1 // Copyright 2020 Google LLC.
2 // Use of this source code is governed by a BSD-style license that can be found in the LICENSE file.
3
4 #ifndef SkVM_opts_DEFINED
5 #define SkVM_opts_DEFINED
6
7 #include "include/private/SkVx.h"
8 #include "src/core/SkVM.h"
9
10 template <int N>
gather32(const int * ptr,const skvx::Vec<N,int> & ix)11 static inline skvx::Vec<N,int> gather32(const int* ptr, const skvx::Vec<N,int>& ix) {
12 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX2
13 if constexpr (N == 8) {
14 return skvx::bit_pun<skvx::Vec<N,int>>(
15 _mm256_i32gather_epi32(ptr, skvx::bit_pun<__m256i>(ix), 4));
16 }
17 #endif
18 // Try to recurse on specializations, falling back on standard scalar map()-based impl.
19 if constexpr (N > 8) {
20 return join(gather32(ptr, ix.lo),
21 gather32(ptr, ix.hi));
22 }
23 return map([&](int i) { return ptr[i]; }, ix);
24 }
25
26 namespace SK_OPTS_NS {
27
interpret_skvm(const skvm::InterpreterInstruction insts[],const int ninsts,const int nregs,const int loop,const int strides[],const int nargs,int n,void * args[])28 inline void interpret_skvm(const skvm::InterpreterInstruction insts[], const int ninsts,
29 const int nregs, const int loop,
30 const int strides[], const int nargs,
31 int n, void* args[]) {
32 using namespace skvm;
33
34 // We'll operate in SIMT style, knocking off K-size chunks from n while possible.
35 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX2
36 constexpr int K = 32; // 1024-bit: 4 ymm or 2 zmm at a time
37 #else
38 constexpr int K = 8; // 256-bit: 2 xmm, 2 v-registers, etc.
39 #endif
40 using I32 = skvx::Vec<K, int>;
41 using I16 = skvx::Vec<K, int16_t>;
42 using F32 = skvx::Vec<K, float>;
43 using U64 = skvx::Vec<K, uint64_t>;
44 using U32 = skvx::Vec<K, uint32_t>;
45 using U16 = skvx::Vec<K, uint16_t>;
46 using U8 = skvx::Vec<K, uint8_t>;
47 union Slot {
48 F32 f32;
49 I32 i32;
50 U32 u32;
51 I16 i16;
52 U16 u16;
53 };
54
55 Slot few_regs[16];
56 std::unique_ptr<char[]> many_regs;
57
58 Slot* r = few_regs;
59
60 if (nregs > (int)SK_ARRAY_COUNT(few_regs)) {
61 // Annoyingly we can't trust that malloc() or new will work with Slot because
62 // the skvx::Vec types may have alignment greater than what they provide.
63 // We'll overallocate one extra register so we can align manually.
64 many_regs.reset(new char[ sizeof(Slot) * (nregs + 1) ]);
65
66 uintptr_t addr = (uintptr_t)many_regs.get();
67 addr += alignof(Slot) -
68 (addr & (alignof(Slot) - 1));
69 SkASSERT((addr & (alignof(Slot) - 1)) == 0);
70 r = (Slot*)addr;
71 }
72
73
74 // Step each argument pointer ahead by its stride a number of times.
75 auto step_args = [&](int times) {
76 for (int i = 0; i < nargs; i++) {
77 args[i] = (void*)( (char*)args[i] + times * strides[i] );
78 }
79 };
80
81 int start = 0,
82 stride;
83 for ( ; n > 0; start = loop, n -= stride, step_args(stride)) {
84 stride = n >= K ? K : 1;
85
86 for (int i = start; i < ninsts; i++) {
87 InterpreterInstruction inst = insts[i];
88
89 // d = op(x,y,z,w, immA,immB)
90 Reg d = inst.d,
91 x = inst.x,
92 y = inst.y,
93 z = inst.z,
94 w = inst.w;
95 int immA = inst.immA,
96 immB = inst.immB;
97
98 // Ops that interact with memory need to know whether we're stride=1 or K,
99 // but all non-memory ops can run the same code no matter the stride.
100 switch (2*(int)inst.op + (stride == K ? 1 : 0)) {
101 default: SkUNREACHABLE;
102
103 #define STRIDE_1(op) case 2*(int)op
104 #define STRIDE_K(op) case 2*(int)op + 1
105 STRIDE_1(Op::store8 ): memcpy(args[immA], &r[x].i32, 1); break;
106 STRIDE_1(Op::store16): memcpy(args[immA], &r[x].i32, 2); break;
107 STRIDE_1(Op::store32): memcpy(args[immA], &r[x].i32, 4); break;
108 STRIDE_1(Op::store64): memcpy((char*)args[immA]+0, &r[x].i32, 4);
109 memcpy((char*)args[immA]+4, &r[y].i32, 4); break;
110
111 STRIDE_K(Op::store8 ): skvx::cast<uint8_t> (r[x].i32).store(args[immA]); break;
112 STRIDE_K(Op::store16): skvx::cast<uint16_t>(r[x].i32).store(args[immA]); break;
113 STRIDE_K(Op::store32): (r[x].i32).store(args[immA]); break;
114 STRIDE_K(Op::store64): (skvx::cast<uint64_t>(r[x].u32) << 0 |
115 skvx::cast<uint64_t>(r[y].u32) << 32).store(args[immA]);
116 break;
117
118 STRIDE_1(Op::load8 ): r[d].i32 = 0; memcpy(&r[d].i32, args[immA], 1); break;
119 STRIDE_1(Op::load16): r[d].i32 = 0; memcpy(&r[d].i32, args[immA], 2); break;
120 STRIDE_1(Op::load32): r[d].i32 = 0; memcpy(&r[d].i32, args[immA], 4); break;
121 STRIDE_1(Op::load64):
122 r[d].i32 = 0; memcpy(&r[d].i32, (char*)args[immA] + 4*immB, 4); break;
123
124 STRIDE_K(Op::load8 ): r[d].i32= skvx::cast<int>(U8 ::Load(args[immA])); break;
125 STRIDE_K(Op::load16): r[d].i32= skvx::cast<int>(U16::Load(args[immA])); break;
126 STRIDE_K(Op::load32): r[d].i32= I32::Load(args[immA]) ; break;
127 STRIDE_K(Op::load64):
128 // Low 32 bits if immB=0, or high 32 bits if immB=1.
129 r[d].i32 = skvx::cast<int>(U64::Load(args[immA]) >> (32*immB)); break;
130
131 // The pointer we base our gather on is loaded indirectly from a uniform:
132 // - args[immA] is the uniform holding our gather base pointer somewhere;
133 // - (const uint8_t*)args[immA] + immB points to the gather base pointer;
134 // - memcpy() loads the gather base and into a pointer of the right type.
135 // After all that we have an ordinary (uniform) pointer `ptr` to load from,
136 // and we then gather from it using the varying indices in r[x].
137 STRIDE_1(Op::gather8): {
138 const uint8_t* ptr;
139 memcpy(&ptr, (const uint8_t*)args[immA] + immB, sizeof(ptr));
140 r[d].i32 = ptr[ r[x].i32[0] ];
141 } break;
142 STRIDE_1(Op::gather16): {
143 const uint16_t* ptr;
144 memcpy(&ptr, (const uint8_t*)args[immA] + immB, sizeof(ptr));
145 r[d].i32 = ptr[ r[x].i32[0] ];
146 } break;
147 STRIDE_1(Op::gather32): {
148 const int* ptr;
149 memcpy(&ptr, (const uint8_t*)args[immA] + immB, sizeof(ptr));
150 r[d].i32 = ptr[ r[x].i32[0] ];
151 } break;
152
153 STRIDE_K(Op::gather8): {
154 const uint8_t* ptr;
155 memcpy(&ptr, (const uint8_t*)args[immA] + immB, sizeof(ptr));
156 r[d].i32 = map([&](int ix) { return (int)ptr[ix]; }, r[x].i32);
157 } break;
158 STRIDE_K(Op::gather16): {
159 const uint16_t* ptr;
160 memcpy(&ptr, (const uint8_t*)args[immA] + immB, sizeof(ptr));
161 r[d].i32 = map([&](int ix) { return (int)ptr[ix]; }, r[x].i32);
162 } break;
163 STRIDE_K(Op::gather32): {
164 const int* ptr;
165 memcpy(&ptr, (const uint8_t*)args[immA] + immB, sizeof(ptr));
166 r[d].i32 = gather32(ptr, r[x].i32);
167 } break;
168
169 #undef STRIDE_1
170 #undef STRIDE_K
171
172 // Ops that don't interact with memory should never care about the stride.
173 #define CASE(op) case 2*(int)op: /*fallthrough*/ case 2*(int)op+1
174
175 // These 128-bit ops are implemented serially for simplicity.
176 CASE(Op::store128): {
177 U64 lo = (skvx::cast<uint64_t>(r[x].u32) << 0 |
178 skvx::cast<uint64_t>(r[y].u32) << 32),
179 hi = (skvx::cast<uint64_t>(r[z].u32) << 0 |
180 skvx::cast<uint64_t>(r[w].u32) << 32);
181 for (int i = 0; i < stride; i++) {
182 memcpy((char*)args[immA] + 16*i + 0, &lo[i], 8);
183 memcpy((char*)args[immA] + 16*i + 8, &hi[i], 8);
184 }
185 } break;
186
187 CASE(Op::load128):
188 r[d].i32 = 0;
189 for (int i = 0; i < stride; i++) {
190 memcpy(&r[d].i32[i], (const char*)args[immA] + 16*i+ 4*immB, 4);
191 } break;
192
193 CASE(Op::assert_true):
194 #ifdef SK_DEBUG
195 if (!all(r[x].i32)) {
196 SkDebugf("inst %d, register %d\n", i, y);
197 for (int i = 0; i < K; i++) {
198 SkDebugf("\t%2d: %08x (%g)\n", i, r[y].i32[i], r[y].f32[i]);
199 }
200 SkASSERT(false);
201 }
202 #endif
203 break;
204
205 CASE(Op::index): {
206 const int iota[] = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9,10,11,12,13,14,15,
207 16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,
208 32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,
209 48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63 };
210 static_assert(K <= SK_ARRAY_COUNT(iota), "");
211
212 r[d].i32 = n - I32::Load(iota);
213 } break;
214
215 CASE(Op::uniform32):
216 r[d].i32 = *(const int*)( (const char*)args[immA] + immB );
217 break;
218
219 CASE(Op::splat): r[d].i32 = immA; break;
220
221 CASE(Op::add_f32): r[d].f32 = r[x].f32 + r[y].f32; break;
222 CASE(Op::sub_f32): r[d].f32 = r[x].f32 - r[y].f32; break;
223 CASE(Op::mul_f32): r[d].f32 = r[x].f32 * r[y].f32; break;
224 CASE(Op::div_f32): r[d].f32 = r[x].f32 / r[y].f32; break;
225 CASE(Op::min_f32): r[d].f32 = min(r[x].f32, r[y].f32); break;
226 CASE(Op::max_f32): r[d].f32 = max(r[x].f32, r[y].f32); break;
227
228 CASE(Op::fma_f32): r[d].f32 = fma( r[x].f32, r[y].f32, r[z].f32); break;
229 CASE(Op::fms_f32): r[d].f32 = fma( r[x].f32, r[y].f32, -r[z].f32); break;
230 CASE(Op::fnma_f32): r[d].f32 = fma(-r[x].f32, r[y].f32, r[z].f32); break;
231
232 CASE(Op::sqrt_f32): r[d].f32 = sqrt(r[x].f32); break;
233
234 CASE(Op::add_i32): r[d].i32 = r[x].i32 + r[y].i32; break;
235 CASE(Op::sub_i32): r[d].i32 = r[x].i32 - r[y].i32; break;
236 CASE(Op::mul_i32): r[d].i32 = r[x].i32 * r[y].i32; break;
237
238 CASE(Op::shl_i32): r[d].i32 = r[x].i32 << immA; break;
239 CASE(Op::sra_i32): r[d].i32 = r[x].i32 >> immA; break;
240 CASE(Op::shr_i32): r[d].u32 = r[x].u32 >> immA; break;
241
242 CASE(Op:: eq_f32): r[d].i32 = r[x].f32 == r[y].f32; break;
243 CASE(Op::neq_f32): r[d].i32 = r[x].f32 != r[y].f32; break;
244 CASE(Op:: gt_f32): r[d].i32 = r[x].f32 > r[y].f32; break;
245 CASE(Op::gte_f32): r[d].i32 = r[x].f32 >= r[y].f32; break;
246
247 CASE(Op:: eq_i32): r[d].i32 = r[x].i32 == r[y].i32; break;
248 CASE(Op:: gt_i32): r[d].i32 = r[x].i32 > r[y].i32; break;
249
250 CASE(Op::bit_and ): r[d].i32 = r[x].i32 & r[y].i32; break;
251 CASE(Op::bit_or ): r[d].i32 = r[x].i32 | r[y].i32; break;
252 CASE(Op::bit_xor ): r[d].i32 = r[x].i32 ^ r[y].i32; break;
253 CASE(Op::bit_clear): r[d].i32 = r[x].i32 & ~r[y].i32; break;
254
255 CASE(Op::select): r[d].i32 = skvx::if_then_else(r[x].i32, r[y].i32, r[z].i32);
256 break;
257
258 CASE(Op::ceil): r[d].f32 = skvx::ceil(r[x].f32) ; break;
259 CASE(Op::floor): r[d].f32 = skvx::floor(r[x].f32) ; break;
260 CASE(Op::to_f32): r[d].f32 = skvx::cast<float>( r[x].i32 ); break;
261 CASE(Op::trunc): r[d].i32 = skvx::cast<int> ( r[x].f32 ); break;
262 CASE(Op::round): r[d].i32 = skvx::cast<int> (skvx::lrint(r[x].f32)); break;
263
264 CASE(Op::to_fp16):
265 r[d].i32 = skvx::cast<int>(skvx::to_half(r[x].f32));
266 break;
267 CASE(Op::from_fp16):
268 r[d].f32 = skvx::from_half(skvx::cast<uint16_t>(r[x].i32));
269 break;
270
271 #undef CASE
272 }
273 }
274 }
275 }
276
277 } // namespace SK_OPTS_NS
278
279 #endif//SkVM_opts_DEFINED
280