N-archive/suyu
1
0
Fork 0
mirror of https://git.suyu.dev/suyu/suyu.git synced 2024-11-27 09:12:46 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions

emit_glsl_warp: Fix ballot related ops for 64-thread warp sizes

Browse source
This commit is contained in:
ameerj 2021-08-31 01:41:47 -04:00
parent a5d9dcf3d9
commit 5b45dfe971
1 changed files with 38 additions and 24 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

62
src/shader_recompiler/backend/glsl/emit_glsl_warp.cpp
View file

@ -11,6 +11,8 @@
namespace Shader::Backend::GLSL { namespace Shader::Backend::GLSL {
namespace { namespace {
constexpr char THREAD_ID[]{"gl_SubGroupInvocationARB"};
void SetInBoundsFlag(EmitContext& ctx, IR::Inst& inst) { void SetInBoundsFlag(EmitContext& ctx, IR::Inst& inst) {
IR::Inst* const in_bounds{inst.GetAssociatedPseudoOperation(IR::Opcode::GetInBoundsFromOp)}; IR::Inst* const in_bounds{inst.GetAssociatedPseudoOperation(IR::Opcode::GetInBoundsFromOp)};
if (!in_bounds) { if (!in_bounds) {
@ -43,69 +45,81 @@ void UseShuffleNv(EmitContext& ctx, IR::Inst& inst, std::string_view shfl_op,
ctx.AddU32("{}={}({},{},{},shfl_in_bounds);", inst, shfl_op, value, index, width); ctx.AddU32("{}={}({},{},{},shfl_in_bounds);", inst, shfl_op, value, index, width);
SetInBoundsFlag(ctx, inst); SetInBoundsFlag(ctx, inst);
} }
std::string_view BallotIndex(EmitContext& ctx) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ".x";
}
return "[gl_SubGroupInvocationARB>>5]";
}
std::string GetMask(EmitContext& ctx, std::string_view mask) {
const auto ballot_index{BallotIndex(ctx)};
return fmt::format("uint(uvec2({}){})", mask, ballot_index);
}
} // Anonymous namespace } // Anonymous namespace
void EmitLaneId(EmitContext& ctx, IR::Inst& inst) { void EmitLaneId(EmitContext& ctx, IR::Inst& inst) {
ctx.AddU32("{}=gl_SubGroupInvocationARB&31u;", inst); ctx.AddU32("{}={}&31u;", inst, THREAD_ID);
} }
void EmitVoteAll(EmitContext& ctx, IR::Inst& inst, std::string_view pred) { void EmitVoteAll(EmitContext& ctx, IR::Inst& inst, std::string_view pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) { if (!ctx.profile.warp_size_potentially_larger_than_guest) {
ctx.AddU1("{}=allInvocationsEqualARB({});", inst, pred); ctx.AddU1("{}=allInvocationsEqualARB({});", inst, pred);
} else { return;
const auto active_mask{fmt::format("uvec2(ballotARB(true))[gl_SubGroupInvocationARB]")};
const auto ballot{fmt::format("uvec2(ballotARB({}))[gl_SubGroupInvocationARB]", pred)};
ctx.AddU1("{}=({}&{})=={};", inst, ballot, active_mask, active_mask);
} }
const auto ballot_index{BallotIndex(ctx)};
const auto active_mask{fmt::format("uvec2(ballotARB(true)){}", ballot_index)};
const auto ballot{fmt::format("uvec2(ballotARB({})){}", pred, ballot_index)};
ctx.AddU1("{}=({}&{})=={};", inst, ballot, active_mask, active_mask);
} }
void EmitVoteAny(EmitContext& ctx, IR::Inst& inst, std::string_view pred) { void EmitVoteAny(EmitContext& ctx, IR::Inst& inst, std::string_view pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) { if (!ctx.profile.warp_size_potentially_larger_than_guest) {
ctx.AddU1("{}=anyInvocationARB({});", inst, pred); ctx.AddU1("{}=anyInvocationARB({});", inst, pred);
} else { return;
const auto active_mask{fmt::format("uvec2(ballotARB(true))[gl_SubGroupInvocationARB]")};
const auto ballot{fmt::format("uvec2(ballotARB({}))[gl_SubGroupInvocationARB]", pred)};
ctx.AddU1("{}=({}&{})!=0u;", inst, ballot, active_mask, active_mask);
} }
const auto ballot_index{BallotIndex(ctx)};
const auto active_mask{fmt::format("uvec2(ballotARB(true)){}", ballot_index)};
const auto ballot{fmt::format("uvec2(ballotARB({})){}", pred, ballot_index)};
ctx.AddU1("{}=({}&{})!=0u;", inst, ballot, active_mask, active_mask);
} }
void EmitVoteEqual(EmitContext& ctx, IR::Inst& inst, std::string_view pred) { void EmitVoteEqual(EmitContext& ctx, IR::Inst& inst, std::string_view pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) { if (!ctx.profile.warp_size_potentially_larger_than_guest) {
ctx.AddU1("{}=allInvocationsEqualARB({});", inst, pred); ctx.AddU1("{}=allInvocationsEqualARB({});", inst, pred);
} else { return;
const auto active_mask{fmt::format("uvec2(ballotARB(true))[gl_SubGroupInvocationARB]")};
const auto ballot{fmt::format("uvec2(ballotARB({}))[gl_SubGroupInvocationARB]", pred)};
const auto value{fmt::format("({}^{})", ballot, active_mask)};
ctx.AddU1("{}=({}==0)||({}=={});", inst, value, value, active_mask);
} }
const auto ballot_index{BallotIndex(ctx)};
const auto active_mask{fmt::format("uvec2(ballotARB(true)){}", ballot_index)};
const auto ballot{fmt::format("uvec2(ballotARB({})){}", pred, ballot_index)};
const auto value{fmt::format("({}^{})", ballot, active_mask)};
ctx.AddU1("{}=({}==0)||({}=={});", inst, value, value, active_mask);
} }
void EmitSubgroupBallot(EmitContext& ctx, IR::Inst& inst, std::string_view pred) { void EmitSubgroupBallot(EmitContext& ctx, IR::Inst& inst, std::string_view pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) { const auto ballot_index{BallotIndex(ctx)};
ctx.AddU32("{}=uvec2(ballotARB({})).x;", inst, pred); ctx.AddU32("{}=uvec2(ballotARB({})){};", inst, pred, ballot_index);
} else {
ctx.AddU32("{}=uvec2(ballotARB({}))[gl_SubGroupInvocationARB];", inst, pred);
}
} }
void EmitSubgroupEqMask(EmitContext& ctx, IR::Inst& inst) { void EmitSubgroupEqMask(EmitContext& ctx, IR::Inst& inst) {
ctx.AddU32("{}=uint(gl_SubGroupEqMaskARB.x);", inst); ctx.AddU32("{}={};", inst, GetMask(ctx, "gl_SubGroupEqMaskARB"));
} }
void EmitSubgroupLtMask(EmitContext& ctx, IR::Inst& inst) { void EmitSubgroupLtMask(EmitContext& ctx, IR::Inst& inst) {
ctx.AddU32("{}=uint(gl_SubGroupLtMaskARB.x);", inst); ctx.AddU32("{}={};", inst, GetMask(ctx, "gl_SubGroupLtMaskARB"));
} }
void EmitSubgroupLeMask(EmitContext& ctx, IR::Inst& inst) { void EmitSubgroupLeMask(EmitContext& ctx, IR::Inst& inst) {
ctx.AddU32("{}=uint(gl_SubGroupLeMaskARB.x);", inst); ctx.AddU32("{}={};", inst, GetMask(ctx, "gl_SubGroupLeMaskARB"));
} }
void EmitSubgroupGtMask(EmitContext& ctx, IR::Inst& inst) { void EmitSubgroupGtMask(EmitContext& ctx, IR::Inst& inst) {
ctx.AddU32("{}=uint(gl_SubGroupGtMaskARB.x);", inst); ctx.AddU32("{}={};", inst, GetMask(ctx, "gl_SubGroupGtMaskARB"));
} }
void EmitSubgroupGeMask(EmitContext& ctx, IR::Inst& inst) { void EmitSubgroupGeMask(EmitContext& ctx, IR::Inst& inst) {
ctx.AddU32("{}=uint(gl_SubGroupGeMaskARB.x);", inst); ctx.AddU32("{}={};", inst, GetMask(ctx, "gl_SubGroupGeMaskARB"));
} }
void EmitShuffleIndex(EmitContext& ctx, IR::Inst& inst, std::string_view value, void EmitShuffleIndex(EmitContext& ctx, IR::Inst& inst, std::string_view value,