add sse2/ssse3 simd optimization

11 files changed, 181 insertions, 45 deletions

diff --git a/fmdsp/fmdsp-vramlookup-c.c b/fmdsp/fmdsp-vramlookup-c.c
index f900c8d..3d06f71 100644
--- a/fmdsp/fmdsp-vramlookup-c.c
+++ b/fmdsp/fmdsp-vramlookup-c.c
@@ -1,14 +1,17 @@
 #include "fmdsp/fmdsp.h"
 
 void fmdsp_vramlookup_c(uint8_t *vram32, const uint8_t *vram, const uint8_t *palette, int stride) {
+  uint32_t palette32[FMDSP_PALETTE_COLORS];
+  for (int i = 0; i < FMDSP_PALETTE_COLORS; i++) {
+    uint8_t r = palette[i*3+0];
+    uint8_t g = palette[i*3+1];
+    uint8_t b = palette[i*3+2];
+    palette32[i] = (((uint32_t)r)<<16) | (((uint32_t)g)<<8) | ((uint32_t)b);
+  }
   for (int y = 0; y < PC98_H; y++) {
     for (int x = 0; x < PC98_W; x++) {
-      uint8_t r = palette[vram[y*PC98_W+x]*3+0];
-      uint8_t g = palette[vram[y*PC98_W+x]*3+1];
-      uint8_t b = palette[vram[y*PC98_W+x]*3+2];
-      uint32_t data = (((uint32_t)r)<<16) | (((uint32_t)g)<<8) | ((uint32_t)b);
       uint32_t *row = (uint32_t *)(vram32 + y*stride);
-      row[x] = data;
+      row[x] = palette32[vram[y*PC98_W+x]];
     }
   }
 }
diff --git a/fmdsp/fmdsp-vramlookup-ssse3.c b/fmdsp/fmdsp-vramlookup-ssse3.c
new file mode 100644
index 0000000..30e7311
--- /dev/null
+++ b/fmdsp/fmdsp-vramlookup-ssse3.c
@@ -0,0 +1,48 @@
+#include "fmdsp/fmdsp.h"
+#include <tmmintrin.h>
+
+void fmdsp_vramlookup_ssse3(uint8_t *vram32, const uint8_t *vram, const uint8_t *palette, int stride) {
+  __m128i z = _mm_setzero_si128();
+  __m128i p[3];
+  {
+    union {
+      __m128i xmm;
+      uint8_t u8[16];
+    } pi[3];
+    for (int i = 0; i < FMDSP_PALETTE_COLORS; i++) {
+      for (int c = 0; c < 3; c++) {
+        pi[c].u8[i] = palette[i*3+c];
+      }
+    }
+    for (int c = 0; c < 3; c++) {
+      p[c] = _mm_load_si128(&pi[c].xmm);
+    }
+  }
+
+  for (int y = 0; y < PC98_H; y++) {
+    for (int x = 0; x < 40; x++) {
+      // 16 pixels
+      __m128i v = _mm_loadu_si128((__m128i *)&vram[y*PC98_W+x*16]);
+
+      __m128i r = _mm_shuffle_epi8(p[0], v);
+      __m128i g = _mm_shuffle_epi8(p[1], v);
+      __m128i b = _mm_shuffle_epi8(p[2], v);
+
+      __m128i gb[2], zr[2];
+      gb[0] = _mm_unpacklo_epi8(b, g);
+      gb[1] = _mm_unpackhi_epi8(b, g);
+      zr[0] = _mm_unpacklo_epi8(r, z);
+      zr[1] = _mm_unpackhi_epi8(r, z);
+      
+      __m128i o[4];
+      o[0] = _mm_unpacklo_epi16(gb[0], zr[0]);
+      o[1] = _mm_unpackhi_epi16(gb[0], zr[0]);
+      o[2] = _mm_unpacklo_epi16(gb[1], zr[1]);
+      o[3] = _mm_unpackhi_epi16(gb[1], zr[1]);
+      for (int i = 0; i < 4; i++) {
+        _mm_storeu_si128((__m128i *)&vram32[(x*4+i)*16], o[i]);
+      }
+    }
+    vram32 += stride;
+  }
+}
diff --git a/fmdsp/fmdsp.h b/fmdsp/fmdsp.h
index a7e4aab..a8c3edc 100644
--- a/fmdsp/fmdsp.h
+++ b/fmdsp/fmdsp.h
@@ -59,6 +59,9 @@ void fmdsp_vramlookup_c(uint8_t *vram32,
                         const uint8_t *vram,
                         const uint8_t *palette,
                         int stride);
+
+void fmdsp_vramlookup_neon(uint8_t *, const uint8_t *, const uint8_t *, int);
+void fmdsp_vramlookup_ssse3(uint8_t *, const uint8_t *, const uint8_t *, int);
 #ifdef __cplusplus
 }
 #endif
diff --git a/gtk/Makefile.am b/gtk/Makefile.am
index 0a9a95b..5d75f84 100644
--- a/gtk/Makefile.am
+++ b/gtk/Makefile.am
@@ -18,12 +18,29 @@ FMDSP_SRC=../fmdsp/fmdsp.c \
           ../fmdsp/font_rom.c \
           ../fmdsp/font_fmdsp_small.c
 
+#fmplayer_CFLAGS=$(CFLAGS)
+#CFLAGS=
+fmplayer_CPPFLAGS=-Wall -Wextra -pedantic \
+                  -I.. \
+                  $(GTK3_CFLAGS) $(PORTAUDIO_CFLAGS)
+fmplayer_LDADD=$(GTK3_LIBS) $(PORTAUDIO_LIBS)
+
 if ENABLE_NEON
 LIBOPNA_SRC+=../libopna/opnassg-sinc-neon.s
 FMDSP_SRC+=../fmdsp/fmdsp-vramlookup-neon.s
 fmplayer_CCASFLAGS=-march=armv8-a -mfpu=crypto-neon-fp-armv8
 endif
 
+if ENABLE_SSE
+noinst_LIBRARIES=libsse.a
+fmplayer_LDADD+=libsse.a
+libsse_a_SOURCES=../libopna/opnassg-sinc-sse2.c \
+                 ../fmdsp/fmdsp-vramlookup-ssse3.c
+libsse_a_CPPFLAGS=$(fmplayer_CPPFLAGS)
+#no way to add -O3?? (always overridden by CFLAGS)
+libsse_a_CFLAGS=-mssse3
+endif
+
 fmplayer_SOURCES=main.c \
                  toneview.c \
                  oscilloview.c \
@@ -34,7 +51,3 @@ fmplayer_SOURCES=main.c \
                  $(FMDRIVER_SRC) \
                  $(FMDSP_SRC)
 
-fmplayer_CPPFLAGS=-Wall -Wextra -pedantic \
-                  -I.. \
-                  $(GTK3_CFLAGS) $(PORTAUDIO_CFLAGS)
-fmplayer_LDADD=$(GTK3_LIBS) $(PORTAUDIO_LIBS)
diff --git a/gtk/configure.ac b/gtk/configure.ac
index 2727888..f49bd74 100644
--- a/gtk/configure.ac
+++ b/gtk/configure.ac
@@ -2,6 +2,8 @@ AC_INIT([fmplayer], [0.1.0])
 AM_INIT_AUTOMAKE([-Wall -Werror foreign subdir-objects])
 AM_SILENT_RULES([yes])
 AC_PROG_CC_C99
+AC_PROG_RANLIB
+AM_PROG_AR
 AM_PROG_AS
 
 dnl AM_PATH_SDL2([2.0.5])
@@ -14,5 +16,11 @@ AS_IF([test "x$enable_neon" = "xyes"], [
   AC_DEFINE([ENABLE_NEON])
 ])
 
+AC_CHECK_HEADER([emmintrin.h], [emmintrin_found=yes])
+AM_CONDITIONAL([ENABLE_SSE], [test "x$emmintrin_found" = "xyes"])
+AS_IF([test "x$emmintrin_found" = "xyes"], [
+  AC_DEFINE([ENABLE_SSE])
+])
+
 AC_CONFIG_FILES([Makefile])
 AC_OUTPUT
diff --git a/gtk/main.c b/gtk/main.c
index fc72d62..855fbf4 100644
--- a/gtk/main.c
+++ b/gtk/main.c
@@ -55,7 +55,9 @@ static struct {
   const char *current_uri;
   bool oscillo_should_update;
   struct oscillodata oscillodata_audiothread[LIBOPNA_OSCILLO_TRACK_COUNT];
-} g;
+} g = {
+  .oscillo_should_update = true
+};
 
 static void quit(void) {
   if (g.pastream) {
@@ -67,18 +69,26 @@ static void quit(void) {
 }
 
 static void on_destroy(GtkWidget *w, gpointer ptr) {
+  (void)w;
+  (void)ptr;
   quit();
 }
 
 static void on_menu_quit(GtkMenuItem *menuitem, gpointer ptr) {
+  (void)menuitem;
+  (void)ptr;
   quit();
 }
 
 static void on_tone_view(GtkMenuItem *menuitem, gpointer ptr) {
+  (void)menuitem;
+  (void)ptr;
   show_toneview();
 }
 
 static void on_oscillo_view(GtkMenuItem *menuitem, gpointer ptr) {
+  (void)menuitem;
+  (void)ptr;
   show_oscilloview();
 }
 
@@ -95,6 +105,9 @@ static int pastream_cb(const void *inptr, void *outptr, unsigned long frames,
                        const PaStreamCallbackTimeInfo *timeinfo,
                        PaStreamCallbackFlags statusFlags,
                        void *userdata) {
+  (void)inptr;
+  (void)timeinfo;
+  (void)statusFlags;
   struct opna_timer *timer = (struct opna_timer *)userdata;
   int16_t *buf = (int16_t *)outptr;
   memset(outptr, 0, sizeof(int16_t)*frames*2);
@@ -133,7 +146,8 @@ static void opna_writereg_libopna(struct fmdriver_work *work, unsigned addr, uns
 }
 
 static unsigned opna_readreg_libopna(struct fmdriver_work *work, unsigned addr) {
-  struct opna_timer *timer = (struct opna_timer *)work->opna;
+  (void)work;
+  //struct opna_timer *timer = (struct opna_timer *)work->opna;
   return opna_readreg(&g.opna, addr);
 }
 
@@ -280,7 +294,7 @@ static bool openfile(const char *uri) {
   g.pa_paused = false;
   {
     const char *turi = strdup(uri);
-    free(g.current_uri);
+    free((void *)g.current_uri);
     g.current_uri = turi;
   }
   return true;
@@ -290,6 +304,7 @@ err:
 }
 
 static void on_file_activated(GtkFileChooser *chooser, gpointer ptr) {
+  (void)ptr;
   gchar *filename = gtk_file_chooser_get_uri(chooser);
   if (filename) {
     openfile(filename);
@@ -326,6 +341,8 @@ static GtkWidget *create_menubar() {
 static gboolean draw_cb(GtkWidget *w,
                  cairo_t *cr,
                  gpointer p) {
+  (void)w;
+  (void)p;
   fmdsp_update(&g.fmdsp, &g.work, &g.opna, g.vram);
   fmdsp_vrampalette(&g.fmdsp, g.vram, g.vram32, g.vram32_stride);
   cairo_surface_t *s = cairo_image_surface_create_for_data(
@@ -341,6 +358,7 @@ static gboolean draw_cb(GtkWidget *w,
 static gboolean tick_cb(GtkWidget *w,
                         GdkFrameClock *frame_clock,
                         gpointer p) {
+  (void)w;
   (void)frame_clock;
   gtk_widget_queue_draw(GTK_WIDGET(p));
   return G_SOURCE_CONTINUE;
@@ -472,6 +490,7 @@ static void drag_data_recv_cb(
   gint x, gint y,
   GtkSelectionData *data,
   guint info, guint time, gpointer ptr) {
+  (void)w;
   (void)x;
   (void)y;
   (void)info;
@@ -484,14 +503,15 @@ static void drag_data_recv_cb(
   gtk_drag_finish(ctx, TRUE, FALSE, time);
 }
 
-void opna_ssg_sinc_calc_neon(unsigned, const int16_t *, int32_t *);
-void fmdsp_vramlookup_neon(uint8_t *, const uint8_t *, const uint8_t *, int);
-
 int main(int argc, char **argv) {
 #ifdef ENABLE_NEON
   opna_ssg_sinc_calc_func = opna_ssg_sinc_calc_neon;
   fmdsp_vramlookup_func = fmdsp_vramlookup_neon;
 #endif
+#ifdef ENABLE_SSE
+  if (__builtin_cpu_supports("sse2")) opna_ssg_sinc_calc_func = opna_ssg_sinc_calc_sse2;
+  if (__builtin_cpu_supports("ssse3")) fmdsp_vramlookup_func = fmdsp_vramlookup_ssse3;
+#endif
   load_fontrom();
   gtk_init(&argc, &argv);
   GtkWidget *w = gtk_window_new(GTK_WINDOW_TOPLEVEL);
diff --git a/libopna/opnassg.h b/libopna/opnassg.h
index 223d542..aadf53e 100644
--- a/libopna/opnassg.h
+++ b/libopna/opnassg.h
@@ -66,6 +66,8 @@ typedef void (*opna_ssg_sinc_calc_func_type)(unsigned resampler_index,
 extern opna_ssg_sinc_calc_func_type opna_ssg_sinc_calc_func;
 void opna_ssg_sinc_calc_c(unsigned resampler_index,
                           const int16_t *inbuf, int32_t *outbuf);
+void opna_ssg_sinc_calc_neon(unsigned, const int16_t *, int32_t *);
+void opna_ssg_sinc_calc_sse2(unsigned, const int16_t *, int32_t *);
 
 extern const int16_t opna_ssg_sinctable[OPNA_SSG_SINCTABLELEN*2];
 
diff --git a/win32/amd64/Makefile b/win32/amd64/Makefile
index 177ef90..22ef073 100644
--- a/win32/amd64/Makefile
+++ b/win32/amd64/Makefile
@@ -9,14 +9,16 @@ vpath %.rc ..
 include ../fmplayer.mak
 
 OBJS=$(addsuffix .o,$(OBJBASE) $(RESBASE))
+OBJS+=$(addsuffix .sse.o,$(SSEOBJBASE))
 ARCH=x86_64
 PREFIX=$(ARCH)-w64-mingw32-
 CC=$(PREFIX)gcc
 WINDRES=$(PREFIX)windres
 STRIP=$(PREFIX)strip
-CFLAGS=-std=c99 -O2 -Wall -Wextra -Werror -pedantic -Wno-unused-parameter -Wno-missing-field-initializers -I../.. \
+CFLAGS=-std=c99 -O2 -Wall -Werror -Wextra -pedantic -Wno-unused-parameter -Wno-missing-field-initializers -I../.. \
        $(addprefix -D,$(DEFINES))
-LIBS=-s -mwindows -municode \
+SSECFLAGS=-mssse3 -O3
+LIBS=-mwindows -municode \
      $(addprefix -l,$(LIBBASE))
 
 $(TARGET):	$(OBJS)
@@ -28,6 +30,10 @@ $(TARGET):	$(OBJS)
 	@echo "  CC       $@"
 	@$(CC) $(CFLAGS) -c $< -o $@
 
+%.sse.o:	%.c
+	@echo "  CC       $@"
+	@$(CC) $(CFLAGS) $(SSECFLAGS) -c $< -o $@
+
 %.o:	%.rc $(ICON)
 	@echo "  WINDRES  $@"
 	@$(WINDRES) -o $@ -i $<
diff --git a/win32/fmplayer.mak b/win32/fmplayer.mak
index 1f68f4d..180de1c 100644
--- a/win32/fmplayer.mak
+++ b/win32/fmplayer.mak
@@ -15,12 +15,16 @@ LIBOPNA_OBJS=opna \
              opnatimer \
              opnafm \
              opnassg \
+             opnassg-sinc-c \
              opnadrum \
              opnaadpcm
 FMDSP_OBJS=fmdsp \
+           fmdsp-vramlookup-c \
            font_rom \
            font_fmdsp_small
 TONEDATA_OBJS=tonedata
+SSEOBJBASE=opnassg-sinc-sse2 \
+           fmdsp-vramlookup-ssse3
 OBJBASE=main \
         toneview \
         oscilloview \
diff --git a/win32/main.c b/win32/main.c
index ab038ae..8dc6bb8 100644
--- a/win32/main.c
+++ b/win32/main.c
@@ -52,6 +52,7 @@ static struct {
   struct fmplayer_file *fmfile;
   struct fmdsp fmdsp;
   uint8_t vram[PC98_W*PC98_H];
+  //uint8_t *vram;
   struct fmdsp_font font;
   uint8_t fontrom[FONT_ROM_FILESIZE];
   bool font_loaded;
@@ -66,6 +67,8 @@ static struct {
   bool fmdsp_2x;
   struct oscillodata oscillodata_audiothread[LIBOPNA_OSCILLO_TRACK_COUNT];
   UINT mmtimer;
+  HBITMAP bitmap_vram;
+  uint8_t *vram32;
 } g;
 
 HWND g_currentdlg;
@@ -445,6 +448,20 @@ static void CALLBACK mmtimer_cb(UINT timerid, UINT msg,
 
 static bool on_create(HWND hwnd, CREATESTRUCT *cs) {
   (void)cs;
+  struct bitmap_info_fmdsp {
+    BITMAPINFOHEADER head;
+    RGBQUAD colors[FMDSP_PALETTE_COLORS];
+  } bmi = {0};
+  bmi.head.biSize = sizeof(bmi.head);
+  bmi.head.biWidth = PC98_W;
+  bmi.head.biHeight = -PC98_H;
+  bmi.head.biPlanes = 1;
+  bmi.head.biBitCount = 32;
+  bmi.head.biCompression = BI_RGB;
+  //bmi.head.biClrUsed = FMDSP_PALETTE_COLORS;
+  g.bitmap_vram = CreateDIBSection(
+    0, (BITMAPINFO *)&bmi, DIB_RGB_COLORS, (void **)&g.vram32, 0, 0
+  );
   HWND button = CreateWindowEx(
     0,
     L"BUTTON",
@@ -563,40 +580,26 @@ static void on_destroy(HWND hwnd) {
 
 static void on_paint(HWND hwnd) {
   fmdsp_update(&g.fmdsp, &g.work, &g.opna, g.vram);
+  fmdsp_vrampalette(&g.fmdsp, g.vram, g.vram32, PC98_W*4);
   PAINTSTRUCT ps;
-  static BITMAPINFO *bi = 0;
-  if (!bi) {
-    bi = HeapAlloc(g.heap, HEAP_ZERO_MEMORY,
-              sizeof(BITMAPINFOHEADER) + sizeof(RGBQUAD)*FMDSP_PALETTE_COLORS);
-    if (!bi) return;
-    bi->bmiHeader.biSize = sizeof(bi->bmiHeader);
-    bi->bmiHeader.biWidth = PC98_W;
-    bi->bmiHeader.biHeight = -PC98_H;
-    bi->bmiHeader.biPlanes = 1;
-    bi->bmiHeader.biBitCount = 8;
-    bi->bmiHeader.biCompression = BI_RGB;
-    bi->bmiHeader.biClrUsed = FMDSP_PALETTE_COLORS;
-  }
-  for (int p = 0; p < FMDSP_PALETTE_COLORS; p++) {
-    bi->bmiColors[p].rgbRed = g.fmdsp.palette[p*3+0];
-    bi->bmiColors[p].rgbGreen = g.fmdsp.palette[p*3+1];
-    bi->bmiColors[p].rgbBlue = g.fmdsp.palette[p*3+2];
-  }
   HDC dc = BeginPaint(hwnd, &ps);
   HDC mdc = CreateCompatibleDC(dc);
-  HBITMAP bitmap = CreateDIBitmap(
-    dc,
-    &bi->bmiHeader, CBM_INIT,
-    g.vram,
-    bi, DIB_RGB_COLORS);
-  SelectObject(mdc, bitmap);
+  SelectObject(mdc, g.bitmap_vram);
+  /*
+  RGBQUAD palette[FMDSP_PALETTE_COLORS];
+  for (int p = 0; p < FMDSP_PALETTE_COLORS; p++) {
+    palette[p].rgbRed = g.fmdsp.palette[p*3+0];
+    palette[p].rgbGreen = g.fmdsp.palette[p*3+1];
+    palette[p].rgbBlue = g.fmdsp.palette[p*3+2];
+  }
+  SetDIBColorTable(mdc, 0, FMDSP_PALETTE_COLORS, palette);
+  */
   if (g.fmdsp_2x) {
     StretchBlt(dc, 0, 80, 1280, 800, mdc, 0, 0, 640, 400, SRCCOPY);
   } else {
     BitBlt(dc, 0, 80, 640, 400, mdc, 0, 0, SRCCOPY);
   }
   DeleteDC(mdc);
-  DeleteObject(bitmap);
   EndPaint(hwnd, &ps);
 }
 
@@ -665,8 +668,20 @@ static LRESULT CALLBACK wndproc(
   HANDLE_MSG(hwnd, WM_SYSKEYUP, on_syskey);
   HANDLE_MSG(hwnd, WM_ACTIVATE, on_activate);
   case WM_USER:
-    InvalidateRect(hwnd, 0, FALSE);
-    return 0;
+    {
+      RECT r = {
+        .left = 0,
+        .top = 80,
+        .right = 640,
+        .bottom = 480,
+      };
+      if (g.fmdsp_2x) {
+        r.right = 1280;
+        r.bottom = 880;
+      }
+      InvalidateRect(hwnd, &r, FALSE);
+      return 0;
+    }
   }
   return DefWindowProc(hwnd, msg, wParam, lParam);
 }
@@ -688,6 +703,9 @@ int CALLBACK wWinMain(HINSTANCE hinst, HINSTANCE hpinst,
   (void)hpinst;
   (void)cmdline_;
 
+  if (__builtin_cpu_supports("sse2")) opna_ssg_sinc_calc_func = opna_ssg_sinc_calc_sse2;
+  if (__builtin_cpu_supports("ssse3")) fmdsp_vramlookup_func = fmdsp_vramlookup_ssse3;
+
   const wchar_t *argfile = 0;
   {
     wchar_t *cmdline = GetCommandLine();
@@ -725,9 +743,14 @@ int CALLBACK wWinMain(HINSTANCE hinst, HINSTANCE hpinst,
   wr.top = 0;
   wr.bottom = 480;
   AdjustWindowRectEx(&wr, style, 0, exStyle);
+#ifdef _WIN64
+#define WIN64STR "(amd64)"
+#else
+#define WIN64STR ""
+#endif
   g.mainwnd = CreateWindowEx(
     exStyle,
-    (wchar_t*)((uintptr_t)wcatom), L"FMPlayer/Win32 v" FMPLAYER_VERSION_STR,
+    (wchar_t*)((uintptr_t)wcatom), L"FMPlayer/Win32 " WIN64STR " v" FMPLAYER_VERSION_STR,
     style,
     CW_USEDEFAULT, CW_USEDEFAULT,
     wr.right-wr.left, wr.bottom-wr.top,
diff --git a/win32/x86/Makefile b/win32/x86/Makefile
index b26f3e1..6146d28 100644
--- a/win32/x86/Makefile
+++ b/win32/x86/Makefile
@@ -9,6 +9,7 @@ vpath %.rc ..
 include ../fmplayer.mak
 
 OBJS=$(addsuffix .o,$(OBJBASE) $(RESBASE))
+OBJS+=$(addsuffix .sse.o,$(SSEOBJBASE))
 ARCH=i686
 PREFIX=$(ARCH)-w64-mingw32-
 CC=$(PREFIX)gcc
@@ -17,6 +18,7 @@ STRIP=$(PREFIX)strip
 CFLAGS=-std=c99 -O2 -Wall -Wextra -Werror -pedantic -I../.. \
        $(addprefix -D,$(DEFINES)) \
        -march=i586 -Wno-unused-parameter -Wno-missing-field-initializers
+SSECFLAGS=-mssse3 -O3
 LIBS=-s -mwindows -municode \
      $(addprefix -l,$(LIBBASE))
 
@@ -29,6 +31,10 @@ $(TARGET):	$(OBJS)
 	@echo "  CC       $@"
 	@$(CC) $(CFLAGS) -c $< -o $@
 
+%.sse.o:	%.c
+	@echo "  CC       $@"
+	@$(CC) $(CFLAGS) $(SSECFLAGS) -c $< -o $@
+
 %.o:	%.rc $(ICON)
 	@echo "  WINDRES  $@"
 	@$(WINDRES) -o $@ -i $<