swr: rewrite resample_common/linear_float_sse/avx in yasm.

Linear interpolation goes from 63 (llvm) or 58 (gcc) to 48 (yasm)
cycles/sample on 64bit, or from 66 (llvm/gcc) to 52 (yasm) cycles/
sample on 32bit. Bon-linear goes from 43 (llvm) or 38 (gcc) to
32 (yasm) cycles/sample on 64bit, or from 46 (llvm) or 44 (gcc) to
38 (yasm) cycles/sample on 32bit (all testing on OSX 10.9.2, llvm
5.1 and gcc 4.8/9).

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

1 files changed, 462 insertions, 0 deletions

diff --git a/libswresample/x86/resample.asm b/libswresample/x86/resample.asm
new file mode 100644
index 0000000000..bb63fc7e8e
--- /dev/null
+++ b/libswresample/x86/resample.asm
@@ -0,0 +1,462 @@
+;******************************************************************************
+;* Copyright (c) 2012 Michael Niedermayer
+;* Copyright (c) 2014 James Almer <jamrial <at> gmail.com>
+;*
+;* This file is part of FFmpeg.
+;*
+;* FFmpeg is free software; you can redistribute it and/or
+;* modify it under the terms of the GNU Lesser General Public
+;* License as published by the Free Software Foundation; either
+;* version 2.1 of the License, or (at your option) any later version.
+;*
+;* FFmpeg is distributed in the hope that it will be useful,
+;* but WITHOUT ANY WARRANTY; without even the implied warranty of
+;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+;* Lesser General Public License for more details.
+;*
+;* You should have received a copy of the GNU Lesser General Public
+;* License along with FFmpeg; if not, write to the Free Software
+;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+;******************************************************************************
+
+%include "libavutil/x86/x86util.asm"
+
+%if ARCH_X86_64
+%define pointer resq
+%else
+%define pointer resd
+%endif
+
+struc ResampleContext
+    .av_class:              pointer 1
+    .filter_bank:           pointer 1
+    .filter_length:         resd 1
+    .filter_alloc:          resd 1
+    .ideal_dst_incr:        resd 1
+    .dst_incr:              resd 1
+    .dst_incr_div:          resd 1
+    .dst_incr_mod:          resd 1
+    .index:                 resd 1
+    .frac:                  resd 1
+    .src_incr:              resd 1
+    .compensation_distance: resd 1
+    .phase_shift:           resd 1
+    .phase_mask:            resd 1
+
+    ; there's a few more here but we only care about the first few
+endstruc
+
+SECTION_RODATA
+
+pf_1: dd 1.0
+
+SECTION .text
+
+%macro RESAMPLE_FLOAT_FNS 0
+; int resample_common_float(ResampleContext *ctx, float *dst,
+;                           const float *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
+
+    ; use red-zone for variable storage
+%define ctx_stackq            [rsp-0x8]
+%define src_stackq            [rsp-0x10]
+%if WIN64
+%define update_context_stackd r4m
+%else ; unix64
+%define update_context_stackd [rsp-0x14]
+%endif
+
+    ; load as many variables in registers as possible; for the rest, store
+    ; on stack so that we have 'ctx' available as one extra register
+    mov                        sized, r3d
+    mov                  phase_maskd, [ctxq+ResampleContext.phase_mask]
+%if UNIX64
+    mov        update_context_stackd, r4d
+%endif
+    mov                       indexd, [ctxq+ResampleContext.index]
+    mov                        fracd, [ctxq+ResampleContext.frac]
+    mov                dst_incr_modd, [ctxq+ResampleContext.dst_incr_mod]
+    mov                 filter_bankq, [ctxq+ResampleContext.filter_bank]
+    mov                    src_incrd, [ctxq+ResampleContext.src_incr]
+    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]
+
+%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, \
+                filter, min_filter_count_x4, min_filter_len_x4, dst_incr_div, \
+                src_incr, phase_mask, 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, \
+                filter, min_filter_count_x4, min_filter_len_x4, dst_incr_div, \
+                src_incr, phase_mask, dst_end, filter_bank
+%endif
+
+    neg           min_filter_len_x4q
+    sub                 filter_bankq, min_filter_len_x4q
+    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
+
+    ; push temp variables to stack
+%define ctx_stackq            r0mp
+%define src_stackq            r2mp
+%define update_context_stackd r4m
+
+    mov                         dstq, r1mp
+    mov                           r3, r3mp
+    lea                           r3, [dstq+r3*4]
+    PUSH                              dword [ctxq+ResampleContext.dst_incr_div]
+    PUSH                              dword [ctxq+ResampleContext.dst_incr_mod]
+    PUSH                              dword [ctxq+ResampleContext.filter_alloc]
+    PUSH                              r3
+    PUSH                              dword [ctxq+ResampleContext.phase_mask]
+    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
+    mov                        fracd, [ctxq+ResampleContext.frac]
+    neg        min_filter_length_x4q
+    mov                 filter_bankq, [ctxq+ResampleContext.filter_bank]
+    sub                         r2mp, min_filter_length_x4q
+    sub                 filter_bankq, min_filter_length_x4q
+    PUSH                              min_filter_length_x4q
+    PUSH                              filter_bankq
+    mov                 phase_shiftd, [ctxq+ResampleContext.phase_shift]
+
+    DEFINE_ARGS src, phase_shift, dst, frac, index, min_filter_count_x4, filter
+
+%define filter_bankq          dword [rsp+0x0]
+%define min_filter_length_x4q dword [rsp+0x4]
+%define src_incrd             dword [rsp+0x8]
+%define phase_maskd           dword [rsp+0xc]
+%define dst_endq              dword [rsp+0x10]
+%define filter_allocd         dword [rsp+0x14]
+%define dst_incr_modd         dword [rsp+0x18]
+%define dst_incr_divd         dword [rsp+0x1c]
+
+    mov                         srcq, r2mp
+%endif
+
+.loop:
+    mov                      filterd, filter_allocd
+    imul                     filterd, indexd
+%if ARCH_X86_64
+    mov         min_filter_count_x4q, min_filter_len_x4q
+    lea                      filterq, [filter_bankq+filterq*4]
+%else ; x86-32
+    mov         min_filter_count_x4q, filter_bankq
+    lea                      filterq, [min_filter_count_x4q+filterq*4]
+    mov         min_filter_count_x4q, min_filter_length_x4q
+%endif
+    xorps                         m0, m0, m0
+
+    align 16
+.inner_loop:
+    movups                        m1, [srcq+min_filter_count_x4q*1]
+    mulps                         m1, m1, [filterq+min_filter_count_x4q*1]
+    addps                         m0, m0, m1
+    add         min_filter_count_x4q, mmsize
+    js .inner_loop
+
+%if cpuflag(avx)
+    vextractf128                 xm1, m0, 0x1
+    addps                        xm0, xm1
+%endif
+
+    ; horizontal sum & store
+    movhlps                      xm1, xm0
+    addps                        xm0, xm1
+    shufps                       xm1, xm0, xm0, q0001
+    add                        fracd, dst_incr_modd
+    addps                        xm0, xm1
+    add                       indexd, dst_incr_divd
+    movss                     [dstq], xm0
+    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, phase_mask, 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, phase_mask, dst_end, filter_bank
+%else ; x86-32
+    DEFINE_ARGS src, phase_shift, dst, frac, index, index_incr
+%endif
+
+.skip:
+    mov                  index_incrd, indexd
+    add                         dstq, 4
+    and                       indexd, phase_maskd
+    sar                  index_incrd, phase_shiftb
+    lea                         srcq, [srcq+index_incrq*4]
+    cmp                         dstq, dst_endq
+    jne .loop
+
+%if ARCH_X86_64
+    DEFINE_ARGS ctx, dst, src, phase_shift, index, frac
+%else ; x86-32
+    DEFINE_ARGS src, ctx, update_context, frac, index
+%endif
+
+    cmp  dword update_context_stackd, 0
+    jz .skip_store
+    ; strictly speaking, the function should always return the consumed
+    ; number of bytes; however, we only use the value if update_context
+    ; is true, so let's just leave it uninitialized otherwise
+    mov                         ctxq, ctx_stackq
+    movifnidn                    rax, srcq
+    mov [ctxq+ResampleContext.frac ], fracd
+    sub                          rax, src_stackq
+    mov [ctxq+ResampleContext.index], indexd
+    shr                          rax, 2
+
+.skip_store:
+%if ARCH_X86_32
+    ADD                          rsp, 0x20
+%endif
+    RET
+
+; int resample_linear_float(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
+
+    ; use red-zone for variable storage
+%define ctx_stackq            [rsp-0x8]
+%define src_stackq            [rsp-0x10]
+%define phase_mask_stackd     [rsp-0x14]
+%if WIN64
+%define update_context_stackd r4m
+%else ; unix64
+%define update_context_stackd [rsp-0x18]
+%endif
+
+    ; load as many variables in registers as possible; for the rest, store
+    ; on stack so that we have 'ctx' available as one extra register
+    mov                        sized, r3d
+    mov                  phase_maskd, [ctxq+ResampleContext.phase_mask]
+%if UNIX64
+    mov        update_context_stackd, r4d
+%endif
+    mov                       indexd, [ctxq+ResampleContext.index]
+    mov                        fracd, [ctxq+ResampleContext.frac]
+    mov                dst_incr_modd, [ctxq+ResampleContext.dst_incr_mod]
+    mov                 filter_bankq, [ctxq+ResampleContext.filter_bank]
+    mov                    src_incrd, [ctxq+ResampleContext.src_incr]
+    mov                   ctx_stackq, ctxq
+    mov            phase_mask_stackd, phase_maskd
+    mov           min_filter_len_x4d, [ctxq+ResampleContext.filter_length]
+    cvtsi2ss                     xm0, src_incrd
+    movss                        xm4, [pf_1]
+    divss                        xm4, xm0
+    mov                dst_incr_divd, [ctxq+ResampleContext.dst_incr_div]
+    shl           min_filter_len_x4d, 2
+    lea                     dst_endq, [dstq+sizeq*4]
+
+%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
+%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
+%endif
+
+    neg           min_filter_len_x4q
+    sub                 filter_bankq, min_filter_len_x4q
+    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
+
+    ; push temp variables to stack
+%define ctx_stackq            r0mp
+%define src_stackq            r2mp
+%define update_context_stackd r4m
+
+    mov                         dstq, r1mp
+    mov                           r3, r3mp
+    lea                           r3, [dstq+r3*4]
+    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
+    PUSH                              r3
+    mov                           r3, dword [ctxq+ResampleContext.src_incr]
+    PUSH                              dword [ctxq+ResampleContext.phase_mask]
+    PUSH                              r3d
+    cvtsi2ss                     xm0, r3d
+    movss                        xm4, [pf_1]
+    divss                        xm4, xm0
+    mov        min_filter_length_x4d, [ctxq+ResampleContext.filter_length]
+    mov                       indexd, [ctxq+ResampleContext.index]
+    shl        min_filter_length_x4d, 2
+    mov                        fracd, [ctxq+ResampleContext.frac]
+    neg        min_filter_length_x4q
+    mov                 filter_bankq, [ctxq+ResampleContext.filter_bank]
+    sub                         r2mp, min_filter_length_x4q
+    sub                 filter_bankq, min_filter_length_x4q
+    PUSH                              min_filter_length_x4q
+    PUSH                              filter_bankq
+    PUSH                              dword [ctxq+ResampleContext.phase_shift]
+
+    DEFINE_ARGS src, filter1, dst, frac, index, min_filter_count_x4, filter2
+
+%define phase_shift_stackd    dword [rsp+0x0]
+%define filter_bankq          dword [rsp+0x4]
+%define min_filter_length_x4q dword [rsp+0x8]
+%define src_incrd             dword [rsp+0xc]
+%define phase_mask_stackd     dword [rsp+0x10]
+%define filter_alloc_x4q      dword [rsp+0x14]
+%define filter_allocd         dword [rsp+0x18]
+%define dst_incr_modd         dword [rsp+0x1c]
+%define dst_endq              dword [rsp+0x20]
+%define dst_incr_divd         dword [rsp+0x24]
+
+    mov                         srcq, r2mp
+%endif
+
+.loop:
+    mov                     filter1d, filter_allocd
+    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]
+%else ; x86-32
+    mov         min_filter_count_x4q, filter_bankq
+    lea                     filter1q, [min_filter_count_x4q+filter1q*4]
+    mov         min_filter_count_x4q, min_filter_length_x4q
+    mov                     filter2q, filter1q
+    add                     filter2q, filter_alloc_x4q
+%endif
+    xorps                         m0, m0, m0
+    xorps                         m2, m2, m2
+
+    align 16
+.inner_loop:
+    movups                        m1, [srcq+min_filter_count_x4q*1]
+    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
+    add         min_filter_count_x4q, mmsize
+    js .inner_loop
+
+%if cpuflag(avx)
+    vextractf128                 xm1, m0, 0x1
+    vextractf128                 xm3, m2, 0x1
+    addps                        xm0, xm1
+    addps                        xm2, xm3
+%endif
+
+    ; val += (v2 - val) * (FELEML) frac / c->src_incr;
+    cvtsi2ss                     xm1, fracd
+    subps                        xm2, xm0
+    mulps                        xm1, xm4
+    shufps                       xm1, xm1, q0000
+    mulps                        xm2, xm1
+    addps                        xm0, xm2
+
+    ; horizontal sum & store
+    movhlps                      xm1, xm0
+    addps                        xm0, xm1
+    shufps                       xm1, xm0, xm0, q0001
+    add                        fracd, dst_incr_modd
+    addps                        xm0, xm1
+    add                       indexd, dst_incr_divd
+    movss                     [dstq], xm0
+    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
+%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
+%else ; x86-32
+    DEFINE_ARGS src, phase_shift, dst, frac, index, index_incr
+%endif
+
+.skip:
+%if ARCH_X86_32
+    mov                 phase_shiftd, phase_shift_stackd
+%endif
+    mov                  index_incrd, indexd
+    add                         dstq, 4
+    and                       indexd, phase_mask_stackd
+    sar                  index_incrd, phase_shiftb
+    lea                         srcq, [srcq+index_incrq*4]
+    cmp                         dstq, dst_endq
+    jne .loop
+
+%if ARCH_X86_64
+    DEFINE_ARGS ctx, dst, src, phase_shift, index, frac
+%else ; x86-32
+    DEFINE_ARGS src, ctx, update_context, frac, index
+%endif
+
+    cmp  dword update_context_stackd, 0
+    jz .skip_store
+    ; strictly speaking, the function should always return the consumed
+    ; number of bytes; however, we only use the value if update_context
+    ; is true, so let's just leave it uninitialized otherwise
+    mov                         ctxq, ctx_stackq
+    movifnidn                    rax, srcq
+    mov [ctxq+ResampleContext.frac ], fracd
+    sub                          rax, src_stackq
+    mov [ctxq+ResampleContext.index], indexd
+    shr                          rax, 2
+
+.skip_store:
+%if ARCH_X86_32
+    ADD                          rsp, 0x28
+%endif
+    RET
+%endmacro
+
+INIT_XMM sse
+RESAMPLE_FLOAT_FNS
+
+%if HAVE_AVX_EXTERNAL
+INIT_YMM avx
+RESAMPLE_FLOAT_FNS
+%endif