[Bug] Fix invalid CUDA kernel launch when num_samples exceeds grid dimension limit

[viiccwen]

Description

reference: comment

the launch_l2_norm_batch function can attempt to launch an invalid CUDA kernel when num_samples exceeds CUDA_MAX_GRID_DIM_1D (65535).

Root Cause

When num_samples > 65535, even with blocks_per_sample = 1, the calculated gridSize = num_samples * 1 = num_samples still exceeds the CUDA 1D grid dimension limit (65535), leading to an invalid kernel launch.

The existing code attempts to reduce blocks_per_sample when gridSize > max_grid:

mahout/qdp/qdp-kernels/src/amplitude.cu

Lines 613 to 620 in ef00f92

	const size_t max_grid = CUDA_MAX_GRID_DIM_1D; // CUDA grid dimension limit for 1D launch
	if (gridSize > max_grid) {
	blocks_per_sample = max_grid / num_samples;
	if (blocks_per_sample == 0) {
	blocks_per_sample = 1;
	}
	gridSize = num_samples * blocks_per_sample;
	}

However, when num_samples > max_grid, even with blocks_per_sample = 1, gridSize = num_samples still exceeds the limit, causing a CUDA error.