Initial messy ARM ASM work

8 files changed, 129 insertions, 19 deletions

diff --git a/configure.ac b/configure.ac
index 5c61f1a..084e484 100644
--- a/configure.ac
+++ b/configure.ac
@@ -74,21 +74,24 @@ AS_IF([test x$enable_asm != xno], [
             X64=true
             BITS=64
             ASFLAGS="$ASFLAGS -DARCH_X86_64=1 -m amd64" ],
+        [armv7*], [
+            ARM=true
+            BITS=32 ],
         )
-    AS_CASE([$host],
-        [*darwin*], [
-            ASFLAGS="$ASFLAGS -f macho$BITS -DPREFIX -DHAVE_ALIGNED_STACK=1" ],
-        [*linux*|*dragonfly*|*bsd*|*solaris*], [
-            ASFLAGS="$ASFLAGS -f elf -DHAVE_ALIGNED_STACK=1" ],
-        [*cygwin*|*mingw*], [
-            ASFLAGS="$ASFLAGS -f win$BITS"
-            AS_IF([test x$BITS = x64], [
-                ASFLAGS="$ASFLAGS -DHAVE_ALIGNED_STACK=1"
-            ], [
-                ASFLAGS="$ASFLAGS -DHAVE_ALIGNED_STACK=0 -DPREFIX"
-            ])
-        ])
     AS_IF([test x$INTEL = xtrue], [
+        AS_CASE([$host],
+            [*darwin*], [
+                ASFLAGS="$ASFLAGS -f macho$BITS -DPREFIX -DHAVE_ALIGNED_STACK=1" ],
+            [*linux*|*dragonfly*|*bsd*|*solaris*], [
+                ASFLAGS="$ASFLAGS -f elf -DHAVE_ALIGNED_STACK=1" ],
+            [*cygwin*|*mingw*], [
+                ASFLAGS="$ASFLAGS -f win$BITS"
+                AS_IF([test x$BITS = x64], [
+                    ASFLAGS="$ASFLAGS -DHAVE_ALIGNED_STACK=1"
+                ], [
+                    ASFLAGS="$ASFLAGS -DHAVE_ALIGNED_STACK=0 -DPREFIX"
+                ])
+            ])
         AC_CHECK_PROG([yasm_check], [$AS], [yes])
         AS_IF([test x$yasm_check != xyes], [
             AC_MSG_WARN(yasm was not found; ASM functions are disabled.)
@@ -116,6 +119,7 @@ AC_SUBST([AS], ["$AS"])
 
 AM_CONDITIONAL([ASM], [test x$enable_asm != xno])
 AM_CONDITIONAL([INTEL], [test x$INTEL = xtrue])
+AM_CONDITIONAL([ARM], [test x$ARM = xtrue])
 AM_CONDITIONAL([X86], [test x$X86 = xtrue])
 AM_CONDITIONAL([X64], [test x$X64 = xtrue])
 
diff --git a/libass/Makefile.am b/libass/Makefile.am
index f2ecb74..1dece0d 100644
--- a/libass/Makefile.am
+++ b/libass/Makefile.am
@@ -18,6 +18,8 @@ SRC_INTEL = x86/blend_bitmaps.asm x86/blur.asm x86/cpuid.asm x86/utils.asm \
 SRC_INTEL64 = x86/be_blur.asm
 SRC_INTEL_RASTERIZER = x86/rasterizer.asm
 
+SRC_ARM = arm/blend_bitmaps.S
+
 SRC_RASTERIZER = ass_rasterizer.h ass_rasterizer.c ass_rasterizer_c.c
 
 SRC_DIRECTWRITE = ass_directwrite.c ass_directwrite.h dwrite_c.h
@@ -58,6 +60,9 @@ if X64
 libass_la_SOURCES += $(SRC_INTEL64)
 endif
 endif
+if ARM
+libass_la_SOURCES += $(SRC_ARM)
+endif
 endif
 
 assheadersdir = $(includedir)/ass
diff --git a/libass/arm/asm.S b/libass/arm/asm.S
new file mode 100644
index 0000000..4b98608
--- /dev/null
+++ b/libass/arm/asm.S
@@ -0,0 +1,11 @@
+.macro function name
+    .macro endfunc
+        .size   ass_\name, . - ass_\name
+        .endfunc
+        .purgem endfunc
+    .endm
+        .global ass_\name
+        .type   ass_\name, %function
+        .func   ass_\name
+ass_\name:
+.endm
diff --git a/libass/arm/blend_bitmaps.S b/libass/arm/blend_bitmaps.S
new file mode 100644
index 0000000..c72d923
--- /dev/null
+++ b/libass/arm/blend_bitmaps.S
@@ -0,0 +1,67 @@
+/*
+ * void add_bitmaps( uint8_t *dst, intptr_t dst_stride,
+ *                   uint8_t *src, intptr_t src_stride,
+ *                   intptr_t width, intptr_t height );
+*/
+
+#include "asm.S"
+
+.macro blend_func type
+function \type\()_bitmaps_arm
+    push   {r4-r7}
+    ldrd   r4, r5, [sp, #16]
+@.Lskip_prologue:
+    mla    r5, r5, r3, r2
+0: @ height loop
+    mov    r6, #0 @ x offset
+1: @ stride loop
+    ldr    r7, [r0, r6]
+    ldr    ip, [r2, r6]
+    uq\type\()8 ip, r7, ip
+    str    ip, [r0, r6]
+    add    r6, #4
+    cmp    r6, r4
+    blo    1b @ still in scan line
+    add    r0, r1
+    add    r2, r3
+    cmp    r2, r5
+    blo    0b
+    pop    {r4-r7}
+    bx     lr
+endfunc
+.endm
+
+blend_func add
+blend_func sub
+
+/*
+ * void mul_bitmaps( uint8_t *dst, intptr_t dst_stride,
+ *                   uint8_t *src1, intptr_t src1_stride,
+ *                   uint8_t *src2, intptr_t src2_stride,
+ *                   intptr_t width, intptr_t height );
+*/
+function mul_bitmaps_arm
+    mov r11, #255
+    push   {r4-r11}
+    ldrd   r4, r5, [sp, #32]
+    ldrd   r6, r7, [sp, #40]
+    mla    r7, r7, r3, r2 @ last address
+0: @ height loop
+    mov    r8, #0 @ x offset
+1: @ stride loop
+    ldrb   r9,  [r2, r8]
+    ldrb   r10, [r4, r8]
+    mla    r9, r9, r10, r11
+    asr    r9, r9, #8
+    strb   r9,  [r0, r8]
+    add r8, r8, #1
+    cmp r8, r6
+    blo 1b @ still in scan line
+    add r0, r1
+    add r2, r3
+    add r4, r5
+    cmp r2, r7
+    blo 0b
+    pop    {r4-r11}
+    bx     lr
+endfunc
diff --git a/libass/ass_bitmap.c b/libass/ass_bitmap.c
index 73620dd..083b0e2 100644
--- a/libass/ass_bitmap.c
+++ b/libass/ass_bitmap.c
@@ -41,7 +41,8 @@
 #undef ALIGN
 #undef DECORATE
 
-#if (defined(__i386__) || defined(__x86_64__)) && CONFIG_ASM
+#if CONFIG_ASM
+#if (defined(__i386__) || defined(__x86_64__))
 
 #define ALIGN           4
 #define DECORATE(func)  ass_##func##_sse2
@@ -54,7 +55,16 @@
 #include "ass_func_template.h"
 #undef ALIGN
 #undef DECORATE
+#elif defined(__arm__)
 
+#define ALIGN           5
+#define DECORATE(func)  ass_##func##_arm // mostly not ported yet
+#define DECORATE2(func)  ass_##func##_arm // mostly not ported yet
+#include "ass_func_template.h"
+#undef ALIGN
+#undef DECORATE
+
+#endif
 #endif
 
 
diff --git a/libass/ass_bitmap.h b/libass/ass_bitmap.h
index e4c1451..84a2bc8 100644
--- a/libass/ass_bitmap.h
+++ b/libass/ass_bitmap.h
@@ -96,6 +96,7 @@ typedef struct {
 extern const BitmapEngine ass_bitmap_engine_c;
 extern const BitmapEngine ass_bitmap_engine_sse2;
 extern const BitmapEngine ass_bitmap_engine_avx2;
+extern const BitmapEngine ass_bitmap_engine_arm;
 
 
 typedef struct {
diff --git a/libass/ass_func_template.h b/libass/ass_func_template.h
index 0eccdfb..6f4755e 100644
--- a/libass/ass_func_template.h
+++ b/libass/ass_func_template.h
@@ -93,10 +93,18 @@ void DECORATE(blur1246_vert)(int16_t *dst, const int16_t *src,
                              uintptr_t src_width, uintptr_t src_height,
                              const int16_t *param);
 
-
 const BitmapEngine DECORATE(bitmap_engine) = {
     .align_order = ALIGN,
 
+#ifdef __arm__
+#undef DECORATE
+#define DECORATE(func) ass_##func##_c
+#endif
+
+#ifndef DECORATE2
+#define DECORATE2(x) DECORATE(x)
+#endif
+
 #if CONFIG_RASTERIZER
 #if CONFIG_LARGE_TILES
     .tile_order = 5,
@@ -111,10 +119,10 @@ const BitmapEngine DECORATE(bitmap_engine) = {
 #endif
 #endif
 
-    .add_bitmaps = DECORATE(add_bitmaps),
-#ifdef __x86_64__
-    .sub_bitmaps = DECORATE(sub_bitmaps),
-    .mul_bitmaps = DECORATE(mul_bitmaps),
+    .add_bitmaps = DECORATE2(add_bitmaps),
+#if defined(__x86_64__) || defined(__arm__)
+    .sub_bitmaps = DECORATE2(sub_bitmaps),
+    .mul_bitmaps = DECORATE2(mul_bitmaps),
 #else
     .sub_bitmaps = ass_sub_bitmaps_c,
     .mul_bitmaps = ass_mul_bitmaps_c,
@@ -139,3 +147,5 @@ const BitmapEngine DECORATE(bitmap_engine) = {
     .main_blur_horz = { DECORATE(blur1234_horz), DECORATE(blur1235_horz), DECORATE(blur1246_horz) },
     .main_blur_vert = { DECORATE(blur1234_vert), DECORATE(blur1235_vert), DECORATE(blur1246_vert) },
 };
+
+#undef DECORATE2
diff --git a/libass/ass_render.c b/libass/ass_render.c
index 77c7e30..42dbf6e 100644
--- a/libass/ass_render.c
+++ b/libass/ass_render.c
@@ -70,6 +70,8 @@ ASS_Renderer *ass_renderer_init(ASS_Library *library)
         priv->engine = &ass_bitmap_engine_sse2;
     else
         priv->engine = &ass_bitmap_engine_c;
+#elif defined(__arm__) && CONFIG_ASM
+    priv->engine = &ass_bitmap_engine_arm;
 #else
     priv->engine = &ass_bitmap_engine_c;
 #endif