Lines Matching refs:f16
24 // Their operands are only sort of f16 operands. Depending on
26 // values are really f16 converted to f32, so we treat these as f16
147 // At least one of the operands needs to be an fpextend of an f16
151 (f32 (fma_like (f32 (VOP3PMadMixModsExt f16:$src0, i32:$src0_mods)),
152 (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_mods)),
153 (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_mods)))),
157 (f32 (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_mods)),
158 (f32 (VOP3PMadMixModsExt f16:$src1, i32:$src1_mods)),
163 (f32 (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_mods)),
165 (f32 (VOP3PMadMixModsExt f16:$src2, i32:$src2_mods)))),
170 (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
171 (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
172 (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers))))),
184 …(build_vector f16:$elt0, (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifi…
185 … (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
186 … (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers)))))),
196 f16:$elt0,
197 (AMDGPUclamp (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
198 (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
199 (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers))))))),
209 (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$lo_src0, i32:$lo_src0_modifiers)),
210 (f32 (VOP3PMadMixMods f16:$lo_src1, i32:$lo_src1_modifiers)),
211 (f32 (VOP3PMadMixMods f16:$lo_src2, i32:$lo_src2_modifiers))))),
212 (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$hi_src0, i32:$hi_src0_modifiers)),
213 (f32 (VOP3PMadMixMods f16:$hi_src1, i32:$hi_src1_modifiers)),
214 (f32 (VOP3PMadMixMods f16:$hi_src2, i32:$hi_src2_modifiers))))))),
227 (f16 (fpround (fmul (f32 (VOP3PMadMixMods f32:$src0, i32:$src0_modifiers)),
237 …(build_vector f16:$elt0, (f16 (fpround (fmul (f32 (VOP3PMadMixMods f32:$src0, i32:$src0_modifiers)…
257 // Clamp modifier is applied after conversion to f16.
278 // Clamp modifier is applied after conversion to f16.
1018 // For f16 and bf16 matrices A and B, each element can be modified by
1021 // neg_lo = 1 (i4 and i8) signed(sext). For f16, bf16 and f32 matrix C each
1027 // wmma f32_bf16 | neg A/B (f16 or bf16) | neg_hi = 1 abs C(f32)
1029 // wmma f16_f16 | both neg_lo,neg_hi = 1 | neg_lo = 1 neg C(f16 or bf16)
1030 // wmma bf16_bf16 | neg A/B (f16 or bf16) | neg_hi = 1 abs C(f16 or bf16)
1039 // swmmac f32_bf16 | neg A/B (f16 or bf16) | A Index - matrix C is in dst
1042 // swmmac bf16_bf16 | neg A/B (f16 or bf16) | A Index - matrix C is in dst
1052 // fp8bf8 wmmas don't use src (0 and 1) modifiers, iu use neg_lo, f16 and bf16
1054 // remaining wmmas(f16, bf16 and f8bf8) use neg_lo and neg_hi for C (C is f32
1055 // f16 or bf16). swmmac use index_key and don't use src 2 modifiers.