added simd_asin function

Geolm · Jan 11, 2024 · 34467c8 · 34467c8
1 parent 8e46e02
commit 34467c8
Show file tree

Hide file tree

Showing 3 changed files with 65 additions and 0 deletions.
diff --git a/simd_math.c b/simd_math.c
@@ -252,5 +252,38 @@ void simd_sincos(simd_vector x, simd_vector* s, simd_vector* c)
     *c = simd_xor(xmm2, sign_bit_cos);
 }
 
+//-----------------------------------------------------------------------------
+// based on https://github.com/jeremybarnes/cephes/blob/master/single/asinf.c
+simd_vector simd_asin(simd_vector xx)
+{
+    simd_vector x = xx;
+    simd_vector sign = simd_sign(xx);
+    simd_vector a  = simd_abs(x);
+    simd_vector greater_one = simd_cmp_gt(a, simd_splat(1.f));
+    simd_vector small_value = simd_cmp_lt(a, simd_splat(1.0e-4f));
+
+    simd_vector z1 = simd_mul(simd_splat(.5f), simd_sub(simd_splat(1.f), a));
+    simd_vector z2 = simd_mul(a, a);
+    simd_vector flag = simd_cmp_gt(a, simd_splat(.5f));
+    simd_vector z = simd_select(z2, z1, flag);
+
+    x = simd_select(a, simd_sqrt(z), flag);
+
+    simd_vector tmp = simd_fmad(simd_splat(4.2163199048E-2f), z, simd_splat(2.4181311049E-2f));
+    tmp = simd_fmad(tmp, z, simd_splat(4.5470025998E-2f));
+    tmp = simd_fmad(tmp, z, simd_splat(7.4953002686E-2f));
+    tmp = simd_fmad(tmp, z, simd_splat(1.6666752422E-1f));
+    tmp = simd_mul(tmp, z);
+    z = simd_fmad(tmp, x, x);
+
+    tmp = simd_add(z, z);
+    tmp = simd_sub(simd_splat(SIMD_MATH_PI2), tmp);
+    z = simd_select(z, tmp, flag);
+
+    z = simd_select(z, a, small_value);
+    z = simd_select(z, simd_splat_zero(), greater_one);
+    z = simd_mul(z, sign);
+    return z;
+}
 
 
diff --git a/simd_math.h b/simd_math.h
@@ -22,6 +22,7 @@ static simd_vector simd_sin(simd_vector x); // output equal to sinf()
 static simd_vector simd_cos(simd_vector a); // output equal to cosf()
 void simd_sincos(simd_vector x, simd_vector* s, simd_vector* c); // output equal to sinf()/cosf()
 static simd_vector simd_acos(simd_vector x); // // max error : 0.000068
+simd_vector simd_asin(simd_vector x);
 simd_vector simd_atan2(simd_vector x, simd_vector y); // max error : 0.000002
 simd_vector simd_log(simd_vector x); // output equal to logf()
 simd_vector simd_exp(simd_vector x); // output equal to expf()

diff --git a/tests/test_simd_math.h b/tests/test_simd_math.h
@@ -102,6 +102,36 @@ TEST approx_sin(void)
     PASS();
 }
 
+TEST arcsin(void)
+{
+    float array[NUM_ELEMENTS];
+    float result[NUM_ELEMENTS];
+    float step = 2.f / (float) NUM_ELEMENTS;
+
+    for(int i=0; i<NUM_ELEMENTS; ++i)
+    {
+        array[i] = (step * (float)i) - 1.f;
+        result[i] = asinf(array[i]);
+    }
+
+    simd_vector epsilon = simd_splat(FLT_MAX);
+    simd_vector max_error = simd_splat_zero();
+
+    for(int i=0; i<NUM_VECTORS; ++i)
+    {
+        simd_vector v_approx = simd_asin(simd_load_offset(array,  i));
+        simd_vector v_result = simd_load_offset(result, i);
+
+        ASSERT(simd_all(simd_equal(v_approx, v_result, epsilon)));
+
+        max_error = simd_max(max_error, simd_abs_diff(v_approx, v_result));
+    }
+
+    printf("simd_asin max error : %f\n", simd_hmax(max_error));
+
+    PASS();
+}
+
 TEST arcos(void)
 {
     float array[NUM_ELEMENTS];
@@ -371,6 +401,7 @@ SUITE(trigonometry)
     RUN_TEST(sinus);
     RUN_TEST(approx_sin);
     RUN_TEST(arcos);
+    RUN_TEST(arcsin);
     RUN_TEST(approx_arcos);
     RUN_TEST(arctan2);
     RUN_TEST(sinuscosinus);