idct_mmx.asm

1.Xvid Mpeg4 1.13 version 2.Building 0 error(s), 0 warning(s) By Jesse Stone Taiwan

  dw   5315,  22654,  22654, -26722 ; w31 w30 w27 w26

;=============================================================================
; Helper macros for the code
;=============================================================================

;-----------------------------------------------------------------------------
; DCT_8_INV_ROW_XMM  INP, OUT, TABLE, ROUNDER
;-----------------------------------------------------------------------------

%macro DCT_8_INV_ROW_XMM 4
  movq mm0, [%1]                ; 0     ; x3 x2 x1 x0
  movq mm1, [%1+8]              ; 1     ; x7 x6 x5 x4
  movq mm2, mm0                 ; 2     ; x3 x2 x1 x0
  movq mm3, [%3]                ; 3     ; w05 w04 w01 w00
  pshufw mm0, mm0, 10001000b    ; x2 x0 x2 x0
  movq mm4, [%3+8]              ; 4     ; w07 w06 w03 w02
  movq mm5, mm1                 ; 5     ; x7 x6 x5 x4
  pmaddwd mm3, mm0              ; x2*w05+x0*w04 x2*w01+x0*w00
  movq mm6, [%3+32]             ; 6     ; w21 w20 w17 w16
  pshufw mm1, mm1, 10001000b    ; x6 x4 x6 x4
  pmaddwd mm4, mm1              ; x6*w07+x4*w06 x6*w03+x4*w02
  movq mm7, [%3+40]             ; 7    ; w23 w22 w19 w18
  pshufw mm2, mm2, 11011101b    ; x3 x1 x3 x1
  pmaddwd mm6, mm2              ; x3*w21+x1*w20 x3*w17+x1*w16
  pshufw mm5, mm5, 11011101b    ; x7 x5 x7 x5
  pmaddwd mm7, mm5              ; x7*w23+x5*w22 x7*w19+x5*w18
  paddd mm3, [%4]               ; +%4
  pmaddwd mm0, [%3+16]          ; x2*w13+x0*w12 x2*w09+x0*w08
  paddd mm3, mm4                ; 4     ; a1=sum(even1) a0=sum(even0)
  pmaddwd mm1, [%3+24]          ; x6*w15+x4*w14 x6*w11+x4*w10
  movq mm4, mm3                 ; 4     ; a1 a0
  pmaddwd mm2, [%3+48]          ; x3*w29+x1*w28 x3*w25+x1*w24
  paddd mm6, mm7                ; 7     ; b1=sum(odd1) b0=sum(odd0)
  pmaddwd mm5, [%3+56]          ; x7*w31+x5*w30 x7*w27+x5*w26
  paddd mm3, mm6                ; a1+b1 a0+b0
  paddd mm0, [%4]               ; +%4
  psrad mm3, SHIFT_INV_ROW      ; y1=a1+b1 y0=a0+b0
  paddd mm0, mm1                ; 1     ; a3=sum(even3) a2=sum(even2)
  psubd mm4, mm6                ; 6     ; a1-b1 a0-b0
  movq mm7, mm0                 ; 7     ; a3 a2
  paddd mm2, mm5                ; 5     ; b3=sum(odd3) b2=sum(odd2)
  paddd mm0, mm2                ; a3+b3 a2+b2
  psrad mm4, SHIFT_INV_ROW      ; y6=a1-b1 y7=a0-b0
  psubd mm7, mm2                ; 2     ; a3-b3 a2-b2
  psrad mm0, SHIFT_INV_ROW      ; y3=a3+b3 y2=a2+b2
  psrad mm7, SHIFT_INV_ROW      ; y4=a3-b3 y5=a2-b2
  packssdw mm3, mm0             ; 0     ; y3 y2 y1 y0
  packssdw mm7, mm4             ; 4     ; y6 y7 y4 y5
  movq [%2], mm3                ; 3     ; save y3 y2 y1 y0
  pshufw mm7, mm7, 10110001b    ; y7 y6 y5 y4
  movq [%2+8], mm7              ; 7     ; save y7 y6 y5 y4
%endmacro

;-----------------------------------------------------------------------------
;
; The first stage DCT 8x8 - forward DCTs of columns
;
; The %2puts are multiplied
; for rows 0,4 - on cos_4_16,
; for rows 1,7 - on cos_1_16,
; for rows 2,6 - on cos_2_16,
; for rows 3,5 - on cos_3_16
; and are shifted to the left for rise of accuracy
;
;-----------------------------------------------------------------------------
;
; The 8-point scaled forward DCT algorithm (26a8m)
;
;-----------------------------------------------------------------------------
;
; #define DCT_8_FRW_COL(x, y)
;{
; short t0, t1, t2, t3, t4, t5, t6, t7;
; short tp03, tm03, tp12, tm12, tp65, tm65;
; short tp465, tm465, tp765, tm765;
;
; t0 = LEFT_SHIFT ( x[0] + x[7] );
; t1 = LEFT_SHIFT ( x[1] + x[6] );
; t2 = LEFT_SHIFT ( x[2] + x[5] );
; t3 = LEFT_SHIFT ( x[3] + x[4] );
; t4 = LEFT_SHIFT ( x[3] - x[4] );
; t5 = LEFT_SHIFT ( x[2] - x[5] );
; t6 = LEFT_SHIFT ( x[1] - x[6] );
; t7 = LEFT_SHIFT ( x[0] - x[7] );
;
; tp03 = t0 + t3;
; tm03 = t0 - t3;
; tp12 = t1 + t2;
; tm12 = t1 - t2;
;
; y[0] = tp03 + tp12;
; y[4] = tp03 - tp12;
;
; y[2] = tm03 + tm12 * tg_2_16;
; y[6] = tm03 * tg_2_16 - tm12;
;
; tp65 =(t6 +t5 )*cos_4_16;
; tm65 =(t6 -t5 )*cos_4_16;
;
; tp765 = t7 + tp65;
; tm765 = t7 - tp65;
; tp465 = t4 + tm65;
; tm465 = t4 - tm65;
;
; y[1] = tp765 + tp465 * tg_1_16;
; y[7] = tp765 * tg_1_16 - tp465;
; y[5] = tm765 * tg_3_16 + tm465;
; y[3] = tm765 - tm465 * tg_3_16;
;}
;
;-----------------------------------------------------------------------------

;-----------------------------------------------------------------------------
; DCT_8_INV_COL_4  INP,OUT
;-----------------------------------------------------------------------------

%macro DCT_8_INV_COL 2
  movq mm0, [tg_3_16 wrt rip]
  movq mm3, [%1+16*3]
  movq mm1, mm0             ; tg_3_16
  movq mm5, [%1+16*5]
  pmulhw mm0, mm3           ; x3*(tg_3_16-1)
  movq mm4, [tg_1_16 wrt rip]
  pmulhw mm1, mm5           ; x5*(tg_3_16-1)
  movq mm7, [%1+16*7]
  movq mm2, mm4             ; tg_1_16
  movq mm6, [%1+16*1]
  pmulhw mm4, mm7           ; x7*tg_1_16
  paddsw mm0, mm3           ; x3*tg_3_16
  pmulhw mm2, mm6           ; x1*tg_1_16
  paddsw mm1, mm3           ; x3+x5*(tg_3_16-1)
  psubsw mm0, mm5           ; x3*tg_3_16-x5 = tm35
  movq mm3, [ocos_4_16 wrt rip]
  paddsw mm1, mm5           ; x3+x5*tg_3_16 = tp35
  paddsw mm4, mm6           ; x1+tg_1_16*x7 = tp17
  psubsw mm2, mm7           ; x1*tg_1_16-x7 = tm17
  movq mm5, mm4             ; tp17
  movq mm6, mm2             ; tm17
  paddsw mm5, mm1           ; tp17+tp35 = b0
  psubsw mm6, mm0           ; tm17-tm35 = b3
  psubsw mm4, mm1           ; tp17-tp35 = t1
  paddsw mm2, mm0           ; tm17+tm35 = t2
  movq mm7, [tg_2_16 wrt rip]
  movq mm1, mm4             ; t1
  movq [%2+3*16], mm5       ; save b0
  paddsw mm1, mm2           ; t1+t2
  movq [%2+5*16], mm6       ; save b3
  psubsw mm4, mm2           ; t1-t2
  movq mm5, [%1+2*16]
  movq mm0, mm7             ; tg_2_16
  movq mm6, [%1+6*16]
  pmulhw mm0, mm5           ; x2*tg_2_16
  pmulhw mm7, mm6           ; x6*tg_2_16
; slot
  pmulhw mm1, mm3           ; ocos_4_16*(t1+t2) = b1/2
; slot
  movq mm2, [%1+0*16]
  pmulhw mm4, mm3           ; ocos_4_16*(t1-t2) = b2/2
  psubsw mm0, mm6           ; t2*tg_2_16-x6 = tm26
  movq mm3, mm2             ; x0
  movq mm6, [%1+4*16]
  paddsw mm7, mm5           ; x2+x6*tg_2_16 = tp26
  paddsw mm2, mm6           ; x0+x4 = tp04
  psubsw mm3, mm6           ; x0-x4 = tm04
  movq mm5, mm2             ; tp04
  movq mm6, mm3             ; tm04
  psubsw mm2, mm7           ; tp04-tp26 = a3
  paddsw mm3, mm0           ; tm04+tm26 = a1
  paddsw mm1, mm1           ; b1
  paddsw mm4, mm4           ; b2
  paddsw mm5, mm7           ; tp04+tp26 = a0
  psubsw mm6, mm0           ; tm04-tm26 = a2
  movq mm7, mm3             ; a1
  movq mm0, mm6             ; a2
  paddsw mm3, mm1           ; a1+b1
  paddsw mm6, mm4           ; a2+b2
  psraw mm3, SHIFT_INV_COL  ; dst1
  psubsw mm7, mm1           ; a1-b1
  psraw mm6, SHIFT_INV_COL  ; dst2
  psubsw mm0, mm4           ; a2-b2
  movq mm1, [%2+3*16]       ; load b0
  psraw mm7, SHIFT_INV_COL  ; dst6
  movq mm4, mm5             ; a0
  psraw mm0, SHIFT_INV_COL  ; dst5
  movq [%2+1*16], mm3
  paddsw mm5, mm1           ; a0+b0
  movq [%2+2*16], mm6
  psubsw mm4, mm1           ; a0-b0
  movq mm3, [%2+5*16]       ; load b3
  psraw mm5, SHIFT_INV_COL  ; dst0
  movq mm6, mm2             ; a3
  psraw mm4, SHIFT_INV_COL  ; dst7
  movq [%2+5*16], mm0
  paddsw mm2, mm3           ; a3+b3
  movq [%2+6*16], mm7
  psubsw mm6, mm3           ; a3-b3
  movq [%2+0*16], mm5
  psraw mm2, SHIFT_INV_COL  ; dst3
  movq [%2+7*16], mm4
  psraw mm6, SHIFT_INV_COL  ; dst4
  movq [%2+3*16], mm2
  movq [%2+4*16], mm6
%endmacro

;=============================================================================
; Code
;=============================================================================

SECTION .text align=16

cglobal idct_x86_64

;-----------------------------------------------------------------------------
; void idct_xmm(uint16_t block[64]);
;-----------------------------------------------------------------------------

ALIGN 16
idct_x86_64:
    mov rax, rdi	; block[]

    ;; Process each row
    lea rsi, [tab_i_04_xmm wrt rip]
    lea rdi, [rounder_0 wrt rip]
    DCT_8_INV_ROW_XMM rax+0*16, rax+0*16, rsi, rdi
    
    lea rsi, [tab_i_17_xmm wrt rip]
    lea rdi, [rounder_1 wrt rip]
    DCT_8_INV_ROW_XMM rax+1*16, rax+1*16, rsi, rdi
    
    lea rsi, [tab_i_26_xmm wrt rip]
    lea rdi, [rounder_2 wrt rip]
    DCT_8_INV_ROW_XMM rax+2*16, rax+2*16, rsi, rdi
    
    lea rsi, [tab_i_35_xmm wrt rip]
    lea rdi, [rounder_3 wrt rip]
    DCT_8_INV_ROW_XMM rax+3*16, rax+3*16, rsi, rdi
    
    lea rsi, [tab_i_04_xmm wrt rip]
    lea rdi, [rounder_4 wrt rip]
    DCT_8_INV_ROW_XMM rax+4*16, rax+4*16, rsi, rdi
    
    lea rsi, [tab_i_35_xmm wrt rip]
    lea rdi, [rounder_5 wrt rip]
    DCT_8_INV_ROW_XMM rax+5*16, rax+5*16, rsi, rdi
    
    lea rsi, [tab_i_26_xmm wrt rip]
    lea rdi, [rounder_6 wrt rip]
    DCT_8_INV_ROW_XMM rax+6*16, rax+6*16, rsi, rdi
    
    lea rsi, [tab_i_17_xmm wrt rip]
    lea rdi, [rounder_7 wrt rip]
    DCT_8_INV_ROW_XMM rax+7*16, rax+7*16, rsi, rdi


    ;; Process the columns (4 at a time)
    DCT_8_INV_COL rax+0, rax+0
    DCT_8_INV_COL rax+8, rax+8

    ret
.endfunc