diff options
Diffstat (limited to 'backends/platform/psp/memory.h')
| -rw-r--r-- | backends/platform/psp/memory.h | 122 |
1 files changed, 60 insertions, 62 deletions
diff --git a/backends/platform/psp/memory.h b/backends/platform/psp/memory.h index 793bc94888..54e9225b2e 100644 --- a/backends/platform/psp/memory.h +++ b/backends/platform/psp/memory.h @@ -27,18 +27,24 @@ #ifndef PSP_MEMORY_H #define PSP_MEMORY_H -#include "backends/platform/psp/psppixelformat.h" -#include "common/list.h" - -#define UNCACHED(x) ((byte *)(((uint32)(x)) | 0x40000000)) /* make an uncached access */ -#define CACHED(x) ((byte *)(((uint32)(x)) & 0xBFFFFFFF)) /* make an uncached access into a cached one */ - #define MIN_AMOUNT_FOR_COMPLEX_COPY 8 #define MIN_AMOUNT_FOR_MISALIGNED_COPY 8 //#define __PSP_DEBUG_PRINT__ -#include "backends/platform/psp/trace.h" +//#include "backends/platform/psp/trace.h" + +// These instructions don't generate automatically but are faster then copying byte by byte +inline void lwl_copy(byte *dst, const byte *src) { + register uint32 data; + asm volatile ("lwr %0,0(%1)\n\t" + "lwl %0,3(%1)\n\t" + : "=&r" (data) : "r" (src), "m" (*src)); + + asm volatile ("swr %1,0(%2)\n\t" + "swl %1,3(%2)\n\t" + : "=m" (*dst) : "r" (data), "r" (dst)); +} /** * Class that does memory copying and swapping if needed @@ -46,42 +52,69 @@ class PspMemory { private: static void testCopy(const byte *debugDst, const byte *debugSrc, uint32 debugBytes); - static void testSwap(const uint16 *debugDst, const uint16 *debugSrc, uint32 debugBytes, PSPPixelFormat &format); static void copy(byte *dst, const byte *src, uint32 bytes); - static void swap(uint16 *dst16, const uint16 *src16, uint32 bytes, PSPPixelFormat &format); static void copy32Aligned(uint32 *dst32, const uint32 *src32, uint32 bytes); - static void swap32Aligned(uint32 *dst32, const uint32 *src32, uint32 bytes, PSPPixelFormat &format); static void copy32Misaligned(uint32 *dst32, const byte *src, uint32 bytes, uint32 alignSrc); - static void swap32Misaligned(uint32 *dst32, const uint16 *src16, uint32 bytes, PSPPixelFormat &format); - static void copy16(uint16 *dst, const uint16 *src, uint32 bytes); - - // For swapping, we know that we have multiples of 16 bits - static void swap16(uint16 *dst16, const uint16 *src16, uint32 bytes, PSPPixelFormat &format) { - PSP_DEBUG_PRINT("swap16 called with dst16[%p], src16[%p], bytes[%d]\n", dst16, src16, bytes); - uint32 shorts = bytes >> 1; - while (shorts--) { - *dst16++ = format.swapRedBlue16(*src16++); + static inline void copy8(byte *dst, const byte *src, int32 bytes) { + //PSP_DEBUG_PRINT("copy8 called with dst[%p], src[%p], bytes[%d]\n", dst, src, bytes); + uint32 words = bytes >> 2; + for (; words; words--) { + lwl_copy(dst, src); + dst += 4; + src += 4; } - } - - static void copy8(byte *dst, const byte *src, uint32 bytes) { - PSP_DEBUG_PRINT("copy8 called with dst[%p], src[%p], bytes[%d]\n", dst, src, bytes); - while (bytes--) { + + uint32 bytesLeft = bytes & 0x3; + for (; bytesLeft; bytesLeft--) { *dst++ = *src++; } } -public: +public: // This is the interface to the outside world - static void fastCopy(byte *dst, const byte *src, uint32 bytes) { + static void *fastCopy(void *dstv, const void *srcv, int32 bytes) { + byte *dst = (byte *)dstv; + byte *src = (byte *)srcv; + if (bytes < MIN_AMOUNT_FOR_COMPLEX_COPY) { copy8(dst, src, bytes); } else { // go to more powerful copy copy(dst, src, bytes); } + + return dstv; } +}; + +inline void *psp_memcpy(void *dst, const void *src, int32 bytes) { + return PspMemory::fastCopy(dst, src, bytes); +} + +#endif /* PSP_MEMORY_H */ + +#if defined(PSP_INCLUDE_SWAP) && !defined(PSP_MEMORY_SWAP_H) +#define PSP_MEMORY_SWAP_H +//#include "backends/platform/psp/psppixelformat.h" + +class PspMemorySwap { +private: + static void testSwap(const uint16 *debugDst, const uint16 *debugSrc, uint32 debugBytes, PSPPixelFormat &format); + static void swap(uint16 *dst16, const uint16 *src16, uint32 bytes, PSPPixelFormat &format); + static void swap32Aligned(uint32 *dst32, const uint32 *src32, uint32 bytes, PSPPixelFormat &format); + static void swap32Misaligned(uint32 *dst32, const uint16 *src16, uint32 bytes, PSPPixelFormat &format); + // For swapping, we know that we have multiples of 16 bits + static void swap16(uint16 *dst16, const uint16 *src16, uint32 bytes, PSPPixelFormat &format) { + PSP_DEBUG_PRINT("swap16 called with dst16[%p], src16[%p], bytes[%d]\n", dst16, src16, bytes); + uint32 shorts = bytes >> 1; + + while (shorts--) { + *dst16++ = format.swapRedBlue16(*src16++); + } +} + +public: static void fastSwap(byte *dst, const byte *src, uint32 bytes, PSPPixelFormat &format) { if (bytes < MIN_AMOUNT_FOR_COMPLEX_COPY * 2) { swap16((uint16 *)dst, (uint16 *)src, bytes, format); @@ -91,41 +124,6 @@ public: } }; -/** - * Class that allocates memory in the VRAM - */ -class VramAllocator : public Common::Singleton<VramAllocator> { -public: - VramAllocator() : _bytesAllocated(0) {} - void *allocate(int32 size, bool smallAllocation = false); // smallAllocation e.g. palettes - void deallocate(void *pointer); - - static inline bool isAddressInVram(void *address) { - if ((uint32)(CACHED(address)) >= VRAM_START_ADDRESS && (uint32)(CACHED(address)) < VRAM_END_ADDRESS) - return true; - return false; - } - +#endif /* PSP_INCLUDE_SWAP */ -private: - /** - * Used to allocate in VRAM - */ - struct Allocation { - byte *address; - uint32 size; - void *getEnd() { return address + size; } - Allocation(void *Address, uint32 Size) : address((byte *)Address), size(Size) {} - Allocation() : address(0), size(0) {} - }; - - enum { - VRAM_START_ADDRESS = 0x04000000, - VRAM_END_ADDRESS = 0x04200000, - VRAM_SMALL_ADDRESS = VRAM_END_ADDRESS - (4 * 1024) // 4K in the end for small allocations - }; - Common::List <Allocation> _allocList; // List of allocations - uint32 _bytesAllocated; -}; -#endif /* PSP_MEMORY_H */ |