OLD | NEW |
1 /* | 1 /* |
2 * Copyright 2015 Google Inc. | 2 * Copyright 2015 Google Inc. |
3 * | 3 * |
4 * Use of this source code is governed by a BSD-style license that can be | 4 * Use of this source code is governed by a BSD-style license that can be |
5 * found in the LICENSE file. | 5 * found in the LICENSE file. |
6 */ | 6 */ |
7 | 7 |
8 #ifndef SkPx_neon_DEFINED | 8 #ifndef SkPx_neon_DEFINED |
9 #define SkPx_neon_DEFINED | 9 #define SkPx_neon_DEFINED |
10 | 10 |
11 // When we have NEON, we like to work 8 pixels at a time. | 11 // When we have NEON, we like to work 8 pixels at a time. |
12 // This lets us exploit vld4/vst4 and represent SkPx as planar uint8x8x4_t, | 12 // This lets us exploit vld4/vst4 and represent SkPx as planar uint8x8x4_t, |
13 // Wide as planar uint16x8x4_t, and Alpha as a single uint8x8_t plane. | 13 // Wide as planar uint16x8x4_t, and Alpha as a single uint8x8_t plane. |
14 | 14 |
15 struct SkPx_neon { | 15 namespace neon { |
| 16 |
| 17 struct SkPx { |
16 static const int N = 8; | 18 static const int N = 8; |
17 | 19 |
18 uint8x8x4_t fVec; | 20 uint8x8x4_t fVec; |
19 SkPx_neon(uint8x8x4_t vec) : fVec(vec) {} | 21 SkPx(uint8x8x4_t vec) : fVec(vec) {} |
20 | 22 |
21 static SkPx_neon Dup(uint32_t px) { return vld4_dup_u8((const uint8_t*)&px);
} | 23 static SkPx Dup(uint32_t px) { return vld4_dup_u8((const uint8_t*)&px); } |
22 static SkPx_neon Load(const uint32_t* px) { return vld4_u8((const uint8_t*)p
x); } | 24 static SkPx Load(const uint32_t* px) { return vld4_u8((const uint8_t*)px); } |
23 static SkPx_neon Load(const uint32_t* px, int n) { | 25 static SkPx Load(const uint32_t* px, int n) { |
24 SkASSERT(0 < n && n < 8); | 26 SkASSERT(0 < n && n < 8); |
25 uint8x8x4_t v = vld4_dup_u8((const uint8_t*)px); // n>=1, so start all
lanes with pixel 0. | 27 uint8x8x4_t v = vld4_dup_u8((const uint8_t*)px); // n>=1, so start all
lanes with pixel 0. |
26 switch (n) { | 28 switch (n) { |
27 case 7: v = vld4_lane_u8((const uint8_t*)(px+6), v, 6); // fall thr
ough | 29 case 7: v = vld4_lane_u8((const uint8_t*)(px+6), v, 6); // fall thr
ough |
28 case 6: v = vld4_lane_u8((const uint8_t*)(px+5), v, 5); // fall thr
ough | 30 case 6: v = vld4_lane_u8((const uint8_t*)(px+5), v, 5); // fall thr
ough |
29 case 5: v = vld4_lane_u8((const uint8_t*)(px+4), v, 4); // fall thr
ough | 31 case 5: v = vld4_lane_u8((const uint8_t*)(px+4), v, 4); // fall thr
ough |
30 case 4: v = vld4_lane_u8((const uint8_t*)(px+3), v, 3); // fall thr
ough | 32 case 4: v = vld4_lane_u8((const uint8_t*)(px+3), v, 3); // fall thr
ough |
31 case 3: v = vld4_lane_u8((const uint8_t*)(px+2), v, 2); // fall thr
ough | 33 case 3: v = vld4_lane_u8((const uint8_t*)(px+2), v, 2); // fall thr
ough |
32 case 2: v = vld4_lane_u8((const uint8_t*)(px+1), v, 1); | 34 case 2: v = vld4_lane_u8((const uint8_t*)(px+1), v, 1); |
33 } | 35 } |
(...skipping 67 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
101 } | 103 } |
102 template <int bits> Wide shr() const { | 104 template <int bits> Wide shr() const { |
103 return (uint16x8x4_t) {{ | 105 return (uint16x8x4_t) {{ |
104 vshrq_n_u16(fVec.val[0], bits), | 106 vshrq_n_u16(fVec.val[0], bits), |
105 vshrq_n_u16(fVec.val[1], bits), | 107 vshrq_n_u16(fVec.val[1], bits), |
106 vshrq_n_u16(fVec.val[2], bits), | 108 vshrq_n_u16(fVec.val[2], bits), |
107 vshrq_n_u16(fVec.val[3], bits), | 109 vshrq_n_u16(fVec.val[3], bits), |
108 }}; | 110 }}; |
109 } | 111 } |
110 | 112 |
111 SkPx_neon addNarrowHi(const SkPx_neon& o) const { | 113 SkPx addNarrowHi(const SkPx& o) const { |
112 return (uint8x8x4_t) {{ | 114 return (uint8x8x4_t) {{ |
113 vshrn_n_u16(vaddw_u8(fVec.val[0], o.fVec.val[0]), 8), | 115 vshrn_n_u16(vaddw_u8(fVec.val[0], o.fVec.val[0]), 8), |
114 vshrn_n_u16(vaddw_u8(fVec.val[1], o.fVec.val[1]), 8), | 116 vshrn_n_u16(vaddw_u8(fVec.val[1], o.fVec.val[1]), 8), |
115 vshrn_n_u16(vaddw_u8(fVec.val[2], o.fVec.val[2]), 8), | 117 vshrn_n_u16(vaddw_u8(fVec.val[2], o.fVec.val[2]), 8), |
116 vshrn_n_u16(vaddw_u8(fVec.val[3], o.fVec.val[3]), 8), | 118 vshrn_n_u16(vaddw_u8(fVec.val[3], o.fVec.val[3]), 8), |
117 }}; | 119 }}; |
118 } | 120 } |
119 }; | 121 }; |
120 | 122 |
121 Alpha alpha() const { return fVec.val[3]; } | 123 Alpha alpha() const { return fVec.val[3]; } |
122 | 124 |
123 Wide widenLo() const { | 125 Wide widenLo() const { |
124 return (uint16x8x4_t) {{ | 126 return (uint16x8x4_t) {{ |
125 vmovl_u8(fVec.val[0]), | 127 vmovl_u8(fVec.val[0]), |
126 vmovl_u8(fVec.val[1]), | 128 vmovl_u8(fVec.val[1]), |
127 vmovl_u8(fVec.val[2]), | 129 vmovl_u8(fVec.val[2]), |
128 vmovl_u8(fVec.val[3]), | 130 vmovl_u8(fVec.val[3]), |
129 }}; | 131 }}; |
130 } | 132 } |
131 // TODO: these two can probably be done faster. | 133 // TODO: these two can probably be done faster. |
132 Wide widenHi() const { return this->widenLo().shl<8>(); } | 134 Wide widenHi() const { return this->widenLo().shl<8>(); } |
133 Wide widenLoHi() const { return this->widenLo() + this->widenHi(); } | 135 Wide widenLoHi() const { return this->widenLo() + this->widenHi(); } |
134 | 136 |
135 SkPx_neon operator+(const SkPx_neon& o) const { | 137 SkPx operator+(const SkPx& o) const { |
136 return (uint8x8x4_t) {{ | 138 return (uint8x8x4_t) {{ |
137 vadd_u8(fVec.val[0], o.fVec.val[0]), | 139 vadd_u8(fVec.val[0], o.fVec.val[0]), |
138 vadd_u8(fVec.val[1], o.fVec.val[1]), | 140 vadd_u8(fVec.val[1], o.fVec.val[1]), |
139 vadd_u8(fVec.val[2], o.fVec.val[2]), | 141 vadd_u8(fVec.val[2], o.fVec.val[2]), |
140 vadd_u8(fVec.val[3], o.fVec.val[3]), | 142 vadd_u8(fVec.val[3], o.fVec.val[3]), |
141 }}; | 143 }}; |
142 } | 144 } |
143 SkPx_neon operator-(const SkPx_neon& o) const { | 145 SkPx operator-(const SkPx& o) const { |
144 return (uint8x8x4_t) {{ | 146 return (uint8x8x4_t) {{ |
145 vsub_u8(fVec.val[0], o.fVec.val[0]), | 147 vsub_u8(fVec.val[0], o.fVec.val[0]), |
146 vsub_u8(fVec.val[1], o.fVec.val[1]), | 148 vsub_u8(fVec.val[1], o.fVec.val[1]), |
147 vsub_u8(fVec.val[2], o.fVec.val[2]), | 149 vsub_u8(fVec.val[2], o.fVec.val[2]), |
148 vsub_u8(fVec.val[3], o.fVec.val[3]), | 150 vsub_u8(fVec.val[3], o.fVec.val[3]), |
149 }}; | 151 }}; |
150 } | 152 } |
151 SkPx_neon saturatedAdd(const SkPx_neon& o) const { | 153 SkPx saturatedAdd(const SkPx& o) const { |
152 return (uint8x8x4_t) {{ | 154 return (uint8x8x4_t) {{ |
153 vqadd_u8(fVec.val[0], o.fVec.val[0]), | 155 vqadd_u8(fVec.val[0], o.fVec.val[0]), |
154 vqadd_u8(fVec.val[1], o.fVec.val[1]), | 156 vqadd_u8(fVec.val[1], o.fVec.val[1]), |
155 vqadd_u8(fVec.val[2], o.fVec.val[2]), | 157 vqadd_u8(fVec.val[2], o.fVec.val[2]), |
156 vqadd_u8(fVec.val[3], o.fVec.val[3]), | 158 vqadd_u8(fVec.val[3], o.fVec.val[3]), |
157 }}; | 159 }}; |
158 } | 160 } |
159 | 161 |
160 Wide operator*(const Alpha& a) const { | 162 Wide operator*(const Alpha& a) const { |
161 return (uint16x8x4_t) {{ | 163 return (uint16x8x4_t) {{ |
162 vmull_u8(fVec.val[0], a.fA), | 164 vmull_u8(fVec.val[0], a.fA), |
163 vmull_u8(fVec.val[1], a.fA), | 165 vmull_u8(fVec.val[1], a.fA), |
164 vmull_u8(fVec.val[2], a.fA), | 166 vmull_u8(fVec.val[2], a.fA), |
165 vmull_u8(fVec.val[3], a.fA), | 167 vmull_u8(fVec.val[3], a.fA), |
166 }}; | 168 }}; |
167 } | 169 } |
168 SkPx_neon approxMulDiv255(const Alpha& a) const { | 170 SkPx approxMulDiv255(const Alpha& a) const { |
169 return (*this * a).addNarrowHi(*this); | 171 return (*this * a).addNarrowHi(*this); |
170 } | 172 } |
171 | 173 |
172 SkPx_neon addAlpha(const Alpha& a) const { | 174 SkPx addAlpha(const Alpha& a) const { |
173 return (uint8x8x4_t) {{ | 175 return (uint8x8x4_t) {{ |
174 fVec.val[0], | 176 fVec.val[0], |
175 fVec.val[1], | 177 fVec.val[1], |
176 fVec.val[2], | 178 fVec.val[2], |
177 vadd_u8(fVec.val[3], a.fA), | 179 vadd_u8(fVec.val[3], a.fA), |
178 }}; | 180 }}; |
179 } | 181 } |
180 }; | 182 }; |
181 typedef SkPx_neon SkPx; | 183 |
| 184 } // namespace neon |
| 185 |
| 186 typedef neon::SkPx SkPx; |
182 | 187 |
183 #endif//SkPx_neon_DEFINED | 188 #endif//SkPx_neon_DEFINED |
OLD | NEW |