1
0
Fork 0
forked from suyu/suyu

Merge pull request #6948 from ameerj/amd-warp-fix

shaders: Fix warp instructions on 64-thread warp devices
This commit is contained in:
Morph 2021-09-12 13:53:29 -04:00 committed by GitHub
commit 9248442bb2
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 109 additions and 54 deletions

View file

@ -11,6 +11,8 @@
namespace Shader::Backend::GLSL {
namespace {
constexpr char THREAD_ID[]{"gl_SubGroupInvocationARB"};
void SetInBoundsFlag(EmitContext& ctx, IR::Inst& inst) {
IR::Inst* const in_bounds{inst.GetAssociatedPseudoOperation(IR::Opcode::GetInBoundsFromOp)};
if (!in_bounds) {
@ -43,84 +45,100 @@ void UseShuffleNv(EmitContext& ctx, IR::Inst& inst, std::string_view shfl_op,
ctx.AddU32("{}={}({},{},{},shfl_in_bounds);", inst, shfl_op, value, index, width);
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
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) {
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_SubGroupInvocationARB]")};
const auto ballot{fmt::format("uvec2(ballotARB({}))[gl_SubGroupInvocationARB]", pred)};
ctx.AddU1("{}=({}&{})=={};", inst, ballot, active_mask, active_mask);
return;
}
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) {
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_SubGroupInvocationARB]")};
const auto ballot{fmt::format("uvec2(ballotARB({}))[gl_SubGroupInvocationARB]", pred)};
ctx.AddU1("{}=({}&{})!=0u;", inst, ballot, active_mask, active_mask);
return;
}
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) {
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_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);
return;
}
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) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
ctx.AddU32("{}=uvec2(ballotARB({})).x;", inst, pred);
} else {
ctx.AddU32("{}=uvec2(ballotARB({}))[gl_SubGroupInvocationARB];", inst, pred);
}
const auto ballot_index{BallotIndex(ctx)};
ctx.AddU32("{}=uvec2(ballotARB({})){};", inst, pred, ballot_index);
}
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) {
ctx.AddU32("{}=uint(gl_SubGroupLtMaskARB.x);", inst);
ctx.AddU32("{}={};", inst, GetMask(ctx, "gl_SubGroupLtMaskARB"));
}
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) {
ctx.AddU32("{}=uint(gl_SubGroupGtMaskARB.x);", inst);
ctx.AddU32("{}={};", inst, GetMask(ctx, "gl_SubGroupGtMaskARB"));
}
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,
std::string_view index, std::string_view clamp,
std::string_view segmentation_mask) {
std::string_view index, std::string_view clamp, std::string_view seg_mask) {
if (ctx.profile.support_gl_warp_intrinsics) {
UseShuffleNv(ctx, inst, "shuffleNV", value, index, clamp, segmentation_mask);
UseShuffleNv(ctx, inst, "shuffleNV", value, index, clamp, seg_mask);
return;
}
const auto not_seg_mask{fmt::format("(~{})", segmentation_mask)};
const auto thread_id{"gl_SubGroupInvocationARB"};
const auto min_thread_id{ComputeMinThreadId(thread_id, segmentation_mask)};
const auto max_thread_id{ComputeMaxThreadId(min_thread_id, clamp, not_seg_mask)};
const bool big_warp{ctx.profile.warp_size_potentially_larger_than_guest};
const auto is_upper_partition{"int(gl_SubGroupInvocationARB)>=32"};
const auto upper_index{fmt::format("{}?{}+32:{}", is_upper_partition, index, index)};
const auto upper_clamp{fmt::format("{}?{}+32:{}", is_upper_partition, clamp, clamp)};
const auto lhs{fmt::format("({}&{})", index, not_seg_mask)};
const auto not_seg_mask{fmt::format("(~{})", seg_mask)};
const auto min_thread_id{ComputeMinThreadId(THREAD_ID, seg_mask)};
const auto max_thread_id{
ComputeMaxThreadId(min_thread_id, big_warp ? upper_clamp : clamp, not_seg_mask)};
const auto lhs{fmt::format("({}&{})", big_warp ? upper_index : index, not_seg_mask)};
const auto src_thread_id{fmt::format("({})|({})", lhs, min_thread_id)};
ctx.Add("shfl_in_bounds=int({})<=int({});", src_thread_id, max_thread_id);
SetInBoundsFlag(ctx, inst);
@ -128,29 +146,34 @@ void EmitShuffleIndex(EmitContext& ctx, IR::Inst& inst, std::string_view value,
}
void EmitShuffleUp(EmitContext& ctx, IR::Inst& inst, std::string_view value, std::string_view index,
std::string_view clamp, std::string_view segmentation_mask) {
std::string_view clamp, std::string_view seg_mask) {
if (ctx.profile.support_gl_warp_intrinsics) {
UseShuffleNv(ctx, inst, "shuffleUpNV", value, index, clamp, segmentation_mask);
UseShuffleNv(ctx, inst, "shuffleUpNV", value, index, clamp, seg_mask);
return;
}
const auto thread_id{"gl_SubGroupInvocationARB"};
const auto max_thread_id{GetMaxThreadId(thread_id, clamp, segmentation_mask)};
const auto src_thread_id{fmt::format("({}-{})", thread_id, index)};
const bool big_warp{ctx.profile.warp_size_potentially_larger_than_guest};
const auto is_upper_partition{"int(gl_SubGroupInvocationARB)>=32"};
const auto upper_clamp{fmt::format("{}?{}+32:{}", is_upper_partition, clamp, clamp)};
const auto max_thread_id{GetMaxThreadId(THREAD_ID, big_warp ? upper_clamp : clamp, seg_mask)};
const auto src_thread_id{fmt::format("({}-{})", THREAD_ID, index)};
ctx.Add("shfl_in_bounds=int({})>=int({});", src_thread_id, max_thread_id);
SetInBoundsFlag(ctx, inst);
ctx.AddU32("{}=shfl_in_bounds?readInvocationARB({},{}):{};", inst, value, src_thread_id, value);
}
void EmitShuffleDown(EmitContext& ctx, IR::Inst& inst, std::string_view value,
std::string_view index, std::string_view clamp,
std::string_view segmentation_mask) {
std::string_view index, std::string_view clamp, std::string_view seg_mask) {
if (ctx.profile.support_gl_warp_intrinsics) {
UseShuffleNv(ctx, inst, "shuffleDownNV", value, index, clamp, segmentation_mask);
UseShuffleNv(ctx, inst, "shuffleDownNV", value, index, clamp, seg_mask);
return;
}
const auto thread_id{"gl_SubGroupInvocationARB"};
const auto max_thread_id{GetMaxThreadId(thread_id, clamp, segmentation_mask)};
const auto src_thread_id{fmt::format("({}+{})", thread_id, index)};
const bool big_warp{ctx.profile.warp_size_potentially_larger_than_guest};
const auto is_upper_partition{"int(gl_SubGroupInvocationARB)>=32"};
const auto upper_clamp{fmt::format("{}?{}+32:{}", is_upper_partition, clamp, clamp)};
const auto max_thread_id{GetMaxThreadId(THREAD_ID, big_warp ? upper_clamp : clamp, seg_mask)};
const auto src_thread_id{fmt::format("({}+{})", THREAD_ID, index)};
ctx.Add("shfl_in_bounds=int({})<=int({});", src_thread_id, max_thread_id);
SetInBoundsFlag(ctx, inst);
ctx.AddU32("{}=shfl_in_bounds?readInvocationARB({},{}):{};", inst, value, src_thread_id, value);
@ -158,14 +181,17 @@ void EmitShuffleDown(EmitContext& ctx, IR::Inst& inst, std::string_view value,
void EmitShuffleButterfly(EmitContext& ctx, IR::Inst& inst, std::string_view value,
std::string_view index, std::string_view clamp,
std::string_view segmentation_mask) {
std::string_view seg_mask) {
if (ctx.profile.support_gl_warp_intrinsics) {
UseShuffleNv(ctx, inst, "shuffleXorNV", value, index, clamp, segmentation_mask);
UseShuffleNv(ctx, inst, "shuffleXorNV", value, index, clamp, seg_mask);
return;
}
const auto thread_id{"gl_SubGroupInvocationARB"};
const auto max_thread_id{GetMaxThreadId(thread_id, clamp, segmentation_mask)};
const auto src_thread_id{fmt::format("({}^{})", thread_id, index)};
const bool big_warp{ctx.profile.warp_size_potentially_larger_than_guest};
const auto is_upper_partition{"int(gl_SubGroupInvocationARB)>=32"};
const auto upper_clamp{fmt::format("{}?{}+32:{}", is_upper_partition, clamp, clamp)};
const auto max_thread_id{GetMaxThreadId(THREAD_ID, big_warp ? upper_clamp : clamp, seg_mask)};
const auto src_thread_id{fmt::format("({}^{})", THREAD_ID, index)};
ctx.Add("shfl_in_bounds=int({})<=int({});", src_thread_id, max_thread_id);
SetInBoundsFlag(ctx, inst);
ctx.AddU32("{}=shfl_in_bounds?readInvocationARB({},{}):{};", inst, value, src_thread_id, value);

View file

@ -7,8 +7,13 @@
namespace Shader::Backend::SPIRV {
namespace {
Id GetThreadId(EmitContext& ctx) {
return ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id);
}
Id WarpExtract(EmitContext& ctx, Id value) {
const Id local_index{ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id)};
const Id thread_id{GetThreadId(ctx)};
const Id local_index{ctx.OpShiftRightArithmetic(ctx.U32[1], thread_id, ctx.Const(5U))};
return ctx.OpVectorExtractDynamic(ctx.U32[1], value, local_index);
}
@ -48,10 +53,17 @@ 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);
}
Id GetUpperClamp(EmitContext& ctx, Id invocation_id, Id clamp) {
const Id thirty_two{ctx.Const(32u)};
const Id is_upper_partition{ctx.OpSGreaterThanEqual(ctx.U1, invocation_id, thirty_two)};
const Id upper_clamp{ctx.OpIAdd(ctx.U32[1], thirty_two, clamp)};
return ctx.OpSelect(ctx.U32[1], is_upper_partition, upper_clamp, clamp);
}
} // Anonymous namespace
Id EmitLaneId(EmitContext& ctx) {
const Id id{ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id)};
const Id id{GetThreadId(ctx)};
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return id;
}
@ -123,7 +135,15 @@ Id EmitSubgroupGeMask(EmitContext& ctx) {
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 thread_id{GetThreadId(ctx)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
const Id thirty_two{ctx.Const(32u)};
const Id is_upper_partition{ctx.OpSGreaterThanEqual(ctx.U1, thread_id, thirty_two)};
const Id upper_index{ctx.OpIAdd(ctx.U32[1], thirty_two, index)};
const Id upper_clamp{ctx.OpIAdd(ctx.U32[1], thirty_two, clamp)};
index = ctx.OpSelect(ctx.U32[1], is_upper_partition, upper_index, index);
clamp = ctx.OpSelect(ctx.U32[1], is_upper_partition, upper_clamp, clamp);
}
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)};
@ -137,7 +157,10 @@ Id EmitShuffleIndex(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id cla
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 thread_id{GetThreadId(ctx)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
clamp = GetUpperClamp(ctx, thread_id, clamp);
}
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)};
@ -148,7 +171,10 @@ Id EmitShuffleUp(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
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 thread_id{GetThreadId(ctx)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
clamp = GetUpperClamp(ctx, thread_id, clamp);
}
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)};
@ -159,7 +185,10 @@ Id EmitShuffleDown(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clam
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 thread_id{GetThreadId(ctx)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
clamp = GetUpperClamp(ctx, thread_id, clamp);
}
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)};