diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h index 427c1146e4..a1b198a15c 100644 --- a/ggml/src/ggml-cpu/arch-fallback.h +++ b/ggml/src/ggml-cpu/arch-fallback.h @@ -176,8 +176,9 @@ #define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0 // repack.cpp +#define ggml_quantize_mat_q8_0_4x1_generic ggml_quantize_mat_q8_0_4x1 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4 -#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8 +#define ggml_quantize_mat_q8_K_4x1_generic ggml_quantize_mat_q8_K_4x1 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 diff --git a/ggml/src/ggml-cpu/arch/riscv/repack.cpp b/ggml/src/ggml-cpu/arch/riscv/repack.cpp index 2a35ff9ad8..8ef3d2b754 100644 --- a/ggml/src/ggml-cpu/arch/riscv/repack.cpp +++ b/ggml/src/ggml-cpu/arch/riscv/repack.cpp @@ -24,6 +24,94 @@ #define UNUSED GGML_UNUSED +void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { + assert(QK8_0 == 32); + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + +#if defined(__riscv_v_intrinsic) + block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy; + const size_t vl_calc = __riscv_vsetvl_e32m8(QK8_0); + const size_t vl_save = __riscv_vsetvl_e64m2(4); + vfloat32m1_t v_scalar_zero = __riscv_vfmv_s_f_f32m1(0.0f, __riscv_vsetvl_e32m1(1)); + + for (int i = 0; i < nb; i++) { + const float *x_block_base = x + i * QK8_0; + vint8m2_t q_r0, q_r1, q_r2, q_r3; + { + vfloat32m8_t v_src = __riscv_vle32_v_f32m8(x_block_base + 0 * k, vl_calc); + vfloat32m8_t v_abs = __riscv_vfabs_v_f32m8(v_src, vl_calc); + vfloat32m1_t v_max = __riscv_vfredmax_vs_f32m8_f32m1(v_abs, v_scalar_zero, vl_calc); + float amax = __riscv_vfmv_f_s_f32m1_f32(v_max); + + float d = amax / 127.0f; + y[i].d[0] = GGML_CPU_FP32_TO_FP16(d); + + float id = d ? 1.0f / d : 0.0f; + vfloat32m8_t v_scaled = __riscv_vfmul_vf_f32m8(v_src, id, vl_calc); + vint16m4_t v_i16 = __riscv_vfncvt_x_f_w_i16m4_rm(v_scaled, 4, vl_calc); + q_r0 = __riscv_vncvt_x_x_w_i8m2(v_i16, vl_calc); + } + asm volatile ("" ::: "memory"); + + { + vfloat32m8_t v_src = __riscv_vle32_v_f32m8(x_block_base + 1 * k, vl_calc); + vfloat32m8_t v_abs = __riscv_vfabs_v_f32m8(v_src, vl_calc); + vfloat32m1_t v_max = __riscv_vfredmax_vs_f32m8_f32m1(v_abs, v_scalar_zero, vl_calc); + float amax = __riscv_vfmv_f_s_f32m1_f32(v_max); + + float d = amax / 127.0f; + y[i].d[1] = GGML_CPU_FP32_TO_FP16(d); + float id = d ? 1.0f / d : 0.0f; + + vfloat32m8_t v_scaled = __riscv_vfmul_vf_f32m8(v_src, id, vl_calc); + vint16m4_t v_i16 = __riscv_vfncvt_x_f_w_i16m4_rm(v_scaled, 4, vl_calc); + q_r1 = __riscv_vncvt_x_x_w_i8m2(v_i16, vl_calc); + } + asm volatile ("" ::: "memory"); + { + vfloat32m8_t v_src = __riscv_vle32_v_f32m8(x_block_base + 2 * k, vl_calc); + vfloat32m8_t v_abs = __riscv_vfabs_v_f32m8(v_src, vl_calc); + vfloat32m1_t v_max = __riscv_vfredmax_vs_f32m8_f32m1(v_abs, v_scalar_zero, vl_calc); + float amax = __riscv_vfmv_f_s_f32m1_f32(v_max); + + float d = amax / 127.0f; + y[i].d[2] = GGML_CPU_FP32_TO_FP16(d); + float id = d ? 1.0f / d : 0.0f; + + vfloat32m8_t v_scaled = __riscv_vfmul_vf_f32m8(v_src, id, vl_calc); + vint16m4_t v_i16 = __riscv_vfncvt_x_f_w_i16m4_rm(v_scaled, 4, vl_calc); + q_r2 = __riscv_vncvt_x_x_w_i8m2(v_i16, vl_calc); + } + asm volatile ("" ::: "memory"); + { + vfloat32m8_t v_src = __riscv_vle32_v_f32m8(x_block_base + 3 * k, vl_calc); + vfloat32m8_t v_abs = __riscv_vfabs_v_f32m8(v_src, vl_calc); + vfloat32m1_t v_max = __riscv_vfredmax_vs_f32m8_f32m1(v_abs, v_scalar_zero, vl_calc); + float amax = __riscv_vfmv_f_s_f32m1_f32(v_max); + + float d = amax / 127.0f; + y[i].d[3] = GGML_CPU_FP32_TO_FP16(d); + float id = d ? 1.0f / d : 0.0f; + + vfloat32m8_t v_scaled = __riscv_vfmul_vf_f32m8(v_src, id, vl_calc); + vint16m4_t v_i16 = __riscv_vfncvt_x_f_w_i16m4_rm(v_scaled, 4, vl_calc); + q_r3 = __riscv_vncvt_x_x_w_i8m2(v_i16, vl_calc); + } + vint64m2_t v_q64_r0 = __riscv_vreinterpret_v_i8m2_i64m2(q_r0); + vint64m2_t v_q64_r1 = __riscv_vreinterpret_v_i8m2_i64m2(q_r1); + vint64m2_t v_q64_r2 = __riscv_vreinterpret_v_i8m2_i64m2(q_r2); + vint64m2_t v_q64_r3 = __riscv_vreinterpret_v_i8m2_i64m2(q_r3); + vint64m2x4_t v_quant_tuple = __riscv_vcreate_v_i64m2x4(v_q64_r0, v_q64_r1, v_q64_r2, v_q64_r3); + __riscv_vsseg4e64_v_i64m2x4((int64_t*)y[i].qs, v_quant_tuple, vl_save); + } +#else + UNUSED(nb); + UNUSED(y); + ggml_quantize_mat_q8_0_4x4_generic(x, vy, k); +#endif +} + void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; @@ -115,6 +203,484 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); } +void ggml_gemv_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined __riscv_v_intrinsic + const vint8mf2_t values = __riscv_vle8_v_i8mf2(kvalues_iq4nl, 16); + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x16 * b_ptr = (const block_q4_0x16 *) vx + (x * nb); + + // 1x16 Accumulator + vfloat32m2_t sumf = __riscv_vfmv_v_f_f32m2(0.0f, 16); + + for (int l = 0; l < nb; l++) { + // 1x16 Integer Accumulator + vint16m1_t sumi_0_lo_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_0_hi_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + + // Accumulation loop. + for (int i = 0; i < QK4_0 / 2; i++) { + // Load `b_ptr`. + const vint8mf2_t b_0_packed = __riscv_vle8_v_i8mf2((const int8_t *)&b_ptr[l].qs[i * 16], 16); + const vint8mf2_t b_0_lo = __riscv_vsra_vx_i8mf2(__riscv_vsll_vx_i8mf2(b_0_packed, 4, 16), 4, 16); + const vint8mf2_t b_0_hi = __riscv_vsra_vx_i8mf2(b_0_packed, 4, 16); + + sumi_0_lo_16 = __riscv_vwmacc_vx_i16m1(sumi_0_lo_16, a_ptr[l].qs[i], b_0_lo, 16); + sumi_0_hi_16 = __riscv_vwmacc_vx_i16m1(sumi_0_hi_16, a_ptr[l].qs[16 + i], b_0_hi, 16); + } + + const vint32m2_t sumi = __riscv_vwadd_vv_i32m2(sumi_0_lo_16, sumi_0_hi_16, 16); + + const vfloat16m1_t b_d = __riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16); + const vfloat32m2_t d_0 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d, 16); + + sumf = __riscv_vfmacc_vv_f32m2(sumf, __riscv_vfcvt_f_x_v_f32m2(sumi, 16), d_0, 16); + } + + __riscv_vse32_v_f32m2(s + x * 16, sumf, 16); + } + return; +#endif + ggml_gemv_q4_0_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + +void ggml_gemv_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK_K; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined __riscv_v_intrinsic + const block_q8_K * a_ptr = (const block_q8_K *) vy; + + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_Kx16 * b_ptr = (const block_q4_Kx16 *) vx + (x * nb); + + // 1x16 Accumulator + vfloat32m2_t sumf = __riscv_vfmv_v_f_f32m2(0.0f, 16); + + for (int l = 0; l < nb; l++) { + vint32m2_t sumi = __riscv_vmv_v_x_i32m2(0, 16); + + // We process 4 sub-blocks at once. + for (int j = 0; j < QK_K / 128; j++) { + // Extract the scales and the mins. + // + // Low bits. + vuint8m2_t scales_mins_lo = __riscv_vle8_v_u8m2(&b_ptr[l].scales[j * 64], 64); + vuint8m2_t scales_lo = __riscv_vand_vx_u8m2(scales_mins_lo, 0x0F, 64); + vuint8m2_t mins_lo = __riscv_vsrl_vx_u8m2(scales_mins_lo, 4, 64); + + // High bits. + vuint8m2_t scales_mins_hi = __riscv_vle8_v_u8m2(&b_ptr[l].scales[128], 64); + vuint8m2_t scales_hi; + vuint8m2_t mins_hi; + if (!j) { + scales_hi = __riscv_vsll_vx_u8m2(__riscv_vand_vx_u8m2(scales_mins_hi, 0x03, 64), 4, 64); + mins_hi = __riscv_vsll_vx_u8m2(__riscv_vand_vx_u8m2(scales_mins_hi, 0x0C, 64), 2, 64); + } else { + scales_hi = __riscv_vand_vx_u8m2(scales_mins_hi, 0x30, 64); + mins_hi = __riscv_vsrl_vx_u8m2(__riscv_vand_vx_u8m2(scales_mins_hi, 0xC0, 64), 2, 64); + } + vuint16m4_t scales = __riscv_vzext_vf2_u16m4(__riscv_vor_vv_u8m2(scales_hi, scales_lo, 64), 64); + vint16m4_t mins = __riscv_vreinterpret_v_u16m4_i16m4(__riscv_vzext_vf2_u16m4(__riscv_vor_vv_u8m2(mins_hi, mins_lo, 64), 64)); + + // Reduce the mins and multiply with `dmin`. + // + // Correct in `sumf`. + vint32m2_t bsums = __riscv_vmv_v_x_i32m2(0, 16); + bsums = __riscv_vwmacc_vx_i32m2(bsums, a_ptr[l].bsums[j * 8] + a_ptr[l].bsums[j * 8 + 1], __riscv_vget_v_i16m4_i16m1(mins, 0), 16); + bsums = __riscv_vwmacc_vx_i32m2(bsums, a_ptr[l].bsums[j * 8 + 2] + a_ptr[l].bsums[j * 8 + 3], __riscv_vget_v_i16m4_i16m1(mins, 1), 16); + bsums = __riscv_vwmacc_vx_i32m2(bsums, a_ptr[l].bsums[j * 8 + 4] + a_ptr[l].bsums[j * 8 + 5], __riscv_vget_v_i16m4_i16m1(mins, 2), 16); + bsums = __riscv_vwmacc_vx_i32m2(bsums, a_ptr[l].bsums[j * 8 + 6] + a_ptr[l].bsums[j * 8 + 7], __riscv_vget_v_i16m4_i16m1(mins, 3), 16); + + const vfloat32m2_t dmins_d = __riscv_vfmul_vf_f32m2( + __riscv_vfwcvt_f_f_v_f32m2(__riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].dmin, 16), 16), a_ptr[l].d, 16); + sumf = __riscv_vfsub_vv_f32m2(sumf, __riscv_vfmul_vv_f32m2(dmins_d, __riscv_vfcvt_f_x_v_f32m2(bsums, 16), 16), 16); + + // Accumulation for 2 sub-blocks. + { + // 4x16 integer accumulators + vint16m1_t sumi_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + + // This might overflow. + // + // Recheck. + for (int i = 0; i < QK4_0; i++) { + // Load `b_ptr`. + const vuint8mf2_t b_0_packed = __riscv_vle8_v_u8mf2(&b_ptr[l].qs[j * 1024 + i * 16], 16); + const vint8mf2_t b_s_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(b_0_packed, 0xF, 16)); + const vint8mf2_t b_s_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(b_0_packed, 4, 16)); + + sumi_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_s_0_16, a_ptr[l].qs[j * 128 + i], b_s_0, 16); + sumi_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_s_1_16, a_ptr[l].qs[j * 128 + 32 + i], b_s_1, 16); + } + + sumi = __riscv_vwmacc_vv_i32m2(sumi, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 0)), + sumi_s_0_16, 16); + sumi = __riscv_vwmacc_vv_i32m2(sumi, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 1)), + sumi_s_1_16, 16); + } + { + // 4x16 integer accumulators + vint16m1_t sumi_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + + // This might overflow. + // + // Recheck. + for (int i = 0; i < QK4_0; i++) { + // Load `b_ptr`. + const vuint8mf2_t b_0_packed = __riscv_vle8_v_u8mf2(&b_ptr[l].qs[j * 1024 + 512 + i * 16], 16); + const vint8mf2_t b_s_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(b_0_packed, 0xF, 16)); + const vint8mf2_t b_s_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(b_0_packed, 4, 16)); + + sumi_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_s_0_16, a_ptr[l].qs[j * 128 + 64 + i], b_s_0, 16); + sumi_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_s_1_16, a_ptr[l].qs[j * 128 + 96 + i], b_s_1, 16); + } + + sumi = __riscv_vwmacc_vv_i32m2(sumi, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 2)), + sumi_s_0_16, 16); + sumi = __riscv_vwmacc_vv_i32m2(sumi, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 3)), + sumi_s_1_16, 16); + } + } + + const vfloat32m2_t b_d = __riscv_vfwcvt_f_f_v_f32m2(__riscv_vle16_v_f16m1((const _Float16 *)&b_ptr[l].d[0], 16), 16); + const vfloat32m2_t d_0 = __riscv_vfmul_vf_f32m2(b_d, a_ptr[l].d, 16); + + sumf = __riscv_vfmacc_vv_f32m2(sumf, __riscv_vfcvt_f_x_v_f32m2(sumi, 16), d_0, 16); + } + + __riscv_vse32_v_f32m2(s + x * 16, sumf, 16); + } + return; +#endif + ggml_gemv_q4_K_16x1_q8_K_generic(n, s, bs, vx, vy, nr, nc); +} + +void ggml_gemv_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined __riscv_v_intrinsic + const vint8mf2_t values = __riscv_vle8_v_i8mf2(kvalues_iq4nl, 16); + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx16 * b_ptr = (const block_iq4_nlx16 *) vx + (x * nb); + + // 1x16 Accumulator1 + vfloat32m2_t sumf = __riscv_vfmv_v_f_f32m2(0.0f, 16); + + for (int l = 0; l < nb; l++) { + // 1x16 integer accumulator + vint32m2_t sumi = __riscv_vmv_v_x_i32m2(0.0f, 16); + + // Accumulation loop. + for (int i = 0; i < QK4_NL / 2; i++) { + // Load `b_ptr`. + const vuint8mf2_t b_0_packed = __riscv_vle8_v_u8mf2((const uint8_t *)&b_ptr[l].qs[i * 16], 16); + const vint8mf2_t b_0_lo = __riscv_vrgather_vv_i8mf2(values, __riscv_vand_vx_u8mf2(b_0_packed, 0xf, 16), 16); + const vint8mf2_t b_0_hi = __riscv_vrgather_vv_i8mf2(values, __riscv_vsrl_vx_u8mf2(b_0_packed, 4, 16), 16); + // const vint16m1_t b_0_lo_16 = __riscv_vwcvt_x_x_v_i16m1(b_0_lo, 16); + // const vint16m1_t b_0_hi_16 = __riscv_vwcvt_x_x_v_i16m1(b_0_hi, 16); + + const vint16m1_t sumi_lo = __riscv_vwmul_vx_i16m1(b_0_lo, a_ptr[l].qs[i], 16); + const vint16m1_t sumi_hi = __riscv_vwmul_vx_i16m1(b_0_hi, a_ptr[l].qs[16 + i], 16); + sumi = __riscv_vadd_vv_i32m2(sumi, __riscv_vwadd_vv_i32m2(sumi_lo, sumi_hi, 16), 16); + } + + const vfloat16m1_t b_d = __riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16); + const vfloat32m2_t d_0 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d, 16); + + sumf = __riscv_vfmacc_vv_f32m2(sumf, __riscv_vfcvt_f_x_v_f32m2(sumi, 16), d_0, 16); + } + + __riscv_vse32_v_f32m2(s + x * 16, sumf, 16); + } + return; +#endif + ggml_gemv_iq4_nl_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + +void ggml_gemv_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + UNUSED(bs); + +#if defined __riscv_v_intrinsic + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q8_0x16 * b_ptr = (const block_q8_0x16 *) vx + (x * nb); + + // 1x16 Accumulator + vfloat32m2_t sumf = __riscv_vfmv_v_f_f32m2(0.0f, 16); + + for (int l = 0; l < nb; l++) { + // 1x16 Integer Accumulator + vint32m2_t sumi = __riscv_vmv_v_x_i32m2(0.0f, 16); + + // Accumulation loop. + for (int i = 0; i < QK8_0; i++) { + // Load `b_ptr`. + const vint8mf2_t b_0 = __riscv_vle8_v_i8mf2((const int8_t *)&b_ptr[l].qs[i * 16], 16); + // const vint16m1_t b_0_16 = __riscv_vwcvt_x_x_v_i16m1(b_0, 16); + + sumi = __riscv_vwadd_wv_i32m2(sumi, __riscv_vwmul_vx_i16m1(b_0, a_ptr[l].qs[i], 16), 16); + } + + const vfloat16m1_t b_d = __riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16); + const vfloat32m2_t d_0 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d, 16); + + sumf = __riscv_vfmacc_vv_f32m2(sumf, __riscv_vfcvt_f_x_v_f32m2(sumi, 16), d_0, 16); + } + + __riscv_vse32_v_f32m2(s + x * 16, sumf, 16); + } + return; +#endif + ggml_gemv_q8_0_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + +void ggml_gemv_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + assert(n % QK_K == 0); + assert(nr == 1); + assert(nc % 16 == 0); + + UNUSED(bs); + + const int N_COLS_TILE = 16; + const int num_k_blocks = n / QK_K; + + const size_t vl = __riscv_vsetvl_e32m2(N_COLS_TILE); + for (int col_tile = 0; col_tile < nc; col_tile += N_COLS_TILE) { + + const block_q8_K* lhs_base_ptr = (const block_q8_K*)vy; + const block_q2_Kx16* rhs_base_ptr = (const block_q2_Kx16*)vx + (col_tile / N_COLS_TILE) * num_k_blocks; + + vfloat32m2_t v_sumf = __riscv_vfmv_v_f_f32m2(0.0f, vl); + + for (int k_block = 0; k_block < num_k_blocks; ++k_block) { + const block_q8_K* lhs_current = &lhs_base_ptr[k_block]; + const block_q2_Kx16* rhs_current = &rhs_base_ptr[k_block]; + + // 1. Prepare Global Min Scales + vfloat16m1_t v_g_min_f16 = __riscv_vle16_v_f16m1((const _Float16*)rhs_current->dmin, vl); + vfloat32m2_t v_g_min_base = __riscv_vfwcvt_f_f_v_f32m2(v_g_min_f16, vl); + + vfloat32m2_t v_g_min_final = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d, vl); + + vint32m2_t v_isum = __riscv_vmv_v_x_i32m2(0, vl); + + const uint8_t* rhs_qs_ptr = rhs_current->qs; + const uint8_t* rhs_sc_ptr = rhs_current->scales; + const int8_t* lhs_qs_ptr = lhs_current->qs; + + // --- Phase Loop (4 phases x 64 elements) --- + for (int phase = 0; phase < 4; ++phase) { + + // A. Load Scales/Mins + vuint16m1_t v_d_sb_0, v_d_sb_1, v_d_sb_2, v_d_sb_3; + vuint16m1_t v_m_sb_0, v_m_sb_1, v_m_sb_2, v_m_sb_3; + + { + vuint8mf2_t v_raw; + // Sub-block 0 + v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 0, vl); + v_d_sb_0 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl); + v_m_sb_0 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl); + + // Sub-block 1 + v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 16, vl); + v_d_sb_1 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl); + v_m_sb_1 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl); + + // Sub-block 2 + v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 32, vl); + v_d_sb_2 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl); + v_m_sb_2 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl); + + // Sub-block 3 + v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 48, vl); + v_d_sb_3 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl); + v_m_sb_3 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl); + + rhs_sc_ptr += 64; + } + + int base_k_phase = (phase < 2) ? (phase * 16) : (128 + (phase-2)*16); + int k_offsets[4] = {0, 32, 64, 96}; + + // B. Inner Dot Product Loop + for (int l = 0; l < 16; ++l) { + vuint8mf2_t v_rhs_data = __riscv_vle8_v_u8mf2(rhs_qs_ptr, vl); + rhs_qs_ptr += 16; + + // Sub-block 0 + { + vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(v_rhs_data, 3, vl); + vint16m1_t v_w = __riscv_vmul_vv_i16m1( + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)), + __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_0), vl); + + int8_t q8 = lhs_qs_ptr[base_k_phase + k_offsets[0] + l]; + v_isum = __riscv_vwmacc_vx_i32m2(v_isum, (int16_t)q8, v_w, vl); + } + // Sub-block 1 + { + vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(v_rhs_data, 2, vl), 3, vl); + vint16m1_t v_w = __riscv_vmul_vv_i16m1( + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)), + __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_1), vl); + + int8_t q8 = lhs_qs_ptr[base_k_phase + k_offsets[1] + l]; + v_isum = __riscv_vwmacc_vx_i32m2(v_isum, (int16_t)q8, v_w, vl); + } + // Sub-block 2 + { + vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(v_rhs_data, 4, vl), 3, vl); + vint16m1_t v_w = __riscv_vmul_vv_i16m1( + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)), + __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_2), vl); + + int8_t q8 = lhs_qs_ptr[base_k_phase + k_offsets[2] + l]; + v_isum = __riscv_vwmacc_vx_i32m2(v_isum, (int16_t)q8, v_w, vl); + } + // Sub-block 3 + { + vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(v_rhs_data, 6, vl), 3, vl); + vint16m1_t v_w = __riscv_vmul_vv_i16m1( + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)), + __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_3), vl); + + int8_t q8 = lhs_qs_ptr[base_k_phase + k_offsets[3] + l]; + v_isum = __riscv_vwmacc_vx_i32m2(v_isum, (int16_t)q8, v_w, vl); + } + } + + // correction + int sb_base_abs = base_k_phase / 16; + + // Sub-block 0 + { + int sb_idx = sb_base_abs + (k_offsets[0] / 16); + int16_t bsum = lhs_current->bsums[sb_idx]; + vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_0); + vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, bsum, vl); + vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min_final, vl); + v_sumf = __riscv_vfsub_vv_f32m2(v_sumf, vf_c, vl); + } + // Sub-block 1 + { + int sb_idx = sb_base_abs + (k_offsets[1] / 16); + int16_t bsum = lhs_current->bsums[sb_idx]; + vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_1); + vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, bsum, vl); + vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min_final, vl); + v_sumf = __riscv_vfsub_vv_f32m2(v_sumf, vf_c, vl); + } + // Sub-block 2 + { + int sb_idx = sb_base_abs + (k_offsets[2] / 16); + int16_t bsum = lhs_current->bsums[sb_idx]; + vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_2); + vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, bsum, vl); + vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min_final, vl); + v_sumf = __riscv_vfsub_vv_f32m2(v_sumf, vf_c, vl); + } + // Sub-block 3 + { + int sb_idx = sb_base_abs + (k_offsets[3] / 16); + int16_t bsum = lhs_current->bsums[sb_idx]; + vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_3); + vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, bsum, vl); + vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min_final, vl); + v_sumf = __riscv_vfsub_vv_f32m2(v_sumf, vf_c, vl); + } + + } // End Phase Loop + + // Apply global Scales + vfloat16m1_t v_g_all_f16 = __riscv_vle16_v_f16m1((const _Float16*)rhs_current->d, vl); + vfloat32m2_t v_g_all_base = __riscv_vfwcvt_f_f_v_f32m2(v_g_all_f16, vl); + + vfloat32m2_t v_g_all_final = __riscv_vfmul_vf_f32m2(v_g_all_base, lhs_current->d, vl); + vfloat32m2_t v_sum = __riscv_vfcvt_f_x_v_f32m2(v_isum, vl); + v_sum = __riscv_vfmul_vv_f32m2(v_sum, v_g_all_final, vl); + v_sumf = __riscv_vfadd_vv_f32m2(v_sumf, v_sum, vl); + + } // End K-Block + __riscv_vse32_v_f32m2(s + col_tile, v_sumf, vl); + + } +} + void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; @@ -340,3 +906,827 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo #endif ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); } + +void ggml_gemm_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined __riscv_v_intrinsic + const vint8mf2_t values = __riscv_vle8_v_i8mf2(kvalues_iq4nl, 16); + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x16 * b_ptr = (const block_q4_0x16 *) vx + (x * nb); + + // 4x16 Accumulators + vfloat32m2_t sumf_0 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + vfloat32m2_t sumf_1 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + vfloat32m2_t sumf_2 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + vfloat32m2_t sumf_3 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + + for (int l = 0; l < nb; l++) { + // 4x16 integer accumulators + vint16m1_t sumi_0_lo_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_1_lo_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_2_lo_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_3_lo_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_0_hi_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_1_hi_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_2_hi_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_3_hi_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + + // Accumulation loop. + for (int i = 0; i < QK4_0 / 2; i++) { + // Load `b_ptr`. + const vint8mf2_t b_0_packed = __riscv_vle8_v_i8mf2((const int8_t *)&b_ptr[l].qs[i * 16], 16); + const vint8mf2_t b_0_lo = __riscv_vsra_vx_i8mf2(__riscv_vsll_vx_i8mf2(b_0_packed, 4, 16), 4, 16); + const vint8mf2_t b_0_hi = __riscv_vsra_vx_i8mf2(b_0_packed, 4, 16); + + sumi_0_lo_16 = __riscv_vwmacc_vx_i16m1(sumi_0_lo_16, a_ptr[l].qs[i * 4], b_0_lo, 16); + sumi_1_lo_16 = __riscv_vwmacc_vx_i16m1(sumi_1_lo_16, a_ptr[l].qs[i * 4 + 1], b_0_lo, 16); + sumi_2_lo_16 = __riscv_vwmacc_vx_i16m1(sumi_2_lo_16, a_ptr[l].qs[i * 4 + 2], b_0_lo, 16); + sumi_3_lo_16 = __riscv_vwmacc_vx_i16m1(sumi_3_lo_16, a_ptr[l].qs[i * 4 + 3], b_0_lo, 16); + + sumi_0_hi_16 = __riscv_vwmacc_vx_i16m1(sumi_0_hi_16, a_ptr[l].qs[64 + i * 4], b_0_hi, 16); + sumi_1_hi_16 = __riscv_vwmacc_vx_i16m1(sumi_1_hi_16, a_ptr[l].qs[64 + i * 4 + 1], b_0_hi, 16); + sumi_2_hi_16 = __riscv_vwmacc_vx_i16m1(sumi_2_hi_16, a_ptr[l].qs[64 + i * 4 + 2], b_0_hi, 16); + sumi_3_hi_16 = __riscv_vwmacc_vx_i16m1(sumi_3_hi_16, a_ptr[l].qs[64 + i * 4 + 3], b_0_hi, 16); + } + + // Do the final accumulation in i32 to prevent overflow. + const vint32m2_t sumi_0 = __riscv_vwadd_vv_i32m2(sumi_0_lo_16, sumi_0_hi_16, 16); + const vint32m2_t sumi_1 = __riscv_vwadd_vv_i32m2(sumi_1_lo_16, sumi_1_hi_16, 16); + const vint32m2_t sumi_2 = __riscv_vwadd_vv_i32m2(sumi_2_lo_16, sumi_2_hi_16, 16); + const vint32m2_t sumi_3 = __riscv_vwadd_vv_i32m2(sumi_3_lo_16, sumi_3_hi_16, 16); + + const vfloat16m1_t b_d = __riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16); + const vfloat32m2_t d_0 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[0], 16); + const vfloat32m2_t d_1 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[1], 16); + const vfloat32m2_t d_2 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[2], 16); + const vfloat32m2_t d_3 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[3], 16); + + sumf_0 = __riscv_vfmacc_vv_f32m2(sumf_0, __riscv_vfcvt_f_x_v_f32m2(sumi_0, 16), d_0, 16); + sumf_1 = __riscv_vfmacc_vv_f32m2(sumf_1, __riscv_vfcvt_f_x_v_f32m2(sumi_1, 16), d_1, 16); + sumf_2 = __riscv_vfmacc_vv_f32m2(sumf_2, __riscv_vfcvt_f_x_v_f32m2(sumi_2, 16), d_2, 16); + sumf_3 = __riscv_vfmacc_vv_f32m2(sumf_3, __riscv_vfcvt_f_x_v_f32m2(sumi_3, 16), d_3, 16); + } + + __riscv_vse32_v_f32m2(s + (y * 4 + 0) * bs + x * 16, sumf_0, 16); + __riscv_vse32_v_f32m2(s + (y * 4 + 1) * bs + x * 16, sumf_1, 16); + __riscv_vse32_v_f32m2(s + (y * 4 + 2) * bs + x * 16, sumf_2, 16); + __riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16); + } + } + return; +#endif + ggml_gemm_q4_0_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + +void ggml_gemm_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK_K; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined __riscv_v_intrinsic + for (int y = 0; y < nr / 4; y++) { + const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_Kx16 * b_ptr = (const block_q4_Kx16 *) vx + (x * nb); + + // 4x16 Accumulators + vfloat32m2_t sumf_0 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + vfloat32m2_t sumf_1 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + vfloat32m2_t sumf_2 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + vfloat32m2_t sumf_3 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + + for (int l = 0; l < nb; l++) { + vint32m2_t sumi_0 = __riscv_vmv_v_x_i32m2(0, 16); + vint32m2_t sumi_1 = __riscv_vmv_v_x_i32m2(0, 16); + vint32m2_t sumi_2 = __riscv_vmv_v_x_i32m2(0, 16); + vint32m2_t sumi_3 = __riscv_vmv_v_x_i32m2(0, 16); + + // We process 4 sub-blocks at once. + for (int j = 0; j < QK_K / 128; j++) { + // Extract the scales and the mins. + // + // Low bits. + vuint8m2_t scales_mins_lo = __riscv_vle8_v_u8m2(&b_ptr[l].scales[j * 64], 64); + vuint8m2_t scales_lo = __riscv_vand_vx_u8m2(scales_mins_lo, 0x0F, 64); + vuint8m2_t mins_lo = __riscv_vsrl_vx_u8m2(scales_mins_lo, 4, 64); + + // High bits. + vuint8m2_t scales_mins_hi = __riscv_vle8_v_u8m2(&b_ptr[l].scales[128], 64); + vuint8m2_t scales_hi; + vuint8m2_t mins_hi; + if (!j) { + scales_hi = __riscv_vsll_vx_u8m2(__riscv_vand_vx_u8m2(scales_mins_hi, 0x03, 64), 4, 64); + mins_hi = __riscv_vsll_vx_u8m2(__riscv_vand_vx_u8m2(scales_mins_hi, 0x0C, 64), 2, 64); + } else { + scales_hi = __riscv_vand_vx_u8m2(scales_mins_hi, 0x30, 64); + mins_hi = __riscv_vsrl_vx_u8m2(__riscv_vand_vx_u8m2(scales_mins_hi, 0xC0, 64), 2, 64); + } + vuint16m4_t scales = __riscv_vzext_vf2_u16m4(__riscv_vor_vv_u8m2(scales_hi, scales_lo, 64), 64); + vint16m4_t mins = __riscv_vreinterpret_v_u16m4_i16m4(__riscv_vzext_vf2_u16m4(__riscv_vor_vv_u8m2(mins_hi, mins_lo, 64), 64)); + + // Reduce the mins and multiply with `dmin`. + // + // Correct in `sumf`. + vint32m2_t bsums_0 = __riscv_vmv_v_x_i32m2(0, 16); + vint32m2_t bsums_1 = __riscv_vmv_v_x_i32m2(0, 16); + vint32m2_t bsums_2 = __riscv_vmv_v_x_i32m2(0, 16); + vint32m2_t bsums_3 = __riscv_vmv_v_x_i32m2(0, 16); + + bsums_0 = __riscv_vwmacc_vx_i32m2(bsums_0, + a_ptr[l].bsums[j * 32] + a_ptr[l].bsums[j * 32 + 4], + __riscv_vget_v_i16m4_i16m1(mins, 0), 16); + bsums_1 = __riscv_vwmacc_vx_i32m2(bsums_1, + a_ptr[l].bsums[j * 32 + 1] + a_ptr[l].bsums[j * 32 + 5], + __riscv_vget_v_i16m4_i16m1(mins, 0), 16); + bsums_2 = __riscv_vwmacc_vx_i32m2(bsums_2, + a_ptr[l].bsums[j * 32 + 2] + a_ptr[l].bsums[j * 32 + 6], + __riscv_vget_v_i16m4_i16m1(mins, 0), 16); + bsums_3 = __riscv_vwmacc_vx_i32m2(bsums_3, + a_ptr[l].bsums[j * 32 + 3] + a_ptr[l].bsums[j * 32 + 7], + __riscv_vget_v_i16m4_i16m1(mins, 0), 16); + bsums_0 = __riscv_vwmacc_vx_i32m2(bsums_0, + a_ptr[l].bsums[j * 32 + 8] + a_ptr[l].bsums[j * 32 + 8 + 4], + __riscv_vget_v_i16m4_i16m1(mins, 1), 16); + bsums_1 = __riscv_vwmacc_vx_i32m2(bsums_1, + a_ptr[l].bsums[j * 32 + 8 + 1] + a_ptr[l].bsums[j * 32 + 8 + 5], + __riscv_vget_v_i16m4_i16m1(mins, 1), 16); + bsums_2 = __riscv_vwmacc_vx_i32m2(bsums_2, + a_ptr[l].bsums[j * 32 + 8 + 2] + a_ptr[l].bsums[j * 32 + 8 + 6], + __riscv_vget_v_i16m4_i16m1(mins, 1), 16); + bsums_3 = __riscv_vwmacc_vx_i32m2(bsums_3, + a_ptr[l].bsums[j * 32 + 8 + 3] + a_ptr[l].bsums[j * 32 + 8 + 7], + __riscv_vget_v_i16m4_i16m1(mins, 1), 16); + bsums_0 = __riscv_vwmacc_vx_i32m2(bsums_0, + a_ptr[l].bsums[j * 32 + 16] + a_ptr[l].bsums[j * 32 + 16 + 4], + __riscv_vget_v_i16m4_i16m1(mins, 2), 16); + bsums_1 = __riscv_vwmacc_vx_i32m2(bsums_1, + a_ptr[l].bsums[j * 32 + 16 + 1] + a_ptr[l].bsums[j * 32 + 16 + 5], + __riscv_vget_v_i16m4_i16m1(mins, 2), 16); + bsums_2 = __riscv_vwmacc_vx_i32m2(bsums_2, + a_ptr[l].bsums[j * 32 + 16 + 2] + a_ptr[l].bsums[j * 32 + 16 + 6], + __riscv_vget_v_i16m4_i16m1(mins, 2), 16); + bsums_3 = __riscv_vwmacc_vx_i32m2(bsums_3, + a_ptr[l].bsums[j * 32 + 16 + 3] + a_ptr[l].bsums[j * 32 + 16 + 7], + __riscv_vget_v_i16m4_i16m1(mins, 2), 16); + bsums_0 = __riscv_vwmacc_vx_i32m2(bsums_0, + a_ptr[l].bsums[j * 32 + 24 + 0] + a_ptr[l].bsums[j * 32 + 24 + 4], + __riscv_vget_v_i16m4_i16m1(mins, 3), 16); + bsums_1 = __riscv_vwmacc_vx_i32m2(bsums_1, + a_ptr[l].bsums[j * 32 + 24 + 1] + a_ptr[l].bsums[j * 32 + 24 + 5], + __riscv_vget_v_i16m4_i16m1(mins, 3), 16); + bsums_2 = __riscv_vwmacc_vx_i32m2(bsums_2, + a_ptr[l].bsums[j * 32 + 24 + 2] + a_ptr[l].bsums[j * 32 + 24 + 6], + __riscv_vget_v_i16m4_i16m1(mins, 3), 16); + bsums_3 = __riscv_vwmacc_vx_i32m2(bsums_3, + a_ptr[l].bsums[j * 32 + 24 + 3] + a_ptr[l].bsums[j * 32 + 24 + 7], + __riscv_vget_v_i16m4_i16m1(mins, 3), 16); + + const vfloat32m2_t dmins_d_0 = __riscv_vfmul_vf_f32m2( + __riscv_vfwcvt_f_f_v_f32m2(__riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].dmin, 16), 16), a_ptr[l].d[0], 16); + const vfloat32m2_t dmins_d_1 = __riscv_vfmul_vf_f32m2( + __riscv_vfwcvt_f_f_v_f32m2(__riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].dmin, 16), 16), a_ptr[l].d[1], 16); + const vfloat32m2_t dmins_d_2 = __riscv_vfmul_vf_f32m2( + __riscv_vfwcvt_f_f_v_f32m2(__riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].dmin, 16), 16), a_ptr[l].d[2], 16); + const vfloat32m2_t dmins_d_3 = __riscv_vfmul_vf_f32m2( + __riscv_vfwcvt_f_f_v_f32m2(__riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].dmin, 16), 16), a_ptr[l].d[3], 16); + + sumf_0 = __riscv_vfsub_vv_f32m2(sumf_0, __riscv_vfmul_vv_f32m2(dmins_d_0, __riscv_vfcvt_f_x_v_f32m2(bsums_0, 16), 16), 16); + sumf_1 = __riscv_vfsub_vv_f32m2(sumf_1, __riscv_vfmul_vv_f32m2(dmins_d_1, __riscv_vfcvt_f_x_v_f32m2(bsums_1, 16), 16), 16); + sumf_2 = __riscv_vfsub_vv_f32m2(sumf_2, __riscv_vfmul_vv_f32m2(dmins_d_2, __riscv_vfcvt_f_x_v_f32m2(bsums_2, 16), 16), 16); + sumf_3 = __riscv_vfsub_vv_f32m2(sumf_3, __riscv_vfmul_vv_f32m2(dmins_d_3, __riscv_vfcvt_f_x_v_f32m2(bsums_3, 16), 16), 16); + + + // Accumulation for 2 sub-blocks. + { + // 4x8 integer accumulators + vint16m1_t sumi_0_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_1_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_2_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_3_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_0_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_1_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_2_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_3_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + + // This might overflow. + // + // Recheck. + for (int i = 0; i < QK4_0; i++) { + // Load `b_ptr`. + const vuint8mf2_t b_0_packed = __riscv_vle8_v_u8mf2(&b_ptr[l].qs[j * 1024 + i * 16], 16); + const vint8mf2_t b_s_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(b_0_packed, 0xF, 16)); + const vint8mf2_t b_s_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(b_0_packed, 4, 16)); + + sumi_0_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_0_s_0_16, a_ptr[l].qs[j * 512 + i * 4], b_s_0, 16); + sumi_1_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_1_s_0_16, a_ptr[l].qs[j * 512 + i * 4 + 1], b_s_0, 16); + sumi_2_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_2_s_0_16, a_ptr[l].qs[j * 512 + i * 4 + 2], b_s_0, 16); + sumi_3_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_3_s_0_16, a_ptr[l].qs[j * 512 + i * 4 + 3], b_s_0, 16); + + sumi_0_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_0_s_1_16, a_ptr[l].qs[j * 512 + 128 + i * 4], b_s_1, 16); + sumi_1_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_1_s_1_16, a_ptr[l].qs[j * 512 + 128 + i * 4 + 1], b_s_1, 16); + sumi_2_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_2_s_1_16, a_ptr[l].qs[j * 512 + 128 + i * 4 + 2], b_s_1, 16); + sumi_3_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_3_s_1_16, a_ptr[l].qs[j * 512 + 128 + i * 4 + 3], b_s_1, 16); + } + + sumi_0 = __riscv_vwmacc_vv_i32m2(sumi_0, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 0)), + sumi_0_s_0_16, 16); + sumi_0 = __riscv_vwmacc_vv_i32m2(sumi_0, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 1)), + sumi_0_s_1_16, 16); + sumi_1 = __riscv_vwmacc_vv_i32m2(sumi_1, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 0)), + sumi_1_s_0_16, 16); + sumi_1 = __riscv_vwmacc_vv_i32m2(sumi_1, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 1)), + sumi_1_s_1_16, 16); + sumi_2 = __riscv_vwmacc_vv_i32m2(sumi_2, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 0)), + sumi_2_s_0_16, 16); + sumi_2 = __riscv_vwmacc_vv_i32m2(sumi_2, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 1)), + sumi_2_s_1_16, 16); + sumi_3 = __riscv_vwmacc_vv_i32m2(sumi_3, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 0)), + sumi_3_s_0_16, 16); + sumi_3 = __riscv_vwmacc_vv_i32m2(sumi_3, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 1)), + sumi_3_s_1_16, 16); + } + { + // 4x8 integer accumulators + vint16m1_t sumi_0_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_1_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_2_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_3_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_0_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_1_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_2_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + vint16m1_t sumi_3_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16); + + // This might overflow. + // + // Recheck. + for (int i = 0; i < QK4_0; i++) { + // Load `b_ptr`. + const vuint8mf2_t b_0_packed = __riscv_vle8_v_u8mf2(&b_ptr[l].qs[j * 1024 + 512 + i * 16], 16); + const vint8mf2_t b_s_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(b_0_packed, 0xF, 16)); + const vint8mf2_t b_s_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(b_0_packed, 4, 16)); + + sumi_0_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_0_s_0_16, a_ptr[l].qs[j * 512 + 256 + i * 4], b_s_0, 16); + sumi_1_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_1_s_0_16, a_ptr[l].qs[j * 512 + 256 + i * 4 + 1], b_s_0, 16); + sumi_2_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_2_s_0_16, a_ptr[l].qs[j * 512 + 256 + i * 4 + 2], b_s_0, 16); + sumi_3_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_3_s_0_16, a_ptr[l].qs[j * 512 + 256 + i * 4 + 3], b_s_0, 16); + + sumi_0_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_0_s_1_16, a_ptr[l].qs[j * 512 + 384 + i * 4], b_s_1, 16); + sumi_1_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_1_s_1_16, a_ptr[l].qs[j * 512 + 384 + i * 4 + 1], b_s_1, 16); + sumi_2_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_2_s_1_16, a_ptr[l].qs[j * 512 + 384 + i * 4 + 2], b_s_1, 16); + sumi_3_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_3_s_1_16, a_ptr[l].qs[j * 512 + 384 + i * 4 + 3], b_s_1, 16); + } + + sumi_0 = __riscv_vwmacc_vv_i32m2(sumi_0, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 2)), + sumi_0_s_0_16, 16); + sumi_0 = __riscv_vwmacc_vv_i32m2(sumi_0, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 3)), + sumi_0_s_1_16, 16); + sumi_1 = __riscv_vwmacc_vv_i32m2(sumi_1, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 2)), + sumi_1_s_0_16, 16); + sumi_1 = __riscv_vwmacc_vv_i32m2(sumi_1, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 3)), + sumi_1_s_1_16, 16); + sumi_2 = __riscv_vwmacc_vv_i32m2(sumi_2, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 2)), + sumi_2_s_0_16, 16); + sumi_2 = __riscv_vwmacc_vv_i32m2(sumi_2, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 3)), + sumi_2_s_1_16, 16); + sumi_3 = __riscv_vwmacc_vv_i32m2(sumi_3, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 2)), + sumi_3_s_0_16, 16); + sumi_3 = __riscv_vwmacc_vv_i32m2(sumi_3, + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 3)), + sumi_3_s_1_16, 16); + } + } + + const vfloat32m2_t b_d = __riscv_vfwcvt_f_f_v_f32m2(__riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16), 16); + const vfloat32m2_t d_0 = __riscv_vfmul_vf_f32m2(b_d, a_ptr[l].d[0], 16); + const vfloat32m2_t d_1 = __riscv_vfmul_vf_f32m2(b_d, a_ptr[l].d[1], 16); + const vfloat32m2_t d_2 = __riscv_vfmul_vf_f32m2(b_d, a_ptr[l].d[2], 16); + const vfloat32m2_t d_3 = __riscv_vfmul_vf_f32m2(b_d, a_ptr[l].d[3], 16); + + sumf_0 = __riscv_vfmacc_vv_f32m2(sumf_0, __riscv_vfcvt_f_x_v_f32m2(sumi_0, 16), d_0, 16); + sumf_1 = __riscv_vfmacc_vv_f32m2(sumf_1, __riscv_vfcvt_f_x_v_f32m2(sumi_1, 16), d_1, 16); + sumf_2 = __riscv_vfmacc_vv_f32m2(sumf_2, __riscv_vfcvt_f_x_v_f32m2(sumi_2, 16), d_2, 16); + sumf_3 = __riscv_vfmacc_vv_f32m2(sumf_3, __riscv_vfcvt_f_x_v_f32m2(sumi_3, 16), d_3, 16); + } + + __riscv_vse32_v_f32m2(s + (y * 4 + 0) * bs + x * 16, sumf_0, 16); + __riscv_vse32_v_f32m2(s + (y * 4 + 1) * bs + x * 16, sumf_1, 16); + __riscv_vse32_v_f32m2(s + (y * 4 + 2) * bs + x * 16, sumf_2, 16); + __riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16); + } + } + return; +#endif + ggml_gemm_q4_K_16x1_q8_K_generic(n, s, bs, vx, vy, nr, nc); +} + +void ggml_gemm_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined __riscv_v_intrinsic + const vint8mf2_t values = __riscv_vle8_v_i8mf2(kvalues_iq4nl, 16); + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx16 * b_ptr = (const block_iq4_nlx16 *) vx + (x * nb); + + // 4x16 Accumulators + vfloat32m2_t sumf_0 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + vfloat32m2_t sumf_1 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + vfloat32m2_t sumf_2 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + vfloat32m2_t sumf_3 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + + for (int l = 0; l < nb; l++) { + // 4x16 integer accumulators + vint32m2_t sumi_0 = __riscv_vmv_v_x_i32m2(0.0f, 16); + vint32m2_t sumi_1 = __riscv_vmv_v_x_i32m2(0.0f, 16); + vint32m2_t sumi_2 = __riscv_vmv_v_x_i32m2(0.0f, 16); + vint32m2_t sumi_3 = __riscv_vmv_v_x_i32m2(0.0f, 16); + + // Accumulation loop. + for (int i = 0; i < QK4_NL / 2; i++) { + // Load `b_ptr`. + const vuint8mf2_t b_0_packed = __riscv_vle8_v_u8mf2((const uint8_t *)&b_ptr[l].qs[i * 16], 16); + const vint8mf2_t b_0_lo = __riscv_vrgather_vv_i8mf2(values, __riscv_vand_vx_u8mf2(b_0_packed, 0xf, 16), 16); + const vint8mf2_t b_0_hi = __riscv_vrgather_vv_i8mf2(values, __riscv_vsrl_vx_u8mf2(b_0_packed, 4, 16), 16); + // const vint16m1_t b_0_lo_16 = __riscv_vwcvt_x_x_v_i16m1(b_0_lo, 16); + // const vint16m1_t b_0_hi_16 = __riscv_vwcvt_x_x_v_i16m1(b_0_hi, 16); + + const vint16m1_t sumi_0_lo = __riscv_vwmul_vx_i16m1(b_0_lo, a_ptr[l].qs[i * 4], 16); + const vint16m1_t sumi_1_lo = __riscv_vwmul_vx_i16m1(b_0_lo, a_ptr[l].qs[i * 4 + 1], 16); + const vint16m1_t sumi_2_lo = __riscv_vwmul_vx_i16m1(b_0_lo, a_ptr[l].qs[i * 4 + 2], 16); + const vint16m1_t sumi_3_lo = __riscv_vwmul_vx_i16m1(b_0_lo, a_ptr[l].qs[i * 4 + 3], 16); + + const vint16m1_t sumi_0_hi = __riscv_vwmul_vx_i16m1(b_0_hi, a_ptr[l].qs[64 + i * 4], 16); + const vint16m1_t sumi_1_hi = __riscv_vwmul_vx_i16m1(b_0_hi, a_ptr[l].qs[64 + i * 4 + 1], 16); + const vint16m1_t sumi_2_hi = __riscv_vwmul_vx_i16m1(b_0_hi, a_ptr[l].qs[64 + i * 4 + 2], 16); + const vint16m1_t sumi_3_hi = __riscv_vwmul_vx_i16m1(b_0_hi, a_ptr[l].qs[64 + i * 4 + 3], 16); + + sumi_0 = __riscv_vadd_vv_i32m2(sumi_0, __riscv_vwadd_vv_i32m2(sumi_0_lo, sumi_0_hi, 16), 16); + sumi_1 = __riscv_vadd_vv_i32m2(sumi_1, __riscv_vwadd_vv_i32m2(sumi_1_lo, sumi_1_hi, 16), 16); + sumi_2 = __riscv_vadd_vv_i32m2(sumi_2, __riscv_vwadd_vv_i32m2(sumi_2_lo, sumi_2_hi, 16), 16); + sumi_3 = __riscv_vadd_vv_i32m2(sumi_3, __riscv_vwadd_vv_i32m2(sumi_3_lo, sumi_3_hi, 16), 16); + } + + const vfloat16m1_t b_d = __riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16); + const vfloat32m2_t d_0 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[0], 16); + const vfloat32m2_t d_1 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[1], 16); + const vfloat32m2_t d_2 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[2], 16); + const vfloat32m2_t d_3 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[3], 16); + + sumf_0 = __riscv_vfmacc_vv_f32m2(sumf_0, __riscv_vfcvt_f_x_v_f32m2(sumi_0, 16), d_0, 16); + sumf_1 = __riscv_vfmacc_vv_f32m2(sumf_1, __riscv_vfcvt_f_x_v_f32m2(sumi_1, 16), d_1, 16); + sumf_2 = __riscv_vfmacc_vv_f32m2(sumf_2, __riscv_vfcvt_f_x_v_f32m2(sumi_2, 16), d_2, 16); + sumf_3 = __riscv_vfmacc_vv_f32m2(sumf_3, __riscv_vfcvt_f_x_v_f32m2(sumi_3, 16), d_3, 16); + } + + __riscv_vse32_v_f32m2(s + (y * 4 + 0) * bs + x * 16, sumf_0, 16); + __riscv_vse32_v_f32m2(s + (y * 4 + 1) * bs + x * 16, sumf_1, 16); + __riscv_vse32_v_f32m2(s + (y * 4 + 2) * bs + x * 16, sumf_2, 16); + __riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16); + } + } + return; +#endif + ggml_gemm_iq4_nl_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + +void ggml_gemm_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined __riscv_v_intrinsic + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q8_0x16 * b_ptr = (const block_q8_0x16 *) vx + (x * nb); + + // 4x16 Accumulators + vfloat32m2_t sumf_0 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + vfloat32m2_t sumf_1 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + vfloat32m2_t sumf_2 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + vfloat32m2_t sumf_3 = __riscv_vfmv_v_f_f32m2(0.0f, 16); + + for (int l = 0; l < nb; l++) { + // 4x16 Integer Accumulators + vint32m2_t sumi_0 = __riscv_vmv_v_x_i32m2(0.0f, 16); + vint32m2_t sumi_1 = __riscv_vmv_v_x_i32m2(0.0f, 16); + vint32m2_t sumi_2 = __riscv_vmv_v_x_i32m2(0.0f, 16); + vint32m2_t sumi_3 = __riscv_vmv_v_x_i32m2(0.0f, 16); + + // Accumulation loop. + for (int i = 0; i < QK8_0; i++) { + // Load `b_ptr`. + const vint8mf2_t b_0 = __riscv_vle8_v_i8mf2((const int8_t *)&b_ptr[l].qs[i * 16], 16); + // const vint16m1_t b_0_16 = __riscv_vwcvt_x_x_v_i16m1(b_0, 16); + + sumi_0 = __riscv_vwadd_wv_i32m2(sumi_0, __riscv_vwmul_vx_i16m1(b_0, a_ptr[l].qs[i * 4 + 0], 16), 16); + sumi_1 = __riscv_vwadd_wv_i32m2(sumi_1, __riscv_vwmul_vx_i16m1(b_0, a_ptr[l].qs[i * 4 + 1], 16), 16); + sumi_2 = __riscv_vwadd_wv_i32m2(sumi_2, __riscv_vwmul_vx_i16m1(b_0, a_ptr[l].qs[i * 4 + 2], 16), 16); + sumi_3 = __riscv_vwadd_wv_i32m2(sumi_3, __riscv_vwmul_vx_i16m1(b_0, a_ptr[l].qs[i * 4 + 3], 16), 16); + } + + const vfloat16m1_t b_d = __riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16); + const vfloat32m2_t d_0 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[0], 16); + const vfloat32m2_t d_1 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[1], 16); + const vfloat32m2_t d_2 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[2], 16); + const vfloat32m2_t d_3 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[3], 16); + + sumf_0 = __riscv_vfmacc_vv_f32m2(sumf_0, __riscv_vfcvt_f_x_v_f32m2(sumi_0, 16), d_0, 16); + sumf_1 = __riscv_vfmacc_vv_f32m2(sumf_1, __riscv_vfcvt_f_x_v_f32m2(sumi_1, 16), d_1, 16); + sumf_2 = __riscv_vfmacc_vv_f32m2(sumf_2, __riscv_vfcvt_f_x_v_f32m2(sumi_2, 16), d_2, 16); + sumf_3 = __riscv_vfmacc_vv_f32m2(sumf_3, __riscv_vfcvt_f_x_v_f32m2(sumi_3, 16), d_3, 16); + } + + __riscv_vse32_v_f32m2(s + (y * 4 + 0) * bs + x * 16, sumf_0, 16); + __riscv_vse32_v_f32m2(s + (y * 4 + 1) * bs + x * 16, sumf_1, 16); + __riscv_vse32_v_f32m2(s + (y * 4 + 2) * bs + x * 16, sumf_2, 16); + __riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16); + } + } + return; +#endif + ggml_gemm_q8_0_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + +void ggml_gemm_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + assert(n % QK_K == 0); + const int num_k_blocks = n / QK_K; + const int N_ROWS_TILE = 4; + const int N_COLS_TILE = 16; + assert(nr % N_ROWS_TILE == 0); + assert(nc % N_COLS_TILE == 0); + + const size_t vl = __riscv_vsetvl_e32m2(N_COLS_TILE); + // --- Tiling Loops --- +#pragma GCC unroll 1 + for (int row_tile = 0; row_tile < nr; row_tile += N_ROWS_TILE) { +#pragma GCC unroll 1 + for (int col_tile = 0; col_tile < nc; col_tile += N_COLS_TILE) { + // Base Pointers + const block_q8_Kx4* lhs_base_ptr = (const block_q8_Kx4*)vy + (row_tile / N_ROWS_TILE) * num_k_blocks; + const block_q2_Kx16* rhs_base_ptr = (const block_q2_Kx16*)vx + (col_tile / N_COLS_TILE) * num_k_blocks; + + // Persistent Float Accumulators + vfloat32m2_t v_sumf_0 = __riscv_vfmv_v_f_f32m2(0.0f, vl); + vfloat32m2_t v_sumf_1 = __riscv_vfmv_v_f_f32m2(0.0f, vl); + vfloat32m2_t v_sumf_2 = __riscv_vfmv_v_f_f32m2(0.0f, vl); + vfloat32m2_t v_sumf_3 = __riscv_vfmv_v_f_f32m2(0.0f, vl); + + // --- Super-Block Loop (K=0..255) --- +#pragma GCC unroll 1 + for (int k_block = 0; k_block < num_k_blocks; ++k_block) { + const block_q8_Kx4* lhs_current = &lhs_base_ptr[k_block]; + const block_q2_Kx16* rhs_current = &rhs_base_ptr[k_block]; + + // 1. Load Global Min Scales (Keep as F16/LMUL=1 to save registers) + vfloat16m1_t v_g_min_f16 = __riscv_vle16_v_f16m1((const _Float16*)rhs_current->dmin, vl); + vfloat32m2_t v_g_min_base = __riscv_vfwcvt_f_f_v_f32m2(v_g_min_f16, vl); + + // 2. Initialize Integer Accumulators + vint32m2_t v_isum_0 = __riscv_vmv_v_x_i32m2(0, vl); + vint32m2_t v_isum_1 = __riscv_vmv_v_x_i32m2(0, vl); + vint32m2_t v_isum_2 = __riscv_vmv_v_x_i32m2(0, vl); + vint32m2_t v_isum_3 = __riscv_vmv_v_x_i32m2(0, vl); + + const uint8_t* rhs_qs_ptr = rhs_current->qs; + const uint8_t* rhs_sc_ptr = rhs_current->scales; + const int8_t* lhs_qs_ptr = lhs_current->qs; + + // --- Phase Loop (4 phases x 64 elements) --- +#pragma GCC unroll 1 + for (int phase = 0; phase < 4; ++phase) { + + // A. Load Scales/Mins for the 4 interleaved sub-blocks + vuint16m1_t v_d_sb_0, v_d_sb_1, v_d_sb_2, v_d_sb_3; + vuint16m1_t v_m_sb_0, v_m_sb_1, v_m_sb_2, v_m_sb_3; + + // Unrolled Load Logic + { + vuint8mf2_t v_raw; + // Sub-block 0 + v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 0, vl); + v_d_sb_0 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl); + v_m_sb_0 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl); + + // Sub-block 1 + v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 16, vl); + v_d_sb_1 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl); + v_m_sb_1 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl); + + // Sub-block 2 + v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 32, vl); + v_d_sb_2 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl); + v_m_sb_2 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl); + + // Sub-block 3 + v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 48, vl); + v_d_sb_3 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl); + v_m_sb_3 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl); + + rhs_sc_ptr += 64; + } + + int base_k_phase = (phase < 2) ? (phase * 16) : (128 + (phase-2)*16); + int k_offsets[4] = {0, 32, 64, 96}; + + // B. Inner Dot Product Loop +#pragma GCC unroll 1 + for (int l = 0; l < 16; ++l) { + vuint8mf2_t v_rhs_data = __riscv_vle8_v_u8mf2(rhs_qs_ptr, vl); + rhs_qs_ptr += 16; + + // Unroll over 4 sub-blocks (0, 1, 2, 3 relative to phase) + + // --- Sub-block 0 --- + { + vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(v_rhs_data, 3, vl); + vint16m1_t v_w = __riscv_vmul_vv_i16m1( + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)), + __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_0), vl); + + const int8_t* q8 = &lhs_qs_ptr[(base_k_phase + k_offsets[0] + l) * 4]; + v_isum_0 = __riscv_vwmacc_vx_i32m2(v_isum_0, (int16_t)q8[0], v_w, vl); + v_isum_1 = __riscv_vwmacc_vx_i32m2(v_isum_1, (int16_t)q8[1], v_w, vl); + v_isum_2 = __riscv_vwmacc_vx_i32m2(v_isum_2, (int16_t)q8[2], v_w, vl); + v_isum_3 = __riscv_vwmacc_vx_i32m2(v_isum_3, (int16_t)q8[3], v_w, vl); + } + // --- Sub-block 1 --- + { + vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(v_rhs_data, 2, vl), 3, vl); + vint16m1_t v_w = __riscv_vmul_vv_i16m1( + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)), + __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_1), vl); + + const int8_t* q8 = &lhs_qs_ptr[(base_k_phase + k_offsets[1] + l) * 4]; + v_isum_0 = __riscv_vwmacc_vx_i32m2(v_isum_0, (int16_t)q8[0], v_w, vl); + v_isum_1 = __riscv_vwmacc_vx_i32m2(v_isum_1, (int16_t)q8[1], v_w, vl); + v_isum_2 = __riscv_vwmacc_vx_i32m2(v_isum_2, (int16_t)q8[2], v_w, vl); + v_isum_3 = __riscv_vwmacc_vx_i32m2(v_isum_3, (int16_t)q8[3], v_w, vl); + } + // --- Sub-block 2 --- + { + vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(v_rhs_data, 4, vl), 3, vl); + vint16m1_t v_w = __riscv_vmul_vv_i16m1( + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)), + __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_2), vl); + + const int8_t* q8 = &lhs_qs_ptr[(base_k_phase + k_offsets[2] + l) * 4]; + v_isum_0 = __riscv_vwmacc_vx_i32m2(v_isum_0, (int16_t)q8[0], v_w, vl); + v_isum_1 = __riscv_vwmacc_vx_i32m2(v_isum_1, (int16_t)q8[1], v_w, vl); + v_isum_2 = __riscv_vwmacc_vx_i32m2(v_isum_2, (int16_t)q8[2], v_w, vl); + v_isum_3 = __riscv_vwmacc_vx_i32m2(v_isum_3, (int16_t)q8[3], v_w, vl); + } + // --- Sub-block 3 --- + { + vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(v_rhs_data, 6, vl), 3, vl); + vint16m1_t v_w = __riscv_vmul_vv_i16m1( + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)), + __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_3), vl); + + const int8_t* q8 = &lhs_qs_ptr[(base_k_phase + k_offsets[3] + l) * 4]; + v_isum_0 = __riscv_vwmacc_vx_i32m2(v_isum_0, (int16_t)q8[0], v_w, vl); + v_isum_1 = __riscv_vwmacc_vx_i32m2(v_isum_1, (int16_t)q8[1], v_w, vl); + v_isum_2 = __riscv_vwmacc_vx_i32m2(v_isum_2, (int16_t)q8[2], v_w, vl); + v_isum_3 = __riscv_vwmacc_vx_i32m2(v_isum_3, (int16_t)q8[3], v_w, vl); + } + } + + // C CORRECTION + int sb_base_abs = base_k_phase / 16; + + // --- Correction Sub-block 0 --- + { + int sb_abs = sb_base_abs + (k_offsets[0] / 16); + vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_0); + + // Row 0 + vfloat32m2_t v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[0], vl); + vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 0], vl); + vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_0 = __riscv_vfsub_vv_f32m2(v_sumf_0, vf_c, vl); + + // Row 1 + v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[1], vl); + v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 1], vl); + vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_1 = __riscv_vfsub_vv_f32m2(v_sumf_1, vf_c, vl); + + // Row 2 + v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[2], vl); + v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 2], vl); + vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_2 = __riscv_vfsub_vv_f32m2(v_sumf_2, vf_c, vl); + + // Row 3 + v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[3], vl); + v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 3], vl); + vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_3 = __riscv_vfsub_vv_f32m2(v_sumf_3, vf_c, vl); + } + + // --- Correction Sub-block 1 --- + { + int sb_abs = sb_base_abs + (k_offsets[1] / 16); + vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_1); + + vfloat32m2_t v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[0], vl); + vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 0], vl); + vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_0 = __riscv_vfsub_vv_f32m2(v_sumf_0, vf_c, vl); + + v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[1], vl); + v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 1], vl); + vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_1 = __riscv_vfsub_vv_f32m2(v_sumf_1, vf_c, vl); + + v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[2], vl); + v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 2], vl); + vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_2 = __riscv_vfsub_vv_f32m2(v_sumf_2, vf_c, vl); + + v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[3], vl); + v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 3], vl); + vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_3 = __riscv_vfsub_vv_f32m2(v_sumf_3, vf_c, vl); + } + + // --- Correction Sub-block 2 --- + { + int sb_abs = sb_base_abs + (k_offsets[2] / 16); + vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_2); + + vfloat32m2_t v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[0], vl); + vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 0], vl); + vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_0 = __riscv_vfsub_vv_f32m2(v_sumf_0, vf_c, vl); + + v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[1], vl); + v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 1], vl); + vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_1 = __riscv_vfsub_vv_f32m2(v_sumf_1, vf_c, vl); + + v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[2], vl); + v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 2], vl); + vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_2 = __riscv_vfsub_vv_f32m2(v_sumf_2, vf_c, vl); + + v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[3], vl); + v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 3], vl); + vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_3 = __riscv_vfsub_vv_f32m2(v_sumf_3, vf_c, vl); + } + + // --- Correction Sub-block 3 --- + { + int sb_abs = sb_base_abs + (k_offsets[3] / 16); + vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_3); + + vfloat32m2_t v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[0], vl); + vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 0], vl); + vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_0 = __riscv_vfsub_vv_f32m2(v_sumf_0, vf_c, vl); + + v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[1], vl); + v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 1], vl); + vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_1 = __riscv_vfsub_vv_f32m2(v_sumf_1, vf_c, vl); + + v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[2], vl); + v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 2], vl); + vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_2 = __riscv_vfsub_vv_f32m2(v_sumf_2, vf_c, vl); + + v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[3], vl); + v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 3], vl); + vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl); + v_sumf_3 = __riscv_vfsub_vv_f32m2(v_sumf_3, vf_c, vl); + } + + } // End Phase Loop + + // --- Apply Main Scales --- + vfloat16m1_t v_g_all_f16 = __riscv_vle16_v_f16m1((const _Float16*)rhs_current->d, vl); + vfloat32m2_t v_g_all_base = __riscv_vfwcvt_f_f_v_f32m2(v_g_all_f16, vl); + + { + vfloat32m2_t v_g_all = __riscv_vfmul_vf_f32m2(v_g_all_base, lhs_current->d[0], vl); + vfloat32m2_t v_sum = __riscv_vfcvt_f_x_v_f32m2(v_isum_0, vl); + v_sum = __riscv_vfmul_vv_f32m2(v_sum, v_g_all, vl); + v_sumf_0 = __riscv_vfadd_vv_f32m2(v_sumf_0, v_sum, vl); + } + // Row 1 + { + vfloat32m2_t v_g_all = __riscv_vfmul_vf_f32m2(v_g_all_base, lhs_current->d[1], vl); + vfloat32m2_t v_sum = __riscv_vfcvt_f_x_v_f32m2(v_isum_1, vl); + v_sum = __riscv_vfmul_vv_f32m2(v_sum, v_g_all, vl); + v_sumf_1 = __riscv_vfadd_vv_f32m2(v_sumf_1, v_sum, vl); + } + // Row 2 + { + vfloat32m2_t v_g_all = __riscv_vfmul_vf_f32m2(v_g_all_base, lhs_current->d[2], vl); + vfloat32m2_t v_sum = __riscv_vfcvt_f_x_v_f32m2(v_isum_2, vl); + v_sum = __riscv_vfmul_vv_f32m2(v_sum, v_g_all, vl); + v_sumf_2 = __riscv_vfadd_vv_f32m2(v_sumf_2, v_sum, vl); + } + // Row 3 + { + vfloat32m2_t v_g_all = __riscv_vfmul_vf_f32m2(v_g_all_base, lhs_current->d[3], vl); + vfloat32m2_t v_sum = __riscv_vfcvt_f_x_v_f32m2(v_isum_3, vl); + v_sum = __riscv_vfmul_vv_f32m2(v_sum, v_g_all, vl); + v_sumf_3 = __riscv_vfadd_vv_f32m2(v_sumf_3, v_sum, vl); + } + + } // End K-Block + + __riscv_vse32_v_f32m2(s + (row_tile + 0) * bs + col_tile, v_sumf_0, vl); + __riscv_vse32_v_f32m2(s + (row_tile + 1) * bs + col_tile, v_sumf_1, vl); + __riscv_vse32_v_f32m2(s + (row_tile + 2) * bs + col_tile, v_sumf_2, vl); + __riscv_vse32_v_f32m2(s + (row_tile + 3) * bs + col_tile, v_sumf_3, vl); + } + } +} diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp index 24e8ab4618..fd72d92901 100644 --- a/ggml/src/ggml-cpu/repack.cpp +++ b/ggml/src/ggml-cpu/repack.cpp @@ -48,6 +48,90 @@ static inline int nearest_int(float fval) { extern "C" { +#if defined __riscv_zvfh +void ggml_quantize_mat_q8_0_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { + assert(QK8_0 == 32); + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy; + + // scalar + const int blck_size_interleave = 1; + float srcv[4][QK8_0]; + float id[4]; + + for (int i = 0; i < nb; i++) { + for (int row_iter = 0; row_iter < 4; row_iter++) { + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_0; j++) { + srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j]; + amax = MAX(amax, fabsf(srcv[row_iter][j])); + } + + const float d = amax / ((1 << 7) - 1); + id[row_iter] = d ? 1.0f / d : 0.0f; + + y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d); + } + + for (int j = 0; j < QK8_0 * 4; j++) { + int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; + int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; + src_offset += (j % blck_size_interleave); + + float x0 = srcv[src_id][src_offset] * id[src_id]; + y[i].qs[j] = roundf(x0); + } + } +} + +void ggml_quantize_mat_q8_K_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { + assert(QK_K == 256); + assert(k % QK_K == 0); + const int nb = k / QK_K; + + block_q8_Kx4 * GGML_RESTRICT y = (block_q8_Kx4 *) vy; + + const int blck_size_interleave = 1; + float srcv[4][QK_K]; + float iscale[4]; + + for (int i = 0; i < nb; i++) { + for (int row_iter = 0; row_iter < 4; row_iter++) { + float amax = 0.0f; // absolute max + float max = 0; + + for (int j = 0; j < QK_K; j++) { + srcv[row_iter][j] = x[row_iter * k + i * QK_K + j]; + // Update the maximum value of the corresponding super block + if(amax < fabsf(srcv[row_iter][j])) { + amax = fabsf(srcv[row_iter][j]); + max = srcv[row_iter][j]; + } + } + + iscale[row_iter] = amax ? -127.f/max : 0; + y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0; + } + + for (int j = 0; j < QK_K / 4; j++) { + y[i].bsums[j] = 0; + } + for (int j = 0; j < QK_K * 4; j++) { + int src_id = j % 4; + int src_offset = j / 4; + int index = ((j >> 6) << 2) + (j & 3); + + float x0 = srcv[src_id][src_offset] * iscale[src_id]; + y[i].qs[j] = nearest_int(x0); + y[i].bsums[index] += y[i].qs[j]; + } + } +} +#endif + void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(QK8_0 == 32); assert(k % QK8_0 == 0); @@ -124,7 +208,6 @@ void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GG } } - void ggml_quantize_mat_q8_K_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(QK_K == 256); assert(k % QK_K == 0); @@ -256,6 +339,20 @@ template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTR ggml_quantize_mat_q8_K_4x8(x, vy, n_per_row); } +#if defined __riscv_zvfh +template <> void ggml_quantize_mat_t<1, GGML_TYPE_Q8_0>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) { + assert(nrow == 4); + UNUSED(nrow); + ggml_quantize_mat_q8_0_4x1(x, vy, n_per_row); +} + +template <> void ggml_quantize_mat_t<1, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) { + assert(nrow == 4); + UNUSED(nrow); + ggml_quantize_mat_q8_K_4x1(x, vy, n_per_row); +} +#endif + extern "C" { void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { @@ -964,6 +1061,292 @@ void ggml_gemv_q8_0_4x8_q8_0_generic(int n, } } +#if defined __riscv_zvfh +void ggml_gemv_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + + float sumf[16]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x16 * b_ptr = (const block_q4_0x16 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; + } + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } +} + +void ggml_gemv_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK_K; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + float sumf[16]; + float sum_minf[16]; + uint8_t scales[128]; + uint8_t mins[128]; + int sumi1; + int sumi2; + int sumi; + const block_q8_K * a_ptr = (const block_q8_K *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_Kx16 * b_ptr = (const block_q4_Kx16 *) vx + (x * nb); + for (int j = 0; j < ncols_interleaved; j++) { + sumf[j] = 0.0f; + sum_minf[j] = 0.0f; + } + for (int l = 0; l < nb; l++) { + for (int i = 0; i < 128; i++) { + scales[i] = b_ptr[l].scales[i] & 0x0F; + mins[i] = b_ptr[l].scales[i] >> 4; + } + for (int i = 0; i < 64; i++) { + scales[i] |= (b_ptr[l].scales[128 + i] & 0x03) << 4; + mins[i] |= (b_ptr[l].scales[128 + i] & 0x0C) << 2; + scales[i + 64] |= (b_ptr[l].scales[128 + i] & 0x30); + mins[i + 64] |= (b_ptr[l].scales[128 + i] & 0xC0) >> 2; + } + for (int sb = 0; sb < 8; sb++) { + uint8_t *min = &mins[sb * 16]; + for (int j = 0; j < ncols_interleaved; j++) { + sum_minf[j] += min[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d; + } + } + for (int sb = 0; sb < 8; sb += 2) { + uint8_t *scales_0 = &scales[sb * 16]; + uint8_t *scales_1 = &scales[(sb + 1) * 16]; + for (int i = 0; i < QK4_0; i++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi1 = 0; + sumi2 = 0; + sumi = 0; + const int v0 = (int8_t) (b_ptr[l].qs[sb * 256 + i * 16 + j] & 0xF); + const int v1 = (int8_t) (b_ptr[l].qs[sb * 256 + i * 16 + j] >> 4); + sumi1 = (v0 * a_ptr[l].qs[sb * 32 + i]); + sumi2 = (v1 * a_ptr[l].qs[sb * 32 + 32 + i]); + sumi1 = sumi1 * scales_0[j]; + sumi2 = sumi2 * scales_1[j]; + sumi += sumi1 + sumi2; + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d; + } + } + } + } + for (int j = 0; j < ncols_interleaved; j++) { + s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j]; + } + } +} + +void ggml_gemv_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert(nr == 1); + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(bs); + UNUSED(nr); + + float sumf[16]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx16 * b_ptr = (const block_iq4_nlx16 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); + } + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } +} + +void ggml_gemv_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert(nr == 1); + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(bs); + UNUSED(nr); + + float sumf[16]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q8_0x16 * b_ptr = (const block_q8_0x16 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) { + sumf[j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / blocklen); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; + sumi += v0 * a_ptr[l].qs[k * blocklen + i]; + } + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) { + s[x * ncols_interleaved + j] = sumf[j]; + } + } +} + +void ggml_gemv_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + assert(n % QK_K == 0); + assert(nr == 1); + assert(nc % 16 == 0); + + const int nb = n / QK_K; + const block_q2_Kx16 * x = (const block_q2_Kx16 *)vx; + const block_q8_K * y = (const block_q8_K *)vy; + + // Layout: Even-Low(0,2,4,6), Odd-Low(1,3,5,7), Even-High(8...), Odd-High(9...) + const int sb_perm[16] = { + 0, 4, 1, 5, 2, 6, 3, 7, // 0-7 + 8, 12, 9, 13, 10, 14, 11, 15 // 8-15 + }; + + for (int col_tile = 0; col_tile < nc; col_tile += 16) { + const block_q2_Kx16 * x_ptr = x + (col_tile / 16) * nb; + const block_q8_K * y_ptr = y; + + float sumf[16] = {0}; + + // Loop over K-blocks + for (int k_block = 0; k_block < nb; ++k_block) { + int32_t isum[16] = {0}; + int32_t summs[16] = {0}; + + const uint8_t * qs_rhs = x_ptr[k_block].qs; + const uint8_t * sc_rhs = x_ptr[k_block].scales; + const int8_t * qs_lhs = y_ptr[k_block].qs; + const int16_t * bs_lhs = y_ptr[k_block].bsums; + + // Iterate over sub-blocks 0..15 + for (int sb = 0; sb < 16; ++sb) { + // Correction Term + int16_t bsum = bs_lhs[sb]; + int scale_offset = sb_perm[sb] * 16; + + for (int col = 0; col < 16; ++col) { + uint8_t sc_val = sc_rhs[scale_offset + col]; + summs[col] += bsum * (sc_val >> 4); // Min is high 4 bits + } + + // Main Dot Product + // Calculate base offsets for Q2 unpacking based on SB + int byte_base; + if (sb < 8) byte_base = (sb % 2 == 0) ? 0 : 16; + else byte_base = (sb % 2 == 0) ? 32 : 48; + + int shift = ((sb / 2) % 4) * 2; + + for (int col = 0; col < 16; ++col) { + uint8_t sc_val = sc_rhs[scale_offset + col]; + int32_t d_sb = sc_val & 0xF; // Scale is low 4 bits + + // Process 16 elements (l=0..15) + for (int l = 0; l < 16; ++l) { + // Q2: Interleaved by column. Byte `l` contains 4 k-values. + int qs_idx = (byte_base + l) * 16 + col; + uint8_t q2_val = (qs_rhs[qs_idx] >> shift) & 3; + + // Q8: Linear access + int k = sb * 16 + l; + int8_t q8_val = qs_lhs[k]; + + isum[col] += q8_val * q2_val * d_sb; + } + } + } + + // Finalize K-Block + for (int col = 0; col < 16; ++col) { + float d_lhs = y_ptr[k_block].d; + float d_rhs = GGML_FP16_TO_FP32(x_ptr[k_block].d[col]); + float dm_rhs = GGML_FP16_TO_FP32(x_ptr[k_block].dmin[col]); + + float d_all = d_lhs * d_rhs; + float d_min = d_lhs * dm_rhs; + + sumf[col] += (isum[col] * d_all) - (summs[col] * d_min); + } + } + + for (int col = 0; col < 16; ++col) { + s[col_tile + col] = sumf[col]; + } + } +} +#endif + void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; @@ -1742,6 +2125,8 @@ void ggml_gemm_q8_0_4x4_q8_0_generic(int n, } } + + void ggml_gemm_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, @@ -1794,6 +2179,342 @@ void ggml_gemm_q8_0_4x8_q8_0_generic(int n, } } +#if defined __riscv_zvfh +void ggml_gemm_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + + float sumf[4][16]; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x16 * b_ptr = (const block_q4_0x16 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; + } + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } +} + +void ggml_gemm_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK_K; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + + float sumf[4][16]; + float sum_minf[4][16]; + uint8_t scales[128]; + uint8_t mins[128]; + int sumi1; + int sumi2; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_Kx16 * b_ptr = (const block_q4_Kx16 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumf[m][j] = 0.0; + sum_minf[m][j] = 0.0; + } + } + for (int l = 0; l < nb; l++) { + for (int i = 0; i < 128; i++) { + scales[i] = b_ptr[l].scales[i] & 0x0F; + mins[i] = b_ptr[l].scales[i] >> 4; + } + for (int i = 0; i < 64; i++) { + scales[i] |= (b_ptr[l].scales[128 + i] & 0x03) << 4; + mins[i] |= (b_ptr[l].scales[128 + i] & 0x0C) << 2; + scales[i + 64] |= (b_ptr[l].scales[128 + i] & 0x30); + mins[i + 64] |= (b_ptr[l].scales[128 + i] & 0xC0) >> 2; + } + + for (int sb = 0; sb < 8; sb++) { + uint8_t *min = &mins[sb * 16]; + for(int m = 0; m < 4; m++) { + const int16_t bsums = a_ptr[l].bsums[sb * 8 + m] + a_ptr[l].bsums[sb * 8 + m + 4]; + for(int j = 0; j < ncols_interleaved; j++) { + sum_minf[m][j] += min[j] * bsums * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; + } + } + } + + for (int sb = 0; sb < 8; sb += 2) { + uint8_t *scales_0 = &scales[sb * 16]; + uint8_t *scales_1 = &scales[(sb + 1) * 16]; + + for (int i = 0; i < QK4_0; i++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi1 = 0; + sumi2 = 0; + sumi = 0; + + const int v0 = (int8_t) (b_ptr[l].qs[sb * 256 + i * 16 + j] & 0xF); + const int v1 = (int8_t) (b_ptr[l].qs[sb * 256 + i * 16 + j] >> 4); + sumi1 = (v0 * a_ptr[l].qs[sb * 4 * 32 + i * 4 + m]); + sumi2 = (v1 * a_ptr[l].qs[sb * 4 * 32 + 32 * 4 + i * 4 + m]); + sumi1 = sumi1 * scales_0[j]; + sumi2 = sumi2 * scales_1[j]; + sumi += sumi1 + sumi2; + + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m]; + } + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j]; + } + } + } + } +} + +void ggml_gemm_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert(n % qk == 0); + assert(nr % 4 == 0); + assert(nc % ncols_interleaved == 0); + + float sumf[4][16]; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx16 * b_ptr = (const block_iq4_nlx16 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + (qk / 2) * 4])); + } + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } +} + +void ggml_gemm_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 16; + const int blocklen = 1; + + assert(n % qk == 0); + assert(nr % 4 == 0); + assert(nc % ncols_interleaved == 0); + + float sumf[4][16]; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q8_0x16 * b_ptr = (const block_q8_0x16 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumf[m][j] = 0.0; + } + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / blocklen); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; + sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]; + } + sumf[m][j] += + sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } + } +} + + +void ggml_gemm_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + assert(n % QK_K == 0); + assert(nr % 4 == 0); + assert(nc % 16 == 0); + const int nb = n / QK_K; + const block_q2_Kx16 * x = (const block_q2_Kx16 *)vx; + const block_q8_Kx4 * y = (const block_q8_Kx4 *)vy; + + const int sb_perm[16] = { + 0, 4, 1, 5, 2, 6, 3, 7, + 8, 12, 9, 13, 10, 14, 11, 15 + }; + + // Iterate Rows in tiles of 4 + for (int row_tile = 0; row_tile < nr; row_tile += 4) { + // Iterate Columns in tiles of 16 + for (int col_tile = 0; col_tile < nc; col_tile += 16) { + + const block_q2_Kx16 * x_ptr = x + (col_tile / 16) * nb; + const block_q8_Kx4 * y_ptr = y + (row_tile / 4) * nb; + + float sumf[4][16]; + memset(sumf, 0, sizeof(sumf)); + + for (int k_block = 0; k_block < nb; ++k_block) { + int32_t isum[4][16]; + int32_t summs[4][16]; + memset(isum, 0, sizeof(isum)); + memset(summs, 0, sizeof(summs)); + + const uint8_t * qs_rhs = x_ptr[k_block].qs; + const uint8_t * sc_rhs = x_ptr[k_block].scales; + const int8_t * qs_lhs = y_ptr[k_block].qs; + const int16_t * bs_lhs = y_ptr[k_block].bsums; + + for (int sb = 0; sb < 16; ++sb) { + int scale_offset = sb_perm[sb] * 16; + + int byte_base; + if (sb < 8) byte_base = (sb % 2 == 0) ? 0 : 16; + else byte_base = (sb % 2 == 0) ? 32 : 48; + int shift = ((sb / 2) % 4) * 2; + + for (int col = 0; col < 16; ++col) { + uint8_t sc_val = sc_rhs[scale_offset + col]; + int32_t d_sb = sc_val & 0xF; + int32_t m_sb = sc_val >> 4; + + // Correction Term + for (int r = 0; r < 4; ++r) { + int bsum_idx = (sb / 4) * 16 + r * 4 + (sb % 4); + summs[r][col] += bs_lhs[bsum_idx] * m_sb; + } + + // Main Dot Product + for (int l = 0; l < 16; ++l) { + int qs_idx = (byte_base + l) * 16 + col; + uint8_t q2_val = (qs_rhs[qs_idx] >> shift) & 3; + + // Calculate Q8 index for this specific k and row + int k = sb * 16 + l; + int q8_idx = (k / 4) * 16 + (k % 4); + + for (int r = 0; r < 4; ++r) { + // Add r*4 to jump to the correct row within the 4x4 chunk + int8_t q8_val = qs_lhs[q8_idx + r * 4]; + isum[r][col] += q8_val * q2_val * d_sb; + } + } + } + } + + // Finalize K-Block + for (int col = 0; col < 16; ++col) { + float d_rhs = GGML_FP16_TO_FP32(x_ptr[k_block].d[col]); + float dm_rhs = GGML_FP16_TO_FP32(x_ptr[k_block].dmin[col]); + + for (int r = 0; r < 4; ++r) { + float d_lhs = y_ptr[k_block].d[r]; + float d_all = d_lhs * d_rhs; + float d_min = d_lhs * dm_rhs; + sumf[r][col] += (isum[r][col] * d_all) - (summs[r][col] * d_min); + } + } + } + + for (int r = 0; r < 4; ++r) { + for (int col = 0; col < 16; ++col) { + s[(row_tile + r) * bs + (col_tile + col)] = sumf[r][col]; + } + } + } + } +} +#endif + } // extern "C" static block_q8_0x4 make_block_q8_0x4(block_q8_0 * in, unsigned int blck_size_interleave) { @@ -1882,6 +2603,31 @@ static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_in return out; } +static block_q4_0x16 make_block_q4_0x16(block_q4_0 * in, unsigned int blck_size_interleave) { + block_q4_0x16 out; + + for (int i = 0; i < 16; i++) { + out.d[i] = in[i].d; + } + + const int end = QK4_0 * 8 / blck_size_interleave; + + if (blck_size_interleave == 1) { + const uint8_t xor_mask = 0x88; + for (int i = 0; i < end; ++i) { + int src_id = i % 16; + int src_offset = i / 16; + int dst_offset = i; + + out.qs[dst_offset] = in[src_id].qs[src_offset] ^ xor_mask; + } + } else { + GGML_ASSERT(false); + } + + return out; +} + static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_interleave) { block_q4_Kx8 out; //Delta(scale) and dmin values of the eight Q4_K structures are copied onto the output interleaved structure @@ -1958,6 +2704,58 @@ static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_in return out; } +static block_q4_Kx16 make_block_q4_Kx16(block_q4_K * in, unsigned int blck_size_interleave) { + block_q4_Kx16 out; + //Delta(scale) and dmin values of the 16 Q4_K structures are copied onto the output interleaved structure + for (int i = 0; i < 16; i++) { + out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d; + } + + for (int i = 0; i < 16; i++) { + out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin; + } + + const int end = QK_K * 8 / blck_size_interleave; + + if (blck_size_interleave == 1) { + for (int i = 0; i < end; ++i) { + int src_id = i % 16; + int src_offset = i / 16; + int dst_offset = i; + + out.qs[dst_offset] = in[src_id].qs[src_offset]; + } + + // RVV repacking. + // + // Extract sums and mins for all 8 sub-blocks for each block of Q4_K. + uint8_t s[128], m[128]; + for (int i = 0; i < 4; i++) { + for (int j = 0; j < 16; j++) { + s[i * 16 + j] = in[j].scales[i] & 63; + m[i * 16 + j] = in[j].scales[i + 4] & 63; + } + } + for (int i = 0; i < 4; i++) { + for (int j = 0; j < 16; j++) { + s[64 + i * 16 + j] = ((in[j].scales[i] & 192) >> 2) | (in[j].scales[i+8] & 15); + m[64 + i * 16 + j] = ((in[j].scales[i + 4] & 192) >> 2) | ((in[j].scales[i+8] & 240) >> 4); + } + } + + for (int i = 0; i < 128; i++) { + out.scales[i] = (s[i] & 15) | ((m[i] & 15) << 4); + } + for (int i = 0; i < 64; i++) { + out.scales[128 + i] = ((s[i] & 48) >> 4) | ((m[i] & 48) >> 2) | (s[64 + i] & 48) | ((m[64 + i] & 48) << 2); + } + } else { + GGML_ASSERT(false); + } + + return out; +} + static block_q2_Kx8 make_block_q2_Kx8(block_q2_K * in, unsigned int blck_size_interleave) { block_q2_Kx8 out; @@ -2135,6 +2933,68 @@ static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_in return out; } +static block_q2_Kx16 make_block_q2_Kx16(const block_q2_K * in, unsigned int blck_size_interleave) { + block_q2_Kx16 out; + constexpr int N_COLS = 16; + + // 1. Copy Super-Scales (d) and Super-Mins (dmin) + for (int i = 0; i < N_COLS; i++) { + out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d; + out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin; + } + + // 2. Interleave Q2_K Data + const int bytes_per_col = 64; + const int total_bytes = N_COLS * bytes_per_col; + const int end = total_bytes / blck_size_interleave; + + for (int i = 0; i < end; ++i) { + int src_col_id = i % N_COLS; + int src_offset = (i / N_COLS) * blck_size_interleave; + int dst_offset = i * blck_size_interleave; + memcpy(&out.qs[dst_offset], &in[src_col_id].qs[src_offset], blck_size_interleave); + } + + // 3. Repack Scales into the Optimized "Sequential-Parallel" Layout + int out_idx = 0; + + // Arrays define the sub-block order for each group + const int even_low_sbs[] = {0, 2, 4, 6}; + const int odd_low_sbs[] = {1, 3, 5, 7}; + const int even_high_sbs[] = {8, 10, 12, 14}; + const int odd_high_sbs[] = {9, 11, 13, 15}; + + // Pack Group 1: Even-Low + for (int sb : even_low_sbs) { + for (int col = 0; col < N_COLS; col++) { + out.scales[out_idx++] = in[col].scales[sb]; + } + } + + // Pack Group 2: Odd-Low + for (int sb : odd_low_sbs) { + for (int col = 0; col < N_COLS; col++) { + out.scales[out_idx++] = in[col].scales[sb]; + } + } + + // Pack Group 3: Even-High + for (int sb : even_high_sbs) { + for (int col = 0; col < N_COLS; col++) { + out.scales[out_idx++] = in[col].scales[sb]; + } + } + + // Pack Group 4: Odd-High + for (int sb : odd_high_sbs) { + for (int col = 0; col < N_COLS; col++) { + out.scales[out_idx++] = in[col].scales[sb]; + } + } + + return out; +} + static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { GGML_ASSERT(t->type == GGML_TYPE_Q4_0); GGML_ASSERT(interleave_block == 4 || interleave_block == 8); @@ -2197,6 +3057,36 @@ static int repack_q4_K_to_q4_K_8_bl(struct ggml_tensor * t, int interleave_block GGML_UNUSED(data_size); } +static int repack_q4_K_to_q4_K_16_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { + GGML_ASSERT(t->type == GGML_TYPE_Q4_K); + constexpr int nrows_interleaved = 16; + + block_q4_Kx16 * dst = (block_q4_Kx16*)t->data; + const block_q4_K * src = (const block_q4_K*) data; + block_q4_K dst_tmp[16]; + int nrow = ggml_nrows(t); + int nblocks = t->ne[0] / QK_K; + + GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_K)); + + if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { + return -1; + } + + for (int b = 0; b < nrow; b += nrows_interleaved) { + for (int64_t x = 0; x < nblocks; x++) { + for (int i = 0; i < nrows_interleaved; i++ ) { + dst_tmp[i] = src[x + i * nblocks]; + } + *dst++ = make_block_q4_Kx16(dst_tmp, interleave_block); + } + src += nrows_interleaved * nblocks; + } + return 0; + + GGML_UNUSED(data_size); +} + static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { GGML_ASSERT(t->type == GGML_TYPE_Q2_K); GGML_ASSERT(interleave_block == 8); @@ -2228,6 +3118,71 @@ static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block GGML_UNUSED(data_size); } +static int repack_q2_K_to_q2_K_16_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { + GGML_ASSERT(t->type == GGML_TYPE_Q2_K); + constexpr int nrows_interleaved = 16; + + block_q2_Kx16 * dst = (block_q2_Kx16*)t->data; + const block_q2_K * src = (const block_q2_K*) data; + + block_q2_K dst_tmp[nrows_interleaved]; + + int nrow = ggml_nrows(t); + int nblocks = t->ne[0] / QK_K; + + GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q2_K)); + + if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { + return -1; + } + + for (int b = 0; b < nrow; b += nrows_interleaved) { + for (int64_t x = 0; x < nblocks; x++) { + // This loop gathers 16 separate blocks (one from each column) + // that correspond to the same K-dimension chunk. + for (int i = 0; i < nrows_interleaved; i++ ) { + dst_tmp[i] = src[x + i * nblocks]; + } + + *dst++ = make_block_q2_Kx16(dst_tmp, interleave_block); + } + src += nrows_interleaved * nblocks; + } + return 0; + + GGML_UNUSED(data_size); +} + +static int repack_q4_0_to_q4_0_16_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { + GGML_ASSERT(t->type == GGML_TYPE_Q4_0); + constexpr int nrows_interleaved = 16; + + block_q4_0x16 * dst = (block_q4_0x16*)t->data; + const block_q4_0 * src = (const block_q4_0*) data; + block_q4_0 dst_tmp[16]; + int nrow = ggml_nrows(t); + int nblocks = t->ne[0] / QK4_0; + + GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0)); + + if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { + return -1; + } + + for (int b = 0; b < nrow; b += nrows_interleaved) { + for (int64_t x = 0; x < nblocks; x++) { + for (int i = 0; i < nrows_interleaved; i++ ) { + dst_tmp[i] = src[x + i * nblocks]; + } + *dst++ = make_block_q4_0x16(dst_tmp, interleave_block); + } + src += nrows_interleaved * nblocks; + } + return 0; + + GGML_UNUSED(data_size); +} + static int repack_q5_K_to_q5_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, @@ -2352,6 +3307,60 @@ static int repack_q8_0_to_q8_0_4_bl(struct ggml_tensor * t, return 0; } +static block_q8_0x16 make_block_q8_0x16(block_q8_0 * in, unsigned int blck_size_interleave) { + block_q8_0x16 out; + + for (int i = 0; i < 16; i++) { + out.d[i] = in[i].d; + } + + const int end = QK8_0 * 16 / blck_size_interleave; + + if (blck_size_interleave == 1) { + for (int i = 0; i < end; ++i) { + int src_id = i % 16; + int src_offset = i / 16; + int dst_offset = i; + out.qs[dst_offset] = in[src_id].qs[src_offset]; + } + } else { + GGML_ASSERT(false); + } + + return out; +} + +static int repack_q8_0_to_q8_0_16_bl(struct ggml_tensor * t, + int interleave_block, + const void * GGML_RESTRICT data, + size_t data_size) { + GGML_ASSERT(t->type == GGML_TYPE_Q8_0); + constexpr int nrows_interleaved = 16; + + block_q8_0x16 * dst = (block_q8_0x16 *) t->data; + const block_q8_0 * src = (const block_q8_0 *) data; + block_q8_0 dst_tmp[16]; + int nrow = ggml_nrows(t); + int nblocks = t->ne[0] / QK8_0; + + GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q8_0)); + + if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { + return -1; + } + + for (int b = 0; b < nrow; b += nrows_interleaved) { + for (int64_t x = 0; x < nblocks; x++) { + for (int i = 0; i < nrows_interleaved; i++) { + dst_tmp[i] = src[x + i * nblocks]; + } + *dst++ = make_block_q8_0x16(dst_tmp, interleave_block); + } + src += nrows_interleaved * nblocks; + } + return 0; +} + static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_size_interleave) { block_iq4_nlx4 out; @@ -2477,6 +3486,63 @@ static int repack_iq4_nl_to_iq4_nl_8_bl(struct ggml_tensor * t, int interleave_b GGML_UNUSED(data_size); } +static block_iq4_nlx16 make_block_iq4_nlx16(block_iq4_nl * in, unsigned int blck_size_interleave) { + block_iq4_nlx16 out; + + for (int i = 0; i < 16; i++) { + out.d[i] = in[i].d; + } + + const int end = QK4_NL * 8 / blck_size_interleave; + + if (blck_size_interleave == 1) { + for (int i = 0; i < end; ++i) { + int src_id = i % 16; + int src_offset = i / 16; + int dst_offset = i; + + out.qs[dst_offset] = in[src_id].qs[src_offset]; + } + } else { + GGML_ASSERT(false); + } + + return out; +} + +static int repack_iq4_nl_to_iq4_nl_16_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { + GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL); + GGML_ASSERT(interleave_block == 1); + + const block_iq4_nl * src = (const block_iq4_nl *)data; + block_iq4_nlx16 * dst = ( block_iq4_nlx16 *)t->data; + + block_iq4_nl dst_tmp[16]; + + int nrow = ggml_nrows(t); + int nrows_interleaved = 16; + int nblocks = t->ne[0] / QK4_NL; + + GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl)); + + if (t->ne[1] % nrows_interleaved != 0) { + return -1; + } + + for (int b = 0; b < nrow; b += nrows_interleaved) { + for (int64_t x = 0; x < nblocks; x++) { + for (int i = 0; i < nrows_interleaved; i++) { + dst_tmp[i] = src[x + i * nblocks]; + } + *dst++ = make_block_iq4_nlx16(dst_tmp, interleave_block); + } + src += nrows_interleaved * nblocks; + } + return 0; + + GGML_UNUSED(data_size); +} + namespace ggml::cpu::repack { // repack template @@ -2527,7 +3593,6 @@ template <> int repack(struct ggml_tensor * t, const void * template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { return repack_iq4_nl_to_iq4_nl_8_bl(t, 8, data, data_size); } - template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { return repack_q8_0_to_q8_0_4_bl(t, 4, data, data_size); } @@ -2536,6 +3601,28 @@ template <> int repack(struct ggml_tensor * t, const void * da return repack_q8_0_to_q8_0_4_bl(t, 8, data, data_size); } +#if defined __riscv_zvfh +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_q4_0_to_q4_0_16_bl(t, 1, data, data_size); +} + +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_q4_K_to_q4_K_16_bl(t, 1, data, data_size); +} + +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_iq4_nl_to_iq4_nl_16_bl(t, 1, data, data_size); +} + +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_q8_0_to_q8_0_16_bl(t, 1, data, data_size); +} + +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_q2_K_to_q2_K_16_bl(t, 1, data, data_size); +} +#endif + // gemv template void gemv(int, float *, size_t, const void *, const void *, int, int); @@ -2586,7 +3673,6 @@ template <> void gemv(int n, float * s, size template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemv_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc); } - template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemv_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); } @@ -2595,6 +3681,28 @@ template <> void gemv(int n, float * s, size_t ggml_gemv_q8_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); } +#if defined __riscv_zvfh +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_q4_0_16x1_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_q4_K_16x1_q8_K(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_iq4_nl_16x1_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_q8_0_16x1_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_q2_K_16x1_q8_K(n, s, bs, vx, vy, nr, nc); +} +#endif + // gemm template void gemm(int, float *, size_t, const void *, const void *, int, int); @@ -2654,6 +3762,28 @@ template <> void gemm(int n, float * s, size_t ggml_gemm_q8_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); } +#if defined __riscv_zvfh +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_q4_0_16x1_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_q4_K_16x1_q8_K(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_iq4_nl_16x1_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_q8_0_16x1_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_q2_K_16x1_q8_K(n, s, bs, vx, vy, nr, nc); +} +#endif + class tensor_traits_base : public ggml::cpu::tensor_traits { public: virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0; @@ -3056,9 +4186,20 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons static const ggml::cpu::repack::tensor_traits q8_0_4x4_q8_0; static const ggml::cpu::repack::tensor_traits q8_0_4x8_q8_0; + // instances for RISC-V + // + // These implement outer-product style matrix multiplication kernels with + // an interleave of 1. +#if defined __riscv_zvfh + static const ggml::cpu::repack::tensor_traits q4_0_16x1_q8_0; + static const ggml::cpu::repack::tensor_traits q4_K_16x1_q8_K; + static const ggml::cpu::repack::tensor_traits iq4_nl_16x1_q8_0; + static const ggml::cpu::repack::tensor_traits q8_0_16x1_q8_0; + static const ggml::cpu::repack::tensor_traits q2_K_16x1_q8_K; +#endif + if (cur->type == GGML_TYPE_Q4_0) { - if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0) - || (ggml_cpu_has_riscv_v() && (ggml_cpu_get_rvv_vlen() >= QK4_0))) { + if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) { if (cur->ne[1] % 8 == 0) { return &q4_0_8x8_q8_0; } @@ -3073,6 +4214,17 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons return &q4_0_4x4_q8_0; } } + if (ggml_cpu_has_riscv_v()) { + #if defined __riscv_zvfh + switch (__riscv_vlenb() * 8) { + case 128: { break; } // TODO + case 256: { if (cur->ne[1] % 16 == 0) { return &q4_0_16x1_q8_0; } break; } + case 512: { break; } // TODO + case 1024: { break; } // TODO + default: { return nullptr; } + } + #endif + } } else if (cur->type == GGML_TYPE_Q4_K) { if (ggml_cpu_has_avx2()) { if (cur->ne[1] % 8 == 0) { @@ -3089,12 +4241,34 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons return &q4_K_8x4_q8_K; } } + if (ggml_cpu_has_riscv_v()) { + #if defined __riscv_zvfh + switch (__riscv_vlenb() * 8) { + case 128: { break; } // TODO + case 256: { if (cur->ne[1] % 16 == 0) { return &q4_K_16x1_q8_K; } break; } + case 512: { break; } // TODO + case 1024: { break; } // TODO + default: { return nullptr; } + } + #endif + } } else if (cur->type == GGML_TYPE_Q2_K) { if (ggml_cpu_has_avx512()) { if (cur->ne[1] % 8 == 0) { return &q2_K_8x8_q8_K; } } + if (ggml_cpu_has_riscv_v()) { + #if defined __riscv_zvfh + switch (__riscv_vlenb() * 8) { + case 128: { break; } // TODO + case 256: { if (cur->ne[1] % 16 == 0) { return &q2_K_16x1_q8_K; } break; } + case 512: { break; } // TODO + case 1024: { break; } // TODO + default: { return nullptr; } + } + #endif + } } else if (cur->type == GGML_TYPE_Q5_K) { if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { if (cur->ne[1] % 8 == 0) { @@ -3118,6 +4292,17 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons return &iq4_nl_4x4_q8_0; } } + if (ggml_cpu_has_riscv_v()) { + #if defined __riscv_zvfh + switch (__riscv_vlenb() * 8) { + case 128: { break; } // TODO + case 256: { if (cur->ne[1] % 16 == 0) { return &iq4_nl_16x1_q8_0; } break; } + case 512: { break; } // TODO + case 1024: { break; } // TODO + default: { return nullptr; } + } + #endif + } } else if (cur->type == GGML_TYPE_Q8_0) { if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { if (cur->ne[1] % 4 == 0) { @@ -3129,6 +4314,17 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons return &q8_0_4x4_q8_0; } } + if (ggml_cpu_has_riscv_v()) { + #if defined __riscv_zvfh + switch (__riscv_vlenb() * 8) { + case 128: { break; } // TODO + case 256: { if (cur->ne[1] % 16 == 0) { return &q8_0_16x1_q8_0; } break; } + case 512: { break; } // TODO + case 1024: { break; } // TODO + default: { return nullptr; } + } + #endif + } } return nullptr; diff --git a/ggml/src/ggml-cpu/repack.h b/ggml/src/ggml-cpu/repack.h index 855320eeeb..2f0521de06 100644 --- a/ggml/src/ggml-cpu/repack.h +++ b/ggml/src/ggml-cpu/repack.h @@ -28,13 +28,17 @@ template struct block { // control size static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding"); static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding"); +static_assert(sizeof(block<4, 16>) == 16 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<4,16> size/padding"); static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding"); static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding"); +static_assert(sizeof(block<8, 16>) == 16 * sizeof(ggml_half) + QK8_0 * 16, "wrong block<8,16> size/padding"); using block_q4_0x4 = block<4, 4>; using block_q4_0x8 = block<4, 8>; +using block_q4_0x16 = block<4, 16>; using block_q8_0x4 = block<8, 4>; using block_q8_0x8 = block<8, 8>; +using block_q8_0x16 = block<8, 16>; struct block_q4_Kx8 { ggml_half d[8]; // super-block scale for quantized scales @@ -44,7 +48,14 @@ struct block_q4_Kx8 { }; static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding"); +struct block_q4_Kx16 { + ggml_half d[16]; // super-block scale for quantized scales + ggml_half dmin[16]; // super-block scale for quantized mins + uint8_t scales[192]; // scales and mins, quantized with 6 bits + uint8_t qs[2048]; // 4--bit quants +}; +static_assert(sizeof(block_q4_Kx16) == sizeof(ggml_half) * 32 + K_SCALE_SIZE * 16 + QK_K * 8, "wrong q4_K block size/padding"); struct block_q2_Kx8 { ggml_half d[8]; // super-block scale for quantized scales ggml_half dmin[8]; // super-block scale for quantized mins @@ -53,6 +64,13 @@ struct block_q2_Kx8 { }; static_assert(sizeof(block_q2_Kx8) == sizeof(ggml_half) * 16 + QK_K/2 + QK_K * 2, "wrong q2_K block size/padding"); +struct block_q2_Kx16 { + ggml_half d[16]; // Super-block scale for quantized scales + ggml_half dmin[16]; // Super-block scale for quantized mins + uint8_t scales[256]; // Sub-block scales (16 cols * 16 sub-blocks) + uint8_t qs[1024]; // Data (16 cols * 64 bytes per block) +}; +static_assert(sizeof(block_q2_Kx16) == sizeof(ggml_half) * 32 + QK_K + QK_K * 4, "wrong q2_K block size/padding"); struct block_q5_Kx8 { ggml_half d[8]; // super-block scale for quantized scales @@ -97,6 +115,14 @@ struct block_iq4_nlx8 { static_assert(sizeof(block_iq4_nlx8) == 8 * sizeof(ggml_half) + QK4_NL * 4, "wrong iq4_nlx8 block size/padding"); +struct block_iq4_nlx16 { + ggml_half d[16]; // deltas for 16 iq4_nl blocks + uint8_t qs[QK4_NL * 8]; // nibbles / quants for 16 iq4_nl blocks +}; + +static_assert(sizeof(block_iq4_nlx16) == 16 * sizeof(ggml_half) + QK4_NL * 8, "wrong iq4_nlx16 block size/padding"); + + #if defined(__cplusplus) extern "C" { #endif @@ -115,6 +141,8 @@ void ggml_gemv_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo void ggml_gemv_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); @@ -125,10 +153,22 @@ void ggml_gemm_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo void ggml_gemm_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); -void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); -void ggml_gemv_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +#if defined __riscv_zvfh +void ggml_quantize_mat_q8_0_4x1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); +void ggml_quantize_mat_q8_K_4x1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); +void ggml_gemv_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +#endif // Native implementations void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); @@ -145,6 +185,8 @@ void ggml_gemv_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); @@ -155,10 +197,22 @@ void ggml_gemm_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); -void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); -void ggml_gemv_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +#if defined __riscv_zvfh +void ggml_quantize_mat_q8_0_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); +void ggml_quantize_mat_q8_K_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); +void ggml_gemv_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +#endif #if defined(__cplusplus) } // extern "C"