From e66d5b88a6f1c2d85c5cd8e351c6ed52c96a0ecf Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Sun, 1 Aug 2021 18:57:45 -0300 Subject: shader: Properly scale image reads and add GL SPIR-V support Thanks for everything! --- .../backend/spirv/emit_spirv_image.cpp | 74 ++++++++++++++++------ 1 file changed, 53 insertions(+), 21 deletions(-) (limited to 'src/shader_recompiler/backend/spirv/emit_spirv_image.cpp') diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp index 7d7c0627e..519ce8b9b 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp +++ b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp @@ -224,6 +224,40 @@ Id Emit(MethodPtrType sparse_ptr, MethodPtrType non_sparse_ptr, EmitContext& ctx Decorate(ctx, inst, sample); return ctx.OpCompositeExtract(result_type, sample, 1U); } + +Id IsScaled(EmitContext& ctx, const IR::Value& index, Id member_index, u32 base_index) { + const Id push_constant_u32{ctx.TypePointer(spv::StorageClass::PushConstant, ctx.U32[1])}; + Id bit{}; + if (index.IsImmediate()) { + // Use BitwiseAnd instead of BitfieldExtract for better codegen on Nvidia OpenGL. + // LOP32I.NZ is used to set the predicate rather than BFE+ISETP. + const u32 index_value{index.U32() + base_index}; + const Id word_index{ctx.Const(index_value / 32)}; + const Id bit_index_mask{ctx.Const(1u << (index_value % 32))}; + const Id pointer{ctx.OpAccessChain(push_constant_u32, ctx.rescaling_push_constants, + member_index, word_index)}; + const Id word{ctx.OpLoad(ctx.U32[1], pointer)}; + bit = ctx.OpBitwiseAnd(ctx.U32[1], word, bit_index_mask); + } else { + Id index_value{ctx.Def(index)}; + if (base_index != 0) { + index_value = ctx.OpIAdd(ctx.U32[1], index_value, ctx.Const(base_index)); + } + const Id word_index{ctx.OpShiftRightArithmetic(ctx.U32[1], index_value, ctx.Const(5u))}; + const Id pointer{ctx.OpAccessChain(push_constant_u32, ctx.rescaling_push_constants, + member_index, word_index)}; + const Id word{ctx.OpLoad(ctx.U32[1], pointer)}; + const Id bit_index{ctx.OpBitwiseAnd(ctx.U32[1], index_value, ctx.Const(31u))}; + bit = ctx.OpBitFieldUExtract(ctx.U32[1], index_value, bit_index, ctx.Const(1u)); + } + return ctx.OpINotEqual(ctx.U1, bit, ctx.u32_zero_value); +} + +Id BitTest(EmitContext& ctx, Id mask, Id bit) { + const Id shifted{ctx.OpShiftRightLogical(ctx.U32[1], mask, bit)}; + const Id bit_value{ctx.OpBitwiseAnd(ctx.U32[1], shifted, ctx.Const(1u))}; + return ctx.OpINotEqual(ctx.U1, bit_value, ctx.u32_zero_value); +} } // Anonymous namespace Id EmitBindlessImageSampleImplicitLod(EmitContext&) { @@ -471,29 +505,27 @@ void EmitImageWrite(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id } Id EmitIsTextureScaled(EmitContext& ctx, const IR::Value& index) { - const Id push_constant_u32{ctx.TypePointer(spv::StorageClass::PushConstant, ctx.U32[1])}; - const Id member_index{ctx.Const(ctx.rescaling_textures_member_index)}; - Id bit{}; - if (index.IsImmediate()) { - // Use BitwiseAnd instead of BitfieldExtract for better codegen on Nvidia OpenGL. - // LOP32I.NZ is used to set the predicate rather than BFE+ISETP. - const u32 index_value{index.U32()}; - const Id word_index{ctx.Const(index_value / 32)}; - const Id bit_index_mask{ctx.Const(1u << (index_value % 32))}; - const Id pointer{ctx.OpAccessChain(push_constant_u32, ctx.rescaling_push_constants, - member_index, word_index)}; - const Id word{ctx.OpLoad(ctx.U32[1], pointer)}; - bit = ctx.OpBitwiseAnd(ctx.U32[1], word, bit_index_mask); + if (ctx.profile.unified_descriptor_binding) { + const Id member_index{ctx.Const(ctx.rescaling_textures_member_index)}; + return IsScaled(ctx, index, member_index, ctx.texture_rescaling_index); } else { - const Id index_value{ctx.Def(index)}; - const Id word_index{ctx.OpShiftRightArithmetic(ctx.U32[1], index_value, ctx.Const(5u))}; - const Id pointer{ctx.OpAccessChain(push_constant_u32, ctx.rescaling_push_constants, - member_index, word_index)}; - const Id word{ctx.OpLoad(ctx.U32[1], pointer)}; - const Id bit_index{ctx.OpBitwiseAnd(ctx.U32[1], index_value, ctx.Const(31u))}; - bit = ctx.OpBitFieldUExtract(ctx.U32[1], index_value, bit_index, ctx.Const(1u)); + const Id composite{ctx.OpLoad(ctx.F32[4], ctx.rescaling_uniform_constant)}; + const Id mask_f32{ctx.OpCompositeExtract(ctx.F32[1], composite, 0u)}; + const Id mask{ctx.OpBitcast(ctx.U32[1], mask_f32)}; + return BitTest(ctx, mask, ctx.Def(index)); + } +} + +Id EmitIsImageScaled(EmitContext& ctx, const IR::Value& index) { + if (ctx.profile.unified_descriptor_binding) { + const Id member_index{ctx.Const(ctx.rescaling_images_member_index)}; + return IsScaled(ctx, index, member_index, ctx.image_rescaling_index); + } else { + const Id composite{ctx.OpLoad(ctx.F32[4], ctx.rescaling_uniform_constant)}; + const Id mask_f32{ctx.OpCompositeExtract(ctx.F32[1], composite, 1u)}; + const Id mask{ctx.OpBitcast(ctx.U32[1], mask_f32)}; + return BitTest(ctx, mask, ctx.Def(index)); } - return ctx.OpINotEqual(ctx.U1, bit, ctx.u32_zero_value); } } // namespace Shader::Backend::SPIRV -- cgit v1.2.3