whispercpp 1.3.5 → 1.3.7

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

@@ -1,109 +1,118 @@
-kernel void kernel_concat_f32_contiguous(
-    global const char * p_src0, ulong off_src0,
-    global const char * p_src1, ulong off_src1,
-    global char * p_dst, ulong off_dst,
-    int d_ne00, int d_ne01, int d_ne02, // src0->ne[0..2] for the slice
-    int d_ne10, int d_ne11, int d_ne12, // src1->ne[0..2] for the slice (d_ne1X must match d_ne0X on non-concat axes)
-    int d_ne0,  int d_ne1,  int d_ne2,  // dst->ne[0..2] for the slice
-    int dim
+kernel void kernel_concat_f32(
+    global  const char * src0,
+    ulong                offset0,
+    global  const 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,
+    ulong           nb10,
+    ulong           nb11,
+    ulong           nb12,
+    ulong           nb13,
+    int             ne0,
+    ulong           nb0,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3,
+    int             dim
 ) {
-    global const float * src0 = (global const float*)((global char*)p_src0 + off_src0);
-    global const float * src1 = (global const float*)((global char*)p_src1 + off_src1);
-    global float * dst        = (global float*)((global char*)p_dst + off_dst);
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
 
-    int i0 = get_global_id(0); // Index along dst's 0th dimension
-    int i1 = get_global_id(1); // Index along dst's 1st dimension
-    int i2 = get_global_id(2); // Index along dst's 2nd dimension
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
 
-    if (i0 >= d_ne0 || i1 >= d_ne1 || i2 >= d_ne2) {
-        return;
-    }
+    int o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
 
-    ulong dst_idx = (ulong)i2 * d_ne0 * d_ne1 + (ulong)i1 * d_ne0 + i0;
-    ulong src_idx;
+    global const float * x;
 
-    if (dim == 0) {
-        if (i0 < d_ne00) { // Data from src0
-            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
-            dst[dst_idx] = src0[src_idx];
-        } else { // Data from src1
-            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + (i0 - d_ne00);
-            dst[dst_idx] = src1[src_idx];
-        }
-    } else if (dim == 1) {
-        if (i1 < d_ne01) { // Data from src0
-            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
-            dst[dst_idx] = src0[src_idx];
-        } else { // Data from src1
-            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)(i1 - d_ne01) * d_ne10 + i0;
-            dst[dst_idx] = src1[src_idx];
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+            x = (global const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
+        } else {
+            x = (global const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
         }
-    } else if (dim == 2) {
-        if (i2 < d_ne02) { // Data from src0
-            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
-            dst[dst_idx] = src0[src_idx];
-        } else { // Data from src1
 
-            src_idx = (ulong)(i2 - d_ne02) * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + i0;
-            dst[dst_idx] = src1[src_idx];
-        }
+        global float * y = (global float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+        *y = *x;
     }
 }
 
-kernel void kernel_concat_f32_non_contiguous(
-    global const char * p_src0, ulong off_src0,
-    global const char * p_src1, ulong off_src1,
-    global char * p_dst, ulong off_dst,
-
-    long ne00, long ne01, long ne02, long ne03,
-    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
-
-    ulong nb10, ulong nb11, ulong nb12, ulong nb13, // Strides for src1
-
-    long d_ne0, long d_ne1, long d_ne2, long d_ne3,
-    ulong d_nb0, ulong d_nb1, ulong d_nb2, ulong d_nb3,
-    int dim
+kernel void kernel_concat_f32_pack(
+    global  const char * src0,
+    ulong                offset0,
+    global  const 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,
+    ulong           nb10,
+    ulong           nb11,
+    ulong           nb12,
+    ulong           nb13,
+    int             ne0,
+    ulong           nb0,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3,
+    int             dim,
+    int             ne1,
+    int             ne2,
+    int             ne3
 ) {
-    global const char * src0_base = p_src0 + off_src0;
-    global const char * src1_base = p_src1 + off_src1;
-    global char * dst_base        = p_dst + off_dst;
-
-    long current_i1 = get_global_id(0); // Index for dst_dim_1
-    long current_i2 = get_global_id(1); // Index for dst_dim_2
-    long current_i3 = get_global_id(2); // Index for dst_dim_3
-
-    if (current_i1 >= d_ne1 || current_i2 >= d_ne2 || current_i3 >= d_ne3) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int lsz = get_local_size(0);
+    int tpr = min(ne0, lsz);          // threads per row
+    int rpw = lsz / tpr;              // rows per workgroup
+    int lid = get_local_id(0);
+    int row = get_group_id(0)*rpw + lid / tpr;
+    int lane = lid - (lid / tpr) * tpr;
+
+    int nrows = ne1*ne2*ne3;
+    if (row >= nrows) {
         return;
     }
 
-    global const float * x_val_ptr;
-    global float * y_val_ptr;
+    int i1 = row % ne1;
+    int t  = row / ne1;
+    int i2 = t % ne2;
+    int i3 = t / ne2;
 
-    for (long current_i0 = 0; current_i0 < d_ne0; ++current_i0) {
-        bool use_src0;
-        long s_i0 = current_i0, s_i1 = current_i1, s_i2 = current_i2, s_i3 = current_i3;
+    int o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
 
-        if (dim == 0) {
-            use_src0 = (current_i0 < ne00);
-            if (!use_src0) { s_i0 = current_i0 - ne00; }
-        } else if (dim == 1) {
-            use_src0 = (current_i1 < ne01);
-            if (!use_src0) { s_i1 = current_i1 - ne01; }
-        } else if (dim == 2) {
-            use_src0 = (current_i2 < ne02);
-            if (!use_src0) { s_i2 = current_i2 - ne02; }
-        } else { // dim == 3
-            use_src0 = (current_i3 < ne03);
-            if (!use_src0) { s_i3 = current_i3 - ne03; }
-        }
-
-        if (use_src0) {
-            x_val_ptr = (global const float *)(src0_base + (ulong)s_i3*nb03 + (ulong)s_i2*nb02 + (ulong)s_i1*nb01 + (ulong)s_i0*nb00);
+    for (int i0 = lane; i0 < ne0; i0 += tpr) {
+        global const float * x;
+        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+            x = (global const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
         } else {
-            x_val_ptr = (global const float *)(src1_base + (ulong)s_i3*nb13 + (ulong)s_i2*nb12 + (ulong)s_i1*nb11 + (ulong)s_i0*nb10);
+            x = (global const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
         }
 
-        y_val_ptr = (global float *)(dst_base + (ulong)current_i3*d_nb3 + (ulong)current_i2*d_nb2 + (ulong)current_i1*d_nb1 + (ulong)current_i0*d_nb0);
-        *y_val_ptr = *x_val_ptr;
+        global float * y = (global float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+        *y = *x;
     }
 }

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

@@ -182,3 +182,107 @@ kernel void kernel_cpy_f32_f32(
         dst_data[i00] = src[0];
     }
 }
+
+kernel void kernel_cpy_f32_f32_pack(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int lsz = get_local_size(0);
+    int tpr = min(ne00, lsz);          // threads per row
+    int rpw = lsz / tpr;               // rows per workgroup
+    int lid = get_local_id(0);
+    int row = get_group_id(0)*rpw + lid / tpr;
+    int lane = lid - (lid / tpr) * tpr;
+
+    int nrows = ne01*ne02*ne03;
+    if (row >= nrows) {
+        return;
+    }
+
+    int i01 = row % ne01;
+    int t   = row / ne01;
+    int i02 = t % ne02;
+    int i03 = t / ne02;
+
+    // linear index of the first element of this row, unflattened over dst dims
+    long n  = (long)row * ne00;
+    int i3  = (int)(n / ((long)ne2*ne1*ne0));
+    long rm = n - (long)i3*ne2*ne1*ne0;
+    int i2  = (int)(rm / ((long)ne1*ne0));
+    rm     -= (long)i2*ne1*ne0;
+    int i1  = (int)(rm / ne0);
+    int i0  = (int)(rm - (long)i1*ne0);
+
+    global float * dst_data = (global float *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = lane; i00 < ne00; i00 += tpr) {
+        global const float * src = (global float *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_i32_i32(
+        global int * src0,
+        ulong offset0,
+        global int * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global int*)((global char*)src0 + offset0);
+    dst = (global int*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global int * dst_data = (global int *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const int * src = (global int *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        dst_data[i00] = src[0];
+    }
+}

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

@@ -0,0 +1,139 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#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
+
+// max workgroup size is usually 1024, this covers various subgroups sizes
+#define MAX_SUBGROUPS 128
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_32
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_cumsum_blk(
+        global char * src0,
+        ulong offset0,
+        global char * tmp,
+        global char * dst,
+        ulong offsetd,
+        int   ne00,
+        int   ne01,
+        int   ne02,
+        int   ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        uint net0,
+        uint net1,
+        uint net2
+) {
+    src0 = src0 + offset0;
+    dst  = dst + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    const int nth = get_local_size(0);
+    const int tid = get_local_id(0);
+
+    const uint sg_size = get_sub_group_size();
+    const uint sg_id = get_sub_group_id();
+    const uint sg_lid = get_sub_group_local_id();
+
+    const int ib = i1 / ne01;
+    const int i00 = ib * nth;
+    const int i01 = i1 % ne01;
+    const int i02 = i2;
+    const int i03 = i3;
+
+    global const float * src0_row = (global const float *)(src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    global       float * tmp_row  = (global float *)tmp + net0 * i01 + net0 * net1 * i02 + net0 * net1 * net2 * i03;
+    global       float * dst_row  = (global float *)dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    __local float partial[MAX_SUBGROUPS];
+
+    float v = 0.0f;
+    if (i00 + tid < ne00) {
+        v = src0_row[i00 + tid];
+    }
+
+    float s = sub_group_scan_inclusive_add(v);
+    if (sg_lid == sg_size - 1) {
+        partial[sg_id] = s;
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    // NB: subgroup size should be larger than number of subgroups
+    // assuming max workgroup size of 1024, subgroup size should be >= 32
+    if (sg_id == 0) {
+        float x = 0.0f;
+        if (sg_lid < get_num_sub_groups()) {
+            x = partial[sg_lid];
+        }
+        float ex = sub_group_scan_exclusive_add(x);
+        if (sg_lid < get_num_sub_groups()) {
+            partial[sg_lid] = ex;
+        }
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    s += partial[sg_id];
+
+    if (i00 + tid < ne00) {
+        dst_row[i00 + tid] = s;
+    }
+    if (ne00 > nth && tid == nth - 1) {
+        tmp_row[ib] = s;
+    }
+}
+
+kernel void kernel_cumsum_add(
+        global char * tmp,
+        global char * dst,
+        ulong offsetd,
+        int   ne00,
+        int   ne01,
+        int   ne02,
+        int   ne03,
+        uint nbt0,
+        uint nbt1,
+        uint nbt2,
+        uint nbt3
+) {
+    dst  = dst + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    const int nth = get_local_size(0);
+    const int tid = get_local_id(0);
+
+    const int ib = i1 / ne01;
+    if (ib == 0) {
+        return;
+    }
+    const int i00 = ib * nth;
+    const int i01 = i1 % ne01;
+    const int i02 = i2;
+    const int i03 = i3;
+
+    global float * tmp_row  = (global float *)(tmp + nbt1 * i01 + nbt2 * i02 + nbt3 * i03);
+    global float * dst_row  = (global float *)dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    if (i00 + tid < ne00) {
+        dst_row[i00 + tid] += tmp_row[ib - 1];
+    }
+}