FHT: SSE_patch_1

git-svn-id: svn://kolibrios.org@2215 a494cfbc-eb01-0410-851d-a64ba20cac60

programs/other/fft/FHT4A.asm

@@ -25,7 +25,7 @@ use32
 
 include '../../macros.inc'
 include '../debug.inc'
-include 'FHT4i.inc'
+include 'fht4code.asm'
 
 
 START:				; start of execution

programs/other/fft/fht4code.asm

@@ -183,7 +183,46 @@ step1:
 	add	ebx, 32
 	cmp	ebx, esi
 	jnz	.loop
-ret
+ret
+
+;=================================================================
+; SSE3 version:	Step1
+;
+;==========================
+
+align 4
+step1_sse:
+	mov	ebx, [esp+8]
+	mov	esi, [esp+4]
+	shl	esi, 3
+	add	esi, ebx
+
+.loop:
+	movddup     xmm0, [ebx]     ;   xmm0: f0 ; f0
+	movddup     xmm1, [ebx+8]   ;   xmm1: f1 ; f1
+	addsubpd    xmm0, xmm1      ;   xmm0: t1 ; t2   ( + - )
+    movddup     xmm1, [ebx+16]  ;   xmm1: f2 ; f2
+    movddup     xmm2, [ebx+24]  ;   xmm2: f3 ; f3
+	addsubpd    xmm1, xmm2      ;   xmm1: t3 ; t4   ( + - )
+
+    movddup     xmm2, xmm0      ;   xmm2: t2 ; t2
+    movddup     xmm3, xmm1      ;   xmm3: t4 ; t4
+	addsubpd    xmm2, xmm3      ;   xmm2: 2+4; 2-4  
+    shufpd      xmm2, xmm2, 1   ;   xmm2: 2-4; 2+4
+    movapd      [ebx+16], xmm2
+
+    shufpd      xmm0, xmm0, 1   ;   xmm0: t2 ; t1
+    shufpd      xmm1, xmm1, 1   ;   xmm1: t4 ; t3
+    movddup     xmm2, xmm0      ;   xmm2: t1 ; t1
+    movddup     xmm3, xmm1      ;   xmm3: t3 ; t3
+	addsubpd    xmm2, xmm3      ;   xmm2: 1+3; 1-3  
+    shufpd      xmm2, xmm2, 1   ;   xmm2: 1-3; 1+3
+    movapd      [ebx], xmm2
+
+	add	ebx, 32
+	cmp	ebx, esi
+	jnz	.loop
+ret
 
 ;       local stack definitions
 ;===========================================================================

programs/other/rtfread/trunk/debug.inc

@@ -1,137 +0,0 @@
-macro debug_print str
-{
-   local ..string, ..label
-
-   jmp ..label
-   ..string db str,0
-  ..label:
-
-   pushf
-   pushad
-   mov  edx,..string
-   call debug_outstr
-   popad
-   popf
-}
-
-dps fix debug_print
-
-macro debug_print_dec arg
-{
-   pushf
-   pushad
-   if ~arg eq eax
-     mov  eax,arg
-   end if
-   call debug_outdec
-   popad
-   popf
-}
-
-dpd fix debug_print_dec
-
-;---------------------------------
-debug_outdec:           ;(eax - num, edi-str)
-        push 10         ;2
-        pop ecx         ;1
-        push -'0'       ;2
-    .l0:
-        xor edx,edx     ;2
-        div ecx         ;2
-        push edx        ;1
-        test eax,eax    ;2
-        jnz .l0         ;2
-    .l1:
-        pop eax         ;1
-        add al,'0'      ;2
-        call debug_outchar ; stosb
-        jnz .l1         ;2
-        ret             ;1
-;---------------------------------
-
-debug_outchar:          ; al - char
-   pushf
-   pushad
-   mov  cl,al
-   mov  eax,63
-   mov  ebx,1
-   mcall
-   popad
-   popf
-ret
-
-debug_outstr:
-   mov  eax,63
-   mov  ebx,1
- @@:
-   mov  cl,[edx]
-   test cl,cl
-   jz   @f
-   mcall
-   inc  edx
-   jmp  @b
- @@:
-   ret
-
-_debug_crlf db 13, 10, 0
-
-macro newline
-{
-  pushf
-  pushad
-  mov edx, _debug_crlf
-  call debug_outstr
-  popad
-  popf
-}
-
-macro print message
-{
-  dps message
-  newline
-}
-
-macro pregs
-{
-  dps "EAX: "
-  dpd eax
-  dps "   EBX: "
-  dpd ebx
-  newline
-  dps "ECX: "
-  dpd ecx
-  dps "   EDX: "
-  dpd edx
-  newline
-}
-
-macro debug_print_hex arg
-{
-    pushf
-    pushad
-    if ~arg eq eax
-      mov eax, arg
-    end if
-    call debug_outhex
-    popad
-    popf
-}
-dph fix debug_print_hex
-
-debug_outhex:
-    ;  eax - number
-    mov   edx, 8
-  .new_char:
-    rol   eax, 4
-    movzx ecx, al
-    and   cl,  0x0f
-    mov   cl,  [__hexdigits + ecx]
-    pushad
-    mcall 63, 1
-    popad
-    dec   edx
-    jnz   .new_char
-ret
-
-__hexdigits:
-  db '0123456789ABCDEF'