[clang] [llvm] [X86][AMX] Support AMX-TRANSPOSE, part 2 (PR #115660)

[Feng Zou via cfe-commits]

================
@@ -0,0 +1,301 @@
+/*===----- amxcomplextransposeintrin.h - AMX-COMPLEX and AMX-TRANSPOSE ------===
+ *
+ * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+ * See https://llvm.org/LICENSE.txt for license information.
+ * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+ *
+ *===------------------------------------------------------------------------===
+ */
+
+#ifndef __IMMINTRIN_H
+#error                                                                         \
+    "Never use <amxcomplextransposeintrin.h> directly; include <immintrin.h> instead."
+#endif // __IMMINTRIN_H
+
+#ifndef __AMX_COMPLEXTRANSPOSEINTRIN_H
+#define __AMX_COMPLEXTRANSPOSEINTRIN_H
+#ifdef __x86_64__
+
+#define __DEFAULT_FN_ATTRS                                                     \
+  __attribute__((__always_inline__, __nodebug__,                               \
+                 __target__("amx-complex,amx-transpose")))
+
+/// Perform matrix multiplication of two tiles containing complex elements and
+///    accumulate the results into a packed single precision tile. Each dword
+///    element in input tiles \a a and \a b is interpreted as a complex number
+///    with FP16 real part and FP16 imaginary part.
+/// Calculates the imaginary part of the result. For each possible combination
+///    of (transposed column of \a a, column of \a b), it performs a set of
+///    multiplication and accumulations on all corresponding complex numbers
+///    (one from \a a and one from \a b). The imaginary part of the \a a element
+///    is multiplied with the real part of the corresponding \a b element, and
+///    the real part of the \a a element is multiplied with the imaginary part
+///    of the corresponding \a b elements. The two accumulated results are
+///    added, and then accumulated into the corresponding row and column of
+///    \a dst.
+///
+/// \headerfile <x86intrin.h>
+///
+/// \code
+/// void _tile_tcmmimfp16ps(__tile dst, __tile a, __tile b);
+/// \endcode
+///
+/// \code{.operation}
+/// FOR m := 0 TO dst.rows - 1
+///	tmp := dst.row[m]
+///	FOR k := 0 TO a.rows - 1
+///		FOR n := 0 TO (dst.colsb / 4) - 1
+///			tmp.fp32[n] += FP32(a.row[m].fp16[2*k+0]) * FP32(b.row[k].fp16[2*n+1])
+///			tmp.fp32[n] += FP32(a.row[m].fp16[2*k+1]) * FP32(b.row[k].fp16[2*n+0])
+///		ENDFOR
+///	ENDFOR
+///	write_row_and_zero(dst, m, tmp, dst.colsb)
+/// ENDFOR
+/// zero_upper_rows(dst, dst.rows)
+/// zero_tileconfig_start()
+/// \endcode
+///
+/// This intrinsic corresponds to the \c TTCMMIMFP16PS instruction.
+///
+/// \param dst
+///    The destination tile. Max size is 1024 Bytes.
+/// \param a
+///    The 1st source tile. Max size is 1024 Bytes.
+/// \param b
+///    The 2nd source tile. Max size is 1024 Bytes.
+#define _tile_tcmmimfp16ps(dst, a, b) __builtin_ia32_ttcmmimfp16ps(dst, a, b)
+
+/// Perform matrix multiplication of two tiles containing complex elements and
+///    accumulate the results into a packed single precision tile. Each dword
+///    element in input tiles \a a and \a b is interpreted as a complex number
+///    with FP16 real part and FP16 imaginary part.
+/// Calculates the real part of the result. For each possible combination
+///    of (rtransposed colum of \a a, column of \a b), it performs a set of
+///    multiplication and accumulations on all corresponding complex numbers
+///    (one from \a a and one from \a b). The real part of the \a a element is
+///    multiplied with the real part of the corresponding \a b element, and the
+///    negated imaginary part of the \a a element is multiplied with the
+///    imaginary part of the corresponding \a b elements. The two accumulated
+///    results are added, and then accumulated into the corresponding row and
+///    column of \a dst.
+///
+/// \headerfile <x86intrin.h>
+///
+/// \code
+/// void _tile_tcmmrlfp16ps(__tile dst, __tile a, __tile b);
+/// \endcode
+///
+/// \code{.operation}
+/// FOR m := 0 TO dst.rows - 1
+///	tmp := dst.row[m]
+///	FOR k := 0 TO a.rows - 1
+///		FOR n := 0 TO (dst.colsb / 4) - 1
+///			tmp.fp32[n] += FP32(a.row[m].fp16[2*k+0]) * FP32(b.row[k].fp16[2*n+0])
+///			tmp.fp32[n] += FP32(-a.row[m].fp16[2*k+1]) * FP32(b.row[k].fp16[2*n+1])
+///		ENDFOR
+///	ENDFOR
+///	write_row_and_zero(dst, m, tmp, dst.colsb)
+/// ENDFOR
+/// zero_upper_rows(dst, dst.rows)
+/// zero_tileconfig_start()
+/// \endcode
+///
+/// This intrinsic corresponds to the \c TTCMMIMFP16PS instruction.
+///
+/// \param dst
+///    The destination tile. Max size is 1024 Bytes.
+/// \param a
+///    The 1st source tile. Max size is 1024 Bytes.
+/// \param b
+///    The 2nd source tile. Max size is 1024 Bytes.
+#define _tile_tcmmrlfp16ps(dst, a, b) __builtin_ia32_ttcmmrlfp16ps(dst, a, b)
+
+/// Perform matrix conjugate transpose and multiplication of two tiles
+///    containing complex elements and accumulate the results into a packed
+///    single precision tile. Each dword element in input tiles \a a and \a b
+///    is interpreted as a complex number with FP16 real part and FP16 imaginary
+///    part.
+/// Calculates the imaginary part of the result. For each possible combination
+///    of (transposed column of \a a, column of \a b), it performs a set of
+///    multiplication and accumulations on all corresponding complex numbers
+///    (one from \a a and one from \a b). The negated imaginary part of the \a a
+///    element is multiplied with the real part of the corresponding \a b
+///    element, and the real part of the \a a element is multiplied with the
+///    imaginary part of the corresponding \a b elements. The two accumulated
+///    results are added, and then accumulated into the corresponding row and
+///    column of \a dst.
+///
+/// \headerfile <x86intrin.h>
+///
+/// \code
+/// void _tile_conjtcmmimfp16ps(__tile dst, __tile a, __tile b);
+/// \endcode
+///
+/// \code{.operation}
+/// FOR m := 0 TO dst.rows - 1
+///	tmp := dst.row[m]
+///	FOR k := 0 TO a.rows - 1
+///		FOR n := 0 TO (dst.colsb / 4) - 1
+///			tmp.fp32[n] += FP32(a.row[m].fp16[2*k+0]) * FP32(b.row[k].fp16[2*n+1])
+///			tmp.fp32[n] += FP32(-a.row[m].fp16[2*k+1]) * FP32(b.row[k].fp16[2*n+0])
+///		ENDFOR
+///	ENDFOR
+///	write_row_and_zero(dst, m, tmp, dst.colsb)
+/// ENDFOR
+/// zero_upper_rows(dst, dst.rows)
+/// zero_tileconfig_start()
+/// \endcode
+///
+/// This intrinsic corresponds to the \c TCONJTCMMIMFP16PS instruction.
+///
+/// \param dst
+///    The destination tile. Max size is 1024 Bytes.
+/// \param a
+///    The 1st source tile. Max size is 1024 Bytes.
+/// \param b
+///    The 2nd source tile. Max size is 1024 Bytes.
+#define _tile_conjtcmmimfp16ps(dst, a, b)                                      \
+  __builtin_ia32_tconjtcmmimfp16ps(dst, a, b)
----------------
fzou1 wrote:

ditto

https://github.com/llvm/llvm-project/pull/115660