improved absdiffX

Author: JishinMaster

avx_mathfun.h

@@ -432,7 +432,7 @@ static inline v8sf cos256_ps(v8sf x)
    it is almost as fast, and gives you a free cosine with your sine */
 static inline void sincos256_ps(v8sf x, v8sf *s, v8sf *c)
 {
-    v8sf xmm1, xmm2, sign_bit_sin, y;
+    v8sf xmm1, xmm2, xmm3, sign_bit_sin, y;
     v8si imm0, imm2, imm4;
 
 #ifndef __AVX2__

simd_utils_altivec_int32.h

@@ -815,9 +815,13 @@ static inline v8ss vec_absdiff(v8ss a, v8ss b)
     cmp = vec_cmpgt(a, b);
     difab = vec_sub(a, b);
     difba = vec_sub(b, a);
+#if 1 // should be faster
+    return vec_sel(difba, difab, cmp);
+#else
     difab = vec_and(*(v8ss *) &cmp, difab);
     difba = vec_andc(difba, *(v8ss *) &cmp);
     return vec_or(difab, difba);
+#endif
 }
 
 static inline void absdiff16s_128s(int16_t *src1, int16_t *src2, int16_t *dst, int len)

simd_utils_avx512_int32.h

@@ -282,44 +282,59 @@ static inline void fast_copy512s_4(int32_t *src, int32_t *dst, int len)
     }
 }
 
-// to be improved?
 static inline __m512i _mm512_absdiff_epi16(__m512i a, __m512i b)
 {
     __m512i cmp, difab, difba;
-    __m512i zero = _mm512_setzero_epi32();
-    __mmask64 cmp_mask = _mm512_cmpgt_epi16_mask(a, b);
-    cmp = _mm512_mask_set1_epi16(zero, cmp_mask, 0xFFFF);
+    __mmask32 cmp_mask = _mm512_cmpgt_epi16_mask(a, b);
+
     difab = _mm512_sub_epi16(a, b);
     difba = _mm512_sub_epi16(b, a);
+#if 1 // should be faster
+    return _mm512_mask_blend_epi16(cmp_mask, difba, difab);
+#else
+    __m512i zero = _mm512_setzero_epi32();
+    cmp = _mm512_mask_set1_epi16(zero, cmp_mask, 0xFFFF);
     difab = _mm512_and_si512(cmp, difab);
     difba = _mm512_andnot_si512(cmp, difba);
     return _mm512_or_si512(difab, difba);
+#endif
+
 }
 
 static inline __m512i _mm512_absdiff_epi32(__m512i a, __m512i b)
 {
     __m512i cmp, difab, difba;
-    __m512i zero = _mm512_setzero_epi32();
-    __mmask64 cmp_mask = _mm512_cmpgt_epi32_mask(a, b);
-    cmp = _mm512_mask_set1_epi32(zero, cmp_mask, 0xFFFFFFFF);
+    __mmask16 cmp_mask = _mm512_cmpgt_epi32_mask(a, b);
+
     difab = _mm512_sub_epi32(a, b);
     difba = _mm512_sub_epi32(b, a);
+#if 1 // should be faster
+    return _mm512_mask_blend_epi32(cmp_mask, difba, difab);
+#else
+    __m512i zero = _mm512_setzero_epi32();
+    cmp = _mm512_mask_set1_epi32(zero, cmp_mask, 0xFFFFFFFF);
     difab = _mm512_and_si512(cmp, difab);
     difba = _mm512_andnot_si512(cmp, difba);
     return _mm512_or_si512(difab, difba);
+#endif
 }
 
 static inline __m512i _mm512_absdiff_epi8(__m512i a, __m512i b)
 {
     __m512i cmp, difab, difba;
-    __m512i zero = _mm512_setzero_epi32();
     __mmask64 cmp_mask = _mm512_cmpgt_epi8_mask(a, b);
-    cmp = _mm512_mask_set1_epi8(zero, cmp_mask, 0xFF);
+
     difab = _mm512_sub_epi8(a, b);
     difba = _mm512_sub_epi8(b, a);
+#if 1 // should be faster
+    return _mm512_mask_blend_epi32(cmp_mask, difba, difab);
+#else
+    __m512i zero = _mm512_setzero_epi32();
+    cmp = _mm512_mask_set1_epi8(zero, cmp_mask, 0xFF);
     difab = _mm512_and_si512(cmp, difab);
     difba = _mm512_andnot_si512(cmp, difba);
     return _mm512_or_si512(difab, difba);
+#endif
 }
 
 static inline void absdiff16s_512s(int16_t *src1, int16_t *src2, int16_t *dst, int len)

simd_utils_avx_int32.h

@@ -282,9 +282,14 @@ static inline __m256i _mm256_absdiff_epi16(__m256i a, __m256i b)
     cmp = _mm256_cmpgt_epi16(a, b);
     difab = _mm256_sub_epi16(a, b);
     difba = _mm256_sub_epi16(b, a);
+#if 1 // should be faster
+    return _mm256_blendv_epi8(difba, difab, cmp);
+#else
     difab = _mm256_and_si256(cmp, difab);
     difba = _mm256_andnot_si256(cmp, difba);
     return _mm256_or_si256(difab, difba);
+#endif
+ 
 }
 
 static inline __m256i _mm256_absdiff_epi32(__m256i a, __m256i b)
@@ -293,9 +298,13 @@ static inline __m256i _mm256_absdiff_epi32(__m256i a, __m256i b)
     cmp = _mm256_cmpgt_epi32(a, b);
     difab = _mm256_sub_epi32(a, b);
     difba = _mm256_sub_epi32(b, a);
+#if 1 // should be faster
+    return _mm256_blendv_epi8(difba, difab, cmp);
+#else
     difab = _mm256_and_si256(cmp, difab);
     difba = _mm256_andnot_si256(cmp, difba);
     return _mm256_or_si256(difab, difba);
+#endif
 }
 
 static inline __m256i _mm256_absdiff_epi8(__m256i a, __m256i b)
@@ -304,9 +313,13 @@ static inline __m256i _mm256_absdiff_epi8(__m256i a, __m256i b)
     cmp = _mm256_cmpgt_epi8(a, b);
     difab = _mm256_sub_epi8(a, b);
     difba = _mm256_sub_epi8(b, a);
+#if 1 // should be faster
+    return _mm256_blendv_epi8(difba, difab, cmp);
+#else
     difab = _mm256_and_si256(cmp, difab);
     difba = _mm256_andnot_si256(cmp, difba);
     return _mm256_or_si256(difab, difba);
+#endif
 }
 
 static inline void absdiff16s_256s(int16_t *src1, int16_t *src2, int16_t *dst, int len)

simd_utils_sse_int32.h

@@ -305,9 +305,14 @@ static inline __m128i _mm_absdiff_epi16(__m128i a, __m128i b)
     cmp = _mm_cmpgt_epi16(a, b);
     difab = _mm_sub_epi16(a, b);
     difba = _mm_sub_epi16(b, a);
+#if 1 // should be faster
+    return _mm_blendv_epi8(difba, difab, cmp);
+#else
     difab = _mm_and_si128(cmp, difab);
     difba = _mm_andnot_si128(cmp, difba);
     return _mm_or_si128(difab, difba);
+#endif
+
 #else
     return vreinterpretq_m128i_s16(vabdq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b)));
 #endif