Filter out fwd igemm kernel where split number is larger than input c…

…hannel (#2187)

* fix a bug in fwd where split is larger than input channel

* format code

* use proper variable

src/solver/conv_asm_implicit_gemm_gtc_fwd_nhwc.cpp

@@ -700,7 +700,8 @@ bool PerformanceConfigAsmImplicitGemmGTCFwdXdlopsNHWC::IsValid(
     if(!(tensor_a_thread_lengths[1] == 1 && tensor_b_thread_lengths[1] == 1))
     {
         // if both 1, indicate padded c support
-        if(((c >> gemm_k_global_split) / group) % gemm_k_per_block != 0)
+        if((c >> gemm_k_global_split == 0) ||
+           (((c >> gemm_k_global_split) / group) % gemm_k_per_block != 0))
             return false;
         // also, add this restriction to k, for vector write out
         if(problem.IsFp16())