From: notaz <notasas@gmail.com>
Date: Tue, 20 Dec 2011 00:38:53 +0000 (+0200)
Subject: gpu_unai: add 4bpp asm sprite blitter
X-Git-Tag: r12~16
X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?p=pcsx_rearmed.git;a=commitdiff_plain;h=b3db94096d7e5b4f60d610a441e370d639b3fd06;ds=inline

gpu_unai: add 4bpp asm sprite blitter
---

diff --git a/plugins/gpu_neon/Makefile b/plugins/gpu_neon/Makefile
index f2277522..66a2bd5c 100644
--- a/plugins/gpu_neon/Makefile
+++ b/plugins/gpu_neon/Makefile
@@ -6,6 +6,9 @@ ARCH = $(shell $(CC) -v 2>&1 | grep -i 'target:' | awk '{print $$2}' | awk -F '-
 
 CFLAGS += -ggdb -Wall -fPIC -O2
 CFLAGS += -fno-strict-aliasing
+ifdef MAEMO
+CFLAGS += -DMAEMO
+endif
 
 SRC += gpu.c
 ifeq "$(ARCH)" "arm"
@@ -38,7 +41,7 @@ gpu_neon.$(EXT): SRC += psx_gpu_if.c
 gpu_neon.$(EXT): CFLAGS += -fno-strict-aliasing
 gpu_peops.$(EXT): SRC += peops_if.c
 gpu_peops.$(EXT): CFLAGS += -fno-strict-aliasing
-gpu_unai.$(EXT): SRC += unai_if.cpp
+gpu_unai.$(EXT): SRC += unai_if.cpp ../gpu_unai/gpu_arm.s
 gpu_unai.$(EXT): CC_ = $(CXX)
 
 all: $(TARGETS)
diff --git a/plugins/gpu_unai/Makefile b/plugins/gpu_unai/Makefile
index 531c9fb6..e4b6e6b9 100644
--- a/plugins/gpu_unai/Makefile
+++ b/plugins/gpu_unai/Makefile
@@ -16,7 +16,7 @@ CFLAGS += -mcpu=arm926ej-s -mtune=arm926ej-s
 SRC += ../../frontend/cspace.c
 endif
 
-SRC += gpu.cpp
+SRC += gpu.cpp gpu_arm.s
 
 TARGET = gpuPCSX4ALL.so
 LDFLAGS += -shared -Wl,-soname,$(TARGET)
diff --git a/plugins/gpu_unai/gpu_arm.h b/plugins/gpu_unai/gpu_arm.h
new file mode 100644
index 00000000..a0b22487
--- /dev/null
+++ b/plugins/gpu_unai/gpu_arm.h
@@ -0,0 +1,9 @@
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void draw_spr16_full(u16 *d, void *s, u16 *pal, int lines);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/plugins/gpu_unai/gpu_arm.s b/plugins/gpu_unai/gpu_arm.s
new file mode 100644
index 00000000..a2fa1741
--- /dev/null
+++ b/plugins/gpu_unai/gpu_arm.s
@@ -0,0 +1,55 @@
+/*
+ * (C) Gražvydas "notaz" Ignotas, 2011
+ *
+ * This work is licensed under the terms of  GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory.
+ */
+
+
+.text
+.align 2
+
+@ in: r0=dst, r2=pal, r12=0x1e
+@ trashes r6-r8,lr,flags
+.macro do_4_pixels rs ibase obase
+.if \ibase - 1 < 0
+    and     r6, r12, \rs, lsl #1
+.else
+    and     r6, r12, \rs, lsr #\ibase-1
+.endif
+    and     r7, r12, \rs, lsr #\ibase+3
+    and     r8, r12, \rs, lsr #\ibase+7
+    and     lr, r12, \rs, lsr #\ibase+11
+    ldrh    r6, [r2, r6]
+    ldrh    r7, [r2, r7]
+    ldrh    r8, [r2, r8]
+    ldrh    lr, [r2, lr]
+    tst     r6, r6
+    strneh  r6, [r0, #\obase+0]
+    tst     r7, r7
+    strneh  r7, [r0, #\obase+2]
+    tst     r8, r8
+    strneh  r8, [r0, #\obase+4]
+    tst     lr, lr
+    strneh  lr, [r0, #\obase+6]
+.endm
+
+.global draw_spr16_full @ (u16 *d, void *s, u16 *pal, int lines)
+draw_spr16_full:
+    stmfd   sp!, {r4-r8,lr}
+    mov     r12, #0x1e             @ empty pixel
+
+0:
+    ldmia   r1, {r4,r5}
+    do_4_pixels r4, 0,  0
+    do_4_pixels r4, 16, 8
+    do_4_pixels r5, 0,  16
+    do_4_pixels r5, 16, 24
+    subs    r3, #1
+    add     r0, #2048
+    add     r1, #2048
+    bgt     0b
+
+    ldmfd   sp!, {r4-r8,pc}
+
+@ vim:filetype=armasm
diff --git a/plugins/gpu_unai/gpu_command.h b/plugins/gpu_unai/gpu_command.h
index d380daec..049b1462 100644
--- a/plugins/gpu_unai/gpu_command.h
+++ b/plugins/gpu_unai/gpu_command.h
@@ -352,6 +352,16 @@ void gpuSendPacketFunction(const int PRIM)
 			break;
 		case 0x7C:
 		case 0x7D:
+#ifdef __arm__
+			if ((GPU_GP1 & 0x180) == 0 && (Masking | PixelMSB) == 0)
+			{
+				gpuSetCLUT    (PacketBuffer.U4[2] >> 16);
+				gpuSetTexture (GPU_GP1);
+				gpuDrawS16();
+				break;
+			}
+			// fallthrough
+#endif
 		case 0x7E:
 		case 0x7F:
 			if (!isSkip)
diff --git a/plugins/gpu_unai/gpu_raster_sprite.h b/plugins/gpu_unai/gpu_raster_sprite.h
index 041fcbdc..5075227e 100644
--- a/plugins/gpu_unai/gpu_raster_sprite.h
+++ b/plugins/gpu_unai/gpu_raster_sprite.h
@@ -84,6 +84,47 @@ void gpuDrawS(const PS gpuSpriteSpanDriver)
 	}
 }
 
+#ifdef __arm__
+#include "gpu_arm.h"
+
+void gpuDrawS16(void)
+{
+	s32 x0, y0;
+	s32 u0, v0;
+	s32 xmin, xmax;
+	s32 ymin, ymax;
+	u32 h = 16;
+
+	x0 = GPU_EXPANDSIGN_SPRT(PacketBuffer.S2[2]) + DrawingOffset[0];
+	y0 = GPU_EXPANDSIGN_SPRT(PacketBuffer.S2[3]) + DrawingOffset[1];
+
+	xmin = DrawingArea[0];	xmax = DrawingArea[2];
+	ymin = DrawingArea[1];	ymax = DrawingArea[3];
+	u0 = PacketBuffer.U1[8];
+	v0 = PacketBuffer.U1[9];
+
+	if (x0 > xmax - 16 || x0 < xmin ||
+	    ((u0 | v0) & 15) || !(TextureWindow[2] & TextureWindow[3] & 8)) {
+		// send corner cases to general handler
+		PacketBuffer.U4[3] = 0x00100010;
+		gpuDrawS(gpuSpriteSpanFn<0x20>);
+		return;
+	}
+
+	if (y0 >= ymax || y0 <= ymin - 16)
+		return;
+	if (y0 < ymin) {
+		h -= ymin - y0;
+		v0 += ymin - y0;
+		y0 = ymin;
+	}
+	else if (ymax - y0 < 16)
+		h = ymax - y0;
+
+	draw_spr16_full(&GPU_FrameBuffer[FRAME_OFFSET(x0, y0)], &TBA[FRAME_OFFSET(u0/4, v0)], CBA, h);
+}
+#endif // __arm__
+
 ///////////////////////////////////////////////////////////////////////////////
 void gpuDrawT(const PT gpuTileSpanDriver)
 {