[llvm] [AMDGPU] Add custom lowering of llvm.convert.from.arbitrary.fp for FP8 (PR #194144)

[Dmitry Sidorov via llvm-commits]

================
@@ -0,0 +1,84 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc < %s -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 | FileCheck -check-prefix=GFX1250 %s
+
+; llvm.convert.from.arbitrary.fp from FP8/BF8 producing f16 result types,
+; lowered via the gfx1250+ unscaled v_cvt_{pk_}f16_{fp8,bf8} instructions.
+
+; Scalar Float8E4M3FN
+define half @from_fp8_f16_s(i8 %x) {
+; GFX1250-LABEL: from_fp8_f16_s:
+; GFX1250:       ; %bb.0:
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_cvt_f16_fp8_e32 v0.l, v0
+; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
+  %r = call half @llvm.convert.from.arbitrary.fp.f16.i8(i8 %x, metadata !"Float8E4M3FN")
+  ret half %r
+}
+
+; Scalar Float8E5M2
+define half @from_bf8_f16_s(i8 %x) {
+; GFX1250-LABEL: from_bf8_f16_s:
+; GFX1250:       ; %bb.0:
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_cvt_f16_bf8_e32 v0.l, v0
+; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
+  %r = call half @llvm.convert.from.arbitrary.fp.f16.i8(i8 %x, metadata !"Float8E5M2")
+  ret half %r
+}
+
+; Vector Float8E4M3FN
+define <2 x half> @from_fp8_v2f16(<2 x i8> %x) {
+; GFX1250-LABEL: from_fp8_v2f16:
+; GFX1250:       ; %bb.0:
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_perm_b32 v0, v1, v0, 0x6050400
+; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX1250-NEXT:    v_cvt_pk_f16_fp8 v0, v0.l
+; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
+  %r = call <2 x half> @llvm.convert.from.arbitrary.fp.v2f16.v2i8(<2 x i8> %x, metadata !"Float8E4M3FN")
+  ret <2 x half> %r
+}
+
+; Vector Float8E5M2
+define <2 x half> @from_bf8_v2f16(<2 x i8> %x) {
+; GFX1250-LABEL: from_bf8_v2f16:
+; GFX1250:       ; %bb.0:
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_perm_b32 v0, v1, v0, 0x6050400
+; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX1250-NEXT:    v_cvt_pk_f16_bf8 v0, v0.l
+; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
+  %r = call <2 x half> @llvm.convert.from.arbitrary.fp.v2f16.v2i8(<2 x i8> %x, metadata !"Float8E5M2")
+  ret <2 x half> %r
+}
+
----------------
MrSidims wrote:

Added

https://github.com/llvm/llvm-project/pull/194144