From 89cbe6eb5df2a1c85ba996caea6479e2434d51a5 Mon Sep 17 00:00:00 2001 From: Siarhei Siamashka Date: Thu, 23 Sep 2010 23:09:46 +0300 Subject: [PATCH 8/9] ARM: added 'neon_composite_out_reverse_0565_8_0565' fast path TODO: That's an initial variant, needs performance tuning --- pixman/pixman-arm-neon-asm.S | 66 ++++++++++++++++++++++++++++++++++++++++++ pixman/pixman-arm-neon.c | 4 ++ 2 files changed, 70 insertions(+), 0 deletions(-) diff --git a/pixman/pixman-arm-neon-asm.S b/pixman/pixman-arm-neon-asm.S index 23ddae9..7a599f0 100644 --- a/pixman/pixman-arm-neon-asm.S +++ b/pixman/pixman-arm-neon-asm.S @@ -1942,3 +1942,69 @@ generate_composite_function \ 10, /* dst_r_basereg */ \ 8, /* src_basereg */ \ 15 /* mask_basereg */ + +/******************************************************************************/ + +.macro pixman_composite_out_reverse_0565_8_0565_process_pixblock_head + /* mask is in d15 */ + convert_0565_to_8888 q4, d3, d2, d1, d0 + convert_0565_to_x888 q5, d6, d5, d4 + /* source pixel data is in {d0, d1, d2, d3} */ + /* destination pixel data is in {d4, d5, d6, xx} */ + + /* 'in' */ + vmull.u8 q8, d15, d3 + vmull.u8 q6, d15, d2 + vmull.u8 q5, d15, d1 + vmull.u8 q4, d15, d0 + vrshr.u16 q13, q8, #8 + vrshr.u16 q12, q6, #8 + vrshr.u16 q11, q5, #8 + vrshr.u16 q10, q4, #8 + vraddhn.u16 d3, q8, q13 + vraddhn.u16 d2, q6, q12 + vraddhn.u16 d1, q5, q11 + vraddhn.u16 d0, q4, q10 + vmvn.8 d24, d3 /* 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_0565_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_0565_8_0565_process_pixblock_tail_head + vld1.8 {d15}, [MASK]! + pixman_composite_out_reverse_0565_8_0565_process_pixblock_tail + vld1.16 {d8, d9}, [SRC]! + vld1.16 {d10, d11}, [DST_R, :128]! + cache_preload 8, 8 + pixman_composite_out_reverse_0565_8_0565_process_pixblock_head + vst1.16 {d28, d29}, [DST_W, :128]! +.endm + +generate_composite_function \ + pixman_composite_out_reverse_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_out_reverse_0565_8_0565_process_pixblock_head, \ + pixman_composite_out_reverse_0565_8_0565_process_pixblock_tail, \ + pixman_composite_out_reverse_0565_8_0565_process_pixblock_tail_head, \ + 28, /* dst_w_basereg */ \ + 10, /* dst_r_basereg */ \ + 8, /* src_basereg */ \ + 15 /* mask_basereg */ diff --git a/pixman/pixman-arm-neon.c b/pixman/pixman-arm-neon.c index a8be7e4..da1fdeb 100644 --- a/pixman/pixman-arm-neon.c +++ b/pixman/pixman-arm-neon.c @@ -94,6 +94,8 @@ PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_8888_8_0565, uint32_t, 1, uint8_t, 1, uint16_t, 1) PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_0565_8_0565, uint16_t, 1, uint8_t, 1, uint16_t, 1) +PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, out_reverse_0565_8_0565, + uint16_t, 1, uint8_t, 1, uint16_t, 1) void pixman_composite_src_n_8_asm_neon (int32_t w, @@ -306,6 +308,8 @@ static const pixman_fast_path_t arm_neon_fast_paths[] = PIXMAN_STD_FAST_PATH (ADD, a8b8g8r8, null, a8b8g8r8, neon_composite_add_8888_8888), PIXMAN_STD_FAST_PATH (OVER_REVERSE, solid, null, a8r8g8b8, neon_composite_over_reverse_n_8888), PIXMAN_STD_FAST_PATH (OVER_REVERSE, solid, null, a8b8g8r8, neon_composite_over_reverse_n_8888), + PIXMAN_STD_FAST_PATH (OUT_REVERSE, r5g6b5, a8, r5g6b5, neon_composite_out_reverse_0565_8_0565), + PIXMAN_STD_FAST_PATH (OUT_REVERSE, b5g6r5, a8, b5g6r5, neon_composite_out_reverse_0565_8_0565), { PIXMAN_OP_NONE }, }; -- 1.6.6.1