ref: 75e88fab368c21cc0089222e5e08c9a15b369885
parent: 22c3594d86273cb5f2c4c7461fb5cd6bbca2df94
author: Victorien Le Couviour--Tuffet <[email protected]>
date: Tue Mar 19 10:29:00 EDT 2019
x86: cdef_filter: use 8-bit arithmetic for SSE Port of c204da0ff33a0d563d6c632b42799e4fbc48f402 for AVX-2 from Kyle Siefring. --------------------- x86_64: ------------------------------------------ before: cdef_filter_4x4_8bpc_ssse3: 141.7 after: cdef_filter_4x4_8bpc_ssse3: 131.6 before: cdef_filter_4x4_8bpc_sse4: 128.3 after: cdef_filter_4x4_8bpc_sse4: 119.0 ------------------------------------------ before: cdef_filter_4x8_8bpc_ssse3: 253.4 after: cdef_filter_4x8_8bpc_ssse3: 236.1 before: cdef_filter_4x8_8bpc_sse4: 228.5 after: cdef_filter_4x8_8bpc_sse4: 213.2 ------------------------------------------ before: cdef_filter_8x8_8bpc_ssse3: 429.6 after: cdef_filter_8x8_8bpc_ssse3: 386.9 before: cdef_filter_8x8_8bpc_sse4: 379.9 after: cdef_filter_8x8_8bpc_sse4: 335.9 ------------------------------------------ --------------------- x86_32: ------------------------------------------ before: cdef_filter_4x4_8bpc_ssse3: 184.3 after: cdef_filter_4x4_8bpc_ssse3: 163.3 before: cdef_filter_4x4_8bpc_sse4: 168.9 after: cdef_filter_4x4_8bpc_sse4: 146.1 ------------------------------------------ before: cdef_filter_4x8_8bpc_ssse3: 335.3 after: cdef_filter_4x8_8bpc_ssse3: 280.7 before: cdef_filter_4x8_8bpc_sse4: 305.1 after: cdef_filter_4x8_8bpc_sse4: 257.9 ------------------------------------------ before: cdef_filter_8x8_8bpc_ssse3: 579.1 after: cdef_filter_8x8_8bpc_ssse3: 500.5 before: cdef_filter_8x8_8bpc_sse4: 517.0 after: cdef_filter_8x8_8bpc_sse4: 455.8 ------------------------------------------
--- a/src/x86/cdef_sse.asm
+++ b/src/x86/cdef_sse.asm
@@ -31,19 +31,25 @@
%if ARCH_X86_32
pb_0: times 16 db 0
+pb_0xFF: times 16 db 0xFF
%endif
pw_128: times 8 dw 128
pw_256: times 8 dw 256
pw_2048: times 8 dw 2048
-pw_0x7FFF: times 8 dw 0x7FFF
%if ARCH_X86_32
-pw_0x8100: times 8 dw 0x8100
+pw_0x7FFF: times 8 dw 0x7FFF
+pw_0x8000: times 8 dw 0x8000
%endif
pd_0to7: dd 0, 4, 2, 6, 1, 5, 3, 7
div_table: dw 840, 840, 420, 420, 280, 280, 210, 210, 168, 168, 140, 140, 120, 120, 105, 105
dw 420, 420, 210, 210, 140, 140, 105, 105, 105, 105, 105, 105, 105, 105, 105, 105
+shufb_lohi: db 0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14, 7, 15
shufw_6543210x: db 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15
-tap_table: dw 4, 2, 3, 3, 2, 1
+tap_table: ; masks for 8-bit shift emulation
+ db 0xFF, 0x7F, 0x3F, 0x1F, 0x0F, 0x07, 0x03, 0x01
+ ; weights
+ db 4, 2, 3, 3, 2, 1
+ ; taps indices
db -1 * 16 + 1, -2 * 16 + 2
db 0 * 16 + 1, -1 * 16 + 2
db 0 * 16 + 1, 0 * 16 + 2
@@ -112,24 +118,23 @@
%endif
%endmacro
-%macro ACCUMULATE_TAP 6 ; tap_offset, shift, strength, mul_tap, w, stride
- %if ARCH_X86_64
+%macro ACCUMULATE_TAP 7 ; tap_offset, shift, shift_mask, strength, mul_tap, w, stride
; load p0/p1
movsx offq, byte [dirq+kq+%1] ; off1
- %if %5 == 4
- movq m5, [stkq+offq*2+%6*0] ; p0
- movhps m5, [stkq+offq*2+%6*1]
- %else
- movu m5, [stkq+offq*2+%6*0] ; p0
- %endif
+ %if %6 == 4
+ movq m5, [stkq+offq*2+%7*0] ; p0
+ movhps m5, [stkq+offq*2+%7*1]
+ %else
+ movu m5, [stkq+offq*2+%7*0] ; p0
+ %endif
neg offq ; -off1
- %if %5 == 4
- movq m6, [stkq+offq*2+%6*0] ; p1
- movhps m6, [stkq+offq*2+%6*1]
- %else
- movu m6, [stkq+offq*2+%6*0] ; p1
- %endif
- %if cpuflag(sse4)
+ %if %6 == 4
+ movq m6, [stkq+offq*2+%7*0] ; p1
+ movhps m6, [stkq+offq*2+%7*1]
+ %else
+ movu m6, [stkq+offq*2+%7*0] ; p1
+ %endif
+ %if cpuflag(sse4)
; out of bounds values are set to a value that is a both a large unsigned
; value and a negative signed value.
; use signed max and unsigned min to remove them
@@ -137,7 +142,8 @@
pminuw m8, m5
pmaxsw m7, m6
pminuw m8, m6
- %else
+ %else
+ %if ARCH_X86_64
pcmpeqw m9, m14, m5
pcmpeqw m10, m14, m6
pandn m9, m5
@@ -146,98 +152,42 @@
pminsw m8, m5 ; min after p0
pmaxsw m7, m10 ; max after p1
pminsw m8, m6 ; min after p1
- %endif
-
- ; accumulate sum[m13] over p0/p1
- psubw m5, m4 ; diff_p0(p0 - px)
- psubw m6, m4 ; diff_p1(p1 - px)
- pabsw m9, m5
- pabsw m10, m6
- psrlw m11, m9, %2
- psignw m12, %4, m5
- psubusw m5, %3, m11
- psrlw m11, m10, %2
- %if cpuflag(sse4)
- pminuw m5, m9
%else
- pminsw m5, m9 ; constrain(diff_p0)
- %endif
- psignw m9, %4, m6
- psubusw m6, %3, m11
- %if cpuflag(sse4)
- pminuw m6, m10
- %else
- pminsw m6, m10 ; constrain(diff_p1)
- %endif
- pmullw m5, m12 ; constrain(diff_p0) * taps
- pmullw m6, m9 ; constrain(diff_p1) * taps
- paddw m13, m5
- paddw m13, m6
- %else
- ; load p0
- movsx offq, byte [dirq+kq+%1] ; off1
- %if %5 == 4
- movq m5, [stkq+offq*2+%6*0] ; p0
- movhps m5, [stkq+offq*2+%6*1]
- %else
- movu m5, [stkq+offq*2+%6*0] ; p0
- %endif
- %if cpuflag(sse4)
- ; out of bounds values are set to a value that is a both a large unsigned
- ; value and a negative signed value.
- ; use signed max and unsigned min to remove them
- pmaxsw m7, m5 ; max after p0
- pminuw m8, m5 ; min after p0
- %else
pcmpeqw m3, m5, OUT_OF_BOUNDS_MEM
pandn m3, m5
pmaxsw m7, m3 ; max after p0
pminsw m8, m5 ; min after p0
- %endif
-
- ; accumulate sum[m7] over p0
- psubw m5, m4 ; diff_p0(p0 - px)
- psignw m6, %4, m5
- pabsw m5, m5
- psrlw m3, m5, %2
- paddsw m3, %3
- pxor m3, [PIC_sym(pw_0x7FFF)]
- pminsw m3, m5 ; constrain(diff_p0)
- pmullw m3, m6 ; constrain(diff_p0) * taps
- paddw m13, m3
-
- ; load p1
- neg offq ; -off1
- %if %5 == 4
- movq m5, [stkq+offq*2+%6*0] ; p1
- movhps m5, [stkq+offq*2+%6*1]
- %else
- movu m5, [stkq+offq*2+%6*0] ; p1
- %endif
- %if cpuflag(sse4)
- ; out of bounds values are set to a value that is a both a large unsigned
- ; value and a negative signed value.
- ; use signed max and unsigned min to remove them
- pmaxsw m7, m5 ; max after p1
- pminuw m8, m5 ; min after p1
- %elif
- pcmpeqw m3, m5, OUT_OF_BOUNDS_MEM
- pandn m3, m5
+ pcmpeqw m3, m6, OUT_OF_BOUNDS_MEM
+ pandn m3, m6
pmaxsw m7, m3 ; max after p1
- pminsw m8, m5 ; min after p1
+ pminsw m8, m6 ; min after p1
%endif
+ %endif
- ; accumulate sum[m7] over p1
- psubw m5, m4 ; diff_p1(p1 - px)
- psignw m6, %4, m5
- pabsw m5, m5
- psrlw m3, m5, %2
- paddsw m3, %3
- pxor m3, [PIC_sym(pw_0x7FFF)]
- pminsw m3, m5 ; constrain(diff_p1)
- pmullw m3, m6 ; constrain(diff_p1) * taps
- paddw m13, m3
+ ; accumulate sum[m13] over p0/p1
+ psubw m5, m4 ; diff_p0(p0 - px)
+ psubw m6, m4 ; diff_p1(p1 - px)
+ packsswb m5, m6 ; convert pixel diff to 8-bit
+ %if ARCH_X86_64 && cpuflag(sse4)
+ pshufb m5, m14 ; group diffs p0 and p1 into pairs
+ %else
+ pshufb m5, [PIC_sym(shufb_lohi)]
%endif
+ pabsb m6, m5
+ psignb m9, %5, m5
+ %if ARCH_X86_64
+ psrlw m10, m6, %2 ; emulate 8-bit shift
+ pand m10, %3
+ psubusb m5, %4, m10
+ %else
+ psrlw m5, m6, %2 ; emulate 8-bit shift
+ pand m5, %3
+ paddusb m5, %4
+ pxor m5, [PIC_sym(pb_0xFF)]
+ %endif
+ pminub m5, m6 ; constrain(diff_p)
+ pmaddubsw m5, m9 ; constrain(diff_p) * taps
+ paddw m13, m5
%endmacro
%macro PMOVZXBW 2-3 0 ; %3 = half
@@ -284,7 +234,7 @@
%macro CDEF_FILTER 3 ; w, h, stride
%if cpuflag(sse4)
- %define OUT_OF_BOUNDS 0x81008100
+ %define OUT_OF_BOUNDS 0x80008000
%else
%define OUT_OF_BOUNDS 0x7FFF7FFF
%endif
@@ -302,7 +252,7 @@
%define px rsp+3*16+2*%3
%else
-cglobal cdef_filter_%1x%2, 2, 7, 8, - 5 * 16 - (%2+4)*%3, \
+cglobal cdef_filter_%1x%2, 2, 7, 8, - 7 * 16 - (%2+4)*%3, \
dst, stride, left, top, stride3, dst4, edge
SAVE_ARG left, 2
SAVE_ARG top, 3
@@ -315,7 +265,7 @@
LEA PIC_reg, PIC_base_offset
%if cpuflag(sse4)
- %define OUT_OF_BOUNDS_MEM [PIC_sym(pw_0x8100)]
+ %define OUT_OF_BOUNDS_MEM [PIC_sym(pw_0x8000)]
%else
%define OUT_OF_BOUNDS_MEM [PIC_sym(pw_0x7FFF)]
%endif
@@ -322,7 +272,7 @@
%define m15 [PIC_sym(pb_0)]
- %define px esp+5*16+2*%3
+ %define px esp+7*16+2*%3
%endif
mov edged, r8m
@@ -640,52 +590,74 @@
cmovl pridmpd, dampingd
neg secdmpd
cmovl secdmpd, dampingd
+ %if ARCH_X86_64
mov [rsp+ 0], pridmpq ; pri_shift
mov [rsp+16], secdmpq ; sec_shift
- %if ARCH_X86_32
- mov dword [esp+ 4], 0 ; zero upper 32 bits of psraw
- mov dword [esp+20], 0 ; source operand in ACCUMULATE_TAP
- %define PIC_reg r6
+ %else
+ mov [esp+0x00], pridmpd
+ mov [esp+0x30], secdmpd
+ mov dword [esp+0x04], 0 ; zero upper 32 bits of psrlw
+ mov dword [esp+0x34], 0 ; source operand in ACCUMULATE_TAP
+ %define PIC_reg r4
LOAD_PIC_REG 8
%endif
+ DEFINE_ARGS dst, stride, pridmp, table, pri, sec, secdmp
+ lea tableq, [PIC_sym(tap_table)]
+ %if ARCH_X86_64
+ SWAP m2, m11
+ SWAP m3, m12
+ %endif
+ movd m2, [tableq+pridmpq]
+ movd m3, [tableq+secdmpq]
+ pshufb m2, m15 ; pri_shift_mask
+ pshufb m3, m15 ; sec_shift_mask
+ %if ARCH_X86_64
+ SWAP m2, m11
+ SWAP m3, m12
+ %else
+ %define PIC_reg r6
+ mov PIC_reg, r4
+ DEFINE_ARGS dst, stride, dir, table, pri, sec, secdmp
+ LOAD_ARG pri
+ LOAD_ARG dir, 1
+ mova [esp+0x10], m2
+ mova [esp+0x40], m3
+ %endif
+
; pri/sec_taps[k] [4 total]
- DEFINE_ARGS dst, stride, tap, dummy, pri, sec
+ DEFINE_ARGS dst, stride, dummy, tap, pri, sec
movd m0, prid
movd m1, secd
%if ARCH_X86_64
- %if cpuflag(sse4)
- psrlw m14, 7
- SWAP m14, m15
- %else
- mova m15, [pw_256]
- %endif
pshufb m0, m15
pshufb m1, m15
%else
- mova m2, [PIC_sym(pw_256)]
- mova m3, [PIC_sym(pw_0x7FFF)]
+ mova m2, m15
+ mova m3, [PIC_sym(pb_0xFF)]
pshufb m0, m2
pshufb m1, m2
pxor m0, m3
pxor m1, m3
mova [esp+0x20], m0
- mova [esp+0x30], m1
+ mova [esp+0x50], m1
%endif
and prid, 1
- lea tapq, [PIC_sym(tap_table)]
- lea priq, [tapq+priq*4] ; pri_taps
- lea secq, [tapq+8] ; sec_taps
+ lea priq, [tapq+8+priq*2] ; pri_taps
+ lea secq, [tapq+12] ; sec_taps
+ %if ARCH_X86_64 && cpuflag(sse4)
+ mova m14, [shufb_lohi]
+ %endif
+
; off1/2/3[k] [6 total] from [tapq+12+(dir+0/2/6)*2+k]
- DEFINE_ARGS dst, stride, tap, dir, pri, sec
+ DEFINE_ARGS dst, stride, dir, tap, pri, sec
%if ARCH_X86_64
mov dird, r6m
- lea tapq, [tapq+dirq*2+12]
+ lea dirq, [tapq+14+dirq*2]
DEFINE_ARGS dst, stride, dir, stk, pri, sec, h, off, k
%else
- LOAD_ARG dir, 1
- lea tapd, [tapd+dird*2+12]
+ lea dird, [tapd+14+dird*2]
DEFINE_ARGS dst, stride, dir, stk, pri, sec
%define hd dword [esp+8]
%define offq dstq
@@ -693,9 +665,9 @@
%endif
mov hd, %1*%2*2/mmsize
lea stkq, [px]
- movif32 [esp+0x1C], strided
+ movif32 [esp+0x3C], strided
.v_loop:
- movif32 [esp+0x18], dstd
+ movif32 [esp+0x38], dstd
mov kq, 1
%if %1 == 4
movq m4, [stkq+%3*0]
@@ -705,7 +677,7 @@
%endif
%if ARCH_X86_32
- %xdefine m11 m6
+ %xdefine m9 m3
%xdefine m13 m7
%xdefine m7 m0
%xdefine m8 m1
@@ -716,46 +688,41 @@
mova m8, m4 ; min
.k_loop:
%if ARCH_X86_64
- movd m2, [priq+kq*2] ; pri_taps
- movd m3, [secq+kq*2] ; sec_taps
+ movd m2, [priq+kq] ; pri_taps
+ movd m3, [secq+kq] ; sec_taps
pshufb m2, m15
pshufb m3, m15
- ACCUMULATE_TAP 0*2, [rsp+ 0], m0, m2, %1, %3
- ACCUMULATE_TAP 2*2, [rsp+16], m1, m3, %1, %3
- ACCUMULATE_TAP 6*2, [rsp+16], m1, m3, %1, %3
+ ACCUMULATE_TAP 0*2, [rsp+ 0], m11, m0, m2, %1, %3
+ ACCUMULATE_TAP 2*2, [rsp+16], m12, m1, m3, %1, %3
+ ACCUMULATE_TAP 6*2, [rsp+16], m12, m1, m3, %1, %3
%else
- movd m2, [priq+kq*2] ; pri_taps
- pshufb m2, [PIC_sym(pw_256)]
- ACCUMULATE_TAP 0*2, [esp+0x00], [esp+0x20], m2, %1, %3
+ movd m2, [priq+kq] ; pri_taps
+ pshufb m2, m15
+ ACCUMULATE_TAP 0*2, [esp+0x00], [esp+0x10], [esp+0x20], m2, %1, %3
- movd m2, [secq+kq*2] ; sec_taps
- pshufb m2, [PIC_sym(pw_256)]
- ACCUMULATE_TAP 2*2, [esp+0x10], [esp+0x30], m2, %1, %3
- ACCUMULATE_TAP 6*2, [esp+0x10], [esp+0x30], m2, %1, %3
+ movd m2, [secq+kq] ; sec_taps
+ pshufb m2, m15
+ ACCUMULATE_TAP 2*2, [esp+0x30], [esp+0x40], [esp+0x50], m2, %1, %3
+ ACCUMULATE_TAP 6*2, [esp+0x30], [esp+0x40], [esp+0x50], m2, %1, %3
%endif
dec kq
jge .k_loop
- %if ARCH_X86_64
- %if cpuflag(sse4)
- SWAP m14, m15
- pcmpgtw m11, m15, m13
- %else
- pxor m11, m11
- pcmpgtw m11, m13
- %endif
+ %if cpuflag(sse4)
+ pcmpgtw m6, m15, m13
%else
- pcmpgtw m11, m15, m13
+ pxor m6, m6
+ pcmpgtw m6, m13
%endif
- paddw m13, m11
+ paddw m13, m6
pmulhrsw m13, [PIC_sym(pw_2048)]
paddw m4, m13
pminsw m4, m7
pmaxsw m4, m8
packuswb m4, m4
- movif32 dstd, [esp+0x18]
- movif32 strided, [esp+0x1C]
+ movif32 dstd, [esp+0x38]
+ movif32 strided, [esp+0x3C]
%if %1 == 4
movd [dstq+strideq*0], m4
psrlq m4, 32