Implement Banshee tiled surfaces

This commit is contained in:
SarahW 2020-08-12 15:13:48 +01:00
commit 133f19fe39
9 changed files with 379 additions and 133 deletions

View file

@ -41,28 +41,36 @@ typedef struct voodoo_x86_data_t
static int last_block[2] = {0, 0};
static int next_block_to_write[2] = {0, 0};
#define addbyte(val) \
code_block[block_pos++] = val; \
if (block_pos >= BLOCK_SIZE) \
fatal("Over!\n")
#define addbyte(val) \
do { \
code_block[block_pos++] = val; \
if (block_pos >= BLOCK_SIZE) \
fatal("Over!\n"); \
} while (0)
#define addword(val) \
*(uint16_t *)&code_block[block_pos] = val; \
block_pos += 2; \
if (block_pos >= BLOCK_SIZE) \
fatal("Over!\n")
#define addword(val) \
do { \
*(uint16_t *)&code_block[block_pos] = val; \
block_pos += 2; \
if (block_pos >= BLOCK_SIZE) \
fatal("Over!\n"); \
} while (0)
#define addlong(val) \
*(uint32_t *)&code_block[block_pos] = val; \
block_pos += 4; \
if (block_pos >= BLOCK_SIZE) \
fatal("Over!\n")
#define addlong(val) \
do { \
*(uint32_t *)&code_block[block_pos] = val; \
block_pos += 4; \
if (block_pos >= BLOCK_SIZE) \
fatal("Over!\n"); \
} while (0)
#define addquad(val) \
*(uint64_t *)&code_block[block_pos] = val; \
block_pos += 8; \
if (block_pos >= BLOCK_SIZE) \
fatal("Over!\n")
#define addquad(val) \
do { \
*(uint64_t *)&code_block[block_pos] = val; \
block_pos += 8; \
if (block_pos >= BLOCK_SIZE) \
fatal("Over!\n"); \
} while (0)
static __m128i xmm_01_w;// = 0x0001000100010001ull;
@ -782,6 +790,28 @@ static inline void voodoo_generate(uint8_t *code_block, voodoo_t *voodoo, voodoo
addbyte(0x4c); /*MOV RSI, R15*/
addbyte(0x89);
addbyte(0xfe);
if (voodoo->col_tiled || voodoo->aux_tiled)
{
addbyte(0x8b); /*MOV EAX, state->x[EDI]*/
addbyte(0x87);
addlong(offsetof(voodoo_state_t, x));
addbyte(0x89); /*MOV EBX, EAX*/
addbyte(0xc3);
addbyte(0x83); /*AND EAX, 63*/
addbyte(0xe0);
addbyte(63);
addbyte(0xc1); /*SHR EBX, 6*/
addbyte(0xeb);
addbyte(6);
addbyte(0xc1); /*SHL EBX, 11 - tile is 128*32, << 12, div 2 because word index*/
addbyte(0xe3);
addbyte(11);
addbyte(0x01); /*ADD EAX, EBX*/
addbyte(0xd8);
addbyte(0x89); /*MOV state->x_tiled[EDI], EAX*/
addbyte(0x87);
addlong(offsetof(voodoo_state_t, x_tiled));
}
addbyte(0x66); /*PXOR XMM2, XMM2*/
addbyte(0x0f);
addbyte(0xef);
@ -899,7 +929,10 @@ static inline void voodoo_generate(uint8_t *code_block, voodoo_t *voodoo, voodoo
{
addbyte(0x8b); /*MOV EBX, state->x[EDI]*/
addbyte(0x9f);
addlong(offsetof(voodoo_state_t, x));
if (voodoo->aux_tiled)
addlong(offsetof(voodoo_state_t, x_tiled));
else
addlong(offsetof(voodoo_state_t, x));
addbyte(0x48); /*MOV RCX, aux_mem[RDI]*/
addbyte(0x8b);
addbyte(0x8f);
@ -2483,7 +2516,10 @@ static inline void voodoo_generate(uint8_t *code_block, voodoo_t *voodoo, voodoo
addquad((uintptr_t)rgb565);
addbyte(0x8b); /*MOV EAX, state->x[EDI]*/
addbyte(0x87);
addlong(offsetof(voodoo_state_t, x));
if (voodoo->col_tiled)
addlong(offsetof(voodoo_state_t, x_tiled));
else
addlong(offsetof(voodoo_state_t, x));
addbyte(0x48); /*MOV RBP, fb_mem*/
addbyte(0x8b);
addbyte(0xaf);
@ -2865,8 +2901,11 @@ static inline void voodoo_generate(uint8_t *code_block, voodoo_t *voodoo, voodoo
addbyte(0x8b); /*MOV EDX, state->x[EDI]*/
addbyte(0x97);
addlong(offsetof(voodoo_state_t, x));
if (voodoo->col_tiled)
addlong(offsetof(voodoo_state_t, x_tiled));
else
addlong(offsetof(voodoo_state_t, x));
addbyte(0x66); /*MOV EAX, XMM0*/
addbyte(0x0f);
addbyte(0x7e);
@ -2932,7 +2971,10 @@ static inline void voodoo_generate(uint8_t *code_block, voodoo_t *voodoo, voodoo
}
addbyte(0x8b); /*MOV EDX, state->x[EDI]*/
addbyte(0x97);
addlong(offsetof(voodoo_state_t, x));
if (voodoo->col_tiled)
addlong(offsetof(voodoo_state_t, x_tiled));
else
addlong(offsetof(voodoo_state_t, x));
addbyte(0x4c); /*ADD RSI, R8*/
addbyte(0x01);
addbyte(0xc6);
@ -3018,6 +3060,12 @@ static inline void voodoo_generate(uint8_t *code_block, voodoo_t *voodoo, voodoo
if ((params->fbzMode & (FBZ_DEPTH_WMASK | FBZ_DEPTH_ENABLE)) == (FBZ_DEPTH_WMASK | FBZ_DEPTH_ENABLE))
{
addbyte(0x8b); /*MOV EDX, state->x[EDI]*/
addbyte(0x97);
if (voodoo->aux_tiled)
addlong(offsetof(voodoo_state_t, x_tiled));
else
addlong(offsetof(voodoo_state_t, x));
addbyte(0x66); /*MOV AX, new_depth*/
addbyte(0x8b);
addbyte(0x87);