whispercpp 1.3.4 → 1.3.5

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

@@ -117,6 +117,27 @@ kernel void kernel_convert_block_q4_0_noshuffle(
     }
 }
 
+kernel void kernel_restore_block_q4_0_noshuffle(
+    global uchar * src_q,
+    global half  * src_d,
+    global struct block_q4_0 * dst,
+    uchar mask_0F,
+    uchar mask_F0
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) src_d + get_global_id(0);
+
+    b->d = *d;
+    for (int i = 0; i < QK4_0/4; ++i) {
+        uchar x0 = q[i + 0      ] ;
+        uchar x1 = q[i + QK4_0/4];
+
+        b->qs[2*i + 0] = convert_uchar((x0 & mask_0F) | ((x1 & mask_0F) << 4));
+        b->qs[2*i + 1] = convert_uchar(((x0 & mask_F0) >> 4) | (x1 & mask_F0));
+    }
+}
+
 //------------------------------------------------------------------------------
 // block_mxfp4
 //------------------------------------------------------------------------------
@@ -147,6 +168,27 @@ kernel void kernel_convert_block_mxfp4(
     }
 }
 
+kernel void kernel_convert_block_mxfp4_trans(
+    global struct block_mxfp4 * src0,
+    __global uint4 * dst_q,
+    __global uchar * dst_e,
+    uint ne00,
+    uint ne01
+) {
+    int i00 = get_global_id(1);
+    uint i01 = get_global_id(0);
+    uint i02 = get_global_id(2);
+
+    uint ne00_blk = ne00 / QK_MXFP4;
+    uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
+    uint dst_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
+
+    global struct block_mxfp4 * b = src0 + src_blk_offset;
+
+    dst_q[dst_blk_offset] = ((global uint4 *)(&(b->qs[0])))[0];
+    dst_e[dst_blk_offset] = b->e;
+}
+
 kernel void kernel_restore_block_mxfp4(
     global uchar * src_q,
     global half  * src_e,
@@ -162,6 +204,27 @@ kernel void kernel_restore_block_mxfp4(
     }
 }
 
+kernel void kernel_restore_block_mxfp4_trans(
+    __global uint4 * src_q,
+    __global uchar * src_e,
+    global struct block_mxfp4 * dst,
+    uint ne00,
+    uint ne01
+) {
+    int i00 = get_global_id(1);
+    uint i01 = get_global_id(0);
+    uint i02 = get_global_id(2);
+
+    uint ne00_blk = ne00 / QK_MXFP4;
+    uint src_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
+    uint dst_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
+
+    global struct block_mxfp4 * b = dst + dst_blk_offset;
+
+    ((global uint4 *)(&(b->qs[0])))[0] = src_q[src_blk_offset];
+    b->e = src_e[src_blk_offset];
+}
+
 //------------------------------------------------------------------------------
 // block_q8_0
 //------------------------------------------------------------------------------

data/ext/sources/ggml/src/ggml-opencl/kernels/expm1.cl

@@ -0,0 +1,82 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// expm1
+//------------------------------------------------------------------------------
+kernel void kernel_expm1_f32_nd(
+        global void * p_src0_base,
+        ulong off_src0_abs,
+        global void * p_dst_base,
+        ulong off_dst_abs,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const float *src_val_ptr = (global const float *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global float *dst_val_ptr = (global float *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = exp(*src_val_ptr) - 1;
+        }
+    }
+}
+
+kernel void kernel_expm1_f16_nd(
+        global void * p_src0_base,
+        ulong off_src0_abs,
+        global void * p_dst_base,
+        ulong off_dst_abs,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const half *src_val_ptr = (global const half *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global half *dst_val_ptr = (global half *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = exp(*src_val_ptr) - 1;
+        }
+    }
+}

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

@@ -0,0 +1,17 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// fill
+//------------------------------------------------------------------------------
+__kernel void kernel_fill_f32(
+        __global float *dst,
+        ulong offsetd,
+        float v,
+        int n
+
+) {
+    dst = (global float*)((global char*)dst + offsetd);
+    if(get_global_id(0) < n){
+        dst[get_global_id(0)] = v;
+    }
+}

data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl

@@ -4,6 +4,7 @@
 #define ACC_TYPE4 float4
 #define DATA_TYPE float
 #define DATA_TYPE4 float4
+#define MASK_DATA_TYPE half
 #define CONVERT_ACC4(x) (x)
 #define CONVERT_DATA4(x) (x)
 
@@ -148,7 +149,7 @@ __kernel void flash_attn_f32(
             if (k_row1 >= n_kv) score1 = -INFINITY;
 
             if (mask_base != NULL) {
-                const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
                 if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
                 if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
             }
@@ -281,7 +282,7 @@ __kernel void flash_attn_f32_q1(
         }
         ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
         if (mask_base != NULL) {
-            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base);
             score += slope * (ACC_TYPE)mask_ptr[k_idx];
         }
         if (logit_softcap > 0.0f) {
@@ -317,7 +318,7 @@ __kernel void flash_attn_f32_q1(
         }
         ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
         if (mask_base != NULL) {
-            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base);
             score += slope * (ACC_TYPE)mask_ptr[k_idx];
         }
         if (logit_softcap > 0.0f) {

data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl

@@ -0,0 +1,162 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#define QK_MXFP4 32
+#define N_SIMDGROUP 2
+#define SIMDGROUP_WIDTH 64
+
+static inline half8 mxfp4_to_fp16_packed8(ushort2 fp4x8) { //, ushort 0x0E00, ushort 0x8000) {
+    ushort2 fp16_packed_a_0, fp16_packed_b_0, bias_a, bias_b, sign_a, sign_b;
+    fp16_packed_a_0.lo = (fp4x8.s0 << 9) & 0x0E00;
+    fp16_packed_a_0.hi = (fp4x8.s0 << 5) & 0x0E00;
+    fp16_packed_b_0.lo = (fp4x8.s0 << 1) & 0x0E00;
+    fp16_packed_b_0.hi = (fp4x8.s0 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a_0.lo != 0) ? 0x3800 : 0x0;
+    bias_a.hi = (fp16_packed_a_0.hi != 0) ? 0x3800 : 0x0;
+    bias_b.lo = (fp16_packed_b_0.lo != 0) ? 0x3800 : 0x0;
+    bias_b.hi = (fp16_packed_b_0.hi != 0) ? 0x3800 : 0x0;
+
+    fp16_packed_a_0.lo = (fp16_packed_a_0.lo != 0x0200) ? fp16_packed_a_0.lo : 0x0;
+    fp16_packed_a_0.hi = (fp16_packed_a_0.hi != 0x0200) ? fp16_packed_a_0.hi : 0x0;
+    fp16_packed_b_0.lo = (fp16_packed_b_0.lo != 0x0200) ? fp16_packed_b_0.lo : 0x0;
+    fp16_packed_b_0.hi = (fp16_packed_b_0.hi != 0x0200) ? fp16_packed_b_0.hi : 0x0;
+
+    sign_a.lo = (fp4x8.s0 << 12) & 0x8000;
+    sign_a.hi = (fp4x8.s0 << 8) & 0x8000;
+    sign_b.lo = (fp4x8.s0 << 4) & 0x8000;
+    sign_b.hi = fp4x8.s0 & 0x8000;
+
+    fp16_packed_a_0 = sign_a + bias_a + fp16_packed_a_0;
+    fp16_packed_b_0 = sign_b + bias_b + fp16_packed_b_0;
+
+    ushort2 fp16_packed_a_1, fp16_packed_b_1;
+    fp16_packed_a_1.lo = (fp4x8.s1 << 9) & 0x0E00;
+    fp16_packed_a_1.hi = (fp4x8.s1 << 5) & 0x0E00;
+    fp16_packed_b_1.lo = (fp4x8.s1 << 1) & 0x0E00;
+    fp16_packed_b_1.hi = (fp4x8.s1 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a_1.lo != 0) ? 0x3800 : 0x0;
+    bias_a.hi = (fp16_packed_a_1.hi != 0) ? 0x3800 : 0x0;
+    bias_b.lo = (fp16_packed_b_1.lo != 0) ? 0x3800 : 0x0;
+    bias_b.hi = (fp16_packed_b_1.hi != 0) ? 0x3800 : 0x0;
+
+    fp16_packed_a_1.lo = (fp16_packed_a_1.lo != 0x0200) ? fp16_packed_a_1.lo : 0x0;
+    fp16_packed_a_1.hi = (fp16_packed_a_1.hi != 0x0200) ? fp16_packed_a_1.hi : 0x0;
+    fp16_packed_b_1.lo = (fp16_packed_b_1.lo != 0x0200) ? fp16_packed_b_1.lo : 0x0;
+    fp16_packed_b_1.hi = (fp16_packed_b_1.hi != 0x0200) ? fp16_packed_b_1.hi : 0x0;
+
+    sign_a.lo = (fp4x8.s1 << 12) & 0x8000;
+    sign_a.hi = (fp4x8.s1 << 8) & 0x8000;
+    sign_b.lo = (fp4x8.s1 << 4) & 0x8000;
+    sign_b.hi = fp4x8.s1 & 0x8000;
+
+    fp16_packed_a_1 = sign_a + bias_a + fp16_packed_a_1;
+    fp16_packed_b_1 = sign_b + bias_b + fp16_packed_b_1;
+
+    return as_half8((ushort8)(fp16_packed_a_0, fp16_packed_b_0, fp16_packed_a_1, fp16_packed_b_1));
+}
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+    bits = (x == 0) ? 0x00400000 : ((uint) x << 23);
+    return as_float(bits);
+}
+
+
+__attribute__((qcom_reqd_sub_group_size("half")))
+__kernel void kernel_gemm_moe_mxfp4_f32(
+    __global uint4 * src0_q,
+    __global uchar * src0_e,
+    __read_only image1d_buffer_t src1,
+    __global ushort4 * src2,
+    __global float * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    int           tile_size
+) {
+    uint i01  = get_global_id(0);
+    uint i20  = get_global_id(2);
+    uint sgid = get_local_id(1);
+    uint slid = get_sub_group_local_id();
+
+    ushort4 router = src2[i20];
+    ushort expert_id = router.x;
+    ushort i11 = router.y;
+    ushort i1 = router.z;
+    ushort tile_id = router.w;
+
+    if (tile_id * tile_size + i01 >= ne01) { // handle edge case when ne01 is not multiple of tile_size
+        return;
+    }
+
+    uint expert_offset = expert_id * ne00 * ne01 / 32;
+    uint tile_offset = expert_offset + tile_id * tile_size + i01;
+
+    __private float sum = 0.0f; // each thread calculate partial sum of one output
+
+    // loop along ne00 in block granularity, skip 4 blocks every iter
+    for (uint ib00 = sgid; ib00 < (ne00 / QK_MXFP4); ib00 += N_SIMDGROUP) {
+        // load one block of q
+        uint4 regQ = src0_q[tile_offset + ib00 * ne01];
+        // convert 8 fp4 to fp16
+        half8 fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s0));
+
+        uint offset = i11 * ne00 / 4 + ib00 * 8;
+        float4 shared_y4;
+        shared_y4 = read_imagef(src1, (offset + 0));
+        float4 acc = shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 4));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s1));
+
+        shared_y4 = read_imagef(src1, (offset + 1));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 5));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s2));
+
+        shared_y4 = read_imagef(src1, (offset + 2));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 6));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s3));
+
+        shared_y4 = read_imagef(src1, (offset + 3));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 7));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+        uchar regE = src0_e[tile_offset + ib00 * ne01];
+        sum += e8m0_to_fp32(regE) * ((acc.s0 + acc.s1) + (acc.s2 + acc.s3));
+    }
+
+    // reduction in local memory, assumes #subgroups=4
+    __local float reduceLM[SIMDGROUP_WIDTH * (N_SIMDGROUP - 1)];
+    if (sgid == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = sum;
+    // if (sgid == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = sum;
+    // if (sgid == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = sum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
+    // if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
+    // if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];
+
+    // 1 outputs per thread in subgroup 0
+    if (sgid == 0) {
+        dst = dst + (offsetd >> 2);
+        dst[i01 + tile_id * tile_size + i1 * ne01] = sum;
+    }
+
+}

data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl

@@ -0,0 +1,156 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#define QK_MXFP4 32
+#define N_SIMDGROUP 4
+#define SIMDGROUP_WIDTH 64
+
+static inline half8 mxfp4_to_fp16_packed8(ushort2 fp4x8) { //, ushort 0x0E00, ushort 0x8000) {
+    ushort2 fp16_packed_a_0, fp16_packed_b_0, bias_a, bias_b, sign_a, sign_b;
+    fp16_packed_a_0.lo = (fp4x8.s0 << 9) & 0x0E00;
+    fp16_packed_a_0.hi = (fp4x8.s0 << 5) & 0x0E00;
+    fp16_packed_b_0.lo = (fp4x8.s0 << 1) & 0x0E00;
+    fp16_packed_b_0.hi = (fp4x8.s0 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a_0.lo != 0) ? 0x3800 : 0x0;
+    bias_a.hi = (fp16_packed_a_0.hi != 0) ? 0x3800 : 0x0;
+    bias_b.lo = (fp16_packed_b_0.lo != 0) ? 0x3800 : 0x0;
+    bias_b.hi = (fp16_packed_b_0.hi != 0) ? 0x3800 : 0x0;
+
+    fp16_packed_a_0.lo = (fp16_packed_a_0.lo != 0x0200) ? fp16_packed_a_0.lo : 0x0;
+    fp16_packed_a_0.hi = (fp16_packed_a_0.hi != 0x0200) ? fp16_packed_a_0.hi : 0x0;
+    fp16_packed_b_0.lo = (fp16_packed_b_0.lo != 0x0200) ? fp16_packed_b_0.lo : 0x0;
+    fp16_packed_b_0.hi = (fp16_packed_b_0.hi != 0x0200) ? fp16_packed_b_0.hi : 0x0;
+
+    sign_a.lo = (fp4x8.s0 << 12) & 0x8000;
+    sign_a.hi = (fp4x8.s0 << 8) & 0x8000;
+    sign_b.lo = (fp4x8.s0 << 4) & 0x8000;
+    sign_b.hi = fp4x8.s0 & 0x8000;
+
+    fp16_packed_a_0 = sign_a + bias_a + fp16_packed_a_0;
+    fp16_packed_b_0 = sign_b + bias_b + fp16_packed_b_0;
+
+    ushort2 fp16_packed_a_1, fp16_packed_b_1;
+    fp16_packed_a_1.lo = (fp4x8.s1 << 9) & 0x0E00;
+    fp16_packed_a_1.hi = (fp4x8.s1 << 5) & 0x0E00;
+    fp16_packed_b_1.lo = (fp4x8.s1 << 1) & 0x0E00;
+    fp16_packed_b_1.hi = (fp4x8.s1 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a_1.lo != 0) ? 0x3800 : 0x0;
+    bias_a.hi = (fp16_packed_a_1.hi != 0) ? 0x3800 : 0x0;
+    bias_b.lo = (fp16_packed_b_1.lo != 0) ? 0x3800 : 0x0;
+    bias_b.hi = (fp16_packed_b_1.hi != 0) ? 0x3800 : 0x0;
+
+    fp16_packed_a_1.lo = (fp16_packed_a_1.lo != 0x0200) ? fp16_packed_a_1.lo : 0x0;
+    fp16_packed_a_1.hi = (fp16_packed_a_1.hi != 0x0200) ? fp16_packed_a_1.hi : 0x0;
+    fp16_packed_b_1.lo = (fp16_packed_b_1.lo != 0x0200) ? fp16_packed_b_1.lo : 0x0;
+    fp16_packed_b_1.hi = (fp16_packed_b_1.hi != 0x0200) ? fp16_packed_b_1.hi : 0x0;
+
+    sign_a.lo = (fp4x8.s1 << 12) & 0x8000;
+    sign_a.hi = (fp4x8.s1 << 8) & 0x8000;
+    sign_b.lo = (fp4x8.s1 << 4) & 0x8000;
+    sign_b.hi = fp4x8.s1 & 0x8000;
+
+    fp16_packed_a_1 = sign_a + bias_a + fp16_packed_a_1;
+    fp16_packed_b_1 = sign_b + bias_b + fp16_packed_b_1;
+
+    return as_half8((ushort8)(fp16_packed_a_0, fp16_packed_b_0, fp16_packed_a_1, fp16_packed_b_1));
+}
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+    bits = (x == 0) ? 0x00400000 : ((uint) x << 23);
+    return as_float(bits);
+}
+
+
+__attribute__((qcom_reqd_sub_group_size("half")))
+__kernel void kernel_gemv_moe_mxfp4_f32(
+    __global uint4 * src0_q,
+    __global uchar * src0_e,
+    __read_only image1d_buffer_t src1,
+    __global uint * src2,
+    __global float * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    int           ne11
+) {
+    uint i01  = get_global_id(0);
+    uint i20  = get_global_id(2);
+    uint sgid = get_local_id(1);
+    uint slid = get_sub_group_local_id();
+
+    uint i11 = i20 % ne11;
+
+    uint expert_id = src2[i20];
+    uint expert_offset = expert_id * ne00 * ne01 / 32;
+
+    __private float sum = 0.0f; // each thread calculate partial sum of one output
+
+    // loop along ne00 in block granularity, skip 4 blocks every iter
+    for (uint ib00 = sgid; ib00 < (ne00 / QK_MXFP4); ib00 += N_SIMDGROUP) {
+
+        // load one block of q
+        uint4 regQ = src0_q[expert_offset + ib00 * ne01 + i01];
+
+        uint offset = i11 * ne00 / 4 + ib00 * 8;
+
+        half8 fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s0));
+
+        float4 shared_y4;
+        shared_y4 = read_imagef(src1, (offset + 0));
+        float4 acc = shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 4));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s1));
+
+        shared_y4 = read_imagef(src1, (offset + 1));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 5));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s2));
+
+        shared_y4 = read_imagef(src1, (offset + 2));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 6));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s3));
+
+        shared_y4 = read_imagef(src1, (offset + 3));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 7));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+        uchar regE = src0_e[ib00 * ne01 + i01 + expert_offset];
+        sum += e8m0_to_fp32(regE) * ((acc.s0 + acc.s1) + (acc.s2 + acc.s3));
+    }
+
+    // reduction in local memory, assumes #subgroups=4
+    __local float reduceLM[SIMDGROUP_WIDTH * (N_SIMDGROUP - 1)];
+    if (sgid == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = sum;
+    if (sgid == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = sum;
+    if (sgid == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = sum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
+    if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
+    if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];
+
+    // 1 outputs per thread in subgroup 0
+    if (sgid == 0) {
+        dst = dst + (offsetd >> 2);
+        dst[i01 + i20 * ne01] = sum;
+    }
+
+}

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

@@ -69,11 +69,14 @@ kernel void kernel_get_rows_f32(
         int ne00,
         ulong nb01,
         ulong nb02,
+        ulong nb03,
         int ne10,
         ulong nb10,
         ulong nb11,
+        ulong nb12,
         ulong nb1,
-        ulong nb2
+        ulong nb2,
+        ulong nb3
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -81,14 +84,19 @@ kernel void kernel_get_rows_f32(
 
     int i10 = get_group_id(0);
     int i11 = get_group_id(1);
+    int i12 = get_group_id(2);
 
-    int r = ((global int *) ((global char *) src1 + i11*nb11 + i10*nb10))[0];
+    int r = ((global int *) ((global char *) src1 + i12*nb12 + i11*nb11 + i10*nb10))[0];
 
     int i02 = i11;
+    int i03 = i12;
 
     for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) {
-        ((global float *) ((global char *) dst + i11*nb2 + i10*nb1))[ind] =
-            ((global float *) ((global char *) src0 + r*nb01 + i02*nb02))[ind];
+        if (ind >= ne00) {
+            return;
+        }
+        ((global float *) ((global char *) dst + i12*nb3 + i11*nb2 + i10*nb1))[ind] =
+            ((global float *) ((global char *) src0 + r*nb01 + i02*nb02 + i03*nb03))[ind];
     }
 }
 
@@ -102,11 +110,14 @@ kernel void kernel_get_rows_f16(
         int ne00,
         ulong nb01,
         ulong nb02,
+        ulong nb03,
         int ne10,
         ulong nb10,
         ulong nb11,
+        ulong nb12,
         ulong nb1,
-        ulong nb2
+        ulong nb2,
+        ulong nb3
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -114,14 +125,19 @@ kernel void kernel_get_rows_f16(
 
     int i10 = get_group_id(0);
     int i11 = get_group_id(1);
+    int i12 = get_group_id(2);
 
-    int r = ((global int32_t *) ((global char *) src1 + i11*nb11 + i10*nb10))[0];
+    int r = ((global int32_t *) ((global char *) src1 + i12*nb12 + i11*nb11 + i10*nb10))[0];
 
     int i02 = i11;
+    int i03 = i12;
 
     for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) {
-        ((global float *) ((global char *) dst + i11*nb2 + i10*nb1))[ind] =
-            ((global half *) ((global char *) src0 + r*nb01 + i02*nb02))[ind];
+        if (ind >= ne00) {
+            return;
+        }
+        ((global float *) ((global char *) dst + i12*nb3 + i11*nb2 + i10*nb1))[ind] =
+            ((global half *) ((global char *) src0 + r*nb01 + i02*nb02 + i03*nb03))[ind];
     }
 }
 
@@ -135,11 +151,14 @@ kernel void kernel_get_rows_q4_0(
         int ne00,
         ulong nb01,
         ulong nb02,
+        ulong nb03,
         int ne10,
         ulong nb10,
         ulong nb11,
+        ulong nb12,
         ulong nb1,
-        ulong nb2
+        ulong nb2,
+        ulong nb3
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -149,15 +168,20 @@ kernel void kernel_get_rows_q4_0(
 
     int i10 = get_group_id(0);
     int i11 = get_group_id(1);
+    int i12 = get_group_id(2);
 
-    int r = ((global int32_t *) ((global char *) src1 + i11*nb11 + i10*nb10))[0];
+    int r = ((global int32_t *) ((global char *) src1 + i12*nb12 + i11*nb11 + i10*nb10))[0];
 
     int i02 = i11;
+    int i03 = i12;
 
     for (int ind = get_local_id(0); ind < ne00/16; ind += get_local_size(0)) {
         float16 temp;
+        if (ind >= ne00) {
+            return;
+        }
         dequantize_q4_0_f32(
-            ((global struct block_q4_0 *) ((global char *) src0 + r*nb01 + i02*nb02)) + ind/NL, ind%NL, &temp);
-        *(((global float16 *) ((global char *) dst + i11*nb2 + i10*nb1)) + ind) = temp;
+            ((global struct block_q4_0 *) ((global char *) src0 + r*nb01 + i02*nb02 + i03*nb03)) + ind/NL, ind%NL, &temp);
+        *(((global float16 *) ((global char *) dst + i12*nb3 + i11*nb2 + i10*nb1)) + ind) = temp;
     }
 }

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

@@ -0,0 +1,39 @@
+
+kernel void kernel_mean_f32(
+    global float *  src0,
+    ulong           offset0,
+    global float *  dst,
+    ulong           offsetd,
+    int             ne00,
+    int             ne01,
+    int             ne02,
+    int             ne03,
+    ulong           nb01,
+    ulong           nb02,
+    ulong           nb03,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    int i3 = get_global_id(2);
+    int i2 = get_global_id(1);
+    int i1 = get_global_id(0);
+
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    global float * src_row = (global float *) ((global char *) src0 + i1*nb01 + i2*nb02 + i3*nb03);
+    global float * dst_row = (global float *) ((global char *) dst  + i1*nb1  + i2*nb2  + i3*nb3);
+
+    float row_sum = 0;
+
+    for (int i0 = 0; i0 < ne00; i0++) {
+        row_sum += src_row[i0];
+    }
+
+    dst_row[0] = row_sum / ne00;
+}