Thu Mar 7 06:12:39 PM EST 2024

diff --git a/src/QAnsel.c b/src/QAnsel.c
index b024b7e18abc9baec5e9fc46bd469094b05ab91e..5f4ef1310ebbe883f95f3d42001071634643ea9e 100644 (file)
--- a/src/QAnsel.c
+++ b/src/QAnsel.c
@@ -228,7 +228,7 @@ void qansel_instruction(cpx_mtx_t* stateVector, unsigned char qubitCount, QInstr
                us2 = get_time();
                printf("\tMetal2x2_R: %lu\n", us2 - us1);
 
-               if (filter.rows / 128 == 0)
+               if (filter.rows / 1024 == 0)
                {
                        printf("\tMetal2x2_Rx4: Invalid\n");
                }

diff --git a/src/complex.c b/src/complex.c
index 3545acea31ec70edce1e7ec9440b16264eed9fe7..ccf5cb5cc1b80b66d5167ec05453bc8dfe2d86a7 100644 (file)
--- a/src/complex.c
+++ b/src/complex.c
@@ -172,7 +172,7 @@ void cpx_mtx_knk_2x2_Rx4(float* ptrR, float* ptrA, float* ptrB, int rowsA, int c
 {
        int rowsR = rowsA * rowsB;
        int colsR = colsA * colsB;
-    for (int i = 0; i < rowsR / (2 * 128); i++)
+    for (int i = 0; i < rowsR / (2 * 1024); i++)
     {
                        kernel_knk_2x2_Rx4(ptrR, ptrA, rowsA, colsA, ptrB[0], ptrB[1], ptrB[2], ptrB[3], ptrB[4], ptrB[5], ptrB[6], ptrB[7], i);
     }

diff --git a/src/kernel.cl b/src/kernel.cl
index 34f291d0047cad9b148922003d9f54f976fda8b5..12646edf017616e28455a8bd7ad891ba1af48b86 100644 (file)
--- a/src/kernel.cl
+++ b/src/kernel.cl
@@ -248,10 +248,10 @@ __kernel void kernel_knk_2x2_Rx4
 {
     const int rowsR = rowsA * 2;
     const int colsR = colsA * 2;
-    const int block = get_global_id(0) * 2 * 128; //{gpu_only}
-    const int block = get_global_id_0 * 2 * 128; //{cpu_only}
+    const int block = get_global_id(0) * 2 * 1024; //{gpu_only}
+    const int block = get_global_id_0 * 2 * 1024; //{cpu_only}
 
-    for (int rowR = block; rowR < block + 2 * 128; rowR += 2)
+    for (int rowR = block; rowR < block + 2 * 1024; rowR += 2)
     {
         for (int colR = 0; colR < colsR; colR += 2)
         {