motorhead/qemu
1
0
Fork 0
mirror of https://github.com/Motorhead1991/qemu.git synced 2025-08-02 23:33:54 -06:00
Code Issues Projects Releases Packages Wiki Activity Actions

target/arm: Pass separate addend to {U, S}DOT helpers

Browse source 
For SVE, we potentially have a 4th argument coming from the
movprfx instruction.  Currently we do not optimize movprfx,
so the problem is not visible.

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210525010358.152808-50-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
This commit is contained in:
Richard Henderson 2021-05-24 18:03:15 -07:00 • committed by Peter Maydell
parent 77e786bb95
commit bc2bd6974e
6 changed files with 109 additions and 76 deletions
Download patch file Download diff file Expand all files Collapse all files

120
target/arm/vec_helper.c
View file

@ -375,71 +375,76 @@ void HELPER(sve2_sqrdmlsh_d)(void *vd, void *vn, void *vm,
* All elements are treated equally, no matter where they are.
*/
void HELPER(gvec_sdot_b)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_sdot_b)(void *vd, void *vn, void *vm, void *va, uint32_t desc)
{
intptr_t i, opr_sz = simd_oprsz(desc);
int32_t *d = vd;
int32_t *d = vd, *a = va;
int8_t *n = vn, *m = vm;
for (i = 0; i < opr_sz / 4; ++i) {
d[i] += n[i * 4 + 0] * m[i * 4 + 0]
+ n[i * 4 + 1] * m[i * 4 + 1]
+ n[i * 4 + 2] * m[i * 4 + 2]
+ n[i * 4 + 3] * m[i * 4 + 3];
d[i] = (a[i] +
n[i * 4 + 0] * m[i * 4 + 0] +
n[i * 4 + 1] * m[i * 4 + 1] +
n[i * 4 + 2] * m[i * 4 + 2] +
n[i * 4 + 3] * m[i * 4 + 3]);
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_udot_b)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_udot_b)(void *vd, void *vn, void *vm, void *va, uint32_t desc)
{
intptr_t i, opr_sz = simd_oprsz(desc);
uint32_t *d = vd;
uint32_t *d = vd, *a = va;
uint8_t *n = vn, *m = vm;
for (i = 0; i < opr_sz / 4; ++i) {
d[i] += n[i * 4 + 0] * m[i * 4 + 0]
+ n[i * 4 + 1] * m[i * 4 + 1]
+ n[i * 4 + 2] * m[i * 4 + 2]
+ n[i * 4 + 3] * m[i * 4 + 3];
d[i] = (a[i] +
n[i * 4 + 0] * m[i * 4 + 0] +
n[i * 4 + 1] * m[i * 4 + 1] +
n[i * 4 + 2] * m[i * 4 + 2] +
n[i * 4 + 3] * m[i * 4 + 3]);
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_sdot_h)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_sdot_h)(void *vd, void *vn, void *vm, void *va, uint32_t desc)
{
intptr_t i, opr_sz = simd_oprsz(desc);
int64_t *d = vd;
int64_t *d = vd, *a = va;
int16_t *n = vn, *m = vm;
for (i = 0; i < opr_sz / 8; ++i) {
d[i] += (int64_t)n[i * 4 + 0] * m[i * 4 + 0]
+ (int64_t)n[i * 4 + 1] * m[i * 4 + 1]
+ (int64_t)n[i * 4 + 2] * m[i * 4 + 2]
+ (int64_t)n[i * 4 + 3] * m[i * 4 + 3];
d[i] = (a[i] +
(int64_t)n[i * 4 + 0] * m[i * 4 + 0] +
(int64_t)n[i * 4 + 1] * m[i * 4 + 1] +
(int64_t)n[i * 4 + 2] * m[i * 4 + 2] +
(int64_t)n[i * 4 + 3] * m[i * 4 + 3]);
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_udot_h)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_udot_h)(void *vd, void *vn, void *vm, void *va, uint32_t desc)
{
intptr_t i, opr_sz = simd_oprsz(desc);
uint64_t *d = vd;
uint64_t *d = vd, *a = va;
uint16_t *n = vn, *m = vm;
for (i = 0; i < opr_sz / 8; ++i) {
d[i] += (uint64_t)n[i * 4 + 0] * m[i * 4 + 0]
+ (uint64_t)n[i * 4 + 1] * m[i * 4 + 1]
+ (uint64_t)n[i * 4 + 2] * m[i * 4 + 2]
+ (uint64_t)n[i * 4 + 3] * m[i * 4 + 3];
d[i] = (a[i] +
(uint64_t)n[i * 4 + 0] * m[i * 4 + 0] +
(uint64_t)n[i * 4 + 1] * m[i * 4 + 1] +
(uint64_t)n[i * 4 + 2] * m[i * 4 + 2] +
(uint64_t)n[i * 4 + 3] * m[i * 4 + 3]);
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_sdot_idx_b)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_sdot_idx_b)(void *vd, void *vn, void *vm,
void *va, uint32_t desc)
{
intptr_t i, segend, opr_sz = simd_oprsz(desc), opr_sz_4 = opr_sz / 4;
intptr_t index = simd_data(desc);
int32_t *d = vd;
int32_t *d = vd, *a = va;
int8_t *n = vn;
int8_t *m_indexed = (int8_t *)vm + H4(index) * 4;
@ -455,10 +460,11 @@ void HELPER(gvec_sdot_idx_b)(void *vd, void *vn, void *vm, uint32_t desc)
int8_t m3 = m_indexed[i * 4 + 3];
do {
d[i] += n[i * 4 + 0] * m0
+ n[i * 4 + 1] * m1
+ n[i * 4 + 2] * m2
+ n[i * 4 + 3] * m3;
d[i] = (a[i] +
n[i * 4 + 0] * m0 +
n[i * 4 + 1] * m1 +
n[i * 4 + 2] * m2 +
n[i * 4 + 3] * m3);
} while (++i < segend);
segend = i + 4;
} while (i < opr_sz_4);
@ -466,11 +472,12 @@ void HELPER(gvec_sdot_idx_b)(void *vd, void *vn, void *vm, uint32_t desc)
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_udot_idx_b)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_udot_idx_b)(void *vd, void *vn, void *vm,
void *va, uint32_t desc)
{
intptr_t i, segend, opr_sz = simd_oprsz(desc), opr_sz_4 = opr_sz / 4;
intptr_t index = simd_data(desc);
uint32_t *d = vd;
uint32_t *d = vd, *a = va;
uint8_t *n = vn;
uint8_t *m_indexed = (uint8_t *)vm + H4(index) * 4;
@ -486,10 +493,11 @@ void HELPER(gvec_udot_idx_b)(void *vd, void *vn, void *vm, uint32_t desc)
uint8_t m3 = m_indexed[i * 4 + 3];
do {
d[i] += n[i * 4 + 0] * m0
+ n[i * 4 + 1] * m1
+ n[i * 4 + 2] * m2
+ n[i * 4 + 3] * m3;
d[i] = (a[i] +
n[i * 4 + 0] * m0 +
n[i * 4 + 1] * m1 +
n[i * 4 + 2] * m2 +
n[i * 4 + 3] * m3);
} while (++i < segend);
segend = i + 4;
} while (i < opr_sz_4);
@ -497,11 +505,12 @@ void HELPER(gvec_udot_idx_b)(void *vd, void *vn, void *vm, uint32_t desc)
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_sdot_idx_h)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_sdot_idx_h)(void *vd, void *vn, void *vm,
void *va, uint32_t desc)
{
intptr_t i, opr_sz = simd_oprsz(desc), opr_sz_8 = opr_sz / 8;
intptr_t index = simd_data(desc);
int64_t *d = vd;
int64_t *d = vd, *a = va;
int16_t *n = vn;
int16_t *m_indexed = (int16_t *)vm + index * 4;
@ -509,30 +518,33 @@ void HELPER(gvec_sdot_idx_h)(void *vd, void *vn, void *vm, uint32_t desc)
* Process the entire segment all at once, writing back the results
* only after we've consumed all of the inputs.
*/
for (i = 0; i < opr_sz_8 ; i += 2) {
uint64_t d0, d1;
for (i = 0; i < opr_sz_8; i += 2) {
int64_t d0, d1;
d0 = n[i * 4 + 0] * (int64_t)m_indexed[i * 4 + 0];
d0 = a[i + 0];
d0 += n[i * 4 + 0] * (int64_t)m_indexed[i * 4 + 0];
d0 += n[i * 4 + 1] * (int64_t)m_indexed[i * 4 + 1];
d0 += n[i * 4 + 2] * (int64_t)m_indexed[i * 4 + 2];
d0 += n[i * 4 + 3] * (int64_t)m_indexed[i * 4 + 3];
d1 = n[i * 4 + 4] * (int64_t)m_indexed[i * 4 + 0];
d1 = a[i + 1];
d1 += n[i * 4 + 4] * (int64_t)m_indexed[i * 4 + 0];
d1 += n[i * 4 + 5] * (int64_t)m_indexed[i * 4 + 1];
d1 += n[i * 4 + 6] * (int64_t)m_indexed[i * 4 + 2];
d1 += n[i * 4 + 7] * (int64_t)m_indexed[i * 4 + 3];
d[i + 0] += d0;
d[i + 1] += d1;
d[i + 0] = d0;
d[i + 1] = d1;
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_udot_idx_h)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_udot_idx_h)(void *vd, void *vn, void *vm,
void *va, uint32_t desc)
{
intptr_t i, opr_sz = simd_oprsz(desc), opr_sz_8 = opr_sz / 8;
intptr_t index = simd_data(desc);
uint64_t *d = vd;
uint64_t *d = vd, *a = va;
uint16_t *n = vn;
uint16_t *m_indexed = (uint16_t *)vm + index * 4;
@ -540,22 +552,24 @@ void HELPER(gvec_udot_idx_h)(void *vd, void *vn, void *vm, uint32_t desc)
* Process the entire segment all at once, writing back the results
* only after we've consumed all of the inputs.
*/
for (i = 0; i < opr_sz_8 ; i += 2) {
for (i = 0; i < opr_sz_8; i += 2) {
uint64_t d0, d1;
d0 = n[i * 4 + 0] * (uint64_t)m_indexed[i * 4 + 0];
d0 = a[i + 0];
d0 += n[i * 4 + 0] * (uint64_t)m_indexed[i * 4 + 0];
d0 += n[i * 4 + 1] * (uint64_t)m_indexed[i * 4 + 1];
d0 += n[i * 4 + 2] * (uint64_t)m_indexed[i * 4 + 2];
d0 += n[i * 4 + 3] * (uint64_t)m_indexed[i * 4 + 3];
d1 = n[i * 4 + 4] * (uint64_t)m_indexed[i * 4 + 0];
d1 = a[i + 1];
d1 += n[i * 4 + 4] * (uint64_t)m_indexed[i * 4 + 0];
d1 += n[i * 4 + 5] * (uint64_t)m_indexed[i * 4 + 1];
d1 += n[i * 4 + 6] * (uint64_t)m_indexed[i * 4 + 2];
d1 += n[i * 4 + 7] * (uint64_t)m_indexed[i * 4 + 3];
d[i + 0] += d0;
d[i + 1] += d1;
d[i + 0] = d0;
d[i + 1] = d1;
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}