cwise-compiler - npm Package Compare versions

Comparing version 1.0.0 to 1.1.0

compiler.js

@@ -9,2 +9,3 @@ "use strict"
   this.arrayArgs = []
+  this.arrayBlockIndices = []
   this.scalarArgs = []
@@ -36,4 +37,6 @@ this.offsetArgs = []
     var arg_type = proc_args[i]
-    if(arg_type === "array") {
+    if(arg_type === "array" || (typeof arg_type === "object" && arg_type.blockIndices)) {
+      proc.argTypes[i] = "array"
       proc.arrayArgs.push(i)
+      proc.arrayBlockIndices.push(arg_type.blockIndices ? arg_type.blockIndices : 0)
       proc.shimArgs.push("array" + i)
@@ -40,0 +43,0 @@ if(i < proc.pre.args.length && proc.pre.args[i].count>0) {

lib/compile.js

@@ -5,2 +5,4 @@ "use strict"
 
+// This function generates very simple loops analogous to how you typically traverse arrays (the outermost loop corresponds to the slowest changing index, the innermost loop to the fastest changing index)
+// TODO: If two arrays have the same strides (and offsets) there is potential for decreasing the number of "pointers" and related variables. The drawback is that the type signature would become more specific and that there would thus be less potential for caching, but it might still be worth it, especially when dealing with large numbers of arguments.
 function innerFill(order, proc, body) {
@@ -13,3 +15,3 @@ var dimension = order.length
     , idx=0, pidx=0, i, j
-  for(i=0; i<dimension; ++i) {
+  for(i=0; i<dimension; ++i) { // Iteration variables
     vars.push(["i",i,"=0"].join(""))
@@ -22,5 +24,5 @@ }
       idx = order[i]
-      if(i === 0) {
+      if(i === 0) { // The innermost/fastest dimension's delta is simply its stride
         vars.push(["d",j,"s",i,"=t",j,"p",idx].join(""))
-      } else {
+      } else { // For other dimensions the delta is basically the stride minus something which essentially "rewinds" the previous (more inner) dimension
         vars.push(["d",j,"s",i,"=(t",j,"p",idx,"-s",pidx,"*t",j,"p",pidx,")"].join(""))
@@ -32,3 +34,3 @@ }
   //Scan loop
-  for(i=dimension-1; i>=0; --i) {
+  for(i=dimension-1; i>=0; --i) { // Start at largest stride and work your way inwards
     idx = order[i]
@@ -57,2 +59,5 @@ code.push(["for(i",i,"=0;i",i,"<s",idx,";++i",i,"){"].join(""))
 
+// Generate "outer" loops that loop over blocks of data, applying "inner" loops to the blocks by manipulating the local variables in such a way that the inner loop only "sees" the current block.
+// TODO: If this is used, then the previous declaration (done by generateCwiseOp) of s* is essentially unnecessary.
+//       I believe the s* are not used elsewhere (in particular, I don't think they're used in the pre/post parts and "shape" is defined independently), so it would be possible to make defining the s* dependent on what loop method is being used.
 function outerFill(matched, order, proc, body) {
@@ -67,6 +72,8 @@ var dimension = order.length
   }
-  //Generate matched loops
+  //Generate loops for unmatched dimensions
+  // The order in which these dimensions are traversed is fairly arbitrary (from small stride to large stride, for the first argument)
+  // TODO: It would be nice if the order in which these loops are placed would also be somehow "optimal" (at the very least we should check that it really doesn't hurt us if they're not).
   for(var i=matched; i<dimension; ++i) {
-    code.push(["for(var j"+i+"=SS[", order[i], "]|0;j", i, ">0;){"].join(""))
-    code.push(["if(j",i,"<",blockSize,"){"].join(""))
+    code.push(["for(var j"+i+"=SS[", order[i], "]|0;j", i, ">0;){"].join("")) // Iterate back to front
+    code.push(["if(j",i,"<",blockSize,"){"].join("")) // Either decrease j by blockSize (s = blockSize), or set it to zero (after setting s = j).
     code.push(["s",order[i],"=j",i].join(""))
@@ -95,2 +102,5 @@ code.push(["j",i,"=0"].join(""))
 //Count the number of compatible inner orders
+// This is the length of the longest common prefix of the arrays in orders.
+// Each array in orders lists the dimensions of the correspond ndarray in order of increasing stride.
+// This is thus the maximum number of dimensions that can be efficiently traversed by simple nested loops for all arrays.
 function countMatches(orders) {
@@ -110,2 +120,3 @@ var matched = 0, dimension = orders[0].length
 //Processes a block according to the given data types
+// Replaces variable names by different ones, either "local" ones (that are then ferried in and out of the given array) or ones matching the arguments that the function performing the ultimate loop will accept.
 function processBlock(block, proc, dtypes) {
@@ -128,3 +139,3 @@ var code = block.body
         arrNum = offArg.array
-        ptrStr = "+q" + offArgIndex
+        ptrStr = "+q" + offArgIndex // Adds offset to the "pointer" in the array
       case "array":
@@ -134,6 +145,39 @@ ptrStr = "p" + arrNum + ptrStr
         var arrStr = "a" + arrNum
-        if(carg.count === 1) {
+        if (proc.arrayBlockIndices[arrNum] === 0) { // Argument to body is just a single value from this array
+          if(carg.count === 1) { // Argument/array used only once(?)
+            if(dtypes[arrNum] === "generic") {
+              if(carg.lvalue) {
+                pre.push(["var ", localStr, "=", arrStr, ".get(", ptrStr, ")"].join("")) // Is this necessary if the argument is ONLY used as an lvalue? (keep in mind that we can have a += something, so we would actually need to check carg.rvalue)
+                code = code.replace(re, localStr)
+                post.push([arrStr, ".set(", ptrStr, ",", localStr,")"].join(""))
+              } else {
+                code = code.replace(re, [arrStr, ".get(", ptrStr, ")"].join(""))
+              }
+            } else {
+              code = code.replace(re, [arrStr, "[", ptrStr, "]"].join(""))
+            }
+          } else if(dtypes[arrNum] === "generic") {
+            pre.push(["var ", localStr, "=", arrStr, ".get(", ptrStr, ")"].join("")) // TODO: Could we optimize by checking for carg.rvalue?
+            code = code.replace(re, localStr)
+            if(carg.lvalue) {
+              post.push([arrStr, ".set(", ptrStr, ",", localStr,")"].join(""))
+            }
+          } else {
+            pre.push(["var ", localStr, "=", arrStr, "[", ptrStr, "]"].join("")) // TODO: Could we optimize by checking for carg.rvalue?
+            code = code.replace(re, localStr)
+            if(carg.lvalue) {
+              post.push([arrStr, "[", ptrStr, "]=", localStr].join(""))
+            }
+          }
+        } else { // Argument to body is a "block"
+          var reStrArr = [carg.name], ptrStrArr = [ptrStr]
+          for(var j=0; j<Math.abs(proc.arrayBlockIndices[arrNum]); j++) {
+            reStrArr.push("\\s*\\[([^\\]]+)\\]")
+            ptrStrArr.push("$" + (j+1) + "*t" + arrNum + "b" + j) // Matched index times stride
+          }
+          re = new RegExp(reStrArr.join(""), "g")
+          ptrStr = ptrStrArr.join("+")
           if(dtypes[arrNum] === "generic") {
-            if(carg.lvalue) {
-              pre.push(["var ", localStr, "=", arrStr, ".get(", ptrStr, ")"].join(""))
+            /*if(carg.lvalue) {
+              pre.push(["var ", localStr, "=", arrStr, ".get(", ptrStr, ")"].join("")) // Is this necessary if the argument is ONLY used as an lvalue? (keep in mind that we can have a += something, so we would actually need to check carg.rvalue)
               code = code.replace(re, localStr)
@@ -143,18 +187,8 @@ post.push([arrStr, ".set(", ptrStr, ",", localStr,")"].join(""))
               code = code.replace(re, [arrStr, ".get(", ptrStr, ")"].join(""))
-            }
+            }*/
+            throw new Error("cwise: Generic arrays not supported in combination with blocks!")
           } else {
+            // This does not produce any local variables, even if variables are used multiple times. It would be possible to do so, but it would complicate things quite a bit.
             code = code.replace(re, [arrStr, "[", ptrStr, "]"].join(""))
           }
-        } else if(dtypes[arrNum] === "generic") {
-          pre.push(["var ", localStr, "=", arrStr, ".get(", ptrStr, ")"].join(""))
-          code = code.replace(re, localStr)
-          if(carg.lvalue) {
-            post.push([arrStr, ".set(", ptrStr, ",", localStr,")"].join(""))
-          }
-        } else {
-          pre.push(["var ", localStr, "=", arrStr, "[", ptrStr, "]"].join(""))
-          code = code.replace(re, localStr)
-          if(carg.lvalue) {
-            post.push([arrStr, "[", ptrStr, "]=", localStr].join(""))
-          }
         }
@@ -206,8 +240,38 @@ break
   //Compute dimension
-  var dimension = typesig[1].length|0
+  // Arrays get put first in typesig, and there are two entries per array (dtype and order), so this gets the number of dimensions in the first array arg.
+  var dimension = (typesig[1].length - Math.abs(proc.arrayBlockIndices[0]))|0
   var orders = new Array(proc.arrayArgs.length)
   var dtypes = new Array(proc.arrayArgs.length)
+  for(var i=0; i<proc.arrayArgs.length; ++i) {
+    dtypes[i] = typesig[2*i]
+    orders[i] = typesig[2*i+1]
+  }
+  
+  //Determine where block and loop indices start and end
+  var blockBegin = [], blockEnd = [] // These indices are exposed as blocks
+  var loopBegin = [], loopEnd = [] // These indices are iterated over
+  var loopOrders = [] // orders restricted to the loop indices
+  for(var i=0; i<proc.arrayArgs.length; ++i) {
+    if (proc.arrayBlockIndices[i]<0) {
+      loopBegin.push(0)
+      loopEnd.push(dimension)
+      blockBegin.push(dimension)
+      blockEnd.push(dimension+proc.arrayBlockIndices[i])
+    } else {
+      loopBegin.push(proc.arrayBlockIndices[i]) // Non-negative
+      loopEnd.push(proc.arrayBlockIndices[i]+dimension)
+      blockBegin.push(0)
+      blockEnd.push(proc.arrayBlockIndices[i])
+    }
+    var newOrder = []
+    for(var j=0; j<orders[i].length; j++) {
+      if (loopBegin[i]<=orders[i][j] && orders[i][j]<loopEnd[i]) {
+        newOrder.push(orders[i][j]-loopBegin[i]) // If this is a loop index, put it in newOrder, subtracting loopBegin, to make sure that all loopOrders are using a common set of indices.
+      }
+    }
+    loopOrders.push(newOrder)
+  }
 
   //First create arguments for procedure
-  var arglist = ["SS"]
+  var arglist = ["SS"] // SS is the overall shape over which we iterate
   var code = ["'use strict'"]
@@ -217,14 +281,16 @@ var vars = []
   for(var j=0; j<dimension; ++j) {
-    vars.push(["s", j, "=SS[", j, "]"].join(""))
+    vars.push(["s", j, "=SS[", j, "]"].join("")) // The limits for each dimension.
   }
   for(var i=0; i<proc.arrayArgs.length; ++i) {
-    arglist.push("a"+i)
-    arglist.push("t"+i)
-    arglist.push("p"+i)
-    dtypes[i] = typesig[2*i]
-    orders[i] = typesig[2*i+1]
+    arglist.push("a"+i) // Actual data array
+    arglist.push("t"+i) // Strides
+    arglist.push("p"+i) // Offset in the array at which the data starts (also used for iterating over the data)
     
-    for(var j=0; j<dimension; ++j) {
-      vars.push(["t",i,"p",j,"=t",i,"[",j,"]"].join(""))
+    for(var j=0; j<dimension; ++j) { // Unpack the strides into vars for looping
+      vars.push(["t",i,"p",j,"=t",i,"[",loopBegin[i]+j,"]"].join(""))
     }
+    
+    for(var j=0; j<Math.abs(proc.arrayBlockIndices[i]); ++j) { // Unpack the strides into vars for block iteration
+      vars.push(["t",i,"b",j,"=t",i,"[",blockBegin[i]+j,"]"].join(""))
+    }
   }
@@ -235,5 +301,6 @@ for(var i=0; i<proc.scalarArgs.length; ++i) {
   if(proc.shapeArgs.length > 0) {
-    vars.push("shape=SS.slice(0)")
+    vars.push("shape=SS.slice(0)") // Makes the shape over which we iterate available to the user defined functions (so you can use width/height for example)
   }
   if(proc.indexArgs.length > 0) {
+    // Prepare an array to keep track of the (logical) indices, initialized to dimension zeroes.
     var zeros = new Array(dimension)
@@ -245,3 +312,3 @@ for(var i=0; i<dimension; ++i) {
   }
-  for(var i=0; i<proc.offsetArgs.length; ++i) {
+  for(var i=0; i<proc.offsetArgs.length; ++i) { // Offset arguments used for stencil operations
     var off_arg = proc.offsetArgs[i]
@@ -282,7 +349,7 @@ var init_string = []
   var body = processBlock(proc.body, proc, dtypes)
-  var matched = countMatches(orders)
+  var matched = countMatches(loopOrders)
   if(matched < dimension) {
-    code.push(outerFill(matched, orders[0], proc, body))
+    code.push(outerFill(matched, loopOrders[0], proc, body)) // TODO: Rather than passing loopOrders[0], it might be interesting to look at passing an order that represents the majority of the arguments for example.
   } else {
-    code.push(innerFill(orders[0], proc, body))
+    code.push(innerFill(loopOrders[0], proc, body))
   }
@@ -296,3 +363,3 @@
   if(proc.debug) {
-    console.log("Generated cwise routine for ", typesig, ":\n\n", code.join("\n"))
+    console.log("-----Generated cwise routine for ", typesig, ":\n" + code.join("\n") + "\n----------")
   }
@@ -304,2 +371,2 @@
 }
-module.exports = generateCWiseOp
+module.exports = generateCWiseOp

lib/thunk.js

 "use strict"
 
+// The function below is called when constructing a cwise function object, and does the following:
+// A function object is constructed which accepts as argument a compilation function and returns another function.
+// It is this other function that is eventually returned by createThunk, and this function is the one that actually
+// checks whether a certain pattern of arguments has already been used before and compiles new loops as needed.
+// The compilation passed to the first function object is used for compiling new functions.
+// Once this function object is created, it is called with compile as argument, where the first argument of compile
+// is bound to "proc" (essentially containing a preprocessed version of the user arguments to cwise).
+// So createThunk roughly works like this:
+// function createThunk(proc) {
+//   var thunk = function(compileBound) {
+//     var CACHED = {}
+//     return function(arrays and scalars) {
+//       if (dtype and order of arrays in CACHED) {
+//         var func = CACHED[dtype and order of arrays]
+//       } else {
+//         var func = CACHED[dtype and order of arrays] = compileBound(dtype and order of arrays)
+//       }
+//       return func(arrays and scalars)
+//     }
+//   }
+//   return thunk(compile.bind1(proc))
+// }
+
 var compile = require("./compile.js")
@@ -14,3 +37,4 @@
   var string_typesig = []
-  var proc_args = [["array",proc.arrayArgs[0],".shape"].join("")]
+  var proc_args = [["array",proc.arrayArgs[0],".shape.slice(", // Slice shape so that we only retain the shape over which we iterate (which gets passed to the cwise operator as SS).
+                    Math.max(0,proc.arrayBlockIndices[i]),proc.arrayBlockIndices[i]<0?(","+proc.arrayBlockIndices[i]+")"):")"].join("")]
   for(var i=0; i<proc.arrayArgs.length; ++i) {
@@ -40,3 +64,3 @@ var j = proc.arrayArgs[i]
   if(proc.debug) {
-    console.log("Generated thunk:", code.join("\n"))
+    console.log("-----Generated thunk:\n" + code.join("\n") + "\n----------")
   }
@@ -43,0 +67,0 @@

package.json

 {
   "name": "cwise-compiler",
-  "version": "1.0.0",
+  "version": "1.1.0",
   "description": "cwise's internal compiler",
@@ -27,3 +27,9 @@ "main": "compiler.js",
     "uniq": "^1.0.0"
+  },
+  "devDependencies": {
+    "ndarray-ops": "^1.2.2",
+    "ndarray": "^1.0.15",
+    "cwise-parser": "^1.0.0",
+    "tape": "^3.0.0"
   }
 }

New alerts

License Policy Violation

License

This package is not allowed per your license policy. Review the package's license to ensure compliance.

Found 1 instance in 1 package

Fixed alerts

License Policy Violation

License

This package is not allowed per your license policy. Review the package's license to ensure compliance.

Found 1 instance in 1 package

Improved metrics

Total package byte prevSize

24178

increased by53.11%

Lines of code

496

increased by23.08%

Worsened metrics

Dev dependency count

4

increased byInfinity%

No dependency changes