64-bit SIMD acceleration
git-svn-id: svn://svn.code.sf.net/p/tigervnc/code/trunk@3858 3789f03b-4d11-0410-bbf8-ca57d06f2519
diff --git a/common/jpeg/simd/jcsamss2-64.asm b/common/jpeg/simd/jcsamss2-64.asm
new file mode 100644
index 0000000..cf7a776
--- /dev/null
+++ b/common/jpeg/simd/jcsamss2-64.asm
@@ -0,0 +1,324 @@
+;
+; jcsamss2.asm - downsampling (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+ SECTION SEG_TEXT
+ BITS 64
+;
+; Downsample pixel values of a single component.
+; This version handles the common case of 2:1 horizontal and 1:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v1_downsample_sse2 (JDIMENSION image_width, int max_v_samp_factor,
+; JDIMENSION v_samp_factor, JDIMENSION width_blocks,
+; JSAMPARRAY input_data, JSAMPARRAY output_data);
+;
+
+; r10 = JDIMENSION image_width
+; r11 = int max_v_samp_factor
+; r12 = JDIMENSION v_samp_factor
+; r13 = JDIMENSION width_blocks
+; r14 = JSAMPARRAY input_data
+; r15 = JSAMPARRAY output_data
+
+ align 16
+ global EXTN(jsimd_h2v1_downsample_sse2)
+
+EXTN(jsimd_h2v1_downsample_sse2):
+ push rbp
+ mov rbp,rsp
+ collect_args
+
+ mov rcx, r13
+ shl rcx,3 ; imul rcx,DCTSIZE (rcx = output_cols)
+ jz near .return
+
+ mov rdx, r10
+
+ ; -- expand_right_edge
+
+ push rcx
+ shl rcx,1 ; output_cols * 2
+ sub rcx,rdx
+ jle short .expand_end
+
+ mov rax, r11
+ test rax,rax
+ jle short .expand_end
+
+ cld
+ mov rsi, r14 ; input_data
+.expandloop:
+ push rax
+ push rcx
+
+ mov rdi, JSAMPROW [rsi]
+ add rdi,rdx
+ mov al, JSAMPLE [rdi-1]
+
+ rep stosb
+
+ pop rcx
+ pop rax
+
+ add rsi, byte SIZEOF_JSAMPROW
+ dec rax
+ jg short .expandloop
+
+.expand_end:
+ pop rcx ; output_cols
+
+ ; -- h2v1_downsample
+
+ mov rax, r12 ; rowctr
+ test eax,eax
+ jle near .return
+
+ mov rdx, 0x00010000 ; bias pattern
+ movd xmm7,edx
+ pcmpeqw xmm6,xmm6
+ pshufd xmm7,xmm7,0x00 ; xmm7={0, 1, 0, 1, 0, 1, 0, 1}
+ psrlw xmm6,BYTE_BIT ; xmm6={0xFF 0x00 0xFF 0x00 ..}
+
+ mov rsi, r14 ; input_data
+ mov rdi, r15 ; output_data
+.rowloop:
+ push rcx
+ push rdi
+ push rsi
+
+ mov rsi, JSAMPROW [rsi] ; inptr
+ mov rdi, JSAMPROW [rdi] ; outptr
+
+ cmp rcx, byte SIZEOF_XMMWORD
+ jae short .columnloop
+
+.columnloop_r8:
+ movdqa xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+ pxor xmm1,xmm1
+ mov rcx, SIZEOF_XMMWORD
+ jmp short .downsample
+
+.columnloop:
+ movdqa xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+ movdqa xmm1, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+
+.downsample:
+ movdqa xmm2,xmm0
+ movdqa xmm3,xmm1
+
+ pand xmm0,xmm6
+ psrlw xmm2,BYTE_BIT
+ pand xmm1,xmm6
+ psrlw xmm3,BYTE_BIT
+
+ paddw xmm0,xmm2
+ paddw xmm1,xmm3
+ paddw xmm0,xmm7
+ paddw xmm1,xmm7
+ psrlw xmm0,1
+ psrlw xmm1,1
+
+ packuswb xmm0,xmm1
+
+ movdqa XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
+
+ sub rcx, byte SIZEOF_XMMWORD ; outcol
+ add rsi, byte 2*SIZEOF_XMMWORD ; inptr
+ add rdi, byte 1*SIZEOF_XMMWORD ; outptr
+ cmp rcx, byte SIZEOF_XMMWORD
+ jae short .columnloop
+ test rcx,rcx
+ jnz short .columnloop_r8
+
+ pop rsi
+ pop rdi
+ pop rcx
+
+ add rsi, byte SIZEOF_JSAMPROW ; input_data
+ add rdi, byte SIZEOF_JSAMPROW ; output_data
+ dec rax ; rowctr
+ jg near .rowloop
+
+.return:
+ uncollect_args
+ pop rbp
+ ret
+
+; --------------------------------------------------------------------------
+;
+; Downsample pixel values of a single component.
+; This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v2_downsample_sse2 (JDIMENSION image_width, int max_v_samp_factor,
+; JDIMENSION v_samp_factor, JDIMENSION width_blocks,
+; JSAMPARRAY input_data, JSAMPARRAY output_data);
+;
+
+; r10 = JDIMENSION image_width
+; r11 = int max_v_samp_factor
+; r12 = JDIMENSION v_samp_factor
+; r13 = JDIMENSION width_blocks
+; r14 = JSAMPARRAY input_data
+; r15 = JSAMPARRAY output_data
+
+ align 16
+ global EXTN(jsimd_h2v2_downsample_sse2)
+
+EXTN(jsimd_h2v2_downsample_sse2):
+ push rbp
+ mov rbp,rsp
+ collect_args
+
+ mov rcx, r13
+ shl rcx,3 ; imul rcx,DCTSIZE (rcx = output_cols)
+ jz near .return
+
+ mov rdx, r10
+
+ ; -- expand_right_edge
+
+ push rcx
+ shl rcx,1 ; output_cols * 2
+ sub rcx,rdx
+ jle short .expand_end
+
+ mov rax, r11
+ test rax,rax
+ jle short .expand_end
+
+ cld
+ mov rsi, r14 ; input_data
+.expandloop:
+ push rax
+ push rcx
+
+ mov rdi, JSAMPROW [rsi]
+ add rdi,rdx
+ mov al, JSAMPLE [rdi-1]
+
+ rep stosb
+
+ pop rcx
+ pop rax
+
+ add rsi, byte SIZEOF_JSAMPROW
+ dec rax
+ jg short .expandloop
+
+.expand_end:
+ pop rcx ; output_cols
+
+ ; -- h2v2_downsample
+
+ mov rax, r12 ; rowctr
+ test rax,rax
+ jle near .return
+
+ mov rdx, 0x00020001 ; bias pattern
+ movd xmm7,edx
+ pcmpeqw xmm6,xmm6
+ pshufd xmm7,xmm7,0x00 ; xmm7={1, 2, 1, 2, 1, 2, 1, 2}
+ psrlw xmm6,BYTE_BIT ; xmm6={0xFF 0x00 0xFF 0x00 ..}
+
+ mov rsi, r14 ; input_data
+ mov rdi, r15 ; output_data
+.rowloop:
+ push rcx
+ push rdi
+ push rsi
+
+ mov rdx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; inptr0
+ mov rsi, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; inptr1
+ mov rdi, JSAMPROW [rdi] ; outptr
+
+ cmp rcx, byte SIZEOF_XMMWORD
+ jae short .columnloop
+
+.columnloop_r8:
+ movdqa xmm0, XMMWORD [rdx+0*SIZEOF_XMMWORD]
+ movdqa xmm1, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+ pxor xmm2,xmm2
+ pxor xmm3,xmm3
+ mov rcx, SIZEOF_XMMWORD
+ jmp short .downsample
+
+.columnloop:
+ movdqa xmm0, XMMWORD [rdx+0*SIZEOF_XMMWORD]
+ movdqa xmm1, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+ movdqa xmm2, XMMWORD [rdx+1*SIZEOF_XMMWORD]
+ movdqa xmm3, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+
+.downsample:
+ movdqa xmm4,xmm0
+ movdqa xmm5,xmm1
+ pand xmm0,xmm6
+ psrlw xmm4,BYTE_BIT
+ pand xmm1,xmm6
+ psrlw xmm5,BYTE_BIT
+ paddw xmm0,xmm4
+ paddw xmm1,xmm5
+
+ movdqa xmm4,xmm2
+ movdqa xmm5,xmm3
+ pand xmm2,xmm6
+ psrlw xmm4,BYTE_BIT
+ pand xmm3,xmm6
+ psrlw xmm5,BYTE_BIT
+ paddw xmm2,xmm4
+ paddw xmm3,xmm5
+
+ paddw xmm0,xmm1
+ paddw xmm2,xmm3
+ paddw xmm0,xmm7
+ paddw xmm2,xmm7
+ psrlw xmm0,2
+ psrlw xmm2,2
+
+ packuswb xmm0,xmm2
+
+ movdqa XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
+
+ sub rcx, byte SIZEOF_XMMWORD ; outcol
+ add rdx, byte 2*SIZEOF_XMMWORD ; inptr0
+ add rsi, byte 2*SIZEOF_XMMWORD ; inptr1
+ add rdi, byte 1*SIZEOF_XMMWORD ; outptr
+ cmp rcx, byte SIZEOF_XMMWORD
+ jae near .columnloop
+ test rcx,rcx
+ jnz near .columnloop_r8
+
+ pop rsi
+ pop rdi
+ pop rcx
+
+ add rsi, byte 2*SIZEOF_JSAMPROW ; input_data
+ add rdi, byte 1*SIZEOF_JSAMPROW ; output_data
+ dec rax ; rowctr
+ jg near .rowloop
+
+.return:
+ uncollect_args
+ pop rbp
+ ret