From cf977208a524a054026f503b7f9d249a36226562 Mon Sep 17 00:00:00 2001
From: Jiajie Chen <c@jia.je>
Date: Thu, 26 Dec 2024 15:46:00 +0800
Subject: [PATCH] Fix vftint_w_d variant's code, reported by Yang Liu

---
 code/vftintrm_w_d.cpp   | 19 +++++++++++++++++++
 code/vftintrm_w_d.h     |  8 ++++----
 code/vftintrne_w_d.cpp  | 19 +++++++++++++++++++
 code/vftintrne_w_d.h    |  8 ++++----
 code/vftintrp_w_d.cpp   | 19 +++++++++++++++++++
 code/vftintrp_w_d.h     |  8 ++++----
 code/vftintrz_w_d.cpp   | 19 +++++++++++++++++++
 code/vftintrz_w_d.h     |  8 ++++----
 code/xvftintrm_w_d.cpp  | 19 +++++++++++++++++++
 code/xvftintrm_w_d.h    | 11 ++++++++---
 code/xvftintrne_w_d.cpp | 19 +++++++++++++++++++
 code/xvftintrne_w_d.h   | 11 ++++++++---
 code/xvftintrp_w_d.cpp  | 19 +++++++++++++++++++
 code/xvftintrp_w_d.h    | 11 ++++++++---
 code/xvftintrz_w_d.cpp  | 19 +++++++++++++++++++
 code/xvftintrz_w_d.h    | 11 ++++++++---
 16 files changed, 200 insertions(+), 28 deletions(-)
 create mode 100644 code/vftintrm_w_d.cpp
 create mode 100644 code/vftintrne_w_d.cpp
 create mode 100644 code/vftintrp_w_d.cpp
 create mode 100644 code/vftintrz_w_d.cpp
 create mode 100644 code/xvftintrm_w_d.cpp
 create mode 100644 code/xvftintrne_w_d.cpp
 create mode 100644 code/xvftintrp_w_d.cpp
 create mode 100644 code/xvftintrz_w_d.cpp

diff --git a/code/vftintrm_w_d.cpp b/code/vftintrm_w_d.cpp
new file mode 100644
index 00000000..4b34472f
--- /dev/null
+++ b/code/vftintrm_w_d.cpp
@@ -0,0 +1,19 @@
+#include "common.h"
+
+v128 vftintrm_w_d(v128 a, v128 b) {
+  v128 dst;
+#include "vftintrm_w_d.h"
+  return dst;
+}
+
+void test() {
+  {
+    __m128d a = {1.0, 2.0};
+    __m128d b = {3.0, 4.0};
+    PRINT(a);
+    PRINT(b);
+    PRINT(__lsx_vftintrm_w_d(a, b));
+    PRINT(vftintrm_w_d(a, b));
+    assert(vftintrm_w_d(a, b) == __lsx_vftintrm_w_d(a, b));
+  }
+}
diff --git a/code/vftintrm_w_d.h b/code/vftintrm_w_d.h
index 6874cf44..20bbde7a 100644
--- a/code/vftintrm_w_d.h
+++ b/code/vftintrm_w_d.h
@@ -1,5 +1,5 @@
-for (int i = 0; i < 2; i++) {
-  dst.dword[i] = (i < 1)
-                     ? (s64)a.fp64[i]
-                     : (s64)b.fp64[i]; // rounding mode is not expressed in C
+for (int i = 0; i < 4; i++) {
+  dst.word[i] = (i < 2)
+                    ? (s64)b.fp64[i]
+                    : (s64)a.fp64[i - 2]; // rounding mode is not expressed in C
 }
diff --git a/code/vftintrne_w_d.cpp b/code/vftintrne_w_d.cpp
new file mode 100644
index 00000000..f1ba0008
--- /dev/null
+++ b/code/vftintrne_w_d.cpp
@@ -0,0 +1,19 @@
+#include "common.h"
+
+v128 vftintrne_w_d(v128 a, v128 b) {
+  v128 dst;
+#include "vftintrne_w_d.h"
+  return dst;
+}
+
+void test() {
+  {
+    __m128d a = {1.0, 2.0};
+    __m128d b = {3.0, 4.0};
+    PRINT(a);
+    PRINT(b);
+    PRINT(__lsx_vftintrne_w_d(a, b));
+    PRINT(vftintrne_w_d(a, b));
+    assert(vftintrne_w_d(a, b) == __lsx_vftintrne_w_d(a, b));
+  }
+}
diff --git a/code/vftintrne_w_d.h b/code/vftintrne_w_d.h
index 6874cf44..20bbde7a 100644
--- a/code/vftintrne_w_d.h
+++ b/code/vftintrne_w_d.h
@@ -1,5 +1,5 @@
-for (int i = 0; i < 2; i++) {
-  dst.dword[i] = (i < 1)
-                     ? (s64)a.fp64[i]
-                     : (s64)b.fp64[i]; // rounding mode is not expressed in C
+for (int i = 0; i < 4; i++) {
+  dst.word[i] = (i < 2)
+                    ? (s64)b.fp64[i]
+                    : (s64)a.fp64[i - 2]; // rounding mode is not expressed in C
 }
diff --git a/code/vftintrp_w_d.cpp b/code/vftintrp_w_d.cpp
new file mode 100644
index 00000000..688a36a6
--- /dev/null
+++ b/code/vftintrp_w_d.cpp
@@ -0,0 +1,19 @@
+#include "common.h"
+
+v128 vftintrp_w_d(v128 a, v128 b) {
+  v128 dst;
+#include "vftintrp_w_d.h"
+  return dst;
+}
+
+void test() {
+  {
+    __m128d a = {1.0, 2.0};
+    __m128d b = {3.0, 4.0};
+    PRINT(a);
+    PRINT(b);
+    PRINT(__lsx_vftintrp_w_d(a, b));
+    PRINT(vftintrp_w_d(a, b));
+    assert(vftintrp_w_d(a, b) == __lsx_vftintrp_w_d(a, b));
+  }
+}
diff --git a/code/vftintrp_w_d.h b/code/vftintrp_w_d.h
index 6874cf44..20bbde7a 100644
--- a/code/vftintrp_w_d.h
+++ b/code/vftintrp_w_d.h
@@ -1,5 +1,5 @@
-for (int i = 0; i < 2; i++) {
-  dst.dword[i] = (i < 1)
-                     ? (s64)a.fp64[i]
-                     : (s64)b.fp64[i]; // rounding mode is not expressed in C
+for (int i = 0; i < 4; i++) {
+  dst.word[i] = (i < 2)
+                    ? (s64)b.fp64[i]
+                    : (s64)a.fp64[i - 2]; // rounding mode is not expressed in C
 }
diff --git a/code/vftintrz_w_d.cpp b/code/vftintrz_w_d.cpp
new file mode 100644
index 00000000..a8a64f19
--- /dev/null
+++ b/code/vftintrz_w_d.cpp
@@ -0,0 +1,19 @@
+#include "common.h"
+
+v128 vftintrz_w_d(v128 a, v128 b) {
+  v128 dst;
+#include "vftintrz_w_d.h"
+  return dst;
+}
+
+void test() {
+  {
+    __m128d a = {1.0, 2.0};
+    __m128d b = {3.0, 4.0};
+    PRINT(a);
+    PRINT(b);
+    PRINT(__lsx_vftintrz_w_d(a, b));
+    PRINT(vftintrz_w_d(a, b));
+    assert(vftintrz_w_d(a, b) == __lsx_vftintrz_w_d(a, b));
+  }
+}
diff --git a/code/vftintrz_w_d.h b/code/vftintrz_w_d.h
index 6874cf44..20bbde7a 100644
--- a/code/vftintrz_w_d.h
+++ b/code/vftintrz_w_d.h
@@ -1,5 +1,5 @@
-for (int i = 0; i < 2; i++) {
-  dst.dword[i] = (i < 1)
-                     ? (s64)a.fp64[i]
-                     : (s64)b.fp64[i]; // rounding mode is not expressed in C
+for (int i = 0; i < 4; i++) {
+  dst.word[i] = (i < 2)
+                    ? (s64)b.fp64[i]
+                    : (s64)a.fp64[i - 2]; // rounding mode is not expressed in C
 }
diff --git a/code/xvftintrm_w_d.cpp b/code/xvftintrm_w_d.cpp
new file mode 100644
index 00000000..29a0ea1a
--- /dev/null
+++ b/code/xvftintrm_w_d.cpp
@@ -0,0 +1,19 @@
+#include "common.h"
+
+v256 xvftintrm_w_d(v256 a, v256 b) {
+  v256 dst;
+#include "xvftintrm_w_d.h"
+  return dst;
+}
+
+void test() {
+  {
+    __m256d a = {1.0, 2.0, 3.0, 4.0};
+    __m256d b = {5.0, 6.0, 7.0, 8.0};
+    PRINT(a);
+    PRINT(b);
+    PRINT(__lasx_xvftintrm_w_d(a, b));
+    PRINT(xvftintrm_w_d(a, b));
+    assert(xvftintrm_w_d(a, b) == __lasx_xvftintrm_w_d(a, b));
+  }
+}
diff --git a/code/xvftintrm_w_d.h b/code/xvftintrm_w_d.h
index 033e5491..1e2eab1a 100644
--- a/code/xvftintrm_w_d.h
+++ b/code/xvftintrm_w_d.h
@@ -1,5 +1,10 @@
 for (int i = 0; i < 4; i++) {
-  dst.dword[i] = (i < 2)
-                     ? (s64)a.fp64[i]
-                     : (s64)b.fp64[i]; // rounding mode is not expressed in C
+  dst.word[i] = (i < 2)
+                    ? (s64)b.fp64[i]
+                    : (s64)a.fp64[i - 2]; // rounding mode is not expressed in C
+}
+for (int i = 4; i < 8; i++) {
+  dst.word[i] = (i < 6)
+                    ? (s64)b.fp64[i - 2]
+                    : (s64)a.fp64[i - 4]; // rounding mode is not expressed in C
 }
diff --git a/code/xvftintrne_w_d.cpp b/code/xvftintrne_w_d.cpp
new file mode 100644
index 00000000..2a4400e4
--- /dev/null
+++ b/code/xvftintrne_w_d.cpp
@@ -0,0 +1,19 @@
+#include "common.h"
+
+v256 xvftintrne_w_d(v256 a, v256 b) {
+  v256 dst;
+#include "xvftintrne_w_d.h"
+  return dst;
+}
+
+void test() {
+  {
+    __m256d a = {1.0, 2.0, 3.0, 4.0};
+    __m256d b = {5.0, 6.0, 7.0, 8.0};
+    PRINT(a);
+    PRINT(b);
+    PRINT(__lasx_xvftintrne_w_d(a, b));
+    PRINT(xvftintrne_w_d(a, b));
+    assert(xvftintrne_w_d(a, b) == __lasx_xvftintrne_w_d(a, b));
+  }
+}
diff --git a/code/xvftintrne_w_d.h b/code/xvftintrne_w_d.h
index 033e5491..1e2eab1a 100644
--- a/code/xvftintrne_w_d.h
+++ b/code/xvftintrne_w_d.h
@@ -1,5 +1,10 @@
 for (int i = 0; i < 4; i++) {
-  dst.dword[i] = (i < 2)
-                     ? (s64)a.fp64[i]
-                     : (s64)b.fp64[i]; // rounding mode is not expressed in C
+  dst.word[i] = (i < 2)
+                    ? (s64)b.fp64[i]
+                    : (s64)a.fp64[i - 2]; // rounding mode is not expressed in C
+}
+for (int i = 4; i < 8; i++) {
+  dst.word[i] = (i < 6)
+                    ? (s64)b.fp64[i - 2]
+                    : (s64)a.fp64[i - 4]; // rounding mode is not expressed in C
 }
diff --git a/code/xvftintrp_w_d.cpp b/code/xvftintrp_w_d.cpp
new file mode 100644
index 00000000..3a05f2d3
--- /dev/null
+++ b/code/xvftintrp_w_d.cpp
@@ -0,0 +1,19 @@
+#include "common.h"
+
+v256 xvftintrp_w_d(v256 a, v256 b) {
+  v256 dst;
+#include "xvftintrp_w_d.h"
+  return dst;
+}
+
+void test() {
+  {
+    __m256d a = {1.0, 2.0, 3.0, 4.0};
+    __m256d b = {5.0, 6.0, 7.0, 8.0};
+    PRINT(a);
+    PRINT(b);
+    PRINT(__lasx_xvftintrp_w_d(a, b));
+    PRINT(xvftintrp_w_d(a, b));
+    assert(xvftintrp_w_d(a, b) == __lasx_xvftintrp_w_d(a, b));
+  }
+}
diff --git a/code/xvftintrp_w_d.h b/code/xvftintrp_w_d.h
index 033e5491..1e2eab1a 100644
--- a/code/xvftintrp_w_d.h
+++ b/code/xvftintrp_w_d.h
@@ -1,5 +1,10 @@
 for (int i = 0; i < 4; i++) {
-  dst.dword[i] = (i < 2)
-                     ? (s64)a.fp64[i]
-                     : (s64)b.fp64[i]; // rounding mode is not expressed in C
+  dst.word[i] = (i < 2)
+                    ? (s64)b.fp64[i]
+                    : (s64)a.fp64[i - 2]; // rounding mode is not expressed in C
+}
+for (int i = 4; i < 8; i++) {
+  dst.word[i] = (i < 6)
+                    ? (s64)b.fp64[i - 2]
+                    : (s64)a.fp64[i - 4]; // rounding mode is not expressed in C
 }
diff --git a/code/xvftintrz_w_d.cpp b/code/xvftintrz_w_d.cpp
new file mode 100644
index 00000000..14ea771c
--- /dev/null
+++ b/code/xvftintrz_w_d.cpp
@@ -0,0 +1,19 @@
+#include "common.h"
+
+v256 xvftintrz_w_d(v256 a, v256 b) {
+  v256 dst;
+#include "xvftintrz_w_d.h"
+  return dst;
+}
+
+void test() {
+  {
+    __m256d a = {1.0, 2.0, 3.0, 4.0};
+    __m256d b = {5.0, 6.0, 7.0, 8.0};
+    PRINT(a);
+    PRINT(b);
+    PRINT(__lasx_xvftintrz_w_d(a, b));
+    PRINT(xvftintrz_w_d(a, b));
+    assert(xvftintrz_w_d(a, b) == __lasx_xvftintrz_w_d(a, b));
+  }
+}
diff --git a/code/xvftintrz_w_d.h b/code/xvftintrz_w_d.h
index 033e5491..1e2eab1a 100644
--- a/code/xvftintrz_w_d.h
+++ b/code/xvftintrz_w_d.h
@@ -1,5 +1,10 @@
 for (int i = 0; i < 4; i++) {
-  dst.dword[i] = (i < 2)
-                     ? (s64)a.fp64[i]
-                     : (s64)b.fp64[i]; // rounding mode is not expressed in C
+  dst.word[i] = (i < 2)
+                    ? (s64)b.fp64[i]
+                    : (s64)a.fp64[i - 2]; // rounding mode is not expressed in C
+}
+for (int i = 4; i < 8; i++) {
+  dst.word[i] = (i < 6)
+                    ? (s64)b.fp64[i - 2]
+                    : (s64)a.fp64[i - 4]; // rounding mode is not expressed in C
 }