[Bug target/104188] New: gcc omitting AVX-512 broadcast instruction
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104188
Bug ID: 104188
Summary: gcc omitting AVX-512 broadcast instruction
Product: gcc
Version: 11.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: kvr000 at gmail dot com
Target Milestone: ---
Hi,
there is a bug when generating AVX-512 instructions from intrinsics. The code
is generated correctly in gcc-10 but gcc-11 completely omits the
vbroadcastf32x4 .
gcc version: 11.2.0-7ubuntu2 - 11.2.0
Source code of minimal working example:
// Matrix 4*4 multiplication:
#ifndef NO_VECTORIZE
#ifdef __x86_64__
#include
#include
#endif
#ifdef __aarch64__
#include
#endif
#endif
union Mat44 {
float m[4][4];
#ifndef NO_VECTORIZE
#ifdef __x86_64__
__m128 row[4];
__m256 rowDuet[2];
__m512 rowQuad;
#endif
#ifdef __aarch64__
float32x4_t row[4];
#endif
#endif
};
void matmult_avx512(union Mat44 *out, union Mat44 *a, union Mat44 *b)
{
__m512 a0123 = _mm512_loadu_ps(a->m[0]);
__m512 b = _mm512_broadcast_f32x4(b->row[0]);
__m512 b = _mm512_broadcast_f32x4(b->row[1]);
__m512 b = _mm512_broadcast_f32x4(b->row[2]);
__m512 b = _mm512_broadcast_f32x4(b->row[3]);
__m512 result = _mm512_mul_ps(_mm512_permute_ps(a0123, 0x00), b);
result = _mm512_fmadd_ps(_mm512_permute_ps(a0123, 0x55), b,
result);
result = _mm512_fmadd_ps(_mm512_permute_ps(a0123, 0xaa), b,
result);
result = _mm512_fmadd_ps(_mm512_permute_ps(a0123, 0xff), b,
result);
_mm512_storeu_ps(out->m[0], result);
}
gcc-10 (correct):
endbr64
vmovups (%rsi), %zmm0
vbroadcastf32x4 (%rdx), %zmm6 // note here
vpermilps $0, %zmm0, %zmm1
vmulps %zmm6, %zmm1, %zmm1
vbroadcastf32x4 16(%rdx), %zmm5 // note here
vpermilps $85, %zmm0, %zmm2
vbroadcastf32x4 32(%rdx), %zmm4 // note here
vbroadcastf32x4 48(%rdx), %zmm3 // note here
vfmadd132ps %zmm5, %zmm1, %zmm2
vpermilps $170, %zmm0, %zmm1
vpermilps $255, %zmm0, %zmm0
vfmadd132ps %zmm4, %zmm2, %zmm1
vfmadd132ps %zmm3, %zmm1, %zmm0
vmovups %zmm0, (%rdi)
vzeroupper
ret
gcc-11 (missing vbroadcasatf32x4) :
endbr64
vmovups (%rsi), %zmm0
vpermilps $0, %zmm0, %zmm1
vmulps (%rdx){1to16}, %zmm1, %zmm1
vpermilps $85, %zmm0, %zmm2
vfmadd132ps 16(%rdx){1to16}, %zmm1, %zmm2
vpermilps $170, %zmm0, %zmm1
vpermilps $255, %zmm0, %zmm0
vfmadd132ps 32(%rdx){1to16}, %zmm2, %zmm1
vfmadd132ps 48(%rdx){1to16}, %zmm1, %zmm0
vmovups %zmm0, (%rdi)
vzeroupper
ret
[Bug target/104188] gcc omitting AVX-512 broadcast instruction
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104188
--- Comment #2 from Zbynek Vyskovsky ---
> {1to16} says to broadcast from first element to all 16.
The vbroadcastf32x4 is supposed to copy first four elmenents to 4-7, 8-11 and
12-15 .
> Why do you think this is wrong code?
It doesn't work. It produces the same number for each column for the same row,
likely as a result of above as it uses single element to multiply instead of
four different elements.
[Bug target/104188] gcc omitting AVX-512 broadcast instruction
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104188
--- Comment #4 from Zbynek Vyskovsky ---
Sure, the code:
#include
#ifndef NO_VECTORIZE
#ifdef __x86_64__
#include
#include
#endif
#ifdef __aarch64__
#include
#endif
#endif
typedef union Mat44 {
float m[4][4];
#ifndef NO_VECTORIZE
#ifdef __x86_64__
__m128 row[4];
__m256 rowDuet[2];
__m512 rowQuad;
#endif
#ifdef __aarch64__
float32x4_t row[4];
#endif
#endif
} Mat44;
__attribute__((noipa)) void matmult_avx512(union Mat44 *out, const Mat44 *a,
const Mat44 *b)
{
__m512 a0123 = _mm512_loadu_ps(a->m[0]);
__m512 b = _mm512_broadcast_f32x4(b->row[0]);
__m512 b = _mm512_broadcast_f32x4(b->row[1]);
__m512 b = _mm512_broadcast_f32x4(b->row[2]);
__m512 b = _mm512_broadcast_f32x4(b->row[3]);
__m512 result = _mm512_mul_ps(_mm512_permute_ps(a0123, 0x00), b);
result = _mm512_fmadd_ps(_mm512_permute_ps(a0123, 0x55), b,
result);
result = _mm512_fmadd_ps(_mm512_permute_ps(a0123, 0xaa), b,
result);
result = _mm512_fmadd_ps(_mm512_permute_ps(a0123, 0xff), b,
result);
_mm512_storeu_ps(out->m[0], result);
}
__attribute__((noipa)) void matmult_ref(Mat44 *out, const Mat44 *a, const Mat44
*b)
{
Mat44 t;
for (int i = 0; i < 4; i++) {
for (int j = 0; j < 4; j++) {
t.m[i][j] = a->m[i][0]*b->m[0][j] +
a->m[i][1]*b->m[1][j] + a->m[i][2]*b->m[2][j] + a->m[i][3]*b->m[3][j];
}
}
*out = t;
}
int main(void)
{
Mat44 in = { m: { { 1, 2, 3, 4 }, { 5, 6, 7, 8 }, { 9, 10, 11, 12 }, {
13, 14, 15, 16 } } };
Mat44 avx512_out;
Mat44 ref_out;
matmult_ref(&ref_out, &in, &in);
matmult_avx512(&avx512_out, &in, &in);
for (int r = 0; r < 4; ++r) {
printf("%5.0f %5.0f %5.0f %5.0f %5.0f %5.0f %5.0f
%5.0f\n", avx512_out.m[r][0], avx512_out.m[r][1], avx512_out.m[r][2],
avx512_out.m[r][3], ref_out.m[r][0], ref_out.m[r][1], ref_out.m[r][2],
ref_out.m[r][3]);
}
return 0;
}
Output (note the repeating first column on first side, caused by duplicating
single element instead of four):
90909090 90 100 110 120
202 202 202 202202 228 254 280
314 314 314 314314 356 398 440
426 426 426 426426 484 542 600
[Bug target/104188] [11/12 Regression] gcc omitting AVX-512 broadcast instruction
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104188 --- Comment #10 from Zbynek Vyskovsky --- Thanks for quick fix!
