Mon Mar 4 12:53:16 AM EST 2024

diff --git a/src/gpu/gpu.c b/src/gpu/gpu.c
index cf8a318f1951828b6a6c2b34b0a109b201a0f969..a9e34ab830d8c8011eab2b5765f4d6a2ca44defa 100644 (file)
--- a/src/gpu/gpu.c
+++ b/src/gpu/gpu.c
@@ -199,12 +199,12 @@ void GPU_knk(float* ptrR, int rowsR, int colsR, float* ptrA, int rowsA, int cols
        //Setup kernel
        cl_kernel kernel = clCreateKernel(program, "gpu_knk", &err); gpuerr(clCreateKernel);
        err = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void*)&memR); gpuerr(clSetKernelArg);
-       err = clSetKernelArg(kernel, 1, sizeof(int), &rowsR); gpuerr(clSetKernelArg);
-       err = clSetKernelArg(kernel, 2, sizeof(int), &colsR); gpuerr(clSetKernelArg);
-       err = clSetKernelArg(kernel, 3, sizeof(cl_mem), (void*)&memA); gpuerr(clSetKernelArg);
-       err = clSetKernelArg(kernel, 4, sizeof(int), &rowsA); gpuerr(clSetKernelArg);
-       err = clSetKernelArg(kernel, 5, sizeof(int), &colsA); gpuerr(clSetKernelArg);
-       err = clSetKernelArg(kernel, 6, sizeof(cl_mem), (void*)&memB); gpuerr(clSetKernelArg);
+       err = clSetKernelArg(kernel, 1, sizeof(cl_mem), (void*)&memA); gpuerr(clSetKernelArg);
+       err = clSetKernelArg(kernel, 2, sizeof(cl_mem), (void*)&memB); gpuerr(clSetKernelArg);
+       err = clSetKernelArg(kernel, 3, sizeof(int), &rowsR); gpuerr(clSetKernelArg);
+       err = clSetKernelArg(kernel, 4, sizeof(int), &colsR); gpuerr(clSetKernelArg);
+       err = clSetKernelArg(kernel, 5, sizeof(int), &rowsA); gpuerr(clSetKernelArg);
+       err = clSetKernelArg(kernel, 6, sizeof(int), &colsA); gpuerr(clSetKernelArg);
        err = clSetKernelArg(kernel, 7, sizeof(int), &rowsB); gpuerr(clSetKernelArg);
        err = clSetKernelArg(kernel, 8, sizeof(int), &colsB); gpuerr(clSetKernelArg);
 

diff --git a/src/gpu/kernel.cl b/src/gpu/kernel.cl
index da234642b708f7b781b9eaab5b3f3972cbe2f176..90a976822fc95155080c8a138640d52b358e7cb2 100644 (file)
--- a/src/gpu/kernel.cl
+++ b/src/gpu/kernel.cl
@@ -31,12 +31,12 @@ __kernel void gpu_mmul
 __kernel void gpu_knk
 (
     __global float* ptrR,
+    __global float* ptrA,
+    __global float* ptrB,
     const int rowsR,
     const int colsR,
-    __global float* ptrA,
     const int rowsA,
     const int colsA,
-    __global float* ptrB,
     const int rowsB,
     const int colsB
 )