From: notaz <notasas@gmail.com>
Date: Mon, 8 Aug 2022 18:16:59 +0000 (+0300)
Subject: gpu_neon: integration for arm64
X-Git-Tag: r24l~433
X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=047899a4f30d4a11e467e6b5aa0ba4cbf94c3f6b;p=pcsx_rearmed.git

gpu_neon: integration for arm64
---

diff --git a/Makefile b/Makefile
index 66675ffc..0cc90ce8 100644
--- a/Makefile
+++ b/Makefile
@@ -81,7 +81,7 @@ endif
 ifeq "$(ARCH)" "arm"
 OBJS += libpcsxcore/gte_arm.o
 endif
-ifeq "$(HAVE_NEON)" "1"
+ifeq "$(HAVE_NEON_ASM)" "1"
 OBJS += libpcsxcore/gte_neon.o
 endif
 libpcsxcore/psxbios.o: CFLAGS += -Wno-nonnull
@@ -186,13 +186,16 @@ OBJS += plugins/gpulib/gpu.o plugins/gpulib/vout_pl.o
 ifeq "$(BUILTIN_GPU)" "neon"
 CFLAGS += -DGPU_NEON
 OBJS += plugins/gpu_neon/psx_gpu_if.o
-ifeq "$(HAVE_NEON)" "1"
-OBJS += plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.o
 plugins/gpu_neon/psx_gpu_if.o: CFLAGS += -DNEON_BUILD -DTEXTURE_CACHE_4BPP -DTEXTURE_CACHE_8BPP
-else
-plugins/gpu_neon/psx_gpu_if.o: CFLAGS += -DTEXTURE_CACHE_4BPP -DTEXTURE_CACHE_8BPP
-endif
 plugins/gpu_neon/psx_gpu_if.o: plugins/gpu_neon/psx_gpu/*.c
+frontend/menu.o frontend/plugin_lib.o: CFLAGS += -DBUILTIN_GPU_NEON
+ ifeq "$(HAVE_NEON_ASM)" "1"
+ OBJS += plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.o
+ else
+ OBJS += plugins/gpu_neon/psx_gpu/psx_gpu_simd.o
+ plugins/gpu_neon/psx_gpu_if.o: CFLAGS += -DSIMD_BUILD
+ plugins/gpu_neon/psx_gpu/psx_gpu_simd.o: CFLAGS += -DSIMD_BUILD
+ endif
 endif
 ifeq "$(BUILTIN_GPU)" "peops"
 CFLAGS += -DGPU_PEOPS
@@ -256,11 +259,13 @@ endif
 
 # frontend/gui
 OBJS += frontend/cspace.o
-ifeq "$(HAVE_NEON)" "1"
+ifeq "$(HAVE_NEON_ASM)" "1"
 OBJS += frontend/cspace_neon.o
+frontend/cspace.o: CFLAGS += -DHAVE_bgr555_to_rgb565 -DHAVE_bgr888_to_x
 else
 ifeq "$(ARCH)" "arm"
 OBJS += frontend/cspace_arm.o
+frontend/cspace.o: CFLAGS += -DHAVE_bgr555_to_rgb565
 endif
 endif
 
@@ -340,7 +345,7 @@ OBJS += frontend/plugin_lib.o
 OBJS += frontend/libpicofe/linux/plat.o
 OBJS += frontend/libpicofe/readpng.o frontend/libpicofe/fonts.o
 frontend/libpicofe/linux/plat.o: CFLAGS += -DNO_HOME_DIR
-ifeq "$(HAVE_NEON)" "1"
+ifeq "$(HAVE_NEON_ASM)" "1"
 OBJS += frontend/libpicofe/arm/neon_scale2x.o
 OBJS += frontend/libpicofe/arm/neon_eagle2x.o
 frontend/libpicofe/arm/neon_scale2x.o: CFLAGS += -DDO_BGR_TO_RGB
diff --git a/configure b/configure
index 4d36f4b5..e79b790b 100755
--- a/configure
+++ b/configure
@@ -50,6 +50,7 @@ have_armv5=""
 have_armv6=""
 have_armv7=""
 have_arm_neon=""
+have_arm_neon_asm=""
 have_tslib=""
 have_gles=""
 have_c64x_dsp=""
@@ -209,7 +210,7 @@ arm*)
   fi
 
   if [ "x$have_arm_neon" = "x" ]; then
-    # detect NEON from user-supplied cflags to enable asm code
+    # detect NEON from user-supplied cflags to enable neon code
     have_arm_neon=`check_define __ARM_NEON__ && echo yes` || true
   fi
   if [ "x$have_armv7" = "x" ]; then
@@ -274,8 +275,14 @@ arm*)
     echo "You probably want to specify -mcpu= or -march= like this:"
     echo "  CFLAGS=-march=armv7-a ./configure ..."
   fi
+  have_arm_neon_asm=$have_arm_neon
   ;;
 aarch64)
+  have_arm_neon="yes"
+  have_arm_neon_asm="no"
+  if [ "x$builtin_gpu" = "x" ]; then
+    builtin_gpu="neon"
+  fi
   ;;
 *)
   # dynarec only available on ARM
@@ -504,6 +511,7 @@ done
 test "x$have_armv6" != "x" || have_armv6="no"
 test "x$have_armv7" != "x" || have_armv7="no"
 test "x$have_arm_neon" != "x" || have_arm_neon="no"
+test "x$have_arm_neon_asm" != "x" || have_arm_neon_asm="no"
 test "x$have_gles" != "x" || have_gles="no"
 test "x$have_c64x_dsp" != "x" || have_c64x_dsp="no"
 
@@ -517,9 +525,11 @@ echo "C compiler flags    $CFLAGS"
 echo "libraries           $MAIN_LDLIBS"
 echo "linker flags        $LDFLAGS$MAIN_LDFLAGS"
 echo "enable dynarec      $enable_dynarec"
+if [ "$ARCH" = "arm" -o "$ARCH" = "aarch64" ]; then
+  echo "enable ARM NEON     $have_arm_neon"
+fi
 if [ "$ARCH" = "arm" ]; then
   echo "ARMv7 optimizations $have_armv7"
-  echo "enable ARM NEON     $have_arm_neon"
   echo "TI C64x DSP support $have_c64x_dsp"
 fi
 echo "tslib support       $have_tslib"
@@ -554,6 +564,9 @@ echo "PLUGINS = $plugins" >> $config_mak
 if [ "$have_arm_neon" = "yes" ]; then
   echo "HAVE_NEON = 1" >> $config_mak
 fi
+if [ "$have_arm_neon_asm" = "yes" ]; then
+  echo "HAVE_NEON_ASM = 1" >> $config_mak
+fi
 if [ "$have_tslib" = "yes" ]; then
   echo "HAVE_TSLIB = 1" >> $config_mak
 fi
diff --git a/frontend/cspace.c b/frontend/cspace.c
index 1dbc8a66..1557f8e7 100644
--- a/frontend/cspace.c
+++ b/frontend/cspace.c
@@ -22,7 +22,7 @@
 #define LE16TOHx2(x) (x)
 #endif
 
-#ifndef __arm__
+#ifndef HAVE_bgr555_to_rgb565
 
 void bgr555_to_rgb565(void *dst_, const void *src_, int bytes)
 {
@@ -43,7 +43,7 @@ void bgr555_to_rgb565(void *dst_, const void *src_, int bytes)
 
 #endif
 
-#if defined(__arm64__) || !defined(__ARM_NEON__)
+#ifndef HAVE_bgr888_to_x
 
 void bgr888_to_rgb565(void *dst_, const void *src_, int bytes)
 {
diff --git a/frontend/menu.c b/frontend/menu.c
index 38683a26..6a81edb2 100644
--- a/frontend/menu.c
+++ b/frontend/menu.c
@@ -1344,7 +1344,7 @@ static int menu_loop_gfx_options(int id, int keys)
 
 // ------------ bios/plugins ------------
 
-#ifdef __ARM_NEON__
+#ifdef BUILTIN_GPU_NEON
 
 static const char h_gpu_neon[] =
 	"Configure built-in NEON GPU plugin";
@@ -1490,7 +1490,7 @@ static const char h_bios[]       = "HLE is simulated BIOS. BIOS selection is sav
 				   "savestates and can't be changed there. Must save\n"
 				   "config and reload the game for change to take effect";
 static const char h_plugin_gpu[] = 
-#ifdef __ARM_NEON__
+#ifdef BUILTIN_GPU_NEON
 				   "builtin_gpu is the NEON GPU, very fast and accurate\n"
 #endif
 				   "gpu_peops is Pete's soft GPU, slow but accurate\n"
@@ -1509,7 +1509,7 @@ static menu_entry e_menu_plugin_options[] =
 	mee_enum_h    ("BIOS",                          0, bios_sel, bioses, h_bios),
 	mee_enum_h    ("GPU plugin",                    0, gpu_plugsel, gpu_plugins, h_plugin_gpu),
 	mee_enum_h    ("SPU plugin",                    0, spu_plugsel, spu_plugins, h_plugin_spu),
-#ifdef __ARM_NEON__
+#ifdef BUILTIN_GPU_NEON
 	mee_handler_h ("Configure built-in GPU plugin", menu_loop_plugin_gpu_neon, h_gpu_neon),
 #endif
 	mee_handler_h ("Configure gpu_peops plugin",    menu_loop_plugin_gpu_peops, h_gpu_peops),
@@ -1964,7 +1964,7 @@ static const char credits_text[] =
 	"(C) 2005-2009 PCSX-df Team\n"
 	"(C) 2009-2011 PCSX-Reloaded Team\n\n"
 	"ARM recompiler (C) 2009-2011 Ari64\n"
-#ifdef __ARM_NEON__
+#ifdef BUILTIN_GPU_NEON
 	"ARM NEON GPU (c) 2011-2012 Exophase\n"
 #endif
 	"PEOpS GPU and SPU by Pete Bernert\n"
diff --git a/frontend/plugin_lib.c b/frontend/plugin_lib.c
index 5a01608d..ca68694e 100644
--- a/frontend/plugin_lib.c
+++ b/frontend/plugin_lib.c
@@ -473,7 +473,7 @@ static int dispmode_default(void)
 	return 1;
 }
 
-#ifdef __ARM_NEON__
+#ifdef BUILTIN_GPU_NEON
 static int dispmode_doubleres(void)
 {
 	if (!(pl_rearmed_cbs.gpu_caps & GPU_CAP_SUPPORTS_2X)
@@ -485,7 +485,9 @@ static int dispmode_doubleres(void)
 	snprintf(hud_msg, sizeof(hud_msg), "double resolution");
 	return 1;
 }
+#endif
 
+#ifdef __ARM_NEON__
 static int dispmode_scale2x(void)
 {
 	if (!resolution_ok(psx_w * 2, psx_h * 2) || psx_bpp != 16)
@@ -511,8 +513,10 @@ static int dispmode_eagle2x(void)
 
 static int (*dispmode_switchers[])(void) = {
 	dispmode_default,
-#ifdef __ARM_NEON__
+#ifdef BUILTIN_GPU_NEON
 	dispmode_doubleres,
+#endif
+#ifdef __ARM_NEON__
 	dispmode_scale2x,
 	dispmode_eagle2x,
 #endif
diff --git a/plugins/gpu_neon/Makefile b/plugins/gpu_neon/Makefile
index 08bf0ee6..955feab6 100644
--- a/plugins/gpu_neon/Makefile
+++ b/plugins/gpu_neon/Makefile
@@ -5,9 +5,14 @@ include ../../config.mak
 SRC += psx_gpu_if.c
 
 CFLAGS += -DTEXTURE_CACHE_4BPP -DTEXTURE_CACHE_8BPP
-ifeq "$(HAVE_NEON)" "1"
-SRC += psx_gpu/psx_gpu_arm_neon.S
 CFLAGS += -DNEON_BUILD
+ifeq "$(HAVE_NEON)" "1"
+ ifeq "$(HAVE_NEON_ASM)" "1"
+ SRC += psx_gpu/psx_gpu_arm_neon.S
+ else
+ OBJS += psx_gpu/psx_gpu_simd.c
+ CFLAGS += -DSIMD_BUILD
+ endif
 else
 CFLAGS += -fno-strict-aliasing
 endif