Name Strings

cl_intel_subgroup_2d_block_io

Contributors

Andrzej Ratajewski, Intel
Bartosz Kościelak, Intel
Ben Ashbaugh, Intel
Fangwen Fu, Intel
Grzegorz Kluczek, Intel
Junjie Gu, Intel
Lukasz Towarek, Intel
Roland Schulz, Intel

Notice

Copyright (c) 2025 Intel Corporation. All rights reserved.

Status

Shipping

Version

Built On: 2025-02-28
Version: 1.1.0

Dependencies

This extension is written against the OpenCL 3.0 C Language specification, V3.0.17.

This extension requires support for subgroups.

This extension depends on cl_intel_required_subgroup_size to query the subgroup sizes supported by a device or to require a subgroup size for a kernel.

Overview

This extension adds additional sub-group functions to read or prefetch two-dimensional blocks of data from a two-dimensional region of memory, or to write two-dimensional blocks of data to a two dimensional region of memory. This is an important operation for many machine learning algorithms, which operate on two-dimensional matrix data as part of a matrix multiplication algorithm.

This extension additionally adds support for two pre-processing operations that may be performed when reading a two-dimensional block of data:

  1. The two-dimensional block may be transposed after reading and before it is written to the instruction’s destination.

  2. The two-dimensional block may be transformed after reading and before it is written to the instruction’s destination. The transform operation converts the two-dimensional block from a row-major layout to a packed layout by combining data elements from multiple block rows into 32-bit values. This layout is used by some matrix multiplication functions.

New API Functions

None.

New API Enums

None.

New OpenCL C Enums

None.

New OpenCL C Functions

These functions are available to devices where the minimum subgroup size is 16. For these devices, the subgroup size must be 16 (the minimum supported subgroup size). Calling these functions on other devices or from kernels with a different subgroup size is undefined behavior.
Add 2d block read functions: 
// 8-bit data, rows in [1, 2, 4, 8, 16, 32], columns in [32]:

void intel_sub_group_2d_block_read_8b_1r32x1c(      // reads one ushort
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_2r32x1c(      // reads two ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_4r32x1c(      // reads four ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_8r32x1c(      // reads eight ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_16r32x1c(     // reads 16 ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_32r32x1c(     // reads 32 ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);

// 16-bit data, rows in [1, 2, 4, 8, 16, 32], columns in [16]:

void intel_sub_group_2d_block_read_16b_1r16x1c(     // reads one ushort
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_2r16x1c(     // reads two ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_4r16x1c(     // reads four ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_8r16x1c(     // reads eight ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_16r16x1c(    // reads 16 ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_32r16x1c(    // reads 32 ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);

// 32-bit data, rows in [1, 2, 4, 8, 16, 32], columns in [8, 16]:

void intel_sub_group_2d_block_read_32b_1r8x1c(      // reads one uint
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_2r8x1c(      // reads one uint
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_4r8x1c(      // reads two uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_8r8x1c(      // reads four uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_16r8x1c(     // reads eight uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_32r8x1c(     // reads 16 uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);

void intel_sub_group_2d_block_read_32b_1r16x1c(     // reads one uint
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_2r16x1c(     // reads two uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_4r16x1c(     // reads four uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_8r16x1c(     // reads eight uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_16r16x1c(    // reads 16 uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_32r16x1c(    // reads 32 uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);

// 8-bit data, rows in [1, 2, 4, 8, 16, 32], columns in [32x2]:

void intel_sub_group_2d_block_read_8b_1r32x2c(      // reads two ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_2r32x2c(      // reads four ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_4r32x2c(      // reads eight ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_8r32x2c(      // reads 16 ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_16r32x2c(     // reads 32 ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_32r32x2c(     // reads 64 ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);

// 16-bit data, rows in [1, 2, 4, 8, 16, 32], columns in [16x2]:

void intel_sub_group_2d_block_read_16b_1r16x2c(     // reads two ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_2r16x2c(     // reads four ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_4r16x2c(     // reads eight ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_8r16x2c(     // reads 16 ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_16r16x2c(    // reads 32 ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_32r16x2c(    // reads 64 ushorts
    global void* base_address,
    int width, int height, int pitch, int2 coord, private ushort* destination);

// 32-bit data, rows in [1, 2, 4, 8, 16, 32], columns in [8x2]:

void intel_sub_group_2d_block_read_32b_1r8x2c(      // reads two uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_2r8x2c(      // reads two uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_4r8x2c(      // reads four uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_8r8x2c(      // reads eight uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_16r8x2c(     // reads 16 uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_32r8x2c(     // reads 32 uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);

// 8-bit data, rows in [8, 16, 32], columns in [16x4]:

void intel_sub_group_2d_block_read_8b_8r16x4c(      // reads 32 uchars
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uchar* destination);
void intel_sub_group_2d_block_read_8b_16r16x4c(     // reads 64 uchars
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uchar* destination);
void intel_sub_group_2d_block_read_8b_32r16x4c(     // reads 128 uchars
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uchar* destination);

// 8-bit data with transform, rows in [32], columns in [16, 16x2, 16x4]:

void intel_sub_group_2d_block_read_transform_8b_32r16x1c(   // reads eight uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_transform_8b_32r16x2c(   // reads 16 uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_transform_8b_32r16x4c(   // reads 32 uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);

// 16-bit data with transform, rows in [16, 32], columns in [16, 16x2]:

void intel_sub_group_2d_block_read_transform_16b_16r16x1c(  // reads eight uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_transform_16b_16r16x2c(  // reads 16 uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_transform_16b_32r16x1c(  // reads 16 uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_transform_16b_32r16x2c(  // reads 32 uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);

// 32-bit data with transpose, rows in [16, 32], columns in [8]:

void intel_sub_group_2d_block_read_transpose_32b_16r8x1c(   // reads eight uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_transpose_32b_32r8x1c(   // reads 16 uints
    global void* base_address,
    int width, int height, int pitch, int2 coord, private uint* destination);
Add 2d block write functions: 
// 8-bit data, rows in [1, 2, 4, 8], columns in [16, 32]:

void intel_sub_group_2d_block_write_8b_1r16x1c(     // stores one uchar
    global void *base_address,
    int width, int height, int pitch, int2 coord, private uchar* value);
void intel_sub_group_2d_block_write_8b_2r16x1c(     // stores two uchars
    global void *base_address,
    int width, int height, int pitch, int2 coord, private uchar* value);
void intel_sub_group_2d_block_write_8b_4r16x1c(     // stores four uchars
    global void *base_address,
    int width, int height, int pitch, int2 coord, private uchar* value);
void intel_sub_group_2d_block_write_8b_8r16x1c(     // stores eight uchars
    global void *base_address,
    int width, int height, int pitch, int2 coord, private uchar* value);

void intel_sub_group_2d_block_write_8b_1r32x1c(     // stores two uchars
    global void *base_address,
    int width, int height, int pitch, int2 coord, private ushort* value);
void intel_sub_group_2d_block_write_8b_2r32x1c(     // stores four uchars
    global void *base_address,
    int width, int height, int pitch, int2 coord, private ushort* value);
void intel_sub_group_2d_block_write_8b_4r32x1c(     // stores eight uchars
    global void *base_address,
    int width, int height, int pitch, int2 coord, private ushort* value);
void intel_sub_group_2d_block_write_8b_8r32x1c(     // stores 16 uchars
    global void *base_address,
    int width, int height, int pitch, int2 coord, private ushort* value);

// 16-bit data, rows in [1, 2, 4, 8], columns in [16]:

void intel_sub_group_2d_block_write_16b_1r16x1c(    // stores one ushort
    global void *base_address,
    int width, int height, int pitch, int2 coord, private ushort*  value);
void intel_sub_group_2d_block_write_16b_2r16x1c(    // stores two ushorts
    global void *base_address,
    int width, int height, int pitch, int2 coord, private ushort* value);
void intel_sub_group_2d_block_write_16b_4r16x1c(    // stores four ushorts
    global void *base_address,
    int width, int height, int pitch, int2 coord, private ushort* value);
void intel_sub_group_2d_block_write_16b_8r16x1c(    // stores eight ushorts
    global void *base_address,
    int width, int height, int pitch, int2 coord, private ushort* value);

// 32-bit data, rows in [1, 2, 4, 8], columns in [16]:

void intel_sub_group_2d_block_write_32b_1r16x1c(    // stores one uint
    global void *base_address,
    int width, int height, int pitch, int2 coord, private uint*  value);
void intel_sub_group_2d_block_write_32b_2r16x1c(    // stores two uints
    global void *base_address,
    int width, int height, int pitch, int2 coord, private uint* value);
void intel_sub_group_2d_block_write_32b_4r16x1c(    // stores four uints
    global void *base_address,
    int width, int height, int pitch, int2 coord, private uint* value);
void intel_sub_group_2d_block_write_32b_8r16x1c(    // stores eight uints
    global void *base_address,
    int width, int height, int pitch, int2 coord, private uint* value);
Add 2d block prefetch functions: 
// 8-bit data, rows in [1, 2, 4, 8, 16, 32], columns in [32, 32x2]:

void intel_sub_group_2d_block_prefetch_8b_1r32x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_2r32x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_4r32x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_8r32x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_16r32x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_32r32x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

void intel_sub_group_2d_block_prefetch_8b_1r32x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_2r32x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_4r32x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_8r32x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_16r32x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_32r32x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

// 8-bit data, rows in [32], columns in [16, 16x2]:

void intel_sub_group_2d_block_prefetch_8b_32r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_32r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

// 8-bit data, rows in [8, 16, 32], columns in [16x4]:

void intel_sub_group_2d_block_prefetch_8b_8r16x4c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_16r16x4c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_32r16x4c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

// 16-bit data, rows in [1, 2, 4, 8, 16, 32], columns in [16, 16x2]:

void intel_sub_group_2d_block_prefetch_16b_1r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_2r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_4r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_8r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_16r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_32r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

void intel_sub_group_2d_block_prefetch_16b_1r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_2r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_4r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_8r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_16r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_32r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

// 32-bit data, rows in [1, 2, 4, 8, 16, 32], columns in [8, 16, 8x2]:

void intel_sub_group_2d_block_prefetch_32b_1r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_2r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_4r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_8r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_16r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_32r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

void intel_sub_group_2d_block_prefetch_32b_1r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_2r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_4r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_8r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_16r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_32r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

void intel_sub_group_2d_block_prefetch_32b_1r8x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_2r8x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_4r8x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_8r8x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_16r8x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_32r8x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

Modifications to the OpenCL C Specification

Add a new Section 6.13.X. "Sub-Group 2D Block IO Functions":

Section 6.13.X.1 Sub-Group 2D Block Read Functions

These functions read one or more 2D blocks of data from a 2D row-major region of global memory. The 2D blocks of data are read collectively, as a sub-group operation. Please refer to the SPV_INTEL_2d_block_io extension for information how the 2D block data is assigned to work-items in the sub-group.

Function Description 
void intel_sub_group_2d_block_read_8b_1r32x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_2r32x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_4r32x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_8r32x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_16r32x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_32r32x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_1r32x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_2r32x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_4r32x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_8r32x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_16r32x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_8b_32r32x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);

Reads one or more row by column blocks of data from the specified region of global memory at the coordinate specified by coord as a sub-group operation. The region of memory to read from is specified by base_address, width, height, and pitch.

The blocks of data are adjacent horizontally, so the total number of columns read is number of columns in one block multiplied by number of blocks.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Since the block has 32 columns, each work item reads two data elements per block row, and packs them into a ushort. Each work item in the sub-group reads 16-bits of data from each row for each block.

 
void intel_sub_group_2d_block_read_8b_8r16x4c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uchar* destination);
void intel_sub_group_2d_block_read_8b_16r16x4c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uchar* destination);
void intel_sub_group_2d_block_read_8b_32r16x4c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uchar* destination);

Reads one or more row by column blocks of data from the specified region of global memory at the coordinate specified by coord as a sub-group operation. The region of memory to read from is specified by base_address, width, height, and pitch.

The blocks of data are read horizontally, so the total number of columns read is the number of columns in one block multiplied by the number of blocks.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Since the block has 16 columns, each work item reads one data element per block row. Each work item in the sub-group reads 8-bits of data from each row, from each block.

 
void intel_sub_group_2d_block_read_16b_1r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_2r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_4r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_8r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_16r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_32r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_1r16x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_2r16x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_4r16x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_8r16x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_16r16x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);
void intel_sub_group_2d_block_read_16b_32r16x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private ushort* destination);

Reads one or more row by column blocks of data from the specified region of global memory at the coordinate specified by coord as a sub-group operation. The region of memory to read from is specified by base_address, width, height, and pitch.

The blocks of data are read horizontally, so the total number of columns read is the number of columns in one block multiplied by the number of blocks.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Since the block has 16 columns, each work item reads one data element per block row. For each block, each work item in the sub-group reads one data element from each row, from each block.

 
void intel_sub_group_2d_block_read_32b_1r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_2r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_4r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_8r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_16r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_32r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);

Reads one or more row by column blocks of data from the specified region of global memory at the coordinate specified by coord as a sub-group operation. The region of memory to read from is specified by base_address, width, height, and pitch.

The blocks of data are read horizontally, so the total number of columns read is the number of columns in one block multiplied by the number of blocks.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Since the block has 16 columns, each work item reads one data element per block row. For each block, each work item in the sub-group reads one data element per block row, from each block.

 
void intel_sub_group_2d_block_read_32b_1r8x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_2r8x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_4r8x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_8r8x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_16r8x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_32r8x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_1r8x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_2r8x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_4r8x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_8r8x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_16r8x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_32b_32r8x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);

Reads one or more row by column blocks of data from the specified region of global memory at the coordinate specified by coord as a sub-group operation. The region of memory to read from is specified by base_address, width, height, and pitch.

The blocks of data are read horizontally, so the total number of columns read is the number of columns in one block multiplied by the number of blocks.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Since the block has 8 columns, the first eight work items receives data from odd rows, and the next eight work-items receives data from even rows. If there is only one row, the data assigned to the last eight work-items is undefined.

 
void intel_sub_group_2d_block_read_transform_8b_32r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_transform_8b_32r16x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);

Reads one or more row by column blocks of data from the specified region of global memory at the coordinate specified by coord as a sub-group operation and performs a packing transformation. The region of memory to read from is specified by base_address, width, height, and pitch.

The blocks of data are read horizontally, so the total number of columns read is the number of columns in one block multiplied by the number of blocks.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Since the block has 16 columns, each work item in the sub-group reads one column of data. Values from first four rows of data are packed into the first component of the return value, then values from next four rows, and so on.

 
void intel_sub_group_2d_block_read_transform_16b_16r16x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_transform_16b_16r16x2c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);

Reads one or more row by column blocks of data from the specified region of global memory at the coordinate specified by coord as a sub-group operation and performs a packing transformation. The region of memory to read from is specified by base_address, width, height, and pitch.

The blocks of data are read horizontally, so the total number of columns read is the number of columns in one block multiplied by the number of blocks.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Since the block has 16 columns, each work item in the sub-group reads one column of data. Values from first two rows of data are packed into the first component of the return value, then values from next two rows, and so on.

 
void intel_sub_group_2d_block_read_transpose_32b_16r8x1c(
    global void* base_address,
    int width, int height, int pitch,
    int2 coord, private uint* destination);
void intel_sub_group_2d_block_read_transpose_32b_32r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord,
    private uint* destination);

Reads a row by column blocks of data from the specified region of global memory at the coordinate specified by coord as a sub-group operation and transposes the data before assigning to work-items. The region of memory to read from is specified by base_address, width, height, and pitch.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Since the block has 16 or 32 rows pre-transpose, which becomes 16 or 32 columns of data post-transpose, each work-item in the subgroup reads one or two columns of data.

Section 6.13.X.2 Sub-Group 2D Block Prefetch Functions

Function Description 
void intel_sub_group_2d_block_prefetch_8b_1r32x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_2r32x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_4r32x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_8r32x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_16r32x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_32r32x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

void intel_sub_group_2d_block_prefetch_8b_1r32x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_2r32x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_4r32x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_8r32x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_16r32x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_32r32x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

void intel_sub_group_2d_block_prefetch_8b_32r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_32r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

void intel_sub_group_2d_block_prefetch_8b_8r16x4c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_16r16x4c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_8b_32r16x4c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);


void intel_sub_group_2d_block_prefetch_16b_1r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_2r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_4r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_8r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_16r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_32r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

void intel_sub_group_2d_block_prefetch_16b_1r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_2r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_4r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_8r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_16r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_16b_32r16x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

void intel_sub_group_2d_block_prefetch_32b_1r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_2r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_4r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_8r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_16r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_32r16x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

void intel_sub_group_2d_block_prefetch_32b_1r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_2r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_4r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_8r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_16r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_32r8x1c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

void intel_sub_group_2d_block_prefetch_32b_1r8x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_2r8x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_4r8x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_8r8x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_16r8x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);
void intel_sub_group_2d_block_prefetch_32b_32r8x2c(
    global void* base_address,
    int width, int height, int pitch, int2 coord);

Prefetches one or more row by column blocks of data from the specified region of global memory at the coordinate specified by coord as a sub-group operation and performs a packing transformation. Prefetching does not affect the functionality of a kernel but may change its performance characteristics. The region of memory to prefetch from is specified by base_address, width, height, and pitch.

The blocks of data are prefetched horizontally, so the total number of columns prefetched is the number of columns in one block multiplied by the number of blocks.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Section 6.13.X.3 Sub-Group 2D Block Write Functions

Function Description 
void intel_sub_group_2d_block_write_8b_1r32x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private ushort* value);
void intel_sub_group_2d_block_write_8b_2r32x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private ushort* value);
void intel_sub_group_2d_block_write_8b_4r32x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private ushort* value);
void intel_sub_group_2d_block_write_8b_8r32x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private ushort* value);

Writes a row by column block of data to the specified region of global memory at the coordinate specified by coord as a sub-group operation. The region of memory to write to is specified by base_address, width, height, and pitch.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Since the block has 32 columns, each work-item writes two data elements per block row.

 
void intel_sub_group_2d_block_write_8b_1r16x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private uchar* value);
void intel_sub_group_2d_block_write_8b_2r16x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private uchar* value);
void intel_sub_group_2d_block_write_8b_4r16x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private uchar* value);
void intel_sub_group_2d_block_write_8b_8r16x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private uchar* value);

Writes a row by column block of data to the specified region of global memory at the coordinate specified by coord as a sub-group operation. The region of memory to write to is specified by base_address, width, height, and pitch.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Since the block has 16 columns, each work-item writes one data element per block row.

 
void intel_sub_group_2d_block_write_16b_1r16x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private ushort*  value);
void intel_sub_group_2d_block_write_16b_2r16x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private ushort* value);
void intel_sub_group_2d_block_write_16b_4r16x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private ushort* value);
void intel_sub_group_2d_block_write_16b_8r16x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private ushort* value);

Writes a row by column block of data to the specified region of global memory at the coordinate specified by coord as a sub-group operation. The region of memory to write to is specified by base_address, width, height, and pitch.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Since the block has 16 columns, each work-item writes one data element per block row.

 
void intel_sub_group_2d_block_write_32b_1r16x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private uint*  value);
void intel_sub_group_2d_block_write_32b_2r16x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private uint* value);
void intel_sub_group_2d_block_write_32b_4r16x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private uint* value);
void intel_sub_group_2d_block_write_32b_8r16x1c(
    global void *base_address, int width, int height,
    int pitch, int2 coord, private uint* value);

Writes a row by column block of data to the specified region of global memory at the coordinate specified by coord as a sub-group operation. The region of memory to write to is specified by base_address, width, height, and pitch.

Note that coord is provided in elements, while width and pitch are provided in bytes.

Since the block has 16 columns, each work-item writes one data element per block row.

6.13.X.6 Restrictions

The following restrictions apply to the sub-group 2d block read, write, and prefetch functions added by this extension:

Behavior is undefined unless:

  • the sub-group size is 16.

  • the first component of coord is a multiple of four for 8-bit data, or a multiple of two for 16-bit data.

  • the per-subgroup base_address is cache-line aligned (64 Bytes).

  • the width is greater than or equal 64 Bytes and less than or equal to 224 bytes.

  • the width is a multiple of four for 8-bit or 16-bit data, or a multiple of the data size otherwise.

  • the height is greater than zero and less than or equal to 224.

  • the pitch is greater than or equal to the width and a multiple of 16 bytes.

  • the sub-group size is equal to the maximum sub-group size; in other words, this is a full sub-group.

Modifications to the OpenCL SPIR-V Environment Specification

SPIR-V support was added in extension version 1.1.0.

Add a new section 5.2.X - cl_intel_subgroup_2d_block_io

If the OpenCL environment supports the extension cl_intel_subgroup_2d_block_io then the environment must accept modules that declare use of the extension SPV_INTEL_2d_block_io and that declare the following SPIR-V capabilities:

  • Subgroup2DBlockIOINTEL

  • Subgroup2DBlockTransformINTEL

  • Subgroup2DBlockTransposeINTEL

The table below describes valid 2D block load and store dimensions for different element sizes:

Instruction Pointer Type Element Size (Bytes) Block Width (Elements) Block Height (Rows) Block Count Notes

Block Loads:

OpSubgroup2DBlockLoadINTEL

uint8_t, void (untyped)

1

32

1, 2, 4, 8, 16, 32

1, 2

OpSubgroup2DBlockLoadINTEL

uint8_t, void (untyped)

1

16

8, 16, 32

4

For loading 8-bit data then up-converting.

OpSubgroup2DBlockLoadINTEL

uint16_t, void (untyped)

2

16

1, 2, 4, 8, 16, 32

1, 2

OpSubgroup2DBlockLoadINTEL

uint32_t, void (untyped)

4

8

1, 2, 4, 8, 16, 32

1, 2

OpSubgroup2DBlockLoadINTEL

uint32_t, void (untyped)

4

16

1, 2, 4, 8, 16, 32

1

Block Loads with Transform:

OpSubgroup2DBlockLoadTransformINTEL

uint8_t, void (untyped)

1

16

32

1, 2, 4

OpSubgroup2DBlockLoadTransformINTEL

uint16_t, void (untyped)

2

16

16, 32

1, 2

Block Loads with Transpose:

OpSubgroup2DBlockLoadTransposeINTEL

uint32_t, void (untyped)

4

8

16, 32

1

Dimensions are in memory, pre-transpose.

Block Stores:

OpSubgroup2DBlockStoreINTEL

uint8_t, void (untyped)

1

16, 32

1, 2, 4, 8

1

OpSubgroup2DBlockStoreINTEL

uint16_t, void (untyped)

2

16

1, 2, 4, 8

1

OpSubgroup2DBlockStoreINTEL

uint32_t, void (untyped)

4

16

1, 2, 4, 8

1

Block Prefetch:

OpSubgroup2DBlockPrefetchINTEL

uint8_t, void (untyped)

1

32

1, 2, 4, 8, 16, 32

1, 2

OpSubgroup2DBlockPrefetchINTEL

uint8_t, void (untyped)

1

16

32

1, 2

OpSubgroup2DBlockPrefetchINTEL

uint8_t, void (untyped)

1

16

8, 16, 32

4

OpSubgroup2DBlockPrefetchINTEL

uint16_t, void (untyped)

2

16

1, 2, 4, 8, 16, 32

1, 2

OpSubgroup2DBlockPrefetchINTEL

uint32_t, void (untyped)

4

8

1, 2, 4, 8, 16, 32

1, 2

OpSubgroup2DBlockPrefetchINTEL

uint32_t, void (untyped)

4

16

1, 2, 4, 8, 16, 32

1

For all instructions:

  • The Memory Width, Memory Height, and Memory Pitch operands must be 32-bit integer type scalars.

  • The Coordinate must be a vector of two 32-bit integer type components.

Issues

None.

Revision History

Rev Date Author Changes

1.0.0

2024-12-03

Bartosz Kościelak

Initial revision

1.1.0

2024-02-28

Ben Ashbaugh

Added SPIR-V support.