[AMDGPU] Fixed some instructions latencies

[arsenm]

LGTM

[foad]

llvm/trunk/lib/Target/AMDGPU/VOP3Instructions.td

383

Hi @rampitec, was it intentional to set Write64Bit for V_MULLIT_F32?

[rampitec]

llvm/trunk/lib/Target/AMDGPU/VOP3Instructions.td

383

No, this is mistake, it is buried inside B64 block. This is full rate opcode.

[foad]

llvm/trunk/lib/Target/AMDGPU/VOP3Instructions.td

383

Fixed by 7d973307d5515b44df92f041916a0be25c1faa4b.

[rampitec]

llvm/trunk/lib/Target/AMDGPU/VOP3Instructions.td

383

Thanks!

llvm/trunk/lib/Target/AMDGPU/VOP1Instructions.td

 defm V_CVT_F64_U32 : VOP1Inst <"v_cvt_f64_u32", VOP1_F64_I32, uint_to_fp>;
 } // End SchedRW = [WriteQuarterRate32]
 
 defm V_FRACT_F32 : VOP1Inst <"v_fract_f32", VOP_F32_F32, AMDGPUfract>;
 defm V_TRUNC_F32 : VOP1Inst <"v_trunc_f32", VOP_F32_F32, ftrunc>;
 defm V_CEIL_F32 : VOP1Inst <"v_ceil_f32", VOP_F32_F32, fceil>;
 defm V_RNDNE_F32 : VOP1Inst <"v_rndne_f32", VOP_F32_F32, frint>;
 defm V_FLOOR_F32 : VOP1Inst <"v_floor_f32", VOP_F32_F32, ffloor>;
-defm V_EXP_F32 : VOP1Inst <"v_exp_f32", VOP_F32_F32, fexp2>;
 
 let SchedRW = [WriteQuarterRate32] in {
+defm V_EXP_F32 : VOP1Inst <"v_exp_f32", VOP_F32_F32, fexp2>;
 defm V_LOG_F32 : VOP1Inst <"v_log_f32", VOP_F32_F32, flog2>;
 defm V_RCP_F32 : VOP1Inst <"v_rcp_f32", VOP_F32_F32, AMDGPUrcp>;
 defm V_RCP_IFLAG_F32 : VOP1Inst <"v_rcp_iflag_f32", VOP_F32_F32>;
 defm V_RSQ_F32 : VOP1Inst <"v_rsq_f32", VOP_F32_F32, AMDGPUrsq>;
+defm V_SQRT_F32 : VOP1Inst <"v_sqrt_f32", VOP_F32_F32, fsqrt>;
 } // End SchedRW = [WriteQuarterRate32]
 
 let SchedRW = [WriteDouble] in {
 defm V_RCP_F64 : VOP1Inst <"v_rcp_f64", VOP_F64_F64, AMDGPUrcp>;
 defm V_RSQ_F64 : VOP1Inst <"v_rsq_f64", VOP_F64_F64, AMDGPUrsq>;
 } // End SchedRW = [WriteDouble];
 
-defm V_SQRT_F32 : VOP1Inst <"v_sqrt_f32", VOP_F32_F32, fsqrt>;
-
 let SchedRW = [WriteDouble] in {
 defm V_SQRT_F64 : VOP1Inst <"v_sqrt_f64", VOP_F64_F64, fsqrt>;
 } // End SchedRW = [WriteDouble]
 
 let SchedRW = [WriteQuarterRate32] in {
 defm V_SIN_F32 : VOP1Inst <"v_sin_f32", VOP_F32_F32, AMDGPUsin>;
 defm V_COS_F32 : VOP1Inst <"v_cos_f32", VOP_F32_F32, AMDGPUcos>;
 } // End SchedRW = [WriteQuarterRate32]
 
 defm V_NOT_B32 : VOP1Inst <"v_not_b32", VOP_I32_I32>;
 defm V_BFREV_B32 : VOP1Inst <"v_bfrev_b32", VOP_I32_I32>;
 defm V_FFBH_U32 : VOP1Inst <"v_ffbh_u32", VOP_I32_I32>;
 defm V_FFBL_B32 : VOP1Inst <"v_ffbl_b32", VOP_I32_I32>;
 defm V_FFBH_I32 : VOP1Inst <"v_ffbh_i32", VOP_I32_I32>;
-defm V_FREXP_EXP_I32_F64 : VOP1Inst <"v_frexp_exp_i32_f64", VOP_I32_F64, int_amdgcn_frexp_exp>;
 
 let SchedRW = [WriteDoubleAdd] in {
+defm V_FREXP_EXP_I32_F64 : VOP1Inst <"v_frexp_exp_i32_f64", VOP_I32_F64, int_amdgcn_frexp_exp>;
 defm V_FREXP_MANT_F64 : VOP1Inst <"v_frexp_mant_f64", VOP_F64_F64, int_amdgcn_frexp_mant>;
 defm V_FRACT_F64 : VOP1Inst <"v_fract_f64", VOP_F64_F64, AMDGPUfract>;
 } // End SchedRW = [WriteDoubleAdd]
 
 defm V_FREXP_EXP_I32_F32 : VOP1Inst <"v_frexp_exp_i32_f32", VOP_I32_F32, int_amdgcn_frexp_exp>;
 defm V_FREXP_MANT_F32 : VOP1Inst <"v_frexp_mant_f32", VOP_F32_F32, int_amdgcn_frexp_mant>;
 
 let VOPAsmPrefer32Bit = 1 in {

  // indexing from. Tablegen seems to not let you define an implicit
  // virtual register output for the super register being written into,
  // so this must have an implicit def of the register added to it.
 defm V_MOVRELD_B32 : VOP1Inst <"v_movreld_b32", VOP_MOVRELD>;
 defm V_MOVRELS_B32 : VOP1Inst <"v_movrels_b32", VOP_I32_VI32_NO_EXT>;
 defm V_MOVRELSD_B32 : VOP1Inst <"v_movrelsd_b32", VOP_NO_EXT<VOP_I32_I32>>;
 } // End Uses = [M0, EXEC]
 
-let SchedRW = [WriteQuarterRate32] in {
 defm V_MOV_FED_B32 : VOP1Inst <"v_mov_fed_b32", VOP_I32_I32>;
-}
 
 // These instruction only exist on SI and CI
 let SubtargetPredicate = isSICI in {
 
 let SchedRW = [WriteQuarterRate32] in {
 defm V_LOG_CLAMP_F32 : VOP1Inst <"v_log_clamp_f32", VOP_F32_F32, int_amdgcn_log_clamp>;
 defm V_RCP_CLAMP_F32 : VOP1Inst <"v_rcp_clamp_f32", VOP_F32_F32>;
 defm V_RCP_LEGACY_F32 : VOP1Inst <"v_rcp_legacy_f32", VOP_F32_F32, AMDGPUrcp_legacy>;

 
 
 let SubtargetPredicate = Has16BitInsts in {
 
 defm V_CVT_F16_U16 : VOP1Inst <"v_cvt_f16_u16", VOP1_F16_I16, uint_to_fp>;
 defm V_CVT_F16_I16 : VOP1Inst <"v_cvt_f16_i16", VOP1_F16_I16, sint_to_fp>;
 defm V_CVT_U16_F16 : VOP1Inst <"v_cvt_u16_f16", VOP_I16_F16, fp_to_uint>;
 defm V_CVT_I16_F16 : VOP1Inst <"v_cvt_i16_f16", VOP_I16_F16, fp_to_sint>;
+let SchedRW = [WriteQuarterRate32] in {
 defm V_RCP_F16 : VOP1Inst <"v_rcp_f16", VOP_F16_F16, AMDGPUrcp>;
 defm V_SQRT_F16 : VOP1Inst <"v_sqrt_f16", VOP_F16_F16, fsqrt>;
 defm V_RSQ_F16 : VOP1Inst <"v_rsq_f16", VOP_F16_F16, AMDGPUrsq>;
 defm V_LOG_F16 : VOP1Inst <"v_log_f16", VOP_F16_F16, flog2>;
 defm V_EXP_F16 : VOP1Inst <"v_exp_f16", VOP_F16_F16, fexp2>;
+defm V_SIN_F16 : VOP1Inst <"v_sin_f16", VOP_F16_F16, AMDGPUsin>;
+defm V_COS_F16 : VOP1Inst <"v_cos_f16", VOP_F16_F16, AMDGPUcos>;
+} // End SchedRW = [WriteQuarterRate32]
 defm V_FREXP_MANT_F16 : VOP1Inst <"v_frexp_mant_f16", VOP_F16_F16, int_amdgcn_frexp_mant>;
 defm V_FREXP_EXP_I16_F16 : VOP1Inst <"v_frexp_exp_i16_f16", VOP_I16_F16, int_amdgcn_frexp_exp>;
 defm V_FLOOR_F16 : VOP1Inst <"v_floor_f16", VOP_F16_F16, ffloor>;
 defm V_CEIL_F16 : VOP1Inst <"v_ceil_f16", VOP_F16_F16, fceil>;
 defm V_TRUNC_F16 : VOP1Inst <"v_trunc_f16", VOP_F16_F16, ftrunc>;
 defm V_RNDNE_F16 : VOP1Inst <"v_rndne_f16", VOP_F16_F16, frint>;
 defm V_FRACT_F16 : VOP1Inst <"v_fract_f16", VOP_F16_F16, AMDGPUfract>;
-defm V_SIN_F16 : VOP1Inst <"v_sin_f16", VOP_F16_F16, AMDGPUsin>;
-defm V_COS_F16 : VOP1Inst <"v_cos_f16", VOP_F16_F16, AMDGPUcos>;
 
 }
 
 let OtherPredicates = [Has16BitInsts] in {
 
 def : GCNPat<
     (f32 (f16_to_fp i16:$src)),
     (V_CVT_F32_F16_e32 $src)

llvm/trunk/lib/Target/AMDGPU/VOP3Instructions.td

 
 let isCommutable = 1 in {
 
 def V_MAD_LEGACY_F32 : VOP3Inst <"v_mad_legacy_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>;
 def V_MAD_F32 : VOP3Inst <"v_mad_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, fmad>;
 def V_MAD_I32_I24 : VOP3Inst <"v_mad_i32_i24", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
 def V_MAD_U32_U24 : VOP3Inst <"v_mad_u32_u24", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
 def V_FMA_F32 : VOP3Inst <"v_fma_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, fma>;
-def V_FMA_F64 : VOP3Inst <"v_fma_f64", VOP3_Profile<VOP_F64_F64_F64_F64>, fma>;
 def V_LERP_U8 : VOP3Inst <"v_lerp_u8", VOP3_Profile<VOP_I32_I32_I32_I32>, int_amdgcn_lerp>;
 
 let SchedRW = [WriteDoubleAdd] in {
+def V_FMA_F64 : VOP3Inst <"v_fma_f64", VOP3_Profile<VOP_F64_F64_F64_F64>, fma>;
 def V_ADD_F64 : VOP3Inst <"v_add_f64", VOP3_Profile<VOP_F64_F64_F64>, fadd, 1>;
 def V_MUL_F64 : VOP3Inst <"v_mul_f64", VOP3_Profile<VOP_F64_F64_F64>, fmul, 1>;
 def V_MIN_F64 : VOP3Inst <"v_min_f64", VOP3_Profile<VOP_F64_F64_F64>, fminnum, 1>;
 def V_MAX_F64 : VOP3Inst <"v_max_f64", VOP3_Profile<VOP_F64_F64_F64>, fmaxnum, 1>;
 } // End SchedRW = [WriteDoubleAdd]
 
 let SchedRW = [WriteQuarterRate32] in {
 def V_MUL_LO_U32 : VOP3Inst <"v_mul_lo_u32", VOP3_Profile<VOP_I32_I32_I32>>;

 let Constraints = "@earlyclobber $vdst" in {
 def V_MQSAD_PK_U16_U8 : VOP3Inst <"v_mqsad_pk_u16_u8", VOP3_Profile<VOP_I64_I64_I32_I64, VOP3_CLAMP>>;
 } // End Constraints = "@earlyclobber $vdst"
 
 def V_TRIG_PREOP_F64 : VOP3Inst <"v_trig_preop_f64", VOP3_Profile<VOP_F64_F64_I32>, AMDGPUtrig_preop> {
   let SchedRW = [WriteDouble];
 }
 
+let SchedRW = [Write64Bit] in {
 // These instructions only exist on SI and CI
 let SubtargetPredicate = isSICI in {
 def V_LSHL_B64 : VOP3Inst <"v_lshl_b64", VOP3_Profile<VOP_I64_I64_I32>>;
 def V_LSHR_B64 : VOP3Inst <"v_lshr_b64", VOP3_Profile<VOP_I64_I64_I32>>;
 def V_ASHR_I64 : VOP3Inst <"v_ashr_i64", VOP3_Profile<VOP_I64_I64_I32>>;
 def V_MULLIT_F32 : VOP3Inst <"v_mullit_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>;
 } // End SubtargetPredicate = isSICI
 
 let SubtargetPredicate = isVI in {
 def V_LSHLREV_B64 : VOP3Inst <"v_lshlrev_b64", VOP3_Profile<VOP_I64_I32_I64>>;
 def V_LSHRREV_B64 : VOP3Inst <"v_lshrrev_b64", VOP3_Profile<VOP_I64_I32_I64>>;
 def V_ASHRREV_I64 : VOP3Inst <"v_ashrrev_i64", VOP3_Profile<VOP_I64_I32_I64>>;
 } // End SubtargetPredicate = isVI
-
+} // End SchedRW = [Write64Bit]
 
 let SubtargetPredicate = isCIVI in {
 
-let Constraints = "@earlyclobber $vdst" in {
+let Constraints = "@earlyclobber $vdst", SchedRW = [WriteQuarterRate32] in {
 def V_QSAD_PK_U16_U8 : VOP3Inst <"v_qsad_pk_u16_u8", VOP3_Profile<VOP_I64_I64_I32_I64, VOP3_CLAMP>>;
 def V_MQSAD_U32_U8 : VOP3Inst <"v_mqsad_u32_u8", VOP3_Profile<VOP_V4I32_I64_I32_V4I32, VOP3_CLAMP>>;
-} // End Constraints = "@earlyclobber $vdst"
+} // End Constraints = "@earlyclobber $vdst", SchedRW = [WriteQuarterRate32]
 
 let isCommutable = 1 in {
-let SchedRW = [WriteDouble, WriteSALU] in {
+let SchedRW = [WriteQuarterRate32, WriteSALU] in {
 def V_MAD_U64_U32 : VOP3Inst <"v_mad_u64_u32", VOP3b_I64_I1_I32_I32_I64>;
 def V_MAD_I64_I32 : VOP3Inst <"v_mad_i64_i32", VOP3b_I64_I1_I32_I32_I64>;
 } // End SchedRW = [WriteDouble, WriteSALU]
 } // End isCommutable = 1
 
 } // End SubtargetPredicate = isCIVI
 
 

llvm/trunk/test/CodeGen/AMDGPU/llvm.exp2.f16.ll

entry:
   %a.val = load half, half addrspace(1)* %a
   %r.val = call half @llvm.exp2.f16(half %a.val)
   store half %r.val, half addrspace(1)* %r
   ret void
 }
 
 ; GCN-LABEL: {{^}}exp2_v2f16
 ; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
 ; SI:  v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
 ; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_exp_f32_e32 v[[R_F32_0:[0-9]+]], v[[A_F32_0]]
-; SI:  v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[R_F32_0]]
+; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
 ; SI:  v_exp_f32_e32 v[[R_F32_1:[0-9]+]], v[[A_F32_1]]
+; SI:  v_exp_f32_e32 v[[R_F32_0:[0-9]+]], v[[A_F32_0]]
 ; SI:  v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[R_F32_1]]
+; SI:  v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[R_F32_0]]
 ; SI-DAG: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[R_F16_1]]
 ; SI-NOT: and
 ; SI:  v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_0]], v[[R_F16_HI]]
 
 ; VI-DAG: v_exp_f16_e32 v[[R_F16_0:[0-9]+]], v[[A_V2_F16]]
 ; VI-DAG: v_exp_f16_sdwa v[[R_F16_1:[0-9]+]], v[[A_V2_F16]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1
 ; VI-NOT: and
 ; VI:     v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_0]], v[[R_F16_1]]

llvm/trunk/test/CodeGen/AMDGPU/llvm.sqrt.f16.ll

entry:
   %a.val = load half, half addrspace(1)* %a
   %r.val = call half @llvm.sqrt.f16(half %a.val)
   store half %r.val, half addrspace(1)* %r
   ret void
 }
 
 ; GCN-LABEL: {{^}}sqrt_v2f16
 ; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
 ; SI:  v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
 ; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_sqrt_f32_e32 v[[R_F32_0:[0-9]+]], v[[A_F32_0]]
-; SI:  v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[R_F32_0]]
+; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
 ; SI:  v_sqrt_f32_e32 v[[R_F32_1:[0-9]+]], v[[A_F32_1]]
+; SI:  v_sqrt_f32_e32 v[[R_F32_0:[0-9]+]], v[[A_F32_0]]
 ; SI:  v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[R_F32_1]]
+; SI:  v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[R_F32_0]]
 ; SI-DAG: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[R_F16_1]]
 ; SI-NOT: v_and_b32
 ; SI: v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_0]], v[[R_F16_HI]]
 
 ; VI-DAG: v_sqrt_f16_e32 v[[R_F16_0:[0-9]+]], v[[A_V2_F16]]
 ; VI-DAG: v_sqrt_f16_sdwa v[[R_F16_1:[0-9]+]], v[[A_V2_F16]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1
 ; VI-NOT: v_and_b32
 ; VI:     v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_0]], v[[R_F16_1]]

llvm/trunk/test/CodeGen/AMDGPU/mad_64_32.ll

define i63 @mad_i64_i32_sextops_i32_i63(i32 %arg0, i32 %arg1, i63 %arg2) #0 {
   %sext1 = sext i32 %arg1 to i63
   %mul = mul i63 %sext0, %sext1
   %mad = add i63 %mul, %arg2
   ret i63 %mad
 }
 
 ; GCN-LABEL: {{^}}mad_i64_i32_sextops_i31_i63:
 ; CI: v_lshl_b64
+; CI: v_bfe_i32 v[[B1:[0-9]+]], v1, 0, 31
 ; CI: v_ashr_i64
-; CI: v_bfe_i32 v1, v1, 0, 31
-; CI: v_bfe_i32 v0, v0, 0, 31
-; CI: v_mad_i64_i32 v[0:1], s[6:7], v0, v1, v[2:3]
+; CI: v_bfe_i32 v[[B2:[0-9]+]], v0, 0, 31
+; CI: v_mad_i64_i32 v[0:1], s[6:7], v[[B2]], v[[B1]], v[1:2]
 define i63 @mad_i64_i32_sextops_i31_i63(i31 %arg0, i31 %arg1, i63 %arg2) #0 {
   %sext0 = sext i31 %arg0 to i63
   %sext1 = sext i31 %arg1 to i63
   %mul = mul i63 %sext0, %sext1
   %mad = add i63 %mul, %arg2
   ret i63 %mad
 }
 

llvm/trunk/test/CodeGen/AMDGPU/mul.ll

 ; SI-DAG: s_mul_i32
 ; SI-DAG: v_mul_hi_u32
 ; SI-DAG: v_mul_hi_u32
 ; SI-DAG: s_mul_i32
 ; SI-DAG: s_mul_i32
 ; SI-DAG: v_mul_hi_u32
 
 ; VI: s_mul_i32
-; VI: v_mad_u64_u32
+; VI: v_mul_hi_u32
 ; VI: s_mul_i32
 ; VI: v_mul_hi_u32
 ; VI: v_mad_u64_u32
-; VI: v_mul_hi_u32
+; VI: v_mad_u64_u32
 ; VI: v_mad_u64_u32
 
 
 
 ; GCN: buffer_store_dwordx4
 define amdgpu_kernel void @s_mul_i128(i128 addrspace(1)* %out, i128 %a, i128 %b) nounwind #0 {
   %mul = mul i128 %a, %b
   store i128 %mul, i128 addrspace(1)* %out