Merge pull request #9694 from ameerj/txq-mips

shader_recompiler: TXQ: Skip QueryLevels when possible

11 files changed, 37 insertions, 29 deletions

diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp
index cf6065208..b7bc11416 100644
--- a/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp
+++ b/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp
@@ -532,7 +532,7 @@ void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
 }
 
 void EmitImageQueryDimensions(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
-                              ScalarS32 lod) {
+                              ScalarS32 lod, [[maybe_unused]] const IR::Value& skip_mips) {
     const auto info{inst.Flags<IR::TextureInstInfo>()};
     const std::string texture{Texture(ctx, info, index)};
     const std::string_view type{TextureType(info)};
diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h b/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h
index 415a249e4..1a1ea61d5 100644
--- a/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h
+++ b/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h
@@ -581,7 +581,7 @@ void EmitImageGatherDref(EmitContext& ctx, IR::Inst& inst, const IR::Value& inde
 void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
                     const IR::Value& coord, const IR::Value& offset, ScalarS32 lod, ScalarS32 ms);
 void EmitImageQueryDimensions(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
-                              ScalarS32 lod);
+                              ScalarS32 lod, const IR::Value& skip_mips);
 void EmitImageQueryLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, Register coord);
 void EmitImageGradient(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
                        const IR::Value& coord, const IR::Value& derivatives,
diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
index d8874b0cc..4be2c25ec 100644
--- a/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
+++ b/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
@@ -460,27 +460,27 @@ void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
 }
 
 void EmitImageQueryDimensions(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
-                              std::string_view lod) {
+                              std::string_view lod, const IR::Value& skip_mips_val) {
     const auto info{inst.Flags<IR::TextureInstInfo>()};
     const auto texture{Texture(ctx, info, index)};
+    const bool skip_mips{skip_mips_val.U1()};
+    const auto mips{
+        [&] { return skip_mips ? "0u" : fmt::format("uint(textureQueryLevels({}))", texture); }};
     switch (info.type) {
     case TextureType::Color1D:
-        return ctx.AddU32x4(
-            "{}=uvec4(uint(textureSize({},int({}))),0u,0u,uint(textureQueryLevels({})));", inst,
-            texture, lod, texture);
+        return ctx.AddU32x4("{}=uvec4(uint(textureSize({},int({}))),0u,0u,{});", inst, texture, lod,
+                            mips());
     case TextureType::ColorArray1D:
     case TextureType::Color2D:
     case TextureType::ColorCube:
     case TextureType::Color2DRect:
-        return ctx.AddU32x4(
-            "{}=uvec4(uvec2(textureSize({},int({}))),0u,uint(textureQueryLevels({})));", inst,
-            texture, lod, texture);
+        return ctx.AddU32x4("{}=uvec4(uvec2(textureSize({},int({}))),0u,{});", inst, texture, lod,
+                            mips());
     case TextureType::ColorArray2D:
     case TextureType::Color3D:
     case TextureType::ColorArrayCube:
-        return ctx.AddU32x4(
-            "{}=uvec4(uvec3(textureSize({},int({}))),uint(textureQueryLevels({})));", inst, texture,
-            lod, texture);
+        return ctx.AddU32x4("{}=uvec4(uvec3(textureSize({},int({}))),{});", inst, texture, lod,
+                            mips());
     case TextureType::Buffer:
         throw NotImplementedException("EmitImageQueryDimensions Texture buffers");
     }
diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h b/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h
index c6df1dba7..8d0a65047 100644
--- a/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h
+++ b/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h
@@ -654,7 +654,7 @@ void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
                     std::string_view coords, std::string_view offset, std::string_view lod,
                     std::string_view ms);
 void EmitImageQueryDimensions(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
-                              std::string_view lod);
+                              std::string_view lod, const IR::Value& skip_mips);
 void EmitImageQueryLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
                        std::string_view coords);
 void EmitImageGradient(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp
index c898ce12f..3b969d915 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp
+++ b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp
@@ -445,11 +445,13 @@ Id EmitImageFetch(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id c
                 TextureImage(ctx, info, index), coords, operands.MaskOptional(), operands.Span());
 }
 
-Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id lod) {
+Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id lod,
+                            const IR::Value& skip_mips_val) {
     const auto info{inst->Flags<IR::TextureInstInfo>()};
     const Id image{TextureImage(ctx, info, index)};
     const Id zero{ctx.u32_zero_value};
-    const auto mips{[&] { return ctx.OpImageQueryLevels(ctx.U32[1], image); }};
+    const bool skip_mips{skip_mips_val.U1()};
+    const auto mips{[&] { return skip_mips ? zero : ctx.OpImageQueryLevels(ctx.U32[1], image); }};
     switch (info.type) {
     case TextureType::Color1D:
         return ctx.OpCompositeConstruct(ctx.U32[4], ctx.OpImageQuerySizeLod(ctx.U32[1], image, lod),
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h b/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h
index db12e8176..a440b557d 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h
+++ b/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h
@@ -539,7 +539,8 @@ Id EmitImageGatherDref(EmitContext& ctx, IR::Inst* inst, const IR::Value& index,
                        const IR::Value& offset, const IR::Value& offset2, Id dref);
 Id EmitImageFetch(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords, Id offset,
                   Id lod, Id ms);
-Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id lod);
+Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id lod,
+                            const IR::Value& skip_mips);
 Id EmitImageQueryLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords);
 Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords,
                      Id derivates, Id offset, Id lod_clamp);
diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.cpp b/src/shader_recompiler/frontend/ir/ir_emitter.cpp
index 430797d23..b7caa4246 100644
--- a/src/shader_recompiler/frontend/ir/ir_emitter.cpp
+++ b/src/shader_recompiler/frontend/ir/ir_emitter.cpp
@@ -1846,15 +1846,16 @@ Value IREmitter::ImageFetch(const Value& handle, const Value& coords, const Valu
     return Inst(op, Flags{info}, handle, coords, offset, lod, multisampling);
 }
 
-Value IREmitter::ImageQueryDimension(const Value& handle, const IR::U32& lod) {
+Value IREmitter::ImageQueryDimension(const Value& handle, const IR::U32& lod,
+                                     const IR::U1& skip_mips) {
     const Opcode op{handle.IsImmediate() ? Opcode::BoundImageQueryDimensions
                                          : Opcode::BindlessImageQueryDimensions};
-    return Inst(op, handle, lod);
+    return Inst(op, handle, lod, skip_mips);
 }
 
 Value IREmitter::ImageQueryDimension(const Value& handle, const IR::U32& lod,
-                                     TextureInstInfo info) {
-    return Inst(Opcode::ImageQueryDimensions, Flags{info}, handle, lod);
+                                     const IR::U1& skip_mips, TextureInstInfo info) {
+    return Inst(Opcode::ImageQueryDimensions, Flags{info}, handle, lod, skip_mips);
 }
 
 Value IREmitter::ImageQueryLod(const Value& handle, const Value& coords, TextureInstInfo info) {
diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.h b/src/shader_recompiler/frontend/ir/ir_emitter.h
index 7aaaa4ab0..df158c928 100644
--- a/src/shader_recompiler/frontend/ir/ir_emitter.h
+++ b/src/shader_recompiler/frontend/ir/ir_emitter.h
@@ -320,9 +320,10 @@ public:
     [[nodiscard]] F32 ImageSampleDrefExplicitLod(const Value& handle, const Value& coords,
                                                  const F32& dref, const F32& lod,
                                                  const Value& offset, TextureInstInfo info);
-    [[nodiscard]] Value ImageQueryDimension(const Value& handle, const IR::U32& lod);
     [[nodiscard]] Value ImageQueryDimension(const Value& handle, const IR::U32& lod,
-                                            TextureInstInfo info);
+                                            const IR::U1& skip_mips);
+    [[nodiscard]] Value ImageQueryDimension(const Value& handle, const IR::U32& lod,
+                                            const IR::U1& skip_mips, TextureInstInfo info);
 
     [[nodiscard]] Value ImageQueryLod(const Value& handle, const Value& coords,
                                       TextureInstInfo info);
diff --git a/src/shader_recompiler/frontend/ir/opcodes.inc b/src/shader_recompiler/frontend/ir/opcodes.inc
index 24e82f802..4447d67b0 100644
--- a/src/shader_recompiler/frontend/ir/opcodes.inc
+++ b/src/shader_recompiler/frontend/ir/opcodes.inc
@@ -482,7 +482,7 @@ OPCODE(BindlessImageSampleDrefExplicitLod,                  F32,            U32,
 OPCODE(BindlessImageGather,                                 F32x4,          U32,            Opaque,         Opaque,         Opaque,                         )
 OPCODE(BindlessImageGatherDref,                             F32x4,          U32,            Opaque,         Opaque,         Opaque,         F32,            )
 OPCODE(BindlessImageFetch,                                  F32x4,          U32,            Opaque,         Opaque,         U32,            Opaque,         )
-OPCODE(BindlessImageQueryDimensions,                        U32x4,          U32,            U32,                                                            )
+OPCODE(BindlessImageQueryDimensions,                        U32x4,          U32,            U32,            U1,                                             )
 OPCODE(BindlessImageQueryLod,                               F32x4,          U32,            Opaque,                                                         )
 OPCODE(BindlessImageGradient,                               F32x4,          U32,            Opaque,         Opaque,         Opaque,         Opaque,         )
 OPCODE(BindlessImageRead,                                   U32x4,          U32,            Opaque,                                                         )
@@ -495,7 +495,7 @@ OPCODE(BoundImageSampleDrefExplicitLod,                     F32,            U32,
 OPCODE(BoundImageGather,                                    F32x4,          U32,            Opaque,         Opaque,         Opaque,                         )
 OPCODE(BoundImageGatherDref,                                F32x4,          U32,            Opaque,         Opaque,         Opaque,         F32,            )
 OPCODE(BoundImageFetch,                                     F32x4,          U32,            Opaque,         Opaque,         U32,            Opaque,         )
-OPCODE(BoundImageQueryDimensions,                           U32x4,          U32,            U32,                                                            )
+OPCODE(BoundImageQueryDimensions,                           U32x4,          U32,            U32,            U1,                                             )
 OPCODE(BoundImageQueryLod,                                  F32x4,          U32,            Opaque,                                                         )
 OPCODE(BoundImageGradient,                                  F32x4,          U32,            Opaque,         Opaque,         Opaque,         Opaque,         )
 OPCODE(BoundImageRead,                                      U32x4,          U32,            Opaque,                                                         )
@@ -508,7 +508,7 @@ OPCODE(ImageSampleDrefExplicitLod,                          F32,            Opaq
 OPCODE(ImageGather,                                         F32x4,          Opaque,         Opaque,         Opaque,         Opaque,                         )
 OPCODE(ImageGatherDref,                                     F32x4,          Opaque,         Opaque,         Opaque,         Opaque,         F32,            )
 OPCODE(ImageFetch,                                          F32x4,          Opaque,         Opaque,         Opaque,         U32,            Opaque,         )
-OPCODE(ImageQueryDimensions,                                U32x4,          Opaque,         U32,                                                            )
+OPCODE(ImageQueryDimensions,                                U32x4,          Opaque,         U32,            U1,                                             )
 OPCODE(ImageQueryLod,                                       F32x4,          Opaque,         Opaque,                                                         )
 OPCODE(ImageGradient,                                       F32x4,          Opaque,         Opaque,         Opaque,         Opaque,         Opaque,         )
 OPCODE(ImageRead,                                           U32x4,          Opaque,         Opaque,                                                         )
diff --git a/src/shader_recompiler/frontend/maxwell/translate/impl/texture_query.cpp b/src/shader_recompiler/frontend/maxwell/translate/impl/texture_query.cpp
index f8cfd4ab6..39af62559 100644
--- a/src/shader_recompiler/frontend/maxwell/translate/impl/texture_query.cpp
+++ b/src/shader_recompiler/frontend/maxwell/translate/impl/texture_query.cpp
@@ -15,11 +15,13 @@ enum class Mode : u64 {
     SamplePos = 5,
 };
 
-IR::Value Query(TranslatorVisitor& v, const IR::U32& handle, Mode mode, IR::Reg src_reg) {
+IR::Value Query(TranslatorVisitor& v, const IR::U32& handle, Mode mode, IR::Reg src_reg, u64 mask) {
     switch (mode) {
     case Mode::Dimension: {
+        const bool needs_num_mips{((mask >> 3) & 1) != 0};
+        const IR::U1 skip_mips{v.ir.Imm1(!needs_num_mips)};
         const IR::U32 lod{v.X(src_reg)};
-        return v.ir.ImageQueryDimension(handle, lod);
+        return v.ir.ImageQueryDimension(handle, lod, skip_mips);
     }
     case Mode::TextureType:
     case Mode::SamplePos:
@@ -46,7 +48,7 @@ void Impl(TranslatorVisitor& v, u64 insn, std::optional<u32> cbuf_offset) {
         handle = v.X(src_reg);
         ++src_reg;
     }
-    const IR::Value query{Query(v, handle, txq.mode, src_reg)};
+    const IR::Value query{Query(v, handle, txq.mode, src_reg, txq.mask)};
     IR::Reg dest_reg{txq.dest_reg};
     for (int element = 0; element < 4; ++element) {
         if (((txq.mask >> element) & 1) == 0) {
diff --git a/src/shader_recompiler/ir_opt/texture_pass.cpp b/src/shader_recompiler/ir_opt/texture_pass.cpp
index 70b620dcb..b28194d0e 100644
--- a/src/shader_recompiler/ir_opt/texture_pass.cpp
+++ b/src/shader_recompiler/ir_opt/texture_pass.cpp
@@ -452,7 +452,8 @@ void PatchImageSampleImplicitLod(IR::Block& block, IR::Inst& inst) {
     const IR::Value coord(inst.Arg(1));
     const IR::Value handle(ir.Imm32(0));
     const IR::U32 lod{ir.Imm32(0)};
-    const IR::Value texture_size = ir.ImageQueryDimension(handle, lod, info);
+    const IR::U1 skip_mips{ir.Imm1(true)};
+    const IR::Value texture_size = ir.ImageQueryDimension(handle, lod, skip_mips, info);
     inst.SetArg(
         1, ir.CompositeConstruct(
                ir.FPMul(IR::F32(ir.CompositeExtract(coord, 0)),