MI200 matrix fused multiply-add (MFMA) behavior specifics

The MI200 MFMA_F16, MFMA_BF16, and MFMA_BF16_1K flush subnormal input/output data to zero. This behavior might affect the convolution operation in certain workloads due to the limited exponent range of the half-precision floating-point datatypes.

MIOpen offers an alternate implementation for the half-precision datatype via conversion instructions to utilize the BFloat16 datatype’s larger exponent range, albeit with reduced accuracy. The following salients apply to this alternate implementation:

• It’s disabled by default in the forward convolution operations.

• It’s enabled by default in the backward data and backward weights convolution operations.

• You can override the default MIOpen behaviors by using the miopenSetConvolutionAttribute API call: Pass the convolution descriptor for the appropriate convolution operation, and the MIOPEN_CONVOLUTION_ATTRIB_FP16_ALT_IMPL convolution attribute (with a non-zero value), to engage the alternate implementation.

• You can also override the behavior using the MIOPEN_DEBUG_CONVOLUTION_ATTRIB_FP16_ALT_IMPL environment variable. When set to 1, MIOPEN_DEBUG_CONVOLUTION_ATTRIB_FP16_ALT_IMPL engages the alternate implementation; when set to 0, it’s disabled. Keep in mind that the environment variable impacts the convolution operation in all directions.