Merge pull request #42 from mmozeiko/v0.4

Fixes for v0.4
cmuratori · Nov 11, 2018 · ed8c632 · ed8c632
2 parents cc046c7 + bc7f41b
commit ed8c632
Show file tree

Hide file tree

Showing 5 changed files with 36 additions and 23 deletions.
diff --git a/.travis.yml b/.travis.yml
@@ -33,7 +33,7 @@ matrix:
     # test if meow_hash.h compiles as C
     - touch test.c
     - |-
-      "${CC}" -c -maes -include meow_intrinsics.h -include meow_hash.h test.c
+      "${CC}" -c -mssse3 -maes -include meow_intrinsics.h -include meow_hash.h test.c
     # build & test
     - cmd.exe /C build.bat
     - build_clang/meow_test.exe
@@ -50,7 +50,7 @@ matrix:
     script:
     # test if meow_hash.h compiles as C
     - touch test.c
-    - ${CC} -c -maes -include meow_intrinsics.h -include meow_hash.h test.c
+    - ${CC} -c -mssse3 -maes -include meow_intrinsics.h -include meow_hash.h test.c
     # build & test
     - ./build.sh
     - build/meow_test.exe
@@ -75,7 +75,7 @@ matrix:
     script:
     # test if meow_hash.h compiles as C
     - touch test.c
-    - ${CC} -c -maes -include meow_intrinsics.h -include meow_hash.h test.c
+    - ${CC} -c -mssse3 -maes -include meow_intrinsics.h -include meow_hash.h test.c
     # build & test
     - ./build.sh
     - ./build/meow_test
@@ -99,7 +99,7 @@ matrix:
     script:
     # test if meow_hash.h compiles as C
     - touch test.c
-    - ${CC} -c -maes -include meow_intrinsics.h -include meow_hash.h test.c
+    - ${CC} -c -mssse3 -maes -include meow_intrinsics.h -include meow_hash.h test.c
     # build & test
     - ./build.sh
     - ./build/meow_test
@@ -149,7 +149,7 @@ matrix:
     script:
     # test if meow_hash.h compiles as C
     - touch test.c
-    - ${CC} -c -maes -include meow_intrinsics.h -include meow_hash.h test.c
+    - ${CC} -c -mssse3 -maes -include meow_intrinsics.h -include meow_hash.h test.c
     # build & test
     - ./build.sh
     - ./build/meow_test

diff --git a/meow_hash.h b/meow_hash.h
@@ -206,17 +206,17 @@ MeowHash_Accelerated(meow_u64 Seed, meow_u64 TotalLengthInBytes, void *SourceIni
                 Align = 0;
             }
 
-            meow_aes_128 PartialState = Meow128_Shuffle_Mem(Overhang - Align, &MeowShiftAdjust[Align]);
+            meow_u128 Partial = Meow128_Shuffle_Mem(Overhang - Align, &MeowShiftAdjust[Align]);
 
-            PartialState = Meow128_And_Mem( PartialState, &MeowMaskLen[16 - Len8] );
-            S3 = Meow128_AESDEC(S3, PartialState);
+            Partial = Meow128_And_Mem( Partial, &MeowMaskLen[16 - Len8] );
+            S3 = Meow128_AESDEC(S3, Partial);
         }
         else
         {
             // NOTE(casey): We don't have to do Jeff's heroics when we know the
             // buffer is aligned, since we cannot span a memory page (by definition).
-            meow_u128 PartialState = Meow128_And_Mem(*(meow_u128 *)Overhang, &MeowMaskLen[16 - Len8]);
-            S3 = Meow128_AESDEC(S3, PartialState);
+            meow_u128 Partial = Meow128_And_Mem(*(meow_u128 *)Overhang, &MeowMaskLen[16 - Len8]);
+            S3 = Meow128_AESDEC(S3, Partial);
         }
     }
 

diff --git a/meow_intrinsics.h b/meow_intrinsics.h
@@ -95,14 +95,13 @@
 #define Meow128_AESDEC_Finalize(A) (A)
 #define Meow128_Set64x2(Low64, High64) _mm_set_epi64x((High64), (Low64))
 #define Meow128_Set64x2_State(Low64, High64) Meow128_Set64x2(Low64, High64)
-#define Meow128_GetAESConstant(Ptr) (*(meow_u128 *)(Ptr));
+#define Meow128_GetAESConstant(Ptr) (*(meow_u128 *)(Ptr))
 
 #define Meow128_And_Mem(A,B) _mm_and_si128((A),_mm_loadu_si128((meow_u128 *)(B)))
 #define Meow128_Shuffle_Mem(Mem,Control) _mm_shuffle_epi8(_mm_loadu_si128((meow_u128 *)(Mem)),_mm_loadu_si128((meow_u128 *)(Control)))
 
 // TODO(casey): Not sure if this should actually be Meow128_Zero(A) ((A) = _mm_setzero_si128()), maybe
 #define Meow128_Zero() _mm_setzero_si128()
-#define Meow128_ZeroState() Meow128_Zero()
 
 #define Meow256_AESDEC(Prior, XOr) _mm256_aesdec_epi128((Prior), (XOr))
 #define Meow256_AESDEC_Mem(Prior, XOr) _mm256_aesdec_epi128((Prior), *(meow_u256 *)(XOr))
@@ -141,8 +140,11 @@ typedef struct {
     meow_u128 B;
 } meow_aes_128;
 
+#define MeowU32From(A, I) (vgetq_lane_u32(vreinterpretq_u32_u8((A)), (I)))
+#define MeowU64From(A, I) (vgetq_lane_u64(vreinterpretq_u64_u8((A)), (I)))
+
 static int
-Meow128_AreEqual(meow_u128 A, meow_u128 B)
+MeowHashesAreEqualImpl(meow_u128 A, meow_u128 B)
 {
     uint8x16_t Powers = {
         1, 2, 4, 8, 16, 32, 64, 128, 1, 2, 4, 8, 16, 32, 64, 128,
@@ -157,6 +159,8 @@ Meow128_AreEqual(meow_u128 A, meow_u128 B)
     return Output == 0xFFFF;
 }
 
+#define MeowHashesAreEqual(A, B) MeowHashesAreEqualImpl((A), (B))
+
 static meow_aes_128
 Meow128_AESDEC(meow_aes_128 Prior, meow_u128 Xor)
 {
@@ -190,10 +194,11 @@ Meow128_Zero()
 }
 
 static meow_aes_128
-Meow128_ZeroState()
+Meow128_GetAESConstant(const meow_u8 *Ptr)
 {
     meow_aes_128 R;
-    R.A = R.B = vdupq_n_u8(0);
+    R.A = vld1q_u8(Ptr);
+    R.B = vdupq_n_u8(0);
     return(R);
 }
 
@@ -213,6 +218,9 @@ Meow128_Set64x2_State(meow_u64 Low64, meow_u64 High64)
    return(R);
 }
 
+#define Meow128_And_Mem(A,B) vandq_u8((A), vld1q_u8((meow_u8 *)B))
+#define Meow128_Shuffle_Mem(Mem,Control) vqtbl1q_u8(vld1q_u8((meow_u8 *)(Mem)),vld1q_u8((meow_u8 *)(Control)))
+
 #endif
 
 #define MEOW_HASH_VERSION 4
@@ -245,6 +253,7 @@ typedef meow_u128 meow_hash;
 
 #endif
 
+typedef struct meow_hash_state meow_hash_state;
 typedef meow_hash meow_hash_implementation(meow_u64 Seed, meow_u64 Len, void *Source);
 typedef void meow_absorb_implementation(struct meow_hash_state *State, meow_u64 Len, void *Source);
 

diff --git a/more/meow_bench.cpp b/more/meow_bench.cpp
@@ -75,7 +75,11 @@ struct input_size_test
     meow_u64 Size;
 };
 
+#ifdef __aarch64__
+#define MAX_SIZE_TO_TEST Gb(1)
+#else
 #define MAX_SIZE_TO_TEST Gb(2)
+#endif
 #define SIZE_TYPE_COUNT 64
 #define SIZE_COUNT_PER_BATCH 16
 struct input_size_tests

diff --git a/more/meow_more.h b/more/meow_more.h
@@ -39,10 +39,10 @@ MeowHashBegin(meow_hash_state *State)
 static void
 MeowHashAbsorbBlocks(meow_hash_state *State, meow_u64 BlockCount, meow_u8 *Source)
 {
-    meow_u128 S0 = State->S0;
-    meow_u128 S1 = State->S1;
-    meow_u128 S2 = State->S2;
-    meow_u128 S3 = State->S3;
+    meow_aes_128 S0 = State->S0;
+    meow_aes_128 S1 = State->S1;
+    meow_aes_128 S2 = State->S2;
+    meow_aes_128 S3 = State->S3;
 
     while(BlockCount--)
     {
@@ -138,11 +138,11 @@ MeowHashEnd(meow_hash_state *State, meow_u64 Seed)
         {
             Align = 0;
         }
+
+        meow_u128 Partial = Meow128_Shuffle_Mem(Source - Align, &MeowShiftAdjust[Align]);
 
-        meow_aes_128 PartialState = Meow128_Shuffle_Mem(Source - Align, &MeowShiftAdjust[Align]);
-
-        PartialState = Meow128_And_Mem( PartialState, &MeowMaskLen[16 - Len] );
-        S3 = Meow128_AESDEC(S3, PartialState);
+        Partial = Meow128_And_Mem( Partial, &MeowMaskLen[16 - Len] );
+        S3 = Meow128_AESDEC(S3, Partial);
     }
 
     meow_u128 Mixer = Meow128_Set64x2(Seed - State->TotalLengthInBytes,