ARM optimised fast_cmyk_to_rgb code.

1 files changed, 207 insertions, 15 deletions

diff --git a/fitz/res_colorspace.c b/fitz/res_colorspace.c
index 50c0a8f7..826909a0 100644
--- a/fitz/res_colorspace.c
+++ b/fitz/res_colorspace.c
@@ -2,6 +2,28 @@
 
 #define SLOWCMYK
 
+#ifdef NDK_PROFILER
+extern void __gnu_mcount_nc(void);
+#define ENTER_PG "push {lr}\nbl __gnu_mcount_nc\n"
+#else
+#define ENTER_PG
+#endif
+
+/* If we're compiling as thumb code, then we need to tell the compiler
+ * to enter and exit ARM mode around our assembly sections. If we move
+ * the ARM functions to a separate file and arrange for it to be compiled
+ * without thumb mode, we can save some time on entry.
+ */
+#ifdef ARCH_ARM
+#ifdef ARCH_THUMB
+#define ENTER_ARM ".balign 4\nmov r12,pc\nbx r12\n0:.arm\n" ENTER_PG
+#define ENTER_THUMB "9:.thumb\n" ENTER_PG
+#else
+#define ENTER_ARM
+#define ENTER_THUMB
+#endif
+#endif
+
 void
 fz_free_colorspace_imp(fz_context *ctx, fz_storable *cs_)
 {
@@ -316,26 +338,192 @@ static void fast_cmyk_to_gray(fz_pixmap *dst, fz_pixmap *src)
 	}
 }
 
-#if 0
+#ifdef ARCH_ARM
+static void
+fast_cmyk_to_rgb_ARM(unsigned char *dst, unsigned char *src, int n)
+__attribute__((naked));
+
 static void
 fast_cmyk_to_rgb_ARM(unsigned char *dst, unsigned char *src, int n)
 {
 	asm volatile(
 	ENTER_ARM
-	"stmfd	r13!,{r4-r12,r14}				\n"
-	"@ r0 = dst						\n"
-	"@ r1 = src						\n"
-	"@ r2 = weights						\n"
-	"mov	r4, #0			@ r4 = CMYK = 0		\n"
-	"mvn	r5, #0xFF000000		@ r5 = RGB = FFFFFF	\n"
-	"1:							\n"
-	"ldr	r3, [r1], #4		@ r3 = cmyk		\n"
-	"cmp	r3, r4			@ if (cmyk == CMYK)	\n"
-	"beq	match			@   goto match		\n"
-	"cmp	r3, #0			@ if (cmyk = 0000)	\n"
-	"beq	black			@
-
-	"ldmfd	r13!,{r4-r7,r9,PC}	@ pop, return to thumb	\n"
+	"stmfd	r13!,{r4-r11,r14}					\n"
+	"@ r0 = dst							\n"
+	"@ r1 = src							\n"
+	"@ r2 = n							\n"
+	"mov	r12, #0			@ r12= CMYK = 0			\n"
+	"b	2f			@ enter loop			\n"
+	"1:				@ White or Black		\n"
+	"@ Cunning trick: On entry r11 = 0 if black, r11 = FF if white	\n"
+	"ldrb	r7, [r1],#1		@ r8 = s[4]			\n"
+	"strb	r11,[r0],#1		@ d[0] = r			\n"
+	"strb	r11,[r0],#1		@ d[1] = g			\n"
+	"strb	r11,[r0],#1		@ d[2] = b			\n"
+	"strb	r7, [r0],#1		@ d[3] = s[4]			\n"
+	"subs	r2, r2, #1		@ r2 = n--			\n"
+	"beq	9f							\n"
+	"2:				@ Main loop starts here		\n"
+	"ldrb	r3, [r1], #4		@ r3 = c			\n"
+	"ldrb	r6, [r1, #-1]		@ r6 = k			\n"
+	"ldrb	r5, [r1, #-2]		@ r5 = y			\n"
+	"ldrb	r4, [r1, #-3]		@ r4 = m			\n"
+	"eors	r11,r6, #0xFF		@ if (k == 255)			\n"
+	"beq	1b			@   goto black			\n"
+	"orr	r7, r3, r4, LSL #8					\n"
+	"orr	r14,r5, r6, LSL #8					\n"
+	"orrs	r7, r7, r14,LSL #16	@ r7 = cmyk			\n"
+	"beq	1b			@ if (cmyk == 0) white		\n"
+	"@ At this point, we have to decode a new pixel			\n"
+	"@ r0 = dst  r1 = src  r2 = n  r7 = cmyk			\n"
+	"3:				@ unmatched			\n"
+	"stmfd	r13!,{r0-r1,r7}		@ stash regs for space		\n"
+	"add	r3, r3, r3, LSR #7	@ r3 = c += c>>7		\n"
+	"add	r4, r4, r4, LSR #7	@ r4 = m += m>>7		\n"
+	"add	r5, r5, r5, LSR #7	@ r5 = y += y>>7		\n"
+	"add	r6, r6, r6, LSR #7	@ r6 = k += k>>7		\n"
+	"mov	r5, r5, LSR #1		@ sacrifice 1 bit of Y		\n"
+	"mul	r8, r3, r4		@ r8 = cm     = c * m		\n"
+	"rsb	r9, r8, r4, LSL #8	@ r9 = c1m    = (m<<8) - cm	\n"
+	"rsb	r3, r8, r3, LSL #8	@ r3 = cm1    = (c<<8) - cm	\n"
+	"rsb	r4, r4, #0x100		@ r4 = 256-m			\n"
+	"rsb	r4, r3, r4, LSL #8	@ r4 = c1m1   =((256-m)<<8)-cm1	\n"
+	"mul	r7, r4, r5		@ r7 = c1m1y  = c1m1 * y	\n"
+	"rsb	r4, r7, r4, LSL #7	@ r4 = c1m1y1 = (c1m1<<7)-c1m1y	\n"
+	"mul	r10,r9, r5		@ r10= c1my   = c1m * y		\n"
+	"rsb	r9, r10,r9, LSL #7	@ r9 = c1my1  = (c1m<<7) - c1my \n"
+	"mul	r11,r3, r5		@ r11= cm1y   = cm1 * y		\n"
+	"rsb	r3, r11,r3, LSL #7	@ r3 = cm1y1  = (cm1<<7) - cm1y	\n"
+	"mul	r5, r8, r5		@ r5 = cmy    = cm * y		\n"
+	"rsb	r8, r5, r8, LSL #7	@ r8 = cmy1   = (cm<<7) - cmy	\n"
+	"@ Register recap:						\n"
+	"@ r3 = cm1y1							\n"
+	"@ r4 = c1m1y1							\n"
+	"@ r5 = cmy							\n"
+	"@ r6 = k							\n"
+	"@ r7 = c1m1y							\n"
+	"@ r8 = cmy1							\n"
+	"@ r9 = c1my1							\n"
+	"@ r10= c1my							\n"
+	"@ r11= cm1y							\n"
+	"@ The actual matrix multiplication				\n"
+	"mul	r14,r4, r6		@ r14= x1 = c1m1y1 * k		\n"
+	"rsb	r4, r14,r4, LSL #8	@ r4 = x0 = (c1m1y1<<8) - x1	\n"
+	"add	r4, r4, r14,LSR #8-5	@ r4 = b = x0 + 32*(x1>>8)	\n"
+	"sub	r1, r4, r14,LSR #8	@ r1 = g = x0 + 31*(x1>>8)	\n"
+	"add	r0, r1, r14,LSR #8-2	@ r0 = r = x0 + 35*(x1>>8)	\n"
+	"								\n"
+	"mul	r14,r7, r6		@ r14= x1 = c1m1y * k		\n"
+	"rsb	r7, r14,r7, LSL #8	@ r7 = x0 = (c1m1y<<8) - x1	\n"
+	"add	r0, r0, r7		@ r0 = r += x0			\n"
+	"add	r1, r1, r7		@ r1 = g += (x0>>8 * 256)	\n"
+	"sub	r1, r1, r7, LSR #8-3	@                    248	\n"
+	"sub	r1, r1, r7, LSR #8-2	@                    244	\n"
+	"sub	r1, r1, r7, LSR #8	@                    243	\n"
+	"sub	r7, r14,r14,LSR #3	@ r7 = 28*(x1>>5)		\n"
+	"add	r0, r0, r7, LSR #8-5	@ r0 = r += 28 * x1		\n"
+	"sub	r7, r7, r14,LSR #4	@ r7 = 26*(x1>>5)		\n"
+	"add	r1, r1, r7, LSR #8-5	@ r1 = g += 26 * x1		\n"
+	"								\n"
+	"mul	r14,r9, r6		@ r14= x1 = c1my1 * k		\n"
+	"sub	r9, r9, r14,LSR #8	@ r9 = x0>>8 = c1my1 - (x1>>8)	\n"
+	"add	r0, r0, r14,LSR #8-5	@ r0 = r += (x1>>8)*32		\n"
+	"add	r0, r0, r14,LSR #8-2	@ r0 = r += (x1>>8)*36		\n"
+	"mov	r14,#237		@ r14= 237			\n"
+	"mla	r0,r14,r9,r0		@ r14= r += x0*237		\n"
+	"mov	r14,#141		@ r14= 141			\n"
+	"mla	r4,r14,r9,r4		@ r14= b += x0*141		\n"
+	"								\n"
+	"mul	r14,r10,r6		@ r14= x1 = c1my * k		\n"
+	"sub	r10,r10,r14,LSR #8	@ r10= x0>>8 = c1my - (x1>>8)	\n"
+	"add	r0, r0, r14,LSR #8-5	@ r0 = r += 32 * x1		\n"
+	"add	r0, r0, r14,LSR #8-1	@ r0 = r += 34 * x1		\n"
+	"mov	r14,#238		@ r14= 238			\n"
+	"mla	r0,r14,r10,r0		@ r0 = r += 238 * x0		\n"
+	"mov	r14,#28			@ r14= 28			\n"
+	"mla	r1,r14,r10,r1		@ r1 = g += 28 * x0		\n"
+	"mov	r14,#36			@ r14= 36			\n"
+	"mla	r4,r14,r10,r4		@ r4 = b += 36 * x0		\n"
+	"								\n"
+	"mul	r14,r3, r6		@ r14= x1 = cm1y1 * k		\n"
+	"sub	r3, r3, r14,LSR #8	@ r3 = x1>>8 = cm1y1 - (x1>>8)	\n"
+	"add	r1, r1, r14,LSR #8-4	@ r1 = g += 16*x1		\n"
+	"sub	r1, r1, r14,LSR #8	@           15*x1		\n"
+	"add	r4, r4, r14,LSR #8-5	@ r4 = b += 32*x1		\n"
+	"add	r4, r4, r14,LSR #8-2	@           36*x1		\n"
+	"mov	r14,#174		@ r14= 174			\n"
+	"mla	r1, r14,r3, r1		@ r1 = g += 174 * x0		\n"
+	"mov	r14,#240		@ r14= 240			\n"
+	"mla	r4, r14,r3, r4		@ r4 = b += 240 * x0		\n"
+	"								\n"
+	"mul	r14,r11,r6		@ r14= x1 = cm1y * k		\n"
+	"sub	r11,r11,r14,LSR #8	@ r11= x0>>8 = cm1y - (x1>>8)	\n"
+	"add	r1, r1, r14,LSR #8-4	@ r1 = g += x1 * 16		\n"
+	"add	r1, r1, r14,LSR #8	@           x1 * 17		\n"
+	"add	r1, r1, r14,LSR #8-1	@           x1 * 19		\n"
+	"mov	r14,#167		@ r14 = 167			\n"
+	"mla	r1, r14,r11,r1		@ r1 = g += 167 * x0		\n"
+	"mov	r14,#80			@ r14 = 80			\n"
+	"mla	r4, r14,r11,r4		@ r4 = b += 80 * x0		\n"
+	"								\n"
+	"mul	r14,r8, r6		@ r14= x1 = cmy1 * k		\n"
+	"sub	r8, r8, r14,LSR #8	@ r8 = x0>>8 = cmy1 - (x1>>8)	\n"
+	"add	r4, r4, r14,LSR #8-1	@ r4 = b += x1 * 2		\n"
+	"mov	r14,#46			@ r14=46			\n"
+	"mla	r0, r14,r8, r0		@ r0 = r += 46 * x0		\n"
+	"mov	r14,#49			@ r14=49			\n"
+	"mla	r1, r14,r8, r1		@ r1 = g += 49 * x0		\n"
+	"mov	r14,#147		@ r14=147			\n"
+	"mla	r4, r14,r8, r4		@ r4 = b += 147 * x0		\n"
+	"								\n"
+	"rsb	r6, r6, #256		@ r6 = k = 256-k		\n"
+	"mul	r14,r5, r6		@ r14= x0 = cmy * (256-k)	\n"
+	"mov	r11,#54			@ r11= 54			\n"
+	"mov	r14,r14,LSR #8		@ r14= (x0>>8)			\n"
+	"mov	r8,#57			@ r8 = 57			\n"
+	"mla	r0,r14,r11,r0		@ r0 = r += 54*x0		\n"
+	"mla	r1,r14,r11,r1		@ r1 = g += 54*x0		\n"
+	"mla	r4,r14,r8, r4		@ r4 = b += 57*x0		\n"
+	"								\n"
+	"sub	r8, r0, r0, LSR #8	@ r8 = r -= (r>>8)		\n"
+	"sub	r9, r1, r1, LSR #8	@ r9 = g -= (r>>8)		\n"
+	"sub	r10,r4, r4, LSR #8	@ r10= b -= (r>>8)		\n"
+	"ldmfd	r13!,{r0-r1,r12}					\n"
+	"mov	r8, r8, LSR #23		@ r8 = r>>23			\n"
+	"mov	r9, r9, LSR #23		@ r9 = g>>23			\n"
+	"mov	r10,r10,LSR #23		@ r10= b>>23			\n"
+	"ldrb	r14,[r1],#1		@ r8 = s[4]			\n"
+	"strb	r8, [r0],#1		@ d[0] = r			\n"
+	"strb	r9, [r0],#1		@ d[1] = g			\n"
+	"strb	r10,[r0],#1		@ d[2] = b			\n"
+	"strb	r14,[r0],#1		@ d[3] = s[4]			\n"
+	"subs	r2, r2, #1		@ r2 = n--			\n"
+	"beq	9f							\n"
+	"@ At this point, we've just decoded a pixel			\n"
+	"@ r0 = dst  r1 = src  r2 = n  r8 = r  r9 = g  r10= b r12= CMYK \n"
+	"4:								\n"
+	"ldrb	r3, [r1], #4		@ r3 = c			\n"
+	"ldrb	r6, [r1, #-1]		@ r6 = k			\n"
+	"ldrb	r5, [r1, #-2]		@ r5 = y			\n"
+	"ldrb	r4, [r1, #-3]		@ r4 = m			\n"
+	"eors	r11,r6, #0xFF		@ if (k == 255)			\n"
+	"beq	1b			@   goto black			\n"
+	"orr	r7, r3, r4, LSL #8					\n"
+	"orr	r14,r5, r6, LSL #8					\n"
+	"orrs	r7, r7, r14,LSL #16	@ r7 = cmyk			\n"
+	"beq	1b			@ if (cmyk == 0) white		\n"
+	"cmp	r7, r12			@ if (cmyk != CMYK)		\n"
+	"bne	3b			@   not the same, loop		\n"
+	"@ If we get here, we just matched a pixel we have just decoded \n"
+	"ldrb	r3, [r1],#1		@ r8 = s[4]			\n"
+	"strb	r8, [r0],#1		@ d[0] = r			\n"
+	"strb	r9, [r0],#1		@ d[1] = g			\n"
+	"strb	r10,[r0],#1		@ d[2] = b			\n"
+	"strb	r3, [r0],#1		@ d[3] = s[4]			\n"
+	"subs	r2, r2, #1		@ r2 = n--			\n"
+	"bne	4b							\n"
+	"9:								\n"
+	"ldmfd	r13!,{r4-r11,PC}	@ pop, return to thumb		\n"
 	ENTER_THUMB
 	);
 }
@@ -346,6 +534,9 @@ static void fast_cmyk_to_rgb(fz_context *ctx, fz_pixmap *dst, fz_pixmap *src)
 	unsigned char *s = src->samples;
 	unsigned char *d = dst->samples;
 	int n = src->w * src->h;
+#ifdef ARCH_ARM
+	fast_cmyk_to_rgb_ARM(d, s, n);
+#else
 	unsigned int C,M,Y,K,r,g,b;
 
 	C = 0;
@@ -488,6 +679,7 @@ static void fast_cmyk_to_rgb(fz_context *ctx, fz_pixmap *dst, fz_pixmap *src)
 		s += 5;
 		d += 4;
 	}
+#endif
 }
 
 static void fast_cmyk_to_bgr(fz_context *ctx, fz_pixmap *dst, fz_pixmap *src)