swr: convert resample_common/linear_int16_mmx2/sse2 to yasm.

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

4 files changed, 187 insertions, 215 deletions

diff --git a/libswresample/resample_template.c b/libswresample/resample_template.c
index 1982992e8a..2a64f50038 100644
--- a/libswresample/resample_template.c
+++ b/libswresample/resample_template.c
@@ -44,17 +44,15 @@
 
 #elif    defined(TEMPLATE_RESAMPLE_FLT)
 
+#    define RENAME(N) N ## _float
 #    define FILTER_SHIFT 0
 #    define DELEM  float
 #    define FELEM  float
 #    define FELEM2 float
 #    define OUT(d, v) d = v
 
-#    if defined(TEMPLATE_RESAMPLE_FLT)
-#        define RENAME(N) N ## _float
-#    endif
-
 #elif defined(TEMPLATE_RESAMPLE_S32)
+
 #    define RENAME(N) N ## _int32
 #    define FILTER_SHIFT 30
 #    define DELEM  int32_t
@@ -65,10 +63,9 @@
 #    define OUT(d, v) v = (v + (1<<(FILTER_SHIFT-1)))>>FILTER_SHIFT;\
                       d = (uint64_t)(v + 0x80000000) > 0xFFFFFFFF ? (v>>63) ^ 0x7FFFFFFF : v
 
-#elif    defined(TEMPLATE_RESAMPLE_S16)      \
-      || defined(TEMPLATE_RESAMPLE_S16_MMX2) \
-      || defined(TEMPLATE_RESAMPLE_S16_SSE2)
+#elif    defined(TEMPLATE_RESAMPLE_S16)
 
+#    define RENAME(N) N ## _int16
 #    define FILTER_SHIFT 15
 #    define DELEM  int16_t
 #    define FELEM  int16_t
@@ -79,18 +76,6 @@
 #    define OUT(d, v) v = (v + (1<<(FILTER_SHIFT-1)))>>FILTER_SHIFT;\
                       d = (unsigned)(v + 32768) > 65535 ? (v>>31) ^ 32767 : v
 
-#    if defined(TEMPLATE_RESAMPLE_S16)
-#        define RENAME(N) N ## _int16
-#    elif defined(TEMPLATE_RESAMPLE_S16_MMX2)
-#        define COMMON_CORE COMMON_CORE_INT16_MMX2
-#        define LINEAR_CORE LINEAR_CORE_INT16_MMX2
-#        define RENAME(N) N ## _int16_mmx2
-#    elif defined(TEMPLATE_RESAMPLE_S16_SSE2)
-#        define COMMON_CORE COMMON_CORE_INT16_SSE2
-#        define LINEAR_CORE LINEAR_CORE_INT16_SSE2
-#        define RENAME(N) N ## _int16_sse2
-#    endif
-
 #endif
 
 #if DO_RESAMPLE_ONE
diff --git a/libswresample/x86/resample.asm b/libswresample/x86/resample.asm
index bb63fc7e8e..2fe03c846b 100644
--- a/libswresample/x86/resample.asm
+++ b/libswresample/x86/resample.asm
@@ -1,6 +1,7 @@
 ;******************************************************************************
 ;* Copyright (c) 2012 Michael Niedermayer
 ;* Copyright (c) 2014 James Almer <jamrial <at> gmail.com>
+;* Copyright (c) 2014 Ronald S. Bultje <rsbultje@gmail.com>
 ;*
 ;* This file is part of FFmpeg.
 ;*
@@ -48,18 +49,19 @@ endstruc
 
 SECTION_RODATA
 
-pf_1: dd 1.0
+pf_1:      dd 1.0
+pd_0x4000: dd 0x4000
 
 SECTION .text
 
-%macro RESAMPLE_FLOAT_FNS 0
-; int resample_common_float(ResampleContext *ctx, float *dst,
-;                           const float *src, int size, int update_ctx)
+%macro RESAMPLE_FNS 3 ; format [float or int16], bps, log2_bps
+; int resample_common_$format(ResampleContext *ctx, $format *dst,
+;                             const $format *src, int size, int update_ctx)
 %if ARCH_X86_64 ; unix64 and win64
-cglobal resample_common_float, 0, 15, 2, ctx, dst, src, phase_shift, index, frac, \
-                                         dst_incr_mod, size, min_filter_count_x4, \
-                                         min_filter_len_x4, dst_incr_div, src_incr, \
-                                         phase_mask, dst_end, filter_bank
+cglobal resample_common_%1, 0, 15, 2, ctx, dst, src, phase_shift, index, frac, \
+                                      dst_incr_mod, size, min_filter_count_x4, \
+                                      min_filter_len_x4, dst_incr_div, src_incr, \
+                                      phase_mask, dst_end, filter_bank
 
     ; use red-zone for variable storage
 %define ctx_stackq            [rsp-0x8]
@@ -85,8 +87,8 @@ cglobal resample_common_float, 0, 15, 2, ctx, dst, src, phase_shift, index, frac
     mov                   ctx_stackq, ctxq
     mov           min_filter_len_x4d, [ctxq+ResampleContext.filter_length]
     mov                dst_incr_divd, [ctxq+ResampleContext.dst_incr_div]
-    shl           min_filter_len_x4d, 2
-    lea                     dst_endq, [dstq+sizeq*4]
+    shl           min_filter_len_x4d, %3
+    lea                     dst_endq, [dstq+sizeq*%2]
 
 %if UNIX64
     mov                          ecx, [ctxq+ResampleContext.phase_shift]
@@ -109,8 +111,8 @@ cglobal resample_common_float, 0, 15, 2, ctx, dst, src, phase_shift, index, frac
     sub                         srcq, min_filter_len_x4q
     mov                   src_stackq, srcq
 %else ; x86-32
-cglobal resample_common_float, 1, 7, 2, ctx, phase_shift, dst, frac, \
-                                        index, min_filter_length_x4, filter_bank
+cglobal resample_common_%1, 1, 7, 2, ctx, phase_shift, dst, frac, \
+                                     index, min_filter_length_x4, filter_bank
 
     ; push temp variables to stack
 %define ctx_stackq            r0mp
@@ -119,7 +121,7 @@ cglobal resample_common_float, 1, 7, 2, ctx, phase_shift, dst, frac, \
 
     mov                         dstq, r1mp
     mov                           r3, r3mp
-    lea                           r3, [dstq+r3*4]
+    lea                           r3, [dstq+r3*%2]
     PUSH                              dword [ctxq+ResampleContext.dst_incr_div]
     PUSH                              dword [ctxq+ResampleContext.dst_incr_mod]
     PUSH                              dword [ctxq+ResampleContext.filter_alloc]
@@ -128,7 +130,7 @@ cglobal resample_common_float, 1, 7, 2, ctx, phase_shift, dst, frac, \
     PUSH                              dword [ctxq+ResampleContext.src_incr]
     mov        min_filter_length_x4d, [ctxq+ResampleContext.filter_length]
     mov                       indexd, [ctxq+ResampleContext.index]
-    shl        min_filter_length_x4d, 2
+    shl        min_filter_length_x4d, %3
     mov                        fracd, [ctxq+ResampleContext.frac]
     neg        min_filter_length_x4q
     mov                 filter_bankq, [ctxq+ResampleContext.filter_bank]
@@ -157,19 +159,28 @@ cglobal resample_common_float, 1, 7, 2, ctx, phase_shift, dst, frac, \
     imul                     filterd, indexd
 %if ARCH_X86_64
     mov         min_filter_count_x4q, min_filter_len_x4q
-    lea                      filterq, [filter_bankq+filterq*4]
+    lea                      filterq, [filter_bankq+filterq*%2]
 %else ; x86-32
     mov         min_filter_count_x4q, filter_bankq
-    lea                      filterq, [min_filter_count_x4q+filterq*4]
+    lea                      filterq, [min_filter_count_x4q+filterq*%2]
     mov         min_filter_count_x4q, min_filter_length_x4q
 %endif
+%ifidn %1, float
     xorps                         m0, m0, m0
+%else ; int16
+    movd                          m0, [pd_0x4000]
+%endif
 
     align 16
 .inner_loop:
-    movups                        m1, [srcq+min_filter_count_x4q*1]
+    movu                          m1, [srcq+min_filter_count_x4q*1]
+%ifidn %1, float
     mulps                         m1, m1, [filterq+min_filter_count_x4q*1]
     addps                         m0, m0, m1
+%else ; int16
+    pmaddwd                       m1, [filterq+min_filter_count_x4q*1]
+    paddd                         m0, m1
+%endif
     add         min_filter_count_x4q, mmsize
     js .inner_loop
 
@@ -179,6 +190,7 @@ cglobal resample_common_float, 1, 7, 2, ctx, phase_shift, dst, frac, \
 %endif
 
     ; horizontal sum & store
+%ifidn %1, float
     movhlps                      xm1, xm0
     addps                        xm0, xm1
     shufps                       xm1, xm0, xm0, q0001
@@ -186,6 +198,21 @@ cglobal resample_common_float, 1, 7, 2, ctx, phase_shift, dst, frac, \
     addps                        xm0, xm1
     add                       indexd, dst_incr_divd
     movss                     [dstq], xm0
+%else ; int16
+%if mmsize == 16
+    pshufd                        m1, m0, q0032
+    paddd                         m0, m1
+    pshufd                        m1, m0, q0001
+%else ; mmsize == 8
+    pshufw                        m1, m0, q0032
+%endif
+    paddd                         m0, m1
+    psrad                         m0, 15
+    add                        fracd, dst_incr_modd
+    packssdw                      m0, m0
+    add                       indexd, dst_incr_divd
+    movd                      [dstq], m0
+%endif
     cmp                        fracd, src_incrd
     jl .skip
     sub                        fracd, src_incrd
@@ -205,10 +232,10 @@ cglobal resample_common_float, 1, 7, 2, ctx, phase_shift, dst, frac, \
 
 .skip:
     mov                  index_incrd, indexd
-    add                         dstq, 4
+    add                         dstq, %2
     and                       indexd, phase_maskd
     sar                  index_incrd, phase_shiftb
-    lea                         srcq, [srcq+index_incrq*4]
+    lea                         srcq, [srcq+index_incrq*%2]
     cmp                         dstq, dst_endq
     jne .loop
 
@@ -228,7 +255,7 @@ cglobal resample_common_float, 1, 7, 2, ctx, phase_shift, dst, frac, \
     mov [ctxq+ResampleContext.frac ], fracd
     sub                          rax, src_stackq
     mov [ctxq+ResampleContext.index], indexd
-    shr                          rax, 2
+    shr                          rax, %3
 
 .skip_store:
 %if ARCH_X86_32
@@ -236,13 +263,24 @@ cglobal resample_common_float, 1, 7, 2, ctx, phase_shift, dst, frac, \
 %endif
     RET
 
-; int resample_linear_float(ResampleContext *ctx, float *dst,
-;                           const float *src, int size, int update_ctx)
+; int resample_linear_$format(ResampleContext *ctx, float *dst,
+;                             const float *src, int size, int update_ctx)
 %if ARCH_X86_64 ; unix64 and win64
-cglobal resample_linear_float, 0, 15, 5, ctx, dst, src, phase_shift, index, frac, \
-                                         dst_incr_mod, size, min_filter_count_x4, \
-                                         min_filter_len_x4, dst_incr_div, src_incr, \
-                                         phase_mask, dst_end, filter_bank
+%if UNIX64
+cglobal resample_linear_%1, 0, 15, 5, ctx, dst, phase_mask, phase_shift, index, frac, \
+                                      size, dst_incr_mod, min_filter_count_x4, \
+                                      min_filter_len_x4, dst_incr_div, src_incr, \
+                                      src, dst_end, filter_bank
+
+    mov                         srcq, r2mp
+%else ; win64
+cglobal resample_linear_%1, 0, 15, 5, ctx, phase_mask, src, phase_shift, index, frac, \
+                                      size, dst_incr_mod, min_filter_count_x4, \
+                                      min_filter_len_x4, dst_incr_div, src_incr, \
+                                      dst, dst_end, filter_bank
+
+    mov                         dstq, r1mp
+%endif
 
     ; use red-zone for variable storage
 %define ctx_stackq            [rsp-0x8]
@@ -269,27 +307,31 @@ cglobal resample_linear_float, 0, 15, 5, ctx, dst, src, phase_shift, index, frac
     mov                   ctx_stackq, ctxq
     mov            phase_mask_stackd, phase_maskd
     mov           min_filter_len_x4d, [ctxq+ResampleContext.filter_length]
+%ifidn %1, float
     cvtsi2ss                     xm0, src_incrd
     movss                        xm4, [pf_1]
     divss                        xm4, xm0
+%else ; int16
+    movd                          m4, [pd_0x4000]
+%endif
     mov                dst_incr_divd, [ctxq+ResampleContext.dst_incr_div]
-    shl           min_filter_len_x4d, 2
-    lea                     dst_endq, [dstq+sizeq*4]
+    shl           min_filter_len_x4d, %3
+    lea                     dst_endq, [dstq+sizeq*%2]
 
 %if UNIX64
     mov                          ecx, [ctxq+ResampleContext.phase_shift]
     mov                          edi, [ctxq+ResampleContext.filter_alloc]
 
-    DEFINE_ARGS filter_alloc, dst, src, phase_shift, index, frac, dst_incr_mod, \
-                filter1, min_filter_count_x4, min_filter_len_x4, dst_incr_div, \
-                src_incr, filter2, dst_end, filter_bank
+    DEFINE_ARGS filter_alloc, dst, filter2, phase_shift, index, frac, filter1, \
+                dst_incr_mod, min_filter_count_x4, min_filter_len_x4, \
+                dst_incr_div, src_incr, src, dst_end, filter_bank
 %elif WIN64
     mov                          R9d, [ctxq+ResampleContext.filter_alloc]
     mov                          ecx, [ctxq+ResampleContext.phase_shift]
 
-    DEFINE_ARGS phase_shift, dst, src, filter_alloc, index, frac, dst_incr_mod, \
-                filter1, min_filter_count_x4, min_filter_len_x4, dst_incr_div, \
-                src_incr, filter2, dst_end, filter_bank
+    DEFINE_ARGS phase_shift, filter2, src, filter_alloc, index, frac, filter1, \
+                dst_incr_mod, min_filter_count_x4, min_filter_len_x4, \
+                dst_incr_div, src_incr, dst, dst_end, filter_bank
 %endif
 
     neg           min_filter_len_x4q
@@ -297,8 +339,8 @@ cglobal resample_linear_float, 0, 15, 5, ctx, dst, src, phase_shift, index, frac
     sub                         srcq, min_filter_len_x4q
     mov                   src_stackq, srcq
 %else ; x86-32
-cglobal resample_linear_float, 1, 7, 5, ctx, filter1, dst, frac, \
-                                        index, min_filter_length_x4, filter_bank
+cglobal resample_linear_%1, 1, 7, 5, ctx, min_filter_length_x4, filter2, \
+                                     frac, index, dst, filter_bank
 
     ; push temp variables to stack
 %define ctx_stackq            r0mp
@@ -307,23 +349,27 @@ cglobal resample_linear_float, 1, 7, 5, ctx, filter1, dst, frac, \
 
     mov                         dstq, r1mp
     mov                           r3, r3mp
-    lea                           r3, [dstq+r3*4]
+    lea                           r3, [dstq+r3*%2]
     PUSH                              dword [ctxq+ResampleContext.dst_incr_div]
     PUSH                              r3
     mov                           r3, dword [ctxq+ResampleContext.filter_alloc]
     PUSH                              dword [ctxq+ResampleContext.dst_incr_mod]
     PUSH                              r3
-    shl                           r3, 2
+    shl                           r3, %3
     PUSH                              r3
     mov                           r3, dword [ctxq+ResampleContext.src_incr]
     PUSH                              dword [ctxq+ResampleContext.phase_mask]
     PUSH                              r3d
+%ifidn %1, float
     cvtsi2ss                     xm0, r3d
     movss                        xm4, [pf_1]
     divss                        xm4, xm0
+%else ; int16
+    movd                          m4, [pd_0x4000]
+%endif
     mov        min_filter_length_x4d, [ctxq+ResampleContext.filter_length]
     mov                       indexd, [ctxq+ResampleContext.index]
-    shl        min_filter_length_x4d, 2
+    shl        min_filter_length_x4d, %3
     mov                        fracd, [ctxq+ResampleContext.frac]
     neg        min_filter_length_x4q
     mov                 filter_bankq, [ctxq+ResampleContext.filter_bank]
@@ -333,7 +379,7 @@ cglobal resample_linear_float, 1, 7, 5, ctx, filter1, dst, frac, \
     PUSH                              filter_bankq
     PUSH                              dword [ctxq+ResampleContext.phase_shift]
 
-    DEFINE_ARGS src, filter1, dst, frac, index, min_filter_count_x4, filter2
+    DEFINE_ARGS filter1, min_filter_count_x4, filter2, frac, index, dst, src
 
 %define phase_shift_stackd    dword [rsp+0x0]
 %define filter_bankq          dword [rsp+0x4]
@@ -354,25 +400,37 @@ cglobal resample_linear_float, 1, 7, 5, ctx, filter1, dst, frac, \
     imul                    filter1d, indexd
 %if ARCH_X86_64
     mov         min_filter_count_x4q, min_filter_len_x4q
-    lea                     filter1q, [filter_bankq+filter1q*4]
-    lea                     filter2q, [filter1q+filter_allocq*4]
+    lea                     filter1q, [filter_bankq+filter1q*%2]
+    lea                     filter2q, [filter1q+filter_allocq*%2]
 %else ; x86-32
     mov         min_filter_count_x4q, filter_bankq
-    lea                     filter1q, [min_filter_count_x4q+filter1q*4]
+    lea                     filter1q, [min_filter_count_x4q+filter1q*%2]
     mov         min_filter_count_x4q, min_filter_length_x4q
     mov                     filter2q, filter1q
     add                     filter2q, filter_alloc_x4q
 %endif
+%ifidn %1, float
     xorps                         m0, m0, m0
     xorps                         m2, m2, m2
+%else ; int16
+    mova                          m0, m4
+    mova                          m2, m4
+%endif
 
     align 16
 .inner_loop:
-    movups                        m1, [srcq+min_filter_count_x4q*1]
+    movu                          m1, [srcq+min_filter_count_x4q*1]
+%ifidn %1, float
     mulps                         m3, m1, [filter2q+min_filter_count_x4q*1]
     mulps                         m1, m1, [filter1q+min_filter_count_x4q*1]
     addps                         m2, m2, m3
     addps                         m0, m0, m1
+%else ; int16
+    pmaddwd                       m3, m1, [filter2q+min_filter_count_x4q*1]
+    pmaddwd                       m1, [filter1q+min_filter_count_x4q*1]
+    paddd                         m2, m3
+    paddd                         m0, m1
+%endif
     add         min_filter_count_x4q, mmsize
     js .inner_loop
 
@@ -383,6 +441,7 @@ cglobal resample_linear_float, 1, 7, 5, ctx, filter1, dst, frac, \
     addps                        xm2, xm3
 %endif
 
+%ifidn %1, float
     ; val += (v2 - val) * (FELEML) frac / c->src_incr;
     cvtsi2ss                     xm1, fracd
     subps                        xm2, xm0
@@ -399,21 +458,55 @@ cglobal resample_linear_float, 1, 7, 5, ctx, filter1, dst, frac, \
     addps                        xm0, xm1
     add                       indexd, dst_incr_divd
     movss                     [dstq], xm0
+%else ; int16
+%if mmsize == 16
+    pshufd                        m3, m2, q0032
+    pshufd                        m1, m0, q0032
+    paddd                         m2, m3
+    paddd                         m0, m1
+    pshufd                        m3, m2, q0001
+    pshufd                        m1, m0, q0001
+%else ; mmsize == 8
+    pshufw                        m3, m2, q0032
+    pshufw                        m1, m0, q0032
+%endif
+    paddd                         m2, m3
+    paddd                         m0, m1
+    psubd                         m2, m0
+    ; This is probably a really bad idea on atom and other machines with a
+    ; long transfer latency between GPRs and XMMs (atom). However, it does
+    ; make the clip a lot simpler...
+    movd                         eax, m2
+    add                       indexd, dst_incr_divd
+    imul                              fracd
+    idiv                              src_incrd
+    movd                          m1, eax
+    add                        fracd, dst_incr_modd
+    paddd                         m0, m1
+    psrad                         m0, 15
+    packssdw                      m0, m0
+    movd                      [dstq], m0
+
+    ; note that for imul/idiv, I need to move filter to edx/eax for each:
+    ; - 32bit: eax=r0[filter1], edx=r2[filter2]
+    ; - win64: eax=r6[filter1], edx=r1[todo]
+    ; - unix64: eax=r6[filter1], edx=r2[todo]
+%endif
     cmp                        fracd, src_incrd
     jl .skip
     sub                        fracd, src_incrd
     inc                       indexd
 
 %if UNIX64
-    DEFINE_ARGS filter_alloc, dst, src, phase_shift, index, frac, dst_incr_mod, \
-                index_incr, min_filter_count_x4, min_filter_len_x4, dst_incr_div, \
-                src_incr, filter2, dst_end, filter_bank
+    DEFINE_ARGS filter_alloc, dst, filter2, phase_shift, index, frac, index_incr, \
+                dst_incr_mod, min_filter_count_x4, min_filter_len_x4, \
+                dst_incr_div, src_incr, src, dst_end, filter_bank
 %elif WIN64
-    DEFINE_ARGS phase_shift, dst, src, filter_alloc, index, frac, dst_incr_mod, \
-                index_incr, min_filter_count_x4, min_filter_len_x4, dst_incr_div, \
-                src_incr, filter2, dst_end, filter_bank
+    DEFINE_ARGS phase_shift, filter2, src, filter_alloc, index, frac, index_incr, \
+                dst_incr_mod, min_filter_count_x4, min_filter_len_x4, \
+                dst_incr_div, src_incr, dst, dst_end, filter_bank
 %else ; x86-32
-    DEFINE_ARGS src, phase_shift, dst, frac, index, index_incr
+    DEFINE_ARGS filter1, phase_shift, index_incr, frac, index, dst, src
 %endif
 
 .skip:
@@ -421,17 +514,23 @@ cglobal resample_linear_float, 1, 7, 5, ctx, filter1, dst, frac, \
     mov                 phase_shiftd, phase_shift_stackd
 %endif
     mov                  index_incrd, indexd
-    add                         dstq, 4
+    add                         dstq, %2
     and                       indexd, phase_mask_stackd
     sar                  index_incrd, phase_shiftb
-    lea                         srcq, [srcq+index_incrq*4]
+    lea                         srcq, [srcq+index_incrq*%2]
     cmp                         dstq, dst_endq
     jne .loop
 
-%if ARCH_X86_64
-    DEFINE_ARGS ctx, dst, src, phase_shift, index, frac
+%if UNIX64
+    DEFINE_ARGS ctx, dst, filter2, phase_shift, index, frac, index_incr, \
+                dst_incr_mod, min_filter_count_x4, min_filter_len_x4, \
+                dst_incr_div, src_incr, src, dst_end, filter_bank
+%elif WIN64
+    DEFINE_ARGS ctx, filter2, src, phase_shift, index, frac, index_incr, \
+                dst_incr_mod, min_filter_count_x4, min_filter_len_x4, \
+                dst_incr_div, src_incr, dst, dst_end, filter_bank
 %else ; x86-32
-    DEFINE_ARGS src, ctx, update_context, frac, index
+    DEFINE_ARGS filter1, ctx, update_context, frac, index, dst, src
 %endif
 
     cmp  dword update_context_stackd, 0
@@ -444,7 +543,7 @@ cglobal resample_linear_float, 1, 7, 5, ctx, filter1, dst, frac, \
     mov [ctxq+ResampleContext.frac ], fracd
     sub                          rax, src_stackq
     mov [ctxq+ResampleContext.index], indexd
-    shr                          rax, 2
+    shr                          rax, %3
 
 .skip_store:
 %if ARCH_X86_32
@@ -454,9 +553,17 @@ cglobal resample_linear_float, 1, 7, 5, ctx, filter1, dst, frac, \
 %endmacro
 
 INIT_XMM sse
-RESAMPLE_FLOAT_FNS
+RESAMPLE_FNS float, 4, 2
 
 %if HAVE_AVX_EXTERNAL
 INIT_YMM avx
-RESAMPLE_FLOAT_FNS
+RESAMPLE_FNS float, 4, 2
 %endif
+
+%if ARCH_X86_32
+INIT_MMX mmxext
+RESAMPLE_FNS int16, 2, 1
+%endif
+
+INIT_XMM sse2
+RESAMPLE_FNS int16, 2, 1
diff --git a/libswresample/x86/resample_mmx.h b/libswresample/x86/resample_mmx.h
index 94237b0507..b0ea496361 100644
--- a/libswresample/x86/resample_mmx.h
+++ b/libswresample/x86/resample_mmx.h
@@ -22,116 +22,6 @@
 #include "libavutil/cpu.h"
 #include "libswresample/swresample_internal.h"
 
-DECLARE_ALIGNED(16, const uint64_t, ff_resample_int16_rounder)[2]    = { 0x0000000000004000ULL, 0x0000000000000000ULL};
-
-#define COMMON_CORE_INT16_MMX2 \
-    x86_reg len= -2*c->filter_length;\
-__asm__ volatile(\
-    "movq "MANGLE(ff_resample_int16_rounder)", %%mm0 \n\t"\
-    "1:                         \n\t"\
-    "movq    (%1, %0), %%mm1    \n\t"\
-    "pmaddwd (%2, %0), %%mm1    \n\t"\
-    "paddd  %%mm1, %%mm0        \n\t"\
-    "add       $8, %0           \n\t"\
-    " js 1b                     \n\t"\
-    "pshufw $0x0E, %%mm0, %%mm1 \n\t"\
-    "paddd %%mm1, %%mm0         \n\t"\
-    "psrad    $15, %%mm0        \n\t"\
-    "packssdw %%mm0, %%mm0      \n\t"\
-    "movd %%mm0, (%3)           \n\t"\
-    : "+r" (len)\
-    : "r" (((uint8_t*)(src+sample_index))-len),\
-      "r" (((uint8_t*)filter)-len),\
-      "r" (dst+dst_index)\
-      NAMED_CONSTRAINTS_ARRAY_ADD(ff_resample_int16_rounder)\
-);
-
-#define LINEAR_CORE_INT16_MMX2 \
-    x86_reg len= -2*c->filter_length;\
-__asm__ volatile(\
-    "pxor          %%mm0, %%mm0 \n\t"\
-    "pxor          %%mm2, %%mm2 \n\t"\
-    "1:                         \n\t"\
-    "movq       (%3, %0), %%mm1 \n\t"\
-    "movq          %%mm1, %%mm3 \n\t"\
-    "pmaddwd    (%4, %0), %%mm1 \n\t"\
-    "pmaddwd    (%5, %0), %%mm3 \n\t"\
-    "paddd         %%mm1, %%mm0 \n\t"\
-    "paddd         %%mm3, %%mm2 \n\t"\
-    "add              $8, %0    \n\t"\
-    " js 1b                     \n\t"\
-    "pshufw $0x0E, %%mm0, %%mm1 \n\t"\
-    "pshufw $0x0E, %%mm2, %%mm3 \n\t"\
-    "paddd         %%mm1, %%mm0 \n\t"\
-    "paddd         %%mm3, %%mm2 \n\t"\
-    "movd          %%mm0, %1    \n\t"\
-    "movd          %%mm2, %2    \n\t"\
-    : "+r" (len),\
-      "=r" (val),\
-      "=r" (v2)\
-    : "r" (((uint8_t*)(src+sample_index))-len),\
-      "r" (((uint8_t*)filter)-len),\
-      "r" (((uint8_t*)(filter+c->filter_alloc))-len)\
-);
-
-#define COMMON_CORE_INT16_SSE2 \
-    x86_reg len= -2*c->filter_length;\
-__asm__ volatile(\
-    "movdqa "MANGLE(ff_resample_int16_rounder)", %%xmm0 \n\t"\
-    "1:                           \n\t"\
-    "movdqu  (%1, %0), %%xmm1     \n\t"\
-    "pmaddwd (%2, %0), %%xmm1     \n\t"\
-    "paddd  %%xmm1, %%xmm0        \n\t"\
-    "add       $16, %0            \n\t"\
-    " js 1b                       \n\t"\
-    "pshufd $0x0E, %%xmm0, %%xmm1 \n\t"\
-    "paddd %%xmm1, %%xmm0         \n\t"\
-    "pshufd $0x01, %%xmm0, %%xmm1 \n\t"\
-    "paddd %%xmm1, %%xmm0         \n\t"\
-    "psrad    $15, %%xmm0         \n\t"\
-    "packssdw %%xmm0, %%xmm0      \n\t"\
-    "movd %%xmm0, (%3)            \n\t"\
-    : "+r" (len)\
-    : "r" (((uint8_t*)(src+sample_index))-len),\
-      "r" (((uint8_t*)filter)-len),\
-      "r" (dst+dst_index)\
-      NAMED_CONSTRAINTS_ARRAY_ADD(ff_resample_int16_rounder)\
-      XMM_CLOBBERS_ONLY("%xmm0", "%xmm1")\
-);
-
-#define LINEAR_CORE_INT16_SSE2 \
-    x86_reg len= -2*c->filter_length;\
-__asm__ volatile(\
-    "pxor          %%xmm0, %%xmm0 \n\t"\
-    "pxor          %%xmm2, %%xmm2 \n\t"\
-    "1:                           \n\t"\
-    "movdqu      (%3, %0), %%xmm1 \n\t"\
-    "movdqa        %%xmm1, %%xmm3 \n\t"\
-    "pmaddwd     (%4, %0), %%xmm1 \n\t"\
-    "pmaddwd     (%5, %0), %%xmm3 \n\t"\
-    "paddd         %%xmm1, %%xmm0 \n\t"\
-    "paddd         %%xmm3, %%xmm2 \n\t"\
-    "add              $16, %0     \n\t"\
-    " js 1b                       \n\t"\
-    "pshufd $0x0E, %%xmm0, %%xmm1 \n\t"\
-    "pshufd $0x0E, %%xmm2, %%xmm3 \n\t"\
-    "paddd         %%xmm1, %%xmm0 \n\t"\
-    "paddd         %%xmm3, %%xmm2 \n\t"\
-    "pshufd $0x01, %%xmm0, %%xmm1 \n\t"\
-    "pshufd $0x01, %%xmm2, %%xmm3 \n\t"\
-    "paddd         %%xmm1, %%xmm0 \n\t"\
-    "paddd         %%xmm3, %%xmm2 \n\t"\
-    "movd          %%xmm0, %1     \n\t"\
-    "movd          %%xmm2, %2     \n\t"\
-    : "+r" (len),\
-      "=r" (val),\
-      "=r" (v2)\
-    : "r" (((uint8_t*)(src+sample_index))-len),\
-      "r" (((uint8_t*)filter)-len),\
-      "r" (((uint8_t*)(filter+c->filter_alloc))-len)\
-    XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3")\
-);
-
 #define COMMON_CORE_DBL_SSE2 \
     x86_reg len= -8*c->filter_length;\
 __asm__ volatile(\
diff --git a/libswresample/x86/resample_x86_dsp.c b/libswresample/x86/resample_x86_dsp.c
index c7d2054f9c..5130ecdd1b 100644
--- a/libswresample/x86/resample_x86_dsp.c
+++ b/libswresample/x86/resample_x86_dsp.c
@@ -27,34 +27,14 @@
 
 #include "libswresample/resample.h"
 
-int swri_resample_common_int16_mmx2 (ResampleContext *c, int16_t *dst, const int16_t *src, int n, int update_ctx);
-int swri_resample_linear_int16_mmx2 (ResampleContext *c, int16_t *dst, const int16_t *src, int n, int update_ctx);
-int swri_resample_common_int16_sse2 (ResampleContext *c, int16_t *dst, const int16_t *src, int n, int update_ctx);
-int swri_resample_linear_int16_sse2 (ResampleContext *c, int16_t *dst, const int16_t *src, int n, int update_ctx);
-int swri_resample_common_float_sse  (ResampleContext *c,   float *dst, const   float *src, int n, int update_ctx);
-int swri_resample_linear_float_sse  (ResampleContext *c,   float *dst, const   float *src, int n, int update_ctx);
-int swri_resample_common_float_avx  (ResampleContext *c,   float *dst, const   float *src, int n, int update_ctx);
-int swri_resample_linear_float_avx  (ResampleContext *c,   float *dst, const   float *src, int n, int update_ctx);
 int swri_resample_common_double_sse2(ResampleContext *c,  double *dst, const  double *src, int n, int update_ctx);
 int swri_resample_linear_double_sse2(ResampleContext *c,  double *dst, const  double *src, int n, int update_ctx);
 
-#if HAVE_MMXEXT_INLINE
-
+#if HAVE_SSE2_INLINE
 #define DO_RESAMPLE_ONE 0
 
 #include "resample_mmx.h"
 
-#if ARCH_X86_32
-#define TEMPLATE_RESAMPLE_S16_MMX2
-#include "libswresample/resample_template.c"
-#undef TEMPLATE_RESAMPLE_S16_MMX2
-#endif
-
-#if HAVE_SSE2_INLINE
-#define TEMPLATE_RESAMPLE_S16_SSE2
-#include "libswresample/resample_template.c"
-#undef TEMPLATE_RESAMPLE_S16_SSE2
-
 #define TEMPLATE_RESAMPLE_DBL_SSE2
 #include "libswresample/resample_template.c"
 #undef TEMPLATE_RESAMPLE_DBL_SSE2
@@ -62,7 +42,15 @@ int swri_resample_linear_double_sse2(ResampleContext *c,  double *dst, const  do
 
 #undef DO_RESAMPLE_ONE
 
-#endif // HAVE_MMXEXT_INLINE
+int ff_resample_common_int16_mmxext(ResampleContext *c, uint8_t *dst,
+                                    const uint8_t *src, int sz, int upd);
+int ff_resample_linear_int16_mmxext(ResampleContext *c, uint8_t *dst,
+                                    const uint8_t *src, int sz, int upd);
+
+int ff_resample_common_int16_sse2(ResampleContext *c, uint8_t *dst,
+                                  const uint8_t *src, int sz, int upd);
+int ff_resample_linear_int16_sse2(ResampleContext *c, uint8_t *dst,
+                                  const uint8_t *src, int sz, int upd);
 
 int ff_resample_common_float_sse(ResampleContext *c, uint8_t *dst,
                                  const uint8_t *src, int sz, int upd);
@@ -79,17 +67,19 @@ void swresample_dsp_x86_init(ResampleContext *c)
     int av_unused mm_flags = av_get_cpu_flags();
 
 #define FNIDX(fmt) (AV_SAMPLE_FMT_##fmt - AV_SAMPLE_FMT_S16P)
-    if (ARCH_X86_32 && HAVE_MMXEXT_INLINE && mm_flags & AV_CPU_FLAG_MMX2) {
-        c->dsp.resample_common[FNIDX(S16P)] = (resample_fn) swri_resample_common_int16_mmx2;
-        c->dsp.resample_linear[FNIDX(S16P)] = (resample_fn) swri_resample_linear_int16_mmx2;
+    if (ARCH_X86_32 && HAVE_MMXEXT_EXTERNAL && mm_flags & AV_CPU_FLAG_MMX2) {
+        c->dsp.resample_common[FNIDX(S16P)] = ff_resample_common_int16_mmxext;
+        c->dsp.resample_linear[FNIDX(S16P)] = ff_resample_linear_int16_mmxext;
     }
     if (HAVE_SSE_EXTERNAL && mm_flags & AV_CPU_FLAG_SSE) {
         c->dsp.resample_common[FNIDX(FLTP)] = ff_resample_common_float_sse;
         c->dsp.resample_linear[FNIDX(FLTP)] = ff_resample_linear_float_sse;
     }
+    if (HAVE_SSE2_EXTERNAL && mm_flags & AV_CPU_FLAG_SSE2) {
+        c->dsp.resample_common[FNIDX(S16P)] = ff_resample_common_int16_sse2;
+        c->dsp.resample_linear[FNIDX(S16P)] = ff_resample_linear_int16_sse2;
+    }
     if (HAVE_SSE2_INLINE && mm_flags & AV_CPU_FLAG_SSE2) {
-        c->dsp.resample_common[FNIDX(S16P)] = (resample_fn) swri_resample_common_int16_sse2;
-        c->dsp.resample_linear[FNIDX(S16P)] = (resample_fn) swri_resample_linear_int16_sse2;
         c->dsp.resample_common[FNIDX(DBLP)] = (resample_fn) swri_resample_common_double_sse2;
         c->dsp.resample_linear[FNIDX(DBLP)] = (resample_fn) swri_resample_linear_double_sse2;
     }