Reland "Improve matrix construction abilities in Metal codegen."

This is a reland of daa573eb91

Original change's description:
> Improve matrix construction abilities in Metal codegen.
>
> GLSL (and thus SkSL) is flexible about the input parameters to a matrix
> constructor. You can mix vectors and scalars freely, and it will
> populate them into your matrix as if it was a flat list of scalars.
>
> Metal does not natively support this, and requires the proper number of
> floatNs to be passed in. However, the Metal code generator will now emit
> constructor helper functions that will fix this up automatically.
>
> Additionally, this CL simplifies the Metal codegen for single-scalar
> matrix construction. This should create a matrix with the passed-in
> scalar running along the matrix diagonal. The Metal codegen previously
> emitted a helper function to do this work on our behalf. However,
> that's not necessary; Metal already contains a single-argument matrix
> constructor that will do this work automatically for us.
>
> Change-Id: I76901bfe167502797aa4cb98d0e8986d9ebc51e5
> Bug: skia:10280
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/292477
> Auto-Submit: John Stiles <johnstiles@google.com>
> Commit-Queue: Brian Osman <brianosman@google.com>
> Reviewed-by: Brian Osman <brianosman@google.com>
> Reviewed-by: Ethan Nicholas <ethannicholas@google.com>

Bug: skia:10280
Change-Id: If5591392bb96e1cfb643d4e3c19a0ee4affec58d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/292689
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This commit is contained in:
John Stiles 2020-05-28 12:17:20 -04:00 committed by Skia Commit-Bot
parent b985b4b67f
commit 1bdafbf016
3 changed files with 282 additions and 145 deletions

View File

@ -379,82 +379,94 @@ void MetalCodeGenerator::writeSpecialIntrinsic(const FunctionCall & c, SpecialIn
}
}
// If it hasn't already been written, writes a constructor for 'matrix' which takes a single value
// of type 'arg'.
String MetalCodeGenerator::getMatrixConstructHelper(const Type& matrix, const Type& arg) {
String key = matrix.name() + arg.name();
auto found = fHelpers.find(key);
if (found != fHelpers.end()) {
return found->second;
}
String name;
// Generates a constructor for 'matrix' which reorganizes the input arguments into the proper shape.
// Keeps track of previously generated constructors so that we won't generate more than one
// constructor for any given permutation of input argument types. Returns the name of the
// generated constructor method.
String MetalCodeGenerator::getMatrixConstructHelper(const Constructor& c) {
const Type& matrix = c.fType;
int columns = matrix.columns();
int rows = matrix.rows();
if (arg.isNumber()) {
// creating a matrix from a single scalar value
name = "float" + to_string(columns) + "x" + to_string(rows) + "_from_float";
fExtraFunctions.printf("float%dx%d %s(float x) {\n",
columns, rows, name.c_str());
fExtraFunctions.printf(" return float%dx%d(", columns, rows);
for (int i = 0; i < columns; ++i) {
if (i > 0) {
fExtraFunctions.writeText(", ");
}
fExtraFunctions.printf("float%d(", rows);
for (int j = 0; j < rows; ++j) {
if (j > 0) {
fExtraFunctions.writeText(", ");
const std::vector<std::unique_ptr<Expression>>& args = c.fArguments;
// Create the helper-method name and use it as our lookup key.
String name;
name.appendf("float%dx%d_from", columns, rows);
for (const std::unique_ptr<Expression>& expr : args) {
name.appendf("_%s", expr->fType.displayName().c_str());
}
// If a helper-method has already been synthesized, we don't need to synthesize it again.
auto [iter, newlyCreated] = fHelpers.insert(name);
if (!newlyCreated) {
return name;
}
// Unlike GLSL, Metal requires that matrices are initialized with exactly R vectors of C
// components apiece. (In Metal 2.0, you can also supply R*C scalars, but you still cannot
// supply a mixture of scalars and vectors.)
fExtraFunctions.printf("float%dx%d %s(", columns, rows, name.c_str());
size_t argIndex = 0;
const char* argSeparator = "";
for (const std::unique_ptr<Expression>& expr : c.fArguments) {
fExtraFunctions.printf("%s%s x%zu", argSeparator,
expr->fType.displayName().c_str(), argIndex++);
argSeparator = ", ";
}
fExtraFunctions.printf(") {\n return float%dx%d(", columns, rows);
argIndex = 0;
int argPosition = 0;
const char* columnSeparator = "";
for (int c = 0; c < columns; ++c) {
fExtraFunctions.printf("%sfloat%d(", columnSeparator, rows);
columnSeparator = "), ";
const char* rowSeparator = "";
for (int r = 0; r < rows; ++r) {
fExtraFunctions.printf("%s", rowSeparator);
rowSeparator = ", ";
const Type& argType = args[argIndex]->fType;
switch (argType.kind()) {
case Type::kScalar_Kind: {
fExtraFunctions.printf("x%zu", argIndex);
break;
}
if (i == j) {
fExtraFunctions.writeText("x");
} else {
fExtraFunctions.writeText("0");
case Type::kVector_Kind: {
fExtraFunctions.printf("x%zu[%d]", argIndex, argPosition);
break;
}
case Type::kMatrix_Kind: {
fExtraFunctions.printf("x%zu[%d][%d]", argIndex,
argPosition / argType.rows(),
argPosition % argType.rows());
break;
}
default: {
SkDEBUGFAIL("incorrect type of argument for matrix constructor");
fExtraFunctions.printf("<error>");
break;
}
}
fExtraFunctions.writeText(")");
++argPosition;
if (argPosition >= argType.columns() * argType.rows()) {
++argIndex;
argPosition = 0;
}
}
fExtraFunctions.writeText(");\n}\n");
} else if (arg.kind() == Type::kMatrix_Kind) {
// creating a matrix from another matrix
int argColumns = arg.columns();
int argRows = arg.rows();
name = "float" + to_string(columns) + "x" + to_string(rows) + "_from_float" +
to_string(argColumns) + "x" + to_string(argRows);
fExtraFunctions.printf("float%dx%d %s(float%dx%d m) {\n",
columns, rows, name.c_str(), argColumns, argRows);
fExtraFunctions.printf(" return float%dx%d(", columns, rows);
for (int i = 0; i < columns; ++i) {
if (i > 0) {
fExtraFunctions.writeText(", ");
}
fExtraFunctions.printf("float%d(", rows);
for (int j = 0; j < rows; ++j) {
if (j > 0) {
fExtraFunctions.writeText(", ");
}
if (i < argColumns && j < argRows) {
fExtraFunctions.printf("m[%d][%d]", i, j);
} else {
fExtraFunctions.writeText("0");
}
}
fExtraFunctions.writeText(")");
}
fExtraFunctions.writeText(");\n}\n");
} else if (matrix.rows() == 2 && matrix.columns() == 2 && arg == *fContext.fFloat4_Type) {
// float2x2(float4) doesn't work, need to split it into float2x2(float2, float2)
name = "float2x2_from_float4";
fExtraFunctions.printf(
"float2x2 %s(float4 v) {\n"
" return float2x2(float2(v[0], v[1]), float2(v[2], v[3]));\n"
"}\n",
name.c_str()
);
} else {
SkASSERT(false);
}
if (argPosition != 0 || argIndex != args.size()) {
SkDEBUGFAIL("incorrect number of arguments for matrix constructor");
name = "<error>";
}
fHelpers[key] = name;
fExtraFunctions.printf("));\n}\n");
return name;
}
@ -468,43 +480,116 @@ bool MetalCodeGenerator::canCoerce(const Type& t1, const Type& t2) {
return t1.isFloat() && t2.isFloat();
}
void MetalCodeGenerator::writeConstructor(const Constructor& c, Precedence parentPrecedence) {
if (c.fArguments.size() == 1 && this->canCoerce(c.fType, c.fArguments[0]->fType)) {
this->writeExpression(*c.fArguments[0], parentPrecedence);
return;
bool MetalCodeGenerator::matrixConstructHelperIsNeeded(const Constructor& c) {
// A matrix construct helper is only necessary if we are, in fact, constructing a matrix.
if (c.fType.kind() != Type::kMatrix_Kind) {
return false;
}
if (c.fType.kind() == Type::kMatrix_Kind && c.fArguments.size() == 1) {
const Expression& arg = *c.fArguments[0];
String name = this->getMatrixConstructHelper(c.fType, arg.fType);
this->write(name);
this->write("(");
this->writeExpression(arg, kSequence_Precedence);
this->write(")");
} else {
this->writeType(c.fType);
// GLSL is fairly free-form about inputs to its matrix constructors, but Metal is not; it
// expects exactly R vectors of C components apiece. (Metal 2.0 also allows a list of R*C
// scalars.) Some cases are simple to translate and so we handle those inline--e.g. a list of
// scalars can be constructed trivially. In more complex cases, we generate a helper function
// that converts our inputs into a properly-shaped matrix.
// A matrix construct helper method is always used if any input argument is a matrix.
// Helper methods are also necessary when any argument would span multiple rows. For instance:
//
// float2 x = (1, 2);
// float3x2(x, 3, 4, 5, 6) = | 1 3 5 | = no helper needed; conversion can be done inline
// | 2 4 6 |
//
// float2 x = (2, 3);
// float3x2(1, x, 4, 5, 6) = | 1 3 5 | = x spans multiple rows; a helper method will be used
// | 2 4 6 |
//
// float4 x = (1, 2, 3, 4);
// float2x2(x) = | 1 3 | = x spans multiple rows; a helper method will be used
// | 2 4 |
//
int position = 0;
for (const std::unique_ptr<Expression>& expr : c.fArguments) {
// If an input argument is a matrix, we need a helper function.
if (expr->fType.kind() == Type::kMatrix_Kind) {
return true;
}
position += expr->fType.columns();
if (position > c.fType.rows()) {
// An input argument would span multiple rows; a helper function is required.
return true;
}
if (position == c.fType.rows()) {
// We've advanced to the end of a row. Wrap to the start of the next row.
position = 0;
}
}
return false;
}
void MetalCodeGenerator::writeConstructor(const Constructor& c, Precedence parentPrecedence) {
// Handle special cases for single-argument constructors.
if (c.fArguments.size() == 1) {
// If the type is coercible, emit it directly.
const Expression& arg = *c.fArguments.front();
if (this->canCoerce(c.fType, arg.fType)) {
this->writeExpression(arg, parentPrecedence);
return;
}
// Metal supports creating matrices with a scalar on the diagonal via the single-argument
// matrix constructor.
if (c.fType.kind() == Type::kMatrix_Kind && arg.fType.isNumber()) {
const Type& matrix = c.fType;
this->write("float");
this->write(to_string(matrix.columns()));
this->write("x");
this->write(to_string(matrix.rows()));
this->write("(");
this->writeExpression(arg, parentPrecedence);
this->write(")");
return;
}
}
// Emit and invoke a matrix-constructor helper method if one is necessary.
if (this->matrixConstructHelperIsNeeded(c)) {
this->write(this->getMatrixConstructHelper(c));
this->write("(");
const char* separator = "";
int scalarCount = 0;
for (const auto& arg : c.fArguments) {
for (const std::unique_ptr<Expression>& expr : c.fArguments) {
this->write(separator);
separator = ", ";
if (Type::kMatrix_Kind == c.fType.kind() && arg->fType.columns() != c.fType.rows()) {
// merge scalars and smaller vectors together
if (!scalarCount) {
this->writeType(c.fType.componentType());
this->write(to_string(c.fType.rows()));
this->write("(");
}
scalarCount += arg->fType.columns();
}
this->writeExpression(*arg, kSequence_Precedence);
if (scalarCount && scalarCount == c.fType.rows()) {
this->write(")");
scalarCount = 0;
}
this->writeExpression(*expr, kSequence_Precedence);
}
this->write(")");
return;
}
// Explicitly invoke the constructor, passing in the necessary arguments.
this->writeType(c.fType);
this->write("(");
const char* separator = "";
int scalarCount = 0;
for (const std::unique_ptr<Expression>& arg : c.fArguments) {
this->write(separator);
separator = ", ";
if (Type::kMatrix_Kind == c.fType.kind() && arg->fType.columns() < c.fType.rows()) {
// Merge scalars and smaller vectors together.
if (!scalarCount) {
this->writeType(c.fType.componentType());
this->write(to_string(c.fType.rows()));
this->write("(");
}
scalarCount += arg->fType.columns();
}
this->writeExpression(*arg, kSequence_Precedence);
if (scalarCount && scalarCount == c.fType.rows()) {
this->write(")");
scalarCount = 0;
}
}
this->write(")");
}
void MetalCodeGenerator::writeFragCoord() {

View File

@ -11,6 +11,7 @@
#include <stack>
#include <tuple>
#include <unordered_map>
#include <unordered_set>
#include "src/sksl/SkSLCodeGenerator.h"
#include "src/sksl/SkSLMemoryLayout.h"
@ -189,7 +190,8 @@ protected:
void writeInverseHack(const Expression& mat);
String getMatrixConstructHelper(const Type& matrix, const Type& arg);
bool matrixConstructHelperIsNeeded(const Constructor& c);
String getMatrixConstructHelper(const Constructor& c);
void writeMatrixTimesEqualHelper(const Type& left, const Type& right, const Type& result);
@ -278,7 +280,7 @@ protected:
std::unordered_map<const FunctionDeclaration*, Requirements> fRequirements;
bool fSetupFragPositionGlobal = false;
bool fSetupFragPositionLocal = false;
std::unordered_map<String, String> fHelpers;
std::unordered_set<String> fHelpers;
int fUniformBuffer = -1;
String fRTHeightName;

View File

@ -60,54 +60,104 @@ DEF_TEST(SkSLMetalHelloWorld, r) {
"}\n");
}
DEF_TEST(SkSLMetal2x2MatrixCopyFromFloat2x2, r) {
test(r, R"__SkSL__(
void main() {
float2x2 m1 = float2x2(float2(1, 2), float2(3, 4));
float2x2 m2 = m1;
float2x2 m3 = float2x2(m1);
sk_FragColor = half4(half(m1[0][0] + m2[0][0] + m3[0][0]));
})__SkSL__",
*SkSL::ShaderCapsFactory::Default(),
R"__MSL__(#include <metal_stdlib>
#include <simd/simd.h>
using namespace metal;
struct Inputs {
};
struct Outputs {
float4 sk_FragColor [[color(0)]];
};
fragment Outputs fragmentMain(Inputs _in [[stage_in]], bool _frontFacing [[front_facing]], float4 _fragCoord [[position]]) {
Outputs _outputStruct;
thread Outputs* _out = &_outputStruct;
_out->sk_FragColor = float4((float2x2(float2(1.0, 2.0), float2(3.0, 4.0))[0][0] + float2x2(float2(1.0, 2.0), float2(3.0, 4.0))[0][0]) + float2x2(float2(1.0, 2.0), float2(3.0, 4.0))[0][0]);
return *_out;
}
)__MSL__");
}
DEF_TEST(SkSLMetal2x2MatrixCopyFromConstantPropagatedFloat4, r) {
test(r, R"__SkSL__(
void main() {
float2x2 m1 = float2x2(float4(1, 2, 3, 4));
float2x2 m2 = m1;
float2x2 m3 = float2x2(m1);
sk_FragColor = half4(half(m1[0][0] + m2[0][0] + m3[0][0]));
})__SkSL__",
*SkSL::ShaderCapsFactory::Default(),
R"__MSL__(#include <metal_stdlib>
#include <simd/simd.h>
using namespace metal;
struct Inputs {
};
struct Outputs {
float4 sk_FragColor [[color(0)]];
};
float2x2 float2x2_from_float4(float4 x0) {
return float2x2(float2(x0[0], x0[1]), float2(x0[2], x0[3]));
}
fragment Outputs fragmentMain(Inputs _in [[stage_in]], bool _frontFacing [[front_facing]], float4 _fragCoord [[position]]) {
Outputs _outputStruct;
thread Outputs* _out = &_outputStruct;
_out->sk_FragColor = float4((float2x2_from_float4(float4(1.0, 2.0, 3.0, 4.0))[0][0] + float2x2_from_float4(float4(1.0, 2.0, 3.0, 4.0))[0][0]) + float2x2_from_float4(float4(1.0, 2.0, 3.0, 4.0))[0][0]);
return *_out;
}
)__MSL__");
}
DEF_TEST(SkSLMetalMatrices, r) {
test(r,
"void main() {"
"float2x2 m1 = float2x2(float4(1, 2, 3, 4));"
"float2x2 m2 = float2x2(float4(0));"
"float2x2 m3 = float2x2(m1);"
"float2x2 m4 = float2x2(1);"
"float2x2 m5 = float2x2(m1[0][0]);"
"float2x2 m6 = float2x2(1, 2, 3, 4);"
"float2x2 m7 = float2x2(5, 6, 7, 8);"
"float3x3 m8 = float3x3(1);"
"float3x3 m9 = float3x3(2);"
"float4x4 m10 = float4x4(1);"
"float4x4 m11 = float4x4(2);"
"sk_FragColor = half4(half(m1[0][0] + m2[0][0] + m3[0][0] + m4[0][0] + m5[0][0] + "
"m6[0][0] + m7[0][0] + m8[0][0] + m9[0][0] + m10[0][0] + m11[0][0]));"
"}",
test(r, R"__SkSL__(
void main() {
float2x2 m1 = float2x2(float4(1, 2, 3, 4));
float2x2 m2 = float2x2(float4(0));
float2x2 m3 = float2x2(m1);
float2x2 m4 = float2x2(1);
float2x2 m5 = float2x2(m1[0][0]);
float2x2 m6 = float2x2(1, 2, 3, 4);
float2x2 m7 = float2x2(5, float3(6, 7, 8));
float3x2 m8 = float3x2(float2(1, 2), 3, float3(4, 5, 6));
float3x3 m9 = float3x3(1);
float4x4 m10 = float4x4(1);
float4x4 m11 = float4x4(2);
sk_FragColor = half4(half(m1[0][0] + m2[0][0] + m3[0][0] + m4[0][0] + m5[0][0] +
m6[0][0] + m7[0][0] + m8[0][0] + m9[0][0] + m10[0][0] + m11[0][0]));
})__SkSL__",
*SkSL::ShaderCapsFactory::Default(),
"#include <metal_stdlib>\n"
"#include <simd/simd.h>\n"
"using namespace metal;\n"
"struct Inputs {\n"
"};\n"
"struct Outputs {\n"
" float4 sk_FragColor [[color(0)]];\n"
"};\n"
"float2x2 float2x2_from_float(float x) {\n"
" return float2x2(float2(x, 0), float2(0, x));\n"
"}\n"
"float2x2 float2x2_from_float4(float4 v) {\n"
" return float2x2(float2(v[0], v[1]), float2(v[2], v[3]));\n"
"}\n"
"float2x2 float2x2_from_float(float x) {\n"
" return float2x2(float2(x, 0), float2(0, x));\n"
"}\n"
"float3x3 float3x3_from_float(float x) {\n"
" return float3x3(float3(x, 0, 0), float3(0, x, 0), float3(0, 0, x));\n"
"}\n"
"float4x4 float4x4_from_float(float x) {\n"
" return float4x4(float4(x, 0, 0, 0), float4(0, x, 0, 0), float4(0, 0, x, 0), float4(0, 0, 0, x));\n"
"}\n"
"fragment Outputs fragmentMain(Inputs _in [[stage_in]], bool _frontFacing [[front_facing]], float4 _fragCoord [[position]]) {\n"
" Outputs _outputStruct;\n"
" thread Outputs* _out = &_outputStruct;\n"
" float2x2 m5 = float2x2_from_float(float2x2_from_float4(float4(1.0, 2.0, 3.0, 4.0))[0][0]);\n"
" _out->sk_FragColor = float4((((((((((float2x2_from_float4(float4(1.0, 2.0, 3.0, 4.0))[0][0] + float2x2_from_float4(float4(0.0))[0][0]) + float2x2_from_float4(float4(1.0, 2.0, 3.0, 4.0))[0][0]) + float2x2_from_float(1.0)[0][0]) + m5[0][0]) + float2x2(float2(1.0, 2.0), float2(3.0, 4.0))[0][0]) + float2x2(float2(5.0, 6.0), float2(7.0, 8.0))[0][0]) + float3x3_from_float(1.0)[0][0]) + float3x3_from_float(2.0)[0][0]) + float4x4_from_float(1.0)[0][0]) + float4x4_from_float(2.0)[0][0]);\n"
" return *_out;\n"
"}\n");
R"__MSL__(#include <metal_stdlib>
#include <simd/simd.h>
using namespace metal;
struct Inputs {
};
struct Outputs {
float4 sk_FragColor [[color(0)]];
};
float2x2 float2x2_from_float4(float4 x0) {
return float2x2(float2(x0[0], x0[1]), float2(x0[2], x0[3]));
}
float2x2 float2x2_from_float_float3(float x0, float3 x1) {
return float2x2(float2(x0, x1[0]), float2(x1[1], x1[2]));
}
float3x2 float3x2_from_float2_float_float3(float2 x0, float x1, float3 x2) {
return float3x2(float2(x0[0], x0[1]), float2(x1, x2[0]), float2(x2[1], x2[2]));
}
fragment Outputs fragmentMain(Inputs _in [[stage_in]], bool _frontFacing [[front_facing]], float4 _fragCoord [[position]]) {
Outputs _outputStruct;
thread Outputs* _out = &_outputStruct;
float2x2 m5 = float2x2(float2x2_from_float4(float4(1.0, 2.0, 3.0, 4.0))[0][0]);
_out->sk_FragColor = float4((((((((((float2x2_from_float4(float4(1.0, 2.0, 3.0, 4.0))[0][0] + float2x2_from_float4(float4(0.0))[0][0]) + float2x2_from_float4(float4(1.0, 2.0, 3.0, 4.0))[0][0]) + float2x2(1.0)[0][0]) + m5[0][0]) + float2x2(float2(1.0, 2.0), float2(3.0, 4.0))[0][0]) + float2x2_from_float_float3(5.0, float3(6.0, 7.0, 8.0))[0][0]) + float3x2_from_float2_float_float3(float2(1.0, 2.0), 3.0, float3(4.0, 5.0, 6.0))[0][0]) + float3x3(1.0)[0][0]) + float4x4(1.0)[0][0]) + float4x4(2.0)[0][0]);
return *_out;
}
)__MSL__");
}
DEF_TEST(SkSLMetalConstantSwizzle, r) {