Home | History | Annotate | Download | only in gpu 
      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 #include "GrProgramDesc.h"
      8 #include "GrPipeline.h"
      9 #include "GrPrimitiveProcessor.h"
     10 #include "GrProcessor.h"
     11 #include "GrRenderTargetPriv.h"
     12 #include "GrShaderCaps.h"
     13 #include "GrTexturePriv.h"
     14 #include "SkChecksum.h"
     15 #include "glsl/GrGLSLFragmentProcessor.h"
     16 #include "glsl/GrGLSLFragmentShaderBuilder.h"
     17 
     18 enum {
     19     kSamplerOrImageTypeKeyBits = 4
     20 };
     21 
     22 static inline uint16_t image_storage_or_sampler_uniform_type_key(GrSLType type ) {
     23     int value = UINT16_MAX;
     24     switch (type) {
     25         case kTexture2DSampler_GrSLType:
     26             value = 0;
     27             break;
     28         case kITexture2DSampler_GrSLType:
     29             value = 1;
     30             break;
     31         case kTextureExternalSampler_GrSLType:
     32             value = 2;
     33             break;
     34         case kTexture2DRectSampler_GrSLType:
     35             value = 3;
     36             break;
     37         case kBufferSampler_GrSLType:
     38             value = 4;
     39             break;
     40         case kImageStorage2D_GrSLType:
     41             value = 5;
     42             break;
     43         case kIImageStorage2D_GrSLType:
     44             value = 6;
     45             break;
     46 
     47         default:
     48             break;
     49     }
     50     SkASSERT((value & ((1 << kSamplerOrImageTypeKeyBits) - 1)) == value);
     51     return value;
     52 }
     53 
     54 static uint16_t sampler_key(GrSLType samplerType, GrPixelConfig config, GrShaderFlags visibility,
     55                             const GrShaderCaps& caps) {
     56     int samplerTypeKey = image_storage_or_sampler_uniform_type_key(samplerType);
     57 
     58     GR_STATIC_ASSERT(1 == sizeof(caps.configTextureSwizzle(config).asKey()));
     59     return SkToU16(samplerTypeKey |
     60                    caps.configTextureSwizzle(config).asKey() << kSamplerOrImageTypeKeyBits |
     61                    (caps.samplerPrecision(config, visibility) << (8 + kSamplerOrImageTypeKeyBits)));
     62 }
     63 
     64 static uint16_t storage_image_key(const GrResourceIOProcessor::ImageStorageAccess& imageAccess) {
     65     GrSLType type = imageAccess.proxy()->imageStorageType();
     66     return image_storage_or_sampler_uniform_type_key(type) |
     67            (int)imageAccess.format() << kSamplerOrImageTypeKeyBits;
     68 }
     69 
     70 static void add_sampler_and_image_keys(GrProcessorKeyBuilder* b, const GrResourceIOProcessor& proc,
     71                                        const GrShaderCaps& caps) {
     72     int numTextureSamplers = proc.numTextureSamplers();
     73     int numBuffers = proc.numBuffers();
     74     int numImageStorages = proc.numImageStorages();
     75     int numUniforms = numTextureSamplers + numBuffers + numImageStorages;
     76     // Need two bytes per key.
     77     int word32Count = (numUniforms + 1) / 2;
     78     if (0 == word32Count) {
     79         return;
     80     }
     81     uint16_t* k16 = SkTCast<uint16_t*>(b->add32n(word32Count));
     82     int j = 0;
     83     for (int i = 0; i < numTextureSamplers; ++i, ++j) {
     84         const GrResourceIOProcessor::TextureSampler& sampler = proc.textureSampler(i);
     85         const GrTexture* tex = sampler.peekTexture();
     86 
     87         k16[j] = sampler_key(tex->texturePriv().samplerType(), tex->config(), sampler.visibility(),
     88                              caps);
     89     }
     90     for (int i = 0; i < numBuffers; ++i, ++j) {
     91         const GrResourceIOProcessor::BufferAccess& access = proc.bufferAccess(i);
     92         k16[j] = sampler_key(kBufferSampler_GrSLType, access.texelConfig(), access.visibility(),
     93                              caps);
     94     }
     95     for (int i = 0; i < numImageStorages; ++i, ++j) {
     96         k16[j] = storage_image_key(proc.imageStorageAccess(i));
     97     }
     98     // zero the last 16 bits if the number of uniforms for samplers and image storages is odd.
     99     if (numUniforms & 0x1) {
    100         k16[numUniforms] = 0;
    101     }
    102 }
    103 
    104 /**
    105  * A function which emits a meta key into the key builder.  This is required because shader code may
    106  * be dependent on properties of the effect that the effect itself doesn't use
    107  * in its key (e.g. the pixel format of textures used). So we create a meta-key for
    108  * every effect using this function. It is also responsible for inserting the effect's class ID
    109  * which must be different for every GrProcessor subclass. It can fail if an effect uses too many
    110  * transforms, etc, for the space allotted in the meta-key.  NOTE, both FPs and GPs share this
    111  * function because it is hairy, though FPs do not have attribs, and GPs do not have transforms
    112  */
    113 static bool gen_meta_key(const GrResourceIOProcessor& proc,
    114                          const GrShaderCaps& shaderCaps,
    115                          uint32_t transformKey,
    116                          GrProcessorKeyBuilder* b) {
    117     size_t processorKeySize = b->size();
    118     uint32_t classID = proc.classID();
    119 
    120     // Currently we allow 16 bits for the class id and the overall processor key size.
    121     static const uint32_t kMetaKeyInvalidMask = ~((uint32_t)SK_MaxU16);
    122     if ((processorKeySize | classID) & kMetaKeyInvalidMask) {
    123         return false;
    124     }
    125 
    126     add_sampler_and_image_keys(b, proc, shaderCaps);
    127 
    128     uint32_t* key = b->add32n(2);
    129     key[0] = (classID << 16) | SkToU32(processorKeySize);
    130     key[1] = transformKey;
    131     return true;
    132 }
    133 
    134 static bool gen_meta_key(const GrXferProcessor& xp,
    135                          const GrShaderCaps& shaderCaps,
    136                          GrProcessorKeyBuilder* b) {
    137     size_t processorKeySize = b->size();
    138     uint32_t classID = xp.classID();
    139 
    140     // Currently we allow 16 bits for the class id and the overall processor key size.
    141     static const uint32_t kMetaKeyInvalidMask = ~((uint32_t)SK_MaxU16);
    142     if ((processorKeySize | classID) & kMetaKeyInvalidMask) {
    143         return false;
    144     }
    145 
    146     b->add32((classID << 16) | SkToU32(processorKeySize));
    147     return true;
    148 }
    149 
    150 static bool gen_frag_proc_and_meta_keys(const GrPrimitiveProcessor& primProc,
    151                                         const GrFragmentProcessor& fp,
    152                                         const GrShaderCaps& shaderCaps,
    153                                         GrProcessorKeyBuilder* b) {
    154     for (int i = 0; i < fp.numChildProcessors(); ++i) {
    155         if (!gen_frag_proc_and_meta_keys(primProc, fp.childProcessor(i), shaderCaps, b)) {
    156             return false;
    157         }
    158     }
    159 
    160     fp.getGLSLProcessorKey(shaderCaps, b);
    161 
    162     return gen_meta_key(fp, shaderCaps, primProc.getTransformKey(fp.coordTransforms(),
    163                                                                  fp.numCoordTransforms()), b);
    164 }
    165 
    166 bool GrProgramDesc::Build(GrProgramDesc* desc,
    167                           const GrPrimitiveProcessor& primProc,
    168                           bool hasPointSize,
    169                           const GrPipeline& pipeline,
    170                           const GrShaderCaps& shaderCaps) {
    171     // The descriptor is used as a cache key. Thus when a field of the
    172     // descriptor will not affect program generation (because of the attribute
    173     // bindings in use or other descriptor field settings) it should be set
    174     // to a canonical value to avoid duplicate programs with different keys.
    175 
    176     GR_STATIC_ASSERT(0 == kProcessorKeysOffset % sizeof(uint32_t));
    177     // Make room for everything up to the effect keys.
    178     desc->key().reset();
    179     desc->key().push_back_n(kProcessorKeysOffset);
    180 
    181     GrProcessorKeyBuilder b(&desc->key());
    182 
    183     primProc.getGLSLProcessorKey(shaderCaps, &b);
    184     if (!gen_meta_key(primProc, shaderCaps, 0, &b)) {
    185         desc->key().reset();
    186         return false;
    187     }
    188     GrProcessor::RequiredFeatures requiredFeatures = primProc.requiredFeatures();
    189 
    190     for (int i = 0; i < pipeline.numFragmentProcessors(); ++i) {
    191         const GrFragmentProcessor& fp = pipeline.getFragmentProcessor(i);
    192         if (!gen_frag_proc_and_meta_keys(primProc, fp, shaderCaps, &b)) {
    193             desc->key().reset();
    194             return false;
    195         }
    196         requiredFeatures |= fp.requiredFeatures();
    197     }
    198 
    199     const GrXferProcessor& xp = pipeline.getXferProcessor();
    200     const GrSurfaceOrigin* originIfDstTexture = nullptr;
    201     GrSurfaceOrigin origin;
    202     if (pipeline.dstTextureProxy()) {
    203         origin = pipeline.dstTextureProxy()->origin();
    204         originIfDstTexture = &origin;
    205     }
    206     xp.getGLSLProcessorKey(shaderCaps, &b, originIfDstTexture);
    207     if (!gen_meta_key(xp, shaderCaps, &b)) {
    208         desc->key().reset();
    209         return false;
    210     }
    211     requiredFeatures |= xp.requiredFeatures();
    212 
    213     // --------DO NOT MOVE HEADER ABOVE THIS LINE--------------------------------------------------
    214     // Because header is a pointer into the dynamic array, we can't push any new data into the key
    215     // below here.
    216     KeyHeader* header = desc->atOffset<KeyHeader, kHeaderOffset>();
    217 
    218     // make sure any padding in the header is zeroed.
    219     memset(header, 0, kHeaderSize);
    220 
    221     GrRenderTarget* rt = pipeline.getRenderTarget();
    222 
    223     if (requiredFeatures & GrProcessor::kSampleLocations_RequiredFeature) {
    224         SkASSERT(pipeline.isHWAntialiasState());
    225         header->fSamplePatternKey =
    226             rt->renderTargetPriv().getMultisampleSpecs(pipeline).fUniqueID;
    227     } else {
    228         header->fSamplePatternKey = 0;
    229     }
    230 
    231     header->fOutputSwizzle = shaderCaps.configOutputSwizzle(rt->config()).asKey();
    232 
    233     header->fSnapVerticesToPixelCenters = pipeline.snapVerticesToPixelCenters();
    234     header->fColorFragmentProcessorCnt = pipeline.numColorFragmentProcessors();
    235     header->fCoverageFragmentProcessorCnt = pipeline.numCoverageFragmentProcessors();
    236     // Fail if the client requested more processors than the key can fit.
    237     if (header->fColorFragmentProcessorCnt != pipeline.numColorFragmentProcessors() ||
    238         header->fCoverageFragmentProcessorCnt != pipeline.numCoverageFragmentProcessors()) {
    239         return false;
    240     }
    241     header->fHasPointSize = hasPointSize ? 1 : 0;
    242     return true;
    243 }
    244