fix(dnn/cuda): fix topk grid oversize

GitOrigin-RevId: d3c811a034

dnn/src/cuda/topk/topk_radix.cu

@@ -470,7 +470,17 @@ static size_t get_scan_workspace(uint32_t size) {
 
 uint32_t topk::find_kth_radix_workspace(uint32_t batch, uint32_t length) {
     using namespace cuda_topk_impl::kth;
-    return (batch * get_grid_dim_x(length) * NR_BUCKET + batch * 2) *
+    int device_id;
+    if (cudaGetDevice(&device_id) != cudaSuccess) {
+        megdnn_trap();
+    }
+    cudaDeviceProp prop;
+    if (cudaGetDeviceProperties(&prop, device_id) != cudaSuccess) {
+        megdnn_trap();
+    }
+    uint32_t grid_dim_y_limit = prop.maxGridSize[1];
+    uint32_t limit = batch > grid_dim_y_limit ? grid_dim_y_limit : batch;
+    return (limit * get_grid_dim_x(length) * NR_BUCKET + limit * 2) *
            sizeof(uint32_t);
 }
 
@@ -491,35 +501,65 @@ cudaError_t topk::find_kth_radix(const ctype* input, ctype* output,
         // assert
         megdnn_trap();
     }
+
+    int device_id;
+    if (cudaGetDevice(&device_id) != cudaSuccess) {
+        megdnn_trap();
+    }
+    cudaDeviceProp prop;
+    if (cudaGetDeviceProperties(&prop, device_id) != cudaSuccess) {
+        megdnn_trap();
+    }
+    uint32_t grid_dim_y_limit = prop.maxGridSize[1];
+
+    uint32_t batch_idx = 0;
     uint32_t grid_dim_x = get_grid_dim_x(length);
-    dim3 grid_dim(grid_dim_x, batch);
-    uint32_t* dev_k = static_cast<uint32_t*>(workspace);
-    uint32_t* dev_prefix = dev_k + batch;
-    uint32_t* bucket_cnt = dev_prefix + batch;
-
-    compute_histogram<ctype, false, 24><<<grid_dim, BLOCK_DIM, 0, stream>>>(
-            input, bucket_cnt, length, lda, nullptr);
-    // use float to make compiler happy; it is not used since last == false
-    update_prefix_and_k<true, false, 24, float>
-            <<<batch, NR_BUCKET, 0, stream>>>(bucket_cnt, dev_prefix, dev_k, k,
-                                              grid_dim_x, nullptr);
-
-    compute_histogram<ctype, true, 16><<<grid_dim, BLOCK_DIM, 0, stream>>>(
-            input, bucket_cnt, length, lda, dev_prefix);
-    update_prefix_and_k<false, false, 16, float>
-            <<<batch, NR_BUCKET, 0, stream>>>(bucket_cnt, dev_prefix, dev_k, k,
-                                              grid_dim_x, nullptr);
-    compute_histogram<ctype, true, 8><<<grid_dim, BLOCK_DIM, 0, stream>>>(
-            input, bucket_cnt, length, lda, dev_prefix);
-    update_prefix_and_k<false, false, 8, float>
-            <<<batch, NR_BUCKET, 0, stream>>>(bucket_cnt, dev_prefix, dev_k, k,
-                                              grid_dim_x, nullptr);
-
-    compute_histogram<ctype, true, 0><<<grid_dim, BLOCK_DIM, 0, stream>>>(
-            input, bucket_cnt, length, lda, dev_prefix);
-    update_prefix_and_k<false, true, 0, ctype><<<batch, NR_BUCKET, 0, stream>>>(
-            bucket_cnt, dev_prefix, dev_k, k, grid_dim_x, output);
+    uint32_t grid_dim_y = 1;
+
+    while (batch_idx < batch) {
+        if (batch - batch_idx >= grid_dim_y_limit) {
+            grid_dim_y = grid_dim_y_limit;
+        } else {
+            grid_dim_y = batch - batch_idx;
+        }
+
+        dim3 grid_dim(grid_dim_x, grid_dim_y);
+        uint32_t* dev_k = static_cast<uint32_t*>(workspace);
+        uint32_t* dev_prefix = dev_k + grid_dim_y;
+        uint32_t* bucket_cnt = dev_prefix + grid_dim_y;
+
+        compute_histogram<ctype, false, 24><<<grid_dim, BLOCK_DIM, 0, stream>>>(
+                input + batch_idx * lda, bucket_cnt, length, lda, nullptr);
+
+        // use float to make compiler happy; it is not used since last == false
+        update_prefix_and_k<true, false, 24, float>
+                <<<grid_dim_y, NR_BUCKET, 0, stream>>>(
+                        bucket_cnt, dev_prefix, dev_k, k, grid_dim_x, nullptr);
+
+        compute_histogram<ctype, true, 16><<<grid_dim, BLOCK_DIM, 0, stream>>>(
+                input + batch_idx * lda, bucket_cnt, length, lda, dev_prefix);
+
+        update_prefix_and_k<false, false, 16, float>
+                <<<grid_dim_y, NR_BUCKET, 0, stream>>>(
+                        bucket_cnt, dev_prefix, dev_k, k, grid_dim_x, nullptr);
+
+        compute_histogram<ctype, true, 8><<<grid_dim, BLOCK_DIM, 0, stream>>>(
+                input + batch_idx * lda, bucket_cnt, length, lda, dev_prefix);
 
+        update_prefix_and_k<false, false, 8, float>
+                <<<grid_dim_y, NR_BUCKET, 0, stream>>>(
+                        bucket_cnt, dev_prefix, dev_k, k, grid_dim_x, nullptr);
+
+        compute_histogram<ctype, true, 0><<<grid_dim, BLOCK_DIM, 0, stream>>>(
+                input + batch_idx * lda, bucket_cnt, length, lda, dev_prefix);
+
+        update_prefix_and_k<false, true, 0, ctype>
+                <<<grid_dim_y, NR_BUCKET, 0, stream>>>(bucket_cnt, dev_prefix,
+                                                       dev_k, k, grid_dim_x,
+                                                       output + batch_idx);
+
+        batch_idx += grid_dim_y;
+    }
     return cudaGetLastError();
 }
 
@@ -530,12 +570,18 @@ cudaError_t topk::topk_select(const ctype* input, const ctype* thresh,
                               int32_t lda, int32_t k, cudaStream_t stream) {
     using namespace cuda_topk_impl;
     using namespace cuda_topk_impl::select;
-    uint32_t length_split = DIVUP(length, REDUCE_SIZE),
-             scan_size = batch * length_split;
-    size_t scan_wk = get_scan_workspace(scan_size);
-    uint64_t *scan_inp = static_cast<uint64_t*>(workspace) +
-                         scan_wk / sizeof(uint64_t),
-             *scan_out = scan_inp + scan_size;
+
+    int device_id;
+    if (cudaGetDevice(&device_id) != cudaSuccess) {
+        megdnn_trap();
+    }
+    cudaDeviceProp prop;
+    if (cudaGetDeviceProperties(&prop, device_id) != cudaSuccess) {
+        megdnn_trap();
+    }
+    uint32_t batch_upper_limit = prop.maxGridSize[1];
+
+    uint32_t length_split = DIVUP(length, REDUCE_SIZE);
 
     void (*kptr_reduce_block_cnt)(const ctype*, const ctype*, uint32_t, int32_t,
                                   uint64_t*, uint32_t);
@@ -585,25 +631,47 @@ cudaError_t topk::topk_select(const ctype* input, const ctype* thresh,
 #undef CASE_SHARD
 #undef CASE_SHARD_ON
 
-    // reduce to scan_inp
-    kptr_reduce_block_cnt<<<dim3(DIVUP(length_split, REDUCE_SHARD), batch),
-                            dim3(REDUCE_WARP_SIZE, REDUCE_SHARD), 0, stream>>>(
-            input, thresh, length, lda, scan_inp, length_split);
+    uint32_t batch_idx = 0;
+    uint32_t batch_real = 1;
 
-    // scan to scan_out
-    scan_out += 1;  // set scan[-1] to 0
-    cudaError_t err = invoke_cub_scan(scan_inp, scan_out, workspace, scan_wk,
-                                      scan_size, stream);
-    if (err != cudaSuccess) {
-        return err;
-    }
-    kern_init_zero<<<1, 1, 0, stream>>>(scan_out - 1);
+    while (batch_idx < batch) {
+        if (batch - batch_idx >= batch_upper_limit) {
+            batch_real = batch_upper_limit;
+        } else {
+            batch_real = batch - batch_idx;
+        }
 
-    // copy result
-    kptr_copy<<<dim3(DIVUP(length_split, kern_copy_shard), batch),
-                dim3(WARP_SIZE, kern_copy_shard), 0, stream>>>(
-            input, thresh, scan_out, length_split, output_value, output_idx,
-            length, k, lda);
+        size_t scan_size = batch_real * length_split;
+        size_t scan_wk = get_scan_workspace(scan_size);
+        uint64_t *scan_inp = static_cast<uint64_t*>(workspace) +
+                             scan_wk / sizeof(uint64_t),
+                 *scan_out = scan_inp + scan_size;
+
+        // reduce to scan_inp
+        kptr_reduce_block_cnt<<<
+                dim3(DIVUP(length_split, REDUCE_SHARD), batch_real),
+                dim3(REDUCE_WARP_SIZE, REDUCE_SHARD), 0, stream>>>(
+                input + batch_idx * lda, thresh + batch_idx, length, lda,
+                scan_inp, length_split);
+
+        // scan to scan_out
+        scan_out += 1;  // set scan[-1] to 0
+        cudaError_t err = invoke_cub_scan(scan_inp, scan_out, workspace,
+                                          scan_wk, scan_size, stream);
+        if (err != cudaSuccess) {
+            return err;
+        }
+        kern_init_zero<<<1, 1, 0, stream>>>(scan_out - 1);
+
+        // copy result
+        kptr_copy<<<dim3(DIVUP(length_split, kern_copy_shard), batch_real),
+                    dim3(WARP_SIZE, kern_copy_shard), 0, stream>>>(
+                input + batch_idx * lda, thresh + batch_idx, scan_out,
+                length_split, output_value + std::abs(k) * batch_idx,
+                output_idx + std::abs(k) * batch_idx, length, k, lda);
+
+        batch_idx += batch_real;
+    }
     return cudaGetLastError();
 }
 

dnn/test/common/topk.cpp

@@ -169,6 +169,12 @@ void test::run_topk_test(Handle* handle) {
         run(5, 123, 3, mode);         // equiv to sort
         run(-5, 123, 3, mode);        // equiv to rev sort
         run(5, 3, 1231, mode, 2000);  // non contig
+
+//! opencl on armv7's CI does not support large batch.
+//! but P30 and MI9 are ok. fix it in the future.
+#if !defined(MEGDNN_ARMV7) && defined(MGB_CUDA)
+        run(3, 70000, 5, mode, 10);  // non contig
+#endif
     }
 
     // special case to check if tie-break is correct