// Copyright (c) 2019 Graphcore Ltd. All rights reserved. #ifndef GFLOAT_CONST_H #define GFLOAT_CONST_H #ifdef __IPU__ #include "poplar/AvailableVTypes.h" #endif #undef POPFLOAT_ENABLE_IPU_VECTORISED_OP #undef POPFLOAT_INITIALISE_VECTORS_FROM_INT #undef POPFLOAT_STORE_HALF_AS_SHORT /*! VERTICES */ #define POPFLOAT_VERTEX_CALC_PACKED_GF_PARAM_STRUCT 0 #define POPFLOAT_VERTEX_CALC_PACKED_GF_PARAM_MAN_BITS 1 #define POPFLOAT_VERTEX_CALC_PACKED_GF_PARAM_EXP_BITS 2 #define POPFLOAT_VERTEX_CALC_PACKED_GF_PARAM_EXP_BIAS 3 #define POPFLOAT_VERTEX_CALC_PACKED_GF_PARAM_EN_DENORM 4 #define POPFLOAT_VERTEX_CALC_PACKED_GF_PARAM_EN_INF 5 #define POPFLOAT_VERTEX_CALC_PACKED_GF_PARAM_PROP_NAN 6 #define POPFLOAT_VERTEX_CALC_PACKED_GF_PARAM_RMODE 7 #define POPFLOAT_VERTEX_CALC_PACKED_GF_PARAM_SR_BITS 8 /*! VBASE POINTER OFFSETS */ #define POPFLOAT_VBASE_GFLOAT_STRUCT_PTR_OFFSET 0 #define POPFLOAT_VBASE_CALC_GFLOAT_PARAM_PTR_OFFSET 1 #define POPFLOAT_VBASE_CAST_GFLOAT_PARAM_PTR_OFFSET 0 #define POPFLOAT_VBASE_CAST_INPUT_BASE_PTR_OFFSET 1 #define POPFLOAT_VBASE_CAST_OUTPUT_BASE_PTR_OFFSET 2 #if defined(VECTOR_AVAIL_SCALED_PTR64) #define POPFLOAT_VBASE_CAST_INPLACE_ELEMENTS_PER_WORKER_OFFSET 2 #define POPFLOAT_VBASE_CAST_INPLACE_LAST_WORKER_PARAM_OFFSET 3 #define POPFLOAT_VBASE_CAST_INPLACE_ROUNDING_PARAM_OFFSET 4 #define POPFLOAT_VBASE_CAST_ELEMENTS_PER_WORKER_OFFSET 3 #define POPFLOAT_VBASE_CAST_LAST_WORKER_PARAM_OFFSET 4 #define POPFLOAT_VBASE_CAST_ROUNDING_PARAM_OFFSET 5 #else #define POPFLOAT_VBASE_CAST_INPLACE_ELEMENTS_PER_WORKER_OFFSET 4 #define POPFLOAT_VBASE_CAST_INPLACE_LAST_WORKER_PARAM_OFFSET 5 #define POPFLOAT_VBASE_CAST_INPLACE_ROUNDING_PARAM_OFFSET 3 #define POPFLOAT_VBASE_CAST_ELEMENTS_PER_WORKER_OFFSET 6 #define POPFLOAT_VBASE_CAST_LAST_WORKER_PARAM_OFFSET 7 #define POPFLOAT_VBASE_CAST_ROUNDING_PARAM_OFFSET 4 #endif /*! CAST PARAMETERS USED BY NATIVE TO GFLOAT CODELETS */ #define POPFLOAT_CAST_PARAMS_SR_MASK_OFFSET 0 #define POPFLOAT_CAST_PARAMS_FP32_SCALE_IN_OFFSET 2 #define POPFLOAT_CAST_PARAMS_FP32_SCALE_OUT_OFFSET 4 #define POPFLOAT_CAST_PARAMS_FP32_CLAMP_OUT_OFFSET 6 #define POPFLOAT_CAST_PARAMS_FP32_DENSITY_PARAM_OFFSET 8 #define POPFLOAT_CAST_PARAMS_FP16_SCALE_IN_OFFSET 2 #define POPFLOAT_CAST_PARAMS_FP16_SCALE_OUT_OFFSET 3 #define POPFLOAT_CAST_PARAMS_FP16_CLAMP_OUT_OFFSET 4 #define POPFLOAT_CAST_PARAMS_FP16_DENSITY_PARAM_OFFSET 5 #define POPFLOAT_ROUND_PARAMS_TOTAL_SIZE 2 #define POPFLOAT_SR_ROUND_PARAMS_FP32_TOTAL_SIZE 9 #define POPFLOAT_SR_ROUND_PARAMS_FP16_TOTAL_SIZE 6 /*! CAST TO GFLOAT16 */ // Scratch: 32 bytes #define POPFLOAT_CAST_TO_GF16_STACK_OUTPUT_OFFSET 0 #define POPFLOAT_CAST_TO_GF16_STACK_CONST_ONE_OFFSET 0 #define POPFLOAT_CAST_TO_GF16_STACK_TRUNCATED_NORM_OFFSET 0 #define POPFLOAT_CAST_TO_GF16_STACK_GF8_EXP_MASK_OFFSET 0 #define POPFLOAT_CAST_TO_GF16_STACK_GF8_INPUT_OFFSET 0 #define POPFLOAT_CAST_TO_GF16_STACK_OUT_BITMASK_OFFSET 2 #define POPFLOAT_CAST_TO_GF16_STACK_GF8_SIGN_MASK_OFFSET 2 #define POPFLOAT_CAST_TO_GF16_STACK_GF8_IN_BASE_PTR_OFFSET 2 #define POPFLOAT_CAST_TO_GP16_STACK_GF8_OUT_BASE_PTR_OFFSET 3 #define POPFLOAT_CAST_TO_GF16_STACK_EXPONENT_OFFSET 4 #define POPFLOAT_CAST_TO_GF16_STACK_GF8_MAX_EXP_OFFSET 4 #define POPFLOAT_CAST_TO_GF16_STACK_SR_MASK_OFFSET 6 #define POPFLOAT_CAST_TO_GF16_STACK_HALF_MIN_OFFSET 6 #define POPFLOAT_CAST_TO_GF16_STACK_GF8_SCALED_IN_OFFSET 6 #define POPFLOAT_CAST_TO_GF16_TOTAL_STACK_SIZE 8 // GF16 Parameters #define POPFLOAT_CAST_TO_GF16_PARAM_EXPONENT_MASK_OFFSET 0 #define POPFLOAT_CAST_TO_GF16_PARAM_SIGN_MASK_OFFSET 2 #define POPFLOAT_CAST_TO_GF16_PARAM_QNAN_OUTPUT_OFFSET 4 #define POPFLOAT_CAST_TO_GF16_PARAM_NORM_MAN_MASK_OFFSET 6 #define POPFLOAT_CAST_TO_GF16_PARAM_MIN_DNRM_OFFSET 8 #define POPFLOAT_CAST_TO_GF16_PARAM_CLAMP_FP32_IN_OFFSET 10 #define POPFLOAT_CAST_TO_GF16_PARAM_SCALE_INPUT_OFFSET 12 #define POPFLOAT_CAST_TO_GF16_PARAM_SCALE_IN_RECIP_OFFSET 14 #define POPFLOAT_CAST_TO_GF16_PARAM_PACK_BITS_MASK_OFFSET 16 #define POPFLOAT_CAST_TO_GF16_PARAM_MAX_EXPONENT_OFFSET 18 #define POPFLOAT_CAST_TO_GF16_PARAM_CLAMP_FP16_IN_OFFSET 20 #define POPFLOAT_CAST_TO_GF16_PARAM_CLAMP_OUTPUT_OFFSET 21 #define POPFLOAT_CAST_TO_GF16_PARAM_MIN_OUTPUT_OFFSET 22 #define POPFLOAT_CAST_TO_GF16_PARAM_POWER2_M_MAN_10_OFFSET 23 #define POPFLOAT_CAST_TO_GF16_PARAM_POWER2_M1_OFFSET 24 #define POPFLOAT_CAST_TO_GF16_PARAM_POWER2_10_OFFSET 25 #define POPFLOAT_CAST_TO_GF16_PARAM_PACK_SHR_ALIGN_OFFSET 26 #define POPFLOAT_CAST_TO_GF16_PARAM_UNPACK_SHR_ALIGN_OFFSET 27 #define POPFLOAT_CAST_TO_GP16_PARAM_GF8_SIGN_MASK_OFFSET 28 #define POPFLOAT_CAST_TO_GF16_TOTAL_PARAM_SIZE 29 /*! CAST TO GFLOAT32 */ // Scratch 32 bytes #define POPFLOAT_CAST_TO_GF32_STACK_CORRECTION_OFFSET 0 #define POPFLOAT_CAST_TO_GF32_STACK_MAN_LSB_OFFSET 0 #define POPFLOAT_CAST_TO_GF32_STACK_HALFMIN_MASK_OFFSET 0 #define POPFLOAT_CAST_TO_GF32_STACK_UNPACK_INPUT_OFFSET 0 #define POPFLOAT_CAST_TO_GF32_STACK_GF8_SHUF8_OFFSET 0 #define POPFLOAT_CAST_TO_GF32_STACK_OUT_BITMASK_OFFSET 2 #define POPFLOAT_CAST_TO_GF32_STACK_IN_BASE_PTR_OFFSET 2 #define POPFLOAT_CAST_TO_GF32_STACK_OUT_BASE_PTR_OFFSET 3 #define POPFLOAT_CAST_TO_GF32_STACK_GF16_SIGN_OFFSET 4 #define POPFLOAT_CAST_TO_GF32_STACK_TRUNCATED_NORM_OFFSET 4 #define POPFLOAT_CAST_TO_GF32_STACK_SR_MASK_OFFSET 6 #define POPFLOAT_CAST_TO_GF32_STACK_TIE_MASK_OFFSET 6 #define POPFLOAT_CAST_TO_GF32_TOTAL_STACK_SIZE 8 // GF32 Parameters #define POPFLOAT_CAST_TO_GF32_PARAM_NORM_MANT_MASK_OFFSET 0 #define POPFLOAT_CAST_TO_GF32_PARAM_EXPONENT_MASK_OFFSET 2 #define POPFLOAT_CAST_TO_GF32_PARAM_SIGN_MASK_OFFSET 4 #define POPFLOAT_CAST_TO_GF32_PARAM_QNAN_MASK_OFFSET 6 #define POPFLOAT_CAST_TO_GF32_PARAM_SIGN_EXP_MASK_OFFSET 8 #define POPFLOAT_CAST_TO_GF32_PARAM_BIT23_MASK_OFFSET 10 #define POPFLOAT_CAST_TO_GF32_PARAM_CLAMP_OUTPUT_OFFSET 12 #define POPFLOAT_CAST_TO_GF32_PARAM_HALF_MIN_OFFSET 14 #define POPFLOAT_CAST_TO_GF32_PARAM_PACK_BITS_MASK_OFFSET 16 #define POPFLOAT_CAST_TO_GF32_PARAM_MIN_NORM_OFFSET 18 #define POPFLOAT_CAST_TO_GF32_PARAM_GF16_EXP_MASK_OFFSET 20 #define POPFLOAT_CAST_TO_GF32_PARAM_MIN_VALUE_OFFSET 22 #define POPFLOAT_CAST_TO_GF32_PARAM_EN_DENORM_OFFSET 23 #define POPFLOAT_CAST_TO_GF32_PARAM_PACK_EXP_ALIGN_OFFSET 24 #define POPFLOAT_CAST_TO_GF32_PARAM_PACK_SHR_ALIGN_OFFSET 25 #define POPFLOAT_CAST_TO_GF32_PARAM_UNPACK_EXP_ALIGN_OFFSET 26 #define POPFLOAT_CAST_TO_GF32_PARAM_UNPACK_SHIFT0_OFFSET 27 #define POPFLOAT_CAST_TO_GF32_PARAM_UNPACK_SHIFT1_OFFSET 28 #define POPFLOAT_CAST_TO_GF32_TOTAL_PARAM_SIZE 29 #define POPFLOAT_IPU_CLAMP_INDEX_MAX 0 #define POPFLOAT_IPU_CLAMP_INDEX_MIN 1 /*! PACKED GFLOAT STRUCTURE 4 bytes: [MANTISSA (8)] [EXPONENT (8)] [BIAS(8)] [|ENINF(1)|ENDNRM(1)] */ #define POPFLOAT_GF_STRUCT_MANTISSA_SIZE_OFFSET 0 #define POPFLOAT_GF_STRUCT_EXPONENT_SIZE_OFFSET 1 #define POPFLOAT_GF_STRUCT_EXP_BIAS_OFFSET 2 #define POPFLOAT_GF_STRUCT_PARAMS_OFFSET 3 #define POPFLOAT_GF_STRUCT_ENDENORM_BIT_OFFSET 0 #define POPFLOAT_GF_STRUCT_ENINF_BIT_OFFSET 1 #define POPFLOAT_GF_STRUCT_ALL_MAN_ZERO_BIT_OFFSET 2 #define POPFLOAT_GF_STRUCT_ALL_MAN_INF_BIT_OFFSET 3 #define POPFLOAT_GF_STRUCT_ENDENORM_MASK \ (1 << POPFLOAT_GF_STRUCT_ENDENORM_BIT_OFFSET) #define POPFLOAT_GF_STRUCT_ENINF_MASK (1 << POPFLOAT_GF_STRUCT_ENINF_BIT_OFFSET) #define POPFLOAT_GF_STRUCT_ALL_MAN_ZERO_MASK \ (1 << POPFLOAT_GF_STRUCT_ALL_MAN_ZERO_BIT_OFFSET) #define POPFLOAT_GF_STRUCT_ALL_MAN_INF_MASK \ (1 << POPFLOAT_GF_STRUCT_ALL_MAN_INF_BIT_OFFSET) #define POPFLOAT_BIT_MASK(BITS) ((1 << BITS) - 1) #define POPFLOAT_MAN_MASK(MAN) (POPFLOAT_BIT_MASK(MAN)) #define POPFLOAT_EXP_MASK(EXP, MAN) (POPFLOAT_BIT_MASK(EXP) << MAN) #define POPFLOAT_SGN_MASK(EXP, MAN) (1 << (EXP + MAN)) #define POPFLOAT_FP8_SIGN_MASK (1 << 7) #define POPFLOAT_FP8_V2_SIGN_MASK \ (POPFLOAT_FP8_SIGN_MASK | (POPFLOAT_FP8_SIGN_MASK << 8)) #define POPFLOAT_FP8_V4_SIGN_MASK \ (POPFLOAT_FP8_V2_SIGN_MASK | (POPFLOAT_FP8_V2_SIGN_MASK << 16)) #define POPFLOAT_NUM_FP16_MANTISSA_BITS (10) #define POPFLOAT_NUM_FP16_EXPONENT_BITS (5) #define POPFLOAT_FP16_EXPONENT_BIAS (15) #define POPFLOAT_FP16_MANTISSA_MASK \ POPFLOAT_MAN_MASK(POPFLOAT_NUM_FP16_MANTISSA_BITS) #define POPFLOAT_FP16_EXPONENT_MASK \ POPFLOAT_EXP_MASK(POPFLOAT_NUM_FP16_EXPONENT_BITS, \ POPFLOAT_NUM_FP16_MANTISSA_BITS) #define POPFLOAT_FP16_SIGN_MASK \ POPFLOAT_SGN_MASK(POPFLOAT_NUM_FP16_EXPONENT_BITS, \ POPFLOAT_NUM_FP16_MANTISSA_BITS) #define POPFLOAT_FP16_MAX_EXP POPFLOAT_FP16_EXPONENT_BIAS #define POPFLOAT_FP16_EXP_MAX_HALF \ (POPFLOAT_FP16_EXPONENT_BIAS + POPFLOAT_FP16_MAX_EXP) #define POPFLOAT_FP16_MIN_NORM (1 - POPFLOAT_FP16_EXPONENT_BIAS) #define POPFLOAT_FP16_MIN_DENORM \ (POPFLOAT_FP16_MIN_NORM - POPFLOAT_NUM_FP16_MANTISSA_BITS) #define POPFLOAT_FP16_GEN_QNAN 0x7C9C #define POPFLOAT_FP16_EXPONENT_SHIFT POPFLOAT_NUM_FP16_MANTISSA_BITS #define POPFLOAT_FP16_INFINITY_MASK \ (POPFLOAT_BIT_MASK(POPFLOAT_NUM_FP16_EXPONENT_BITS)) #define POPFLOAT_FP16_SIGN_SHIFT \ (POPFLOAT_FP16_EXPONENT_SHIFT + POPFLOAT_NUM_FP16_EXPONENT_BITS) #define POPFLOAT_FP16_Q_SHIFT (POPFLOAT_FP16_EXPONENT_SHIFT - 1) #define POPFLOAT_FP16_EXPONENT(v) \ (((v)&POPFLOAT_FP16_EXPONENT_MASK) >> POPFLOAT_FP16_EXPONENT_SHIFT) #define POPFLOAT_FP16_MANTISSA(v) ((v)&POPFLOAT_FP16_MANTISSA_MASK) #define POPFLOAT_FP16_SIGN(v) (((v) >> POPFLOAT_FP16_SIGN_SHIFT) & 1) #define POPFLOAT_FP16_IS_NEG(v) (POPFLOAT_FP16_SIGN(v) != 0) #define POPFLOAT_FP16_IS_ZERO(v) \ ((POPFLOAT_FP16_EXPONENT(v) == 0) && (POPFLOAT_FP16_MANTISSA(v) == 0)) #define POPFLOAT_FP16_IS_SUBNORM(v) \ ((POPFLOAT_FP16_EXPONENT(v) == 0) && (POPFLOAT_FP16_MANTISSA(v) != 0)) #define POPFLOAT_FP16_IS_INFINITY(v) \ ((POPFLOAT_FP16_EXPONENT(v) == POPFLOAT_FP16_INFINITY_MASK) && \ (POPFLOAT_FP16_MANTISSA(v) == 0)) #define POPFLOAT_FP16_IS_NAN(v) \ ((POPFLOAT_FP16_EXPONENT(v) == POPFLOAT_FP16_INFINITY_MASK) && \ (POPFLOAT_FP16_MANTISSA(v) != 0)) #define POPFLOAT_FP16_IS_QNAN(v) \ (POPFLOAT_FP16_IS_NAN(v) && (((v >> POPFLOAT_FP16_Q_SHIFT) & 1) == 1)) #define POPFLOAT_FP16_IS_SNAN(v) \ (POPFLOAT_FP16_IS_NAN(v) && (((v >> POPFLOAT_FP16_Q_SHIFT) & 1) == 0)) #define POPFLOAT_FP16_INFINITY \ (POPFLOAT_FP16_INFINITY_MASK << POPFLOAT_FP16_EXPONENT_SHIFT) #define POPFLOAT_NUM_FP32_MANTISSA_BITS (23) #define POPFLOAT_NUM_FP32_EXPONENT_BITS (8) #define POPFLOAT_FP32_EXPONENT_BIAS (127) #define POPFLOAT_FP32_MANTISSA_MASK \ POPFLOAT_MAN_MASK(POPFLOAT_NUM_FP32_MANTISSA_BITS) #define POPFLOAT_FP32_EXPONENT_MASK \ POPFLOAT_EXP_MASK(POPFLOAT_NUM_FP32_EXPONENT_BITS, \ POPFLOAT_NUM_FP32_MANTISSA_BITS) #define POPFLOAT_FP32_SIGN_MASK \ POPFLOAT_SGN_MASK(POPFLOAT_NUM_FP32_EXPONENT_BITS, \ POPFLOAT_NUM_FP32_MANTISSA_BITS) #define POPFLOAT_FP32_MAX_EXP POPFLOAT_FP32_EXPONENT_BIAS #define POPFLOAT_FP32_EXP_MAX_HALF \ (POPFLOAT_FP32_EXPONENT_BIAS + POPFLOAT_FP16_MAX_EXP) #define POPFLOAT_FP32_GEN_QNAN 0x7FD9C07E #define POPFLOAT_FP32_POWER2(N) \ ((N + POPFLOAT_FP32_EXPONENT_BIAS) << POPFLOAT_NUM_FP32_MANTISSA_BITS) #define POPFLOAT_FP32_EXPONENT_SHIFT POPFLOAT_NUM_FP32_MANTISSA_BITS #define POPFLOAT_FP32_INFINITY_MASK \ (POPFLOAT_BIT_MASK(POPFLOAT_NUM_FP32_EXPONENT_BITS)) #define POPFLOAT_FP32_SIGN_SHIFT \ (POPFLOAT_FP32_EXPONENT_SHIFT + POPFLOAT_NUM_FP32_EXPONENT_BITS) #define POPFLOAT_FP32_Q_SHIFT (POPFLOAT_FP32_EXPONENT_SHIFT - 1) #define POPFLOAT_FP32_EXPONENT(v) \ (((v)&POPFLOAT_FP32_EXPONENT_MASK) >> POPFLOAT_FP32_EXPONENT_SHIFT) #define POPFLOAT_FP32_MANTISSA(v) ((v)&POPFLOAT_FP32_MANTISSA_MASK) #define POPFLOAT_FP32_SIGN(v) (((v) >> POPFLOAT_FP32_SIGN_SHIFT) & 1) #define POPFLOAT_FP32_IS_NEG(v) (POPFLOAT_FP32_SIGN(v) != 0) #define POPFLOAT_FP32_IS_ZERO(v) \ ((POPFLOAT_FP32_EXPONENT(v) == 0) && (POPFLOAT_FP32_MANTISSA(v) == 0)) #define POPFLOAT_FP32_IS_SUBNORM(v) \ ((POPFLOAT_FP32_EXPONENT(v) == 0) && (POPFLOAT_FP32_MANTISSA(v) != 0)) #define POPFLOAT_FP32_IS_INFINITY(v) \ ((POPFLOAT_FP32_EXPONENT(v) == POPFLOAT_FP32_INFINITY_MASK) && \ (POPFLOAT_FP32_MANTISSA(v) == 0)) #define POPFLOAT_FP32_IS_NAN(v) \ ((POPFLOAT_FP32_EXPONENT(v) == POPFLOAT_FP32_INFINITY_MASK) && \ (POPFLOAT_FP32_MANTISSA(v) != 0)) #define POPFLOAT_FP32_IS_QNAN(v) \ (POPFLOAT_FP32_IS_NAN(v) && ((((v) >> POPFLOAT_FP32_Q_SHIFT) & 1) == 1)) #define POPFLOAT_FP32_IS_SNAN(v) \ (POPFLOAT_FP32_IS_NAN(v) && ((((v) >> POPFLOAT_FP32_Q_SHIFT) & 1) == 0)) #define POPFLOAT_FP32_INFINITY \ (POPFLOAT_FP32_INFINITY_MASK << POPFLOAT_FP32_EXPONENT_SHIFT) #define POPFLOAT_GF16_VEC_SIZE (4) #define POPFLOAT_GF32_VEC_SIZE (2) /*! Rounding modes */ #define POPFLOAT_ROUND_INVALID (0) #define POPFLOAT_ROUND_TO_ZERO (1) #define POPFLOAT_ROUND_TO_NEAREST_EVEN (2) #define POPFLOAT_ROUND_TO_NEAREST_AWAY (3) #define POPFLOAT_ROUND_TO_POS_INF (4) #define POPFLOAT_ROUND_TO_NEG_INF (5) #define POPFLOAT_ROUND_STOCHASTIC (6) #define POPFLOAT_ROUND_STOCHASTIC_EXTEND (7) /*! SR noise densities */ #define POPFLOAT_SR_DENSITY_INVALID (0) #define POPFLOAT_SR_DENSITY_UNIFORM (1) #define POPFLOAT_SR_DENSITY_NORMAL (2) #define POPFLOAT_SR_DENSITY_TRUNCATED_NORMAL (3) #define POPFLOAT_SR_DENSITY_BERNOULLI (4) #define POPFLOAT_SR_DENSITY_LOGISTIC (5) #define POPFLOAT_SR_DENSITY_TRUNCATED_LOGISTIC (6) #define POPFLOAT_SR_DENSITY_LAPLACE (7) #define POPFLOAT_SR_DENSITY_TRUNCATED_LAPLACE (8) #define POPFLOAT_SR_DENSITY_LOGIT_NORMAL (9) #define POPFLOAT_SR_DENSITY_TRUNCATED_LOGIT_NORMAL (10) /*! GF16 storage formats */ #define POPFLOAT_ENABLE_GF16_CLASS_FP16 #define POPFLOAT_ENABLE_GF16_CLASS_GF8 #define POPFLOAT_GF16_CLASS_FP16 (0) #ifdef POPFLOAT_ENABLE_GF16_CLASS_GF8 #define POPFLOAT_GF16_CLASS_FP8_1_5_2 (1) #define POPFLOAT_GF16_CLASS_FP8_MIN_NORM_ALIGN (2) #define POPFLOAT_GF16_CLASS_FP8_MAX_NORM_ALIGN (3) #define POPFLOAT_GF16_CLASS_COUNT (4) #define POPFLOAT_ENABLE_GF16_CLASS_FP8_1_5_2 #define POPFLOAT_ENABLE_GF16_CLASS_FP8_MIN_NORM_ALIGN #define POPFLOAT_ENABLE_GF16_CLASS_FP8_MAX_NORM_ALIGN #endif /*! GF32 storage formats */ #define POPFLOAT_ENABLE_GF32_CLASS_FP32 #define POPFLOAT_ENABLE_GF32_CLASS_FP16 #define POPFLOAT_ENABLE_GF32_CLASS_GF16 #define POPFLOAT_GF32_CLASS_FP32 (0) #define POPFLOAT_GF32_CLASS_FP16 (1) #ifdef POPFLOAT_ENABLE_GF32_CLASS_GF16 #define POPFLOAT_GF32_CLASS_BFLOAT (2) #define POPFLOAT_GF32_CLASS_GF16_EN_DENORM (3) #define POPFLOAT_GF32_CLASS_GF16_NO_DENORM (4) #define POPFLOAT_GF32_CLASS_COUNT (5) #define POPFLOAT_ENABLE_GF32_CLASS_BFLOAT #define POPFLOAT_ENABLE_GF32_CLASS_GF16_EN_DENORM #define POPFLOAT_ENABLE_GF32_CLASS_GF16_NO_DENORM #endif #endif