SkJumper: use AVX2 mask loads and stores for U32
SkRasterPipeline_f16: 63 -> 58 (8888+f16 loads, f16 store) SkRasterPipeline_srgb: 96 -> 84 (2x 8888 loads, 8888 store) PS3 has a simpler way to build the mask, in a uint64_t. Timing is still roughlt the same. Change-Id: Ie278611dff02281e5a0f3a57185050bbe852bff0 Reviewed-on: https://skia-review.googlesource.com/9165 Commit-Queue: Mike Klein <mtklein@chromium.org> Reviewed-by: Herb Derby <herb@google.com>
This commit is contained in:
Mike Klein
parent
8e48c1e1d3
commit
767c7e7a0b
@ -2300,18 +2300,20 @@ _sk_lerp_565_hsw:
|
||||
|
||||
.globl _sk_load_tables_hsw
|
||||
_sk_load_tables_hsw:
|
||||
.byte 73,137,200 // mov %rcx,%r8
|
||||
.byte 72,173 // lods %ds:(%rsi),%rax
|
||||
.byte 76,139,0 // mov (%rax),%r8
|
||||
.byte 72,133,201 // test %rcx,%rcx
|
||||
.byte 117,104 // jne 666 <_sk_load_tables_hsw+0x72>
|
||||
.byte 196,193,126,111,28,184 // vmovdqu (%r8,%rdi,4),%ymm3
|
||||
.byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
|
||||
.byte 76,3,8 // add (%rax),%r9
|
||||
.byte 77,133,192 // test %r8,%r8
|
||||
.byte 117,106 // jne 673 <_sk_load_tables_hsw+0x7f>
|
||||
.byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3
|
||||
.byte 196,226,125,88,82,16 // vpbroadcastd 0x10(%rdx),%ymm2
|
||||
.byte 197,237,219,203 // vpand %ymm3,%ymm2,%ymm1
|
||||
.byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8
|
||||
.byte 76,139,64,8 // mov 0x8(%rax),%r8
|
||||
.byte 72,139,72,8 // mov 0x8(%rax),%rcx
|
||||
.byte 76,139,72,16 // mov 0x10(%rax),%r9
|
||||
.byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9
|
||||
.byte 196,194,53,146,4,136 // vgatherdps %ymm9,(%r8,%ymm1,4),%ymm0
|
||||
.byte 196,226,53,146,4,137 // vgatherdps %ymm9,(%rcx,%ymm1,4),%ymm0
|
||||
.byte 197,245,114,211,8 // vpsrld $0x8,%ymm3,%ymm1
|
||||
.byte 197,109,219,201 // vpand %ymm1,%ymm2,%ymm9
|
||||
.byte 196,65,45,118,210 // vpcmpeqd %ymm10,%ymm10,%ymm10
|
||||
@ -2325,56 +2327,17 @@ _sk_load_tables_hsw:
|
||||
.byte 196,98,125,24,66,12 // vbroadcastss 0xc(%rdx),%ymm8
|
||||
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
|
||||
.byte 72,173 // lods %ds:(%rsi),%rax
|
||||
.byte 76,137,193 // mov %r8,%rcx
|
||||
.byte 255,224 // jmpq *%rax
|
||||
.byte 65,137,201 // mov %ecx,%r9d
|
||||
.byte 65,128,225,7 // and $0x7,%r9b
|
||||
.byte 197,229,239,219 // vpxor %ymm3,%ymm3,%ymm3
|
||||
.byte 65,254,201 // dec %r9b
|
||||
.byte 69,15,182,201 // movzbl %r9b,%r9d
|
||||
.byte 65,128,249,6 // cmp $0x6,%r9b
|
||||
.byte 119,134 // ja 604 <_sk_load_tables_hsw+0x10>
|
||||
.byte 76,141,21,131,0,0,0 // lea 0x83(%rip),%r10 # 708 <_sk_load_tables_hsw+0x114>
|
||||
.byte 79,99,12,138 // movslq (%r10,%r9,4),%r9
|
||||
.byte 77,1,209 // add %r10,%r9
|
||||
.byte 65,255,225 // jmpq *%r9
|
||||
.byte 196,193,121,110,68,184,24 // vmovd 0x18(%r8,%rdi,4),%xmm0
|
||||
.byte 196,226,125,89,192 // vpbroadcastq %xmm0,%ymm0
|
||||
.byte 197,245,239,201 // vpxor %ymm1,%ymm1,%ymm1
|
||||
.byte 196,227,117,2,216,64 // vpblendd $0x40,%ymm0,%ymm1,%ymm3
|
||||
.byte 196,227,125,57,216,1 // vextracti128 $0x1,%ymm3,%xmm0
|
||||
.byte 196,195,121,34,68,184,20,1 // vpinsrd $0x1,0x14(%r8,%rdi,4),%xmm0,%xmm0
|
||||
.byte 196,227,101,56,216,1 // vinserti128 $0x1,%xmm0,%ymm3,%ymm3
|
||||
.byte 196,227,125,57,216,1 // vextracti128 $0x1,%ymm3,%xmm0
|
||||
.byte 196,195,121,34,68,184,16,0 // vpinsrd $0x0,0x10(%r8,%rdi,4),%xmm0,%xmm0
|
||||
.byte 196,227,101,56,216,1 // vinserti128 $0x1,%xmm0,%ymm3,%ymm3
|
||||
.byte 196,195,97,34,68,184,12,3 // vpinsrd $0x3,0xc(%r8,%rdi,4),%xmm3,%xmm0
|
||||
.byte 196,227,101,2,216,15 // vpblendd $0xf,%ymm0,%ymm3,%ymm3
|
||||
.byte 196,195,97,34,68,184,8,2 // vpinsrd $0x2,0x8(%r8,%rdi,4),%xmm3,%xmm0
|
||||
.byte 196,227,101,2,216,15 // vpblendd $0xf,%ymm0,%ymm3,%ymm3
|
||||
.byte 196,195,97,34,68,184,4,1 // vpinsrd $0x1,0x4(%r8,%rdi,4),%xmm3,%xmm0
|
||||
.byte 196,227,101,2,216,15 // vpblendd $0xf,%ymm0,%ymm3,%ymm3
|
||||
.byte 196,193,121,110,4,184 // vmovd (%r8,%rdi,4),%xmm0
|
||||
.byte 196,227,101,2,216,1 // vpblendd $0x1,%ymm0,%ymm3,%ymm3
|
||||
.byte 233,252,254,255,255 // jmpq 604 <_sk_load_tables_hsw+0x10>
|
||||
.byte 239 // out %eax,(%dx)
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255,225 // jmpq *%rcx
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255,211 // callq *%rbx
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255,197 // inc %ebp
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255,177,255,255,255,157 // pushq -0x62000001(%rcx)
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // .byte 0xff
|
||||
.byte 135,255 // xchg %edi,%edi
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // .byte 0xff
|
||||
.byte 185,8,0,0,0 // mov $0x8,%ecx
|
||||
.byte 68,41,193 // sub %r8d,%ecx
|
||||
.byte 192,225,3 // shl $0x3,%cl
|
||||
.byte 73,199,194,255,255,255,255 // mov $0xffffffffffffffff,%r10
|
||||
.byte 73,211,234 // shr %cl,%r10
|
||||
.byte 196,193,249,110,194 // vmovq %r10,%xmm0
|
||||
.byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0
|
||||
.byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3
|
||||
.byte 233,114,255,255,255 // jmpq 60e <_sk_load_tables_hsw+0x1a>
|
||||
|
||||
.globl _sk_load_a8_hsw
|
||||
_sk_load_a8_hsw:
|
||||
@ -2383,7 +2346,7 @@ _sk_load_a8_hsw:
|
||||
.byte 72,139,0 // mov (%rax),%rax
|
||||
.byte 72,1,248 // add %rdi,%rax
|
||||
.byte 77,133,192 // test %r8,%r8
|
||||
.byte 117,42 // jne 75e <_sk_load_a8_hsw+0x3a>
|
||||
.byte 117,42 // jne 6d6 <_sk_load_a8_hsw+0x3a>
|
||||
.byte 197,251,16,0 // vmovsd (%rax),%xmm0
|
||||
.byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
|
||||
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
|
||||
@ -2404,9 +2367,9 @@ _sk_load_a8_hsw:
|
||||
.byte 77,9,217 // or %r11,%r9
|
||||
.byte 72,131,193,8 // add $0x8,%rcx
|
||||
.byte 73,255,202 // dec %r10
|
||||
.byte 117,234 // jne 766 <_sk_load_a8_hsw+0x42>
|
||||
.byte 117,234 // jne 6de <_sk_load_a8_hsw+0x42>
|
||||
.byte 196,193,249,110,193 // vmovq %r9,%xmm0
|
||||
.byte 235,181 // jmp 738 <_sk_load_a8_hsw+0x14>
|
||||
.byte 235,181 // jmp 6b0 <_sk_load_a8_hsw+0x14>
|
||||
|
||||
.globl _sk_store_a8_hsw
|
||||
_sk_store_a8_hsw:
|
||||
@ -2419,7 +2382,7 @@ _sk_store_a8_hsw:
|
||||
.byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
|
||||
.byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8
|
||||
.byte 72,133,201 // test %rcx,%rcx
|
||||
.byte 117,10 // jne 7b6 <_sk_store_a8_hsw+0x33>
|
||||
.byte 117,10 // jne 72e <_sk_store_a8_hsw+0x33>
|
||||
.byte 196,65,123,17,4,57 // vmovsd %xmm8,(%r9,%rdi,1)
|
||||
.byte 72,173 // lods %ds:(%rsi),%rax
|
||||
.byte 255,224 // jmpq *%rax
|
||||
@ -2428,9 +2391,9 @@ _sk_store_a8_hsw:
|
||||
.byte 254,200 // dec %al
|
||||
.byte 68,15,182,192 // movzbl %al,%r8d
|
||||
.byte 65,128,248,6 // cmp $0x6,%r8b
|
||||
.byte 119,236 // ja 7b2 <_sk_store_a8_hsw+0x2f>
|
||||
.byte 119,236 // ja 72a <_sk_store_a8_hsw+0x2f>
|
||||
.byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8
|
||||
.byte 76,141,21,66,0,0,0 // lea 0x42(%rip),%r10 # 814 <_sk_store_a8_hsw+0x91>
|
||||
.byte 76,141,21,66,0,0,0 // lea 0x42(%rip),%r10 # 78c <_sk_store_a8_hsw+0x91>
|
||||
.byte 75,99,4,130 // movslq (%r10,%r8,4),%rax
|
||||
.byte 76,1,208 // add %r10,%rax
|
||||
.byte 255,224 // jmpq *%rax
|
||||
@ -2441,7 +2404,7 @@ _sk_store_a8_hsw:
|
||||
.byte 196,67,121,20,68,57,2,4 // vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1)
|
||||
.byte 196,67,121,20,68,57,1,2 // vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1)
|
||||
.byte 196,67,121,20,4,57,0 // vpextrb $0x0,%xmm8,(%r9,%rdi,1)
|
||||
.byte 235,158 // jmp 7b2 <_sk_store_a8_hsw+0x2f>
|
||||
.byte 235,158 // jmp 72a <_sk_store_a8_hsw+0x2f>
|
||||
.byte 247,255 // idiv %edi
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
@ -2470,7 +2433,7 @@ _sk_load_565_hsw:
|
||||
.byte 72,173 // lods %ds:(%rsi),%rax
|
||||
.byte 76,139,16 // mov (%rax),%r10
|
||||
.byte 72,133,201 // test %rcx,%rcx
|
||||
.byte 117,92 // jne 896 <_sk_load_565_hsw+0x66>
|
||||
.byte 117,92 // jne 80e <_sk_load_565_hsw+0x66>
|
||||
.byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0
|
||||
.byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2
|
||||
.byte 196,226,125,88,66,104 // vpbroadcastd 0x68(%rdx),%ymm0
|
||||
@ -2497,8 +2460,8 @@ _sk_load_565_hsw:
|
||||
.byte 65,254,200 // dec %r8b
|
||||
.byte 69,15,182,192 // movzbl %r8b,%r8d
|
||||
.byte 65,128,248,6 // cmp $0x6,%r8b
|
||||
.byte 119,146 // ja 840 <_sk_load_565_hsw+0x10>
|
||||
.byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 900 <_sk_load_565_hsw+0xd0>
|
||||
.byte 119,146 // ja 7b8 <_sk_load_565_hsw+0x10>
|
||||
.byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 878 <_sk_load_565_hsw+0xd0>
|
||||
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
|
||||
.byte 76,1,200 // add %r9,%rax
|
||||
.byte 255,224 // jmpq *%rax
|
||||
@ -2510,7 +2473,7 @@ _sk_load_565_hsw:
|
||||
.byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
|
||||
.byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
|
||||
.byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
|
||||
.byte 233,66,255,255,255 // jmpq 840 <_sk_load_565_hsw+0x10>
|
||||
.byte 233,66,255,255,255 // jmpq 7b8 <_sk_load_565_hsw+0x10>
|
||||
.byte 102,144 // xchg %ax,%ax
|
||||
.byte 242,255 // repnz (bad)
|
||||
.byte 255 // (bad)
|
||||
@ -2555,7 +2518,7 @@ _sk_store_565_hsw:
|
||||
.byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9
|
||||
.byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
|
||||
.byte 72,133,201 // test %rcx,%rcx
|
||||
.byte 117,10 // jne 97e <_sk_store_565_hsw+0x62>
|
||||
.byte 117,10 // jne 8f6 <_sk_store_565_hsw+0x62>
|
||||
.byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2)
|
||||
.byte 72,173 // lods %ds:(%rsi),%rax
|
||||
.byte 255,224 // jmpq *%rax
|
||||
@ -2564,8 +2527,8 @@ _sk_store_565_hsw:
|
||||
.byte 254,200 // dec %al
|
||||
.byte 68,15,182,192 // movzbl %al,%r8d
|
||||
.byte 65,128,248,6 // cmp $0x6,%r8b
|
||||
.byte 119,236 // ja 97a <_sk_store_565_hsw+0x5e>
|
||||
.byte 76,141,21,71,0,0,0 // lea 0x47(%rip),%r10 # 9dc <_sk_store_565_hsw+0xc0>
|
||||
.byte 119,236 // ja 8f2 <_sk_store_565_hsw+0x5e>
|
||||
.byte 76,141,21,71,0,0,0 // lea 0x47(%rip),%r10 # 954 <_sk_store_565_hsw+0xc0>
|
||||
.byte 75,99,4,130 // movslq (%r10,%r8,4),%rax
|
||||
.byte 76,1,208 // add %r10,%rax
|
||||
.byte 255,224 // jmpq *%rax
|
||||
@ -2577,7 +2540,7 @@ _sk_store_565_hsw:
|
||||
.byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
|
||||
.byte 197,121,126,192 // vmovd %xmm8,%eax
|
||||
.byte 102,65,137,4,121 // mov %ax,(%r9,%rdi,2)
|
||||
.byte 235,161 // jmp 97a <_sk_store_565_hsw+0x5e>
|
||||
.byte 235,161 // jmp 8f2 <_sk_store_565_hsw+0x5e>
|
||||
.byte 15,31,0 // nopl (%rax)
|
||||
.byte 242,255 // repnz (bad)
|
||||
.byte 255 // (bad)
|
||||
@ -2604,11 +2567,13 @@ _sk_store_565_hsw:
|
||||
|
||||
.globl _sk_load_8888_hsw
|
||||
_sk_load_8888_hsw:
|
||||
.byte 73,137,200 // mov %rcx,%r8
|
||||
.byte 72,173 // lods %ds:(%rsi),%rax
|
||||
.byte 76,139,16 // mov (%rax),%r10
|
||||
.byte 72,133,201 // test %rcx,%rcx
|
||||
.byte 117,83 // jne a55 <_sk_load_8888_hsw+0x5d>
|
||||
.byte 196,193,126,111,28,186 // vmovdqu (%r10,%rdi,4),%ymm3
|
||||
.byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
|
||||
.byte 76,3,8 // add (%rax),%r9
|
||||
.byte 77,133,192 // test %r8,%r8
|
||||
.byte 117,85 // jne 9da <_sk_load_8888_hsw+0x6a>
|
||||
.byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3
|
||||
.byte 196,226,125,88,82,16 // vpbroadcastd 0x10(%rdx),%ymm2
|
||||
.byte 197,237,219,195 // vpand %ymm3,%ymm2,%ymm0
|
||||
.byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
|
||||
@ -2626,62 +2591,24 @@ _sk_load_8888_hsw:
|
||||
.byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
|
||||
.byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
|
||||
.byte 72,173 // lods %ds:(%rsi),%rax
|
||||
.byte 76,137,193 // mov %r8,%rcx
|
||||
.byte 255,224 // jmpq *%rax
|
||||
.byte 65,137,200 // mov %ecx,%r8d
|
||||
.byte 65,128,224,7 // and $0x7,%r8b
|
||||
.byte 197,229,239,219 // vpxor %ymm3,%ymm3,%ymm3
|
||||
.byte 65,254,200 // dec %r8b
|
||||
.byte 69,15,182,192 // movzbl %r8b,%r8d
|
||||
.byte 65,128,248,6 // cmp $0x6,%r8b
|
||||
.byte 119,155 // ja a08 <_sk_load_8888_hsw+0x10>
|
||||
.byte 76,141,13,132,0,0,0 // lea 0x84(%rip),%r9 # af8 <_sk_load_8888_hsw+0x100>
|
||||
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
|
||||
.byte 76,1,200 // add %r9,%rax
|
||||
.byte 255,224 // jmpq *%rax
|
||||
.byte 196,193,121,110,68,186,24 // vmovd 0x18(%r10,%rdi,4),%xmm0
|
||||
.byte 196,226,125,89,192 // vpbroadcastq %xmm0,%ymm0
|
||||
.byte 197,245,239,201 // vpxor %ymm1,%ymm1,%ymm1
|
||||
.byte 196,227,117,2,216,64 // vpblendd $0x40,%ymm0,%ymm1,%ymm3
|
||||
.byte 196,227,125,57,216,1 // vextracti128 $0x1,%ymm3,%xmm0
|
||||
.byte 196,195,121,34,68,186,20,1 // vpinsrd $0x1,0x14(%r10,%rdi,4),%xmm0,%xmm0
|
||||
.byte 196,227,101,56,216,1 // vinserti128 $0x1,%xmm0,%ymm3,%ymm3
|
||||
.byte 196,227,125,57,216,1 // vextracti128 $0x1,%ymm3,%xmm0
|
||||
.byte 196,195,121,34,68,186,16,0 // vpinsrd $0x0,0x10(%r10,%rdi,4),%xmm0,%xmm0
|
||||
.byte 196,227,101,56,216,1 // vinserti128 $0x1,%xmm0,%ymm3,%ymm3
|
||||
.byte 196,195,97,34,68,186,12,3 // vpinsrd $0x3,0xc(%r10,%rdi,4),%xmm3,%xmm0
|
||||
.byte 196,227,101,2,216,15 // vpblendd $0xf,%ymm0,%ymm3,%ymm3
|
||||
.byte 196,195,97,34,68,186,8,2 // vpinsrd $0x2,0x8(%r10,%rdi,4),%xmm3,%xmm0
|
||||
.byte 196,227,101,2,216,15 // vpblendd $0xf,%ymm0,%ymm3,%ymm3
|
||||
.byte 196,195,97,34,68,186,4,1 // vpinsrd $0x1,0x4(%r10,%rdi,4),%xmm3,%xmm0
|
||||
.byte 196,227,101,2,216,15 // vpblendd $0xf,%ymm0,%ymm3,%ymm3
|
||||
.byte 196,193,121,110,4,186 // vmovd (%r10,%rdi,4),%xmm0
|
||||
.byte 196,227,101,2,216,1 // vpblendd $0x1,%ymm0,%ymm3,%ymm3
|
||||
.byte 233,18,255,255,255 // jmpq a08 <_sk_load_8888_hsw+0x10>
|
||||
.byte 102,144 // xchg %ax,%ax
|
||||
.byte 237 // in (%dx),%eax
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 223,255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255,209 // callq *%rcx
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255,195 // inc %ebx
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255,175,255,255,255,155 // ljmp *-0x64000001(%rdi)
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // .byte 0xff
|
||||
.byte 133,255 // test %edi,%edi
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // .byte 0xff
|
||||
.byte 185,8,0,0,0 // mov $0x8,%ecx
|
||||
.byte 68,41,193 // sub %r8d,%ecx
|
||||
.byte 192,225,3 // shl $0x3,%cl
|
||||
.byte 72,199,192,255,255,255,255 // mov $0xffffffffffffffff,%rax
|
||||
.byte 72,211,232 // shr %cl,%rax
|
||||
.byte 196,225,249,110,192 // vmovq %rax,%xmm0
|
||||
.byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0
|
||||
.byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3
|
||||
.byte 235,138 // jmp 98a <_sk_load_8888_hsw+0x1a>
|
||||
|
||||
.globl _sk_store_8888_hsw
|
||||
_sk_store_8888_hsw:
|
||||
.byte 73,137,200 // mov %rcx,%r8
|
||||
.byte 72,173 // lods %ds:(%rsi),%rax
|
||||
.byte 76,139,8 // mov (%rax),%r9
|
||||
.byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
|
||||
.byte 76,3,8 // add (%rax),%r9
|
||||
.byte 196,98,125,24,66,8 // vbroadcastss 0x8(%rdx),%ymm8
|
||||
.byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
|
||||
.byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
|
||||
@ -2697,59 +2624,28 @@ _sk_store_8888_hsw:
|
||||
.byte 196,193,61,114,240,24 // vpslld $0x18,%ymm8,%ymm8
|
||||
.byte 196,65,45,235,192 // vpor %ymm8,%ymm10,%ymm8
|
||||
.byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8
|
||||
.byte 72,133,201 // test %rcx,%rcx
|
||||
.byte 117,10 // jne b73 <_sk_store_8888_hsw+0x5f>
|
||||
.byte 196,65,126,127,4,185 // vmovdqu %ymm8,(%r9,%rdi,4)
|
||||
.byte 77,133,192 // test %r8,%r8
|
||||
.byte 117,12 // jne a6c <_sk_store_8888_hsw+0x6c>
|
||||
.byte 196,65,126,127,1 // vmovdqu %ymm8,(%r9)
|
||||
.byte 72,173 // lods %ds:(%rsi),%rax
|
||||
.byte 76,137,193 // mov %r8,%rcx
|
||||
.byte 255,224 // jmpq *%rax
|
||||
.byte 137,200 // mov %ecx,%eax
|
||||
.byte 36,7 // and $0x7,%al
|
||||
.byte 254,200 // dec %al
|
||||
.byte 68,15,182,192 // movzbl %al,%r8d
|
||||
.byte 65,128,248,6 // cmp $0x6,%r8b
|
||||
.byte 119,236 // ja b6f <_sk_store_8888_hsw+0x5b>
|
||||
.byte 76,141,21,82,0,0,0 // lea 0x52(%rip),%r10 # bdc <_sk_store_8888_hsw+0xc8>
|
||||
.byte 75,99,4,130 // movslq (%r10,%r8,4),%rax
|
||||
.byte 76,1,208 // add %r10,%rax
|
||||
.byte 255,224 // jmpq *%rax
|
||||
.byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9
|
||||
.byte 196,67,121,22,76,185,24,2 // vpextrd $0x2,%xmm9,0x18(%r9,%rdi,4)
|
||||
.byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9
|
||||
.byte 196,67,121,22,76,185,20,1 // vpextrd $0x1,%xmm9,0x14(%r9,%rdi,4)
|
||||
.byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9
|
||||
.byte 196,65,121,126,76,185,16 // vmovd %xmm9,0x10(%r9,%rdi,4)
|
||||
.byte 196,67,121,22,68,185,12,3 // vpextrd $0x3,%xmm8,0xc(%r9,%rdi,4)
|
||||
.byte 196,67,121,22,68,185,8,2 // vpextrd $0x2,%xmm8,0x8(%r9,%rdi,4)
|
||||
.byte 196,67,121,22,68,185,4,1 // vpextrd $0x1,%xmm8,0x4(%r9,%rdi,4)
|
||||
.byte 196,65,121,126,4,185 // vmovd %xmm8,(%r9,%rdi,4)
|
||||
.byte 235,147 // jmp b6f <_sk_store_8888_hsw+0x5b>
|
||||
.byte 248 // clc
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255,240 // push %rax
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 232,255,255,255,224 // callq ffffffffe1000be8 <_sk_linear_gradient_2stops_hsw+0xffffffffe0fffbf2>
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255,211 // callq *%rbx
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255,197 // inc %ebp
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // .byte 0xff
|
||||
.byte 183,255 // mov $0xff,%bh
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // .byte 0xff
|
||||
.byte 185,8,0,0,0 // mov $0x8,%ecx
|
||||
.byte 68,41,193 // sub %r8d,%ecx
|
||||
.byte 192,225,3 // shl $0x3,%cl
|
||||
.byte 72,199,192,255,255,255,255 // mov $0xffffffffffffffff,%rax
|
||||
.byte 72,211,232 // shr %cl,%rax
|
||||
.byte 196,97,249,110,200 // vmovq %rax,%xmm9
|
||||
.byte 196,66,125,33,201 // vpmovsxbd %xmm9,%ymm9
|
||||
.byte 196,66,53,142,1 // vpmaskmovd %ymm8,%ymm9,(%r9)
|
||||
.byte 235,211 // jmp a65 <_sk_store_8888_hsw+0x65>
|
||||
|
||||
.globl _sk_load_f16_hsw
|
||||
_sk_load_f16_hsw:
|
||||
.byte 72,173 // lods %ds:(%rsi),%rax
|
||||
.byte 72,139,0 // mov (%rax),%rax
|
||||
.byte 72,133,201 // test %rcx,%rcx
|
||||
.byte 117,97 // jne c63 <_sk_load_f16_hsw+0x6b>
|
||||
.byte 117,97 // jne afd <_sk_load_f16_hsw+0x6b>
|
||||
.byte 197,249,16,12,248 // vmovupd (%rax,%rdi,8),%xmm1
|
||||
.byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2
|
||||
.byte 197,249,16,92,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm3
|
||||
@ -2775,35 +2671,35 @@ _sk_load_f16_hsw:
|
||||
.byte 197,251,16,12,248 // vmovsd (%rax,%rdi,8),%xmm1
|
||||
.byte 196,65,57,87,192 // vxorpd %xmm8,%xmm8,%xmm8
|
||||
.byte 72,131,249,1 // cmp $0x1,%rcx
|
||||
.byte 117,6 // jne c79 <_sk_load_f16_hsw+0x81>
|
||||
.byte 117,6 // jne b13 <_sk_load_f16_hsw+0x81>
|
||||
.byte 197,250,126,201 // vmovq %xmm1,%xmm1
|
||||
.byte 235,30 // jmp c97 <_sk_load_f16_hsw+0x9f>
|
||||
.byte 235,30 // jmp b31 <_sk_load_f16_hsw+0x9f>
|
||||
.byte 197,241,22,76,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm1,%xmm1
|
||||
.byte 72,131,249,3 // cmp $0x3,%rcx
|
||||
.byte 114,18 // jb c97 <_sk_load_f16_hsw+0x9f>
|
||||
.byte 114,18 // jb b31 <_sk_load_f16_hsw+0x9f>
|
||||
.byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2
|
||||
.byte 72,131,249,3 // cmp $0x3,%rcx
|
||||
.byte 117,19 // jne ca4 <_sk_load_f16_hsw+0xac>
|
||||
.byte 117,19 // jne b3e <_sk_load_f16_hsw+0xac>
|
||||
.byte 197,250,126,210 // vmovq %xmm2,%xmm2
|
||||
.byte 235,46 // jmp cc5 <_sk_load_f16_hsw+0xcd>
|
||||
.byte 235,46 // jmp b5f <_sk_load_f16_hsw+0xcd>
|
||||
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
|
||||
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
|
||||
.byte 233,117,255,255,255 // jmpq c19 <_sk_load_f16_hsw+0x21>
|
||||
.byte 233,117,255,255,255 // jmpq ab3 <_sk_load_f16_hsw+0x21>
|
||||
.byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
|
||||
.byte 72,131,249,5 // cmp $0x5,%rcx
|
||||
.byte 114,21 // jb cc5 <_sk_load_f16_hsw+0xcd>
|
||||
.byte 114,21 // jb b5f <_sk_load_f16_hsw+0xcd>
|
||||
.byte 197,251,16,92,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm3
|
||||
.byte 72,131,249,5 // cmp $0x5,%rcx
|
||||
.byte 117,18 // jne cce <_sk_load_f16_hsw+0xd6>
|
||||
.byte 117,18 // jne b68 <_sk_load_f16_hsw+0xd6>
|
||||
.byte 197,250,126,219 // vmovq %xmm3,%xmm3
|
||||
.byte 233,84,255,255,255 // jmpq c19 <_sk_load_f16_hsw+0x21>
|
||||
.byte 233,84,255,255,255 // jmpq ab3 <_sk_load_f16_hsw+0x21>
|
||||
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
|
||||
.byte 233,75,255,255,255 // jmpq c19 <_sk_load_f16_hsw+0x21>
|
||||
.byte 233,75,255,255,255 // jmpq ab3 <_sk_load_f16_hsw+0x21>
|
||||
.byte 197,225,22,92,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3
|
||||
.byte 72,131,249,7 // cmp $0x7,%rcx
|
||||
.byte 15,130,59,255,255,255 // jb c19 <_sk_load_f16_hsw+0x21>
|
||||
.byte 15,130,59,255,255,255 // jb ab3 <_sk_load_f16_hsw+0x21>
|
||||
.byte 197,123,16,68,248,48 // vmovsd 0x30(%rax,%rdi,8),%xmm8
|
||||
.byte 233,48,255,255,255 // jmpq c19 <_sk_load_f16_hsw+0x21>
|
||||
.byte 233,48,255,255,255 // jmpq ab3 <_sk_load_f16_hsw+0x21>
|
||||
|
||||
.globl _sk_store_f16_hsw
|
||||
_sk_store_f16_hsw:
|
||||
@ -2822,7 +2718,7 @@ _sk_store_f16_hsw:
|
||||
.byte 196,65,57,98,205 // vpunpckldq %xmm13,%xmm8,%xmm9
|
||||
.byte 196,65,57,106,197 // vpunpckhdq %xmm13,%xmm8,%xmm8
|
||||
.byte 72,133,201 // test %rcx,%rcx
|
||||
.byte 117,27 // jne d4e <_sk_store_f16_hsw+0x65>
|
||||
.byte 117,27 // jne be8 <_sk_store_f16_hsw+0x65>
|
||||
.byte 197,120,17,28,248 // vmovups %xmm11,(%rax,%rdi,8)
|
||||
.byte 197,120,17,84,248,16 // vmovups %xmm10,0x10(%rax,%rdi,8)
|
||||
.byte 197,120,17,76,248,32 // vmovups %xmm9,0x20(%rax,%rdi,8)
|
||||
@ -2831,22 +2727,22 @@ _sk_store_f16_hsw:
|
||||
.byte 255,224 // jmpq *%rax
|
||||
.byte 197,121,214,28,248 // vmovq %xmm11,(%rax,%rdi,8)
|
||||
.byte 72,131,249,1 // cmp $0x1,%rcx
|
||||
.byte 116,241 // je d4a <_sk_store_f16_hsw+0x61>
|
||||
.byte 116,241 // je be4 <_sk_store_f16_hsw+0x61>
|
||||
.byte 197,121,23,92,248,8 // vmovhpd %xmm11,0x8(%rax,%rdi,8)
|
||||
.byte 72,131,249,3 // cmp $0x3,%rcx
|
||||
.byte 114,229 // jb d4a <_sk_store_f16_hsw+0x61>
|
||||
.byte 114,229 // jb be4 <_sk_store_f16_hsw+0x61>
|
||||
.byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8)
|
||||
.byte 116,221 // je d4a <_sk_store_f16_hsw+0x61>
|
||||
.byte 116,221 // je be4 <_sk_store_f16_hsw+0x61>
|
||||
.byte 197,121,23,84,248,24 // vmovhpd %xmm10,0x18(%rax,%rdi,8)
|
||||
.byte 72,131,249,5 // cmp $0x5,%rcx
|
||||
.byte 114,209 // jb d4a <_sk_store_f16_hsw+0x61>
|
||||
.byte 114,209 // jb be4 <_sk_store_f16_hsw+0x61>
|
||||
.byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8)
|
||||
.byte 116,201 // je d4a <_sk_store_f16_hsw+0x61>
|
||||
.byte 116,201 // je be4 <_sk_store_f16_hsw+0x61>
|
||||
.byte 197,121,23,76,248,40 // vmovhpd %xmm9,0x28(%rax,%rdi,8)
|
||||
.byte 72,131,249,7 // cmp $0x7,%rcx
|
||||
.byte 114,189 // jb d4a <_sk_store_f16_hsw+0x61>
|
||||
.byte 114,189 // jb be4 <_sk_store_f16_hsw+0x61>
|
||||
.byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8)
|
||||
.byte 235,181 // jmp d4a <_sk_store_f16_hsw+0x61>
|
||||
.byte 235,181 // jmp be4 <_sk_store_f16_hsw+0x61>
|
||||
|
||||
.globl _sk_clamp_x_hsw
|
||||
_sk_clamp_x_hsw:
|
||||
|
||||
@ -514,18 +514,20 @@ _sk_lerp_565_hsw LABEL PROC
|
||||
|
||||
PUBLIC _sk_load_tables_hsw
|
||||
_sk_load_tables_hsw LABEL PROC
|
||||
DB 73,137,200 ; mov %rcx,%r8
|
||||
DB 72,173 ; lods %ds:(%rsi),%rax
|
||||
DB 76,139,0 ; mov (%rax),%r8
|
||||
DB 72,133,201 ; test %rcx,%rcx
|
||||
DB 117,104 ; jne 6fe <_sk_load_tables_hsw+0x72>
|
||||
DB 196,193,126,111,28,184 ; vmovdqu (%r8,%rdi,4),%ymm3
|
||||
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
|
||||
DB 76,3,8 ; add (%rax),%r9
|
||||
DB 77,133,192 ; test %r8,%r8
|
||||
DB 117,106 ; jne 70b <_sk_load_tables_hsw+0x7f>
|
||||
DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3
|
||||
DB 196,226,125,88,82,16 ; vpbroadcastd 0x10(%rdx),%ymm2
|
||||
DB 197,237,219,203 ; vpand %ymm3,%ymm2,%ymm1
|
||||
DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8
|
||||
DB 76,139,64,8 ; mov 0x8(%rax),%r8
|
||||
DB 72,139,72,8 ; mov 0x8(%rax),%rcx
|
||||
DB 76,139,72,16 ; mov 0x10(%rax),%r9
|
||||
DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9
|
||||
DB 196,194,53,146,4,136 ; vgatherdps %ymm9,(%r8,%ymm1,4),%ymm0
|
||||
DB 196,226,53,146,4,137 ; vgatherdps %ymm9,(%rcx,%ymm1,4),%ymm0
|
||||
DB 197,245,114,211,8 ; vpsrld $0x8,%ymm3,%ymm1
|
||||
DB 197,109,219,201 ; vpand %ymm1,%ymm2,%ymm9
|
||||
DB 196,65,45,118,210 ; vpcmpeqd %ymm10,%ymm10,%ymm10
|
||||
@ -539,56 +541,17 @@ _sk_load_tables_hsw LABEL PROC
|
||||
DB 196,98,125,24,66,12 ; vbroadcastss 0xc(%rdx),%ymm8
|
||||
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
|
||||
DB 72,173 ; lods %ds:(%rsi),%rax
|
||||
DB 76,137,193 ; mov %r8,%rcx
|
||||
DB 255,224 ; jmpq *%rax
|
||||
DB 65,137,201 ; mov %ecx,%r9d
|
||||
DB 65,128,225,7 ; and $0x7,%r9b
|
||||
DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3
|
||||
DB 65,254,201 ; dec %r9b
|
||||
DB 69,15,182,201 ; movzbl %r9b,%r9d
|
||||
DB 65,128,249,6 ; cmp $0x6,%r9b
|
||||
DB 119,134 ; ja 69c <_sk_load_tables_hsw+0x10>
|
||||
DB 76,141,21,131,0,0,0 ; lea 0x83(%rip),%r10 # 7a0 <_sk_load_tables_hsw+0x114>
|
||||
DB 79,99,12,138 ; movslq (%r10,%r9,4),%r9
|
||||
DB 77,1,209 ; add %r10,%r9
|
||||
DB 65,255,225 ; jmpq *%r9
|
||||
DB 196,193,121,110,68,184,24 ; vmovd 0x18(%r8,%rdi,4),%xmm0
|
||||
DB 196,226,125,89,192 ; vpbroadcastq %xmm0,%ymm0
|
||||
DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1
|
||||
DB 196,227,117,2,216,64 ; vpblendd $0x40,%ymm0,%ymm1,%ymm3
|
||||
DB 196,227,125,57,216,1 ; vextracti128 $0x1,%ymm3,%xmm0
|
||||
DB 196,195,121,34,68,184,20,1 ; vpinsrd $0x1,0x14(%r8,%rdi,4),%xmm0,%xmm0
|
||||
DB 196,227,101,56,216,1 ; vinserti128 $0x1,%xmm0,%ymm3,%ymm3
|
||||
DB 196,227,125,57,216,1 ; vextracti128 $0x1,%ymm3,%xmm0
|
||||
DB 196,195,121,34,68,184,16,0 ; vpinsrd $0x0,0x10(%r8,%rdi,4),%xmm0,%xmm0
|
||||
DB 196,227,101,56,216,1 ; vinserti128 $0x1,%xmm0,%ymm3,%ymm3
|
||||
DB 196,195,97,34,68,184,12,3 ; vpinsrd $0x3,0xc(%r8,%rdi,4),%xmm3,%xmm0
|
||||
DB 196,227,101,2,216,15 ; vpblendd $0xf,%ymm0,%ymm3,%ymm3
|
||||
DB 196,195,97,34,68,184,8,2 ; vpinsrd $0x2,0x8(%r8,%rdi,4),%xmm3,%xmm0
|
||||
DB 196,227,101,2,216,15 ; vpblendd $0xf,%ymm0,%ymm3,%ymm3
|
||||
DB 196,195,97,34,68,184,4,1 ; vpinsrd $0x1,0x4(%r8,%rdi,4),%xmm3,%xmm0
|
||||
DB 196,227,101,2,216,15 ; vpblendd $0xf,%ymm0,%ymm3,%ymm3
|
||||
DB 196,193,121,110,4,184 ; vmovd (%r8,%rdi,4),%xmm0
|
||||
DB 196,227,101,2,216,1 ; vpblendd $0x1,%ymm0,%ymm3,%ymm3
|
||||
DB 233,252,254,255,255 ; jmpq 69c <_sk_load_tables_hsw+0x10>
|
||||
DB 239 ; out %eax,(%dx)
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255,225 ; jmpq *%rcx
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255,211 ; callq *%rbx
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255,197 ; inc %ebp
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255,177,255,255,255,157 ; pushq -0x62000001(%rcx)
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; .byte 0xff
|
||||
DB 135,255 ; xchg %edi,%edi
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; .byte 0xff
|
||||
DB 185,8,0,0,0 ; mov $0x8,%ecx
|
||||
DB 68,41,193 ; sub %r8d,%ecx
|
||||
DB 192,225,3 ; shl $0x3,%cl
|
||||
DB 73,199,194,255,255,255,255 ; mov $0xffffffffffffffff,%r10
|
||||
DB 73,211,234 ; shr %cl,%r10
|
||||
DB 196,193,249,110,194 ; vmovq %r10,%xmm0
|
||||
DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0
|
||||
DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3
|
||||
DB 233,114,255,255,255 ; jmpq 6a6 <_sk_load_tables_hsw+0x1a>
|
||||
|
||||
PUBLIC _sk_load_a8_hsw
|
||||
_sk_load_a8_hsw LABEL PROC
|
||||
@ -597,7 +560,7 @@ _sk_load_a8_hsw LABEL PROC
|
||||
DB 72,139,0 ; mov (%rax),%rax
|
||||
DB 72,1,248 ; add %rdi,%rax
|
||||
DB 77,133,192 ; test %r8,%r8
|
||||
DB 117,42 ; jne 7f6 <_sk_load_a8_hsw+0x3a>
|
||||
DB 117,42 ; jne 76e <_sk_load_a8_hsw+0x3a>
|
||||
DB 197,251,16,0 ; vmovsd (%rax),%xmm0
|
||||
DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0
|
||||
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
|
||||
@ -618,9 +581,9 @@ _sk_load_a8_hsw LABEL PROC
|
||||
DB 77,9,217 ; or %r11,%r9
|
||||
DB 72,131,193,8 ; add $0x8,%rcx
|
||||
DB 73,255,202 ; dec %r10
|
||||
DB 117,234 ; jne 7fe <_sk_load_a8_hsw+0x42>
|
||||
DB 117,234 ; jne 776 <_sk_load_a8_hsw+0x42>
|
||||
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
|
||||
DB 235,181 ; jmp 7d0 <_sk_load_a8_hsw+0x14>
|
||||
DB 235,181 ; jmp 748 <_sk_load_a8_hsw+0x14>
|
||||
|
||||
PUBLIC _sk_store_a8_hsw
|
||||
_sk_store_a8_hsw LABEL PROC
|
||||
@ -633,7 +596,7 @@ _sk_store_a8_hsw LABEL PROC
|
||||
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
|
||||
DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8
|
||||
DB 72,133,201 ; test %rcx,%rcx
|
||||
DB 117,10 ; jne 84e <_sk_store_a8_hsw+0x33>
|
||||
DB 117,10 ; jne 7c6 <_sk_store_a8_hsw+0x33>
|
||||
DB 196,65,123,17,4,57 ; vmovsd %xmm8,(%r9,%rdi,1)
|
||||
DB 72,173 ; lods %ds:(%rsi),%rax
|
||||
DB 255,224 ; jmpq *%rax
|
||||
@ -642,9 +605,9 @@ _sk_store_a8_hsw LABEL PROC
|
||||
DB 254,200 ; dec %al
|
||||
DB 68,15,182,192 ; movzbl %al,%r8d
|
||||
DB 65,128,248,6 ; cmp $0x6,%r8b
|
||||
DB 119,236 ; ja 84a <_sk_store_a8_hsw+0x2f>
|
||||
DB 119,236 ; ja 7c2 <_sk_store_a8_hsw+0x2f>
|
||||
DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8
|
||||
DB 76,141,21,66,0,0,0 ; lea 0x42(%rip),%r10 # 8ac <_sk_store_a8_hsw+0x91>
|
||||
DB 76,141,21,66,0,0,0 ; lea 0x42(%rip),%r10 # 824 <_sk_store_a8_hsw+0x91>
|
||||
DB 75,99,4,130 ; movslq (%r10,%r8,4),%rax
|
||||
DB 76,1,208 ; add %r10,%rax
|
||||
DB 255,224 ; jmpq *%rax
|
||||
@ -655,7 +618,7 @@ _sk_store_a8_hsw LABEL PROC
|
||||
DB 196,67,121,20,68,57,2,4 ; vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1)
|
||||
DB 196,67,121,20,68,57,1,2 ; vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1)
|
||||
DB 196,67,121,20,4,57,0 ; vpextrb $0x0,%xmm8,(%r9,%rdi,1)
|
||||
DB 235,158 ; jmp 84a <_sk_store_a8_hsw+0x2f>
|
||||
DB 235,158 ; jmp 7c2 <_sk_store_a8_hsw+0x2f>
|
||||
DB 247,255 ; idiv %edi
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
@ -684,7 +647,7 @@ _sk_load_565_hsw LABEL PROC
|
||||
DB 72,173 ; lods %ds:(%rsi),%rax
|
||||
DB 76,139,16 ; mov (%rax),%r10
|
||||
DB 72,133,201 ; test %rcx,%rcx
|
||||
DB 117,92 ; jne 92e <_sk_load_565_hsw+0x66>
|
||||
DB 117,92 ; jne 8a6 <_sk_load_565_hsw+0x66>
|
||||
DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0
|
||||
DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2
|
||||
DB 196,226,125,88,66,104 ; vpbroadcastd 0x68(%rdx),%ymm0
|
||||
@ -711,8 +674,8 @@ _sk_load_565_hsw LABEL PROC
|
||||
DB 65,254,200 ; dec %r8b
|
||||
DB 69,15,182,192 ; movzbl %r8b,%r8d
|
||||
DB 65,128,248,6 ; cmp $0x6,%r8b
|
||||
DB 119,146 ; ja 8d8 <_sk_load_565_hsw+0x10>
|
||||
DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 998 <_sk_load_565_hsw+0xd0>
|
||||
DB 119,146 ; ja 850 <_sk_load_565_hsw+0x10>
|
||||
DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 910 <_sk_load_565_hsw+0xd0>
|
||||
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
|
||||
DB 76,1,200 ; add %r9,%rax
|
||||
DB 255,224 ; jmpq *%rax
|
||||
@ -724,7 +687,7 @@ _sk_load_565_hsw LABEL PROC
|
||||
DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
|
||||
DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
|
||||
DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
|
||||
DB 233,66,255,255,255 ; jmpq 8d8 <_sk_load_565_hsw+0x10>
|
||||
DB 233,66,255,255,255 ; jmpq 850 <_sk_load_565_hsw+0x10>
|
||||
DB 102,144 ; xchg %ax,%ax
|
||||
DB 242,255 ; repnz (bad)
|
||||
DB 255 ; (bad)
|
||||
@ -769,7 +732,7 @@ _sk_store_565_hsw LABEL PROC
|
||||
DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9
|
||||
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
|
||||
DB 72,133,201 ; test %rcx,%rcx
|
||||
DB 117,10 ; jne a16 <_sk_store_565_hsw+0x62>
|
||||
DB 117,10 ; jne 98e <_sk_store_565_hsw+0x62>
|
||||
DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2)
|
||||
DB 72,173 ; lods %ds:(%rsi),%rax
|
||||
DB 255,224 ; jmpq *%rax
|
||||
@ -778,8 +741,8 @@ _sk_store_565_hsw LABEL PROC
|
||||
DB 254,200 ; dec %al
|
||||
DB 68,15,182,192 ; movzbl %al,%r8d
|
||||
DB 65,128,248,6 ; cmp $0x6,%r8b
|
||||
DB 119,236 ; ja a12 <_sk_store_565_hsw+0x5e>
|
||||
DB 76,141,21,71,0,0,0 ; lea 0x47(%rip),%r10 # a74 <_sk_store_565_hsw+0xc0>
|
||||
DB 119,236 ; ja 98a <_sk_store_565_hsw+0x5e>
|
||||
DB 76,141,21,71,0,0,0 ; lea 0x47(%rip),%r10 # 9ec <_sk_store_565_hsw+0xc0>
|
||||
DB 75,99,4,130 ; movslq (%r10,%r8,4),%rax
|
||||
DB 76,1,208 ; add %r10,%rax
|
||||
DB 255,224 ; jmpq *%rax
|
||||
@ -791,7 +754,7 @@ _sk_store_565_hsw LABEL PROC
|
||||
DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
|
||||
DB 197,121,126,192 ; vmovd %xmm8,%eax
|
||||
DB 102,65,137,4,121 ; mov %ax,(%r9,%rdi,2)
|
||||
DB 235,161 ; jmp a12 <_sk_store_565_hsw+0x5e>
|
||||
DB 235,161 ; jmp 98a <_sk_store_565_hsw+0x5e>
|
||||
DB 15,31,0 ; nopl (%rax)
|
||||
DB 242,255 ; repnz (bad)
|
||||
DB 255 ; (bad)
|
||||
@ -818,11 +781,13 @@ _sk_store_565_hsw LABEL PROC
|
||||
|
||||
PUBLIC _sk_load_8888_hsw
|
||||
_sk_load_8888_hsw LABEL PROC
|
||||
DB 73,137,200 ; mov %rcx,%r8
|
||||
DB 72,173 ; lods %ds:(%rsi),%rax
|
||||
DB 76,139,16 ; mov (%rax),%r10
|
||||
DB 72,133,201 ; test %rcx,%rcx
|
||||
DB 117,83 ; jne aed <_sk_load_8888_hsw+0x5d>
|
||||
DB 196,193,126,111,28,186 ; vmovdqu (%r10,%rdi,4),%ymm3
|
||||
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
|
||||
DB 76,3,8 ; add (%rax),%r9
|
||||
DB 77,133,192 ; test %r8,%r8
|
||||
DB 117,85 ; jne a72 <_sk_load_8888_hsw+0x6a>
|
||||
DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3
|
||||
DB 196,226,125,88,82,16 ; vpbroadcastd 0x10(%rdx),%ymm2
|
||||
DB 197,237,219,195 ; vpand %ymm3,%ymm2,%ymm0
|
||||
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
|
||||
@ -840,62 +805,24 @@ _sk_load_8888_hsw LABEL PROC
|
||||
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
|
||||
DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3
|
||||
DB 72,173 ; lods %ds:(%rsi),%rax
|
||||
DB 76,137,193 ; mov %r8,%rcx
|
||||
DB 255,224 ; jmpq *%rax
|
||||
DB 65,137,200 ; mov %ecx,%r8d
|
||||
DB 65,128,224,7 ; and $0x7,%r8b
|
||||
DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3
|
||||
DB 65,254,200 ; dec %r8b
|
||||
DB 69,15,182,192 ; movzbl %r8b,%r8d
|
||||
DB 65,128,248,6 ; cmp $0x6,%r8b
|
||||
DB 119,155 ; ja aa0 <_sk_load_8888_hsw+0x10>
|
||||
DB 76,141,13,132,0,0,0 ; lea 0x84(%rip),%r9 # b90 <_sk_load_8888_hsw+0x100>
|
||||
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
|
||||
DB 76,1,200 ; add %r9,%rax
|
||||
DB 255,224 ; jmpq *%rax
|
||||
DB 196,193,121,110,68,186,24 ; vmovd 0x18(%r10,%rdi,4),%xmm0
|
||||
DB 196,226,125,89,192 ; vpbroadcastq %xmm0,%ymm0
|
||||
DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1
|
||||
DB 196,227,117,2,216,64 ; vpblendd $0x40,%ymm0,%ymm1,%ymm3
|
||||
DB 196,227,125,57,216,1 ; vextracti128 $0x1,%ymm3,%xmm0
|
||||
DB 196,195,121,34,68,186,20,1 ; vpinsrd $0x1,0x14(%r10,%rdi,4),%xmm0,%xmm0
|
||||
DB 196,227,101,56,216,1 ; vinserti128 $0x1,%xmm0,%ymm3,%ymm3
|
||||
DB 196,227,125,57,216,1 ; vextracti128 $0x1,%ymm3,%xmm0
|
||||
DB 196,195,121,34,68,186,16,0 ; vpinsrd $0x0,0x10(%r10,%rdi,4),%xmm0,%xmm0
|
||||
DB 196,227,101,56,216,1 ; vinserti128 $0x1,%xmm0,%ymm3,%ymm3
|
||||
DB 196,195,97,34,68,186,12,3 ; vpinsrd $0x3,0xc(%r10,%rdi,4),%xmm3,%xmm0
|
||||
DB 196,227,101,2,216,15 ; vpblendd $0xf,%ymm0,%ymm3,%ymm3
|
||||
DB 196,195,97,34,68,186,8,2 ; vpinsrd $0x2,0x8(%r10,%rdi,4),%xmm3,%xmm0
|
||||
DB 196,227,101,2,216,15 ; vpblendd $0xf,%ymm0,%ymm3,%ymm3
|
||||
DB 196,195,97,34,68,186,4,1 ; vpinsrd $0x1,0x4(%r10,%rdi,4),%xmm3,%xmm0
|
||||
DB 196,227,101,2,216,15 ; vpblendd $0xf,%ymm0,%ymm3,%ymm3
|
||||
DB 196,193,121,110,4,186 ; vmovd (%r10,%rdi,4),%xmm0
|
||||
DB 196,227,101,2,216,1 ; vpblendd $0x1,%ymm0,%ymm3,%ymm3
|
||||
DB 233,18,255,255,255 ; jmpq aa0 <_sk_load_8888_hsw+0x10>
|
||||
DB 102,144 ; xchg %ax,%ax
|
||||
DB 237 ; in (%dx),%eax
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 223,255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255,209 ; callq *%rcx
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255,195 ; inc %ebx
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255,175,255,255,255,155 ; ljmp *-0x64000001(%rdi)
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; .byte 0xff
|
||||
DB 133,255 ; test %edi,%edi
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; .byte 0xff
|
||||
DB 185,8,0,0,0 ; mov $0x8,%ecx
|
||||
DB 68,41,193 ; sub %r8d,%ecx
|
||||
DB 192,225,3 ; shl $0x3,%cl
|
||||
DB 72,199,192,255,255,255,255 ; mov $0xffffffffffffffff,%rax
|
||||
DB 72,211,232 ; shr %cl,%rax
|
||||
DB 196,225,249,110,192 ; vmovq %rax,%xmm0
|
||||
DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0
|
||||
DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3
|
||||
DB 235,138 ; jmp a22 <_sk_load_8888_hsw+0x1a>
|
||||
|
||||
PUBLIC _sk_store_8888_hsw
|
||||
_sk_store_8888_hsw LABEL PROC
|
||||
DB 73,137,200 ; mov %rcx,%r8
|
||||
DB 72,173 ; lods %ds:(%rsi),%rax
|
||||
DB 76,139,8 ; mov (%rax),%r9
|
||||
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
|
||||
DB 76,3,8 ; add (%rax),%r9
|
||||
DB 196,98,125,24,66,8 ; vbroadcastss 0x8(%rdx),%ymm8
|
||||
DB 197,60,89,200 ; vmulps %ymm0,%ymm8,%ymm9
|
||||
DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9
|
||||
@ -911,59 +838,28 @@ _sk_store_8888_hsw LABEL PROC
|
||||
DB 196,193,61,114,240,24 ; vpslld $0x18,%ymm8,%ymm8
|
||||
DB 196,65,45,235,192 ; vpor %ymm8,%ymm10,%ymm8
|
||||
DB 196,65,53,235,192 ; vpor %ymm8,%ymm9,%ymm8
|
||||
DB 72,133,201 ; test %rcx,%rcx
|
||||
DB 117,10 ; jne c0b <_sk_store_8888_hsw+0x5f>
|
||||
DB 196,65,126,127,4,185 ; vmovdqu %ymm8,(%r9,%rdi,4)
|
||||
DB 77,133,192 ; test %r8,%r8
|
||||
DB 117,12 ; jne b04 <_sk_store_8888_hsw+0x6c>
|
||||
DB 196,65,126,127,1 ; vmovdqu %ymm8,(%r9)
|
||||
DB 72,173 ; lods %ds:(%rsi),%rax
|
||||
DB 76,137,193 ; mov %r8,%rcx
|
||||
DB 255,224 ; jmpq *%rax
|
||||
DB 137,200 ; mov %ecx,%eax
|
||||
DB 36,7 ; and $0x7,%al
|
||||
DB 254,200 ; dec %al
|
||||
DB 68,15,182,192 ; movzbl %al,%r8d
|
||||
DB 65,128,248,6 ; cmp $0x6,%r8b
|
||||
DB 119,236 ; ja c07 <_sk_store_8888_hsw+0x5b>
|
||||
DB 76,141,21,82,0,0,0 ; lea 0x52(%rip),%r10 # c74 <_sk_store_8888_hsw+0xc8>
|
||||
DB 75,99,4,130 ; movslq (%r10,%r8,4),%rax
|
||||
DB 76,1,208 ; add %r10,%rax
|
||||
DB 255,224 ; jmpq *%rax
|
||||
DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9
|
||||
DB 196,67,121,22,76,185,24,2 ; vpextrd $0x2,%xmm9,0x18(%r9,%rdi,4)
|
||||
DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9
|
||||
DB 196,67,121,22,76,185,20,1 ; vpextrd $0x1,%xmm9,0x14(%r9,%rdi,4)
|
||||
DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9
|
||||
DB 196,65,121,126,76,185,16 ; vmovd %xmm9,0x10(%r9,%rdi,4)
|
||||
DB 196,67,121,22,68,185,12,3 ; vpextrd $0x3,%xmm8,0xc(%r9,%rdi,4)
|
||||
DB 196,67,121,22,68,185,8,2 ; vpextrd $0x2,%xmm8,0x8(%r9,%rdi,4)
|
||||
DB 196,67,121,22,68,185,4,1 ; vpextrd $0x1,%xmm8,0x4(%r9,%rdi,4)
|
||||
DB 196,65,121,126,4,185 ; vmovd %xmm8,(%r9,%rdi,4)
|
||||
DB 235,147 ; jmp c07 <_sk_store_8888_hsw+0x5b>
|
||||
DB 248 ; clc
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255,240 ; push %rax
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 232,255,255,255,224 ; callq ffffffffe1000c80 <_sk_linear_gradient_2stops_hsw+0xffffffffe0fffbf2>
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255,211 ; callq *%rbx
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255,197 ; inc %ebp
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; .byte 0xff
|
||||
DB 183,255 ; mov $0xff,%bh
|
||||
DB 255 ; (bad)
|
||||
DB 255 ; .byte 0xff
|
||||
DB 185,8,0,0,0 ; mov $0x8,%ecx
|
||||
DB 68,41,193 ; sub %r8d,%ecx
|
||||
DB 192,225,3 ; shl $0x3,%cl
|
||||
DB 72,199,192,255,255,255,255 ; mov $0xffffffffffffffff,%rax
|
||||
DB 72,211,232 ; shr %cl,%rax
|
||||
DB 196,97,249,110,200 ; vmovq %rax,%xmm9
|
||||
DB 196,66,125,33,201 ; vpmovsxbd %xmm9,%ymm9
|
||||
DB 196,66,53,142,1 ; vpmaskmovd %ymm8,%ymm9,(%r9)
|
||||
DB 235,211 ; jmp afd <_sk_store_8888_hsw+0x65>
|
||||
|
||||
PUBLIC _sk_load_f16_hsw
|
||||
_sk_load_f16_hsw LABEL PROC
|
||||
DB 72,173 ; lods %ds:(%rsi),%rax
|
||||
DB 72,139,0 ; mov (%rax),%rax
|
||||
DB 72,133,201 ; test %rcx,%rcx
|
||||
DB 117,97 ; jne cfb <_sk_load_f16_hsw+0x6b>
|
||||
DB 117,97 ; jne b95 <_sk_load_f16_hsw+0x6b>
|
||||
DB 197,249,16,12,248 ; vmovupd (%rax,%rdi,8),%xmm1
|
||||
DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2
|
||||
DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3
|
||||
@ -989,35 +885,35 @@ _sk_load_f16_hsw LABEL PROC
|
||||
DB 197,251,16,12,248 ; vmovsd (%rax,%rdi,8),%xmm1
|
||||
DB 196,65,57,87,192 ; vxorpd %xmm8,%xmm8,%xmm8
|
||||
DB 72,131,249,1 ; cmp $0x1,%rcx
|
||||
DB 117,6 ; jne d11 <_sk_load_f16_hsw+0x81>
|
||||
DB 117,6 ; jne bab <_sk_load_f16_hsw+0x81>
|
||||
DB 197,250,126,201 ; vmovq %xmm1,%xmm1
|
||||
DB 235,30 ; jmp d2f <_sk_load_f16_hsw+0x9f>
|
||||
DB 235,30 ; jmp bc9 <_sk_load_f16_hsw+0x9f>
|
||||
DB 197,241,22,76,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm1,%xmm1
|
||||
DB 72,131,249,3 ; cmp $0x3,%rcx
|
||||
DB 114,18 ; jb d2f <_sk_load_f16_hsw+0x9f>
|
||||
DB 114,18 ; jb bc9 <_sk_load_f16_hsw+0x9f>
|
||||
DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2
|
||||
DB 72,131,249,3 ; cmp $0x3,%rcx
|
||||
DB 117,19 ; jne d3c <_sk_load_f16_hsw+0xac>
|
||||
DB 117,19 ; jne bd6 <_sk_load_f16_hsw+0xac>
|
||||
DB 197,250,126,210 ; vmovq %xmm2,%xmm2
|
||||
DB 235,46 ; jmp d5d <_sk_load_f16_hsw+0xcd>
|
||||
DB 235,46 ; jmp bf7 <_sk_load_f16_hsw+0xcd>
|
||||
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
|
||||
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
|
||||
DB 233,117,255,255,255 ; jmpq cb1 <_sk_load_f16_hsw+0x21>
|
||||
DB 233,117,255,255,255 ; jmpq b4b <_sk_load_f16_hsw+0x21>
|
||||
DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
|
||||
DB 72,131,249,5 ; cmp $0x5,%rcx
|
||||
DB 114,21 ; jb d5d <_sk_load_f16_hsw+0xcd>
|
||||
DB 114,21 ; jb bf7 <_sk_load_f16_hsw+0xcd>
|
||||
DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3
|
||||
DB 72,131,249,5 ; cmp $0x5,%rcx
|
||||
DB 117,18 ; jne d66 <_sk_load_f16_hsw+0xd6>
|
||||
DB 117,18 ; jne c00 <_sk_load_f16_hsw+0xd6>
|
||||
DB 197,250,126,219 ; vmovq %xmm3,%xmm3
|
||||
DB 233,84,255,255,255 ; jmpq cb1 <_sk_load_f16_hsw+0x21>
|
||||
DB 233,84,255,255,255 ; jmpq b4b <_sk_load_f16_hsw+0x21>
|
||||
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
|
||||
DB 233,75,255,255,255 ; jmpq cb1 <_sk_load_f16_hsw+0x21>
|
||||
DB 233,75,255,255,255 ; jmpq b4b <_sk_load_f16_hsw+0x21>
|
||||
DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3
|
||||
DB 72,131,249,7 ; cmp $0x7,%rcx
|
||||
DB 15,130,59,255,255,255 ; jb cb1 <_sk_load_f16_hsw+0x21>
|
||||
DB 15,130,59,255,255,255 ; jb b4b <_sk_load_f16_hsw+0x21>
|
||||
DB 197,123,16,68,248,48 ; vmovsd 0x30(%rax,%rdi,8),%xmm8
|
||||
DB 233,48,255,255,255 ; jmpq cb1 <_sk_load_f16_hsw+0x21>
|
||||
DB 233,48,255,255,255 ; jmpq b4b <_sk_load_f16_hsw+0x21>
|
||||
|
||||
PUBLIC _sk_store_f16_hsw
|
||||
_sk_store_f16_hsw LABEL PROC
|
||||
@ -1036,7 +932,7 @@ _sk_store_f16_hsw LABEL PROC
|
||||
DB 196,65,57,98,205 ; vpunpckldq %xmm13,%xmm8,%xmm9
|
||||
DB 196,65,57,106,197 ; vpunpckhdq %xmm13,%xmm8,%xmm8
|
||||
DB 72,133,201 ; test %rcx,%rcx
|
||||
DB 117,27 ; jne de6 <_sk_store_f16_hsw+0x65>
|
||||
DB 117,27 ; jne c80 <_sk_store_f16_hsw+0x65>
|
||||
DB 197,120,17,28,248 ; vmovups %xmm11,(%rax,%rdi,8)
|
||||
DB 197,120,17,84,248,16 ; vmovups %xmm10,0x10(%rax,%rdi,8)
|
||||
DB 197,120,17,76,248,32 ; vmovups %xmm9,0x20(%rax,%rdi,8)
|
||||
@ -1045,22 +941,22 @@ _sk_store_f16_hsw LABEL PROC
|
||||
DB 255,224 ; jmpq *%rax
|
||||
DB 197,121,214,28,248 ; vmovq %xmm11,(%rax,%rdi,8)
|
||||
DB 72,131,249,1 ; cmp $0x1,%rcx
|
||||
DB 116,241 ; je de2 <_sk_store_f16_hsw+0x61>
|
||||
DB 116,241 ; je c7c <_sk_store_f16_hsw+0x61>
|
||||
DB 197,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%rax,%rdi,8)
|
||||
DB 72,131,249,3 ; cmp $0x3,%rcx
|
||||
DB 114,229 ; jb de2 <_sk_store_f16_hsw+0x61>
|
||||
DB 114,229 ; jb c7c <_sk_store_f16_hsw+0x61>
|
||||
DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8)
|
||||
DB 116,221 ; je de2 <_sk_store_f16_hsw+0x61>
|
||||
DB 116,221 ; je c7c <_sk_store_f16_hsw+0x61>
|
||||
DB 197,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%rax,%rdi,8)
|
||||
DB 72,131,249,5 ; cmp $0x5,%rcx
|
||||
DB 114,209 ; jb de2 <_sk_store_f16_hsw+0x61>
|
||||
DB 114,209 ; jb c7c <_sk_store_f16_hsw+0x61>
|
||||
DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8)
|
||||
DB 116,201 ; je de2 <_sk_store_f16_hsw+0x61>
|
||||
DB 116,201 ; je c7c <_sk_store_f16_hsw+0x61>
|
||||
DB 197,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%rax,%rdi,8)
|
||||
DB 72,131,249,7 ; cmp $0x7,%rcx
|
||||
DB 114,189 ; jb de2 <_sk_store_f16_hsw+0x61>
|
||||
DB 114,189 ; jb c7c <_sk_store_f16_hsw+0x61>
|
||||
DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8)
|
||||
DB 235,181 ; jmp de2 <_sk_store_f16_hsw+0x61>
|
||||
DB 235,181 ; jmp c7c <_sk_store_f16_hsw+0x61>
|
||||
|
||||
PUBLIC _sk_clamp_x_hsw
|
||||
_sk_clamp_x_hsw LABEL PROC
|
||||
|
||||
@ -258,23 +258,6 @@ static inline V load(const T* src, size_t tail) {
|
||||
return unaligned_load<V>(src);
|
||||
}
|
||||
|
||||
#if 1 && defined(JUMPER) && defined(__AVX__)
|
||||
template <>
|
||||
inline U8 load(const uint8_t* src, size_t tail) {
|
||||
if (__builtin_expect(tail, 0)) {
|
||||
uint64_t v = 0;
|
||||
size_t shift = 0;
|
||||
#pragma nounroll
|
||||
while (tail --> 0) {
|
||||
v |= (uint64_t)*src++ << shift;
|
||||
shift += 8;
|
||||
}
|
||||
return unaligned_load<U8>(&v);
|
||||
}
|
||||
return unaligned_load<U8>(src);
|
||||
}
|
||||
#endif
|
||||
|
||||
template <typename V, typename T>
|
||||
static inline void store(T* dst, V v, size_t tail) {
|
||||
#if defined(JUMPER)
|
||||
@ -295,6 +278,52 @@ static inline void store(T* dst, V v, size_t tail) {
|
||||
memcpy(dst, &v, sizeof(v));
|
||||
}
|
||||
|
||||
#if 1 && defined(JUMPER) && defined(__AVX__)
|
||||
template <>
|
||||
inline U8 load(const uint8_t* src, size_t tail) {
|
||||
if (__builtin_expect(tail, 0)) {
|
||||
uint64_t v = 0;
|
||||
size_t shift = 0;
|
||||
#pragma nounroll
|
||||
while (tail --> 0) {
|
||||
v |= (uint64_t)*src++ << shift;
|
||||
shift += 8;
|
||||
}
|
||||
return unaligned_load<U8>(&v);
|
||||
}
|
||||
return unaligned_load<U8>(src);
|
||||
}
|
||||
#endif
|
||||
|
||||
#if 1 && defined(JUMPER) && defined(__AVX2__)
|
||||
static inline U32 mask(size_t tail) {
|
||||
// It's easiest to build the mask as 8 8-bit values, either 0x00 or 0xff.
|
||||
// Start fully on, then shift away lanes from the top until we've got our mask.
|
||||
uint64_t mask = 0xffffffffffffffff >> 8*(kStride-tail);
|
||||
|
||||
// Sign-extend each mask lane to its full width, 0x00000000 or 0xffffffff.
|
||||
return _mm256_cvtepi8_epi32(_mm_cvtsi64_si128((int64_t)mask));
|
||||
}
|
||||
|
||||
template <>
|
||||
inline U32 load(const uint32_t* src, size_t tail) {
|
||||
__builtin_assume(tail < kStride);
|
||||
if (__builtin_expect(tail, 0)) {
|
||||
return _mm256_maskload_epi32((const int*)src, mask(tail));
|
||||
}
|
||||
return unaligned_load<U32>(src);
|
||||
}
|
||||
|
||||
template <>
|
||||
inline void store(uint32_t* dst, U32 v, size_t tail) {
|
||||
__builtin_assume(tail < kStride);
|
||||
if (__builtin_expect(tail, 0)) {
|
||||
return _mm256_maskstore_epi32((int*)dst, mask(tail), v);
|
||||
}
|
||||
memcpy(dst, &v, sizeof(v));
|
||||
}
|
||||
#endif
|
||||
|
||||
|
||||
static F lerp(F from, F to, F t) {
|
||||
return mad(to-from, t, from);
|
||||
|
||||
Loading…
Reference in New Issue
Block a user