我有一个紧密的内部循环，它正在消耗相当多的 CPU 功率。所以我正在尝试优化它。我有两个版本的代码，一个在浮点数上运行，另一个在uint8_t. 浮点版本要快一点，因为它是矢量化的，而整数数学则没有。反正有没有强制整数数学矢量化？那是可能的吗？最后，有用吗？还是整数数学被高估了？
#include <algorithm>
#include <array>
typedef std::array<float, 3> Vec3F;
Vec3F v3fSpread(Vec3F const& source, Vec3F dest, float drop) {
if (source[0] <= dest[0] + drop && source[1] <= dest[1] + drop && source[2] <= dest[2] + drop) {
return dest;
} else {
float denom = std::max(source[0], std::max(source[1], source[2]));
dest[0] = std::max(source[0] - drop * source[0] / denom, dest[0]);
dest[1] = std::max(source[1] - drop * source[1] / denom, dest[1]);
dest[2] = std::max(source[2] - drop * source[2] / denom, dest[2]);
return dest;
}
}
组装成（56行）：
v3fSpread(std::array<float, 3ul> const&, std::array<float, 3ul>, float):
movq %xmm0, -40(%rsp)
movaps %xmm2, %xmm0
movd %xmm1, %rax
movss -40(%rsp), %xmm6
movl %eax, -32(%rsp)
movss (%rdi), %xmm1
addss %xmm6, %xmm0
movss -36(%rsp), %xmm7
movss -32(%rsp), %xmm8
movss 4(%rdi), %xmm3
ucomiss %xmm1, %xmm0
jb .L24
movaps %xmm2, %xmm0
movss 8(%rdi), %xmm4
addss %xmm7, %xmm0
ucomiss %xmm3, %xmm0
jae .L4
.L5:
movaps %xmm4, %xmm0
movaps %xmm1, %xmm5
maxss %xmm3, %xmm0
mulss %xmm2, %xmm5
maxss %xmm1, %xmm0
divss %xmm0, %xmm5
subss %xmm5, %xmm1
movaps %xmm2, %xmm5
mulss %xmm3, %xmm5
mulss %xmm4, %xmm2
maxss %xmm1, %xmm6
divss %xmm0, %xmm5
movss %xmm6, -24(%rsp)
divss %xmm0, %xmm2
subss %xmm5, %xmm3
maxss %xmm3, %xmm7
movss %xmm7, -20(%rsp)
movq -24(%rsp), %xmm0
subss %xmm2, %xmm4
maxss %xmm4, %xmm8
movss %xmm8, -16(%rsp)
movd -16(%rsp), %xmm1
ret
.L24:
movss 8(%rdi), %xmm4
jmp .L5
.L4:
movaps %xmm2, %xmm0
addss %xmm8, %xmm0
ucomiss %xmm4, %xmm0
jb .L5
movss %xmm6, -24(%rsp)
movss %xmm7, -20(%rsp)
movss %xmm8, -16(%rsp)
movq -24(%rsp), %xmm0
movd -16(%rsp), %xmm1
ret
和：
#include <algorithm>
#include <array>
#include <inttypes.h>
typedef std::array<uint8_t, 3> Vec3B;
typedef std::array<int32_t, 3> Vec3I;
Vec3B v3bSpread(Vec3B const& source, Vec3B dest, int32_t drop) {
Vec3I intSource = {source[0], source[1], source[2]};
Vec3I intDest = {dest[0], dest[1], dest[2]};
if (intSource[0] <= intDest[0] + drop && intSource[1] <= intDest[1] + drop && intSource[2] <= intDest[2] + drop) {
return dest;
} else {
int32_t denom = std::max(intSource[0], std::max(intSource[1], intSource[2]));
dest[0] = (uint8_t)std::max<int32_t>(intSource[0] - drop * intSource[0] / denom, intDest[0]);
dest[1] = (uint8_t)std::max<int32_t>(intSource[1] - drop * intSource[1] / denom, intDest[1]);
dest[2] = (uint8_t)std::max<int32_t>(intSource[2] - drop * intSource[2] / denom, intDest[2]);
return dest;
}
}
组装成（68行）：
v3bSpread(std::array<unsigned char, 3ul> const&, std::array<unsigned char, 3ul>, unsigned int):
pushq %rbx
movzbl %sil, %r11d
movl %esi, %ebx
movzbl (%rdi), %r8d
movzbl %r11b, %eax
shrw $8, %bx
addl %edx, %eax
shrl $16, %esi
movzbl 1(%rdi), %r10d
movl %edx, %r9d
movzbl 2(%rdi), %edi
cmpl %eax, %r8d
ja .L4
movzbl %bl, %eax
addl %edx, %eax
cmpl %eax, %r10d
jbe .L10
.L4:
cmpl %edi, %r10d
movl %edi, %ecx
movl %r8d, %eax
cmovge %r10d, %ecx
cmpl %ecx, %r8d
cmovge %r8d, %ecx
imull %r9d, %eax
xorl %edx, %edx
divl %ecx
subl %eax, %r8d
movl %r10d, %eax
cmpl %r11d, %r8d
cmovge %r8d, %r11d
imull %r9d, %eax
xorl %edx, %edx
movb %r11b, -32(%rsp)
divl %ecx
movzbl %bl, %edx
subl %eax, %r10d
movl %edi, %eax
cmpl %edx, %r10d
cmovl %edx, %r10d
imull %r9d, %eax
xorl %edx, %edx
movb %r10b, -31(%rsp)
divl %ecx
subl %eax, %edi
movzbl %sil, %eax
cmpl %eax, %edi
cmovl %eax, %edi
movb %dil, -30(%rsp)
.L6:
movzbl -31(%rsp), %eax
movzbl -32(%rsp), %edx
movb %al, %dh
movzbl -30(%rsp), %eax
popq %rbx
salq $16, %rax
orq %rdx, %rax
ret
.L10:
movzbl %sil, %eax
addl %edx, %eax
cmpl %eax, %edi
ja .L4
movb %r11b, -32(%rsp)
movb %bl, -31(%rsp)
movb %sil, -30(%rsp)
jmp .L6