Transposing a 8x8 matrix can be achieved by making four 4x4 matrices, and transposing each of them. This is not want I'm going for.
In another question, one answer gave a solution that would only require 24 instructions for an 8x8 matrix. However, this does not apply to floats.
Since the AVX2 contains registers of 256 bits, each register would fit eight 32 bits integers (floats). But the question is:
How to transpose an 8x8 float matrix, using AVX/AVX2, with the smallest instructions possible?
Answer
I already answered this question Fast memory transpose with SSE, AVX, and OpenMP.
Let me repeat the solution for transposing an 8x8 float matrix with AVX. Let me know if this is any faster than using 4x4 blocks and _MM_TRANSPOSE4_PS. I used it for a kernel in a larger matrix transpose which was memory bound so that was probably not a fair test.
inline void transpose8_ps(__m256 &row0, __m256 &row1, __m256 &row2, __m256 &row3, __m256 &row4, __m256 &row5, __m256 &row6, __m256 &row7) {
__m256 __t0, __t1, __t2, __t3, __t4, __t5, __t6, __t7;
__m256 __tt0, __tt1, __tt2, __tt3, __tt4, __tt5, __tt6, __tt7;
__t0 = _mm256_unpacklo_ps(row0, row1);
__t1 = _mm256_unpackhi_ps(row0, row1);
__t2 = _mm256_unpacklo_ps(row2, row3);
__t3 = _mm256_unpackhi_ps(row2, row3);
__t4 = _mm256_unpacklo_ps(row4, row5);
__t5 = _mm256_unpackhi_ps(row4, row5);
__t6 = _mm256_unpacklo_ps(row6, row7);
__t7 = _mm256_unpackhi_ps(row6, row7);
__tt0 = _mm256_shuffle_ps(__t0,__t2,_MM_SHUFFLE(1,0,1,0));
__tt1 = _mm256_shuffle_ps(__t0,__t2,_MM_SHUFFLE(3,2,3,2));
__tt2 = _mm256_shuffle_ps(__t1,__t3,_MM_SHUFFLE(1,0,1,0));
__tt3 = _mm256_shuffle_ps(__t1,__t3,_MM_SHUFFLE(3,2,3,2));
__tt4 = _mm256_shuffle_ps(__t4,__t6,_MM_SHUFFLE(1,0,1,0));
__tt5 = _mm256_shuffle_ps(__t4,__t6,_MM_SHUFFLE(3,2,3,2));
__tt6 = _mm256_shuffle_ps(__t5,__t7,_MM_SHUFFLE(1,0,1,0));
__tt7 = _mm256_shuffle_ps(__t5,__t7,_MM_SHUFFLE(3,2,3,2));
row0 = _mm256_permute2f128_ps(__tt0, __tt4, 0x20);
row1 = _mm256_permute2f128_ps(__tt1, __tt5, 0x20);
row2 = _mm256_permute2f128_ps(__tt2, __tt6, 0x20);
row3 = _mm256_permute2f128_ps(__tt3, __tt7, 0x20);
row4 = _mm256_permute2f128_ps(__tt0, __tt4, 0x31);
row5 = _mm256_permute2f128_ps(__tt1, __tt5, 0x31);
row6 = _mm256_permute2f128_ps(__tt2, __tt6, 0x31);
row7 = _mm256_permute2f128_ps(__tt3, __tt7, 0x31);
}
Based on this comment I learned that there are more efficient methods which to do the 8x8 transpose. See Example 11-19 and and 11-20 in the Intel optimization manual under section "11.11 Handling Port 5 Pressure". Example 11-19 uses the same number of instructions but reduces the pressure on port5 by using blends which go to port0 as well. I may implement this with intrinsics at some point but I don't have a need for this at this point.
I looked more carefully into Example 11-19 and 11-20 in the Intel Manuals I mentioned above. It turns out that example 11-19 uses 4 more shuffle operations than necessary. It has 8 unpack, 12 shuffles, and 8 128-bit permutes. My method uses 4 fewer shuffles. They replace 8 of the shuffles with blends. So 4 shuffles and 8 blends. I doubt that's better than my method with only eight shuffles.
Example 11-20 is, however, an improvement if you need to load the matrix from memory. This uses 8 unpacks, 8 inserts, 8 shuffles, 8 128-bit loads, and 8 stores. The 128-bit loads reduce the port pressure. I went ahead and implemented this using intrinsics.
//Example 11-20. 8x8 Matrix Transpose Using VINSERTF128 loads
void tran(float* mat, float* matT) {
__m256 r0, r1, r2, r3, r4, r5, r6, r7;
__m256 t0, t1, t2, t3, t4, t5, t6, t7;
r0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[0*8+0])), _mm_load_ps(&mat[4*8+0]), 1);
r1 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[1*8+0])), _mm_load_ps(&mat[5*8+0]), 1);
r2 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[2*8+0])), _mm_load_ps(&mat[6*8+0]), 1);
r3 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[3*8+0])), _mm_load_ps(&mat[7*8+0]), 1);
r4 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[0*8+4])), _mm_load_ps(&mat[4*8+4]), 1);
r5 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[1*8+4])), _mm_load_ps(&mat[5*8+4]), 1);
r6 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[2*8+4])), _mm_load_ps(&mat[6*8+4]), 1);
r7 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[3*8+4])), _mm_load_ps(&mat[7*8+4]), 1);
t0 = _mm256_unpacklo_ps(r0,r1);
t1 = _mm256_unpackhi_ps(r0,r1);
t2 = _mm256_unpacklo_ps(r2,r3);
t3 = _mm256_unpackhi_ps(r2,r3);
t4 = _mm256_unpacklo_ps(r4,r5);
t5 = _mm256_unpackhi_ps(r4,r5);
t6 = _mm256_unpacklo_ps(r6,r7);
t7 = _mm256_unpackhi_ps(r6,r7);
r0 = _mm256_shuffle_ps(t0,t2, 0x44);
r1 = _mm256_shuffle_ps(t0,t2, 0xEE);
r2 = _mm256_shuffle_ps(t1,t3, 0x44);
r3 = _mm256_shuffle_ps(t1,t3, 0xEE);
r4 = _mm256_shuffle_ps(t4,t6, 0x44);
r5 = _mm256_shuffle_ps(t4,t6, 0xEE);
r6 = _mm256_shuffle_ps(t5,t7, 0x44);
r7 = _mm256_shuffle_ps(t5,t7, 0xEE);
_mm256_store_ps(&matT[0*8], r0);
_mm256_store_ps(&matT[1*8], r1);
_mm256_store_ps(&matT[2*8], r2);
_mm256_store_ps(&matT[3*8], r3);
_mm256_store_ps(&matT[4*8], r4);
_mm256_store_ps(&matT[5*8], r5);
_mm256_store_ps(&matT[6*8], r6);
_mm256_store_ps(&matT[7*8], r7);
}
So I looked into example 11-19 again. The basic idea as far as I can tell is that two shuffle instructions (shufps) can be replaced by one shuffle and two blends. For example
r0 = _mm256_shuffle_ps(t0,t2, 0x44);
r1 = _mm256_shuffle_ps(t0,t2, 0xEE);
can be replace with
v = _mm256_shuffle_ps(t0,t2, 0x4E);
r0 = _mm256_blend_ps(t0, v, 0xCC);
r1 = _mm256_blend_ps(t2, v, 0x33);
This explains why my original code used 8 shuffles and Example 11-19 uses 4 shuffles and eight blends.
The blends are good for throughput because shuffles only go to one port (creating a bottleneck on the shuffle port), but blends can run on multiple ports and thus don't compete. But what is better: 8 shuffles or 4 shuffles and 8 blends?
This has to be tested, and can depend on surrounding code. If you mostly bottleneck on total uop throughput with a lot of other uops in the loop that don't need port 5, you might go for the pure shuffle version. Ideally you should do some computation on the transposed data before storing it, while it's already in registers. See https://agner.org/optimize/ and other performance links in the x86 tag wiki.
I don't, however, see a way to replace the unpack instructions with blends.
Here is full code which combines Example 11-19 converting 2 shuffles to 1 shuffle and two blends and Example 11-20 which uses vinsertf128 loads (which on Intel Haswell/Skylake CPUs are 2 uops: one ALU for any port, one memory. They unfortunately don't micro-fuse. vinsertf128 with all register operands is 1 uop for the shuffle port on Intel, so this is good because the compiler folds the load into a memory operand for vinsertf128.) This has the advantage of only needing the source data 16-byte aligned for maximum performance, avoiding any cache-line splits.
#include <stdio.h>
#include <x86intrin.h>
#include <omp.h>
/*
void tran(float* mat, float* matT) {
__m256 r0, r1, r2, r3, r4, r5, r6, r7;
__m256 t0, t1, t2, t3, t4, t5, t6, t7;
r0 = _mm256_load_ps(&mat[0*8]);
r1 = _mm256_load_ps(&mat[1*8]);
r2 = _mm256_load_ps(&mat[2*8]);
r3 = _mm256_load_ps(&mat[3*8]);
r4 = _mm256_load_ps(&mat[4*8]);
r5 = _mm256_load_ps(&mat[5*8]);
r6 = _mm256_load_ps(&mat[6*8]);
r7 = _mm256_load_ps(&mat[7*8]);
t0 = _mm256_unpacklo_ps(r0, r1);
t1 = _mm256_unpackhi_ps(r0, r1);
t2 = _mm256_unpacklo_ps(r2, r3);
t3 = _mm256_unpackhi_ps(r2, r3);
t4 = _mm256_unpacklo_ps(r4, r5);
t5 = _mm256_unpackhi_ps(r4, r5);
t6 = _mm256_unpacklo_ps(r6, r7);
t7 = _mm256_unpackhi_ps(r6, r7);
r0 = _mm256_shuffle_ps(t0,t2,_MM_SHUFFLE(1,0,1,0));
r1 = _mm256_shuffle_ps(t0,t2,_MM_SHUFFLE(3,2,3,2));
r2 = _mm256_shuffle_ps(t1,t3,_MM_SHUFFLE(1,0,1,0));
r3 = _mm256_shuffle_ps(t1,t3,_MM_SHUFFLE(3,2,3,2));
r4 = _mm256_shuffle_ps(t4,t6,_MM_SHUFFLE(1,0,1,0));
r5 = _mm256_shuffle_ps(t4,t6,_MM_SHUFFLE(3,2,3,2));
r6 = _mm256_shuffle_ps(t5,t7,_MM_SHUFFLE(1,0,1,0));
r7 = _mm256_shuffle_ps(t5,t7,_MM_SHUFFLE(3,2,3,2));
t0 = _mm256_permute2f128_ps(r0, r4, 0x20);
t1 = _mm256_permute2f128_ps(r1, r5, 0x20);
t2 = _mm256_permute2f128_ps(r2, r6, 0x20);
t3 = _mm256_permute2f128_ps(r3, r7, 0x20);
t4 = _mm256_permute2f128_ps(r0, r4, 0x31);
t5 = _mm256_permute2f128_ps(r1, r5, 0x31);
t6 = _mm256_permute2f128_ps(r2, r6, 0x31);
t7 = _mm256_permute2f128_ps(r3, r7, 0x31);
_mm256_store_ps(&matT[0*8], t0);
_mm256_store_ps(&matT[1*8], t1);
_mm256_store_ps(&matT[2*8], t2);
_mm256_store_ps(&matT[3*8], t3);
_mm256_store_ps(&matT[4*8], t4);
_mm256_store_ps(&matT[5*8], t5);
_mm256_store_ps(&matT[6*8], t6);
_mm256_store_ps(&matT[7*8], t7);
}
*/
void tran(float* mat, float* matT) {
__m256 r0, r1, r2, r3, r4, r5, r6, r7;
__m256 t0, t1, t2, t3, t4, t5, t6, t7;
r0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[0*8+0])), _mm_load_ps(&mat[4*8+0]), 1);
r1 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[1*8+0])), _mm_load_ps(&mat[5*8+0]), 1);
r2 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[2*8+0])), _mm_load_ps(&mat[6*8+0]), 1);
r3 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[3*8+0])), _mm_load_ps(&mat[7*8+0]), 1);
r4 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[0*8+4])), _mm_load_ps(&mat[4*8+4]), 1);
r5 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[1*8+4])), _mm_load_ps(&mat[5*8+4]), 1);
r6 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[2*8+4])), _mm_load_ps(&mat[6*8+4]), 1);
r7 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&mat[3*8+4])), _mm_load_ps(&mat[7*8+4]), 1);
t0 = _mm256_unpacklo_ps(r0,r1);
t1 = _mm256_unpackhi_ps(r0,r1);
t2 = _mm256_unpacklo_ps(r2,r3);
t3 = _mm256_unpackhi_ps(r2,r3);
t4 = _mm256_unpacklo_ps(r4,r5);
t5 = _mm256_unpackhi_ps(r4,r5);
t6 = _mm256_unpacklo_ps(r6,r7);
t7 = _mm256_unpackhi_ps(r6,r7);
__m256 v;
//r0 = _mm256_shuffle_ps(t0,t2, 0x44);
//r1 = _mm256_shuffle_ps(t0,t2, 0xEE);
v = _mm256_shuffle_ps(t0,t2, 0x4E);
r0 = _mm256_blend_ps(t0, v, 0xCC);
r1 = _mm256_blend_ps(t2, v, 0x33);
//r2 = _mm256_shuffle_ps(t1,t3, 0x44);
//r3 = _mm256_shuffle_ps(t1,t3, 0xEE);
v = _mm256_shuffle_ps(t1,t3, 0x4E);
r2 = _mm256_blend_ps(t1, v, 0xCC);
r3 = _mm256_blend_ps(t3, v, 0x33);
//r4 = _mm256_shuffle_ps(t4,t6, 0x44);
//r5 = _mm256_shuffle_ps(t4,t6, 0xEE);
v = _mm256_shuffle_ps(t4,t6, 0x4E);
r4 = _mm256_blend_ps(t4, v, 0xCC);
r5 = _mm256_blend_ps(t6, v, 0x33);
//r6 = _mm256_shuffle_ps(t5,t7, 0x44);
//r7 = _mm256_shuffle_ps(t5,t7, 0xEE);
v = _mm256_shuffle_ps(t5,t7, 0x4E);
r6 = _mm256_blend_ps(t5, v, 0xCC);
r7 = _mm256_blend_ps(t7, v, 0x33);
_mm256_store_ps(&matT[0*8], r0);
_mm256_store_ps(&matT[1*8], r1);
_mm256_store_ps(&matT[2*8], r2);
_mm256_store_ps(&matT[3*8], r3);
_mm256_store_ps(&matT[4*8], r4);
_mm256_store_ps(&matT[5*8], r5);
_mm256_store_ps(&matT[6*8], r6);
_mm256_store_ps(&matT[7*8], r7);
}
int verify(float *mat) {
int i,j;
int error = 0;
for(i=0; i<8; i++) {
for(j=0; j<8; j++) {
if(mat[j*8+i] != 1.0f*i*8+j) error++;
}
}
return error;
}
void print_mat(float *mat) {
int i,j;
for(i=0; i<8; i++) {
for(j=0; j<8; j++) printf("%2.0f ", mat[i*8+j]);
puts("");
}
puts("");
}
int main(void) {
int i,j, rep;
float mat[64] __attribute__((aligned(64)));
float matT[64] __attribute__((aligned(64)));
double dtime;
rep = 10000000;
for(i=0; i<64; i++) mat[i] = i;
print_mat(mat);
tran(mat,matT);
//dtime = -omp_get_wtime();
//tran(mat, matT, rep);
//dtime += omp_get_wtime();
printf("errors %d\n", verify(matT));
//printf("dtime %f\n", dtime);
print_mat(matT);
}