shader: Implement VOTE

src/shader_recompiler/CMakeLists.txt

@@ -15,6 +15,7 @@ add_library(shader_recompiler STATIC
     backend/spirv/emit_spirv_memory.cpp
     backend/spirv/emit_spirv_select.cpp
     backend/spirv/emit_spirv_undefined.cpp
+    backend/spirv/emit_spirv_vote.cpp
     environment.h
     exception.h
     file_environment.cpp
@@ -122,6 +123,7 @@ add_library(shader_recompiler STATIC
     frontend/maxwell/translate/impl/select_source_with_predicate.cpp
     frontend/maxwell/translate/impl/texture_fetch.cpp
     frontend/maxwell/translate/impl/texture_fetch_swizzled.cpp
+    frontend/maxwell/translate/impl/vote.cpp
     frontend/maxwell/translate/translate.cpp
     frontend/maxwell/translate/translate.h
     ir_opt/collect_shader_info_pass.cpp

src/shader_recompiler/backend/spirv/emit_context.cpp

@@ -259,6 +259,10 @@ void EmitContext::DefineInputs(const Info& info, Stage stage) {
     if (info.uses_local_invocation_id) {
         local_invocation_id = DefineInput(*this, U32[3], spv::BuiltIn::LocalInvocationId);
     }
+    if (profile.warp_size_potentially_larger_than_guest && info.uses_subgroup_vote) {
+        subgroup_local_invocation_id =
+            DefineInput(*this, U32[1], spv::BuiltIn::SubgroupLocalInvocationId);
+    }
     if (info.loads_position) {
         const bool is_fragment{stage != Stage::Fragment};
         const spv::BuiltIn built_in{is_fragment ? spv::BuiltIn::Position : spv::BuiltIn::FragCoord};

src/shader_recompiler/backend/spirv/emit_context.h

@@ -82,6 +82,7 @@ public:
 
     Id workgroup_id{};
     Id local_invocation_id{};
+    Id subgroup_local_invocation_id{};
     Id instance_id{};
     Id instance_index{};
     Id base_instance{};
@@ -96,7 +97,7 @@ public:
     std::array<Id, 32> output_generics{};
 
     std::array<Id, 8> frag_color{};
-    Id frag_depth {};
+    Id frag_depth{};
 
     std::vector<Id> interfaces;
 

src/shader_recompiler/backend/spirv/emit_spirv.cpp

@@ -224,6 +224,15 @@ void SetupCapabilities(const Profile& profile, const Info& info, EmitContext& ct
         ctx.AddExtension("SPV_KHR_shader_draw_parameters");
         ctx.AddCapability(spv::Capability::DrawParameters);
     }
+    if (info.uses_subgroup_vote && profile.support_vote) {
+        ctx.AddExtension("SPV_KHR_shader_ballot");
+        ctx.AddCapability(spv::Capability::SubgroupBallotKHR);
+        if (!profile.warp_size_potentially_larger_than_guest) {
+            // vote ops are only used when not taking the long path
+            ctx.AddExtension("SPV_KHR_subgroup_vote");
+            ctx.AddCapability(spv::Capability::SubgroupVoteKHR);
+        }
+    }
     // TODO: Track this usage
     ctx.AddCapability(spv::Capability::ImageGatherExtended);
 }

src/shader_recompiler/backend/spirv/emit_spirv.h

@@ -346,5 +346,9 @@ Id EmitImageSampleDrefImplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Va
                                   Id coords, Id dref, Id bias_lc, Id offset);
 Id EmitImageSampleDrefExplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index,
                                   Id coords, Id dref, Id lod_lc, Id offset);
+Id EmitVoteAll(EmitContext& ctx, Id pred);
+Id EmitVoteAny(EmitContext& ctx, Id pred);
+Id EmitVoteEqual(EmitContext& ctx, Id pred);
+Id EmitSubgroupBallot(EmitContext& ctx, Id pred);
 
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/backend/spirv/emit_spirv_vote.cpp

@@ -0,0 +1,58 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "shader_recompiler/backend/spirv/emit_spirv.h"
+
+namespace Shader::Backend::SPIRV {
+namespace {
+Id LargeWarpBallot(EmitContext& ctx, Id ballot) {
+    const Id shift{ctx.Constant(ctx.U32[1], 5)};
+    const Id local_index{ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id)};
+    return ctx.OpVectorExtractDynamic(ctx.U32[1], ballot, local_index);
+}
+} // Anonymous namespace
+
+Id EmitVoteAll(EmitContext& ctx, Id pred) {
+    if (!ctx.profile.warp_size_potentially_larger_than_guest) {
+        return ctx.OpSubgroupAllKHR(ctx.U1, pred);
+    }
+    const Id mask_ballot{ctx.OpSubgroupBallotKHR(ctx.U32[4], ctx.true_value)};
+    const Id active_mask{LargeWarpBallot(ctx, mask_ballot)};
+    const Id ballot{LargeWarpBallot(ctx, ctx.OpSubgroupBallotKHR(ctx.U32[4], pred))};
+    const Id lhs{ctx.OpBitwiseAnd(ctx.U32[1], ballot, active_mask)};
+    return ctx.OpIEqual(ctx.U1, lhs, active_mask);
+}
+
+Id EmitVoteAny(EmitContext& ctx, Id pred) {
+    if (!ctx.profile.warp_size_potentially_larger_than_guest) {
+        return ctx.OpSubgroupAnyKHR(ctx.U1, pred);
+    }
+    const Id mask_ballot{ctx.OpSubgroupBallotKHR(ctx.U32[4], ctx.true_value)};
+    const Id active_mask{LargeWarpBallot(ctx, mask_ballot)};
+    const Id ballot{LargeWarpBallot(ctx, ctx.OpSubgroupBallotKHR(ctx.U32[4], pred))};
+    const Id lhs{ctx.OpBitwiseAnd(ctx.U32[1], ballot, active_mask)};
+    return ctx.OpINotEqual(ctx.U1, lhs, ctx.u32_zero_value);
+}
+
+Id EmitVoteEqual(EmitContext& ctx, Id pred) {
+    if (!ctx.profile.warp_size_potentially_larger_than_guest) {
+        return ctx.OpSubgroupAllEqualKHR(ctx.U1, pred);
+    }
+    const Id mask_ballot{ctx.OpSubgroupBallotKHR(ctx.U32[4], ctx.true_value)};
+    const Id active_mask{LargeWarpBallot(ctx, mask_ballot)};
+    const Id ballot{LargeWarpBallot(ctx, ctx.OpSubgroupBallotKHR(ctx.U32[4], pred))};
+    const Id lhs{ctx.OpBitwiseXor(ctx.U32[1], ballot, active_mask)};
+    return ctx.OpLogicalOr(ctx.U1, ctx.OpIEqual(ctx.U1, lhs, ctx.u32_zero_value),
+                           ctx.OpIEqual(ctx.U1, lhs, active_mask));
+}
+
+Id EmitSubgroupBallot(EmitContext& ctx, Id pred) {
+    const Id ballot{ctx.OpSubgroupBallotKHR(ctx.U32[4], pred)};
+    if (!ctx.profile.warp_size_potentially_larger_than_guest) {
+        return ctx.OpCompositeExtract(ctx.U32[1], ballot, 0U);
+    }
+    return LargeWarpBallot(ctx, ballot);
+}
+
+} // namespace Shader::Backend::SPIRV

src/shader_recompiler/frontend/ir/ir_emitter.cpp

@@ -1444,4 +1444,20 @@ F32 IREmitter::ImageSampleDrefExplicitLod(const Value& handle, const Value& coor
     return Inst<F32>(op, Flags{info}, handle, coords, dref, lod_lc, offset);
 }
 
+U1 IREmitter::VoteAll(const U1& value) {
+    return Inst<U1>(Opcode::VoteAll, value);
+}
+
+U1 IREmitter::VoteAny(const U1& value) {
+    return Inst<U1>(Opcode::VoteAny, value);
+}
+
+U1 IREmitter::VoteEqual(const U1& value) {
+    return Inst<U1>(Opcode::VoteEqual, value);
+}
+
+U32 IREmitter::SubgroupBallot(const U1& value) {
+    return Inst<U32>(Opcode::SubgroupBallot, value);
+}
+
 } // namespace Shader::IR

src/shader_recompiler/frontend/ir/ir_emitter.h

@@ -234,6 +234,11 @@ public:
                                                  const Value& offset, const F32& lod_clamp,
                                                  TextureInstInfo info);
 
+    [[nodiscard]] U1 VoteAll(const U1& value);
+    [[nodiscard]] U1 VoteAny(const U1& value);
+    [[nodiscard]] U1 VoteEqual(const U1& value);
+    [[nodiscard]] U32 SubgroupBallot(const U1& value);
+
 private:
     IR::Block::iterator insertion_point;
 

src/shader_recompiler/frontend/ir/opcodes.inc

@@ -355,3 +355,9 @@ OPCODE(ImageSampleImplicitLod,                              F32x4,          U32,
 OPCODE(ImageSampleExplicitLod,                              F32x4,          U32,            Opaque,         Opaque,         Opaque,                         )
 OPCODE(ImageSampleDrefImplicitLod,                          F32,            U32,            Opaque,         F32,            Opaque,         Opaque,         )
 OPCODE(ImageSampleDrefExplicitLod,                          F32,            U32,            Opaque,         F32,            Opaque,         Opaque,         )
+
+// Vote operations
+OPCODE(VoteAll,                                             U1,             U1,                                                                             )
+OPCODE(VoteAny,                                             U1,             U1,                                                                             )
+OPCODE(VoteEqual,                                           U1,             U1,                                                                             )
+OPCODE(SubgroupBallot,                                      U32,            U1,                                                                             )

src/shader_recompiler/frontend/maxwell/translate/impl/not_implemented.cpp

@@ -417,10 +417,6 @@ void TranslatorVisitor::VMNMX(u64) {
     ThrowNotImplemented(Opcode::VMNMX);
 }
 
-void TranslatorVisitor::VOTE(u64) {
-    ThrowNotImplemented(Opcode::VOTE);
-}
-
 void TranslatorVisitor::VOTE_vtg(u64) {
     ThrowNotImplemented(Opcode::VOTE_vtg);
 }

src/shader_recompiler/frontend/maxwell/translate/impl/vote.cpp

@@ -0,0 +1,52 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <optional>
+
+#include "common/bit_field.h"
+#include "common/common_types.h"
+#include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
+
+namespace Shader::Maxwell {
+namespace {
+enum class VoteOp : u64 {
+    ALL,
+    ANY,
+    EQ,
+};
+
+[[nodiscard]] IR::U1 VoteOperation(IR::IREmitter& ir, const IR::U1& pred, VoteOp vote_op) {
+    switch (vote_op) {
+    case VoteOp::ALL:
+        return ir.VoteAll(pred);
+    case VoteOp::ANY:
+        return ir.VoteAny(pred);
+    case VoteOp::EQ:
+        return ir.VoteEqual(pred);
+    default:
+        throw NotImplementedException("Invalid VOTE op {}", vote_op);
+    }
+}
+
+void Vote(TranslatorVisitor& v, u64 insn) {
+    union {
+        u64 insn;
+        BitField<0, 8, IR::Reg> dest_reg;
+        BitField<39, 3, IR::Pred> pred_a;
+        BitField<42, 1, u64> neg_pred_a;
+        BitField<45, 3, IR::Pred> pred_b;
+        BitField<48, 2, VoteOp> vote_op;
+    } const vote{insn};
+
+    const IR::U1 vote_pred{v.ir.GetPred(vote.pred_a, vote.neg_pred_a != 0)};
+    v.ir.SetPred(vote.pred_b, VoteOperation(v.ir, vote_pred, vote.vote_op));
+    v.X(vote.dest_reg, v.ir.SubgroupBallot(vote_pred));
+}
+} // Anonymous namespace
+
+void TranslatorVisitor::VOTE(u64 insn) {
+    Vote(*this, insn);
+}
+
+} // namespace Shader::Maxwell

src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp

@@ -359,6 +359,12 @@ void VisitUsages(Info& info, IR::Inst& inst) {
             inst.GetAssociatedPseudoOperation(IR::Opcode::GetSparseFromOp) != nullptr;
         break;
     }
+    case IR::Opcode::VoteAll:
+    case IR::Opcode::VoteAny:
+    case IR::Opcode::VoteEqual:
+    case IR::Opcode::SubgroupBallot:
+        info.uses_subgroup_vote = true;
+        break;
     default:
         break;
     }

src/shader_recompiler/profile.h

@@ -19,6 +19,8 @@ struct Profile {
     bool support_fp16_signed_zero_nan_preserve{};
     bool support_fp32_signed_zero_nan_preserve{};
     bool support_fp64_signed_zero_nan_preserve{};
+    bool support_vote{};
+    bool warp_size_potentially_larger_than_guest{};
 
     // FClamp is broken and OpFMax + OpFMin should be used instead
     bool has_broken_spirv_clamp{};

src/shader_recompiler/shader_info.h

@@ -80,6 +80,7 @@ struct Info {
     bool uses_sampled_1d{};
     bool uses_sparse_residency{};
     bool uses_demote_to_helper_invocation{};
+    bool uses_subgroup_vote{};
 
     IR::Type used_constant_buffer_types{};
 

src/video_core/renderer_vulkan/vk_compute_pipeline.cpp

@@ -36,13 +36,18 @@ ComputePipeline::ComputePipeline(const Device& device, VKDescriptorPool& descrip
     descriptor_update_template = std::move(tuple.descriptor_update_template);
     descriptor_allocator = DescriptorAllocator(descriptor_pool, *descriptor_set_layout);
 
+    const VkPipelineShaderStageRequiredSubgroupSizeCreateInfoEXT subgroup_size_ci{
+        .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_REQUIRED_SUBGROUP_SIZE_CREATE_INFO_EXT,
+        .pNext = nullptr,
+        .requiredSubgroupSize = GuestWarpSize,
+    };
     pipeline = device.GetLogical().CreateComputePipeline({
         .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
         .pNext = nullptr,
         .flags = 0,
         .stage{
             .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
-            .pNext = nullptr,
+            .pNext = device.IsExtSubgroupSizeControlSupported() ? &subgroup_size_ci : nullptr,
             .flags = 0,
             .stage = VK_SHADER_STAGE_COMPUTE_BIT,
             .module = *spv_module,

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -455,6 +455,8 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, Tegra::GPU& gpu_,
             float_control.shaderSignedZeroInfNanPreserveFloat32 != VK_FALSE,
         .support_fp64_signed_zero_nan_preserve =
             float_control.shaderSignedZeroInfNanPreserveFloat64 != VK_FALSE,
+        .support_vote = true,
+        .warp_size_potentially_larger_than_guest = device.IsWarpSizePotentiallyBiggerThanGuest(),
         .has_broken_spirv_clamp = driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS_KHR,
     };
 }

src/video_core/vulkan_common/vulkan_device.cpp

@@ -737,6 +737,7 @@ std::vector<const char*> Device::LoadExtensions(bool requires_surface) {
             subgroup_properties.maxSubgroupSize >= GuestWarpSize) {
             extensions.push_back(VK_EXT_SUBGROUP_SIZE_CONTROL_EXTENSION_NAME);
             guest_warp_stages = subgroup_properties.requiredSubgroupSizeStages;
+            ext_subgroup_size_control = true;
         }
     } else {
         is_warp_potentially_bigger = true;

src/video_core/vulkan_common/vulkan_device.h

@@ -193,6 +193,11 @@ public:
         return ext_shader_viewport_index_layer;
     }
 
+    /// Returns true if the device supports VK_EXT_subgroup_size_control.
+    bool IsExtSubgroupSizeControlSupported() const {
+        return ext_subgroup_size_control;
+    }
+
     /// Returns true if the device supports VK_EXT_transform_feedback.
     bool IsExtTransformFeedbackSupported() const {
         return ext_transform_feedback;
@@ -297,6 +302,7 @@ private:
     bool ext_depth_range_unrestricted{};        ///< Support for VK_EXT_depth_range_unrestricted.
     bool ext_shader_viewport_index_layer{};     ///< Support for VK_EXT_shader_viewport_index_layer.
     bool ext_tooling_info{};                    ///< Support for VK_EXT_tooling_info.
+    bool ext_subgroup_size_control{};           ///< Support for VK_EXT_subgroup_size_control.
     bool ext_transform_feedback{};              ///< Support for VK_EXT_transform_feedback.
     bool ext_custom_border_color{};             ///< Support for VK_EXT_custom_border_color.
     bool ext_extended_dynamic_state{};          ///< Support for VK_EXT_extended_dynamic_state.