Merge branch 'dev'v6.63

13 files changed, 115 insertions, 23 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index bb4fbe1..238664f 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 2.6...3.0.2)
 
-project(xbyak LANGUAGES CXX VERSION 6.62)
+project(xbyak LANGUAGES CXX VERSION 6.63)
 
 file(GLOB headers xbyak/*.h)
 
diff --git a/doc/changelog.md b/doc/changelog.md
index 31fd939..484b6bd 100644
--- a/doc/changelog.md
+++ b/doc/changelog.md
@@ -1,5 +1,7 @@
 # History
 
+* 2022/Oct/06 vpmadd52{h,l}uq support AVX-IFMA
+* 2022/Oct/05 support amx_fp16/avx_vnni_int8/avx_ne_convert and add setDefaultEncoding()
 * 2022/Aug/15 ver 6.62 add serialize instruction
 * 2022/Aug/02 ver 6.61.1 noexcept is supported by Visual Studio 2015 or later
 * 2022/Jul/29 ver 6.61 fix exception of movzx eax, ah in 64-bit mode
diff --git a/doc/usage.md b/doc/usage.md
index 7dad245..7b5678e 100644
--- a/doc/usage.md
+++ b/doc/usage.md
@@ -110,7 +110,15 @@ vfpclassps k5{k3}, [rax+64]{1to4}, 5    --> vfpclassps(k5|k3, yword_b [rax+64],
 vpdpbusd(xm0, xm1, xm2); // default encoding is EVEX
 vpdpbusd(xm0, xm1, xm2, EvexEncoding); // same as the above
 vpdpbusd(xm0, xm1, xm2, VexEncoding); // VEX encoding
+setDefaultEncoding(VexEncoding); // default encoding is VEX
+vpdpbusd(xm0, xm1, xm2); // VEX encoding
 ```
+
+- setDefaultEncoding(PreferredEncoding encoding);
+  - Set the default encoding to select EVEX or VEX.
+  - The default value is EvexEncoding.
+  - This function affects only an instruction that has a PreferredEncoding argument such as vpdpbusd.
+
 ### Remark
 * `k1`, ..., `k7` are opmask registers.
   - `k0` is dealt as no mask.
diff --git a/gen/gen_avx512.cpp b/gen/gen_avx512.cpp
index 35960bb..41b28e9 100644
--- a/gen/gen_avx512.cpp
+++ b/gen/gen_avx512.cpp
@@ -387,9 +387,6 @@ void putX_X_XM_IMM()
 		{ 0x57, "vreducess", T_66 | T_0F3A | T_MUST_EVEX | T_EW0 | T_SAE_X | T_N4, true },
 		{ 0x57, "vreducesh", T_0F3A | T_MUST_EVEX | T_EW0 | T_SAE_X | T_N2, true },
 
-		{ 0xB4, "vpmadd52luq", T_66 | T_0F38 | T_YMM | T_MUST_EVEX | T_EW1 | T_B64, false },
-		{ 0xB5, "vpmadd52huq", T_66 | T_0F38 | T_YMM | T_MUST_EVEX | T_EW1 | T_B64, false },
-
 		{ 0x70, "vpshldw", T_66 | T_0F3A | T_YMM | T_MUST_EVEX | T_EW1 | T_SAE_Z, true },
 		{ 0x71, "vpshldd", T_66 | T_0F3A | T_YMM | T_MUST_EVEX | T_EW0 | T_SAE_Z | T_B32, true },
 		{ 0x71, "vpshldq", T_66 | T_0F3A | T_YMM | T_MUST_EVEX | T_EW1 | T_SAE_Z | T_B64, true },
@@ -827,7 +824,6 @@ void putMisc()
 	puts("void vfpclasssh(const Opmask& k, const Operand& op, uint8_t imm) { if (!op.isXMEM()) XBYAK_THROW(ERR_BAD_MEM_SIZE) opVex(k, 0, op, T_0F3A | T_MUST_EVEX | T_EW0 | T_N2, 0x67, imm); }");
 
 	puts("void vpshufbitqmb(const Opmask& k, const Xmm& x, const Operand& op) { opVex(k, &x, op, T_66 | T_0F38 | T_EW0 | T_YMM | T_MUST_EVEX, 0x8F); }");
-	puts("void vcvtneps2bf16(const Xmm& x, const Operand& op) { opCvt2(x, op, T_F3 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_MUST_EVEX | T_B32, 0x72); }");
 
 	puts("void vp2intersectd(const Opmask& k, const Xmm& x, const Operand& op) { if (k.getOpmaskIdx() != 0) XBYAK_THROW(ERR_OPMASK_IS_ALREADY_SET) opAVX_K_X_XM(k, x, op, T_F2 | T_0F38 | T_YMM | T_EVEX | T_EW0 | T_B32, 0x68); }");
 	puts("void vp2intersectq(const Opmask& k, const Xmm& x, const Operand& op) { if (k.getOpmaskIdx() != 0) XBYAK_THROW(ERR_OPMASK_IS_ALREADY_SET) opAVX_K_X_XM(k, x, op, T_F2 | T_0F38 | T_YMM | T_EVEX | T_EW1 | T_B64, 0x68); }");
diff --git a/gen/gen_code.cpp b/gen/gen_code.cpp
index 9e2b808..b89009b 100644
--- a/gen/gen_code.cpp
+++ b/gen/gen_code.cpp
@@ -1667,6 +1667,25 @@ void put()
 		puts("void vcvtps2ph(const Operand& op, const Xmm& x, uint8_t imm) { checkCvt1(x, op); opVex(x, 0, op, T_0F3A | T_66 | T_W0 | T_EVEX | T_EW0 | T_N8 | T_N_VL | T_SAE_Y | T_M_K, 0x1D, imm); }");
 
 	}
+	{
+		const struct Tbl {
+			const char *name;
+			int type;
+			uint8_t code;
+		} tbl[] = {
+			{ "vbcstnebf162ps", T_F3 | T_0F38 | T_W0 | T_B16 | T_YMM, 0xB1 },
+			{ "vbcstnesh2ps", T_66 | T_0F38 | T_W0 | T_B16 | T_YMM, 0xB1 },
+			{ "vcvtneebf162ps", T_F3 | T_0F38 | T_W0 | T_YMM, 0xB0 },
+			{ "vcvtneeph2ps", T_66 | T_0F38 | T_W0 | T_YMM, 0xB0 },
+			{ "vcvtneobf162ps", T_F2 | T_0F38 | T_W0 | T_YMM, 0xB0 },
+			{ "vcvtneoph2ps", T_0F38 | T_W0 | T_YMM, 0xB0 }
+		};
+		for (size_t i = 0; i < NUM_OF_ARRAY(tbl); i++) {
+			const Tbl& p = tbl[i];
+			printf("void %s(const Xmm& x, const Address& addr) { opVex(x, 0, addr, %s, 0x%02X); }\n", p.name, type2String(p.type).c_str(), p.code);
+		}
+		puts("void vcvtneps2bf16(const Xmm& x, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opCvt2(x, op, T_F3 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_B32 | orEvexIf(encoding), 0x72); }");
+	}
 	// haswell gpr(reg, reg, r/m)
 	{
 		const struct Tbl {
@@ -1756,11 +1775,33 @@ void put()
 			{ 0x51, "vpdpbusds", T_66 | T_0F38 | T_YMM | T_EW0 | T_SAE_Z | T_B32},
 			{ 0x52, "vpdpwssd", T_66 | T_0F38 | T_YMM | T_EW0 | T_SAE_Z | T_B32},
 			{ 0x53, "vpdpwssds", T_66 | T_0F38 | T_YMM | T_EW0 | T_SAE_Z | T_B32},
+			{ 0xB4, "vpmadd52luq", T_66 | T_0F38 | T_YMM | T_EW1 | T_B64 },
+			{ 0xB5, "vpmadd52huq", T_66 | T_0F38 | T_YMM | T_EW1 | T_B64 },
+		};
+		for (size_t i = 0; i < NUM_OF_ARRAY(tbl); i++) {
+			const Tbl *p = &tbl[i];
+			std::string type = type2String(p->type);
+			printf("void %s(const Xmm& x1, const Xmm& x2, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opEncoding(x1, x2, op, %s, 0x%02X, encoding); }\n", p->name, type.c_str(), p->code);
+		}
+	}
+	// avx-vnni-int8
+	{
+		const struct Tbl {
+			uint8_t code;
+			const char *name;
+			int type;
+		} tbl[] = {
+			{ 0x50, "vpdpbssd", T_F2 | T_0F38 | T_W0 | T_YMM },
+			{ 0x51, "vpdpbssds", T_F2 | T_0F38 | T_W0 | T_YMM },
+			{ 0x50, "vpdpbsud", T_F3 | T_0F38 | T_W0 | T_YMM },
+			{ 0x51, "vpdpbsuds", T_F3 | T_0F38 | T_W0 | T_YMM },
+			{ 0x50, "vpdpbuud", T_0F38 | T_W0 | T_YMM },
+			{ 0x51, "vpdpbuuds", T_0F38 | T_W0 | T_YMM },
 		};
 		for (size_t i = 0; i < NUM_OF_ARRAY(tbl); i++) {
 			const Tbl *p = &tbl[i];
 			std::string type = type2String(p->type);
-			printf("void %s(const Xmm& x1, const Xmm& x2, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opVnni(x1, x2, op, %s, 0x%02X, encoding); }\n", p->name, type.c_str(), p->code);
+			printf("void %s(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, %s, 0x%02X); }\n", p->name, type.c_str(), p->code);
 		}
 	}
 }
@@ -1843,6 +1884,7 @@ void putAMX_INT8()
 	puts("void tdpbsud(const Tmm& x1, const Tmm& x2, const Tmm& x3) { opVex(x1, &x3, x2, T_F3 | T_0F38 | T_W0, 0x5e); }");
 	puts("void tdpbusd(const Tmm& x1, const Tmm& x2, const Tmm& x3) { opVex(x1, &x3, x2, T_66 | T_0F38 | T_W0, 0x5e); }");
 	puts("void tdpbuud(const Tmm& x1, const Tmm& x2, const Tmm& x3) { opVex(x1, &x3, x2, T_0F38 | T_W0, 0x5e); }");
+	puts("void tdpfp16ps(const Tmm &x1, const Tmm &x2, const Tmm &x3) { opVex(x1, &x3, x2, T_F2 | T_0F38 | T_W0, 0x5c); }");
 }
 void putAMX_BF16()
 {
diff --git a/meson.build b/meson.build
index c708887..cb06c03 100644
--- a/meson.build
+++ b/meson.build
@@ -5,7 +5,7 @@
 project(
 	'xbyak',
 	'cpp',
-	version: '6.62',
+	version: '6.63',
 	license: 'BSD-3-Clause',
 	default_options: 'b_ndebug=if-release'
 )
diff --git a/readme.md b/readme.md
index a497791..20f08e8 100644
--- a/readme.md
+++ b/readme.md
@@ -1,5 +1,5 @@
 
-# Xbyak 6.62 [![Badge Build]][Build Status]
+# Xbyak 6.63 [![Badge Build]][Build Status]
 
 *A C++ JIT assembler for x86 (IA32), x64 (AMD64, x86-64)*
 
@@ -28,6 +28,7 @@ If you want to use them, then specify `-fno-operator-names` option to gcc/clang.
 
 ### News
 
+- add amx_fp16/avx_vnni_int8/avx_ne_convert/avx-ifma
 - add movdiri, movdir64b, clwb, cldemote
 - WAITPKG instructions (tpause, umonitor, umwait) are supported.
 - MmapAllocator supports memfd with user-defined strings. see sample/memfd.cpp
diff --git a/readme.txt b/readme.txt
index 712b6a1..bf7875e 100644
--- a/readme.txt
+++ b/readme.txt
@@ -1,5 +1,5 @@
 

-    C++用x86(IA-32), x64(AMD64, x86-64) JITアセンブラ Xbyak 6.62

+    C++用x86(IA-32), x64(AMD64, x86-64) JITアセンブラ Xbyak 6.63

 

 -----------------------------------------------------------------------------

 ◎概要

@@ -166,13 +166,15 @@ vfpclassps k5{k3}, [rax+64]{1to4}, 5    --> vfpclassps(k5|k3, xword_b [rax+64],
 vpdpbusd(xm0, xm1, xm2); // default encoding is EVEX

 vpdpbusd(xm0, xm1, xm2, EvexEncoding); // same as the above

 vpdpbusd(xm0, xm1, xm2, VexEncoding); // VEX encoding

-

+setDefaultEncoding(VexEncoding); // default encoding is VEX

+vpdpbusd(xm0, xm1, xm2); // VEX encoding

 注意

 * k1, ..., k7 は新しいopmaskレジスタです。

 * z, sae, rn-sae, rd-sae, ru-sae, rz-saeの代わりにT_z, T_sae, T_rn_sae, T_rd_sae, T_ru_sae, T_rz_saeを使ってください。

 * `k4 | k3`と`k3 | k4`は意味が異なります。

 * {1toX}の代わりにptr_bを使ってください。Xは自動的に決まります。

 * 一部の命令はメモリサイズを指定するためにxword/yword/zword(_b)を使ってください。

+* setDefaultEncoding()でencoding省略時のEVEX/VEXを設定できます。

 

 ・ラベル

 

@@ -400,6 +402,8 @@ sample/{echo,hello}.bfは http://www.kmonos.net/alang/etc/brainfuck.php から
 -----------------------------------------------------------------------------

 ◎履歴

 

+2022/10/06 ver 6.63 AVX-IFMA用のvpmadd52{h,l}uq対応

+2022/10/05          amx_fp16/avx_vnni_int8/avx_ne_convertt対応とsetDefaultEncoding()追加

 2022/09/15 ver 6.62 serialize追加

 2022/08/02 ver 6.61.1 noexceptはVisual Studio 2015以降対応

 2022/07/29 ver 6.61 movzx eax, ahがエラーになるのを修正

diff --git a/sample/test_util.cpp b/sample/test_util.cpp
index 2488ce1..a20d2df 100644
--- a/sample/test_util.cpp
+++ b/sample/test_util.cpp
@@ -89,6 +89,10 @@ void putCPUinfo(bool onlyCpuidFeature)
 		{ Cpu::tMOVDIRI, "movdiri" },
 		{ Cpu::tMOVDIR64B, "movdir64b" },
 		{ Cpu::tCLZERO, "clzero" },
+		{ Cpu::tAMX_FP16, "amx_fp16" },
+		{ Cpu::tAVX_VNNI_INT8, "avx_vnni_int8" },
+		{ Cpu::tAVX_NE_CONVERT, "avx_ne_convert" },
+		{ Cpu::tAVX_IFMA, "avx_ifma" },
 	};
 	for (size_t i = 0; i < NUM_OF_ARRAY(tbl); i++) {
 		if (cpu.has(tbl[i].type)) printf(" %s", tbl[i].str);
diff --git a/test/misc.cpp b/test/misc.cpp
index e86c953..5f7db7e 100644
--- a/test/misc.cpp
+++ b/test/misc.cpp
@@ -876,6 +876,10 @@ CYBOZU_TEST_AUTO(vnni)
 			vpdpbusd(xm0, xm1, xm2);
 			vpdpbusd(xm0, xm1, xm2, EvexEncoding); // EVEX
 			vpdpbusd(xm0, xm1, xm2, VexEncoding); // VEX
+			setDefaultEncoding(VexEncoding);
+			vpdpbusd(xm0, xm1, xm2); // VEX
+			setDefaultEncoding(EvexEncoding);
+			vpdpbusd(xm0, xm1, xm2); // EVEX
 		}
 		void badVex()
 		{
@@ -886,6 +890,8 @@ CYBOZU_TEST_AUTO(vnni)
 		0x62, 0xF2, 0x75, 0x08, 0x50, 0xC2,
 		0x62, 0xF2, 0x75, 0x08, 0x50, 0xC2,
 		0xC4, 0xE2, 0x71, 0x50, 0xC2,
+		0xC4, 0xE2, 0x71, 0x50, 0xC2,
+		0x62, 0xF2, 0x75, 0x08, 0x50, 0xC2,
 	};
 	const size_t n = sizeof(tbl) / sizeof(tbl[0]);
 	CYBOZU_TEST_EQUAL(c.getSize(), n);
diff --git a/xbyak/xbyak.h b/xbyak/xbyak.h
index d7955bf..16a0ed8 100644
--- a/xbyak/xbyak.h
+++ b/xbyak/xbyak.h
@@ -155,7 +155,7 @@ namespace Xbyak {
 
 enum {
 	DEFAULT_MAX_CODE_SIZE = 4096,
-	VERSION = 0x6620 /* 0xABCD = A.BC(.D) */
+	VERSION = 0x6630 /* 0xABCD = A.BC(.D) */
 };
 
 #ifndef MIE_INTEGER_TYPE_DEFINED
@@ -2419,18 +2419,21 @@ private:
 		if (addr.getRegExp().getIndex().getKind() != kind) XBYAK_THROW(ERR_BAD_VSIB_ADDRESSING)
 		opVex(x, 0, addr, type, code);
 	}
-	void opVnni(const Xmm& x1, const Xmm& x2, const Operand& op, int type, int code0, PreferredEncoding encoding)
+	void opEncoding(const Xmm& x1, const Xmm& x2, const Operand& op, int type, int code0, PreferredEncoding encoding)
 	{
+		opAVX_X_X_XM(x1, x2, op, type | orEvexIf(encoding), code0);
+	}
+	int orEvexIf(PreferredEncoding encoding) {
 		if (encoding == DefaultEncoding) {
-			encoding = EvexEncoding;
+			encoding = defaultEncoding_;
 		}
 		if (encoding == EvexEncoding) {
 #ifdef XBYAK_DISABLE_AVX512
 			XBYAK_THROW(ERR_EVEX_IS_INVALID)
 #endif
-			type |= T_MUST_EVEX;
+			return T_MUST_EVEX;
 		}
-		opAVX_X_X_XM(x1, x2, op, type, code0);
+		return 0;
 	}
 	void opInOut(const Reg& a, const Reg& d, uint8_t code)
 	{
@@ -2515,6 +2518,7 @@ public:
 #endif
 private:
 	bool isDefaultJmpNEAR_;
+	PreferredEncoding defaultEncoding_;
 public:
 	void L(const std::string& label) { labelMgr_.defineSlabel(label); }
 	void L(Label& label) { labelMgr_.defineClabel(label); }
@@ -2794,6 +2798,7 @@ public:
 		, es(Segment::es), cs(Segment::cs), ss(Segment::ss), ds(Segment::ds), fs(Segment::fs), gs(Segment::gs)
 #endif
 		, isDefaultJmpNEAR_(false)
+		, defaultEncoding_(EvexEncoding)
 	{
 		labelMgr_.set(this);
 	}
@@ -2830,6 +2835,9 @@ public:
 	#undef jnl
 #endif
 
+	// set default encoding to select Vex or Evex
+	void setDefaultEncoding(PreferredEncoding encoding) { defaultEncoding_ = encoding; }
+
 	/*
 		use single byte nop if useMultiByteNop = false
 	*/
diff --git a/xbyak/xbyak_mnemonic.h b/xbyak/xbyak_mnemonic.h
index c63ca2b..bf2ab70 100644
--- a/xbyak/xbyak_mnemonic.h
+++ b/xbyak/xbyak_mnemonic.h
@@ -1,4 +1,4 @@
-const char *getVersionString() const { return "6.62"; }
+const char *getVersionString() const { return "6.63"; }
 void adc(const Operand& op, uint32_t imm) { opRM_I(op, imm, 0x10, 2); }
 void adc(const Operand& op1, const Operand& op2) { opRM_RM(op1, op2, 0x10); }
 void adcx(const Reg32e& reg, const Operand& op) { opGen(reg, op, 0xF6, 0x66, isREG32_REG32orMEM, NONE, 0x38); }
@@ -845,6 +845,8 @@ void vandnpd(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand())
 void vandnps(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand()) { opAVX_X_X_XM(xmm, op1, op2, T_0F | T_EW0 | T_YMM | T_EVEX | T_ER_Z | T_B32, 0x55); }
 void vandpd(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand()) { opAVX_X_X_XM(xmm, op1, op2, T_0F | T_66 | T_EW1 | T_YMM | T_EVEX | T_ER_Z | T_B64, 0x54); }
 void vandps(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand()) { opAVX_X_X_XM(xmm, op1, op2, T_0F | T_EW0 | T_YMM | T_EVEX | T_ER_Z | T_B32, 0x54); }
+void vbcstnebf162ps(const Xmm& x, const Address& addr) { opVex(x, 0, addr, T_F3 | T_0F38 | T_W0 | T_YMM | T_B16, 0xB1); }
+void vbcstnesh2ps(const Xmm& x, const Address& addr) { opVex(x, 0, addr, T_66 | T_0F38 | T_W0 | T_YMM | T_B16, 0xB1); }
 void vblendpd(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F3A | T_W0 | T_YMM, 0x0D, imm); }
 void vblendps(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F3A | T_W0 | T_YMM, 0x0C, imm); }
 void vblendvpd(const Xmm& x1, const Xmm& x2, const Operand& op, const Xmm& x4) { opAVX_X_X_XM(x1, x2, op, T_0F3A | T_66 | T_YMM, 0x4B, x4.getIdx() << 4); }
@@ -989,6 +991,11 @@ void vcomisd(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, T_N8 | T
 void vcomiss(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, T_N4 | T_0F | T_EW0 | T_EVEX | T_SAE_X, 0x2F); }
 void vcvtdq2pd(const Xmm& x, const Operand& op) { checkCvt1(x, op); opVex(x, 0, op, T_0F | T_F3 | T_YMM | T_EVEX | T_EW0 | T_B32 | T_N8 | T_N_VL, 0xE6); }
 void vcvtdq2ps(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, T_0F | T_EW0 | T_YMM | T_EVEX | T_ER_Z | T_B32, 0x5B); }
+void vcvtneebf162ps(const Xmm& x, const Address& addr) { opVex(x, 0, addr, T_F3 | T_0F38 | T_W0 | T_YMM, 0xB0); }
+void vcvtneeph2ps(const Xmm& x, const Address& addr) { opVex(x, 0, addr, T_66 | T_0F38 | T_W0 | T_YMM, 0xB0); }
+void vcvtneobf162ps(const Xmm& x, const Address& addr) { opVex(x, 0, addr, T_F2 | T_0F38 | T_W0 | T_YMM, 0xB0); }
+void vcvtneoph2ps(const Xmm& x, const Address& addr) { opVex(x, 0, addr, T_0F38 | T_W0 | T_YMM, 0xB0); }
+void vcvtneps2bf16(const Xmm& x, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opCvt2(x, op, T_F3 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_B32 | orEvexIf(encoding), 0x72); }
 void vcvtpd2dq(const Xmm& x, const Operand& op) { opCvt2(x, op, T_0F | T_F2 | T_YMM | T_EVEX | T_EW1 | T_B64 | T_ER_Z, 0xE6); }
 void vcvtpd2ps(const Xmm& x, const Operand& op) { opCvt2(x, op, T_0F | T_66 | T_YMM | T_EVEX | T_EW1 | T_B64 | T_ER_Z, 0x5A); }
 void vcvtph2ps(const Xmm& x, const Operand& op) { checkCvt1(x, op); opVex(x, 0, op, T_0F38 | T_66 | T_W0 | T_EVEX | T_EW0 | T_N8 | T_N_VL | T_SAE_Y, 0x13); }
@@ -1192,10 +1199,16 @@ void vpcmpgtq(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1
 void vpcmpgtw(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F | T_YMM, 0x65); }
 void vpcmpistri(const Xmm& xm, const Operand& op, uint8_t imm) { opAVX_X_XM_IMM(xm, op, T_66 | T_0F3A, 0x63, imm); }
 void vpcmpistrm(const Xmm& xm, const Operand& op, uint8_t imm) { opAVX_X_XM_IMM(xm, op, T_66 | T_0F3A, 0x62, imm); }
-void vpdpbusd(const Xmm& x1, const Xmm& x2, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opVnni(x1, x2, op, T_66 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_B32, 0x50, encoding); }
-void vpdpbusds(const Xmm& x1, const Xmm& x2, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opVnni(x1, x2, op, T_66 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_B32, 0x51, encoding); }
-void vpdpwssd(const Xmm& x1, const Xmm& x2, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opVnni(x1, x2, op, T_66 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_B32, 0x52, encoding); }
-void vpdpwssds(const Xmm& x1, const Xmm& x2, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opVnni(x1, x2, op, T_66 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_B32, 0x53, encoding); }
+void vpdpbssd(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_F2 | T_0F38 | T_W0 | T_YMM, 0x50); }
+void vpdpbssds(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_F2 | T_0F38 | T_W0 | T_YMM, 0x51); }
+void vpdpbsud(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_F3 | T_0F38 | T_W0 | T_YMM, 0x50); }
+void vpdpbsuds(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_F3 | T_0F38 | T_W0 | T_YMM, 0x51); }
+void vpdpbusd(const Xmm& x1, const Xmm& x2, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opEncoding(x1, x2, op, T_66 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_B32, 0x50, encoding); }
+void vpdpbusds(const Xmm& x1, const Xmm& x2, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opEncoding(x1, x2, op, T_66 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_B32, 0x51, encoding); }
+void vpdpbuud(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_0F38 | T_W0 | T_YMM, 0x50); }
+void vpdpbuuds(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_0F38 | T_W0 | T_YMM, 0x51); }
+void vpdpwssd(const Xmm& x1, const Xmm& x2, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opEncoding(x1, x2, op, T_66 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_B32, 0x52, encoding); }
+void vpdpwssds(const Xmm& x1, const Xmm& x2, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opEncoding(x1, x2, op, T_66 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_B32, 0x53, encoding); }
 void vperm2f128(const Ymm& y1, const Ymm& y2, const Operand& op, uint8_t imm) { if (!(y1.isYMM() && y2.isYMM() && op.isYMEM())) XBYAK_THROW(ERR_BAD_COMBINATION) opVex(y1, &y2, op, T_0F3A | T_66 | T_W0 | T_YMM, 0x06, imm); }
 void vperm2i128(const Ymm& y1, const Ymm& y2, const Operand& op, uint8_t imm) { if (!(y1.isYMM() && y2.isYMM() && op.isYMEM())) XBYAK_THROW(ERR_BAD_COMBINATION) opVex(y1, &y2, op, T_0F3A | T_66 | T_W0 | T_YMM, 0x46, imm); }
 void vpermd(const Ymm& y1, const Ymm& y2, const Operand& op) { opAVX_X_X_XM(y1, y2, op, T_66 | T_0F38 | T_W0 | T_EW0 | T_YMM | T_EVEX | T_B32, 0x36); }
@@ -1227,6 +1240,8 @@ void vpinsrb(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { if
 void vpinsrd(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { if (!(x1.isXMM() && x2.isXMM() && (op.isREG(32) || op.isMEM()))) XBYAK_THROW(ERR_BAD_COMBINATION) opVex(x1, &x2, op, T_0F3A | T_66 | T_W0 | T_EVEX | T_EW0 | T_N4, 0x22, imm); }
 void vpinsrq(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { if (!(x1.isXMM() && x2.isXMM() && (op.isREG(64) || op.isMEM()))) XBYAK_THROW(ERR_BAD_COMBINATION) opVex(x1, &x2, op, T_0F3A | T_66 | T_W1 | T_EVEX | T_EW1 | T_N8, 0x22, imm); }
 void vpinsrw(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { if (!(x1.isXMM() && x2.isXMM() && (op.isREG(32) || op.isMEM()))) XBYAK_THROW(ERR_BAD_COMBINATION) opVex(x1, &x2, op, T_0F | T_66 | T_EVEX | T_N2, 0xC4, imm); }
+void vpmadd52huq(const Xmm& x1, const Xmm& x2, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opEncoding(x1, x2, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_B64, 0xB5, encoding); }
+void vpmadd52luq(const Xmm& x1, const Xmm& x2, const Operand& op, PreferredEncoding encoding = DefaultEncoding) { opEncoding(x1, x2, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_B64, 0xB4, encoding); }
 void vpmaddubsw(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F38 | T_YMM | T_EVEX, 0x04); }
 void vpmaddwd(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F | T_YMM | T_EVEX, 0xF5); }
 void vpmaskmovd(const Address& addr, const Xmm& x1, const Xmm& x2) { opAVX_X_X_XM(x2, x1, addr, T_0F38 | T_66 | T_W0 | T_YMM, 0x8E); }
@@ -1654,6 +1669,7 @@ void tdpbssd(const Tmm& x1, const Tmm& x2, const Tmm& x3) { opVex(x1, &x3, x2, T
 void tdpbsud(const Tmm& x1, const Tmm& x2, const Tmm& x3) { opVex(x1, &x3, x2, T_F3 | T_0F38 | T_W0, 0x5e); }
 void tdpbusd(const Tmm& x1, const Tmm& x2, const Tmm& x3) { opVex(x1, &x3, x2, T_66 | T_0F38 | T_W0, 0x5e); }
 void tdpbuud(const Tmm& x1, const Tmm& x2, const Tmm& x3) { opVex(x1, &x3, x2, T_0F38 | T_W0, 0x5e); }
+void tdpfp16ps(const Tmm &x1, const Tmm &x2, const Tmm &x3) { opVex(x1, &x3, x2, T_F2 | T_0F38 | T_W0, 0x5c); }
 void tdpbf16ps(const Tmm& x1, const Tmm& x2, const Tmm& x3) { opVex(x1, &x3, x2, T_F3 | T_0F38 | T_W0, 0x5c); }
 #else
 void jcxz(std::string label) { db(0x67); opJmp(label, T_SHORT, 0xe3, 0, 0); }
@@ -1908,7 +1924,6 @@ void vcompressps(const Operand& op, const Xmm& x) { opAVX_X_XM_IMM(x, op, T_N4 |
 void vcompressw(const Operand& op, const Xmm& x) { opAVX_X_XM_IMM(x, op, T_N2 | T_66 | T_0F38 | T_EW1 | T_YMM | T_MUST_EVEX, 0x63); }
 void vcvtdq2ph(const Xmm& x, const Operand& op) { checkCvt4(x, op); opCvt(x, op, T_N16 | T_N_VL | T_MAP5 | T_EW0 | T_YMM | T_ER_Z | T_MUST_EVEX | T_B32, 0x5B); }
 void vcvtne2ps2bf16(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_F2 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_MUST_EVEX | T_B32, 0x72); }
-void vcvtneps2bf16(const Xmm& x, const Operand& op) { opCvt2(x, op, T_F3 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_MUST_EVEX | T_B32, 0x72); }
 void vcvtpd2ph(const Xmm& x, const Operand& op) { opCvt5(x, op, T_N16 | T_N_VL | T_66 | T_MAP5 | T_EW1 | T_ER_Z | T_MUST_EVEX | T_B64, 0x5A); }
 void vcvtpd2qq(const Xmm& x, const Operand& op) { opAVX_X_XM_IMM(x, op, T_66 | T_0F | T_EW1 | T_YMM | T_ER_Z | T_MUST_EVEX | T_B64, 0x7B); }
 void vcvtpd2udq(const Xmm& x, const Operand& op) { opCvt2(x, op, T_0F | T_EW1 | T_YMM | T_ER_Z | T_MUST_EVEX | T_B64, 0x79); }
@@ -2142,8 +2157,6 @@ void vpgatherqd(const Xmm& x, const Address& addr) { opGather2(x, addr, T_N4 | T
 void vpgatherqq(const Xmm& x, const Address& addr) { opGather2(x, addr, T_N8 | T_66 | T_0F38 | T_EW1 | T_YMM | T_MUST_EVEX | T_VSIB, 0x91, 0); }
 void vplzcntd(const Xmm& x, const Operand& op) { opAVX_X_XM_IMM(x, op, T_66 | T_0F38 | T_EW0 | T_YMM | T_MUST_EVEX | T_B32, 0x44); }
 void vplzcntq(const Xmm& x, const Operand& op) { opAVX_X_XM_IMM(x, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_MUST_EVEX | T_B64, 0x44); }
-void vpmadd52huq(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_MUST_EVEX | T_B64, 0xB5); }
-void vpmadd52luq(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_MUST_EVEX | T_B64, 0xB4); }
 void vpmaxsq(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_MUST_EVEX | T_B64, 0x3D); }
 void vpmaxuq(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_MUST_EVEX | T_B64, 0x3F); }
 void vpminsq(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_MUST_EVEX | T_B64, 0x39); }
diff --git a/xbyak/xbyak_util.h b/xbyak/xbyak_util.h
index 12c0068..8431925 100644
--- a/xbyak/xbyak_util.h
+++ b/xbyak/xbyak_util.h
@@ -410,6 +410,10 @@ public:
 	XBYAK_DEFINE_TYPE(65, tMOVDIRI);
 	XBYAK_DEFINE_TYPE(66, tMOVDIR64B);
 	XBYAK_DEFINE_TYPE(67, tCLZERO); // AMD Zen
+	XBYAK_DEFINE_TYPE(68, tAMX_FP16);
+	XBYAK_DEFINE_TYPE(69, tAVX_VNNI_INT8);
+	XBYAK_DEFINE_TYPE(70, tAVX_NE_CONVERT);
+	XBYAK_DEFINE_TYPE(71, tAVX_IFMA);
 
 #undef XBYAK_SPLIT_ID
 #undef XBYAK_DEFINE_TYPE
@@ -553,6 +557,10 @@ public:
 				if (type_ & tAVX512F) {
 					if (EAX & (1U << 5)) type_ |= tAVX512_BF16;
 				}
+				if (EAX & (1U << 21)) type_ |= tAMX_FP16;
+				if (EAX & (1U << 23)) type_ |= tAVX_IFMA;
+				if (EDX & (1U << 4)) type_ |= tAVX_VNNI_INT8;
+				if (EDX & (1U << 5)) type_ |= tAVX_NE_CONVERT;
 			}
 		}
 		setFamily();