Implement Float16 formats and additional blend modes.
The float16 loads and stores are horrendously unoptimized, but this can be addressed as a later pass.
Tests: dEQP-VK.renderpass.suballocation.formats.r16_sfloat.*
Tests: dEQP-VK.renderpass.suballocation.formats.r16g16_sfloat.*
Tests: dEQP-VK.renderpass.suballocation.formats.r16g16b16a16_sfloat.*
Tests: dEQP-VK.pipeline.blend.format.r16_sfloat.*
Tests: dEQP-VK.pipeline.blend.format.r16g16_sfloat.*
Tests: dEQP-VK.pipeline.blend.format.r16g16b16a16_sfloat.*
Bug: b/132434966
Bug: b/132433217
Change-Id: Ifa8feaeecefa1926b1f500e6c9d23e6c242a6844
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/31113
Tested-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
diff --git a/src/Pipeline/PixelProgram.cpp b/src/Pipeline/PixelProgram.cpp
index 57aead8..a90a398 100644
--- a/src/Pipeline/PixelProgram.cpp
+++ b/src/Pipeline/PixelProgram.cpp
@@ -187,6 +187,9 @@
}
}
break;
+ case VK_FORMAT_R16_SFLOAT:
+ case VK_FORMAT_R16G16_SFLOAT:
+ case VK_FORMAT_R16G16B16A16_SFLOAT:
case VK_FORMAT_R32_SFLOAT:
case VK_FORMAT_R32G32_SFLOAT:
case VK_FORMAT_R32G32B32A32_SFLOAT:
@@ -268,6 +271,9 @@
case VK_FORMAT_R32_UINT:
case VK_FORMAT_R32G32_UINT:
case VK_FORMAT_R32G32B32A32_UINT:
+ case VK_FORMAT_R16_SFLOAT:
+ case VK_FORMAT_R16G16_SFLOAT:
+ case VK_FORMAT_R16G16B16A16_SFLOAT:
case VK_FORMAT_R16_SINT:
case VK_FORMAT_R16G16_SINT:
case VK_FORMAT_R16G16B16A16_SINT:
diff --git a/src/Pipeline/PixelRoutine.cpp b/src/Pipeline/PixelRoutine.cpp
index 1cb8b23..616647c 100644
--- a/src/Pipeline/PixelRoutine.cpp
+++ b/src/Pipeline/PixelRoutine.cpp
@@ -1728,11 +1728,22 @@
blendFactor.y = *Pointer<Float4>(data + OFFSET(DrawData,factor.blendConstant4F[1]));
blendFactor.z = *Pointer<Float4>(data + OFFSET(DrawData,factor.blendConstant4F[2]));
break;
+ case VK_BLEND_FACTOR_CONSTANT_ALPHA:
+ blendFactor.x = *Pointer<Float4>(data + OFFSET(DrawData,factor.blendConstant4F[3]));
+ blendFactor.y = *Pointer<Float4>(data + OFFSET(DrawData,factor.blendConstant4F[3]));
+ blendFactor.z = *Pointer<Float4>(data + OFFSET(DrawData,factor.blendConstant4F[3]));
+ break;
case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR:
blendFactor.x = *Pointer<Float4>(data + OFFSET(DrawData,factor.invBlendConstant4F[0]));
blendFactor.y = *Pointer<Float4>(data + OFFSET(DrawData,factor.invBlendConstant4F[1]));
blendFactor.z = *Pointer<Float4>(data + OFFSET(DrawData,factor.invBlendConstant4F[2]));
break;
+ case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA:
+ blendFactor.x = *Pointer<Float4>(data + OFFSET(DrawData,factor.invBlendConstant4F[3]));
+ blendFactor.y = *Pointer<Float4>(data + OFFSET(DrawData,factor.invBlendConstant4F[3]));
+ blendFactor.z = *Pointer<Float4>(data + OFFSET(DrawData,factor.invBlendConstant4F[3]));
+ break;
+
default:
UNIMPLEMENTED("VkBlendFactor: %d", int(blendFactorActive));
}
@@ -1776,9 +1787,11 @@
blendFactor.w = Float4(1.0f);
break;
case VK_BLEND_FACTOR_CONSTANT_COLOR:
+ case VK_BLEND_FACTOR_CONSTANT_ALPHA:
blendFactor.w = *Pointer<Float4>(data + OFFSET(DrawData,factor.blendConstant4F[3]));
break;
case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR:
+ case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA:
blendFactor.w = *Pointer<Float4>(data + OFFSET(DrawData,factor.invBlendConstant4F[3]));
break;
default:
@@ -1794,6 +1807,11 @@
}
Pointer<Byte> buffer;
+
+ // pixel holds four texel color values.
+ // Note: Despite the type being Vector4f, the colors may be stored as
+ // integers. Half-floats are stored as full 32-bit floats.
+ // Non-float and non-fixed point formats are not alpha blended.
Vector4f pixel;
Vector4s color;
@@ -1850,6 +1868,48 @@
pixel.w = *Pointer<Float4>(buffer + 16 * x + 16, 16);
transpose4x4(pixel.x, pixel.y, pixel.z, pixel.w);
break;
+ case VK_FORMAT_R16_SFLOAT:
+ buffer = cBuffer;
+ pixel.x.x = Float(*Pointer<Half>(buffer + 2 * x + 0));
+ pixel.x.y = Float(*Pointer<Half>(buffer + 2 * x + 2));
+ buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
+ pixel.x.z = Float(*Pointer<Half>(buffer + 2 * x + 0));
+ pixel.x.w = Float(*Pointer<Half>(buffer + 2 * x + 2));
+ pixel.y = pixel.z = pixel.w = one;
+ break;
+ case VK_FORMAT_R16G16_SFLOAT:
+ buffer = cBuffer;
+ pixel.x.x = Float(*Pointer<Half>(buffer + 4 * x + 0));
+ pixel.y.x = Float(*Pointer<Half>(buffer + 4 * x + 2));
+ pixel.x.y = Float(*Pointer<Half>(buffer + 4 * x + 4));
+ pixel.y.y = Float(*Pointer<Half>(buffer + 4 * x + 6));
+ buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
+ pixel.x.z = Float(*Pointer<Half>(buffer + 4 * x + 0));
+ pixel.y.z = Float(*Pointer<Half>(buffer + 4 * x + 2));
+ pixel.x.w = Float(*Pointer<Half>(buffer + 4 * x + 4));
+ pixel.y.w = Float(*Pointer<Half>(buffer + 4 * x + 6));
+ pixel.z = pixel.w = one;
+ break;
+ case VK_FORMAT_R16G16B16A16_SFLOAT:
+ buffer = cBuffer;
+ pixel.x.x = Float(*Pointer<Half>(buffer + 8 * x + 0x0));
+ pixel.y.x = Float(*Pointer<Half>(buffer + 8 * x + 0x2));
+ pixel.z.x = Float(*Pointer<Half>(buffer + 8 * x + 0x4));
+ pixel.w.x = Float(*Pointer<Half>(buffer + 8 * x + 0x6));
+ pixel.x.y = Float(*Pointer<Half>(buffer + 8 * x + 0x8));
+ pixel.y.y = Float(*Pointer<Half>(buffer + 8 * x + 0xa));
+ pixel.z.y = Float(*Pointer<Half>(buffer + 8 * x + 0xc));
+ pixel.w.y = Float(*Pointer<Half>(buffer + 8 * x + 0xe));
+ buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
+ pixel.x.z = Float(*Pointer<Half>(buffer + 8 * x + 0x0));
+ pixel.y.z = Float(*Pointer<Half>(buffer + 8 * x + 0x2));
+ pixel.z.z = Float(*Pointer<Half>(buffer + 8 * x + 0x4));
+ pixel.w.z = Float(*Pointer<Half>(buffer + 8 * x + 0x6));
+ pixel.x.w = Float(*Pointer<Half>(buffer + 8 * x + 0x8));
+ pixel.y.w = Float(*Pointer<Half>(buffer + 8 * x + 0xa));
+ pixel.z.w = Float(*Pointer<Half>(buffer + 8 * x + 0xc));
+ pixel.w.w = Float(*Pointer<Half>(buffer + 8 * x + 0xe));
+ break;
default:
UNIMPLEMENTED("VkFormat: %d", int(state.targetFormat[index]));
}
@@ -1975,6 +2035,7 @@
{
switch(state.targetFormat[index])
{
+ case VK_FORMAT_R16_SFLOAT:
case VK_FORMAT_R32_SFLOAT:
case VK_FORMAT_R32_SINT:
case VK_FORMAT_R32_UINT:
@@ -1984,6 +2045,7 @@
case VK_FORMAT_R8_UINT:
case VK_FORMAT_A2B10G10R10_UINT_PACK32:
break;
+ case VK_FORMAT_R16G16_SFLOAT:
case VK_FORMAT_R32G32_SFLOAT:
case VK_FORMAT_R32G32_SINT:
case VK_FORMAT_R32G32_UINT:
@@ -1996,6 +2058,7 @@
oC.z = UnpackHigh(oC.z, oC.y);
oC.y = oC.z;
break;
+ case VK_FORMAT_R16G16B16A16_SFLOAT:
case VK_FORMAT_R32G32B32A32_SFLOAT:
case VK_FORMAT_R32G32B32A32_SINT:
case VK_FORMAT_R32G32B32A32_UINT:
@@ -2029,10 +2092,12 @@
xMask &= sMask;
}
+ auto targetFormat = state.targetFormat[index];
+
Pointer<Byte> buffer;
Float4 value;
- switch(state.targetFormat[index])
+ switch(targetFormat)
{
case VK_FORMAT_R32_SFLOAT:
case VK_FORMAT_R32_SINT:
@@ -2066,6 +2131,32 @@
*Pointer<Float>(buffer + 4) = oC.x.y;
}
break;
+ case VK_FORMAT_R16_SFLOAT:
+ if(rgbaWriteMask & 0x00000001)
+ {
+ buffer = cBuffer + 2 * x;
+
+ value = Insert(value, Float(*Pointer<Half>(buffer + 0)), 0);
+ value = Insert(value, Float(*Pointer<Half>(buffer + 2)), 1);
+
+ buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
+
+ value = Insert(value, Float(*Pointer<Half>(buffer + 0)), 2);
+ value = Insert(value, Float(*Pointer<Half>(buffer + 2)), 3);
+
+ oC.x = As<Float4>(As<Int4>(oC.x) & *Pointer<Int4>(constants + OFFSET(Constants, maskD4X) + xMask * 16, 16));
+ value = As<Float4>(As<Int4>(value) & *Pointer<Int4>(constants + OFFSET(Constants, invMaskD4X) + xMask * 16, 16));
+ oC.x = As<Float4>(As<Int4>(oC.x) | As<Int4>(value));
+
+ *Pointer<Half>(buffer + 0) = Half(oC.x.z);
+ *Pointer<Half>(buffer + 2) = Half(oC.x.w);
+
+ buffer -= *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
+
+ *Pointer<Half>(buffer + 0) = Half(oC.x.x);
+ *Pointer<Half>(buffer + 2) = Half(oC.x.y);
+ }
+ break;
case VK_FORMAT_R16_SINT:
case VK_FORMAT_R16_UINT:
if(rgbaWriteMask & 0x00000001)
@@ -2084,7 +2175,7 @@
value = As<Float4>(As<Int4>(value) & *Pointer<Int4>(constants + OFFSET(Constants, invMaskD4X) + xMask * 16, 16));
oC.x = As<Float4>(As<Int4>(oC.x) | As<Int4>(value));
- if(state.targetFormat[index] == VK_FORMAT_R16_SINT)
+ if(targetFormat == VK_FORMAT_R16_SINT)
{
Float component = oC.x.z;
*Pointer<Short>(buffer + 0) = Short(As<Int>(component));
@@ -2127,7 +2218,7 @@
xyzw |= UInt(*Pointer<UShort>(buffer)) << 16;
Short4 tmpCol = Short4(As<Int4>(oC.x));
- if(state.targetFormat[index] == VK_FORMAT_R8_SINT)
+ if(targetFormat == VK_FORMAT_R8_SINT)
{
tmpCol = As<Short4>(PackSigned(tmpCol, tmpCol));
}
@@ -2184,6 +2275,39 @@
oC.y = As<Float4>(As<Int4>(oC.y) | As<Int4>(value));
*Pointer<Float4>(buffer) = oC.y;
break;
+ case VK_FORMAT_R16G16_SFLOAT:
+ if((rgbaWriteMask & 0x00000003) != 0x0)
+ {
+ buffer = cBuffer + 4 * x;
+
+ UInt2 rgbaMask;
+ UInt2 packedCol;
+ packedCol = Insert(packedCol, (UInt(As<UShort>(Half(oC.x.y))) << 16) | UInt(As<UShort>(Half(oC.x.x))), 0);
+ packedCol = Insert(packedCol, (UInt(As<UShort>(Half(oC.x.w))) << 16) | UInt(As<UShort>(Half(oC.x.z))), 1);
+
+ UShort4 value = *Pointer<UShort4>(buffer);
+ UInt2 mergedMask = *Pointer<UInt2>(constants + OFFSET(Constants, maskD01Q) + xMask * 8);
+ if((rgbaWriteMask & 0x3) != 0x3)
+ {
+ Int tmpMask = *Pointer<Int>(constants + OFFSET(Constants, maskW4Q[rgbaWriteMask & 0x3][0]));
+ rgbaMask = As<UInt2>(Int2(tmpMask, tmpMask));
+ mergedMask &= rgbaMask;
+ }
+ *Pointer<UInt2>(buffer) = (packedCol & mergedMask) | (As<UInt2>(value) & ~mergedMask);
+
+ buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
+
+ packedCol = Insert(packedCol, (UInt(As<UShort>(Half(oC.y.y))) << 16) | UInt(As<UShort>(Half(oC.y.x))), 0);
+ packedCol = Insert(packedCol, (UInt(As<UShort>(Half(oC.y.w))) << 16) | UInt(As<UShort>(Half(oC.y.z))), 1);
+ value = *Pointer<UShort4>(buffer);
+ mergedMask = *Pointer<UInt2>(constants + OFFSET(Constants, maskD23Q) + xMask * 8);
+ if((rgbaWriteMask & 0x3) != 0x3)
+ {
+ mergedMask &= rgbaMask;
+ }
+ *Pointer<UInt2>(buffer) = (packedCol & mergedMask) | (As<UInt2>(value) & ~mergedMask);
+ }
+ break;
case VK_FORMAT_R16G16_SINT:
case VK_FORMAT_R16G16_UINT:
if((rgbaWriteMask & 0x00000003) != 0x0)
@@ -2226,7 +2350,7 @@
buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
xyzw = Insert(xyzw, *Pointer<Int>(buffer), 1);
- if(state.targetFormat[index] == VK_FORMAT_R8G8_SINT)
+ if(targetFormat == VK_FORMAT_R8G8_SINT)
{
packedCol = As<Int2>(PackSigned(Short4(As<Int4>(oC.x)), Short4(As<Int4>(oC.y))));
}
@@ -2325,6 +2449,42 @@
*Pointer<Float4>(buffer + 16, 16) = oC.w;
}
break;
+ case VK_FORMAT_R16G16B16A16_SFLOAT:
+ if((rgbaWriteMask & 0x0000000F) != 0x0)
+ {
+ buffer = cBuffer + 8 * x;
+
+ UInt4 rgbaMask;
+ UInt4 value = *Pointer<UInt4>(buffer);
+ UInt4 packedCol;
+ packedCol = Insert(packedCol, (UInt(As<UShort>(Half(oC.x.y))) << 16) | UInt(As<UShort>(Half(oC.x.x))), 0);
+ packedCol = Insert(packedCol, (UInt(As<UShort>(Half(oC.x.w))) << 16) | UInt(As<UShort>(Half(oC.x.z))), 1);
+ packedCol = Insert(packedCol, (UInt(As<UShort>(Half(oC.y.y))) << 16) | UInt(As<UShort>(Half(oC.y.x))), 2);
+ packedCol = Insert(packedCol, (UInt(As<UShort>(Half(oC.y.w))) << 16) | UInt(As<UShort>(Half(oC.y.z))), 3);
+ UInt4 mergedMask = *Pointer<UInt4>(constants + OFFSET(Constants, maskQ01X) + xMask * 16);
+ if((rgbaWriteMask & 0xF) != 0xF)
+ {
+ UInt2 tmpMask = *Pointer<UInt2>(constants + OFFSET(Constants, maskW4Q[rgbaWriteMask][0]));
+ rgbaMask = UInt4(tmpMask, tmpMask);
+ mergedMask &= rgbaMask;
+ }
+ *Pointer<UInt4>(buffer) = (packedCol & mergedMask) | (As<UInt4>(value) & ~mergedMask);
+
+ buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
+
+ value = *Pointer<UInt4>(buffer);
+ packedCol = Insert(packedCol, (UInt(As<UShort>(Half(oC.z.y))) << 16) | UInt(As<UShort>(Half(oC.z.x))), 0);
+ packedCol = Insert(packedCol, (UInt(As<UShort>(Half(oC.z.w))) << 16) | UInt(As<UShort>(Half(oC.z.z))), 1);
+ packedCol = Insert(packedCol, (UInt(As<UShort>(Half(oC.w.y))) << 16) | UInt(As<UShort>(Half(oC.w.x))), 2);
+ packedCol = Insert(packedCol, (UInt(As<UShort>(Half(oC.w.w))) << 16) | UInt(As<UShort>(Half(oC.w.z))), 3);
+ mergedMask = *Pointer<UInt4>(constants + OFFSET(Constants, maskQ23X) + xMask * 16);
+ if((rgbaWriteMask & 0xF) != 0xF)
+ {
+ mergedMask &= rgbaMask;
+ }
+ *Pointer<UInt4>(buffer) = (packedCol & mergedMask) | (As<UInt4>(value) & ~mergedMask);
+ }
+ break;
case VK_FORMAT_R16G16B16A16_SINT:
case VK_FORMAT_R16G16B16A16_UINT:
if((rgbaWriteMask & 0x0000000F) != 0x0)
@@ -2365,7 +2525,7 @@
buffer = cBuffer + 4 * x;
- bool isSigned = state.targetFormat[index] == VK_FORMAT_R8G8B8A8_SINT || state.targetFormat[index] == VK_FORMAT_A8B8G8R8_SINT_PACK32;
+ bool isSigned = targetFormat == VK_FORMAT_R8G8B8A8_SINT || targetFormat == VK_FORMAT_A8B8G8R8_SINT_PACK32;
if(isSigned)
{
@@ -2432,7 +2592,7 @@
}
break;
default:
- UNIMPLEMENTED("VkFormat: %d", int(state.targetFormat[index]));
+ UNIMPLEMENTED("VkFormat: %d", int(targetFormat));
}
}
diff --git a/src/Reactor/Reactor.cpp b/src/Reactor/Reactor.cpp
index c4740a7..f9a0072 100644
--- a/src/Reactor/Reactor.cpp
+++ b/src/Reactor/Reactor.cpp
@@ -3176,6 +3176,16 @@
return RValue<UInt2>(Nucleus::createNot(val.value));
}
+ RValue<UInt> Extract(RValue<UInt2> val, int i)
+ {
+ return RValue<UInt>(Nucleus::createExtractElement(val.value, UInt::getType(), i));
+ }
+
+ RValue<UInt2> Insert(RValue<UInt2> val, RValue<UInt> element, int i)
+ {
+ return RValue<UInt2>(Nucleus::createInsertElement(val.value, element.value, i));
+ }
+
Int4::Int4() : XYZW(this)
{
}
diff --git a/src/Reactor/Reactor.hpp b/src/Reactor/Reactor.hpp
index 6fbc061..5cf1be0 100644
--- a/src/Reactor/Reactor.hpp
+++ b/src/Reactor/Reactor.hpp
@@ -1399,6 +1399,8 @@
// RValue<Bool> operator==(RValue<UInt2> lhs, RValue<UInt2> rhs);
// RValue<UInt2> RoundInt(RValue<Float4> cast);
+ RValue<UInt> Extract(RValue<UInt2> val, int i);
+ RValue<UInt2> Insert(RValue<UInt2> val, RValue<UInt> element, int i);
template<class T>
struct Scalar;
diff --git a/src/Reactor/SubzeroReactor.cpp b/src/Reactor/SubzeroReactor.cpp
index 4310aba..ba18fd2 100644
--- a/src/Reactor/SubzeroReactor.cpp
+++ b/src/Reactor/SubzeroReactor.cpp
@@ -2689,16 +2689,6 @@
return T(Type_v2i32);
}
- RValue<UInt> Extract(RValue<UInt2> val, int i)
- {
- return RValue<UInt>(Nucleus::createExtractElement(val.value, UInt::getType(), i));
- }
-
- RValue<UInt2> Insert(RValue<UInt2> val, RValue<UInt> element, int i)
- {
- return RValue<UInt2>(Nucleus::createInsertElement(val.value, element.value, i));
- }
-
RValue<UInt2> operator<<(RValue<UInt2> lhs, unsigned char rhs)
{
if(emulateIntrinsics)
