Thu Mar 7 06:40:46 PM EST 2024

diff --git a/src/.kernel.tmp.1 b/src/.kernel.tmp.1
index 1da95816c44ededf150d07c2cba4676f78d3c39a..04c0834df7c3007bc6287331ef314cffb3fe37a4 100644 (file)
--- a/src/.kernel.tmp.1
+++ b/src/.kernel.tmp.1
@@ -135,12 +135,12 @@ __kernel void kernel_knk_2x2
                     ptrR[(rowR + 0) * (colsR * 2) + ((colR + 1) * 2) + 1] = outer + inner;
                 break;
                 case 2:
-                    ptrR[(rowR + 1) * (colsR * 2) + ((colR + 0) * 2)] = first + lasts;
-                    ptrR[(rowR + 1) * (colsR * 2) + ((colR + 0) * 2) + 1] = outer + inner;
+                    ptrR[(rowR + 0) * (colsR * 2) + ((colR + 0) * 2)] = first + lasts;
+                    ptrR[(rowR + 0) * (colsR * 2) + ((colR + 0) * 2) + 1] = outer + inner;
                 break;
                 case 3:
-                    ptrR[(rowR + 1) * (colsR * 2) + ((colR + 1) * 2)] = first + lasts;
-                    ptrR[(rowR + 1) * (colsR * 2) + ((colR + 1) * 2) + 1] = outer + inner;
+                    ptrR[(rowR + 0) * (colsR * 2) + ((colR + 1) * 2)] = first + lasts;
+                    ptrR[(rowR + 0) * (colsR * 2) + ((colR + 1) * 2) + 1] = outer + inner;
                 break;
             }
 

diff --git a/src/.kernel.tmp.2 b/src/.kernel.tmp.2
index 1e791cc72a16245bb1931b4a70dfb487507ea110..32bfc81c314627764b9396871127e92642f20c4b 100644 (file)
Binary files a/src/.kernel.tmp.2 and b/src/.kernel.tmp.2 differ

diff --git a/src/kernel.cl b/src/kernel.cl
index d11e287bbf726d21261deb8b2ae881aba2ecc76c..d9800779956ca17aa24242c8932ecead1517b673 100644 (file)
--- a/src/kernel.cl
+++ b/src/kernel.cl
@@ -143,12 +143,12 @@ __kernel void kernel_knk_2x2
                     ptrR[(rowR + 0) * (colsR * 2) + ((colR + 1) * 2) + 1] = outer + inner;
                 break;
                 case 2:
-                    ptrR[(rowR + 1) * (colsR * 2) + ((colR + 0) * 2)] = first + lasts;
-                    ptrR[(rowR + 1) * (colsR * 2) + ((colR + 0) * 2) + 1] = outer + inner;
+                    ptrR[(rowR + 0) * (colsR * 2) + ((colR + 0) * 2)] = first + lasts;
+                    ptrR[(rowR + 0) * (colsR * 2) + ((colR + 0) * 2) + 1] = outer + inner;
                 break;
                 case 3:
-                    ptrR[(rowR + 1) * (colsR * 2) + ((colR + 1) * 2)] = first + lasts;
-                    ptrR[(rowR + 1) * (colsR * 2) + ((colR + 1) * 2) + 1] = outer + inner;
+                    ptrR[(rowR + 0) * (colsR * 2) + ((colR + 1) * 2)] = first + lasts;
+                    ptrR[(rowR + 0) * (colsR * 2) + ((colR + 1) * 2) + 1] = outer + inner;
                 break;
             }
 

diff --git a/src/kernel_cpu.cl b/src/kernel_cpu.cl
index f8a12f23f1378239ee431e2b4277cf1db16a9c45..7376958621439305809d29e25c5c0e65f0a386a7 100644 (file)
--- a/src/kernel_cpu.cl
+++ b/src/kernel_cpu.cl
@@ -139,12 +139,12 @@ void kernel_knk_2x2
                     ptrR[(rowR + 0) * (colsR * 2) + ((colR + 1) * 2) + 1] = outer + inner;
                 break;
                 case 2:
-                    ptrR[(rowR + 1) * (colsR * 2) + ((colR + 0) * 2)] = first + lasts;
-                    ptrR[(rowR + 1) * (colsR * 2) + ((colR + 0) * 2) + 1] = outer + inner;
+                    ptrR[(rowR + 0) * (colsR * 2) + ((colR + 0) * 2)] = first + lasts;
+                    ptrR[(rowR + 0) * (colsR * 2) + ((colR + 0) * 2) + 1] = outer + inner;
                 break;
                 case 3:
-                    ptrR[(rowR + 1) * (colsR * 2) + ((colR + 1) * 2)] = first + lasts;
-                    ptrR[(rowR + 1) * (colsR * 2) + ((colR + 1) * 2) + 1] = outer + inner;
+                    ptrR[(rowR + 0) * (colsR * 2) + ((colR + 1) * 2)] = first + lasts;
+                    ptrR[(rowR + 0) * (colsR * 2) + ((colR + 1) * 2) + 1] = outer + inner;
                 break;
             }
 

diff --git a/src/kernel_gpu.cl b/src/kernel_gpu.cl
index e5e8787456c614b6b2a193478cbc256cafb82112..6345e20c53d4a934cd6ae887be8dfb96cf92c094 100644 (file)
--- a/src/kernel_gpu.cl
+++ b/src/kernel_gpu.cl
@@ -332,14 +332,14 @@ unsigned char kernel_gpu[] = {
   0x20, 0x63, 0x61, 0x73, 0x65, 0x20, 0x32, 0x3a, 0x0a, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x74, 0x72, 0x52, 0x5b, 0x28, 0x72,
-  0x6f, 0x77, 0x52, 0x20, 0x2b, 0x20, 0x31, 0x29, 0x20, 0x2a, 0x20, 0x28,
+  0x6f, 0x77, 0x52, 0x20, 0x2b, 0x20, 0x30, 0x29, 0x20, 0x2a, 0x20, 0x28,
   0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b,
   0x20, 0x28, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2b, 0x20, 0x30, 0x29,
   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, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x74, 0x72, 0x52,
-  0x5b, 0x28, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2b, 0x20, 0x31, 0x29, 0x20,
+  0x5b, 0x28, 0x72, 0x6f, 0x77, 0x52, 0x20, 0x2b, 0x20, 0x30, 0x29, 0x20,
   0x2a, 0x20, 0x28, 0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32,
   0x29, 0x20, 0x2b, 0x20, 0x28, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2b,
   0x20, 0x30, 0x29, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b, 0x20, 0x31,
@@ -351,14 +351,14 @@ unsigned char kernel_gpu[] = {
   0x73, 0x65, 0x20, 0x33, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x70, 0x74, 0x72, 0x52, 0x5b, 0x28, 0x72, 0x6f, 0x77, 0x52,
-  0x20, 0x2b, 0x20, 0x31, 0x29, 0x20, 0x2a, 0x20, 0x28, 0x63, 0x6f, 0x6c,
+  0x20, 0x2b, 0x20, 0x30, 0x29, 0x20, 0x2a, 0x20, 0x28, 0x63, 0x6f, 0x6c,
   0x73, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b, 0x20, 0x28, 0x28,
   0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2b, 0x20, 0x31, 0x29, 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, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x74, 0x72, 0x52, 0x5b, 0x28, 0x72,
-  0x6f, 0x77, 0x52, 0x20, 0x2b, 0x20, 0x31, 0x29, 0x20, 0x2a, 0x20, 0x28,
+  0x6f, 0x77, 0x52, 0x20, 0x2b, 0x20, 0x30, 0x29, 0x20, 0x2a, 0x20, 0x28,
   0x63, 0x6f, 0x6c, 0x73, 0x52, 0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b,
   0x20, 0x28, 0x28, 0x63, 0x6f, 0x6c, 0x52, 0x20, 0x2b, 0x20, 0x31, 0x29,
   0x20, 0x2a, 0x20, 0x32, 0x29, 0x20, 0x2b, 0x20, 0x31, 0x5d, 0x20, 0x3d,