📄 3dnow_normal.s
字号:
MOVD ( REGOFF(8, EDX), MM5 ) /* | x2 */
PFMUL ( MM0, MM4 ) /* x1*m5 | x0*m0 */
ADD_L ( STRIDE, EDX) /* next normal */
PREFETCH ( REGIND(EDX) )
PFMUL ( MM2, MM5 ) /* | x2*m10 */
ADD_L ( CONST(16), EAX ) /* next r */
DEC_L ( EDI ) /* decrement normal counter */
MOVQ ( MM4, REGOFF(-16, EAX) ) /* write r0, r1 */
MOVD ( MM5, REGOFF(-8, EAX) ) /* write r2 */
JA ( LLBL (G3TNR_transform) )
FEMMS
LLBL (G3TNR_end):
POP_L ( ESI )
POP_L ( EDI )
RET
ALIGNTEXT16
GLOBL GLNAME(_mesa_3dnow_transform_normals)
HIDDEN(_mesa_3dnow_transform_normals)
GLNAME(_mesa_3dnow_transform_normals):
#undef FRAME_OFFSET
#define FRAME_OFFSET 8
PUSH_L ( EDI )
PUSH_L ( ESI )
MOV_L ( ARG_IN, ESI )
MOV_L ( ARG_DEST, EAX )
MOV_L ( ARG_MAT, ECX )
MOV_L ( REGOFF(V4F_COUNT, ESI), EDI ) /* dest->count = in->count */
MOV_L ( EDI, REGOFF(V4F_COUNT, EAX) )
MOV_L ( REGOFF(V4F_START, EAX), EAX ) /* dest->start */
MOV_L ( REGOFF(V4F_START, ESI), EDX ) /* in->start */
MOV_L ( REGOFF(MATRIX_INV, ECX), ECX ) /* mat->inv */
CMP_L ( CONST(0), EDI ) /* count > 0 ?? */
JE ( LLBL (G3T_end) )
FEMMS
MOVQ ( REGIND(ECX), MM3 ) /* m1 | m0 */
MOVQ ( REGOFF(16, ECX), MM4 ) /* m5 | m4 */
MOVD ( REGOFF(8, ECX), MM5 ) /* | m2 */
PUNPCKLDQ ( REGOFF(24, ECX), MM5 ) /* m6 | m2 */
MOVQ ( REGOFF(32, ECX), MM6 ) /* m9 | m8 */
MOVD ( REGOFF(40, ECX), MM7 ) /* | m10 */
ALIGNTEXT32
LLBL (G3T_transform):
PREFETCHW ( REGIND(EAX) )
MOVQ ( REGIND(EDX), MM0 ) /* x1 | x0 */
MOVD ( REGOFF(8, EDX), MM2 ) /* | x2 */
MOVQ ( MM0, MM1 ) /* x1 | x0 */
PUNPCKLDQ ( MM2, MM2 ) /* x2 | x2 */
PFMUL ( MM3, MM0 ) /* x1*m1 | x0*m0 */
ADD_L ( CONST(16), EAX ) /* next r */
PFMUL ( MM4, MM1 ) /* x1*m5 | x0*m4 */
PFACC ( MM1, MM0 ) /* x0*m4+x1*m5 | x0*m0+x1*m1 */
PFMUL ( MM5, MM2 ) /* x2*m6 | x2*m2 */
PFADD ( MM2, MM0 ) /* x0*m4...+x2*m6| x0*m0+x1*m1+x2*m2 */
MOVQ ( REGIND(EDX), MM1 ) /* x1 | x0 */
MOVQ ( MM0, REGOFF(-16, EAX) ) /* write r0, r1 */
PFMUL ( MM6, MM1 ) /* x1*m9 | x0*m8 */
MOVD ( REGOFF(8, EDX), MM2 ) /* | x2 */
PFMUL ( MM7, MM2 ) /* | x2*m10 */
ADD_L ( STRIDE, EDX ) /* next normal */
PREFETCH ( REGIND(EDX) )
PFACC ( MM1, MM1 ) /* *not used* | x0*m8+x1*m9 */
PFADD ( MM2, MM1 ) /* *not used* | x0*m8+x1*m9+x2*m10 */
MOVD ( MM1, REGOFF(-8, EAX) ) /* write r2 */
DEC_L ( EDI ) /* decrement normal counter */
JA ( LLBL (G3T_transform) )
FEMMS
LLBL (G3T_end):
POP_L ( ESI )
POP_L ( EDI )
RET
ALIGNTEXT16
GLOBL GLNAME(_mesa_3dnow_normalize_normals)
HIDDEN(_mesa_3dnow_normalize_normals)
GLNAME(_mesa_3dnow_normalize_normals):
#undef FRAME_OFFSET
#define FRAME_OFFSET 12
PUSH_L ( EDI )
PUSH_L ( ESI )
PUSH_L ( EBP )
MOV_L ( ARG_IN, ESI )
MOV_L ( ARG_DEST, EAX )
MOV_L ( REGOFF(V4F_COUNT, ESI), EBP ) /* dest->count = in->count */
MOV_L ( EBP, REGOFF(V4F_COUNT, EAX) )
MOV_L ( REGOFF(V4F_START, EAX), EAX ) /* dest->start */
MOV_L ( REGOFF(V4F_START, ESI), ECX ) /* in->start */
MOV_L ( ARG_LENGTHS, EDX )
CMP_L ( CONST(0), EBP ) /* count > 0 ?? */
JE ( LLBL (G3N_end) )
FEMMS
CMP_L ( CONST(0), EDX ) /* lengths == 0 ? */
JE ( LLBL (G3N_norm2) ) /* calculate lengths */
ALIGNTEXT32
LLBL (G3N_norm1): /* use precalculated lengths */
PREFETCH ( REGIND(EAX) )
MOVQ ( REGIND(ECX), MM0 ) /* x1 | x0 */
MOVD ( REGOFF(8, ECX), MM1 ) /* | x2 */
MOVD ( REGIND(EDX), MM3 ) /* | length (x) */
PFMUL ( MM3, MM1 ) /* | x2 (normalized) */
PUNPCKLDQ ( MM3, MM3 ) /* length (x) | length (x) */
ADD_L ( STRIDE, ECX ) /* next normal */
PREFETCH ( REGIND(ECX) )
PFMUL ( MM3, MM0 ) /* x1 (normalized) | x0 (normalized) */
MOVQ ( MM0, REGIND(EAX) ) /* write new x0, x1 */
MOVD ( MM1, REGOFF(8, EAX) ) /* write new x2 */
ADD_L ( CONST(16), EAX ) /* next r */
ADD_L ( CONST(4), EDX ) /* next length */
DEC_L ( EBP ) /* decrement normal counter */
JA ( LLBL (G3N_norm1) )
JMP ( LLBL (G3N_end1) )
ALIGNTEXT32
LLBL (G3N_norm2): /* need to calculate lengths */
PREFETCHW ( REGIND(EAX) )
PREFETCH ( REGIND(ECX) )
MOVQ ( REGIND(ECX), MM0 ) /* x1 | x0 */
MOVD ( REGOFF(8, ECX), MM1 ) /* | x2 */
MOVQ ( MM0, MM3 ) /* x1 | x0 */
ADD_L ( STRIDE, ECX ) /* next normal */
PFMUL ( MM0, MM3 ) /* x1*x1 | x0*x0 */
MOVQ ( MM1, MM4 ) /* | x2 */
ADD_L ( CONST(16), EAX ) /* next r */
PFMUL ( MM1, MM4 ) /* | x2*x2 */
PFADD ( MM4, MM3 ) /* | x0*x0+x2*x2 */
PFACC ( MM3, MM3 ) /* x0*x0+...+x2*x2 | x0*x0+x1*x1+x2*x2*/
PFRSQRT ( MM3, MM5 ) /* 1/sqrt (x0*x0+x1*x1+x2*x2) */
MOVQ ( MM5, MM4 )
PUNPCKLDQ ( MM3, MM3 )
PFMUL ( MM5, MM5 )
PFRSQIT1 ( MM3, MM5 )
DEC_L ( EBP ) /* decrement normal counter */
PFRCPIT2 ( MM4, MM5 )
PFMUL ( MM5, MM0 ) /* x1 (normalized) | x0 (normalized) */
MOVQ ( MM0, REGOFF(-16, EAX) ) /* write new x0, x1 */
PFMUL ( MM5, MM1 ) /* | x2 (normalized) */
MOVD ( MM1, REGOFF(-8, EAX) ) /* write new x2 */
JA ( LLBL (G3N_norm2) )
LLBL (G3N_end1):
FEMMS
LLBL (G3N_end):
POP_L ( EBP )
POP_L ( ESI )
POP_L ( EDI )
RET
ALIGNTEXT16
GLOBL GLNAME(_mesa_3dnow_rescale_normals)
HIDDEN(_mesa_3dnow_rescale_normals)
GLNAME(_mesa_3dnow_rescale_normals):
#undef FRAME_OFFSET
#define FRAME_OFFSET 8
PUSH_L ( EDI )
PUSH_L ( ESI )
MOV_L ( ARG_IN, ESI )
MOV_L ( ARG_DEST, EAX )
MOV_L ( REGOFF(V4F_COUNT, ESI), EDX ) /* dest->count = in->count */
MOV_L ( EDX, REGOFF(V4F_COUNT, EAX) )
MOV_L ( REGOFF(V4F_START, EAX), EAX ) /* dest->start */
MOV_L ( REGOFF(V4F_START, ESI), ECX ) /* in->start */
CMP_L ( CONST(0), EDX )
JE ( LLBL (G3R_end) )
FEMMS
MOVD ( ARG_SCALE, MM0 ) /* scale */
PUNPCKLDQ ( MM0, MM0 )
ALIGNTEXT32
LLBL (G3R_rescale):
PREFETCHW ( REGIND(EAX) )
MOVQ ( REGIND(ECX), MM1 ) /* x1 | x0 */
MOVD ( REGOFF(8, ECX), MM2 ) /* | x2 */
PFMUL ( MM0, MM1 ) /* x1*scale | x0*scale */
ADD_L ( STRIDE, ECX ) /* next normal */
PREFETCH ( REGIND(ECX) )
PFMUL ( MM0, MM2 ) /* | x2*scale */
ADD_L ( CONST(16), EAX ) /* next r */
MOVQ ( MM1, REGOFF(-16, EAX) ) /* write r0, r1 */
MOVD ( MM2, REGOFF(-8, EAX) ) /* write r2 */
DEC_L ( EDX ) /* decrement normal counter */
JA ( LLBL (G3R_rescale) )
FEMMS
LLBL (G3R_end):
POP_L ( ESI )
POP_L ( EDI )
RET
#endif
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -