +1,12 @@
 ! Copyright (C) 2019 HMC Clinic.
 ! See http://factorcode.org/license.txt for BSD license.
+
 USING: accessors alien.c-types alien.data arrays
 concurrency.combinators grouping kernel locals math.functions
 math.ranges math.statistics math multi-methods quotations sequences
-sequences.private specialized-arrays tensors.tensor-slice typed ;
+sequences.extras sequences.private specialized-arrays
+tensors.tensor-slice typed ;
+
 QUALIFIED-WITH: alien.c-types c
 SPECIALIZED-ARRAY: c:float
 IN: tensors
@@ -47,7 +50,7 @@ PRIVATE>
 ! Construct a one-dimensional tensor with values start, start+step,
 ! ..., stop (inclusive)
 : arange ( a b step -- tensor )
-    <range> [ length 1array ] keep >float-array <tensor> ;
+    <range> [ length >fixnum 1array ] keep >float-array <tensor> ;
 
 ! Construct a tensors with vec { 0 1 2 ... } and reshape to the desired shape
 : naturals ( shape -- tensor )
@@ -148,20 +151,24 @@ METHOD: t% { number tensor } swap [ swap mod ] curry t-uop ;
 
 ! Perform matrix multiplication muliplying an
 ! mxn matrix with a nxp matrix
-TYPED:: 2d-matmul ( vec1: slice vec2: slice res: slice n: number p: number -- )
+TYPED:: 2d-matmul ( vec1: float-array start1: fixnum
+                      vec2: float-array start2: fixnum
+                      res: float-array start3: fixnum
+                      m: fixnum n: fixnum p: fixnum -- )
     ! For each element in the range, we want to compute the dot product of the
     ! corresponding row and column
-    res
+    m [ :> i
-    [   >fixnum
+        p [ :> j
-        ! Get the row
+            0.0 ! This is the sum
-        [ [ vec1 n ] dip p row ]
+            n [ :> k
-        ! Get the column
+                ! Add to the sum
-        ! [ p mod vec2 swap p every ] bi
+                i n * k + start1 + vec1 nth-unsafe
-        [ p mod f p vec2 <step-slice> ] bi
+                k p * j + start2 + vec2 nth-unsafe
-        ! Take the dot product
+                * +
-        [ * ] [ + ] 2map-reduce
+            ] each-integer
-    ]
+            i p * j + start3 + res set-nth-unsafe
-    map! drop ;
+        ] each-integer
+    ] each-integer ;
 
 PRIVATE>
 
@@ -176,70 +183,52 @@ TYPED:: matmul ( tensor1: tensor tensor2: tensor -- tensor3: tensor )
     tensor1 shape>> unclip-last-slice :> n
     unclip-last-slice :> m :> top-shape
     tensor2 shape>> last :> p
-    top-shape product :> rest
+    top-shape product :> top-prod
 
-    ! Now create the new tensor with { 0 ... m*p-1 } repeating
+    ! Create the shape of the resulting tensor
-    top-shape { m p } append naturals m p * t% :> tensor3
+    top-shape { m p } append
+
+    ! Now create the new float array to store the underlying result
+    dup product c:float (c-array) :> vec3
 
     ! Now update the tensor3 to contain the multiplied matricies
-    rest [0,b)
+    top-prod [
-    [
         :> i
-        ! First make vec1
+        ! Compute vec1 and start1
-        m n * i * dup m n * + tensor1 vec>> <slice>
+        tensor1 vec>> m n * i *
-        ! Now make vec2
+        ! Compute vec2 and start2
-        n p * i * dup n p * + tensor2 vec>> <slice>
+        tensor2 vec>> n p * i *
-        ! Now make the resulting vector
+        ! Compute the result
-        m p * i * dup m p * + tensor3 vec>> <slice>
+        vec3 m p * i *
-        ! Push n and p and multiply the clices
+        ! Push m, n, and p and multiply the arrays
-        n p 2d-matmul
+        m n p 2d-matmul
-        0
+    ] each-integer
-    ] map drop
+    vec3 <tensor> ;
-    tensor3 ;
+
 
 <PRIVATE
-! helper for transpose: gets the turns a shape into a list of things
+! helper for transpose: turns a shape into a list of things
 ! by which to multiply indices to get a full index
 : ind-mults ( shape -- seq )
-    rest-slice <reversed> cum-product { 1 } prepend ;
+    <reversed> 1 swap [ swap [ * ] keep ] map nip ;
 
 ! helper for transpose: given shape, flat index, & mults for the shape, gives nd index
-:: trans-index ( ind shape mults -- seq )
+: transpose-index ( i shape -- seq )
-    ! what we use to divide things
+    <reversed> [ /mod ] map reverse nip ;
-    shape reverse :> S
-    ! accumulator
-    V{ } clone
-    ! loop thru elements & indices of S (mod by elment m)
-    S [| m i |
-        ! we divide by the product of the 1st n elements of S
-        S i head-slice product :> div
-        ! do not mod on the last index
-        i S length 1 - = not :> mod?
-        ! multiply accumulator by mults & sum
-        dup mults [ * ] 2map sum
-        ! subtract from ind & divide
-        ind swap - div /
-        ! mod if necessary
-        mod? [ m mod ] [ ] if
-        ! append to accumulator
-        [ dup ] dip swap push
-    ] each-index
-    reverse ;
 PRIVATE>
 
-! Transpose an n-dimensional tensor
+! Transpose an n-dimensional tensor by flipping the axes
 TYPED:: transpose ( tensor: tensor -- tensor': tensor )
-    ! new shape
+    tensor shape>> :> old-shape
-    tensor shape>> reverse :> newshape
+    tensor vec>> :> vec
-    ! what we multiply by to get indices in the old tensor
+    old-shape reverse :> new-shape
-    tensor shape>> ind-mults :> old-mults
+    ! check that the size is fine
-    ! what we multiply to get indices in new tensor
+    new-shape product vec length assert=
-    newshape ind-mults :> mults
+    old-shape ind-mults reverse :> mults
-    ! new tensor of correct shape
+    ! loop through new tensor
-    newshape naturals dup vec>>
+    new-shape dup product <iota> [
-    [ ! go thru each index
         ! find index in original tensor
-        newshape mults trans-index old-mults [ * ] 2map sum >fixnum
+        old-shape mults [ [ /mod ] dip * ] 2map-sum nip
         ! get that index in original tensor
-        tensor vec>> nth
+        vec nth-unsafe
-    ] map! >>vec ;
+    ] float-array{ } map-as <tensor> ;