SIMD, ENH: Use logical bitwise to implement comparison functions (bool_)

8 files changed, 99 insertions, 9 deletions

diff --git a/numpy/core/src/_simd/_simd.dispatch.c.src b/numpy/core/src/_simd/_simd.dispatch.c.src
index ad8c538e4..f8a0a3196 100644
--- a/numpy/core/src/_simd/_simd.dispatch.c.src
+++ b/numpy/core/src/_simd/_simd.dispatch.c.src
@@ -462,6 +462,9 @@ SIMD_IMPL_INTRIN_2(or_@bsfx@,  v@bsfx@, v@bsfx@, v@bsfx@)
 SIMD_IMPL_INTRIN_2(xor_@bsfx@, v@bsfx@, v@bsfx@, v@bsfx@)
 SIMD_IMPL_INTRIN_1(not_@bsfx@, v@bsfx@, v@bsfx@)
 /**end repeat**/
+SIMD_IMPL_INTRIN_2(andc_b8, vb8, vb8, vb8)
+SIMD_IMPL_INTRIN_2(orc_b8,  vb8, vb8, vb8)
+SIMD_IMPL_INTRIN_2(xnor_b8, vb8, vb8, vb8)
 /***************************
  * Conversions
  ***************************/
@@ -710,6 +713,9 @@ SIMD_INTRIN_DEF(or_@bsfx@)
 SIMD_INTRIN_DEF(xor_@bsfx@)
 SIMD_INTRIN_DEF(not_@bsfx@)
 /**end repeat**/
+SIMD_INTRIN_DEF(andc_b8)
+SIMD_INTRIN_DEF(orc_b8)
+SIMD_INTRIN_DEF(xnor_b8)
 /***************************
  * Conversions
  ***************************/
diff --git a/numpy/core/src/common/simd/avx2/operators.h b/numpy/core/src/common/simd/avx2/operators.h
index 5fc7719e9..0e77fc6be 100644
--- a/numpy/core/src/common/simd/avx2/operators.h
+++ b/numpy/core/src/common/simd/avx2/operators.h
@@ -114,6 +114,11 @@ NPY_FINLINE __m256i npyv_shr_s64(__m256i a, int c)
 #define npyv_not_b32 npyv_not_u8
 #define npyv_not_b64 npyv_not_u8
 
+// ANDC, ORC and XNOR
+#define npyv_andc_b8(A, B) _mm256_andnot_si256(A, B)
+#define npyv_orc_b8(A, B) npyv_or_b8(npyv_not_b8(A), B)
+#define npyv_xnor_b8(A, B) npyv_not_b8(npyv_xor_b8(A, B))
+
 /***************************
  * Comparison
  ***************************/
diff --git a/numpy/core/src/common/simd/avx512/operators.h b/numpy/core/src/common/simd/avx512/operators.h
index d53932fa8..8c98b72dd 100644
--- a/numpy/core/src/common/simd/avx512/operators.h
+++ b/numpy/core/src/common/simd/avx512/operators.h
@@ -152,6 +152,9 @@
     #define npyv_xor_b16 _kxor_mask32
     #define npyv_not_b8  _knot_mask64
     #define npyv_not_b16 _knot_mask32
+    #define npyv_andc_b8 _kandn_mask64
+    #define npyv_orc_b8(A, B) npyv_or_b8(npyv_not_b8(A), B)
+    #define npyv_xnor_b8 _kxnor_mask64
 #elif defined(NPY_HAVE_AVX512BW)
     NPY_FINLINE npyv_b8  npyv_and_b8(npyv_b8 a, npyv_b8 b)
     { return a & b; }
@@ -169,6 +172,12 @@
     { return ~a; }
     NPY_FINLINE npyv_b16 npyv_not_b16(npyv_b16 a)
     { return ~a; }
+    NPY_FINLINE npyv_b8  npyv_andc_b8(npyv_b8 a, npyv_b8 b)
+    { return (~a) & b; }
+    NPY_FINLINE npyv_b8  npyv_orc_b8(npyv_b8 a, npyv_b8 b)
+    { return (~a) | b; }
+    NPY_FINLINE npyv_b8  npyv_xnor_b8(npyv_b8 a, npyv_b8 b)
+    { return ~(a ^ b); }
 #else
     #define npyv_and_b8  _mm512_and_si512
     #define npyv_and_b16 _mm512_and_si512
@@ -178,6 +187,9 @@
     #define npyv_xor_b16 _mm512_xor_si512
     #define npyv_not_b8  npyv_not_u8
     #define npyv_not_b16 npyv_not_u8
+    #define npyv_andc_b8 _mm512_andnot_si512
+    #define npyv_orc_b8(A, B) npyv_or_b8(npyv_not_b8(A), B)
+    #define npyv_xnor_b8(A, B) npyv_not_b8(npyv_xor_b8(A, B))
 #endif
 
 #define npyv_and_b32 _mm512_kand
diff --git a/numpy/core/src/common/simd/neon/operators.h b/numpy/core/src/common/simd/neon/operators.h
index b43ba3653..6c155fc67 100644
--- a/numpy/core/src/common/simd/neon/operators.h
+++ b/numpy/core/src/common/simd/neon/operators.h
@@ -116,6 +116,11 @@
 #define npyv_not_b32  vmvnq_u32
 #define npyv_not_b64  npyv_not_u64
 
+// ANDC, ORC and XNOR
+#define npyv_andc_b8(A, B) vbicq_u8(B, A)
+#define npyv_orc_b8(A, B) vornq_u8(B, A)
+#define npyv_xnor_b8 vceqq_u8
+
 /***************************
  * Comparison
  ***************************/
diff --git a/numpy/core/src/common/simd/sse/operators.h b/numpy/core/src/common/simd/sse/operators.h
index 51c84fb4e..51bdca356 100644
--- a/numpy/core/src/common/simd/sse/operators.h
+++ b/numpy/core/src/common/simd/sse/operators.h
@@ -115,6 +115,11 @@ NPY_FINLINE __m128i npyv_shr_s64(__m128i a, int c)
 #define npyv_not_b32 npyv_not_u8
 #define npyv_not_b64 npyv_not_u8
 
+// ANDC, ORC and XNOR
+#define npyv_andc_b8(A, B) _mm_andnot_si128(A, B)
+#define npyv_orc_b8(A, B) npyv_or_b8(npyv_not_b8(A), B)
+#define npyv_xnor_b8(A, B) npyv_not_b8(npyv_xor_b8(A, B))
+
 /***************************
  * Comparison
  ***************************/
diff --git a/numpy/core/src/common/simd/vsx/operators.h b/numpy/core/src/common/simd/vsx/operators.h
index d34057ff3..fc29ba920 100644
--- a/numpy/core/src/common/simd/vsx/operators.h
+++ b/numpy/core/src/common/simd/vsx/operators.h
@@ -133,6 +133,11 @@ NPY_FINLINE npyv_f32 npyv_not_f32(npyv_f32 a)
 NPY_FINLINE npyv_f64 npyv_not_f64(npyv_f64 a)
 { return vec_nor(a, a); }
 
+// ANDC, ORC and XNOR
+#define npyv_andc_b8(A, B) vec_andc(B, A)
+#define npyv_orc_b8(A, B) vec_orc(B, A)
+#define npyv_xnor_b8 vec_eqv
+
 /***************************
  * Comparison
  ***************************/
diff --git a/numpy/core/src/umath/loops_comparison.dispatch.c.src b/numpy/core/src/umath/loops_comparison.dispatch.c.src
index e5518afb3..07bbf0354 100644
--- a/numpy/core/src/umath/loops_comparison.dispatch.c.src
+++ b/numpy/core/src/umath/loops_comparison.dispatch.c.src
@@ -207,7 +207,8 @@ static void simd_binary_scalar2_@kind@_@sfx@(char **args, npy_intp len)
 /**begin repeat
  * #kind = equal, not_equal, less, less_equal, greater, greater_equal#
  * #OP = ==, !=, <, <=, >, >=#
- * #VOP = cmpeq, cmpneq, cmplt, cmple, cmpgt, cmpge#
+ * #VOP = xnor, xor, andc, orc, andc, orc#
+ * #rev = 0, 0, 0, 0, 1, 1#
  */
 
 #if NPY_SIMD
@@ -225,7 +226,11 @@ static void simd_binary_@kind@_b8(char **args, npy_intp len)
         // Whatever element in src != 0x0 is converted to 0xFF
         npyv_b8 a = npyv_cmpneq_u8(npyv_load_u8(src1), vzero);
         npyv_b8 b = npyv_cmpneq_u8(npyv_load_u8(src2), vzero);
-        npyv_b8 c = npyv_@VOP@_u8(npyv_cvt_u8_b8(a), npyv_cvt_u8_b8(b));
+#if !@rev@
+        npyv_b8 c = npyv_@VOP@_b8(a, b);
+#else
+        npyv_b8 c = npyv_@VOP@_b8(b, a);
+#endif
         npyv_store_u8(dst, npyv_and_u8(npyv_cvt_u8_b8(c), truemask));
     }
 
@@ -243,13 +248,17 @@ static void simd_binary_scalar1_@kind@_b8(char **args, npy_intp len)
     npyv_lanetype_u8 *dst   = (npyv_lanetype_u8 *) args[2];
     const npyv_u8 vzero     = npyv_setall_u8(0x0);
     const npyv_u8 vscalar   = npyv_setall_u8(scalar);
-    const npyv_u8 a         = npyv_cvt_u8_b8(npyv_cmpneq_u8(vscalar, vzero));
+    const npyv_b8 a         = npyv_cmpneq_u8(vscalar, vzero);
     const npyv_u8 truemask  = npyv_setall_u8(0x1);
     const int vstep         = npyv_nlanes_u8;
 
     for (; len >= vstep; len -= vstep, src += vstep, dst += vstep) {
         npyv_b8 b = npyv_cmpneq_u8(npyv_load_u8(src), vzero);
-        npyv_b8 c = npyv_@VOP@_u8(a, npyv_cvt_u8_b8(b));
+#if !@rev@
+        npyv_b8 c = npyv_@VOP@_b8(a, b);
+#else
+        npyv_b8 c = npyv_@VOP@_b8(b, a);
+#endif
         npyv_store_u8(dst, npyv_and_u8(npyv_cvt_u8_b8(c), truemask));
     }
 
@@ -266,13 +275,17 @@ static void simd_binary_scalar2_@kind@_b8(char **args, npy_intp len)
     npyv_lanetype_u8 *dst   = (npyv_lanetype_u8 *) args[2];
     const npyv_u8 vzero     = npyv_setall_u8(0x0);
     const npyv_u8 vscalar   = npyv_setall_u8(scalar);
-    const npyv_u8 b         = npyv_cvt_u8_b8(npyv_cmpneq_u8(vscalar, vzero));
+    const npyv_b8 b         = npyv_cmpneq_u8(vscalar, vzero);
     const npyv_u8 truemask  = npyv_setall_u8(0x1);
     const int vstep         = npyv_nlanes_u8;
 
     for (; len >= vstep; len -= vstep, src += vstep, dst += vstep) {
         npyv_b8 a = npyv_cmpneq_u8(npyv_load_u8(src), vzero);
-        npyv_b8 c = npyv_@VOP@_u8(npyv_cvt_u8_b8(a), b);
+#if !@rev@
+        npyv_b8 c = npyv_@VOP@_b8(a, b);
+#else
+        npyv_b8 c = npyv_@VOP@_b8(b, a);
+#endif
         npyv_store_u8(dst, npyv_and_u8(npyv_cvt_u8_b8(c), truemask));
     }
 
diff --git a/numpy/core/tests/test_simd.py b/numpy/core/tests/test_simd.py
index b0b62b4a2..81d759ad4 100644
--- a/numpy/core/tests/test_simd.py
+++ b/numpy/core/tests/test_simd.py
@@ -156,6 +156,48 @@ class _SIMD_BOOL(_Test_Utility):
             tobits = bin(self.tobits(vdata))
             assert tobits == bin(data_bits)
 
+    def test_andc(self):
+        if self.sfx not in ("b8"):
+            return
+        andc_simd = getattr(self.npyv, f"andc_b8")
+        # create the vectors
+        data = self._data()
+        rdata = self._data(reverse=True)
+        vdata = self._load_b(data)
+        vrdata = self._load_b(rdata)
+        # check andc
+        sandc = [(~x & y) & 0xFF for x, y in zip(rdata, data)]
+        vandc = andc_simd(vrdata, vdata)
+        assert sandc == vandc
+
+    def test_orc(self):
+        if self.sfx not in ("b8"):
+            return
+        orc_simd = getattr(self.npyv, f"orc_b8")
+        # create the vectors
+        data = self._data()
+        rdata = self._data(reverse=True)
+        vdata = self._load_b(data)
+        vrdata = self._load_b(rdata)
+        # check orc
+        sorc = [(~x | y) & 0xFF for x, y in zip(rdata, data)]
+        vorc = orc_simd(vrdata, vdata)
+        assert sorc == vorc
+
+    def test_xnor(self):
+        if self.sfx not in ("b8"):
+            return
+        xnor_simd = getattr(self.npyv, f"xnor_b8")
+        # create the vectors
+        data = self._data()
+        rdata = self._data(reverse=True)
+        vdata = self._load_b(data)
+        vrdata = self._load_b(rdata)
+        # check orc
+        sxnor = [~(x ^ y) & 0xFF for x, y in zip(rdata, data)]
+        vxnor = xnor_simd(vrdata, vdata)
+        assert sxnor == vxnor
+
     def test_pack(self):
         """
         Pack multiple vectors into one
@@ -166,15 +208,12 @@ class _SIMD_BOOL(_Test_Utility):
         """
         if self.sfx not in ("b16", "b32", "b64"):
             return
-
         # create the vectors
         data = self._data()
         rdata = self._data(reverse=True)
         vdata = self._load_b(data)
         vrdata = self._load_b(rdata)
-
         pack_simd = getattr(self.npyv, f"pack_b8_{self.sfx}")
-
         # for scalar execution, concatenate the elements of the multiple lists
         # into a single list (spack) and then iterate over the elements of
         # the created list applying a mask to capture the first byte of them.