diff options
Diffstat (limited to 'third_party/aom/aom_dsp/x86/intrapred_sse2.asm')
| -rw-r--r-- | third_party/aom/aom_dsp/x86/intrapred_sse2.asm | 146 |
1 files changed, 0 insertions, 146 deletions
diff --git a/third_party/aom/aom_dsp/x86/intrapred_sse2.asm b/third_party/aom/aom_dsp/x86/intrapred_sse2.asm index 02567db49..9aece27be 100644 --- a/third_party/aom/aom_dsp/x86/intrapred_sse2.asm +++ b/third_party/aom/aom_dsp/x86/intrapred_sse2.asm @@ -623,149 +623,3 @@ cglobal h_predictor_32x32, 2, 5, 3, dst, stride, line, left lea dstq, [dstq+strideq*4] jnz .loop REP_RET - -INIT_XMM sse2 -cglobal tm_predictor_4x4, 4, 4, 5, dst, stride, above, left - pxor m1, m1 - movq m0, [aboveq-1]; [63:0] tl t1 t2 t3 t4 x x x - punpcklbw m0, m1 - pshuflw m2, m0, 0x0 ; [63:0] tl tl tl tl [word] - psrldq m0, 2 - psubw m0, m2 ; [63:0] t1-tl t2-tl t3-tl t4-tl [word] - movd m2, [leftq] - punpcklbw m2, m1 - pshuflw m4, m2, 0x0 ; [63:0] l1 l1 l1 l1 [word] - pshuflw m3, m2, 0x55 ; [63:0] l2 l2 l2 l2 [word] - paddw m4, m0 - paddw m3, m0 - packuswb m4, m4 - packuswb m3, m3 - movd [dstq ], m4 - movd [dstq+strideq], m3 - lea dstq, [dstq+strideq*2] - pshuflw m4, m2, 0xaa - pshuflw m3, m2, 0xff - paddw m4, m0 - paddw m3, m0 - packuswb m4, m4 - packuswb m3, m3 - movd [dstq ], m4 - movd [dstq+strideq], m3 - RET - -INIT_XMM sse2 -cglobal tm_predictor_8x8, 4, 4, 5, dst, stride, above, left - pxor m1, m1 - movd m2, [aboveq-1] - movq m0, [aboveq] - punpcklbw m2, m1 - punpcklbw m0, m1 ; t1 t2 t3 t4 t5 t6 t7 t8 [word] - pshuflw m2, m2, 0x0 ; [63:0] tl tl tl tl [word] - DEFINE_ARGS dst, stride, line, left - mov lineq, -4 - punpcklqdq m2, m2 ; tl tl tl tl tl tl tl tl [word] - psubw m0, m2 ; t1-tl t2-tl ... t8-tl [word] - movq m2, [leftq] - punpcklbw m2, m1 ; l1 l2 l3 l4 l5 l6 l7 l8 [word] -.loop: - pshuflw m4, m2, 0x0 ; [63:0] l1 l1 l1 l1 [word] - pshuflw m3, m2, 0x55 ; [63:0] l2 l2 l2 l2 [word] - punpcklqdq m4, m4 ; l1 l1 l1 l1 l1 l1 l1 l1 [word] - punpcklqdq m3, m3 ; l2 l2 l2 l2 l2 l2 l2 l2 [word] - paddw m4, m0 - paddw m3, m0 - packuswb m4, m3 - movq [dstq ], m4 - movhps [dstq+strideq], m4 - lea dstq, [dstq+strideq*2] - psrldq m2, 4 - inc lineq - jnz .loop - REP_RET - -INIT_XMM sse2 -cglobal tm_predictor_16x16, 4, 5, 8, dst, stride, above, left - pxor m1, m1 - mova m2, [aboveq-16]; - mova m0, [aboveq] ; t1 t2 ... t16 [byte] - punpckhbw m2, m1 ; [127:112] tl [word] - punpckhbw m4, m0, m1 - punpcklbw m0, m1 ; m0:m4 t1 t2 ... t16 [word] - DEFINE_ARGS dst, stride, line, left, stride8 - mov lineq, -8 - pshufhw m2, m2, 0xff - mova m3, [leftq] ; l1 l2 ... l16 [byte] - punpckhqdq m2, m2 ; tl repeated 8 times [word] - psubw m0, m2 - psubw m4, m2 ; m0:m4 t1-tl t2-tl ... t16-tl [word] - punpckhbw m5, m3, m1 - punpcklbw m3, m1 ; m3:m5 l1 l2 ... l16 [word] - lea stride8q, [strideq*8] -.loop: - pshuflw m6, m3, 0x0 - pshuflw m7, m5, 0x0 - punpcklqdq m6, m6 ; l1 repeated 8 times [word] - punpcklqdq m7, m7 ; l8 repeated 8 times [word] - paddw m1, m6, m0 - paddw m6, m4 ; m1:m6 ti-tl+l1 [i=1,15] [word] - psrldq m5, 2 - packuswb m1, m6 - mova [dstq ], m1 - paddw m1, m7, m0 - paddw m7, m4 ; m1:m7 ti-tl+l8 [i=1,15] [word] - psrldq m3, 2 - packuswb m1, m7 - mova [dstq+stride8q], m1 - inc lineq - lea dstq, [dstq+strideq] - jnz .loop - REP_RET - -INIT_XMM sse2 -cglobal tm_predictor_32x32, 4, 4, 8, dst, stride, above, left - pxor m1, m1 - movd m2, [aboveq-1] - mova m0, [aboveq] - mova m4, [aboveq+16] - punpcklbw m2, m1 - punpckhbw m3, m0, m1 - punpckhbw m5, m4, m1 - punpcklbw m0, m1 - punpcklbw m4, m1 - pshuflw m2, m2, 0x0 - DEFINE_ARGS dst, stride, line, left - mov lineq, -16 - punpcklqdq m2, m2 - add leftq, 32 - psubw m0, m2 - psubw m3, m2 - psubw m4, m2 - psubw m5, m2 -.loop: - movd m2, [leftq+lineq*2] - pxor m1, m1 - punpcklbw m2, m1 - pshuflw m7, m2, 0x55 - pshuflw m2, m2, 0x0 - punpcklqdq m2, m2 - punpcklqdq m7, m7 - paddw m6, m2, m3 - paddw m1, m2, m0 - packuswb m1, m6 - mova [dstq ], m1 - paddw m6, m2, m5 - paddw m1, m2, m4 - packuswb m1, m6 - mova [dstq+16 ], m1 - paddw m6, m7, m3 - paddw m1, m7, m0 - packuswb m1, m6 - mova [dstq+strideq ], m1 - paddw m6, m7, m5 - paddw m1, m7, m4 - packuswb m1, m6 - mova [dstq+strideq+16], m1 - lea dstq, [dstq+strideq*2] - inc lineq - jnz .loop - REP_RET |