x86: af_volume: add SSE2/SSSE3/AVX-optimized s32 volume scaling

2 files changed, 105 insertions, 3 deletions

diff --git a/libavfilter/x86/af_volume.asm b/libavfilter/x86/af_volume.asm
index 922ad5dcc4..4e5ad2258c 100644
--- a/libavfilter/x86/af_volume.asm
+++ b/libavfilter/x86/af_volume.asm
@@ -19,12 +19,15 @@
 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 ;******************************************************************************
 
-%include "libavutil/x86/x86inc.asm"
+%include "libavutil/x86/x86util.asm"
 
 SECTION_RODATA 32
 
-pw_1:   times 8 dw 1
-pw_128: times 8 dw 128
+pd_1_256:     times 4 dq 0x3F70000000000000
+pd_int32_max: times 4 dq 0x41DFFFFFFFC00000
+pw_1:         times 8 dw 1
+pw_128:       times 8 dw 128
+pq_128:       times 2 dq 128
 
 SECTION_TEXT
 
@@ -54,3 +57,82 @@ cglobal scale_samples_s16, 4,4,4, dst, src, len, volume
     sub       lenq, mmsize
     jge .loop
     REP_RET
+
+;------------------------------------------------------------------------------
+; void ff_scale_samples_s32(uint8_t *dst, const uint8_t *src, int len,
+;                           int volume)
+;------------------------------------------------------------------------------
+
+%macro SCALE_SAMPLES_S32 0
+cglobal scale_samples_s32, 4,4,4, dst, src, len, volume
+%if ARCH_X86_32 && cpuflag(avx)
+    vbroadcastss   xmm2, volumem
+%else
+    movd           xmm2, volumed
+    pshufd         xmm2, xmm2, 0
+%endif
+    CVTDQ2PD         m2, xmm2
+    mulpd            m2, m2, [pd_1_256]
+    mova             m3, [pd_int32_max]
+    lea            lenq, [lend*4-mmsize]
+.loop:
+    CVTDQ2PD         m0, [srcq+lenq         ]
+    CVTDQ2PD         m1, [srcq+lenq+mmsize/2]
+    mulpd            m0, m0, m2
+    mulpd            m1, m1, m2
+    minpd            m0, m0, m3
+    minpd            m1, m1, m3
+    cvtpd2dq       xmm0, m0
+    cvtpd2dq       xmm1, m1
+%if cpuflag(avx)
+    vmovdqa [dstq+lenq         ], xmm0
+    vmovdqa [dstq+lenq+mmsize/2], xmm1
+%else
+    movq    [dstq+lenq         ], xmm0
+    movq    [dstq+lenq+mmsize/2], xmm1
+%endif
+    sub            lenq, mmsize
+    jge .loop
+    REP_RET
+%endmacro
+
+INIT_XMM sse2
+%define CVTDQ2PD cvtdq2pd
+SCALE_SAMPLES_S32
+%define CVTDQ2PD vcvtdq2pd
+INIT_YMM avx
+SCALE_SAMPLES_S32
+%undef CVTDQ2PD
+
+; NOTE: This is not bit-identical with the C version because it clips to
+;       [-INT_MAX, INT_MAX] instead of [INT_MIN, INT_MAX]
+
+INIT_XMM ssse3, atom
+cglobal scale_samples_s32, 4,4,8, dst, src, len, volume
+    movd        m4, volumem
+    pshufd      m4, m4, 0
+    mova        m5, [pq_128]
+    pxor        m6, m6
+    lea       lenq, [lend*4-mmsize]
+.loop:
+    ; src[i] = av_clipl_int32((src[i] * volume + 128) >> 8);
+    mova        m7, [srcq+lenq]
+    pabsd       m3, m7
+    pshufd      m0, m3, q0100
+    pshufd      m1, m3, q0302
+    pmuludq     m0, m4
+    pmuludq     m1, m4
+    paddq       m0, m5
+    paddq       m1, m5
+    psrlq       m0, 7
+    psrlq       m1, 7
+    shufps      m2, m0, m1, q3131
+    shufps      m0, m0, m1, q2020
+    pcmpgtd     m2, m6
+    por         m0, m2
+    psrld       m0, 1
+    psignd      m0, m7
+    mova  [dstq+lenq], m0
+    sub       lenq, mmsize
+    jge .loop
+    REP_RET
diff --git a/libavfilter/x86/af_volume_init.c b/libavfilter/x86/af_volume_init.c
index 00103df216..02bedd2c57 100644
--- a/libavfilter/x86/af_volume_init.c
+++ b/libavfilter/x86/af_volume_init.c
@@ -25,6 +25,13 @@
 void ff_scale_samples_s16_sse2(uint8_t *dst, const uint8_t *src, int len,
                                int volume);
 
+void ff_scale_samples_s32_sse2(uint8_t *dst, const uint8_t *src, int len,
+                               int volume);
+void ff_scale_samples_s32_ssse3_atom(uint8_t *dst, const uint8_t *src, int len,
+                                     int volume);
+void ff_scale_samples_s32_avx(uint8_t *dst, const uint8_t *src, int len,
+                              int volume);
+
 void ff_volume_init_x86(VolumeContext *vol)
 {
     int mm_flags = av_get_cpu_flags();
@@ -35,5 +42,18 @@ void ff_volume_init_x86(VolumeContext *vol)
             vol->scale_samples = ff_scale_samples_s16_sse2;
             vol->samples_align = 8;
         }
+    } else if (sample_fmt == AV_SAMPLE_FMT_S32) {
+        if (EXTERNAL_SSE2(mm_flags)) {
+            vol->scale_samples = ff_scale_samples_s32_sse2;
+            vol->samples_align = 4;
+        }
+        if (EXTERNAL_SSSE3(mm_flags) && mm_flags & AV_CPU_FLAG_ATOM) {
+            vol->scale_samples = ff_scale_samples_s32_ssse3_atom;
+            vol->samples_align = 4;
+        }
+        if (EXTERNAL_AVX(mm_flags)) {
+            vol->scale_samples = ff_scale_samples_s32_avx;
+            vol->samples_align = 8;
+        }
     }
 }