whispercpp 1.3.2 → 1.3.4

data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl

@@ -0,0 +1,176 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+
+static inline half4 mxfp4_to_fp16_packed(ushort fp4x4) {
+    ushort2 fp16_packed_a, fp16_packed_b, bias_a, bias_b, sign_a, sign_b;
+    fp16_packed_a.lo = (fp4x4 << 9) & 0x0E00;
+    fp16_packed_a.hi = (fp4x4 << 5) & 0x0E00;
+    fp16_packed_b.lo = (fp4x4 << 1) & 0x0E00;
+    fp16_packed_b.hi = (fp4x4 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a.lo == 0) ? 0x0 : 0x3800;
+    bias_a.hi = (fp16_packed_a.hi == 0) ? 0x0 : 0x3800;
+    bias_b.lo = (fp16_packed_b.lo == 0) ? 0x0 : 0x3800;
+    bias_b.hi = (fp16_packed_b.hi == 0) ? 0x0 : 0x3800;
+
+    fp16_packed_a.lo = (fp16_packed_a.lo == 0x0200) ? 0x0 : fp16_packed_a.lo;
+    fp16_packed_a.hi = (fp16_packed_a.hi == 0x0200) ? 0x0 : fp16_packed_a.hi;
+    fp16_packed_b.lo = (fp16_packed_b.lo == 0x0200) ? 0x0 : fp16_packed_b.lo;
+    fp16_packed_b.hi = (fp16_packed_b.hi == 0x0200) ? 0x0 : fp16_packed_b.hi;
+
+    sign_a.lo = (fp4x4 << 12) & 0x8000;
+    sign_a.hi = (fp4x4 << 8) & 0x8000;
+    sign_b.lo = (fp4x4 << 4) & 0x8000;
+    sign_b.hi = fp4x4 & 0x8000;
+
+    fp16_packed_a = sign_a + bias_a + fp16_packed_a;
+    fp16_packed_b = sign_b + bias_b + fp16_packed_b;
+
+    return as_half4((ushort4)(fp16_packed_a, fp16_packed_b));
+}
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+    bits = (x == 0) ? 0x00400000 : ((uint) x << 23);
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 4
+#define N_SG_MXFP4 1
+#define N_SIMDWIDTH 64
+#define SRC0Q_IMG
+#endif
+
+kernel void kernel_mul_mv_id_mxfp4_f32_flat(
+#ifdef SRC0Q_IMG
+    __read_only image1d_buffer_t src0_q,
+#else
+    global uchar * src0_q,
+#endif
+    global uchar * src0_e,
+    global uchar * src1,
+    ulong         offset1,
+    global uchar * src2,
+    ulong         offset2,
+    global uchar * dst,
+    ulong         offsetd,
+    int           ne00,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb03,
+    int           ne11,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    ulong         nb13,
+    int           ne20,
+    int           ne21,
+    ulong         nb21,
+    int           ne0,
+    int           ne1,
+    int           r2,
+    int           r3
+) {
+    dst  = dst  + offsetd;
+
+    const int iid1 = get_group_id(2) / ne20;
+    const int idx  = get_group_id(2) % ne20;
+
+    uint i02 = ((global uint *) (src2 + offset2 + iid1 * nb21))[idx];
+
+    int i11 = idx % ne11;
+
+    int nb = ne00 / QK_MXFP4;
+
+    uint src0_off = i02*nb02;
+    src0_off /= 17; // 17 = sizeof(block_mxfp4)
+
+    src0_e = src0_e + src0_off;
+
+    dst = dst + (idx * ne0 + iid1 * ne1 * ne0) * sizeof(float);
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint offset_src0 = first_row*nb01;
+    offset_src0 /= 17; // 17 = sizeof(block_mxfp4)
+#ifdef SRC0Q_IMG
+    ulong offset_q = src0_off + offset_src0;
+#else
+    src0_q = src0_q + src0_off*16;
+    global uchar16 * x_q = (global uchar16 *)(src0_q) + offset_src0;
+#endif
+    global uchar * x_e = src0_e + offset_src0;
+
+    const short ix = get_sub_group_local_id() >> 1;
+    const short it = get_sub_group_local_id() & 1;
+
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    src1 = src1 + offset1 + i11 * nb11 + iid1 * nb12;
+    global float * y   = (global float *) (src1 + r1 * nb11);
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH / 2) {
+        global float4 * y4 = (global float4 *)yb;
+
+        #pragma unroll
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            uchar xb_e = x_e[row * nb + ib];
+#ifdef SRC0Q_IMG
+            ushort4 xb_q = as_ushort4(read_imageui(src0_q, (offset_q + row * nb + ib) * 2 + it).xy);
+#else
+            ushort4 xb_q = vload4(0, (global ushort *)((global uchar *)(x_q + row * nb + ib) + 8 * it));
+#endif
+
+            half4 fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s0);
+            half4 fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s1);
+            float4 acc1 = y4[0] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2);
+            acc1 += y4[4] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3);
+
+            fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s2);
+            fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s3);
+            acc1 += y4[1] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2);
+            acc1 += y4[5] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3);
+
+            sumf[row] += e8m0_to_fp32(xb_e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH / 2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *)dst + (ulong)r1 * ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl

@@ -0,0 +1,283 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+// This function requires the original shuffled weights.
+// As a reminder, the original weights are shuffled so that (q[0], q[16]) are
+// packed together in a byte, so are (q[1], q[17]) and so on.
+inline float block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+//
+// This variant outputs 8 values.
+//
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each SIMD group works on 8 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_8x_flat(
+        global char  * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = 0;
+
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float8 sumf = 0.f;
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il);
+
+        sumf.s4 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il);
+        sumf.s5 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il);
+        sumf.s6 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il);
+        sumf.s7 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float8 tot = (float8)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_q4_0_f32_8x_flat(
+        global char  *  src0_q,
+        global half  *  src0_d,
+        global float *  src1,
+        ulong           offset1,
+        global char  *  src2,
+        ulong           offset2,
+        global float *  dst,
+        ulong           offsetd,
+        int             ne00,
+        int             ne01,
+        int             ne02,
+        ulong           nb00,
+        ulong           nb02,
+        int             ne10,
+        int             ne11,
+        int             ne12,
+        ulong           nb11,
+        ulong           nb12,
+        int             ne20,
+        int             ne21,
+        ulong           nb21,
+        int             ne0,
+        int             ne1,
+        int             r2,
+        int             r3
+) {
+    src1 = (global float *)((global char *)src1 + offset1);
+    src2 = (global char  *)((global char *)src2 + offset2);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    const int iid1 = get_group_id(2)/ne20;
+    const int idx  = get_group_id(2)%ne20;
+
+    const int i02 = ((global int *)(src2 + iid1*nb21))[idx];
+
+    const int i11 = idx%ne11;
+    const int i12 = iid1;
+
+    const int i1 = idx;
+    const int i2 = i12;
+
+    global char  * src0_q_cur = src0_q + (i02*nb02/nb00)*(QK4_0/2);
+    global half  * src0_d_cur = src0_d + (i02*nb02/nb00);
+    global float * src1_cur   = (global float *)((global char *) src1  + i11*nb11 + i12*nb12);
+    global float * dst_cur    = dst + i1*ne0 + i2*ne1*ne0;
+
+    mul_vec_q_n_f32_8x_flat(src0_q_cur, src0_d_cur, src1_cur, dst_cur, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl

@@ -0,0 +1,140 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK8_0 32
+typedef struct {
+    half d;       // delta
+    char qs[QK8_0]; // quants
+} block_q8_0;
+
+#define NB_Q8_0 8
+
+#ifdef INTEL_GPU
+#define N_R0_Q8_0 4 // number of rows each subgroup works on
+#define N_SG_Q8_0 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_Q8_0 4
+#define N_SG_Q8_0 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_q8_0_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * src2,
+    ulong         offset2,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    ulong         nb01,
+    ulong         nb02,
+    int           ne11,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    int           ne20,
+    int           ne21,
+    ulong         nb21,
+    int           ne0,
+    int           ne1
+) {
+    src0 = (global char *)((global char *)src0 + offset0);
+    src1 = (global char *)((global char *)src1 + offset1);
+    src2 = (global char *)((global char *)src2 + offset2);
+    dst  = (global char *)((global char *)dst  + offsetd);
+
+    int iid1 = get_group_id(2)/ne20;
+    int idx  = get_group_id(2)%ne20;
+
+    int i02 = ((global int *) (src2 + iid1*nb21))[idx];
+
+    int i11_ = idx % ne11;
+    int i12_ = iid1;
+
+    int i1 = idx;
+    int i2 = i12_;
+
+    global char * src0_cur = src0 + i02*nb02;
+    global char * src1_cur = src1 + i11_*nb11 + i12_*nb12;
+
+    global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float);
+
+    int nb = ne00/QK8_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+
+    int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0;
+
+    ulong offset_src1 = r1*nb11;
+    global float * y  = (global float *) (src1_cur + offset_src1);
+
+    // pointers to src0 rows
+    global block_q8_0 * ax[N_R0_Q8_0];
+    for (int row = 0; row < N_R0_Q8_0; ++row) {
+        ulong offset_src0 = (first_row + row)*nb01;
+        ax[row] = (global block_q8_0 *) ((global char *) src0_cur + offset_src0);
+    }
+
+    float yl[NB_Q8_0];
+    float sumf[N_R0_Q8_0] = { 0.f };
+
+    const short ix = get_sub_group_local_id()/4;
+    const short il = get_sub_group_local_id()%4;
+
+    global float * yb = y + ix*QK8_0 + il*NB_Q8_0;
+
+    // each thread handles NB_Q8_0 quants at a time
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) {
+        for (short i = 0; i < NB_Q8_0; ++i) {
+            yl[i] = yb[i];
+        }
+
+        for (short row = 0; row < N_R0_Q8_0; row++) {
+            global char * qs = ax[row][ib].qs + il*NB_Q8_0;
+            float sumq = 0.f;
+            for (short iq = 0; iq < NB_Q8_0; ++iq) {
+                sumq += qs[iq] * yl[iq];
+            }
+            sumf[row] += sumq*ax[row][ib].d;
+        }
+
+        yb += N_SIMDWIDTH*NB_Q8_0;
+    }
+
+    global float * dst_f32 = (global float *) dst_cur + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_Q8_0; ++row) {
+        float tot = sub_group_reduce_add(sumf[row]);
+
+        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
+            dst_f32[first_row + row] = tot;
+        }
+    }
+}