ShuriZma
/
duckstation
mirror of https://github.com/stenzek/duckstation
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

GPU: Move software fill/write/copy into rasterizer namespace

This commit is contained in:
Stenzek 2024-09-26 18:29:22 +10:00
parent 495a0da8d4
commit c46ec398dc
No known key found for this signature in database
8 changed files with 282 additions and 411 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

191
src/core/gpu.cpp
View File

@ -4,6 +4,7 @@
#include "gpu.h" #include "gpu.h"
#include "dma.h" #include "dma.h"
#include "gpu_shadergen.h" #include "gpu_shadergen.h"
#include "gpu_sw_rasterizer.h"
#include "host.h" #include "host.h"
#include "interrupt_controller.h" #include "interrupt_controller.h"
#include "settings.h" #include "settings.h"
@ -72,6 +73,7 @@ static void JoinScreenshotThreads();
GPU::GPU() GPU::GPU()
{ {
GPU_SW_Rasterizer::SelectImplementation();
ResetStatistics(); ResetStatistics();
} }
@ -1529,195 +1531,6 @@ void GPU::ClearDisplay()
DestroyDeinterlaceTextures(); DestroyDeinterlaceTextures();
} }
void GPU::ReadVRAM(u32 x, u32 y, u32 width, u32 height)
{
}
void GPU::FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color)
{
const u16 color16 = VRAMRGBA8888ToRGBA5551(color);
const GSVector4i fill = GSVector4i(color16, color16, color16, color16, color16, color16, color16, color16);
constexpr u32 vector_width = 8;
const u32 aligned_width = Common::AlignDownPow2(width, vector_width);
if ((x + width) <= VRAM_WIDTH && !IsInterlacedRenderingEnabled())
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
u16* row_ptr = &g_vram[row * VRAM_WIDTH + x];
u32 xoffs = 0;
for (; xoffs < aligned_width; xoffs += vector_width, row_ptr += vector_width)
GSVector4i::store<false>(row_ptr, fill);
for (; xoffs < width; xoffs++)
*(row_ptr++) = color16;
}
}
else if (IsInterlacedRenderingEnabled())
{
// Hardware tests show that fills seem to break on the first two lines when the offset matches the displayed field.
if (IsCRTCScanlinePending())
SynchronizeCRTC();
const u32 active_field = GetActiveLineLSB();
if ((x + width) <= VRAM_WIDTH)
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
if ((row & u32(1)) == active_field)
continue;
u16* row_ptr = &g_vram[row * VRAM_WIDTH + x];
u32 xoffs = 0;
for (; xoffs < aligned_width; xoffs += vector_width, row_ptr += vector_width)
GSVector4i::store<false>(row_ptr, fill);
for (; xoffs < width; xoffs++)
*(row_ptr++) = color16;
}
}
else
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
if ((row & u32(1)) == active_field)
continue;
u16* row_ptr = &g_vram[row * VRAM_WIDTH];
for (u32 xoffs = 0; xoffs < width; xoffs++)
{
const u32 col = (x + xoffs) % VRAM_WIDTH;
row_ptr[col] = color16;
}
}
}
}
else
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
u16* row_ptr = &g_vram[row * VRAM_WIDTH];
for (u32 xoffs = 0; xoffs < width; xoffs++)
{
const u32 col = (x + xoffs) % VRAM_WIDTH;
row_ptr[col] = color16;
}
}
}
}
void GPU::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask)
{
// Fast path when the copy is not oversized.
if ((x + width) <= VRAM_WIDTH && (y + height) <= VRAM_HEIGHT && !set_mask && !check_mask)
{
const u16* src_ptr = static_cast<const u16*>(data);
u16* dst_ptr = &g_vram[y * VRAM_WIDTH + x];
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
std::copy_n(src_ptr, width, dst_ptr);
src_ptr += width;
dst_ptr += VRAM_WIDTH;
}
}
else
{
// Slow path when we need to handle wrap-around.
// During transfer/render operations, if ((dst_pixel & mask_and) == 0) { pixel = src_pixel | mask_or }
const u16* src_ptr = static_cast<const u16*>(data);
const u16 mask_and = check_mask ? 0x8000 : 0;
const u16 mask_or = set_mask ? 0x8000 : 0;
for (u32 row = 0; row < height;)
{
u16* dst_row_ptr = &g_vram[((y + row++) % VRAM_HEIGHT) * VRAM_WIDTH];
for (u32 col = 0; col < width;)
{
// TODO: Handle unaligned reads...
u16* pixel_ptr = &dst_row_ptr[(x + col++) % VRAM_WIDTH];
if (((*pixel_ptr) & mask_and) == 0)
*pixel_ptr = *(src_ptr++) | mask_or;
}
}
}
}
void GPU::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height)
{
// Break up oversized copies. This behavior has not been verified on console.
if ((src_x + width) > VRAM_WIDTH || (dst_x + width) > VRAM_WIDTH)
{
u32 remaining_rows = height;
u32 current_src_y = src_y;
u32 current_dst_y = dst_y;
while (remaining_rows > 0)
{
const u32 rows_to_copy =
std::min<u32>(remaining_rows, std::min<u32>(VRAM_HEIGHT - current_src_y, VRAM_HEIGHT - current_dst_y));
u32 remaining_columns = width;
u32 current_src_x = src_x;
u32 current_dst_x = dst_x;
while (remaining_columns > 0)
{
const u32 columns_to_copy =
std::min<u32>(remaining_columns, std::min<u32>(VRAM_WIDTH - current_src_x, VRAM_WIDTH - current_dst_x));
CopyVRAM(current_src_x, current_src_y, current_dst_x, current_dst_y, columns_to_copy, rows_to_copy);
current_src_x = (current_src_x + columns_to_copy) % VRAM_WIDTH;
current_dst_x = (current_dst_x + columns_to_copy) % VRAM_WIDTH;
remaining_columns -= columns_to_copy;
}
current_src_y = (current_src_y + rows_to_copy) % VRAM_HEIGHT;
current_dst_y = (current_dst_y + rows_to_copy) % VRAM_HEIGHT;
remaining_rows -= rows_to_copy;
}
return;
}
// This doesn't have a fast path, but do we really need one? It's not common.
const u16 mask_and = m_GPUSTAT.GetMaskAND();
const u16 mask_or = m_GPUSTAT.GetMaskOR();
// Copy in reverse when src_x < dst_x, this is verified on console.
if (src_x < dst_x || ((src_x + width - 1) % VRAM_WIDTH) < ((dst_x + width - 1) % VRAM_WIDTH))
{
for (u32 row = 0; row < height; row++)
{
const u16* src_row_ptr = &g_vram[((src_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
u16* dst_row_ptr = &g_vram[((dst_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
for (s32 col = static_cast<s32>(width - 1); col >= 0; col--)
{
const u16 src_pixel = src_row_ptr[(src_x + static_cast<u32>(col)) % VRAM_WIDTH];
u16* dst_pixel_ptr = &dst_row_ptr[(dst_x + static_cast<u32>(col)) % VRAM_WIDTH];
if ((*dst_pixel_ptr & mask_and) == 0)
*dst_pixel_ptr = src_pixel | mask_or;
}
}
}
else
{
for (u32 row = 0; row < height; row++)
{
const u16* src_row_ptr = &g_vram[((src_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
u16* dst_row_ptr = &g_vram[((dst_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
for (u32 col = 0; col < width; col++)
{
const u16 src_pixel = src_row_ptr[(src_x + col) % VRAM_WIDTH];
u16* dst_pixel_ptr = &dst_row_ptr[(dst_x + col) % VRAM_WIDTH];
if ((*dst_pixel_ptr & mask_and) == 0)
*dst_pixel_ptr = src_pixel | mask_or;
}
}
}
}
void GPU::SetClampedDrawingArea() void GPU::SetClampedDrawingArea()
{ {
if (m_drawing_area.left > m_drawing_area.right || m_drawing_area.top > m_drawing_area.bottom) [[unlikely]] if (m_drawing_area.left > m_drawing_area.right || m_drawing_area.top > m_drawing_area.bottom) [[unlikely]]

22
src/core/gpu.h
View File

@ -322,10 +322,10 @@ protected:
bool IsCLUTValid() const; bool IsCLUTValid() const;
// Rendering in the backend // Rendering in the backend
virtual void ReadVRAM(u32 x, u32 y, u32 width, u32 height); virtual void ReadVRAM(u32 x, u32 y, u32 width, u32 height) = 0;
virtual void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color); virtual void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color) = 0;
virtual void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask); virtual void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask) = 0;
virtual void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height); virtual void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height) = 0;
virtual void DispatchRenderCommand() = 0; virtual void DispatchRenderCommand() = 0;
virtual void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) = 0; virtual void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) = 0;
virtual void UpdateDisplay() = 0; virtual void UpdateDisplay() = 0;
@ -416,6 +416,8 @@ protected:
union GPUSTAT union GPUSTAT
{ {
// During transfer/render operations, if ((dst_pixel & mask_and) == 0) { pixel = src_pixel | mask_or }
u32 bits; u32 bits;
BitField<u32, u8, 0, 4> texture_page_x_base; BitField<u32, u8, 0, 4> texture_page_x_base;
BitField<u32, u8, 4, 1> texture_page_y_base; BitField<u32, u8, 4, 1> texture_page_y_base;
@ -459,18 +461,6 @@ protected:
static constexpr u32 ACTIVE = (1 << 19) | (1 << 22); static constexpr u32 ACTIVE = (1 << 19) | (1 << 22);
return ((bits & ACTIVE) == ACTIVE); return ((bits & ACTIVE) == ACTIVE);
} }
// During transfer/render operations, if ((dst_pixel & mask_and) == 0) { pixel = src_pixel | mask_or }
ALWAYS_INLINE u16 GetMaskAND() const
{
// return check_mask_before_draw ? 0x8000 : 0x0000;
return Truncate16((bits << 3) & 0x8000);
}
ALWAYS_INLINE u16 GetMaskOR() const
{
// return set_mask_while_drawing ? 0x8000 : 0x0000;
return Truncate16((bits << 4) & 0x8000);
}
} m_GPUSTAT = {}; } m_GPUSTAT = {};
struct DrawMode struct DrawMode

1
src/core/gpu_hw.cpp
View File

@ -6,6 +6,7 @@
#include "cpu_pgxp.h" #include "cpu_pgxp.h"
#include "gpu_hw_shadergen.h" #include "gpu_hw_shadergen.h"
#include "gpu_sw_backend.h" #include "gpu_sw_backend.h"
#include "gpu_sw_rasterizer.h"
#include "host.h" #include "host.h"
#include "settings.h" #include "settings.h"
#include "system.h" #include "system.h"

176
src/core/gpu_sw_backend.cpp
View File

@ -16,8 +16,6 @@ GPU_SW_Backend::~GPU_SW_Backend() = default;
bool GPU_SW_Backend::Initialize(bool force_thread) bool GPU_SW_Backend::Initialize(bool force_thread)
{ {
GPU_SW_Rasterizer::SelectImplementation();
return GPUBackend::Initialize(force_thread); return GPUBackend::Initialize(force_thread);
} }
@ -59,186 +57,20 @@ void GPU_SW_Backend::DrawLine(const GPUBackendDrawLineCommand* cmd)
void GPU_SW_Backend::FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color, GPUBackendCommandParameters params) void GPU_SW_Backend::FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color, GPUBackendCommandParameters params)
{ {
const u16 color16 = VRAMRGBA8888ToRGBA5551(color); GPU_SW_Rasterizer::FillVRAM(x, y, width, height, color, params.interlaced_rendering, params.active_line_lsb);
const GSVector4i fill = GSVector4i(color16, color16, color16, color16, color16, color16, color16, color16);
constexpr u32 vector_width = 8;
const u32 aligned_width = Common::AlignDownPow2(width, vector_width);
if ((x + width) <= VRAM_WIDTH && !params.interlaced_rendering)
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
u16* row_ptr = &g_vram[row * VRAM_WIDTH + x];
u32 xoffs = 0;
for (; xoffs < aligned_width; xoffs += vector_width, row_ptr += vector_width)
GSVector4i::store<false>(row_ptr, fill);
for (; xoffs < width; xoffs++)
*(row_ptr++) = color16;
}
}
else if (params.interlaced_rendering)
{
// Hardware tests show that fills seem to break on the first two lines when the offset matches the displayed field.
const u32 active_field = params.active_line_lsb;
if ((x + width) <= VRAM_WIDTH)
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
if ((row & u32(1)) == active_field)
continue;
u16* row_ptr = &g_vram[row * VRAM_WIDTH + x];
u32 xoffs = 0;
for (; xoffs < aligned_width; xoffs += vector_width, row_ptr += vector_width)
GSVector4i::store<false>(row_ptr, fill);
for (; xoffs < width; xoffs++)
*(row_ptr++) = color16;
}
}
else
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
if ((row & u32(1)) == active_field)
continue;
u16* row_ptr = &g_vram[row * VRAM_WIDTH];
for (u32 xoffs = 0; xoffs < width; xoffs++)
{
const u32 col = (x + xoffs) % VRAM_WIDTH;
row_ptr[col] = color16;
}
}
}
}
else
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
u16* row_ptr = &g_vram[row * VRAM_WIDTH];
for (u32 xoffs = 0; xoffs < width; xoffs++)
{
const u32 col = (x + xoffs) % VRAM_WIDTH;
row_ptr[col] = color16;
}
}
}
} }
void GPU_SW_Backend::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, void GPU_SW_Backend::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data,
GPUBackendCommandParameters params) GPUBackendCommandParameters params)
{ {
// Fast path when the copy is not oversized. GPU_SW_Rasterizer::WriteVRAM(x, y, width, height, data, params.set_mask_while_drawing, params.check_mask_before_draw);
if ((x + width) <= VRAM_WIDTH && (y + height) <= VRAM_HEIGHT && !params.IsMaskingEnabled())
{
const u16* src_ptr = static_cast<const u16*>(data);
u16* dst_ptr = &g_vram[y * VRAM_WIDTH + x];
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
std::copy_n(src_ptr, width, dst_ptr);
src_ptr += width;
dst_ptr += VRAM_WIDTH;
}
}
else
{
// Slow path when we need to handle wrap-around.
const u16* src_ptr = static_cast<const u16*>(data);
const u16 mask_and = params.GetMaskAND();
const u16 mask_or = params.GetMaskOR();
for (u32 row = 0; row < height;)
{
u16* dst_row_ptr = &g_vram[((y + row++) % VRAM_HEIGHT) * VRAM_WIDTH];
for (u32 col = 0; col < width;)
{
// TODO: Handle unaligned reads...
u16* pixel_ptr = &dst_row_ptr[(x + col++) % VRAM_WIDTH];
if (((*pixel_ptr) & mask_and) == 0)
*pixel_ptr = *(src_ptr++) | mask_or;
}
}
}
} }
void GPU_SW_Backend::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height, void GPU_SW_Backend::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height,
GPUBackendCommandParameters params) GPUBackendCommandParameters params)
{ {
// Break up oversized copies. This behavior has not been verified on console. GPU_SW_Rasterizer::CopyVRAM(src_x, src_y, dst_x, dst_y, width, height, params.set_mask_while_drawing,
if ((src_x + width) > VRAM_WIDTH || (dst_x + width) > VRAM_WIDTH) params.check_mask_before_draw);
{
u32 remaining_rows = height;
u32 current_src_y = src_y;
u32 current_dst_y = dst_y;
while (remaining_rows > 0)
{
const u32 rows_to_copy =
std::min<u32>(remaining_rows, std::min<u32>(VRAM_HEIGHT - current_src_y, VRAM_HEIGHT - current_dst_y));
u32 remaining_columns = width;
u32 current_src_x = src_x;
u32 current_dst_x = dst_x;
while (remaining_columns > 0)
{
const u32 columns_to_copy =
std::min<u32>(remaining_columns, std::min<u32>(VRAM_WIDTH - current_src_x, VRAM_WIDTH - current_dst_x));
CopyVRAM(current_src_x, current_src_y, current_dst_x, current_dst_y, columns_to_copy, rows_to_copy, params);
current_src_x = (current_src_x + columns_to_copy) % VRAM_WIDTH;
current_dst_x = (current_dst_x + columns_to_copy) % VRAM_WIDTH;
remaining_columns -= columns_to_copy;
}
current_src_y = (current_src_y + rows_to_copy) % VRAM_HEIGHT;
current_dst_y = (current_dst_y + rows_to_copy) % VRAM_HEIGHT;
remaining_rows -= rows_to_copy;
}
return;
}
// This doesn't have a fast path, but do we really need one? It's not common.
const u16 mask_and = params.GetMaskAND();
const u16 mask_or = params.GetMaskOR();
// Copy in reverse when src_x < dst_x, this is verified on console.
if (src_x < dst_x || ((src_x + width - 1) % VRAM_WIDTH) < ((dst_x + width - 1) % VRAM_WIDTH))
{
for (u32 row = 0; row < height; row++)
{
const u16* src_row_ptr = &g_vram[((src_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
u16* dst_row_ptr = &g_vram[((dst_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
for (s32 col = static_cast<s32>(width - 1); col >= 0; col--)
{
const u16 src_pixel = src_row_ptr[(src_x + static_cast<u32>(col)) % VRAM_WIDTH];
u16* dst_pixel_ptr = &dst_row_ptr[(dst_x + static_cast<u32>(col)) % VRAM_WIDTH];
if ((*dst_pixel_ptr & mask_and) == 0)
*dst_pixel_ptr = src_pixel | mask_or;
}
}
}
else
{
for (u32 row = 0; row < height; row++)
{
const u16* src_row_ptr = &g_vram[((src_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
u16* dst_row_ptr = &g_vram[((dst_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
for (u32 col = 0; col < width; col++)
{
const u16 src_pixel = src_row_ptr[(src_x + col) % VRAM_WIDTH];
u16* dst_pixel_ptr = &dst_row_ptr[(dst_x + col) % VRAM_WIDTH];
if ((*dst_pixel_ptr & mask_and) == 0)
*dst_pixel_ptr = src_pixel | mask_or;
}
}
}
} }
void GPU_SW_Backend::UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) void GPU_SW_Backend::UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit)

44
src/core/gpu_sw_rasterizer.cpp
View File

@ -12,11 +12,6 @@
LOG_CHANNEL(GPU_SW_Rasterizer); LOG_CHANNEL(GPU_SW_Rasterizer);
namespace GPU_SW_Rasterizer { namespace GPU_SW_Rasterizer {
// Default implementation, compatible with all ISAs.
extern const DrawRectangleFunctionTable DrawRectangleFunctions;
extern const DrawTriangleFunctionTable DrawTriangleFunctions;
extern const DrawLineFunctionTable DrawLineFunctions;
constinit const DitherLUT g_dither_lut = []() constexpr { constinit const DitherLUT g_dither_lut = []() constexpr {
DitherLUT lut = {}; DitherLUT lut = {};
for (u32 i = 0; i < DITHER_MATRIX_SIZE; i++) for (u32 i = 0; i < DITHER_MATRIX_SIZE; i++)
@ -33,30 +28,33 @@ constinit const DitherLUT g_dither_lut = []() constexpr {
return lut; return lut;
}(); }();
const DrawRectangleFunctionTable* DrawRectangleFunctions = nullptr;
const DrawTriangleFunctionTable* DrawTriangleFunctions = nullptr;
const DrawLineFunctionTable* DrawLineFunctions = nullptr;
FillVRAMFunction FillVRAM = nullptr;
WriteVRAMFunction WriteVRAM = nullptr;
CopyVRAMFunction CopyVRAM = nullptr;
GPUDrawingArea g_drawing_area = {}; GPUDrawingArea g_drawing_area = {};
} // namespace GPU_SW_Rasterizer } // namespace GPU_SW_Rasterizer
// Default implementation definitions. // Default scalar implementation definitions.
namespace GPU_SW_Rasterizer { namespace GPU_SW_Rasterizer::Scalar {
namespace {
#include "gpu_sw_rasterizer.inl" #include "gpu_sw_rasterizer.inl"
} }
} // namespace GPU_SW_Rasterizer::Scalar
// Default vector implementation definitions. // Default vector implementation definitions.
#if defined(CPU_ARCH_SSE) || defined(CPU_ARCH_NEON) #if defined(CPU_ARCH_SSE) || defined(CPU_ARCH_NEON)
namespace GPU_SW_Rasterizer::SIMD { namespace GPU_SW_Rasterizer::SIMD {
namespace {
#define USE_VECTOR 1 #define USE_VECTOR 1
#include "gpu_sw_rasterizer.inl" #include "gpu_sw_rasterizer.inl"
#undef USE_VECTOR #undef USE_VECTOR
} // namespace
} // namespace GPU_SW_Rasterizer::SIMD } // namespace GPU_SW_Rasterizer::SIMD
#endif #endif
// Initialize with default implementation.
namespace GPU_SW_Rasterizer {
const DrawRectangleFunctionTable* SelectedDrawRectangleFunctions = &DrawRectangleFunctions;
const DrawTriangleFunctionTable* SelectedDrawTriangleFunctions = &DrawTriangleFunctions;
const DrawLineFunctionTable* SelectedDrawLineFunctions = &DrawLineFunctions;
} // namespace GPU_SW_Rasterizer
// Declare alternative implementations. // Declare alternative implementations.
void GPU_SW_Rasterizer::SelectImplementation() void GPU_SW_Rasterizer::SelectImplementation()
{ {
@ -66,13 +64,16 @@ void GPU_SW_Rasterizer::SelectImplementation()
selected = true; selected = true;
#define SELECT_ALTERNATIVE_RASTERIZER(isa) \ #define SELECT_IMPLEMENTATION(isa) \
do \ do \
{ \ { \
INFO_LOG("Using " #isa " software rasterizer implementation."); \ INFO_LOG("Using " #isa " software rasterizer implementation."); \
SelectedDrawRectangleFunctions = &isa::DrawRectangleFunctions; \ DrawRectangleFunctions = &isa::DrawRectangleFunctions; \
SelectedDrawTriangleFunctions = &isa::DrawTriangleFunctions; \ DrawTriangleFunctions = &isa::DrawTriangleFunctions; \
SelectedDrawLineFunctions = &isa::DrawLineFunctions; \ DrawLineFunctions = &isa::DrawLineFunctions; \
FillVRAM = &isa::FillVRAMImpl; \
WriteVRAM = &isa::WriteVRAMImpl; \
CopyVRAM = &isa::CopyVRAMImpl; \
} while (0) } while (0)
#if defined(CPU_ARCH_SSE) || defined(CPU_ARCH_NEON) #if defined(CPU_ARCH_SSE) || defined(CPU_ARCH_NEON)
@ -83,19 +84,20 @@ void GPU_SW_Rasterizer::SelectImplementation()
#if defined(CPU_ARCH_SSE) && defined(_MSC_VER) && 0 #if defined(CPU_ARCH_SSE) && defined(_MSC_VER) && 0
if (cpuinfo_has_x86_avx2() && (!use_isa || StringUtil::Strcasecmp(use_isa, "AVX2") == 0)) if (cpuinfo_has_x86_avx2() && (!use_isa || StringUtil::Strcasecmp(use_isa, "AVX2") == 0))
{ {
SELECT_ALTERNATIVE_RASTERIZER(AVX2); SELECT_IMPLEMENTATION(AVX2);
return; return;
} }
#endif #endif
if (!use_isa || StringUtil::Strcasecmp(use_isa, "SIMD") == 0) if (!use_isa || StringUtil::Strcasecmp(use_isa, "SIMD") == 0)
{ {
SELECT_ALTERNATIVE_RASTERIZER(SIMD); SELECT_IMPLEMENTATION(SIMD);
return; return;
} }
#endif #endif
INFO_LOG("Using scalar software rasterizer implementation."); INFO_LOG("Using scalar software rasterizer implementation.");
SELECT_IMPLEMENTATION(Scalar);
#undef SELECT_ALTERNATIVE_RASTERIZER #undef SELECT_IMPLEMENTATION
} }

26
src/core/gpu_sw_rasterizer.h
View File

@ -34,34 +34,38 @@ using DrawLineFunction = void (*)(const GPUBackendDrawLineCommand* cmd, const GP
const GPUBackendDrawLineCommand::Vertex* p1); const GPUBackendDrawLineCommand::Vertex* p1);
typedef const DrawLineFunction DrawLineFunctionTable[2][2]; typedef const DrawLineFunction DrawLineFunctionTable[2][2];
// Default implementation, compatible with all ISAs. using FillVRAMFunction = void (*)(u32 x, u32 y, u32 width, u32 height, u32 color, bool interlaced, u8 active_line_lsb);
extern const DrawRectangleFunctionTable DrawRectangleFunctions; using WriteVRAMFunction = void (*)(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask,
extern const DrawTriangleFunctionTable DrawTriangleFunctions; bool check_mask);
extern const DrawLineFunctionTable DrawLineFunctions; using CopyVRAMFunction = void (*)(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height, bool set_mask,
bool check_mask);
// Current implementation, selected at runtime. // Current implementation, selected at runtime.
extern const DrawRectangleFunctionTable* SelectedDrawRectangleFunctions; extern const DrawRectangleFunctionTable* DrawRectangleFunctions;
extern const DrawTriangleFunctionTable* SelectedDrawTriangleFunctions; extern const DrawTriangleFunctionTable* DrawTriangleFunctions;
extern const DrawLineFunctionTable* SelectedDrawLineFunctions; extern const DrawLineFunctionTable* DrawLineFunctions;
extern FillVRAMFunction FillVRAM;
extern WriteVRAMFunction WriteVRAM;
extern CopyVRAMFunction CopyVRAM;
extern void SelectImplementation(); extern void SelectImplementation();
ALWAYS_INLINE static DrawLineFunction GetDrawLineFunction(bool shading_enable, bool transparency_enable) ALWAYS_INLINE static DrawLineFunction GetDrawLineFunction(bool shading_enable, bool transparency_enable)
{ {
return (*SelectedDrawLineFunctions)[u8(shading_enable)][u8(transparency_enable)]; return (*DrawLineFunctions)[u8(shading_enable)][u8(transparency_enable)];
} }
ALWAYS_INLINE static DrawRectangleFunction GetDrawRectangleFunction(bool texture_enable, bool raw_texture_enable, ALWAYS_INLINE static DrawRectangleFunction GetDrawRectangleFunction(bool texture_enable, bool raw_texture_enable,
bool transparency_enable) bool transparency_enable)
{ {
return (*SelectedDrawRectangleFunctions)[u8(texture_enable)][u8(raw_texture_enable)][u8(transparency_enable)]; return (*DrawRectangleFunctions)[u8(texture_enable)][u8(raw_texture_enable)][u8(transparency_enable)];
} }
ALWAYS_INLINE static DrawTriangleFunction GetDrawTriangleFunction(bool shading_enable, bool texture_enable, ALWAYS_INLINE static DrawTriangleFunction GetDrawTriangleFunction(bool shading_enable, bool texture_enable,
bool raw_texture_enable, bool transparency_enable) bool raw_texture_enable, bool transparency_enable)
{ {
return (*SelectedDrawTriangleFunctions)[u8(shading_enable)][u8(texture_enable)][u8(raw_texture_enable)] return (
[u8(transparency_enable)]; *DrawTriangleFunctions)[u8(shading_enable)][u8(texture_enable)][u8(raw_texture_enable)][u8(transparency_enable)];
} }
#define DECLARE_ALTERNATIVE_RASTERIZER(isa) \ #define DECLARE_ALTERNATIVE_RASTERIZER(isa) \

231
src/core/gpu_sw_rasterizer.inl
View File

@ -1519,6 +1519,237 @@ constinit const DrawTriangleFunctionTable DrawTriangleFunctions = {
{{&DrawTriangle<true, true, false, false>, &DrawTriangle<true, true, false, true>}, {{&DrawTriangle<true, true, false, false>, &DrawTriangle<true, true, false, true>},
{&DrawTriangle<true, true, true, false>, &DrawTriangle<true, true, true, true>}}}}; {&DrawTriangle<true, true, true, false>, &DrawTriangle<true, true, true, true>}}}};
static void FillVRAMImpl(u32 x, u32 y, u32 width, u32 height, u32 color, bool interlaced, u8 active_line_lsb)
{
#ifdef USE_VECTOR
const u16 color16 = VRAMRGBA8888ToRGBA5551(color);
const GSVector4i fill = GSVector4i(color16, color16, color16, color16, color16, color16, color16, color16);
constexpr u32 vector_width = 8;
const u32 aligned_width = Common::AlignDownPow2(width, vector_width);
if ((x + width) <= VRAM_WIDTH && !interlaced)
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
u16* row_ptr = &g_vram[row * VRAM_WIDTH + x];
u32 xoffs = 0;
for (; xoffs < aligned_width; xoffs += vector_width, row_ptr += vector_width)
GSVector4i::store<false>(row_ptr, fill);
for (; xoffs < width; xoffs++)
*(row_ptr++) = color16;
}
}
else if (interlaced)
{
// Hardware tests show that fills seem to break on the first two lines when the offset matches the displayed field.
const u32 active_field = active_line_lsb;
if ((x + width) <= VRAM_WIDTH)
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
if ((row & u32(1)) == active_field)
continue;
u16* row_ptr = &g_vram[row * VRAM_WIDTH + x];
u32 xoffs = 0;
for (; xoffs < aligned_width; xoffs += vector_width, row_ptr += vector_width)
GSVector4i::store<false>(row_ptr, fill);
for (; xoffs < width; xoffs++)
*(row_ptr++) = color16;
}
}
else
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
if ((row & u32(1)) == active_field)
continue;
u16* row_ptr = &g_vram[row * VRAM_WIDTH];
for (u32 xoffs = 0; xoffs < width; xoffs++)
{
const u32 col = (x + xoffs) % VRAM_WIDTH;
row_ptr[col] = color16;
}
}
}
}
else
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
u16* row_ptr = &g_vram[row * VRAM_WIDTH];
for (u32 xoffs = 0; xoffs < width; xoffs++)
{
const u32 col = (x + xoffs) % VRAM_WIDTH;
row_ptr[col] = color16;
}
}
}
#else
const u16 color16 = VRAMRGBA8888ToRGBA5551(color);
if ((x + width) <= VRAM_WIDTH && !interlaced)
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
std::fill_n(&g_vram[row * VRAM_WIDTH + x], width, color16);
}
}
else if (interlaced)
{
// Hardware tests show that fills seem to break on the first two lines when the offset matches the displayed field.
const u32 active_field = active_line_lsb;
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
if ((row & u32(1)) == active_field)
continue;
u16* row_ptr = &g_vram[row * VRAM_WIDTH];
for (u32 xoffs = 0; xoffs < width; xoffs++)
{
const u32 col = (x + xoffs) % VRAM_WIDTH;
row_ptr[col] = color16;
}
}
}
else
{
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
const u32 row = (y + yoffs) % VRAM_HEIGHT;
u16* row_ptr = &g_vram[row * VRAM_WIDTH];
for (u32 xoffs = 0; xoffs < width; xoffs++)
{
const u32 col = (x + xoffs) % VRAM_WIDTH;
row_ptr[col] = color16;
}
}
}
#endif
}
static void WriteVRAMImpl(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask)
{
// TODO: Vector implementation
// Fast path when the copy is not oversized.
if ((x + width) <= VRAM_WIDTH && (y + height) <= VRAM_HEIGHT && !set_mask && !check_mask)
{
const u16* src_ptr = static_cast<const u16*>(data);
u16* dst_ptr = &g_vram[y * VRAM_WIDTH + x];
for (u32 yoffs = 0; yoffs < height; yoffs++)
{
std::copy_n(src_ptr, width, dst_ptr);
src_ptr += width;
dst_ptr += VRAM_WIDTH;
}
}
else
{
// Slow path when we need to handle wrap-around.
// During transfer/render operations, if ((dst_pixel & mask_and) == 0) { pixel = src_pixel | mask_or }
const u16* src_ptr = static_cast<const u16*>(data);
const u16 mask_and = check_mask ? 0x8000u : 0x0000u;
const u16 mask_or = set_mask ? 0x8000u : 0x0000u;
for (u32 row = 0; row < height;)
{
u16* dst_row_ptr = &g_vram[((y + row++) % VRAM_HEIGHT) * VRAM_WIDTH];
for (u32 col = 0; col < width;)
{
// TODO: Handle unaligned reads...
u16* pixel_ptr = &dst_row_ptr[(x + col++) % VRAM_WIDTH];
if (((*pixel_ptr) & mask_and) == 0)
*pixel_ptr = *(src_ptr++) | mask_or;
}
}
}
}
static void CopyVRAMImpl(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height, bool set_mask,
bool check_mask)
{
// TODO: Vector implementation.
// Break up oversized copies. This behavior has not been verified on console.
if ((src_x + width) > VRAM_WIDTH || (dst_x + width) > VRAM_WIDTH)
{
u32 remaining_rows = height;
u32 current_src_y = src_y;
u32 current_dst_y = dst_y;
while (remaining_rows > 0)
{
const u32 rows_to_copy =
std::min<u32>(remaining_rows, std::min<u32>(VRAM_HEIGHT - current_src_y, VRAM_HEIGHT - current_dst_y));
u32 remaining_columns = width;
u32 current_src_x = src_x;
u32 current_dst_x = dst_x;
while (remaining_columns > 0)
{
const u32 columns_to_copy =
std::min<u32>(remaining_columns, std::min<u32>(VRAM_WIDTH - current_src_x, VRAM_WIDTH - current_dst_x));
CopyVRAM(current_src_x, current_src_y, current_dst_x, current_dst_y, columns_to_copy, rows_to_copy, set_mask,
check_mask);
current_src_x = (current_src_x + columns_to_copy) % VRAM_WIDTH;
current_dst_x = (current_dst_x + columns_to_copy) % VRAM_WIDTH;
remaining_columns -= columns_to_copy;
}
current_src_y = (current_src_y + rows_to_copy) % VRAM_HEIGHT;
current_dst_y = (current_dst_y + rows_to_copy) % VRAM_HEIGHT;
remaining_rows -= rows_to_copy;
}
return;
}
// This doesn't have a fast path, but do we really need one? It's not common.
const u16 mask_and = check_mask ? 0x8000u : 0x0000u;
const u16 mask_or = set_mask ? 0x8000u : 0x0000u;
// Copy in reverse when src_x < dst_x, this is verified on console.
if (src_x < dst_x || ((src_x + width - 1) % VRAM_WIDTH) < ((dst_x + width - 1) % VRAM_WIDTH))
{
for (u32 row = 0; row < height; row++)
{
const u16* src_row_ptr = &g_vram[((src_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
u16* dst_row_ptr = &g_vram[((dst_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
for (s32 col = static_cast<s32>(width - 1); col >= 0; col--)
{
const u16 src_pixel = src_row_ptr[(src_x + static_cast<u32>(col)) % VRAM_WIDTH];
u16* dst_pixel_ptr = &dst_row_ptr[(dst_x + static_cast<u32>(col)) % VRAM_WIDTH];
*dst_pixel_ptr = ((*dst_pixel_ptr & mask_and) == 0) ? (src_pixel | mask_or) : *dst_pixel_ptr;
}
}
}
else
{
for (u32 row = 0; row < height; row++)
{
const u16* src_row_ptr = &g_vram[((src_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
u16* dst_row_ptr = &g_vram[((dst_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
for (u32 col = 0; col < width; col++)
{
const u16 src_pixel = src_row_ptr[(src_x + col) % VRAM_WIDTH];
u16* dst_pixel_ptr = &dst_row_ptr[(dst_x + col) % VRAM_WIDTH];
*dst_pixel_ptr = ((*dst_pixel_ptr & mask_and) == 0) ? (src_pixel | mask_or) : *dst_pixel_ptr;
}
}
}
}
#ifdef __INTELLISENSE__ #ifdef __INTELLISENSE__
} }
#endif #endif

2
src/core/gpu_types.h
View File

@ -275,8 +275,6 @@ union GPUBackendCommandParameters
BitField<u8, bool, 2, 1> set_mask_while_drawing; BitField<u8, bool, 2, 1> set_mask_while_drawing;
BitField<u8, bool, 3, 1> check_mask_before_draw; BitField<u8, bool, 3, 1> check_mask_before_draw;
ALWAYS_INLINE bool IsMaskingEnabled() const { return (bits & 12u) != 0u; }
// During transfer/render operations, if ((dst_pixel & mask_and) == 0) { pixel = src_pixel | mask_or } // During transfer/render operations, if ((dst_pixel & mask_and) == 0) { pixel = src_pixel | mask_or }
u16 GetMaskAND() const u16 GetMaskAND() const
{ {