Exverge/suyu
3
0
Fork 0
forked from suyu/suyu
Code Pull requests Activity

shader_recompiler: Fix shuffle partitioning for >64 invoc-per-subgroup GPUs

Browse source 
The existing implementation only supports 64 invoc-per-subgroup GPUs, and misbehaves on adreno when invocations need to be split into 4 emulated subgroups.
This commit is contained in:
Billy Laws 2022-12-30 13:45:00 +00:00
parent 6c812a0c84
commit 68ed60cee4
1 changed files with 28 additions and 30 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

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

@ -58,11 +58,10 @@ Id SelectValue(EmitContext& ctx, Id in_range, Id value, Id src_thread_id) {
ctx.OpGroupNonUniformShuffle(ctx.U32[1], SubgroupScope(ctx), 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);
Id AddPartitionBase(EmitContext& ctx, Id thread_id) {
const Id partition_idx{ctx.OpShiftRightLogical(ctx.U32[1], GetThreadId(ctx), ctx.Const(5u))};
const Id partition_base{ctx.OpShiftLeftLogical(ctx.U32[1], partition_idx, ctx.Const(5u))};
return ctx.OpIAdd(ctx.U32[1], thread_id, partition_base);
}
} // Anonymous namespace
@ -145,64 +144,63 @@ 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{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 thread_id{EmitLaneId(ctx)};
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)};
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)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
src_thread_id = AddPartitionBase(ctx, src_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{GetThreadId(ctx)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
clamp = GetUpperClamp(ctx, thread_id, clamp);
}
const Id thread_id{EmitLaneId(ctx)};
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)};
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)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
src_thread_id = AddPartitionBase(ctx, src_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{GetThreadId(ctx)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
clamp = GetUpperClamp(ctx, thread_id, clamp);
}
const Id thread_id{EmitLaneId(ctx)};
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)};
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)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
src_thread_id = AddPartitionBase(ctx, src_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{GetThreadId(ctx)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
clamp = GetUpperClamp(ctx, thread_id, clamp);
}
const Id thread_id{EmitLaneId(ctx)};
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)};
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)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
src_thread_id = AddPartitionBase(ctx, src_thread_id);
}
SetInBoundsFlag(inst, in_range);
return SelectValue(ctx, in_range, value, src_thread_id);
}