Merge pull request #256 from fjtrujy/master

[PS2] Improvement paddings

Makefile.libretro

@@ -14,6 +14,7 @@ filter_out1 = $(filter-out $(firstword $1),$1)
 filter_out2 = $(call filter_out1,$(call filter_out1,$1))
 unixpath = $(subst \,/,$1)
 unixcygpath = /$(subst :,,$(call unixpath,$1))
+RETROARCH_OBJECTS :=
 
 ifeq ($(platform),)
 	platform = unix
@@ -163,10 +164,13 @@ else ifeq ($(platform), ps2)
 	FCEU_DEFINES := -DPATH_MAX=1024 -DINLINE=inline -DPSS_STYLE=1 -DFCEU_VERSION_NUMERIC=9813 -DHAVE_ASPRINTF
 	ENDIANNESS_DEFINES := -DLSB_FIRST -DLOCAL_LE=1
 	PLATFORM_DEFINES := -DPS2 -G0 -DFRONTEND_SUPPORTS_ABGR1555 -DRENDER_GSKIT_PS2
+	PLATFORM_DEFINES += -Dmemcpy=mips_memcpy -Dmemset=mips_memset
 	PLATFORM_DEFINES += -I$(PS2SDK)/ee/include -I$(PS2SDK)/common/include -Ips2/ -I$(PS2DEV)/gsKit/include
 	STATIC_LINKING=1
 	EXTERNAL_ZLIB=1
 
+	RETROARCH_OBJECTS +=  ps2/SMS_Utils.o
+
 # PSP
 else ifeq ($(platform), psp1)
 	TARGET := $(TARGET_NAME)_libretro_$(platform).a
@@ -672,6 +676,7 @@ include Makefile.common
 CFLAGS += -DWANT_GRIFFIN
 
 OBJECTS := $(SOURCES_C:.c=.o)
+OBJECTS += $(RETROARCH_OBJECTS)
 
 DEFINES := $(COREDEFINES) $(PLATFORM_DEFINES)
 
@@ -739,6 +744,11 @@ endif
 %.o: %.c
 	$(CC) -c $(OBJOUT)$@ $< $(CFLAGS) $(INCFLAGS)
 
+ifeq ($(platform), ps2)
+%.o: %.s
+	ee-as -G0 $< -o $@
+endif
+
 clean-objs:
 	rm -f $(OBJECTS)
 

ps2/SMS_Utils.s

@@ -0,0 +1,202 @@
+/*
+#     ___  _ _      ___
+#    |    | | |    |
+# ___|    |   | ___|    PS2DEV Open Source Project.
+#----------------------------------------------------------
+# MUL64 is pulled from some binary library (I don't remember which one).
+# mips_memcpy routine is pulled from 'sde' library from MIPS.
+#
+*/
+.set noat
+.set noreorder
+.set nomacro
+
+.globl MUL64
+.globl mips_memcpy
+.globl mips_memset
+
+.text
+
+MUL64:
+    pmultuw	$v0, $a0, $a1
+    dsra32	$a2, $a0, 0
+    dsra32  $v1, $a1, 0
+    mult    $v1, $a0, $v1
+    mult1   $a2, $a2, $a1
+    addu    $v1, $v1, $a2
+    dsll32  $v1, $v1, 0
+    jr      $ra
+    daddu   $v0, $v0, $v1
+
+mips_memcpy:
+    addu    $v0, $a0, $zero
+    beqz    $a2, 1f
+    sltiu   $t2, $a2, 12
+    bnez    $t2, 2f
+    xor     $v1, $a1, $a0
+    andi    $v1, $v1, 7
+    negu    $a3, $a0
+    beqz    $v1, 3f
+    andi    $a3, $a3, 7
+    beqz    $a3, 4f
+    subu    $a2, $a2, $a3
+    ldr     $v1, 0($a1)
+    ldl     $v1, 7($a1)
+    addu    $a1, $a1, $a3
+    sdr     $v1, 0($a0)
+    addu    $a0, $a0, $a3
+4:
+    andi    $v1, $a2, 31
+    subu    $a3, $a2, $v1
+    beqz    $a3, 5f
+    addu    $a2, $v1, $zero
+    addu    $a3, $a3, $a1
+6:
+    ldr     $v1,  0($a1)
+    ldl     $v1,  7($a1)
+    ldr     $t0,  8($a1)
+    ldl     $t0, 15($a1)
+    ldr     $t1, 16($a1)
+    ldl     $t1, 23($a1)
+    ldr     $t2, 24($a1)
+    ldl     $t2, 31($a1)
+    sd      $v1,  0($a0)
+    sd      $t0,  8($a0)
+    sd      $t1, 16($a0)
+    addiu   $a1, $a1, 32
+    addiu   $a0, $a0, 32
+    bne     $a1, $a3, 6b
+    sd      $t2, -8($a0)
+5:
+    andi    $v1, $a2, 7
+    subu    $a3, $a2, $v1
+    beqz    $a3, 2f
+    addu    $a2, $v1, $zero
+    addu    $a3, $a3, $a1
+7:
+    ldr     $v1, 0($a1)
+    ldl     $v1, 7($a1)
+    addiu   $a1, $a1, 8
+    addiu   $a0, $a0, 8
+    nop
+    bne     $a1, $a3, 7b
+    sd      $v1, -8($a0)
+    beq     $zero, $zero, 2f
+    nop
+3:
+    beqz    $a3, 8f
+    subu    $a2, $a2, $a3
+    ldr     $v1, 0($a1)
+    addu    $a1, $a1, $a3
+    sdr     $v1, 0($a0)
+    addu    $a0, $a0, $a3
+8:
+    andi    $v1, $a2, 31
+    subu    $a3, $a2, $v1
+    beqz    $a3, 9f
+    addu    $a2, $v1, $zero
+    addu    $a3, $a3, $a1
+10:
+    ld      $v1,  0($a1)
+    ld      $t0,  8($a1)
+    ld      $t1, 16($a1)
+    ld      $t2, 24($a1)
+    sd      $v1,  0($a0)
+    sd      $t0,  8($a0)
+    sd      $t1, 16($a0)
+    addiu   $a1, $a1, 32
+    addiu   $a0, $a0, 32
+    bne     $a1, $a3, 10b
+    sd      $t2, -8($a0)
+9:
+    andi    $v1, $a2, 7
+    subu    $a3, $a2, $v1
+    beqz    $a3, 2f
+    addu    $a2, $v1, $zero
+    addu    $a3, $a3, $a1
+11:
+    ld      $v1, 0($a1)
+    addiu   $a1, $a1, 8
+    addiu   $a0, $a0, 8
+    nop
+    nop
+    bne     $a1, $a3, 11b
+    sd      $v1, -8($a0)
+2:
+    beqz    $a2, 1f
+    addu    $a3, $a2, $a1
+12:
+    lbu     $v1, 0($a1)
+    addiu   $a1, $a1, 1
+    addiu   $a0, $a0, 1
+    nop
+    nop
+    bne     $a1, $a3, 12b
+    sb      $v1, -1($a0)
+1:
+    jr      $ra
+    nop
+
+mips_memset:
+    beqz    $a2, 1f
+    sltiu   $at, $a2, 16
+    bnez    $at, 2f
+    andi    $a1, $a1, 0xFF
+    dsll    $at, $a1, 0x8
+    or      $a1, $a1, $at
+    dsll    $at, $a1, 0x10
+    or      $a1, $a1, $at
+    dsll32  $at, $a1, 0x0
+    or      $a1, $a1, $at
+    andi    $v1, $a0, 0x7
+    beqz    $v1, 3f
+    li      $a3, 8
+    subu    $a3, $a3, $v1
+    subu    $a2, $a2, $a3
+    sdr     $a1, 0($a0)
+    addu    $a0, $a0, $a3
+3:
+    andi    $v1, $a2, 0x1f
+    subu    $a3, $a2, $v1
+    beqz    $a3, 4f
+    move    $a2, $v1
+    addu    $a3, $a3, $a0
+5:
+    sd      $a1,  0($a0)
+    sd      $a1,  8($a0)
+    sd      $a1, 16($a0)
+    addiu   $a0, $a0, 32
+    sd      $a1, -8($a0)
+    bne     $a0, $a3, 5b
+4:
+    andi    $v1, $a2, 0x7
+    subu    $a3, $a2, $v1
+    beqz    $a3, 2f
+    move    $a2, $v1
+    addu    $a3, $a3, $a0
+6:
+    addiu   $a0, $a0, 8
+    beq     $a0, $a3, 2f
+    sd      $a1, -8($a0)
+    addiu   $a0, $a0, 8
+    beq     $a0, $a3, 2f
+    sd      $a1, -8($a0)
+    addiu   $a0, $a0, 8
+    bne     $a0, $a3, 6b
+    sd      $a1, -8($a0)
+2:
+    beqz    $a2, 1f
+    addu    $a3, $a2, $a0
+7:
+    addiu   $a0, $a0, 1
+    beq     $a0, $a3, 1f
+    sb      $a1, -1($a0)
+    addiu   $a0, $a0, 1
+    beq     $a0, $a3, 1f
+    sb      $a1, -1($a0)
+    addiu   $a0, $a0, 1
+    bne     $a0, $a3, 7b
+    sb      $a1, -1($a0)
+1:
+    jr  $ra
+    nop

src/drivers/libretro/libretro-common/include/libretro_gskit_ps2.h

@@ -35,6 +35,16 @@
 
 #define RETRO_HW_RENDER_INTERFACE_GSKIT_PS2_VERSION 1
 
+struct retro_hw_ps2_insets
+{
+  float top;
+  float left;
+  float bottom;
+  float right;
+};
+
+#define empty_ps2_insets (struct retro_hw_ps2_insets){0.f, 0.f, 0.f, 0.f}
+
 struct retro_hw_render_interface_gskit_ps2
 {
   /* Must be set to RETRO_HW_RENDER_INTERFACE_GSKIT_PS2. */
@@ -48,6 +58,8 @@ struct retro_hw_render_interface_gskit_ps2
    */
    GSTEXTURE *coreTexture;
    bool clearTexture;
+   bool updatedPalette;
+   struct retro_hw_ps2_insets padding;
 };
 typedef struct retro_hw_render_interface_gskit_ps2 RETRO_HW_RENDER_INTEFACE_GSKIT_PS2;
 

src/drivers/libretro/libretro.c

@@ -55,6 +55,10 @@ void* linearMemAlign(size_t size, size_t alignment);
 void linearFree(void* mem);
 #endif
 
+#if defined(RENDER_GSKIT_PS2)
+RETRO_HW_RENDER_INTEFACE_GSKIT_PS2 *ps2 = NULL;
+#endif
+
 static retro_video_refresh_t video_cb = NULL;
 static retro_input_poll_t poll_cb = NULL;
 static retro_input_state_t input_cb = NULL;
@@ -1564,16 +1568,9 @@ static void retro_run_blit(uint8_t *gfx)
 
    video_cb(texture_vram_p, width, height, 256);
 #elif defined(RENDER_GSKIT_PS2)
-
-   RETRO_HW_RENDER_INTEFACE_GSKIT_PS2 *ps2 = NULL;
    uint32_t *buf = (uint32_t *)RETRO_HW_FRAME_BUFFER_VALID;
 
-   incr   += (overscan_h ? 16 : 0);
+   if (!ps2) {
-   width  -= (overscan_h ? 16 : 0);
-   height -= (overscan_v ? 16 : 0);
-   pitch  -= (overscan_h ? 32 : 0);
-   gfx    += (overscan_v ? ((overscan_h ? 8 : 0) + 256 * 8) : (overscan_h ? 8 : 0));
-
       if (!environ_cb(RETRO_ENVIRONMENT_GET_HW_RENDER_INTERFACE, (void **)&ps2) || !ps2) {
          printf("Failed to get HW rendering interface!\n");
          return;
@@ -1585,21 +1582,17 @@ static void retro_run_blit(uint8_t *gfx)
          return;
       }
 
-   if (ps2->clearTexture || !ps2->coreTexture->Clut || !ps2->coreTexture->Mem ) {
-      /* If it is empty we need to create it */
       ps2->coreTexture->Width = width;
       ps2->coreTexture->Height = height;
       ps2->coreTexture->PSM = GS_PSM_T8;
       ps2->coreTexture->ClutPSM = GS_PSM_CT16;
       ps2->coreTexture->Filter = GS_FILTER_LINEAR;
+      ps2->coreTexture->Clut = (u32*)retro_palette;
+      ps2->updatedPalette = true;
+      ps2->padding = (struct retro_hw_ps2_insets){8.0f, 8.0f, 8.0f, 8.0f};
    }
 
-   for (y = 0; y < height; y++, gfx += incr)
+   ps2->coreTexture->Mem = (u32*)gfx;
-         for ( x = 0; x < width; x++, gfx++)
-            fceu_video_out[y * width + x] = *gfx;
-
-   ps2->coreTexture->Clut = (u32*)retro_palette;
-   ps2->coreTexture->Mem = (u32*)fceu_video_out;
 
    video_cb(buf, width, height, pitch);
 #else