📄 3dnow_xform3.s
字号:
LLBL( G3TP3R_2 ):
FEMMS
POP_L ( EDI )
POP_L ( ESI )
RET
ALIGNTEXT16
GLOBL GLNAME( _mesa_3dnow_transform_points3_3d_no_rot )
HIDDEN(_mesa_3dnow_transform_points3_3d_no_rot)
GLNAME( _mesa_3dnow_transform_points3_3d_no_rot ):
PUSH_L ( ESI )
MOV_L ( ARG_DEST, ECX )
MOV_L ( ARG_MATRIX, ESI )
MOV_L ( ARG_SOURCE, EAX )
MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
PUSH_L ( EDI )
MOV_L ( REGOFF(V4F_START, ECX), EDX )
MOV_L ( ESI, ECX )
MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
MOV_L ( REGOFF(V4F_START, EAX), EAX )
TEST_L ( ESI, ESI )
JZ ( LLBL( G3TP3NRR_2 ) )
PREFETCH ( REGIND(EAX) )
PREFETCHW ( REGIND(EDX) )
MOVD ( REGIND(ECX), MM0 ) /* | m00 */
PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */
PUNPCKLDQ ( MM2, MM2 ) /* m22 | m22 */
MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
PUNPCKLDQ ( MM3, MM3 ) /* m32 | m32 */
ALIGNTEXT16
LLBL( G3TP3NRR_1 ):
PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */
ADD_L ( EDI, EAX ) /* next vertex */
PREFETCHW ( REGIND(EAX) )
PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
PFADD ( MM1, MM4 ) /* x1*m11+m31 | x0*m00+m30 */
PFMUL ( MM2, MM5 ) /* | x2*m22 */
PFADD ( MM3, MM5 ) /* | x2*m22+m32 */
MOVQ ( MM4, REGIND(EDX) ) /* write r0, r1 */
ADD_L ( CONST(16), EDX ) /* next r */
DEC_L ( ESI ) /* decrement vertex counter */
MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 */
JNZ ( LLBL( G3TP3NRR_1 ) ) /* cnt > 0 ? -> process next vertex */
LLBL( G3TP3NRR_2 ):
FEMMS
POP_L ( EDI )
POP_L ( ESI )
RET
ALIGNTEXT16
GLOBL GLNAME( _mesa_3dnow_transform_points3_2d )
HIDDEN(_mesa_3dnow_transform_points3_2d)
GLNAME( _mesa_3dnow_transform_points3_2d ):
PUSH_L ( ESI )
MOV_L ( ARG_DEST, ECX )
MOV_L ( ARG_MATRIX, ESI )
MOV_L ( ARG_SOURCE, EAX )
MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
PUSH_L ( EDI )
MOV_L ( REGOFF(V4F_START, ECX), EDX )
MOV_L ( ESI, ECX )
MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
MOV_L ( REGOFF(V4F_START, EAX), EAX )
TEST_L ( ESI, ESI )
JZ ( LLBL( G3TP2R_3) )
PREFETCH ( REGIND(EAX) )
PREFETCHW ( REGIND(EDX) )
MOVD ( REGIND(ECX), MM0 ) /* | m00 */
PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
ALIGNTEXT16
LLBL( G3TP2R_2 ):
PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */
MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */
ADD_L ( EDI, EAX ) /* next vertex */
PREFETCH ( REGIND(EAX) )
MOVQ ( MM3, MM4 ) /* x1 | x0 */
PFMUL ( MM0, MM3 ) /* x1*m10 | x0*m00 */
ADD_L ( CONST(16), EDX ) /* next r */
PFMUL ( MM1, MM4 ) /* x1*m11 | x0*m01 */
PFACC ( MM4, MM3 ) /* x0*m00+x1*m10 | x0*m01+x1*m11 */
MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 (=x2) */
PFADD ( MM2, MM3 ) /* x0*...*m10+m30 | x0*...*m11+m31 */
MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */
DEC_L ( ESI ) /* decrement vertex counter */
JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */
LLBL( G3TP2R_3 ):
FEMMS
POP_L ( EDI )
POP_L ( ESI )
RET
ALIGNTEXT16
GLOBL GLNAME( _mesa_3dnow_transform_points3_2d_no_rot )
HIDDEN(_mesa_3dnow_transform_points3_2d_no_rot)
GLNAME( _mesa_3dnow_transform_points3_2d_no_rot ):
PUSH_L ( ESI )
MOV_L ( ARG_DEST, ECX )
MOV_L ( ARG_MATRIX, ESI )
MOV_L ( ARG_SOURCE, EAX )
MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
PUSH_L ( EDI )
MOV_L ( REGOFF(V4F_START, ECX), EDX )
MOV_L ( ESI, ECX )
MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
MOV_L ( REGOFF(V4F_START, EAX), EAX )
TEST_L ( ESI, ESI )
JZ ( LLBL( G3TP2NRR_2 ) )
PREFETCH ( REGIND(EAX) )
PREFETCHW ( REGIND(EDX) )
MOVD ( REGIND(ECX), MM0 ) /* | m00 */
PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
ALIGNTEXT16
LLBL( G3TP2NRR_1 ):
PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */
ADD_L ( EDI, EAX ) /* next vertex */
PREFETCH ( REGIND(EAX) )
PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
ADD_L ( CONST(16), EDX ) /* next r */
PFADD ( MM1, MM4 ) /* x1*m11+m31 | x0*m00+m30 */
MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */
MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 (=x2) */
DEC_L ( ESI ) /* decrement vertex counter */
JNZ ( LLBL( G3TP2NRR_1 ) ) /* cnt > 0 ? -> process next vertex */
LLBL( G3TP2NRR_2 ):
FEMMS
POP_L ( EDI )
POP_L ( ESI )
RET
ALIGNTEXT16
GLOBL GLNAME( _mesa_3dnow_transform_points3_identity )
HIDDEN(_mesa_3dnow_transform_points3_identity)
GLNAME( _mesa_3dnow_transform_points3_identity ):
PUSH_L ( ESI )
MOV_L ( ARG_DEST, ECX )
MOV_L ( ARG_MATRIX, ESI )
MOV_L ( ARG_SOURCE, EAX )
MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
PUSH_L ( EDI )
MOV_L ( REGOFF(V4F_START, ECX), EDX )
MOV_L ( ESI, ECX )
MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
MOV_L ( REGOFF(V4F_START, EAX), EAX )
TEST_L ( ESI, ESI )
JZ ( LLBL( G3TPIR_2 ) )
PREFETCHW ( REGIND(EDX) )
ALIGNTEXT16
LLBL( G3TPIR_1 ):
PREFETCHW ( REGOFF(32, EDX) )
MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
MOVD ( REGOFF(8, EAX), MM1 ) /* | x2 */
ADD_L ( EDI, EAX ) /* next vertex */
ADD_L ( CONST(16), EDX ) /* next r */
DEC_L ( ESI ) /* decrement vertex counter */
MOVQ ( MM0, REGOFF(-16, EDX) ) /* r1 | r0 */
MOVD ( MM1, REGOFF(-8, EDX) ) /* | r2 */
JNZ ( LLBL( G3TPIR_1 ) ) /* cnt > 0 ? -> process next vertex */
LLBL( G3TPIR_2 ):
FEMMS
POP_L ( EDI )
POP_L ( ESI )
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 + -