opencl: generalize Adreno MoE kernels on M (#23449)

ggml/src/ggml-opencl/kernels/cvt.cl

@@ -220,6 +220,10 @@ kernel void kernel_convert_block_q4_0_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK4_0;
     uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
     uint dst_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
@@ -263,6 +267,10 @@ kernel void kernel_restore_block_q4_0_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK4_0;
     uint dst_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
     uint src_d_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
@@ -401,6 +409,10 @@ kernel void kernel_convert_block_q4_1_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK4_1;
     uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
     uint dst_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
@@ -446,6 +458,10 @@ kernel void kernel_restore_block_q4_1_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK4_1;
     uint dst_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
     uint src_dm_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
@@ -491,6 +507,10 @@ kernel void kernel_convert_block_q5_0_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK5_0;
     uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
     uint dst_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
@@ -536,6 +556,10 @@ kernel void kernel_restore_block_q5_0_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK5_0;
     uint dst_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
     uint src_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
@@ -583,6 +607,10 @@ kernel void kernel_convert_block_q5_1_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK5_1;
     uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
     uint dst_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
@@ -630,6 +658,10 @@ kernel void kernel_restore_block_q5_1_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK5_1;
     uint dst_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
     uint src_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
@@ -679,6 +711,10 @@ kernel void kernel_convert_block_q4_k_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK_K;
     uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
     uint dst_blk_offset = i01 + i00 * ne01     + i02 * ne00_blk * ne01;
@@ -732,6 +768,10 @@ kernel void kernel_restore_block_q4_k_trans4_ns(
     uint i01 = get_global_id(0);  // row index
     uint i02 = get_global_id(2);  // batch index
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK_K;
 
     uint src_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
@@ -784,6 +824,10 @@ kernel void kernel_convert_block_q5_k_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK_K;
     uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
     uint dst_blk_offset = i01 + i00 * ne01     + i02 * ne00_blk * ne01;
@@ -850,6 +894,10 @@ kernel void kernel_restore_block_q5_k_trans4_ns(
     uint i01 = get_global_id(0);  // row index
     uint i02 = get_global_id(2);  // batch index
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK_K;
 
     uint src_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
@@ -916,6 +964,10 @@ kernel void kernel_convert_block_q6_k_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK_K;
 
     uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
@@ -993,6 +1045,10 @@ kernel void kernel_restore_block_q6_k_trans4_ns(
     uint i01 = get_global_id(0);  // row index
     uint i02 = get_global_id(2);  // batch index
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK_K;
 
     uint src_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
@@ -1147,6 +1203,10 @@ kernel void kernel_convert_block_mxfp4_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     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;
@@ -1190,6 +1250,10 @@ kernel void kernel_restore_block_mxfp4_trans4_ns(
     uint i01 = get_global_id(0);
     uint i02 = get_global_id(2);
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint ne00_blk = ne00 / QK_MXFP4;
     uint dst_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
     uint src_d_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;

ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl

@@ -163,7 +163,7 @@ kernel void kernel_gemm_moe_mxfp4_f32_ns(
     uint block_id_n = get_global_id(2); // n_tile
 
     // Boundary check
-    if (((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) || (block_id_n >= total_tiles[0])) {
+    if (block_id_n >= total_tiles[0]) {
         return;
     }
 
@@ -248,6 +248,10 @@ kernel void kernel_gemm_moe_mxfp4_f32_ns(
         dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
     }
 
+    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
+        return;
+    }
+
     // Load poster router and share in LM
     __local uint out_idx[TILESIZE_N];
 

ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl

@@ -115,7 +115,7 @@ kernel void kernel_gemm_moe_q4_0_f32_ns(
     uint block_id_n = get_global_id(2); // n_tile
 
     // Boundary check
-    if (((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) || (block_id_n >= total_tiles[0])) {
+    if (block_id_n >= total_tiles[0]) {
         return;
     }
 
@@ -198,6 +198,10 @@ kernel void kernel_gemm_moe_q4_0_f32_ns(
         dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
     }
 
+    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
+        return;
+    }
+
     // Load poster router and share in LM
     __local uint out_idx[TILESIZE_N];
 

ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl

@@ -116,7 +116,7 @@ kernel void kernel_gemm_moe_q4_1_f32_ns(
     uint block_id_n = get_global_id(2); // n_tile
 
     // Boundary check
-    if (((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) || (block_id_n >= total_tiles[0])) {
+    if (block_id_n >= total_tiles[0]) {
         return;
     }
 
@@ -200,6 +200,10 @@ kernel void kernel_gemm_moe_q4_1_f32_ns(
         dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
     }
 
+    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
+        return;
+    }
+
     // Load poster router and share in LM
     __local uint out_idx[TILESIZE_N];
 

ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl

@@ -133,7 +133,7 @@ kernel void kernel_gemm_moe_q4_k_f32_ns(
     uint block_id_n = get_global_id(2); // n_tile
 
     // Boundary check
-    if (((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) || (block_id_n >= total_tiles[0])) {
+    if (block_id_n >= total_tiles[0]) {
         return;
     }
 
@@ -225,6 +225,10 @@ kernel void kernel_gemm_moe_q4_k_f32_ns(
         dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
     }
 
+    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
+        return;
+    }
+
     // Load post router and share in LM
     __local uint out_idx[TILESIZE_N];
 

ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl

@@ -116,7 +116,7 @@ kernel void kernel_gemm_moe_q5_0_f32_ns(
     uint block_id_n = get_global_id(2); // n_tile
 
     // Boundary check
-    if (((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) || (block_id_n >= total_tiles[0])) {
+    if (block_id_n >= total_tiles[0]) {
         return;
     }
 
@@ -202,6 +202,10 @@ kernel void kernel_gemm_moe_q5_0_f32_ns(
         dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
     }
 
+    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
+        return;
+    }
+
     // Load poster router and share in LM
     __local uint out_idx[TILESIZE_N];
 

ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl

@@ -117,7 +117,7 @@ kernel void kernel_gemm_moe_q5_1_f32_ns(
     uint block_id_n = get_global_id(2); // n_tile
 
     // Boundary check
-    if (((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) || (block_id_n >= total_tiles[0])) {
+    if (block_id_n >= total_tiles[0]) {
         return;
     }
 
@@ -204,6 +204,10 @@ kernel void kernel_gemm_moe_q5_1_f32_ns(
         dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
     }
 
+    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
+        return;
+    }
+
     // Load poster router and share in LM
     __local uint out_idx[TILESIZE_N];
 

ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl

@@ -134,7 +134,7 @@ kernel void kernel_gemm_moe_q5_k_f32_ns(
     uint block_id_n = get_global_id(2); // n_tile
 
     // Boundary check
-    if (((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) || (block_id_n >= total_tiles[0])) {
+    if (block_id_n >= total_tiles[0]) {
         return;
     }
 
@@ -230,6 +230,10 @@ kernel void kernel_gemm_moe_q5_k_f32_ns(
         dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
     }
 
+    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
+        return;
+    }
+
     // Load post router and share in LM
     __local uint out_idx[TILESIZE_N];
 

ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl

@@ -117,7 +117,7 @@ kernel void kernel_gemm_moe_q6_k_f32_ns(
     uint block_id_n = get_global_id(2); // n_tile
 
     // Boundary check
-    if (((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) || (block_id_n >= total_tiles[0])) {
+    if (block_id_n >= total_tiles[0]) {
         return;
     }
 
@@ -209,6 +209,10 @@ kernel void kernel_gemm_moe_q6_k_f32_ns(
         dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
     }
 
+    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
+        return;
+    }
+
     // Load post router and share in LM
     __local uint out_idx[TILESIZE_N];
 

ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl

@@ -82,6 +82,10 @@ __kernel void kernel_gemv_moe_mxfp4_f32_ns(
     uint sgid = get_local_id(1);
     uint slid = get_sub_group_local_id();
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint i11 = i20 % ne11;
 
     uint expert_id = src2[i20];

ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl

@@ -37,6 +37,10 @@ __kernel void kernel_gemv_moe_q4_0_f32_ns(
     uint sgid = get_local_id(1);
     uint slid = get_sub_group_local_id();
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint i11 = i20 % ne11;
 
     uint expert_id = src2[i20];

ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl

@@ -38,6 +38,10 @@ __kernel void kernel_gemv_moe_q4_1_f32_ns(
     uint sgid = get_local_id(1);
     uint slid = get_sub_group_local_id();
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint i11 = i20 % ne11;
 
     uint expert_id = src2[i20];

ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl

@@ -54,6 +54,10 @@ __kernel void kernel_gemv_moe_q4_k_f32_ns(
     uint sgid = get_local_id(1);
     uint slid = get_sub_group_local_id();
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint i11 = i20 % ne11;
 
     uint expert_id = src2[i20];

ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl

@@ -38,6 +38,10 @@ __kernel void kernel_gemv_moe_q5_0_f32_ns(
     uint sgid = get_local_id(1);
     uint slid = get_sub_group_local_id();
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint i11 = i20 % ne11;
 
     uint expert_id = src2[i20];

ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl

@@ -39,6 +39,10 @@ __kernel void kernel_gemv_moe_q5_1_f32_ns(
     uint sgid = get_local_id(1);
     uint slid = get_sub_group_local_id();
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint i11 = i20 % ne11;
 
     uint expert_id = src2[i20];

ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl

@@ -55,6 +55,10 @@ __kernel void kernel_gemv_moe_q5_k_f32_ns(
     uint sgid = get_local_id(1);
     uint slid = get_sub_group_local_id();
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint i11 = i20 % ne11;
 
     uint expert_id = src2[i20];

ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl

@@ -38,6 +38,10 @@ __kernel void kernel_gemv_moe_q6_k_f32_ns(
     uint sgid = get_local_id(1);
     uint slid = get_sub_group_local_id();
 
+    if (i01 >= ne01) {
+        return;
+    }
+
     uint i11 = i20 % ne11;
 
     uint expert_id = src2[i20];