whispercpp 1.3.5 → 1.3.6

data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl

@@ -0,0 +1,229 @@
+#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_1                   32
+
+struct block_q4_1 {
+    half d; // delta
+    half m; // min
+    uchar qs[QK4_1 / 2]; // nibbles / quants
+};
+
+inline float block_q4_1_dot_y_flat(
+    global const uchar * x,
+    global const half  * dh,
+    global const half  * mh,
+    float sumy,
+    float16 yl,
+    int il
+) {
+    float                 d   = *dh;
+    float                 m   = *mh;
+    global const ushort * qs = ((global const ushort *) x + il/2);
+
+    float4 acc = (float4)(0.0f, 0.0f, 0.0f, 0.0f);
+
+    acc.s0 += yl.s0 * (qs[0] & 0x000F);
+    acc.s0 += yl.s1 * (qs[0] & 0x0F00);
+    acc.s0 += yl.s8 * (qs[0] & 0x00F0);
+    acc.s3 += yl.s9 * (qs[0] & 0xF000);
+
+    acc.s0 += yl.s2 * (qs[1] & 0x000F);
+    acc.s1 += yl.s3 * (qs[1] & 0x0F00);
+    acc.s2 += yl.sa * (qs[1] & 0x00F0);
+    acc.s3 += yl.sb * (qs[1] & 0xF000);
+
+    acc.s0 += yl.s4 * (qs[2] & 0x000F);
+    acc.s1 += yl.s5 * (qs[2] & 0x0F00);
+    acc.s2 += yl.sc * (qs[2] & 0x00F0);
+    acc.s3 += yl.sd * (qs[2] & 0xF000);
+
+    acc.s0 += yl.s6 * (qs[3] & 0x000F);
+    acc.s1 += yl.s7 * (qs[3] & 0x0F00);
+    acc.s2 += yl.se * (qs[3] & 0x00F0);
+    acc.s3 += yl.sf * (qs[3] & 0xF000);
+
+    return d * (acc.s0 + acc.s1 + acc.s2 + acc.s3) + sumy * m;
+}
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4 // each subgroup works on 4 rows
+#define N_SIMDGROUP 1 // number of subgroups in a thread group
+#define N_SIMDWIDTH 16 // assuming subgroup size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_flat(
+        global void * src0_q,
+        global void * src0_d,
+        global void * src0_m,
+        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_1;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    // The number of scales/mins is the same as the number of blocks.
+    ulong offset0_dm = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02));
+    // Each block contains QK4_1/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q  = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_1/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_dm;
+    global half  * m = (global half  *) src0_m + offset0_dm;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float4 sumf = (float4)(0.f, 0.f, 0.f, 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_1 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0;
+
+        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_q4_1_dot_y_flat(x + ib*QK4_1/2 + 0*nb*QK4_1/2, d + ib + 0*nb, m + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += block_q4_1_dot_y_flat(x + ib*QK4_1/2 + 1*nb*QK4_1/2, d + ib + 1*nb, m + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += block_q4_1_dot_y_flat(x + ib*QK4_1/2 + 2*nb*QK4_1/2, d + ib + 2*nb, m + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += block_q4_1_dot_y_flat(x + ib*QK4_1/2 + 3*nb*QK4_1/2, d + ib + 3*nb, m + ib + 3*nb, sumy, yl, il);
+
+        yb += QK4_1 * (N_SIMDWIDTH/2);
+    }
+
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3)
+    );
+
+    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;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q4_1_f32_flat(
+        global void * src0_q,
+        global void * src0_d,
+        global void * src0_m,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32_flat(src0_q, src0_d, src0_m, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl

@@ -0,0 +1,180 @@
+#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
+
+//------------------------------------------------------------------------------
+// block_q4_K
+//------------------------------------------------------------------------------
+#define QK_K            256
+#define K_SCALE_SIZE    12
+
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+typedef struct {
+    half d;    // super-block scale for quantized scales
+    half dmin; // super-block scale for quantized mins
+
+    uchar scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uchar qs[QK_K/2];           // 4-bit quants
+} block_q4_K;
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4 // number of rows each SIMD group works on
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // SIMD group size
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+#undef  BLOCK_STRIDE
+// number of (super) blocks each subgroup processes
+// each thread in a subgroup processes a block (32 weights)
+#define BLOCK_STRIDE (N_SIMDWIDTH/8)
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q4_K_f32(
+        global char * src0,
+        int offset0,
+        global char * src1,
+        int offset1,
+        global char * dst,
+        int offsetd,
+        int ne00,
+        int ne01,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    ushort kmask1 = 0x3f3f;
+    ushort kmask2 = 0x0f0f;
+    ushort kmask3 = 0xc0c0;
+
+    int ix = get_sub_group_local_id()/8;  // super block index
+    int it = get_sub_group_local_id()%8;  // block index (inside super block)
+    int iq = it/4;     // 0 or 1 - first or second half of the super block
+    int ir = it%4;     // 0...3 - block index in the half super block
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    int offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    int offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_q4_K * x = (global block_q4_K *) (src0 + offset_src0);
+    global float      * y = (global float      *) (src1 + offset_src1);
+
+    float yl[16];
+    float yh[16];
+    float sumf[N_DST] = {0.f};
+    float all_sum;
+
+    global float * y4 = y + ix * QK_K + 64 * iq + 8 * ir;
+
+    ushort  sc16[4];
+    uchar * sc8 = (uchar *)sc16;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        float4 sumy = {0.f, 0.f, 0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+0] = y4[i+0];
+            sumy.s0 += yl[i+0];
+
+            yl[i+8] = y4[i+32];
+            sumy.s1 += yl[i+8];
+
+            yh[i+0] = y4[i+128];
+            sumy.s2 += yh[i+0];
+
+            yh[i+8] = y4[i+160];
+            sumy.s3 += yh[i+8];
+        }
+
+        global ushort * sc = (global ushort *)x[ib].scales + iq;
+        global ushort * q1 = (global ushort *)x[ib].qs + 16 * iq + 4 * ir;
+        global half     * dh = &x[ib].d;
+
+        for (int row = 0; row < N_DST; row++) {
+            sc16[0] = sc[0] & kmask1;
+            sc16[1] = sc[2] & kmask1;
+            sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
+            sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);
+
+            global ushort * q2 = q1 + 32;
+
+            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
+            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                acc1.s0 += yl[i+0] * (q1[i/2] & 0x000F);
+                acc1.s1 += yl[i+1] * (q1[i/2] & 0x0F00);
+                acc1.s2 += yl[i+8] * (q1[i/2] & 0x00F0);
+                acc1.s3 += yl[i+9] * (q1[i/2] & 0xF000);
+                acc2.s0 += yh[i+0] * (q2[i/2] & 0x000F);
+                acc2.s1 += yh[i+1] * (q2[i/2] & 0x0F00);
+                acc2.s2 += yh[i+8] * (q2[i/2] & 0x00F0);
+                acc2.s3 += yh[i+9] * (q2[i/2] & 0xF000);
+            }
+
+            float dall = dh[0];
+            float dmin = dh[1];
+            sumf[row] += dall * ((acc1.s0 + 1.f/256.f * acc1.s1) * sc8[0] +
+                                 (acc1.s2 + 1.f/256.f * acc1.s3) * sc8[1] * 1.f/16.f +
+                                 (acc2.s0 + 1.f/256.f * acc2.s1) * sc8[4] +
+                                 (acc2.s2 + 1.f/256.f * acc2.s3) * sc8[5] * 1.f/16.f) -
+                         dmin * (sumy.s0 * sc8[2] + sumy.s1 * sc8[3] + sumy.s2 * sc8[6] + sumy.s3 * sc8[7]);
+
+            q1 += nb01/2;
+            sc += nb01/2;
+            dh += nb01/2;
+        }
+
+        y4 += BLOCK_STRIDE * QK_K;
+    }
+
+    global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0;
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = sub_group_reduce_add(sumf[row]);
+        if (first_row + row < ne01) {
+            if (get_sub_group_local_id() == 0) {
+                dst_f32[first_row + row] = all_sum;
+            }
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mv_q6_k.cl → mul_mv_q6_k_f32.cl}

@@ -111,6 +111,10 @@ kernel void kernel_mul_mv_q6_K_f32(
 
     int row = N_SIMDGROUP * r0 + get_sub_group_id();
 
+    if (row >= ne01) {
+        return;
+    }
+
     int i12 = im%ne12;
     int i13 = im/ne12;
 

data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl

@@ -0,0 +1,194 @@
+#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
+
+//------------------------------------------------------------------------------
+// kernel_mul_mv_q6_K_f32_flat
+//------------------------------------------------------------------------------
+#define Q6_K_MASK1 0x03
+#define Q6_K_MASK2 0x0C
+#define Q6_K_MASK3 0x30
+#define Q6_K_MASK4 0xC0
+
+#define QK_K       256
+
+inline float block_q_6_K_dot_y_flat(
+    global uchar * blk_ql,
+    global uchar * blk_qh,
+    global char  * blk_scales,
+    global half  * blk_d,
+    global float * yy,
+    int ib,
+    int ip,
+    int is,
+    int l0
+) {
+    int y_offset   = 128*ip + l0;
+    int q_offset_l =  64*ip + l0;
+    int q_offset_h =  32*ip + l0;
+
+    global uchar * q1 = blk_ql     + ib*128 + q_offset_l;
+    global uchar * q2 = q1         + QK_K/8;
+    global uchar * qh = blk_qh     + ib*64 + q_offset_h;
+    global char  * sc = blk_scales + ib*16 + is;
+
+    global float * y = yy + ib * QK_K + y_offset;
+
+    float dall = blk_d[ib];
+
+    float  sumf = 0;
+    float4 sums = {0.f, 0.f, 0.f, 0.f};
+
+    sums.s0 += y[0+ 0] * ((float)((q1[0] & 0xF) | ((qh[0] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[0+32] * ((float)((q2[0] & 0xF) | ((qh[0] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[0+64] * ((float)((q1[0]  >> 4) | ((qh[0] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[0+96] * ((float)((q2[0]  >> 4) | ((qh[0] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sums.s0 += y[1+ 0] * ((float)((q1[1] & 0xF) | ((qh[1] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[1+32] * ((float)((q2[1] & 0xF) | ((qh[1] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[1+64] * ((float)((q1[1]  >> 4) | ((qh[1] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[1+96] * ((float)((q2[1]  >> 4) | ((qh[1] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sums.s0 += y[2+ 0] * ((float)((q1[2] & 0xF) | ((qh[2] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[2+32] * ((float)((q2[2] & 0xF) | ((qh[2] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[2+64] * ((float)((q1[2]  >> 4) | ((qh[2] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[2+96] * ((float)((q2[2]  >> 4) | ((qh[2] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sums.s0 += y[3+ 0] * ((float)((q1[3] & 0xF) | ((qh[3] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[3+32] * ((float)((q2[3] & 0xF) | ((qh[3] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[3+64] * ((float)((q1[3]  >> 4) | ((qh[3] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[3+96] * ((float)((q2[3]  >> 4) | ((qh[3] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sumf += dall * (sums.s0 * sc[0] + sums.s1 * sc[2] + sums.s2 * sc[4] + sums.s3 * sc[6]);
+
+    return sumf;
+}
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4
+#define N_SIMDGROUP 2
+#define N_SIMDWIDTH 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 2
+#define N_SIMDWIDTH 64
+#endif
+
+#define BLOCK_STRIDE (N_SIMDWIDTH/16) // number of blocks each subgroup processes
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q6_K_f32_flat(
+        global uchar * src0_ql,
+        global uchar * src0_qh,
+        global char  * src0_s,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    int first_row = (N_SIMDGROUP * r0 + get_sub_group_id()) * N_DST;
+
+    ulong offset_src0    = first_row*nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    ulong offset_src0_ql = offset_src0 * 128;
+    ulong offset_src0_qh = offset_src0 * 64;
+    ulong offset_src0_s  = offset_src0 * 16;
+    ulong offset_src0_d  = offset_src0;
+
+    global uchar * blk_ql     = (global uchar *) src0_ql + offset_src0_ql;
+    global uchar * blk_qh     = (global uchar *) src0_qh + offset_src0_qh;
+    global char  * blk_scales = (global char  *) src0_s  + offset_src0_s;
+    global half  * blk_d      = (global half  *) src0_d  + offset_src0_d;
+    global float * yy         = (global float *) src1    + r1*ne10 + im*ne00*ne1;
+
+    int tid = get_sub_group_local_id()/BLOCK_STRIDE; // first block_stride groups have tid=0
+    int ix  = get_sub_group_local_id()%BLOCK_STRIDE; // first block is 0..block_stride-1
+    int ip  = tid/8;   // first or second half of (super) block (0 or 1)
+    int il  = tid%8;   // each half has 8 parts, one per scale
+    int n   = 4;       // 4 scales at a time (and 4 sums)
+    int l0  = n*il;    // offset into half-block, 0..28
+    int is  = 8*ip + l0/16; // 0, 1, 8, 9
+
+    float4 sumf = 0;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        if (first_row + 0 < ne01) {
+            sumf.s0 += block_q_6_K_dot_y_flat(blk_ql + 0*nb*128, blk_qh + 0*nb*64, blk_scales + 0*nb*16, blk_d + 0*nb, yy, ib, ip, is, l0);
+        }
+        if (first_row + 1 < ne01) {
+            sumf.s1 += block_q_6_K_dot_y_flat(blk_ql + 1*nb*128, blk_qh + 1*nb*64, blk_scales + 1*nb*16, blk_d + 1*nb, yy, ib, ip, is, l0);
+        }
+        if (first_row + 2 < ne01) {
+            sumf.s2 += block_q_6_K_dot_y_flat(blk_ql + 2*nb*128, blk_qh + 2*nb*64, blk_scales + 2*nb*16, blk_d + 2*nb, yy, ib, ip, is, l0);
+        }
+        if (first_row + 3 < ne01) {
+            sumf.s3 += block_q_6_K_dot_y_flat(blk_ql + 3*nb*128, blk_qh + 3*nb*64, blk_scales + 3*nb*16, blk_d + 3*nb, yy, ib, ip, is, l0);
+        }
+    }
+
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0),
+        sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2),
+        sub_group_reduce_add(sumf.s3)
+    );
+    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;
+        }
+    }
+}