From f64de08203aaee95ca755336de3e1db85d990198 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Wed, 4 Dec 2024 15:20:07 +0100 Subject: IQ4_XS_R4 (#123) * Adding iq4_xs_r4 This is a 1st working version on Zen4. We get PP-512(LLaMA-3.1-8B) = 226 t/s, so 16% slower than iq4_nl_x4. * iq4_xs_r4: WIP * iq4_xs_r4: Use AVX2 version for matrix x vector on Zen4 * iq4_xs_r4: NEON We get PP-512(LLaMA-3.1-8B) = 115.6 t/s on M2-Max, up from 68.2 t/s for iq4_xs! * DRY --------- Co-authored-by: Iwan Kawrakow --- ggml/src/iqk/iqk_quantize.cpp | 108 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 108 insertions(+) (limited to 'ggml/src/iqk/iqk_quantize.cpp') diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp index f2e6a45e..acef04db 100644 --- a/ggml/src/iqk/iqk_quantize.cpp +++ b/ggml/src/iqk/iqk_quantize.cpp @@ -3572,3 +3572,111 @@ void vec_dot_q6_0_r4_q8_0(int n, float * s, size_t bs, const void * vx, size_t b GGML_UNUSED(bx); GGML_UNUSED(by); } + +// +// ========================================= iq4_xs_r4 +// + +void quantize_row_iq4_xs_r4_ref(const float * x, block_iq4_xs_r4 * y, int64_t k) { + quantize_iq4_xs_r4(x, (void *)y, 4, k/4, nullptr); +} + +void quantize_row_iq4_xs_r4(const float * x, void * y, int64_t k) { + quantize_iq4_xs_r4(x, y, 4, k/4, nullptr); +} + +static void repack_iq4_xs(int nrows, int n_per_row, const block_iq4_xs * x, block_iq4_xs_r4 * y) { + GGML_ASSERT(nrows%4 == 0); + GGML_ASSERT(n_per_row%QK_K == 0); + int nblock = n_per_row/QK_K; + const block_iq4_xs * x4[4]; + for (int row = 0; row < nrows; row += 4) { + for (int k = 0; k < 4; ++k) x4[k] = x + nblock*k; + for (int ibl = 0; ibl < nblock; ++ibl) { + std::memset(y[ibl].scales_l, 0, QK_K/16); + std::memset(y[ibl].scales_h, 0, QK_K/32); + for (int k = 0; k < 4; ++k) { + y[ibl].d[k] = x4[k][ibl].d; + for (int ib = 0; ib < QK_K/32; ++ib) { + uint8_t sl = (x4[k][ibl].scales_l[ib/2] >> 4*(ib%2)) & 0xf; + uint8_t sh = (x4[k][ibl].scales_h >> 2*ib) & 3; + int i = 4*ib + k; + y[ibl].scales_l[i%16] |= (sl << 4*(i/16)); + y[ibl].scales_h[i%8 ] |= (sh << 2*(i/8)); + } + } + for (int ib = 0; ib < QK_K/32; ++ib) { + for (int k = 0; k < 4; ++k) for (int i = 0; i < 4; ++i) { + y[ibl].qs[64*ib+4*k+i+ 0] = (x4[k][ibl].qs[16*ib+i+0] & 0xf) | ((x4[k][ibl].qs[16*ib+i+ 8] & 0x0f) << 4); // 0....3 + 8...11 from each row + y[ibl].qs[64*ib+4*k+i+16] = (x4[k][ibl].qs[16*ib+i+0] >> 4) | ((x4[k][ibl].qs[16*ib+i+ 8] & 0xf0)); // 16...19 + 24...27 from each row + y[ibl].qs[64*ib+4*k+i+32] = (x4[k][ibl].qs[16*ib+i+4] & 0xf) | ((x4[k][ibl].qs[16*ib+i+12] & 0x0f) << 4); // 4....7 + 12...15 from each row + y[ibl].qs[64*ib+4*k+i+48] = (x4[k][ibl].qs[16*ib+i+4] >> 4) | ((x4[k][ibl].qs[16*ib+i+12] & 0xf0)); // 20...23 + 28...31 from each row + } + } + } + x += 4*nblock; + y += nblock; + } +} + +size_t quantize_iq4_xs_r4(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) { + GGML_ASSERT(nrows%4 == 0); + GGML_ASSERT(n_per_row%QK_K == 0); + char * qcur = (char *)dst; + auto row_size = ggml_row_size(GGML_TYPE_IQ4_XS, n_per_row); + std::vector qtmp(4*row_size); + for (int row = 0; row < nrows; row += 4) { + quantize_iq4_xs(src, (void *)qtmp.data(), 4, n_per_row, imatrix); + repack_iq4_xs(4, n_per_row, (const block_iq4_xs *)qtmp.data(), (block_iq4_xs_r4 *)qcur); + qcur += 4*row_size; + src += 4*n_per_row; + } + return nrows*row_size; +} + +void dequantize_row_iq4_xs_r4(const block_iq4_xs_r4 * x, float * y, int64_t k) { + auto n_per_row = k/4; + float * y4[4] = {y, y + n_per_row, y + 2*n_per_row, y + 3*n_per_row}; + int nblock = n_per_row/QK_K; + for (int ibl = 0; ibl < nblock; ++ibl) { + for (int k = 0; k < 4; ++k) { + const float d = GGML_FP16_TO_FP32(x[ibl].d[k]); + for (int ib = 0; ib < QK_K/32; ++ib) { + int is = 4*ib + k; + float dl = d * ((((x[ibl].scales_l[is%16] >> 4*(is/16)) & 0xf) | (((x[ibl].scales_h[is%8] >> 2*(is/8)) & 3) << 4)) - 32); + for (int i = 0; i < 4; ++i) { + y4[k][QK_K*ibl+32*ib+i+ 0] = dl * iq4k_values[x[ibl].qs[64*ib+4*k+i+ 0] & 0xf]; + y4[k][QK_K*ibl+32*ib+i+ 8] = dl * iq4k_values[x[ibl].qs[64*ib+4*k+i+ 0] >> 4]; + y4[k][QK_K*ibl+32*ib+i+16] = dl * iq4k_values[x[ibl].qs[64*ib+4*k+i+16] & 0xf]; + y4[k][QK_K*ibl+32*ib+i+24] = dl * iq4k_values[x[ibl].qs[64*ib+4*k+i+16] >> 4]; + y4[k][QK_K*ibl+32*ib+i+ 4] = dl * iq4k_values[x[ibl].qs[64*ib+4*k+i+32] & 0xf]; + y4[k][QK_K*ibl+32*ib+i+12] = dl * iq4k_values[x[ibl].qs[64*ib+4*k+i+32] >> 4]; + y4[k][QK_K*ibl+32*ib+i+20] = dl * iq4k_values[x[ibl].qs[64*ib+4*k+i+48] & 0xf]; + y4[k][QK_K*ibl+32*ib+i+28] = dl * iq4k_values[x[ibl].qs[64*ib+4*k+i+48] >> 4]; + } + } + } + //dequantize_row_iq4_xs(x + ib, ytmp, QK_K); + //for (int k = 0; k < 4; ++k) { + // for (int l = 0; l < 16; ++l) { + // for (int i = 0; i < 4; ++i) { + // //y4[k][ib*kBlockSize + i + 16*(l%4) + 4*(l/4)] = ytmp[16*l + 4*k + i]; + // y4[k][ib*kBlockSize + i + 8*(l%8) + 4*(l/8)] = ytmp[16*l + 4*k + i]; + // } + // } + //} + } +} + +void vec_dot_iq4_xs_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) { +#if GGML_USE_IQK_MULMAT + if (iqk_mul_mat(1, 1, n, GGML_TYPE_IQ4_XS_R4, vx, 0, GGML_TYPE_Q8_K, vy, 0, s, 0, 0, 1)) { + return; + } +#endif + GGML_ASSERT(n%QK4_NL == 0); + GGML_ASSERT(nrc == 1); + GGML_UNUSED(bs); + GGML_UNUSED(bx); + GGML_UNUSED(by); +} -- cgit v1.2.3