Qualcomm Incorporated

Huawei Technologies Co., Limited

Lin Chen

Andrew Evan Gruber

Yun Du

Guofang Jiao

Chihong Zhang

Chun Yu

Alexei Vladimirovich Bourd

Lingjun Chen

Ning Bi

David Rigel Garcia Garcia

Chunhui Mei

Liang Han

Long Chen

Junhong Sun

Zhou Hong

Liang Li

Eric Demers

Xiangdong Jin

Zilin Ying

Jing Wu

Pramod Vasant Argade

Richard Hammerstone

Sumesh Udayakumaran

Yufei Zhang

Hongjiang Shang

Chiente Ho

Jiaji Liu

Stewart Griffin Hall

Piyush Agarwal

Guofang Jia

A method of storing data in general purpose registers (GPRs) includes packing a tile of data items into GPRs, where the tile includes multiple channels. The tile of data items is read from memory. At least two channels of the data are stored in a first GPR, and at least two additional channels are stored in a second GPR. Auxiliary data is loaded into a third GPR. The auxiliary data and the tile data can be used together for performing convolution operations.

Storing complex data in warp GPRS

Techniques are described to perform a shuffle operation. Rather than using an all-lane to all-lane cross bar, a shuffler circuit having a smaller cross bar is described. The shuffler circuit performs the shuffle operation piecewise by reordering data received from processing lanes and outputting the reordered data.

Shuffler circuit for lane shuffle in SIMD architecture

Systems and techniques are disclosed for general purpose register dynamic allocation based on latency associated with of instructions in processor threads. A streaming processor can include a general purpose registers configured to stored data associated with threads, and a thread scheduler configured to receive allocation information for the general purpose registers, the information describing general purpose registers that are to be assigned as persistent general purpose registers (pGPRs) and volatile general purpose registers (vGPRs). The plurality of general purpose registers can be allocated according to the received information. The streaming processor can include the general purpose registers allocated according to the received information, the allocated based on execution latencies of instructions included in the threads.

General purpose register allocation in streaming processor

A GPU may be configured to detect and nullify unnecessary instructions. Nullifying unnecessary instructions include overwriting a detected unnecessary instruction with a no operation (NOP) instruction. In another example, nullifying unnecessary instructions may include writing a value to a 1-bit instruction memory. Each bit of the 1-bit instruction memory may be associated with a particular instruction of the draw call. If the 1-bit instruction memory has a true value (e.g., 1), the GPU is configured to not execute the particular instruction.

Dynamic shader instruction nullification for graphics processing

A SIMD processor may be configured to determine one or more active threads from a plurality of threads, select one active thread from the one or more active threads, and perform a divergent operation on the selected active thread. The divergent operation may be a serial operation.

Techniques for serialized execution in a SIMD processing system

A texture unit of a graphics processing unit (GPU) may receive a texture data. The texture unit may receive the texture data from the memory. The texture unit may also multiply, by a multiplier circuit of the texture unit, the texture data by at least one constant, where the constant is not associated with a filtering operation, and where the texture data comprises at least one texel. The texture unit may also output, by the texture unit, a result of multiplying the texture data by the at least one constant.

Constant multiplication with texture unit of graphics processing unit

A method for processing data in a graphics processing unit (GPU) including receiving an instance identifier for an instance and a shader program comprising a preamble code block and a main shader code block, assigning, the instance identifier to a general purpose register at wave creation, allocating address space within the constant memory for instance uniforms, and determining the preamble code block has not been executed and the wave is a first wave of the instance to be executed, based on determining the preamble code block has not been executed and the wave is the first wave to be executed, executing the preamble code block to store the plurality of instance uniforms in the constant memory and based, at least in part, on executing the preamble code block, executing the wave of the plurality of waves using at least one of the plurality of instance constants stored inconstant memory.

Per-instance preamble for graphics processing

A method for processing data in a graphics processing unit including receiving a code block of instructions common to a plurality of groups of threads of a shader, executing the code block of instructions common to the plurality of groups of threads of the shader creating a result by a first group of threads of the plurality of groups of threads, storing the result of the code block of instructions common to the plurality of groups of threads of the shader in on-chip random access memory (RAM), the on-chip RAM accessible by each of the plurality of groups of threads, and upon a determination that storing the result of the code block of instructions common to the plurality of groups of threads of the shader has completed, returning the result of the code block of instructions common to the plurality of groups of threads of the shader from on-chip RAM.

Per-shader preamble for graphics processing

In one example, a method includes responsive to receiving, by a processing unit, one or more instructions requesting that a first value be moved from a first general purpose register (GPR) to a third GPR and that a second value be moved from a second GPR to a fourth GPR, copying, by an initial logic unit and during a first clock cycle, the first value to an initial pipeline register, copying, by the initial logic and during a second clock cycle, the second value to the initial pipeline register, copying, by a final logic unit and during a third clock cycle, the first value from a final pipeline register to the third GPR, and copying, by the final logic unit and during a fourth clock cycle, the second value from the final pipeline register to the fourth GPR.

Utilizing pipeline registers as intermediate storage

Techniques are described in which an indication is included to indicate a last use of an intermediate value generated as part of determining a final value is not be stored in a general purpose register (GPR). A processing unit avoids storing the intermediate value in the GPR based on the indication because the intermediate value is no longer needed for determining the final value.

Skipping of data storage

A device includes a memory, and at least one programmable processor configured to determine, for each warp of a plurality of warps, whether a Boolean expression is true for a corresponding thread of each warp, pause execution of each warp having a corresponding thread for which the expression is true, determine a number of active threads for each of the plurality of warps for which the expression is true, sort the plurality of warps for which the expression is true based on the number of active threads in each of the plurality of warps, swap thread data of an active thread of a first warp of the plurality of warps with thread data of an inactive thread of a second warp of the plurality of warps, and resume execution of the at least one of the plurality of warps for which the expression is true.

GPU divergence barrier

At least one processor may emulate a fused multiply-add operation for a first operand, a second operand, and a third operand. The at least one processor may determine an intermediate value based at least in part on multiplying the first operand with the second operand, determine at least one of an upper intermediate value or a lower intermediate value, wherein determining the upper intermediate value comprises rounding, towards zero, the intermediate value by a specified number of bits, and wherein determining the lower intermediate value comprises subtracting the intermediate value by the upper intermediate value, determine an upper value and a lower value based at least in part on adding or subtracting the third operand to one of the upper intermediate value or the lower intermediate value, and determine an emulated fused multiply-add result by adding the upper value and the lower value.

Emulation of fused multiply-add operations

Techniques are described for determining whether data of a variable for each of a plurality of graphics items is same. If determined that the data is the same, the techniques store the data in a storage location of a specialized shared general purpose register that is associated with the variable.

Load scheme for shared register in GPU

Techniques described in the disclosure are generally related to reserving padding bytes in system memory when storing data in the system memory. The reserving of padding bytes may allow a memory interface to efficiently utilize the channels to the system memory when storing or subsequently retrieving the data.

Padding for multi-channel memory

This disclosure describes techniques for selectively activating a resume check operation in a single instruction, multiple data (SIMD) processing system. A processor is described that is configured to selectively enable or disable a resume check operation for a particular instruction based on information included in the instruction that indicates whether a resume check operation is to be performed for the instruction. A compiler is also described that is configured to generate compiled code which, when executed, causes a resume check operation to be selectively enabled or disabled for particular instructions. The compiled code may include one or more instructions that each specify whether a resume check operation is to be performed for the respective instruction. The techniques of this disclosure may be used to reduce the power consumption of and/or improve the performance of a SIMD system that utilizes a resume check operation to manage the reactivation of deactivated threads.

Growing community of inventors

San Diego, CA, United States of America

Lin Chen