[BLAS] fix row transpose arg

EnzymeAD · Sep 19, 2023 · 47d2f54 · 47d2f54
1 parent 3870f89
commit 47d2f54
Show file tree

Hide file tree

Showing 4 changed files with 7 additions and 7 deletions.
diff --git a/enzyme/test/Enzyme/ReverseMode/blas/gemv_f_c_split_blascpy.ll b/enzyme/test/Enzyme/ReverseMode/blas/gemv_f_c_split_blascpy.ll
@@ -187,7 +187,7 @@ entry:
 ; CHECK-NEXT:   store i64 1, i64* %byref.constant.int.1
 ; CHECK-NEXT:   %intcast.constant.int.1 = bitcast i64* %byref.constant.int.1 to i8*
 ; CHECK-NEXT:   call void @dgemv_64_(i8* %malloccall, i8* %m_p, i8* %n_p, i8* %fpcast.constant.fp.1.0, i8* %A, i8* %lda_p, i8* %20, i8* %intcast.int.one, i8* %fpcast.constant.fp.0.0, i8* %19, i8* %intcast.constant.int.1)
-; CHECK-NEXT:   %ld.row.trans = load i8, i8* %byref.transpose.transa
+; CHECK-NEXT:   %ld.row.trans = load i8, i8* %malloccall
 ; CHECK-DAG:    %[[c1:.+]] = icmp eq i8 %ld.row.trans, 110
 ; CHECK-DAG:    %[[c2:.+]] = icmp eq i8 %ld.row.trans, 78
 ; CHECK-NEXT:   %[[c3:.+]] = or i1 %[[c2]], %[[c1]]
@@ -203,7 +203,7 @@ entry:
 ; CHECK-NEXT:   store double 1.000000e+00, double* %byref.constant.fp.1.09
 ; CHECK-NEXT:   %fpcast.constant.fp.1.010 = bitcast double* %byref.constant.fp.1.09 to i8*
 ; CHECK-NEXT:   call void @dgemv_64_(i8* %byref.transpose.transa, i8* %m_p, i8* %n_p, i8* %alpha, i8* %A, i8* %lda_p, i8* %"y'", i8* %incy_p, i8* %fpcast.constant.fp.1.010, i8* %"x'", i8* %incx_p)
-; CHECK-NEXT:   %ld.row.trans11 = load i8, i8* %byref.transpose.transa
+; CHECK-NEXT:   %ld.row.trans11 = load i8, i8* %malloccall
 ; CHECK-DAG:   %[[r39:.+]] = icmp eq i8 %ld.row.trans11, 110
 ; CHECK-DAG:   %[[r40:.+]] = icmp eq i8 %ld.row.trans11, 78
 ; CHECK-NEXT:   %41 = or i1 %[[r40]], %[[r39]]
@@ -213,7 +213,7 @@ entry:
 ; CHECK-NEXT:   %45 = load double, double* %44
 ; CHECK-NEXT:   %46 = fadd fast double %45, %43
 ; CHECK-NEXT:   store double %46, double* %44
-; CHECK-NEXT:   %ld.row.trans12 = load i8, i8* %byref.transpose.transa
+; CHECK-NEXT:   %ld.row.trans12 = load i8, i8* %malloccall
 ; CHECK-DAG:   %[[r47:.+]] = icmp eq i8 %ld.row.trans12, 110
 ; CHECK-DAG:   %[[r48:.+]] = icmp eq i8 %ld.row.trans12, 78
 ; CHECK-NEXT:   %49 = or i1 %[[r48]], %[[r47]]

diff --git a/enzyme/test/Enzyme/ReverseMode/blas/gemv_f_c_split_blascpy_runtime_act.ll b/enzyme/test/Enzyme/ReverseMode/blas/gemv_f_c_split_blascpy_runtime_act.ll
@@ -178,7 +178,7 @@ entry:
 ; CHECK-NEXT:   br i1 %rt.inactive.beta, label %invertentry.beta.done, label %invertentry.beta.active
 
 ; CHECK: invertentry.beta.active:                          ; preds = %invertentry.A.done
-; CHECK-NEXT:   %ld.row.trans = load i8, i8* %byref.transpose.transa
+; CHECK-NEXT:   %ld.row.trans = load i8, i8* %malloccall
 ; CHECK-DAG:   %[[r39:.+]] = icmp eq i8 %ld.row.trans, 110
 ; CHECK-DAG:   %[[r40:.+]] = icmp eq i8 %ld.row.trans, 78
 ; CHECK-NEXT:   %[[r41:.+]] = or i1 %[[r40]], %[[r39]]
@@ -194,7 +194,7 @@ entry:
 ; CHECK-NEXT:   br i1 %rt.inactive.y, label %invertentry.y.done, label %invertentry.y.active
 
 ; CHECK: invertentry.y.active:                             ; preds = %invertentry.beta.done
-; CHECK-NEXT:   %ld.row.trans2 = load i8, i8* %byref.transpose.transa
+; CHECK-NEXT:   %ld.row.trans2 = load i8, i8* %malloccall
 ; CHECK-DAG:   %[[r47:.+]] = icmp eq i8 %ld.row.trans2, 110
 ; CHECK-DAG:   %[[r48:.+]] = icmp eq i8 %ld.row.trans2, 78
 ; CHECK-NEXT:   %[[r49:.+]] = or i1 %[[r48]], %[[r47]]

diff --git a/enzyme/test/Enzyme/ReverseMode/blas/gemv_f_c_split_memcpy.ll b/enzyme/test/Enzyme/ReverseMode/blas/gemv_f_c_split_memcpy.ll
@@ -194,7 +194,7 @@ entry:
 ; CHECK-NEXT:   store double 1.000000e+00, double* %byref.constant.fp.1.0
 ; CHECK-NEXT:   %fpcast.constant.fp.1.0 = bitcast double* %byref.constant.fp.1.0 to i8*
 ; CHECK-NEXT:   call void @dgemv_64_(i8* %byref.transpose.transa, i8* %m_p, i8* %n_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %"y'", i8* %incy_p, i8* %fpcast.constant.fp.1.0, i8* %"x'", i8* %incx_p)
-; CHECK-NEXT:   %ld.row.trans = load i8, i8* %byref.transpose.transa
+; CHECK-NEXT:   %ld.row.trans = load i8, i8* %malloccall
 ; CHECK-DAG:    %[[r2:.+]] = icmp eq i8 %ld.row.trans, 110
 ; CHECK-DAG:    %[[r3:.+]] = icmp eq i8 %ld.row.trans, 78
 ; CHECK-NEXT:   %26 = or i1 %[[r3]], %[[r2]]

diff --git a/enzyme/tools/enzyme-tblgen/blas-tblgen.cpp b/enzyme/tools/enzyme-tblgen/blas-tblgen.cpp
@@ -1042,7 +1042,7 @@ void rev_call_arg(StringRef argName, DagInit *ruleDag, Rule &rule,
       auto tname = Dag->getArgNameStr(0);
       auto rname = Dag->getArgNameStr(1);
       auto cname = Dag->getArgNameStr(2);
-      os << "get_blas_row(Builder2, arg_transposed_" << tname << ", arg_"
+      os << "get_blas_row(Builder2, arg_" << tname << ", arg_"
          << rname << ", arg_" << cname << ", byRef)";
     } else if (Def->isSubClassOf("MagicInst") && Def->getName() == "ld") {
       assert(Dag->getNumArgs() == 5);