1/* NEON optimized code (C) COPYRIGHT 2009 Motorola
2 *
3 * Use of this source code is governed by a BSD-style license that can be
4 * found in the LICENSE file.
5 */
6
7#include "SkBitmapProcState.h"
8#include "SkPerspIter.h"
9#include "SkShader.h"
10#include "SkUtilsArm.h"
11
12extern const SkBitmapProcState::MatrixProc ClampX_ClampY_Procs_neon[];
13extern const SkBitmapProcState::MatrixProc RepeatX_RepeatY_Procs_neon[];
14
15static void decal_nofilter_scale_neon(uint32_t dst[], SkFixed fx, SkFixed dx, int count);
16static void decal_filter_scale_neon(uint32_t dst[], SkFixed fx, SkFixed dx, int count);
17
18static unsigned SK_USHIFT16(unsigned x) {
19    return x >> 16;
20}
21
22#define MAKENAME(suffix)        ClampX_ClampY ## suffix ## _neon
23#define TILEX_PROCF(fx, max)    SkClampMax((fx) >> 16, max)
24#define TILEY_PROCF(fy, max)    SkClampMax((fy) >> 16, max)
25#define TILEX_LOW_BITS(fx, max) (((fx) >> 12) & 0xF)
26#define TILEY_LOW_BITS(fy, max) (((fy) >> 12) & 0xF)
27#define CHECK_FOR_DECAL
28#include "SkBitmapProcState_matrix_clamp_neon.h"
29
30#define MAKENAME(suffix)        RepeatX_RepeatY ## suffix ## _neon
31#define TILEX_PROCF(fx, max)    SK_USHIFT16(((fx) & 0xFFFF) * ((max) + 1))
32#define TILEY_PROCF(fy, max)    SK_USHIFT16(((fy) & 0xFFFF) * ((max) + 1))
33#define TILEX_LOW_BITS(fx, max) ((((fx) & 0xFFFF) * ((max) + 1) >> 12) & 0xF)
34#define TILEY_LOW_BITS(fy, max) ((((fy) & 0xFFFF) * ((max) + 1) >> 12) & 0xF)
35#include "SkBitmapProcState_matrix_repeat_neon.h"
36
37
38void decal_nofilter_scale_neon(uint32_t dst[], SkFixed fx, SkFixed dx, int count)
39{
40    int i;
41
42    if (count >= 8) {
43        /* SkFixed is 16.16 fixed point */
44        SkFixed dx2 = dx+dx;
45        SkFixed dx4 = dx2+dx2;
46        SkFixed dx8 = dx4+dx4;
47
48        /* now build fx/fx+dx/fx+2dx/fx+3dx */
49        SkFixed fx1, fx2, fx3;
50        int32x4_t lbase, hbase;
51        uint16_t *dst16 = (uint16_t *)dst;
52
53        fx1 = fx+dx;
54        fx2 = fx1+dx;
55        fx3 = fx2+dx;
56
57        /* avoid an 'lbase unitialized' warning */
58        lbase = vdupq_n_s32(fx);
59        lbase = vsetq_lane_s32(fx1, lbase, 1);
60        lbase = vsetq_lane_s32(fx2, lbase, 2);
61        lbase = vsetq_lane_s32(fx3, lbase, 3);
62        hbase = vaddq_s32(lbase, vdupq_n_s32(dx4));
63
64        /* take upper 16 of each, store, and bump everything */
65        do {
66            int32x4_t lout, hout;
67            uint16x8_t hi16;
68
69            lout = lbase;
70            hout = hbase;
71            /* gets hi's of all louts then hi's of all houts */
72            asm ("vuzpq.16 %q0, %q1" : "+w" (lout), "+w" (hout));
73            hi16 = vreinterpretq_u16_s32(hout);
74            vst1q_u16(dst16, hi16);
75
76            /* on to the next */
77            lbase = vaddq_s32 (lbase, vdupq_n_s32(dx8));
78            hbase = vaddq_s32 (hbase, vdupq_n_s32(dx8));
79            dst16 += 8;
80            count -= 8;
81            fx += dx8;
82        } while (count >= 8);
83        dst = (uint32_t *) dst16;
84    }
85
86    uint16_t* xx = (uint16_t*)dst;
87    for (i = count; i > 0; --i) {
88        *xx++ = SkToU16(fx >> 16); fx += dx;
89    }
90}
91
92void decal_filter_scale_neon(uint32_t dst[], SkFixed fx, SkFixed dx, int count)
93{
94    if (count >= 8) {
95        int32x4_t wide_fx;
96        int32x4_t wide_fx2;
97        int32x4_t wide_dx8 = vdupq_n_s32(dx*8);
98
99        wide_fx = vdupq_n_s32(fx);
100        wide_fx = vsetq_lane_s32(fx+dx, wide_fx, 1);
101        wide_fx = vsetq_lane_s32(fx+dx+dx, wide_fx, 2);
102        wide_fx = vsetq_lane_s32(fx+dx+dx+dx, wide_fx, 3);
103
104        wide_fx2 = vaddq_s32(wide_fx, vdupq_n_s32(dx+dx+dx+dx));
105
106        while (count >= 8) {
107            int32x4_t wide_out;
108            int32x4_t wide_out2;
109
110            wide_out = vshlq_n_s32(vshrq_n_s32(wide_fx, 12), 14);
111            wide_out = vorrq_s32(wide_out,
112            vaddq_s32(vshrq_n_s32(wide_fx,16), vdupq_n_s32(1)));
113
114            wide_out2 = vshlq_n_s32(vshrq_n_s32(wide_fx2, 12), 14);
115            wide_out2 = vorrq_s32(wide_out2,
116            vaddq_s32(vshrq_n_s32(wide_fx2,16), vdupq_n_s32(1)));
117
118            vst1q_u32(dst, vreinterpretq_u32_s32(wide_out));
119            vst1q_u32(dst+4, vreinterpretq_u32_s32(wide_out2));
120
121            dst += 8;
122            fx += dx*8;
123            wide_fx = vaddq_s32(wide_fx, wide_dx8);
124            wide_fx2 = vaddq_s32(wide_fx2, wide_dx8);
125            count -= 8;
126        }
127    }
128
129    if (count & 1)
130    {
131        SkASSERT((fx >> (16 + 14)) == 0);
132        *dst++ = (fx >> 12 << 14) | ((fx >> 16) + 1);
133        fx += dx;
134    }
135    while ((count -= 2) >= 0)
136    {
137        SkASSERT((fx >> (16 + 14)) == 0);
138        *dst++ = (fx >> 12 << 14) | ((fx >> 16) + 1);
139        fx += dx;
140
141        *dst++ = (fx >> 12 << 14) | ((fx >> 16) + 1);
142        fx += dx;
143    }
144}
145