Mirrors/yuzu
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

shader: Implement SHFL

Browse source
This commit is contained in:
ameerj 2021-03-25 11:31:37 -04:00
parent 49e87ea8ab
commit 32c5483beb
16 changed files with 284 additions and 69 deletions
Download patch file Download diff file Expand all files Collapse all files

3
src/shader_recompiler/backend/spirv/emit_context.cpp
View file

@ -274,7 +274,8 @@ void EmitContext::DefineInputs(const Info& info) {
if (info.uses_local_invocation_id) {
local_invocation_id = DefineInput(*this, U32[3], spv::BuiltIn::LocalInvocationId);
}
if (profile.warp_size_potentially_larger_than_guest && info.uses_subgroup_vote) {
if (info.uses_subgroup_invocation_id ||
(profile.warp_size_potentially_larger_than_guest && info.uses_subgroup_vote)) {
subgroup_local_invocation_id =
DefineInput(*this, U32[1], spv::BuiltIn::SubgroupLocalInvocationId);
}

6
src/shader_recompiler/backend/spirv/emit_spirv.cpp
View file

@ -224,7 +224,7 @@ void SetupCapabilities(const Profile& profile, const Info& info, EmitContext& ct
ctx.AddExtension("SPV_KHR_shader_draw_parameters");
ctx.AddCapability(spv::Capability::DrawParameters);
}
if (info.uses_subgroup_vote && profile.support_vote) {
if ((info.uses_subgroup_vote || info.uses_subgroup_invocation_id) && profile.support_vote) {
ctx.AddExtension("SPV_KHR_shader_ballot");
ctx.AddCapability(spv::Capability::SubgroupBallotKHR);
if (!profile.warp_size_potentially_larger_than_guest) {
@ -315,4 +315,8 @@ void EmitGetSparseFromOp(EmitContext&) {
throw LogicError("Unreachable instruction");
}
void EmitGetInBoundsFromOp(EmitContext&) {
throw LogicError("Unreachable instruction");
}
} // namespace Shader::Backend::SPIRV

9
src/shader_recompiler/backend/spirv/emit_spirv.h
View file

@ -158,6 +158,7 @@ void EmitGetSignFromOp(EmitContext& ctx);
void EmitGetCarryFromOp(EmitContext& ctx);
void EmitGetOverflowFromOp(EmitContext& ctx);
void EmitGetSparseFromOp(EmitContext& ctx);
void EmitGetInBoundsFromOp(EmitContext& ctx);
Id EmitFPAbs16(EmitContext& ctx, Id value);
Id EmitFPAbs32(EmitContext& ctx, Id value);
Id EmitFPAbs64(EmitContext& ctx, Id value);
@ -355,5 +356,13 @@ Id EmitVoteAll(EmitContext& ctx, Id pred);
Id EmitVoteAny(EmitContext& ctx, Id pred);
Id EmitVoteEqual(EmitContext& ctx, Id pred);
Id EmitSubgroupBallot(EmitContext& ctx, Id pred);
Id EmitShuffleIndex(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
Id segmentation_mask);
Id EmitShuffleUp(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
Id segmentation_mask);
Id EmitShuffleDown(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
Id segmentation_mask);
Id EmitShuffleButterfly(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
Id segmentation_mask);
} // namespace Shader::Backend::SPIRV

58
src/shader_recompiler/backend/spirv/emit_spirv_vote.cpp
View file

@ -1,58 +0,0 @@
// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include "shader_recompiler/backend/spirv/emit_spirv.h"
namespace Shader::Backend::SPIRV {
namespace {
Id LargeWarpBallot(EmitContext& ctx, Id ballot) {
const Id shift{ctx.Constant(ctx.U32[1], 5)};
const Id local_index{ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id)};
return ctx.OpVectorExtractDynamic(ctx.U32[1], ballot, local_index);
}
} // Anonymous namespace
Id EmitVoteAll(EmitContext& ctx, Id pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpSubgroupAllKHR(ctx.U1, pred);
}
const Id mask_ballot{ctx.OpSubgroupBallotKHR(ctx.U32[4], ctx.true_value)};
const Id active_mask{LargeWarpBallot(ctx, mask_ballot)};
const Id ballot{LargeWarpBallot(ctx, ctx.OpSubgroupBallotKHR(ctx.U32[4], pred))};
const Id lhs{ctx.OpBitwiseAnd(ctx.U32[1], ballot, active_mask)};
return ctx.OpIEqual(ctx.U1, lhs, active_mask);
}
Id EmitVoteAny(EmitContext& ctx, Id pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpSubgroupAnyKHR(ctx.U1, pred);
}
const Id mask_ballot{ctx.OpSubgroupBallotKHR(ctx.U32[4], ctx.true_value)};
const Id active_mask{LargeWarpBallot(ctx, mask_ballot)};
const Id ballot{LargeWarpBallot(ctx, ctx.OpSubgroupBallotKHR(ctx.U32[4], pred))};
const Id lhs{ctx.OpBitwiseAnd(ctx.U32[1], ballot, active_mask)};
return ctx.OpINotEqual(ctx.U1, lhs, ctx.u32_zero_value);
}
Id EmitVoteEqual(EmitContext& ctx, Id pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpSubgroupAllEqualKHR(ctx.U1, pred);
}
const Id mask_ballot{ctx.OpSubgroupBallotKHR(ctx.U32[4], ctx.true_value)};
const Id active_mask{LargeWarpBallot(ctx, mask_ballot)};
const Id ballot{LargeWarpBallot(ctx, ctx.OpSubgroupBallotKHR(ctx.U32[4], pred))};
const Id lhs{ctx.OpBitwiseXor(ctx.U32[1], ballot, active_mask)};
return ctx.OpLogicalOr(ctx.U1, ctx.OpIEqual(ctx.U1, lhs, ctx.u32_zero_value),
ctx.OpIEqual(ctx.U1, lhs, active_mask));
}
Id EmitSubgroupBallot(EmitContext& ctx, Id pred) {
const Id ballot{ctx.OpSubgroupBallotKHR(ctx.U32[4], pred)};
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpCompositeExtract(ctx.U32[1], ballot, 0U);
}
return LargeWarpBallot(ctx, ballot);
}
} // namespace Shader::Backend::SPIRV

135
src/shader_recompiler/backend/spirv/emit_spirv_warp.cpp Normal file
View file

@ -0,0 +1,135 @@
// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include "shader_recompiler/backend/spirv/emit_spirv.h"
namespace Shader::Backend::SPIRV {
namespace {
Id LargeWarpBallot(EmitContext& ctx, Id ballot) {
const Id shift{ctx.Constant(ctx.U32[1], 5)};
const Id local_index{ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id)};
return ctx.OpVectorExtractDynamic(ctx.U32[1], ballot, local_index);
}
void SetInBoundsFlag(IR::Inst* inst, Id result) {
IR::Inst* const in_bounds{inst->GetAssociatedPseudoOperation(IR::Opcode::GetInBoundsFromOp)};
if (!in_bounds) {
return;
}
in_bounds->SetDefinition(result);
in_bounds->Invalidate();
}
Id ComputeMinThreadId(EmitContext& ctx, Id thread_id, Id segmentation_mask) {
return ctx.OpBitwiseAnd(ctx.U32[1], thread_id, segmentation_mask);
}
Id ComputeMaxThreadId(EmitContext& ctx, Id min_thread_id, Id clamp, Id not_seg_mask) {
return ctx.OpBitwiseOr(ctx.U32[1], min_thread_id,
ctx.OpBitwiseAnd(ctx.U32[1], clamp, not_seg_mask));
}
Id GetMaxThreadId(EmitContext& ctx, Id thread_id, Id clamp, Id segmentation_mask) {
const Id not_seg_mask{ctx.OpNot(ctx.U32[1], segmentation_mask)};
const Id min_thread_id{ComputeMinThreadId(ctx, thread_id, segmentation_mask)};
return ComputeMaxThreadId(ctx, min_thread_id, clamp, not_seg_mask);
}
Id SelectValue(EmitContext& ctx, Id in_range, Id value, Id src_thread_id) {
return ctx.OpSelect(ctx.U32[1], in_range,
ctx.OpSubgroupReadInvocationKHR(ctx.U32[1], value, src_thread_id), value);
}
} // Anonymous namespace
Id EmitVoteAll(EmitContext& ctx, Id pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpSubgroupAllKHR(ctx.U1, pred);
}
const Id mask_ballot{ctx.OpSubgroupBallotKHR(ctx.U32[4], ctx.true_value)};
const Id active_mask{LargeWarpBallot(ctx, mask_ballot)};
const Id ballot{LargeWarpBallot(ctx, ctx.OpSubgroupBallotKHR(ctx.U32[4], pred))};
const Id lhs{ctx.OpBitwiseAnd(ctx.U32[1], ballot, active_mask)};
return ctx.OpIEqual(ctx.U1, lhs, active_mask);
}
Id EmitVoteAny(EmitContext& ctx, Id pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpSubgroupAnyKHR(ctx.U1, pred);
}
const Id mask_ballot{ctx.OpSubgroupBallotKHR(ctx.U32[4], ctx.true_value)};
const Id active_mask{LargeWarpBallot(ctx, mask_ballot)};
const Id ballot{LargeWarpBallot(ctx, ctx.OpSubgroupBallotKHR(ctx.U32[4], pred))};
const Id lhs{ctx.OpBitwiseAnd(ctx.U32[1], ballot, active_mask)};
return ctx.OpINotEqual(ctx.U1, lhs, ctx.u32_zero_value);
}
Id EmitVoteEqual(EmitContext& ctx, Id pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpSubgroupAllEqualKHR(ctx.U1, pred);
}
const Id mask_ballot{ctx.OpSubgroupBallotKHR(ctx.U32[4], ctx.true_value)};
const Id active_mask{LargeWarpBallot(ctx, mask_ballot)};
const Id ballot{LargeWarpBallot(ctx, ctx.OpSubgroupBallotKHR(ctx.U32[4], pred))};
const Id lhs{ctx.OpBitwiseXor(ctx.U32[1], ballot, active_mask)};
return ctx.OpLogicalOr(ctx.U1, ctx.OpIEqual(ctx.U1, lhs, ctx.u32_zero_value),
ctx.OpIEqual(ctx.U1, lhs, active_mask));
}
Id EmitSubgroupBallot(EmitContext& ctx, Id pred) {
const Id ballot{ctx.OpSubgroupBallotKHR(ctx.U32[4], pred)};
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpCompositeExtract(ctx.U32[1], ballot, 0U);
}
return LargeWarpBallot(ctx, ballot);
}
Id EmitShuffleIndex(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
Id segmentation_mask) {
const Id not_seg_mask{ctx.OpNot(ctx.U32[1], segmentation_mask)};
const Id thread_id{ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id)};
const Id min_thread_id{ComputeMinThreadId(ctx, thread_id, segmentation_mask)};
const Id max_thread_id{ComputeMaxThreadId(ctx, min_thread_id, clamp, not_seg_mask)};
const Id lhs{ctx.OpBitwiseAnd(ctx.U32[1], index, not_seg_mask)};
const Id src_thread_id{ctx.OpBitwiseOr(ctx.U32[1], lhs, min_thread_id)};
const Id in_range{ctx.OpSLessThanEqual(ctx.U1, src_thread_id, max_thread_id)};
SetInBoundsFlag(inst, in_range);
return SelectValue(ctx, in_range, value, src_thread_id);
}
Id EmitShuffleUp(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
Id segmentation_mask) {
const Id thread_id{ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id)};
const Id max_thread_id{GetMaxThreadId(ctx, thread_id, clamp, segmentation_mask)};
const Id src_thread_id{ctx.OpISub(ctx.U32[1], thread_id, index)};
const Id in_range{ctx.OpSGreaterThanEqual(ctx.U1, src_thread_id, max_thread_id)};
SetInBoundsFlag(inst, in_range);
return SelectValue(ctx, in_range, value, src_thread_id);
}
Id EmitShuffleDown(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
Id segmentation_mask) {
const Id thread_id{ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id)};
const Id max_thread_id{GetMaxThreadId(ctx, thread_id, clamp, segmentation_mask)};
const Id src_thread_id{ctx.OpIAdd(ctx.U32[1], thread_id, index)};
const Id in_range{ctx.OpSLessThanEqual(ctx.U1, src_thread_id, max_thread_id)};
SetInBoundsFlag(inst, in_range);
return SelectValue(ctx, in_range, value, src_thread_id);
}
Id EmitShuffleButterfly(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
Id segmentation_mask) {
const Id thread_id{ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id)};
const Id max_thread_id{GetMaxThreadId(ctx, thread_id, clamp, segmentation_mask)};
const Id src_thread_id{ctx.OpBitwiseXor(ctx.U32[1], thread_id, index)};
const Id in_range{ctx.OpSLessThanEqual(ctx.U1, src_thread_id, max_thread_id)};
SetInBoundsFlag(inst, in_range);
return SelectValue(ctx, in_range, value, src_thread_id);
}
} // namespace Shader::Backend::SPIRV