ggml : cleanup

Author: ggerganov

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -1009,6 +1009,64 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort_merge(ggml_metal_l
     return res;
 }
 
+// note: reuse the argsort kernel for top_k
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_TOP_K);
+
+    char base[256];
+    char name[256];
+
+    // note: the top_k kernel is always descending order
+    ggml_sort_order order = GGML_SORT_ORDER_DESC;
+
+    const char * order_str = "undefined";
+    switch (order) {
+        case GGML_SORT_ORDER_ASC:  order_str = "asc";  break;
+        case GGML_SORT_ORDER_DESC: order_str = "desc"; break;
+        default: GGML_ABORT("fatal error");
+    };
+
+    snprintf(base, 256, "kernel_argsort_%s_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->type), order_str);
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k_merge(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_TOP_K);
+
+    char base[256];
+    char name[256];
+
+    ggml_sort_order order = GGML_SORT_ORDER_DESC;
+
+    const char * order_str = "undefined";
+    switch (order) {
+        case GGML_SORT_ORDER_ASC:  order_str = "asc";  break;
+        case GGML_SORT_ORDER_DESC: order_str = "desc"; break;
+        default: GGML_ABORT("fatal error");
+    };
+
+    snprintf(base, 256, "kernel_argsort_merge_%s_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->type), order_str);
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_pad(
         ggml_metal_library_t lib,
         const struct ggml_tensor * op,

ggml/src/ggml-metal/ggml-metal-device.h

@@ -128,6 +128,8 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_id         (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argmax            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort_merge     (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k             (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k_merge       (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, enum ggml_op op, int32_t n_fuse, bool row);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_l2_norm           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_group_norm        (ggml_metal_library_t lib, const struct ggml_tensor * op);

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -3764,18 +3764,17 @@ int ggml_metal_op_top_k(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
     GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_argsort(lib, op);
+    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_top_k(lib, op);
 
     // bitonic sort requires the number of elements to be power of 2
     int nth = 1;
     while (nth < ne00 && 2*nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
         nth *= 2;
     }
 
+    // blocks per row
     const int npr = (ne00 + nth - 1)/nth;
 
-    // Metal kernels require the buffer size to be multiple of 16 bytes
-    // https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/1443142-setthreadgroupmemorylength
     const size_t smem = GGML_PAD(nth*sizeof(int32_t), 16);
 
     ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
@@ -3803,7 +3802,7 @@ int ggml_metal_op_top_k(ggml_metal_op_t ctx, int idx) {
         /*.ne1   =*/ ne1,
         /*.ne2   =*/ ne2,
         /*.ne3   =*/ ne3,
-        /*.top_k =*/ std::min(nth, top_k),
+        /*.top_k =*/ std::min(nth, top_k), // for each block, keep just the top_k indices
     };
 
     if (npr > 1) {
@@ -3819,13 +3818,18 @@ int ggml_metal_op_top_k(ggml_metal_op_t ctx, int idx) {
 
     ggml_metal_encoder_dispatch_threadgroups(enc, npr*ne01, ne02, ne03, nth, 1, 1);
 
-    ggml_metal_pipeline_t pipeline_merge = ggml_metal_library_get_pipeline_argsort_merge(lib, op);
+    ggml_metal_pipeline_t pipeline_merge = ggml_metal_library_get_pipeline_top_k_merge(lib, op);
 
     int len = args.top_k;
 
     while (len < args.ne0) {
         ggml_metal_op_concurrency_reset(ctx);
 
+        // merges per row
+        const int nm = (args.ne0 + 2*len - 1) / (2*len);
+
+        const int nth = std::min(512, std::min(len, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_merge)));
+
         ggml_metal_kargs_argsort_merge args_merge = {
             /*.ne00  =*/ ne00,
             /*.ne01  =*/ ne01,
@@ -3839,15 +3843,10 @@ int ggml_metal_op_top_k(ggml_metal_op_t ctx, int idx) {
             /*.ne1   =*/ ne1,
             /*.ne2   =*/ ne2,
             /*.ne3   =*/ ne3,
-            /*.top_k =*/ 2*len >= args.ne0 ? top_k : args.ne0,
+            /*.top_k =*/ nm == 1 ? top_k : args.ne0, // the final merge outputs top_k elements
             /*.len   =*/ len,
         };
 
-        // merges per row
-        const int nm = (args.ne0 + 2*len - 1) / (2*len);
-
-        const int nth = std::min(512, std::min(len, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_merge)));
-
         ggml_metal_encoder_set_pipeline(enc, pipeline_merge);
         ggml_metal_encoder_set_bytes   (enc, &args_merge, sizeof(args_merge), 0);
         ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);

ggml/src/ggml-metal/ggml-metal.metal

@@ -4779,7 +4779,11 @@ kernel void kernel_argsort_merge_f32_i32(
     const int chunk = (total + ntg.x - 1) / ntg.x;
 
     const int k0 = tpitg.x * chunk;
-    const int k1 = min(k0 + chunk, total);
+    const int k1 = MIN(MIN(k0 + chunk, total), args.top_k);
+
+    if (k0 >= args.top_k) {
+        return;
+    }
 
     if (k0 >= total) {
         return;
@@ -4830,16 +4834,16 @@ kernel void kernel_argsort_merge_f32_i32(
         val1 = src0_row[idx1];
     }
 
-    for (int k = k0; k < k1 && k < args.top_k; ++k) {
+    for (int k = k0; k < k1; ++k) {
         int32_t out_idx;
 
         if (i >= len0) {
-            while (k < k1 && k < args.top_k) {
+            while (k < k1) {
                 dst[k++] = tmp1[j++];
             }
             break;
         } else if (j >= len1) {
-            while (k < k1 && k < args.top_k) {
+            while (k < k1) {
                 dst[k++] = tmp0[i++];
             }
             break;

ggml/src/ggml.c

@@ -5158,9 +5158,6 @@ struct ggml_tensor * ggml_top_k(
 
     struct ggml_tensor * result = ggml_new_tensor_4d(ctx, GGML_TYPE_I32, k, a->ne[1], a->ne[2], a->ne[3]);
 
-    // TODO: tmp
-    ggml_set_op_params_i32(result, 0, (int32_t) GGML_SORT_ORDER_DESC);
-
     result->op     = GGML_OP_TOP_K;
     result->src[0] = a;