From f1ac3622e991d976005f70823cb40612898ff3a8 Mon Sep 17 00:00:00 2001 From: miha-q <> Date: Mon, 4 Mar 2024 00:53:16 -0500 Subject: [PATCH] Mon Mar 4 12:53:16 AM EST 2024 --- src/gpu/gpu.c | 12 ++++++------ src/gpu/kernel.cl | 4 ++-- 2 files changed, 8 insertions(+), 8 deletions(-) diff --git a/src/gpu/gpu.c b/src/gpu/gpu.c index cf8a318..a9e34ab 100644 --- 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 da23464..90a9768 100644 --- 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 ) -- 2.39.5