矩阵转置的SSE汇编优化艺术以及ARM cortext 汇编优化 .
原文地址：http://blog.csdn.net/feixiang_john/article/details/8438658
平时我们做图像处理或者视频处理, 很多地方会用到矩阵转置:
比如: DCT变换, 图像旋转, 图像滤波, 以及一些数据的内存行和列的交换等, 会大量使用转置这个动作.
然而由于数据量很大,处理速度很慢!如何来提高处理速度呢?
下面看看分析:
HEVC中有个地方是如下这样实现(直接行和列对应的位置交换):
- Pel tmp;
- for (k=0;k<blkSize-1;k++)
- {
- for (l=k+1;l<blkSize;l++)
- {
- tmp = pDst[k*dstStride+l];
- pDst[k*dstStride+l] = pDst[l*dstStride+k];
- pDst[l*dstStride+k] = tmp;
- }
- }
Pel tmp;
for (k=0;k<blkSize-1;k++)
{
for (l=k+1;l<blkSize;l++)
{
tmp = pDst[k*dstStride+l];
pDst[k*dstStride+l] = pDst[l*dstStride+k];
pDst[l*dstStride+k] = tmp;
}
}
如何用汇编来实现呢?
我们先用SSE汇编来实现一个8X8的矩阵转置吧: 这里输入地址pSrc_128[i] 和输出地址pDst_128[i]可以相同也可以不同:
相同的话就是原地转置, 不同的话就是非原地转置.
- __m128i* m_pSrc_tmp = pSrc_128[i];
- __m128i* m_pDst_tmp = pDst_128[i];
- __m128i Org_8_0,Org_8_1, Org_8_2, Org_8_3;
- __m128i tttt1,tttt2,tttt3,tttt4,tttt33,tttt44;
- __m128i tttt5,tttt6, tttt7, tttt8;
- int stride_ii = dstStride>>3;
- //one
- Org_8_0 = _mm_load_si128(m_pSrc_tmp);
- m_pSrc_tmp+=8;
- Org_8_1 = _mm_load_si128(m_pSrc_tmp);
- m_pSrc_tmp+=8;
- Org_8_2 = _mm_load_si128(m_pSrc_tmp);
- m_pSrc_tmp+=8;
- Org_8_3 = _mm_load_si128(m_pSrc_tmp);
- m_pSrc_tmp+=8;
- tttt1 = _mm_unpacklo_epi16(Org_8_0, Org_8_1);
- tttt2 = _mm_unpacklo_epi16(Org_8_2, Org_8_3);
- tttt3 = _mm_unpackhi_epi16(Org_8_0, Org_8_1);
- tttt4 = _mm_unpackhi_epi16(Org_8_2, Org_8_3);
- tttt5 = _mm_unpacklo_epi32(tttt1, tttt2);
- tttt6 = _mm_unpackhi_epi32(tttt1, tttt2);
- Org_8_0 = _mm_load_si128(m_pSrc_tmp);
- m_pSrc_tmp+=8;;
- Org_8_1 = _mm_load_si128(m_pSrc_tmp);
- m_pSrc_tmp+=8;
- Org_8_2 = _mm_load_si128(m_pSrc_tmp);
- m_pSrc_tmp+=8;
- Org_8_3 = _mm_load_si128(m_pSrc_tmp);
- //m_pSrc_tmp+=8;
- tttt1 = _mm_unpacklo_epi16(Org_8_0, Org_8_1);
- tttt2 = _mm_unpacklo_epi16(Org_8_2, Org_8_3);
- tttt33 = _mm_unpackhi_epi16(Org_8_0, Org_8_1);
- tttt44 = _mm_unpackhi_epi16(Org_8_2, Org_8_3);
- tttt7 = _mm_unpacklo_epi32(tttt1, tttt2);
- tttt8 = _mm_unpackhi_epi32(tttt1, tttt2);
- tttt1 = _mm_unpacklo_epi64(tttt5, tttt7);
- tttt2 = _mm_unpackhi_epi64(tttt5, tttt7);
- _mm_storeu_si128(m_pDst_tmp, tttt1);
- m_pDst_tmp+=stride_ii;
- _mm_storeu_si128(m_pDst_tmp, tttt2);
- m_pDst_tmp+=stride_ii;
- tttt5 = _mm_unpacklo_epi64(tttt6, tttt8);
- tttt7 = _mm_unpackhi_epi64(tttt6, tttt8);
- _mm_storeu_si128(m_pDst_tmp, tttt5);
- m_pDst_tmp+=stride_ii;
- _mm_storeu_si128(m_pDst_tmp, tttt7);
- m_pDst_tmp+=stride_ii;
- //tow
- tttt5 = _mm_unpacklo_epi32(tttt3, tttt4);
- tttt6 = _mm_unpackhi_epi32(tttt3, tttt4);
- tttt7 = _mm_unpacklo_epi32(tttt33, tttt44);
- tttt8 = _mm_unpackhi_epi32(tttt33, tttt44);
- tttt1 = _mm_unpacklo_epi64(tttt5, tttt7);
- tttt2 = _mm_unpackhi_epi64(tttt5, tttt7);
- _mm_storeu_si128(m_pDst_tmp, tttt1);
- m_pDst_tmp+=stride_ii;
- _mm_storeu_si128(m_pDst_tmp, tttt2);
- m_pDst_tmp+=stride_ii;
- tttt5 = _mm_unpacklo_epi64(tttt6, tttt8);
- tttt7 = _mm_unpackhi_epi64(tttt6, tttt8);
- _mm_storeu_si128(m_pDst_tmp, tttt5);
- m_pDst_tmp+=stride_ii;
- _mm_storeu_si128(m_pDst_tmp, tttt7);
__m128i* m_pSrc_tmp = pSrc_128[i]; __m128i* m_pDst_tmp = pDst_128[i]; __m128i Org_8_0,Org_8_1, Org_8_2, Org_8_3; __m128i tttt1,tttt2,tttt3,tttt4,tttt33,tttt44; __m128i tttt5,tttt6, tttt7, tttt8; int stride_ii = dstStride>>3; //one Org_8_0 = _mm_load_si128(m_pSrc_tmp); m_pSrc_tmp+=8; Org_8_1 = _mm_load_si128(m_pSrc_tmp); m_pSrc_tmp+=8; Org_8_2 = _mm_load_si128(m_pSrc_tmp); m_pSrc_tmp+=8; Org_8_3 = _mm_load_si128(m_pSrc_tmp); m_pSrc_tmp+=8; tttt1 = _mm_unpacklo_epi16(Org_8_0, Org_8_1); tttt2 = _mm_unpacklo_epi16(Org_8_2, Org_8_3); tttt3 = _mm_unpackhi_epi16(Org_8_0, Org_8_1); tttt4 = _mm_unpackhi_epi16(Org_8_2, Org_8_3); tttt5 = _mm_unpacklo_epi32(tttt1, tttt2); tttt6 = _mm_unpackhi_epi32(tttt1, tttt2); Org_8_0 = _mm_load_si128(m_pSrc_tmp); m_pSrc_tmp+=8;; Org_8_1 = _mm_load_si128(m_pSrc_tmp); m_pSrc_tmp+=8; Org_8_2 = _mm_load_si128(m_pSrc_tmp); m_pSrc_tmp+=8; Org_8_3 = _mm_load_si128(m_pSrc_tmp); //m_pSrc_tmp+=8; tttt1 = _mm_unpacklo_epi16(Org_8_0, Org_8_1); tttt2 = _mm_unpacklo_epi16(Org_8_2, Org_8_3); tttt33 = _mm_unpackhi_epi16(Org_8_0, Org_8_1); tttt44 = _mm_unpackhi_epi16(Org_8_2, Org_8_3); tttt7 = _mm_unpacklo_epi32(tttt1, tttt2); tttt8 = _mm_unpackhi_epi32(tttt1, tttt2); tttt1 = _mm_unpacklo_epi64(tttt5, tttt7); tttt2 = _mm_unpackhi_epi64(tttt5, tttt7); _mm_storeu_si128(m_pDst_tmp, tttt1); m_pDst_tmp+=stride_ii; _mm_storeu_si128(m_pDst_tmp, tttt2); m_pDst_tmp+=stride_ii; tttt5 = _mm_unpacklo_epi64(tttt6, tttt8); tttt7 = _mm_unpackhi_epi64(tttt6, tttt8); _mm_storeu_si128(m_pDst_tmp, tttt5); m_pDst_tmp+=stride_ii; _mm_storeu_si128(m_pDst_tmp, tttt7); m_pDst_tmp+=stride_ii; //tow tttt5 = _mm_unpacklo_epi32(tttt3, tttt4); tttt6 = _mm_unpackhi_epi32(tttt3, tttt4); tttt7 = _mm_unpacklo_epi32(tttt33, tttt44); tttt8 = _mm_unpackhi_epi32(tttt33, tttt44); tttt1 = _mm_unpacklo_epi64(tttt5, tttt7); tttt2 = _mm_unpackhi_epi64(tttt5, tttt7); _mm_storeu_si128(m_pDst_tmp, tttt1); m_pDst_tmp+=stride_ii; _mm_storeu_si128(m_pDst_tmp, tttt2); m_pDst_tmp+=stride_ii; tttt5 = _mm_unpacklo_epi64(tttt6, tttt8); tttt7 = _mm_unpackhi_epi64(tttt6, tttt8); _mm_storeu_si128(m_pDst_tmp, tttt5); m_pDst_tmp+=stride_ii; _mm_storeu_si128(m_pDst_tmp, tttt7);
要实现的是NXN的转置,如何实现呢:
基于8X8来实现NXN的块或者图像的转置:
这里先把NXN划分为size_case 个8X8, 然后循环调用8X8的转置!
- __m128i* pDst_128[64];
- __m128i* pSrc_128[64];
- int size_case = (blkSize>>3);
- dstStride = dstStride_tmp;
- for(int y = 0; y<size_case; y++)//对所有8x8的块进行地址映射
- for(int x = 0; x<size_case; x++)
- {
- pSrc_128[y*size_case + x] = (__m128i*)(pDst + 8*x + y*8*64);
- pDst_128[y*size_case + x] = (__m128i*)(rpDst + 8*y + x*8*dstStride);
- }
- size_case = size_case*size_case;
- for(int i = 0;i <size_case; i++)//开始转置
- {
- 8x8转置的代码:
- }
__m128i* pDst_128[64];
__m128i* pSrc_128[64];
int size_case = (blkSize>>3);
dstStride = dstStride_tmp;
for(int y = 0; y<size_case; y++)//对所有8x8的块进行地址映射
for(int x = 0; x<size_case; x++)
{
pSrc_128[y*size_case + x] = (__m128i*)(pDst + 8*x + y*8*64);
pDst_128[y*size_case + x] = (__m128i*)(rpDst + 8*y + x*8*dstStride);
}
size_case = size_case*size_case;
for(int i = 0;i <size_case; i++)//开始转置
{
8x8转置的代码:
}
通过比较, 用SSE汇编优化实现转置比用纯 C代码实现的转置速度快5倍左右!
同样在ARM cortext上的汇编优化也是基于这个原理:
主要循环体代码如下:
- VTRN.16 q8, q9
- VTRN.16 q10, q11
- VTRN.16 q4, q5
- VTRN.16 q6, q7
- VTRN.32 q8, q10
- VTRN.32 q9, q11
- VTRN.32 q4, q6
- VTRN.32 q5, q7
- VSWP d17, d8
- VSWP d19, d10
- VSWP d21, d12
- VSWP d23, d14
VTRN.16 q8, q9 VTRN.16 q10, q11 VTRN.16 q4, q5 VTRN.16 q6, q7 VTRN.32 q8, q10 VTRN.32 q9, q11 VTRN.32 q4, q6 VTRN.32 q5, q7 VSWP d17, d8 VSWP d19, d10 VSWP d21, d12 VSWP d23, d14
感兴趣的可以自己调试下!
当然DSP上也是同样的方法, 只是涉及到的指令不同而已!