arm_neon: fix the color conversion to use the actual pitch of the pictures...

arm_neon: fix the color conversion to use the actual pitch of the pictures instead of trying to guess it.

modules/arm_neon/i420_yuy2.c

@@ -36,16 +36,18 @@ vlc_module_begin ()
 vlc_module_end ()
 
 void i420_yuyv_neon (uint8_t *out, const uint8_t **in,
-                     uintptr_t pitch, uintptr_t height);
+                     unsigned int pitch, unsigned int s_off,
+                     unsigned int height);
 
 static void I420_YUYV (filter_t *filter, picture_t *src, picture_t *dst)
 {
     uint8_t *out = dst->p->p_pixels;
     const uint8_t *yuv[3] = { src->Y_PIXELS, src->U_PIXELS, src->V_PIXELS, };
-    size_t pitch = (filter->fmt_in.video.i_width + 15) & ~15;
     size_t height = filter->fmt_in.video.i_height;
+    int i_pitch = (dst->p->i_pitch >> 1) & ~0xF;
+    int s_offset = src->p->i_pitch - i_pitch;
 
-    i420_yuyv_neon (out, yuv, pitch, height);
+    i420_yuyv_neon (out, yuv, i_pitch, s_offset, height);
 }
 
 void i420_uyvy_neon (uint8_t *out, const uint8_t **in,
@@ -55,10 +57,11 @@ static void I420_UYVY (filter_t *filter, picture_t *src, picture_t *dst)
 {
     uint8_t *out = dst->p->p_pixels;
     const uint8_t *yuv[3] = { src->Y_PIXELS, src->U_PIXELS, src->V_PIXELS, };
-    size_t pitch = (filter->fmt_in.video.i_width + 15) & ~15;
     size_t height = filter->fmt_in.video.i_height;
+    int i_pitch = (dst->p->i_pitch >> 1) & ~0xF;
+    int s_offset = src->p->i_pitch - i_pitch;
 
-    i420_yuyv_neon (out, yuv, pitch, height);
+    i420_yuyv_neon (out, yuv, i_pitch, s_offset, height);
 }
 
 VIDEO_FILTER_WRAPPER (I420_YUYV)

modules/arm_neon/i420_yuyv.S

@@ -24,21 +24,24 @@
 #define O1	r0
 #define	O2	r1
 #define	PITCH	r2
-#define	HEIGHT	r3
+#define	S_OFF	r3
 #define	Y1	r4
 #define	Y2	r5
 #define	U	r6
 #define	V	r7
+#define	HEIGHT	r8
 #define	END_O1	r12
 
 	.align
 	.global i420_yuyv_neon
 	.type	i420_yuyv_neon, %function
 i420_yuyv_neon:
-	push		{r4-r7, lr}
+	push		{r4-r8, lr}
+	ldr		HEIGHT, [sp, #(4*6)]
 	ldmia		r1,	{Y1, U, V}
 	add		O2,	O1,	PITCH, lsl #1
 	add		Y2,	Y1,	PITCH
+	add		Y2,	S_OFF
 1:
 	mov		END_O1,	O2
 	pld		[Y2]
@@ -64,21 +67,27 @@ i420_yuyv_neon:
 	sub		HEIGHT,	#2
 	mov		O1,	O2
 	add		O2,	PITCH,	lsl #1
+	add		Y2,	S_OFF
 	mov		Y1,	Y2
 	add		Y2,	PITCH
+	add		Y2,	S_OFF
+	add		U,	S_OFF,	lsr #1
+	add		V,	S_OFF,	lsr #1
 
 	cmp		HEIGHT,	#0
 	bne		1b
 
-	pop		{r4-r7, pc}
+	pop		{r4-r8, pc}
 
 	.global i420_uyvy_neon
 	.type	i420_uyvy_neon, %function
 i420_uyvy_neon:
-	push		{r4-r7, lr}
+	push		{r4-r8, lr}
+	ldr		HEIGHT, [sp, #(4*6)]
 	ldmia		r1,	{Y1, U, V}
 	add		O2,	O1,	PITCH, lsl #1
 	add		Y2,	Y1,	PITCH
+	add		Y2,	S_OFF
 1:
 	mov		END_O1,	O2
 2:
@@ -103,10 +112,14 @@ i420_uyvy_neon:
 	sub		HEIGHT,	#2
 	mov		O1,	O2
 	add		O2,	PITCH,	lsl #1
+	add		Y2,	S_OFF
 	mov		Y1,	Y2
 	add		Y2,	PITCH
+	add		Y2,	S_OFF
+	add		U,	S_OFF,	lsr #1
+	add		V,	S_OFF,	lsr #1
 
 	cmp		HEIGHT,	#0
 	bne		1b
 
-	pop		{r4-r7, pc}
+	pop		{r4-r8, pc}