chore : correct typos [no ci] (#20041)
* fix(docs): correct typos found during code review Non-functional changes only: - Fixed minor spelling mistakes in comments - Corrected typos in user-facing strings - No variables, logic, or functional code was modified. Signed-off-by: Marcel Petrick <mail@marcelpetrick.it> * Update docs/backend/CANN.md Co-authored-by: Aaron Teo <taronaeo@gmail.com> * Revert "Auxiliary commit to revert individual files from 846d1c301281178efbc6ce6060ad34c1ebe45af8" This reverts commit 02fcf0c7db661d5ff3eff96b2b2db9fdb7213256. * Update tests/test-backend-ops.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update tests/test-backend-ops.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> --------- Signed-off-by: Marcel Petrick <mail@marcelpetrick.it> Co-authored-by: Aaron Teo <taronaeo@gmail.com> Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
Marcel Petrick
co-authored by
Aaron Teo
Sigbjørn Skjæret
Aaron Teo
Sigbjørn Skjæret
parent
7a99dc85e2
commit
92f7da00b4
@@ -968,7 +968,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
|
||||
|
||||
const int vector_length = ggml_cpu_get_sve_cnt()*8;
|
||||
|
||||
//VLA Implemenation for SVE
|
||||
//VLA Implementation for SVE
|
||||
switch (vector_length) {
|
||||
case 128:
|
||||
{
|
||||
|
||||
@@ -781,7 +781,7 @@ void ggml_gemv_q4_K_8x8_q8_K(int n,
|
||||
|
||||
const uint8_t * q4_base = q4_ptr[b].qs + sb * QK_K;
|
||||
|
||||
// Load the 64 quants from q8K duplicated to use vecdots with the interelaved columns
|
||||
// Load the 64 quants from q8K duplicated to use vecdots with the interleaved columns
|
||||
// but still need the qs to use the low and hi bits from q4
|
||||
const int8_t * q8_base = q8_ptr[b].qs + sb * 64;
|
||||
int8x16_t q8_qs[8];
|
||||
@@ -3796,7 +3796,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n,
|
||||
|
||||
for (int b = 0; b < nb; b++) {
|
||||
// bsums pairs belongs to the same q8_k subblock
|
||||
// 64 elemnts loaded and made sum of 0-7 and 8-15 sum || 16-23 and 24 - 31 sum
|
||||
// 64 elements loaded and made sum of 0-7 and 8-15 sum || 16-23 and 24 - 31 sum
|
||||
const int16x8_t bsums[4]{
|
||||
vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
|
||||
vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
|
||||
|
||||
@@ -423,7 +423,7 @@ void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
|
||||
quants_interleaved[j] = i0;
|
||||
}
|
||||
|
||||
// Masks to shuffle the quants of corresonding sub blocks for rearraning quants for vectorized bsums computation
|
||||
// Masks to shuffle the quants of corresponding sub blocks for rearranging quants for vectorized bsums computation
|
||||
__m256i shuffle_mask_sb2 = _mm256_castsi128_si256(_mm_setr_epi8(0, 1, 0, 1, 4, 5, 6, 7, 8, 9, 8, 9, 12, 13, 14, 15));
|
||||
shuffle_mask_sb2 = _mm256_permute2f128_si256(shuffle_mask_sb2, shuffle_mask_sb2, 0);
|
||||
__m256i shuffle_mask_sb3 = _mm256_castsi128_si256(_mm_setr_epi8(0, 1, 2, 3, 0, 1, 6, 7, 8, 9, 10, 11, 8, 9, 14, 15));
|
||||
@@ -625,7 +625,7 @@ static void gemv_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
|
||||
iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(rhs_vec_0123_3 ,_mm256_shuffle_epi32(rhs_vec_4567_3, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170));
|
||||
iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_3, 177) ,rhs_vec_4567_3, 170), _mm256_shuffle_epi32(lhs_vec_1, 255));
|
||||
|
||||
// Accumulated values multipled with appropriate scales
|
||||
// Accumulated values multiplied with appropriate scales
|
||||
acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
|
||||
}
|
||||
|
||||
@@ -868,7 +868,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
|
||||
const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
|
||||
const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
|
||||
|
||||
// Multiply with appropiate scales and accumulate
|
||||
// Multiply with appropriate scales and accumulate
|
||||
acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
|
||||
acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
|
||||
acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
|
||||
@@ -1076,7 +1076,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
|
||||
const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
|
||||
const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
|
||||
|
||||
// Multiply with appropiate scales and accumulate
|
||||
// Multiply with appropriate scales and accumulate
|
||||
acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
|
||||
acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
|
||||
acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
|
||||
@@ -1257,7 +1257,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
|
||||
// Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
|
||||
const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
|
||||
|
||||
// Multiply with appropiate scales and accumulate
|
||||
// Multiply with appropriate scales and accumulate
|
||||
acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
|
||||
acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
|
||||
acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
|
||||
@@ -1428,7 +1428,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
|
||||
// Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
|
||||
const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
|
||||
|
||||
// Multiply with appropiate scales and accumulate
|
||||
// Multiply with appropriate scales and accumulate
|
||||
acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
|
||||
acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
|
||||
acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
|
||||
@@ -1612,7 +1612,7 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0);
|
||||
|
||||
// Dot product done within 32 bit lanes and accumulated in the same vector
|
||||
// First done for first sub block and thenn for second sub block in each sb
|
||||
// First done for first sub block and then for second sub block in each sb
|
||||
// B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
|
||||
// B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
|
||||
// ...........................................................................
|
||||
@@ -2422,7 +2422,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
|
||||
const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
|
||||
acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
|
||||
acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
|
||||
@@ -2785,7 +2785,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
|
||||
const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
|
||||
acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
|
||||
acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
|
||||
@@ -2802,7 +2802,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
acc_min_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]);
|
||||
}
|
||||
}
|
||||
// Store accumlated values
|
||||
// Store accumulated values
|
||||
for (int i = 0; i < 4; i++) {
|
||||
_mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
|
||||
}
|
||||
@@ -3130,7 +3130,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
|
||||
const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);//GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
|
||||
acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
|
||||
acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
|
||||
@@ -3460,7 +3460,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
|
||||
const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse); //GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
|
||||
acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
|
||||
acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
|
||||
@@ -4268,7 +4268,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
|
||||
const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
|
||||
acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
|
||||
acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
|
||||
@@ -5035,7 +5035,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
acc_min_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]);
|
||||
}
|
||||
}
|
||||
// Store accumlated values
|
||||
// Store accumulated values
|
||||
for (int i = 0; i < 4; i++) {
|
||||
_mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
|
||||
}
|
||||
@@ -5677,7 +5677,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
|
||||
const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
|
||||
acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
|
||||
acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
|
||||
@@ -6349,7 +6349,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
|
||||
const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
|
||||
acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
|
||||
acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
|
||||
|
||||
Reference in New Issue
Block a user