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

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

Browse source 
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
Download patch file Download diff file Expand all files Collapse all files

41
src/shader_recompiler/backend/spirv/emit_spirv_warp.cpp
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)};