6.41. Tensor Map Object Managment

This section describes the tensor map object management functions of the low-level CUDA driver application programming interface.

The tensor core API is only supported on devices of compute capability 9.0 or higher.

Functions

CUresult cuTensorMapEncodeIm2col(CUtensorMap *tensorMap, CUtensorMapDataType tensorDataType, cuuint32_t tensorRank, void *globalAddress, const cuuint64_t *globalDim, const cuuint64_t *globalStrides, const int *pixelBoxLowerCorner, const int *pixelBoxUpperCorner, cuuint32_t channelsPerPixel, cuuint32_t pixelsPerColumn, const cuuint32_t *elementStrides, CUtensorMapInterleave interleave, CUtensorMapSwizzle swizzle, CUtensorMapL2promotion l2Promotion, CUtensorMapFloatOOBfill oobFill)

Create a tensor map descriptor object representing im2col memory region.

CUresult cuTensorMapEncodeIm2colWide(CUtensorMap *tensorMap, CUtensorMapDataType tensorDataType, cuuint32_t tensorRank, void *globalAddress, const cuuint64_t *globalDim, const cuuint64_t *globalStrides, int pixelBoxLowerCornerWidth, int pixelBoxUpperCornerWidth, cuuint32_t channelsPerPixel, cuuint32_t pixelsPerColumn, const cuuint32_t *elementStrides, CUtensorMapInterleave interleave, CUtensorMapIm2ColWideMode mode, CUtensorMapSwizzle swizzle, CUtensorMapL2promotion l2Promotion, CUtensorMapFloatOOBfill oobFill)

Create a tensor map descriptor object representing im2col memory region, but where the elements are exclusively loaded along the W dimension.

CUresult cuTensorMapEncodeTiled(CUtensorMap *tensorMap, CUtensorMapDataType tensorDataType, cuuint32_t tensorRank, void *globalAddress, const cuuint64_t *globalDim, const cuuint64_t *globalStrides, const cuuint32_t *boxDim, const cuuint32_t *elementStrides, CUtensorMapInterleave interleave, CUtensorMapSwizzle swizzle, CUtensorMapL2promotion l2Promotion, CUtensorMapFloatOOBfill oobFill)

Create a tensor map descriptor object representing tiled memory region.

CUresult cuTensorMapReplaceAddress(CUtensorMap *tensorMap, void *globalAddress)

Modify an existing tensor map descriptor with an updated global address.

6.41.1. Functions

CUresult cuTensorMapEncodeIm2col(CUtensorMap *tensorMap, CUtensorMapDataType tensorDataType, cuuint32_t tensorRank, void *globalAddress, const cuuint64_t *globalDim, const cuuint64_t *globalStrides, const int *pixelBoxLowerCorner, const int *pixelBoxUpperCorner, cuuint32_t channelsPerPixel, cuuint32_t pixelsPerColumn, const cuuint32_t *elementStrides, CUtensorMapInterleave interleave, CUtensorMapSwizzle swizzle, CUtensorMapL2promotion l2Promotion, CUtensorMapFloatOOBfill oobFill)

Create a tensor map descriptor object representing im2col memory region.

Creates a descriptor for Tensor Memory Access (TMA) object specified by the parameters describing a im2col memory layout and returns it in tensorMap.

Tensor map objects are only supported on devices of compute capability 9.0 or higher. Additionally, a tensor map object is an opaque value, and, as such, should only be accessed through CUDA APIs and PTX.

The parameters passed are bound to the following requirements:

  • tensorMap address must be aligned to 64 bytes.

  • tensorDataType has to be an enum from CUtensorMapDataType which is defined as:

    typedef enum CUtensorMapDataType_enum {
        CU_TENSOR_MAP_DATA_TYPE_UINT8 = 0,       // 1 byte
        CU_TENSOR_MAP_DATA_TYPE_UINT16,          // 2 bytes
        CU_TENSOR_MAP_DATA_TYPE_UINT32,          // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_INT32,           // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_UINT64,          // 8 bytes
        CU_TENSOR_MAP_DATA_TYPE_INT64,           // 8 bytes
        CU_TENSOR_MAP_DATA_TYPE_FLOAT16,         // 2 bytes
        CU_TENSOR_MAP_DATA_TYPE_FLOAT32,         // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_FLOAT64,         // 8 bytes
        CU_TENSOR_MAP_DATA_TYPE_BFLOAT16,        // 2 bytes
        CU_TENSOR_MAP_DATA_TYPE_FLOAT32_FTZ,     // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_TFLOAT32,        // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_TFLOAT32_FTZ     // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B,    // 4 bits
        CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B,   // 4 bits
        CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B    // 6 bits
    } CUtensorMapDataType;
    
    CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B copies ‘16 x U4’ packed values to memory aligned as 8 bytes. There are no gaps between packed values. CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B copies ‘16 x U4’ packed values to memory aligned as 16 bytes. There are 8 byte gaps between every 8 byte chunk of packed values. CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B copies ‘16 x U6’ packed values to memory aligned as 16 bytes. There are 4 byte gaps between every 12 byte chunk of packed values.

  • tensorRank, which specifies the number of tensor dimensions, must be 3, 4, or 5.

  • globalAddress, which specifies the starting address of the memory region described, must be 16 byte aligned. The following requirements need to also be met:

    • When interleave is CU_TENSOR_MAP_INTERLEAVE_32B, globalAddress must be 32 byte aligned.

    • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B or CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, globalAddress must be 32 byte aligned.

  • globalDim array, which specifies tensor size of each of the tensorRank dimensions, must be non-zero and less than or equal to 2^32. Additionally, the following requirements need to be met for the packed data types:

    • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B or CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, globalDim[0] must be a multiple of 128.

    • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B, globalDim[0] must be a multiple of 2.

    • Dimension for the packed data types must reflect the number of individual U# values.

  • globalStrides array, which specifies tensor stride of each of the lower tensorRank - 1 dimensions in bytes, must be a multiple of 16 and less than 2^40. Additionally, the following requirements need to be met:

    • When interleave is CU_TENSOR_MAP_INTERLEAVE_32B, the strides must be a multiple of 32.

    • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B or CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, the strides must be a multiple of 32. Each following dimension specified includes previous dimension stride:

      globalStrides[0] = globalDim[0] * elementSizeInBytes(tensorDataType) + padding[0];
      for (i = 1; i < tensorRank - 1; i++)
          globalStrides[i] = globalStrides[i – 1] * (globalDim[i] + padding[i]);
          assert(globalStrides[i] >= globalDim[i]);
      

  • pixelBoxLowerCorner array specifies the coordinate offsets {D, H, W} of the bounding box from top/left/front corner. The number of offsets and their precision depend on the tensor dimensionality:

    • When tensorRank is 3, one signed offset within range [-32768, 32767] is supported.

    • When tensorRank is 4, two signed offsets each within range [-128, 127] are supported.

    • When tensorRank is 5, three offsets each within range [-16, 15] are supported.

  • pixelBoxUpperCorner array specifies the coordinate offsets {D, H, W} of the bounding box from bottom/right/back corner. The number of offsets and their precision depend on the tensor dimensionality:

    • When tensorRank is 3, one signed offset within range [-32768, 32767] is supported.

    • When tensorRank is 4, two signed offsets each within range [-128, 127] are supported.

    • When tensorRank is 5, three offsets each within range [-16, 15] are supported. The bounding box specified by pixelBoxLowerCorner and pixelBoxUpperCorner must have non-zero area.

  • channelsPerPixel, which specifies the number of elements which must be accessed along C dimension, must be less than or equal to 256. Additionally, when tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B or CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, channelsPerPixel must be 128.

  • pixelsPerColumn, which specifies the number of elements that must be accessed along the {N, D, H, W} dimensions, must be less than or equal to 1024.

  • elementStrides array, which specifies the iteration step along each of the tensorRank dimensions, must be non-zero and less than or equal to 8. Note that when interleave is CU_TENSOR_MAP_INTERLEAVE_NONE, the first element of this array is ignored since TMA doesn’t support the stride for dimension zero. When all elements of the elementStrides array are one, boxDim specifies the number of elements to load. However, if elementStrides[i] is not equal to one for some i, then TMA loads ceil( boxDim[i] / elementStrides[i]) number of elements along i-th dimension. To load N elements along i-th dimension, boxDim[i] must be set to N * elementStrides[i].

  • interleave specifies the interleaved layout of type CUtensorMapInterleave, which is defined as:

    typedef enum CUtensorMapInterleave_enum {
        CU_TENSOR_MAP_INTERLEAVE_NONE = 0,
        CU_TENSOR_MAP_INTERLEAVE_16B,
        CU_TENSOR_MAP_INTERLEAVE_32B
    } CUtensorMapInterleave;
    
    TMA supports interleaved layouts like NC/8HWC8 where C8 utilizes 16 bytes in memory assuming 2 byte per channel or NC/16HWC16 where C16 uses 32 bytes. When interleave is CU_TENSOR_MAP_INTERLEAVE_NONE and swizzle is not CU_TENSOR_MAP_SWIZZLE_NONE, the bounding box inner dimension (computed as channelsPerPixel multiplied by element size in bytes derived from tensorDataType) must be less than or equal to the swizzle size.
    • CU_TENSOR_MAP_SWIZZLE_32B requires the bounding box inner dimension to be <= 32.

    • CU_TENSOR_MAP_SWIZZLE_64B requires the bounding box inner dimension to be <= 64.

    • CU_TENSOR_MAP_SWIZZLE_128B* require the bounding box inner dimension to be <= 128. Additionally, tensorDataType of CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B requires interleave to be CU_TENSOR_MAP_INTERLEAVE_NONE.

  • swizzle, which specifies the shared memory bank swizzling pattern, has to be of type CUtensorMapSwizzle which is defined as:

    typedef enum CUtensorMapSwizzle_enum {
        CU_TENSOR_MAP_SWIZZLE_NONE = 0,
        CU_TENSOR_MAP_SWIZZLE_32B,                   // Swizzle 16B chunks within 32B  span
        CU_TENSOR_MAP_SWIZZLE_64B,                   // Swizzle 16B chunks within 64B  span
        CU_TENSOR_MAP_SWIZZLE_128B,                  // Swizzle 16B chunks within 128B span
        CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B,         // Swizzle 32B chunks within 128B span
        CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B_FLIP_8B, // Swizzle 32B chunks within 128B span, additionally swap lower 8B with upper 8B within each 16B for every alternate row
        CU_TENSOR_MAP_SWIZZLE_128B_ATOM_64B,         // Swizzle 64B chunks within 128B span
        CU_TENSOR_MAP_SWIZZLE_96B,                   // Swizzle 16B chunks within 96B  span
    } CUtensorMapSwizzle;
    
    Data are organized in a specific order in global memory; however, this may not match the order in which the application accesses data in shared memory. This difference in data organization may cause bank conflicts when shared memory is accessed. In order to avoid this problem, data can be loaded to shared memory with shuffling across shared memory banks. When interleave is CU_TENSOR_MAP_INTERLEAVE_32B, swizzle must be CU_TENSOR_MAP_SWIZZLE_32B. Other interleave modes can have any swizzling pattern. When the tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B, only the following swizzle modes are supported:
    • CU_TENSOR_MAP_SWIZZLE_NONE (Load & Store)

    • CU_TENSOR_MAP_SWIZZLE_128B (Load & Store)

    • CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B (Load & Store)

    • CU_TENSOR_MAP_SWIZZLE_128B_ATOM_64B (Store only) When the tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, only the following swizzle modes are supported:

    • CU_TENSOR_MAP_SWIZZLE_NONE (Load only)

    • CU_TENSOR_MAP_SWIZZLE_128B (Load only)

    • CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B (Load only)

  • l2Promotion specifies L2 fetch size which indicates the byte granularity at which L2 requests are filled from DRAM. It must be of type CUtensorMapL2promotion, which is defined as:

    typedef enum CUtensorMapL2promotion_enum {
        CU_TENSOR_MAP_L2_PROMOTION_NONE = 0,
        CU_TENSOR_MAP_L2_PROMOTION_L2_64B,
        CU_TENSOR_MAP_L2_PROMOTION_L2_128B,
        CU_TENSOR_MAP_L2_PROMOTION_L2_256B
    } CUtensorMapL2promotion;
    

  • oobFill, which indicates whether zero or a special NaN constant should be used to fill out-of-bound elements, must be of type CUtensorMapFloatOOBfill which is defined as:

    typedef enum CUtensorMapFloatOOBfill_enum {
        CU_TENSOR_MAP_FLOAT_OOB_FILL_NONE = 0,
        CU_TENSOR_MAP_FLOAT_OOB_FILL_NAN_REQUEST_ZERO_FMA
    } CUtensorMapFloatOOBfill;
    
    Note that CU_TENSOR_MAP_FLOAT_OOB_FILL_NAN_REQUEST_ZERO_FMA can only be used when tensorDataType represents a floating-point data type, and when tensorDataType is not CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B, CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, and CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B.

Parameters
  • tensorMap – - Tensor map object to create

  • tensorDataType – - Tensor data type

  • tensorRank – - Dimensionality of tensor; must be at least 3

  • globalAddress – - Starting address of memory region described by tensor

  • globalDim – - Array containing tensor size (number of elements) along each of the tensorRank dimensions

  • globalStrides – - Array containing stride size (in bytes) along each of the tensorRank - 1 dimensions

  • pixelBoxLowerCorner – - Array containing DHW dimensions of lower box corner

  • pixelBoxUpperCorner – - Array containing DHW dimensions of upper box corner

  • channelsPerPixel – - Number of channels per pixel

  • pixelsPerColumn – - Number of pixels per column

  • elementStrides – - Array containing traversal stride in each of the tensorRank dimensions

  • interleave – - Type of interleaved layout the tensor addresses

  • swizzle – - Bank swizzling pattern inside shared memory

  • l2Promotion – - L2 promotion size

  • oobFill – - Indicate whether zero or special NaN constant will be used to fill out-of-bound elements

Returns

CUDA_SUCCESS, CUDA_ERROR_DEINITIALIZED, CUDA_ERROR_NOT_INITIALIZED, CUDA_ERROR_INVALID_CONTEXT, CUDA_ERROR_INVALID_VALUE

CUresult cuTensorMapEncodeIm2colWide(CUtensorMap *tensorMap, CUtensorMapDataType tensorDataType, cuuint32_t tensorRank, void *globalAddress, const cuuint64_t *globalDim, const cuuint64_t *globalStrides, int pixelBoxLowerCornerWidth, int pixelBoxUpperCornerWidth, cuuint32_t channelsPerPixel, cuuint32_t pixelsPerColumn, const cuuint32_t *elementStrides, CUtensorMapInterleave interleave, CUtensorMapIm2ColWideMode mode, CUtensorMapSwizzle swizzle, CUtensorMapL2promotion l2Promotion, CUtensorMapFloatOOBfill oobFill)

Create a tensor map descriptor object representing im2col memory region, but where the elements are exclusively loaded along the W dimension.

Creates a descriptor for Tensor Memory Access (TMA) object specified by the parameters describing a im2col memory layout and where the row is always loaded along the W dimensuin and returns it in tensorMap. This assumes the tensor layout in memory is either NDHWC, NHWC, or NWC.

This API is only supported on devices of compute capability 10.0 or higher. Additionally, a tensor map object is an opaque value, and, as such, should only be accessed through CUDA APIs and PTX.

The parameters passed are bound to the following requirements:

  • tensorMap address must be aligned to 64 bytes.

  • tensorDataType has to be an enum from CUtensorMapDataType which is defined as:

    typedef enum CUtensorMapDataType_enum {
        CU_TENSOR_MAP_DATA_TYPE_UINT8 = 0,       // 1 byte
        CU_TENSOR_MAP_DATA_TYPE_UINT16,          // 2 bytes
        CU_TENSOR_MAP_DATA_TYPE_UINT32,          // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_INT32,           // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_UINT64,          // 8 bytes
        CU_TENSOR_MAP_DATA_TYPE_INT64,           // 8 bytes
        CU_TENSOR_MAP_DATA_TYPE_FLOAT16,         // 2 bytes
        CU_TENSOR_MAP_DATA_TYPE_FLOAT32,         // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_FLOAT64,         // 8 bytes
        CU_TENSOR_MAP_DATA_TYPE_BFLOAT16,        // 2 bytes
        CU_TENSOR_MAP_DATA_TYPE_FLOAT32_FTZ,     // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_TFLOAT32,        // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_TFLOAT32_FTZ     // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B,    // 4 bits
        CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B,   // 4 bits
        CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B    // 6 bits
    } CUtensorMapDataType;
    
    CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B copies ‘16 x U4’ packed values to memory aligned as 8 bytes. There are no gaps between packed values. CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B copies ‘16 x U4’ packed values to memory aligned as 16 bytes. There are 8 byte gaps between every 8 byte chunk of packed values. CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B copies ‘16 x U6’ packed values to memory aligned as 16 bytes. There are 4 byte gaps between every 12 byte chunk of packed values.

  • tensorRank, which specifies the number of tensor dimensions, must be 3, 4, or 5.

  • globalAddress, which specifies the starting address of the memory region described, must be 16 byte aligned. The following requirements need to also be met:

    • When interleave is CU_TENSOR_MAP_INTERLEAVE_32B, globalAddress must be 32 byte aligned.

    • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B or CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, globalAddress must be 32 byte aligned.

globalDim array, which specifies tensor size of each of the tensorRank dimensions, must be non-zero and less than or equal to 2^32. Additionally, the following requirements need to be met for the packed data types:

  • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B or CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, globalDim[0] must be a multiple of 128.

  • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B, globalDim[0] must be a multiple of 2.

  • Dimension for the packed data types must reflect the number of individual U# values.

globalStrides array, which specifies tensor stride of each of the lower tensorRank - 1 dimensions in bytes, must be a multiple of 16 and less than 2^40. Additionally, the following requirements need to be met:

  • When interleave is CU_TENSOR_MAP_INTERLEAVE_32B, the strides must be a multiple of 32.

  • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B or CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, the strides must be a multiple of 32. Each following dimension specified includes previous dimension stride:

    globalStrides[0] = globalDim[0] * elementSizeInBytes(tensorDataType) + padding[0];
    for (i = 1; i < tensorRank - 1; i++)
        globalStrides[i] = globalStrides[i – 1] * (globalDim[i] + padding[i]);
        assert(globalStrides[i] >= globalDim[i]);
    

pixelBoxLowerCornerWidth specifies the coordinate offset W of the bounding box from left corner. The offset must be within range [-32768, 32767].

  • pixelBoxUpperCornerWidth specifies the coordinate offset W of the bounding box from right corner. The offset must be within range [-32768, 32767].

The bounding box specified by pixelBoxLowerCornerWidth and pixelBoxUpperCornerWidth must have non-zero area. Note that the size of the box along D and H dimensions is always equal to one.

  • channelsPerPixel, which specifies the number of elements which must be accessed along C dimension, must be less than or equal to 256. Additionally, when tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B or CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, channelsPerPixel must be 128.

  • pixelsPerColumn, which specifies the number of elements that must be accessed along the W dimension, must be less than or equal to 1024. This field is ignored when mode is CU_TENSOR_MAP_IM2COL_WIDE_MODE_W128.

  • elementStrides array, which specifies the iteration step along each of the tensorRank dimensions, must be non-zero and less than or equal to 8. Note that when interleave is CU_TENSOR_MAP_INTERLEAVE_NONE, the first element of this array is ignored since TMA doesn’t support the stride for dimension zero. When all elements of the elementStrides array are one, boxDim specifies the number of elements to load. However, if elementStrides[i] is not equal to one for some i, then TMA loads ceil( boxDim[i] / elementStrides[i]) number of elements along i-th dimension. To load N elements along i-th dimension, boxDim[i] must be set to N * elementStrides[i].

  • interleave specifies the interleaved layout of type CUtensorMapInterleave, which is defined as:

    typedef enum CUtensorMapInterleave_enum {
        CU_TENSOR_MAP_INTERLEAVE_NONE = 0,
        CU_TENSOR_MAP_INTERLEAVE_16B,
        CU_TENSOR_MAP_INTERLEAVE_32B
    } CUtensorMapInterleave;
    
    TMA supports interleaved layouts like NC/8HWC8 where C8 utilizes 16 bytes in memory assuming 2 byte per channel or NC/16HWC16 where C16 uses 32 bytes. When interleave is CU_TENSOR_MAP_INTERLEAVE_NONE, the bounding box inner dimension (computed as channelsPerPixel multiplied by element size in bytes derived from tensorDataType) must be less than or equal to the swizzle size.
    • CU_TENSOR_MAP_SWIZZLE_64B requires the bounding box inner dimension to be <= 64.

    • CU_TENSOR_MAP_SWIZZLE_128B* require the bounding box inner dimension to be <= 128. Additionally, tensorDataType of CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B requires interleave to be CU_TENSOR_MAP_INTERLEAVE_NONE.

  • mode, which describes loading of elements loaded along the W dimension, has to be one of the following CUtensorMapIm2ColWideMode types:

    CU_TENSOR_MAP_IM2COL_WIDE_MODE_W,
    CU_TENSOR_MAP_IM2COL_WIDE_MODE_W128
    
    CU_TENSOR_MAP_IM2COL_WIDE_MODE_W allows the number of elements loaded along the W dimension to be specified via the pixelsPerColumn field.

  • swizzle, which specifies the shared memory bank swizzling pattern, must be one of the following CUtensorMapSwizzle modes (other swizzle modes are not supported):

    typedef enum CUtensorMapSwizzle_enum {
        CU_TENSOR_MAP_SWIZZLE_64B,                   // Swizzle 16B chunks within 64B  span
        CU_TENSOR_MAP_SWIZZLE_128B,                  // Swizzle 16B chunks within 128B span
        CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B,         // Swizzle 32B chunks within 128B span
        CU_TENSOR_MAP_SWIZZLE_96B,                   // Swizzle 16B chunks within 96B  span
    } CUtensorMapSwizzle;
    
    Data are organized in a specific order in global memory; however, this may not match the order in which the application accesses data in shared memory. This difference in data organization may cause bank conflicts when shared memory is accessed. In order to avoid this problem, data can be loaded to shared memory with shuffling across shared memory banks. When the tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B, only the following swizzle modes are supported:
    • CU_TENSOR_MAP_SWIZZLE_64B (Store only)

    • CU_TENSOR_MAP_SWIZZLE_128B (Load & Store)

    • CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B (Load & Store) When the tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, only the following swizzle modes are supported:

    • CU_TENSOR_MAP_SWIZZLE_128B (Load only)

    • CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B (Load only)

Additionally, CU_TENSOR_MAP_SWIZZLE_96B is supported only when mode is CU_TENSOR_MAP_IM2COL_WIDE_MODE_W.

  • l2Promotion specifies L2 fetch size which indicates the byte granularity at which L2 requests are filled from DRAM. It must be of type CUtensorMapL2promotion, which is defined as:

    typedef enum CUtensorMapL2promotion_enum {
        CU_TENSOR_MAP_L2_PROMOTION_NONE = 0,
        CU_TENSOR_MAP_L2_PROMOTION_L2_64B,
        CU_TENSOR_MAP_L2_PROMOTION_L2_128B,
        CU_TENSOR_MAP_L2_PROMOTION_L2_256B
    } CUtensorMapL2promotion;
    

  • oobFill, which indicates whether zero or a special NaN constant should be used to fill out-of-bound elements, must be of type CUtensorMapFloatOOBfill which is defined as:

    typedef enum CUtensorMapFloatOOBfill_enum {
        CU_TENSOR_MAP_FLOAT_OOB_FILL_NONE = 0,
        CU_TENSOR_MAP_FLOAT_OOB_FILL_NAN_REQUEST_ZERO_FMA
    } CUtensorMapFloatOOBfill;
    
    Note that CU_TENSOR_MAP_FLOAT_OOB_FILL_NAN_REQUEST_ZERO_FMA can only be used when tensorDataType represents a floating-point data type, and when tensorDataType is not CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B, CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, and CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B.

Parameters
  • tensorMap – - Tensor map object to create

  • tensorDataType – - Tensor data type

  • tensorRank – - Dimensionality of tensor; must be at least 3

  • globalAddress – - Starting address of memory region described by tensor

  • globalDim – - Array containing tensor size (number of elements) along each of the tensorRank dimensions

  • globalStrides – - Array containing stride size (in bytes) along each of the tensorRank - 1 dimensions

  • pixelBoxLowerCornerWidth – - Width offset of left box corner

  • pixelBoxUpperCornerWidth – - Width offset of right box corner

  • channelsPerPixel – - Number of channels per pixel

  • pixelsPerColumn – - Number of pixels per column

  • elementStrides – - Array containing traversal stride in each of the tensorRank dimensions

  • interleave – - Type of interleaved layout the tensor addresses

  • mode – - W or W128 mode

  • swizzle – - Bank swizzling pattern inside shared memory

  • l2Promotion – - L2 promotion size

  • oobFill – - Indicate whether zero or special NaN constant will be used to fill out-of-bound elements

Returns

CUDA_SUCCESS, CUDA_ERROR_DEINITIALIZED, CUDA_ERROR_NOT_INITIALIZED, CUDA_ERROR_INVALID_CONTEXT, CUDA_ERROR_INVALID_VALUE

CUresult cuTensorMapEncodeTiled(CUtensorMap *tensorMap, CUtensorMapDataType tensorDataType, cuuint32_t tensorRank, void *globalAddress, const cuuint64_t *globalDim, const cuuint64_t *globalStrides, const cuuint32_t *boxDim, const cuuint32_t *elementStrides, CUtensorMapInterleave interleave, CUtensorMapSwizzle swizzle, CUtensorMapL2promotion l2Promotion, CUtensorMapFloatOOBfill oobFill)

Create a tensor map descriptor object representing tiled memory region.

Creates a descriptor for Tensor Memory Access (TMA) object specified by the parameters describing a tiled region and returns it in tensorMap.

Tensor map objects are only supported on devices of compute capability 9.0 or higher. Additionally, a tensor map object is an opaque value, and, as such, should only be accessed through CUDA APIs and PTX.

The parameters passed are bound to the following requirements:

  • tensorMap address must be aligned to 64 bytes.

  • tensorDataType has to be an enum from CUtensorMapDataType which is defined as:

    typedef enum CUtensorMapDataType_enum {
        CU_TENSOR_MAP_DATA_TYPE_UINT8 = 0,       // 1 byte
        CU_TENSOR_MAP_DATA_TYPE_UINT16,          // 2 bytes
        CU_TENSOR_MAP_DATA_TYPE_UINT32,          // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_INT32,           // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_UINT64,          // 8 bytes
        CU_TENSOR_MAP_DATA_TYPE_INT64,           // 8 bytes
        CU_TENSOR_MAP_DATA_TYPE_FLOAT16,         // 2 bytes
        CU_TENSOR_MAP_DATA_TYPE_FLOAT32,         // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_FLOAT64,         // 8 bytes
        CU_TENSOR_MAP_DATA_TYPE_BFLOAT16,        // 2 bytes
        CU_TENSOR_MAP_DATA_TYPE_FLOAT32_FTZ,     // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_TFLOAT32,        // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_TFLOAT32_FTZ,    // 4 bytes
        CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B,    // 4 bits
        CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B,   // 4 bits
        CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B    // 6 bits
    } CUtensorMapDataType;
    
    CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B copies ‘16 x U4’ packed values to memory aligned as 8 bytes. There are no gaps between packed values. CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B copies ‘16 x U4’ packed values to memory aligned as 16 bytes. There are 8 byte gaps between every 8 byte chunk of packed values. CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B copies ‘16 x U6’ packed values to memory aligned as 16 bytes. There are 4 byte gaps between every 12 byte chunk of packed values.

  • tensorRank must be non-zero and less than or equal to the maximum supported dimensionality of 5. If interleave is not CU_TENSOR_MAP_INTERLEAVE_NONE, then tensorRank must additionally be greater than or equal to 3.

  • globalAddress, which specifies the starting address of the memory region described, must be 16 byte aligned. The following requirements need to also be met:

    • When interleave is CU_TENSOR_MAP_INTERLEAVE_32B, globalAddress must be 32 byte aligned.

    • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B or CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, globalAddress must be 32 byte aligned.

globalDim array, which specifies tensor size of each of the tensorRank dimensions, must be non-zero and less than or equal to 2^32. Additionally, the following requirements need to be met for the packed data types:

  • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B or CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, globalDim[0] must be a multiple of 128.

  • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B, globalDim[0] must be a multiple of 2.

  • Dimension for the packed data types must reflect the number of individual U# values.

globalStrides array, which specifies tensor stride of each of the lower tensorRank - 1 dimensions in bytes, must be a multiple of 16 and less than 2^40. Additionally, the following requirements need to be met:

  • When interleave is CU_TENSOR_MAP_INTERLEAVE_32B, the strides must be a multiple of 32.

  • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B or CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, the strides must be a multiple of 32. Each following dimension specified includes previous dimension stride:

    globalStrides[0] = globalDim[0] * elementSizeInBytes(tensorDataType) + padding[0];
    for (i = 1; i < tensorRank - 1; i++)
        globalStrides[i] = globalStrides[i – 1] * (globalDim[i] + padding[i]);
        assert(globalStrides[i] >= globalDim[i]);
    

boxDim array, which specifies number of elements to be traversed along each of the tensorRank dimensions, must be non-zero and less than or equal to 256. Additionally, the following requirements need to be met:

  • When interleave is CU_TENSOR_MAP_INTERLEAVE_NONE, { boxDim[0] * elementSizeInBytes( tensorDataType ) } must be a multiple of 16 bytes.

  • When tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B or CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, boxDim[0] must be 128.

elementStrides array, which specifies the iteration step along each of the tensorRank dimensions, must be non-zero and less than or equal to 8. Note that when interleave is CU_TENSOR_MAP_INTERLEAVE_NONE, the first element of this array is ignored since TMA doesn’t support the stride for dimension zero. When all elements of elementStrides array is one, boxDim specifies the number of elements to load. However, if the elementStrides[i] is not equal to one, then TMA loads ceil( boxDim[i] / elementStrides[i]) number of elements along i-th dimension. To load N elements along i-th dimension, boxDim[i] must be set to N * elementStrides[i].

  • interleave specifies the interleaved layout of type CUtensorMapInterleave, which is defined as:

    typedef enum CUtensorMapInterleave_enum {
        CU_TENSOR_MAP_INTERLEAVE_NONE = 0,
        CU_TENSOR_MAP_INTERLEAVE_16B,
        CU_TENSOR_MAP_INTERLEAVE_32B
    } CUtensorMapInterleave;
    
    TMA supports interleaved layouts like NC/8HWC8 where C8 utilizes 16 bytes in memory assuming 2 byte per channel or NC/16HWC16 where C16 uses 32 bytes. When interleave is CU_TENSOR_MAP_INTERLEAVE_NONE and swizzle is not CU_TENSOR_MAP_SWIZZLE_NONE, the bounding box inner dimension (computed as boxDim[0] multiplied by element size derived from tensorDataType) must be less than or equal to the swizzle size.
    • CU_TENSOR_MAP_SWIZZLE_32B requires the bounding box inner dimension to be <= 32.

    • CU_TENSOR_MAP_SWIZZLE_64B requires the bounding box inner dimension to be <= 64.

    • CU_TENSOR_MAP_SWIZZLE_128B* require the bounding box inner dimension to be <= 128. Additionally, tensorDataType of CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B requires interleave to be CU_TENSOR_MAP_INTERLEAVE_NONE.

  • swizzle, which specifies the shared memory bank swizzling pattern, has to be of type CUtensorMapSwizzle which is defined as:

    typedef enum CUtensorMapSwizzle_enum {
        CU_TENSOR_MAP_SWIZZLE_NONE = 0,
        CU_TENSOR_MAP_SWIZZLE_32B,                   // Swizzle 16B chunks within 32B  span
        CU_TENSOR_MAP_SWIZZLE_64B,                   // Swizzle 16B chunks within 64B  span
        CU_TENSOR_MAP_SWIZZLE_128B,                  // Swizzle 16B chunks within 128B span
        CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B,         // Swizzle 32B chunks within 128B span
        CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B_FLIP_8B, // Swizzle 32B chunks within 128B span, additionally swap lower 8B with upper 8B within each 16B for every alternate row
        CU_TENSOR_MAP_SWIZZLE_128B_ATOM_64B,         // Swizzle 64B chunks within 128B span
        CU_TENSOR_MAP_SWIZZLE_96B,                   // Swizzle 16B chunks within 96B  span
    } CUtensorMapSwizzle;
    
    Data are organized in a specific order in global memory; however, this may not match the order in which the application accesses data in shared memory. This difference in data organization may cause bank conflicts when shared memory is accessed. In order to avoid this problem, data can be loaded to shared memory with shuffling across shared memory banks. When interleave is CU_TENSOR_MAP_INTERLEAVE_32B, swizzle must be CU_TENSOR_MAP_SWIZZLE_32B. Other interleave modes can have any swizzling pattern. When the tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B, only the following swizzle modes are supported:
    • CU_TENSOR_MAP_SWIZZLE_NONE (Load & Store)

    • CU_TENSOR_MAP_SWIZZLE_128B (Load & Store)

    • CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B (Load & Store)

    • CU_TENSOR_MAP_SWIZZLE_128B_ATOM_64B (Store only) When the tensorDataType is CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, only the following swizzle modes are supported:

    • CU_TENSOR_MAP_SWIZZLE_NONE (Load only)

    • CU_TENSOR_MAP_SWIZZLE_128B (Load only)

    • CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B (Load only)

  • l2Promotion specifies L2 fetch size which indicates the byte granurality at which L2 requests is filled from DRAM. It must be of type CUtensorMapL2promotion, which is defined as:

    typedef enum CUtensorMapL2promotion_enum {
        CU_TENSOR_MAP_L2_PROMOTION_NONE = 0,
        CU_TENSOR_MAP_L2_PROMOTION_L2_64B,
        CU_TENSOR_MAP_L2_PROMOTION_L2_128B,
        CU_TENSOR_MAP_L2_PROMOTION_L2_256B
    } CUtensorMapL2promotion;
    

  • oobFill, which indicates whether zero or a special NaN constant should be used to fill out-of-bound elements, must be of type CUtensorMapFloatOOBfill which is defined as:

    typedef enum CUtensorMapFloatOOBfill_enum {
        CU_TENSOR_MAP_FLOAT_OOB_FILL_NONE = 0,
        CU_TENSOR_MAP_FLOAT_OOB_FILL_NAN_REQUEST_ZERO_FMA
    } CUtensorMapFloatOOBfill;
    
    Note that CU_TENSOR_MAP_FLOAT_OOB_FILL_NAN_REQUEST_ZERO_FMA can only be used when tensorDataType represents a floating-point data type, and when tensorDataType is not CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B, CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN16B, and CU_TENSOR_MAP_DATA_TYPE_16U6_ALIGN16B.

Parameters
  • tensorMap – - Tensor map object to create

  • tensorDataType – - Tensor data type

  • tensorRank – - Dimensionality of tensor

  • globalAddress – - Starting address of memory region described by tensor

  • globalDim – - Array containing tensor size (number of elements) along each of the tensorRank dimensions

  • globalStrides – - Array containing stride size (in bytes) along each of the tensorRank - 1 dimensions

  • boxDim – - Array containing traversal box size (number of elments) along each of the tensorRank dimensions. Specifies how many elements to be traversed along each tensor dimension.

  • elementStrides – - Array containing traversal stride in each of the tensorRank dimensions

  • interleave – - Type of interleaved layout the tensor addresses

  • swizzle – - Bank swizzling pattern inside shared memory

  • l2Promotion – - L2 promotion size

  • oobFill – - Indicate whether zero or special NaN constant must be used to fill out-of-bound elements

Returns

CUDA_SUCCESS, CUDA_ERROR_DEINITIALIZED, CUDA_ERROR_NOT_INITIALIZED, CUDA_ERROR_INVALID_CONTEXT, CUDA_ERROR_INVALID_VALUE

CUresult cuTensorMapReplaceAddress(CUtensorMap *tensorMap, void *globalAddress)

Modify an existing tensor map descriptor with an updated global address.

Modifies the descriptor for Tensor Memory Access (TMA) object passed in tensorMap with an updated globalAddress.

Tensor map objects are only supported on devices of compute capability 9.0 or higher. Additionally, a tensor map object is an opaque value, and, as such, should only be accessed through CUDA API calls.

Parameters
  • tensorMap – - Tensor map object to modify

  • globalAddress – - Starting address of memory region described by tensor, must follow previous alignment requirements

Returns

CUDA_SUCCESS, CUDA_ERROR_DEINITIALIZED, CUDA_ERROR_NOT_INITIALIZED, CUDA_ERROR_INVALID_CONTEXT, CUDA_ERROR_INVALID_VALUE