diff options
Diffstat (limited to 'src/video_core/shader')
| -rw-r--r-- | src/video_core/shader/shader.cpp | 18 | ||||
| -rw-r--r-- | src/video_core/shader/shader_interpreter.cpp | 8 |
2 files changed, 17 insertions, 9 deletions
diff --git a/src/video_core/shader/shader.cpp b/src/video_core/shader/shader.cpp index 509558fc0..eb1db0778 100644 --- a/src/video_core/shader/shader.cpp +++ b/src/video_core/shader/shader.cpp @@ -121,15 +121,23 @@ OutputVertex Run(UnitState<false>& state, const InputVertex& input, int num_attr OutputVertex ret; // TODO(neobrain): Under some circumstances, up to 16 attributes may be output. We need to // figure out what those circumstances are and enable the remaining outputs then. - for (int i = 0; i < 7; ++i) { - const auto& output_register_map = g_state.regs.vs_output_attributes[i]; // TODO: Don't hardcode VS here + unsigned index = 0; + for (unsigned i = 0; i < 7; ++i) { + + if (index >= g_state.regs.vs_output_total) + break; + + if ((g_state.regs.vs.output_mask & (1 << i)) == 0) + continue; + + const auto& output_register_map = g_state.regs.vs_output_attributes[index]; // TODO: Don't hardcode VS here u32 semantics[4] = { output_register_map.map_x, output_register_map.map_y, output_register_map.map_z, output_register_map.map_w }; - for (int comp = 0; comp < 4; ++comp) { + for (unsigned comp = 0; comp < 4; ++comp) { float24* out = ((float24*)&ret) + semantics[comp]; if (semantics[comp] != Regs::VSOutputAttributes::INVALID) { *out = state.registers.output[i][comp]; @@ -139,10 +147,12 @@ OutputVertex Run(UnitState<false>& state, const InputVertex& input, int num_attr memset(out, 0, sizeof(*out)); } } + + index++; } // The hardware takes the absolute and saturates vertex colors like this, *before* doing interpolation - for (int i = 0; i < 4; ++i) { + for (unsigned i = 0; i < 4; ++i) { ret.color[i] = float24::FromFloat32( std::fmin(std::fabs(ret.color[i].ToFloat32()), 1.0f)); } diff --git a/src/video_core/shader/shader_interpreter.cpp b/src/video_core/shader/shader_interpreter.cpp index 02e1a1cb1..9b978583e 100644 --- a/src/video_core/shader/shader_interpreter.cpp +++ b/src/video_core/shader/shader_interpreter.cpp @@ -2,10 +2,10 @@ // Licensed under GPLv2 or any later version // Refer to the license.txt file included. -#include <common/file_util.h> - +#include <numeric> #include <nihstro/shader_bytecode.h> +#include "common/file_util.h" #include "video_core/pica.h" #include "video_core/pica_state.h" #include "video_core/shader/shader.h" @@ -214,10 +214,8 @@ void RunInterpreter(UnitState<Debug>& state) { if (opcode == OpCode::Id::DPH || opcode == OpCode::Id::DPHI) src1[3] = float24::FromFloat32(1.0f); - float24 dot = float24::FromFloat32(0.f); int num_components = (opcode == OpCode::Id::DP3) ? 3 : 4; - for (int i = 0; i < num_components; ++i) - dot = dot + src1[i] * src2[i]; + float24 dot = std::inner_product(src1, src1 + num_components, src2, float24::FromFloat32(0.f)); for (int i = 0; i < 4; ++i) { if (!swizzle.DestComponentEnabled(i)) |