Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Simplify kc remainder processing for qs8-x8c8 avxvnni #7297

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Simplify kc remainder processing for qs8-x8c8 avxvnni #7297

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
448 changes: 80 additions & 368 deletions src/qs8-packw/gen/qs8-packw-x16c8-gemm-goi-avx256vnni-prfm.c
View file
Open in desktop

Large diffs are not rendered by default.

448 changes: 80 additions & 368 deletions src/qs8-packw/gen/qs8-packw-x16c8-gemm-goi-avx256vnni.c
View file
Open in desktop

Large diffs are not rendered by default.

448 changes: 80 additions & 368 deletions src/qs8-packw/gen/qs8-packw-x16c8-gemm-goi-avxvnni-prfm.c
View file
Open in desktop

Large diffs are not rendered by default.

448 changes: 80 additions & 368 deletions src/qs8-packw/gen/qs8-packw-x16c8-gemm-goi-avxvnni.c
View file
Open in desktop

Large diffs are not rendered by default.

248 changes: 44 additions & 204 deletions src/qs8-packw/gen/qs8-packw-x8c8-gemm-goi-avx256vnni-prfm.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -130,108 +130,28 @@ void xnn_qs8_packw_gemm_goi_ukernel_x8c8__avx256vnni_prfm(
// KC remainder of 1..7
if (k != 0) {
assert(k >= 1 && k <= 7);
__m256i v0 = _mm256_setzero_si256();
__m256i v4 = _mm256_setzero_si256();

if (k & 4) {
v0 = _mm256_insert_epi32(v0, *(const int32_t *)w0, 0);
v0 = _mm256_insert_epi32(v0, *(const int32_t *)w1, 2);
v0 = _mm256_insert_epi32(v0, *(const int32_t *)w2, 4);
v0 = _mm256_insert_epi32(v0, *(const int32_t *)w3, 6);
v4 = _mm256_insert_epi32(v4, *(const int32_t *)w4, 0);
v4 = _mm256_insert_epi32(v4, *(const int32_t *)w5, 2);
v4 = _mm256_insert_epi32(v4, *(const int32_t *)w6, 4);
v4 = _mm256_insert_epi32(v4, *(const int32_t *)w7, 6);
w0 += 4;
w1 += 4;
w2 += 4;
w3 += 4;
w4 += 4;
w5 += 4;
w6 += 4;
w7 += 4;
}
if (k & 2) {
if (k & 4) {
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w0, 2);
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w1, 6);
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w2, 10);
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w3, 14);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w4, 2);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w5, 6);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w6, 10);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w7, 14);
} else {
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w0, 0);
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w1, 4);
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w2, 8);
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w3, 12);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w4, 0);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w5, 4);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w6, 8);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w7, 12);
}

w0 += 2;
w1 += 2;
w2 += 2;
w3 += 2;
w4 += 2;
w5 += 2;
w6 += 2;
w7 += 2;
}
if (k & 1) {
if ((k & 4) && (k & 2)) {
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w0, 6);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w1, 14);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w2, 22);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w3, 30);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w4, 6);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w5, 14);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w6, 22);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w7, 30);
}
else if (k & 4) {
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w0, 4);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w1, 12);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w2, 20);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w3, 28);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w4, 4);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w5, 12);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w6, 20);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w7, 28);
}
else if (k & 2) {
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w0, 2);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w1, 10);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w2, 18);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w3, 26);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w4, 2);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w5, 10);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w6, 18);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w7, 26);
}
else {
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w0, 0);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w1, 8);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w2, 16);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w3, 24);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w4, 0);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w5, 8);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w6, 16);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w7, 24);
}

w0 += 1;
w1 += 1;
w2 += 1;
w3 += 1;
w4 += 1;
w5 += 1;
w6 += 1;
w7 += 1;
}

__m256i vmask = _mm256_srli_epi64(_mm256_set1_epi32(-1), (8 - k) * sizeof(int8_t) * 8);

__m256i v0 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(w0));
v0 = _mm256_blend_epi32(v0, _mm256_set1_epi64x((int64_t) unaligned_load_u64(w1)), 0x0C);
v0 = _mm256_blend_epi32(v0, _mm256_set1_epi64x((int64_t) unaligned_load_u64(w2)), 0x30);
v0 = _mm256_blend_epi32(v0, _mm256_set1_epi64x((int64_t) unaligned_load_u64(w3)), 0xC0);
v0 = _mm256_and_si256(v0, vmask);
__m256i v4 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(w4));
v4 = _mm256_blend_epi32(v4, _mm256_set1_epi64x((int64_t) unaligned_load_u64(w5)), 0x0C);
v4 = _mm256_blend_epi32(v4, _mm256_set1_epi64x((int64_t) unaligned_load_u64(w6)), 0x30);
v4 = _mm256_blend_epi32(v4, _mm256_set1_epi64x((int64_t) unaligned_load_u64(w7)), 0xC0);
v4 = _mm256_and_si256(v4, vmask);

w0 += k;
w1 += k;
w2 += k;
w3 += k;
w4 += k;
w5 += k;
w6 += k;
w7 += k;

vacc0 = _mm256_dpbusd_epi32(vacc0, vone, v0);
vacc4 = _mm256_dpbusd_epi32(vacc4, vone, v4);
Expand Down Expand Up @@ -360,108 +280,28 @@ void xnn_qs8_packw_gemm_goi_ukernel_x8c8__avx256vnni_prfm(
// KC remainder of 1..7
if (k != 0) {
assert(k >= 1 && k <= 7);
__m256i v0 = _mm256_setzero_si256();
__m256i v4 = _mm256_setzero_si256();

if (k & 4) {
v0 = _mm256_insert_epi32(v0, *(const int32_t *)w0, 0);
v0 = _mm256_insert_epi32(v0, *(const int32_t *)w1, 2);
v0 = _mm256_insert_epi32(v0, *(const int32_t *)w2, 4);
v0 = _mm256_insert_epi32(v0, *(const int32_t *)w3, 6);
v4 = _mm256_insert_epi32(v4, *(const int32_t *)w4, 0);
v4 = _mm256_insert_epi32(v4, *(const int32_t *)w5, 2);
v4 = _mm256_insert_epi32(v4, *(const int32_t *)w6, 4);
v4 = _mm256_insert_epi32(v4, *(const int32_t *)w7, 6);
w0 += 4;
w1 += 4;
w2 += 4;
w3 += 4;
w4 += 4;
w5 += 4;
w6 += 4;
w7 += 4;
}
if (k & 2) {
if (k & 4) {
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w0, 2);
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w1, 6);
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w2, 10);
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w3, 14);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w4, 2);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w5, 6);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w6, 10);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w7, 14);
} else {
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w0, 0);
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w1, 4);
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w2, 8);
v0 = _mm256_insert_epi16(v0, *(const int16_t *)w3, 12);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w4, 0);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w5, 4);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w6, 8);
v4 = _mm256_insert_epi16(v4, *(const int16_t *)w7, 12);
}

w0 += 2;
w1 += 2;
w2 += 2;
w3 += 2;
w4 += 2;
w5 += 2;
w6 += 2;
w7 += 2;
}
if (k & 1) {
if ((k & 4) && (k & 2)) {
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w0, 6);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w1, 14);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w2, 22);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w3, 30);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w4, 6);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w5, 14);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w6, 22);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w7, 30);
}
else if (k & 4) {
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w0, 4);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w1, 12);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w2, 20);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w3, 28);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w4, 4);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w5, 12);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w6, 20);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w7, 28);
}
else if (k & 2) {
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w0, 2);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w1, 10);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w2, 18);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w3, 26);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w4, 2);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w5, 10);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w6, 18);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w7, 26);
}
else {
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w0, 0);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w1, 8);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w2, 16);
v0 = _mm256_insert_epi8(v0, *(const int8_t *)w3, 24);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w4, 0);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w5, 8);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w6, 16);
v4 = _mm256_insert_epi8(v4, *(const int8_t *)w7, 24);
}

w0 += 1;
w1 += 1;
w2 += 1;
w3 += 1;
w4 += 1;
w5 += 1;
w6 += 1;
w7 += 1;
}

__m256i vmask = _mm256_srli_epi64(_mm256_set1_epi32(-1), (8 - k) * sizeof(int8_t) * 8);

__m256i v0 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(w0));
v0 = _mm256_blend_epi32(v0, _mm256_set1_epi64x((int64_t) unaligned_load_u64(w1)), 0x0C);
v0 = _mm256_blend_epi32(v0, _mm256_set1_epi64x((int64_t) unaligned_load_u64(w2)), 0x30);
v0 = _mm256_blend_epi32(v0, _mm256_set1_epi64x((int64_t) unaligned_load_u64(w3)), 0xC0);
v0 = _mm256_and_si256(v0, vmask);
__m256i v4 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(w4));
v4 = _mm256_blend_epi32(v4, _mm256_set1_epi64x((int64_t) unaligned_load_u64(w5)), 0x0C);
v4 = _mm256_blend_epi32(v4, _mm256_set1_epi64x((int64_t) unaligned_load_u64(w6)), 0x30);
v4 = _mm256_blend_epi32(v4, _mm256_set1_epi64x((int64_t) unaligned_load_u64(w7)), 0xC0);
v4 = _mm256_and_si256(v4, vmask);

w0 += k;
w1 += k;
w2 += k;
w3 += k;
w4 += k;
w5 += k;
w6 += k;
w7 += k;

vacc0 = _mm256_dpbusd_epi32(vacc0, vone, v0);
vacc4 = _mm256_dpbusd_epi32(vacc4, vone, v4);
Expand Down
Loading