Eigen の matmul 呼んでるだけ
define void @cluster_0__XlaCompiledKernel_true__XlaNumConstantArgs_0__XlaNumResourceArgs_0_.v6(i8* nocapture align 8 dereferenceable(8) %retval, i8* noalias %run_options, i8** noalias nocapture readonly %params, i8** noalias nocapture readonly %temps, i64* noalias nocapture readnone %prof_counters) local_unnamed_addr #0 {
entry:
%0 = getelementptr inbounds i8*, i8** %params, i64 1
%1 = bitcast i8** %0 to float**
%arg1.untyped12 = load float*, float** %1, align 8, !invariant.load !0, !dereferenceable !1, !align !2
%2 = bitcast i8** %params to float**
%arg0.untyped34 = load float*, float** %2, align 8, !invariant.load !0, !dereferenceable !1, !align !2
%3 = load i8*, i8** %temps, align 8, !invariant.load !0, !dereferenceable !1, !align !2
%4 = bitcast i8* %3 to float*
tail call void @__xla_cpu_runtime_EigenMatMulF32(i8* %run_options, float* %4, float* %arg0.untyped34, float* %arg1.untyped12, i64 1024, i64 1024, i64 1024, i32 0, i32 0)
%5 = bitcast i8* %retval to i8**
store i8* %3, i8** %5, align 8, !alias.scope !3, !noalias !6
ret void
}