diff --git a/src/video_core/clipper.cpp b/src/video_core/clipper.cpp index 3a09d62f4..a385589d2 100644 --- a/src/video_core/clipper.cpp +++ b/src/video_core/clipper.cpp @@ -59,12 +59,12 @@ static void InitScreenCoordinates(OutputVertex& vtx) } viewport; const auto& regs = g_state.regs; - viewport.halfsize_x = float24::FromRawFloat24(regs.viewport_size_x); - viewport.halfsize_y = float24::FromRawFloat24(regs.viewport_size_y); + viewport.halfsize_x = float24::FromRaw(regs.viewport_size_x); + viewport.halfsize_y = float24::FromRaw(regs.viewport_size_y); viewport.offset_x = float24::FromFloat32(static_cast(regs.viewport_corner.x)); viewport.offset_y = float24::FromFloat32(static_cast(regs.viewport_corner.y)); - viewport.zscale = float24::FromRawFloat24(regs.viewport_depth_range); - viewport.offset_z = float24::FromRawFloat24(regs.viewport_depth_far_plane); + viewport.zscale = float24::FromRaw(regs.viewport_depth_range); + viewport.offset_z = float24::FromRaw(regs.viewport_depth_far_plane); float24 inv_w = float24::FromFloat32(1.f) / vtx.pos.w; vtx.color *= inv_w; diff --git a/src/video_core/command_processor.cpp b/src/video_core/command_processor.cpp index 6540ccb26..5dfedfe31 100644 --- a/src/video_core/command_processor.cpp +++ b/src/video_core/command_processor.cpp @@ -98,10 +98,10 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) { Math::Vec4& attribute = g_state.vs.default_attributes[setup.index]; // NOTE: The destination component order indeed is "backwards" - attribute.w = float24::FromRawFloat24(default_attr_write_buffer[0] >> 8); - attribute.z = float24::FromRawFloat24(((default_attr_write_buffer[0] & 0xFF) << 16) | ((default_attr_write_buffer[1] >> 16) & 0xFFFF)); - attribute.y = float24::FromRawFloat24(((default_attr_write_buffer[1] & 0xFFFF) << 8) | ((default_attr_write_buffer[2] >> 24) & 0xFF)); - attribute.x = float24::FromRawFloat24(default_attr_write_buffer[2] & 0xFFFFFF); + attribute.w = float24::FromRaw(default_attr_write_buffer[0] >> 8); + attribute.z = float24::FromRaw(((default_attr_write_buffer[0] & 0xFF) << 16) | ((default_attr_write_buffer[1] >> 16) & 0xFFFF)); + attribute.y = float24::FromRaw(((default_attr_write_buffer[1] & 0xFFFF) << 8) | ((default_attr_write_buffer[2] >> 24) & 0xFF)); + attribute.x = float24::FromRaw(default_attr_write_buffer[2] & 0xFFFFFF); LOG_TRACE(HW_GPU, "Set default VS attribute %x to (%f %f %f %f)", (int)setup.index, attribute.x.ToFloat32(), attribute.y.ToFloat32(), attribute.z.ToFloat32(), @@ -418,10 +418,10 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) { uniform[3 - i] = float24::FromFloat32(*(float*)(&uniform_write_buffer[i])); } else { // TODO: Untested - uniform.w = float24::FromRawFloat24(uniform_write_buffer[0] >> 8); - uniform.z = float24::FromRawFloat24(((uniform_write_buffer[0] & 0xFF)<<16) | ((uniform_write_buffer[1] >> 16) & 0xFFFF)); - uniform.y = float24::FromRawFloat24(((uniform_write_buffer[1] & 0xFFFF)<<8) | ((uniform_write_buffer[2] >> 24) & 0xFF)); - uniform.x = float24::FromRawFloat24(uniform_write_buffer[2] & 0xFFFFFF); + uniform.w = float24::FromRaw(uniform_write_buffer[0] >> 8); + uniform.z = float24::FromRaw(((uniform_write_buffer[0] & 0xFF) << 16) | ((uniform_write_buffer[1] >> 16) & 0xFFFF)); + uniform.y = float24::FromRaw(((uniform_write_buffer[1] & 0xFFFF) << 8) | ((uniform_write_buffer[2] >> 24) & 0xFF)); + uniform.x = float24::FromRaw(uniform_write_buffer[2] & 0xFFFFFF); } LOG_TRACE(HW_GPU, "Set uniform %x to (%f %f %f %f)", (int)uniform_setup.index, diff --git a/src/video_core/pica_types.h b/src/video_core/pica_types.h index a34421c5d..53f61f287 100644 --- a/src/video_core/pica_types.h +++ b/src/video_core/pica_types.h @@ -4,35 +4,51 @@ #pragma once +#include + #include "common/common_types.h" namespace Pica { -struct float24 { - static float24 FromFloat32(float val) { - float24 ret; +/** + * Template class for converting arbitrary Pica float types to IEEE 754 32-bit single-precision + * floating point. + * + * When decoding, format is as follows: + * - The first `M` bits are the mantissa + * - The next `E` bits are the exponent + * - The last bit is the sign bit + * + * @todo Verify on HW if this conversion is sufficently accurate. + */ +template +struct Float { +public: + static Float FromFloat32(float val) { + Float ret; ret.value = val; return ret; } - // 16 bit mantissa, 7 bit exponent, 1 bit sign - // TODO: No idea if this works as intended - static float24 FromRawFloat24(u32 hex) { - float24 ret; - if ((hex & 0xFFFFFF) == 0) { - ret.value = 0; - } else { - u32 mantissa = hex & 0xFFFF; - u32 exponent = (hex >> 16) & 0x7F; - u32 sign = hex >> 23; - ret.value = std::pow(2.0f, (float)exponent-63.0f) * (1.0f + mantissa * std::pow(2.0f, -16.f)); - if (sign) - ret.value = -ret.value; - } - return ret; + static Float FromRaw(u32 hex) { + Float res; + + const int width = M + E + 1; + const int bias = 128 - (1 << (E - 1)); + const int exponent = (hex >> M) & ((1 << E) - 1); + const unsigned mantissa = hex & ((1 << M) - 1); + + if (hex & ((1 << (width - 1)) - 1)) + hex = ((hex >> (E + M)) << 31) | (mantissa << (23 - M)) | ((exponent + bias) << 23); + else + hex = ((hex >> (E + M)) << 31); + + std::memcpy(&res.value, &hex, sizeof(float)); + + return res; } - static float24 Zero() { + static Float Zero() { return FromFloat32(0.f); } @@ -41,27 +57,27 @@ struct float24 { return value; } - float24 operator * (const float24& flt) const { + Float operator * (const Float& flt) const { if ((this->value == 0.f && !std::isnan(flt.value)) || (flt.value == 0.f && !std::isnan(this->value))) // PICA gives 0 instead of NaN when multiplying by inf return Zero(); - return float24::FromFloat32(ToFloat32() * flt.ToFloat32()); + return Float::FromFloat32(ToFloat32() * flt.ToFloat32()); } - float24 operator / (const float24& flt) const { - return float24::FromFloat32(ToFloat32() / flt.ToFloat32()); + Float operator / (const Float& flt) const { + return Float::FromFloat32(ToFloat32() / flt.ToFloat32()); } - float24 operator + (const float24& flt) const { - return float24::FromFloat32(ToFloat32() + flt.ToFloat32()); + Float operator + (const Float& flt) const { + return Float::FromFloat32(ToFloat32() + flt.ToFloat32()); } - float24 operator - (const float24& flt) const { - return float24::FromFloat32(ToFloat32() - flt.ToFloat32()); + Float operator - (const Float& flt) const { + return Float::FromFloat32(ToFloat32() - flt.ToFloat32()); } - float24& operator *= (const float24& flt) { + Float& operator *= (const Float& flt) { if ((this->value == 0.f && !std::isnan(flt.value)) || (flt.value == 0.f && !std::isnan(this->value))) // PICA gives 0 instead of NaN when multiplying by inf @@ -70,111 +86,61 @@ struct float24 { return *this; } - float24& operator /= (const float24& flt) { + Float& operator /= (const Float& flt) { value /= flt.ToFloat32(); return *this; } - float24& operator += (const float24& flt) { + Float& operator += (const Float& flt) { value += flt.ToFloat32(); return *this; } - float24& operator -= (const float24& flt) { + Float& operator -= (const Float& flt) { value -= flt.ToFloat32(); return *this; } - float24 operator - () const { - return float24::FromFloat32(-ToFloat32()); + Float operator - () const { + return Float::FromFloat32(-ToFloat32()); } - bool operator < (const float24& flt) const { + bool operator < (const Float& flt) const { return ToFloat32() < flt.ToFloat32(); } - bool operator > (const float24& flt) const { + bool operator > (const Float& flt) const { return ToFloat32() > flt.ToFloat32(); } - bool operator >= (const float24& flt) const { + bool operator >= (const Float& flt) const { return ToFloat32() >= flt.ToFloat32(); } - bool operator <= (const float24& flt) const { + bool operator <= (const Float& flt) const { return ToFloat32() <= flt.ToFloat32(); } - bool operator == (const float24& flt) const { + bool operator == (const Float& flt) const { return ToFloat32() == flt.ToFloat32(); } - bool operator != (const float24& flt) const { + bool operator != (const Float& flt) const { return ToFloat32() != flt.ToFloat32(); } private: + static const unsigned MASK = (1 << (M + E + 1)) - 1; + static const unsigned MANTISSA_MASK = (1 << M) - 1; + static const unsigned EXPONENT_MASK = (1 << E) - 1; + // Stored as a regular float, merely for convenience // TODO: Perform proper arithmetic on this! float value; }; -static_assert(sizeof(float24) == sizeof(float), "Shader JIT assumes float24 is implemented as a 32-bit float"); - -struct float16 { - // 10 bit mantissa, 5 bit exponent, 1 bit sign - // TODO: No idea if this works as intended - static float16 FromRawFloat16(u32 hex) { - float16 ret; - if ((hex & 0xFFFF) == 0) { - ret.value = 0; - } else { - u32 mantissa = hex & 0x3FF; - u32 exponent = (hex >> 10) & 0x1F; - u32 sign = (hex >> 15) & 1; - ret.value = std::pow(2.0f, (float)exponent - 15.0f) * (1.0f + mantissa * std::pow(2.0f, -10.f)); - if (sign) - ret.value = -ret.value; - } - return ret; - } - - float ToFloat32() const { - return value; - } - -private: - // Stored as a regular float, merely for convenience - // TODO: Perform proper arithmetic on this! - float value; -}; - -struct float20 { - // 12 bit mantissa, 7 bit exponent, 1 bit sign - // TODO: No idea if this works as intended - static float20 FromRawFloat20(u32 hex) { - float20 ret; - if ((hex & 0xFFFFF) == 0) { - ret.value = 0; - } else { - u32 mantissa = hex & 0xFFF; - u32 exponent = (hex >> 12) & 0x7F; - u32 sign = (hex >> 19) & 1; - ret.value = std::pow(2.0f, (float)exponent - 63.0f) * (1.0f + mantissa * std::pow(2.0f, -12.f)); - if (sign) - ret.value = -ret.value; - } - return ret; - } - - float ToFloat32() const { - return value; - } - -private: - // Stored as a regular float, merely for convenience - // TODO: Perform proper arithmetic on this! - float value; -}; +using float24 = Float<16, 7>; +using float20 = Float<12, 7>; +using float16 = Float<10, 5>; } // namespace Pica diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp index 6e7d6a40d..d70d62ede 100644 --- a/src/video_core/renderer_opengl/gl_rasterizer.cpp +++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp @@ -810,8 +810,8 @@ void RasterizerOpenGL::SyncCullMode() { } void RasterizerOpenGL::SyncDepthModifiers() { - float depth_scale = -Pica::float24::FromRawFloat24(Pica::g_state.regs.viewport_depth_range).ToFloat32(); - float depth_offset = Pica::float24::FromRawFloat24(Pica::g_state.regs.viewport_depth_far_plane).ToFloat32() / 2.0f; + float depth_scale = -Pica::float24::FromRaw(Pica::g_state.regs.viewport_depth_range).ToFloat32(); + float depth_offset = Pica::float24::FromRaw(Pica::g_state.regs.viewport_depth_far_plane).ToFloat32() / 2.0f; // TODO: Implement scale modifier uniform_block_data.data.depth_offset = depth_offset; @@ -948,9 +948,9 @@ void RasterizerOpenGL::SyncLightAmbient(int light_index) { void RasterizerOpenGL::SyncLightPosition(int light_index) { std::array position = { - Pica::float16::FromRawFloat16(Pica::g_state.regs.lighting.light[light_index].x).ToFloat32(), - Pica::float16::FromRawFloat16(Pica::g_state.regs.lighting.light[light_index].y).ToFloat32(), - Pica::float16::FromRawFloat16(Pica::g_state.regs.lighting.light[light_index].z).ToFloat32() }; + Pica::float16::FromRaw(Pica::g_state.regs.lighting.light[light_index].x).ToFloat32(), + Pica::float16::FromRaw(Pica::g_state.regs.lighting.light[light_index].y).ToFloat32(), + Pica::float16::FromRaw(Pica::g_state.regs.lighting.light[light_index].z).ToFloat32() }; if (position != uniform_block_data.data.light_src[light_index].position) { uniform_block_data.data.light_src[light_index].position = position; @@ -962,8 +962,8 @@ void RasterizerOpenGL::SyncDrawState() { const auto& regs = Pica::g_state.regs; // Sync the viewport - GLsizei viewport_width = (GLsizei)Pica::float24::FromRawFloat24(regs.viewport_size_x).ToFloat32() * 2; - GLsizei viewport_height = (GLsizei)Pica::float24::FromRawFloat24(regs.viewport_size_y).ToFloat32() * 2; + GLsizei viewport_width = (GLsizei)Pica::float24::FromRaw(regs.viewport_size_x).ToFloat32() * 2; + GLsizei viewport_height = (GLsizei)Pica::float24::FromRaw(regs.viewport_size_y).ToFloat32() * 2; // OpenGL uses different y coordinates, so negate corner offset and flip origin // TODO: Ensure viewport_corner.x should not be negated or origin flipped diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h index 4e681f9ea..b9c1d61bd 100644 --- a/src/video_core/renderer_opengl/gl_rasterizer.h +++ b/src/video_core/renderer_opengl/gl_rasterizer.h @@ -83,8 +83,8 @@ struct PicaShaderConfig { res.lighting.light[light_index].directional = light.directional != 0; res.lighting.light[light_index].two_sided_diffuse = light.two_sided_diffuse != 0; res.lighting.light[light_index].dist_atten_enable = regs.lighting.IsDistAttenEnabled(num); - res.lighting.light[light_index].dist_atten_bias = Pica::float20::FromRawFloat20(light.dist_atten_bias).ToFloat32(); - res.lighting.light[light_index].dist_atten_scale = Pica::float20::FromRawFloat20(light.dist_atten_scale).ToFloat32(); + res.lighting.light[light_index].dist_atten_bias = Pica::float20::FromRaw(light.dist_atten_bias).ToFloat32(); + res.lighting.light[light_index].dist_atten_scale = Pica::float20::FromRaw(light.dist_atten_scale).ToFloat32(); } res.lighting.lut_d0.enable = regs.lighting.lut_enable_d0 == 0;