Special case for Shuffle test added as well as comments for respectives implementations

diff --git a/src/util/arch/ppc64el/simd_utils.h b/src/util/arch/ppc64el/simd_utils.h
index 107ca110627a40f8a53c0dd966bd0337794a6061..6e93651e59b77e36d9606cdaee283ebadb801bc8 100644 (file)
--- a/src/util/arch/ppc64el/simd_utils.h
+++ b/src/util/arch/ppc64el/simd_utils.h
@@ -462,6 +462,9 @@ char testbit128(m128 val, unsigned int n) {
 
 static really_inline
 m128 pshufb_m128(m128 a, m128 b) {
+    /* On Intel, if bit 0x80 is set, then result is zero, otherwise which the lane it is &0xf.
+       In NEON or PPC, if >=16, then the result is zero, otherwise it is that lane.
+       below is the version that is converted from Intel to PPC.  */
     uint8x16_t mask =(uint8x16_t)vec_cmpge((uint8x16_t)b, (uint8x16_t)vec_splats((uint8_t)0x80));
     uint8x16_t res = vec_perm ((uint8x16_t)a, (uint8x16_t)a, (uint8x16_t)b);
     return (m128) vec_sel((uint8x16_t)res, (uint8x16_t)zeroes128(), (uint8x16_t)mask);

diff --git a/src/util/supervector/arch/ppc64el/impl.cpp b/src/util/supervector/arch/ppc64el/impl.cpp
index 0af136a55684457193e603f2f0f980e62d12f177..ce975cec6c3f03804d5c7babff39fe27dff8be14 100644 (file)
--- a/src/util/supervector/arch/ppc64el/impl.cpp
+++ b/src/util/supervector/arch/ppc64el/impl.cpp
@@ -603,6 +603,9 @@ template<>
 template<>
 really_inline SuperVector<16> SuperVector<16>::pshufb<false>(SuperVector<16> b)
 {
+    /* On Intel, if bit 0x80 is set, then result is zero, otherwise which the lane it is &0xf.
+       In NEON or PPC, if >=16, then the result is zero, otherwise it is that lane.
+       below is the version that is converted from Intel to PPC.  */
     uint8x16_t mask =(uint8x16_t)vec_cmpge((uint8x16_t)b.u.v128[0], (uint8x16_t)vec_splats((uint8_t)0x80));
     uint8x16_t res = vec_perm ((uint8x16_t)u.v128[0], (uint8x16_t)u.v128[0], (uint8x16_t)b.u.v128[0]);
     return (m128) vec_sel((uint8x16_t)res, (uint8x16_t)vec_splat_s8(0), (uint8x16_t)mask);

diff --git a/unit/internal/simd_utils.cpp b/unit/internal/simd_utils.cpp
index 037230d0a32077a9ec168ac482f44d742de39121..1fc6224b19d78169730f39fba7edf5c3ac8dfba2 100644 (file)
--- a/unit/internal/simd_utils.cpp
+++ b/unit/internal/simd_utils.cpp
@@ -845,26 +845,63 @@ TEST(SimdUtilsTest, pshufb_m128) {
     srand (time(NULL));
     u8 vec[16];
     for (int i=0; i<16; i++) {
-        vec[i] = rand() % 100 + 1;
+        vec[i] = rand() % 1000 + 1;
     }
     u8 vec2[16];
     for (int i=0; i<16; i++) {
-        vec2[i]=i + (rand() % 15 + 0);
+        vec2[i]=i + (rand() % 100 + 0);
     }
 
+    /* On Intel, if bit 0x80 is set, then result is zero, otherwise which the lane it is &0xf.
+       In NEON or PPC, if >=16, then the result is zero, otherwise it is that lane.
+       Thus bellow we have to check thah case to NEON or PPC.  */
+    
+    /*Insure that vec2 has at least 1 or more 0x80*/
+    u8 vec3[16] = {0};
+    vec3[15] = 0x80;
+
+    for (int i=0; i<15; i++) {
+        int l = rand() % 1000 + 0;
+        if (l % 16 ==0){
+            vec3[i]= 0x80;
+        } else{
+            vec3[i]= vec2[i];
+        }
+    }
+    /*
+        printf("vec3: ");
+        for(int i=15; i>=0; i--) { printf("%02x, ", vec3[i]); }
+        printf("\n");
+    */
+
+    /*Test Special Case*/
     m128 v1 = loadu128(vec);
-    m128 v2 = loadu128(vec2);
+    m128 v2 = loadu128(vec3);
     m128 vres = pshufb_m128(v1, v2); 
     
     u8 res[16];
     storeu128(res, vres);
 
+    for (int i=0; i<16; i++) {
+       if(vec3[i] & 0x80){
+          ASSERT_EQ(res[i], 0);
+        }else{    
+           ASSERT_EQ(vec[vec3[i] % 16 ], res[i]);
+           }
+    }
+       
+    /*Test Other Cases*/
+    v1 = loadu128(vec);
+    v2 = loadu128(vec2);
+    vres = pshufb_m128(v1, v2); 
+    storeu128(res, vres);
+
     for (int i=0; i<16; i++) {
        if(vec2[i] & 0x80){
           ASSERT_EQ(res[i], 0);
         }else{    
            ASSERT_EQ(vec[vec2[i] % 16 ], res[i]);
-       }
+           }
     }
 }