Data types

Data types#

hipSPARSELt: Data types
Data Structures | Enumerations
Data types

Data Structures

struct  hipsparseLtHandle_t
 Handle to the hipSPARSELt library context queue. More...
 
struct  hipsparseLtMatDescriptor_t
 Descriptor of the matrix. More...
 
struct  hipsparseLtMatmulDescriptor_t
 Descriptor of the matrix multiplication operation. More...
 
struct  hipsparseLtMatmulAlgSelection_t
 Descriptor of the matrix multiplication algorithm. More...
 
struct  hipsparseLtMatmulPlan_t
 Descriptor of the matrix multiplication execution plan. More...
 

Enumerations

enum  hipsparseLtSparsity_t { HIPSPARSELT_SPARSITY_50_PERCENT }
 Specify the sparsity of the structured matrix. More...
 
enum  hipsparseLtMatDescAttribute_t {
  HIPSPARSELT_MAT_NUM_BATCHES ,
  HIPSPARSELT_MAT_BATCH_STRIDE
}
 Specify the additional attributes of a matrix descriptor. More...
 
enum  hipsparseLtComputetype_t {
  HIPSPARSELT_COMPUTE_16F = 0 ,
  HIPSPARSELT_COMPUTE_32I ,
  HIPSPARSELT_COMPUTE_32F ,
  HIPSPARSELT_COMPUTE_TF32 ,
  HIPSPARSELT_COMPUTE_TF32_FAST
}
 Specify the compute precision modes of the matrix. More...
 
enum  hipsparseLtMatmulDescAttribute_t {
  HIPSPARSELT_MATMUL_ACTIVATION_RELU = 0 ,
  HIPSPARSELT_MATMUL_ACTIVATION_RELU_UPPERBOUND = 1 ,
  HIPSPARSELT_MATMUL_ACTIVATION_RELU_THRESHOLD = 2 ,
  HIPSPARSELT_MATMUL_ACTIVATION_GELU = 3 ,
  HIPSPARSELT_MATMUL_ACTIVATION_GELU_SCALING = 4 ,
  HIPSPARSELT_MATMUL_ALPHA_VECTOR_SCALING = 5 ,
  HIPSPARSELT_MATMUL_BETA_VECTOR_SCALING = 6 ,
  HIPSPARSELT_MATMUL_BIAS_STRIDE = 7 ,
  HIPSPARSELT_MATMUL_BIAS_POINTER = 8 ,
  HIPSPARSELT_MATMUL_ACTIVATION_ABS = 9 ,
  HIPSPARSELT_MATMUL_ACTIVATION_LEAKYRELU = 10 ,
  HIPSPARSELT_MATMUL_ACTIVATION_LEAKYRELU_ALPHA = 11 ,
  HIPSPARSELT_MATMUL_ACTIVATION_SIGMOID = 12 ,
  HIPSPARSELT_MATMUL_ACTIVATION_TANH = 13 ,
  HIPSPARSELT_MATMUL_ACTIVATION_TANH_ALPHA = 14 ,
  HIPSPARSELT_MATMUL_ACTIVATION_TANH_BETA = 15 ,
  HIPSPARSELT_MATMUL_BIAS_TYPE = 16 ,
  HIPSPARSELT_MATMUL_SPARSE_MAT_POINTER = 17
}
 Specify the additional attributes of a matrix multiplication descriptor. More...
 
enum  hipsparseLtMatmulAlg_t { HIPSPARSELT_MATMUL_ALG_DEFAULT }
 Specify the algorithm for matrix-matrix multiplication. More...
 
enum  hipsparseLtMatmulAlgAttribute_t {
  HIPSPARSELT_MATMUL_ALG_CONFIG_ID = 0 ,
  HIPSPARSELT_MATMUL_ALG_CONFIG_MAX_ID = 1 ,
  HIPSPARSELT_MATMUL_SEARCH_ITERATIONS = 2 ,
  HIPSPARSELT_MATMUL_SPLIT_K = 3 ,
  HIPSPARSELT_MATMUL_SPLIT_K_MODE = 4 ,
  HIPSPARSELT_MATMUL_SPLIT_K_BUFFERS = 5
}
 Specify the matrix multiplication algorithm attributes. More...
 
enum  hipsparseLtPruneAlg_t {
  HIPSPARSELT_PRUNE_SPMMA_TILE = 0 ,
  HIPSPARSELT_PRUNE_SPMMA_STRIP = 1
}
 Specify the pruning algorithm to apply to the structured matrix before the compression. More...
 
enum  hipsparseLtSplitKMode_t {
  HIPSPARSELT_SPLIT_K_MODE_ONE_KERNEL = 0 ,
  HIPSPARSELT_SPLIT_K_MODE_TWO_KERNELS = 1
}
 Specify the Split-K mode value. More...
 

Detailed Description

HIP = Heterogeneous-compute Interface for Portability

Define a extremely thin runtime layer that allows source code to be compiled unmodified through either AMD HCC or NVCC. Key features tend to be in the spirit and terminology of CUDA, but with a portable path to other GPUs as well.

This is the master include file for hipSPARSELt, wrapping around hipSPARSELt and cuSPARSELt "version 0.3".

Enumeration Type Documentation

◆ hipsparseLtComputetype_t

enum hipsparseLtComputetype_t

Specify the compute precision modes of the matrix.

Enumerator
HIPSPARSELT_COMPUTE_16F 

16-bit floating-point precision. CUDA backend only.

HIPSPARSELT_COMPUTE_32I 

32-bit integer precision.

HIPSPARSELT_COMPUTE_32F 

32-bit floating-point precision. HIP backend only.

HIPSPARSELT_COMPUTE_TF32 

32-bit floating point value, rounded to TF32 before the computation. CUDA backend only.

HIPSPARSELT_COMPUTE_TF32_FAST 

32-bit floating point value, truncated to TF32 before the computation. CUDA backend only.

◆ hipsparseLtMatDescAttribute_t

enum hipsparseLtMatDescAttribute_t

Specify the additional attributes of a matrix descriptor.

The hipsparseLtMatDescAttribute_t enumeration is used in the hipsparseLtMatDescSetAttribute and hipsparseLtMatDescGetAttribute functions.

Enumerator
HIPSPARSELT_MAT_NUM_BATCHES 

Number of matrices in a batch. READ/WRITE.

HIPSPARSELT_MAT_BATCH_STRIDE 

Stride between consecutive matrices in a batch expressed in terms of matrix elements. READ/WRITE.

◆ hipsparseLtMatmulAlg_t

enum hipsparseLtMatmulAlg_t

Specify the algorithm for matrix-matrix multiplication.

The hipsparseLtMatmulAlg_t enumeration is used in the hipsparseLtMatmulAlgSelectionInit function.

Enumerator
HIPSPARSELT_MATMUL_ALG_DEFAULT 

◆ hipsparseLtMatmulAlgAttribute_t

enum hipsparseLtMatmulAlgAttribute_t

Specify the matrix multiplication algorithm attributes.

The hipsparseLtMatmulAlgAttribute_t enumeration is used in the hipsparseLtMatmulAlgGetAttribute and hipsparseLtMatmulAlgSetAttribute functions.

Enumerator
HIPSPARSELT_MATMUL_ALG_CONFIG_ID 
HIPSPARSELT_MATMUL_ALG_CONFIG_MAX_ID 
HIPSPARSELT_MATMUL_SEARCH_ITERATIONS 
HIPSPARSELT_MATMUL_SPLIT_K 
HIPSPARSELT_MATMUL_SPLIT_K_MODE 
HIPSPARSELT_MATMUL_SPLIT_K_BUFFERS 

◆ hipsparseLtMatmulDescAttribute_t

enum hipsparseLtMatmulDescAttribute_t

Specify the additional attributes of a matrix multiplication descriptor.

The hipsparseLtMatmulDescAttribute_t enumeration is used in the hipsparseLtMatmulDescSetAttribute and hipsparseLtMatmulDescGetAttribute functions.

Enumerator
HIPSPARSELT_MATMUL_ACTIVATION_RELU 

ReLU activation function.

HIPSPARSELT_MATMUL_ACTIVATION_RELU_UPPERBOUND 

Upper bound of the ReLU activation function.

HIPSPARSELT_MATMUL_ACTIVATION_RELU_THRESHOLD 

Lower threshold of the ReLU activation function.

HIPSPARSELT_MATMUL_ACTIVATION_GELU 

GeLU activation function.

HIPSPARSELT_MATMUL_ACTIVATION_GELU_SCALING 

Scaling coefficient for the GeLU activation function. It implies GeLU is endable.

HIPSPARSELT_MATMUL_ALPHA_VECTOR_SCALING 

Enable/disable alpha vector (per-channel) scaling.

HIPSPARSELT_MATMUL_BETA_VECTOR_SCALING 

Enable/disable beta vector (per-channel) scaling.

HIPSPARSELT_MATMUL_BIAS_STRIDE 

Bias pointer. The Bias vector size must be equal to the number of rows of the output matrix (D).

HIPSPARSELT_MATMUL_BIAS_POINTER 

Bias stride between consecutive Bias vectors. 0 means broadcast the first bias vector.

HIPSPARSELT_MATMUL_ACTIVATION_ABS 

ABS activation function. HIP backend only.

HIPSPARSELT_MATMUL_ACTIVATION_LEAKYRELU 

LeakyReLU activation function. HIP backend only.

HIPSPARSELT_MATMUL_ACTIVATION_LEAKYRELU_ALPHA 

Alpha value of the LeakyReLU activation function. HIP backend only.

HIPSPARSELT_MATMUL_ACTIVATION_SIGMOID 

Sigmoid activation function. HIP backend only.

HIPSPARSELT_MATMUL_ACTIVATION_TANH 

Tanh activation function. HIP backend only.

HIPSPARSELT_MATMUL_ACTIVATION_TANH_ALPHA 

Alpha value of the Tanh activation function. HIP backend only.

HIPSPARSELT_MATMUL_ACTIVATION_TANH_BETA 

Beta value of the Tanh activation function. HIP backend only.

HIPSPARSELT_MATMUL_BIAS_TYPE 

Precision of Bias. HIP backend only.

  • When the input datatype is FP16, the Bias type can be FP16 or FP32. (default FP16)
  • When the input datatype is BF16, the Bias type can be BF16 or FP32. (default BF16)
  • In other cases, the Bias type is FP32.
HIPSPARSELT_MATMUL_SPARSE_MAT_POINTER 

Pointer to the pruned sparse matrix.

◆ hipsparseLtPruneAlg_t

enum hipsparseLtPruneAlg_t

Specify the pruning algorithm to apply to the structured matrix before the compression.

The hipsparseLtPruneAlg_t enumeration is used in the hipsparseLtSpMMAPrune and hipsparseLtSpMMAPrune2 functions.

Enumerator
HIPSPARSELT_PRUNE_SPMMA_TILE 

Zero out eight elements in a 4x4 tile. Non-zero elements have the maximum L1-norm value in all combinations in the tile. Exactly two elements in each row and column.

HIPSPARSELT_PRUNE_SPMMA_STRIP 

Zero out two elements in a 1x4 strip. Non-zero elements have the maximum L1-norm value in all combinations in the strip.

◆ hipsparseLtSparsity_t

enum hipsparseLtSparsity_t

Specify the sparsity of the structured matrix.

The enumerator specifies the sparsity ratio of the structured matrix as sparsity = nnz / total elements. The sparsity property is used in the hipsparseLtStructuredDescriptorInit function.

Enumerator
HIPSPARSELT_SPARSITY_50_PERCENT 

50% sparsity ratio:

  • 1:2 for TF32 and float.
  • 2:4 for half, bfloat16, and int.

◆ hipsparseLtSplitKMode_t

enum hipsparseLtSplitKMode_t

Specify the Split-K mode value.

The hipsparseLtSplitKMode_t enumeration is used by the HIPSPARSELT_MATMUL_SPLIT_K_MODE attribute in hipsparseLtMatmulAlgAttribute_t.

Enumerator
HIPSPARSELT_SPLIT_K_MODE_ONE_KERNEL 

Use the same SP-MM kernel to do the final reduction.

HIPSPARSELT_SPLIT_K_MODE_TWO_KERNELS 

Use another kernel to do the final reduction.