AuroraMiddleware/skia2
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Revert "Fix precision caps and rrect/ellipse effect precisions"

Browse Source 
This reverts commit e421800227.

Reason for revert: Also may be responsible for layout test failures? Playing it safe.

Original change's description:
> Fix precision caps and rrect/ellipse effect precisions
> 
> Replaces all the complex precision caps with a single flag that says
> whether "float" == fp32. Updates the ellipse and rrect effects to
> use float coords, and use the scale workaround when float != fp32.
> 
> Bug: skia:7190
> Change-Id: Ieccff9f38acd05e5cec78fe90d01a5da901a9307
> Reviewed-on: https://skia-review.googlesource.com/70961
> Commit-Queue: Chris Dalton <csmartdalton@google.com>
> Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
> Reviewed-by: Brian Salomon <bsalomon@google.com>

TBR=egdaniel@google.com,bsalomon@google.com,csmartdalton@google.com,ethannicholas@google.com

Change-Id: Idca2f0390e7a0eb85010255183f2f27332b8d26d
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: skia:7190
Reviewed-on: https://skia-review.googlesource.com/72540
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
This commit is contained in:
Brian Osman 2017-11-16 14:02:11 +00:00 committed by Skia Commit-Bot
parent 554c1f0508
commit 24f9c19172
17 changed files with 359 additions and 141 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

71
include/gpu/GrShaderCaps.h
View File

@ -20,6 +20,40 @@ class SkJSONWriter;
class GrShaderCaps : public SkRefCnt {
public:
/** Info about shader variable precision within a given shader stage. That is, this info
is relevant to a float (or vecNf) variable declared with a GrSLPrecision
in a given GrShaderType. The info here is hoisted from the OpenGL spec. */
struct PrecisionInfo {
PrecisionInfo() {
fLogRangeLow = 0;
fLogRangeHigh = 0;
fBits = 0;
}
/** Is this precision level allowed in the shader stage? */
bool supported() const { return 0 != fBits; }
bool operator==(const PrecisionInfo& that) const {
return fLogRangeLow == that.fLogRangeLow && fLogRangeHigh == that.fLogRangeHigh &&
fBits == that.fBits;
}
bool operator!=(const PrecisionInfo& that) const { return !(*this == that); }
/** floor(log2(|min_value|)) */
int fLogRangeLow;
/** floor(log2(|max_value|)) */
int fLogRangeHigh;
/** Number of bits of precision. As defined in OpenGL (with names modified to reflect this
struct) :
"""
If the smallest representable value greater than 1 is 1 + e, then fBits will
contain floor(log2(e)), and every value in the range [2^fLogRangeLow,
2^fLogRangeHigh] can be represented to at least one part in 2^fBits.
"""
*/
int fBits;
};
/**
* Indicates how GLSL must interact with advanced blend equations. The KHR extension requires
* special layout qualifiers in the fragment shader.
@ -47,6 +81,24 @@ public:
bool texelBufferSupport() const { return fTexelBufferSupport; }
int imageLoadStoreSupport() const { return fImageLoadStoreSupport; }
/**
* Get the precision info for a variable of type kFloat_GrSLType, kFloat2_GrSLType, etc in a
* given shader type. If the shader type is not supported or the precision level is not
* supported in that shader type then the returned struct will report false when supported() is
* called.
*/
const PrecisionInfo& getFloatShaderPrecisionInfo(GrShaderType shaderType,
GrSLPrecision precision) const {
return fFloatPrecisions[shaderType][precision];
}
/**
* Is there any difference between the float shader variable precision types? If this is true
* then unless the shader type is not supported, any call to getFloatShaderPrecisionInfo() would
* report the same info for all precisions in all shader types.
*/
bool floatPrecisionVaries() const { return fShaderPrecisionVaries; }
/**
* Some helper functions for encapsulating various extensions to read FB Buffer on openglES
*
@ -84,10 +136,6 @@ public:
bool vertexIDSupport() const { return fVertexIDSupport; }
bool floatIs32Bits() const { return fFloatIs32Bits; }
bool halfIs32Bits() const { return fHalfIs32Bits; }
AdvBlendEqInteraction advBlendEqInteraction() const { return fAdvBlendEqInteraction; }
bool mustEnableAdvBlendEqs() const {
@ -212,9 +260,17 @@ public:
return fConfigOutputSwizzle[config];
}
/** Precision qualifier that should be used with a sampler, given its config and visibility. */
GrSLPrecision samplerPrecision(GrPixelConfig config, GrShaderFlags visibility) const {
return static_cast<GrSLPrecision>(fSamplerPrecisions[visibility][config]);
}
GrGLSLGeneration generation() const { return fGLSLGeneration; }
private:
/** GrCaps subclasses must call this after filling in the shader precision table. */
void initSamplerPrecisionTable();
void applyOptionsOverrides(const GrContextOptions& options);
GrGLSLGeneration fGLSLGeneration;
@ -228,6 +284,7 @@ private:
bool fIntegerSupport : 1;
bool fTexelBufferSupport : 1;
bool fImageLoadStoreSupport : 1;
bool fShaderPrecisionVaries : 1;
bool fDropsTileOnZeroDivide : 1;
bool fFBFetchSupport : 1;
bool fFBFetchNeedsCustomOutput : 1;
@ -243,8 +300,6 @@ private:
bool fExternalTextureSupport : 1;
bool fTexelFetchSupport : 1;
bool fVertexIDSupport : 1;
bool fFloatIs32Bits : 1;
bool fHalfIs32Bits : 1;
bool fDisableImageMultitexturing : 1;
// Used for specific driver bug work arounds
@ -256,6 +311,8 @@ private:
bool fMustObfuscateUniformColor : 1;
bool fMustGuardDivisionEvenAfterExplicitZeroCheck : 1;
PrecisionInfo fFloatPrecisions[kGrShaderTypeCount][kGrSLPrecisionCount];
const char* fVersionDeclString;
const char* fShaderDerivativeExtensionString;
@ -282,6 +339,8 @@ private:
GrSwizzle fConfigTextureSwizzle[kGrPixelConfigCnt];
GrSwizzle fConfigOutputSwizzle[kGrPixelConfigCnt];
uint8_t fSamplerPrecisions[(1 << kGrShaderTypeCount)][kGrPixelConfigCnt];
friend class GrGLCaps; // For initialization.
friend class GrMockCaps;
friend class GrMtlCaps;

27
include/private/GrTypesPriv.h
View File

@ -909,33 +909,6 @@ static inline bool GrPixelConfigIsUnorm(GrPixelConfig config) {
return false;
}
/**
* Precision qualifier that should be used with a sampler.
*/
static inline GrSLPrecision GrSLSamplerPrecision(GrPixelConfig config) {
switch (config) {
case kUnknown_GrPixelConfig:
case kAlpha_8_GrPixelConfig:
case kGray_8_GrPixelConfig:
case kRGB_565_GrPixelConfig:
case kRGBA_4444_GrPixelConfig:
case kRGBA_8888_GrPixelConfig:
case kBGRA_8888_GrPixelConfig:
case kSRGBA_8888_GrPixelConfig:
case kSBGRA_8888_GrPixelConfig:
case kRGBA_8888_sint_GrPixelConfig:
return kLow_GrSLPrecision;
case kRGBA_float_GrPixelConfig:
case kRG_float_GrPixelConfig:
return kHigh_GrSLPrecision;
case kAlpha_half_GrPixelConfig:
case kRGBA_half_GrPixelConfig:
return kMedium_GrSLPrecision;
}
SK_ABORT("Unexpected type");
return kHigh_GrSLPrecision;
}
static inline GrPixelConfigIsClamped GrGetPixelConfigIsClamped(GrPixelConfig config) {
return GrPixelConfigIsFloatingPoint(config) ? GrPixelConfigIsClamped::kNo
: GrPixelConfigIsClamped::kYes;

2
src/gpu/GrProgramDesc.cpp
View File

@ -52,7 +52,7 @@ static uint16_t sampler_key(GrSLType samplerType, GrPixelConfig config, GrShader
GR_STATIC_ASSERT(1 == sizeof(caps.configTextureSwizzle(config).asKey()));
return SkToU16(samplerTypeKey |
caps.configTextureSwizzle(config).asKey() << kSamplerOrImageTypeKeyBits |
(GrSLSamplerPrecision(config) << (8 + kSamplerOrImageTypeKeyBits)));
(caps.samplerPrecision(config, visibility) << (8 + kSamplerOrImageTypeKeyBits)));
}
static void add_sampler_and_image_keys(GrProcessorKeyBuilder* b, const GrResourceIOProcessor& proc,

104
src/gpu/GrShaderCaps.cpp
View File

@ -13,6 +13,32 @@
////////////////////////////////////////////////////////////////////////////////////////////
static const char* shader_type_to_string(GrShaderType type) {
switch (type) {
case kVertex_GrShaderType:
return "vertex";
case kGeometry_GrShaderType:
return "geometry";
case kFragment_GrShaderType:
return "fragment";
}
return "";
}
static const char* precision_to_string(GrSLPrecision p) {
switch (p) {
case kLow_GrSLPrecision:
return "low";
case kMedium_GrSLPrecision:
return "medium";
case kHigh_GrSLPrecision:
return "high";
default:
SK_ABORT("Unexpected precision type.");
return "";
}
}
GrShaderCaps::GrShaderCaps(const GrContextOptions& options) {
fGLSLGeneration = k330_GrGLSLGeneration;
fShaderDerivativeSupport = false;
@ -24,6 +50,7 @@ GrShaderCaps::GrShaderCaps(const GrContextOptions& options) {
fIntegerSupport = false;
fTexelBufferSupport = false;
fImageLoadStoreSupport = false;
fShaderPrecisionVaries = false;
fDropsTileOnZeroDivide = false;
fFBFetchSupport = false;
fFBFetchNeedsCustomOutput = false;
@ -46,8 +73,6 @@ GrShaderCaps::GrShaderCaps(const GrContextOptions& options) {
fExternalTextureSupport = false;
fTexelFetchSupport = false;
fVertexIDSupport = false;
fFloatIs32Bits = true;
fHalfIs32Bits = false;
fVersionDeclString = nullptr;
fShaderDerivativeExtensionString = nullptr;
@ -88,6 +113,26 @@ void GrShaderCaps::dumpJSON(SkJSONWriter* writer) const {
writer->appendBool("Texel Buffer Support", fTexelBufferSupport);
writer->appendBool("Image Load Store Support", fImageLoadStoreSupport);
writer->appendBool("Variable Precision", fShaderPrecisionVaries);
for (int s = 0; s < kGrShaderTypeCount; ++s) {
GrShaderType shaderType = static_cast<GrShaderType>(s);
writer->beginArray(SkStringPrintf("%s precisions",
shader_type_to_string(shaderType)).c_str());
for (int p = 0; p < kGrSLPrecisionCount; ++p) {
if (fFloatPrecisions[s][p].supported()) {
GrSLPrecision precision = static_cast<GrSLPrecision>(p);
writer->beginObject(nullptr, false);
writer->appendString("precision", precision_to_string(precision));
writer->appendS32("log_low", fFloatPrecisions[s][p].fLogRangeLow);
writer->appendS32("log_high", fFloatPrecisions[s][p].fLogRangeHigh);
writer->appendS32("bits", fFloatPrecisions[s][p].fBits);
writer->endObject();
}
}
writer->endArray();
}
static const char* kAdvBlendEqInteractionStr[] = {
"Not Supported",
"Automatic",
@ -121,8 +166,6 @@ void GrShaderCaps::dumpJSON(SkJSONWriter* writer) const {
writer->appendBool("External texture support", fExternalTextureSupport);
writer->appendBool("texelFetch support", fTexelFetchSupport);
writer->appendBool("sk_VertexID support", fVertexIDSupport);
writer->appendBool("float == fp32", fFloatIs32Bits);
writer->appendBool("half == fp32", fHalfIs32Bits);
writer->appendS32("Max VS Samplers", fMaxVertexSamplers);
writer->appendS32("Max GS Samplers", fMaxGeometrySamplers);
@ -135,6 +178,59 @@ void GrShaderCaps::dumpJSON(SkJSONWriter* writer) const {
writer->endObject();
}
void GrShaderCaps::initSamplerPrecisionTable() {
// Determine the largest precision qualifiers that are effectively the same as lowp/mediump.
// e.g. if lowp == mediump, then use mediump instead of lowp.
GrSLPrecision effectiveMediumP[kGrShaderTypeCount];
GrSLPrecision effectiveLowP[kGrShaderTypeCount];
for (int s = 0; s < kGrShaderTypeCount; ++s) {
const PrecisionInfo* info = fFloatPrecisions[s];
effectiveMediumP[s] = info[kHigh_GrSLPrecision] == info[kMedium_GrSLPrecision] ?
kHigh_GrSLPrecision : kMedium_GrSLPrecision;
effectiveLowP[s] = info[kMedium_GrSLPrecision] == info[kLow_GrSLPrecision] ?
effectiveMediumP[s] : kLow_GrSLPrecision;
}
// Determine which precision qualifiers should be used with samplers.
for (int visibility = 0; visibility < (1 << kGrShaderTypeCount); ++visibility) {
GrSLPrecision mediump = kHigh_GrSLPrecision;
GrSLPrecision lowp = kHigh_GrSLPrecision;
for (int s = 0; s < kGrShaderTypeCount; ++s) {
if (visibility & (1 << s)) {
mediump = SkTMin(mediump, effectiveMediumP[s]);
lowp = SkTMin(lowp, effectiveLowP[s]);
}
GR_STATIC_ASSERT(0 == kLow_GrSLPrecision);
GR_STATIC_ASSERT(1 == kMedium_GrSLPrecision);
GR_STATIC_ASSERT(2 == kHigh_GrSLPrecision);
GR_STATIC_ASSERT((1 << kVertex_GrShaderType) == kVertex_GrShaderFlag);
GR_STATIC_ASSERT((1 << kGeometry_GrShaderType) == kGeometry_GrShaderFlag);
GR_STATIC_ASSERT((1 << kFragment_GrShaderType) == kFragment_GrShaderFlag);
GR_STATIC_ASSERT(3 == kGrShaderTypeCount);
}
uint8_t* table = fSamplerPrecisions[visibility];
table[kUnknown_GrPixelConfig] = lowp;
table[kAlpha_8_GrPixelConfig] = lowp;
table[kGray_8_GrPixelConfig] = lowp;
table[kRGB_565_GrPixelConfig] = lowp;
table[kRGBA_4444_GrPixelConfig] = lowp;
table[kRGBA_8888_GrPixelConfig] = lowp;
table[kBGRA_8888_GrPixelConfig] = lowp;
table[kSRGBA_8888_GrPixelConfig] = lowp;
table[kSBGRA_8888_GrPixelConfig] = lowp;
table[kRGBA_8888_sint_GrPixelConfig] = lowp;
table[kRGBA_float_GrPixelConfig] = kHigh_GrSLPrecision;
table[kRG_float_GrPixelConfig] = kHigh_GrSLPrecision;
table[kAlpha_half_GrPixelConfig] = mediump;
table[kRGBA_half_GrPixelConfig] = mediump;
GR_STATIC_ASSERT(14 == kGrPixelConfigCnt);
}
}
void GrShaderCaps::applyOptionsOverrides(const GrContextOptions& options) {
#if GR_TEST_UTILS
fDualSourceBlendingSupport = fDualSourceBlendingSupport && !options.fSuppressDualSourceBlending;

33
src/gpu/effects/GrEllipseEffect.cpp
View File

@ -28,33 +28,34 @@ public:
(void)center;
auto radii = _outer.radii();
(void)radii;
prevRadii = float2(-1.0);
useScale = !sk_Caps.floatIs32Bits;
prevRadii = half2(-1.0);
useScale = sk_Caps.floatPrecisionVaries;
fEllipseVar = args.fUniformHandler->addUniform(kFragment_GrShaderFlag, kFloat4_GrSLType,
kDefault_GrSLPrecision, "ellipse");
if (useScale) {
fScaleVar = args.fUniformHandler->addUniform(kFragment_GrShaderFlag, kFloat2_GrSLType,
fScaleVar = args.fUniformHandler->addUniform(kFragment_GrShaderFlag, kHalf2_GrSLType,
kDefault_GrSLPrecision, "scale");
}
fragBuilder->codeAppendf(
"float2 prevCenter;\nfloat2 prevRadii = float2(%f, %f);\nbool useScale = "
"%s;\nfloat2 d = sk_FragCoord.xy - %s.xy;\n@if (useScale) {\n d *= "
"%s.y;\n}\nfloat2 Z = d * %s.zw;\nfloat implicit = dot(Z, d) - 1.0;\nfloat "
"grad_dot = 4.0 * dot(Z, Z);\ngrad_dot = max(grad_dot, 0.0001);\nfloat approx_dist "
"= implicit * inversesqrt(grad_dot);\n@if (useScale) {\n approx_dist *= "
"%s.x;\n}\nhalf alpha;\n@switch (%d) {\n case 0:\n alpha = "
"half(approx_dist > 0.0 ? 0.0 : 1.0);\n break;\n case 1:\n ",
"half2 prevCenter;\nhalf2 prevRadii = half2(%f, %f);\nbool useScale = %s;\nhalf2 d "
"= half2(sk_FragCoord.xy - %s.xy);\n@if (useScale) {\n d *= %s.y;\n}\nhalf2 Z = "
"d * half2(%s.zw);\nhalf implicit = dot(Z, d) - 1.0;\nhalf grad_dot = 4.0 * dot(Z, "
"Z);\ngrad_dot = half(max(float(grad_dot), 0.0001));\nhalf approx_dist = "
"float(implicit) * inversesqrt(float(grad_dot));\n@if (useScale) {\n "
"approx_dist *= %s.x;\n}\nhalf alpha;\n@switch (%d) {\n case 0:\n alpha "
"= half(float(approx_dist) > 0.0 ? 0.0 : 1.",
prevRadii.fX, prevRadii.fY, (useScale ? "true" : "false"),
args.fUniformHandler->getUniformCStr(fEllipseVar),
fScaleVar.isValid() ? args.fUniformHandler->getUniformCStr(fScaleVar) : "float2(0)",
fScaleVar.isValid() ? args.fUniformHandler->getUniformCStr(fScaleVar) : "half2(0)",
args.fUniformHandler->getUniformCStr(fEllipseVar),
fScaleVar.isValid() ? args.fUniformHandler->getUniformCStr(fScaleVar) : "float2(0)",
fScaleVar.isValid() ? args.fUniformHandler->getUniformCStr(fScaleVar) : "half2(0)",
(int)_outer.edgeType());
fragBuilder->codeAppendf(
" alpha = half(clamp(0.5 - approx_dist, 0.0, 1.0));\n break;\n case "
"2:\n alpha = half(approx_dist > 0.0 ? 1.0 : 0.0);\n break;\n "
"case 3:\n alpha = half(clamp(0.5 + approx_dist, 0.0, 1.0));\n "
"break;\n default:\n discard;\n}\n%s = %s * alpha;\n",
"0);\n break;\n case 1:\n alpha = half(clamp(0.5 - "
"float(approx_dist), 0.0, 1.0));\n break;\n case 2:\n alpha = "
"half(float(approx_dist) > 0.0 ? 1.0 : 0.0);\n break;\n case 3:\n "
"alpha = half(clamp(0.5 + float(approx_dist), 0.0, 1.0));\n break;\n "
"default:\n discard;\n}\n%s = %s * alpha;\n",
args.fOutputColor, args.fInputColor ? args.fInputColor : "half4(1)");
}

26
src/gpu/effects/GrEllipseEffect.fp
View File

@ -6,17 +6,17 @@
*/
layout(key) in GrClipEdgeType edgeType;
in float2 center;
in float2 radii;
in half2 center;
in half2 radii;
float2 prevCenter;
float2 prevRadii = float2(-1);
half2 prevCenter;
half2 prevRadii = half2(-1);
// The ellipse uniform is (center.x, center.y, 1 / rx^2, 1 / ry^2)
// The last two terms can underflow when float != fp32, so we also provide a workaround.
// The last two terms can underflow with halfs, so we use floats.
uniform float4 ellipse;
bool useScale = !sk_Caps.floatIs32Bits;
layout(when=useScale) uniform float2 scale;
bool useScale = sk_Caps.floatPrecisionVaries;
layout(when=useScale) uniform half2 scale;
@optimizationFlags { kCompatibleWithCoverageAsAlpha_OptimizationFlag }
@ -50,7 +50,7 @@ layout(when=useScale) uniform float2 scale;
void main() {
// d is the offset to the ellipse center
float2 d = sk_FragCoord.xy - ellipse.xy;
half2 d = sk_FragCoord.xy - ellipse.xy;
// If we're on a device with a "real" mediump then we'll do the distance computation in a space
// that is normalized by the larger radius. The scale uniform will be scale, 1/scale. The
// inverse squared radii uniform values are already in this normalized space. The center is
@ -58,14 +58,14 @@ void main() {
@if (useScale) {
d *= scale.y;
}
float2 Z = d * ellipse.zw;
half2 Z = d * ellipse.zw;
// implicit is the evaluation of (x/rx)^2 + (y/ry)^2 - 1.
float implicit = dot(Z, d) - 1;
half implicit = dot(Z, d) - 1;
// grad_dot is the squared length of the gradient of the implicit.
float grad_dot = 4 * dot(Z, Z);
half grad_dot = 4 * dot(Z, Z);
// Avoid calling inversesqrt on zero.
grad_dot = max(grad_dot, 1e-4);
float approx_dist = implicit * inversesqrt(grad_dot);
half approx_dist = implicit * inversesqrt(grad_dot);
@if (useScale) {
approx_dist *= scale.x;
}
@ -102,4 +102,4 @@ void main() {
et = (GrClipEdgeType) testData->fRandom->nextULessThan(kGrClipEdgeTypeCnt);
} while (GrClipEdgeType::kHairlineAA == et);
return GrEllipseEffect::Make(et, center, SkPoint::Make(rx, ry));
}
}

68
src/gpu/effects/GrRRectEffect.cpp
View File

@ -165,9 +165,9 @@ void GLCircularRRectEffect::emitCode(EmitArgs& args) {
fRadiusPlusHalfUniform = uniformHandler->addUniform(kFragment_GrShaderFlag, kHalf2_GrSLType,
"radiusPlusHalf", &radiusPlusHalfName);
// If we're on a device where float != fp32 then the length calculation could overflow.
// If we're on a device with a "real" mediump then the length calculation could overflow.
SkString clampedCircleDistance;
if (!args.fShaderCaps->floatIs32Bits()) {
if (args.fShaderCaps->floatPrecisionVaries()) {
clampedCircleDistance.printf("clamp(%s.x * (1.0 - length(dxy * %s.y)), 0.0, 1.0);",
radiusPlusHalfName, radiusPlusHalfName);
} else {
@ -192,13 +192,13 @@ void GLCircularRRectEffect::emitCode(EmitArgs& args) {
// alphas together.
switch (crre.getCircularCornerFlags()) {
case CircularRRectEffect::kAll_CornerFlags:
fragBuilder->codeAppendf("float2 dxy0 = %s.xy - sk_FragCoord.xy;", rectName);
fragBuilder->codeAppendf("float2 dxy1 = sk_FragCoord.xy - %s.zw;", rectName);
fragBuilder->codeAppend("float2 dxy = max(max(dxy0, dxy1), 0.0);");
fragBuilder->codeAppendf("half2 dxy0 = %s.xy - sk_FragCoord.xy;", rectName);
fragBuilder->codeAppendf("half2 dxy1 = sk_FragCoord.xy - %s.zw;", rectName);
fragBuilder->codeAppend("half2 dxy = max(max(dxy0, dxy1), 0.0);");
fragBuilder->codeAppendf("half alpha = %s;", clampedCircleDistance.c_str());
break;
case CircularRRectEffect::kTopLeft_CornerFlag:
fragBuilder->codeAppendf("float2 dxy = max(%s.xy - sk_FragCoord.xy, 0.0);",
fragBuilder->codeAppendf("half2 dxy = max(%s.xy - sk_FragCoord.xy, 0.0);",
rectName);
fragBuilder->codeAppendf("half rightAlpha = clamp(%s.z - sk_FragCoord.x, 0.0, 1.0);",
rectName);
@ -208,8 +208,8 @@ void GLCircularRRectEffect::emitCode(EmitArgs& args) {
clampedCircleDistance.c_str());
break;
case CircularRRectEffect::kTopRight_CornerFlag:
fragBuilder->codeAppendf("float2 dxy = max(float2(sk_FragCoord.x - %s.z, "
"%s.y - sk_FragCoord.y), 0.0);",
fragBuilder->codeAppendf("half2 dxy = max(half2(sk_FragCoord.x - %s.z, "
"%s.y - sk_FragCoord.y), 0.0);",
rectName, rectName);
fragBuilder->codeAppendf("half leftAlpha = clamp(sk_FragCoord.x - %s.x, 0.0, 1.0);",
rectName);
@ -219,7 +219,7 @@ void GLCircularRRectEffect::emitCode(EmitArgs& args) {
clampedCircleDistance.c_str());
break;
case CircularRRectEffect::kBottomRight_CornerFlag:
fragBuilder->codeAppendf("float2 dxy = max(sk_FragCoord.xy - %s.zw, 0.0);",
fragBuilder->codeAppendf("half2 dxy = max(sk_FragCoord.xy - %s.zw, 0.0);",
rectName);
fragBuilder->codeAppendf("half leftAlpha = clamp(sk_FragCoord.x - %s.x, 0.0, 1.0);",
rectName);
@ -229,8 +229,8 @@ void GLCircularRRectEffect::emitCode(EmitArgs& args) {
clampedCircleDistance.c_str());
break;
case CircularRRectEffect::kBottomLeft_CornerFlag:
fragBuilder->codeAppendf("float2 dxy = max(float2(%s.x - sk_FragCoord.x, "
"sk_FragCoord.y - %s.w), 0.0);",
fragBuilder->codeAppendf("half2 dxy = max(half2(%s.x - sk_FragCoord.x, sk_FragCoord.y - "
"%s.w), 0.0);",
rectName, rectName);
fragBuilder->codeAppendf("half rightAlpha = clamp(%s.z - sk_FragCoord.x, 0.0, 1.0);",
rectName);
@ -240,36 +240,36 @@ void GLCircularRRectEffect::emitCode(EmitArgs& args) {
clampedCircleDistance.c_str());
break;
case CircularRRectEffect::kLeft_CornerFlags:
fragBuilder->codeAppendf("float2 dxy0 = %s.xy - sk_FragCoord.xy;", rectName);
fragBuilder->codeAppendf("float dy1 = sk_FragCoord.y - %s.w;", rectName);
fragBuilder->codeAppend("float2 dxy = max(float2(dxy0.x, max(dxy0.y, dy1)), 0.0);");
fragBuilder->codeAppendf("half2 dxy0 = %s.xy - sk_FragCoord.xy;", rectName);
fragBuilder->codeAppendf("half dy1 = sk_FragCoord.y - %s.w;", rectName);
fragBuilder->codeAppend("half2 dxy = max(half2(dxy0.x, max(dxy0.y, dy1)), 0.0);");
fragBuilder->codeAppendf("half rightAlpha = clamp(%s.z - sk_FragCoord.x, 0.0, 1.0);",
rectName);
fragBuilder->codeAppendf("half alpha = rightAlpha * %s;",
clampedCircleDistance.c_str());
break;
case CircularRRectEffect::kTop_CornerFlags:
fragBuilder->codeAppendf("float2 dxy0 = %s.xy - sk_FragCoord.xy;", rectName);
fragBuilder->codeAppendf("float dx1 = sk_FragCoord.x - %s.z;", rectName);
fragBuilder->codeAppend("float2 dxy = max(float2(max(dxy0.x, dx1), dxy0.y), 0.0);");
fragBuilder->codeAppendf("half2 dxy0 = %s.xy - sk_FragCoord.xy;", rectName);
fragBuilder->codeAppendf("half dx1 = sk_FragCoord.x - %s.z;", rectName);
fragBuilder->codeAppend("half2 dxy = max(half2(max(dxy0.x, dx1), dxy0.y), 0.0);");
fragBuilder->codeAppendf("half bottomAlpha = clamp(%s.w - sk_FragCoord.y, 0.0, 1.0);",
rectName);
fragBuilder->codeAppendf("half alpha = bottomAlpha * %s;",
clampedCircleDistance.c_str());
break;
case CircularRRectEffect::kRight_CornerFlags:
fragBuilder->codeAppendf("float dy0 = %s.y - sk_FragCoord.y;", rectName);
fragBuilder->codeAppendf("float2 dxy1 = sk_FragCoord.xy - %s.zw;", rectName);
fragBuilder->codeAppend("float2 dxy = max(float2(dxy1.x, max(dy0, dxy1.y)), 0.0);");
fragBuilder->codeAppendf("half dy0 = %s.y - sk_FragCoord.y;", rectName);
fragBuilder->codeAppendf("half2 dxy1 = sk_FragCoord.xy - %s.zw;", rectName);
fragBuilder->codeAppend("half2 dxy = max(half2(dxy1.x, max(dy0, dxy1.y)), 0.0);");
fragBuilder->codeAppendf("half leftAlpha = clamp(sk_FragCoord.x - %s.x, 0.0, 1.0);",
rectName);
fragBuilder->codeAppendf("half alpha = leftAlpha * %s;",
clampedCircleDistance.c_str());
break;
case CircularRRectEffect::kBottom_CornerFlags:
fragBuilder->codeAppendf("float dx0 = %s.x - sk_FragCoord.x;", rectName);
fragBuilder->codeAppendf("float2 dxy1 = sk_FragCoord.xy - %s.zw;", rectName);
fragBuilder->codeAppend("float2 dxy = max(float2(max(dx0, dxy1.x), dxy1.y), 0.0);");
fragBuilder->codeAppendf("half dx0 = %s.x - sk_FragCoord.x;", rectName);
fragBuilder->codeAppendf("half2 dxy1 = sk_FragCoord.xy - %s.zw;", rectName);
fragBuilder->codeAppend("half2 dxy = max(half2(max(dx0, dxy1.x), dxy1.y), 0.0);");
fragBuilder->codeAppendf("half topAlpha = clamp(sk_FragCoord.y - %s.y, 0.0, 1.0);",
rectName);
fragBuilder->codeAppendf("half alpha = topAlpha * %s;",
@ -523,14 +523,14 @@ void GLEllipticalRRectEffect::emitCode(EmitArgs& args) {
// The code below is a simplified version of the above that performs maxs on the vector
// components before computing distances and alpha values so that only one distance computation
// need be computed to determine the min alpha.
fragBuilder->codeAppendf("float2 dxy0 = %s.xy - sk_FragCoord.xy;", rectName);
fragBuilder->codeAppendf("float2 dxy1 = sk_FragCoord.xy - %s.zw;", rectName);
fragBuilder->codeAppendf("half2 dxy0 = %s.xy - sk_FragCoord.xy;", rectName);
fragBuilder->codeAppendf("half2 dxy1 = sk_FragCoord.xy - %s.zw;", rectName);
// If we're on a device where float != fp32 then we'll do the distance computation in a space
// If we're on a device with a "real" mediump then we'll do the distance computation in a space
// that is normalized by the largest radius. The scale uniform will be scale, 1/scale. The
// radii uniform values are already in this normalized space.
const char* scaleName = nullptr;
if (!args.fShaderCaps->floatIs32Bits()) {
if (args.fShaderCaps->floatPrecisionVaries()) {
fScaleUniform = uniformHandler->addUniform(kFragment_GrShaderFlag, kHalf2_GrSLType, "scale",
&scaleName);
}
@ -543,12 +543,12 @@ void GLEllipticalRRectEffect::emitCode(EmitArgs& args) {
kHalf2_GrSLType,
"invRadiiXY",
&invRadiiXYSqdName);
fragBuilder->codeAppend("float2 dxy = max(max(dxy0, dxy1), 0.0);");
fragBuilder->codeAppend("half2 dxy = max(max(dxy0, dxy1), 0.0);");
if (scaleName) {
fragBuilder->codeAppendf("dxy *= %s.y;", scaleName);
}
// Z is the x/y offsets divided by squared radii.
fragBuilder->codeAppendf("float2 Z = dxy * %s.xy;", invRadiiXYSqdName);
fragBuilder->codeAppendf("half2 Z = dxy * %s.xy;", invRadiiXYSqdName);
break;
}
case SkRRect::kNinePatch_Type: {
@ -561,11 +561,11 @@ void GLEllipticalRRectEffect::emitCode(EmitArgs& args) {
fragBuilder->codeAppendf("dxy0 *= %s.y;", scaleName);
fragBuilder->codeAppendf("dxy1 *= %s.y;", scaleName);
}
fragBuilder->codeAppend("float2 dxy = max(max(dxy0, dxy1), 0.0);");
fragBuilder->codeAppend("half2 dxy = max(max(dxy0, dxy1), 0.0);");
// Z is the x/y offsets divided by squared radii. We only care about the (at most) one
// corner where both the x and y offsets are positive, hence the maxes. (The inverse
// squared radii will always be positive.)
fragBuilder->codeAppendf("float2 Z = max(max(dxy0 * %s.xy, dxy1 * %s.zw), 0.0);",
fragBuilder->codeAppendf("half2 Z = max(max(dxy0 * %s.xy, dxy1 * %s.zw), 0.0);",
invRadiiLTRBSqdName, invRadiiLTRBSqdName);
break;
@ -574,12 +574,12 @@ void GLEllipticalRRectEffect::emitCode(EmitArgs& args) {
SK_ABORT("RRect should always be simple or nine-patch.");
}
// implicit is the evaluation of (x/a)^2 + (y/b)^2 - 1.
fragBuilder->codeAppend("float implicit = dot(Z, dxy) - 1.0;");
fragBuilder->codeAppend("half implicit = dot(Z, dxy) - 1.0;");
// grad_dot is the squared length of the gradient of the implicit.
fragBuilder->codeAppend("float grad_dot = 4.0 * dot(Z, Z);");
fragBuilder->codeAppend("half grad_dot = 4.0 * dot(Z, Z);");
// avoid calling inversesqrt on zero.
fragBuilder->codeAppend("grad_dot = max(grad_dot, 1.0e-4);");
fragBuilder->codeAppend("float approx_dist = implicit * inversesqrt(grad_dot);");
fragBuilder->codeAppend("half approx_dist = implicit * inversesqrt(grad_dot);");
if (scaleName) {
fragBuilder->codeAppendf("approx_dist *= %s.x;", scaleName);
}

115
src/gpu/gl/GrGLCaps.cpp
View File

@ -291,7 +291,7 @@ void GrGLCaps::init(const GrContextOptions& contextOptions,
**************************************************************************/
// This must be called after fCoreProfile is set on the GrGLCaps
this->initGLSL(ctxInfo, gli);
this->initGLSL(ctxInfo);
GrShaderCaps* shaderCaps = fShaderCaps.get();
shaderCaps->fPathRenderingSupport = this->hasPathRenderingSupport(ctxInfo, gli);
@ -634,6 +634,8 @@ void GrGLCaps::init(const GrContextOptions& contextOptions,
fDrawRangeElementsSupport = version >= GR_GL_VER(3,0);
}
this->initShaderPrecisionTable(ctxInfo, gli, shaderCaps);
if (kGL_GrGLStandard == standard) {
if ((version >= GR_GL_VER(4, 0) || ctxInfo.hasExtension("GL_ARB_sample_shading")) &&
ctxInfo.vendor() != kIntel_GrGLVendor) {
@ -758,27 +760,7 @@ const char* get_glsl_version_decl_string(GrGLStandard standard, GrGLSLGeneration
return "<no version>";
}
bool is_float_fp32(const GrGLContextInfo& ctxInfo, const GrGLInterface* gli, GrGLenum precision) {
if (kGLES_GrGLStandard != ctxInfo.standard() &&
ctxInfo.version() < GR_GL_VER(4,1) &&
!ctxInfo.hasExtension("GL_ARB_ES2_compatibility")) {
// We're on a desktop GL that doesn't have precision info. Assume they're all 32bit float.
return true;
}
// glGetShaderPrecisionFormat doesn't accept GL_GEOMETRY_SHADER as a shader type. Hopefully the
// geometry shaders don't have lower precision than vertex and fragment.
for (GrGLenum shader : {GR_GL_FRAGMENT_SHADER, GR_GL_VERTEX_SHADER}) {
GrGLint range[2];
GrGLint bits;
GR_GL_GetShaderPrecisionFormat(gli, shader, precision, range, &bits);
if (range[0] < 127 || range[1] < 127 || bits < 23) {
return false;
}
}
return true;
}
void GrGLCaps::initGLSL(const GrGLContextInfo& ctxInfo, const GrGLInterface* gli) {
void GrGLCaps::initGLSL(const GrGLContextInfo& ctxInfo) {
GrGLStandard standard = ctxInfo.standard();
GrGLVersion version = ctxInfo.version();
@ -955,9 +937,6 @@ void GrGLCaps::initGLSL(const GrGLContextInfo& ctxInfo, const GrGLInterface* gli
shaderCaps->fVertexIDSupport = ctxInfo.glslGeneration() >= k330_GrGLSLGeneration;
}
shaderCaps->fFloatIs32Bits = is_float_fp32(ctxInfo, gli, GR_GL_HIGH_FLOAT);
shaderCaps->fHalfIs32Bits = is_float_fp32(ctxInfo, gli, GR_GL_MEDIUM_FLOAT);
if (kTegra3_GrGLRenderer == ctxInfo.renderer()) {
// The Tegra3 compiler will sometimes never return if we have min(abs(x), 1.0),
// so we must do the abs first in a separate expression.
@ -1429,6 +1408,92 @@ void GrGLCaps::onDumpJSON(SkJSONWriter* writer) const {
writer->endObject();
}
static GrGLenum precision_to_gl_float_type(GrSLPrecision p) {
switch (p) {
case kLow_GrSLPrecision:
return GR_GL_LOW_FLOAT;
case kMedium_GrSLPrecision:
return GR_GL_MEDIUM_FLOAT;
case kHigh_GrSLPrecision:
return GR_GL_HIGH_FLOAT;
default:
SK_ABORT("Unexpected precision type.");
return -1;
}
}
static GrGLenum shader_type_to_gl_shader(GrShaderType type) {
switch (type) {
case kVertex_GrShaderType:
return GR_GL_VERTEX_SHADER;
case kGeometry_GrShaderType:
return GR_GL_GEOMETRY_SHADER;
case kFragment_GrShaderType:
return GR_GL_FRAGMENT_SHADER;
}
SK_ABORT("Unknown shader type.");
return -1;
}
void GrGLCaps::initShaderPrecisionTable(const GrGLContextInfo& ctxInfo,
const GrGLInterface* intf,
GrShaderCaps* shaderCaps) {
if (kGLES_GrGLStandard == ctxInfo.standard() || ctxInfo.version() >= GR_GL_VER(4, 1) ||
ctxInfo.hasExtension("GL_ARB_ES2_compatibility")) {
for (int s = 0; s < kGrShaderTypeCount; ++s) {
if (kGeometry_GrShaderType != s) {
GrShaderType shaderType = static_cast<GrShaderType>(s);
GrGLenum glShader = shader_type_to_gl_shader(shaderType);
GrShaderCaps::PrecisionInfo* first = nullptr;
shaderCaps->fShaderPrecisionVaries = false;
for (int p = 0; p < kGrSLPrecisionCount; ++p) {
GrSLPrecision precision = static_cast<GrSLPrecision>(p);
GrGLenum glPrecision = precision_to_gl_float_type(precision);
GrGLint range[2];
GrGLint bits;
GR_GL_GetShaderPrecisionFormat(intf, glShader, glPrecision, range, &bits);
if (bits) {
shaderCaps->fFloatPrecisions[s][p].fLogRangeLow = range[0];
shaderCaps->fFloatPrecisions[s][p].fLogRangeHigh = range[1];
shaderCaps->fFloatPrecisions[s][p].fBits = bits;
if (!first) {
first = &shaderCaps->fFloatPrecisions[s][p];
}
else if (!shaderCaps->fShaderPrecisionVaries) {
shaderCaps->fShaderPrecisionVaries =
(*first != shaderCaps->fFloatPrecisions[s][p]);
}
}
}
}
}
}
else {
// We're on a desktop GL that doesn't have precision info. Assume they're all 32bit float.
shaderCaps->fShaderPrecisionVaries = false;
for (int s = 0; s < kGrShaderTypeCount; ++s) {
if (kGeometry_GrShaderType != s) {
for (int p = 0; p < kGrSLPrecisionCount; ++p) {
shaderCaps->fFloatPrecisions[s][p].fLogRangeLow = 127;
shaderCaps->fFloatPrecisions[s][p].fLogRangeHigh = 127;
shaderCaps->fFloatPrecisions[s][p].fBits = 23;
}
}
}
}
// GetShaderPrecisionFormat doesn't accept GL_GEOMETRY_SHADER as a shader type. Assume they're
// the same as the vertex shader. Only fragment shaders were ever allowed to omit support for
// highp. GS was added after GetShaderPrecisionFormat was added to the list of features that
// are recommended against.
if (shaderCaps->fGeometryShaderSupport) {
for (int p = 0; p < kGrSLPrecisionCount; ++p) {
shaderCaps->fFloatPrecisions[kGeometry_GrShaderType][p] =
shaderCaps->fFloatPrecisions[kVertex_GrShaderType][p];
}
}
shaderCaps->initSamplerPrecisionTable();
}
bool GrGLCaps::bgraIsInternalFormat() const {
return fConfigTable[kBGRA_8888_GrPixelConfig].fFormats.fBaseInternalFormat == GR_GL_BGRA;
}

4
src/gpu/gl/GrGLCaps.h
View File

@ -428,7 +428,7 @@ private:
GrGLenum* externalType) const;
void init(const GrContextOptions&, const GrGLContextInfo&, const GrGLInterface*);
void initGLSL(const GrGLContextInfo&, const GrGLInterface*);
void initGLSL(const GrGLContextInfo&);
bool hasPathRenderingSupport(const GrGLContextInfo&, const GrGLInterface*);
void onApplyOptionsOverrides(const GrContextOptions& options) override;
@ -441,6 +441,8 @@ private:
void initConfigTable(const GrContextOptions&, const GrGLContextInfo&, const GrGLInterface*,
GrShaderCaps*);
void initShaderPrecisionTable(const GrGLContextInfo&, const GrGLInterface*, GrShaderCaps*);
GrGLStandard fStandard;
SkTArray<StencilFormat, true> fStencilFormats;

4
src/gpu/glsl/GrGLSLProgramBuilder.cpp
View File

@ -303,7 +303,7 @@ GrGLSLProgramBuilder::SamplerHandle GrGLSLProgramBuilder::emitSampler(GrSLType s
const char* name,
GrShaderFlags visibility) {
this->updateSamplerCounts(visibility);
GrSLPrecision precision = GrSLSamplerPrecision(config);
GrSLPrecision precision = this->shaderCaps()->samplerPrecision(config, visibility);
GrSwizzle swizzle = this->shaderCaps()->configTextureSwizzle(config);
return this->uniformHandler()->addSampler(visibility, swizzle, samplerType, precision, name);
}
@ -311,7 +311,7 @@ GrGLSLProgramBuilder::SamplerHandle GrGLSLProgramBuilder::emitSampler(GrSLType s
GrGLSLProgramBuilder::TexelBufferHandle GrGLSLProgramBuilder::emitTexelBuffer(
GrPixelConfig config, const char* name, GrShaderFlags visibility) {
this->updateSamplerCounts(visibility);
GrSLPrecision precision = GrSLSamplerPrecision(config);
GrSLPrecision precision = this->shaderCaps()->samplerPrecision(config, visibility);
return this->uniformHandler()->addTexelBuffer(visibility, precision, name);
}

17
src/gpu/mtl/GrMtlCaps.mm
View File

@ -229,10 +229,21 @@ void GrMtlCaps::initShaderCaps() {
shaderCaps->fTexelFetchSupport = false;
shaderCaps->fVertexIDSupport = false;
shaderCaps->fImageLoadStoreSupport = false;
shaderCaps->fShaderPrecisionVaries = false; // ???
// Metal uses IEEE float and half floats so assuming those values here.
shaderCaps->fFloatIs32Bits = true;
shaderCaps->fHalfIs32Bits = false;
// Metal uses IEEE float and half floats so using those values here.
for (int s = 0; s < kGrShaderTypeCount; ++s) {
auto& highp = shaderCaps->fFloatPrecisions[s][kHigh_GrSLPrecision];
highp.fLogRangeLow = highp.fLogRangeHigh = 127;
highp.fBits = 23;
auto& mediump = shaderCaps->fFloatPrecisions[s][kMedium_GrSLPrecision];
mediump.fLogRangeLow = mediump.fLogRangeHigh = 15;
mediump.fBits = 10;
shaderCaps->fFloatPrecisions[s][kLow_GrSLPrecision] = mediump;
}
shaderCaps->initSamplerPrecisionTable();
shaderCaps->fMaxVertexSamplers =
shaderCaps->fMaxFragmentSamplers = 16;

15
src/gpu/vk/GrVkCaps.cpp
View File

@ -259,8 +259,19 @@ void GrVkCaps::initShaderCaps(const VkPhysicalDeviceProperties& properties, uint
shaderCaps->fVertexIDSupport = true;
// Assume the minimum precisions mandated by the SPIR-V spec.
shaderCaps->fFloatIs32Bits = true;
shaderCaps->fHalfIs32Bits = false;
shaderCaps->fShaderPrecisionVaries = true;
for (int s = 0; s < kGrShaderTypeCount; ++s) {
auto& highp = shaderCaps->fFloatPrecisions[s][kHigh_GrSLPrecision];
highp.fLogRangeLow = highp.fLogRangeHigh = 127;
highp.fBits = 23;
auto& mediump = shaderCaps->fFloatPrecisions[s][kMedium_GrSLPrecision];
mediump.fLogRangeLow = mediump.fLogRangeHigh = 14;
mediump.fBits = 10;
shaderCaps->fFloatPrecisions[s][kLow_GrSLPrecision] = mediump;
}
shaderCaps->initSamplerPrecisionTable();
shaderCaps->fMaxVertexSamplers =
shaderCaps->fMaxGeometrySamplers =

2
src/sksl/SkSLCPP.h
View File

@ -18,7 +18,7 @@ using std::abs;
// macros to make sk_Caps.<cap name> work from C++ code
#define sk_Caps (*args.fShaderCaps)
#define floatIs32Bits floatIs32Bits()
#define floatPrecisionVaries floatPrecisionVaries()
// functions to make GLSL constructors work from C++ code
inline SkPoint float2(float xy) { return SkPoint::Make(xy, xy); }

2
src/sksl/SkSLIRGenerator.cpp
View File

@ -140,7 +140,7 @@ static void fill_caps(const SKSL_CAPS_CLASS& caps,
CAP(mustEnableSpecificAdvBlendEqs);
CAP(mustDeclareFragmentShaderOutput);
CAP(canUseAnyFunctionInShader);
CAP(floatIs32Bits);
CAP(floatPrecisionVaries);
CAP(integerSupport);
#undef CAP
}

4
src/sksl/SkSLUtil.h
View File

@ -135,8 +135,8 @@ public:
return false;
}
bool floatIs32Bits() const {
return true;
bool floatPrecisionVaries() const {
return false;
}
bool integerSupport() const {

2
tests/ApplyGammaTest.cpp
View File

@ -105,7 +105,7 @@ DEF_GPUTEST_FOR_RENDERING_CONTEXTS(ApplyGamma, reporter, ctxInfo) {
SkAutoTMalloc<uint32_t> read(kW * kH);
// We allow more error on GPUs with lower precision shader variables.
float error = context->caps()->shaderCaps()->halfIs32Bits() ? 0.5f : 1.2f;
float error = context->caps()->shaderCaps()->floatPrecisionVaries() ? 1.2f : 0.5f;
for (auto toSRGB : { false, true }) {
sk_sp<SkSurface> dst(SkSurface::MakeRenderTarget(context, SkBudgeted::kNo, ii));

4
tests/SRGBReadWritePixelsTest.cpp
View File

@ -184,14 +184,14 @@ DEF_GPUTEST_FOR_RENDERING_CONTEXTS(SRGBReadWritePixels, reporter, ctxInfo) {
return;
}
float error = context->caps()->shaderCaps()->halfIs32Bits() ? 0.5f : 1.2f;
float error = context->caps()->shaderCaps()->floatPrecisionVaries() ? 1.2f : 0.5f;
// Write srgba data and read as srgba and then as rgba
if (sContext->writePixels(iiSRGBA, origData, 0, 0, 0)) {
// For the all-srgba case, we allow a small error only for devices that have
// precision variation because the srgba data gets converted to linear and back in
// the shader.
float smallError = context->caps()->shaderCaps()->halfIs32Bits() ? 0.0f : 1.f;
float smallError = context->caps()->shaderCaps()->floatPrecisionVaries() ? 1.f : 0.0f;
read_and_check_pixels(reporter, sContext.get(), origData, iiSRGBA,
check_srgb_to_linear_to_srgb_conversion, smallError,
"write/read srgba to srgba texture");