[FFmpeg-cvslog] x86/flacdsp: add a SSE4 version of lpc16

[James Almer]

ffmpeg | branch: master | James Almer <jamrial at gmail.com> | Sat May 11 23:57:57 2024 -0300| [c00c8679ed35ccf053c3d0602932dda3adac752e] | committer: James Almer

x86/flacdsp: add a SSE4 version of lpc16

flac_lpc_16_13_c: 2841.3
flac_lpc_16_13_sse4: 2151.8
flac_lpc_16_16_c: 3382.8
flac_lpc_16_16_sse4: 2228.3
flac_lpc_16_29_c: 5800.3
flac_lpc_16_29_sse4: 3727.3
flac_lpc_16_32_c: 5972.8
flac_lpc_16_32_sse4: 4052.3

Signed-off-by: James Almer <jamrial at gmail.com>

> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=c00c8679ed35ccf053c3d0602932dda3adac752e
---

 libavcodec/x86/flacdsp.asm    | 13 +++++++------
 libavcodec/x86/flacdsp_init.c |  3 +++
 2 files changed, 10 insertions(+), 6 deletions(-)

diff --git a/libavcodec/x86/flacdsp.asm b/libavcodec/x86/flacdsp.asm
index 4b2fd65435..f38eb7db76 100644
--- a/libavcodec/x86/flacdsp.asm
+++ b/libavcodec/x86/flacdsp.asm
@@ -38,9 +38,9 @@ SECTION .text
 %endif
 %endmacro
 
-%macro LPC_32 1
+%macro LPC_32 3
 INIT_XMM %1
-cglobal flac_lpc_32, 5,6,5, decoded, coeffs, pred_order, qlevel, len, j
+cglobal flac_lpc_%2, 5,6,5, decoded, coeffs, pred_order, qlevel, len, j
     sub    lend, pred_orderd
     jle .ret
     movsxdifnidn pred_orderq, pred_orderd
@@ -67,14 +67,14 @@ ALIGN 16
     jl .loop_order
 .end_order:
     PMACSDQL m2, m0, m1, m2, m0
-    psrlq  m2, m4
+    %3     m2, m4
     movd   m0, [decodedq]
     paddd  m0, m2
     movd   [decodedq], m0
     sub  lend, 2
     jl .ret
     PMACSDQL m3, m1, m0, m3, m1
-    psrlq  m3, m4
+    %3     m3, m4
     movd   m1, [decodedq+4]
     paddd  m1, m3
     movd   [decodedq+4], m1
@@ -83,10 +83,11 @@ ALIGN 16
     RET
 %endmacro
 
+LPC_32 sse4, 16, psrad
+LPC_32 sse4, 32, psrlq
 %if HAVE_XOP_EXTERNAL
-LPC_32 xop
+LPC_32 xop,  32, psrlq
 %endif
-LPC_32 sse4
 
 ;----------------------------------------------------------------------------------
 ;void ff_flac_decorrelate_[lrm]s_16_sse2(uint8_t **out, int32_t **in, int channels,
diff --git a/libavcodec/x86/flacdsp_init.c b/libavcodec/x86/flacdsp_init.c
index 87daed7005..dee4bf88fc 100644
--- a/libavcodec/x86/flacdsp_init.c
+++ b/libavcodec/x86/flacdsp_init.c
@@ -23,6 +23,8 @@
 #include "libavutil/x86/cpu.h"
 #include "config.h"
 
+void ff_flac_lpc_16_sse4(int32_t *samples, const int coeffs[32], int order,
+                         int qlevel, int len);
 void ff_flac_lpc_32_sse4(int32_t *samples, const int coeffs[32], int order,
                          int qlevel, int len);
 void ff_flac_lpc_32_xop(int32_t *samples, const int coeffs[32], int order,
@@ -93,6 +95,7 @@ av_cold void ff_flacdsp_init_x86(FLACDSPContext *c, enum AVSampleFormat fmt, int
         }
     }
     if (EXTERNAL_SSE4(cpu_flags)) {
+        c->lpc16 = ff_flac_lpc_16_sse4;
         c->lpc32 = ff_flac_lpc_32_sse4;
     }
     if (EXTERNAL_AVX(cpu_flags)) {