diff options
author | Chris Dalton <csmartdalton@google.com> | 2017-11-15 18:27:09 -0700 |
---|---|---|
committer | Skia Commit-Bot <skia-commit-bot@chromium.org> | 2017-11-16 03:31:45 +0000 |
commit | e42180022720f2fcfd3c634cad855506a7940591 (patch) | |
tree | 76fb7187cd7dd509da717a221b3c309f9b703d3b /src/gpu | |
parent | ae9d6189ec03c4a188e67b42d755c42024397f44 (diff) |
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>
Diffstat (limited to 'src/gpu')
-rw-r--r-- | src/gpu/GrProgramDesc.cpp | 2 | ||||
-rw-r--r-- | src/gpu/GrShaderCaps.cpp | 104 | ||||
-rw-r--r-- | src/gpu/effects/GrEllipseEffect.cpp | 33 | ||||
-rw-r--r-- | src/gpu/effects/GrEllipseEffect.fp | 26 | ||||
-rw-r--r-- | src/gpu/effects/GrRRectEffect.cpp | 68 | ||||
-rw-r--r-- | src/gpu/gl/GrGLCaps.cpp | 115 | ||||
-rw-r--r-- | src/gpu/gl/GrGLCaps.h | 4 | ||||
-rw-r--r-- | src/gpu/glsl/GrGLSLProgramBuilder.cpp | 4 | ||||
-rw-r--r-- | src/gpu/mtl/GrMtlCaps.mm | 17 | ||||
-rw-r--r-- | src/gpu/vk/GrVkCaps.cpp | 15 |
10 files changed, 101 insertions, 287 deletions
diff --git a/src/gpu/GrProgramDesc.cpp b/src/gpu/GrProgramDesc.cpp index a32e0cc28b..9c0add66e6 100644 --- a/src/gpu/GrProgramDesc.cpp +++ b/src/gpu/GrProgramDesc.cpp @@ -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 | - (caps.samplerPrecision(config, visibility) << (8 + kSamplerOrImageTypeKeyBits))); + (GrSLSamplerPrecision(config) << (8 + kSamplerOrImageTypeKeyBits))); } static void add_sampler_and_image_keys(GrProcessorKeyBuilder* b, const GrResourceIOProcessor& proc, diff --git a/src/gpu/GrShaderCaps.cpp b/src/gpu/GrShaderCaps.cpp index 124fd4de40..2d9cd0a39b 100644 --- a/src/gpu/GrShaderCaps.cpp +++ b/src/gpu/GrShaderCaps.cpp @@ -13,32 +13,6 @@ //////////////////////////////////////////////////////////////////////////////////////////// -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; @@ -50,7 +24,6 @@ GrShaderCaps::GrShaderCaps(const GrContextOptions& options) { fIntegerSupport = false; fTexelBufferSupport = false; fImageLoadStoreSupport = false; - fShaderPrecisionVaries = false; fDropsTileOnZeroDivide = false; fFBFetchSupport = false; fFBFetchNeedsCustomOutput = false; @@ -73,6 +46,8 @@ GrShaderCaps::GrShaderCaps(const GrContextOptions& options) { fExternalTextureSupport = false; fTexelFetchSupport = false; fVertexIDSupport = false; + fFloatIs32Bits = true; + fHalfIs32Bits = false; fVersionDeclString = nullptr; fShaderDerivativeExtensionString = nullptr; @@ -113,26 +88,6 @@ 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", @@ -166,6 +121,8 @@ 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); @@ -178,59 +135,6 @@ 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; diff --git a/src/gpu/effects/GrEllipseEffect.cpp b/src/gpu/effects/GrEllipseEffect.cpp index 4a0f37665c..6b1a8cc59f 100644 --- a/src/gpu/effects/GrEllipseEffect.cpp +++ b/src/gpu/effects/GrEllipseEffect.cpp @@ -28,34 +28,33 @@ public: (void)center; auto radii = _outer.radii(); (void)radii; - prevRadii = half2(-1.0); - useScale = sk_Caps.floatPrecisionVaries; + prevRadii = float2(-1.0); + useScale = !sk_Caps.floatIs32Bits; fEllipseVar = args.fUniformHandler->addUniform(kFragment_GrShaderFlag, kFloat4_GrSLType, kDefault_GrSLPrecision, "ellipse"); if (useScale) { - fScaleVar = args.fUniformHandler->addUniform(kFragment_GrShaderFlag, kHalf2_GrSLType, + fScaleVar = args.fUniformHandler->addUniform(kFragment_GrShaderFlag, kFloat2_GrSLType, kDefault_GrSLPrecision, "scale"); } fragBuilder->codeAppendf( - "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.", + "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 ", prevRadii.fX, prevRadii.fY, (useScale ? "true" : "false"), args.fUniformHandler->getUniformCStr(fEllipseVar), - fScaleVar.isValid() ? args.fUniformHandler->getUniformCStr(fScaleVar) : "half2(0)", + fScaleVar.isValid() ? args.fUniformHandler->getUniformCStr(fScaleVar) : "float2(0)", args.fUniformHandler->getUniformCStr(fEllipseVar), - fScaleVar.isValid() ? args.fUniformHandler->getUniformCStr(fScaleVar) : "half2(0)", + fScaleVar.isValid() ? args.fUniformHandler->getUniformCStr(fScaleVar) : "float2(0)", (int)_outer.edgeType()); fragBuilder->codeAppendf( - "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", + " 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", args.fOutputColor, args.fInputColor ? args.fInputColor : "half4(1)"); } diff --git a/src/gpu/effects/GrEllipseEffect.fp b/src/gpu/effects/GrEllipseEffect.fp index 54641b72c6..27beb85d2f 100644 --- a/src/gpu/effects/GrEllipseEffect.fp +++ b/src/gpu/effects/GrEllipseEffect.fp @@ -6,17 +6,17 @@ */ layout(key) in GrClipEdgeType edgeType; -in half2 center; -in half2 radii; +in float2 center; +in float2 radii; -half2 prevCenter; -half2 prevRadii = half2(-1); +float2 prevCenter; +float2 prevRadii = float2(-1); // The ellipse uniform is (center.x, center.y, 1 / rx^2, 1 / ry^2) -// The last two terms can underflow with halfs, so we use floats. +// The last two terms can underflow when float != fp32, so we also provide a workaround. uniform float4 ellipse; -bool useScale = sk_Caps.floatPrecisionVaries; -layout(when=useScale) uniform half2 scale; +bool useScale = !sk_Caps.floatIs32Bits; +layout(when=useScale) uniform float2 scale; @optimizationFlags { kCompatibleWithCoverageAsAlpha_OptimizationFlag } @@ -50,7 +50,7 @@ layout(when=useScale) uniform half2 scale; void main() { // d is the offset to the ellipse center - half2 d = sk_FragCoord.xy - ellipse.xy; + float2 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; } - half2 Z = d * ellipse.zw; + float2 Z = d * ellipse.zw; // implicit is the evaluation of (x/rx)^2 + (y/ry)^2 - 1. - half implicit = dot(Z, d) - 1; + float implicit = dot(Z, d) - 1; // grad_dot is the squared length of the gradient of the implicit. - half grad_dot = 4 * dot(Z, Z); + float grad_dot = 4 * dot(Z, Z); // Avoid calling inversesqrt on zero. grad_dot = max(grad_dot, 1e-4); - half approx_dist = implicit * inversesqrt(grad_dot); + float 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)); -}
\ No newline at end of file +} diff --git a/src/gpu/effects/GrRRectEffect.cpp b/src/gpu/effects/GrRRectEffect.cpp index 4b56443b0d..510a26b425 100644 --- a/src/gpu/effects/GrRRectEffect.cpp +++ b/src/gpu/effects/GrRRectEffect.cpp @@ -165,9 +165,9 @@ void GLCircularRRectEffect::emitCode(EmitArgs& args) { fRadiusPlusHalfUniform = uniformHandler->addUniform(kFragment_GrShaderFlag, kHalf2_GrSLType, "radiusPlusHalf", &radiusPlusHalfName); - // If we're on a device with a "real" mediump then the length calculation could overflow. + // If we're on a device where float != fp32 then the length calculation could overflow. SkString clampedCircleDistance; - if (args.fShaderCaps->floatPrecisionVaries()) { + if (!args.fShaderCaps->floatIs32Bits()) { 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("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("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("half alpha = %s;", clampedCircleDistance.c_str()); break; case CircularRRectEffect::kTopLeft_CornerFlag: - fragBuilder->codeAppendf("half2 dxy = max(%s.xy - sk_FragCoord.xy, 0.0);", + fragBuilder->codeAppendf("float2 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("half2 dxy = max(half2(sk_FragCoord.x - %s.z, " - "%s.y - sk_FragCoord.y), 0.0);", + fragBuilder->codeAppendf("float2 dxy = max(float2(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("half2 dxy = max(sk_FragCoord.xy - %s.zw, 0.0);", + fragBuilder->codeAppendf("float2 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("half2 dxy = max(half2(%s.x - sk_FragCoord.x, sk_FragCoord.y - " - "%s.w), 0.0);", + fragBuilder->codeAppendf("float2 dxy = max(float2(%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("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("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("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("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("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("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("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("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 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("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("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 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("half2 dxy0 = %s.xy - sk_FragCoord.xy;", rectName); - fragBuilder->codeAppendf("half2 dxy1 = sk_FragCoord.xy - %s.zw;", rectName); + fragBuilder->codeAppendf("float2 dxy0 = %s.xy - sk_FragCoord.xy;", rectName); + fragBuilder->codeAppendf("float2 dxy1 = sk_FragCoord.xy - %s.zw;", rectName); - // If we're on a device with a "real" mediump then we'll do the distance computation in a space + // If we're on a device where float != fp32 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->floatPrecisionVaries()) { + if (!args.fShaderCaps->floatIs32Bits()) { fScaleUniform = uniformHandler->addUniform(kFragment_GrShaderFlag, kHalf2_GrSLType, "scale", &scaleName); } @@ -543,12 +543,12 @@ void GLEllipticalRRectEffect::emitCode(EmitArgs& args) { kHalf2_GrSLType, "invRadiiXY", &invRadiiXYSqdName); - fragBuilder->codeAppend("half2 dxy = max(max(dxy0, dxy1), 0.0);"); + fragBuilder->codeAppend("float2 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("half2 Z = dxy * %s.xy;", invRadiiXYSqdName); + fragBuilder->codeAppendf("float2 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("half2 dxy = max(max(dxy0, dxy1), 0.0);"); + fragBuilder->codeAppend("float2 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("half2 Z = max(max(dxy0 * %s.xy, dxy1 * %s.zw), 0.0);", + fragBuilder->codeAppendf("float2 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("half implicit = dot(Z, dxy) - 1.0;"); + fragBuilder->codeAppend("float implicit = dot(Z, dxy) - 1.0;"); // grad_dot is the squared length of the gradient of the implicit. - fragBuilder->codeAppend("half grad_dot = 4.0 * dot(Z, Z);"); + fragBuilder->codeAppend("float grad_dot = 4.0 * dot(Z, Z);"); // avoid calling inversesqrt on zero. fragBuilder->codeAppend("grad_dot = max(grad_dot, 1.0e-4);"); - fragBuilder->codeAppend("half approx_dist = implicit * inversesqrt(grad_dot);"); + fragBuilder->codeAppend("float approx_dist = implicit * inversesqrt(grad_dot);"); if (scaleName) { fragBuilder->codeAppendf("approx_dist *= %s.x;", scaleName); } diff --git a/src/gpu/gl/GrGLCaps.cpp b/src/gpu/gl/GrGLCaps.cpp index fe9bce8528..0c185e4a78 100644 --- a/src/gpu/gl/GrGLCaps.cpp +++ b/src/gpu/gl/GrGLCaps.cpp @@ -291,7 +291,7 @@ void GrGLCaps::init(const GrContextOptions& contextOptions, **************************************************************************/ // This must be called after fCoreProfile is set on the GrGLCaps - this->initGLSL(ctxInfo); + this->initGLSL(ctxInfo, gli); GrShaderCaps* shaderCaps = fShaderCaps.get(); shaderCaps->fPathRenderingSupport = this->hasPathRenderingSupport(ctxInfo, gli); @@ -634,8 +634,6 @@ 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) { @@ -760,7 +758,27 @@ const char* get_glsl_version_decl_string(GrGLStandard standard, GrGLSLGeneration return "<no version>"; } -void GrGLCaps::initGLSL(const GrGLContextInfo& ctxInfo) { +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) { GrGLStandard standard = ctxInfo.standard(); GrGLVersion version = ctxInfo.version(); @@ -937,6 +955,9 @@ void GrGLCaps::initGLSL(const GrGLContextInfo& ctxInfo) { 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. @@ -1408,92 +1429,6 @@ 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; } diff --git a/src/gpu/gl/GrGLCaps.h b/src/gpu/gl/GrGLCaps.h index 923b3f416d..3a50e8a393 100644 --- a/src/gpu/gl/GrGLCaps.h +++ b/src/gpu/gl/GrGLCaps.h @@ -428,7 +428,7 @@ private: GrGLenum* externalType) const; void init(const GrContextOptions&, const GrGLContextInfo&, const GrGLInterface*); - void initGLSL(const GrGLContextInfo&); + void initGLSL(const GrGLContextInfo&, const GrGLInterface*); bool hasPathRenderingSupport(const GrGLContextInfo&, const GrGLInterface*); void onApplyOptionsOverrides(const GrContextOptions& options) override; @@ -441,8 +441,6 @@ private: void initConfigTable(const GrContextOptions&, const GrGLContextInfo&, const GrGLInterface*, GrShaderCaps*); - void initShaderPrecisionTable(const GrGLContextInfo&, const GrGLInterface*, GrShaderCaps*); - GrGLStandard fStandard; SkTArray<StencilFormat, true> fStencilFormats; diff --git a/src/gpu/glsl/GrGLSLProgramBuilder.cpp b/src/gpu/glsl/GrGLSLProgramBuilder.cpp index 6f8d1be736..118b904540 100644 --- a/src/gpu/glsl/GrGLSLProgramBuilder.cpp +++ b/src/gpu/glsl/GrGLSLProgramBuilder.cpp @@ -303,7 +303,7 @@ GrGLSLProgramBuilder::SamplerHandle GrGLSLProgramBuilder::emitSampler(GrSLType s const char* name, GrShaderFlags visibility) { this->updateSamplerCounts(visibility); - GrSLPrecision precision = this->shaderCaps()->samplerPrecision(config, visibility); + GrSLPrecision precision = GrSLSamplerPrecision(config); 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 = this->shaderCaps()->samplerPrecision(config, visibility); + GrSLPrecision precision = GrSLSamplerPrecision(config); return this->uniformHandler()->addTexelBuffer(visibility, precision, name); } diff --git a/src/gpu/mtl/GrMtlCaps.mm b/src/gpu/mtl/GrMtlCaps.mm index d431f0a1e1..2354b85e64 100644 --- a/src/gpu/mtl/GrMtlCaps.mm +++ b/src/gpu/mtl/GrMtlCaps.mm @@ -229,21 +229,10 @@ void GrMtlCaps::initShaderCaps() { shaderCaps->fTexelFetchSupport = false; shaderCaps->fVertexIDSupport = false; shaderCaps->fImageLoadStoreSupport = false; - shaderCaps->fShaderPrecisionVaries = 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(); + // Metal uses IEEE float and half floats so assuming those values here. + shaderCaps->fFloatIs32Bits = true; + shaderCaps->fHalfIs32Bits = false; shaderCaps->fMaxVertexSamplers = shaderCaps->fMaxFragmentSamplers = 16; diff --git a/src/gpu/vk/GrVkCaps.cpp b/src/gpu/vk/GrVkCaps.cpp index 2b24205af8..f5c4b342ed 100644 --- a/src/gpu/vk/GrVkCaps.cpp +++ b/src/gpu/vk/GrVkCaps.cpp @@ -259,19 +259,8 @@ void GrVkCaps::initShaderCaps(const VkPhysicalDeviceProperties& properties, uint shaderCaps->fVertexIDSupport = true; // Assume the minimum precisions mandated by the SPIR-V spec. - 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->fFloatIs32Bits = true; + shaderCaps->fHalfIs32Bits = false; shaderCaps->fMaxVertexSamplers = shaderCaps->fMaxGeometrySamplers = |