From c15688ca3a8ab11a93166292b37bf590357627a3 Mon Sep 17 00:00:00 2001 From: Iwan Kawrakow Date: Mon, 23 Jun 2025 10:55:22 +0200 Subject: [PATCH] iq3_k 47.3 t/s -> 166.5 t/s. iq3_k_r4 is at 96.5 t/s. --- ggml/src/iqk/iqk_gemm_iqk_quants.cpp | 67 +++++++++++++++++++++++++++- ggml/src/iqk/iqk_mul_mat.cpp | 1 + 2 files changed, 67 insertions(+), 1 deletion(-) diff --git a/ggml/src/iqk/iqk_gemm_iqk_quants.cpp b/ggml/src/iqk/iqk_gemm_iqk_quants.cpp index db0120d5..41c79855 100644 --- a/ggml/src/iqk/iqk_gemm_iqk_quants.cpp +++ b/ggml/src/iqk/iqk_gemm_iqk_quants.cpp @@ -4137,6 +4137,71 @@ void iqk_convert_iq5_ks_q8_k_r8(int n, const void * vx, size_t bx, void * vy, in } } +void iqk_convert_iq3_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) { + GGML_ASSERT(n%QK_K == 0); + GGML_ASSERT(nrc_x%8 == 0); + + int nb = n/QK_K; + + const block_iq3_k * x8[8]; + + block_q8_k_r8 * y = (block_q8_k_r8 *)vy; + + int8x16x2_t values; + { + auto v1 = vld1_s8(iq3nl_values+0); + auto v2 = vld1_s8(iq3nl_values+8); + values.val[0] = vcombine_s8(v1, v1); + values.val[1] = vcombine_s8(v2, v2); + } + + int8x16x2_t xv[8]; + uint32_t block[8]; + int8_t ls[16]; + + auto ml = vdupq_n_u8(0x03); + auto mh = vdupq_n_u8(0x04); + + for (int ix = 0; ix < nrc_x; ix += 8) { + for (int k = 0; k < 8; ++k) x8[k] = (const block_iq3_k *)((const char *)vx + (ix+k)*bx); + for (int i = 0; i < nb; ++i) { + for (int k = 0; k < 8; ++k) { + float d = GGML_FP16_TO_FP32(x8[k][i].d); + auto extra = x8[k][i].extra; + auto hbits = vld1q_u8_x2(x8[k][i].qh); + uint16_t sh = x8[k][i].scales_h; + for (int i128 = 0; i128 < 2; ++i128) { + + ls[8*i128+0] = ((2*(x8[k][i].scales_l[4*i128+0] & 0xf) + 1) * ((sh & 0x01) ? -1 : 1)); + ls[8*i128+1] = ((2*(x8[k][i].scales_l[4*i128+0] >> 4) + 1) * ((sh & 0x02) ? -1 : 1)); + ls[8*i128+2] = ((2*(x8[k][i].scales_l[4*i128+1] & 0xf) + 1) * ((sh & 0x04) ? -1 : 1)); + ls[8*i128+3] = ((2*(x8[k][i].scales_l[4*i128+1] >> 4) + 1) * ((sh & 0x08) ? -1 : 1)); + ls[8*i128+4] = ((2*(x8[k][i].scales_l[4*i128+2] & 0xf) + 1) * ((sh & 0x10) ? -1 : 1)); + ls[8*i128+5] = ((2*(x8[k][i].scales_l[4*i128+2] >> 4) + 1) * ((sh & 0x20) ? -1 : 1)); + ls[8*i128+6] = ((2*(x8[k][i].scales_l[4*i128+3] & 0xf) + 1) * ((sh & 0x40) ? -1 : 1)); + ls[8*i128+7] = ((2*(x8[k][i].scales_l[4*i128+3] >> 4) + 1) * ((sh & 0x80) ? -1 : 1)); + + auto bits = vld1q_u8_x2(x8[k][i].qs+32*i128); + xv[4*i128+0].val[0] = vqtbl1q_s8(values.val[extra & 1], vorrq_u8(vandq_u8(bits.val[0], ml), vandq_u8(vshlq_n_u8(hbits.val[0], 2), mh))); extra >>= 1; + xv[4*i128+0].val[1] = vqtbl1q_s8(values.val[extra & 1], vorrq_u8(vandq_u8(bits.val[1], ml), vandq_u8(vshlq_n_u8(hbits.val[1], 2), mh))); extra >>= 1; + xv[4*i128+1].val[0] = vqtbl1q_s8(values.val[extra & 1], vorrq_u8(vandq_u8(vshrq_n_u8(bits.val[0], 2), ml), vandq_u8(vshlq_n_u8(hbits.val[0], 1), mh))); extra >>= 1; + xv[4*i128+1].val[1] = vqtbl1q_s8(values.val[extra & 1], vorrq_u8(vandq_u8(vshrq_n_u8(bits.val[1], 2), ml), vandq_u8(vshlq_n_u8(hbits.val[1], 1), mh))); extra >>= 1; + xv[4*i128+2].val[0] = vqtbl1q_s8(values.val[extra & 1], vorrq_u8(vandq_u8(vshrq_n_u8(bits.val[0], 4), ml), vandq_u8(hbits.val[0], mh))); extra >>= 1; + xv[4*i128+2].val[1] = vqtbl1q_s8(values.val[extra & 1], vorrq_u8(vandq_u8(vshrq_n_u8(bits.val[1], 4), ml), vandq_u8(hbits.val[1], mh))); extra >>= 1; + xv[4*i128+3].val[0] = vqtbl1q_s8(values.val[extra & 1], vorrq_u8(vshrq_n_u8(bits.val[0], 6), vandq_u8(vshrq_n_u8(hbits.val[0], 1), mh))); extra >>= 1; + xv[4*i128+3].val[1] = vqtbl1q_s8(values.val[extra & 1], vorrq_u8(vshrq_n_u8(bits.val[1], 6), vandq_u8(vshrq_n_u8(hbits.val[1], 1), mh))); extra >>= 1; + hbits.val[0] = vshrq_n_u8(hbits.val[0], 4); + hbits.val[1] = vshrq_n_u8(hbits.val[1], 4); + sh >>= 8; + } + float dnew = convert_to_q8_k_r8(1.f/127, xv, ls, block, (uint32_t *)y[i].qs + k); + y[i].d[k] = GGML_FP32_TO_FP16(d*dnew); + } + } + y += nb; + } +} + void iqk_convert_iq4_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) { GGML_ASSERT(n%QK_K == 0); GGML_ASSERT(nrc_x%8 == 0); @@ -4281,7 +4346,7 @@ bool iqk_convert_iqk_quants_q80_r8(int type, int n, const void * vx, size_t bx, switch (ggml_type(type)) { case GGML_TYPE_IQ2_KS : iqk_convert_iq2_ks_q8_k_r8(n, vx, bx, vy, nrc_x); break; // case GGML_TYPE_IQ2_K : iqk_convert_iq2_k_q8_k_r8 (n, vx, bx, vy, nrc_x); break; - // case GGML_TYPE_IQ3_K : iqk_convert_iq3_k_q8_k_r8 (n, vx, bx, vy, nrc_x); break; + case GGML_TYPE_IQ3_K : iqk_convert_iq3_k_q8_k_r8 (n, vx, bx, vy, nrc_x); break; case GGML_TYPE_IQ4_KS : iqk_convert_iq4_ks_q8_k_r8(n, vx, bx, vy, nrc_x); break; case GGML_TYPE_IQ4_K : iqk_convert_iq4_k_q8_k_r8 (n, vx, bx, vy, nrc_x); break; case GGML_TYPE_IQ5_KS : iqk_convert_iq5_ks_q8_k_r8(n, vx, bx, vy, nrc_x); break; diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp index 718089a5..e03fd7aa 100644 --- a/ggml/src/iqk/iqk_mul_mat.cpp +++ b/ggml/src/iqk/iqk_mul_mat.cpp @@ -284,6 +284,7 @@ struct MulMat { case GGML_TYPE_IQ2_KS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type; case GGML_TYPE_IQ4_KS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type; case GGML_TYPE_IQ5_KS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type; + case GGML_TYPE_IQ3_K : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type; case GGML_TYPE_IQ4_K : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type; case GGML_TYPE_IQ5_K : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type; case GGML_TYPE_IQ6_K : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;