{ "blit.comp": { "useFP16": true, "macro": [ "C4" ] }, "blitregion.comp": { "useFP16": true }, "nchwTonc4hw4.comp": { "useFP16": true }, "nc4hw4Tonchw.comp": { "useFP16": true }, "reduce.comp": { "useFP16": true, "macro": [ "VMAX", "VMIN", "MEAN", "PROD", "SUM" ] }, "reduce_int.comp": { "useFP16": false, "macro": [ "VMAX", "VMIN", "MEAN", "PROD", "SUM" ] }, "binary.comp": { "useFP16": true, "macro": [ "ADD", "ATAN2", "SUB", "MUL", "DIV", "POW", "VMAX", "SQUDIFF", "VMIN", "LESS", "LESSEQUAL", "GREATER", "GREATEREQUAL", "EQUAL", "NOTEQUAL", "VMOD", "FLOORDIV", "FLOORMOD" ] }, "binary_blit.comp": { "useFP16": true, "macro": [ "ADD", "ATAN2", "SUB", "MUL", "DIV", "POW", "VMAX", "SQUDIFF", "VMIN", "LESS", "LESSEQUAL", "GREATER", "GREATEREQUAL", "EQUAL", "NOTEQUAL", "VMOD", "FLOORDIV", "FLOORMOD" ] }, "binary_int.comp": { "useFP16": false, "macro": [ "ADD", "SUB", "MUL", "DIV", "VMAX", "SQUDIFF", "VMIN", "LESS", "LESSEQUAL", "GREATER", "GREATEREQUAL", "EQUAL", "NOTEQUAL", "VMOD" ] }, "binary_blit_int.comp": { "useFP16": false, "macro": [ "ADD", "SUB", "MUL", "DIV", "VMAX", "SQUDIFF", "VMIN", "LESS", "LESSEQUAL", "GREATER", "GREATEREQUAL", "EQUAL", "NOTEQUAL", "VMOD" ] }, "unary.comp": { "useFP16": true, "macro": [ "SIGMOID", "TANH", "ABS", "SQRT", "RSQRT", "NEG", "SQUARE", "EXP", "SIGN", "LOG", "TAN", "COS", "SIN", "CEIL", "FLOOR", "EXPM1", "RECIPROCAL", "SINH", "ASINH", "ASIN", "COSH", "ACOS", "ACOSH", "ATAN", "ATANH", "LOG1P", "ROUND", "RELU", "CLAMP", "HARDSWISH", "GELU", "SILU" ] }, "unary_int.comp": { "useFP16": false, "macro": [ "ABS", "NEG", "SQUARE", "SIGN" ] }, "unPackImage4x4.comp": { "useFP16": false, "macro": [ "TRANSPOSE" ] }, "packAsImage4x4.comp": { "useFP16": false, "macro": [ "TRANSPOSE" ] }, "winogradDestTransform8x8.comp": { "useFP16": false, "macro": [ "RELU", "RELU6" ] }, "winogradTransformDest2_3_1.comp": { "useFP16": false, "macro": [ "RELU", "RELU6" ] }, "deconvIm2Col.comp": { "useFP16": false, "macro": [ "RELU", "RELU6" ] }, "col2Im.comp": { "useFP16": false, "macro": [ "RELU", "RELU6" ] }, "col2ImMali.comp": { "useFP16": false, "macro": [ "RELU", "RELU6" ] }, "convolution.comp": { "useFP16": true, "macro": [ "RELU", "RELU6" ] }, "convolutionint8.comp": { "useFP16": true, "macro": [ "RELU", "RELU6" ] }, "convolutionint4.comp": { "useFP16": true, "macro": [ "RELU", "RELU6" ] }, "deconvolution.comp": { "useFP16": true, "macro": [ "RELU", "RELU6" ] }, "convolutionDepthwise.comp": { "useFP16": true, "macro": [ "RELU", "RELU6" ] }, "dwweightcopy.comp": { "useFP16": true }, "deconvolutionDepthwise.comp": { "useFP16": true, "macro": [ "RELU", "RELU6" ] }, "norm.comp": { "useFP16": true, "macro": [ "LAYERNORM_SCALE" ] }, "norm_opt.comp": { "useFP16": true, "macro": [ "LAYERNORM_SCALE" ] }, "argmax.comp": { "useFP16": true, "macro": [ "ARGMIN" ] }, "convolutionDepthwiseMali.comp": { "useFP16": false, "macro": [ "RELU", "RELU6" ] }, "gridSampleNearest.comp": { "useFP16": true, "macro": [ "PAD_MODE_ZEROS" ] }, "gridSampleBilinear.comp": { "useFP16": true, "macro": [ "PAD_MODE_ZEROS" ] }, "resizeNearest.comp": { "useFP16": true, "macro": [ "NEAREST_ROUND" ] }, "resizeBilinear.comp": { "useFP16": true }, "range.comp": { "useFP16": true, "macro": [ "USE_INT" ] }, "cast_float_int.comp": { "useFP16": true, "macro": [ "REVERT" ] }, "scale.comp": { "useFP16": true }, "matmulunit.comp": { "useFP16": true }, "matmulunit_HAS_BIAS.comp": { "useFP16": true }, "softmaxHeight_NHWC.comp": { "useFP16": true }, "select.comp": { "useFP16": true }, "maxpool.comp": { "useFP16": true }, "avgpool.comp": { "useFP16": true }, "preluWithChannel.comp": { "useFP16": true }, "onehot.comp": { "useFP16": true }, "attention_fused.comp": { "useFP16": true }, "attention_fused_packed.comp": { "useFP16": true }, "attention_decode_q1_subgroup.comp": { "useFP16": true }, "attention_decode_q1_subgroup_hd128.comp": { "useFP16": true }, "attention_kvcache_update.comp": { "useFP16": true }, "attention_prefill_rearrange_q.comp": { "useFP16": true }, "attention_prefill_kblock_init_state.comp": { "useFP16": true }, "attention_prefill_kblock_qk.comp": { "useFP16": true }, "attention_prefill_kblock_qk_full.comp": { "useFP16": true }, "attention_prefill_kblock_softmax_online.comp": { "useFP16": true }, "attention_prefill_kblock_qkv_acc.comp": { "useFP16": true }, "attention_prefill_kblock_qkv_acc_full.comp": { "useFP16": true }, "attention_prefill_kblock_finalize.comp": { "useFP16": true }, "linear_attn_conv_silu.comp": { "useFP16": true }, "linear_attn_conv_state_update.comp": { "useFP16": true }, "linear_attn_qkv_prep.comp": { "useFP16": true }, "linear_attn_gated_delta_rule_prefill.comp": { "useFP16": true }, "linear_attn_gated_delta_rule_decode.comp": { "useFP16": true }, "linear_attn_gated_delta_rule_prefill_nosubgroup.comp": { "useFP16": true }, "linear_attn_gated_delta_rule_decode_nosubgroup.comp": { "useFP16": true }, "roipooling.comp": { "useFP16": true }, "C4_to_COOP.comp": { "useFP16": true }, "COOP_to_C4.comp": { "useFP16": true }, "matmul_coop.comp": { "useFP16": true }, "int4_weight_to_coop.comp": { "useFP16": true }, "int8_weight_to_coop.comp": { "useFP16": true }, "gemv_dequant_int4.comp": { "useFP16": true }, "gemv_dequant_int4_nosubgroup.comp": { "useFP16": true }, "gemv_dequant_int2.comp": { "useFP16": true }, "gemv_dequant_int2_nosubgroup.comp": { "useFP16": true }, "gemv_dequant_int3.comp": { "useFP16": true }, "gemv_dequant_int3_nosubgroup.comp": { "useFP16": true }, "gemv_dequant_int8.comp": { "useFP16": true }, "gemv_dequant_int8_nosubgroup.comp": { "useFP16": true }, "pack_a_k4m4_to_m64k4.comp": { "useFP16": true }, "int4_weight_to_pack.comp": { "useFP16": true }, "int8_weight_to_pack.comp": { "useFP16": true }, "int2_weight_to_pack.comp": { "useFP16": true }, "int3_weight_to_pack.comp": { "useFP16": true }, "conv1x1_int4_weight_prepare.comp": { "useFP16": false }, "conv1x1_int8_weight_prepare.comp": { "useFP16": false }, "conv1x1_int2_weight_prepare.comp": { "useFP16": false }, "conv1x1_int3_weight_prepare.comp": { "useFP16": false }, "conv1x1_quant_meta_prepare.comp": { "useFP16": true }, "gemm_m8n4.comp": { "useFP16": true }, "conv1x1_int8_weight_sumK.comp": { "useFP16": false }, "conv1x1_int4_weight_sumK.comp": { "useFP16": false }, "dynamic_int4_to_int8_unpack.comp": { "useFP16": false }, "dynamic_quant_minmax.comp": { "useFP16": true }, "dynamic_quant_reduce_minmax.comp": { "useFP16": true }, "dynamic_quant_pack.comp": { "useFP16": true }, "dynamic_quant_reduce_sum.comp": { "useFP16": false }, "dynamic_w8a8_coop_gemm.comp": { "useFP16": false }, "dynamic_w8a8_dequant_correction.comp": { "useFP16": true }, "topkv2.comp": { "useFP16": true, "macro": [ "SORT_DESC" ] } }