x/dav1d
1
0
Fork 0
mirror of https://code.videolan.org/videolan/dav1d synced 2026-06-11 04:03:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

x86/itx: share pass2 loop between {16,32}x32 dct^2 functions

Browse Source
This commit is contained in:
Ronald S. Bultje
2021-08-12 14:47:14 -04:00
parent 9cf9d4a613
commit 5455e8250c
1 changed files with 35 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/x86/itx16_sse.asm
+35 -49
View File
@@ -5305,15 +5305,13 @@ cglobal inv_txfm_add_dct_dct_16x32_16bpc, 4, 7, 16, 0-77*16, \
jge .loop_pass1
; pass=2
add rsp, 9*16
%if ARCH_X86_64
mov r6, dstq
%if WIN64
mov r7, [rsp+gprsize*1+76*16]
%endif
%else
mov dstq, [rsp+gprsize*1+76*16]
mov dstq, [rsp+gprsize*1+67*16]
%endif
mov eobd, [rsp+gprsize*0+76*16]
mov eobd, [rsp+gprsize*0+67*16]
cmp eobd, 44
jl .load_veryfast
cmp eobd, 151
@@ -5328,21 +5326,38 @@ cglobal inv_txfm_add_dct_dct_16x32_16bpc, 4, 7, 16, 0-77*16, \
lea r4, [o(m_suffix(idct_8x32_internal_8bpc, _ssse3).main_veryfast)]
; fall-through
.run:
add rsp, 9*16
call m(inv_txfm_add_dct_dct_8x32_16bpc).pass2
%if ARCH_X86_64
lea dstq, [r6+16]
lea r2, [dstq+32]
mov r7, -4
%else
mov dstq, [rsp+gprsize*1+67*16]
add dstq, 16
lea r2, [rsp+67*16]
mov dword [r2+0*gprsize], 2
%endif
add rsp, 32*16
jmp m(inv_txfm_add_dct_dct_16x32_16bpc).loop_pass2_entry
.loop_pass2:
mova m0, [rsp+16* 3]
.loop_pass2_entry:
%if ARCH_X86_32
mov dstq, [r2+1*gprsize]
%endif
call m(inv_txfm_add_dct_dct_8x32_16bpc).pass2
%assign stack_size (stack_size-41*16)
add rsp, 32*16
%if ARCH_X86_64
add r7, 2
lea dstq, [r2+r7*8]
jl .loop_pass2
%if WIN64
mov r7, [rsp+gprsize*1+3*16]
%endif
%else
add dword [r2+1*gprsize], 16
dec dword [r2+0*gprsize]
jg .loop_pass2
%endif
%assign stack_size (stack_size-73*16)
%if STACK_ALIGNMENT >= 16
%assign stack_size_padded (stack_size_padded-41*16)
%assign stack_offset (stack_offset-41*16)
%assign stack_size_padded (stack_size_padded-73*16)
%assign stack_offset (stack_offset-73*16)
%else
%xdefine rstkm [rsp + stack_size]
%endif
@@ -6375,9 +6390,9 @@ cglobal inv_txfm_add_dct_dct_32x32_16bpc, 4, 7, 16, 0-(5*32+1)*16, \
; remove entirely-zero iterations
%if ARCH_X86_32
mov [rsp+5*32*16+0*gprsize], dstq
mov [rsp+5*32*16+1*gprsize], dstq
%elif WIN64
mov [rsp+5*32*16+0*gprsize], r7
mov [rsp+5*32*16+1*gprsize], r7
%endif
%undef cmp
mov r5d, 14
@@ -6410,7 +6425,7 @@ cglobal inv_txfm_add_dct_dct_32x32_16bpc, 4, 7, 16, 0-(5*32+1)*16, \
.end_zero_loop:
; actual first pass after skipping all-zero data
mov [rsp+gprsize*1+5*32*16], eobd
mov [rsp+gprsize*0+5*32*16], eobd
.loop_pass1:
mova m0, [cq+128* 1+r5*8]
mova m1, [cq+128* 7+r5*8]
@@ -6552,7 +6567,7 @@ cglobal inv_txfm_add_dct_dct_32x32_16bpc, 4, 7, 16, 0-(5*32+1)*16, \
jge .loop_pass1
; pass=2 code starts here
mov eobd, [rsp+gprsize*1+5*32*16]
mov eobd, [rsp+gprsize*0+5*32*16]
add rsp, 29*16
cmp eobd, 36
jl .load_veryfast
@@ -6573,38 +6588,9 @@ cglobal inv_txfm_add_dct_dct_32x32_16bpc, 4, 7, 16, 0-(5*32+1)*16, \
mov r7, -8
%else
lea r2, [rsp+(4*32+3)*16]
mov dword [r2+1*gprsize], 4
mov dword [r2+0*gprsize], 4
%endif
jmp .loop_pass2_entry
.loop_pass2:
mova m0, [rsp+16* 3]
.loop_pass2_entry:
%if ARCH_X86_32
mov dstq, [r2+0*gprsize]
%endif
call m(inv_txfm_add_dct_dct_8x32_16bpc).pass2
add rsp, 32*16
%if ARCH_X86_64
add r7, 2
lea dstq, [r2+r7*8]
jl .loop_pass2
%else
mov dstq, [r2+0*gprsize]
add dword [r2+0*gprsize], 16
dec dword [r2+1*gprsize]
jg .loop_pass2
%endif
%assign stack_size (stack_size-(4*32+29)*16)
%if STACK_ALIGNMENT >= 16
%assign stack_size_padded (stack_size_padded-(4*32+29)*16)
%assign stack_offset (stack_offset-(4*32+29)*16)
%else
%xdefine rstkm [rsp + stack_size]
%endif
%if WIN64
mov r7, [rsp+3*16+gprsize*0]
%endif
RET
jmp m(inv_txfm_add_dct_dct_16x32_16bpc).loop_pass2_entry
.dconly:
imul r5d, [cq], 2896