Open topic with navigation

Overview

Texture memory is read-only device memory, and can be accessed using the device functions described in Texture Functions. Reading a texture using one of these functions is called a texture fetch. Texture memory traffic is routed through the texture cache (which is independent of the L1 data cache) and the L2 cache. The texture memory system is designed to be efficient for spatially-localized accesses in 2D arrays. The GPU's hardware support for texturing provides features beyond typical memory systems, such as customizable behavior when reading out-of-bounds, and filtered interpolation when reading from coordinates between array elements. Texture memory is designed for streaming fetches with a constant latency; a texture cache hit reduces device memory bandwidth usage, but not fetch latency.

Chart Texture Each time a warp executes a texture function to read from texture memory, this counts as a single fetch. The Texture chart shows the number of fetches and the fetch rate. A texture fetch is a memory request, which incurs transactions from the texture cache. Transactions from the texture cache are 32 byte units. Note that this chart does not include global read only data accesses (using const __restrict__ or the __ldg() intrinsic), which also use the texture cache. For those, see Global Memory.

Analysis

• If the Texture Cache Hit Rate is low …
  • … the access pattern is not a good match for the texture cache's optimizations for 2D spatial locality. Try to modify the access patterns to group spatially-collocated reads more closely together. Alternatively, for 1D linear access patterns, global memory may be more efficient than texture memory.
• If the texture memory bandwidth is a performance bottleneck …
  • … consider using a simpler texture filtering mode, or a lower resolution texture.

of

NVIDIA GameWorks Documentation Rev. 1.0.150630 ©2015. NVIDIA Corporation. All Rights Reserved.