Make aie and aie2 directories for aievec_tests, replace uses of xches…

…scc with xchesscc_wrapper (Xilinx#1841)
fifield · Oct 15, 2024 · 4b1f36b · 4b1f36b
1 parent c0b15bb
commit 4b1f36b
Show file tree

Hide file tree

Showing 451 changed files with 219 additions and 219 deletions.
diff --git a/...ized_memref/conv2d_uij_f32_unbounded.mlir → ...ized_memref/conv2d_uij_f32_unbounded.mlir b/...ized_memref/conv2d_uij_f32_unbounded.mlir → ...ized_memref/conv2d_uij_f32_unbounded.mlir
@@ -2,7 +2,7 @@
 // RUN: aie-opt %s -affine-super-vectorize="virtual-vector-size=8" --aie-vectorize | aie-translate --aievec-to-cpp -o gen.cc
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. -c %S/kernel.cc -o kernel.o
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. %S/float.cc work/kernel.o
-// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../profiling.tcl ./work/a.out"
+// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../../profiling.tcl ./work/a.out"
 
 func.func @conv2d (%A: memref<?x?xf32>, %B: memref<?xf32>, %C: memref<?x?xf32>) {
     %c0 = arith.constant 0 : index

diff --git a/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt b/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt
diff --git a/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt b/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt
diff --git a/...tests/float_dynamic_sized_memref/float.cc → ...s/aie/float_dynamic_sized_memref/float.cc b/...tests/float_dynamic_sized_memref/float.cc → ...s/aie/float_dynamic_sized_memref/float.cc
diff --git a/..._tests/float_dynamic_sized_memref/float.h → ...ts/aie/float_dynamic_sized_memref/float.h b/..._tests/float_dynamic_sized_memref/float.h → ...ts/aie/float_dynamic_sized_memref/float.h
diff --git a/...ests/float_dynamic_sized_memref/kernel.cc → .../aie/float_dynamic_sized_memref/kernel.cc b/...ests/float_dynamic_sized_memref/kernel.cc → .../aie/float_dynamic_sized_memref/kernel.cc
diff --git a/...s/float_dynamic_sized_memref/reference.cc → ...e/float_dynamic_sized_memref/reference.cc b/...s/float_dynamic_sized_memref/reference.cc → ...e/float_dynamic_sized_memref/reference.cc
diff --git a/..._tests/float_static_sized_memref/conv2d.c → ...ts/aie/float_static_sized_memref/conv2d.c b/..._tests/float_static_sized_memref/conv2d.c → ...ts/aie/float_static_sized_memref/conv2d.c
diff --git a/...c_sized_memref/conv2d_uij_f32_noinit.mlir → ...c_sized_memref/conv2d_uij_f32_noinit.mlir b/...c_sized_memref/conv2d_uij_f32_noinit.mlir → ...c_sized_memref/conv2d_uij_f32_noinit.mlir
@@ -2,7 +2,7 @@
 // RUN: aie-opt %s -affine-super-vectorize="virtual-vector-size=8" --aie-vectorize | aie-translate --aievec-to-cpp -o gen.cc
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. -c %S/kernel.cc -o kernel.o
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. %S/float.cc work/kernel.o
-// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../profiling.tcl ./work/a.out"
+// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../../profiling.tcl ./work/a.out"
 
 func.func @conv2d (%A: memref<18x272xf32>, %B: memref<9xf32>, %C: memref<16x256xf32>) {
     affine.for %arg3 = 0 to 16 {

diff --git a/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt b/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt
diff --git a/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt b/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt
diff --git a/..._tests/float_static_sized_memref/float.cc → ...ts/aie/float_static_sized_memref/float.cc b/..._tests/float_static_sized_memref/float.cc → ...ts/aie/float_static_sized_memref/float.cc
diff --git a/...c_tests/float_static_sized_memref/float.h → ...sts/aie/float_static_sized_memref/float.h b/...c_tests/float_static_sized_memref/float.h → ...sts/aie/float_static_sized_memref/float.h
diff --git a/...tests/float_static_sized_memref/kernel.cc → ...s/aie/float_static_sized_memref/kernel.cc b/...tests/float_static_sized_memref/kernel.cc → ...s/aie/float_static_sized_memref/kernel.cc
diff --git a/...ts/float_static_sized_memref/reference.cc → ...ie/float_static_sized_memref/reference.cc b/...ts/float_static_sized_memref/reference.cc → ...ie/float_static_sized_memref/reference.cc
diff --git a/...ized_memref/conv2d_uij_i16_unbounded.mlir → ...ized_memref/conv2d_uij_i16_unbounded.mlir b/...ized_memref/conv2d_uij_i16_unbounded.mlir → ...ized_memref/conv2d_uij_i16_unbounded.mlir
@@ -6,7 +6,7 @@
 // RUN: aie-opt %s -affine-super-vectorize="virtual-vector-size=16" --aie-vectorize="shift=10 zero-offset=4" | aie-translate --aievec-to-cpp -o gen.cc
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. -c %S/kernel.cc -o kernel.o
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. %S/i16xi16.cc work/kernel.o
-// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../profiling.tcl ./work/a.out"
+// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../../profiling.tcl ./work/a.out"
 
 func.func @conv2d(%A: memref<?x?xi16>, %B: memref<?xi16>, %C: memref<?x?xi16>) {
     %c0 = arith.constant 0 : index

diff --git a/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt b/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt
diff --git a/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt b/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt
diff --git a/...s/i16xi16_dynamic_sized_memref/i16xi16.cc → ...e/i16xi16_dynamic_sized_memref/i16xi16.cc b/...s/i16xi16_dynamic_sized_memref/i16xi16.cc → ...e/i16xi16_dynamic_sized_memref/i16xi16.cc
diff --git a/...ts/i16xi16_dynamic_sized_memref/i16xi16.h → ...ie/i16xi16_dynamic_sized_memref/i16xi16.h b/...ts/i16xi16_dynamic_sized_memref/i16xi16.h → ...ie/i16xi16_dynamic_sized_memref/i16xi16.h
diff --git a/...ts/i16xi16_dynamic_sized_memref/kernel.cc → ...ie/i16xi16_dynamic_sized_memref/kernel.cc b/...ts/i16xi16_dynamic_sized_memref/kernel.cc → ...ie/i16xi16_dynamic_sized_memref/kernel.cc
diff --git a/...i16xi16_dynamic_sized_memref/reference.cc → ...i16xi16_dynamic_sized_memref/reference.cc b/...i16xi16_dynamic_sized_memref/reference.cc → ...i16xi16_dynamic_sized_memref/reference.cc
diff --git a/...c_sized_memref/conv2d_uij_i16_noinit.mlir → ...c_sized_memref/conv2d_uij_i16_noinit.mlir b/...c_sized_memref/conv2d_uij_i16_noinit.mlir → ...c_sized_memref/conv2d_uij_i16_noinit.mlir
@@ -2,7 +2,7 @@
 // RUN: aie-opt %s -affine-super-vectorize="virtual-vector-size=16" --aie-vectorize="shift=10 zero-offset=4" | aie-translate --aievec-to-cpp -o gen.cc
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. -c %S/kernel.cc -o kernel.o
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. %S/i16xi16.cc work/kernel.o
-// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../profiling.tcl ./work/a.out"
+// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../../profiling.tcl ./work/a.out"
 
 func.func @conv2d (%A: memref<18x288xi16>, %B: memref<9xi16>, %C: memref<16x256xi16>) {
     affine.for %arg3 = 0 to 16 {

diff --git a/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt b/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt
diff --git a/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt b/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt
diff --git a/...ts/i16xi16_static_sized_memref/i16xi16.cc → ...ie/i16xi16_static_sized_memref/i16xi16.cc b/...ts/i16xi16_static_sized_memref/i16xi16.cc → ...ie/i16xi16_static_sized_memref/i16xi16.cc
diff --git a/...sts/i16xi16_static_sized_memref/i16xi16.h → ...aie/i16xi16_static_sized_memref/i16xi16.h b/...sts/i16xi16_static_sized_memref/i16xi16.h → ...aie/i16xi16_static_sized_memref/i16xi16.h
diff --git a/...sts/i16xi16_static_sized_memref/kernel.cc → ...aie/i16xi16_static_sized_memref/kernel.cc b/...sts/i16xi16_static_sized_memref/kernel.cc → ...aie/i16xi16_static_sized_memref/kernel.cc
diff --git a/.../i16xi16_static_sized_memref/reference.cc → .../i16xi16_static_sized_memref/reference.cc b/.../i16xi16_static_sized_memref/reference.cc → .../i16xi16_static_sized_memref/reference.cc
diff --git a/...ized_memref/conv2d_uij_i32_unbounded.mlir → ...ized_memref/conv2d_uij_i32_unbounded.mlir b/...ized_memref/conv2d_uij_i32_unbounded.mlir → ...ized_memref/conv2d_uij_i32_unbounded.mlir
@@ -2,7 +2,7 @@
 // RUN: aie-opt %s -affine-super-vectorize="virtual-vector-size=8" --aie-vectorize="shift=10" | aie-translate --aievec-to-cpp -o gen.cc
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. -c %S/kernel.cc -o kernel.o
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. %S/i32xi32.cc work/kernel.o
-// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../profiling.tcl ./work/a.out"
+// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../../profiling.tcl ./work/a.out"
 
 func.func @conv2d (%A: memref<?x?xi32>, %B: memref<?xi32>, %C: memref<?x?xi32>) {
     %c0 = arith.constant 0 : index

diff --git a/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt b/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt
diff --git a/...-ml/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt b/...-ml/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt
diff --git a/...s/i32xi32_dynamic_sized_memref/i32xi32.cc → ...e/i32xi32_dynamic_sized_memref/i32xi32.cc b/...s/i32xi32_dynamic_sized_memref/i32xi32.cc → ...e/i32xi32_dynamic_sized_memref/i32xi32.cc
diff --git a/...ts/i32xi32_dynamic_sized_memref/i32xi32.h → ...ie/i32xi32_dynamic_sized_memref/i32xi32.h b/...ts/i32xi32_dynamic_sized_memref/i32xi32.h → ...ie/i32xi32_dynamic_sized_memref/i32xi32.h
diff --git a/...ts/i32xi32_dynamic_sized_memref/kernel.cc → ...ie/i32xi32_dynamic_sized_memref/kernel.cc b/...ts/i32xi32_dynamic_sized_memref/kernel.cc → ...ie/i32xi32_dynamic_sized_memref/kernel.cc
diff --git a/...i32xi32_dynamic_sized_memref/reference.cc → ...i32xi32_dynamic_sized_memref/reference.cc b/...i32xi32_dynamic_sized_memref/reference.cc → ...i32xi32_dynamic_sized_memref/reference.cc
diff --git a/...c_sized_memref/conv2d_uij_i32_noinit.mlir → ...c_sized_memref/conv2d_uij_i32_noinit.mlir b/...c_sized_memref/conv2d_uij_i32_noinit.mlir → ...c_sized_memref/conv2d_uij_i32_noinit.mlir
@@ -2,7 +2,7 @@
 // RUN: aie-opt %s -affine-super-vectorize="virtual-vector-size=8" --aie-vectorize="shift=10" | aie-translate --aievec-to-cpp -o gen.cc
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. -c %S/kernel.cc -o kernel.o
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. %S/i32xi32.cc work/kernel.o
-// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../profiling.tcl ./work/a.out"
+// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../../profiling.tcl ./work/a.out"
 
 func.func @conv2d (%A: memref<18x272xi32>, %B: memref<9xi32>, %C: memref<16x256xi32>) {
     affine.for %arg3 = 0 to 16 {

diff --git a/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt b/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt
diff --git a/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt b/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt
diff --git a/...ts/i32xi32_static_sized_memref/i32xi32.cc → ...ie/i32xi32_static_sized_memref/i32xi32.cc b/...ts/i32xi32_static_sized_memref/i32xi32.cc → ...ie/i32xi32_static_sized_memref/i32xi32.cc
diff --git a/...sts/i32xi32_static_sized_memref/i32xi32.h → ...aie/i32xi32_static_sized_memref/i32xi32.h b/...sts/i32xi32_static_sized_memref/i32xi32.h → ...aie/i32xi32_static_sized_memref/i32xi32.h
diff --git a/...sts/i32xi32_static_sized_memref/kernel.cc → ...aie/i32xi32_static_sized_memref/kernel.cc b/...sts/i32xi32_static_sized_memref/kernel.cc → ...aie/i32xi32_static_sized_memref/kernel.cc
diff --git a/.../i32xi32_static_sized_memref/reference.cc → .../i32xi32_static_sized_memref/reference.cc b/.../i32xi32_static_sized_memref/reference.cc → .../i32xi32_static_sized_memref/reference.cc
diff --git a/...sized_memref/conv2d_uij_i8_unbounded.mlir → ...sized_memref/conv2d_uij_i8_unbounded.mlir b/...sized_memref/conv2d_uij_i8_unbounded.mlir → ...sized_memref/conv2d_uij_i8_unbounded.mlir
@@ -2,7 +2,7 @@
 // RUN: aie-opt %s -affine-super-vectorize="virtual-vector-size=16" --aie-vectorize="shift=0 dup-factor=2" | aie-translate --aievec-to-cpp -o gen.cc
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. -c %S/kernel.cc -o kernel.o
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. %S/i8xi8.cc work/kernel.o
-// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../profiling.tcl ./work/a.out"
+// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../../profiling.tcl ./work/a.out"
 
 func.func @conv2d (%A: memref<?x?xi8>, %B: memref<?xi8>, %C: memref<?x?xi8>) {
     %c0 = arith.constant 0 : index

diff --git a/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt b/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt
diff --git a/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt b/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt
diff --git a/...tests/i8xi8_dynamic_sized_memref/i8xi8.cc → ...s/aie/i8xi8_dynamic_sized_memref/i8xi8.cc b/...tests/i8xi8_dynamic_sized_memref/i8xi8.cc → ...s/aie/i8xi8_dynamic_sized_memref/i8xi8.cc
diff --git a/..._tests/i8xi8_dynamic_sized_memref/i8xi8.h → ...ts/aie/i8xi8_dynamic_sized_memref/i8xi8.h b/..._tests/i8xi8_dynamic_sized_memref/i8xi8.h → ...ts/aie/i8xi8_dynamic_sized_memref/i8xi8.h
diff --git a/...ests/i8xi8_dynamic_sized_memref/kernel.cc → .../aie/i8xi8_dynamic_sized_memref/kernel.cc b/...ests/i8xi8_dynamic_sized_memref/kernel.cc → .../aie/i8xi8_dynamic_sized_memref/kernel.cc
diff --git a/...ic_sized_memref/conv2d_uij_i8_noinit.mlir → ...ic_sized_memref/conv2d_uij_i8_noinit.mlir b/...ic_sized_memref/conv2d_uij_i8_noinit.mlir → ...ic_sized_memref/conv2d_uij_i8_noinit.mlir
@@ -2,7 +2,7 @@
 // RUN: aie-opt %s -affine-super-vectorize="virtual-vector-size=16" --aie-vectorize="shift=0 dup-factor=2" | aie-translate --aievec-to-cpp -o gen.cc
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. -c %S/kernel.cc -o kernel.o
 // RUN: xchesscc_wrapper aie -f -g +s +w work +o work -I%S -I. %S/i8xi8.cc work/kernel.o
-// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../profiling.tcl ./work/a.out"
+// RUN: cp -r %S/data . && xca_udm_dbg -qf -T -P %aietools/data/versal_prod/lib -t "%S/../../profiling.tcl ./work/a.out"
 
 func.func @conv2d (%A: memref<18x288xi8>, %B: memref<48xi8>, %C: memref<16x256xi8>) {
     affine.for %arg3 = 0 to 16 {

diff --git a/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt b/.../kernelAndInImage_256x16_k3_gaussblur.txt → .../kernelAndInImage_256x16_k3_gaussblur.txt
diff --git a/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt b/...ref/data/refImage_256x16_k3_gaussblur.txt → ...ref/data/refImage_256x16_k3_gaussblur.txt
diff --git a/..._tests/i8xi8_static_sized_memref/i8xi8.cc → ...ts/aie/i8xi8_static_sized_memref/i8xi8.cc b/..._tests/i8xi8_static_sized_memref/i8xi8.cc → ...ts/aie/i8xi8_static_sized_memref/i8xi8.cc
diff --git a/...c_tests/i8xi8_static_sized_memref/i8xi8.h → ...sts/aie/i8xi8_static_sized_memref/i8xi8.h b/...c_tests/i8xi8_static_sized_memref/i8xi8.h → ...sts/aie/i8xi8_static_sized_memref/i8xi8.h
diff --git a/...tests/i8xi8_static_sized_memref/kernel.cc → ...s/aie/i8xi8_static_sized_memref/kernel.cc b/...tests/i8xi8_static_sized_memref/kernel.cc → ...s/aie/i8xi8_static_sized_memref/kernel.cc
diff --git a/...ests/bf16_add_reduce/bf16_add_reduce.mlir → ...aie2/bf16_add_reduce/bf16_add_reduce.mlir b/...ests/bf16_add_reduce/bf16_add_reduce.mlir → ...aie2/bf16_add_reduce/bf16_add_reduce.mlir
@@ -2,7 +2,7 @@
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I. %S/testbench.cc work/dut.o
 // RUN: mkdir -p data
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 

diff --git a/...ts/aievec_tests/bf16_add_reduce/defines.h → ...evec_tests/aie2/bf16_add_reduce/defines.h b/...ts/aievec_tests/bf16_add_reduce/defines.h → ...evec_tests/aie2/bf16_add_reduce/defines.h
diff --git a/...tests/aievec_tests/bf16_add_reduce/dut.cc → .../aievec_tests/aie2/bf16_add_reduce/dut.cc b/...tests/aievec_tests/bf16_add_reduce/dut.cc → .../aievec_tests/aie2/bf16_add_reduce/dut.cc
diff --git a/...aievec_tests/bf16_add_reduce/testbench.cc → ...c_tests/aie2/bf16_add_reduce/testbench.cc b/...aievec_tests/bf16_add_reduce/testbench.cc → ...c_tests/aie2/bf16_add_reduce/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...tests/bf16_exp_lut/bf16_exp_lut-llvm.mlir → .../aie2/bf16_exp_lut/bf16_exp_lut-llvm.mlir b/...tests/bf16_exp_lut/bf16_exp_lut-llvm.mlir → .../aie2/bf16_exp_lut/bf16_exp_lut-llvm.mlir
@@ -12,7 +12,7 @@
 // RUN: llvm-link -S lut_constants.ll dut_functions.ll -o dut.ll
 // RUN: %PEANO_INSTALL_DIR/bin/clang %clang_aie2_args -c dut.ll -o dut.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -DTO_LLVM -D__AIEARCH__=20 -D__AIENGINE__ -I. %S/testbench.cc dut.o 
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 

diff --git a/...evec_tests/bf16_exp_lut/bf16_exp_lut.mlir → ...tests/aie2/bf16_exp_lut/bf16_exp_lut.mlir b/...evec_tests/bf16_exp_lut/bf16_exp_lut.mlir → ...tests/aie2/bf16_exp_lut/bf16_exp_lut.mlir
@@ -7,7 +7,7 @@
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIEARCH__=20 -D__AIENGINE__ -I. -c %aie_runtime_lib%/AIE2/lut_based_ops.cpp -o lut_based_ops.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIEARCH__=20 -D__AIENGINE__ -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIEARCH__=20 -D__AIENGINE__ -I. %S/testbench.cc work/dut.o work/lut_based_ops.o
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 

diff --git a/...tests/aievec_tests/bf16_exp_lut/defines.h → .../aievec_tests/aie2/bf16_exp_lut/defines.h b/...tests/aievec_tests/bf16_exp_lut/defines.h → .../aievec_tests/aie2/bf16_exp_lut/defines.h
diff --git a/...it_tests/aievec_tests/bf16_exp_lut/dut.cc → ...sts/aievec_tests/aie2/bf16_exp_lut/dut.cc b/...it_tests/aievec_tests/bf16_exp_lut/dut.cc → ...sts/aievec_tests/aie2/bf16_exp_lut/dut.cc
diff --git a/...s/aievec_tests/bf16_exp_lut/dut_simple.cc → ...vec_tests/aie2/bf16_exp_lut/dut_simple.cc b/...s/aievec_tests/bf16_exp_lut/dut_simple.cc → ...vec_tests/aie2/bf16_exp_lut/dut_simple.cc
diff --git a/...ts/aievec_tests/bf16_exp_lut/testbench.cc → ...evec_tests/aie2/bf16_exp_lut/testbench.cc b/...ts/aievec_tests/bf16_exp_lut/testbench.cc → ...evec_tests/aie2/bf16_exp_lut/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...oat_add_reduce/bf16_float_add_reduce.mlir → ...oat_add_reduce/bf16_float_add_reduce.mlir b/...oat_add_reduce/bf16_float_add_reduce.mlir → ...oat_add_reduce/bf16_float_add_reduce.mlir
@@ -2,7 +2,7 @@
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I. %S/testbench.cc work/dut.o
 // RUN: mkdir -p data
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 

diff --git a/...vec_tests/bf16_float_add_reduce/defines.h → ...ests/aie2/bf16_float_add_reduce/defines.h b/...vec_tests/bf16_float_add_reduce/defines.h → ...ests/aie2/bf16_float_add_reduce/defines.h
diff --git a/...aievec_tests/bf16_float_add_reduce/dut.cc → ...c_tests/aie2/bf16_float_add_reduce/dut.cc b/...aievec_tests/bf16_float_add_reduce/dut.cc → ...c_tests/aie2/bf16_float_add_reduce/dut.cc
diff --git a/..._tests/bf16_float_add_reduce/testbench.cc → ...s/aie2/bf16_float_add_reduce/testbench.cc b/..._tests/bf16_float_add_reduce/testbench.cc → ...s/aie2/bf16_float_add_reduce/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...evec_tests/bf16_inv_lut/bf16_inv_lut.mlir → ...tests/aie2/bf16_inv_lut/bf16_inv_lut.mlir b/...evec_tests/bf16_inv_lut/bf16_inv_lut.mlir → ...tests/aie2/bf16_inv_lut/bf16_inv_lut.mlir
@@ -7,7 +7,7 @@
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIEARCH__=20 -D__AIENGINE__ -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIEARCH__=20 -D__AIENGINE__ -I. %S/testbench.cc work/dut.o work/lut_based_ops.o
 // RUN: mkdir -p data
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 

diff --git a/...tests/aievec_tests/bf16_inv_lut/defines.h → .../aievec_tests/aie2/bf16_inv_lut/defines.h b/...tests/aievec_tests/bf16_inv_lut/defines.h → .../aievec_tests/aie2/bf16_inv_lut/defines.h
diff --git a/...it_tests/aievec_tests/bf16_inv_lut/dut.cc → ...sts/aievec_tests/aie2/bf16_inv_lut/dut.cc b/...it_tests/aievec_tests/bf16_inv_lut/dut.cc → ...sts/aievec_tests/aie2/bf16_inv_lut/dut.cc
diff --git a/...ts/aievec_tests/bf16_inv_lut/testbench.cc → ...evec_tests/aie2/bf16_inv_lut/testbench.cc b/...ts/aievec_tests/bf16_inv_lut/testbench.cc → ...evec_tests/aie2/bf16_inv_lut/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...bf16_max_reduce/bf16_max_reduce-llvm.mlir → ...bf16_max_reduce/bf16_max_reduce-llvm.mlir b/...bf16_max_reduce/bf16_max_reduce-llvm.mlir → ...bf16_max_reduce/bf16_max_reduce-llvm.mlir
@@ -8,7 +8,7 @@
 // RUN: aie-translate llvmir.mlir %llvmir-to-ll% -o dut.ll
 // RUN: %PEANO_INSTALL_DIR/bin/clang %clang_aie2_args -c dut.ll -o dut.o
 // RUN: xchesscc_wrapper %xchesscc_aie2_args -DTO_LLVM +w work +o work -I%S -I. %S/testbench.cc dut.o
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 

diff --git a/...ests/bf16_max_reduce/bf16_max_reduce.mlir → ...aie2/bf16_max_reduce/bf16_max_reduce.mlir b/...ests/bf16_max_reduce/bf16_max_reduce.mlir → ...aie2/bf16_max_reduce/bf16_max_reduce.mlir
@@ -6,7 +6,7 @@
 // RUN: aie-opt %s --convert-vector-to-aievec="aie-target=aie2" -lower-affine | aie-translate -aie2=true --aievec-to-cpp -o dut.cc
 // RUN: xchesscc_wrapper %xchesscc_aie2_args +w work +o work -I%S -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper %xchesscc_aie2_args +w work +o work -I%S -I. %S/testbench.cc work/dut.o
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 

diff --git a/...ts/aievec_tests/bf16_max_reduce/defines.h → ...evec_tests/aie2/bf16_max_reduce/defines.h b/...ts/aievec_tests/bf16_max_reduce/defines.h → ...evec_tests/aie2/bf16_max_reduce/defines.h
diff --git a/...tests/aievec_tests/bf16_max_reduce/dut.cc → .../aievec_tests/aie2/bf16_max_reduce/dut.cc b/...tests/aievec_tests/bf16_max_reduce/dut.cc → .../aievec_tests/aie2/bf16_max_reduce/dut.cc
diff --git a/...aievec_tests/bf16_max_reduce/testbench.cc → ...c_tests/aie2/bf16_max_reduce/testbench.cc b/...aievec_tests/bf16_max_reduce/testbench.cc → ...c_tests/aie2/bf16_max_reduce/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...bf16_min_reduce/bf16_min_reduce-llvm.mlir → ...bf16_min_reduce/bf16_min_reduce-llvm.mlir b/...bf16_min_reduce/bf16_min_reduce-llvm.mlir → ...bf16_min_reduce/bf16_min_reduce-llvm.mlir
@@ -8,7 +8,7 @@
 // RUN: aie-translate llvmir.mlir %llvmir-to-ll% -o dut.ll
 // RUN: %PEANO_INSTALL_DIR/bin/clang %clang_aie2_args -c dut.ll -o dut.o
 // RUN: xchesscc_wrapper %xchesscc_aie2_args -DTO_LLVM +w work +o work -I%S -I. %S/testbench.cc dut.o
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 

diff --git a/...ests/bf16_min_reduce/bf16_min_reduce.mlir → ...aie2/bf16_min_reduce/bf16_min_reduce.mlir b/...ests/bf16_min_reduce/bf16_min_reduce.mlir → ...aie2/bf16_min_reduce/bf16_min_reduce.mlir
@@ -6,7 +6,7 @@
 // RUN: aie-opt %s --convert-vector-to-aievec="aie-target=aie2" -lower-affine | aie-translate -aie2=true --aievec-to-cpp -o dut.cc
 // RUN: xchesscc_wrapper %xchesscc_aie2_args +w work +o work -I%S -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper %xchesscc_aie2_args +w work +o work -I%S -I. %S/testbench.cc work/dut.o
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 

diff --git a/...ts/aievec_tests/bf16_min_reduce/defines.h → ...evec_tests/aie2/bf16_min_reduce/defines.h b/...ts/aievec_tests/bf16_min_reduce/defines.h → ...evec_tests/aie2/bf16_min_reduce/defines.h
diff --git a/...tests/aievec_tests/bf16_min_reduce/dut.cc → .../aievec_tests/aie2/bf16_min_reduce/dut.cc b/...tests/aievec_tests/bf16_min_reduce/dut.cc → .../aievec_tests/aie2/bf16_min_reduce/dut.cc
diff --git a/...aievec_tests/bf16_min_reduce/testbench.cc → ...c_tests/aie2/bf16_min_reduce/testbench.cc b/...aievec_tests/bf16_min_reduce/testbench.cc → ...c_tests/aie2/bf16_min_reduce/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...evec_tests/bf16_rsqrt_v16/bf16_rsqrt.mlir → ...tests/aie2/bf16_rsqrt_v16/bf16_rsqrt.mlir b/...evec_tests/bf16_rsqrt_v16/bf16_rsqrt.mlir → ...tests/aie2/bf16_rsqrt_v16/bf16_rsqrt.mlir
@@ -6,7 +6,7 @@
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIE_ARCH__=20 -D__AIENGINE__ -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIE_ARCH__=20 -D__AIENGINE__ -I. %S/testbench.cc work/dut.o
 // RUN: mkdir -p data
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 // Cycle count: 1708

diff --git a/...sts/aievec_tests/bf16_rsqrt_v16/defines.h → ...ievec_tests/aie2/bf16_rsqrt_v16/defines.h b/...sts/aievec_tests/bf16_rsqrt_v16/defines.h → ...ievec_tests/aie2/bf16_rsqrt_v16/defines.h
diff --git a/..._tests/aievec_tests/bf16_rsqrt_v16/dut.cc → ...s/aievec_tests/aie2/bf16_rsqrt_v16/dut.cc b/..._tests/aievec_tests/bf16_rsqrt_v16/dut.cc → ...s/aievec_tests/aie2/bf16_rsqrt_v16/dut.cc
diff --git a/.../aievec_tests/bf16_rsqrt_v16/testbench.cc → ...ec_tests/aie2/bf16_rsqrt_v16/testbench.cc b/.../aievec_tests/bf16_rsqrt_v16/testbench.cc → ...ec_tests/aie2/bf16_rsqrt_v16/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...evec_tests/bf16_rsqrt_v32/bf16_rsqrt.mlir → ...tests/aie2/bf16_rsqrt_v32/bf16_rsqrt.mlir b/...evec_tests/bf16_rsqrt_v32/bf16_rsqrt.mlir → ...tests/aie2/bf16_rsqrt_v32/bf16_rsqrt.mlir
@@ -6,7 +6,7 @@
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIE_ARCH__=20 -D__AIENGINE__ -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIE_ARCH__=20 -D__AIENGINE__ -I. %S/testbench.cc work/dut.o
 // RUN: mkdir -p data
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 // Cycle count: 1040

diff --git a/...sts/aievec_tests/bf16_rsqrt_v32/defines.h → ...ievec_tests/aie2/bf16_rsqrt_v32/defines.h b/...sts/aievec_tests/bf16_rsqrt_v32/defines.h → ...ievec_tests/aie2/bf16_rsqrt_v32/defines.h
diff --git a/..._tests/aievec_tests/bf16_rsqrt_v32/dut.cc → ...s/aievec_tests/aie2/bf16_rsqrt_v32/dut.cc b/..._tests/aievec_tests/bf16_rsqrt_v32/dut.cc → ...s/aievec_tests/aie2/bf16_rsqrt_v32/dut.cc
diff --git a/.../aievec_tests/bf16_rsqrt_v32/testbench.cc → ...ec_tests/aie2/bf16_rsqrt_v32/testbench.cc b/.../aievec_tests/bf16_rsqrt_v32/testbench.cc → ...ec_tests/aie2/bf16_rsqrt_v32/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...evec_tests/bf16_softmax/bf16_softmax.mlir → ...tests/aie2/bf16_softmax/bf16_softmax.mlir b/...evec_tests/bf16_softmax/bf16_softmax.mlir → ...tests/aie2/bf16_softmax/bf16_softmax.mlir
@@ -3,7 +3,7 @@
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIEARCH__=20 -D__AIENGINE__ -I. -c dut.cc -o dut.cc.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIEARCH__=20 -D__AIENGINE__ -I. %S/testbench.cc work/dut.cc.o work/lut_based_ops.cpp.o
 // RUN: mkdir -p data
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 module {

diff --git a/...tests/aievec_tests/bf16_softmax/defines.h → .../aievec_tests/aie2/bf16_softmax/defines.h b/...tests/aievec_tests/bf16_softmax/defines.h → .../aievec_tests/aie2/bf16_softmax/defines.h
diff --git a/...it_tests/aievec_tests/bf16_softmax/dut.cc → ...sts/aievec_tests/aie2/bf16_softmax/dut.cc b/...it_tests/aievec_tests/bf16_softmax/dut.cc → ...sts/aievec_tests/aie2/bf16_softmax/dut.cc
diff --git a/...ts/aievec_tests/bf16_softmax/testbench.cc → ...evec_tests/aie2/bf16_softmax/testbench.cc b/...ts/aievec_tests/bf16_softmax/testbench.cc → ...evec_tests/aie2/bf16_softmax/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...ec_tests/bf16_softmax_2/bf16_softmax.mlir → ...sts/aie2/bf16_softmax_2/bf16_softmax.mlir b/...ec_tests/bf16_softmax_2/bf16_softmax.mlir → ...sts/aie2/bf16_softmax_2/bf16_softmax.mlir
@@ -3,7 +3,7 @@
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIEARCH__=20 -D__AIENGINE__ -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIEARCH__=20 -D__AIENGINE__ -I. %S/testbench.cc work/dut.o work/lut_based_ops.o
 // RUN: mkdir -p data
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 module {

diff --git a/...sts/aievec_tests/bf16_softmax_2/defines.h → ...ievec_tests/aie2/bf16_softmax_2/defines.h b/...sts/aievec_tests/bf16_softmax_2/defines.h → ...ievec_tests/aie2/bf16_softmax_2/defines.h
diff --git a/..._tests/aievec_tests/bf16_softmax_2/dut.cc → ...s/aievec_tests/aie2/bf16_softmax_2/dut.cc b/..._tests/aievec_tests/bf16_softmax_2/dut.cc → ...s/aievec_tests/aie2/bf16_softmax_2/dut.cc
diff --git a/.../aievec_tests/bf16_softmax_2/testbench.cc → ...ec_tests/aie2/bf16_softmax_2/testbench.cc b/.../aievec_tests/bf16_softmax_2/testbench.cc → ...ec_tests/aie2/bf16_softmax_2/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...aievec_tests/bf16_sqrt_v16/bf16_sqrt.mlir → ...c_tests/aie2/bf16_sqrt_v16/bf16_sqrt.mlir b/...aievec_tests/bf16_sqrt_v16/bf16_sqrt.mlir → ...c_tests/aie2/bf16_sqrt_v16/bf16_sqrt.mlir
@@ -6,7 +6,7 @@
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIE_ARCH__=20 -D__AIENGINE__ -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIE_ARCH__=20 -D__AIENGINE__ -I. %S/testbench.cc work/dut.o
 // RUN: mkdir -p data
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 // Cycle count: 2724

diff --git a/...ests/aievec_tests/bf16_sqrt_v16/defines.h → ...aievec_tests/aie2/bf16_sqrt_v16/defines.h b/...ests/aievec_tests/bf16_sqrt_v16/defines.h → ...aievec_tests/aie2/bf16_sqrt_v16/defines.h
diff --git a/...t_tests/aievec_tests/bf16_sqrt_v16/dut.cc → ...ts/aievec_tests/aie2/bf16_sqrt_v16/dut.cc b/...t_tests/aievec_tests/bf16_sqrt_v16/dut.cc → ...ts/aievec_tests/aie2/bf16_sqrt_v16/dut.cc
diff --git a/...s/aievec_tests/bf16_sqrt_v32/testbench.cc → ...vec_tests/aie2/bf16_sqrt_v16/testbench.cc b/...s/aievec_tests/bf16_sqrt_v32/testbench.cc → ...vec_tests/aie2/bf16_sqrt_v16/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...aievec_tests/bf16_sqrt_v32/bf16_sqrt.mlir → ...c_tests/aie2/bf16_sqrt_v32/bf16_sqrt.mlir b/...aievec_tests/bf16_sqrt_v32/bf16_sqrt.mlir → ...c_tests/aie2/bf16_sqrt_v32/bf16_sqrt.mlir
@@ -6,7 +6,7 @@
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIE_ARCH__=20 -D__AIENGINE__ -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I%aie_runtime_lib%/AIE2 -I %aietools/include -D__AIE_ARCH__=20 -D__AIENGINE__ -I. %S/testbench.cc work/dut.o
 // RUN: mkdir -p data
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 // Cycle count: 2122

diff --git a/...ests/aievec_tests/bf16_sqrt_v32/defines.h → ...aievec_tests/aie2/bf16_sqrt_v32/defines.h b/...ests/aievec_tests/bf16_sqrt_v32/defines.h → ...aievec_tests/aie2/bf16_sqrt_v32/defines.h
diff --git a/...t_tests/aievec_tests/bf16_sqrt_v32/dut.cc → ...ts/aievec_tests/aie2/bf16_sqrt_v32/dut.cc b/...t_tests/aievec_tests/bf16_sqrt_v32/dut.cc → ...ts/aievec_tests/aie2/bf16_sqrt_v32/dut.cc
diff --git a/...s/aievec_tests/bf16_sqrt_v16/testbench.cc → ...vec_tests/aie2/bf16_sqrt_v32/testbench.cc b/...s/aievec_tests/bf16_sqrt_v16/testbench.cc → ...vec_tests/aie2/bf16_sqrt_v32/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...f16xbf16_add_elem/bf16xbf16_add_elem.mlir → ...f16xbf16_add_elem/bf16xbf16_add_elem.mlir b/...f16xbf16_add_elem/bf16xbf16_add_elem.mlir → ...f16xbf16_add_elem/bf16xbf16_add_elem.mlir
@@ -6,7 +6,7 @@
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper aie2 -f -g +s +w work +o work -I%S -I. %S/testbench.cc work/dut.o
 // RUN: mkdir -p data
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 

diff --git a/...aievec_tests/bf16xbf16_add_elem/defines.h → ...c_tests/aie2/bf16xbf16_add_elem/defines.h b/...aievec_tests/bf16xbf16_add_elem/defines.h → ...c_tests/aie2/bf16xbf16_add_elem/defines.h
diff --git a/...ts/aievec_tests/bf16xbf16_add_elem/dut.cc → ...evec_tests/aie2/bf16xbf16_add_elem/dut.cc b/...ts/aievec_tests/bf16xbf16_add_elem/dut.cc → ...evec_tests/aie2/bf16xbf16_add_elem/dut.cc
diff --git a/...vec_tests/bf16xbf16_add_elem/testbench.cc → ...ests/aie2/bf16xbf16_add_elem/testbench.cc b/...vec_tests/bf16xbf16_add_elem/testbench.cc → ...ests/aie2/bf16xbf16_add_elem/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>

diff --git a/...f16_max_elem/bf16xbf16_max_elem-llvm.mlir → ...f16_max_elem/bf16xbf16_max_elem-llvm.mlir b/...f16_max_elem/bf16xbf16_max_elem-llvm.mlir → ...f16_max_elem/bf16xbf16_max_elem-llvm.mlir
@@ -8,7 +8,7 @@
 // RUN: aie-translate llvmir.mlir %llvmir-to-ll% -o dut.ll
 // RUN: %PEANO_INSTALL_DIR/bin/clang %clang_aie2_args -c dut.ll -o dut.o
 // RUN: xchesscc_wrapper %xchesscc_aie2_args -DTO_LLVM +w work +o work -I%S -I. %S/testbench.cc dut.o
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 

diff --git a/...f16xbf16_max_elem/bf16xbf16_max_elem.mlir → ...f16xbf16_max_elem/bf16xbf16_max_elem.mlir b/...f16xbf16_max_elem/bf16xbf16_max_elem.mlir → ...f16xbf16_max_elem/bf16xbf16_max_elem.mlir
@@ -6,7 +6,7 @@
 // RUN: aie-opt %s -affine-super-vectorize="virtual-vector-size=32" --convert-vector-to-aievec="aie-target=aie2" -lower-affine | aie-translate -aie2=true --aievec-to-cpp -o dut.cc
 // RUN: xchesscc_wrapper %xchesscc_aie2_args +w work +o work -I%S -I. -c dut.cc -o dut.o
 // RUN: xchesscc_wrapper %xchesscc_aie2_args +w work +o work -I%S -I. %S/testbench.cc work/dut.o
-// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
+// RUN: xca_udm_dbg --aiearch aie-ml -qf -T -P %aietools/data/aie_ml/lib/ -t "%S/../../profiling.tcl ./work/a.out" >& xca_udm_dbg.stdout
 // RUN: FileCheck --input-file=./xca_udm_dbg.stdout %s
 // CHECK: TEST PASSED
 

diff --git a/...aievec_tests/bf16xbf16_max_elem/defines.h → ...c_tests/aie2/bf16xbf16_max_elem/defines.h b/...aievec_tests/bf16xbf16_max_elem/defines.h → ...c_tests/aie2/bf16xbf16_max_elem/defines.h
diff --git a/...ts/aievec_tests/bf16xbf16_max_elem/dut.cc → ...evec_tests/aie2/bf16xbf16_max_elem/dut.cc b/...ts/aievec_tests/bf16xbf16_max_elem/dut.cc → ...evec_tests/aie2/bf16xbf16_max_elem/dut.cc
diff --git a/...vec_tests/bf16xbf16_max_elem/testbench.cc → ...ests/aie2/bf16xbf16_max_elem/testbench.cc b/...vec_tests/bf16xbf16_max_elem/testbench.cc → ...ests/aie2/bf16xbf16_max_elem/testbench.cc
@@ -1,4 +1,4 @@
-#include "../common/testbench.h"
+#include "../../common/testbench.h"
 #include "defines.h"
 #include <algorithm>
 #include <cstdint>