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

glsl: Implement VOTE for subgroup size potentially larger

Browse source
This commit is contained in:
ameerj 2021-05-30 00:53:26 -04:00
parent 770b754afd
commit e35ffbbeb0
5 changed files with 43 additions and 20 deletions
Download patch file Download diff file Expand all files Collapse all files

12
src/shader_recompiler/backend/glsl/emit_context.cpp
View file

@ -122,9 +122,11 @@ EmitContext::EmitContext(IR::Program& program, Bindings& bindings, const Profile
void EmitContext::SetupExtensions(std::string&) {
header += "#extension GL_ARB_separate_shader_objects : enable\n";
header += "#extension GL_ARB_sparse_texture2 : enable\n";
header += "#extension GL_EXT_texture_shadow_lod : enable\n";
// header += "#extension GL_ARB_texture_cube_map_array : enable\n";
if (stage != Stage::Compute) {
// TODO: track this usage
header += "#extension GL_ARB_sparse_texture2 : enable\n";
header += "#extension GL_EXT_texture_shadow_lod : enable\n";
}
if (info.uses_int64) {
header += "#extension GL_ARB_gpu_shader_int64 : enable\n";
}
@ -149,6 +151,10 @@ void EmitContext::SetupExtensions(std::string&) {
info.uses_subgroup_shuffles || info.uses_fswzadd) {
header += "#extension GL_ARB_shader_ballot : enable\n";
header += "#extension GL_ARB_shader_group_vote : enable\n";
header += "#extension GL_KHR_shader_subgroup_basic : enable\n";
if (!info.uses_int64) {
header += "#extension GL_ARB_gpu_shader_int64 : enable\n";
}
}
}

43
src/shader_recompiler/backend/glsl/emit_glsl_warp.cpp
View file

@ -42,31 +42,42 @@ void EmitLaneId([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& in
}
void EmitVoteAll(EmitContext& ctx, IR::Inst& inst, std::string_view pred) {
ctx.AddU1("{}=allInvocationsEqualARB({});", inst, pred);
// TODO:
// if (ctx.profile.warp_size_potentially_larger_than_guest) {
// }
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
ctx.AddU1("{}=allInvocationsEqualARB({});", inst, pred);
} else {
const auto active_mask{fmt::format("uvec2(ballotARB(true))[gl_SubgroupInvocationID]")};
const auto ballot{fmt::format("uvec2(ballotARB({}))[gl_SubgroupInvocationID]", pred)};
ctx.AddU1("{}=({}&{})=={};", inst, ballot, active_mask, active_mask);
}
}
void EmitVoteAny(EmitContext& ctx, IR::Inst& inst, std::string_view pred) {
ctx.AddU1("{}=anyInvocationARB({});", inst, pred);
// TODO:
// if (ctx.profile.warp_size_potentially_larger_than_guest) {
// }
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
ctx.AddU1("{}=anyInvocationARB({});", inst, pred);
} else {
const auto active_mask{fmt::format("uvec2(ballotARB(true))[gl_SubgroupInvocationID]")};
const auto ballot{fmt::format("uvec2(ballotARB({}))[gl_SubgroupInvocationID]", pred)};
ctx.AddU1("{}=({}&{})!=0u;", inst, ballot, active_mask, active_mask);
}
}
void EmitVoteEqual(EmitContext& ctx, IR::Inst& inst, std::string_view pred) {
ctx.AddU1("{}=allInvocationsEqualARB({});", inst, pred);
// TODO:
// if (ctx.profile.warp_size_potentially_larger_than_guest) {
// }
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
ctx.AddU1("{}=allInvocationsEqualARB({});", inst, pred);
} else {
const auto active_mask{fmt::format("uvec2(ballotARB(true))[gl_SubgroupInvocationID]")};
const auto ballot{fmt::format("uvec2(ballotARB({}))[gl_SubgroupInvocationID]", pred)};
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) {
ctx.AddU32("{}=uvec2(ballotARB({})).x;", inst, pred);
// TODO:
// if (ctx.profile.warp_size_potentially_larger_than_guest) {
// }
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
ctx.AddU32("{}=uvec2(ballotARB({})).x;", inst, pred);
} else {
ctx.AddU32("{}=uvec2(ballotARB({}))[gl_SubgroupInvocationID];", inst, pred);
}
}
void EmitSubgroupEqMask(EmitContext& ctx, IR::Inst& inst) {