Tue Mar 5 06:22:43 PM EST 2024

diff --git a/src/QAnsel.c b/src/QAnsel.c
index 16bdbd57dd447b1d7d277ad46dfb6c2e8967648a..0eca8d24a0e59aaa35bef4d2712ad857550b8366 100644 (file)
--- a/src/QAnsel.c
+++ b/src/QAnsel.c
@@ -238,7 +238,7 @@ void qansel_instruction(cpx_mtx_t* stateVector, unsigned char qubitCount, QInstr
                }
                else
                {
-                       cpx_mtx_knk(tmp.ptr, filter.ptr, gate.ptr, filter.rows, filter.cols, gate.rows, gate.cols);
+                       cpx_mtx_knk_2(tmp.ptr, filter.ptr, gate.ptr, filter.rows, filter.cols, gate.rows, gate.cols);
                }
                #endif
 
@@ -266,7 +266,7 @@ void qansel_instruction(cpx_mtx_t* stateVector, unsigned char qubitCount, QInstr
        us2 = get_time();
        printf("\tBare: %lu\n", us2 - us1);
        #else
-       if (USE_GPU && tmp.cols >= 32)
+       if (USE_GPU && tmp.cols >= 64)
        {
                cpx_mtx_dot_metal(tmp.ptr, stateVector->ptr, filter.ptr, stateVector->rows, stateVector->cols, filter.rows, filter.cols);
        }
@@ -1404,6 +1404,8 @@ void main(int argc, char** argv)
        USE_GPU = cpx_mtx_begin();
        RANDOM_FILE = fopen("/dev/TrueRNG0", "r");
        if (!RANDOM_FILE) RANDOM_FILE = fopen("/dev/random", "r");
+       USE_GPU = 0;
+       USE_THREADS = 0;
        process(argc, argv);
        fclose(RANDOM_FILE);
        if (USE_GPU) cpx_mtx_clean();

diff --git a/src/complex.c b/src/complex.c
index 931428fde3ce2907c5a605caaec9ec9ae744625f..e1eb9931184bc716900938bc3d459dd69dec2e78 100644 (file)
--- a/src/complex.c
+++ b/src/complex.c
@@ -143,6 +143,19 @@ void cpx_mtx_knk(float* ptrR, float* ptrA, float* ptrB, int rowsA, int colsA, in
     }
 }
 
+void cpx_mtx_knk_2(float* ptrR, float* ptrA, float* ptrB, int rowsA, int colsA, int rowsB, int colsB)
+{
+       int rowsR = rowsA * rowsB;
+       int colsR = colsA * colsB;
+    for (int i = 0; i < rowsR; i++)
+    {
+               for (int j = 0; j < colsR; j++)
+               {
+                       kernel_knk_2(ptrR, ptrA, ptrB, rowsA, colsA, rowsB, colsB, i, j);
+               }
+    }
+}
+
 /*-----------------------------------------------------------------------------------*/
 /*THREADED*/
 /*-----------------------------------------------------------------------------------*/

diff --git a/src/kernel.cl b/src/kernel.cl
index 2d11a44a7f6a38f448243198678a33f0f208e34d..66c40dbb51eb2a74b7627c6b5c9312d552c8e5ed 100644 (file)
--- a/src/kernel.cl
+++ b/src/kernel.cl
@@ -18,26 +18,25 @@ __kernel void kernel_dot
     const int colR = get_global_id(1); //{gpu_only}
     //const int rowR = get_global_id_0; {cpu_only}
     //const int colR = get_global_id_1; {cpu_only}
-    int posA, posB;
     float rR = 0;
     float iR = 0;
     const int posR = rowR * (colsR * 2) + (colR * 2);
 
     for (int i = 0; i < colsA; i++)
     {
-        int posA = rowR * (colsA * 2) + (i * 2);
-        int posB = i * (colsB * 2) + (colR * 2);
+        const int posA = rowR * (colsA * 2) + (i * 2);
+        const int posB = i * (colsB * 2) + (colR * 2);
 
-        float rA = ptrA[posA];
-        float iA = ptrA[posA + 1];
-        float rB = ptrB[posB];
-        float iB = ptrB[posB + 1];
+        const float rA = ptrA[posA];
+        const float iA = ptrA[posA + 1];
+        const float rB = ptrB[posB];
+        const float iB = ptrB[posB + 1];
 
         //(rA + iA)(rB + iB)
-        float first = rA * rB;
-        float outer = rA * iB;
-        float inner = iA * rB;
-        float lasts = iA * iB;
+        const float first = rA * rB;
+        const float outer = rA * iB;
+        const float inner = iA * rB;
+        const float lasts = iA * iB;
 
         rR += first + lasts;
         iR += outer + inner;
@@ -87,3 +86,83 @@ __kernel void kernel_knk
         ptrR[rowR * (colsR * 2) + (colR * 2) + 1] = outer + inner;
     }
 }
+
+__kernel void kernel_knk_2
+(
+    __global float* ptrR,
+    __global float* ptrA,
+    __global float* ptrB,
+    const int rowsA,
+    const int colsA,
+    const int rowsB,
+    const int colsB //{gpu_only}
+    //const int colsB, {cpu_only}
+    //const int get_global_id_0, {cpu_only}
+    //const int get_global_id_1 {cpu_only}
+)
+{
+    //This is based off the equality that
+    //  (A knk I) dot (J knk B) = A knk B
+    //  If I and J are identity matrices, then
+    //      we don't even need to compute them
+    //      to know their values at a given
+    //      index, meaning we just need to
+    //      carry out effectively a copy of
+    //      the dot product procedure.
+    const int rowsI = rowsB;
+    const int colsI = rowsB;
+    const int rowsJ = colsA;
+    const int colsJ = colsA;
+    const int rowsX = rowsA * rowsI;
+    const int colsX = colsA * colsI;
+    const int rowsY = rowsJ * rowsB;
+    const int colsY = colsJ * colsB;
+    const int rowsR = rowsX;
+    const int colsR = colsY;
+
+    const int rowR = get_global_id(0); //{gpu_only}
+    const int colR = get_global_id(1); //{gpu_only}
+    //const int rowR = get_global_id_0; {cpu_only}
+    //const int colR = get_global_id_1; {cpu_only}
+    int posA, posB;
+    float rR = 0;
+    float iR = 0;
+    const int posR = rowR * (colsR * 2) + (colR * 2);
+
+    for (int i = 0; i < colsX; i++)
+    {
+        const int rowX = rowR;
+        const int colX = i;
+        const int rowY = i;
+        const int colY = colR;
+
+        const int rowA = rowX / rowsI;
+        const int colA = colX / colsI;
+        const int rowI = rowX % rowsI;
+        const int colI = colX % colsI;
+
+        const int rowJ = rowY / rowsB;
+        const int colJ = colY / colsB;
+        const int rowB = rowY % rowsB;
+        const int colB = colY % colsB;
+
+        const int posA = rowA * (colsA * 2) + (colsA * 2);
+        const int posB = rowB * (colsB * 2) + (colsB * 2);
+
+        const float rA = rowI == colI ? ptrA[posA] : 0;
+        const float iA = rowI == colI ? ptrA[posA + 1] : 0;
+        const float rB = rowJ == colJ ? ptrB[posB] : 0;
+        const float iB = rowJ == colJ ? ptrB[posB + 1] : 0;
+
+        //(rA + iA)(rB + iB)
+        const float first = rA * rB;
+        const float outer = rA * iB;
+        const float inner = iA * rB;
+        const float lasts = iA * iB;
+
+        rR += first + lasts;
+        iR += outer + inner;
+    }
+    ptrR[rowR * (colsR * 2) + (colR * 2)] = rR;
+    ptrR[rowR * (colsR * 2) + (colR * 2) + 1] = iR;
+}

diff --git a/src/kernel.cl.c b/src/kernel.cl.c
index 34ea4ade92d5257004462cc242888d900d822537..0b2ee4e360a41e6ef3131d55fa6af04ecc01c0b2 100644 (file)
--- a/src/kernel.cl.c
+++ b/src/kernel.cl.c
@@ -48,169 +48,380 @@ unsigned char src_kernel_cl[] = {
   0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x3d, 0x20, 0x67, 0x65,
   0x74, 0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x5f, 0x69, 0x64, 0x5f,
   0x31, 0x3b, 0x20, 0x7b, 0x63, 0x70, 0x75, 0x5f, 0x6f, 0x6e, 0x6c, 0x79,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x70, 0x6f,
-  0x73, 0x41, 0x2c, 0x20, 0x70, 0x6f, 0x73, 0x42, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x72, 0x52, 0x20, 0x3d,
-  0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6c, 0x6f, 0x61,
-  0x74, 0x20, 0x69, 0x52, 0x20, 0x3d, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20,
+  0x72, 0x52, 0x20, 0x3d, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x69, 0x52, 0x20, 0x3d, 0x20, 0x30,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x69, 0x6e, 0x74, 0x20, 0x70, 0x6f, 0x73, 0x52, 0x20, 0x3d, 0x20, 0x72,
+  0x6f, 0x77, 0x52, 0x20, 0x2a, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x52,
+  0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c,
+  0x52, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x69, 0x6e, 0x74, 0x20, 0x69, 0x20,
+  0x3d, 0x20, 0x30, 0x3b, 0x20, 0x69, 0x20, 0x3c, 0x20, 0x63, 0x6f, 0x6c,
+  0x73, 0x41, 0x3b, 0x20, 0x69, 0x2b, 0x2b, 0x29, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x70, 0x6f, 0x73,
+  0x41, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2a, 0x20, 0x28,
+  0x63, 0x6f, 0x6c, 0x73, 0x41, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b,
+  0x20, 0x28, 0x69, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x69, 0x6e, 0x74, 0x20, 0x70, 0x6f, 0x73, 0x42, 0x20, 0x3d, 0x20, 0x69,
+  0x20, 0x2a, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x20, 0x2a, 0x20,
+  0x32, 0x29, 0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2a,
+  0x20, 0x32, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61,
+  0x74, 0x20, 0x72, 0x41, 0x20, 0x3d, 0x20, 0x70, 0x74, 0x72, 0x41, 0x5b,
+  0x70, 0x6f, 0x73, 0x41, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f,
+  0x61, 0x74, 0x20, 0x69, 0x41, 0x20, 0x3d, 0x20, 0x70, 0x74, 0x72, 0x41,
+  0x5b, 0x70, 0x6f, 0x73, 0x41, 0x20, 0x2b, 0x20, 0x31, 0x5d, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x72, 0x42, 0x20, 0x3d,
+  0x20, 0x70, 0x74, 0x72, 0x42, 0x5b, 0x70, 0x6f, 0x73, 0x42, 0x5d, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x69, 0x42, 0x20,
+  0x3d, 0x20, 0x70, 0x74, 0x72, 0x42, 0x5b, 0x70, 0x6f, 0x73, 0x42, 0x20,
+  0x2b, 0x20, 0x31, 0x5d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x28, 0x72, 0x41, 0x20, 0x2b, 0x20, 0x69,
+  0x41, 0x29, 0x28, 0x72, 0x42, 0x20, 0x2b, 0x20, 0x69, 0x42, 0x29, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x66, 0x69, 0x72, 0x73,
+  0x74, 0x20, 0x3d, 0x20, 0x72, 0x41, 0x20, 0x2a, 0x20, 0x72, 0x42, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x6f, 0x75, 0x74,
+  0x65, 0x72, 0x20, 0x3d, 0x20, 0x72, 0x41, 0x20, 0x2a, 0x20, 0x69, 0x42,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x69, 0x6e,
+  0x6e, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x69, 0x41, 0x20, 0x2a, 0x20, 0x72,
+  0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x6c,
+  0x61, 0x73, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x69, 0x41, 0x20, 0x2a, 0x20,
+  0x69, 0x42, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x72, 0x52, 0x20, 0x2b, 0x3d, 0x20, 0x66, 0x69, 0x72, 0x73, 0x74,
+  0x20, 0x2b, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x73, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x52, 0x20, 0x2b, 0x3d, 0x20,
+  0x6f, 0x75, 0x74, 0x65, 0x72, 0x20, 0x2b, 0x20, 0x69, 0x6e, 0x6e, 0x65,
+  0x72, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x70, 0x74, 0x72, 0x52, 0x5b, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2a,
+  0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29,
+  0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2a, 0x20, 0x32,
+  0x29, 0x5d, 0x20, 0x3d, 0x20, 0x72, 0x52, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x70, 0x74, 0x72, 0x52, 0x5b, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2a,
+  0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29,
+  0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2a, 0x20, 0x32,
+  0x29, 0x20, 0x2b, 0x20, 0x31, 0x5d, 0x20, 0x3d, 0x20, 0x69, 0x52, 0x3b,
+  0x0a, 0x7d, 0x0a, 0x0a, 0x5f, 0x5f, 0x6b, 0x65, 0x72, 0x6e, 0x65, 0x6c,
+  0x20, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x6b, 0x65, 0x72, 0x6e, 0x65, 0x6c,
+  0x5f, 0x6b, 0x6e, 0x6b, 0x0a, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x5f,
+  0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x20, 0x66, 0x6c, 0x6f, 0x61,
+  0x74, 0x2a, 0x20, 0x70, 0x74, 0x72, 0x52, 0x2c, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x5f, 0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x20, 0x66, 0x6c,
+  0x6f, 0x61, 0x74, 0x2a, 0x20, 0x70, 0x74, 0x72, 0x41, 0x2c, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x5f, 0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x20,
+  0x66, 0x6c, 0x6f, 0x61, 0x74, 0x2a, 0x20, 0x70, 0x74, 0x72, 0x42, 0x2c,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69,
+  0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x41, 0x2c, 0x0a, 0x20, 0x20,
   0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20,
-  0x70, 0x6f, 0x73, 0x52, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x52, 0x20,
-  0x2a, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32,
-  0x29, 0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2a, 0x20,
-  0x32, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72,
-  0x20, 0x28, 0x69, 0x6e, 0x74, 0x20, 0x69, 0x20, 0x3d, 0x20, 0x30, 0x3b,
-  0x20, 0x69, 0x20, 0x3c, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x41, 0x3b, 0x20,
-  0x69, 0x2b, 0x2b, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x70,
-  0x6f, 0x73, 0x41, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2a,
-  0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x41, 0x20, 0x2a, 0x20, 0x32, 0x29,
-  0x20, 0x2b, 0x20, 0x28, 0x69, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x3b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6e, 0x74, 0x20,
-  0x70, 0x6f, 0x73, 0x42, 0x20, 0x3d, 0x20, 0x69, 0x20, 0x2a, 0x20, 0x28,
-  0x63, 0x6f, 0x6c, 0x73, 0x42, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b,
-  0x20, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x3b,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6c,
-  0x6f, 0x61, 0x74, 0x20, 0x72, 0x41, 0x20, 0x3d, 0x20, 0x70, 0x74, 0x72,
-  0x41, 0x5b, 0x70, 0x6f, 0x73, 0x41, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x69,
-  0x41, 0x20, 0x3d, 0x20, 0x70, 0x74, 0x72, 0x41, 0x5b, 0x70, 0x6f, 0x73,
-  0x41, 0x20, 0x2b, 0x20, 0x31, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x72, 0x42,
+  0x63, 0x6f, 0x6c, 0x73, 0x41, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77,
+  0x73, 0x42, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x20,
+  0x2f, 0x2f, 0x7b, 0x67, 0x70, 0x75, 0x5f, 0x6f, 0x6e, 0x6c, 0x79, 0x7d,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x2c, 0x20,
+  0x7b, 0x63, 0x70, 0x75, 0x5f, 0x6f, 0x6e, 0x6c, 0x79, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69,
+  0x6e, 0x74, 0x20, 0x67, 0x65, 0x74, 0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61,
+  0x6c, 0x5f, 0x69, 0x64, 0x5f, 0x30, 0x7b, 0x63, 0x70, 0x75, 0x5f, 0x6f,
+  0x6e, 0x6c, 0x79, 0x7d, 0x0a, 0x29, 0x0a, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72,
+  0x6f, 0x77, 0x73, 0x52, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x41,
+  0x20, 0x2a, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x42, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20,
+  0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6c, 0x73,
+  0x41, 0x20, 0x2a, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74,
+  0x20, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x3d, 0x20, 0x67, 0x65, 0x74, 0x5f,
+  0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x5f, 0x69, 0x64, 0x28, 0x30, 0x29,
+  0x3b, 0x20, 0x2f, 0x2f, 0x7b, 0x67, 0x70, 0x75, 0x5f, 0x6f, 0x6e, 0x6c,
+  0x79, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x52, 0x20,
+  0x3d, 0x20, 0x67, 0x65, 0x74, 0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c,
+  0x5f, 0x69, 0x64, 0x5f, 0x30, 0x3b, 0x20, 0x7b, 0x63, 0x70, 0x75, 0x5f,
+  0x6f, 0x6e, 0x6c, 0x79, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f,
+  0x72, 0x20, 0x28, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x52, 0x20,
+  0x3d, 0x20, 0x30, 0x3b, 0x20, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x3c, 0x20,
+  0x63, 0x6f, 0x6c, 0x73, 0x52, 0x3b, 0x20, 0x63, 0x6f, 0x6c, 0x52, 0x2b,
+  0x2b, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69,
+  0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x41, 0x20, 0x3d, 0x20, 0x72, 0x6f,
+  0x77, 0x52, 0x20, 0x2f, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x42, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x41, 0x20, 0x3d,
+  0x20, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2f, 0x20, 0x63, 0x6f, 0x6c, 0x73,
+  0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77,
+  0x42, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x25, 0x20, 0x72,
+  0x6f, 0x77, 0x73, 0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20,
+  0x63, 0x6f, 0x6c, 0x42, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6c, 0x52, 0x20,
+  0x25, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x3b, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x69, 0x6e, 0x74, 0x20, 0x70, 0x6f, 0x73, 0x41, 0x20, 0x3d, 0x20, 0x72,
+  0x6f, 0x77, 0x41, 0x20, 0x2a, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x41,
+  0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c,
+  0x41, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e,
+  0x74, 0x20, 0x70, 0x6f, 0x73, 0x42, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77,
+  0x42, 0x20, 0x2a, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x20, 0x2a,
+  0x20, 0x32, 0x29, 0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x42, 0x20,
+  0x2a, 0x20, 0x32, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f,
+  0x61, 0x74, 0x20, 0x72, 0x41, 0x20, 0x3d, 0x20, 0x70, 0x74, 0x72, 0x41,
+  0x5b, 0x70, 0x6f, 0x73, 0x41, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c,
+  0x6f, 0x61, 0x74, 0x20, 0x69, 0x41, 0x20, 0x3d, 0x20, 0x70, 0x74, 0x72,
+  0x41, 0x5b, 0x70, 0x6f, 0x73, 0x41, 0x20, 0x2b, 0x20, 0x31, 0x5d, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x72, 0x42, 0x20,
+  0x3d, 0x20, 0x70, 0x74, 0x72, 0x42, 0x5b, 0x70, 0x6f, 0x73, 0x42, 0x5d,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x69, 0x42,
   0x20, 0x3d, 0x20, 0x70, 0x74, 0x72, 0x42, 0x5b, 0x70, 0x6f, 0x73, 0x42,
-  0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66,
-  0x6c, 0x6f, 0x61, 0x74, 0x20, 0x69, 0x42, 0x20, 0x3d, 0x20, 0x70, 0x74,
-  0x72, 0x42, 0x5b, 0x70, 0x6f, 0x73, 0x42, 0x20, 0x2b, 0x20, 0x31, 0x5d,
-  0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
-  0x2f, 0x28, 0x72, 0x41, 0x20, 0x2b, 0x20, 0x69, 0x41, 0x29, 0x28, 0x72,
-  0x42, 0x20, 0x2b, 0x20, 0x69, 0x42, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x66, 0x69,
-  0x72, 0x73, 0x74, 0x20, 0x3d, 0x20, 0x72, 0x41, 0x20, 0x2a, 0x20, 0x72,
-  0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66,
-  0x6c, 0x6f, 0x61, 0x74, 0x20, 0x6f, 0x75, 0x74, 0x65, 0x72, 0x20, 0x3d,
-  0x20, 0x72, 0x41, 0x20, 0x2a, 0x20, 0x69, 0x42, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20,
-  0x69, 0x6e, 0x6e, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x69, 0x41, 0x20, 0x2a,
-  0x20, 0x72, 0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x73,
-  0x20, 0x3d, 0x20, 0x69, 0x41, 0x20, 0x2a, 0x20, 0x69, 0x42, 0x3b, 0x0a,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x52, 0x20,
-  0x2b, 0x3d, 0x20, 0x66, 0x69, 0x72, 0x73, 0x74, 0x20, 0x2b, 0x20, 0x6c,
-  0x61, 0x73, 0x74, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x69, 0x52, 0x20, 0x2b, 0x3d, 0x20, 0x6f, 0x75, 0x74, 0x65,
-  0x72, 0x20, 0x2b, 0x20, 0x69, 0x6e, 0x6e, 0x65, 0x72, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x70, 0x74, 0x72,
-  0x52, 0x5b, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2a, 0x20, 0x28, 0x63, 0x6f,
-  0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b, 0x20, 0x28,
-  0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x5d, 0x20, 0x3d,
-  0x20, 0x72, 0x52, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x70, 0x74, 0x72,
-  0x52, 0x5b, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2a, 0x20, 0x28, 0x63, 0x6f,
-  0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b, 0x20, 0x28,
-  0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b, 0x20,
-  0x31, 0x5d, 0x20, 0x3d, 0x20, 0x69, 0x52, 0x3b, 0x0a, 0x7d, 0x0a, 0x0a,
+  0x20, 0x2b, 0x20, 0x31, 0x5d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x28, 0x72, 0x41, 0x20, 0x2b, 0x20,
+  0x69, 0x41, 0x29, 0x28, 0x72, 0x42, 0x20, 0x2b, 0x20, 0x69, 0x42, 0x29,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x66, 0x69, 0x72,
+  0x73, 0x74, 0x20, 0x3d, 0x20, 0x72, 0x41, 0x20, 0x2a, 0x20, 0x72, 0x42,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x6f, 0x75,
+  0x74, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x72, 0x41, 0x20, 0x2a, 0x20, 0x69,
+  0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x69,
+  0x6e, 0x6e, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x69, 0x41, 0x20, 0x2a, 0x20,
+  0x72, 0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20,
+  0x6c, 0x61, 0x73, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x69, 0x41, 0x20, 0x2a,
+  0x20, 0x69, 0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x70, 0x74, 0x72, 0x52, 0x5b, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2a,
+  0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29,
+  0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2a, 0x20, 0x32,
+  0x29, 0x5d, 0x20, 0x3d, 0x20, 0x66, 0x69, 0x72, 0x73, 0x74, 0x20, 0x2b,
+  0x20, 0x6c, 0x61, 0x73, 0x74, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x70, 0x74, 0x72, 0x52, 0x5b, 0x72, 0x6f, 0x77,
+  0x52, 0x20, 0x2a, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20, 0x2a,
+  0x20, 0x32, 0x29, 0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20,
+  0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b, 0x20, 0x31, 0x5d, 0x20, 0x3d, 0x20,
+  0x6f, 0x75, 0x74, 0x65, 0x72, 0x20, 0x2b, 0x20, 0x69, 0x6e, 0x6e, 0x65,
+  0x72, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x0a, 0x0a,
   0x5f, 0x5f, 0x6b, 0x65, 0x72, 0x6e, 0x65, 0x6c, 0x20, 0x76, 0x6f, 0x69,
   0x64, 0x20, 0x6b, 0x65, 0x72, 0x6e, 0x65, 0x6c, 0x5f, 0x6b, 0x6e, 0x6b,
-  0x0a, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x5f, 0x5f, 0x67, 0x6c, 0x6f,
-  0x62, 0x61, 0x6c, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x2a, 0x20, 0x70,
-  0x74, 0x72, 0x52, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x5f, 0x5f, 0x67,
+  0x5f, 0x32, 0x0a, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x5f, 0x5f, 0x67,
   0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x2a,
-  0x20, 0x70, 0x74, 0x72, 0x41, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x5f,
+  0x20, 0x70, 0x74, 0x72, 0x52, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x5f,
   0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x20, 0x66, 0x6c, 0x6f, 0x61,
-  0x74, 0x2a, 0x20, 0x70, 0x74, 0x72, 0x42, 0x2c, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72,
-  0x6f, 0x77, 0x73, 0x41, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x73,
-  0x41, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
-  0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x42, 0x2c, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e,
-  0x74, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x20, 0x2f, 0x2f, 0x7b, 0x67,
+  0x74, 0x2a, 0x20, 0x70, 0x74, 0x72, 0x41, 0x2c, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x5f, 0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x20, 0x66, 0x6c,
+  0x6f, 0x61, 0x74, 0x2a, 0x20, 0x70, 0x74, 0x72, 0x42, 0x2c, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74,
+  0x20, 0x72, 0x6f, 0x77, 0x73, 0x41, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f,
+  0x6c, 0x73, 0x41, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x42,
+  0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x20, 0x2f, 0x2f,
+  0x7b, 0x67, 0x70, 0x75, 0x5f, 0x6f, 0x6e, 0x6c, 0x79, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69,
+  0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x2c, 0x20, 0x7b, 0x63,
   0x70, 0x75, 0x5f, 0x6f, 0x6e, 0x6c, 0x79, 0x7d, 0x0a, 0x20, 0x20, 0x20,
   0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74,
-  0x20, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x2c, 0x20, 0x7b, 0x63, 0x70, 0x75,
-  0x5f, 0x6f, 0x6e, 0x6c, 0x79, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f,
-  0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x67,
-  0x65, 0x74, 0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x5f, 0x69, 0x64,
-  0x5f, 0x30, 0x7b, 0x63, 0x70, 0x75, 0x5f, 0x6f, 0x6e, 0x6c, 0x79, 0x7d,
-  0x0a, 0x29, 0x0a, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x52,
-  0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x41, 0x20, 0x2a, 0x20, 0x72,
-  0x6f, 0x77, 0x73, 0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x73,
-  0x52, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x41, 0x20, 0x2a, 0x20,
-  0x63, 0x6f, 0x6c, 0x73, 0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77,
+  0x20, 0x67, 0x65, 0x74, 0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x5f,
+  0x69, 0x64, 0x5f, 0x30, 0x2c, 0x20, 0x7b, 0x63, 0x70, 0x75, 0x5f, 0x6f,
+  0x6e, 0x6c, 0x79, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x67, 0x65, 0x74,
+  0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x5f, 0x69, 0x64, 0x5f, 0x31,
+  0x20, 0x7b, 0x63, 0x70, 0x75, 0x5f, 0x6f, 0x6e, 0x6c, 0x79, 0x7d, 0x0a,
+  0x29, 0x0a, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x54, 0x68,
+  0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x62, 0x61, 0x73, 0x65, 0x64, 0x20,
+  0x6f, 0x66, 0x66, 0x20, 0x74, 0x68, 0x65, 0x20, 0x65, 0x71, 0x75, 0x61,
+  0x6c, 0x69, 0x74, 0x79, 0x20, 0x74, 0x68, 0x61, 0x74, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x2f, 0x2f, 0x20, 0x20, 0x28, 0x41, 0x20, 0x6b, 0x6e, 0x6b,
+  0x20, 0x49, 0x29, 0x20, 0x64, 0x6f, 0x74, 0x20, 0x28, 0x4a, 0x20, 0x6b,
+  0x6e, 0x6b, 0x20, 0x42, 0x29, 0x20, 0x3d, 0x20, 0x41, 0x20, 0x6b, 0x6e,
+  0x6b, 0x20, 0x42, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x20,
+  0x49, 0x66, 0x20, 0x49, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x4a, 0x20, 0x61,
+  0x72, 0x65, 0x20, 0x69, 0x64, 0x65, 0x6e, 0x74, 0x69, 0x74, 0x79, 0x20,
+  0x6d, 0x61, 0x74, 0x72, 0x69, 0x63, 0x65, 0x73, 0x2c, 0x20, 0x74, 0x68,
+  0x65, 0x6e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x77, 0x65, 0x20, 0x64, 0x6f, 0x6e, 0x27, 0x74, 0x20,
+  0x65, 0x76, 0x65, 0x6e, 0x20, 0x6e, 0x65, 0x65, 0x64, 0x20, 0x74, 0x6f,
+  0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x20, 0x74, 0x68, 0x65,
+  0x6d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x74, 0x6f, 0x20, 0x6b, 0x6e, 0x6f, 0x77, 0x20, 0x74, 0x68,
+  0x65, 0x69, 0x72, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x73, 0x20, 0x61,
+  0x74, 0x20, 0x61, 0x20, 0x67, 0x69, 0x76, 0x65, 0x6e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6e,
+  0x64, 0x65, 0x78, 0x2c, 0x20, 0x6d, 0x65, 0x61, 0x6e, 0x69, 0x6e, 0x67,
+  0x20, 0x77, 0x65, 0x20, 0x6a, 0x75, 0x73, 0x74, 0x20, 0x6e, 0x65, 0x65,
+  0x64, 0x20, 0x74, 0x6f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, 0x72, 0x72, 0x79, 0x20, 0x6f,
+  0x75, 0x74, 0x20, 0x65, 0x66, 0x66, 0x65, 0x63, 0x74, 0x69, 0x76, 0x65,
+  0x6c, 0x79, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x70, 0x79, 0x20, 0x6f, 0x66,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x64, 0x6f, 0x74, 0x20, 0x70, 0x72, 0x6f,
+  0x64, 0x75, 0x63, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x63, 0x65, 0x64, 0x75,
+  0x72, 0x65, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x49, 0x20,
+  0x3d, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63,
+  0x6f, 0x6c, 0x73, 0x49, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x42,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x4a, 0x20, 0x3d, 0x20,
+  0x63, 0x6f, 0x6c, 0x73, 0x41, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c,
+  0x73, 0x4a, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x41, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e,
+  0x74, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x58, 0x20, 0x3d, 0x20, 0x72, 0x6f,
+  0x77, 0x73, 0x41, 0x20, 0x2a, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x49, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69,
+  0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x58, 0x20, 0x3d, 0x20, 0x63,
+  0x6f, 0x6c, 0x73, 0x41, 0x20, 0x2a, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x49,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x59, 0x20, 0x3d, 0x20,
+  0x72, 0x6f, 0x77, 0x73, 0x4a, 0x20, 0x2a, 0x20, 0x72, 0x6f, 0x77, 0x73,
+  0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x59, 0x20, 0x3d,
+  0x20, 0x63, 0x6f, 0x6c, 0x73, 0x4a, 0x20, 0x2a, 0x20, 0x63, 0x6f, 0x6c,
+  0x73, 0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x52, 0x20,
+  0x3d, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x58, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63,
+  0x6f, 0x6c, 0x73, 0x52, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x59,
+  0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x3d, 0x20,
+  0x67, 0x65, 0x74, 0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x5f, 0x69,
+  0x64, 0x28, 0x30, 0x29, 0x3b, 0x20, 0x2f, 0x2f, 0x7b, 0x67, 0x70, 0x75,
+  0x5f, 0x6f, 0x6e, 0x6c, 0x79, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c,
   0x52, 0x20, 0x3d, 0x20, 0x67, 0x65, 0x74, 0x5f, 0x67, 0x6c, 0x6f, 0x62,
-  0x61, 0x6c, 0x5f, 0x69, 0x64, 0x28, 0x30, 0x29, 0x3b, 0x20, 0x2f, 0x2f,
+  0x61, 0x6c, 0x5f, 0x69, 0x64, 0x28, 0x31, 0x29, 0x3b, 0x20, 0x2f, 0x2f,
   0x7b, 0x67, 0x70, 0x75, 0x5f, 0x6f, 0x6e, 0x6c, 0x79, 0x7d, 0x0a, 0x20,
   0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69,
   0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x3d, 0x20, 0x67, 0x65,
   0x74, 0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x5f, 0x69, 0x64, 0x5f,
   0x30, 0x3b, 0x20, 0x7b, 0x63, 0x70, 0x75, 0x5f, 0x6f, 0x6e, 0x6c, 0x79,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x69,
-  0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x3d, 0x20, 0x30, 0x3b,
-  0x20, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x3c, 0x20, 0x63, 0x6f, 0x6c, 0x73,
-  0x52, 0x3b, 0x20, 0x63, 0x6f, 0x6c, 0x52, 0x2b, 0x2b, 0x29, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72,
-  0x6f, 0x77, 0x41, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2f,
-  0x20, 0x72, 0x6f, 0x77, 0x73, 0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x3d,
+  0x20, 0x67, 0x65, 0x74, 0x5f, 0x67, 0x6c, 0x6f, 0x62, 0x61, 0x6c, 0x5f,
+  0x69, 0x64, 0x5f, 0x31, 0x3b, 0x20, 0x7b, 0x63, 0x70, 0x75, 0x5f, 0x6f,
+  0x6e, 0x6c, 0x79, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6e, 0x74,
+  0x20, 0x70, 0x6f, 0x73, 0x41, 0x2c, 0x20, 0x70, 0x6f, 0x73, 0x42, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x72,
+  0x52, 0x20, 0x3d, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66,
+  0x6c, 0x6f, 0x61, 0x74, 0x20, 0x69, 0x52, 0x20, 0x3d, 0x20, 0x30, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69,
+  0x6e, 0x74, 0x20, 0x70, 0x6f, 0x73, 0x52, 0x20, 0x3d, 0x20, 0x72, 0x6f,
+  0x77, 0x52, 0x20, 0x2a, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20,
+  0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x52,
+  0x20, 0x2a, 0x20, 0x32, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x66, 0x6f, 0x72, 0x20, 0x28, 0x69, 0x6e, 0x74, 0x20, 0x69, 0x20, 0x3d,
+  0x20, 0x30, 0x3b, 0x20, 0x69, 0x20, 0x3c, 0x20, 0x63, 0x6f, 0x6c, 0x73,
+  0x58, 0x3b, 0x20, 0x69, 0x2b, 0x2b, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x58,
+  0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x52, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69,
+  0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x58, 0x20, 0x3d, 0x20, 0x69, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x59, 0x20,
+  0x3d, 0x20, 0x69, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63,
+  0x6f, 0x6c, 0x59, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6c, 0x52, 0x3b, 0x0a,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x41, 0x20,
+  0x3d, 0x20, 0x72, 0x6f, 0x77, 0x58, 0x20, 0x2f, 0x20, 0x72, 0x6f, 0x77,
+  0x73, 0x49, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f,
+  0x6c, 0x41, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6c, 0x58, 0x20, 0x2f, 0x20,
+  0x63, 0x6f, 0x6c, 0x73, 0x49, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74,
+  0x20, 0x72, 0x6f, 0x77, 0x49, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x58,
+  0x20, 0x25, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x49, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x49, 0x20, 0x3d, 0x20, 0x63,
+  0x6f, 0x6c, 0x58, 0x20, 0x25, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x49, 0x3b,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x4a,
+  0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x59, 0x20, 0x2f, 0x20, 0x72, 0x6f,
+  0x77, 0x73, 0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63,
+  0x6f, 0x6c, 0x4a, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6c, 0x59, 0x20, 0x2f,
+  0x20, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e,
-  0x74, 0x20, 0x63, 0x6f, 0x6c, 0x41, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6c,
-  0x52, 0x20, 0x2f, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x3b, 0x0a, 0x20,
+  0x74, 0x20, 0x72, 0x6f, 0x77, 0x42, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77,
+  0x59, 0x20, 0x25, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x42, 0x3b, 0x0a, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
-  0x20, 0x69, 0x6e, 0x74, 0x20, 0x72, 0x6f, 0x77, 0x42, 0x20, 0x3d, 0x20,
-  0x72, 0x6f, 0x77, 0x52, 0x20, 0x25, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x42,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x42,
-  0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x25, 0x20, 0x63, 0x6f,
-  0x6c, 0x73, 0x42, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20,
-  0x70, 0x6f, 0x73, 0x41, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x41, 0x20,
-  0x2a, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x41, 0x20, 0x2a, 0x20, 0x32,
-  0x29, 0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x41, 0x20, 0x2a, 0x20,
-  0x32, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x70, 0x6f,
-  0x73, 0x42, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x42, 0x20, 0x2a, 0x20,
-  0x28, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20,
-  0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x42, 0x20, 0x2a, 0x20, 0x32, 0x29,
+  0x20, 0x69, 0x6e, 0x74, 0x20, 0x63, 0x6f, 0x6c, 0x42, 0x20, 0x3d, 0x20,
+  0x63, 0x6f, 0x6c, 0x59, 0x20, 0x25, 0x20, 0x63, 0x6f, 0x6c, 0x73, 0x42,
   0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x72,
-  0x41, 0x20, 0x3d, 0x20, 0x70, 0x74, 0x72, 0x41, 0x5b, 0x70, 0x6f, 0x73,
-  0x41, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20,
-  0x69, 0x41, 0x20, 0x3d, 0x20, 0x70, 0x74, 0x72, 0x41, 0x5b, 0x70, 0x6f,
-  0x73, 0x41, 0x20, 0x2b, 0x20, 0x31, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66,
-  0x6c, 0x6f, 0x61, 0x74, 0x20, 0x72, 0x42, 0x20, 0x3d, 0x20, 0x70, 0x74,
-  0x72, 0x42, 0x5b, 0x70, 0x6f, 0x73, 0x42, 0x5d, 0x3b, 0x0a, 0x20, 0x20,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x70, 0x6f, 0x73,
+  0x41, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x41, 0x20, 0x2a, 0x20, 0x28,
+  0x63, 0x6f, 0x6c, 0x73, 0x41, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b,
+  0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x41, 0x20, 0x2a, 0x20, 0x32, 0x29,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x20, 0x70, 0x6f, 0x73, 0x42,
+  0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x42, 0x20, 0x2a, 0x20, 0x28, 0x63,
+  0x6f, 0x6c, 0x73, 0x42, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b, 0x20,
+  0x28, 0x63, 0x6f, 0x6c, 0x73, 0x42, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x3b,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x72, 0x41,
+  0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x49, 0x20, 0x3d, 0x3d, 0x20, 0x63,
+  0x6f, 0x6c, 0x49, 0x20, 0x3f, 0x20, 0x70, 0x74, 0x72, 0x41, 0x5b, 0x70,
+  0x6f, 0x73, 0x41, 0x5d, 0x20, 0x3a, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x69, 0x42, 0x20, 0x3d, 0x20, 0x70,
-  0x74, 0x72, 0x42, 0x5b, 0x70, 0x6f, 0x73, 0x42, 0x20, 0x2b, 0x20, 0x31,
-  0x5d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x2f, 0x2f, 0x28, 0x72, 0x41, 0x20, 0x2b, 0x20, 0x69, 0x41, 0x29, 0x28,
-  0x72, 0x42, 0x20, 0x2b, 0x20, 0x69, 0x42, 0x29, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66,
-  0x6c, 0x6f, 0x61, 0x74, 0x20, 0x66, 0x69, 0x72, 0x73, 0x74, 0x20, 0x3d,
-  0x20, 0x72, 0x41, 0x20, 0x2a, 0x20, 0x72, 0x42, 0x3b, 0x0a, 0x20, 0x20,
+  0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x69, 0x41, 0x20, 0x3d, 0x20, 0x72,
+  0x6f, 0x77, 0x49, 0x20, 0x3d, 0x3d, 0x20, 0x63, 0x6f, 0x6c, 0x49, 0x20,
+  0x3f, 0x20, 0x70, 0x74, 0x72, 0x41, 0x5b, 0x70, 0x6f, 0x73, 0x41, 0x20,
+  0x2b, 0x20, 0x31, 0x5d, 0x20, 0x3a, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x6f, 0x75, 0x74, 0x65, 0x72, 0x20,
-  0x3d, 0x20, 0x72, 0x41, 0x20, 0x2a, 0x20, 0x69, 0x42, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
-  0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x69, 0x6e, 0x6e, 0x65, 0x72,
-  0x20, 0x3d, 0x20, 0x69, 0x41, 0x20, 0x2a, 0x20, 0x72, 0x42, 0x3b, 0x0a,
+  0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x72, 0x42, 0x20, 0x3d, 0x20, 0x72,
+  0x6f, 0x77, 0x4a, 0x20, 0x3d, 0x3d, 0x20, 0x63, 0x6f, 0x6c, 0x4a, 0x20,
+  0x3f, 0x20, 0x70, 0x74, 0x72, 0x42, 0x5b, 0x70, 0x6f, 0x73, 0x42, 0x5d,
+  0x20, 0x3a, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61,
+  0x74, 0x20, 0x69, 0x42, 0x20, 0x3d, 0x20, 0x72, 0x6f, 0x77, 0x4a, 0x20,
+  0x3d, 0x3d, 0x20, 0x63, 0x6f, 0x6c, 0x4a, 0x20, 0x3f, 0x20, 0x70, 0x74,
+  0x72, 0x42, 0x5b, 0x70, 0x6f, 0x73, 0x42, 0x20, 0x2b, 0x20, 0x31, 0x5d,
+  0x20, 0x3a, 0x20, 0x30, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x28, 0x72, 0x41, 0x20, 0x2b, 0x20, 0x69,
+  0x41, 0x29, 0x28, 0x72, 0x42, 0x20, 0x2b, 0x20, 0x69, 0x42, 0x29, 0x0a,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x6c, 0x61, 0x73, 0x74,
-  0x73, 0x20, 0x3d, 0x20, 0x69, 0x41, 0x20, 0x2a, 0x20, 0x69, 0x42, 0x3b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x74, 0x72,
-  0x52, 0x5b, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2a, 0x20, 0x28, 0x63, 0x6f,
-  0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b, 0x20, 0x28,
-  0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x5d, 0x20, 0x3d,
-  0x20, 0x66, 0x69, 0x72, 0x73, 0x74, 0x20, 0x2b, 0x20, 0x6c, 0x61, 0x73,
-  0x74, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x70, 0x74, 0x72, 0x52, 0x5b, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2a, 0x20,
-  0x28, 0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20,
-  0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29,
-  0x20, 0x2b, 0x20, 0x31, 0x5d, 0x20, 0x3d, 0x20, 0x6f, 0x75, 0x74, 0x65,
-  0x72, 0x20, 0x2b, 0x20, 0x69, 0x6e, 0x6e, 0x65, 0x72, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x00
+  0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x66, 0x69, 0x72, 0x73,
+  0x74, 0x20, 0x3d, 0x20, 0x72, 0x41, 0x20, 0x2a, 0x20, 0x72, 0x42, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x6f, 0x75, 0x74,
+  0x65, 0x72, 0x20, 0x3d, 0x20, 0x72, 0x41, 0x20, 0x2a, 0x20, 0x69, 0x42,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x69, 0x6e,
+  0x6e, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x69, 0x41, 0x20, 0x2a, 0x20, 0x72,
+  0x42, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x6c,
+  0x61, 0x73, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x69, 0x41, 0x20, 0x2a, 0x20,
+  0x69, 0x42, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x72, 0x52, 0x20, 0x2b, 0x3d, 0x20, 0x66, 0x69, 0x72, 0x73, 0x74,
+  0x20, 0x2b, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x73, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x52, 0x20, 0x2b, 0x3d, 0x20,
+  0x6f, 0x75, 0x74, 0x65, 0x72, 0x20, 0x2b, 0x20, 0x69, 0x6e, 0x6e, 0x65,
+  0x72, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x70, 0x74, 0x72, 0x52, 0x5b, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2a,
+  0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29,
+  0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2a, 0x20, 0x32,
+  0x29, 0x5d, 0x20, 0x3d, 0x20, 0x72, 0x52, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x70, 0x74, 0x72, 0x52, 0x5b, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2a,
+  0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29,
+  0x20, 0x2b, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2a, 0x20, 0x32,
+  0x29, 0x20, 0x2b, 0x20, 0x31, 0x5d, 0x20, 0x3d, 0x20, 0x69, 0x52, 0x3b,
+  0x0a, 0x7d, 0x00
 };
-unsigned int src_kernel_cl_len = 2551;
+unsigned int src_kernel_cl_len = 5079;

diff --git a/src/kernel_cpu.cl.c b/src/kernel_cpu.cl.c
index 434782e908360ec818b2d66a525e6517c1a68260..c304adc625831db0be8088696bc852a0f5cc69b2 100644 (file)
--- a/src/kernel_cpu.cl.c
+++ b/src/kernel_cpu.cl.c
@@ -15,26 +15,25 @@ void kernel_dot
     const int colsR = colsB;
     const int rowR = get_global_id_0; 
     const int colR = get_global_id_1; 
-    int posA, posB;
     float rR = 0;
     float iR = 0;
     const int posR = rowR * (colsR * 2) + (colR * 2);
 
     for (int i = 0; i < colsA; i++)
     {
-        int posA = rowR * (colsA * 2) + (i * 2);
-        int posB = i * (colsB * 2) + (colR * 2);
+        const int posA = rowR * (colsA * 2) + (i * 2);
+        const int posB = i * (colsB * 2) + (colR * 2);
 
-        float rA = ptrA[posA];
-        float iA = ptrA[posA + 1];
-        float rB = ptrB[posB];
-        float iB = ptrB[posB + 1];
+        const float rA = ptrA[posA];
+        const float iA = ptrA[posA + 1];
+        const float rB = ptrB[posB];
+        const float iB = ptrB[posB + 1];
 
         //(rA + iA)(rB + iB)
-        float first = rA * rB;
-        float outer = rA * iB;
-        float inner = iA * rB;
-        float lasts = iA * iB;
+        const float first = rA * rB;
+        const float outer = rA * iB;
+        const float inner = iA * rB;
+        const float lasts = iA * iB;
 
         rR += first + lasts;
         iR += outer + inner;
@@ -82,3 +81,80 @@ void kernel_knk
         ptrR[rowR * (colsR * 2) + (colR * 2) + 1] = outer + inner;
     }
 }
+
+void kernel_knk_2
+(
+    float* ptrR,
+    float* ptrA,
+    float* ptrB,
+    const int rowsA,
+    const int colsA,
+    const int rowsB,
+    const int colsB, 
+    const int get_global_id_0, 
+    const int get_global_id_1 
+)
+{
+    //This is based off the equality that
+    //  (A knk I) dot (J knk B) = A knk B
+    //  If I and J are identity matrices, then
+    //      we don't even need to compute them
+    //      to know their values at a given
+    //      index, meaning we just need to
+    //      carry out effectively a copy of
+    //      the dot product procedure.
+    const int rowsI = rowsB;
+    const int colsI = rowsB;
+    const int rowsJ = colsA;
+    const int colsJ = colsA;
+    const int rowsX = rowsA * rowsI;
+    const int colsX = colsA * colsI;
+    const int rowsY = rowsJ * rowsB;
+    const int colsY = colsJ * colsB;
+    const int rowsR = rowsX;
+    const int colsR = colsY;
+
+    const int rowR = get_global_id_0; 
+    const int colR = get_global_id_1; 
+    int posA, posB;
+    float rR = 0;
+    float iR = 0;
+    const int posR = rowR * (colsR * 2) + (colR * 2);
+
+    for (int i = 0; i < colsX; i++)
+    {
+        const int rowX = rowR;
+        const int colX = i;
+        const int rowY = i;
+        const int colY = colR;
+
+        const int rowA = rowX / rowsI;
+        const int colA = colX / colsI;
+        const int rowI = rowX % rowsI;
+        const int colI = colX % colsI;
+
+        const int rowJ = rowY / rowsB;
+        const int colJ = colY / colsB;
+        const int rowB = rowY % rowsB;
+        const int colB = colY % colsB;
+
+        const int posA = rowA * (colsA * 2) + (colsA * 2);
+        const int posB = rowB * (colsB * 2) + (colsB * 2);
+
+        const float rA = rowI == colI ? ptrA[posA] : 0;
+        const float iA = rowI == colI ? ptrA[posA + 1] : 0;
+        const float rB = rowJ == colJ ? ptrB[posB] : 0;
+        const float iB = rowJ == colJ ? ptrB[posB + 1] : 0;
+
+        //(rA + iA)(rB + iB)
+        const float first = rA * rB;
+        const float outer = rA * iB;
+        const float inner = iA * rB;
+        const float lasts = iA * iB;
+
+        rR += first + lasts;
+        iR += outer + inner;
+    }
+    ptrR[rowR * (colsR * 2) + (colR * 2)] = rR;
+    ptrR[rowR * (colsR * 2) + (colR * 2) + 1] = iR;
+}