diff options
Diffstat (limited to 'pixman/pixman/pixman-arm-neon-asm.S')
| -rw-r--r-- | pixman/pixman/pixman-arm-neon-asm.S | 103 |
1 files changed, 103 insertions, 0 deletions
diff --git a/pixman/pixman/pixman-arm-neon-asm.S b/pixman/pixman/pixman-arm-neon-asm.S index 108abacd1..8ddbefc87 100644 --- a/pixman/pixman/pixman-arm-neon-asm.S +++ b/pixman/pixman/pixman-arm-neon-asm.S @@ -388,6 +388,7 @@ generate_composite_function \ vld1.16 {d4, d5}, [DST_R, :128]!
vst1.16 {d28, d29}, [DST_W, :128]!
pixman_composite_over_n_0565_process_pixblock_head
+ cache_preload 8, 8
.endm
.macro pixman_composite_over_n_0565_init
@@ -710,6 +711,7 @@ generate_composite_function_single_scanline \ vld4.8 {d4, d5, d6, d7}, [DST_R, :128]!
vst4.8 {d28, d29, d30, d31}, [DST_W, :128]!
pixman_composite_over_8888_8888_process_pixblock_head
+ cache_preload 8, 8
.endm
.macro pixman_composite_over_n_8888_init
@@ -1870,3 +1872,104 @@ generate_composite_function \ 10, /* dst_r_basereg */ \
8, /* src_basereg */ \
15 /* mask_basereg */
+
+/******************************************************************************/
+
+.macro pixman_composite_add_0565_8_0565_process_pixblock_head
+ /* mask is in d15 */
+ convert_0565_to_x888 q4, d2, d1, d0
+ convert_0565_to_x888 q5, d6, d5, d4
+ /* source pixel data is in {d0, d1, d2, XX} */
+ /* destination pixel data is in {d4, d5, d6, XX} */
+ vmull.u8 q6, d15, d2
+ vmull.u8 q5, d15, d1
+ vmull.u8 q4, d15, d0
+ vrshr.u16 q12, q6, #8
+ vrshr.u16 q11, q5, #8
+ vrshr.u16 q10, q4, #8
+ vraddhn.u16 d2, q6, q12
+ vraddhn.u16 d1, q5, q11
+ vraddhn.u16 d0, q4, q10
+.endm
+
+.macro pixman_composite_add_0565_8_0565_process_pixblock_tail
+ vqadd.u8 q0, q0, q2
+ vqadd.u8 q1, q1, q3
+ /* 32bpp result is in {d0, d1, d2, XX} */
+ convert_8888_to_0565 d2, d1, d0, q14, q15, q3
+.endm
+
+/* TODO: expand macros and do better instructions scheduling */
+.macro pixman_composite_add_0565_8_0565_process_pixblock_tail_head
+ vld1.8 {d15}, [MASK]!
+ pixman_composite_add_0565_8_0565_process_pixblock_tail
+ vld1.16 {d8, d9}, [SRC]!
+ vld1.16 {d10, d11}, [DST_R, :128]!
+ cache_preload 8, 8
+ pixman_composite_add_0565_8_0565_process_pixblock_head
+ vst1.16 {d28, d29}, [DST_W, :128]!
+.endm
+
+generate_composite_function \
+ pixman_composite_add_0565_8_0565_asm_neon, 16, 8, 16, \
+ FLAG_DST_READWRITE, \
+ 8, /* number of pixels, processed in a single block */ \
+ 5, /* prefetch distance */ \
+ default_init_need_all_regs, \
+ default_cleanup_need_all_regs, \
+ pixman_composite_add_0565_8_0565_process_pixblock_head, \
+ pixman_composite_add_0565_8_0565_process_pixblock_tail, \
+ pixman_composite_add_0565_8_0565_process_pixblock_tail_head, \
+ 28, /* dst_w_basereg */ \
+ 10, /* dst_r_basereg */ \
+ 8, /* src_basereg */ \
+ 15 /* mask_basereg */
+
+/******************************************************************************/
+
+.macro pixman_composite_out_reverse_8_0565_process_pixblock_head
+ /* mask is in d15 */
+ convert_0565_to_x888 q5, d6, d5, d4
+ /* destination pixel data is in {d4, d5, d6, xx} */
+ vmvn.8 d24, d15 /* get inverted alpha */
+ /* now do alpha blending */
+ vmull.u8 q8, d24, d4
+ vmull.u8 q9, d24, d5
+ vmull.u8 q10, d24, d6
+.endm
+
+.macro pixman_composite_out_reverse_8_0565_process_pixblock_tail
+ vrshr.u16 q14, q8, #8
+ vrshr.u16 q15, q9, #8
+ vrshr.u16 q12, q10, #8
+ vraddhn.u16 d0, q14, q8
+ vraddhn.u16 d1, q15, q9
+ vraddhn.u16 d2, q12, q10
+ /* 32bpp result is in {d0, d1, d2, XX} */
+ convert_8888_to_0565 d2, d1, d0, q14, q15, q3
+.endm
+
+/* TODO: expand macros and do better instructions scheduling */
+.macro pixman_composite_out_reverse_8_0565_process_pixblock_tail_head
+ vld1.8 {d15}, [SRC]!
+ pixman_composite_out_reverse_8_0565_process_pixblock_tail
+ vld1.16 {d10, d11}, [DST_R, :128]!
+ cache_preload 8, 8
+ pixman_composite_out_reverse_8_0565_process_pixblock_head
+ vst1.16 {d28, d29}, [DST_W, :128]!
+.endm
+
+generate_composite_function \
+ pixman_composite_out_reverse_8_0565_asm_neon, 8, 0, 16, \
+ FLAG_DST_READWRITE, \
+ 8, /* number of pixels, processed in a single block */ \
+ 5, /* prefetch distance */ \
+ default_init_need_all_regs, \
+ default_cleanup_need_all_regs, \
+ pixman_composite_out_reverse_8_0565_process_pixblock_head, \
+ pixman_composite_out_reverse_8_0565_process_pixblock_tail, \
+ pixman_composite_out_reverse_8_0565_process_pixblock_tail_head, \
+ 28, /* dst_w_basereg */ \
+ 10, /* dst_r_basereg */ \
+ 15, /* src_basereg */ \
+ 0 /* mask_basereg */
|