Changeset 3945

Timestamp:

09/29/08 14:43:38 (2 months ago)

Author:

Don Clugston

Message:

Bigint: Simple asm performance tests. Preparation for using cpuid cacheinfo.

Files:

• trunk/tango/math/internal/BignumX86.d (modified) (4 diffs)
• trunk/tango/math/internal/BiguintCore.d (modified) (2 diffs)

trunk/tango/math/internal/BignumX86.d

@@ -30 +30 @@
- *
- *  Timing results (cycles per int)
-
-
-
-
-
-
-
+
+
+
+
+
+
+
+
+
+
+
- */
-
@@ -605 +609 @@
-        mov EDI, [ESP + LASTPARAM + 4*5]; // dest.ptr
-        mov EBX, [ESP + LASTPARAM + 4*2]; // left.length
-        align 16;
-        nop;
-        nop;
-        
-        mov ESI, [ESP + LASTPARAM + 4*3];  // left.ptr
-        lea EDI, [EDI + 4*EBX]; // EDI = end of dest for first pass
@@ -680 +680 @@
-        adc EBP, 0;
-        mov [-4+EDI+4*EBX], EBP;
-        
-        add EDI, 4;
-        cmp EDI, [ESP + LASTPARAM + 4*0]; // is EDI = &dest[$]?
@@ -837 +836 @@
-}
-
+version(TangoPerformanceTest) {
+import tango.stdc.stdio;
+int clock() { asm { rdtsc; } }
+
+uint [2000] X1;
+uint [2000] Y1;
+uint [4000] Z1;
+
+void testPerformance()
+{
+    // The performance results at the top of this file were obtained using
+    // a Windows device driver to access the CPU performance counters.
+    // The code below is less accurate but more widely usable.
+    // The value for division is quite inconsistent.
+    for (int i=0; i<X1.length; ++i) { X1[i]=i; Y1[i]=i; Z1[i]=i; }
+    int t, t0;    
+    multibyteShr(Z1[0..2000], X1[0..2000], 7);
+    t0 = clock();
+    multibyteShr(Z1[0..1000], X1[0..1000], 7);
+    t = clock();
+    multibyteShr(Z1[0..2000], X1[0..2000], 7);
+    auto shrtime = (clock() - t) - (t - t0);
+    t0 = clock();
+    multibyteAddSub!('+')(Z1[0..1000], X1[0..1000], Y1[0..1000], 0);
+    t = clock();
+    multibyteAddSub!('+')(Z1[0..2000], X1[0..2000], Y1[0..2000], 0);
+    auto addtime = (clock() - t) - (t-t0);
+    t0 = clock();
+    multibyteMul(Z1[0..1000], X1[0..1000], 7, 0);
+    t = clock();
+    multibyteMul(Z1[0..2000], X1[0..2000], 7, 0);
+    auto multime = (clock() - t) - (t - t0);
+    multibyteMulAdd!('+')(Z1[0..2000], X1[0..2000], 217, 0);
+    t0 = clock();
+    multibyteMulAdd!('+')(Z1[0..1000], X1[0..1000], 217, 0);
+    t = clock();
+    multibyteMulAdd!('+')(Z1[0..2000], X1[0..2000], 217, 0);
+    auto muladdtime = (clock() - t) - (t - t0);        
+    multibyteMultiplyAccumulate(Z1[0..64], X1[0..32], Y1[0..32]);
+    t = clock();
+    multibyteMultiplyAccumulate(Z1[0..64], X1[0..32], Y1[0..32]);
+    auto accumtime = clock() - t;
+    t0 = clock();
+    multibyteDivAssign(Z1[0..2000], 217, 0);
+    t = clock();
+    multibyteDivAssign(Z1[0..1000], 37, 0);
+    auto divtime = (t - t0) - (clock() - t);
+    
+    printf("-- BigInt asm performance (cycles/int) --\n");    
+    printf("Add:        %.2f\n", addtime/1000.0);
+    printf("Shr:        %.2f\n", shrtime/1000.0);
+    printf("Mul:        %.2f\n", multime/1000.0);
+    printf("MulAdd:     %.2f\n", muladdtime/1000.0);
+    printf("Div:        %.2f\n", divtime/1000.0);
+    printf("MulAccum32: %.2f*n*n (total %d)\n\n", accumtime/(32.0*32.0), accumtime);
+}
+
+static this()
+{
+    testPerformance();
+}
+}
+
+
-} // version(D_InlineAsm_X86)

trunk/tango/math/internal/BiguintCore.d

@@ -21 +21 @@
-}
-
+// private import tango.core.Cpuid;
+static this()
+{
+    CACHELIMIT = 8000; // tango.core.Cpuid.datacache[0].size/2;
+}
+
-private:
+// Limits for when to switch between multiplication algorithms.
+const int CACHELIMIT;   // Half the size of the data cache.
+
-const uint [] ZERO = [0];
-const uint [] ONE = [1];
@@ -764 +773 @@
-}
-
-// Limits for when to switch between multiplication algorithms.
-const int CACHELIMIT = 8000;   // Half the size of the data cache.
-
-/* Determine how much space is required for the temporaries