Changeset 172

Timestamp:

01/10/08 02:48:44 (8 months ago)

Author:

Don Clugston

Message:

Added sqrt intrinsic, with codegen for both X87 and SSE.

Files:

• trunk/blade/BladeDemo.d (modified) (1 diff)
• trunk/blade/BladeRank.d (modified) (3 diffs)
• trunk/blade/BladeSimplify.d (modified) (4 diffs)
• trunk/blade/CodegenX86.d (modified) (12 diffs)
• trunk/blade/PostfixX86.d (modified) (2 diffs)

trunk/blade/BladeDemo.d

@@ -60 +60 @@
-   mixin(vectorize("u = sum(abs(-p))"));
-   mixin(vectorize("a = -a"));
+   mixin(vectorize("u = sum(sqrt(abs(p))) + sum(sqrt(abs(q)))"));
-
-    writefln("a=", a);

trunk/blade/BladeRank.d

@@ -111 +111 @@
-            "Extra characters after array literal",
-            "Rank mismatch in array literal",
-
+, sqrt
-            ][-err-1];
-}
@@ -125 +125 @@
-    }
-    ReturnType onVisitFunction(This this_, char [] func, char [][] args) {
-
+
+
-            if (args.length!=2) return RankError.CommaExpected;
-            auto lrank = doVisit(this_,args[0]);
@@ -133 +134 @@
-            if (lrank!=1 || rrank!=1) return RankError.RankMismatchDotProduct;       
-            return 0;
-} else if (func=="sum") {
+case "sum":
-            auto lrank = doVisit(this_,args[0]);
-            if (lrank<0) return lrank; // propagate errors
-            return 0;
-
+
+
-            auto lrank = doVisit(this_,args[0]);
-            if (lrank>1) return RankError.AbsDimensionality;
-            return 0;
+        default:
+            assert(0, "BLADE ICE: Unsupported function:" ~ func);
+            return 0;
-        }        
-        assert(0, "BLADE ICE: Unsupported function:" ~ func);
-        return 0;
-    }
-    ReturnType onVisitPrefix(This this_, char [] op, char [] expr) {

trunk/blade/BladeSimplify.d

@@ -51 +51 @@
-bool isBladeIntrinsic(char [] str)
-{
-
+ || str=="sqrt"
-}
-
@@ -284 +284 @@
-        switch (func) {
-            case "dot":
-Each element is reduced seperately
+2-argument functions
-                char [] left = wrapInParens(doVisit(this_,args[0]));
-                char [] right = wrapInParens(doVisit(this_, args[1]));
-                return func ~ "(" ~ left ~ "," ~ right ~ ")";
-            case "sum":
+            case "sqrt":
-            case "abs":
+                // 1-argument functions
-                char [] left = wrapInParens(doVisit(this_,args[0]));
-                return func ~ "(" ~ left ~ ")";            
@@ -465 +467 @@
-        switch(func) {
-        case "dot":
+            // dot(a*v1, c*v2) == (a*b)*dot(v1, v2)
+            // this is a scalar, but the dot is a nested expression
-            ScalarFold left = doVisit(this_,args[0]);
-            ScalarFold right = doVisit(this_, args[1]);
-            return ScalarFold("", combineMul(combineMul(left.multiplier, right.multiplier), "{" ~ func ~ "(" ~ wrapInParens(left.expr) ~ "," ~ wrapInParens(right.expr) ~ ")}"));
-        case "sum": 
+            // Functions where f(v) is a scalar.
-            ScalarFold left = doVisit(this_,args[0]);
-            return ScalarFold("", combineMul(left.multiplier, "{" ~ func ~ "(" ~ wrapInParens(left.expr) ~ ")}"));
-
+
+
-            ScalarFold left = doVisit(this_,args[0]);
-            return ScalarFold((left.expr!="" ? func ~ "(" ~ wrapInParens(left.expr) ~ ")" : ""), (left.multiplier.length>0 && left.multiplier!="-")? func ~ "(" ~ wrapInParens(left.multiplier) ~ ")":"");
+        case "sqrt":
+            // There's a reduction in precision if we use f(a*v)==f(a)*f(v); also
+            // a and b could both be negative.
+            ScalarFold left = doVisit(this_,args[0]);
+            return ScalarFold((left.expr!="" ? func ~ "(" ~ wrapInParens(left.expr) ~ ")" : ""), (left.multiplier.length>0)? func ~ "(" ~ wrapInParens(left.multiplier) ~ ")":"");
-        default:
-            assert(0, "BLADE: Unsupported function");
@@ -600 +611 @@
-    assert(foldScalars("A*=-(-B)", "10")=="A*={(-(-B))}");
-    assert(foldScalars("sum((A*sum(A)))", "1")=="({sum(A)})*({sum(A)})");
-
+
+
+

trunk/blade/CodegenX86.d

@@ -308 +308 @@
-            ranklist~="1";
-            typelist ~= typeof(T.data[0]).stringof;
-
+ 
-            stridelist~="0";
-            ranklist~="0";
@@ -392 +392 @@
-    while(done<operations.length) {
-        char [] next;
-
+  
-            // Perform an arithmetic operation on the top two FPU stack items.
-            next = "  " ~ opToX87[operations[done]] ~ "p ST(1), ST;  //" ~ operations[done] ~ \n;
@@ -398 +398 @@
-            ++done;
-            numOnStack--;
-
-
-
-
-
-
-
+
+
+
+
+
+
+
+
+
+
-            // load a vector onto the FPU stack, to begin a new subexpression.
-            int u  = operations[done]-'A';
@@ -415 +418 @@
-            ++done;
-            numOnStack++;
-
-
-
-
+  
+  
+  
+  
-             // An operation will be performed between the stack top and a vector.
-             // If it's a float or double, we can combine the load+arithmetic op
@@ -432 +435 @@
-                next = "  fld real ptr ["  ~ vectorRegister[vectorNum(ranklist, operations[done])] ~ "]; //" ~ operations[done] ~ \n
-                    ~ "  " ~ opToX87[operations[done+1]] ~ "p ST(1), ST; //" ~ operations[done+1] ~\n;
- 
+
-                next = "  " ~ opToX87[operations[done+1]] ~ " "
-                  ~ indexedVector(typelist, ranklist, stridelist, operations[done] ) ~ comment;
@@ -438 +441 @@
-            mainbody ~= next;
-            done+=2;
-
-
- 
- 
- 
-
+  
+  
+
+
+
+  
-            // For scalar float or double values, we can multiply directly, saving one slot on the FP stack.
-            next = "  fmul " ~ operandSize(typelist[operations[done]-'A']) ~ "values[" ~ itoa(operations[done]-'A') ~"]; // * " ~ operations[done..done+1] ~ "\n";
-            mainbody ~= next; //firstbody ~= next;
-
-
-
+  
+    
+  
-    }
-        
@@ -562 +565 @@
-        "  jz short L2; // test for length==0"\n;
-    while(done<operations.length) {
-
-
+  
+  
-            // Perform an arithmetic operation on the top two items.
-            comment = ";  //" ~ operations[done] ~ \n;
@@ -570 +573 @@
-            ++done;
-            numOnStack--;
-
-
-
-
-
-
-
-
-
+
+
+
+
+
+
+
+
+
+
+
+
+
-            // load a vector onto the FPU stack, to begin a new subexpression.
-            comment = ";  // " ~ operations[done] ~ \n;
@@ -585 +592 @@
-            ++done;
-            numOnStack++;
-
+  
-                // operation on self, eg XX+ --> don't need to load it again.
-                comment = ";  // " ~ operations[done..done+2] ~ \n;
@@ -593 +600 @@
-                    ~ XMM(numOnStack-1) ~ comment;
-                done +=2;          
-
+  
-             // An operation will be performed between the stack top and a vector.
-            comment = ";  // " ~ operations[done..done+2] ~ \n;
@@ -606 +613 @@
-            }
-            done+=2;
-
+  
-            comment = "; // " ~operations[done..done+2] ~ \n;
-            int u = scalarNum(ranklist, operations[done]-'A');
@@ -612 +619 @@
-            extra ~= "  " ~ opToSSESingle[operations[done+1]] ~ suffix ~ XMM(numOnStack-1) ~ ", " ~ XMM(u) ~ comment;
-            done +=2;
-
+  
-    }
-        

trunk/blade/PostfixX86.d

@@ -57 +57 @@
-    }
-    ReturnType onVisitFunction(This this_, char [] func, char [][] args) {
-
-
-
-
-
-
+
+
+
+
+
+
+
+
+
+
+
+
-    }
-    ReturnType onVisitPrefix(This this_, char [] op, char [] expr) {
@@ -136 +142 @@
-    }
-    ReturnType onVisitFunction(This this_, char [] func, char [][] args) {      
-
-
-
-
-
-
+
+
+
+
+
+
+
+
+
+
+
+
-    }
-    ReturnType onVisitPrefix(This this_, char [] op, char [] expr) {