SpirvShader: Implement GroupNonUniformShuffleRelative capability
Implements the following opcodes:
• OpGroupNonUniformShuffleUp
• OpGroupNonUniformShuffleDown
Bug: b/133510501
Tests: dEQP-VK.subgroups.*
Change-Id: Ia0c8b3f90b9247e7052c287601e110107dbaa17e
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/35090
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
Tested-by: Ben Clayton <bclayton@google.com>
diff --git a/src/Pipeline/SpirvShader.cpp b/src/Pipeline/SpirvShader.cpp
index 594f8d7..57fabdd 100644
--- a/src/Pipeline/SpirvShader.cpp
+++ b/src/Pipeline/SpirvShader.cpp
@@ -846,6 +846,7 @@
case spv::CapabilityGroupNonUniformVote: capabilities.GroupNonUniformVote = true; break;
case spv::CapabilityGroupNonUniformBallot: capabilities.GroupNonUniformBallot = true; break;
case spv::CapabilityGroupNonUniformShuffle: capabilities.GroupNonUniformShuffle = true; break;
+ case spv::CapabilityGroupNonUniformShuffleRelative: capabilities.GroupNonUniformShuffleRelative = true; break;
default:
UNSUPPORTED("Unsupported capability %u", insn.word(1));
}
@@ -1100,6 +1101,8 @@
case spv::OpGroupNonUniformBallotFindMSB:
case spv::OpGroupNonUniformShuffle:
case spv::OpGroupNonUniformShuffleXor:
+ case spv::OpGroupNonUniformShuffleUp:
+ case spv::OpGroupNonUniformShuffleDown:
case spv::OpCopyObject:
case spv::OpArrayLength:
// Instructions that yield an intermediate value or divergent pointer
@@ -2759,6 +2762,8 @@
case spv::OpGroupNonUniformBallotFindMSB:
case spv::OpGroupNonUniformShuffle:
case spv::OpGroupNonUniformShuffleXor:
+ case spv::OpGroupNonUniformShuffleUp:
+ case spv::OpGroupNonUniformShuffleDown:
return EmitGroupNonUniform(insn, state);
case spv::OpArrayLength:
@@ -6193,6 +6198,38 @@
break;
}
+ case spv::OpGroupNonUniformShuffleUp:
+ {
+ GenericValue value(this, state, insn.word(4));
+ GenericValue delta(this, state, insn.word(5));
+ auto d0 = CmpEQ(SIMD::Int(0), delta.Int(0));
+ auto d1 = CmpEQ(SIMD::Int(1), delta.Int(0));
+ auto d2 = CmpEQ(SIMD::Int(2), delta.Int(0));
+ auto d3 = CmpEQ(SIMD::Int(3), delta.Int(0));
+ for (auto i = 0u; i < type.sizeInComponents; i++)
+ {
+ SIMD::Int v = value.Int(i);
+ dst.move(i, (d0 & v.xyzw) | (d1 & v.xxyz) | (d2 & v.xxxy) | (d3 & v.xxxx));
+ }
+ break;
+ }
+
+ case spv::OpGroupNonUniformShuffleDown:
+ {
+ GenericValue value(this, state, insn.word(4));
+ GenericValue delta(this, state, insn.word(5));
+ auto d0 = CmpEQ(SIMD::Int(0), delta.Int(0));
+ auto d1 = CmpEQ(SIMD::Int(1), delta.Int(0));
+ auto d2 = CmpEQ(SIMD::Int(2), delta.Int(0));
+ auto d3 = CmpEQ(SIMD::Int(3), delta.Int(0));
+ for (auto i = 0u; i < type.sizeInComponents; i++)
+ {
+ SIMD::Int v = value.Int(i);
+ dst.move(i, (d0 & v.xyzw) | (d1 & v.yzww) | (d2 & v.zwww) | (d3 & v.wwww));
+ }
+ break;
+ }
+
default:
UNIMPLEMENTED("EmitGroupNonUniform op: %s", OpcodeName(type.opcode()).c_str());
}
diff --git a/src/Pipeline/SpirvShader.hpp b/src/Pipeline/SpirvShader.hpp
index c668f8f..862235b 100644
--- a/src/Pipeline/SpirvShader.hpp
+++ b/src/Pipeline/SpirvShader.hpp
@@ -740,6 +740,7 @@
bool GroupNonUniformVote : 1;
bool GroupNonUniformBallot : 1;
bool GroupNonUniformShuffle : 1;
+ bool GroupNonUniformShuffleRelative : 1;
};
Capabilities const &getUsedCapabilities() const
diff --git a/src/Vulkan/VkPhysicalDevice.cpp b/src/Vulkan/VkPhysicalDevice.cpp
index ec022be..2103634 100644
--- a/src/Vulkan/VkPhysicalDevice.cpp
+++ b/src/Vulkan/VkPhysicalDevice.cpp
@@ -317,7 +317,8 @@
VK_SUBGROUP_FEATURE_BASIC_BIT |
VK_SUBGROUP_FEATURE_VOTE_BIT |
VK_SUBGROUP_FEATURE_BALLOT_BIT |
- VK_SUBGROUP_FEATURE_SHUFFLE_BIT;
+ VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
+ VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT;
properties->quadOperationsInAllStages = VK_FALSE;
}
