Merge pull request #2192 from Tilka/sse2

VertexLoaderX64: support SSE2 as a fallback
This commit is contained in:
skidau 2015-03-21 12:58:24 +11:00
commit 39c41f5c70
10 changed files with 326 additions and 228 deletions

View File

@ -1638,6 +1638,7 @@ void XEmitter::PACKUSWB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x67, dest, ar
void XEmitter::PUNPCKLBW(X64Reg dest, const OpArg &arg) {WriteSSEOp(0x66, 0x60, dest, arg);} void XEmitter::PUNPCKLBW(X64Reg dest, const OpArg &arg) {WriteSSEOp(0x66, 0x60, dest, arg);}
void XEmitter::PUNPCKLWD(X64Reg dest, const OpArg &arg) {WriteSSEOp(0x66, 0x61, dest, arg);} void XEmitter::PUNPCKLWD(X64Reg dest, const OpArg &arg) {WriteSSEOp(0x66, 0x61, dest, arg);}
void XEmitter::PUNPCKLDQ(X64Reg dest, const OpArg &arg) {WriteSSEOp(0x66, 0x62, dest, arg);} void XEmitter::PUNPCKLDQ(X64Reg dest, const OpArg &arg) {WriteSSEOp(0x66, 0x62, dest, arg);}
void XEmitter::PUNPCKLQDQ(X64Reg dest, const OpArg &arg) {WriteSSEOp(0x66, 0x6C, dest, arg);}
void XEmitter::PSRLW(X64Reg reg, int shift) void XEmitter::PSRLW(X64Reg reg, int shift)
{ {

View File

@ -680,6 +680,7 @@ public:
void PUNPCKLBW(X64Reg dest, const OpArg &arg); void PUNPCKLBW(X64Reg dest, const OpArg &arg);
void PUNPCKLWD(X64Reg dest, const OpArg &arg); void PUNPCKLWD(X64Reg dest, const OpArg &arg);
void PUNPCKLDQ(X64Reg dest, const OpArg &arg); void PUNPCKLDQ(X64Reg dest, const OpArg &arg);
void PUNPCKLQDQ(X64Reg dest, const OpArg &arg);
void PTEST(X64Reg dest, OpArg arg); void PTEST(X64Reg dest, OpArg arg);
void PAND(X64Reg dest, OpArg arg); void PAND(X64Reg dest, OpArg arg);

View File

@ -100,16 +100,17 @@ static T ReadNormalized(I value)
} }
template <typename T, bool swap = false> template <typename T, bool swap = false>
static void ReadVertexAttribute(T* dst, DataReader src, const AttributeFormat& format, int base_component, int max_components, bool reverse) static void ReadVertexAttribute(T* dst, DataReader src, const AttributeFormat& format, int base_component, int components, bool reverse)
{ {
if (format.enable) if (format.enable)
{ {
src.Skip(format.offset); src.Skip(format.offset);
src.Skip(base_component * (1<<(format.type>>1))); src.Skip(base_component * (1<<(format.type>>1)));
for (int i = 0; i < std::min(format.components - base_component, max_components); i++) int i;
for (i = 0; i < std::min(format.components - base_component, components); i++)
{ {
int i_dst = reverse ? max_components - i - 1 : i; int i_dst = reverse ? components - i - 1 : i;
switch (format.type) switch (format.type)
{ {
case VAR_UNSIGNED_BYTE: case VAR_UNSIGNED_BYTE:
@ -131,6 +132,11 @@ static void ReadVertexAttribute(T* dst, DataReader src, const AttributeFormat& f
_assert_msg_(VIDEO, !format.integer || format.type != VAR_FLOAT, "only non-float values are allowed to be streamed as integer"); _assert_msg_(VIDEO, !format.integer || format.type != VAR_FLOAT, "only non-float values are allowed to be streamed as integer");
} }
for (; i < components; i++)
{
int i_dst = reverse ? components - i - 1 : i;
dst[i_dst] = i == 3;
}
} }
} }

View File

@ -131,12 +131,13 @@ void VertexLoader::CompileVertexTranslator()
WriteCall(VertexLoader_Position::GetFunction(m_VtxDesc.Position, m_VtxAttr.PosFormat, m_VtxAttr.PosElements)); WriteCall(VertexLoader_Position::GetFunction(m_VtxDesc.Position, m_VtxAttr.PosFormat, m_VtxAttr.PosElements));
m_VertexSize += VertexLoader_Position::GetSize(m_VtxDesc.Position, m_VtxAttr.PosFormat, m_VtxAttr.PosElements); m_VertexSize += VertexLoader_Position::GetSize(m_VtxDesc.Position, m_VtxAttr.PosFormat, m_VtxAttr.PosElements);
m_native_vtx_decl.position.components = 3; int pos_elements = m_VtxAttr.PosElements + 2;
m_native_vtx_decl.position.components = pos_elements;
m_native_vtx_decl.position.enable = true; m_native_vtx_decl.position.enable = true;
m_native_vtx_decl.position.offset = nat_offset; m_native_vtx_decl.position.offset = nat_offset;
m_native_vtx_decl.position.type = VAR_FLOAT; m_native_vtx_decl.position.type = VAR_FLOAT;
m_native_vtx_decl.position.integer = false; m_native_vtx_decl.position.integer = false;
nat_offset += 12; nat_offset += pos_elements * sizeof(float);
// Normals // Normals
if (m_VtxDesc.Normal != NOT_PRESENT) if (m_VtxDesc.Normal != NOT_PRESENT)

View File

@ -370,8 +370,9 @@ void VertexLoaderARM64::GenerateVertexLoader()
load_size <<= 3; load_size <<= 3;
s32 offset = GetAddressImm(ARRAY_POSITION, m_VtxDesc.Position, EncodeRegTo64(scratch1_reg), load_size); s32 offset = GetAddressImm(ARRAY_POSITION, m_VtxDesc.Position, EncodeRegTo64(scratch1_reg), load_size);
ReadVertex(m_VtxDesc.Position, m_VtxAttr.PosFormat, m_VtxAttr.PosElements + 2, 3, int pos_elements = m_VtxAttr.PosElements + 2;
m_VtxAttr.ByteDequant, m_VtxAttr.PosFrac, &m_native_vtx_decl.position, offset); ReadVertex(m_VtxDesc.Position, m_VtxAttr.PosFormat, pos_elements, pos_elements,
m_VtxAttr.ByteDequant, m_VtxAttr.PosFrac, &m_native_vtx_decl.position, offset);
} }
if (m_VtxDesc.Normal) if (m_VtxDesc.Normal)

View File

@ -1,3 +1,4 @@
#include "Common/BitSet.h"
#include "Common/CPUDetect.h" #include "Common/CPUDetect.h"
#include "Common/Intrinsics.h" #include "Common/Intrinsics.h"
#include "Common/JitRegister.h" #include "Common/JitRegister.h"
@ -6,8 +7,6 @@
using namespace Gen; using namespace Gen;
#define VERTEX_LOADER_REGS {XMM0+16}
static const X64Reg src_reg = ABI_PARAM1; static const X64Reg src_reg = ABI_PARAM1;
static const X64Reg dst_reg = ABI_PARAM2; static const X64Reg dst_reg = ABI_PARAM2;
static const X64Reg scratch1 = RAX; static const X64Reg scratch1 = RAX;
@ -66,7 +65,7 @@ OpArg VertexLoaderX64::GetVertexAddr(int array, u64 attribute)
int VertexLoaderX64::ReadVertex(OpArg data, u64 attribute, int format, int count_in, int count_out, bool dequantize, u8 scaling_exponent, AttributeFormat* native_format) int VertexLoaderX64::ReadVertex(OpArg data, u64 attribute, int format, int count_in, int count_out, bool dequantize, u8 scaling_exponent, AttributeFormat* native_format)
{ {
static const __m128i shuffle_lut[5][3] = { static const __m128i shuffle_lut[4][3] = {
{_mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFF00L), // 1x u8 {_mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFF00L), // 1x u8
_mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFF01L, 0xFFFFFF00L), // 2x u8 _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFF01L, 0xFFFFFF00L), // 2x u8
_mm_set_epi32(0xFFFFFFFFL, 0xFFFFFF02L, 0xFFFFFF01L, 0xFFFFFF00L)}, // 3x u8 _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFF02L, 0xFFFFFF01L, 0xFFFFFF00L)}, // 3x u8
@ -79,9 +78,6 @@ int VertexLoaderX64::ReadVertex(OpArg data, u64 attribute, int format, int count
{_mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFFFFL, 0x0001FFFFL), // 1x s16 {_mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFFFFL, 0x0001FFFFL), // 1x s16
_mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0x0203FFFFL, 0x0001FFFFL), // 2x s16 _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0x0203FFFFL, 0x0001FFFFL), // 2x s16
_mm_set_epi32(0xFFFFFFFFL, 0x0405FFFFL, 0x0203FFFFL, 0x0001FFFFL)}, // 3x s16 _mm_set_epi32(0xFFFFFFFFL, 0x0405FFFFL, 0x0203FFFFL, 0x0001FFFFL)}, // 3x s16
{_mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFFFFL, 0x00010203L), // 1x float
_mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0x04050607L, 0x00010203L), // 2x float
_mm_set_epi32(0xFFFFFFFFL, 0x08090A0BL, 0x04050607L, 0x00010203L)}, // 3x float
}; };
static const __m128 scale_factors[32] = { static const __m128 scale_factors[32] = {
_mm_set_ps1(1./(1u<< 0)), _mm_set_ps1(1./(1u<< 1)), _mm_set_ps1(1./(1u<< 2)), _mm_set_ps1(1./(1u<< 3)), _mm_set_ps1(1./(1u<< 0)), _mm_set_ps1(1./(1u<< 1)), _mm_set_ps1(1./(1u<< 2)), _mm_set_ps1(1./(1u<< 3)),
@ -98,47 +94,109 @@ int VertexLoaderX64::ReadVertex(OpArg data, u64 attribute, int format, int count
int elem_size = 1 << (format / 2); int elem_size = 1 << (format / 2);
int load_bytes = elem_size * count_in; int load_bytes = elem_size * count_in;
if (load_bytes > 8)
MOVDQU(coords, data);
else if (load_bytes > 4)
MOVQ_xmm(coords, data);
else
MOVD_xmm(coords, data);
PSHUFB(coords, M(&shuffle_lut[format][count_in - 1]));
if (format != FORMAT_FLOAT)
{
// Sign extend
if (format == FORMAT_BYTE)
PSRAD(coords, 24);
if (format == FORMAT_SHORT)
PSRAD(coords, 16);
CVTDQ2PS(coords, R(coords));
if (dequantize && scaling_exponent)
MULPS(coords, M(&scale_factors[scaling_exponent]));
}
OpArg dest = MDisp(dst_reg, m_dst_ofs); OpArg dest = MDisp(dst_reg, m_dst_ofs);
switch (count_out)
{
case 1: MOVSS(dest, coords); break;
case 2: MOVLPS(dest, coords); break;
case 3: MOVUPS(dest, coords); break;
}
native_format->components = count_out; native_format->components = count_out;
native_format->enable = true; native_format->enable = true;
native_format->offset = m_dst_ofs; native_format->offset = m_dst_ofs;
native_format->type = VAR_FLOAT; native_format->type = VAR_FLOAT;
native_format->integer = false; native_format->integer = false;
m_dst_ofs += sizeof(float) * count_out; m_dst_ofs += sizeof(float) * count_out;
if (attribute == DIRECT) if (attribute == DIRECT)
m_src_ofs += load_bytes; m_src_ofs += load_bytes;
if (format == FORMAT_FLOAT)
{
// Floats don't need to be scaled or converted,
// so we can just load/swap/store them directly
// and return early.
for (int i = 0; i < count_in; i++)
{
LoadAndSwap(32, scratch3, data);
MOV(32, dest, R(scratch3));
data.offset += sizeof(float);
dest.offset += sizeof(float);
}
return load_bytes;
}
if (cpu_info.bSSSE3)
{
if (load_bytes > 8)
MOVDQU(coords, data);
else if (load_bytes > 4)
MOVQ_xmm(coords, data);
else
MOVD_xmm(coords, data);
PSHUFB(coords, M(&shuffle_lut[format][count_in - 1]));
// Sign-extend.
if (format == FORMAT_BYTE)
PSRAD(coords, 24);
if (format == FORMAT_SHORT)
PSRAD(coords, 16);
}
else
{
// SSE2
X64Reg temp = XMM1;
switch (format)
{
case FORMAT_UBYTE:
MOVD_xmm(coords, data);
PXOR(temp, R(temp));
PUNPCKLBW(coords, R(temp));
PUNPCKLWD(coords, R(temp));
break;
case FORMAT_BYTE:
MOVD_xmm(coords, data);
PUNPCKLBW(coords, R(coords));
PUNPCKLWD(coords, R(coords));
PSRAD(coords, 24);
break;
case FORMAT_USHORT:
case FORMAT_SHORT:
switch (count_in)
{
case 1:
LoadAndSwap(32, scratch3, data);
MOVD_xmm(coords, R(scratch3)); // ......X.
break;
case 2:
LoadAndSwap(32, scratch3, data);
MOVD_xmm(coords, R(scratch3)); // ......XY
PSHUFLW(coords, R(coords), 0x24); // ....Y.X.
break;
case 3:
LoadAndSwap(64, scratch3, data);
MOVQ_xmm(coords, R(scratch3)); // ....XYZ.
PUNPCKLQDQ(coords, R(coords)); // ..Z.XYZ.
PSHUFLW(coords, R(coords), 0xAC); // ..Z.Y.X.
break;
}
if (format == FORMAT_SHORT)
PSRAD(coords, 16);
else
PSRLD(coords, 16);
break;
}
}
CVTDQ2PS(coords, R(coords));
if (dequantize && scaling_exponent)
MULPS(coords, M(&scale_factors[scaling_exponent]));
switch (count_out)
{
case 1: MOVSS(dest, coords); break;
case 2: MOVLPS(dest, coords); break;
case 3: MOVUPS(dest, coords); break;
}
return load_bytes; return load_bytes;
} }
@ -290,7 +348,10 @@ void VertexLoaderX64::ReadColor(OpArg data, u64 attribute, int format)
void VertexLoaderX64::GenerateVertexLoader() void VertexLoaderX64::GenerateVertexLoader()
{ {
ABI_PushRegistersAndAdjustStack(VERTEX_LOADER_REGS, 8); BitSet32 xmm_regs;
xmm_regs[XMM0+16] = true;
xmm_regs[XMM1+16] = !cpu_info.bSSSE3;
ABI_PushRegistersAndAdjustStack(xmm_regs, 8);
// Backup count since we're going to count it down. // Backup count since we're going to count it down.
PUSH(32, R(ABI_PARAM3)); PUSH(32, R(ABI_PARAM3));
@ -332,7 +393,8 @@ void VertexLoaderX64::GenerateVertexLoader()
} }
OpArg data = GetVertexAddr(ARRAY_POSITION, m_VtxDesc.Position); OpArg data = GetVertexAddr(ARRAY_POSITION, m_VtxDesc.Position);
ReadVertex(data, m_VtxDesc.Position, m_VtxAttr.PosFormat, m_VtxAttr.PosElements + 2, 3, int pos_elements = 2 + m_VtxAttr.PosElements;
ReadVertex(data, m_VtxDesc.Position, m_VtxAttr.PosFormat, pos_elements, pos_elements,
m_VtxAttr.ByteDequant, m_VtxAttr.PosFrac, &m_native_vtx_decl.position); m_VtxAttr.ByteDequant, m_VtxAttr.PosFrac, &m_native_vtx_decl.position);
if (m_VtxDesc.Normal) if (m_VtxDesc.Normal)
@ -408,7 +470,7 @@ void VertexLoaderX64::GenerateVertexLoader()
m_native_vtx_decl.texcoords[i].offset = m_dst_ofs; m_native_vtx_decl.texcoords[i].offset = m_dst_ofs;
PXOR(XMM0, R(XMM0)); PXOR(XMM0, R(XMM0));
CVTSI2SS(XMM0, R(scratch1)); CVTSI2SS(XMM0, R(scratch1));
SHUFPS(XMM0, R(XMM0), 0x45); SHUFPS(XMM0, R(XMM0), 0x45); // 000X -> 0X00
MOVUPS(MDisp(dst_reg, m_dst_ofs), XMM0); MOVUPS(MDisp(dst_reg, m_dst_ofs), XMM0);
m_dst_ofs += sizeof(float) * 3; m_dst_ofs += sizeof(float) * 3;
} }
@ -426,7 +488,7 @@ void VertexLoaderX64::GenerateVertexLoader()
// Get the original count. // Get the original count.
POP(32, R(ABI_RETURN)); POP(32, R(ABI_RETURN));
ABI_PopRegistersAndAdjustStack(VERTEX_LOADER_REGS, 8); ABI_PopRegistersAndAdjustStack(xmm_regs, 8);
if (m_VtxDesc.Position & MASK_INDEXED) if (m_VtxDesc.Position & MASK_INDEXED)
{ {
@ -446,12 +508,6 @@ void VertexLoaderX64::GenerateVertexLoader()
m_native_vtx_decl.stride = m_dst_ofs; m_native_vtx_decl.stride = m_dst_ofs;
} }
bool VertexLoaderX64::IsInitialized()
{
// Uses PSHUFB.
return cpu_info.bSSSE3;
}
int VertexLoaderX64::RunVertices(DataReader src, DataReader dst, int count, int primitive) int VertexLoaderX64::RunVertices(DataReader src, DataReader dst, int count, int primitive)
{ {
m_numLoadedVertices += count; m_numLoadedVertices += count;

View File

@ -8,7 +8,7 @@ public:
protected: protected:
std::string GetName() const override { return "VertexLoaderX64"; } std::string GetName() const override { return "VertexLoaderX64"; }
bool IsInitialized() override; bool IsInitialized() override { return true; }
int RunVertices(DataReader src, DataReader dst, int count, int primitive) override; int RunVertices(DataReader src, DataReader dst, int count, int primitive) override;
private: private:

View File

@ -30,8 +30,8 @@ void LOADERDECL Pos_ReadDirect(VertexLoader* loader)
DataReader dst(g_vertex_manager_write_ptr, nullptr); DataReader dst(g_vertex_manager_write_ptr, nullptr);
DataReader src(g_video_buffer_read_ptr, nullptr); DataReader src(g_video_buffer_read_ptr, nullptr);
for (int i = 0; i < 3; ++i) for (int i = 0; i < N; ++i)
dst.Write(i < N ? PosScale(src.Read<T>(), scale) : 0.f); dst.Write(PosScale(src.Read<T>(), scale));
g_vertex_manager_write_ptr = dst.GetPointer(); g_vertex_manager_write_ptr = dst.GetPointer();
g_video_buffer_read_ptr = src.GetPointer(); g_video_buffer_read_ptr = src.GetPointer();
@ -50,8 +50,8 @@ void LOADERDECL Pos_ReadIndex(VertexLoader* loader)
auto const scale = loader->m_posScale; auto const scale = loader->m_posScale;
DataReader dst(g_vertex_manager_write_ptr, nullptr); DataReader dst(g_vertex_manager_write_ptr, nullptr);
for (int i = 0; i < 3; ++i) for (int i = 0; i < N; ++i)
dst.Write(i < N ? PosScale(Common::FromBigEndian(data[i]), scale) : 0.f); dst.Write(PosScale(Common::FromBigEndian(data[i]), scale));
g_vertex_manager_write_ptr = dst.GetPointer(); g_vertex_manager_write_ptr = dst.GetPointer();
LOG_VTX(); LOG_VTX();

View File

@ -810,6 +810,7 @@ TWO_OP_SSE_TEST(PACKUSWB, "dqword")
TWO_OP_SSE_TEST(PUNPCKLBW, "dqword") TWO_OP_SSE_TEST(PUNPCKLBW, "dqword")
TWO_OP_SSE_TEST(PUNPCKLWD, "dqword") TWO_OP_SSE_TEST(PUNPCKLWD, "dqword")
TWO_OP_SSE_TEST(PUNPCKLDQ, "dqword") TWO_OP_SSE_TEST(PUNPCKLDQ, "dqword")
TWO_OP_SSE_TEST(PUNPCKLQDQ, "dqword")
TWO_OP_SSE_TEST(PTEST, "dqword") TWO_OP_SSE_TEST(PTEST, "dqword")
TWO_OP_SSE_TEST(PAND, "dqword") TWO_OP_SSE_TEST(PAND, "dqword")

View File

@ -1,13 +1,18 @@
#include <limits>
#include <memory>
#include <tuple>
#include <type_traits>
#include <unordered_set> #include <unordered_set>
#include "Common/Common.h"
#include "VideoCommon/DataReader.h"
#include "VideoCommon/VertexLoaderBase.h"
// Needs to be included later because it defines a TEST macro that conflicts
// with a TEST method definition in x64Emitter.h.
#include <gtest/gtest.h> // NOLINT #include <gtest/gtest.h> // NOLINT
#include "Common/Common.h"
#include "Common/MathUtil.h"
#include "VideoCommon/CPMemory.h"
#include "VideoCommon/DataReader.h"
#include "VideoCommon/OpcodeDecoding.h"
#include "VideoCommon/VertexLoaderBase.h"
TEST(VertexLoaderUID, UniqueEnough) TEST(VertexLoaderUID, UniqueEnough)
{ {
std::unordered_set<VertexLoaderUID> uids; std::unordered_set<VertexLoaderUID> uids;
@ -38,181 +43,207 @@ protected:
void SetUp() override void SetUp() override
{ {
memset(&input_memory[0], 0, sizeof(input_memory)); memset(input_memory, 0, sizeof(input_memory));
memset(&output_memory[0], 0, sizeof(input_memory)); memset(output_memory, 0xFF, sizeof(input_memory));
memset(&m_vtx_desc, 0, sizeof(m_vtx_desc)); memset(&m_vtx_desc, 0, sizeof(m_vtx_desc));
memset(&m_vtx_attr, 0, sizeof(m_vtx_attr)); memset(&m_vtx_attr, 0, sizeof(m_vtx_attr));
m_loader = nullptr;
ResetPointers(); ResetPointers();
} }
// Pushes a value to the input stream. void CreateAndCheckSizes(size_t input_size, size_t output_size)
{
m_loader.reset(VertexLoaderBase::CreateVertexLoader(m_vtx_desc, m_vtx_attr));
ASSERT_EQ((int)input_size, m_loader->m_VertexSize);
ASSERT_EQ((int)output_size, m_loader->m_native_vtx_decl.stride);
}
template <typename T> template <typename T>
void Input(T val) void Input(T val)
{ {
// Converts *to* big endian, not from. // Write swapped.
*(T*)(&input_memory[m_input_pos]) = Common::FromBigEndian(val); m_src.Write<T, true>(val);
m_input_pos += sizeof(val);
} }
// Reads a value from the output stream. void ExpectOut(float val)
template <typename T>
T Output()
{ {
T out = *(T*)&output_memory[m_output_pos]; // Read unswapped.
m_output_pos += sizeof(out); MathUtil::IntFloat expected(val), actual(m_dst.Read<float, false>());
return out; if (!actual.f || actual.f != actual.f)
EXPECT_EQ(expected.i, actual.i);
else
EXPECT_EQ(expected.f, actual.f);
} }
// Combination of EXPECT_EQ and Output. void RunVertices(int count, int expected_count = -1)
template <typename T>
void ExpectOut(T val)
{ {
EXPECT_EQ(val, Output<T>()); if (expected_count == -1)
expected_count = count;
ResetPointers();
int actual_count = m_loader->RunVertices(m_src, m_dst, count, GX_DRAW_POINTS);
EXPECT_EQ(actual_count, expected_count);
} }
void ResetPointers() void ResetPointers()
{ {
m_input_pos = m_output_pos = 0; m_src = DataReader(input_memory, input_memory + sizeof(input_memory));
src = DataReader(input_memory, input_memory + sizeof(input_memory)); m_dst = DataReader(output_memory, output_memory + sizeof(output_memory));
dst = DataReader(output_memory, output_memory + sizeof(output_memory));
} }
u32 m_input_pos, m_output_pos; DataReader m_src;
DataReader src; DataReader m_dst;
DataReader dst;
TVtxDesc m_vtx_desc; TVtxDesc m_vtx_desc;
VAT m_vtx_attr; VAT m_vtx_attr;
std::unique_ptr<VertexLoaderBase> m_loader;
}; };
TEST_F(VertexLoaderTest, PositionDirectFloatXYZ) class VertexLoaderParamTest : public VertexLoaderTest, public ::testing::WithParamInterface<std::tuple<int, int, int, int>> {};
extern int gtest_AllCombinationsVertexLoaderParamTest_dummy_;
INSTANTIATE_TEST_CASE_P(
AllCombinations, VertexLoaderParamTest,
::testing::Combine(
::testing::Values(DIRECT, INDEX8, INDEX16),
::testing::Values(FORMAT_UBYTE, FORMAT_BYTE, FORMAT_USHORT, FORMAT_SHORT, FORMAT_FLOAT),
::testing::Values(0, 1), // elements
::testing::Values(0, 1, 31) // frac
)
);
TEST_P(VertexLoaderParamTest, PositionAll)
{ {
m_vtx_desc.Position = 1; // Direct int addr, format, elements, frac;
m_vtx_attr.g0.PosElements = 1; // XYZ std::tie(addr, format, elements, frac) = GetParam();
m_vtx_attr.g0.PosFormat = 4; // Float this->m_vtx_desc.Position = addr;
this->m_vtx_attr.g0.PosFormat = format;
this->m_vtx_attr.g0.PosElements = elements;
this->m_vtx_attr.g0.PosFrac = frac;
this->m_vtx_attr.g0.ByteDequant = true;
elements += 2;
VertexLoaderBase* loader = VertexLoaderBase::CreateVertexLoader(m_vtx_desc, m_vtx_attr); std::vector<float> values = {
std::numeric_limits<float>::lowest(),
std::numeric_limits<float>::denorm_min(),
std::numeric_limits<float>::min(),
std::numeric_limits<float>::max(),
std::numeric_limits<float>::quiet_NaN(),
std::numeric_limits<float>::infinity(),
-0x8000, -0x80, -1, -0, 0, 1, 123, 0x7F, 0xFF, 0x7FFF, 0xFFFF, 12345678,
};
ASSERT_EQ(0u, values.size() % 2);
ASSERT_EQ(0u, values.size() % 3);
ASSERT_EQ(3 * sizeof(float), (u32)loader->m_native_vtx_decl.stride); int count = (int)values.size() / elements;
ASSERT_EQ(3 * sizeof(float), (u32)loader->m_VertexSize); u32 elem_size = 1 << (format / 2);
size_t input_size = elements * elem_size;
// Write some vertices. if (addr & MASK_INDEXED)
Input(0.0f); Input(0.0f); Input(0.0f);
Input(1.0f); Input(0.0f); Input(0.0f);
Input(0.0f); Input(1.0f); Input(0.0f);
Input(0.0f); Input(0.0f); Input(1.0f);
// Convert 4 points. "7" -> primitive are points.
int count = loader->RunVertices(src, dst, 4, 7);
src.Skip(4 * loader->m_VertexSize);
dst.Skip(count * loader->m_native_vtx_decl.stride);
delete loader;
ExpectOut(0.0f); ExpectOut(0.0f); ExpectOut(0.0f);
ExpectOut(1.0f); ExpectOut(0.0f); ExpectOut(0.0f);
ExpectOut(0.0f); ExpectOut(1.0f); ExpectOut(0.0f);
ExpectOut(0.0f); ExpectOut(0.0f); ExpectOut(1.0f);
// Test that scale does nothing for floating point inputs.
Input(1.0f); Input(2.0f); Input(4.0f);
m_vtx_attr.g0.PosFrac = 1;
loader = VertexLoaderBase::CreateVertexLoader(m_vtx_desc, m_vtx_attr);
count = loader->RunVertices(src, dst, 1, 7);
src.Skip(1 * loader->m_VertexSize);
dst.Skip(count * loader->m_native_vtx_decl.stride);
ExpectOut(1.0f); ExpectOut(2.0f); ExpectOut(4.0f);
delete loader;
}
TEST_F(VertexLoaderTest, PositionDirectU16XY)
{
m_vtx_desc.Position = 1; // Direct
m_vtx_attr.g0.PosElements = 0; // XY
m_vtx_attr.g0.PosFormat = 2; // U16
VertexLoaderBase* loader = VertexLoaderBase::CreateVertexLoader(m_vtx_desc, m_vtx_attr);
ASSERT_EQ(3 * sizeof(float), (u32)loader->m_native_vtx_decl.stride);
ASSERT_EQ(2 * sizeof(u16), (u32)loader->m_VertexSize);
// Write some vertices.
Input<u16>(0); Input<u16>(0);
Input<u16>(1); Input<u16>(2);
Input<u16>(256); Input<u16>(257);
Input<u16>(65535); Input<u16>(65534);
Input<u16>(12345); Input<u16>(54321);
// Convert 5 points. "7" -> primitive are points.
int count = loader->RunVertices(src, dst, 5, 7);
src.Skip(5 * loader->m_VertexSize);
dst.Skip(count * loader->m_native_vtx_decl.stride);
delete loader;
ExpectOut(0.0f); ExpectOut(0.0f); ExpectOut(0.0f);
ExpectOut(1.0f); ExpectOut(2.0f); ExpectOut(0.0f);
ExpectOut(256.0f); ExpectOut(257.0f); ExpectOut(0.0f);
ExpectOut(65535.0f); ExpectOut(65534.0f); ExpectOut(0.0f);
ExpectOut(12345.0f); ExpectOut(54321.0f); ExpectOut(0.0f);
// Test that scale works on U16 inputs.
Input<u16>(42); Input<u16>(24);
m_vtx_attr.g0.PosFrac = 1;
m_vtx_attr.g0.ByteDequant = 1;
loader = VertexLoaderBase::CreateVertexLoader(m_vtx_desc, m_vtx_attr);
count = loader->RunVertices(src, dst, 1, 7);
src.Skip(1 * loader->m_VertexSize);
dst.Skip(count * loader->m_native_vtx_decl.stride);
ExpectOut(21.0f); ExpectOut(12.0f); ExpectOut(0.0f);
delete loader;
}
TEST_F(VertexLoaderTest, PositionDirectFloatXYZSpeed)
{
m_vtx_desc.Position = 1; // Direct
m_vtx_attr.g0.PosElements = 1; // XYZ
m_vtx_attr.g0.PosFormat = 4; // Float
VertexLoaderBase* loader = VertexLoaderBase::CreateVertexLoader(m_vtx_desc, m_vtx_attr);
ASSERT_EQ(3 * sizeof(float), (u32)loader->m_native_vtx_decl.stride);
ASSERT_EQ(3 * sizeof(float), (u32)loader->m_VertexSize);
for (int i = 0; i < 1000; ++i)
{ {
ResetPointers(); input_size = addr - 1;
int count = loader->RunVertices(src, dst, 100000, 7); for (int i = 0; i < count; i++)
src.Skip(100000 * loader->m_VertexSize); if (addr == INDEX8)
dst.Skip(count * loader->m_native_vtx_decl.stride); Input<u8>(i);
else
Input<u16>(i);
cached_arraybases[ARRAY_POSITION] = m_src.GetPointer();
g_main_cp_state.array_strides[ARRAY_POSITION] = elements * elem_size;
}
CreateAndCheckSizes(input_size, elements * sizeof(float));
for (float value : values)
{
switch (format)
{
case FORMAT_UBYTE: Input((u8)value); break;
case FORMAT_BYTE: Input((s8)value); break;
case FORMAT_USHORT: Input((u16)value); break;
case FORMAT_SHORT: Input((s16)value); break;
case FORMAT_FLOAT: Input(value); break;
}
}
RunVertices(count);
float scale = 1.f / (1u << (format == FORMAT_FLOAT ? 0 : frac));
for (auto iter = values.begin(); iter != values.end();)
{
float f, g;
switch (format)
{
case FORMAT_UBYTE: f = (u8)*iter++; g = (u8)*iter++; break;
case FORMAT_BYTE: f = (s8)*iter++; g = (s8)*iter++; break;
case FORMAT_USHORT: f = (u16)*iter++; g = (u16)*iter++; break;
case FORMAT_SHORT: f = (s16)*iter++; g = (s16)*iter++; break;
case FORMAT_FLOAT: f = *iter++; g = *iter++; break;
}
ExpectOut(f * scale);
ExpectOut(g * scale);
} }
delete loader;
} }
TEST_F(VertexLoaderTest, PositionDirectU16XYSpeed) TEST_F(VertexLoaderTest, PositionIndex16FloatXY)
{ {
m_vtx_desc.Position = 1; // Direct m_vtx_desc.Position = INDEX16;
m_vtx_attr.g0.PosElements = 0; // XY m_vtx_attr.g0.PosFormat = FORMAT_FLOAT;
m_vtx_attr.g0.PosFormat = 2; // U16 CreateAndCheckSizes(sizeof(u16), 2 * sizeof(float));
Input<u16>(1); Input<u16>(0);
cached_arraybases[ARRAY_POSITION] = m_src.GetPointer();
g_main_cp_state.array_strides[ARRAY_POSITION] = sizeof(float); // ;)
Input(1.f); Input(2.f); Input(3.f);
RunVertices(2);
ExpectOut(2); ExpectOut(3);
ExpectOut(1); ExpectOut(2);
}
VertexLoaderBase* loader = VertexLoaderBase::CreateVertexLoader(m_vtx_desc, m_vtx_attr); class VertexLoaderSpeedTest : public VertexLoaderTest, public ::testing::WithParamInterface<std::tuple<int, int>> {};
extern int gtest_FormatsAndElementsVertexLoaderSpeedTest_dummy_;
ASSERT_EQ(3 * sizeof(float), (u32)loader->m_native_vtx_decl.stride); INSTANTIATE_TEST_CASE_P(
ASSERT_EQ(2 * sizeof(u16), (u32)loader->m_VertexSize); FormatsAndElements, VertexLoaderSpeedTest,
::testing::Combine(
::testing::Values(FORMAT_UBYTE, FORMAT_BYTE, FORMAT_USHORT, FORMAT_SHORT, FORMAT_FLOAT),
::testing::Values(0, 1) // elements
)
);
TEST_P(VertexLoaderSpeedTest, PositionDirectAll)
{
int format, elements;
std::tie(format, elements) = GetParam();
const char* map[] = { "u8", "s8", "u16", "s16", "float" };
printf("format: %s, elements: %d\n", map[format], elements);
m_vtx_desc.Position = DIRECT;
m_vtx_attr.g0.PosFormat = format;
m_vtx_attr.g0.PosElements = elements;
elements += 2;
size_t elem_size = 1 << (format / 2);
CreateAndCheckSizes(elements * elem_size, elements * sizeof(float));
for (int i = 0; i < 1000; ++i) for (int i = 0; i < 1000; ++i)
{ RunVertices(100000);
ResetPointers(); }
int count = loader->RunVertices(src, dst, 100000, 7);
src.Skip(100000 * loader->m_VertexSize); TEST_P(VertexLoaderSpeedTest, TexCoordSingleElement)
dst.Skip(count * loader->m_native_vtx_decl.stride); {
} int format, elements;
delete loader; std::tie(format, elements) = GetParam();
const char* map[] = { "u8", "s8", "u16", "s16", "float" };
printf("format: %s, elements: %d\n", map[format], elements);
m_vtx_desc.Position = DIRECT;
m_vtx_attr.g0.PosFormat = FORMAT_BYTE;
m_vtx_desc.Tex0Coord = DIRECT;
m_vtx_attr.g0.Tex0CoordFormat = format;
m_vtx_attr.g0.Tex0CoordElements = elements;
elements += 1;
size_t elem_size = 1 << (format / 2);
CreateAndCheckSizes(2 * sizeof(s8) + elements * elem_size,
2 * sizeof(float) + elements * sizeof(float));
for (int i = 0; i < 1000; ++i)
RunVertices(100000);
} }
TEST_F(VertexLoaderTest, LargeFloatVertexSpeed) TEST_F(VertexLoaderTest, LargeFloatVertexSpeed)
{ {
// Enables most attributes in floating point direct mode to test speed. // Enables most attributes in floating point indexed mode to test speed.
m_vtx_desc.PosMatIdx = 1; m_vtx_desc.PosMatIdx = 1;
m_vtx_desc.Tex0MatIdx = 1; m_vtx_desc.Tex0MatIdx = 1;
m_vtx_desc.Tex1MatIdx = 1; m_vtx_desc.Tex1MatIdx = 1;
@ -222,54 +253,54 @@ TEST_F(VertexLoaderTest, LargeFloatVertexSpeed)
m_vtx_desc.Tex5MatIdx = 1; m_vtx_desc.Tex5MatIdx = 1;
m_vtx_desc.Tex6MatIdx = 1; m_vtx_desc.Tex6MatIdx = 1;
m_vtx_desc.Tex7MatIdx = 1; m_vtx_desc.Tex7MatIdx = 1;
m_vtx_desc.Position = 1; m_vtx_desc.Position = INDEX16;
m_vtx_desc.Normal = 1; m_vtx_desc.Normal = INDEX16;
m_vtx_desc.Color0 = 1; m_vtx_desc.Color0 = INDEX16;
m_vtx_desc.Color1 = 1; m_vtx_desc.Color1 = INDEX16;
m_vtx_desc.Tex0Coord = 1; m_vtx_desc.Tex0Coord = INDEX16;
m_vtx_desc.Tex1Coord = 1; m_vtx_desc.Tex1Coord = INDEX16;
m_vtx_desc.Tex2Coord = 1; m_vtx_desc.Tex2Coord = INDEX16;
m_vtx_desc.Tex3Coord = 1; m_vtx_desc.Tex3Coord = INDEX16;
m_vtx_desc.Tex4Coord = 1; m_vtx_desc.Tex4Coord = INDEX16;
m_vtx_desc.Tex5Coord = 1; m_vtx_desc.Tex5Coord = INDEX16;
m_vtx_desc.Tex6Coord = 1; m_vtx_desc.Tex6Coord = INDEX16;
m_vtx_desc.Tex7Coord = 1; m_vtx_desc.Tex7Coord = INDEX16;
m_vtx_attr.g0.PosElements = 1; // XYZ m_vtx_attr.g0.PosElements = 1; // XYZ
m_vtx_attr.g0.PosFormat = 4; // Float m_vtx_attr.g0.PosFormat = FORMAT_FLOAT;
m_vtx_attr.g0.NormalElements = 1; // NBT m_vtx_attr.g0.NormalElements = 1; // NBT
m_vtx_attr.g0.NormalFormat = 4; // Float m_vtx_attr.g0.NormalFormat = FORMAT_FLOAT;
m_vtx_attr.g0.Color0Elements = 1; // Has Alpha m_vtx_attr.g0.Color0Elements = 1; // Has Alpha
m_vtx_attr.g0.Color0Comp = 5; // RGBA8888 m_vtx_attr.g0.Color0Comp = FORMAT_32B_8888;
m_vtx_attr.g0.Color1Elements = 1; // Has Alpha m_vtx_attr.g0.Color1Elements = 1; // Has Alpha
m_vtx_attr.g0.Color1Comp = 5; // RGBA8888 m_vtx_attr.g0.Color1Comp = FORMAT_32B_8888;
m_vtx_attr.g0.Tex0CoordElements = 1; // ST m_vtx_attr.g0.Tex0CoordElements = 1; // ST
m_vtx_attr.g0.Tex0CoordFormat = 4; // Float m_vtx_attr.g0.Tex0CoordFormat = FORMAT_FLOAT;
m_vtx_attr.g1.Tex1CoordElements = 1; // ST m_vtx_attr.g1.Tex1CoordElements = 1; // ST
m_vtx_attr.g1.Tex1CoordFormat = 4; // Float m_vtx_attr.g1.Tex1CoordFormat = FORMAT_FLOAT;
m_vtx_attr.g1.Tex2CoordElements = 1; // ST m_vtx_attr.g1.Tex2CoordElements = 1; // ST
m_vtx_attr.g1.Tex2CoordFormat = 4; // Float m_vtx_attr.g1.Tex2CoordFormat = FORMAT_FLOAT;
m_vtx_attr.g1.Tex3CoordElements = 1; // ST m_vtx_attr.g1.Tex3CoordElements = 1; // ST
m_vtx_attr.g1.Tex3CoordFormat = 4; // Float m_vtx_attr.g1.Tex3CoordFormat = FORMAT_FLOAT;
m_vtx_attr.g1.Tex4CoordElements = 1; // ST m_vtx_attr.g1.Tex4CoordElements = 1; // ST
m_vtx_attr.g1.Tex4CoordFormat = 4; // Float m_vtx_attr.g1.Tex4CoordFormat = FORMAT_FLOAT;
m_vtx_attr.g2.Tex5CoordElements = 1; // ST m_vtx_attr.g2.Tex5CoordElements = 1; // ST
m_vtx_attr.g2.Tex5CoordFormat = 4; // Float m_vtx_attr.g2.Tex5CoordFormat = FORMAT_FLOAT;
m_vtx_attr.g2.Tex6CoordElements = 1; // ST m_vtx_attr.g2.Tex6CoordElements = 1; // ST
m_vtx_attr.g2.Tex6CoordFormat = 4; // Float m_vtx_attr.g2.Tex6CoordFormat = FORMAT_FLOAT;
m_vtx_attr.g2.Tex7CoordElements = 1; // ST m_vtx_attr.g2.Tex7CoordElements = 1; // ST
m_vtx_attr.g2.Tex7CoordFormat = 4; // Float m_vtx_attr.g2.Tex7CoordFormat = FORMAT_FLOAT;
VertexLoaderBase* loader = VertexLoaderBase::CreateVertexLoader(m_vtx_desc, m_vtx_attr); CreateAndCheckSizes(33, 156);
for (int i = 0; i < 16; i++)
{
cached_arraybases[i] = m_src.GetPointer();
g_main_cp_state.array_strides[i] = 129;
}
// This test is only done 100x in a row since it's ~20x slower using the // This test is only done 100x in a row since it's ~20x slower using the
// current vertex loader implementation. // current vertex loader implementation.
for (int i = 0; i < 100; ++i) for (int i = 0; i < 100; ++i)
{ RunVertices(100000);
ResetPointers();
int count = loader->RunVertices(src, dst, 100000, 7);
src.Skip(100000 * loader->m_VertexSize);
dst.Skip(count * loader->m_native_vtx_decl.stride);
}
delete loader;
} }