whispercpp 1.3.3 → 1.3.4

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

@@ -77,3 +77,76 @@ kernel void kernel_mul_row(
     uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
     dst[gid] = src0[gid] * src1[idx1];
 }
+
+kernel void kernel_mul_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        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 ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) * *((global half *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_mul_row_f16(
+        global half4 * src0,
+        ulong offset0,
+        global half4 * src1,
+        ulong offset1,
+        global half4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    src1 = (global half4*)((global char*)src1 + offset1);
+    dst = (global half4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] * src1[idx1];
+}

data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl

@@ -0,0 +1,130 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+#define REQD_SUBGROUP_SIZE_128
+#endif
+
+#define OPWM 64
+#define OPWN 64
+#define CPWK 8
+#define OPTM 4
+#define OPTN 8
+
+#define WG_M (OPWM / OPTM)
+#define WG_N (OPWN / OPTN)
+#define VEC_K (CPWK / 4)
+
+REQD_SUBGROUP_SIZE_128
+__kernel void mul_mat_f16_f32(
+    const int M, const int N, const int K,
+    __global const void* A_void, ulong A_offset,
+    __global const void* B_void, ulong B_offset,
+    __global       void* C_void, ulong C_offset) {
+
+    __global const half*  A = (__global const half* )((__global const char*)A_void + A_offset);
+    __global const float* B = (__global const float*)((__global const char*)B_void + B_offset);
+    __global       float* C = (__global       float*)((__global       char*)C_void + C_offset);
+
+    const int lidm = get_local_id(0);
+    const int lidn = get_local_id(1);
+    const int lid = lidn * WG_M + lidm;
+
+    const int offsetM = get_group_id(0) * OPWM;
+    const int offsetN = get_group_id(1) * OPWN;
+
+    __local half4  Alocal[OPWM][VEC_K];
+    __local float4 Blocal[OPWN][VEC_K];
+
+    float sum[OPTM][OPTN];
+
+    for (int wm = 0; wm < OPTM; wm++) {
+        for (int wn = 0; wn < OPTN; wn++) {
+            sum[wm][wn] = 0.0f;
+        }
+    }
+
+    const int numTiles = (K + CPWK - 1) / CPWK;
+
+    const int load_row_a = lid % OPWM;
+    const int load_vec_k_a = lid / OPWM;
+    const int global_row_a = offsetM + load_row_a;
+
+    const int load_row_b = lid % OPWN;
+    const int load_vec_k_b = lid / OPWN;
+    const int global_row_b = offsetN + load_row_b;
+
+    for (int t = 0; t < numTiles; t++) {
+        const int k_start = t * CPWK;
+        const int k_vec_start_a = k_start + load_vec_k_a * 4;
+        const int k_vec_start_b = k_start + load_vec_k_b * 4;
+
+        if (global_row_a < M && k_vec_start_a < K) {
+            if (k_vec_start_a + 3 < K) {
+                Alocal[load_row_a][load_vec_k_a] = vload4(0, A + global_row_a * K + k_vec_start_a);
+            } else {
+                half4 tempA = (half4)(0.0h);
+                if (k_vec_start_a < K) tempA.s0 = A[global_row_a * K + k_vec_start_a];
+                if (k_vec_start_a + 1 < K) tempA.s1 = A[global_row_a * K + k_vec_start_a + 1];
+                if (k_vec_start_a + 2 < K) tempA.s2 = A[global_row_a * K + k_vec_start_a + 2];
+                Alocal[load_row_a][load_vec_k_a] = tempA;
+            }
+        } else {
+            Alocal[load_row_a][load_vec_k_a] = (half4)(0.0h);
+        }
+
+        if (global_row_b < N && k_vec_start_b < K) {
+            if (k_vec_start_b + 3 < K) {
+                Blocal[load_row_b][load_vec_k_b] = vload4(0, B + global_row_b * K + k_vec_start_b);
+            } else {
+                float4 tempB = (float4)(0.0f);
+                if (k_vec_start_b < K) tempB.s0 = B[global_row_b * K + k_vec_start_b];
+                if (k_vec_start_b + 1 < K) tempB.s1 = B[global_row_b * K + k_vec_start_b + 1];
+                if (k_vec_start_b + 2 < K) tempB.s2 = B[global_row_b * K + k_vec_start_b + 2];
+                Blocal[load_row_b][load_vec_k_b] = tempB;
+            }
+        } else {
+            Blocal[load_row_b][load_vec_k_b] = (float4)(0.0f);
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        #pragma unroll
+        for (int k_vec = 0; k_vec < VEC_K; k_vec++) {
+            float4 a_fvecs[OPTM];
+            int current_row_a = lidm;
+            for (int wm = 0; wm < OPTM; wm++) {
+                a_fvecs[wm] = convert_float4(Alocal[current_row_a][k_vec]);
+                current_row_a += WG_M;
+            }
+
+            float4 b_fvecs[OPTN];
+            int current_row_b = lidn;
+            for (int wn = 0; wn < OPTN; wn++) {
+                b_fvecs[wn] = Blocal[current_row_b][k_vec];
+                current_row_b += WG_N;
+            }
+
+            for (int wm = 0; wm < OPTM; wm++) {
+                for (int wn = 0; wn < OPTN; wn++) {
+                    sum[wm][wn] += dot(a_fvecs[wm], b_fvecs[wn]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    for (int wm = 0; wm < OPTM; wm++) {
+        int globalRow = offsetM + lidm + wm * WG_M;
+        if (globalRow < M) {
+            for (int wn = 0; wn < OPTN; wn++) {
+                int globalCol = offsetN + lidn + wn * WG_N;
+                if (globalCol < N) {
+                    C[globalCol * M + globalRow] = sum[wm][wn];
+                }
+            }
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl

@@ -0,0 +1,132 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 4
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 16
+#define TM 4
+#define TN 8
+
+kernel void kernel_mul_mm_f16_f32_l4_lm(
+    global half4 * src0,
+    ulong offset0,
+    global float4 * src1,
+    ulong offset1,
+    global float * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    local half  buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    half  cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+            buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = src0[idx].s0;
+            buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = src0[idx].s1;
+            buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = src0[idx].s2;
+            buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = src0[idx].s3;
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            const int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+            buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+            buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+            buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+            buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(convert_float(cache_a[cr]), cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl

@@ -0,0 +1,133 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 4
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 16
+#define TM 4
+#define TN 8
+
+kernel void kernel_mul_mm_f32_f32_l4_lm(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * src1,
+    ulong offset1,
+    global float * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+            buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = src0[idx].s0;
+            buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = src0[idx].s1;
+            buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = src0[idx].s2;
+            buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = src0[idx].s3;
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            const int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+            buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+            buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+            buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+            buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}

data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl

@@ -0,0 +1,189 @@
+#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
+typedef struct {
+    uchar e; // E8M0
+    uchar qs[QK_MXFP4/2];
+} block_mxfp4;
+
+constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (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 2
+#define N_SG_MXFP4 2
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_mv_mxfp4_f32(
+    global char * src0,
+    global char * src1,
+    global char * dst,
+    int ne00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne12,
+    ulong nb11,
+    ulong nb12,
+    ulong nb13,
+    int ne0,
+    int ne1,
+    int r2,
+    int r3,
+    local  char * shmem
+) {
+    local float * shmem_f32 = (local float *) shmem;
+    int nb = ne00/QK_MXFP4;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = 0;
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_mxfp4 * x = (global block_mxfp4 *) (src0 + offset_src0);
+    global float       * y = (global float       *) (src1 + offset_src1);
+
+    const short ix = get_sub_group_local_id()/2;  // 0...15
+    const short it = get_sub_group_local_id()%2;  // 0 or 1
+
+    shmem_f32[get_sub_group_local_id()] = kvalues_mxfp4_f[get_sub_group_local_id()%16];
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    float4 yl[4];
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    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;
+        yl[0] = y4[0];
+        yl[1] = y4[4];
+        yl[2] = y4[1];
+        yl[3] = y4[5];
+
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            global block_mxfp4 * xb = x + row*nb + ib;
+            global uchar       * q2 = (global uchar *)(xb->qs + 8*it);
+
+            float4 acc1 = yl[0]*(float4)(shmem_f32[q2[0] &  0x0F], shmem_f32[q2[1] &  0x0F], shmem_f32[q2[2] &  0x0F], shmem_f32[q2[3] &  0x0F]);
+            float4 acc2 = yl[1]*(float4)(shmem_f32[q2[0] >> 4   ], shmem_f32[q2[1] >> 4   ], shmem_f32[q2[2] >> 4   ], shmem_f32[q2[3] >> 4   ]);
+            float4 acc3 = yl[2]*(float4)(shmem_f32[q2[4] &  0x0F], shmem_f32[q2[5] &  0x0F], shmem_f32[q2[6] &  0x0F], shmem_f32[q2[7] &  0x0F]);
+            float4 acc4 = yl[3]*(float4)(shmem_f32[q2[4] >> 4   ], shmem_f32[q2[5] >> 4   ], shmem_f32[q2[6] >> 4   ], shmem_f32[q2[7] >> 4   ]);
+
+            acc1 = (acc1 + acc3) + (acc2 + acc4);
+
+            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)im*ne0*ne1 + (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;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_mxfp4_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * src2,
+    ulong         offset2,
+    global char * 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,
+    local  char * shmem
+) {
+    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);
+
+    const int iid1 = get_group_id(2)/ne20;
+    const 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);
+
+    mul_mv_mxfp4_f32(src0_cur, src1_cur, dst_cur,
+        ne00, nb01, nb02, nb03, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shmem);
+}