diff options
author | Fernando Sahmkow <fsahmkow27@gmail.com> | 2023-08-18 18:30:31 +0200 |
---|---|---|
committer | Liam <byteslice@airmail.cc> | 2023-08-19 04:17:02 +0200 |
commit | c03f0b3c893f2bc2ae4f1e1825c5ac1453c36710 (patch) | |
tree | 3f7f14fad6574878a7c0f94fbe16d574b61cb95a /src/shader_recompiler/backend | |
parent | Merge pull request #11278 from Kelebek1/dma_sync (diff) | |
download | yuzu-c03f0b3c893f2bc2ae4f1e1825c5ac1453c36710.tar yuzu-c03f0b3c893f2bc2ae4f1e1825c5ac1453c36710.tar.gz yuzu-c03f0b3c893f2bc2ae4f1e1825c5ac1453c36710.tar.bz2 yuzu-c03f0b3c893f2bc2ae4f1e1825c5ac1453c36710.tar.lz yuzu-c03f0b3c893f2bc2ae4f1e1825c5ac1453c36710.tar.xz yuzu-c03f0b3c893f2bc2ae4f1e1825c5ac1453c36710.tar.zst yuzu-c03f0b3c893f2bc2ae4f1e1825c5ac1453c36710.zip |
Diffstat (limited to 'src/shader_recompiler/backend')
-rw-r--r-- | src/shader_recompiler/backend/glasm/emit_glasm_image.cpp | 23 | ||||
-rw-r--r-- | src/shader_recompiler/backend/glsl/emit_glsl_image.cpp | 8 |
2 files changed, 26 insertions, 5 deletions
diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp index 85ee27333..d0e308124 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp +++ b/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp @@ -558,12 +558,15 @@ void EmitImageGradient(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, const IR::Value& coord, const IR::Value& derivatives, const IR::Value& offset, const IR::Value& lod_clamp) { const auto info{inst.Flags<IR::TextureInstInfo>()}; - ScopedRegister dpdx, dpdy; + ScopedRegister dpdx, dpdy, coords; const bool multi_component{info.num_derivates > 1 || info.has_lod_clamp}; if (multi_component) { // Allocate this early to avoid aliasing other registers dpdx = ScopedRegister{ctx.reg_alloc}; dpdy = ScopedRegister{ctx.reg_alloc}; + if (info.num_derivates >= 3) { + coords = ScopedRegister{ctx.reg_alloc}; + } } const auto sparse_inst{PrepareSparse(inst)}; const std::string_view sparse_mod{sparse_inst ? ".SPARSE" : ""}; @@ -580,15 +583,27 @@ void EmitImageGradient(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, "MOV.F {}.y,{}.w;", dpdx.reg, derivatives_vec, dpdx.reg, derivatives_vec, dpdy.reg, derivatives_vec, dpdy.reg, derivatives_vec); + Register final_coord; + if (info.num_derivates >= 3) { + ctx.Add("MOV.F {}.z,{}.x;" + "MOV.F {}.z,{}.y;", + dpdx.reg, coord_vec, dpdy.reg, coord_vec); + ctx.Add("MOV.F {}.x,0;" + "MOV.F {}.y,0;", + "MOV.F {}.z,0;", coords.reg, coords.reg, coords.reg); + final_coord = coords.reg; + } else { + final_coord = coord_vec; + } if (info.has_lod_clamp) { const ScalarF32 lod_clamp_value{ctx.reg_alloc.Consume(lod_clamp)}; ctx.Add("MOV.F {}.w,{};" "TXD.F.LODCLAMP{} {},{},{},{},{},{}{};", - dpdy.reg, lod_clamp_value, sparse_mod, ret, coord_vec, dpdx.reg, dpdy.reg, + dpdy.reg, lod_clamp_value, sparse_mod, ret, final_coord, dpdx.reg, dpdy.reg, texture, type, offset_vec); } else { - ctx.Add("TXD.F{} {},{},{},{},{},{}{};", sparse_mod, ret, coord_vec, dpdx.reg, dpdy.reg, - texture, type, offset_vec); + ctx.Add("TXD.F{} {},{},{},{},{},{}{};", sparse_mod, ret, final_coord, dpdx.reg, + dpdy.reg, texture, type, offset_vec); } } else { ctx.Add("TXD.F{} {},{},{}.x,{}.y,{},{}{};", sparse_mod, ret, coord_vec, derivatives_vec, diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp index 418505475..3ad668a47 100644 --- a/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp +++ b/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp @@ -548,7 +548,7 @@ void EmitImageGradient(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, if (sparse_inst) { throw NotImplementedException("EmitImageGradient Sparse"); } - if (!offset.IsEmpty()) { + if (!offset.IsEmpty() && info.num_derivates <= 2) { throw NotImplementedException("EmitImageGradient offset"); } const auto texture{Texture(ctx, info, index)}; @@ -556,6 +556,12 @@ void EmitImageGradient(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, const bool multi_component{info.num_derivates > 1 || info.has_lod_clamp}; const auto derivatives_vec{ctx.var_alloc.Consume(derivatives)}; if (multi_component) { + if (info.num_derivates >= 3) { + const auto offset_vec{ctx.var_alloc.Consume(offset)}; + ctx.Add("{}=textureGrad({},{},vec3({}.xz, {}.x),vec3({}.yz, {}.y));", texel, texture, + coords, derivatives_vec, offset_vec, derivatives_vec, offset_vec); + return; + } ctx.Add("{}=textureGrad({},{},vec2({}.xz),vec2({}.yz));", texel, texture, coords, derivatives_vec, derivatives_vec); } else { |