lavr: x86: optimized 2-channel s16p to s16 conversion

libavresample/x86/audio_convert.asm

@@ -223,6 +223,43 @@ INIT_YMM avx
 CONV_FLT_TO_S32
 %endif
 
+;------------------------------------------------------------------------------
+; void ff_conv_s16p_to_s16_2ch(int16_t *dst, int16_t *const *src, int len,
+;                              int channels);
+;------------------------------------------------------------------------------
+
+%macro CONV_S16P_TO_S16_2CH 0
+cglobal conv_s16p_to_s16_2ch, 3,4,5, dst, src0, len, src1
+    mov       src1q, [src0q+gprsize]
+    mov       src0q, [src0q        ]
+    lea        lenq, [2*lend]
+    add       src0q, lenq
+    add       src1q, lenq
+    lea        dstq, [dstq+2*lenq]
+    neg        lenq
+.loop
+    mova         m0, [src0q+lenq       ]
+    mova         m1, [src1q+lenq       ]
+    mova         m2, [src0q+lenq+mmsize]
+    mova         m3, [src1q+lenq+mmsize]
+    SBUTTERFLY2  wd, 0, 1, 4
+    SBUTTERFLY2  wd, 2, 3, 4
+    mova  [dstq+2*lenq+0*mmsize], m0
+    mova  [dstq+2*lenq+1*mmsize], m1
+    mova  [dstq+2*lenq+2*mmsize], m2
+    mova  [dstq+2*lenq+3*mmsize], m3
+    add        lenq, 2*mmsize
+    jl .loop
+    REP_RET
+%endmacro
+
+INIT_XMM sse2
+CONV_S16P_TO_S16_2CH
+%if HAVE_AVX
+INIT_XMM avx
+CONV_S16P_TO_S16_2CH
+%endif
+
 ;-----------------------------------------------------------------------------
 ; void ff_conv_fltp_to_flt_6ch(float *dst, float *const *src, int len,
 ;                              int channels);

libavresample/x86/audio_convert_init.c

@@ -22,6 +22,8 @@
 #include "libavutil/cpu.h"
 #include "libavresample/audio_convert.h"
 
+/* flat conversions */
+
 extern void ff_conv_s16_to_s32_sse2(int16_t *dst, const int32_t *src, int len);
 
 extern void ff_conv_s16_to_flt_sse2(float *dst, const int16_t *src, int len);
@@ -38,6 +40,13 @@ extern void ff_conv_flt_to_s16_sse2(int16_t *dst, const float *src, int len);
 extern void ff_conv_flt_to_s32_sse2(int32_t *dst, const float *src, int len);
 extern void ff_conv_flt_to_s32_avx (int32_t *dst, const float *src, int len);
 
+/* interleave conversions */
+
+extern void ff_conv_s16p_to_s16_2ch_sse2(int16_t *dst, int16_t *const *src,
+                                         int len, int channels);
+extern void ff_conv_s16p_to_s16_2ch_avx (int16_t *dst, int16_t *const *src,
+                                         int len, int channels);
+
 extern void ff_conv_fltp_to_flt_6ch_mmx (float *dst, float *const *src, int len,
                                          int channels);
 extern void ff_conv_fltp_to_flt_6ch_sse4(float *dst, float *const *src, int len,
@@ -71,6 +80,8 @@ av_cold void ff_audio_convert_init_x86(AudioConvert *ac)
                                   0, 16, 16, "SSE2", ff_conv_flt_to_s16_sse2);
         ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S32, AV_SAMPLE_FMT_FLT,
                                   0, 16, 16, "SSE2", ff_conv_flt_to_s32_sse2);
+        ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S16P,
+                                  2, 16, 16, "SSE2", ff_conv_s16p_to_s16_2ch_sse2);
     }
     if (mm_flags & AV_CPU_FLAG_SSE4 && HAVE_SSE) {
         ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S16,
@@ -83,6 +94,8 @@ av_cold void ff_audio_convert_init_x86(AudioConvert *ac)
                                   0, 32, 16, "AVX", ff_conv_s32_to_flt_avx);
         ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S32, AV_SAMPLE_FMT_FLT,
                                   0, 32, 32, "AVX", ff_conv_flt_to_s32_avx);
+        ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S16P,
+                                  2, 16, 16, "AVX", ff_conv_s16p_to_s16_2ch_avx);
         ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_FLTP,
                                   6, 16, 4, "AVX", ff_conv_fltp_to_flt_6ch_avx);
     }