[llvm-branch-commits] [clang] [llvm] AMDGPU: Builtins & CodeGen support for v_cvt_scalef32_pk_{f|bf}16_fp4 for gfx950 (PR #117744)

Matt Arsenault via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Tue Nov 26 12:01:02 PST 2024


https://github.com/arsenm updated https://github.com/llvm/llvm-project/pull/117744

>From b5c2479e4d0e87b5a866b783b70cf371c3b94aff Mon Sep 17 00:00:00 2001
From: Pravin Jagtap <Pravin.Jagtap at amd.com>
Date: Mon, 22 Apr 2024 01:58:18 -0400
Subject: [PATCH] AMDGPU: Builtins & CodeGen support for
 v_cvt_scalef32_pk_{f|bf}16_fp4 for gfx950

OPSEL ASM Syntax for v_cvt_scalef32_pk_{f|bf}16_fp4 : opsel:[x,y,z]
where, x & y i.e. OPSEL[1 : 0] selects which src_byte to read.

Note: Conventional Inst{13} i.e. OPSEL[2] is ignored in asm syntax.

Co-authored-by: Pravin Jagtap <Pravin.Jagtap at amd.com>
---
 clang/include/clang/Basic/BuiltinsAMDGPU.def  |  2 +
 .../builtins-amdgcn-gfx950-err.cl             |  4 +-
 .../CodeGenOpenCL/builtins-amdgcn-gfx950.cl   | 76 +++++++++++++++++
 .../builtins-amdgcn-error-gfx950-param.cl     |  4 +-
 llvm/include/llvm/IR/IntrinsicsAMDGPU.td      | 12 ++-
 .../Target/AMDGPU/AMDGPURegisterBankInfo.cpp  |  2 +
 llvm/lib/Target/AMDGPU/VOP3Instructions.td    |  2 +
 .../llvm.amdgcn.cvt.scalef32.pk.gfx950.ll     | 82 +++++++++++++++++++
 8 files changed, 178 insertions(+), 6 deletions(-)

diff --git a/clang/include/clang/Basic/BuiltinsAMDGPU.def b/clang/include/clang/Basic/BuiltinsAMDGPU.def
index 629a26edbb7ca0..d55cdcb8513dab 100644
--- a/clang/include/clang/Basic/BuiltinsAMDGPU.def
+++ b/clang/include/clang/Basic/BuiltinsAMDGPU.def
@@ -584,6 +584,8 @@ TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk_bf8_f16, "V2sV2sV2hfIb", "nc", "
 TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk_bf8_bf16, "V2sV2sV2yfIb", "nc", "bf8-cvt-scale-insts")
 TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk_f32_fp4, "V2fUifIi", "nc", "fp4-cvt-scale-insts")
 TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk_fp4_f32, "UiUifffIi", "nc", "fp4-cvt-scale-insts")
+TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk_f16_fp4, "V2hUifIi", "nc", "fp4-cvt-scale-insts")
+TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk_bf16_fp4, "V2yUifIi", "nc", "fp4-cvt-scale-insts")
 
 #undef BUILTIN
 #undef TARGET_BUILTIN
diff --git a/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx950-err.cl b/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx950-err.cl
index b304aa2b5d014c..12b472018f147f 100644
--- a/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx950-err.cl
+++ b/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx950-err.cl
@@ -18,7 +18,7 @@ typedef float __attribute__((ext_vector_type(2))) float2;
 typedef __bf16 __attribute__((ext_vector_type(2))) bfloat2;
 
 void test(global uint* out, global uint2* out_v2u32, uint a, uint b, global half2* out_v2f16, global float* out_f32, float scale, global short2* out_v2i16, float src0, float src1,
-          global float2* out_v2f32, half2 src0_v2f16, bfloat2 src0_v2bf16) {
+          global float2* out_v2f32, half2 src0_v2f16, bfloat2 src0_v2bf16, global bfloat2* out_v2bf16) {
   *out = __builtin_amdgcn_prng_b32(a); // expected-error{{'__builtin_amdgcn_prng_b32' needs target feature prng-inst}}
   *out_v2u32 = __builtin_amdgcn_permlane16_swap(a, b, false, false); // expected-error{{'__builtin_amdgcn_permlane16_swap' needs target feature permlane16-swap}}
   *out_v2u32 = __builtin_amdgcn_permlane32_swap(a, b, false, false); // expected-error{{'__builtin_amdgcn_permlane32_swap' needs target feature permlane32-swap}}
@@ -36,4 +36,6 @@ void test(global uint* out, global uint2* out_v2u32, uint a, uint b, global half
   *out_v2i16 = __builtin_amdgcn_cvt_scalef32_pk_bf8_bf16(*out_v2i16, src0_v2bf16, scale, true); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk_bf8_bf16' needs target feature bf8-cvt-scale-insts}}
   *out_v2f32 = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(a, scale, 3); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk_f32_fp4' needs target feature fp4-cvt-scale-insts}}
   *out = __builtin_amdgcn_cvt_scalef32_pk_fp4_f32(*out, src0, src1, scale, 3); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk_fp4_f32' needs target feature fp4-cvt-scale-insts}}
+  *out_v2f16 = __builtin_amdgcn_cvt_scalef32_pk_f16_fp4(a, scale, 3); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk_f16_fp4' needs target feature fp4-cvt-scale-insts}}
+  *out_v2bf16 = __builtin_amdgcn_cvt_scalef32_pk_bf16_fp4(a, scale, 3); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk_bf16_fp4' needs target feature fp4-cvt-scale-insts}}
 }
diff --git a/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx950.cl b/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx950.cl
index e1b1010a7ef852..4ec62505e823b5 100644
--- a/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx950.cl
+++ b/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx950.cl
@@ -829,3 +829,79 @@ void test_cvt_scalef32_pk_fp4_f32(global unsigned int* out, float src0, float sr
   *out = __builtin_amdgcn_cvt_scalef32_pk_fp4_f32(*out, src0, src1, scale, 2);
   *out = __builtin_amdgcn_cvt_scalef32_pk_fp4_f32(*out, src0, src1, scale, 3);
 }
+
+// CHECK-LABEL: @test_cvt_scalef32_pk_f16_fp4(
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    [[OUT_ADDR:%.*]] = alloca ptr addrspace(1), align 8, addrspace(5)
+// CHECK-NEXT:    [[SRC_ADDR:%.*]] = alloca i32, align 4, addrspace(5)
+// CHECK-NEXT:    [[SCALE_ADDR:%.*]] = alloca float, align 4, addrspace(5)
+// CHECK-NEXT:    store ptr addrspace(1) [[OUT:%.*]], ptr addrspace(5) [[OUT_ADDR]], align 8
+// CHECK-NEXT:    store i32 [[SRC:%.*]], ptr addrspace(5) [[SRC_ADDR]], align 4
+// CHECK-NEXT:    store float [[SCALE:%.*]], ptr addrspace(5) [[SCALE_ADDR]], align 4
+// CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr addrspace(5) [[SRC_ADDR]], align 4
+// CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[SCALE_ADDR]], align 4
+// CHECK-NEXT:    [[TMP2:%.*]] = call <2 x half> @llvm.amdgcn.cvt.scalef32.pk.f16.fp4(i32 [[TMP0]], float [[TMP1]], i32 0)
+// CHECK-NEXT:    [[TMP3:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
+// CHECK-NEXT:    store <2 x half> [[TMP2]], ptr addrspace(1) [[TMP3]], align 4
+// CHECK-NEXT:    [[TMP4:%.*]] = load i32, ptr addrspace(5) [[SRC_ADDR]], align 4
+// CHECK-NEXT:    [[TMP5:%.*]] = load float, ptr addrspace(5) [[SCALE_ADDR]], align 4
+// CHECK-NEXT:    [[TMP6:%.*]] = call <2 x half> @llvm.amdgcn.cvt.scalef32.pk.f16.fp4(i32 [[TMP4]], float [[TMP5]], i32 1)
+// CHECK-NEXT:    [[TMP7:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
+// CHECK-NEXT:    store <2 x half> [[TMP6]], ptr addrspace(1) [[TMP7]], align 4
+// CHECK-NEXT:    [[TMP8:%.*]] = load i32, ptr addrspace(5) [[SRC_ADDR]], align 4
+// CHECK-NEXT:    [[TMP9:%.*]] = load float, ptr addrspace(5) [[SCALE_ADDR]], align 4
+// CHECK-NEXT:    [[TMP10:%.*]] = call <2 x half> @llvm.amdgcn.cvt.scalef32.pk.f16.fp4(i32 [[TMP8]], float [[TMP9]], i32 2)
+// CHECK-NEXT:    [[TMP11:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
+// CHECK-NEXT:    store <2 x half> [[TMP10]], ptr addrspace(1) [[TMP11]], align 4
+// CHECK-NEXT:    [[TMP12:%.*]] = load i32, ptr addrspace(5) [[SRC_ADDR]], align 4
+// CHECK-NEXT:    [[TMP13:%.*]] = load float, ptr addrspace(5) [[SCALE_ADDR]], align 4
+// CHECK-NEXT:    [[TMP14:%.*]] = call <2 x half> @llvm.amdgcn.cvt.scalef32.pk.f16.fp4(i32 [[TMP12]], float [[TMP13]], i32 3)
+// CHECK-NEXT:    [[TMP15:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
+// CHECK-NEXT:    store <2 x half> [[TMP14]], ptr addrspace(1) [[TMP15]], align 4
+// CHECK-NEXT:    ret void
+//
+void test_cvt_scalef32_pk_f16_fp4(global half2* out, uint src, float scale)
+{
+  *out = __builtin_amdgcn_cvt_scalef32_pk_f16_fp4(src, scale, 0);
+  *out = __builtin_amdgcn_cvt_scalef32_pk_f16_fp4(src, scale, 1);
+  *out = __builtin_amdgcn_cvt_scalef32_pk_f16_fp4(src, scale, 2);
+  *out = __builtin_amdgcn_cvt_scalef32_pk_f16_fp4(src, scale, 3);
+}
+
+// CHECK-LABEL: @test_cvt_scalef32_pk_bf16_fp4(
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    [[OUT_ADDR:%.*]] = alloca ptr addrspace(1), align 8, addrspace(5)
+// CHECK-NEXT:    [[SRC_ADDR:%.*]] = alloca i32, align 4, addrspace(5)
+// CHECK-NEXT:    [[SCALE_ADDR:%.*]] = alloca float, align 4, addrspace(5)
+// CHECK-NEXT:    store ptr addrspace(1) [[OUT:%.*]], ptr addrspace(5) [[OUT_ADDR]], align 8
+// CHECK-NEXT:    store i32 [[SRC:%.*]], ptr addrspace(5) [[SRC_ADDR]], align 4
+// CHECK-NEXT:    store float [[SCALE:%.*]], ptr addrspace(5) [[SCALE_ADDR]], align 4
+// CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr addrspace(5) [[SRC_ADDR]], align 4
+// CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[SCALE_ADDR]], align 4
+// CHECK-NEXT:    [[TMP2:%.*]] = call <2 x bfloat> @llvm.amdgcn.cvt.scalef32.pk.bf16.fp4(i32 [[TMP0]], float [[TMP1]], i32 0)
+// CHECK-NEXT:    [[TMP3:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
+// CHECK-NEXT:    store <2 x bfloat> [[TMP2]], ptr addrspace(1) [[TMP3]], align 4
+// CHECK-NEXT:    [[TMP4:%.*]] = load i32, ptr addrspace(5) [[SRC_ADDR]], align 4
+// CHECK-NEXT:    [[TMP5:%.*]] = load float, ptr addrspace(5) [[SCALE_ADDR]], align 4
+// CHECK-NEXT:    [[TMP6:%.*]] = call <2 x bfloat> @llvm.amdgcn.cvt.scalef32.pk.bf16.fp4(i32 [[TMP4]], float [[TMP5]], i32 1)
+// CHECK-NEXT:    [[TMP7:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
+// CHECK-NEXT:    store <2 x bfloat> [[TMP6]], ptr addrspace(1) [[TMP7]], align 4
+// CHECK-NEXT:    [[TMP8:%.*]] = load i32, ptr addrspace(5) [[SRC_ADDR]], align 4
+// CHECK-NEXT:    [[TMP9:%.*]] = load float, ptr addrspace(5) [[SCALE_ADDR]], align 4
+// CHECK-NEXT:    [[TMP10:%.*]] = call <2 x bfloat> @llvm.amdgcn.cvt.scalef32.pk.bf16.fp4(i32 [[TMP8]], float [[TMP9]], i32 2)
+// CHECK-NEXT:    [[TMP11:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
+// CHECK-NEXT:    store <2 x bfloat> [[TMP10]], ptr addrspace(1) [[TMP11]], align 4
+// CHECK-NEXT:    [[TMP12:%.*]] = load i32, ptr addrspace(5) [[SRC_ADDR]], align 4
+// CHECK-NEXT:    [[TMP13:%.*]] = load float, ptr addrspace(5) [[SCALE_ADDR]], align 4
+// CHECK-NEXT:    [[TMP14:%.*]] = call <2 x bfloat> @llvm.amdgcn.cvt.scalef32.pk.bf16.fp4(i32 [[TMP12]], float [[TMP13]], i32 3)
+// CHECK-NEXT:    [[TMP15:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
+// CHECK-NEXT:    store <2 x bfloat> [[TMP14]], ptr addrspace(1) [[TMP15]], align 4
+// CHECK-NEXT:    ret void
+//
+void test_cvt_scalef32_pk_bf16_fp4(global bfloat2* out, uint src, float scale)
+{
+  *out = __builtin_amdgcn_cvt_scalef32_pk_bf16_fp4(src, scale, 0);
+  *out = __builtin_amdgcn_cvt_scalef32_pk_bf16_fp4(src, scale, 1);
+  *out = __builtin_amdgcn_cvt_scalef32_pk_bf16_fp4(src, scale, 2);
+  *out = __builtin_amdgcn_cvt_scalef32_pk_bf16_fp4(src, scale, 3);
+}
diff --git a/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950-param.cl b/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950-param.cl
index fdc42f53432718..183a640e1d14c5 100644
--- a/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950-param.cl
+++ b/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950-param.cl
@@ -165,7 +165,7 @@ void test_permlane32_swap(__global int* out, int old, int src, bool X) {
 
 void test_cvt_scalef32(global half2* out_v2f16, global float* out_f32, uint src, float scale, int index, bool X,
                        global short2* out_v2i16, float src0, float src1, global float2* out_v2f32,
-                       half2 src0_v2f16, bfloat2 src0_v2bf16, global uint* out) {
+                       half2 src0_v2f16, bfloat2 src0_v2bf16, global uint* out, global bfloat2* out_v2bf16) {
   *out_v2f16 = __builtin_amdgcn_cvt_scalef32_f16_fp8(*out_v2f16, src, scale, index, X); // expected-error{{argument to '__builtin_amdgcn_cvt_scalef32_f16_fp8' must be a constant integer}}
   *out_f32 = __builtin_amdgcn_cvt_scalef32_f32_fp8(src, scale, index); // // expected-error{{argument to '__builtin_amdgcn_cvt_scalef32_f32_fp8' must be a constant integer}}
   *out_v2f16 = __builtin_amdgcn_cvt_scalef32_f16_bf8(*out_v2f16, src, scale, index, X); // expected-error{{argument to '__builtin_amdgcn_cvt_scalef32_f16_bf8' must be a constant integer}}
@@ -180,4 +180,6 @@ void test_cvt_scalef32(global half2* out_v2f16, global float* out_f32, uint src,
   *out_v2i16 = __builtin_amdgcn_cvt_scalef32_pk_bf8_bf16(*out_v2i16, src0_v2bf16, scale, X); // expected-error{{argument to '__builtin_amdgcn_cvt_scalef32_pk_bf8_bf16' must be a constant integer}}
   *out_v2f32 = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(src, scale, index); // expected-error{{argument to '__builtin_amdgcn_cvt_scalef32_pk_f32_fp4' must be a constant integer}}
   *out = __builtin_amdgcn_cvt_scalef32_pk_fp4_f32(*out, src0, src1, scale, index); // expected-error{{argument to '__builtin_amdgcn_cvt_scalef32_pk_fp4_f32' must be a constant integer}}
+  *out_v2f16 = __builtin_amdgcn_cvt_scalef32_pk_f16_fp4(src, scale, index); // expected-error{{argument to '__builtin_amdgcn_cvt_scalef32_pk_f16_fp4' must be a constant integer}}
+  *out_v2bf16 = __builtin_amdgcn_cvt_scalef32_pk_bf16_fp4(src, scale, index); // expected-error{{argument to '__builtin_amdgcn_cvt_scalef32_pk_bf16_fp4' must be a constant integer}}
 }
diff --git a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
index 44c291d54e70ab..7f1e7103104d86 100644
--- a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
+++ b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
@@ -609,7 +609,7 @@ def int_amdgcn_cvt_scalef32_pk32_bf6_bf16 : AMDGPUCvtScaleF32Intrinsic<llvm_v6i3
 def int_amdgcn_cvt_scalef32_2xpk16_fp6_f32 : AMDGPUCvtScaleF32ToFP6BF6Intrinsic<llvm_v6i32_ty, llvm_v16f32_ty, llvm_v16f32_ty, "cvt_scalef32_2xpk16_fp6_f32">;
 def int_amdgcn_cvt_scalef32_2xpk16_bf6_f32 : AMDGPUCvtScaleF32ToFP6BF6Intrinsic<llvm_v6i32_ty, llvm_v16f32_ty, llvm_v16f32_ty, "cvt_scalef32_2xpk16_bf6_f32">;
 
-class AMDGPUCvtScaleFP8BF8ToF32Intrinsic<LLVMType DstTy, string name> : DefaultAttrsIntrinsic<
+class AMDGPUCvtScaleFP4FP8BF8ToF1632Intrinsic<LLVMType DstTy, string name> : DefaultAttrsIntrinsic<
   [DstTy],
   [llvm_i32_ty,   // src
    llvm_float_ty, // scale
@@ -669,8 +669,8 @@ def int_amdgcn_cvt_scalef32_f16_fp8  : AMDGPUCvtScaleFP8BF8ToF16TiedInputIntrins
 def int_amdgcn_cvt_scalef32_f16_bf8  : AMDGPUCvtScaleFP8BF8ToF16TiedInputIntrinsic<llvm_v2f16_ty, "cvt_scalef32_f16_bf8">;
 
 // llvm.amdgcn.cvt.scalef32.f32.fp8 int src, float scale, int src_sel_index [0..3]
-def int_amdgcn_cvt_scalef32_f32_fp8  : AMDGPUCvtScaleFP8BF8ToF32Intrinsic<llvm_float_ty, "cvt_scalef32_f32_fp8">;
-def int_amdgcn_cvt_scalef32_f32_bf8  : AMDGPUCvtScaleFP8BF8ToF32Intrinsic<llvm_float_ty, "cvt_scalef32_f32_bf8">;
+def int_amdgcn_cvt_scalef32_f32_fp8  : AMDGPUCvtScaleFP4FP8BF8ToF1632Intrinsic<llvm_float_ty, "cvt_scalef32_f32_fp8">;
+def int_amdgcn_cvt_scalef32_f32_bf8  : AMDGPUCvtScaleFP4FP8BF8ToF1632Intrinsic<llvm_float_ty, "cvt_scalef32_f32_bf8">;
 
 // llvm.amdgcn.cvt.scalef32.pk.fp8.f32 v2i16 old_vdst, float srcA, float srcB, float scale, bool dst_lo_hi_sel
 def int_amdgcn_cvt_scalef32_pk_fp8_f32 : AMDGPUCvtScaleF32ToFP8BF8TiedInputIntrinsic<"cvt_scalef32_pk_fp8_f32">;
@@ -687,11 +687,15 @@ def int_amdgcn_cvt_scalef32_pk_bf8_f16 : AMDGPUCvtScaleF16BF16ToFP8BF8TiedInputI
 def int_amdgcn_cvt_scalef32_pk_bf8_bf16: AMDGPUCvtScaleF16BF16ToFP8BF8TiedInputIntrinsic<llvm_v2bf16_ty, "cvt_scalef32_pk_bf8_bf16">;
 
 // llvm.amdgcn.cvt.scalef32.pk.f32.fp4 int src, float scale, int src_sel_index [0..3]
-def int_amdgcn_cvt_scalef32_pk_f32_fp4 : AMDGPUCvtScaleFP8BF8ToF32Intrinsic<llvm_v2f32_ty, "cvt_scalef32_pk_f32_fp4">;
+def int_amdgcn_cvt_scalef32_pk_f32_fp4 : AMDGPUCvtScaleFP4FP8BF8ToF1632Intrinsic<llvm_v2f32_ty, "cvt_scalef32_pk_f32_fp4">;
 
 // llvm.amdgcn.cvt.scalef32.pk.fp4.f32 i32 old_vdst, float srcA, float srcB, float scale, int dst_sel_index[0..3]
 def int_amdgcn_cvt_scalef32_pk_fp4_f32 : AMDGPUCvtScaleF32ToFP4Intrinsic<"cvt_scalef32_pk_fp4_f32">;
 
+// llvm.amdgcn.cvt.scalef32.pk.f32.fp4 int src, float scale, int src_sel_index [0..3]
+def int_amdgcn_cvt_scalef32_pk_f16_fp4 : AMDGPUCvtScaleFP4FP8BF8ToF1632Intrinsic<llvm_v2f16_ty, "cvt_scalef32_pk_f16_fp4">;
+def int_amdgcn_cvt_scalef32_pk_bf16_fp4: AMDGPUCvtScaleFP4FP8BF8ToF1632Intrinsic<llvm_v2bf16_ty, "cvt_scalef32_pk_bf16_fp4">;
+
 def int_amdgcn_prng_b32 : DefaultAttrsIntrinsic<
   [llvm_i32_ty], [llvm_i32_ty], [IntrNoMem]
 >, ClangBuiltin<"__builtin_amdgcn_prng_b32">;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
index f5d2104299a23c..d9d62e4804a7ce 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -4561,6 +4561,8 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
     case Intrinsic::amdgcn_cvt_scalef32_pk_bf8_bf16:
     case Intrinsic::amdgcn_cvt_scalef32_pk_f32_fp4:
     case Intrinsic::amdgcn_cvt_scalef32_pk_fp4_f32:
+    case Intrinsic::amdgcn_cvt_scalef32_pk_f16_fp4:
+    case Intrinsic::amdgcn_cvt_scalef32_pk_bf16_fp4:
     case Intrinsic::amdgcn_ashr_pk_i8_i32:
     case Intrinsic::amdgcn_ashr_pk_u8_i32:
     case Intrinsic::amdgcn_cvt_scalef32_2xpk16_fp6_f32:
diff --git a/llvm/lib/Target/AMDGPU/VOP3Instructions.td b/llvm/lib/Target/AMDGPU/VOP3Instructions.td
index 3a70ad1192cf5b..1013bbe05dbd28 100644
--- a/llvm/lib/Target/AMDGPU/VOP3Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP3Instructions.td
@@ -1090,6 +1090,8 @@ class Cvt_Scale_FP4FP8BF8ToF16F32_Pat<SDPatternOperator node, VOP3_Pseudo inst,
 >;
 def : Cvt_Scale_FP4FP8BF8ToF16F32_Pat<int_amdgcn_cvt_scalef32_f32_fp8, V_CVT_SCALEF32_F32_FP8_e64, f32>;
 def : Cvt_Scale_FP4FP8BF8ToF16F32_Pat<int_amdgcn_cvt_scalef32_f32_bf8, V_CVT_SCALEF32_F32_BF8_e64, f32>;
+def : Cvt_Scale_FP4FP8BF8ToF16F32_Pat<int_amdgcn_cvt_scalef32_pk_f16_fp4, V_CVT_SCALEF32_PK_F16_FP4_e64, v2f16>;
+def : Cvt_Scale_FP4FP8BF8ToF16F32_Pat<int_amdgcn_cvt_scalef32_pk_bf16_fp4, V_CVT_SCALEF32_PK_BF16_FP4_e64, v2bf16>;
 
 class Cvt_Scale_FP8BF8ToF16_Pat<SDPatternOperator node, VOP3_Pseudo inst, int dst_sel> : GCNPat<
     (v2f16 (node v2f16:$vdst_in, i32:$src0, f32:$src1, timm:$src_sel, dst_sel)),
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.cvt.scalef32.pk.gfx950.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.cvt.scalef32.pk.gfx950.ll
index a28268da8983ac..cee5ee77a9510d 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.cvt.scalef32.pk.gfx950.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.cvt.scalef32.pk.gfx950.ll
@@ -18,6 +18,8 @@ declare <2 x i16> @llvm.amdgcn.cvt.scalef32.pk.bf8.f16(<2 x i16>, <2 x half>, fl
 declare <2 x i16> @llvm.amdgcn.cvt.scalef32.pk.bf8.fb16(<2 x i16>, <2 x bfloat>, float, i1)
 declare <2 x float> @llvm.amdgcn.cvt.scalef32.pk.f32.fp4(i32, float, i32)
 declare i32 @llvm.amdgcn.cvt.scalef32.pk.fp4.f32(i32, float, float, float, i32)
+declare <2 x half> @llvm.amdgcn.cvt.scalef32.pk.f16.fp4(i32, float, i32)
+declare <2 x bfloat> @llvm.amdgcn.cvt.scalef32.pk.bf16.fp4(i32, float, i32)
 
 define amdgpu_ps void @test_scalef32_pk32_fp6_f32_vv(<16 x float> %src, float %scale, ptr addrspace(1) %out) {
 ; GFX950-SDAG-LABEL: test_scalef32_pk32_fp6_f32_vv:
@@ -768,3 +770,83 @@ define i32 @test_cvt_scale_fp4_f32_byte3_fabs_fneg(i32 %old, float %src0, float
   %ret = tail call i32 @llvm.amdgcn.cvt.scalef32.pk.fp4.f32(i32 %old, float %fabs.src0, float %fneg.src1, float %scale, i32 3)
   ret i32 %ret
 }
+
+define <2 x half> @test_cvt_scale_f16_fp4_byte0(i32 %src, float %scale) {
+; GCN-LABEL: test_cvt_scale_f16_fp4_byte0:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_cvt_scalef32_pk_f16_fp4 v0, v0, v1
+; GCN-NEXT:    s_setpc_b64 s[30:31]
+  %ret = tail call <2 x half> @llvm.amdgcn.cvt.scalef32.pk.f16.fp4(i32 %src, float %scale, i32 0)
+  ret <2 x half> %ret
+}
+
+define <2 x half> @test_cvt_scale_f16_fp4_byte1(i32 %src, float %scale) {
+; GCN-LABEL: test_cvt_scale_f16_fp4_byte1:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_cvt_scalef32_pk_f16_fp4 v0, v0, v1 op_sel:[0,1,0]
+; GCN-NEXT:    s_setpc_b64 s[30:31]
+  %ret = tail call <2 x half> @llvm.amdgcn.cvt.scalef32.pk.f16.fp4(i32 %src, float %scale, i32 1)
+  ret <2 x half> %ret
+}
+
+define <2 x half> @test_cvt_scale_f16_fp4_byte2(i32 %src, float %scale) {
+; GCN-LABEL: test_cvt_scale_f16_fp4_byte2:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_cvt_scalef32_pk_f16_fp4 v0, v0, v1 op_sel:[1,0,0]
+; GCN-NEXT:    s_setpc_b64 s[30:31]
+  %ret = tail call <2 x half> @llvm.amdgcn.cvt.scalef32.pk.f16.fp4(i32 %src, float %scale, i32 2)
+  ret <2 x half> %ret
+}
+
+define <2 x half> @test_cvt_scale_f16_fp4_byte3(i32 %src, float %scale) {
+; GCN-LABEL: test_cvt_scale_f16_fp4_byte3:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_cvt_scalef32_pk_f16_fp4 v0, v0, v1 op_sel:[1,1,0]
+; GCN-NEXT:    s_setpc_b64 s[30:31]
+  %ret = tail call <2 x half> @llvm.amdgcn.cvt.scalef32.pk.f16.fp4(i32 %src, float %scale, i32 3)
+  ret <2 x half> %ret
+}
+
+define <2 x bfloat> @test_cvt_scale_bf16_fp4_byte0(i32 %src, float %scale) {
+; GCN-LABEL: test_cvt_scale_bf16_fp4_byte0:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_cvt_scalef32_pk_bf16_fp4 v0, v0, v1
+; GCN-NEXT:    s_setpc_b64 s[30:31]
+  %ret = tail call <2 x bfloat> @llvm.amdgcn.cvt.scalef32.pk.bf16.fp4(i32 %src, float %scale, i32 0)
+  ret <2 x bfloat> %ret
+}
+
+define <2 x bfloat> @test_cvt_scale_bf16_fp4_byte1(i32 %src, float %scale) {
+; GCN-LABEL: test_cvt_scale_bf16_fp4_byte1:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_cvt_scalef32_pk_bf16_fp4 v0, v0, v1 op_sel:[0,1,0]
+; GCN-NEXT:    s_setpc_b64 s[30:31]
+  %ret = tail call <2 x bfloat> @llvm.amdgcn.cvt.scalef32.pk.bf16.fp4(i32 %src, float %scale, i32 1)
+  ret <2 x bfloat> %ret
+}
+
+define <2 x bfloat> @test_cvt_scale_bf16_fp4_byte2(i32 %src, float %scale) {
+; GCN-LABEL: test_cvt_scale_bf16_fp4_byte2:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_cvt_scalef32_pk_bf16_fp4 v0, v0, v1 op_sel:[1,0,0]
+; GCN-NEXT:    s_setpc_b64 s[30:31]
+  %ret = tail call <2 x bfloat> @llvm.amdgcn.cvt.scalef32.pk.bf16.fp4(i32 %src, float %scale, i32 2)
+  ret <2 x bfloat> %ret
+}
+
+define <2 x bfloat> @test_cvt_scale_bf16_fp4_byte3(i32 %src, float %scale) {
+; GCN-LABEL: test_cvt_scale_bf16_fp4_byte3:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_cvt_scalef32_pk_bf16_fp4 v0, v0, v1 op_sel:[1,1,0]
+; GCN-NEXT:    s_setpc_b64 s[30:31]
+  %ret = tail call <2 x bfloat> @llvm.amdgcn.cvt.scalef32.pk.bf16.fp4(i32 %src, float %scale, i32 3)
+  ret <2 x bfloat> %ret
+}



More information about the llvm-branch-commits mailing list