MMX

diff --git a/Makefile b/Makefile
index 9ff0e58..8f198e5 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -1,22 +1,27 @@
 !ifdef __UNIX__
 dosobj = src/dos/main.obj src/dos/gfx.obj src/dos/timer.obj src/dos/watdpmi.obj &
-                src/dos/vbe.obj src/dos/vga.obj src/dos/keyb.obj src/dos/mouse.obj &
-                src/dos/logger.obj
-scrobj = src/introscr.obj src/menuscr.obj
-gameobj = src/game.obj src/util.obj src/gfxutil.obj src/dynarr.obj src/rbtree.obj
+       src/dos/vbe.obj src/dos/vga.obj src/dos/keyb.obj src/dos/mouse.obj &
+       src/dos/logger.obj
+scrobj = src/introscr.obj src/intro_s.obj src/menuscr.obj
+gameobj = src/game.obj src/util.obj src/gfxutil.obj src/dynarr.obj &
+       src/rbtree.obj
 gfxobj = src/3dgfx/3dgfx.obj src/3dgfx/mesh.obj src/3dgfx/meshload.obj &
-               src/3dgfx/polyfill.obj src/3dgfx/polyclip.obj src/sprite.obj
+       src/3dgfx/polyfill.obj src/3dgfx/polyclip.obj src/sprite.obj
+
 incpath = -Isrc -Isrc/dos -Ilibs/imago/src
 libpath = libpath libs/imago
+
 !else
-dosobj = src\dos\main.obj src\dos\gfx.obj src\dos\timer.obj &
-       src\dos\watdpmi.obj src\dos\vbe.obj src\dos\vga.obj src\dos\keyb.obj &
-       src\dos\mouse.obj src\dos\logger.obj
-scrobj = src\introscr.obj src\menuscr.obj
+
+dosobj = src\dos\main.obj src\dos\gfx.obj src\dos\timer.obj src\dos\watdpmi.obj &
+       src\dos\vbe.obj src\dos\vga.obj src\dos\keyb.obj src\dos\mouse.obj &
+       src\dos\logger.obj
+scrobj = src\introscr.obj src\intro_s.obj src\menuscr.obj
 gameobj = src\game.obj src\util.obj src\gfxutil.obj src\dynarr.obj &
        src\rbtree.obj
 gfxobj = src\3dgfx\3dgfx.obj src\3dgfx\mesh.obj src\3dgfx\meshload.obj &
        src\3dgfx\polyfill.obj src\3dgfx\polyclip.obj src\sprite.obj
+
 incpath = -Isrc -Isrc\dos -Ilibs\imago\src
 libpath = libpath libs\imago
 !endif

diff --git a/libs/imago/src/conv.c b/libs/imago/src/conv.c
index 51ecd1f..501b3c4 100644 (file)
--- a/libs/imago/src/conv.c
+++ b/libs/imago/src/conv.c
@@ -32,6 +32,7 @@ struct pixel {
 static void unpack_grey8(struct pixel *unp, void *pptr, int count);
 static void unpack_rgb24(struct pixel *unp, void *pptr, int count);
 static void unpack_rgba32(struct pixel *unp, void *pptr, int count);
+static void unpack_bgra32(struct pixel *unp, void *pptr, int count);
 static void unpack_greyf(struct pixel *unp, void *pptr, int count);
 static void unpack_rgbf(struct pixel *unp, void *pptr, int count);
 static void unpack_rgbaf(struct pixel *unp, void *pptr, int count);
@@ -40,6 +41,7 @@ static void unpack_rgb565(struct pixel *unp, void *pptr, int count);
 static void pack_grey8(void *pptr, struct pixel *unp, int count);
 static void pack_rgb24(void *pptr, struct pixel *unp, int count);
 static void pack_rgba32(void *pptr, struct pixel *unp, int count);
+static void pack_bgra32(void *pptr, struct pixel *unp, int count);
 static void pack_greyf(void *pptr, struct pixel *unp, int count);
 static void pack_rgbf(void *pptr, struct pixel *unp, int count);
 static void pack_rgbaf(void *pptr, struct pixel *unp, int count);
@@ -50,6 +52,7 @@ static void (*unpack[])(struct pixel*, void*, int) = {
        unpack_grey8,
        unpack_rgb24,
        unpack_rgba32,
+       unpack_bgra32,
        unpack_greyf,
        unpack_rgbf,
        unpack_rgbaf,
@@ -61,6 +64,7 @@ static void (*pack[])(void*, struct pixel*, int) = {
        pack_grey8,
        pack_rgb24,
        pack_rgba32,
+       pack_bgra32,
        pack_greyf,
        pack_rgbf,
        pack_rgbaf,
@@ -145,6 +149,20 @@ static void unpack_rgba32(struct pixel *unp, void *pptr, int count)
        }
 }
 
+static void unpack_bgra32(struct pixel *unp, void *pptr, int count)
+{
+       int i;
+       unsigned char *pix = pptr;
+
+       for(i=0; i<count; i++) {
+               unp->b = (float)*pix++ / 255.0;
+               unp->g = (float)*pix++ / 255.0;
+               unp->r = (float)*pix++ / 255.0;
+               unp->a = (float)*pix++ / 255.0;
+               unp++;
+       }
+}
+
 static void unpack_greyf(struct pixel *unp, void *pptr, int count)
 {
        int i;
@@ -256,6 +274,25 @@ static void pack_rgba32(void *pptr, struct pixel *unp, int count)
        }
 }
 
+static void pack_bgra32(void *pptr, struct pixel *unp, int count)
+{
+       int i;
+       unsigned char *pix = pptr;
+
+       for(i=0; i<count; i++) {
+               int r = (int)(unp->r * 255.0);
+               int g = (int)(unp->g * 255.0);
+               int b = (int)(unp->b * 255.0);
+               int a = (int)(unp->a * 255.0);
+
+               *pix++ = CLAMP(b, 0, 255);
+               *pix++ = CLAMP(g, 0, 255);
+               *pix++ = CLAMP(r, 0, 255);
+               *pix++ = CLAMP(a, 0, 255);
+               unp++;
+       }
+}
+
 static void pack_greyf(void *pptr, struct pixel *unp, int count)
 {
        int i;

diff --git a/libs/imago/src/imago2.c b/libs/imago/src/imago2.c
index 816f8c0..65c554d 100644 (file)
--- a/libs/imago/src/imago2.c
+++ b/libs/imago/src/imago2.c
@@ -419,6 +419,7 @@ static int pixel_size(enum img_fmt fmt)
        case IMG_FMT_RGB24:
                return 3;
        case IMG_FMT_RGBA32:
+       case IMG_FMT_BGRA32:
                return 4;
        case IMG_FMT_GREYF:
                return sizeof(float);

diff --git a/libs/imago/src/imago2.h b/libs/imago/src/imago2.h
index 2cd1481..8086d28 100644 (file)
--- a/libs/imago/src/imago2.h
+++ b/libs/imago/src/imago2.h
@@ -32,6 +32,7 @@ enum img_fmt {
        IMG_FMT_GREY8,
        IMG_FMT_RGB24,
        IMG_FMT_RGBA32,
+       IMG_FMT_BGRA32,
        IMG_FMT_GREYF,
        IMG_FMT_RGBF,
        IMG_FMT_RGBAF,

diff --git a/src/dos/gfx.c b/src/dos/gfx.c
index 671cd8a..c588742 100644 (file)
--- a/src/dos/gfx.c
+++ b/src/dos/gfx.c
@@ -1,10 +1,12 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include "game.h"
 #include "cdpmi.h"
 #include "gfx.h"
 #include "vbe.h"
 #include "vga.h"
+#include "util.h"
 
 #define SAME_BPP(a, b) \
        ((a) == (b) || ((a) == 16 && (b) == 15) || ((a) == 15 && (b) == 16) || \
@@ -238,7 +240,7 @@ static void blit_frame_lfb(void *pixels, int vsync)
 {
        dbg_fps(pixels);
        if(vsync) wait_vsync();
-       memcpy(vpgaddr[frontidx], pixels, pgsize);
+       memcpy64(vpgaddr[frontidx], pixels, pgsize >> 3);
 }
 
 static void blit_frame_banked(void *pixels, int vsync)
@@ -256,7 +258,7 @@ static void blit_frame_banked(void *pixels, int vsync)
        pending = pgsize;
        while(pending > 0) {
                sz = pending > curmode->bank_size ? curmode->bank_size : pending;
-               memcpy((void*)0xa0000, pptr, sz);
+               memcpy64((void*)0xa0000, pptr, sz >> 3);
                pptr += sz;
                pending -= sz;
                vbe_setwin(0, ++offs);

diff --git a/src/dos/main.c b/src/dos/main.c
index 7d7806b..4a3dc1a 100644 (file)
--- a/src/dos/main.c
+++ b/src/dos/main.c
@@ -10,7 +10,6 @@
 
 static struct video_mode *vmode;
 static int quit;
-static void *vmem;
 
 int main(int argc, char **argv)
 {

diff --git a/src/game.c b/src/game.c
index 86170cb..b1f87f4 100644 (file)
--- a/src/game.c
+++ b/src/game.c
@@ -6,7 +6,7 @@
 
 int fb_width, fb_height;
 long fb_size;
-uint16_t *fb_pixels;
+uint16_t *fb_pixels, *vmem;
 
 long time_msec;
 

diff --git a/src/game.h b/src/game.h
index a986389..a08ea51 100644 (file)
--- a/src/game.h
+++ b/src/game.h
@@ -8,7 +8,7 @@
 extern int fb_width;
 extern int fb_height;
 extern long fb_size;
-extern uint16_t *fb_pixels;
+extern uint16_t *fb_pixels, *vmem;
 
 extern long time_msec;
 

diff --git a/src/intro_s.asm b/src/intro_s.asm
new file mode 100644 (file)
index 0000000..9cb4598
--- /dev/null
+++ b/src/intro_s.asm
@@ -0,0 +1,68 @@
+; vi:ft=nasm:
+       section .text
+       bits 32
+
+       global fade_image
+fade_image:
+       push ebp
+       mov ebp, esp
+       push ebx
+
+       mov eax, [ebp + 8]
+       mov edx, [ebp + 12]
+       mov ebx, [ebp + 16]
+       call fade_image_
+
+       pop ebx
+       pop ebp
+       ret
+
+       ; void fade_image(uint16_t *dest, uint32_t *src, uint16_t fade)
+       ;  - dest points to a 16bit 565 framebuffer
+       ;  - src points to a 32bit RGBX image
+       ;  - fade is 24.8 fixed point [0, 1]
+       ; watcom register calling convention arguments: eax, edx, ebx
+       global fade_image_
+fade_image_:
+       push ecx
+       push edi
+
+       mov ecx, 640 * 480
+       mov edi, eax
+
+       ; take fade and duplicate it across all words of mm1
+       movd mm1, ebx           ; mm1 [00|00|00|VV]
+       punpckldq mm1, mm1      ; mm1 [00|VV|00|VV]
+       packssdw mm1, mm1       ; mm1 [VV|VV|VV|VV]
+.loop:
+       ; grab RGB32 pixel and unpack it to zero-extended words in mm0
+       movd mm0, [edx]         ; mm0 [??|??|?R|GB]
+       add edx, 4
+       pxor mm7, mm7
+       punpcklbw mm0, mm7      ; mm0 [0?|0R|0G|0B]
+       ; multiply by fade and divide by 256 to drop the decimal part
+       pmullw mm0, mm1
+       psrlw mm0, 10   ; 8 for the div + 2 to make them 666, easier 565 packing
+       ; pack result into 565 in ax
+       packuswb mm0, mm0
+       movd eax, mm0
+       mov ebx, eax
+       shr al, 1       ; blue in position [........|00RRRRRR|00GGGGGG|000BBBBB]
+       xor bl, bl
+       shr bx, 3
+       xor ah, ah
+       or ax, bx       ; green in position ...|00RRRRRR|00000GGG|GGGBBBBB]
+       shr ebx, 6
+       and ebx, 0f800h
+       or eax, ebx     ; done [RRRRRGGG|GGGBBBBB]
+       mov [edi], ax
+       add edi, 2
+
+       dec ecx
+       jnz .loop
+
+       emms            ; clear fpu state
+
+       pop edi
+       pop ecx
+       ret

diff --git a/src/introscr.c b/src/introscr.c
index 4d0db19..a836650 100644 (file)
--- a/src/introscr.c
+++ b/src/introscr.c
@@ -5,6 +5,8 @@
 #include "gfxutil.h"
 #include "game.h"
 
+#define USE_MMX
+
 #define FADE_DUR       800
 
 static void *logo;
@@ -13,7 +15,7 @@ static long start_time;
 
 int intro_init(void)
 {
-       if(!(logo = img_load_pixels("data/msglogo.jpg", &logo_width, &logo_height, IMG_FMT_RGB24))) {
+       if(!(logo = img_load_pixels("data/msglogo.jpg", &logo_width, &logo_height, IMG_FMT_BGRA32))) {
                fprintf(stderr, "failed to load logo image\n");
                return -1;
        }
@@ -37,13 +39,12 @@ void intro_stop(void)
 {
 }
 
+void fade_image(void *dest, void *src, uint16_t fade);
+
 void intro_draw(void)
 {
-       int i, j;
        long tm;
        uint16_t fade;
-       unsigned char *src = logo;
-       uint16_t *dest = fb_pixels;
 
        tm = time_msec - start_time;
        if(tm < FADE_DUR) {
@@ -57,16 +58,24 @@ void intro_draw(void)
                //menu_start();
        }
 
-       for(i=0; i<fb_height; i++) {
-               for(j=0; j<fb_width; j++) {
-                       uint16_t r = (uint16_t)*src++ * fade / 256;
-                       uint16_t g = (uint16_t)*src++ * fade / 256;
-                       uint16_t b = (uint16_t)*src++ * fade / 256;
+#ifdef USE_MMX
+       fade_image(fb_pixels, logo, fade);
+#else
+       {
+               int i, j;
+               uint32_t *src = logo;
+               uint16_t *dest = fb_pixels;
+               for(i=0; i<640*480; i++) {
+                       uint32_t pix = *src++;
+                       uint16_t r = (uint16_t)UNPACK_R32(pix) * fade / 256;
+                       uint16_t g = (uint16_t)UNPACK_G32(pix) * fade / 256;
+                       uint16_t b = (uint16_t)UNPACK_B32(pix) * fade / 256;
                        *dest++ = PACK_RGB16(r, g, b);
                }
        }
+#endif
 
-       blit_frame(fb_pixels, 1);
+       blit_frame(fb_pixels, 0);
 }
 
 void intro_keyb(int key, int pressed)

diff --git a/src/sdl/main.c b/src/sdl/main.c
index 57cc7f2..80fab1a 100644 (file)
--- a/src/sdl/main.c
+++ b/src/sdl/main.c
@@ -44,7 +44,7 @@ int main(int argc, char **argv)
                fprintf(stderr, "failed to allocate virtual framebuffer\n");
                return 1;
        }
-       fb_pixels = (uint16_t*)((char*)fb_buf + FB_WIDTH * 2);
+       vmem = fb_pixels = (uint16_t*)((char*)fb_buf + FB_WIDTH * 2);
 
        SDL_Init(SDL_INIT_VIDEO | SDL_INIT_TIMER | SDL_INIT_NOPARACHUTE);
        if(!(fbsurf = SDL_SetVideoMode(xsz, ysz, FB_BPP, sdl_flags))) {

diff --git a/src/util.h b/src/util.h
index 6d869eb..d56bcf5 100644 (file)
--- a/src/util.h
+++ b/src/util.h
@@ -30,6 +30,18 @@ static INLINE int32_t cround64(double val)
 extern uint32_t perf_start_count, perf_interval_count;
 
 #ifdef __WATCOMC__
+void memcpy64(void *dest, void *src, int count);
+#pragma aux memcpy64 = \
+       "cploop:" \
+       "movq mm0, [edx]" \
+       "movq [ebx], mm0" \
+       "add edx, 8" \
+       "add ebx, 8" \
+       "dec ecx" \
+       "jnz cploop" \
+       "emms" \
+       parm[ebx][edx][ecx];
+
 void perf_start(void);
 #pragma aux perf_start = \
        "xor eax, eax" \
@@ -52,6 +64,17 @@ void debug_break(void);
 #endif
 
 #ifdef __GNUC__
+#define memcpy64(dest, src, count) asm volatile ( \
+       "0:\n\t" \
+       "movq (%1), %%mm0\n\t" \
+       "movq %%mm0, (%0)\n\t" \
+       "add $8, %1\n\t" \
+       "add $8, %0\n\t" \
+       "dec %2\n\t" \
+       "jnz 0b\n\t" \
+       "emms\n\t" \
+       :: "r"(dest), "r"(src), "r"(count))
+
 #define perf_start()  asm volatile ( \
        "xor %%eax, %%eax\n" \
        "cpuid\n" \