From 1083fe499973fb363ab4325ec3b4a905cdb5c0e1 Mon Sep 17 00:00:00 2001 From: "Ronald S. Bultje" Date: Wed, 27 Apr 2011 10:05:10 -0700 Subject: SSE2/SSSE3 optimizations for build_predictors_mbuv{,_s}(). decoding before 10.425 10.432 10.423 =10.426 after: 10.405 10.416 10.398 =10.406, 0.2% faster encoding before 14.252 14.331 14.250 14.223 14.241 14.220 14.221 =14.248 after 14.095 14.090 14.085 14.095 14.064 14.081 14.089 =14.086, 1.1% faster Change-Id: I483d3d8f0deda8ad434cea76e16028380722aee2 --- vp8/common/x86/recon_sse2.asm | 408 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 408 insertions(+) (limited to 'vp8/common/x86/recon_sse2.asm') diff --git a/vp8/common/x86/recon_sse2.asm b/vp8/common/x86/recon_sse2.asm index 67b6420a9..a1cc2b1fe 100644 --- a/vp8/common/x86/recon_sse2.asm +++ b/vp8/common/x86/recon_sse2.asm @@ -229,3 +229,411 @@ sym(vp8_copy_mem16x16_sse2): UNSHADOW_ARGS pop rbp ret + + +;void vp8_intra_pred_uv_dc_mmx2( +; unsigned char *dst, +; int dst_stride +; unsigned char *src, +; int src_stride, +; ) +global sym(vp8_intra_pred_uv_dc_mmx2) +sym(vp8_intra_pred_uv_dc_mmx2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 4 + push rsi + push rdi + ; end prolog + + ; from top + mov rsi, arg(2) ;src; + movsxd rax, dword ptr arg(3) ;src_stride; + sub rsi, rax + pxor mm0, mm0 + movd mm1, [rsi] + movd mm2, [rsi+4] + punpcklbw mm1, mm0 + punpcklbw mm2, mm0 + paddw mm1, mm2 + pshufw mm2, mm1, 0x0e + paddw mm1, mm2 + pshufw mm2, mm1, 0x01 + paddw mm1, mm2 + + ; from left + dec rsi + lea rdi, [rax*3] + movzx ecx, byte [rsi+rax] + movzx edx, byte [rsi+rax*2] + add ecx, edx + movzx edx, byte [rsi+rdi] + add ecx, edx + lea rsi, [rsi+rax*4] + movzx edx, byte [rsi] + add ecx, edx + movzx edx, byte [rsi+rax] + add ecx, edx + movzx edx, byte [rsi+rax*2] + add ecx, edx + movzx edx, byte [rsi+rdi] + add ecx, edx + movzx edx, byte [rsi+rax*4] + add ecx, edx + + ; add up + pextrw edx, mm1, 0x0 + lea edx, [edx+ecx+8] + sar edx, 4 + movd mm1, edx + pshufw mm1, mm1, 0x0 + packuswb mm1, mm1 + + ; write out + mov rdi, arg(0) ;dst; + movsxd rcx, dword ptr arg(1) ;dst_stride + lea rax, [rcx*3] + + movq [rdi ], mm1 + movq [rdi+rcx ], mm1 + movq [rdi+rcx*2], mm1 + movq [rdi+rax ], mm1 + lea rdi, [rdi+rcx*4] + movq [rdi ], mm1 + movq [rdi+rcx ], mm1 + movq [rdi+rcx*2], mm1 + movq [rdi+rax ], mm1 + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +;void vp8_intra_pred_uv_dctop_mmx2( +; unsigned char *dst, +; int dst_stride +; unsigned char *src, +; int src_stride, +; ) +global sym(vp8_intra_pred_uv_dctop_mmx2) +sym(vp8_intra_pred_uv_dctop_mmx2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 4 + push rsi + push rdi + ; end prolog + + ; from top + mov rsi, arg(2) ;src; + movsxd rax, dword ptr arg(3) ;src_stride; + sub rsi, rax + pxor mm0, mm0 + movd mm1, [rsi] + movd mm2, [rsi+4] + punpcklbw mm1, mm0 + punpcklbw mm2, mm0 + paddw mm1, mm2 + pshufw mm2, mm1, 0x0e + paddw mm1, mm2 + pshufw mm2, mm1, 0x01 + paddw mm1, mm2 + + ; add up + paddw mm1, [GLOBAL(dc_4)] + psraw mm1, 3 + pshufw mm1, mm1, 0x0 + packuswb mm1, mm1 + + ; write out + mov rdi, arg(0) ;dst; + movsxd rcx, dword ptr arg(1) ;dst_stride + lea rax, [rcx*3] + + movq [rdi ], mm1 + movq [rdi+rcx ], mm1 + movq [rdi+rcx*2], mm1 + movq [rdi+rax ], mm1 + lea rdi, [rdi+rcx*4] + movq [rdi ], mm1 + movq [rdi+rcx ], mm1 + movq [rdi+rcx*2], mm1 + movq [rdi+rax ], mm1 + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +;void vp8_intra_pred_uv_dcleft_mmx2( +; unsigned char *dst, +; int dst_stride +; unsigned char *src, +; int src_stride, +; ) +global sym(vp8_intra_pred_uv_dcleft_mmx2) +sym(vp8_intra_pred_uv_dcleft_mmx2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 4 + push rsi + push rdi + ; end prolog + + ; from left + mov rsi, arg(2) ;src; + movsxd rax, dword ptr arg(3) ;src_stride; + dec rsi + lea rdi, [rax*3] + movzx ecx, byte [rsi] + movzx edx, byte [rsi+rax] + add ecx, edx + movzx edx, byte [rsi+rax*2] + add ecx, edx + movzx edx, byte [rsi+rdi] + add ecx, edx + lea rsi, [rsi+rax*4] + movzx edx, byte [rsi] + add ecx, edx + movzx edx, byte [rsi+rax] + add ecx, edx + movzx edx, byte [rsi+rax*2] + add ecx, edx + movzx edx, byte [rsi+rdi] + lea edx, [ecx+edx+4] + + ; add up + shr edx, 3 + movd mm1, edx + pshufw mm1, mm1, 0x0 + packuswb mm1, mm1 + + ; write out + mov rdi, arg(0) ;dst; + movsxd rcx, dword ptr arg(1) ;dst_stride + lea rax, [rcx*3] + + movq [rdi ], mm1 + movq [rdi+rcx ], mm1 + movq [rdi+rcx*2], mm1 + movq [rdi+rax ], mm1 + lea rdi, [rdi+rcx*4] + movq [rdi ], mm1 + movq [rdi+rcx ], mm1 + movq [rdi+rcx*2], mm1 + movq [rdi+rax ], mm1 + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +;void vp8_intra_pred_uv_dc128_mmx( +; unsigned char *dst, +; int dst_stride +; unsigned char *src, +; int src_stride, +; ) +global sym(vp8_intra_pred_uv_dc128_mmx) +sym(vp8_intra_pred_uv_dc128_mmx): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 4 + ; end prolog + + ; write out + movq mm1, [GLOBAL(dc_128)] + mov rax, arg(0) ;dst; + movsxd rdx, dword ptr arg(1) ;dst_stride + lea rcx, [rdx*3] + + movq [rax ], mm1 + movq [rax+rdx ], mm1 + movq [rax+rdx*2], mm1 + movq [rax+rcx ], mm1 + lea rax, [rax+rdx*4] + movq [rax ], mm1 + movq [rax+rdx ], mm1 + movq [rax+rdx*2], mm1 + movq [rax+rcx ], mm1 + + ; begin epilog + UNSHADOW_ARGS + pop rbp + ret + +;void vp8_intra_pred_uv_tm_sse2( +; unsigned char *dst, +; int dst_stride +; unsigned char *src, +; int src_stride, +; ) +%macro vp8_intra_pred_uv_tm 1 +global sym(vp8_intra_pred_uv_tm_%1) +sym(vp8_intra_pred_uv_tm_%1): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 4 + push rsi + push rdi + ; end prolog + + ; read top row + mov edx, 4 + mov rsi, arg(2) ;src; + movsxd rax, dword ptr arg(3) ;src_stride; + sub rsi, rax + pxor xmm0, xmm0 +%ifidn %1, ssse3 + movdqa xmm2, [GLOBAL(dc_1024)] +%endif + movq xmm1, [rsi] + punpcklbw xmm1, xmm0 + + ; set up left ptrs ans subtract topleft + movd xmm3, [rsi-1] + lea rsi, [rsi+rax-1] +%ifidn %1, sse2 + punpcklbw xmm3, xmm0 + pshuflw xmm3, xmm3, 0x0 + punpcklqdq xmm3, xmm3 +%else + pshufb xmm3, xmm2 +%endif + psubw xmm1, xmm3 + + ; set up dest ptrs + mov rdi, arg(0) ;dst; + movsxd rcx, dword ptr arg(1) ;dst_stride + +vp8_intra_pred_uv_tm_%1_loop: + movd xmm3, [rsi] + movd xmm5, [rsi+rax] +%ifidn %1, sse2 + punpcklbw xmm3, xmm0 + punpcklbw xmm5, xmm0 + pshuflw xmm3, xmm3, 0x0 + pshuflw xmm5, xmm5, 0x0 + punpcklqdq xmm3, xmm3 + punpcklqdq xmm5, xmm5 +%else + pshufb xmm3, xmm2 + pshufb xmm5, xmm2 +%endif + paddw xmm3, xmm1 + paddw xmm5, xmm1 + packuswb xmm3, xmm5 + movq [rdi ], xmm3 + movhps[rdi+rcx], xmm3 + lea rsi, [rsi+rax*2] + lea rdi, [rdi+rcx*2] + dec edx + jnz vp8_intra_pred_uv_tm_%1_loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret +%endmacro + +vp8_intra_pred_uv_tm sse2 +vp8_intra_pred_uv_tm ssse3 + +;void vp8_intra_pred_uv_ve_mmx( +; unsigned char *dst, +; int dst_stride +; unsigned char *src, +; int src_stride, +; ) +global sym(vp8_intra_pred_uv_ve_mmx) +sym(vp8_intra_pred_uv_ve_mmx): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 4 + ; end prolog + + ; read from top + mov rax, arg(2) ;src; + movsxd rdx, dword ptr arg(3) ;src_stride; + sub rax, rdx + movq mm1, [rax] + + ; write out + mov rax, arg(0) ;dst; + movsxd rdx, dword ptr arg(1) ;dst_stride + lea rcx, [rdx*3] + + movq [rax ], mm1 + movq [rax+rdx ], mm1 + movq [rax+rdx*2], mm1 + movq [rax+rcx ], mm1 + lea rax, [rax+rdx*4] + movq [rax ], mm1 + movq [rax+rdx ], mm1 + movq [rax+rdx*2], mm1 + movq [rax+rcx ], mm1 + + ; begin epilog + UNSHADOW_ARGS + pop rbp + ret + +;void vp8_intra_pred_uv_ho_mmx2( +; unsigned char *dst, +; int dst_stride +; unsigned char *src, +; int src_stride, +; ) +global sym(vp8_intra_pred_uv_ho_mmx2) +sym(vp8_intra_pred_uv_ho_mmx2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 4 + push rsi + push rdi + ; end prolog + + ; read from left and write out + mov edx, 4 + mov rsi, arg(2) ;src; + movsxd rax, dword ptr arg(3) ;src_stride; + mov rdi, arg(0) ;dst; + movsxd rcx, dword ptr arg(1) ;dst_stride + dec rsi +vp8_intra_pred_uv_ho_mmx2_loop: + movd mm0, [rsi] + movd mm1, [rsi+rax] + punpcklbw mm0, mm0 + punpcklbw mm1, mm1 + pshufw mm0, mm0, 0x0 + pshufw mm1, mm1, 0x0 + movq [rdi ], mm0 + movq [rdi+rcx], mm1 + lea rsi, [rsi+rax*2] + lea rdi, [rdi+rcx*2] + dec edx + jnz vp8_intra_pred_uv_ho_mmx2_loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +SECTION_RODATA +dc_128: + times 8 db 128 +dc_4: + times 4 dw 4 +align 16 +dc_1024: + times 8 dw 0x400 -- cgit v1.2.3