Add const-eval for pack and unpack of 2x16float.

This CL adds const-eval for pack and unpack of 2x16 float values.

Bug: tint:1581
Change-Id: I59a1925148124e628c3771ca96d309fad045f27d
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/109280
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Dan Sinclair <dsinclair@chromium.org>

src/tint/intrinsics.def

@@ -507,7 +507,7 @@ fn mix<N: num, T: f32_f16>(vec<N, T>, vec<N, T>, T) -> vec<N, T>
 fn modf<T: f32_f16>(T) -> __modf_result<T>
 fn modf<N: num, T: f32_f16>(vec<N, T>) -> __modf_result_vec<N, T>
 fn normalize<N: num, T: f32_f16>(vec<N, T>) -> vec<N, T>
-fn pack2x16float(vec2<f32>) -> u32
+@const fn pack2x16float(vec2<f32>) -> u32
 @const fn pack2x16snorm(vec2<f32>) -> u32
 @const fn pack2x16unorm(vec2<f32>) -> u32
 @const fn pack4x8snorm(vec4<f32>) -> u32
@@ -549,7 +549,7 @@ fn tanh<N: num, T: f32_f16>(vec<N, T>) -> vec<N, T>
 fn transpose<M: num, N: num, T: f32_f16>(mat<M, N, T>) -> mat<N, M, T>
 fn trunc<T: f32_f16>(T) -> T
 fn trunc<N: num, T: f32_f16>(vec<N, T>) -> vec<N, T>
-fn unpack2x16float(u32) -> vec2<f32>
+@const fn unpack2x16float(u32) -> vec2<f32>
 @const fn unpack2x16snorm(u32) -> vec2<f32>
 @const fn unpack2x16unorm(u32) -> vec2<f32>
 @const fn unpack4x8snorm(u32) -> vec4<f32>

src/tint/resolver/const_eval.cc

@@ -2076,6 +2076,34 @@ ConstEval::Result ConstEval::insertBits(const sem::Type* ty,
     return TransformElements(builder, ty, transform, args[0], args[1]);
 }
 
+ConstEval::Result ConstEval::pack2x16float(const sem::Type* ty,
+                                           utils::VectorRef<const sem::Constant*> args,
+                                           const Source& source) {
+    auto convert = [&](f32 val) -> utils::Result<uint32_t> {
+        auto conv = CheckedConvert<f16>(val);
+        if (!conv) {
+            AddError(OverflowErrorMessage(val, "f16"), source);
+            return utils::Failure;
+        }
+        uint16_t v = conv.Get().BitsRepresentation();
+        return utils::Result<uint32_t>{v};
+    };
+
+    auto* e = args[0];
+    auto e0 = convert(e->Index(0)->As<f32>());
+    if (!e0) {
+        return utils::Failure;
+    }
+
+    auto e1 = convert(e->Index(1)->As<f32>());
+    if (!e1) {
+        return utils::Failure;
+    }
+
+    u32 ret = u32((e0.Get() & 0x0000'ffff) | (e1.Get() << 16));
+    return CreateElement(builder, ty, ret);
+}
+
 ConstEval::Result ConstEval::pack2x16snorm(const sem::Type* ty,
                                            utils::VectorRef<const sem::Constant*> args,
                                            const Source&) {
@@ -2254,6 +2282,26 @@ ConstEval::Result ConstEval::step(const sem::Type* ty,
     return TransformElements(builder, ty, transform, args[0], args[1]);
 }
 
+ConstEval::Result ConstEval::unpack2x16float(const sem::Type* ty,
+                                             utils::VectorRef<const sem::Constant*> args,
+                                             const Source& source) {
+    auto* inner_ty = sem::Type::DeepestElementOf(ty);
+    auto e = args[0]->As<u32>().value;
+
+    utils::Vector<const sem::Constant*, 2> els;
+    els.Reserve(2);
+    for (size_t i = 0; i < 2; ++i) {
+        auto in = f16::FromBits(uint16_t((e >> (16 * i)) & 0x0000'ffff));
+        auto val = CheckedConvert<f32>(in);
+        if (!val) {
+            AddError(OverflowErrorMessage(in, "f32"), source);
+            return utils::Failure;
+        }
+        els.Push(CreateElement(builder, inner_ty, val.Get()));
+    }
+    return CreateComposite(builder, ty, std::move(els));
+}
+
 ConstEval::Result ConstEval::unpack2x16snorm(const sem::Type* ty,
                                              utils::VectorRef<const sem::Constant*> args,
                                              const Source&) {

src/tint/resolver/const_eval.h

@@ -557,6 +557,15 @@ class ConstEval {
                       utils::VectorRef<const sem::Constant*> args,
                       const Source& source);
 
+    /// pack2x16float builtin
+    /// @param ty the expression type
+    /// @param args the input arguments
+    /// @param source the source location of the conversion
+    /// @return the result value, or null if the value cannot be calculated
+    Result pack2x16float(const sem::Type* ty,
+                         utils::VectorRef<const sem::Constant*> args,
+                         const Source& source);
+
     /// pack2x16snorm builtin
     /// @param ty the expression type
     /// @param args the input arguments
@@ -647,6 +656,15 @@ class ConstEval {
                 utils::VectorRef<const sem::Constant*> args,
                 const Source& source);
 
+    /// unpack2x16float builtin
+    /// @param ty the expression type
+    /// @param args the input arguments
+    /// @param source the source location of the conversion
+    /// @return the result value, or null if the value cannot be calculated
+    Result unpack2x16float(const sem::Type* ty,
+                           utils::VectorRef<const sem::Constant*> args,
+                           const Source& source);
+
     /// unpack2x16snorm builtin
     /// @param ty the expression type
     /// @param args the input arguments

src/tint/resolver/const_eval_builtin_test.cc

@@ -1265,6 +1265,25 @@ INSTANTIATE_TEST_SUITE_P(  //
     testing::Combine(testing::Values(sem::BuiltinType::kPack4X8Unorm),
                      testing::ValuesIn(Pack4x8unormCases())));
 
+std::vector<Case> Pack2x16floatCases() {
+    return {
+        C({Vec(f32(f16::Lowest()), f32(f16::Highest()))}, Val(u32(0x7bff'fbff))),
+        C({Vec(f32(1), f32(-1))}, Val(u32(0xbc00'3c00))),
+        C({Vec(f32(0), f32(0))}, Val(u32(0x0000'0000))),
+        C({Vec(f32(10), f32(-10.5))}, Val(u32(0xc940'4900))),
+
+        E({Vec(f32(0), f32::Highest())},
+          "12:34 error: value 3.4028234663852885981e+38 cannot be represented as 'f16'"),
+        E({Vec(f32::Lowest(), f32(0))},
+          "12:34 error: value -3.4028234663852885981e+38 cannot be represented as 'f16'"),
+    };
+}
+INSTANTIATE_TEST_SUITE_P(  //
+    Pack2x16float,
+    ResolverConstEvalBuiltinTest,
+    testing::Combine(testing::Values(sem::BuiltinType::kPack2X16Float),
+                     testing::ValuesIn(Pack2x16floatCases())));
+
 std::vector<Case> Pack2x16snormCases() {
     return {
         C({Vec(f32(0), f32(0))}, Val(u32(0x0000'0000))),
@@ -1508,6 +1527,20 @@ INSTANTIATE_TEST_SUITE_P(  //
     testing::Combine(testing::Values(sem::BuiltinType::kUnpack4X8Unorm),
                      testing::ValuesIn(Unpack4x8unormCases())));
 
+std::vector<Case> Unpack2x16floatCases() {
+    return {
+        C({Val(u32(0x7bff'fbff))}, Vec(f32(f16::Lowest()), f32(f16::Highest()))),
+        C({Val(u32(0xbc00'3c00))}, Vec(f32(1), f32(-1))),
+        C({Val(u32(0x0000'0000))}, Vec(f32(0), f32(0))),
+        C({Val(u32(0xc940'4900))}, Vec(f32(10), f32(-10.5))),
+    };
+}
+INSTANTIATE_TEST_SUITE_P(  //
+    Unpack2x16float,
+    ResolverConstEvalBuiltinTest,
+    testing::Combine(testing::Values(sem::BuiltinType::kUnpack2X16Float),
+                     testing::ValuesIn(Unpack2x16floatCases())));
+
 std::vector<Case> Unpack2x16snormCases() {
     return {
         C({Val(u32(0x0000'0000))}, Vec(f32(0), f32(0))),

test/tint/builtins/gen/literal/pack2x16float/0e97b3.wgsl.expected.dxc.hlsl

@@ -1,10 +1,5 @@
-uint tint_pack2x16float(float2 param_0) {
-  uint2 i = f32tof16(param_0);
-  return i.x | (i.y << 16);
-}
-
 void pack2x16float_0e97b3() {
-  uint res = tint_pack2x16float((1.0f).xx);
+  uint res = 1006648320u;
 }
 
 struct tint_symbol {

test/tint/builtins/gen/literal/pack2x16float/0e97b3.wgsl.expected.fxc.hlsl

@@ -1,10 +1,5 @@
-uint tint_pack2x16float(float2 param_0) {
-  uint2 i = f32tof16(param_0);
-  return i.x | (i.y << 16);
-}
-
 void pack2x16float_0e97b3() {
-  uint res = tint_pack2x16float((1.0f).xx);
+  uint res = 1006648320u;
 }
 
 struct tint_symbol {

test/tint/builtins/gen/literal/pack2x16float/0e97b3.wgsl.expected.glsl

@@ -1,7 +1,7 @@
 #version 310 es
 
 void pack2x16float_0e97b3() {
-  uint res = packHalf2x16(vec2(1.0f));
+  uint res = 1006648320u;
 }
 
 vec4 vertex_main() {
@@ -21,7 +21,7 @@ void main() {
 precision mediump float;
 
 void pack2x16float_0e97b3() {
-  uint res = packHalf2x16(vec2(1.0f));
+  uint res = 1006648320u;
 }
 
 void fragment_main() {
@@ -35,7 +35,7 @@ void main() {
 #version 310 es
 
 void pack2x16float_0e97b3() {
-  uint res = packHalf2x16(vec2(1.0f));
+  uint res = 1006648320u;
 }
 
 void compute_main() {

test/tint/builtins/gen/literal/pack2x16float/0e97b3.wgsl.expected.msl

@@ -2,7 +2,7 @@
 
 using namespace metal;
 void pack2x16float_0e97b3() {
-  uint res = as_type<uint>(half2(float2(1.0f)));
+  uint res = 1006648320u;
 }
 
 struct tint_symbol {

test/tint/builtins/gen/literal/pack2x16float/0e97b3.wgsl.expected.spvasm

@@ -1,10 +1,9 @@
 ; SPIR-V
 ; Version: 1.3
 ; Generator: Google Tint Compiler; 0
-; Bound: 35
+; Bound: 32
 ; Schema: 0
                OpCapability Shader
-         %15 = OpExtInstImport "GLSL.std.450"
                OpMemoryModel Logical GLSL450
                OpEntryPoint Vertex %vertex_main "vertex_main" %value %vertex_point_size
                OpEntryPoint Fragment %fragment_main "fragment_main"
@@ -32,38 +31,36 @@
        %void = OpTypeVoid
           %9 = OpTypeFunction %void
        %uint = OpTypeInt 32 0
-    %v2float = OpTypeVector %float 2
+%uint_1006648320 = OpConstant %uint 1006648320
-    %float_1 = OpConstant %float 1
-         %18 = OpConstantComposite %v2float %float_1 %float_1
 %_ptr_Function_uint = OpTypePointer Function %uint
-         %21 = OpConstantNull %uint
+         %17 = OpConstantNull %uint
-         %22 = OpTypeFunction %v4float
+         %18 = OpTypeFunction %v4float
+    %float_1 = OpConstant %float 1
 %pack2x16float_0e97b3 = OpFunction %void None %9
          %12 = OpLabel
-        %res = OpVariable %_ptr_Function_uint Function %21
+        %res = OpVariable %_ptr_Function_uint Function %17
-         %13 = OpExtInst %uint %15 PackHalf2x16 %18
+               OpStore %res %uint_1006648320
-               OpStore %res %13
                OpReturn
                OpFunctionEnd
-%vertex_main_inner = OpFunction %v4float None %22
+%vertex_main_inner = OpFunction %v4float None %18
-         %24 = OpLabel
+         %20 = OpLabel
-         %25 = OpFunctionCall %void %pack2x16float_0e97b3
+         %21 = OpFunctionCall %void %pack2x16float_0e97b3
                OpReturnValue %5
                OpFunctionEnd
 %vertex_main = OpFunction %void None %9
-         %27 = OpLabel
+         %23 = OpLabel
-         %28 = OpFunctionCall %v4float %vertex_main_inner
+         %24 = OpFunctionCall %v4float %vertex_main_inner
-               OpStore %value %28
+               OpStore %value %24
                OpStore %vertex_point_size %float_1
                OpReturn
                OpFunctionEnd
 %fragment_main = OpFunction %void None %9
+         %27 = OpLabel
+         %28 = OpFunctionCall %void %pack2x16float_0e97b3
+               OpReturn
+               OpFunctionEnd
+%compute_main = OpFunction %void None %9
          %30 = OpLabel
          %31 = OpFunctionCall %void %pack2x16float_0e97b3
                OpReturn
                OpFunctionEnd
-%compute_main = OpFunction %void None %9
-         %33 = OpLabel
-         %34 = OpFunctionCall %void %pack2x16float_0e97b3
-               OpReturn
-               OpFunctionEnd

test/tint/builtins/gen/literal/unpack2x16float/32a5cf.wgsl.expected.dxc.hlsl

@@ -1,10 +1,5 @@
-float2 tint_unpack2x16float(uint param_0) {
-  uint i = param_0;
-  return f16tof32(uint2(i & 0xffff, i >> 16));
-}
-
 void unpack2x16float_32a5cf() {
-  float2 res = tint_unpack2x16float(1u);
+  float2 res = float2(5.96046448e-08f, 0.0f);
 }
 
 struct tint_symbol {

test/tint/builtins/gen/literal/unpack2x16float/32a5cf.wgsl.expected.fxc.hlsl

@@ -1,10 +1,5 @@
-float2 tint_unpack2x16float(uint param_0) {
-  uint i = param_0;
-  return f16tof32(uint2(i & 0xffff, i >> 16));
-}
-
 void unpack2x16float_32a5cf() {
-  float2 res = tint_unpack2x16float(1u);
+  float2 res = float2(5.96046448e-08f, 0.0f);
 }
 
 struct tint_symbol {

test/tint/builtins/gen/literal/unpack2x16float/32a5cf.wgsl.expected.glsl

@@ -1,7 +1,7 @@
 #version 310 es
 
 void unpack2x16float_32a5cf() {
-  vec2 res = unpackHalf2x16(1u);
+  vec2 res = vec2(5.96046448e-08f, 0.0f);
 }
 
 vec4 vertex_main() {
@@ -21,7 +21,7 @@ void main() {
 precision mediump float;
 
 void unpack2x16float_32a5cf() {
-  vec2 res = unpackHalf2x16(1u);
+  vec2 res = vec2(5.96046448e-08f, 0.0f);
 }
 
 void fragment_main() {
@@ -35,7 +35,7 @@ void main() {
 #version 310 es
 
 void unpack2x16float_32a5cf() {
-  vec2 res = unpackHalf2x16(1u);
+  vec2 res = vec2(5.96046448e-08f, 0.0f);
 }
 
 void compute_main() {

test/tint/builtins/gen/literal/unpack2x16float/32a5cf.wgsl.expected.msl

@@ -2,7 +2,7 @@
 
 using namespace metal;
 void unpack2x16float_32a5cf() {
-  float2 res = float2(as_type<half2>(1u));
+  float2 res = float2(5.96046448e-08f, 0.0f);
 }
 
 struct tint_symbol {

test/tint/builtins/gen/literal/unpack2x16float/32a5cf.wgsl.expected.spvasm

@@ -1,10 +1,9 @@
 ; SPIR-V
 ; Version: 1.3
 ; Generator: Google Tint Compiler; 0
-; Bound: 35
+; Bound: 33
 ; Schema: 0
                OpCapability Shader
-         %15 = OpExtInstImport "GLSL.std.450"
                OpMemoryModel Logical GLSL450
                OpEntryPoint Vertex %vertex_main "vertex_main" %value %vertex_point_size
                OpEntryPoint Fragment %fragment_main "fragment_main"
@@ -32,38 +31,37 @@
        %void = OpTypeVoid
           %9 = OpTypeFunction %void
     %v2float = OpTypeVector %float 2
-       %uint = OpTypeInt 32 0
+%float_5_96046448en08 = OpConstant %float 5.96046448e-08
-     %uint_1 = OpConstant %uint 1
+         %15 = OpConstantComposite %v2float %float_5_96046448en08 %8
 %_ptr_Function_v2float = OpTypePointer Function %v2float
-         %20 = OpConstantNull %v2float
+         %18 = OpConstantNull %v2float
-         %21 = OpTypeFunction %v4float
+         %19 = OpTypeFunction %v4float
     %float_1 = OpConstant %float 1
 %unpack2x16float_32a5cf = OpFunction %void None %9
          %12 = OpLabel
-        %res = OpVariable %_ptr_Function_v2float Function %20
+        %res = OpVariable %_ptr_Function_v2float Function %18
-         %13 = OpExtInst %v2float %15 UnpackHalf2x16 %uint_1
+               OpStore %res %15
-               OpStore %res %13
                OpReturn
                OpFunctionEnd
-%vertex_main_inner = OpFunction %v4float None %21
+%vertex_main_inner = OpFunction %v4float None %19
-         %23 = OpLabel
+         %21 = OpLabel
-         %24 = OpFunctionCall %void %unpack2x16float_32a5cf
+         %22 = OpFunctionCall %void %unpack2x16float_32a5cf
                OpReturnValue %5
                OpFunctionEnd
 %vertex_main = OpFunction %void None %9
-         %26 = OpLabel
+         %24 = OpLabel
-         %27 = OpFunctionCall %v4float %vertex_main_inner
+         %25 = OpFunctionCall %v4float %vertex_main_inner
-               OpStore %value %27
+               OpStore %value %25
                OpStore %vertex_point_size %float_1
                OpReturn
                OpFunctionEnd
 %fragment_main = OpFunction %void None %9
-         %30 = OpLabel
+         %28 = OpLabel
-         %31 = OpFunctionCall %void %unpack2x16float_32a5cf
+         %29 = OpFunctionCall %void %unpack2x16float_32a5cf
                OpReturn
                OpFunctionEnd
 %compute_main = OpFunction %void None %9
-         %33 = OpLabel
+         %31 = OpLabel
-         %34 = OpFunctionCall %void %unpack2x16float_32a5cf
+         %32 = OpFunctionCall %void %unpack2x16float_32a5cf
                OpReturn
                OpFunctionEnd