fixed some stuff

libs/lospecs/examples/spec.txt

@@ -41,9 +41,10 @@ VPBROADCAST_16u16(w1@256) -> @256 =
 
 # REF: https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#techs=MMX,SSE_ALL,AVX_ALL&ig_expand=324,324,101,6663,4774&text=vpmulh
 # We want to get the high 16 bits of product, check endianness later
+# Lower index = LSB
 VPMULH_16u16(w1@256, w2@256) -> @256 =
   map<16, 16>(
-    fun x@16 y@16 . mult<16>(x, y)[0:15],
+    fun x@16 y@16 . mult<16>(x, y)[16:31],
     w1,
     w2
   )
@@ -54,15 +55,17 @@ VPMULH_16u16(w1@256, w2@256) -> @256 =
 # rshift<n>(x,y): right shift logical x by y (x >> y) as n-bit ints
 # sext<n>(x): sign extend x to n bits
 # Can be implemented differently from this
+# REF: https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#techs=MMX,SSE_ALL,AVX_ALL&ig_expand=324,324,101,6663,4774,4785&text=vpmulhrs
+# REF: https://www.felixcloutier.com/x86/pmulhrsw
+# Need to check mult here
+# mult<16> : 16, 16 -> 32
 VPMULHRS_16u16(w1@256, w2@256) -> @256 =
   map<16, 16>(
     fun x@16 y@16 . add<32>(
                       rshift<32>(
-                        mult<32>(sext<32>(x), 
-                                 sext<32>(y)
-                                 ), 
+                        mult<16>(x, y), 
                         14),
-                      1)[16:31],
+                      1)[1:16],
     w1,
     w2)
 
@@ -103,23 +106,23 @@ VPMADDUBSW_256(w1@256, w2@256) -> @256 =
 # map<n, x> has type: ((...(a -> b) -> ...) -> c) -> d) -> a[x] -> b[x] -> ... -> c[x] -> d[x]
 #                    | Function that takes args      | | Lists of said args        |   | List of ret type |
 # where if a is a type then a[x] is the type of lists of length x with elements with type a
-#PACKUS_16u16(w1@256, w2@256) -> @256 =
-# map<128, 2>(
-#   fun v1@128 v2@128 . or<128>(
-#                             map<16,8>(
-#                               fun a@16 . SatToUW<8>(a),
-#                                 v1
-#                             ),
-#                    sla<128>(map<16,8>(
-#                               fun a@16 . SatToUW<8>(a),
-#                                 v2
-#                             ),
-#                             64
-#                         )
-#                       )            
-#   w1,
-#   w2,
-# )
+PACKUS_16u16(w1@256, w2@256) -> @256 =
+ map<128, 2>(
+   fun v1@128 v2@128 . or<128>(
+                             map<16,8>(
+                               fun a@16 . SatToUW<8>(a),
+                                 v1
+                             ),
+                             sla<128>(map<16,8>(
+                               fun a@16 . SatToUW<8>(a),
+                                 v2
+                             ),
+                             64
+                            )
+                       ),            
+   w1,
+   w2
+ )
 
 # TODO:
 # Extend map to take in existing functions