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 12 extern const SkBitmapProcState::MatrixProc ClampX_ClampY_Procs_neon[]; 13 extern const SkBitmapProcState::MatrixProc RepeatX_RepeatY_Procs_neon[]; 14 15 static void decal_nofilter_scale_neon(uint32_t dst[], SkFixed fx, SkFixed dx, int count); 16 static void decal_filter_scale_neon(uint32_t dst[], SkFixed fx, SkFixed dx, int count); 17 18 static 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 38 void 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 int32x2_t lower, upper; 51 int32x4_t lbase, hbase; 52 uint16_t *dst16 = (uint16_t *)dst; 53 54 fx1 = fx+dx; 55 fx2 = fx1+dx; 56 fx3 = fx2+dx; 57 58 /* avoid an 'lbase unitialized' warning */ 59 lbase = vdupq_n_s32(fx); 60 lbase = vsetq_lane_s32(fx1, lbase, 1); 61 lbase = vsetq_lane_s32(fx2, lbase, 2); 62 lbase = vsetq_lane_s32(fx3, lbase, 3); 63 hbase = vaddq_s32(lbase, vdupq_n_s32(dx4)); 64 65 /* take upper 16 of each, store, and bump everything */ 66 do { 67 int32x4_t lout, hout; 68 uint16x8_t hi16; 69 70 lout = lbase; 71 hout = hbase; 72 /* gets hi's of all louts then hi's of all houts */ 73 asm ("vuzpq.16 %q0, %q1" : "+w" (lout), "+w" (hout)); 74 hi16 = vreinterpretq_u16_s32(hout); 75 vst1q_u16(dst16, hi16); 76 77 /* on to the next */ 78 lbase = vaddq_s32 (lbase, vdupq_n_s32(dx8)); 79 hbase = vaddq_s32 (hbase, vdupq_n_s32(dx8)); 80 dst16 += 8; 81 count -= 8; 82 fx += dx8; 83 } while (count >= 8); 84 dst = (uint32_t *) dst16; 85 } 86 87 uint16_t* xx = (uint16_t*)dst; 88 for (i = count; i > 0; --i) { 89 *xx++ = SkToU16(fx >> 16); fx += dx; 90 } 91 } 92 93 void decal_filter_scale_neon(uint32_t dst[], SkFixed fx, SkFixed dx, int count) 94 { 95 if (count >= 8) { 96 int32x4_t wide_fx; 97 int32x4_t wide_fx2; 98 int32x4_t wide_dx8 = vdupq_n_s32(dx*8); 99 100 wide_fx = vdupq_n_s32(fx); 101 wide_fx = vsetq_lane_s32(fx+dx, wide_fx, 1); 102 wide_fx = vsetq_lane_s32(fx+dx+dx, wide_fx, 2); 103 wide_fx = vsetq_lane_s32(fx+dx+dx+dx, wide_fx, 3); 104 105 wide_fx2 = vaddq_s32(wide_fx, vdupq_n_s32(dx+dx+dx+dx)); 106 107 while (count >= 8) { 108 int32x4_t wide_out; 109 int32x4_t wide_out2; 110 111 wide_out = vshlq_n_s32(vshrq_n_s32(wide_fx, 12), 14); 112 wide_out = vorrq_s32(wide_out, 113 vaddq_s32(vshrq_n_s32(wide_fx,16), vdupq_n_s32(1))); 114 115 wide_out2 = vshlq_n_s32(vshrq_n_s32(wide_fx2, 12), 14); 116 wide_out2 = vorrq_s32(wide_out2, 117 vaddq_s32(vshrq_n_s32(wide_fx2,16), vdupq_n_s32(1))); 118 119 vst1q_u32(dst, vreinterpretq_u32_s32(wide_out)); 120 vst1q_u32(dst+4, vreinterpretq_u32_s32(wide_out2)); 121 122 dst += 8; 123 fx += dx*8; 124 wide_fx = vaddq_s32(wide_fx, wide_dx8); 125 wide_fx2 = vaddq_s32(wide_fx2, wide_dx8); 126 count -= 8; 127 } 128 } 129 130 if (count & 1) 131 { 132 SkASSERT((fx >> (16 + 14)) == 0); 133 *dst++ = (fx >> 12 << 14) | ((fx >> 16) + 1); 134 fx += dx; 135 } 136 while ((count -= 2) >= 0) 137 { 138 SkASSERT((fx >> (16 + 14)) == 0); 139 *dst++ = (fx >> 12 << 14) | ((fx >> 16) + 1); 140 fx += dx; 141 142 *dst++ = (fx >> 12 << 14) | ((fx >> 16) + 1); 143 fx += dx; 144 } 145 } 146
