From 73aa6c6d6abb557e31e6214ff9b7bb4bfbd04d38 Mon Sep 17 00:00:00 2001
From: "Sergey Semyonov (Serge)" <ion2@mail.ru>
Date: Wed, 24 Jan 2007 15:51:55 +0000
Subject: [PATCH] sse2 mixers

git-svn-id: svn://kolibrios.org@293 a494cfbc-eb01-0410-851d-a64ba20cac60
---
 kernel/trunk/drivers/infinity.asm |  72 +++--
 kernel/trunk/drivers/mix_mmx.inc  | 241 ++++++++++++++++
 kernel/trunk/drivers/mix_sse2.inc | 139 +++++++++
 kernel/trunk/drivers/mixer.asm    | 457 ++++++------------------------
 4 files changed, 524 insertions(+), 385 deletions(-)
 create mode 100644 kernel/trunk/drivers/mix_mmx.inc
 create mode 100644 kernel/trunk/drivers/mix_sse2.inc

diff --git a/kernel/trunk/drivers/infinity.asm b/kernel/trunk/drivers/infinity.asm
index 7b82a07c2d..fd24c81599 100644
--- a/kernel/trunk/drivers/infinity.asm
+++ b/kernel/trunk/drivers/infinity.asm
@@ -19,17 +19,21 @@ include 'proc32.inc'
 include 'main.inc'
 include 'imports.inc'
 
-USE_MMX      equ 0
-USE_MMX_128  equ 0
-USE_SSE      equ 0
+FORCE_MMX        equ 0  ;set to 1 to force use mmx or
+FORCE_MMX_128    equ 0  ;integer sse2 extensions
+                        ;and reduce driver size
+;USE_SSE         equ 0
 
-DEBUG               equ 1
+DEBUG            equ 1
 
-EVENT_NOTIFY        equ 0x00000200
+EVENT_NOTIFY     equ 0x00000200
+
+OS_BASE          equ 0
+new_app_base     equ 0x60400000
+PROC_BASE        equ OS_BASE+0x0080000
+
+CAPS_SSE2        equ 26
 
-OS_BASE               equ 0;  0x80400000
-new_app_base          equ 0x60400000;   0x01000000
-PROC_BASE             equ OS_BASE+0x0080000
 
 public START
 public service_proc
@@ -79,6 +83,42 @@ proc START stdcall, state:dword
            mov [str.fd], eax
            mov [str.bk], eax
 
+if FORCE_MMX
+ if FORCE_MMX_128
+  display 'Use only FORCE_MMX or FORCE_MMX_128 not both together',13,10
+  stop
+ end if
+           mov [mix_2_core], mmx_mix_2
+           mov [mix_3_core], mmx_mix_3
+           mov [mix_4_core], mmx_mix_4
+end if
+
+if FORCE_MMX_128
+ if FORCE_MMX
+  display 'Use only FORCE_MMX or FORCE_MMX_128 not both together',13,10
+  stop
+ end if
+           mov [mix_2_core], mmx128_mix_2
+           mov [mix_3_core], mmx128_mix_3
+           mov [mix_4_core], mmx128_mix_4
+end if
+
+if ~(FORCE_MMX or FORCE_MMX_128)  ;autodetect
+           mov eax, 1
+           cpuid
+           bt edx, CAPS_SSE2
+           jc .mmx128
+                                           ;old 64-bit mmx
+           mov [mix_2_core], mmx_mix_2
+           mov [mix_3_core], mmx_mix_3
+           mov [mix_4_core], mmx_mix_4
+           jmp @F
+.mmx128:                                   ;new 128-bit sse2 extensions
+           mov [mix_2_core], mmx128_mix_2
+           mov [mix_3_core], mmx128_mix_3
+           mov [mix_4_core], mmx128_mix_4
+@@:
+end if
            stdcall set_handler, [hSound], new_mix
            stdcall RegService, szInfinity, service_proc
            ret
@@ -563,14 +603,8 @@ proc dev_play stdcall, hsrv:dword
 endp
 
 include 'mixer.asm'
-
-;if USE_MMX
-; include 'mix_mmx.inc'
-;end if
-
-if USE_MMX_128
- include 'mix_sse2.inc'
-end if
+include 'mix_mmx.inc'
+include 'mix_sse2.inc'
 
 ;if USE_SSE
 ; include 'mix_sse.inc'
@@ -664,7 +698,7 @@ mix_buff_map  rd 1
 str.fd        rd 1
 str.bk        rd 1
 
-mix_2_1.core  rd 1
-mix_3_1.core  rd 1
-mix_4_1.core  rd 1
+mix_2_core    rd 1
+mix_3_core    rd 1
+mix_4_core    rd 1
 
diff --git a/kernel/trunk/drivers/mix_mmx.inc b/kernel/trunk/drivers/mix_mmx.inc
new file mode 100644
index 0000000000..9413c142d1
--- /dev/null
+++ b/kernel/trunk/drivers/mix_mmx.inc
@@ -0,0 +1,241 @@
+
+; params
+;  edi= output
+;  eax= input stream 1
+;  ebx= input stream 2
+
+if used mmx_mix_2
+
+align 4
+mmx_mix_2:
+           movq mm0, [eax]
+           movq mm1, [eax+8]
+           movq mm2, [eax+16]
+           movq mm3, [eax+24]
+           movq mm4, [eax+32]
+           movq mm5, [eax+40]
+           movq mm6, [eax+48]
+           movq mm7, [eax+56]
+
+           paddsw mm0, [ebx]
+           movq [edi], mm0
+           paddsw mm1,[ebx+8]
+           movq [edi+8], mm1
+           paddsw mm2, [ebx+16]
+           movq [edi+16], mm2
+           paddsw mm3, [ebx+24]
+           movq [edi+24], mm3
+           paddsw mm4, [ebx+32]
+           movq [edi+32], mm4
+           paddsw mm5, [ebx+40]
+           movq [edi+40], mm5
+           paddsw mm6, [ebx+48]
+           movq [edi+48], mm6
+           paddsw mm7, [ebx+56]
+           movq [edi+56], mm7
+
+           movq mm0, [eax+64]
+           movq mm1, [eax+72]
+           movq mm2, [eax+80]
+           movq mm3, [eax+88]
+           movq mm4, [eax+96]
+           movq mm5, [eax+104]
+           movq mm6, [eax+112]
+           movq mm7, [eax+120]
+
+           paddsw mm0, [ebx+64]
+           movq [edi+64], mm0
+           paddsw mm1, [ebx+72]
+           movq [edi+72], mm1
+           paddsw mm2, [ebx+80]
+           movq [edi+80], mm2
+           paddsw mm3, [ebx+88]
+           movq [edi+88], mm3
+           paddsw mm4, [ebx+96]
+           movq [edi+96], mm4
+           paddsw mm5, [ecx+104]
+           movq [edx+104], mm5
+           paddsw mm6, [ebx+112]
+           movq [edi+112], mm6
+           paddsw mm7, [ebx+120]
+           movq [edi+120], mm7
+           ret
+
+align 4
+mmx_mix_3:
+           movq mm0, [eax]
+           movq mm1, [eax+8]
+           movq mm2, [eax+16]
+           movq mm3, [eax+24]
+           movq mm4, [eax+32]
+           movq mm5, [eax+40]
+           movq mm6, [eax+48]
+           movq mm7, [eax+56]
+
+           paddsw mm0, [ebx]
+           paddsw mm1, [ebx+8]
+           paddsw mm2, [ebx+16]
+           paddsw mm3, [ebx+24]
+           paddsw mm4, [ebx+32]
+           paddsw mm5, [ebx+40]
+           paddsw mm6, [ebx+48]
+           paddsw mm7, [ebx+56]
+           paddsw mm0, [ecx]
+           movq [edi], mm0
+           paddsw mm1,[ecx+8]
+           movq [edi+8], mm1
+           paddsw mm2, [ecx+16]
+           movq [edi+16], mm2
+           paddsw mm3, [ecx+24]
+           movq [edi+24], mm3
+           paddsw mm4, [ecx+32]
+           movq [edi+32], mm4
+           paddsw mm5, [ecx+40]
+           movq [edi+40], mm5
+           paddsw mm6, [ecx+48]
+           movq [edi+48], mm6
+           paddsw mm7, [ecx+56]
+           movq [edi+56], mm7
+
+           movq mm0, [eax+64]
+           movq mm1, [eax+72]
+           movq mm2, [eax+80]
+           movq mm3, [eax+88]
+           movq mm4, [eax+96]
+           movq mm5, [eax+104]
+           movq mm6, [eax+112]
+           movq mm7, [eax+120]
+           paddsw mm0, [ebx+64]
+           paddsw mm1, [ebx+72]
+           paddsw mm2, [ebx+80]
+           paddsw mm3, [ebx+88]
+           paddsw mm4, [ebx+96]
+           paddsw mm5, [ebx+104]
+           paddsw mm6, [ebx+112]
+           paddsw mm7, [ebx+120]
+           paddsw mm0, [ecx+64]
+           movq [edi+64], mm0
+           paddsw mm1, [ecx+72]
+           movq [edi+72], mm1
+           paddsw mm2, [ecx+80]
+           movq [edi+80], mm2
+           paddsw mm3, [ecx+88]
+           movq [edi+88], mm3
+           paddsw mm4, [ecx+96]
+           movq [edi+96], mm4
+           paddsw mm5, [ecx+104]
+           movq [edi+104], mm5
+           paddsw mm6, [ecx+112]
+           movq [edi+112], mm6
+           paddsw mm7, [ecx+120]
+           movq [edi+120], mm7
+           ret
+
+align 4
+mmx_mix_4:
+
+           movq mm0, [eax]
+           movq mm2, [eax+8]
+           movq mm4, [eax+16]
+           movq mm6, [eax+24]
+           movq mm1, [ebx]
+           movq mm3, [ebx+8]
+           movq mm5, [ebx+16]
+           movq mm7, [ebx+24]
+           paddsw mm0, [ecx]
+           paddsw mm2, [ecx+8]
+           paddsw mm4, [ecx+16]
+           paddsw mm6, [ecx+24]
+           paddsw mm1, [edx]
+           paddsw mm3, [edx+8]
+           paddsw mm5, [edx+16]
+           paddsw mm7, [edx+24]
+
+           paddsw mm0, mm1
+           movq [edi], mm0
+           paddsw mm2, mm3
+           movq [edi+8], mm2
+           paddsw mm4, mm5
+           movq [edi+16], mm4
+           paddsw mm5, mm6
+           movq [edi+24], mm6
+
+           movq mm0, [eax+32]
+           movq mm2, [eax+40]
+           movq mm4, [eax+48]
+           movq mm6, [eax+56]
+           movq mm1, [ebx+32]
+           movq mm3, [ebx+40]
+           movq mm5, [ebx+48]
+           movq mm7, [ebx+56]
+           paddsw mm0, [ecx+32]
+           paddsw mm2, [ecx+40]
+           paddsw mm4, [ecx+48]
+           paddsw mm6, [ecx+56]
+           paddsw mm1, [edx+32]
+           paddsw mm3, [edx+40]
+           paddsw mm5, [edx+48]
+           paddsw mm7, [edx+56]
+
+           paddsw mm0, mm1
+           movq [edi+32], mm0
+           paddsw mm2, mm2
+           movq [edi+40], mm2
+           paddsw mm4, mm5
+           movq [edi+48], mm4
+           paddsw mm6, mm7
+           movq [edi+56], mm6
+
+           movq mm0, [eax+64]
+           movq mm2, [eax+72]
+           movq mm4, [eax+80]
+           movq mm6, [eax+88]
+           movq mm1, [ebx+64]
+           movq mm3, [ebx+72]
+           movq mm5, [ebx+80]
+           movq mm7, [ebx+88]
+           paddsw mm0, [ecx+64]
+           paddsw mm2, [ecx+72]
+           paddsw mm4, [ecx+80]
+           paddsw mm6, [ecx+88]
+           paddsw mm1, [edx+64]
+           paddsw mm3, [edx+72]
+           paddsw mm5, [edx+80]
+           paddsw mm7, [edx+88]
+
+           paddsw mm0, mm1
+           movq [edi+64], mm0
+           paddsw mm2, mm3
+           movq [edi+72], mm2
+           paddsw mm4, mm5
+           movq [edi+80], mm4
+           paddsw mm6, mm5
+           movq [edi+88], mm7
+
+           movq mm0, [eax+96]
+           movq mm2, [eax+104]
+           movq mm4, [eax+112]
+           movq mm6, [eax+120]
+           movq mm1, [ebx+96]
+           movq mm3, [ebx+104]
+           movq mm5, [ebx+112]
+           movq mm7, [ebx+120]
+           paddsw mm0, [ecx+96]
+           paddsw mm2, [ecx+104]
+           paddsw mm4, [ecx+112]
+           paddsw mm6, [ecx+120]
+           paddsw mm1, [edx+96]
+           paddsw mm3, [edx+104]
+           paddsw mm5, [edx+112]
+           paddsw mm7, [edx+120]
+           paddsw mm0, mm1
+           movq [eax+96], mm0
+           paddsw mm2, mm3
+           movq [edi+104], mm2
+           paddsw mm4, mm5
+           movq [edi+112], mm4
+           paddsw mm6, mm7
+           movq [edi+120], mm6
+           ret
+
+end if
diff --git a/kernel/trunk/drivers/mix_sse2.inc b/kernel/trunk/drivers/mix_sse2.inc
new file mode 100644
index 0000000000..4ca0a9e63d
--- /dev/null
+++ b/kernel/trunk/drivers/mix_sse2.inc
@@ -0,0 +1,139 @@
+
+if used mmx128_mix_2
+
+align 4
+mmx128_mix_2:
+           prefetcht1 [eax+128]
+           prefetcht1 [ebx+128]
+
+           movaps xmm0, [eax]
+           movaps xmm1, [eax+16]
+           movaps xmm2, [eax+32]
+           movaps xmm3, [eax+48]
+           movaps xmm4, [eax+64]
+           movaps xmm5, [eax+80]
+           movaps xmm6, [eax+96]
+           movaps xmm7, [eax+112]
+
+           paddsw xmm0, [ebx]
+           movaps [edi], xmm0
+           paddsw xmm1,[ebx+16]
+           movaps [edi+16], xmm1
+           paddsw xmm2, [ebx+32]
+           movaps [edi+32], xmm2
+           paddsw xmm3, [ebx+48]
+           movaps [edi+48], xmm3
+           paddsw xmm4, [ebx+64]
+           movaps [edi+64], xmm4
+           paddsw xmm5, [ebx+80]
+           movaps [edi+80], xmm5
+           paddsw xmm6, [ebx+96]
+           movaps [edi+96], xmm6
+           paddsw xmm7, [ebx+112]
+           movaps [edi+112], xmm7
+           ret
+
+align 4
+mmx128_mix_3:
+           prefetcht1 [eax+128]
+           prefetcht1 [ebx+128]
+           prefetcht1 [ecx+128]
+
+           movaps xmm0, [eax]
+           movaps xmm1, [eax+16]
+           movaps xmm2, [eax+32]
+           movaps xmm3, [eax+48]
+           movaps xmm4, [eax+64]
+           movaps xmm5, [eax+80]
+           movaps xmm6, [eax+96]
+           movaps xmm7, [eax+112]
+
+           paddsw xmm0, [ebx]
+           paddsw xmm1, [ebx+16]
+           paddsw xmm2, [ebx+32]
+           paddsw xmm3, [ebx+48]
+           paddsw xmm4, [ebx+64]
+           paddsw xmm5, [ebx+80]
+           paddsw xmm6, [ebx+96]
+           paddsw xmm7, [ebx+112]
+
+           paddsw xmm0, [ecx]
+           movaps [edi], xmm0
+           paddsw xmm1, [ecx+16]
+           movaps [edi+16], xmm1
+           paddsw xmm2, [ecx+32]
+           movaps [edi+32], xmm2
+           paddsw xmm3, [ecx+48]
+           movaps [edi+48], xmm3
+           paddsw xmm4, [ecx+64]
+           movaps [edi+64], xmm4
+           paddsw xmm5, [ecx+80]
+           movaps [edi+80], xmm5
+           paddsw xmm6, [ecx+96]
+           movaps [edi+96], xmm6
+           paddsw xmm7, [ecx+112]
+           movaps [edi+112], xmm7
+           ret
+
+align 4
+mmx128_mix_4:
+           prefetcht1 [eax+128]
+           prefetcht1 [ebx+128]
+           prefetcht1 [ecx+128]
+           prefetcht1 [edx+128]
+
+           movaps xmm0, [eax]
+           movaps xmm2, [eax+16]
+           movaps xmm4, [eax+32]
+           movaps xmm6, [eax+48]
+           movaps xmm1, [ebx]
+           movaps xmm3, [ebx+16]
+           movaps xmm5, [ebx+32]
+           movaps xmm7, [ebx+48]
+
+           paddsw xmm0, [ecx]
+           paddsw xmm2, [ecx+16]
+           paddsw xmm4, [ecx+32]
+           paddsw xmm6, [ecx+48]
+           paddsw xmm1, [edx]
+           paddsw xmm3, [edx+16]
+           paddsw xmm5, [edx+32]
+           paddsw xmm7, [edx+48]
+
+           paddsw xmm0, xmm1
+           movaps [edi], xmm0
+           paddsw xmm2, xmm3
+           movaps [edi+16], xmm2
+           paddsw xmm4, xmm5
+           movaps [edi+32], xmm4
+           paddsw xmm6, xmm7
+           movaps [edi+48], xmm6
+
+           movaps xmm0, [eax+64]
+           movaps xmm2, [eax+80]
+           movaps xmm4, [eax+96]
+           movaps xmm6, [eax+112]
+
+           movaps xmm1, [ebx+64]
+           movaps xmm3, [ebx+80]
+           movaps xmm5, [ebx+96]
+           movaps xmm7, [ebx+112]
+           paddsw xmm0, [ecx+64]
+           paddsw xmm2, [ecx+80]
+           paddsw xmm4, [ecx+96]
+           paddsw xmm6, [ecx+112]
+
+           paddsw xmm1, [edx+64]
+           paddsw xmm3, [edx+80]
+           paddsw xmm5, [edx+96]
+           paddsw xmm7, [edx+112]
+           paddsw xmm0, xmm1
+           movaps [edi+64], xmm0
+           paddsw xmm2, xmm3
+           movaps [edi+80], xmm2
+           paddsw xmm4, xmm5
+           movaps [edi+96], xmm4
+           paddsw xmm6, xmm7
+           movaps [edi+112], xmm6
+           ret
+end if
diff --git a/kernel/trunk/drivers/mixer.asm b/kernel/trunk/drivers/mixer.asm
index f7c1226179..cc70e8908f 100644
--- a/kernel/trunk/drivers/mixer.asm
+++ b/kernel/trunk/drivers/mixer.asm
@@ -90,7 +90,7 @@ proc new_mix stdcall, output:dword
 .m3:
            add [output],512
 
-           sub [main_count], 1
+           dec [main_count]
            jnz .l00
 
            call update_stream
@@ -324,8 +324,8 @@ align 16
            mov [edi], ebx
            add edi, 4
 
-    add eax, [esp+16]
-    cmp eax, [esp+24]
+           add eax, [esp+16]
+           cmp eax, [esp+24]
            jb .l1
 
            mov ebp, esp
@@ -382,8 +382,8 @@ align 16
            mov [edi], ebx
            add edi, 4
 
-    add esi, [esp+16]
-    cmp esi, [esp+24]
+           add esi, [esp+16]
+           cmp esi, [esp+24]
            jb .l1
 
            mov ebp, esp
@@ -622,6 +622,7 @@ proc alloc_mix_buff
            ret
 endp
 
+align 4
 proc m16_s_mmx
 
            movq    mm0, [esi]
@@ -777,56 +778,59 @@ proc m8_s_mmx
            ret
 endp
 
-
 align 4
 proc mix_2_1 stdcall, output:dword, str0:dword, str1:dword
 
            mov edi, [output]
+           mov eax, [str0]
+           mov ebx, [str1]
+           mov esi, 128
+           call [mix_2_core]   ;edi, eax, ebx
 
-           stdcall mix_2_1_mmx, edi, [str0],[str1]
-;           stdcall mix_2_1_sse, edi, [str0],[str1]
-           add edi, 128
-           add [str0], 128
-           add [str1], 128
-           stdcall mix_2_1_mmx, edi, [str0],[str1]
-;           stdcall mix_2_1_sse, edi, [str0],[str1]
-           add edi, 128
-           add [str0], 128
-           add [str1], 128
-           stdcall mix_2_1_mmx, edi, [str0],[str1]
-;           stdcall mix_2_1_sse, edi, [str0],[str1]
-           add edi, 128
-           add [str0], 128
-           add [str1], 128
-           stdcall mix_2_1_mmx, edi, [str0],[str1]
-;           stdcall mix_2_1_sse, edi, [str0],[str1]
+           add edi, esi
+           add eax, esi
+           add ebx, esi
+           call [mix_2_core]   ;edi, eax, ebx
 
+           add edi, esi
+           add eax, esi
+           add ebx, esi
+           call [mix_2_core]   ;edi, eax, ebx
+
+           add edi, esi
+           add eax, esi
+           add ebx, esi
+           call [mix_2_core]   ;edi, eax, ebx
            ret
 endp
 
-
 align 4
 proc mix_3_1 stdcall, output:dword, str0:dword, str1:dword, str2:dword
 
            mov edi, [output]
+           mov eax, [str0]
+           mov ebx, [str1]
+           mov ecx, [str2]
+           mov esi, 128
+           call [mix_3_core]
 
-           stdcall mix_3_1_mmx, edi, [str0],[str1],[str2]
-           add edi, 128
-           add [str0], 128
-           add [str1], 128
-           add [str2], 128
-           stdcall mix_3_1_mmx, edi, [str0],[str1],[str2]
-           add edi, 128
-           add [str0], 128
-           add [str1], 128
-           add [str2], 128
-           stdcall mix_3_1_mmx, edi, [str0],[str1],[str2]
-           add edi, 128
-           add [str0], 128
-           add [str1], 128
-           add [str2], 128
-           stdcall mix_3_1_mmx, edi, [str0],[str1],[str2]
+           add edi, esi
+           add eax, esi
+           add ebx, esi
+           add ecx, esi
+           call [mix_3_core]
 
+           add edi, esi
+           add eax, esi
+           add ebx, esi
+           add ecx, esi
+           call [mix_3_core]
+
+           add edi, esi
+           add eax, esi
+           add ebx, esi
+           add ecx, esi
+           call [mix_3_core]
            ret
 endp
 
@@ -839,29 +843,35 @@ proc mix_4_1 stdcall, str0:dword, str1:dword,\
            call alloc_mix_buff
            and eax, eax
            jz .err
-           mov [output], eax
 
            mov edi, eax
+           mov eax, [str0]
+           mov ebx, [str1]
+           mov ecx, [str2]
+           mov edx, [str3]
+           mov esi, 128
+           call [mix_4_core]  ;edi, eax, ebx, ecx, edx
 
-           stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3]
-           add edi, 128
-           add [str0], 128
-           add [str1], 128
-           add [str2], 128
-           add [str3], 128
-           stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3]
-           add edi, 128
-           add [str0], 128
-           add [str1], 128
-           add [str2], 128
-           add [str3], 128
-           stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3]
-           add edi, 128
-           add [str0], 128
-           add [str1], 128
-           add [str2], 128
-           add [str3], 128
-           stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3]
+           add edi, esi
+           add eax, esi
+           add ebx, esi
+           add ecx, esi
+           add edx, esi
+           call [mix_4_core]  ;edi, eax, ebx, ecx, edx
+
+           add edi, esi
+           add eax, esi
+           add ebx, esi
+           add ecx, esi
+           add edx, esi
+           call [mix_4_core]  ;edi, eax, ebx, ecx, edx
+
+           add edi, esi
+           add eax, esi
+           add ebx, esi
+           add ecx, esi
+           add edx, esi
+           call [mix_4_core]  ;edi, eax, ebx, ecx, edx
            mov eax, [output]
            ret
 .err:
@@ -876,318 +886,33 @@ proc final_mix stdcall, output:dword, str0:dword, str1:dword,\
 
            mov edi, [output]
 
-           stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3]
-           add edi, 128
-           add [str0], 128
-           add [str1], 128
-           add [str2], 128
-           add [str3], 128
-           stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3]
-           add edi, 128
-           add [str0], 128
-           add [str1], 128
-           add [str2], 128
-           add [str3], 128
-           stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3]
-           add edi, 128
-           add [str0], 128
-           add [str1], 128
-           add [str2], 128
-           add [str3], 128
-           stdcall mix_4_1_mmx, edi, [str0],[str1],[str2],[str3]
-
-           ret
-endp
-
-align 4
-proc mix_2_1_mmx stdcall, output:dword, str0:dword, str1:dword
-
-           mov edx, [output]
-           mov eax, [str0]
-           mov ecx, [str1]
-
-           movq mm0, [eax]
-           paddsw mm0, [ecx]
-           movq [edx], mm0
-
-           movq mm1, [eax+8]
-           paddsw mm1,[ecx+8]
-           movq [edx+8], mm1
-
-           movq mm2, [eax+16]
-           paddsw mm2, [ecx+16]
-           movq [edx+16], mm2
-
-           movq mm3, [eax+24]
-           paddsw mm3, [ecx+24]
-           movq [edx+24], mm3
-
-           movq mm0, [eax+32]
-           paddsw mm0, [ecx+32]
-           movq [edx+32], mm0
-
-           movq mm1, [eax+40]
-           paddsw mm1, [ecx+40]
-           movq [edx+40], mm1
-
-           movq mm2, [eax+48]
-           paddsw mm2, [ecx+48]
-           movq [edx+48], mm2
-
-           movq mm3, [eax+56]
-           paddsw mm3, [ecx+56]
-           movq [edx+56], mm3
-
-           movq mm0, [eax+64]
-           paddsw mm0, [ecx+64]
-           movq [edx+64], mm0
-
-           movq mm1, [eax+72]
-           paddsw mm1, [ecx+72]
-           movq [edx+72], mm1
-
-           movq mm2, [eax+80]
-           paddsw mm2, [ecx+80]
-           movq [edx+80], mm2
-
-           movq mm3, [eax+88]
-           paddsw mm3, [ecx+88]
-           movq [edx+88], mm3
-
-           movq mm0, [eax+96]
-           paddsw mm0, [ecx+96]
-           movq [edx+96], mm0
-
-           movq mm1, [eax+104]
-           paddsw mm1, [ecx+104]
-           movq [edx+104], mm1
-
-           movq mm2, [eax+112]
-           paddsw mm2, [ecx+112]
-           movq [edx+112], mm2
-
-           movq mm3, [eax+120]
-           paddsw mm3, [ecx+120]
-           movq [edx+120], mm3
-
-           ret
-endp
-
-
-
-align 4
-proc mix_3_1_mmx stdcall, output:dword, str0:dword, str1:dword, str2:dword
-
-           mov edx, [output]
            mov eax, [str0]
            mov ebx, [str1]
            mov ecx, [str2]
+           mov edx, [str3]
+           mov esi, 128
+           call [mix_4_core]  ;edi, eax, ebx, ecx, edx
 
-           movq mm0, [eax]
-           paddsw mm0, [ebx]
-           paddsw mm0, [ecx]
-           movq [edx], mm0
+           add edi, esi
+           add eax, esi
+           add ebx, esi
+           add ecx, esi
+           add edx, esi
+           call [mix_4_core]  ;edi, eax, ebx, ecx, edx
 
-           movq mm1, [eax+8]
-           paddsw mm1,[ebx+8]
-           paddsw mm1,[ecx+8]
-           movq [edx+8], mm1
-
-           movq mm2, [eax+16]
-           paddsw mm2, [ebx+16]
-           paddsw mm2, [ecx+16]
-           movq [edx+16], mm2
-
-           movq mm3, [eax+24]
-           paddsw mm3, [ebx+24]
-           paddsw mm3, [ecx+24]
-           movq [edx+24], mm3
-
-           movq mm0, [eax+32]
-           paddsw mm0, [ebx+32]
-           paddsw mm0, [ecx+32]
-           movq [edx+32], mm0
-
-           movq mm1, [eax+40]
-           paddsw mm1, [ebx+40]
-           paddsw mm1, [ecx+40]
-           movq [edx+40], mm1
-
-           movq mm2, [eax+48]
-           paddsw mm2, [ebx+48]
-           paddsw mm2, [ecx+48]
-           movq [edx+48], mm2
-
-           movq mm3, [eax+56]
-           paddsw mm3, [ebx+56]
-           paddsw mm3, [ecx+56]
-           movq [edx+56], mm3
-
-           movq mm0, [eax+64]
-           paddsw mm0, [ebx+64]
-           paddsw mm0, [ecx+64]
-           movq [edx+64], mm0
-
-           movq mm1, [eax+72]
-           paddsw mm1, [ebx+72]
-           paddsw mm1, [ecx+72]
-           movq [edx+72], mm1
-
-           movq mm2, [eax+80]
-           paddsw mm2, [ebx+80]
-           paddsw mm2, [ecx+80]
-           movq [edx+80], mm2
-
-           movq mm3, [eax+88]
-           paddsw mm3, [ebx+88]
-           paddsw mm3, [ecx+88]
-           movq [edx+88], mm3
-
-           movq mm0, [eax+96]
-           paddsw mm0, [ebx+96]
-           paddsw mm0, [ecx+96]
-           movq [edx+96], mm0
-
-           movq mm1, [eax+104]
-           paddsw mm1, [ebx+104]
-           paddsw mm1, [ecx+104]
-           movq [edx+104], mm1
-
-           movq mm2, [eax+112]
-           paddsw mm2, [ebx+112]
-           paddsw mm2, [ecx+112]
-           movq [edx+112], mm2
-
-           movq mm3, [eax+120]
-           paddsw mm3, [ebx+120]
-           paddsw mm3, [ecx+120]
-           movq [edx+120], mm3
-
-           ret
-endp
-
-align 4
-proc mix_4_1_mmx stdcall, output:dword, str0:dword, str1:dword,\
-                          str2:dword, str3:dword
-
-           mov edx, [output]
-           mov esi, [str0]
-           mov eax, [str1]
-           mov ebx, [str2]
-           mov ecx, [str3]
-
-           movq mm0, [esi]
-           movq mm1, [eax]
-           paddsw mm0, [ebx]
-           paddsw mm1, [ecx]
-           paddsw mm0, mm1
-           movq [edx], mm0
-
-           movq mm2, [esi+8]
-           movq mm3, [eax+8]
-           paddsw mm2, [ebx+8]
-           paddsw mm3, [ecx+8]
-           paddsw mm2, mm3
-           movq [edx+8], mm2
-
-           movq mm0, [esi+16]
-           movq mm1, [eax+16]
-           paddsw mm0, [ebx+16]
-           paddsw mm1, [ecx+16]
-           paddsw mm0, mm1
-           movq [edx+16], mm0
-
-           movq mm2, [esi+24]
-           movq mm3, [eax+24]
-           paddsw mm2, [ebx+24]
-           paddsw mm3, [ecx+24]
-           paddsw mm2, mm3
-           movq [edx+24], mm2
-
-           movq mm0, [esi+32]
-           movq mm1, [eax+32]
-           paddsw mm0, [ebx+32]
-           paddsw mm1, [ecx+32]
-           paddsw mm0, mm1
-           movq [edx+32], mm0
-
-           movq mm2, [esi+40]
-           movq mm3, [eax+40]
-           paddsw mm2, [ebx+40]
-           paddsw mm3, [ecx+40]
-           paddsw mm2, mm3
-           movq [edx+40], mm2
-
-           movq mm0, [esi+48]
-           movq mm1, [eax+48]
-           paddsw mm0, [ebx+48]
-           paddsw mm1, [ecx+48]
-           paddsw mm0, mm1
-           movq [edx+48], mm0
-
-           movq mm2, [esi+56]
-           movq mm3, [eax+56]
-           paddsw mm2, [ebx+56]
-           paddsw mm3, [ecx+56]
-           paddsw mm2, mm3
-           movq [edx+56], mm2
-
-           movq mm0, [esi+64]
-           movq mm1, [eax+64]
-           paddsw mm0, [ebx+64]
-           paddsw mm1, [ecx+64]
-           paddsw mm0, mm1
-           movq [edx+64], mm0
-
-           movq mm2, [esi+72]
-           movq mm3, [eax+72]
-           paddsw mm2, [ebx+72]
-           paddsw mm3, [ecx+72]
-           paddsw mm2, mm3
-           movq [edx+72], mm2
-
-           movq mm2, [esi+80]
-           movq mm3, [eax+80]
-           paddsw mm2, [ebx+80]
-           paddsw mm3, [ecx+80]
-           paddsw mm2, mm3
-           movq [edx+80], mm2
-
-           movq mm2, [esi+88]
-           movq mm3, [eax+88]
-           paddsw mm2, [ebx+88]
-           paddsw mm3, [ecx+88]
-           paddsw mm2, mm3
-           movq [edx+88], mm2
-
-           movq mm2, [esi+96]
-           movq mm3, [eax+96]
-           paddsw mm2, [ebx+96]
-           paddsw mm3, [ecx+96]
-           paddsw mm2, mm3
-           movq [edx+96], mm2
-
-           movq mm2, [esi+104]
-           movq mm3, [eax+104]
-           paddsw mm2, [ebx+104]
-           paddsw mm3, [ecx+104]
-           paddsw mm2, mm3
-           movq [edx+104], mm2
-
-           movq mm2, [esi+112]
-           movq mm3, [eax+112]
-           paddsw mm2, [ebx+112]
-           paddsw mm3, [ecx+112]
-           paddsw mm2, mm3
-           movq [edx+112], mm2
-
-           movq mm2, [esi+120]
-           movq mm3, [eax+120]
-           paddsw mm2, [ebx+120]
-           paddsw mm3, [ecx+120]
-           paddsw mm2, mm3
-           movq [edx+120], mm2
+           add edi, esi
+           add eax, esi
+           add ebx, esi
+           add ecx, esi
+           add edx, esi
+           call [mix_4_core]  ;edi, eax, ebx, ecx, edx
 
+           add edi, esi
+           add eax, esi
+           add ebx, esi
+           add ecx, esi
+           add edx, esi
+           call [mix_4_core]  ;edi, eax, ebx, ecx, edx
            ret
 endp