add avx10 minmax

herumi · Oct 13, 2024 · 86e532f · 86e532f
1 parent aabf2ab
commit 86e532f
Show file tree

Hide file tree

Showing 3 changed files with 80 additions and 0 deletions.
diff --git a/gen/gen_avx512.cpp b/gen/gen_avx512.cpp
@@ -447,6 +447,13 @@ void putX_X_XM_IMM()
 		{ 0x1B, "vcvtne2ph2hf8s", T_MUST_EVEX | T_F2 | T_MAP5 | T_EW0 | T_YMM | T_B16 | T_N1, false },
 
 		{ 0x52, "vdpphps", T_MUST_EVEX | T_0F38 | T_EW0 | T_YMM | T_B32, false },
+		{ 0x52, "vminmaxnepbf16", T_MUST_EVEX | T_F2 | T_0F3A | T_EW0 | T_YMM | T_B16, true },
+		{ 0x52, "vminmaxpd", T_MUST_EVEX | T_66 | T_0F3A | T_EW1 | T_YMM | T_B64 | T_SAE_Y | T_SAE_Z, true },
+		{ 0x52, "vminmaxph", T_MUST_EVEX | T_0F3A | T_EW0 | T_YMM | T_B16 | T_SAE_Y | T_SAE_Z, true },
+		{ 0x52, "vminmaxps", T_MUST_EVEX | T_66 | T_0F3A | T_EW0 | T_YMM | T_B32 | T_SAE_Y | T_SAE_Z, true },
+		{ 0x53, "vminmaxsd", T_MUST_EVEX | T_66 | T_0F3A | T_EW1 | T_SAE_X | T_N8, true },
+		{ 0x53, "vminmaxsh", T_MUST_EVEX | T_0F3A | T_EW0 | T_SAE_X | T_N2, true },
+		{ 0x53, "vminmaxss", T_MUST_EVEX | T_66 | T_0F3A | T_EW0 | T_SAE_X | T_N4, true },
 	};
 	for (size_t i = 0; i < NUM_OF_ARRAY(tbl); i++) {
 		const Tbl *p = &tbl[i];

diff --git a/test/avx10/minmax.txt b/test/avx10/minmax.txt
@@ -0,0 +1,66 @@
+vminmaxnepbf16(xm1|k3|T_z, xm2, xm3, 5);
+vminmaxnepbf16(xm1|k3|T_z, xm2, ptr[rax+128], 5);
+vminmaxnepbf16(xm1|k3|T_z, xm2, ptr_b[rax+128], 5);
+
+vminmaxnepbf16(ym1|k3|T_z, ym2, ym3, 5);
+vminmaxnepbf16(ym1|k3|T_z, ym2, ptr[rax+128], 5);
+vminmaxnepbf16(ym1|k3|T_z, ym2, ptr_b[rax+128], 5);
+
+vminmaxnepbf16(zm1|k3|T_z, zm2, zm3, 5);
+vminmaxnepbf16(zm1|k3|T_z, zm2, ptr[rax+128], 5);
+vminmaxnepbf16(zm1|k3|T_z, zm2, ptr_b[rax+128], 5);
+//
+vminmaxpd(xm1|k3|T_z, xm2, xm3, 5);
+vminmaxpd(xm1|k3|T_z, xm2, ptr[rax+128], 5);
+vminmaxpd(xm1|k3|T_z, xm2, ptr_b[rax+128], 5);
+
+vminmaxpd(ym1|k3|T_z, ym2, ym3, 5);
+vminmaxpd(ym1|k3|T_z, ym2, ym3|T_sae, 5);
+vminmaxpd(ym1|k3|T_z, ym2, ptr[rax+128], 5);
+vminmaxpd(ym1|k3|T_z, ym2, ptr_b[rax+128], 5);
+
+vminmaxpd(zm1|k3|T_z, zm2, zm3, 5);
+vminmaxpd(zm1|k3|T_z, zm2, zm3|T_sae, 5);
+vminmaxpd(zm1|k3|T_z, zm2, ptr[rax+128], 5);
+vminmaxpd(zm1|k3|T_z, zm2, ptr_b[rax+128], 5);
+//
+vminmaxph(xm1|k3|T_z, xm2, xm3, 5);
+vminmaxph(xm1|k3|T_z, xm2, ptr[rax+128], 5);
+vminmaxph(xm1|k3|T_z, xm2, ptr[rax+128], 5);
+vminmaxph(xm1|k3|T_z, xm2, ptr_b[rax+128], 5);
+
+vminmaxph(ym1|k3|T_z, ym2, ym3, 5);
+vminmaxph(ym1|k3|T_z, ym2, ym3|T_sae, 5);
+vminmaxph(ym1|k3|T_z, ym2, ptr[rax+128], 5);
+vminmaxph(ym1|k3|T_z, ym2, ptr_b[rax+128], 5);
+
+vminmaxph(zm1|k3|T_z, zm2, zm3, 5);
+vminmaxph(zm1|k3|T_z, zm2, zm3|T_sae, 5);
+vminmaxph(zm1|k3|T_z, zm2, ptr[rax+128], 5);
+vminmaxph(zm1|k3|T_z, zm2, ptr_b[rax+128], 5);
+//
+vminmaxps(xm1|k3|T_z, xm2, xm3, 5);
+vminmaxps(xm1|k3|T_z, xm2, ptr[rax+128], 5);
+vminmaxps(xm1|k3|T_z, xm2, ptr_b[rax+128], 5);
+
+vminmaxps(ym1|k3|T_z, ym2, ym3, 5);
+vminmaxps(ym1|k3|T_z, ym2, ym3|T_sae, 5);
+vminmaxps(ym1|k3|T_z, ym2, ptr[rax+128], 5);
+vminmaxps(ym1|k3|T_z, ym2, ptr_b[rax+128], 5);
+
+vminmaxps(zm1|k3|T_z, zm2, zm3, 5);
+vminmaxps(zm1|k3|T_z, zm2, zm3|T_sae, 5);
+vminmaxps(zm1|k3|T_z, zm2, ptr[rax+128], 5);
+vminmaxps(zm1|k3|T_z, zm2, ptr_b[rax+128], 5);
+//
+vminmaxsd(xm1|k3|T_z, xm2, xm3, 5);
+vminmaxsd(xm1|k3|T_z, xm2, xm3|T_sae, 5);
+vminmaxsd(xm1|k3|T_z, xm2, ptr[rax+128], 5);
+//
+vminmaxsh(xm1|k3|T_z, xm2, xm3, 5);
+vminmaxsh(xm1|k3|T_z, xm2, xm3|T_sae, 5);
+vminmaxsh(xm1|k3|T_z, xm2, ptr[rax+128], 5);
+//
+vminmaxss(xm1|k3|T_z, xm2, xm3, 5);
+vminmaxss(xm1|k3|T_z, xm2, xm3|T_sae, 5);
+vminmaxss(xm1|k3|T_z, xm2, ptr[rax+128], 5);
diff --git a/xbyak/xbyak_mnemonic.h b/xbyak/xbyak_mnemonic.h
@@ -2374,6 +2374,13 @@ void vinserti64x4(const Zmm& r1, const Zmm& r2, const Operand& op, uint8_t imm)
 void vmaxpbf16(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66|T_MAP5|T_EW0|T_YMM|T_MUST_EVEX|T_B16, 0x5F); }
 void vmaxph(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand()) { opAVX_X_X_XM(xmm, op1, op2, T_MAP5 | T_EW0 | T_YMM | T_MUST_EVEX | T_SAE_Z | T_B16, 0x5F); }
 void vmaxsh(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand()) { opAVX_X_X_XM(xmm, op1, op2, T_MAP5 | T_F3 | T_EW0 | T_MUST_EVEX | T_SAE_X | T_N2, 0x5F); }
+void vminmaxnepbf16(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { opAVX_X_X_XM(x1, x2, op, T_F2|T_0F3A|T_EW0|T_YMM|T_MUST_EVEX|T_B16, 0x52, imm); }
+void vminmaxpd(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { opAVX_X_X_XM(x1, x2, op, T_66|T_0F3A|T_EW1|T_YMM|T_SAE_Y|T_SAE_Z|T_MUST_EVEX|T_B64, 0x52, imm); }
+void vminmaxph(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { opAVX_X_X_XM(x1, x2, op, T_0F3A|T_EW0|T_YMM|T_SAE_Y|T_SAE_Z|T_MUST_EVEX|T_B16, 0x52, imm); }
+void vminmaxps(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { opAVX_X_X_XM(x1, x2, op, T_66|T_0F3A|T_EW0|T_YMM|T_SAE_Y|T_SAE_Z|T_MUST_EVEX|T_B32, 0x52, imm); }
+void vminmaxsd(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { opAVX_X_X_XM(x1, x2, op, T_N8|T_66|T_0F3A|T_EW1|T_SAE_X|T_MUST_EVEX, 0x53, imm); }
+void vminmaxsh(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { opAVX_X_X_XM(x1, x2, op, T_N2|T_0F3A|T_EW0|T_SAE_X|T_MUST_EVEX, 0x53, imm); }
+void vminmaxss(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { opAVX_X_X_XM(x1, x2, op, T_N4|T_66|T_0F3A|T_EW0|T_SAE_X|T_MUST_EVEX, 0x53, imm); }
 void vminpbf16(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66|T_MAP5|T_EW0|T_YMM|T_MUST_EVEX|T_B16, 0x5D); }
 void vminph(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand()) { opAVX_X_X_XM(xmm, op1, op2, T_MAP5 | T_EW0 | T_YMM | T_MUST_EVEX | T_SAE_Z | T_B16, 0x5D); }
 void vminsh(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand()) { opAVX_X_X_XM(xmm, op1, op2, T_MAP5 | T_F3 | T_EW0 | T_MUST_EVEX | T_SAE_X | T_N2, 0x5D); }