Performance improvement: replaced vgatherdps instructions with multiple

scalar loads

Author: sadko4u

include/private/dsp/arch/x86/avx512/fft.h

@@ -29,22 +29,19 @@
 #include <private/dsp/arch/x86/avx512/fft/const.h>
 #include <private/dsp/arch/x86/avx512/fft/normalize.h>
 #include <private/dsp/arch/x86/avx512/fft/butterfly.h>
-#include <private/dsp/arch/x86/avx512/fft/rbits.h>
 
 #define FFT_SCRAMBLE_SELF_DIRECT_NAME   scramble_self_direct8
 #define FFT_SCRAMBLE_SELF_REVERSE_NAME  scramble_self_reverse8
 #define FFT_SCRAMBLE_COPY_DIRECT_NAME   scramble_copy_direct8
 #define FFT_SCRAMBLE_COPY_REVERSE_NAME  scramble_copy_reverse8
 #define FFT_TYPE                        uint8_t
-#define FFT_REVERSE_BITS                FFT_REVERSE_BITS8
 #include <private/dsp/arch/x86/avx512/fft/scramble.h>
 
 #define FFT_SCRAMBLE_SELF_DIRECT_NAME   scramble_self_direct16
 #define FFT_SCRAMBLE_SELF_REVERSE_NAME  scramble_self_reverse16
 #define FFT_SCRAMBLE_COPY_DIRECT_NAME   scramble_copy_direct16
 #define FFT_SCRAMBLE_COPY_REVERSE_NAME  scramble_copy_reverse16
 #define FFT_TYPE                        uint16_t
-#define FFT_REVERSE_BITS                FFT_REVERSE_BITS16
 #include <private/dsp/arch/x86/avx512/fft/scramble.h>
 
 namespace lsp
@@ -156,10 +153,10 @@ namespace lsp
             }
             else
             {
-                if (rank <= 8)
-                    scramble_copy_direct8(dst_re, dst_im, src_re, src_im, rank);
+                if (rank <= 13)
+                    scramble_copy_direct8(dst_re, dst_im, src_re, src_im, rank - 5);
                 else
-                    scramble_copy_direct16(dst_re, dst_im, src_re, src_im, rank);
+                    scramble_copy_direct16(dst_re, dst_im, src_re, src_im, rank - 5);
             }
 
             for (size_t i=4; i < rank; ++i)
@@ -186,10 +183,10 @@ namespace lsp
             }
             else
             {
-                if (rank <= 8)
-                    scramble_copy_reverse8(dst_re, dst_im, src_re, src_im, rank);
+                if (rank <= 13)
+                    scramble_copy_reverse8(dst_re, dst_im, src_re, src_im, rank - 5);
                 else
-                    scramble_copy_reverse16(dst_re, dst_im, src_re, src_im, rank);
+                    scramble_copy_reverse16(dst_re, dst_im, src_re, src_im, rank - 5);
             }
 
             for (size_t i=4; i < rank; ++i)

include/private/dsp/arch/x86/avx512/fft/const.h

@@ -103,7 +103,7 @@ namespace lsp
             LSP_DSP_VEC16(0.9999999954041073), LSP_DSP_VEC16(0.0000958737990960),   // rank = 18
         };
 
-        static const uint32_t FFT_SCRAMBLE_DIRECT_INDICES[] __lsp_aligned64 =
+        static const uint32_t FFT_SCRAMBLE_INDICES[] __lsp_aligned64 =
         {
             // butterfly 1
             0,  2,  4,  6,  8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30,          // + 0x000
@@ -119,15 +119,8 @@ namespace lsp
             // butterfly 4
             0,  1,  2,  3, 16, 17, 18, 19,  8,  9, 10, 11, 24, 25, 26, 27,          // + 0x200
             4,  5,  6,  7, 20, 21, 22, 23, 12, 13, 14, 15, 28, 29, 30, 31,          // + 0x240
-        };
-
-        static const uint32_t FFT_RBITS[] __lsp_aligned64 =
-        {
-            0xff00ff00,             // step2
-            0xf0f0f0f0,             // step3
-            0xcccccccc,             // step4
-            0xaaaaaaaa,             // step5
-            32,                     // increment
+            // permute 16x
+            0,  8,  4, 12,  2, 10,  6, 14,  1,  9,  5, 13,  3, 11,  7, 15,          // + 0x280
         };
 
     } /* namespace avx512 */