[mono] Implement Rdm and Dp (#49737)

See #42322 and #42280. Tested manually on an arm64 Linux VM running on an M1 Mac Mini. Does not enable RDM or DP when AOT-compiling the associated runtime tests; our CI hardware doesn't support these extensions yet.
dotnet · Jun 24, 2021 · 4aeef0d · 4aeef0d
1 parent db9cfd2
commit 4aeef0d
Show file tree

Hide file tree

Showing 6 changed files with 162 additions and 7 deletions.
diff --git a/src/mono/mono/mini/aot-compiler.c b/src/mono/mono/mini/aot-compiler.c
@@ -8240,7 +8240,7 @@ parse_cpu_features (const gchar *attr)
 	// if we disable a feature from the SSE-AVX tree we also need to disable all dependencies
 	if (!enabled && (feature & MONO_CPU_X86_FULL_SSEAVX_COMBINED))
 		feature = (MonoCPUFeatures) (MONO_CPU_X86_FULL_SSEAVX_COMBINED & ~feature);
-	
+
 #elif defined(TARGET_ARM64)
 	// MONO_CPU_ARM64_BASE is unconditionally set in mini_get_cpu_features.
 	if (!strcmp (attr + prefix, "crc"))
@@ -8249,6 +8249,10 @@ parse_cpu_features (const gchar *attr)
 		feature = MONO_CPU_ARM64_CRYPTO;
 	else if (!strcmp (attr + prefix, "neon"))
 		feature = MONO_CPU_ARM64_NEON;
+	else if (!strcmp (attr + prefix, "rdm"))
+		feature = MONO_CPU_ARM64_RDM;
+	else if (!strcmp (attr + prefix, "dotprod"))
+		feature = MONO_CPU_ARM64_DP;
 #elif defined(TARGET_WASM)
 	if (!strcmp (attr + prefix, "simd"))
 		feature = MONO_CPU_WASM_SIMD;

diff --git a/src/mono/mono/mini/llvm-intrinsics.h b/src/mono/mono/mini/llvm-intrinsics.h
@@ -25,6 +25,7 @@
 #define Widen INTRIN_kind_widen
 #define WidenAcross INTRIN_kind_widen_across
 #define Across INTRIN_kind_across
+#define Arm64DotProd INTRIN_kind_arm64_dot_prod
 #if !defined(Generic)
 #define Generic
 #endif
@@ -466,6 +467,10 @@ INTRINS_OVR_TAG(AARCH64_ADV_SIMD_SRI, aarch64_neon_vsri, Arm64, V64 | V128 | I1
 
 INTRINS_OVR_TAG(AARCH64_ADV_SIMD_TBX1, aarch64_neon_tbx1, Arm64, V64 | V128 | I1)
 INTRINS_OVR_TAG(AARCH64_ADV_SIMD_TBL1, aarch64_neon_tbl1, Arm64, V64 | V128 | I1)
+
+INTRINS_OVR_TAG_KIND(AARCH64_ADV_SIMD_SDOT, aarch64_neon_sdot, Arm64, Arm64DotProd, V64 | V128 | I4)
+INTRINS_OVR_TAG_KIND(AARCH64_ADV_SIMD_UDOT, aarch64_neon_udot, Arm64, Arm64DotProd, V64 | V128 | I4)
+
 #endif
 
 #undef INTRINS
@@ -486,6 +491,7 @@ INTRINS_OVR_TAG(AARCH64_ADV_SIMD_TBL1, aarch64_neon_tbl1, Arm64, V64 | V128 | I1
 #undef Ftoi
 #undef WidenAcross
 #undef Across
+#undef Arm64DotProd
 #undef Generic
 #undef X86
 #undef Arm64

diff --git a/src/mono/mono/mini/mini-llvm.c b/src/mono/mono/mini/mini-llvm.c
@@ -360,6 +360,7 @@ enum {
 	INTRIN_kind_widen,
 	INTRIN_kind_widen_across,
 	INTRIN_kind_across,
+	INTRIN_kind_arm64_dot_prod,
 };
 
 static const uint8_t intrin_kind [] = {
@@ -9660,6 +9661,21 @@ process_bb (EmitContext *ctx, MonoBasicBlock *bb)
 			values [ins->dreg] = result;
 			break;
 		}
+		case OP_ARM64_SELECT_QUAD: {
+			LLVMTypeRef src_type = simd_class_to_llvm_type (ctx, ins->data.op [1].klass);
+			LLVMTypeRef ret_type = simd_class_to_llvm_type (ctx, ins->klass);
+			unsigned int src_type_bits = mono_llvm_get_prim_size_bits (src_type);
+			unsigned int ret_type_bits = mono_llvm_get_prim_size_bits (ret_type);
+			unsigned int src_intermediate_elems = src_type_bits / 32;
+			unsigned int ret_intermediate_elems = ret_type_bits / 32;
+			LLVMTypeRef intermediate_type = LLVMVectorType (i4_t, src_intermediate_elems);
+			LLVMValueRef result = LLVMBuildBitCast (builder, lhs, intermediate_type, "arm64_select_quad");
+			result = LLVMBuildExtractElement (builder, result, rhs, "arm64_select_quad");
+			result = broadcast_element (ctx, result, ret_intermediate_elems);
+			result = LLVMBuildBitCast (builder, result, ret_type, "arm64_select_quad");
+			values [ins->dreg] = result;
+			break;
+		}
 		case OP_LSCNT32:
 		case OP_LSCNT64: {
 			// %shr = ashr i32 %x, 31
@@ -9683,6 +9699,43 @@ process_bb (EmitContext *ctx, MonoBasicBlock *bb)
 			values [ins->dreg] = LLVMBuildCall (builder, get_intrins (ctx, ins->opcode == OP_LSCNT32 ? INTRINS_CTLZ_I32 : INTRINS_CTLZ_I64), args, 2, "");
 			break;
 		}
+		case OP_ARM64_SQRDMLAH:
+		case OP_ARM64_SQRDMLAH_BYSCALAR:
+		case OP_ARM64_SQRDMLAH_SCALAR:
+		case OP_ARM64_SQRDMLSH:
+		case OP_ARM64_SQRDMLSH_BYSCALAR:
+		case OP_ARM64_SQRDMLSH_SCALAR: {
+			gboolean byscalar = FALSE;
+			gboolean scalar = FALSE;
+			gboolean subtract = FALSE;
+			switch (ins->opcode) {
+			case OP_ARM64_SQRDMLAH_BYSCALAR: byscalar = TRUE; break;
+			case OP_ARM64_SQRDMLAH_SCALAR: scalar = TRUE; break;
+			case OP_ARM64_SQRDMLSH: subtract = TRUE; break;
+			case OP_ARM64_SQRDMLSH_BYSCALAR: subtract = TRUE; byscalar = TRUE; break;
+			case OP_ARM64_SQRDMLSH_SCALAR: subtract = TRUE; scalar = TRUE; break;
+			}
+			int acc_iid = subtract ? INTRINS_AARCH64_ADV_SIMD_SQSUB : INTRINS_AARCH64_ADV_SIMD_SQADD;
+			LLVMTypeRef ret_t = simd_class_to_llvm_type (ctx, ins->klass);
+			llvm_ovr_tag_t ovr_tag = ovr_tag_from_llvm_type (ret_t);
+			ScalarOpFromVectorOpCtx sctx = scalar_op_from_vector_op (ctx, ret_t, ins);
+			LLVMValueRef args [] = { lhs, rhs, arg3 };
+			if (byscalar) {
+				unsigned int elems = LLVMGetVectorSize (ret_t);
+				args [2] = broadcast_element (ctx, scalar_from_vector (ctx, args [2]), elems);
+			}
+			if (scalar) {
+				ovr_tag = sctx.ovr_tag;
+				scalar_op_from_vector_op_process_args (&sctx, args, 3);
+			}
+			LLVMValueRef result = call_overloaded_intrins (ctx, INTRINS_AARCH64_ADV_SIMD_SQRDMULH, ovr_tag, &args [1], "arm64_sqrdmlxh");
+			args [1] = result;
+			result = call_overloaded_intrins (ctx, acc_iid, ovr_tag, &args [0], "arm64_sqrdmlxh");
+			if (scalar)
+				result = scalar_op_from_vector_op_process_result (&sctx, result);
+			values [ins->dreg] = result;
+			break;
+		}
 		case OP_ARM64_SMULH:
 		case OP_ARM64_UMULH: {
 			LLVMValueRef op1, op2;
@@ -12136,6 +12189,13 @@ add_intrinsic (LLVMModuleRef module, int id)
 						int associated_prim = MAX(ew, 2);
 						LLVMTypeRef associated_scalar_type = intrin_types [0][associated_prim];
 						intrins = add_intrins2 (module, id, associated_scalar_type, distinguishing_type);
+					} else if (kind == INTRIN_kind_arm64_dot_prod) {
+						/*
+						 * @llvm.aarch64.neon.sdot.v2i32.v8i8
+						 * @llvm.aarch64.neon.sdot.v4i32.v16i8
+						 */
+						 LLVMTypeRef associated_type = intrin_types [vw][0];
+						 intrins = add_intrins2 (module, id, distinguishing_type, associated_type);
 					} else
 						intrins = add_intrins1 (module, id, distinguishing_type);
 					int key = key_from_id_and_tag (id, test);
@@ -13530,9 +13590,11 @@ MonoCPUFeatures mono_llvm_get_cpu_features (void)
 		{ "bmi2",	MONO_CPU_X86_BMI2 },
 #endif
 #if defined(TARGET_ARM64)
-		{ "crc",	MONO_CPU_ARM64_CRC },
-		{ "crypto",	MONO_CPU_ARM64_CRYPTO },
-		{ "neon",	MONO_CPU_ARM64_NEON }
+		{ "crc",     MONO_CPU_ARM64_CRC },
+		{ "crypto",  MONO_CPU_ARM64_CRYPTO },
+		{ "neon",    MONO_CPU_ARM64_NEON },
+		{ "rdm",     MONO_CPU_ARM64_RDM },
+		{ "dotprod", MONO_CPU_ARM64_DP },
 #endif
 #if defined(TARGET_WASM)
 		{ "simd",	MONO_CPU_WASM_SIMD },

diff --git a/src/mono/mono/mini/mini-ops.h b/src/mono/mono/mini/mini-ops.h
@@ -1740,6 +1740,7 @@ MINI_OP(OP_ARM64_UQXTN2, "arm64_uqxtn2", XREG, XREG, XREG)
 MINI_OP(OP_ARM64_SQXTUN2, "arm64_sqxtun2", XREG, XREG, XREG)
 
 MINI_OP(OP_ARM64_SELECT_SCALAR, "arm64_select_scalar", XREG, XREG, IREG)
+MINI_OP(OP_ARM64_SELECT_QUAD, "arm64_select_quad", XREG, XREG, IREG)
 
 MINI_OP(OP_ARM64_FCVTZU, "arm64_fcvtzu", XREG, XREG, NONE)
 MINI_OP(OP_ARM64_FCVTZS, "arm64_fcvtzs", XREG, XREG, NONE)
@@ -1807,4 +1808,11 @@ MINI_OP(OP_ARM64_XNARROW_SCALAR, "arm64_xnarrow_scalar", XREG, XREG, NONE)
 
 MINI_OP3(OP_ARM64_EXT, "arm64_ext", XREG, XREG, XREG, IREG)
 
+MINI_OP3(OP_ARM64_SQRDMLAH, "arm64_sqrdmlah", XREG, XREG, XREG, XREG)
+MINI_OP3(OP_ARM64_SQRDMLAH_BYSCALAR, "arm64_sqrdmlah_byscalar", XREG, XREG, XREG, XREG)
+MINI_OP3(OP_ARM64_SQRDMLAH_SCALAR, "arm64_sqrdmlah_scalar", XREG, XREG, XREG, XREG)
+MINI_OP3(OP_ARM64_SQRDMLSH, "arm64_sqrdmlsh", XREG, XREG, XREG, XREG)
+MINI_OP3(OP_ARM64_SQRDMLSH_BYSCALAR, "arm64_sqrdmlsh_byscalar", XREG, XREG, XREG, XREG)
+MINI_OP3(OP_ARM64_SQRDMLSH_SCALAR, "arm64_sqrdmlsh_scalar", XREG, XREG, XREG, XREG)
+
 #endif // TARGET_ARM64
diff --git a/src/mono/mono/mini/simd-intrinsics.c b/src/mono/mono/mini/simd-intrinsics.c
@@ -1428,13 +1428,31 @@ static SimdIntrinsic advsimd_methods [] = {
 	{SN_get_IsSupported},
 };
 
+static const SimdIntrinsic rdm_methods [] = {
+	{SN_MultiplyRoundedDoublingAndAddSaturateHigh, OP_ARM64_SQRDMLAH},
+	{SN_MultiplyRoundedDoublingAndAddSaturateHighScalar, OP_ARM64_SQRDMLAH_SCALAR},
+	{SN_MultiplyRoundedDoublingAndSubtractSaturateHigh, OP_ARM64_SQRDMLSH},
+	{SN_MultiplyRoundedDoublingAndSubtractSaturateHighScalar, OP_ARM64_SQRDMLSH_SCALAR},
+	{SN_MultiplyRoundedDoublingBySelectedScalarAndAddSaturateHigh},
+	{SN_MultiplyRoundedDoublingBySelectedScalarAndSubtractSaturateHigh},
+	{SN_MultiplyRoundedDoublingScalarBySelectedScalarAndAddSaturateHigh},
+	{SN_MultiplyRoundedDoublingScalarBySelectedScalarAndSubtractSaturateHigh},
+	{SN_get_IsSupported},
+};
+
+static const SimdIntrinsic dp_methods [] = {
+	{SN_DotProduct, OP_XOP_OVR_X_X_X_X, INTRINS_AARCH64_ADV_SIMD_SDOT, OP_XOP_OVR_X_X_X_X, INTRINS_AARCH64_ADV_SIMD_UDOT},
+	{SN_DotProductBySelectedQuadruplet},
+	{SN_get_IsSupported},
+};
+
 static const IntrinGroup supported_arm_intrinsics [] = {
 	{ "AdvSimd", MONO_CPU_ARM64_NEON, advsimd_methods, sizeof (advsimd_methods) },
 	{ "Aes", MONO_CPU_ARM64_CRYPTO, crypto_aes_methods, sizeof (crypto_aes_methods) },
 	{ "ArmBase", MONO_CPU_ARM64_BASE, armbase_methods, sizeof (armbase_methods) },
 	{ "Crc32", MONO_CPU_ARM64_CRC, crc32_methods, sizeof (crc32_methods) },
-	{ "Dp", MONO_CPU_ARM64_DP, unsupported, sizeof (unsupported) },
-	{ "Rdm", MONO_CPU_ARM64_RDM, unsupported, sizeof (unsupported) },
+	{ "Dp", MONO_CPU_ARM64_DP, dp_methods, sizeof (dp_methods) },
+	{ "Rdm", MONO_CPU_ARM64_RDM, rdm_methods, sizeof (rdm_methods) },
 	{ "Sha1", MONO_CPU_ARM64_CRYPTO, sha1_methods, sizeof (sha1_methods) },
 	{ "Sha256", MONO_CPU_ARM64_CRYPTO, sha256_methods, sizeof (sha256_methods) },
 };
@@ -1740,6 +1758,51 @@ emit_arm64_intrinsics (
 		}
 	}
 
+	if (feature == MONO_CPU_ARM64_RDM) {
+		switch (id) {
+		case SN_MultiplyRoundedDoublingBySelectedScalarAndAddSaturateHigh:
+		case SN_MultiplyRoundedDoublingBySelectedScalarAndSubtractSaturateHigh:
+		case SN_MultiplyRoundedDoublingScalarBySelectedScalarAndAddSaturateHigh:
+		case SN_MultiplyRoundedDoublingScalarBySelectedScalarAndSubtractSaturateHigh: {
+			MonoClass *ret_klass = mono_class_from_mono_type_internal (fsig->ret);
+			int opcode = 0;
+			switch (id) {
+			case SN_MultiplyRoundedDoublingBySelectedScalarAndAddSaturateHigh: opcode = OP_ARM64_SQRDMLAH_BYSCALAR; break;
+			case SN_MultiplyRoundedDoublingBySelectedScalarAndSubtractSaturateHigh: opcode = OP_ARM64_SQRDMLSH_BYSCALAR; break;
+			case SN_MultiplyRoundedDoublingScalarBySelectedScalarAndAddSaturateHigh: opcode = OP_ARM64_SQRDMLAH_SCALAR; break;
+			case SN_MultiplyRoundedDoublingScalarBySelectedScalarAndSubtractSaturateHigh: opcode = OP_ARM64_SQRDMLSH_SCALAR; break;
+			}
+			MonoInst *scalar = emit_simd_ins (cfg, ret_klass, OP_ARM64_SELECT_SCALAR, args [2]->dreg, args [3]->dreg);
+			MonoInst *ret = emit_simd_ins (cfg, ret_klass, opcode, args [0]->dreg, args [1]->dreg);
+			ret->inst_c1 = arg0_type;
+			ret->sreg3 = scalar->dreg;
+			return ret;
+		}
+		default:
+			g_assert_not_reached ();
+		}
+	}
+
+	if (feature == MONO_CPU_ARM64_DP) {
+		switch (id) {
+		case SN_DotProductBySelectedQuadruplet: {
+			MonoClass *ret_klass = mono_class_from_mono_type_internal (fsig->ret);
+			MonoClass *arg_klass = mono_class_from_mono_type_internal (fsig->params [1]);
+			MonoClass *quad_klass = mono_class_from_mono_type_internal (fsig->params [2]);
+			gboolean is_unsigned = type_is_unsigned (fsig->ret);
+			int iid = is_unsigned ? INTRINS_AARCH64_ADV_SIMD_UDOT : INTRINS_AARCH64_ADV_SIMD_SDOT;
+			MonoInst *quad = emit_simd_ins (cfg, arg_klass, OP_ARM64_SELECT_QUAD, args [2]->dreg, args [3]->dreg);
+			quad->data.op [1].klass = quad_klass;
+			MonoInst *ret = emit_simd_ins (cfg, ret_klass, OP_XOP_OVR_X_X_X_X, args [0]->dreg, args [1]->dreg);
+			ret->sreg3 = quad->dreg;
+			ret->inst_c0 = iid;
+			return ret;
+		}
+		default:
+			g_assert_not_reached ();
+		}
+	}
+
 	return NULL;
 }
 #endif // TARGET_ARM64

diff --git a/src/mono/mono/mini/simd-methods.h b/src/mono/mono/mini/simd-methods.h
@@ -248,7 +248,7 @@ METHOD(HashUpdate2)
 METHOD(ScheduleUpdate0)
 METHOD(ScheduleUpdate1)
 METHOD(MixColumns)
-//AdvSimd
+// AdvSimd
 METHOD(AbsSaturate)
 METHOD(AbsSaturateScalar)
 METHOD(AbsScalar)
@@ -559,3 +559,15 @@ METHOD(ZeroExtendWideningLower)
 METHOD(ZeroExtendWideningUpper)
 METHOD(ZipHigh)
 METHOD(ZipLow)
+// Arm.Rdm
+METHOD(MultiplyRoundedDoublingAndAddSaturateHigh)
+METHOD(MultiplyRoundedDoublingAndSubtractSaturateHigh)
+METHOD(MultiplyRoundedDoublingBySelectedScalarAndAddSaturateHigh)
+METHOD(MultiplyRoundedDoublingBySelectedScalarAndSubtractSaturateHigh)
+// Arm.Rdm.Arm64
+METHOD(MultiplyRoundedDoublingAndAddSaturateHighScalar)
+METHOD(MultiplyRoundedDoublingAndSubtractSaturateHighScalar)
+METHOD(MultiplyRoundedDoublingScalarBySelectedScalarAndAddSaturateHigh)
+METHOD(MultiplyRoundedDoublingScalarBySelectedScalarAndSubtractSaturateHigh)
+// Arm.Dp
+METHOD(DotProductBySelectedQuadruplet)