TF_SRGB: use HWY_REP4 (#3587)

libjxl · May 15, 2024 · 64b12f0 · 64b12f0
1 parent 715b442
commit 64b12f0
Showing 1 changed file with 12 additions and 24 deletions.
diff --git a/lib/jxl/cms/transfer_functions-inl.h b/lib/jxl/cms/transfer_functions-inl.h
@@ -199,20 +199,14 @@ class TF_SRGB {
     // TODO(janwas): range reduction
     // Computed via af_cheb_rational (k=100); replicated 4x.
     HWY_ALIGN constexpr float p[(4 + 1) * 4] = {
-        2.200248328e-04f, 2.200248328e-04f, 2.200248328e-04f, 2.200248328e-04f,
-        1.043637593e-02f, 1.043637593e-02f, 1.043637593e-02f, 1.043637593e-02f,
-        1.624820318e-01f, 1.624820318e-01f, 1.624820318e-01f, 1.624820318e-01f,
-        7.961564959e-01f, 7.961564959e-01f, 7.961564959e-01f, 7.961564959e-01f,
-        8.210152774e-01f, 8.210152774e-01f, 8.210152774e-01f, 8.210152774e-01f,
+        HWY_REP4(2.200248328e-04f), HWY_REP4(1.043637593e-02f),
+        HWY_REP4(1.624820318e-01f), HWY_REP4(7.961564959e-01f),
+        HWY_REP4(8.210152774e-01f),
     };
     HWY_ALIGN constexpr float q[(4 + 1) * 4] = {
-        2.631846970e-01f,  2.631846970e-01f,  2.631846970e-01f,
-        2.631846970e-01f,  1.076976492e+00f,  1.076976492e+00f,
-        1.076976492e+00f,  1.076976492e+00f,  4.987528350e-01f,
-        4.987528350e-01f,  4.987528350e-01f,  4.987528350e-01f,
-        -5.512498495e-02f, -5.512498495e-02f, -5.512498495e-02f,
-        -5.512498495e-02f, 6.521209011e-03f,  6.521209011e-03f,
-        6.521209011e-03f,  6.521209011e-03f,
+        HWY_REP4(2.631846970e-01f), HWY_REP4(1.076976492e+00f),
+        HWY_REP4(4.987528350e-01f), HWY_REP4(-5.512498495e-02f),
+        HWY_REP4(6.521209011e-03f),
     };
     const V linear = Mul(x, Set(d, kLowDivInv));
     const V poly = EvalRationalPolynomial(d, x, p, q);
@@ -231,20 +225,14 @@ class TF_SRGB {
 
     // Computed via af_cheb_rational (k=100); replicated 4x.
     HWY_ALIGN constexpr float p[(4 + 1) * 4] = {
-        -5.135152395e-04f, -5.135152395e-04f, -5.135152395e-04f,
-        -5.135152395e-04f, 5.287254571e-03f,  5.287254571e-03f,
-        5.287254571e-03f,  5.287254571e-03f,  3.903842876e-01f,
-        3.903842876e-01f,  3.903842876e-01f,  3.903842876e-01f,
-        1.474205315e+00f,  1.474205315e+00f,  1.474205315e+00f,
-        1.474205315e+00f,  7.352629620e-01f,  7.352629620e-01f,
-        7.352629620e-01f,  7.352629620e-01f,
+        HWY_REP4(-5.135152395e-04f), HWY_REP4(5.287254571e-03f),
+        HWY_REP4(3.903842876e-01f),  HWY_REP4(1.474205315e+00f),
+        HWY_REP4(7.352629620e-01f),
     };
     HWY_ALIGN constexpr float q[(4 + 1) * 4] = {
-        1.004519624e-02f, 1.004519624e-02f, 1.004519624e-02f, 1.004519624e-02f,
-        3.036675394e-01f, 3.036675394e-01f, 3.036675394e-01f, 3.036675394e-01f,
-        1.340816930e+00f, 1.340816930e+00f, 1.340816930e+00f, 1.340816930e+00f,
-        9.258482155e-01f, 9.258482155e-01f, 9.258482155e-01f, 9.258482155e-01f,
-        2.424867759e-02f, 2.424867759e-02f, 2.424867759e-02f, 2.424867759e-02f,
+        HWY_REP4(1.004519624e-02f), HWY_REP4(3.036675394e-01f),
+        HWY_REP4(1.340816930e+00f), HWY_REP4(9.258482155e-01f),
+        HWY_REP4(2.424867759e-02f),
     };
     const V linear = Mul(x, Set(d, kLowDiv));
     const V poly = EvalRationalPolynomial(d, Sqrt(x), p, q);