csrmv LRB: fix unsigned 32-bit overflow in grid_size (AISPARSE-662)

projects/rocsparse/clients/unittests/CMakeLists.txt

@@ -148,6 +148,8 @@ set(ROCSPARSE_UNIT_TEST_DEVICE_SOURCES
   unit_test_internal_collectives_readfirstlane.cpp
   # Per-axis grid-size helpers in rocsparse_grid.hpp (append-only leaf)
   unit_test_internal_grid.cpp
+  # dispatch_grid_stride_x variant selection in rocsparse_grid.hpp (append-only leaf)
+  unit_test_internal_grid_dispatch.cpp
   unit_test_internal_collectives_boost.cpp
   unit_test_internal_infostructs.cpp
   #10156: distinct anchor to avoid stacked-branch merge conflicts
@@ -203,6 +205,8 @@ set(ROCSPARSE_UNIT_TEST_DEVICE_SOURCES
   #10158: distinct anchor to avoid stacked-branch merge conflicts
   unit_test_host_level3.cpp
   unit_test_internal_level2_extras.cpp
+  # csrmv LRB / adaptive partial_scale_y forced grid.x clamp [AISPARSE-662]
+  unit_test_internal_level2_csrmv_lrb_grids.cpp
   # --- Level 2 (SpMV/gebsrmv) public C-API host tests [AISPARSE-642] --------
   # Distinct anchor (end of the level-2 group) so this leaf does not conflict
   # with sibling unit-test branches that append after rocsparse_unit_test_main.

projects/rocsparse/clients/unittests/unit_test_internal_grid_dispatch.cpp

@@ -0,0 +1,157 @@
+/*! \file */
+/* ************************************************************************
+ * Copyright (C) 2026 Advanced Micro Devices, Inc. All rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ *
+ * ************************************************************************ */
+
+//
+// Unit tests for rocsparse::dispatch_grid_stride_x in rocsparse_grid.hpp.
+//
+// dispatch_grid_stride_x picks the straight-line variant (std::false_type)
+// while count fits the grid.x cap and the grid-stride variant (std::true_type)
+// once it does not. These tests check that choice and the grid it passes at
+// the boundary, count == cap and count == cap + 1, against both caps: the
+// derived (2^32 - 1) / block_size bound, which binds at block size 32 and up,
+// and maxGridSize[0], which binds at block size 1. No kernel is launched.
+//
+#include "unit_test_utils.hpp"
+
+#include "rocsparse_grid.hpp"
+
+#include <gtest/gtest.h>
+
+#include <algorithm>
+#include <cstdint>
+#include <type_traits>
+
+using namespace rocsparse_ut;
+
+namespace
+{
+    constexpr int64_t BLOCK_SIZES[] = {1, 32, 64, 128, 256, 512, 1024};
+
+    struct Dispatch
+    {
+        int      calls       = 0;
+        bool     grid_stride = false;
+        uint32_t grid        = 0;
+    };
+
+    template <typename J>
+    Dispatch dispatch(rocsparse_handle handle, J count, int64_t block_size)
+    {
+        Dispatch         d;
+        rocsparse_status status = rocsparse::dispatch_grid_stride_x(
+            handle, count, block_size, [&](auto grid_stride, uint32_t grid) {
+                ++d.calls;
+                d.grid_stride = decltype(grid_stride)::value;
+                d.grid        = grid;
+                return rocsparse_status_success;
+            });
+        EXPECT_EQ(status, rocsparse_status_success);
+        return d;
+    }
+
+    int64_t grid_x_cap(rocsparse_handle handle, int64_t block_size)
+    {
+        return std::min(static_cast<int64_t>(handle->properties.maxGridSize[0]),
+                        rocsparse::dispatch_limit_x(block_size));
+    }
+}
+
+class internal_grid_dispatch : public HandleTest
+{
+};
+
+// count == cap launches the straight-line variant on the full grid; cap + 1
+// launches the grid-stride variant on the clamped grid.
+TEST_F(internal_grid_dispatch, boundary)
+{
+    for(const int64_t bs : BLOCK_SIZES)
+    {
+        const int64_t cap = grid_x_cap(handle, bs);
+
+        const Dispatch at = dispatch(handle, cap, bs);
+        EXPECT_EQ(at.calls, 1) << "block size " << bs;
+        EXPECT_FALSE(at.grid_stride) << "block size " << bs << ", count " << cap;
+        EXPECT_EQ(static_cast<int64_t>(at.grid), cap) << "block size " << bs;
+
+        const Dispatch over = dispatch(handle, cap + 1, bs);
+        EXPECT_EQ(over.calls, 1) << "block size " << bs;
+        EXPECT_TRUE(over.grid_stride) << "block size " << bs << ", count " << cap + 1;
+        EXPECT_EQ(static_cast<int64_t>(over.grid), cap) << "block size " << bs;
+    }
+}
+
+// The derived bound binds below maxGridSize[0] at 256 threads, and the device
+// cap binds at 1 thread.
+TEST_F(internal_grid_dispatch, boundary_concrete_caps)
+{
+    ASSERT_GT(static_cast<int64_t>(handle->properties.maxGridSize[0]), 16777215);
+    EXPECT_EQ(grid_x_cap(handle, 256), 16777215);
+    EXPECT_EQ(grid_x_cap(handle, 1), static_cast<int64_t>(handle->properties.maxGridSize[0]));
+
+    EXPECT_FALSE(dispatch(handle, int32_t(16777215), 256).grid_stride);
+    EXPECT_TRUE(dispatch(handle, int32_t(16777216), 256).grid_stride);
+}
+
+// The choice does not depend on the count type. A 32-bit count crosses the
+// bound as readily as a 64-bit one.
+TEST_F(internal_grid_dispatch, count_types)
+{
+    constexpr int64_t bs  = 1024;
+    const int64_t     cap = grid_x_cap(handle, bs);
+
+    EXPECT_FALSE(dispatch(handle, static_cast<int32_t>(cap), bs).grid_stride);
+    EXPECT_TRUE(dispatch(handle, static_cast<int32_t>(cap + 1), bs).grid_stride);
+    EXPECT_FALSE(dispatch(handle, static_cast<uint32_t>(cap), bs).grid_stride);
+    EXPECT_TRUE(dispatch(handle, static_cast<uint32_t>(cap + 1), bs).grid_stride);
+    EXPECT_FALSE(dispatch(handle, static_cast<int64_t>(cap), bs).grid_stride);
+    EXPECT_TRUE(dispatch(handle, static_cast<int64_t>(cap + 1), bs).grid_stride);
+    EXPECT_FALSE(dispatch(handle, static_cast<uint64_t>(cap), bs).grid_stride);
+    EXPECT_TRUE(dispatch(handle, static_cast<uint64_t>(cap + 1), bs).grid_stride);
+}
+
+// Small counts and a zero count stay on the straight-line variant with the
+// grid equal to count.
+TEST_F(internal_grid_dispatch, small_counts)
+{
+    for(const int64_t bs : BLOCK_SIZES)
+    {
+        for(const int64_t count : {int64_t(0), int64_t(1), int64_t(1000)})
+        {
+            const Dispatch d = dispatch(handle, count, bs);
+            EXPECT_FALSE(d.grid_stride) << "block size " << bs << ", count " << count;
+            EXPECT_EQ(static_cast<int64_t>(d.grid), count) << "block size " << bs;
+        }
+    }
+}
+
+// The launcher's status is returned to the caller.
+TEST_F(internal_grid_dispatch, returns_launch_status)
+{
+    EXPECT_EQ(rocsparse::dispatch_grid_stride_x(
+                  handle,
+                  int64_t(1),
+                  256,
+                  [](auto, uint32_t) { return rocsparse_status_internal_error; }),
+              rocsparse_status_internal_error);
+}