1 | /* |
2 | * Copyright 2016 Google Inc. |
3 | * |
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 | #include "src/gpu/GrProgramDesc.h" |
9 | |
10 | #include "include/private/SkChecksum.h" |
11 | #include "include/private/SkTo.h" |
12 | #include "src/gpu/GrPipeline.h" |
13 | #include "src/gpu/GrPrimitiveProcessor.h" |
14 | #include "src/gpu/GrProcessor.h" |
15 | #include "src/gpu/GrProgramInfo.h" |
16 | #include "src/gpu/GrRenderTargetPriv.h" |
17 | #include "src/gpu/GrShaderCaps.h" |
18 | #include "src/gpu/GrTexturePriv.h" |
19 | #include "src/gpu/glsl/GrGLSLFragmentProcessor.h" |
20 | #include "src/gpu/glsl/GrGLSLFragmentShaderBuilder.h" |
21 | |
22 | enum { |
23 | kSamplerOrImageTypeKeyBits = 4 |
24 | }; |
25 | |
26 | static inline uint16_t texture_type_key(GrTextureType type) { |
27 | int value = UINT16_MAX; |
28 | switch (type) { |
29 | case GrTextureType::k2D: |
30 | value = 0; |
31 | break; |
32 | case GrTextureType::kExternal: |
33 | value = 1; |
34 | break; |
35 | case GrTextureType::kRectangle: |
36 | value = 2; |
37 | break; |
38 | default: |
39 | SK_ABORT("Unexpected texture type" ); |
40 | value = 3; |
41 | break; |
42 | } |
43 | SkASSERT((value & ((1 << kSamplerOrImageTypeKeyBits) - 1)) == value); |
44 | return SkToU16(value); |
45 | } |
46 | |
47 | static uint32_t sampler_key(GrTextureType textureType, const GrSwizzle& swizzle, |
48 | const GrCaps& caps) { |
49 | int samplerTypeKey = texture_type_key(textureType); |
50 | |
51 | static_assert(2 == sizeof(swizzle.asKey())); |
52 | uint16_t swizzleKey = 0; |
53 | if (caps.shaderCaps()->textureSwizzleAppliedInShader()) { |
54 | swizzleKey = swizzle.asKey(); |
55 | } |
56 | return SkToU32(samplerTypeKey | swizzleKey << kSamplerOrImageTypeKeyBits); |
57 | } |
58 | |
59 | static void add_fp_sampler_keys(GrProcessorKeyBuilder* b, const GrFragmentProcessor& fp, |
60 | const GrCaps& caps) { |
61 | int numTextureSamplers = fp.numTextureSamplers(); |
62 | if (!numTextureSamplers) { |
63 | return; |
64 | } |
65 | for (int i = 0; i < numTextureSamplers; ++i) { |
66 | const GrFragmentProcessor::TextureSampler& sampler = fp.textureSampler(i); |
67 | const GrBackendFormat& backendFormat = sampler.view().proxy()->backendFormat(); |
68 | |
69 | uint32_t samplerKey = sampler_key(backendFormat.textureType(), sampler.view().swizzle(), |
70 | caps); |
71 | b->add32(samplerKey); |
72 | |
73 | caps.addExtraSamplerKey(b, sampler.samplerState(), backendFormat); |
74 | } |
75 | } |
76 | |
77 | static void add_pp_sampler_keys(GrProcessorKeyBuilder* b, const GrPrimitiveProcessor& pp, |
78 | const GrCaps& caps) { |
79 | int numTextureSamplers = pp.numTextureSamplers(); |
80 | if (!numTextureSamplers) { |
81 | return; |
82 | } |
83 | for (int i = 0; i < numTextureSamplers; ++i) { |
84 | const GrPrimitiveProcessor::TextureSampler& sampler = pp.textureSampler(i); |
85 | const GrBackendFormat& backendFormat = sampler.backendFormat(); |
86 | |
87 | uint32_t samplerKey = sampler_key(backendFormat.textureType(), sampler.swizzle(), caps); |
88 | b->add32(samplerKey); |
89 | |
90 | caps.addExtraSamplerKey(b, sampler.samplerState(), backendFormat); |
91 | } |
92 | } |
93 | |
94 | /** |
95 | * A function which emits a meta key into the key builder. This is required because shader code may |
96 | * be dependent on properties of the effect that the effect itself doesn't use |
97 | * in its key (e.g. the pixel format of textures used). So we create a meta-key for |
98 | * every effect using this function. It is also responsible for inserting the effect's class ID |
99 | * which must be different for every GrProcessor subclass. It can fail if an effect uses too many |
100 | * transforms, etc, for the space allotted in the meta-key. NOTE, both FPs and GPs share this |
101 | * function because it is hairy, though FPs do not have attribs, and GPs do not have transforms |
102 | */ |
103 | static bool gen_fp_meta_key(const GrFragmentProcessor& fp, |
104 | const GrCaps& caps, |
105 | uint32_t transformKey, |
106 | GrProcessorKeyBuilder* b) { |
107 | size_t processorKeySize = b->size(); |
108 | uint32_t classID = fp.classID(); |
109 | |
110 | // Currently we allow 16 bits for the class id and the overall processor key size. |
111 | static const uint32_t kMetaKeyInvalidMask = ~((uint32_t)UINT16_MAX); |
112 | if ((processorKeySize | classID) & kMetaKeyInvalidMask) { |
113 | return false; |
114 | } |
115 | |
116 | add_fp_sampler_keys(b, fp, caps); |
117 | |
118 | uint32_t* key = b->add32n(2); |
119 | key[0] = (classID << 16) | SkToU32(processorKeySize); |
120 | key[1] = transformKey; |
121 | return true; |
122 | } |
123 | |
124 | static bool gen_pp_meta_key(const GrPrimitiveProcessor& pp, |
125 | const GrCaps& caps, |
126 | uint32_t transformKey, |
127 | GrProcessorKeyBuilder* b) { |
128 | size_t processorKeySize = b->size(); |
129 | uint32_t classID = pp.classID(); |
130 | |
131 | // Currently we allow 16 bits for the class id and the overall processor key size. |
132 | static const uint32_t kMetaKeyInvalidMask = ~((uint32_t)UINT16_MAX); |
133 | if ((processorKeySize | classID) & kMetaKeyInvalidMask) { |
134 | return false; |
135 | } |
136 | |
137 | add_pp_sampler_keys(b, pp, caps); |
138 | |
139 | uint32_t* key = b->add32n(2); |
140 | key[0] = (classID << 16) | SkToU32(processorKeySize); |
141 | key[1] = transformKey; |
142 | return true; |
143 | } |
144 | |
145 | static bool gen_xp_meta_key(const GrXferProcessor& xp, GrProcessorKeyBuilder* b) { |
146 | size_t processorKeySize = b->size(); |
147 | uint32_t classID = xp.classID(); |
148 | |
149 | // Currently we allow 16 bits for the class id and the overall processor key size. |
150 | static const uint32_t kMetaKeyInvalidMask = ~((uint32_t)UINT16_MAX); |
151 | if ((processorKeySize | classID) & kMetaKeyInvalidMask) { |
152 | return false; |
153 | } |
154 | |
155 | b->add32((classID << 16) | SkToU32(processorKeySize)); |
156 | return true; |
157 | } |
158 | |
159 | static bool gen_frag_proc_and_meta_keys(const GrPrimitiveProcessor& primProc, |
160 | const GrFragmentProcessor& fp, |
161 | const GrCaps& caps, |
162 | GrProcessorKeyBuilder* b) { |
163 | for (int i = 0; i < fp.numChildProcessors(); ++i) { |
164 | if (!gen_frag_proc_and_meta_keys(primProc, fp.childProcessor(i), caps, b)) { |
165 | return false; |
166 | } |
167 | } |
168 | |
169 | fp.getGLSLProcessorKey(*caps.shaderCaps(), b); |
170 | |
171 | return gen_fp_meta_key(fp, caps, primProc.computeCoordTransformsKey(fp), b); |
172 | } |
173 | |
174 | bool GrProgramDesc::Build(GrProgramDesc* desc, const GrRenderTarget* renderTarget, |
175 | const GrProgramInfo& programInfo, const GrCaps& caps) { |
176 | |
177 | #ifdef SK_DEBUG |
178 | if (renderTarget) { |
179 | SkASSERT(programInfo.backendFormat() == renderTarget->backendFormat()); |
180 | } |
181 | #endif |
182 | |
183 | // The descriptor is used as a cache key. Thus when a field of the |
184 | // descriptor will not affect program generation (because of the attribute |
185 | // bindings in use or other descriptor field settings) it should be set |
186 | // to a canonical value to avoid duplicate programs with different keys. |
187 | |
188 | static_assert(0 == kProcessorKeysOffset % sizeof(uint32_t)); |
189 | // Make room for everything up to the effect keys. |
190 | desc->key().reset(); |
191 | desc->key().push_back_n(kProcessorKeysOffset); |
192 | |
193 | GrProcessorKeyBuilder b(&desc->key()); |
194 | |
195 | programInfo.primProc().getGLSLProcessorKey(*caps.shaderCaps(), &b); |
196 | programInfo.primProc().getAttributeKey(&b); |
197 | if (!gen_pp_meta_key(programInfo.primProc(), caps, 0, &b)) { |
198 | desc->key().reset(); |
199 | return false; |
200 | } |
201 | |
202 | for (int i = 0; i < programInfo.pipeline().numFragmentProcessors(); ++i) { |
203 | const GrFragmentProcessor& fp = programInfo.pipeline().getFragmentProcessor(i); |
204 | if (!gen_frag_proc_and_meta_keys(programInfo.primProc(), fp, caps, &b)) { |
205 | desc->key().reset(); |
206 | return false; |
207 | } |
208 | } |
209 | |
210 | const GrXferProcessor& xp = programInfo.pipeline().getXferProcessor(); |
211 | const GrSurfaceOrigin* originIfDstTexture = nullptr; |
212 | GrSurfaceOrigin origin; |
213 | if (programInfo.pipeline().dstProxyView().proxy()) { |
214 | origin = programInfo.pipeline().dstProxyView().origin(); |
215 | originIfDstTexture = &origin; |
216 | } |
217 | xp.getGLSLProcessorKey(*caps.shaderCaps(), &b, originIfDstTexture); |
218 | if (!gen_xp_meta_key(xp, &b)) { |
219 | desc->key().reset(); |
220 | return false; |
221 | } |
222 | |
223 | if (programInfo.requestedFeatures() & GrProcessor::CustomFeatures::kSampleLocations) { |
224 | SkASSERT(programInfo.pipeline().isHWAntialiasState()); |
225 | b.add32(renderTarget->renderTargetPriv().getSamplePatternKey()); |
226 | } |
227 | |
228 | // --------DO NOT MOVE HEADER ABOVE THIS LINE-------------------------------------------------- |
229 | // Because header is a pointer into the dynamic array, we can't push any new data into the key |
230 | // below here. |
231 | KeyHeader* = desc->atOffset<KeyHeader, kHeaderOffset>(); |
232 | |
233 | // make sure any padding in the header is zeroed. |
234 | memset(header, 0, kHeaderSize); |
235 | header->fWriteSwizzle = programInfo.pipeline().writeSwizzle().asKey(); |
236 | header->fColorFragmentProcessorCnt = programInfo.pipeline().numColorFragmentProcessors(); |
237 | header->fCoverageFragmentProcessorCnt = programInfo.pipeline().numCoverageFragmentProcessors(); |
238 | // Fail if the client requested more processors than the key can fit. |
239 | if (header->fColorFragmentProcessorCnt != programInfo.pipeline().numColorFragmentProcessors() || |
240 | header->fCoverageFragmentProcessorCnt != |
241 | programInfo.pipeline().numCoverageFragmentProcessors()) { |
242 | desc->key().reset(); |
243 | return false; |
244 | } |
245 | // If we knew the shader won't depend on origin, we could skip this (and use the same program |
246 | // for both origins). Instrumenting all fragment processors would be difficult and error prone. |
247 | header->fSurfaceOriginKey = |
248 | GrGLSLFragmentShaderBuilder::KeyForSurfaceOrigin(programInfo.origin()); |
249 | header->fProcessorFeatures = (uint8_t)programInfo.requestedFeatures(); |
250 | // Ensure enough bits. |
251 | SkASSERT(header->fProcessorFeatures == (int) programInfo.requestedFeatures()); |
252 | header->fSnapVerticesToPixelCenters = programInfo.pipeline().snapVerticesToPixelCenters(); |
253 | // The base descriptor only stores whether or not the primitiveType is kPoints. Backend- |
254 | // specific versions (e.g., Vulkan) require more detail |
255 | header->fHasPointSize = (programInfo.primitiveType() == GrPrimitiveType::kPoints); |
256 | |
257 | header->fInitialKeyLength = desc->keyLength(); |
258 | // Fail if the initial key length won't fit in 27 bits. |
259 | if (header->fInitialKeyLength != desc->keyLength()) { |
260 | desc->key().reset(); |
261 | return false; |
262 | } |
263 | |
264 | return true; |
265 | } |
266 | |