aboutsummaryrefslogtreecommitdiff
path: root/pixman/pixman/pixman-arm-neon-asm.S
diff options
context:
space:
mode:
Diffstat (limited to 'pixman/pixman/pixman-arm-neon-asm.S')
-rw-r--r--pixman/pixman/pixman-arm-neon-asm.S114
1 files changed, 72 insertions, 42 deletions
diff --git a/pixman/pixman/pixman-arm-neon-asm.S b/pixman/pixman/pixman-arm-neon-asm.S
index c2a8bd6ea..fe128aa94 100644
--- a/pixman/pixman/pixman-arm-neon-asm.S
+++ b/pixman/pixman/pixman-arm-neon-asm.S
@@ -892,21 +892,13 @@ generate_composite_function \
vst1.16 {d28, d29}, [DST_W, :128]!
.endm
-.macro pixman_composite_over_8888_8_0565_init
- vpush {d8-d15}
-.endm
-
-.macro pixman_composite_over_8888_8_0565_cleanup
- vpop {d8-d15}
-.endm
-
generate_composite_function \
pixman_composite_over_8888_8_0565_asm_neon, 32, 8, 16, \
FLAG_DST_READWRITE | FLAG_DEINTERLEAVE_32BPP, \
8, /* number of pixels, processed in a single block */ \
5, /* prefetch distance */ \
- pixman_composite_over_8888_8_0565_init, \
- pixman_composite_over_8888_8_0565_cleanup, \
+ default_init_need_all_regs, \
+ default_cleanup_need_all_regs, \
pixman_composite_over_n_8_0565_process_pixblock_head, \
pixman_composite_over_n_8_0565_process_pixblock_tail, \
pixman_composite_over_8888_8_0565_process_pixblock_tail_head, \
@@ -1519,14 +1511,6 @@ generate_composite_function_single_scanline \
vraddhn.u16 d31, q13, q11
.endm
-.macro pixman_composite_out_reverse_8888_8888_8888_init
- vpush {d8-d15}
-.endm
-
-.macro pixman_composite_out_reverse_8888_8888_8888_cleanup
- vpop {d8-d15}
-.endm
-
/* TODO: expand macros and do better instructions scheduling */
.macro pixman_composite_out_reverse_8888_8888_8888_process_pixblock_tail_head
vld4.8 {d4, d5, d6, d7}, [DST_R, :128]!
@@ -1542,8 +1526,8 @@ generate_composite_function_single_scanline \
pixman_composite_scanline_out_reverse_mask_asm_neon, 32, 32, 32, \
FLAG_DST_READWRITE | FLAG_DEINTERLEAVE_32BPP, \
8, /* number of pixels, processed in a single block */ \
- pixman_composite_out_reverse_8888_8888_8888_init, \
- pixman_composite_out_reverse_8888_8888_8888_cleanup, \
+ default_init_need_all_regs, \
+ default_cleanup_need_all_regs, \
pixman_composite_out_reverse_8888_n_8888_process_pixblock_head, \
pixman_composite_out_reverse_8888_n_8888_process_pixblock_tail, \
pixman_composite_out_reverse_8888_8888_8888_process_pixblock_tail_head \
@@ -1609,21 +1593,13 @@ generate_composite_function \
vst4.8 {d28, d29, d30, d31}, [DST_W, :128]!
.endm
-.macro pixman_composite_over_8888_8888_8888_init
- vpush {d8-d15}
-.endm
-
-.macro pixman_composite_over_8888_8888_8888_cleanup
- vpop {d8-d15}
-.endm
-
generate_composite_function \
pixman_composite_over_8888_8888_8888_asm_neon, 32, 32, 32, \
FLAG_DST_READWRITE | FLAG_DEINTERLEAVE_32BPP, \
8, /* number of pixels, processed in a single block */ \
5, /* prefetch distance */ \
- pixman_composite_over_8888_8888_8888_init, \
- pixman_composite_over_8888_8888_8888_cleanup, \
+ default_init_need_all_regs, \
+ default_cleanup_need_all_regs, \
pixman_composite_over_8888_n_8888_process_pixblock_head, \
pixman_composite_over_8888_n_8888_process_pixblock_tail, \
pixman_composite_over_8888_8888_8888_process_pixblock_tail_head \
@@ -1636,8 +1612,8 @@ generate_composite_function_single_scanline \
pixman_composite_scanline_over_mask_asm_neon, 32, 32, 32, \
FLAG_DST_READWRITE | FLAG_DEINTERLEAVE_32BPP, \
8, /* number of pixels, processed in a single block */ \
- pixman_composite_over_8888_8888_8888_init, \
- pixman_composite_over_8888_8888_8888_cleanup, \
+ default_init_need_all_regs, \
+ default_cleanup_need_all_regs, \
pixman_composite_over_8888_n_8888_process_pixblock_head, \
pixman_composite_over_8888_n_8888_process_pixblock_tail, \
pixman_composite_over_8888_8888_8888_process_pixblock_tail_head \
@@ -1659,21 +1635,13 @@ generate_composite_function_single_scanline \
vst4.8 {d28, d29, d30, d31}, [DST_W, :128]!
.endm
-.macro pixman_composite_over_8888_8_8888_init
- vpush {d8-d15}
-.endm
-
-.macro pixman_composite_over_8888_8_8888_cleanup
- vpop {d8-d15}
-.endm
-
generate_composite_function \
pixman_composite_over_8888_8_8888_asm_neon, 32, 8, 32, \
FLAG_DST_READWRITE | FLAG_DEINTERLEAVE_32BPP, \
8, /* number of pixels, processed in a single block */ \
5, /* prefetch distance */ \
- pixman_composite_over_8888_8_8888_init, \
- pixman_composite_over_8888_8_8888_cleanup, \
+ default_init_need_all_regs, \
+ default_cleanup_need_all_regs, \
pixman_composite_over_8888_n_8888_process_pixblock_head, \
pixman_composite_over_8888_n_8888_process_pixblock_tail, \
pixman_composite_over_8888_8_8888_process_pixblock_tail_head \
@@ -1840,3 +1808,65 @@ generate_composite_function \
0, /* dst_r_basereg */ \
0, /* src_basereg */ \
0 /* mask_basereg */
+
+/******************************************************************************/
+
+.macro pixman_composite_over_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} */
+ vmvn.8 d7, d15
+ vmull.u8 q6, d15, d2
+ vmull.u8 q5, d15, d1
+ vmull.u8 q4, d15, d0
+ vmull.u8 q8, d7, d4
+ vmull.u8 q9, d7, d5
+ vmull.u8 q13, d7, d6
+ 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_over_0565_8_0565_process_pixblock_tail
+ vrshr.u16 q14, q8, #8
+ vrshr.u16 q15, q9, #8
+ vrshr.u16 q12, q13, #8
+ vraddhn.u16 d28, q14, q8
+ vraddhn.u16 d29, q15, q9
+ vraddhn.u16 d30, q12, q13
+ vqadd.u8 q0, q0, q14
+ vqadd.u8 q1, q1, q15
+ /* 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_over_0565_8_0565_process_pixblock_tail_head
+ vld1.8 {d15}, [MASK]!
+ pixman_composite_over_0565_8_0565_process_pixblock_tail
+ vld1.16 {d8, d9}, [SRC]!
+ vld1.16 {d10, d11}, [DST_R, :128]!
+ cache_preload 8, 8
+ pixman_composite_over_0565_8_0565_process_pixblock_head
+ vst1.16 {d28, d29}, [DST_W, :128]!
+.endm
+
+generate_composite_function \
+ pixman_composite_over_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_over_0565_8_0565_process_pixblock_head, \
+ pixman_composite_over_0565_8_0565_process_pixblock_tail, \
+ pixman_composite_over_0565_8_0565_process_pixblock_tail_head, \
+ 28, /* dst_w_basereg */ \
+ 10, /* dst_r_basereg */ \
+ 8, /* src_basereg */ \
+ 15 /* mask_basereg */
generated by cgit v1.2.3 (git 2.25.1) at 2024-07-14 04:50:16 +0000