MAINT: AVX512 intrinsics implementation for float64 input np.log

Author: xiegengxin

benchmarks/benchmarks/bench_avx.py

@@ -39,6 +39,19 @@ def setup(self, ufuncname, stride, dtype):
     def time_ufunc(self, ufuncname, stride, dtype):
         self.f(self.arr[::stride])
 
+class AVX_UFunc_log(Benchmark):
+    params = [stride, dtype]
+    param_names = ['stride', 'dtype']
+    timeout = 10
+
+    def setup(self, stride, dtype):
+        np.seterr(all='ignore')
+        N = 10000
+        self.arr = np.array(np.random.random_sample(stride*N), dtype=dtype)
+
+    def time_log(self, stride, dtype):
+        np.log(self.arr[::stride])
+
 avx_bfuncs = ['maximum',
               'minimum']
 

numpy/core/code_generators/generate_umath.py

@@ -726,6 +726,7 @@ def english_upper(s):
           None,
           TD('e', f='log', astype={'e':'f'}),
           TD('f', simd=[('fma', 'f'), ('avx512f', 'f')]),
+          TD('d', simd=[('avx512f', 'd')]),
           TD('fdg' + cmplx, f='log'),
           TD(P, f='log'),
           ),

numpy/core/src/umath/loops.c.src

@@ -1558,6 +1558,14 @@ DOUBLE_exp(char **args, npy_intp const *dimensions, npy_intp const *steps, void
         *(npy_double *)op1 = npy_exp(in1);
     }
 }
+NPY_NO_EXPORT NPY_GCC_OPT_3 void
+DOUBLE_log(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data))
+{
+    UNARY_LOOP {
+        const npy_double in1 = *(npy_double *)ip1;
+        *(npy_double *)op1 = npy_log(in1);
+    }
+}
 
 /**begin repeat
  * #isa = avx512f, fma#
@@ -1700,6 +1708,16 @@ DOUBLE_exp_avx512f(char **args, npy_intp const *dimensions, npy_intp const *step
     }
 }
 
+NPY_NO_EXPORT NPY_GCC_OPT_3 void
+DOUBLE_log_avx512f(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data))
+{
+    if (!run_unary_avx512f_log_DOUBLE(args, dimensions, steps)) {
+        UNARY_LOOP {
+            const npy_double in1 = *(npy_double *)ip1;
+            *(npy_double *)op1 = npy_log(in1);
+        }
+    }
+}
 
 /**begin repeat
  * Float types

numpy/core/src/umath/loops.h.src

@@ -202,6 +202,12 @@ DOUBLE_exp(char **args, npy_intp const *dimensions, npy_intp const *steps, void
 NPY_NO_EXPORT void
 DOUBLE_exp_avx512f(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
 
+NPY_NO_EXPORT void
+DOUBLE_log(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
+
+NPY_NO_EXPORT void
+DOUBLE_log_avx512f(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
+
 /**begin repeat
  *  #func = sin, cos, exp, log#
  */

numpy/core/src/umath/npy_simd_data.h

@@ -134,4 +134,149 @@ static npy_uint64 EXP_Table_tail[32] = {
 #define NPY_COEFF_INVF7_SINEf -0x1.a06bbap-13f
 #define NPY_COEFF_INVF9_SINEf 0x1.7d3bbcp-19f
 
+/*
+ * Lookup table of log(c_k)
+ */
+static npy_uint64 LOG_TABLE_TOP[64] = {
+    0x0000000000000000,
+    0x3F8FC0A8B1000000,
+    0x3F9F829B0E780000,
+    0x3FA77458F6340000,
+    0x3FAF0A30C0100000,
+    0x3FB341D7961C0000,
+    0x3FB6F0D28AE60000,
+    0x3FBA926D3A4A0000,
+    0x3FBE27076E2A0000,
+    0x3FC0D77E7CD10000,
+    0x3FC29552F8200000,
+    0x3FC44D2B6CCB0000,
+    0x3FC5FF3070A80000,
+    0x3FC7AB8902110000,
+    0x3FC9525A9CF40000,
+    0x3FCAF3C94E810000,
+    0x3FCC8FF7C79B0000,
+    0x3FCE27076E2B0000,
+    0x3FCFB9186D5E0000,
+    0x3FD0A324E2738000,
+    0x3FD1675CABAB8000,
+    0x3FD22941FBCF8000,
+    0x3FD2E8E2BAE10000,
+    0x3FD3A64C55698000,
+    0x3FD4618BC21C8000,
+    0x3FD51AAD872E0000,
+    0x3FD5D1BDBF580000,
+    0x3FD686C81E9B0000,
+    0x3FD739D7F6BC0000,
+    0x3FD7EAF83B828000,
+    0x3FD89A3386C18000,
+    0x3FD947941C210000,
+    0x3FD9F323ECBF8000,
+    0x3FDA9CEC9A9A0000,
+    0x3FDB44F77BCC8000,
+    0x3FDBEB4D9DA70000,
+    0x3FDC8FF7C79A8000,
+    0x3FDD32FE7E010000,
+    0x3FDDD46A04C20000,
+    0x3FDE744261D68000,
+    0x3FDF128F5FAF0000,
+    0x3FDFAF588F790000,
+    0x3FE02552A5A5C000,
+    0x3FE0723E5C1CC000,
+    0x3FE0BE72E4254000,
+    0x3FE109F39E2D4000,
+    0x3FE154C3D2F4C000,
+    0x3FE19EE6B467C000,
+    0x3FE1E85F5E704000,
+    0x3FE23130D7BEC000,
+    0x3FE2795E1289C000,
+    0x3FE2C0E9ED448000,
+    0x3FE307D7334F0000,
+    0x3FE34E289D9D0000,
+    0x3FE393E0D3564000,
+    0x3FE3D9026A714000,
+    0x3FE41D8FE8468000,
+    0x3FE4618BC21C4000,
+    0x3FE4A4F85DB04000,
+    0x3FE4E7D811B74000,
+    0x3FE52A2D265BC000,
+    0x3FE56BF9D5B40000,
+    0x3FE5AD404C358000,
+    0x3FE5EE02A9240000,
+};
+
+static npy_uint64 LOG_TABLE_TAIL[64] = {
+    0x0000000000000000,
+    0xBD5FE0E183092C59,
+    0x3D2980267C7E09E4,
+    0xBD62303B9CB0D5E1,
+    0x3D662A6617CC9717,
+    0xBD4717B6B33E44F8,
+    0xBD62968C836CC8C2,
+    0x3D6AAC6CA17A4554,
+    0x3D6E5CBD3D50FFFC,
+    0xBD6C69A65A23A170,
+    0xBD35B967F4471DFC,
+    0x3D6F4799F4F6543E,
+    0xBD6B0B0DE3077D7E,
+    0xBD537B720E4A694B,
+    0x3D65AD1D904C1D4E,
+    0xBD600349CC67F9B2,
+    0xBD697794F689F843,
+    0xBD3A342C2AF0003C,
+    0x3D5F1546AAA3361C,
+    0x3D50E35F73F7A018,
+    0x3D630701CE63EAB9,
+    0xBD3A6976F5EB0963,
+    0x3D5D309C2CC91A85,
+    0xBD6D0B1C68651946,
+    0xBD609EC17A426426,
+    0xBD3F4BD8DB0A7CC1,
+    0x3D4394A11B1C1EE4,
+    0x3D54AEC442BE1015,
+    0xBD67FCB18ED9D603,
+    0x3D67E1B259D2F3DA,
+    0xBD6ED2A52C73BF78,
+    0x3D56FABA4CDD147D,
+    0x3D584BF2B68D766F,
+    0x3D40931A909FEA5E,
+    0x3D4EC5197DDB55D3,
+    0x3D5B7BF7861D37AC,
+    0x3D5A21AC25DB1EF3,
+    0xBD542A9E21373414,
+    0xBD6DAFA08CECADB1,
+    0x3D3E1F8DF68DBCF3,
+    0x3D3BB2CD720EC44C,
+    0xBD49C24CA098362B,
+    0x3D60FEC69C695D7F,
+    0x3D6F404E57963891,
+    0xBD657D49676844CC,
+    0x3D592DFBC7D93617,
+    0x3D65E9A98F33A396,
+    0x3D52DD98B97BAEF0,
+    0x3D1A07BD8B34BE7C,
+    0xBD17AFA4392F1BA7,
+    0xBD5DCA290F818480,
+    0x3D5D1772F5386374,
+    0x3D60BE1FB590A1F5,
+    0xBD6E2CE9146D271A,
+    0xBD65E6563BBD9FC9,
+    0x3D66FAA404263D0B,
+    0xBD5AA33736867A17,
+    0x3D6EC27D0B7B37B3,
+    0xBD244FDD840B8591,
+    0x3D6BB09CB0985646,
+    0x3D46ABB9DF22BC57,
+    0xBD58CD7DC73BD194,
+    0x3D6F2CFB29AAA5F0,
+    0x3D66757006095FD2,
+};
+
+#define NPY_TANG_LOG_A1 0x1.55555555554e6p-4
+#define NPY_TANG_LOG_A2 0x1.9999999bac6d4p-7
+#define NPY_TANG_LOG_A3 0x1.2492307f1519fp-9
+#define NPY_TANG_LOG_A4 0x1.c8034c85dfffp-12
+
+#define NPY_TANG_LOG_LN2HI 0x1.62e42fefa4p-1
+#define NPY_TANG_LOG_LN2LO -0x1.8432a1b0e2634p-43
+
 #endif