1/* 2 * Copyright 2015 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 "GrVkCaps.h" 9#include "GrRenderTarget.h" 10#include "GrShaderCaps.h" 11#include "GrVkUtil.h" 12#include "vk/GrVkBackendContext.h" 13#include "vk/GrVkInterface.h" 14 15GrVkCaps::GrVkCaps(const GrContextOptions& contextOptions, const GrVkInterface* vkInterface, 16 VkPhysicalDevice physDev, uint32_t featureFlags, uint32_t extensionFlags) 17 : INHERITED(contextOptions) { 18 fCanUseGLSLForShaderModule = false; 19 fMustDoCopiesFromOrigin = false; 20 fSupportsCopiesAsDraws = false; 21 fMustSubmitCommandsBeforeCopyOp = false; 22 fMustSleepOnTearDown = false; 23 fNewSecondaryCBOnPipelineChange = false; 24 25 /************************************************************************** 26 * GrDrawTargetCaps fields 27 **************************************************************************/ 28 fMipMapSupport = true; // always available in Vulkan 29 fSRGBSupport = true; // always available in Vulkan 30 fNPOTTextureTileSupport = true; // always available in Vulkan 31 fTwoSidedStencilSupport = true; // always available in Vulkan 32 fStencilWrapOpsSupport = true; // always available in Vulkan 33 fDiscardRenderTargetSupport = true; 34 fReuseScratchTextures = true; //TODO: figure this out 35 fGpuTracingSupport = false; //TODO: figure this out 36 fCompressedTexSubImageSupport = false; //TODO: figure this out 37 fOversizedStencilSupport = false; //TODO: figure this out 38 39 fUseDrawInsteadOfClear = false; 40 fFenceSyncSupport = true; // always available in Vulkan 41 fCrossContextTextureSupport = false; // TODO: Add thread-safe memory pools so we can enable this 42 43 fMapBufferFlags = kNone_MapFlags; //TODO: figure this out 44 fBufferMapThreshold = SK_MaxS32; //TODO: figure this out 45 46 fMaxRenderTargetSize = 4096; // minimum required by spec 47 fMaxTextureSize = 4096; // minimum required by spec 48 fMaxColorSampleCount = 4; // minimum required by spec 49 fMaxStencilSampleCount = 4; // minimum required by spec 50 51 fShaderCaps.reset(new GrShaderCaps(contextOptions)); 52 53 this->init(contextOptions, vkInterface, physDev, featureFlags, extensionFlags); 54} 55 56bool GrVkCaps::initDescForDstCopy(const GrRenderTarget* src, GrSurfaceDesc* desc) const { 57 // We can always succeed here with either a CopyImage (none msaa src) or ResolveImage (msaa). 58 // For CopyImage we can make a simple texture, for ResolveImage we require the dst to be a 59 // render target as well. 60 desc->fOrigin = src->origin(); 61 desc->fConfig = src->config(); 62 if (src->numColorSamples() > 1 || (src->asTexture() && this->supportsCopiesAsDraws())) { 63 desc->fFlags = kRenderTarget_GrSurfaceFlag; 64 } else { 65 // Just going to use CopyImage here 66 desc->fFlags = kNone_GrSurfaceFlags; 67 } 68 69 return true; 70} 71 72void GrVkCaps::init(const GrContextOptions& contextOptions, const GrVkInterface* vkInterface, 73 VkPhysicalDevice physDev, uint32_t featureFlags, uint32_t extensionFlags) { 74 75 VkPhysicalDeviceProperties properties; 76 GR_VK_CALL(vkInterface, GetPhysicalDeviceProperties(physDev, &properties)); 77 78 VkPhysicalDeviceMemoryProperties memoryProperties; 79 GR_VK_CALL(vkInterface, GetPhysicalDeviceMemoryProperties(physDev, &memoryProperties)); 80 81 this->initGrCaps(properties, memoryProperties, featureFlags); 82 this->initShaderCaps(properties, featureFlags); 83 this->initConfigTable(vkInterface, physDev); 84 this->initStencilFormat(vkInterface, physDev); 85 86 if (SkToBool(extensionFlags & kNV_glsl_shader_GrVkExtensionFlag)) { 87 // Currently disabling this feature since it does not play well with validation layers which 88 // expect a SPIR-V shader 89 // fCanUseGLSLForShaderModule = true; 90 } 91 92 if (kQualcomm_VkVendor == properties.vendorID) { 93 fMustDoCopiesFromOrigin = true; 94 } 95 96 if (kNvidia_VkVendor == properties.vendorID) { 97 fSupportsCopiesAsDraws = true; 98 fMustSubmitCommandsBeforeCopyOp = true; 99 } 100 101#if defined(SK_BUILD_FOR_WIN) 102 if (kNvidia_VkVendor == properties.vendorID) { 103 fMustSleepOnTearDown = true; 104 } 105#elif defined(SK_BUILD_FOR_ANDROID) 106 if (kImagination_VkVendor == properties.vendorID) { 107 fMustSleepOnTearDown = true; 108 } 109#endif 110 111 this->applyOptionsOverrides(contextOptions); 112 fShaderCaps->applyOptionsOverrides(contextOptions); 113} 114 115int get_max_sample_count(VkSampleCountFlags flags) { 116 SkASSERT(flags & VK_SAMPLE_COUNT_1_BIT); 117 if (!(flags & VK_SAMPLE_COUNT_2_BIT)) { 118 return 0; 119 } 120 if (!(flags & VK_SAMPLE_COUNT_4_BIT)) { 121 return 2; 122 } 123 if (!(flags & VK_SAMPLE_COUNT_8_BIT)) { 124 return 4; 125 } 126 if (!(flags & VK_SAMPLE_COUNT_16_BIT)) { 127 return 8; 128 } 129 if (!(flags & VK_SAMPLE_COUNT_32_BIT)) { 130 return 16; 131 } 132 if (!(flags & VK_SAMPLE_COUNT_64_BIT)) { 133 return 32; 134 } 135 return 64; 136} 137 138void GrVkCaps::initSampleCount(const VkPhysicalDeviceProperties& properties) { 139 VkSampleCountFlags colorSamples = properties.limits.framebufferColorSampleCounts; 140 VkSampleCountFlags stencilSamples = properties.limits.framebufferStencilSampleCounts; 141 142 fMaxColorSampleCount = get_max_sample_count(colorSamples); 143 fMaxStencilSampleCount = get_max_sample_count(stencilSamples); 144} 145 146void GrVkCaps::initGrCaps(const VkPhysicalDeviceProperties& properties, 147 const VkPhysicalDeviceMemoryProperties& memoryProperties, 148 uint32_t featureFlags) { 149 // So GPUs, like AMD, are reporting MAX_INT support vertex attributes. In general, there is no 150 // need for us ever to support that amount, and it makes tests which tests all the vertex 151 // attribs timeout looping over that many. For now, we'll cap this at 64 max and can raise it if 152 // we ever find that need. 153 static const uint32_t kMaxVertexAttributes = 64; 154 fMaxVertexAttributes = SkTMin(properties.limits.maxVertexInputAttributes, kMaxVertexAttributes); 155 // AMD advertises support for MAX_UINT vertex input attributes, but in reality only supports 32. 156 if (kAMD_VkVendor == properties.vendorID) { 157 fMaxVertexAttributes = SkTMin(fMaxVertexAttributes, 32); 158 } 159 160 // We could actually query and get a max size for each config, however maxImageDimension2D will 161 // give the minimum max size across all configs. So for simplicity we will use that for now. 162 fMaxRenderTargetSize = SkTMin(properties.limits.maxImageDimension2D, (uint32_t)INT_MAX); 163 fMaxTextureSize = SkTMin(properties.limits.maxImageDimension2D, (uint32_t)INT_MAX); 164 165 this->initSampleCount(properties); 166 167 // Assuming since we will always map in the end to upload the data we might as well just map 168 // from the get go. There is no hard data to suggest this is faster or slower. 169 fBufferMapThreshold = 0; 170 171 fMapBufferFlags = kCanMap_MapFlag | kSubset_MapFlag; 172 173 fStencilWrapOpsSupport = true; 174 fOversizedStencilSupport = true; 175 fSampleShadingSupport = SkToBool(featureFlags & kSampleRateShading_GrVkFeatureFlag); 176 177 // AMD seems to have issues binding new VkPipelines inside a secondary command buffer. 178 // Current workaround is to use a different secondary command buffer for each new VkPipeline. 179 if (kAMD_VkVendor == properties.vendorID) { 180 fNewSecondaryCBOnPipelineChange = true; 181 } 182} 183 184void GrVkCaps::initShaderCaps(const VkPhysicalDeviceProperties& properties, uint32_t featureFlags) { 185 GrShaderCaps* shaderCaps = fShaderCaps.get(); 186 shaderCaps->fVersionDeclString = "#version 330\n"; 187 188 189 // fConfigOutputSwizzle will default to RGBA so we only need to set it for alpha only config. 190 for (int i = 0; i < kGrPixelConfigCnt; ++i) { 191 GrPixelConfig config = static_cast<GrPixelConfig>(i); 192 if (GrPixelConfigIsAlphaOnly(config)) { 193 shaderCaps->fConfigTextureSwizzle[i] = GrSwizzle::RRRR(); 194 shaderCaps->fConfigOutputSwizzle[i] = GrSwizzle::AAAA(); 195 } else { 196 if (kGray_8_GrPixelConfig == config) { 197 shaderCaps->fConfigTextureSwizzle[i] = GrSwizzle::RRRA(); 198 } else if (kRGBA_4444_GrPixelConfig == config) { 199 // The vulkan spec does not require R4G4B4A4 to be supported for texturing so we 200 // store the data in a B4G4R4A4 texture and then swizzle it when doing texture reads 201 // or writing to outputs. Since we're not actually changing the data at all, the 202 // only extra work is the swizzle in the shader for all operations. 203 shaderCaps->fConfigTextureSwizzle[i] = GrSwizzle::BGRA(); 204 shaderCaps->fConfigOutputSwizzle[i] = GrSwizzle::BGRA(); 205 } else { 206 shaderCaps->fConfigTextureSwizzle[i] = GrSwizzle::RGBA(); 207 } 208 } 209 } 210 211 if (kImagination_VkVendor == properties.vendorID) { 212 shaderCaps->fAtan2ImplementedAsAtanYOverX = true; 213 } 214 215 // Vulkan is based off ES 3.0 so the following should all be supported 216 shaderCaps->fUsesPrecisionModifiers = true; 217 shaderCaps->fFlatInterpolationSupport = true; 218 219 // GrShaderCaps 220 221 shaderCaps->fShaderDerivativeSupport = true; 222 shaderCaps->fGeometryShaderSupport = SkToBool(featureFlags & kGeometryShader_GrVkFeatureFlag); 223 224 shaderCaps->fDualSourceBlendingSupport = SkToBool(featureFlags & kDualSrcBlend_GrVkFeatureFlag); 225 if (kAMD_VkVendor == properties.vendorID) { 226 // Currently DualSourceBlending is not working on AMD. vkCreateGraphicsPipeline fails when 227 // using a draw with dual source. Looking into whether it is driver bug or issue with our 228 // SPIR-V. Bug skia:6405 229 shaderCaps->fDualSourceBlendingSupport = false; 230 } 231 232 shaderCaps->fIntegerSupport = true; 233 234 // Assume the minimum precisions mandated by the SPIR-V spec. 235 shaderCaps->fShaderPrecisionVaries = true; 236 for (int s = 0; s < kGrShaderTypeCount; ++s) { 237 auto& highp = shaderCaps->fFloatPrecisions[s][kHigh_GrSLPrecision]; 238 highp.fLogRangeLow = highp.fLogRangeHigh = 127; 239 highp.fBits = 23; 240 241 auto& mediump = shaderCaps->fFloatPrecisions[s][kMedium_GrSLPrecision]; 242 mediump.fLogRangeLow = mediump.fLogRangeHigh = 14; 243 mediump.fBits = 10; 244 245 shaderCaps->fFloatPrecisions[s][kLow_GrSLPrecision] = mediump; 246 } 247 shaderCaps->initSamplerPrecisionTable(); 248 249 shaderCaps->fMaxVertexSamplers = 250 shaderCaps->fMaxGeometrySamplers = 251 shaderCaps->fMaxFragmentSamplers = SkTMin( 252 SkTMin(properties.limits.maxPerStageDescriptorSampledImages, 253 properties.limits.maxPerStageDescriptorSamplers), 254 (uint32_t)INT_MAX); 255 shaderCaps->fMaxCombinedSamplers = SkTMin( 256 SkTMin(properties.limits.maxDescriptorSetSampledImages, 257 properties.limits.maxDescriptorSetSamplers), 258 (uint32_t)INT_MAX); 259} 260 261bool stencil_format_supported(const GrVkInterface* interface, 262 VkPhysicalDevice physDev, 263 VkFormat format) { 264 VkFormatProperties props; 265 memset(&props, 0, sizeof(VkFormatProperties)); 266 GR_VK_CALL(interface, GetPhysicalDeviceFormatProperties(physDev, format, &props)); 267 return SkToBool(VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT & props.optimalTilingFeatures); 268} 269 270void GrVkCaps::initStencilFormat(const GrVkInterface* interface, VkPhysicalDevice physDev) { 271 // List of legal stencil formats (though perhaps not supported on 272 // the particular gpu/driver) from most preferred to least. We are guaranteed to have either 273 // VK_FORMAT_D24_UNORM_S8_UINT or VK_FORMAT_D32_SFLOAT_S8_UINT. VK_FORMAT_D32_SFLOAT_S8_UINT 274 // can optionally have 24 unused bits at the end so we assume the total bits is 64. 275 static const StencilFormat 276 // internal Format stencil bits total bits packed? 277 gS8 = { VK_FORMAT_S8_UINT, 8, 8, false }, 278 gD24S8 = { VK_FORMAT_D24_UNORM_S8_UINT, 8, 32, true }, 279 gD32S8 = { VK_FORMAT_D32_SFLOAT_S8_UINT, 8, 64, true }; 280 281 if (stencil_format_supported(interface, physDev, VK_FORMAT_S8_UINT)) { 282 fPreferedStencilFormat = gS8; 283 } else if (stencil_format_supported(interface, physDev, VK_FORMAT_D24_UNORM_S8_UINT)) { 284 fPreferedStencilFormat = gD24S8; 285 } else { 286 SkASSERT(stencil_format_supported(interface, physDev, VK_FORMAT_D32_SFLOAT_S8_UINT)); 287 fPreferedStencilFormat = gD32S8; 288 } 289} 290 291void GrVkCaps::initConfigTable(const GrVkInterface* interface, VkPhysicalDevice physDev) { 292 for (int i = 0; i < kGrPixelConfigCnt; ++i) { 293 VkFormat format; 294 if (GrPixelConfigToVkFormat(static_cast<GrPixelConfig>(i), &format)) { 295 fConfigTable[i].init(interface, physDev, format); 296 } 297 } 298 299 // We currently do not support compressed textures in Vulkan 300 const uint16_t kFlagsToRemove = ConfigInfo::kTextureable_Flag|ConfigInfo::kRenderable_Flag; 301 fConfigTable[kETC1_GrPixelConfig].fOptimalFlags &= ~kFlagsToRemove; 302 fConfigTable[kETC1_GrPixelConfig].fLinearFlags &= ~kFlagsToRemove; 303} 304 305void GrVkCaps::ConfigInfo::InitConfigFlags(VkFormatFeatureFlags vkFlags, uint16_t* flags) { 306 if (SkToBool(VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT & vkFlags) && 307 SkToBool(VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT & vkFlags)) { 308 *flags = *flags | kTextureable_Flag; 309 } 310 311 if (SkToBool(VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BLEND_BIT & vkFlags)) { 312 *flags = *flags | kRenderable_Flag; 313 } 314 315 if (SkToBool(VK_FORMAT_FEATURE_BLIT_SRC_BIT & vkFlags)) { 316 *flags = *flags | kBlitSrc_Flag; 317 } 318 319 if (SkToBool(VK_FORMAT_FEATURE_BLIT_DST_BIT & vkFlags)) { 320 *flags = *flags | kBlitDst_Flag; 321 } 322} 323 324void GrVkCaps::ConfigInfo::init(const GrVkInterface* interface, 325 VkPhysicalDevice physDev, 326 VkFormat format) { 327 VkFormatProperties props; 328 memset(&props, 0, sizeof(VkFormatProperties)); 329 GR_VK_CALL(interface, GetPhysicalDeviceFormatProperties(physDev, format, &props)); 330 InitConfigFlags(props.linearTilingFeatures, &fLinearFlags); 331 InitConfigFlags(props.optimalTilingFeatures, &fOptimalFlags); 332} 333