|
1 #include "precompiled.h" |
|
2 // |
|
3 // Copyright (c) 2012-2013 The ANGLE Project Authors. All rights reserved. |
|
4 // Use of this source code is governed by a BSD-style license that can be |
|
5 // found in the LICENSE file. |
|
6 // |
|
7 |
|
8 // RenderStateCache.cpp: Defines rx::RenderStateCache, a cache of Direct3D render |
|
9 // state objects. |
|
10 |
|
11 #include "libGLESv2/renderer/RenderStateCache.h" |
|
12 #include "libGLESv2/renderer/renderer11_utils.h" |
|
13 |
|
14 #include "common/debug.h" |
|
15 #include "third_party/murmurhash/MurmurHash3.h" |
|
16 |
|
17 namespace rx |
|
18 { |
|
19 |
|
20 // MSDN's documentation of ID3D11Device::CreateBlendState, ID3D11Device::CreateRasterizerState, |
|
21 // ID3D11Device::CreateDepthStencilState and ID3D11Device::CreateSamplerState claims the maximum |
|
22 // number of unique states of each type an application can create is 4096 |
|
23 const unsigned int RenderStateCache::kMaxBlendStates = 4096; |
|
24 const unsigned int RenderStateCache::kMaxRasterizerStates = 4096; |
|
25 const unsigned int RenderStateCache::kMaxDepthStencilStates = 4096; |
|
26 const unsigned int RenderStateCache::kMaxSamplerStates = 4096; |
|
27 |
|
28 RenderStateCache::RenderStateCache() : mDevice(NULL), mCounter(0), |
|
29 mBlendStateCache(kMaxBlendStates, hashBlendState, compareBlendStates), |
|
30 mRasterizerStateCache(kMaxRasterizerStates, hashRasterizerState, compareRasterizerStates), |
|
31 mDepthStencilStateCache(kMaxDepthStencilStates, hashDepthStencilState, compareDepthStencilStates), |
|
32 mSamplerStateCache(kMaxSamplerStates, hashSamplerState, compareSamplerStates) |
|
33 { |
|
34 } |
|
35 |
|
36 RenderStateCache::~RenderStateCache() |
|
37 { |
|
38 clear(); |
|
39 } |
|
40 |
|
41 void RenderStateCache::initialize(ID3D11Device *device) |
|
42 { |
|
43 clear(); |
|
44 mDevice = device; |
|
45 } |
|
46 |
|
47 void RenderStateCache::clear() |
|
48 { |
|
49 for (BlendStateMap::iterator i = mBlendStateCache.begin(); i != mBlendStateCache.end(); i++) |
|
50 { |
|
51 i->second.first->Release(); |
|
52 } |
|
53 mBlendStateCache.clear(); |
|
54 |
|
55 for (RasterizerStateMap::iterator i = mRasterizerStateCache.begin(); i != mRasterizerStateCache.end(); i++) |
|
56 { |
|
57 i->second.first->Release(); |
|
58 } |
|
59 mRasterizerStateCache.clear(); |
|
60 |
|
61 for (DepthStencilStateMap::iterator i = mDepthStencilStateCache.begin(); i != mDepthStencilStateCache.end(); i++) |
|
62 { |
|
63 i->second.first->Release(); |
|
64 } |
|
65 mDepthStencilStateCache.clear(); |
|
66 |
|
67 for (SamplerStateMap::iterator i = mSamplerStateCache.begin(); i != mSamplerStateCache.end(); i++) |
|
68 { |
|
69 i->second.first->Release(); |
|
70 } |
|
71 mSamplerStateCache.clear(); |
|
72 } |
|
73 |
|
74 std::size_t RenderStateCache::hashBlendState(const gl::BlendState &blendState) |
|
75 { |
|
76 static const unsigned int seed = 0xABCDEF98; |
|
77 |
|
78 std::size_t hash = 0; |
|
79 MurmurHash3_x86_32(&blendState, sizeof(gl::BlendState), seed, &hash); |
|
80 return hash; |
|
81 } |
|
82 |
|
83 bool RenderStateCache::compareBlendStates(const gl::BlendState &a, const gl::BlendState &b) |
|
84 { |
|
85 return memcmp(&a, &b, sizeof(gl::BlendState)) == 0; |
|
86 } |
|
87 |
|
88 ID3D11BlendState *RenderStateCache::getBlendState(const gl::BlendState &blendState) |
|
89 { |
|
90 if (!mDevice) |
|
91 { |
|
92 ERR("RenderStateCache is not initialized."); |
|
93 return NULL; |
|
94 } |
|
95 |
|
96 BlendStateMap::iterator i = mBlendStateCache.find(blendState); |
|
97 if (i != mBlendStateCache.end()) |
|
98 { |
|
99 BlendStateCounterPair &state = i->second; |
|
100 state.second = mCounter++; |
|
101 return state.first; |
|
102 } |
|
103 else |
|
104 { |
|
105 if (mBlendStateCache.size() >= kMaxBlendStates) |
|
106 { |
|
107 TRACE("Overflowed the limit of %u blend states, removing the least recently used " |
|
108 "to make room.", kMaxBlendStates); |
|
109 |
|
110 BlendStateMap::iterator leastRecentlyUsed = mBlendStateCache.begin(); |
|
111 for (BlendStateMap::iterator i = mBlendStateCache.begin(); i != mBlendStateCache.end(); i++) |
|
112 { |
|
113 if (i->second.second < leastRecentlyUsed->second.second) |
|
114 { |
|
115 leastRecentlyUsed = i; |
|
116 } |
|
117 } |
|
118 leastRecentlyUsed->second.first->Release(); |
|
119 mBlendStateCache.erase(leastRecentlyUsed); |
|
120 } |
|
121 |
|
122 // Create a new blend state and insert it into the cache |
|
123 D3D11_BLEND_DESC blendDesc = { 0 }; |
|
124 blendDesc.AlphaToCoverageEnable = blendState.sampleAlphaToCoverage; |
|
125 blendDesc.IndependentBlendEnable = FALSE; |
|
126 |
|
127 for (unsigned int i = 0; i < D3D11_SIMULTANEOUS_RENDER_TARGET_COUNT; i++) |
|
128 { |
|
129 D3D11_RENDER_TARGET_BLEND_DESC &rtBlend = blendDesc.RenderTarget[i]; |
|
130 |
|
131 rtBlend.BlendEnable = blendState.blend; |
|
132 if (blendState.blend) |
|
133 { |
|
134 rtBlend.SrcBlend = gl_d3d11::ConvertBlendFunc(blendState.sourceBlendRGB, false); |
|
135 rtBlend.DestBlend = gl_d3d11::ConvertBlendFunc(blendState.destBlendRGB, false); |
|
136 rtBlend.BlendOp = gl_d3d11::ConvertBlendOp(blendState.blendEquationRGB); |
|
137 |
|
138 rtBlend.SrcBlendAlpha = gl_d3d11::ConvertBlendFunc(blendState.sourceBlendAlpha, true); |
|
139 rtBlend.DestBlendAlpha = gl_d3d11::ConvertBlendFunc(blendState.destBlendAlpha, true); |
|
140 rtBlend.BlendOpAlpha = gl_d3d11::ConvertBlendOp(blendState.blendEquationAlpha); |
|
141 } |
|
142 |
|
143 rtBlend.RenderTargetWriteMask = gl_d3d11::ConvertColorMask(blendState.colorMaskRed, |
|
144 blendState.colorMaskGreen, |
|
145 blendState.colorMaskBlue, |
|
146 blendState.colorMaskAlpha); |
|
147 } |
|
148 |
|
149 ID3D11BlendState *dx11BlendState = NULL; |
|
150 HRESULT result = mDevice->CreateBlendState(&blendDesc, &dx11BlendState); |
|
151 if (FAILED(result) || !dx11BlendState) |
|
152 { |
|
153 ERR("Unable to create a ID3D11BlendState, HRESULT: 0x%X.", result); |
|
154 return NULL; |
|
155 } |
|
156 |
|
157 mBlendStateCache.insert(std::make_pair(blendState, std::make_pair(dx11BlendState, mCounter++))); |
|
158 |
|
159 return dx11BlendState; |
|
160 } |
|
161 } |
|
162 |
|
163 std::size_t RenderStateCache::hashRasterizerState(const RasterizerStateKey &rasterState) |
|
164 { |
|
165 static const unsigned int seed = 0xABCDEF98; |
|
166 |
|
167 std::size_t hash = 0; |
|
168 MurmurHash3_x86_32(&rasterState, sizeof(RasterizerStateKey), seed, &hash); |
|
169 return hash; |
|
170 } |
|
171 |
|
172 bool RenderStateCache::compareRasterizerStates(const RasterizerStateKey &a, const RasterizerStateKey &b) |
|
173 { |
|
174 return memcmp(&a, &b, sizeof(RasterizerStateKey)) == 0; |
|
175 } |
|
176 |
|
177 ID3D11RasterizerState *RenderStateCache::getRasterizerState(const gl::RasterizerState &rasterState, |
|
178 bool scissorEnabled, unsigned int depthSize) |
|
179 { |
|
180 if (!mDevice) |
|
181 { |
|
182 ERR("RenderStateCache is not initialized."); |
|
183 return NULL; |
|
184 } |
|
185 |
|
186 RasterizerStateKey key; |
|
187 key.rasterizerState = rasterState; |
|
188 key.scissorEnabled = scissorEnabled; |
|
189 key.depthSize = depthSize; |
|
190 |
|
191 RasterizerStateMap::iterator i = mRasterizerStateCache.find(key); |
|
192 if (i != mRasterizerStateCache.end()) |
|
193 { |
|
194 RasterizerStateCounterPair &state = i->second; |
|
195 state.second = mCounter++; |
|
196 return state.first; |
|
197 } |
|
198 else |
|
199 { |
|
200 if (mRasterizerStateCache.size() >= kMaxRasterizerStates) |
|
201 { |
|
202 TRACE("Overflowed the limit of %u rasterizer states, removing the least recently used " |
|
203 "to make room.", kMaxRasterizerStates); |
|
204 |
|
205 RasterizerStateMap::iterator leastRecentlyUsed = mRasterizerStateCache.begin(); |
|
206 for (RasterizerStateMap::iterator i = mRasterizerStateCache.begin(); i != mRasterizerStateCache.end(); i++) |
|
207 { |
|
208 if (i->second.second < leastRecentlyUsed->second.second) |
|
209 { |
|
210 leastRecentlyUsed = i; |
|
211 } |
|
212 } |
|
213 leastRecentlyUsed->second.first->Release(); |
|
214 mRasterizerStateCache.erase(leastRecentlyUsed); |
|
215 } |
|
216 |
|
217 D3D11_CULL_MODE cullMode = gl_d3d11::ConvertCullMode(rasterState.cullFace, rasterState.cullMode); |
|
218 |
|
219 // Disable culling if drawing points |
|
220 if (rasterState.pointDrawMode) |
|
221 { |
|
222 cullMode = D3D11_CULL_NONE; |
|
223 } |
|
224 |
|
225 D3D11_RASTERIZER_DESC rasterDesc; |
|
226 rasterDesc.FillMode = D3D11_FILL_SOLID; |
|
227 rasterDesc.CullMode = cullMode; |
|
228 rasterDesc.FrontCounterClockwise = (rasterState.frontFace == GL_CCW) ? FALSE: TRUE; |
|
229 rasterDesc.DepthBias = ldexp(rasterState.polygonOffsetUnits, -static_cast<int>(depthSize)); |
|
230 rasterDesc.DepthBiasClamp = 0.0f; // MSDN documentation of DepthBiasClamp implies a value of zero will preform no clamping, must be tested though. |
|
231 rasterDesc.SlopeScaledDepthBias = rasterState.polygonOffsetFactor; |
|
232 rasterDesc.DepthClipEnable = TRUE; |
|
233 rasterDesc.ScissorEnable = scissorEnabled ? TRUE : FALSE; |
|
234 rasterDesc.MultisampleEnable = rasterState.multiSample; |
|
235 rasterDesc.AntialiasedLineEnable = FALSE; |
|
236 |
|
237 ID3D11RasterizerState *dx11RasterizerState = NULL; |
|
238 HRESULT result = mDevice->CreateRasterizerState(&rasterDesc, &dx11RasterizerState); |
|
239 if (FAILED(result) || !dx11RasterizerState) |
|
240 { |
|
241 ERR("Unable to create a ID3D11RasterizerState, HRESULT: 0x%X.", result); |
|
242 return NULL; |
|
243 } |
|
244 |
|
245 mRasterizerStateCache.insert(std::make_pair(key, std::make_pair(dx11RasterizerState, mCounter++))); |
|
246 |
|
247 return dx11RasterizerState; |
|
248 } |
|
249 } |
|
250 |
|
251 std::size_t RenderStateCache::hashDepthStencilState(const gl::DepthStencilState &dsState) |
|
252 { |
|
253 static const unsigned int seed = 0xABCDEF98; |
|
254 |
|
255 std::size_t hash = 0; |
|
256 MurmurHash3_x86_32(&dsState, sizeof(gl::DepthStencilState), seed, &hash); |
|
257 return hash; |
|
258 } |
|
259 |
|
260 bool RenderStateCache::compareDepthStencilStates(const gl::DepthStencilState &a, const gl::DepthStencilState &b) |
|
261 { |
|
262 return memcmp(&a, &b, sizeof(gl::DepthStencilState)) == 0; |
|
263 } |
|
264 |
|
265 ID3D11DepthStencilState *RenderStateCache::getDepthStencilState(const gl::DepthStencilState &dsState) |
|
266 { |
|
267 if (!mDevice) |
|
268 { |
|
269 ERR("RenderStateCache is not initialized."); |
|
270 return NULL; |
|
271 } |
|
272 |
|
273 DepthStencilStateMap::iterator i = mDepthStencilStateCache.find(dsState); |
|
274 if (i != mDepthStencilStateCache.end()) |
|
275 { |
|
276 DepthStencilStateCounterPair &state = i->second; |
|
277 state.second = mCounter++; |
|
278 return state.first; |
|
279 } |
|
280 else |
|
281 { |
|
282 if (mDepthStencilStateCache.size() >= kMaxDepthStencilStates) |
|
283 { |
|
284 TRACE("Overflowed the limit of %u depth stencil states, removing the least recently used " |
|
285 "to make room.", kMaxDepthStencilStates); |
|
286 |
|
287 DepthStencilStateMap::iterator leastRecentlyUsed = mDepthStencilStateCache.begin(); |
|
288 for (DepthStencilStateMap::iterator i = mDepthStencilStateCache.begin(); i != mDepthStencilStateCache.end(); i++) |
|
289 { |
|
290 if (i->second.second < leastRecentlyUsed->second.second) |
|
291 { |
|
292 leastRecentlyUsed = i; |
|
293 } |
|
294 } |
|
295 leastRecentlyUsed->second.first->Release(); |
|
296 mDepthStencilStateCache.erase(leastRecentlyUsed); |
|
297 } |
|
298 |
|
299 D3D11_DEPTH_STENCIL_DESC dsDesc = { 0 }; |
|
300 dsDesc.DepthEnable = dsState.depthTest ? TRUE : FALSE; |
|
301 dsDesc.DepthWriteMask = gl_d3d11::ConvertDepthMask(dsState.depthMask); |
|
302 dsDesc.DepthFunc = gl_d3d11::ConvertComparison(dsState.depthFunc); |
|
303 dsDesc.StencilEnable = dsState.stencilTest ? TRUE : FALSE; |
|
304 dsDesc.StencilReadMask = gl_d3d11::ConvertStencilMask(dsState.stencilMask); |
|
305 dsDesc.StencilWriteMask = gl_d3d11::ConvertStencilMask(dsState.stencilWritemask); |
|
306 dsDesc.FrontFace.StencilFailOp = gl_d3d11::ConvertStencilOp(dsState.stencilFail); |
|
307 dsDesc.FrontFace.StencilDepthFailOp = gl_d3d11::ConvertStencilOp(dsState.stencilPassDepthFail); |
|
308 dsDesc.FrontFace.StencilPassOp = gl_d3d11::ConvertStencilOp(dsState.stencilPassDepthPass); |
|
309 dsDesc.FrontFace.StencilFunc = gl_d3d11::ConvertComparison(dsState.stencilFunc); |
|
310 dsDesc.BackFace.StencilFailOp = gl_d3d11::ConvertStencilOp(dsState.stencilBackFail); |
|
311 dsDesc.BackFace.StencilDepthFailOp = gl_d3d11::ConvertStencilOp(dsState.stencilBackPassDepthFail); |
|
312 dsDesc.BackFace.StencilPassOp = gl_d3d11::ConvertStencilOp(dsState.stencilBackPassDepthPass); |
|
313 dsDesc.BackFace.StencilFunc = gl_d3d11::ConvertComparison(dsState.stencilBackFunc); |
|
314 |
|
315 ID3D11DepthStencilState *dx11DepthStencilState = NULL; |
|
316 HRESULT result = mDevice->CreateDepthStencilState(&dsDesc, &dx11DepthStencilState); |
|
317 if (FAILED(result) || !dx11DepthStencilState) |
|
318 { |
|
319 ERR("Unable to create a ID3D11DepthStencilState, HRESULT: 0x%X.", result); |
|
320 return NULL; |
|
321 } |
|
322 |
|
323 mDepthStencilStateCache.insert(std::make_pair(dsState, std::make_pair(dx11DepthStencilState, mCounter++))); |
|
324 |
|
325 return dx11DepthStencilState; |
|
326 } |
|
327 } |
|
328 |
|
329 std::size_t RenderStateCache::hashSamplerState(const gl::SamplerState &samplerState) |
|
330 { |
|
331 static const unsigned int seed = 0xABCDEF98; |
|
332 |
|
333 std::size_t hash = 0; |
|
334 MurmurHash3_x86_32(&samplerState, sizeof(gl::SamplerState), seed, &hash); |
|
335 return hash; |
|
336 } |
|
337 |
|
338 bool RenderStateCache::compareSamplerStates(const gl::SamplerState &a, const gl::SamplerState &b) |
|
339 { |
|
340 return memcmp(&a, &b, sizeof(gl::SamplerState)) == 0; |
|
341 } |
|
342 |
|
343 ID3D11SamplerState *RenderStateCache::getSamplerState(const gl::SamplerState &samplerState) |
|
344 { |
|
345 if (!mDevice) |
|
346 { |
|
347 ERR("RenderStateCache is not initialized."); |
|
348 return NULL; |
|
349 } |
|
350 |
|
351 SamplerStateMap::iterator i = mSamplerStateCache.find(samplerState); |
|
352 if (i != mSamplerStateCache.end()) |
|
353 { |
|
354 SamplerStateCounterPair &state = i->second; |
|
355 state.second = mCounter++; |
|
356 return state.first; |
|
357 } |
|
358 else |
|
359 { |
|
360 if (mSamplerStateCache.size() >= kMaxSamplerStates) |
|
361 { |
|
362 TRACE("Overflowed the limit of %u sampler states, removing the least recently used " |
|
363 "to make room.", kMaxSamplerStates); |
|
364 |
|
365 SamplerStateMap::iterator leastRecentlyUsed = mSamplerStateCache.begin(); |
|
366 for (SamplerStateMap::iterator i = mSamplerStateCache.begin(); i != mSamplerStateCache.end(); i++) |
|
367 { |
|
368 if (i->second.second < leastRecentlyUsed->second.second) |
|
369 { |
|
370 leastRecentlyUsed = i; |
|
371 } |
|
372 } |
|
373 leastRecentlyUsed->second.first->Release(); |
|
374 mSamplerStateCache.erase(leastRecentlyUsed); |
|
375 } |
|
376 |
|
377 D3D11_SAMPLER_DESC samplerDesc; |
|
378 samplerDesc.Filter = gl_d3d11::ConvertFilter(samplerState.minFilter, samplerState.magFilter, samplerState.maxAnisotropy); |
|
379 samplerDesc.AddressU = gl_d3d11::ConvertTextureWrap(samplerState.wrapS); |
|
380 samplerDesc.AddressV = gl_d3d11::ConvertTextureWrap(samplerState.wrapT); |
|
381 samplerDesc.AddressW = D3D11_TEXTURE_ADDRESS_CLAMP; |
|
382 samplerDesc.MipLODBias = static_cast<float>(samplerState.lodOffset); |
|
383 samplerDesc.MaxAnisotropy = samplerState.maxAnisotropy; |
|
384 samplerDesc.ComparisonFunc = D3D11_COMPARISON_NEVER; |
|
385 samplerDesc.BorderColor[0] = 0.0f; |
|
386 samplerDesc.BorderColor[1] = 0.0f; |
|
387 samplerDesc.BorderColor[2] = 0.0f; |
|
388 samplerDesc.BorderColor[3] = 0.0f; |
|
389 samplerDesc.MinLOD = gl_d3d11::ConvertMinLOD(samplerState.minFilter, samplerState.lodOffset); |
|
390 samplerDesc.MaxLOD = gl_d3d11::ConvertMaxLOD(samplerState.minFilter, samplerState.lodOffset); |
|
391 |
|
392 ID3D11SamplerState *dx11SamplerState = NULL; |
|
393 HRESULT result = mDevice->CreateSamplerState(&samplerDesc, &dx11SamplerState); |
|
394 if (FAILED(result) || !dx11SamplerState) |
|
395 { |
|
396 ERR("Unable to create a ID3D11DepthStencilState, HRESULT: 0x%X.", result); |
|
397 return NULL; |
|
398 } |
|
399 |
|
400 mSamplerStateCache.insert(std::make_pair(samplerState, std::make_pair(dx11SamplerState, mCounter++))); |
|
401 |
|
402 return dx11SamplerState; |
|
403 } |
|
404 } |
|
405 |
|
406 } |