From 8f819530dcbc4277fa58f58301a83f61827eb1c5 Mon Sep 17 00:00:00 2001
From: Jiajie Chen <c@jia.je>
Date: Thu, 14 Dec 2023 02:43:17 +0800
Subject: [PATCH] Fix xvssran/xvssrln

---
 code/gen_impl.py    | 71 ++++++++++++++++++++++++++++++++++-----------
 code/xvssran_b_h.h  | 12 ++++++--
 code/xvssran_bu_h.h | 12 ++++++--
 code/xvssran_h_w.h  | 12 ++++++--
 code/xvssran_hu_w.h | 12 ++++++--
 code/xvssran_w_d.h  | 12 ++++++--
 code/xvssran_wu_d.h | 12 ++++++--
 code/xvssrln_b_h.h  | 12 ++++++--
 code/xvssrln_bu_h.h | 12 ++++++--
 code/xvssrln_h_w.h  | 12 ++++++--
 code/xvssrln_hu_w.h | 12 ++++++--
 code/xvssrln_w_d.h  | 12 ++++++--
 code/xvssrln_wu_d.h | 12 ++++++--
 13 files changed, 174 insertions(+), 41 deletions(-)

diff --git a/code/gen_impl.py b/code/gen_impl.py
index 385a41cb..eea6ae36 100644
--- a/code/gen_impl.py
+++ b/code/gen_impl.py
@@ -239,23 +239,60 @@
                 with open(
                     f"{prefix}s{name}n_{width}_{double_width_signed}.h", "w"
                 ) as f:
-                    print(f"for (int i = 0;i < {vlen // w};i++) {{", file=f)
-                    print(f"if (i < {vlen // 2 // w}) {{", file=f)
-                    print(
-                        f"  {shift_sign}{double_w} temp = ({shift_sign}{double_w})a.{double_m}[i] >> (b.{double_m}[i] & {double_w-1});",
-                        file=f,
-                    )
-                    print(
-                        f"  dst.{m}[i] = clamp<{shift_sign}{double_w}>(temp, {min}, {max});",
-                        file=f,
-                    )
-                    print(f"}} else {{", file=f)
-                    print(
-                        f"  dst.{m}[i] = 0;",
-                        file=f,
-                    )
-                    print(f"}}", file=f)
-                    print(f"}}", file=f)
+                    if prefix == "v":
+                        print(f"for (int i = 0;i < {vlen // w};i++) {{", file=f)
+                        print(f"if (i < {vlen // 2 // w}) {{", file=f)
+                        print(
+                            f"  {shift_sign}{double_w} temp = ({shift_sign}{double_w})a.{double_m}[i] >> (b.{double_m}[i] & {double_w-1});",
+                            file=f,
+                        )
+                        print(
+                            f"  dst.{m}[i] = clamp<{shift_sign}{double_w}>(temp, {min}, {max});",
+                            file=f,
+                        )
+                        print(f"}} else {{", file=f)
+                        print(
+                            f"  dst.{m}[i] = 0;",
+                            file=f,
+                        )
+                        print(f"}}", file=f)
+                        print(f"}}", file=f)
+                    else:
+                        print(f"for (int i = 0;i < {vlen // 2 // w};i++) {{", file=f)
+                        print(f"if (i < {vlen // 4 // w}) {{", file=f)
+                        print(
+                            f"  {shift_sign}{double_w} temp = ({shift_sign}{double_w})a.{double_m}[i] >> (b.{double_m}[i] & {double_w-1});",
+                            file=f,
+                        )
+                        print(
+                            f"  dst.{m}[i] = clamp<{shift_sign}{double_w}>(temp, {min}, {max});",
+                            file=f,
+                        )
+                        print(f"}} else {{", file=f)
+                        print(
+                            f"  dst.{m}[i] = 0;",
+                            file=f,
+                        )
+                        print(f"}}", file=f)
+                        print(f"}}", file=f)
+
+                        print(f"for (int i = {vlen // 2 // w};i < {vlen // w};i++) {{", file=f)
+                        print(f"if (i < {3 * vlen // 4 // w}) {{", file=f)
+                        print(
+                            f"  {shift_sign}{double_w} temp = ({shift_sign}{double_w})a.{double_m}[i - {vlen // 4 // w}] >> (b.{double_m}[i - {vlen // 4 // w}] & {double_w-1});",
+                            file=f,
+                        )
+                        print(
+                            f"  dst.{m}[i] = clamp<{shift_sign}{double_w}>(temp, {min}, {max});",
+                            file=f,
+                        )
+                        print(f"}} else {{", file=f)
+                        print(
+                            f"  dst.{m}[i] = 0;",
+                            file=f,
+                        )
+                        print(f"}}", file=f)
+                        print(f"}}", file=f)
                 with open(
                     f"{prefix}s{name}rn_{width}_{double_width_signed}.h", "w"
                 ) as f:
diff --git a/code/xvssran_b_h.h b/code/xvssran_b_h.h
index fac704cc..75f383a4 100644
--- a/code/xvssran_b_h.h
+++ b/code/xvssran_b_h.h
@@ -1,8 +1,16 @@
-for (int i = 0; i < 32; i++) {
-  if (i < 16) {
+for (int i = 0; i < 16; i++) {
+  if (i < 8) {
     s16 temp = (s16)a.half[i] >> (b.half[i] & 15);
     dst.byte[i] = clamp<s16>(temp, -128, 127);
   } else {
     dst.byte[i] = 0;
   }
 }
+for (int i = 16; i < 32; i++) {
+  if (i < 24) {
+    s16 temp = (s16)a.half[i - 8] >> (b.half[i - 8] & 15);
+    dst.byte[i] = clamp<s16>(temp, -128, 127);
+  } else {
+    dst.byte[i] = 0;
+  }
+}
diff --git a/code/xvssran_bu_h.h b/code/xvssran_bu_h.h
index 547a2ad5..8b1914c6 100644
--- a/code/xvssran_bu_h.h
+++ b/code/xvssran_bu_h.h
@@ -1,8 +1,16 @@
-for (int i = 0; i < 32; i++) {
-  if (i < 16) {
+for (int i = 0; i < 16; i++) {
+  if (i < 8) {
     s16 temp = (s16)a.half[i] >> (b.half[i] & 15);
     dst.byte[i] = clamp<s16>(temp, 0, 255);
   } else {
     dst.byte[i] = 0;
   }
 }
+for (int i = 16; i < 32; i++) {
+  if (i < 24) {
+    s16 temp = (s16)a.half[i - 8] >> (b.half[i - 8] & 15);
+    dst.byte[i] = clamp<s16>(temp, 0, 255);
+  } else {
+    dst.byte[i] = 0;
+  }
+}
diff --git a/code/xvssran_h_w.h b/code/xvssran_h_w.h
index fc4c47af..ed27287f 100644
--- a/code/xvssran_h_w.h
+++ b/code/xvssran_h_w.h
@@ -1,8 +1,16 @@
-for (int i = 0; i < 16; i++) {
-  if (i < 8) {
+for (int i = 0; i < 8; i++) {
+  if (i < 4) {
     s32 temp = (s32)a.word[i] >> (b.word[i] & 31);
     dst.half[i] = clamp<s32>(temp, -32768, 32767);
   } else {
     dst.half[i] = 0;
   }
 }
+for (int i = 8; i < 16; i++) {
+  if (i < 12) {
+    s32 temp = (s32)a.word[i - 4] >> (b.word[i - 4] & 31);
+    dst.half[i] = clamp<s32>(temp, -32768, 32767);
+  } else {
+    dst.half[i] = 0;
+  }
+}
diff --git a/code/xvssran_hu_w.h b/code/xvssran_hu_w.h
index f99435fb..c3568676 100644
--- a/code/xvssran_hu_w.h
+++ b/code/xvssran_hu_w.h
@@ -1,8 +1,16 @@
-for (int i = 0; i < 16; i++) {
-  if (i < 8) {
+for (int i = 0; i < 8; i++) {
+  if (i < 4) {
     s32 temp = (s32)a.word[i] >> (b.word[i] & 31);
     dst.half[i] = clamp<s32>(temp, 0, 65535);
   } else {
     dst.half[i] = 0;
   }
 }
+for (int i = 8; i < 16; i++) {
+  if (i < 12) {
+    s32 temp = (s32)a.word[i - 4] >> (b.word[i - 4] & 31);
+    dst.half[i] = clamp<s32>(temp, 0, 65535);
+  } else {
+    dst.half[i] = 0;
+  }
+}
diff --git a/code/xvssran_w_d.h b/code/xvssran_w_d.h
index 9108b1a4..92562f55 100644
--- a/code/xvssran_w_d.h
+++ b/code/xvssran_w_d.h
@@ -1,8 +1,16 @@
-for (int i = 0; i < 8; i++) {
-  if (i < 4) {
+for (int i = 0; i < 4; i++) {
+  if (i < 2) {
     s64 temp = (s64)a.dword[i] >> (b.dword[i] & 63);
     dst.word[i] = clamp<s64>(temp, -2147483648, 2147483647);
   } else {
     dst.word[i] = 0;
   }
 }
+for (int i = 4; i < 8; i++) {
+  if (i < 6) {
+    s64 temp = (s64)a.dword[i - 2] >> (b.dword[i - 2] & 63);
+    dst.word[i] = clamp<s64>(temp, -2147483648, 2147483647);
+  } else {
+    dst.word[i] = 0;
+  }
+}
diff --git a/code/xvssran_wu_d.h b/code/xvssran_wu_d.h
index d939ff03..56802a92 100644
--- a/code/xvssran_wu_d.h
+++ b/code/xvssran_wu_d.h
@@ -1,8 +1,16 @@
-for (int i = 0; i < 8; i++) {
-  if (i < 4) {
+for (int i = 0; i < 4; i++) {
+  if (i < 2) {
     s64 temp = (s64)a.dword[i] >> (b.dword[i] & 63);
     dst.word[i] = clamp<s64>(temp, 0, 4294967295);
   } else {
     dst.word[i] = 0;
   }
 }
+for (int i = 4; i < 8; i++) {
+  if (i < 6) {
+    s64 temp = (s64)a.dword[i - 2] >> (b.dword[i - 2] & 63);
+    dst.word[i] = clamp<s64>(temp, 0, 4294967295);
+  } else {
+    dst.word[i] = 0;
+  }
+}
diff --git a/code/xvssrln_b_h.h b/code/xvssrln_b_h.h
index 33172401..6919f864 100644
--- a/code/xvssrln_b_h.h
+++ b/code/xvssrln_b_h.h
@@ -1,8 +1,16 @@
-for (int i = 0; i < 32; i++) {
-  if (i < 16) {
+for (int i = 0; i < 16; i++) {
+  if (i < 8) {
     u16 temp = (u16)a.half[i] >> (b.half[i] & 15);
     dst.byte[i] = clamp<u16>(temp, 0, 127);
   } else {
     dst.byte[i] = 0;
   }
 }
+for (int i = 16; i < 32; i++) {
+  if (i < 24) {
+    u16 temp = (u16)a.half[i - 8] >> (b.half[i - 8] & 15);
+    dst.byte[i] = clamp<u16>(temp, 0, 127);
+  } else {
+    dst.byte[i] = 0;
+  }
+}
diff --git a/code/xvssrln_bu_h.h b/code/xvssrln_bu_h.h
index 9be937d0..ef433475 100644
--- a/code/xvssrln_bu_h.h
+++ b/code/xvssrln_bu_h.h
@@ -1,8 +1,16 @@
-for (int i = 0; i < 32; i++) {
-  if (i < 16) {
+for (int i = 0; i < 16; i++) {
+  if (i < 8) {
     u16 temp = (u16)a.half[i] >> (b.half[i] & 15);
     dst.byte[i] = clamp<u16>(temp, 0, 255);
   } else {
     dst.byte[i] = 0;
   }
 }
+for (int i = 16; i < 32; i++) {
+  if (i < 24) {
+    u16 temp = (u16)a.half[i - 8] >> (b.half[i - 8] & 15);
+    dst.byte[i] = clamp<u16>(temp, 0, 255);
+  } else {
+    dst.byte[i] = 0;
+  }
+}
diff --git a/code/xvssrln_h_w.h b/code/xvssrln_h_w.h
index 870d96ee..136e0d8c 100644
--- a/code/xvssrln_h_w.h
+++ b/code/xvssrln_h_w.h
@@ -1,8 +1,16 @@
-for (int i = 0; i < 16; i++) {
-  if (i < 8) {
+for (int i = 0; i < 8; i++) {
+  if (i < 4) {
     u32 temp = (u32)a.word[i] >> (b.word[i] & 31);
     dst.half[i] = clamp<u32>(temp, 0, 32767);
   } else {
     dst.half[i] = 0;
   }
 }
+for (int i = 8; i < 16; i++) {
+  if (i < 12) {
+    u32 temp = (u32)a.word[i - 4] >> (b.word[i - 4] & 31);
+    dst.half[i] = clamp<u32>(temp, 0, 32767);
+  } else {
+    dst.half[i] = 0;
+  }
+}
diff --git a/code/xvssrln_hu_w.h b/code/xvssrln_hu_w.h
index 27a7bdcc..2afe20ce 100644
--- a/code/xvssrln_hu_w.h
+++ b/code/xvssrln_hu_w.h
@@ -1,8 +1,16 @@
-for (int i = 0; i < 16; i++) {
-  if (i < 8) {
+for (int i = 0; i < 8; i++) {
+  if (i < 4) {
     u32 temp = (u32)a.word[i] >> (b.word[i] & 31);
     dst.half[i] = clamp<u32>(temp, 0, 65535);
   } else {
     dst.half[i] = 0;
   }
 }
+for (int i = 8; i < 16; i++) {
+  if (i < 12) {
+    u32 temp = (u32)a.word[i - 4] >> (b.word[i - 4] & 31);
+    dst.half[i] = clamp<u32>(temp, 0, 65535);
+  } else {
+    dst.half[i] = 0;
+  }
+}
diff --git a/code/xvssrln_w_d.h b/code/xvssrln_w_d.h
index 6cf2000a..6a74bd45 100644
--- a/code/xvssrln_w_d.h
+++ b/code/xvssrln_w_d.h
@@ -1,8 +1,16 @@
-for (int i = 0; i < 8; i++) {
-  if (i < 4) {
+for (int i = 0; i < 4; i++) {
+  if (i < 2) {
     u64 temp = (u64)a.dword[i] >> (b.dword[i] & 63);
     dst.word[i] = clamp<u64>(temp, 0, 2147483647);
   } else {
     dst.word[i] = 0;
   }
 }
+for (int i = 4; i < 8; i++) {
+  if (i < 6) {
+    u64 temp = (u64)a.dword[i - 2] >> (b.dword[i - 2] & 63);
+    dst.word[i] = clamp<u64>(temp, 0, 2147483647);
+  } else {
+    dst.word[i] = 0;
+  }
+}
diff --git a/code/xvssrln_wu_d.h b/code/xvssrln_wu_d.h
index 81790d15..ab2146aa 100644
--- a/code/xvssrln_wu_d.h
+++ b/code/xvssrln_wu_d.h
@@ -1,8 +1,16 @@
-for (int i = 0; i < 8; i++) {
-  if (i < 4) {
+for (int i = 0; i < 4; i++) {
+  if (i < 2) {
     u64 temp = (u64)a.dword[i] >> (b.dword[i] & 63);
     dst.word[i] = clamp<u64>(temp, 0, 4294967295);
   } else {
     dst.word[i] = 0;
   }
 }
+for (int i = 4; i < 8; i++) {
+  if (i < 6) {
+    u64 temp = (u64)a.dword[i - 2] >> (b.dword[i - 2] & 63);
+    dst.word[i] = clamp<u64>(temp, 0, 4294967295);
+  } else {
+    dst.word[i] = 0;
+  }
+}