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/vk/GrVkUniformHandler.h"
9
10 #include "src/gpu/GrTexture.h"
11 #include "src/gpu/GrUtil.h"
12 #include "src/gpu/glsl/GrGLSLProgramBuilder.h"
13 #include "src/gpu/vk/GrVkGpu.h"
14 #include "src/gpu/vk/GrVkPipelineStateBuilder.h"
15 #include "src/gpu/vk/GrVkTexture.h"
16
17 // To determine whether a current offset is aligned, we can just 'and' the lowest bits with the
18 // alignment mask. A value of 0 means aligned, any other value is how many bytes past alignment we
19 // are. This works since all alignments are powers of 2. The mask is always (alignment - 1).
20 // This alignment mask will give correct alignments for using the std430 block layout. If you want
21 // the std140 alignment, you can use this, but then make sure if you have an array type it is
22 // aligned to 16 bytes (i.e. has mask of 0xF).
23 // These are designated in the Vulkan spec, section 14.5.4 "Offset and Stride Assignment".
24 // https://www.khronos.org/registry/vulkan/specs/1.0-wsi_extensions/html/vkspec.html#interfaces-resources-layout
sksltype_to_alignment_mask(SkSLType type)25 static uint32_t sksltype_to_alignment_mask(SkSLType type) {
26 switch(type) {
27 case SkSLType::kShort: // fall through
28 case SkSLType::kUShort:
29 return 0x1;
30 case SkSLType::kShort2: // fall through
31 case SkSLType::kUShort2:
32 return 0x3;
33 case SkSLType::kShort3: // fall through
34 case SkSLType::kShort4:
35 case SkSLType::kUShort3:
36 case SkSLType::kUShort4:
37 return 0x7;
38 case SkSLType::kInt:
39 case SkSLType::kUInt:
40 return 0x3;
41 case SkSLType::kInt2:
42 case SkSLType::kUInt2:
43 return 0x7;
44 case SkSLType::kInt3:
45 case SkSLType::kUInt3:
46 case SkSLType::kInt4:
47 case SkSLType::kUInt4:
48 return 0xF;
49 case SkSLType::kHalf: // fall through
50 case SkSLType::kFloat:
51 return 0x3;
52 case SkSLType::kHalf2: // fall through
53 case SkSLType::kFloat2:
54 return 0x7;
55 case SkSLType::kHalf3: // fall through
56 case SkSLType::kFloat3:
57 return 0xF;
58 case SkSLType::kHalf4: // fall through
59 case SkSLType::kFloat4:
60 return 0xF;
61 case SkSLType::kHalf2x2: // fall through
62 case SkSLType::kFloat2x2:
63 return 0x7;
64 case SkSLType::kHalf3x3: // fall through
65 case SkSLType::kFloat3x3:
66 return 0xF;
67 case SkSLType::kHalf4x4: // fall through
68 case SkSLType::kFloat4x4:
69 return 0xF;
70
71 // This query is only valid for certain types.
72 case SkSLType::kVoid:
73 case SkSLType::kBool:
74 case SkSLType::kBool2:
75 case SkSLType::kBool3:
76 case SkSLType::kBool4:
77 case SkSLType::kTexture2DSampler:
78 case SkSLType::kTextureExternalSampler:
79 case SkSLType::kTexture2DRectSampler:
80 case SkSLType::kSampler:
81 case SkSLType::kTexture2D:
82 case SkSLType::kInput:
83 break;
84 }
85 SK_ABORT("Unexpected type");
86 }
87
88 /** Returns the size in bytes taken up in vulkanbuffers for SkSLTypes. */
sksltype_to_vk_size(SkSLType type,int layout)89 static inline uint32_t sksltype_to_vk_size(SkSLType type, int layout) {
90 switch(type) {
91 case SkSLType::kShort:
92 return sizeof(int16_t);
93 case SkSLType::kShort2:
94 return 2 * sizeof(int16_t);
95 case SkSLType::kShort3:
96 return 3 * sizeof(int16_t);
97 case SkSLType::kShort4:
98 return 4 * sizeof(int16_t);
99 case SkSLType::kUShort:
100 return sizeof(uint16_t);
101 case SkSLType::kUShort2:
102 return 2 * sizeof(uint16_t);
103 case SkSLType::kUShort3:
104 return 3 * sizeof(uint16_t);
105 case SkSLType::kUShort4:
106 return 4 * sizeof(uint16_t);
107 case SkSLType::kHalf: // fall through
108 case SkSLType::kFloat:
109 return sizeof(float);
110 case SkSLType::kHalf2: // fall through
111 case SkSLType::kFloat2:
112 return 2 * sizeof(float);
113 case SkSLType::kHalf3: // fall through
114 case SkSLType::kFloat3:
115 return 3 * sizeof(float);
116 case SkSLType::kHalf4: // fall through
117 case SkSLType::kFloat4:
118 return 4 * sizeof(float);
119 case SkSLType::kInt: // fall through
120 case SkSLType::kUInt:
121 return sizeof(int32_t);
122 case SkSLType::kInt2: // fall through
123 case SkSLType::kUInt2:
124 return 2 * sizeof(int32_t);
125 case SkSLType::kInt3: // fall through
126 case SkSLType::kUInt3:
127 return 3 * sizeof(int32_t);
128 case SkSLType::kInt4: // fall through
129 case SkSLType::kUInt4:
130 return 4 * sizeof(int32_t);
131 case SkSLType::kHalf2x2: // fall through
132 case SkSLType::kFloat2x2:
133 if (layout == GrVkUniformHandler::kStd430Layout) {
134 return 4 * sizeof(float);
135 } else {
136 return 8 * sizeof(float);
137 }
138 case SkSLType::kHalf3x3: // fall through
139 case SkSLType::kFloat3x3:
140 return 12 * sizeof(float);
141 case SkSLType::kHalf4x4: // fall through
142 case SkSLType::kFloat4x4:
143 return 16 * sizeof(float);
144
145 // This query is only valid for certain types.
146 case SkSLType::kVoid:
147 case SkSLType::kBool:
148 case SkSLType::kBool2:
149 case SkSLType::kBool3:
150 case SkSLType::kBool4:
151 case SkSLType::kTexture2DSampler:
152 case SkSLType::kTextureExternalSampler:
153 case SkSLType::kTexture2DRectSampler:
154 case SkSLType::kSampler:
155 case SkSLType::kTexture2D:
156 case SkSLType::kInput:
157 break;
158 }
159 SK_ABORT("Unexpected type");
160 }
161
162 // Given the current offset into the ubo data, calculate the offset for the uniform we're trying to
163 // add taking into consideration all alignment requirements. The uniformOffset is set to the offset
164 // for the new uniform, and currentOffset is updated to be the offset to the end of the new uniform.
get_aligned_offset(uint32_t * currentOffset,SkSLType type,int arrayCount,int layout)165 static uint32_t get_aligned_offset(uint32_t* currentOffset,
166 SkSLType type,
167 int arrayCount,
168 int layout) {
169 uint32_t alignmentMask = sksltype_to_alignment_mask(type);
170 // For std140 layout we must make arrays align to 16 bytes.
171 if (layout == GrVkUniformHandler::kStd140Layout && (arrayCount || type == SkSLType::kFloat2x2)) {
172 alignmentMask = 0xF;
173 }
174 uint32_t offsetDiff = *currentOffset & alignmentMask;
175 if (offsetDiff != 0) {
176 offsetDiff = alignmentMask - offsetDiff + 1;
177 }
178 int32_t uniformOffset = *currentOffset + offsetDiff;
179 SkASSERT(sizeof(float) == 4);
180 if (arrayCount) {
181 // TODO: this shouldn't be necessary for std430
182 uint32_t elementSize = std::max<uint32_t>(16, sksltype_to_vk_size(type, layout));
183 SkASSERT(0 == (elementSize & 0xF));
184 *currentOffset = uniformOffset + elementSize * arrayCount;
185 } else {
186 *currentOffset = uniformOffset + sksltype_to_vk_size(type, layout);
187 }
188 return uniformOffset;
189 }
190
~GrVkUniformHandler()191 GrVkUniformHandler::~GrVkUniformHandler() {
192 for (VkUniformInfo& sampler : fSamplers.items()) {
193 if (sampler.fImmutableSampler) {
194 sampler.fImmutableSampler->unref();
195 sampler.fImmutableSampler = nullptr;
196 }
197 }
198 }
199
internalAddUniformArray(const GrFragmentProcessor * owner,uint32_t visibility,SkSLType type,const char * name,bool mangleName,int arrayCount,const char ** outName)200 GrGLSLUniformHandler::UniformHandle GrVkUniformHandler::internalAddUniformArray(
201 const GrFragmentProcessor* owner,
202 uint32_t visibility,
203 SkSLType type,
204 const char* name,
205 bool mangleName,
206 int arrayCount,
207 const char** outName) {
208 SkASSERT(name && strlen(name));
209 SkASSERT(SkSLTypeCanBeUniformValue(type));
210
211 // TODO this is a bit hacky, lets think of a better way. Basically we need to be able to use
212 // the uniform view matrix name in the GP, and the GP is immutable so it has to tell the PB
213 // exactly what name it wants to use for the uniform view matrix. If we prefix anythings, then
214 // the names will mismatch. I think the correct solution is to have all GPs which need the
215 // uniform view matrix, they should upload the view matrix in their setData along with regular
216 // uniforms.
217 char prefix = 'u';
218 if ('u' == name[0] || !strncmp(name, GR_NO_MANGLE_PREFIX, strlen(GR_NO_MANGLE_PREFIX))) {
219 prefix = '\0';
220 }
221 SkString resolvedName = fProgramBuilder->nameVariable(prefix, name, mangleName);
222
223 VkUniformInfo tempInfo;
224 tempInfo.fVariable = GrShaderVar{std::move(resolvedName),
225 type,
226 GrShaderVar::TypeModifier::None,
227 arrayCount};
228
229 tempInfo.fVisibility = visibility;
230 tempInfo.fOwner = owner;
231 tempInfo.fRawName = SkString(name);
232
233 for (int layout = 0; layout < kLayoutCount; ++layout) {
234 tempInfo.fOffsets[layout] = get_aligned_offset(&fCurrentOffsets[layout],
235 type,
236 arrayCount,
237 layout);
238 }
239
240 fUniforms.push_back(tempInfo);
241
242 if (outName) {
243 *outName = fUniforms.back().fVariable.c_str();
244 }
245
246 return GrGLSLUniformHandler::UniformHandle(fUniforms.count() - 1);
247 }
248
addSampler(const GrBackendFormat & backendFormat,GrSamplerState state,const skgpu::Swizzle & swizzle,const char * name,const GrShaderCaps * shaderCaps)249 GrGLSLUniformHandler::SamplerHandle GrVkUniformHandler::addSampler(
250 const GrBackendFormat& backendFormat, GrSamplerState state, const skgpu::Swizzle& swizzle,
251 const char* name, const GrShaderCaps* shaderCaps) {
252 SkASSERT(name && strlen(name));
253
254 const char prefix = 'u';
255 SkString mangleName = fProgramBuilder->nameVariable(prefix, name, /*mangle=*/true);
256
257 SkString layoutQualifier;
258 layoutQualifier.appendf("set=%d, binding=%d", kSamplerDescSet, fSamplers.count());
259
260 VkUniformInfo tempInfo;
261 tempInfo.fVariable =
262 GrShaderVar{std::move(mangleName),
263 SkSLCombinedSamplerTypeForTextureType(backendFormat.textureType()),
264 GrShaderVar::TypeModifier::Uniform,
265 GrShaderVar::kNonArray,
266 std::move(layoutQualifier),
267 SkString()};
268
269 tempInfo.fVisibility = kFragment_GrShaderFlag;
270 tempInfo.fOwner = nullptr;
271 tempInfo.fRawName = SkString(name);
272 tempInfo.fOffsets[0] = 0;
273 tempInfo.fOffsets[1] = 0;
274
275 fSamplers.push_back(tempInfo);
276
277 // Check if we are dealing with an external texture and store the needed information if so.
278 auto ycbcrInfo = backendFormat.getVkYcbcrConversionInfo();
279 if (ycbcrInfo && ycbcrInfo->isValid()) {
280 GrVkGpu* gpu = static_cast<GrVkPipelineStateBuilder*>(fProgramBuilder)->gpu();
281 GrVkSampler* sampler = gpu->resourceProvider().findOrCreateCompatibleSampler(state,
282 *ycbcrInfo);
283 fSamplers.back().fImmutableSampler = sampler;
284 if (!sampler) {
285 return {};
286 }
287 }
288
289 fSamplerSwizzles.push_back(swizzle);
290 SkASSERT(fSamplerSwizzles.count() == fSamplers.count());
291 return GrGLSLUniformHandler::SamplerHandle(fSamplers.count() - 1);
292 }
293
addInputSampler(const skgpu::Swizzle & swizzle,const char * name)294 GrGLSLUniformHandler::SamplerHandle GrVkUniformHandler::addInputSampler(
295 const skgpu::Swizzle& swizzle, const char* name) {
296 SkASSERT(name && strlen(name));
297 SkASSERT(fInputUniform.fVariable.getType() == SkSLType::kVoid);
298
299 const char prefix = 'u';
300 SkString mangleName = fProgramBuilder->nameVariable(prefix, name, /*mangle=*/true);
301
302 SkString layoutQualifier;
303 layoutQualifier.appendf("input_attachment_index=%d, set=%d, binding=%d",
304 kDstInputAttachmentIndex, kInputDescSet, kInputBinding);
305
306 fInputUniform = {
307 GrShaderVar{std::move(mangleName), SkSLType::kInput, GrShaderVar::TypeModifier::Uniform,
308 GrShaderVar::kNonArray, std::move(layoutQualifier), SkString()},
309 kFragment_GrShaderFlag, nullptr, SkString(name)};
310 fInputSwizzle = swizzle;
311 return GrGLSLUniformHandler::SamplerHandle(0);
312 }
313
appendUniformDecls(GrShaderFlags visibility,SkString * out) const314 void GrVkUniformHandler::appendUniformDecls(GrShaderFlags visibility, SkString* out) const {
315 for (const VkUniformInfo& sampler : fSamplers.items()) {
316 SkASSERT(sampler.fVariable.getType() == SkSLType::kTexture2DSampler ||
317 sampler.fVariable.getType() == SkSLType::kTextureExternalSampler);
318 if (visibility == sampler.fVisibility) {
319 sampler.fVariable.appendDecl(fProgramBuilder->shaderCaps(), out);
320 out->append(";\n");
321 }
322 }
323 if (fInputUniform.fVariable.getType() == SkSLType::kInput) {
324 if (visibility == fInputUniform.fVisibility) {
325 SkASSERT(visibility == kFragment_GrShaderFlag);
326 fInputUniform.fVariable.appendDecl(fProgramBuilder->shaderCaps(), out);
327 out->append(";\n");
328 }
329 }
330
331 #ifdef SK_DEBUG
332 bool firstOffsetCheck = false;
333 for (const VkUniformInfo& localUniform : fUniforms.items()) {
334 if (!firstOffsetCheck) {
335 // Check to make sure we are starting our offset at 0 so the offset qualifier we
336 // set on each variable in the uniform block is valid.
337 SkASSERT(0 == localUniform.fOffsets[kStd140Layout] &&
338 0 == localUniform.fOffsets[kStd430Layout]);
339 firstOffsetCheck = true;
340 }
341 }
342 #endif
343
344 // At this point we determine whether we'll be using push constants based on the
345 // uniforms set so far. Later checks will use the internal bool we set here to
346 // keep things consistent.
347 this->determineIfUsePushConstants();
348 SkString uniformsString;
349 for (const VkUniformInfo& localUniform : fUniforms.items()) {
350 if (visibility & localUniform.fVisibility) {
351 if (SkSLTypeCanBeUniformValue(localUniform.fVariable.getType())) {
352 Layout layout = fUsePushConstants ? kStd430Layout : kStd140Layout;
353 uniformsString.appendf("layout(offset=%d) ", localUniform.fOffsets[layout]);
354 localUniform.fVariable.appendDecl(fProgramBuilder->shaderCaps(), &uniformsString);
355 uniformsString.append(";\n");
356 }
357 }
358 }
359
360 if (!uniformsString.isEmpty()) {
361 if (fUsePushConstants) {
362 out->append("layout (push_constant) ");
363 } else {
364 out->appendf("layout (set=%d, binding=%d) ",
365 kUniformBufferDescSet, kUniformBinding);
366 }
367 out->append("uniform uniformBuffer\n{\n");
368 out->appendf("%s\n};\n", uniformsString.c_str());
369 }
370 }
371
getRTFlipOffset() const372 uint32_t GrVkUniformHandler::getRTFlipOffset() const {
373 Layout layout = fUsePushConstants ? kStd430Layout : kStd140Layout;
374 uint32_t currentOffset = fCurrentOffsets[layout];
375 return get_aligned_offset(¤tOffset, SkSLType::kFloat2, 0, layout);
376 }
377
determineIfUsePushConstants() const378 void GrVkUniformHandler::determineIfUsePushConstants() const {
379 // We may insert a uniform for flipping origin-sensitive language features (e.g. sk_FragCoord).
380 // We won't know that for sure until then but we need to make this determination now,
381 // so assume we will need it.
382 static constexpr uint32_t kPad = 2*sizeof(float);
383 fUsePushConstants =
384 fCurrentOffsets[kStd430Layout] > 0 &&
385 fCurrentOffsets[kStd430Layout] + kPad <= fProgramBuilder->caps()->maxPushConstantsSize();
386 }
387