under developing saturation

2 files changed, 203 insertions, 1 deletions

diff --git a/test/Makefile b/test/Makefile
index 336dcaf..d5613e4 100644
--- a/test/Makefile
+++ b/test/Makefile
@@ -60,7 +60,7 @@ apx: apx.cpp $(XBYAK_INC)
 avx10_test: avx10_test.cpp $(XBYAK_INC)
 	$(CXX) $(CFLAGS) avx10_test.cpp -o $@ -DXBYAK64
 
-TEST_FILES=old.txt new-ymm.txt bf16.txt comp.txt convert.txt
+TEST_FILES=old.txt new-ymm.txt bf16.txt comp.txt convert.txt minmax.txt saturation.txt
 xed_test:
 	@for target in $(addprefix avx10/, $(TEST_FILES)); do ./test_by_xed.sh $$target; done
 
diff --git a/test/avx10/saturation.txt b/test/avx10/saturation.txt
new file mode 100644
index 0000000..591960e
--- /dev/null
+++ b/test/avx10/saturation.txt
@@ -0,0 +1,202 @@
+//
+vcvtnebf162ibs(xm1, xm2);
+vcvtnebf162ibs(xm1, ptr[rax+128]);
+vcvtnebf162ibs(xm1, ptr_b[rax+128]);
+
+vcvtnebf162ibs(ym1, ym2);
+vcvtnebf162ibs(ym1, ptr[rax+128]);
+vcvtnebf162ibs(ym1, ptr_b[rax+128]);
+
+vcvtnebf162ibs(zm1, zm2);
+vcvtnebf162ibs(zm1, ptr[rax+128]);
+vcvtnebf162ibs(zm1, ptr_b[rax+128]);
+//
+vcvtnebf162iubs(xm1, xm2);
+vcvtnebf162iubs(xm1, ptr[rax+128]);
+vcvtnebf162iubs(xm1, ptr_b[rax+128]);
+
+vcvtnebf162iubs(ym1, ym2);
+vcvtnebf162iubs(ym1, ptr[rax+128]);
+vcvtnebf162iubs(ym1, ptr_b[rax+128]);
+
+vcvtnebf162iubs(zm1, zm2);
+vcvtnebf162iubs(zm1, ptr[rax+128]);
+vcvtnebf162iubs(zm1, ptr_b[rax+128]);
+//
+vcvttnebf162ibs(xm1, xm2);
+vcvttnebf162ibs(xm1, ptr[rax+128]);
+vcvttnebf162ibs(xm1, ptr_b[rax+128]);
+
+vcvttnebf162ibs(ym1, ym2);
+vcvttnebf162ibs(ym1, ptr[rax+128]);
+vcvttnebf162ibs(ym1, ptr_b[rax+128]);
+
+vcvttnebf162ibs(zm1, zm2);
+vcvttnebf162ibs(zm1, ptr[rax+128]);
+vcvttnebf162ibs(zm1, ptr_b[rax+128]);
+//
+vcvttnebf162iubs(xm1, xm2);
+vcvttnebf162iubs(xm1, ptr[rax+128]);
+vcvttnebf162iubs(xm1, ptr_b[rax+128]);
+
+vcvttnebf162iubs(ym1, ym2);
+vcvttnebf162iubs(ym1, ptr[rax+128]);
+vcvttnebf162iubs(ym1, ptr_b[rax+128]);
+
+vcvttnebf162iubs(zm1, zm2);
+vcvttnebf162iubs(zm1, ptr[rax+128]);
+vcvttnebf162iubs(zm1, ptr_b[rax+128]);
+//
+vcvttpd2qqs(xm1, xm2);
+vcvttpd2qqs(xm1, ptr[rax+128]);
+vcvttpd2qqs(xm1, ptr_b[rax+128]);
+
+vcvttpd2qqs(ym1, ym2);
+vcvttpd2qqs(ym1, ym2|T_sae);
+vcvttpd2qqs(ym1, ptr[rax+128]);
+vcvttpd2qqs(ym1, ptr_b[rax+128]);
+
+vcvttpd2qqs(zm1, zm2);
+vcvttpd2qqs(zm1, zm2|T_sae);
+vcvttpd2qqs(zm1, ptr[rax+128]);
+vcvttpd2qqs(zm1, ptr_b[rax+128]);
+//
+vcvttpd2uqqs(xm1, xm2);
+vcvttpd2uqqs(xm1, ptr[rax+128]);
+vcvttpd2uqqs(xm1, ptr_b[rax+128]);
+
+vcvttpd2uqqs(ym1, ym2);
+vcvttpd2uqqs(ym1, ym2|T_sae);
+vcvttpd2uqqs(ym1, ptr[rax+128]);
+vcvttpd2uqqs(ym1, ptr_b[rax+128]);
+
+vcvttpd2uqqs(zm1, zm2);
+vcvttpd2uqqs(zm1, zm2|T_sae);
+vcvttpd2uqqs(zm1, ptr[rax+128]);
+vcvttpd2uqqs(zm1, ptr_b[rax+128]);
+//
+vcvtph2ibs(xm1, xm2);
+vcvtph2ibs(xm1, ptr[rax+128]);
+vcvtph2ibs(xm1, ptr_b[rax+128]);
+
+vcvtph2ibs(ym1, ym2);
+vcvtph2ibs(ym1, ym2|T_rd_sae);
+vcvtph2ibs(ym1, ptr[rax+128]);
+vcvtph2ibs(ym1, ptr_b[rax+128]);
+
+vcvtph2ibs(zm1, zm2);
+vcvtph2ibs(zm1, zm2|T_ru_sae);
+vcvtph2ibs(zm1, ptr[rax+128]);
+vcvtph2ibs(zm1, ptr_b[rax+128]);
+//
+vcvtph2iubs(xm1, xm2);
+vcvtph2iubs(xm1, ptr[rax+128]);
+vcvtph2iubs(xm1, ptr_b[rax+128]);
+
+vcvtph2iubs(ym1, ym2);
+vcvtph2iubs(ym1, ym2|T_rd_sae);
+vcvtph2iubs(ym1, ptr[rax+128]);
+vcvtph2iubs(ym1, ptr_b[rax+128]);
+
+vcvtph2iubs(zm1, zm2);
+vcvtph2iubs(zm1, zm2|T_ru_sae);
+vcvtph2iubs(zm1, ptr[rax+128]);
+vcvtph2iubs(zm1, ptr_b[rax+128]);
+//
+vcvttph2ibs(xm1, xm2);
+vcvttph2ibs(xm1, ptr[rax+128]);
+vcvttph2ibs(xm1, ptr_b[rax+128]);
+
+vcvttph2ibs(ym1, ym2);
+vcvttph2ibs(ym1, ym2|T_rd_sae);
+vcvttph2ibs(ym1, ptr[rax+128]);
+vcvttph2ibs(ym1, ptr_b[rax+128]);
+
+vcvttph2ibs(zm1, zm2);
+vcvttph2ibs(zm1, zm2|T_ru_sae);
+vcvttph2ibs(zm1, ptr[rax+128]);
+vcvttph2ibs(zm1, ptr_b[rax+128]);
+//
+vcvttph2iubs(xm1, xm2);
+vcvttph2iubs(xm1, ptr[rax+128]);
+vcvttph2iubs(xm1, ptr_b[rax+128]);
+
+vcvttph2iubs(ym1, ym2);
+vcvttph2iubs(ym1, ym2|T_rd_sae);
+vcvttph2iubs(ym1, ptr[rax+128]);
+vcvttph2iubs(ym1, ptr_b[rax+128]);
+
+vcvttph2iubs(zm1, zm2);
+vcvttph2iubs(zm1, zm2|T_ru_sae);
+vcvttph2iubs(zm1, ptr[rax+128]);
+vcvttph2iubs(zm1, ptr_b[rax+128]);
+//
+vcvttps2dqs(xm1, xm2);
+vcvttps2dqs(xm1, ptr[rax+128]);
+vcvttps2dqs(xm1, ptr_b[rax+128]);
+
+vcvttps2dqs(ym1, ym2);
+vcvttps2dqs(ym1, ym2|T_sae);
+vcvttps2dqs(ym1, ptr[rax+128]);
+vcvttps2dqs(ym1, ptr_b[rax+128]);
+
+vcvttps2dqs(zm1, zm2);
+vcvttps2dqs(zm1, zm2|T_sae);
+vcvttps2dqs(zm1, ptr[rax+128]);
+vcvttps2dqs(zm1, ptr_b[rax+128]);
+//
+vcvtps2ibs(xm1, xm2);
+vcvtps2ibs(xm1, ptr[rax+128]);
+vcvtps2ibs(xm1, ptr_b[rax+128]);
+
+vcvtps2ibs(ym1, ym2);
+vcvtps2ibs(ym1, ym2|T_rd_sae);
+vcvtps2ibs(ym1, ptr[rax+128]);
+vcvtps2ibs(ym1, ptr_b[rax+128]);
+
+vcvtps2ibs(zm1, zm2);
+vcvtps2ibs(zm1, zm2|T_ru_sae);
+vcvtps2ibs(zm1, ptr[rax+128]);
+vcvtps2ibs(zm1, ptr_b[rax+128]);
+//
+vcvtps2iubs(xm1, xm2);
+vcvtps2iubs(xm1, ptr[rax+128]);
+vcvtps2iubs(xm1, ptr_b[rax+128]);
+
+vcvtps2iubs(ym1, ym2);
+vcvtps2iubs(ym1, ym2|T_rd_sae);
+vcvtps2iubs(ym1, ptr[rax+128]);
+vcvtps2iubs(ym1, ptr_b[rax+128]);
+
+vcvtps2iubs(zm1, zm2);
+vcvtps2iubs(zm1, zm2|T_ru_sae);
+vcvtps2iubs(zm1, ptr[rax+128]);
+vcvtps2iubs(zm1, ptr_b[rax+128]);
+//
+vcvttps2ibs(xm1, xm2);
+vcvttps2ibs(xm1, ptr[rax+128]);
+vcvttps2ibs(xm1, ptr_b[rax+128]);
+
+vcvttps2ibs(ym1, ym2);
+vcvttps2ibs(ym1, ym2|T_rd_sae);
+vcvttps2ibs(ym1, ptr[rax+128]);
+vcvttps2ibs(ym1, ptr_b[rax+128]);
+
+vcvttps2ibs(zm1, zm2);
+vcvttps2ibs(zm1, zm2|T_ru_sae);
+vcvttps2ibs(zm1, ptr[rax+128]);
+vcvttps2ibs(zm1, ptr_b[rax+128]);
+//
+vcvttps2iubs(xm1, xm2);
+vcvttps2iubs(xm1, ptr[rax+128]);
+vcvttps2iubs(xm1, ptr_b[rax+128]);
+
+vcvttps2iubs(ym1, ym2);
+vcvttps2iubs(ym1, ym2|T_rd_sae);
+vcvttps2iubs(ym1, ptr[rax+128]);
+vcvttps2iubs(ym1, ptr_b[rax+128]);
+
+vcvttps2iubs(zm1, zm2);
+vcvttps2iubs(zm1, zm2|T_ru_sae);
+vcvttps2iubs(zm1, ptr[rax+128]);
+vcvttps2iubs(zm1, ptr_b[rax+128]);