Submitted AVX-optimized pmix functions

Author: sadko4u

include/private/dsp/arch/x86/avx/pmath/pmix.h

@@ -496,6 +496,11 @@
                 EXPORT1(clamp_kk1);
                 EXPORT1(clamp_kk2);
 
+                EXPORT1(pmix_v1);
+                EXPORT1(pmix_v2);
+                EXPORT1(pmix_k1);
+                EXPORT1(pmix_k2);
+
                 // FMA3 support?
                 if (f->features & CPU_OPTION_FMA3)
                 {
@@ -620,6 +625,11 @@
 
                     CEXPORT2(favx, check_point3d_on_triangle_p3p, check_point3d_on_triangle_p3p_fma3);
                     CEXPORT2(favx, check_point3d_on_triangle_pvp, check_point3d_on_triangle_pvp_fma3);
+
+                    CEXPORT2(favx, pmix_v1, pmix_v1_fma3);
+                    CEXPORT2(favx, pmix_v2, pmix_v2_fma3);
+                    CEXPORT2(favx, pmix_k1, pmix_k1_fma3);
+                    CEXPORT2(favx, pmix_k2, pmix_k2_fma3);
                 }
             }
 

src/main/x86/avx.cpp

@@ -110,8 +110,8 @@ PTEST_BEGIN("dsp.pmath", pmix_k1, 5, 1000)
 
             CALL(generic::pmix_k1);
             IF_ARCH_X86(CALL(sse::pmix_k1));
-//            IF_ARCH_X86(CALL(avx::pmix_k1));
-//            IF_ARCH_X86(CALL(avx::pmix_k1_fma3));
+            IF_ARCH_X86(CALL(avx::pmix_k1));
+            IF_ARCH_X86(CALL(avx::pmix_k1_fma3));
 //            IF_ARCH_X86(CALL(avx512::pmix_k1));
 //            IF_ARCH_ARM(CALL(neon_d32::pmix_k1));
 //            IF_ARCH_AARCH64(CALL(asimd::pmix_k1));

src/test/ptest/pmath/pmix_k1.cpp

@@ -110,8 +110,8 @@ PTEST_BEGIN("dsp.pmath", pmix_k2, 5, 1000)
 
             CALL(generic::pmix_k2);
             IF_ARCH_X86(CALL(sse::pmix_k2));
-//            IF_ARCH_X86(CALL(avx::pmix_k2));
-//            IF_ARCH_X86(CALL(avx::pmix_k2_fma3));
+            IF_ARCH_X86(CALL(avx::pmix_k2));
+            IF_ARCH_X86(CALL(avx::pmix_k2_fma3));
 //            IF_ARCH_X86(CALL(avx512::pmix_k2));
 //            IF_ARCH_ARM(CALL(neon_d32::pmix_k2));
 //            IF_ARCH_AARCH64(CALL(asimd::pmix_k2));

src/test/ptest/pmath/pmix_k2.cpp

@@ -114,8 +114,8 @@ PTEST_BEGIN("dsp.pmath", pmix_v1, 5, 1000)
 
             CALL(generic::pmix_v1);
             IF_ARCH_X86(CALL(sse::pmix_v1));
-//            IF_ARCH_X86(CALL(avx::pmix_v1));
-//            IF_ARCH_X86(CALL(avx::pmix_v1_fma3));
+            IF_ARCH_X86(CALL(avx::pmix_v1));
+            IF_ARCH_X86(CALL(avx::pmix_v1_fma3));
 //            IF_ARCH_X86(CALL(avx512::pmix_v1));
 //            IF_ARCH_ARM(CALL(neon_d32::pmix_v1));
 //            IF_ARCH_AARCH64(CALL(asimd::pmix_v1));

src/test/ptest/pmath/pmix_v1.cpp

@@ -112,8 +112,8 @@ PTEST_BEGIN("dsp.pmath", pmix_v2, 5, 1000)
 
             CALL(generic::pmix_v2);
             IF_ARCH_X86(CALL(sse::pmix_v2));
-//            IF_ARCH_X86(CALL(avx::pmix_v2));
-//            IF_ARCH_X86(CALL(avx::pmix_v2_fma3));
+            IF_ARCH_X86(CALL(avx::pmix_v2));
+            IF_ARCH_X86(CALL(avx::pmix_v2_fma3));
 //            IF_ARCH_X86(CALL(avx512::pmix_v2));
 //            IF_ARCH_ARM(CALL(neon_d32::pmix_v2));
 //            IF_ARCH_AARCH64(CALL(asimd::pmix_v2));

src/test/ptest/pmath/pmix_v2.cpp

@@ -120,8 +120,8 @@ UTEST_BEGIN("dsp.pmath", pmix_k1)
             call(#func, align, generic, func)
 
         IF_ARCH_X86(CALL(generic::pmix_k1, sse::pmix_k1, 16));
-//        IF_ARCH_X86(CALL(generic::pmix_k1, avx::pmix_k1, 32));
-//        IF_ARCH_X86(CALL(generic::pmix_k1, avx::pmix_k1_fma3, 32));
+        IF_ARCH_X86(CALL(generic::pmix_k1, avx::pmix_k1, 32));
+        IF_ARCH_X86(CALL(generic::pmix_k1, avx::pmix_k1_fma3, 32));
 //        IF_ARCH_X86(CALL(generic::pmix_k1, avx512::pmix_k1, 64));
 //        IF_ARCH_ARM(CALL(generic::pmix_k1, neon_d32::pmix_k1, 16));
 //        IF_ARCH_AARCH64(CALL(generic::pmix_k1, asimd::pmix_k1, 16));

src/test/utest/pmath/pmix_k1.cpp

@@ -120,8 +120,8 @@ UTEST_BEGIN("dsp.pmath", pmix_k2)
             call(#func, align, generic, func)
 
         IF_ARCH_X86(CALL(generic::pmix_k2, sse::pmix_k2, 16));
-//        IF_ARCH_X86(CALL(generic::pmix_k2, avx::pmix_k2, 32));
-//        IF_ARCH_X86(CALL(generic::pmix_k2, avx::pmix_k2_fma3, 32));
+        IF_ARCH_X86(CALL(generic::pmix_k2, avx::pmix_k2, 32));
+        IF_ARCH_X86(CALL(generic::pmix_k2, avx::pmix_k2_fma3, 32));
 //        IF_ARCH_X86(CALL(generic::pmix_k2, avx512::pmix_k2, 64));
 //        IF_ARCH_ARM(CALL(generic::pmix_k2, neon_d32::pmix_k2, 16));
 //        IF_ARCH_AARCH64(CALL(generic::pmix_k2, asimd::pmix_k1, 16));

src/test/utest/pmath/pmix_k2.cpp

@@ -123,8 +123,8 @@ UTEST_BEGIN("dsp.pmath", pmix_v1)
             call(#func, align, generic, func)
 
         IF_ARCH_X86(CALL(generic::pmix_v1, sse::pmix_v1, 16));
-//        IF_ARCH_X86(CALL(generic::pmix_v1, avx::pmix_v1, 32));
-//        IF_ARCH_X86(CALL(generic::pmix_v1, avx::pmix_v1_fma3, 32));
+        IF_ARCH_X86(CALL(generic::pmix_v1, avx::pmix_v1, 32));
+        IF_ARCH_X86(CALL(generic::pmix_v1, avx::pmix_v1_fma3, 32));
 //        IF_ARCH_X86(CALL(generic::pmix_v1, avx512::pmix_v1, 64));
 //        IF_ARCH_ARM(CALL(generic::pmix_v1, neon_d32::pmix_v1, 16));
 //        IF_ARCH_AARCH64(CALL(generic::pmix_v1, asimd::pmix_v1, 16));

src/test/utest/pmath/pmix_v1.cpp

@@ -123,8 +123,8 @@ UTEST_BEGIN("dsp.pmath", pmix_v2)
             call(#func, align, generic, func)
 
         IF_ARCH_X86(CALL(generic::pmix_v2, sse::pmix_v2, 16));
-//        IF_ARCH_X86(CALL(generic::pmix_v2, avx::pmix_v2, 32));
-//        IF_ARCH_X86(CALL(generic::pmix_v2, avx::pmix_v2_fma3, 32));
+        IF_ARCH_X86(CALL(generic::pmix_v2, avx::pmix_v2, 32));
+        IF_ARCH_X86(CALL(generic::pmix_v2, avx::pmix_v2_fma3, 32));
 //        IF_ARCH_X86(CALL(generic::pmix_v2, avx512::pmix_v2, 64));
 //        IF_ARCH_ARM(CALL(generic::pmix_v2, neon_d32::pmix_v2, 16));
 //        IF_ARCH_AARCH64(CALL(generic::pmix_v2, asimd::pmix_v2, 16));