fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Moving byte order/memory access to poly.

Browse Source
This commit is contained in:
Ben Vanik
2014-07-17 19:20:17 -07:00
parent ce70978ef6
commit ec4f41fec4
42 changed files with 608 additions and 486 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
src/xenia/gpu/command_processor.cc
View File

@@ -118,7 +118,7 @@ void CommandProcessor::Pump() {
// TODO(benvanik): use read_ptr_update_freq_ and only issue after moving
// that many indices.
if (read_ptr_writeback_ptr_) {
XESETUINT32BE(p + read_ptr_writeback_ptr_, read_ptr_index_);
poly::store_and_swap<uint32_t>(p + read_ptr_writeback_ptr_, read_ptr_index_);
}
}
@@ -174,7 +174,7 @@ void CommandProcessor::ExecuteIndirectBuffer(uint32_t ptr, uint32_t length) {
for (uint32_t __m = 0; __m < count; __m++) { \
XETRACECP("[%.8X] %.8X", \
packet_ptr + (1 + __m) * 4, \
XEGETUINT32BE(packet_base + 1 * 4 + __m * 4)); \
poly::load_and_swap<uint32_t>(packet_base + 1 * 4 + __m * 4)); \
}
void CommandProcessor::AdvancePtr(PacketArgs& args, uint32_t n) {
@@ -186,9 +186,9 @@ void CommandProcessor::AdvancePtr(PacketArgs& args, uint32_t n) {
}
#define ADVANCE_PTR(n) AdvancePtr(args, n)
#define PEEK_PTR() \
XEGETUINT32BE(p + args.ptr)
poly::load_and_swap<uint32_t>(p + args.ptr)
#define READ_PTR() \
XEGETUINT32BE(p + args.ptr); ADVANCE_PTR(1);
poly::load_and_swap<uint32_t>(p + args.ptr); ADVANCE_PTR(1);
uint32_t CommandProcessor::ExecutePacket(PacketArgs& args) {
uint8_t* p = memory_->membase();
@@ -339,7 +339,7 @@ uint32_t CommandProcessor::ExecutePacket(PacketArgs& args) {
// Memory.
XE_GPU_ENDIAN endianness = (XE_GPU_ENDIAN)(poll_reg_addr & 0x3);
poll_reg_addr &= ~0x3;
value = XEGETUINT32LE(p + GpuToCpu(packet_ptr, poll_reg_addr));
value = poly::load<uint32_t>(p + GpuToCpu(packet_ptr, poll_reg_addr));
value = GpuSwap(value, endianness);
} else {
// Register.
@@ -434,7 +434,7 @@ uint32_t CommandProcessor::ExecutePacket(PacketArgs& args) {
// Memory.
XE_GPU_ENDIAN endianness = (XE_GPU_ENDIAN)(poll_reg_addr & 0x3);
poll_reg_addr &= ~0x3;
value = XEGETUINT32LE(p + GpuToCpu(packet_ptr, poll_reg_addr));
value = poly::load<uint32_t>(p + GpuToCpu(packet_ptr, poll_reg_addr));
value = GpuSwap(value, endianness);
} else {
// Register.
@@ -475,8 +475,8 @@ uint32_t CommandProcessor::ExecutePacket(PacketArgs& args) {
XE_GPU_ENDIAN endianness = (XE_GPU_ENDIAN)(write_reg_addr & 0x3);
write_reg_addr &= ~0x3;
write_data = GpuSwap(write_data, endianness);
XESETUINT32LE(p + GpuToCpu(packet_ptr, write_reg_addr),
write_data);
poly::store(p + GpuToCpu(packet_ptr, write_reg_addr),
write_data);
} else {
// Register.
WriteRegister(packet_ptr, write_reg_addr, write_data);
@@ -524,7 +524,7 @@ uint32_t CommandProcessor::ExecutePacket(PacketArgs& args) {
XE_GPU_ENDIAN endianness = (XE_GPU_ENDIAN)(address & 0x3);
address &= ~0x3;
data_value = GpuSwap(data_value, endianness);
XESETUINT32LE(p + GpuToCpu(address), data_value);
poly::store(p + GpuToCpu(address), data_value);
}
break;
@@ -637,7 +637,7 @@ uint32_t CommandProcessor::ExecutePacket(PacketArgs& args) {
size &= 0xFFF;
index += 0x4000; // alu constants
for (uint32_t n = 0; n < size; n++, index++) {
uint32_t data = XEGETUINT32BE(
uint32_t data = poly::load_and_swap<uint32_t>(
p + GpuToCpu(packet_ptr, address + n * 4));
const char* reg_name = regs->GetRegisterName(index);
XETRACECP("[%.8X] %.8X -> %.4X %s",
@@ -769,7 +769,7 @@ void CommandProcessor::WriteRegister(
uint8_t* p = memory_->membase();
uint32_t scratch_addr = regs->values[XE_GPU_REG_SCRATCH_ADDR].u32;
uint32_t mem_addr = scratch_addr + (scratch_reg * 4);
XESETUINT32BE(p + GpuToCpu(primary_buffer_ptr_, mem_addr), value);
poly::store_and_swap<uint32_t>(p + GpuToCpu(primary_buffer_ptr_, mem_addr), value);
}
}
}

6
src/xenia/gpu/d3d11/d3d11_buffer_resource.cc
View File

@@ -69,7 +69,7 @@ int D3D11IndexBufferResource::InvalidateRegion(
memory_range_.host_base);
uint32_t* dest = reinterpret_cast<uint32_t*>(res.pData);
for (uint32_t n = 0; n < index_count; n++) {
dest[n] = XESWAP32(src[n]);
dest[n] = poly::byte_swap(src[n]);
}
} else {
uint32_t index_count = memory_range_.length / 2;
@@ -77,7 +77,7 @@ int D3D11IndexBufferResource::InvalidateRegion(
memory_range_.host_base);
uint16_t* dest = reinterpret_cast<uint16_t*>(res.pData);
for (uint32_t n = 0; n < index_count; n++) {
dest[n] = XESWAP16(src[n]);
dest[n] = poly::byte_swap(src[n]);
}
}
resource_cache_->context()->Unmap(handle_, 0);
@@ -138,7 +138,7 @@ int D3D11VertexBufferResource::InvalidateRegion(
uint32_t o = 0;
for (uint32_t i = 0; i < count; i++) {
for (uint32_t j = 0; j < el.size_words; j++) {
dest_ptr[o + j] = XESWAP32(src_ptr[o + j]);
dest_ptr[o + j] = poly::byte_swap(src_ptr[o + j]);
}
o += stride;
}

2
src/xenia/gpu/shader_resource.cc
View File

@@ -35,7 +35,7 @@ ShaderResource::ShaderResource(const MemoryRange& memory_range,
size_t byte_size = dword_count_ * sizeof(uint32_t);
dwords_ = (uint32_t*)xe_malloc(byte_size);
for (uint32_t n = 0; n < dword_count_; n++) {
dwords_[n] = XEGETUINT32BE(memory_range.host_base + n * 4);
dwords_[n] = poly::load_and_swap<uint32_t>(memory_range.host_base + n * 4);
}
// Disassemble, for debugging.

4
src/xenia/gpu/texture_resource.cc
View File

@@ -307,12 +307,12 @@ void TextureResource::TextureSwap(uint8_t* dest, const uint8_t* src,
switch (info_.endianness) {
case XE_GPU_ENDIAN_8IN16:
for (uint32_t i = 0; i < pitch; i += 2, src += 2, dest += 2) {
*(uint16_t*)dest = XESWAP16(*(uint16_t*)src);
*(uint16_t*)dest = poly::byte_swap(*(uint16_t*)src);
}
break;
case XE_GPU_ENDIAN_8IN32: // Swap bytes.
for (uint32_t i = 0; i < pitch; i += 4, src += 4, dest += 4) {
*(uint32_t*)dest = XESWAP32(*(uint32_t*)src);
*(uint32_t*)dest = poly::byte_swap(*(uint32_t*)src);
}
break;
case XE_GPU_ENDIAN_16IN32: // Swap half words.

2
src/xenia/gpu/xenos/xenos.h
View File

@@ -83,7 +83,7 @@ XEFORCEINLINE uint32_t GpuSwap(uint32_t value, XE_GPU_ENDIAN endianness) {
((value >> 8) & 0x00FF00FF);
case XE_GPU_ENDIAN_8IN32: // Swap bytes.
// NOTE: we are likely doing two swaps here. Wasteful. Oh well.
return XESWAP32(value);
return poly::byte_swap(value);
case XE_GPU_ENDIAN_16IN32: // Swap half words.
return ((value >> 16) & 0xFFFF) | (value << 16);
}