[FFmpeg-devel] [PATCH 2/3] x86/ac3dsp: add ff_float_to_fixed24_avx2()

[James Almer]

Signed-off-by: James Almer <jamrial at gmail.com>
---
 libavcodec/ac3dsp.h          |  4 ++--
 libavcodec/ac3enc_template.c |  2 +-
 libavcodec/x86/ac3dsp.asm    | 28 ++++++++++++++++++++++++++--
 libavcodec/x86/ac3dsp_init.c |  4 ++++
 4 files changed, 33 insertions(+), 5 deletions(-)

diff --git a/libavcodec/ac3dsp.h b/libavcodec/ac3dsp.h
index a01bff3d11..25341f3396 100644
--- a/libavcodec/ac3dsp.h
+++ b/libavcodec/ac3dsp.h
@@ -47,9 +47,9 @@ typedef struct AC3DSPContext {
      * [-(1<<24),(1<<24)]
      *
      * @param dst destination array of int32_t.
-     *            constraints: 16-byte aligned
+     *            constraints: 32-byte aligned
      * @param src source array of float.
-     *            constraints: 16-byte aligned
+     *            constraints: 32-byte aligned
      * @param len number of elements to convert.
      *            constraints: multiple of 32 greater than zero
      */
diff --git a/libavcodec/ac3enc_template.c b/libavcodec/ac3enc_template.c
index be4ecebc9c..a16faea681 100644
--- a/libavcodec/ac3enc_template.c
+++ b/libavcodec/ac3enc_template.c
@@ -112,7 +112,7 @@ static void apply_channel_coupling(AC3EncodeContext *s)
 {
     LOCAL_ALIGNED_16(CoefType, cpl_coords,      [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
 #if AC3ENC_FLOAT
-    LOCAL_ALIGNED_16(int32_t, fixed_cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
+    LOCAL_ALIGNED_32(int32_t, fixed_cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
 #else
     int32_t (*fixed_cpl_coords)[AC3_MAX_CHANNELS][16] = cpl_coords;
 #endif
diff --git a/libavcodec/x86/ac3dsp.asm b/libavcodec/x86/ac3dsp.asm
index 42c8310462..e31c58e1c1 100644
--- a/libavcodec/x86/ac3dsp.asm
+++ b/libavcodec/x86/ac3dsp.asm
@@ -21,10 +21,10 @@
 
 %include "libavutil/x86/x86util.asm"
 
-SECTION_RODATA
+SECTION_RODATA 32
 
 ; 16777216.0f - used in ff_float_to_fixed24()
-pf_1_24: times 4 dd 0x4B800000
+pf_1_24: times 8 dd 0x4B800000
 
 ; used in ff_ac3_compute_mantissa_size()
 cextern ac3_bap_bits
@@ -128,6 +128,30 @@ cglobal float_to_fixed24, 3, 3, 9, dst, src, len
     jl .loop
     RET
 
+INIT_YMM avx2
+cglobal float_to_fixed24, 3, 3, 5, dst, src, len
+    movaps     m0, [pf_1_24]
+    shl      lenq, 2
+    add      srcq, lenq
+    add      dstq, lenq
+    neg      lenq
+.loop:
+    mulps      m1, m0, [srcq+lenq+mmsize*0]
+    mulps      m2, m0, [srcq+lenq+mmsize*1]
+    mulps      m3, m0, [srcq+lenq+mmsize*2]
+    mulps      m4, m0, [srcq+lenq+mmsize*3]
+    cvtps2dq   m1, m1
+    cvtps2dq   m2, m2
+    cvtps2dq   m3, m3
+    cvtps2dq   m4, m4
+    movdqa  [dstq+lenq+mmsize*0], m1
+    movdqa  [dstq+lenq+mmsize*1], m2
+    movdqa  [dstq+lenq+mmsize*2], m3
+    movdqa  [dstq+lenq+mmsize*3], m4
+    add      lenq, mmsize*4
+    jl .loop
+    RET
+
 ;------------------------------------------------------------------------------
 ; int ff_ac3_compute_mantissa_size(uint16_t mant_cnt[6][16])
 ;------------------------------------------------------------------------------
diff --git a/libavcodec/x86/ac3dsp_init.c b/libavcodec/x86/ac3dsp_init.c
index 43b3b4ac85..106121b5b9 100644
--- a/libavcodec/x86/ac3dsp_init.c
+++ b/libavcodec/x86/ac3dsp_init.c
@@ -27,6 +27,7 @@
 void ff_ac3_exponent_min_sse2  (uint8_t *exp, int num_reuse_blocks, int nb_coefs);
 
 void ff_float_to_fixed24_sse2 (int32_t *dst, const float *src, unsigned int len);
+void ff_float_to_fixed24_avx2 (int32_t *dst, const float *src, unsigned int len);
 
 int ff_ac3_compute_mantissa_size_sse2(uint16_t mant_cnt[6][16]);
 
@@ -48,6 +49,9 @@ av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c)
         if (!(cpu_flags & AV_CPU_FLAG_ATOM))
             c->extract_exponents = ff_ac3_extract_exponents_ssse3;
     }
+    if (EXTERNAL_AVX2_FAST(cpu_flags)) {
+        c->float_to_fixed24 = ff_float_to_fixed24_avx2;
+    }
 }
 
 #define DOWNMIX_FUNC_OPT(ch, opt)                                       \
-- 
2.42.1