Author: Paul Robinson Date: 2023-05-31T06:33:18-07:00 New Revision: e5399f1d7cabfca90030ca03f52818e892aa389f
URL: https://github.com/llvm/llvm-project/commit/e5399f1d7cabfca90030ca03f52818e892aa389f DIFF: https://github.com/llvm/llvm-project/commit/e5399f1d7cabfca90030ca03f52818e892aa389f.diff LOG: [Headers][doc] Add shuffle-like intrinsic descriptions to avx2intrin.h Differential Revision: https://reviews.llvm.org/D151749 Added: Modified: clang/lib/Headers/avx2intrin.h Removed: ################################################################################ diff --git a/clang/lib/Headers/avx2intrin.h b/clang/lib/Headers/avx2intrin.h index be4d30a98195e..409401f46ba9b 100644 --- a/clang/lib/Headers/avx2intrin.h +++ b/clang/lib/Headers/avx2intrin.h @@ -41,24 +41,126 @@ _mm256_abs_epi32(__m256i __a) return (__m256i)__builtin_elementwise_abs((__v8si)__a); } +/// Converts the elements of two 256-bit vectors of [16 x i16] to 8-bit +/// integers using signed saturation, and returns the 256-bit result. +/// +/// \code{.operation} +/// FOR i := 0 TO 7 +/// j := i*16 +/// k := i*8 +/// result[7+k:k] := SATURATE8(__a[15+j:j]) +/// result[71+k:64+k] := SATURATE8(__b[15+j:j]) +/// result[135+k:128+k] := SATURATE8(__a[143+j:128+j]) +/// result[199+k:192+k] := SATURATE8(__b[143+j:128+j]) +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPACKSSWB instruction. +/// +/// \param __a +/// A 256-bit vector of [16 x i16] used to generate result[63:0] and +/// result[191:128]. +/// \param __b +/// A 256-bit vector of [16 x i16] used to generate result[127:64] and +/// result[255:192]. +/// \returns A 256-bit integer vector containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_packs_epi16(__m256i __a, __m256i __b) { return (__m256i)__builtin_ia32_packsswb256((__v16hi)__a, (__v16hi)__b); } +/// Converts the elements of two 256-bit vectors of [8 x i32] to 16-bit +/// integers using signed saturation, and returns the resulting 256-bit +/// vector of [16 x i16]. +/// +/// \code{.operation} +/// FOR i := 0 TO 3 +/// j := i*32 +/// k := i*16 +/// result[15+k:k] := SATURATE16(__a[31+j:j]) +/// result[79+k:64+k] := SATURATE16(__b[31+j:j]) +/// result[143+k:128+k] := SATURATE16(__a[159+j:128+j]) +/// result[207+k:192+k] := SATURATE16(__b[159+j:128+j]) +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPACKSSDW instruction. +/// +/// \param __a +/// A 256-bit vector of [8 x i32] used to generate result[63:0] and +/// result[191:128]. +/// \param __b +/// A 256-bit vector of [8 x i32] used to generate result[127:64] and +/// result[255:192]. +/// \returns A 256-bit vector of [16 x i16] containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_packs_epi32(__m256i __a, __m256i __b) { return (__m256i)__builtin_ia32_packssdw256((__v8si)__a, (__v8si)__b); } +/// Converts elements from two 256-bit vectors of [16 x i16] to 8-bit integers +/// using unsigned saturation, and returns the 256-bit result. +/// +/// \code{.operation} +/// FOR i := 0 TO 7 +/// j := i*16 +/// k := i*8 +/// result[7+k:k] := SATURATE8U(__a[15+j:j]) +/// result[71+k:64+k] := SATURATE8U(__b[15+j:j]) +/// result[135+k:128+k] := SATURATE8U(__a[143+j:128+j]) +/// result[199+k:192+k] := SATURATE8U(__b[143+j:128+j]) +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPACKUSWB instruction. +/// +/// \param __a +/// A 256-bit vector of [16 x i16] used to generate result[63:0] and +/// result[191:128]. +/// \param __b +/// A 256-bit vector of [16 x i16] used to generate result[127:64] and +/// result[255:192]. +/// \returns A 256-bit integer vector containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_packus_epi16(__m256i __a, __m256i __b) { return (__m256i)__builtin_ia32_packuswb256((__v16hi)__a, (__v16hi)__b); } +/// Converts elements from two 256-bit vectors of [8 x i32] to 16-bit integers +/// using unsigned saturation, and returns the resulting 256-bit vector of +/// [16 x i16]. +/// +/// \code{.operation} +/// FOR i := 0 TO 3 +/// j := i*32 +/// k := i*16 +/// result[15+k:k] := SATURATE16U(__V1[31+j:j]) +/// result[79+k:64+k] := SATURATE16U(__V2[31+j:j]) +/// result[143+k:128+k] := SATURATE16U(__V1[159+j:128+j]) +/// result[207+k:192+k] := SATURATE16U(__V2[159+j:128+j]) +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPACKUSDW instruction. +/// +/// \param __V1 +/// A 256-bit vector of [8 x i32] used to generate result[63:0] and +/// result[191:128]. +/// \param __V2 +/// A 256-bit vector of [8 x i32] used to generate result[127:64] and +/// result[255:192]. +/// \returns A 256-bit vector of [16 x i16] containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_packus_epi32(__m256i __V1, __m256i __V2) { @@ -215,6 +317,30 @@ _mm256_adds_epu16(__m256i __a, __m256i __b) return (__m256i)__builtin_elementwise_add_sat((__v16hu)__a, (__v16hu)__b); } +/// Uses the lower half of the 256-bit vector \a a as the upper half of a +/// temporary 256-bit value, and the lower half of the 256-bit vector \a b +/// as the lower half of the temporary value. Right-shifts the temporary +/// value by \a n bytes, and uses the lower 16 bytes of the shifted value +/// as the lower 16 bytes of the result. Uses the upper halves of \a a and +/// \a b to make another temporary value, right shifts by \a n, and uses +/// the lower 16 bytes of the shifted value as the upper 16 bytes of the +/// result. +/// +/// \headerfile <immintrin.h> +/// +/// \code +/// __m256i _mm256_alignr_epi8(__m256i a, __m256i b, const int n); +/// \endcode +/// +/// This intrinsic corresponds to the \c VPALIGNR instruction. +/// +/// \param a +/// A 256-bit integer vector containing source values. +/// \param b +/// A 256-bit integer vector containing source values. +/// \param n +/// An immediate value specifying the number of bytes to shift. +/// \returns A 256-bit integer vector containing the result. #define _mm256_alignr_epi8(a, b, n) \ ((__m256i)__builtin_ia32_palignr256((__v32qi)(__m256i)(a), \ (__v32qi)(__m256i)(b), (n))) @@ -243,6 +369,35 @@ _mm256_avg_epu16(__m256i __a, __m256i __b) return (__m256i)__builtin_ia32_pavgw256((__v16hi)__a, (__v16hi)__b); } +/// Merges 8-bit integer values from either of the two 256-bit vectors +/// \a __V1 or \a __V2, as specified by the 256-bit mask \a __M and returns +/// the resulting 256-bit integer vector. +/// +/// \code{.operation} +/// FOR i := 0 TO 31 +/// j := i*8 +/// IF __M[7+i] == 0 +/// result[7+j:j] := __V1[7+j:j] +/// ELSE +/// result[7+j:j] := __V2[7+j:j] +/// FI +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPBLENDVB instruction. +/// +/// \param __V1 +/// A 256-bit integer vector containing source values. +/// \param __V2 +/// A 256-bit integer vector containing source values. +/// \param __M +/// A 256-bit integer vector, with bit [7] of each byte specifying the +/// source for each corresponding byte of the result. When the mask bit +/// is 0, the byte is copied from \a __V1; otherwise, it is copied from +/// \a __V2. +/// \returns A 256-bit integer vector containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_blendv_epi8(__m256i __V1, __m256i __V2, __m256i __M) { @@ -250,6 +405,43 @@ _mm256_blendv_epi8(__m256i __V1, __m256i __V2, __m256i __M) (__v32qi)__M); } +/// Merges 16-bit integer values from either of the two 256-bit vectors +/// \a V1 or \a V2, as specified by the immediate integer operand \a M, +/// and returns the resulting 256-bit vector of [16 x i16]. +/// +/// \code{.operation} +/// FOR i := 0 TO 7 +/// j := i*16 +/// IF M[i] == 0 +/// result[7+j:j] := V1[7+j:j] +/// result[135+j:128+j] := V1[135+j:128+j] +/// ELSE +/// result[7+j:j] := V2[7+j:j] +/// result[135+j:128+j] := V2[135+j:128+j] +/// FI +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// \code +/// __m256i _mm256_blend_epi16(__m256i V1, __m256i V2, const int M); +/// \endcode +/// +/// This intrinsic corresponds to the \c VPBLENDW instruction. +/// +/// \param V1 +/// A 256-bit vector of [16 x i16] containing source values. +/// \param V2 +/// A 256-bit vector of [16 x i16] containing source values. +/// \param M +/// An immediate 8-bit integer operand, with bits [7:0] specifying the +/// source for each element of the result. The position of the mask bit +/// corresponds to the index of a copied value. When a mask bit is 0, the +/// element is copied from \a V1; otherwise, it is copied from \a V2. +/// \a M[0] determines the source for elements 0 and 8, \a M[1] for +/// elements 1 and 9, and so forth. +/// \returns A 256-bit vector of [16 x i16] containing the result. #define _mm256_blend_epi16(V1, V2, M) \ ((__m256i)__builtin_ia32_pblendw256((__v16hi)(__m256i)(V1), \ (__v16hi)(__m256i)(V2), (int)(M))) @@ -896,18 +1088,148 @@ _mm256_sad_epu8(__m256i __a, __m256i __b) return __builtin_ia32_psadbw256((__v32qi)__a, (__v32qi)__b); } +/// Shuffles 8-bit integers in the 256-bit integer vector \a __a acccording +/// to control information in the 256-bit integer vector \a __b, and +/// returns the 256-bit result. In effect there are two separate 128-bit +/// shuffles in the lower and upper halves. +/// +/// \code{.operation} +/// FOR i := 0 TO 31 +/// j := i*8 +/// IF __b[j+7] == 1 +/// result[j+7:j] := 0 +/// ELSE +/// k := __b[j+3:j] * 8 +/// IF i > 15 +/// k := k + 128 +/// FI +/// result[j+7:j] := __a[k+7:k] +/// FI +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPSHUFB instruction. +/// +/// \param __a +/// A 256-bit integer vector containing source values. +/// \param __b +/// A 256-bit integer vector containing control information to determine +/// what goes into the corresponding byte of the result. If bit 7 of the +/// control byte is 1, the result byte is 0; otherwise, bits 3:0 of the +/// control byte specify the index (within the same 128-bit half) of \a __a +/// to copy to the result byte. +/// \returns A 256-bit integer vector containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_shuffle_epi8(__m256i __a, __m256i __b) { return (__m256i)__builtin_ia32_pshufb256((__v32qi)__a, (__v32qi)__b); } +/// Shuffles 32-bit integers from the 256-bit vector of [8 x i32] in \a a +/// acccording to control information in the integer literal \a imm, and +/// returns the 256-bit result. In effect there are two parallel 128-bit +/// shuffles in the lower and upper halves. +/// +/// \code{.operation} +/// FOR i := 0 to 3 +/// j := i*32 +/// k := (imm >> i*2)[1:0] * 32 +/// result[j+31:j] := a[k+31:k] +/// result[128+j+31:128+j] := a[128+k+31:128+k] +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// \code +/// __m256i _mm256_shuffle_epi32(__m256i a, const int imm); +/// \endcode +/// +/// This intrinsic corresponds to the \c VPSHUFB instruction. +/// +/// \param a +/// A 256-bit vector of [8 x i32] containing source values. +/// \param imm +/// An immediate 8-bit value specifying which elements to copy from \a a. +/// \a imm[1:0] specifies the index in \a a for elements 0 and 4 of the +/// result, \a imm[3:2] specifies the index for elements 1 and 5, and so +/// forth. +/// \returns A 256-bit vector of [8 x i32] containing the result. #define _mm256_shuffle_epi32(a, imm) \ ((__m256i)__builtin_ia32_pshufd256((__v8si)(__m256i)(a), (int)(imm))) +/// Shuffles 16-bit integers from the 256-bit vector of [16 x i16] in \a a +/// acccording to control information in the integer literal \a imm, and +/// returns the 256-bit result. The upper 64 bits of each 128-bit half +/// are shuffled in parallel; the lower 64 bits of each 128-bit half are +/// copied from \a a unchanged. +/// +/// \code{.operation} +/// result[63:0] := a[63:0] +/// result[191:128] := a[191:128] +/// FOR i := 0 TO 3 +/// j := i * 16 + 64 +/// k := (imm >> i*2)[1:0] * 16 + 64 +/// result[j+15:j] := a[k+15:k] +/// result[128+j+15:128+j] := a[128+k+15:128+k] +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// \code +/// __m256i _mm256_shufflehi_epi16(__m256i a, const int imm); +/// \endcode +/// +/// This intrinsic corresponds to the \c VPSHUFHW instruction. +/// +/// \param a +/// A 256-bit vector of [16 x i16] containing source values. +/// \param imm +/// An immediate 8-bit value specifying which elements to copy from \a a. +/// \a imm[1:0] specifies the index in \a a for elements 4 and 8 of the +/// result, \a imm[3:2] specifies the index for elements 5 and 9, and so +/// forth. Indexes are offset by 4 (so 0 means index 4, and so forth). +/// \returns A 256-bit vector of [16 x i16] containing the result. #define _mm256_shufflehi_epi16(a, imm) \ ((__m256i)__builtin_ia32_pshufhw256((__v16hi)(__m256i)(a), (int)(imm))) +/// Shuffles 16-bit integers from the 256-bit vector of [16 x i16] \a a +/// acccording to control information in the integer literal \a imm, and +/// returns the 256-bit [16 x i16] result. The lower 64 bits of each +/// 128-bit half are shuffled; the upper 64 bits of each 128-bit half are +/// copied from \a a unchanged. +/// +/// \code{.operation} +/// result[127:64] := a[127:64] +/// result[255:192] := a[255:192] +/// FOR i := 0 TO 3 +/// j := i * 16 +/// k := (imm >> i*2)[1:0] * 16 +/// result[j+15:j] := a[k+15:k] +/// result[128+j+15:128+j] := a[128+k+15:128+k] +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// \code +/// __m256i _mm256_shufflelo_epi16(__m256i a, const int imm); +/// \endcode +/// +/// This intrinsic corresponds to the \c VPSHUFLW instruction. +/// +/// \param a +/// A 256-bit vector of [16 x i16] to use as a source of data for the +/// result. +/// \param imm +/// An immediate 8-bit value specifying which elements to copy from \a a. +/// \a imm[1:0] specifies the index in \a a for elements 0 and 8 of the +/// result, \a imm[3:2] specifies the index for elements 1 and 9, and so +/// forth. +/// \returns A 256-bit vector of [16 x i16] containing the result. #define _mm256_shufflelo_epi16(a, imm) \ ((__m256i)__builtin_ia32_pshuflw256((__v16hi)(__m256i)(a), (int)(imm))) @@ -1544,48 +1866,266 @@ _mm256_subs_epu16(__m256i __a, __m256i __b) return (__m256i)__builtin_elementwise_sub_sat((__v16hu)__a, (__v16hu)__b); } +/// Unpacks and interleaves 8-bit integers from parts of the 256-bit integer +/// vectors in \a __a and \a __b to form the 256-bit result. Specifically, +/// uses the upper 64 bits of each 128-bit half of \a __a and \a __b as +/// input; other bits in these parameters are ignored. +/// +/// \code{.operation} +/// result[7:0] := __a[71:64] +/// result[15:8] := __b[71:64] +/// result[23:16] := __a[79:72] +/// result[31:24] := __b[79:72] +/// . . . +/// result[127:120] := __b[127:120] +/// result[135:128] := __a[199:192] +/// . . . +/// result[255:248] := __b[255:248] +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPUNPCKHBW instruction. +/// +/// \param __a +/// A 256-bit integer vector used as the source for the even-numbered bytes +/// of the result. +/// \param __b +/// A 256-bit integer vector used as the source for the odd-numbered bytes +/// of the result. +/// \returns A 256-bit integer vector containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_unpackhi_epi8(__m256i __a, __m256i __b) { return (__m256i)__builtin_shufflevector((__v32qi)__a, (__v32qi)__b, 8, 32+8, 9, 32+9, 10, 32+10, 11, 32+11, 12, 32+12, 13, 32+13, 14, 32+14, 15, 32+15, 24, 32+24, 25, 32+25, 26, 32+26, 27, 32+27, 28, 32+28, 29, 32+29, 30, 32+30, 31, 32+31); } +/// Unpacks and interleaves 16-bit integers from parts of the 256-bit vectors +/// of [16 x i16] in \a __a and \a __b to return the resulting 256-bit +/// vector of [16 x i16]. Specifically, uses the upper 64 bits of each +/// 128-bit half of \a __a and \a __b as input; other bits in these +/// parameters are ignored. +/// +/// \code{.operation} +/// result[15:0] := __a[79:64] +/// result[31:16] := __b[79:64] +/// result[47:32] := __a[95:80] +/// result[63:48] := __b[95:80] +/// . . . +/// result[127:112] := __b[127:112] +/// result[143:128] := __a[211:196] +/// . . . +/// result[255:240] := __b[255:240] +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPUNPCKHWD instruction. +/// +/// \param __a +/// A 256-bit vector of [16 x i16] used as the source for the even-numbered +/// elements of the result. +/// \param __b +/// A 256-bit vector of [16 x i16] used as the source for the odd-numbered +/// elements of the result. +/// \returns A 256-bit vector of [16 x i16] containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_unpackhi_epi16(__m256i __a, __m256i __b) { return (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)__b, 4, 16+4, 5, 16+5, 6, 16+6, 7, 16+7, 12, 16+12, 13, 16+13, 14, 16+14, 15, 16+15); } +/// Unpacks and interleaves 32-bit integers from parts of the 256-bit vectors +/// of [8 x i32] in \a __a and \a __b to return the resulting 256-bit vector +/// of [8 x i32]. Specifically, uses the upper 64 bits of each 128-bit half +/// of \a __a and \a __b as input; other bits in these parameters are +/// ignored. +/// +/// \code{.operation} +/// result[31:0] := __a[95:64] +/// result[63:32] := __b[95:64] +/// result[95:64] := __a[127:96] +/// result[127:96] := __b[127:96] +/// result[159:128] := __a[223:192] +/// result[191:160] := __b[223:192] +/// result[223:192] := __a[255:224] +/// result[255:224] := __b[255:224] +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPUNPCKHDQ instruction. +/// +/// \param __a +/// A 256-bit vector of [8 x i32] used as the source for the even-numbered +/// elements of the result. +/// \param __b +/// A 256-bit vector of [8 x i32] used as the source for the odd-numbered +/// elements of the result. +/// \returns A 256-bit vector of [8 x i32] containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_unpackhi_epi32(__m256i __a, __m256i __b) { return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 2, 8+2, 3, 8+3, 6, 8+6, 7, 8+7); } +/// Unpacks and interleaves 64-bit integers from parts of the 256-bit vectors +/// of [4 x i64] in \a __a and \a __b to return the resulting 256-bit vector +/// of [4 x i64]. Specifically, uses the upper 64 bits of each 128-bit half +/// of \a __a and \a __b as input; other bits in these parameters are +/// ignored. +/// +/// \code{.operation} +/// result[63:0] := __a[127:64] +/// result[127:64] := __b[127:64] +/// result[191:128] := __a[255:192] +/// result[255:192] := __b[255:192] +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPUNPCKHQDQ instruction. +/// +/// \param __a +/// A 256-bit vector of [4 x i64] used as the source for the even-numbered +/// elements of the result. +/// \param __b +/// A 256-bit vector of [4 x i64] used as the source for the odd-numbered +/// elements of the result. +/// \returns A 256-bit vector of [4 x i64] containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_unpackhi_epi64(__m256i __a, __m256i __b) { return (__m256i)__builtin_shufflevector((__v4di)__a, (__v4di)__b, 1, 4+1, 3, 4+3); } +/// Unpacks and interleaves 8-bit integers from parts of the 256-bit integer +/// vectors in \a __a and \a __b to form the 256-bit result. Specifically, +/// uses the lower 64 bits of each 128-bit half of \a __a and \a __b as +/// input; other bits in these parameters are ignored. +/// +/// \code{.operation} +/// result[7:0] := __a[7:0] +/// result[15:8] := __b[7:0] +/// result[23:16] := __a[15:8] +/// result[31:24] := __b[15:8] +/// . . . +/// result[127:120] := __b[63:56] +/// result[135:128] := __a[135:128] +/// . . . +/// result[255:248] := __b[191:184] +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPUNPCKLBW instruction. +/// +/// \param __a +/// A 256-bit integer vector used as the source for the even-numbered bytes +/// of the result. +/// \param __b +/// A 256-bit integer vector used as the source for the odd-numbered bytes +/// of the result. +/// \returns A 256-bit integer vector containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_unpacklo_epi8(__m256i __a, __m256i __b) { return (__m256i)__builtin_shufflevector((__v32qi)__a, (__v32qi)__b, 0, 32+0, 1, 32+1, 2, 32+2, 3, 32+3, 4, 32+4, 5, 32+5, 6, 32+6, 7, 32+7, 16, 32+16, 17, 32+17, 18, 32+18, 19, 32+19, 20, 32+20, 21, 32+21, 22, 32+22, 23, 32+23); } +/// Unpacks and interleaves 16-bit integers from parts of the 256-bit vectors +/// of [16 x i16] in \a __a and \a __b to return the resulting 256-bit +/// vector of [16 x i16]. Specifically, uses the lower 64 bits of each +/// 128-bit half of \a __a and \a __b as input; other bits in these +/// parameters are ignored. +/// +/// \code{.operation} +/// result[15:0] := __a[15:0] +/// result[31:16] := __b[15:0] +/// result[47:32] := __a[31:16] +/// result[63:48] := __b[31:16] +/// . . . +/// result[127:112] := __b[63:48] +/// result[143:128] := __a[143:128] +/// . . . +/// result[255:239] := __b[191:176] +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPUNPCKLWD instruction. +/// +/// \param __a +/// A 256-bit vector of [16 x i16] used as the source for the even-numbered +/// elements of the result. +/// \param __b +/// A 256-bit vector of [16 x i16] used as the source for the odd-numbered +/// elements of the result. +/// \returns A 256-bit vector of [16 x i16] containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_unpacklo_epi16(__m256i __a, __m256i __b) { return (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)__b, 0, 16+0, 1, 16+1, 2, 16+2, 3, 16+3, 8, 16+8, 9, 16+9, 10, 16+10, 11, 16+11); } +/// Unpacks and interleaves 32-bit integers from parts of the 256-bit vectors +/// of [8 x i32] in \a __a and \a __b to return the resulting 256-bit vector +/// of [8 x i32]. Specifically, uses the lower 64 bits of each 128-bit half +/// of \a __a and \a __b as input; other bits in these parameters are +/// ignored. +/// +/// \code{.operation} +/// result[31:0] := __a[31:0] +/// result[63:32] := __b[31:0] +/// result[95:64] := __a[63:32] +/// result[127:96] := __b[63:32] +/// result[159:128] := __a[159:128] +/// result[191:160] := __b[159:128] +/// result[223:192] := __a[191:160] +/// result[255:224] := __b[191:190] +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPUNPCKLDQ instruction. +/// +/// \param __a +/// A 256-bit vector of [8 x i32] used as the source for the even-numbered +/// elements of the result. +/// \param __b +/// A 256-bit vector of [8 x i32] used as the source for the odd-numbered +/// elements of the result. +/// \returns A 256-bit vector of [8 x i32] containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_unpacklo_epi32(__m256i __a, __m256i __b) { return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 0, 8+0, 1, 8+1, 4, 8+4, 5, 8+5); } +/// Unpacks and interleaves 64-bit integers from parts of the 256-bit vectors +/// of [4 x i64] in \a __a and \a __b to return the resulting 256-bit vector +/// of [4 x i64]. Specifically, uses the lower 64 bits of each 128-bit half +/// of \a __a and \a __b as input; other bits in these parameters are +/// ignored. +/// +/// \code{.operation} +/// result[63:0] := __a[63:0] +/// result[127:64] := __b[63:0] +/// result[191:128] := __a[191:128] +/// result[255:192] := __b[191:128] +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPUNPCKLQDQ instruction. +/// +/// \param __a +/// A 256-bit vector of [4 x i64] used as the source for the even-numbered +/// elements of the result. +/// \param __b +/// A 256-bit vector of [4 x i64] used as the source for the odd-numbered +/// elements of the result. +/// \returns A 256-bit vector of [4 x i64] containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_unpacklo_epi64(__m256i __a, __m256i __b) { @@ -1605,30 +2145,84 @@ _mm256_stream_load_si256(__m256i const *__V) return (__m256i)__builtin_nontemporal_load((const __v4di_aligned *)__V); } +/// Broadcasts the 32-bit floating-point value from the low element of the +/// 128-bit vector of [4 x float] in \a __X to all elements of the result's +/// 128-bit vector of [4 x float]. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VBROADCASTSS instruction. +/// +/// \param __X +/// A 128-bit vector of [4 x float] whose low element will be broadcast. +/// \returns A 128-bit vector of [4 x float] containing the result. static __inline__ __m128 __DEFAULT_FN_ATTRS128 _mm_broadcastss_ps(__m128 __X) { return (__m128)__builtin_shufflevector((__v4sf)__X, (__v4sf)__X, 0, 0, 0, 0); } +/// Broadcasts the 64-bit floating-point value from the low element of the +/// 128-bit vector of [2 x double] in \a __a to both elements of the +/// result's 128-bit vector of [2 x double]. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c MOVDDUP instruction. +/// +/// \param __a +/// A 128-bit vector of [2 x double] whose low element will be broadcast. +/// \returns A 128-bit vector of [2 x double] containing the result. static __inline__ __m128d __DEFAULT_FN_ATTRS128 _mm_broadcastsd_pd(__m128d __a) { return __builtin_shufflevector((__v2df)__a, (__v2df)__a, 0, 0); } +/// Broadcasts the 32-bit floating-point value from the low element of the +/// 128-bit vector of [4 x float] in \a __X to all elements of the +/// result's 256-bit vector of [8 x float]. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VBROADCASTSS instruction. +/// +/// \param __X +/// A 128-bit vector of [4 x float] whose low element will be broadcast. +/// \returns A 256-bit vector of [8 x float] containing the result. static __inline__ __m256 __DEFAULT_FN_ATTRS256 _mm256_broadcastss_ps(__m128 __X) { return (__m256)__builtin_shufflevector((__v4sf)__X, (__v4sf)__X, 0, 0, 0, 0, 0, 0, 0, 0); } +/// Broadcasts the 64-bit floating-point value from the low element of the +/// 128-bit vector of [2 x double] in \a __X to all elements of the +/// result's 256-bit vector of [4 x double]. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VBROADCASTSD instruction. +/// +/// \param __X +/// A 128-bit vector of [2 x double] whose low element will be broadcast. +/// \returns A 256-bit vector of [4 x double] containing the result. static __inline__ __m256d __DEFAULT_FN_ATTRS256 _mm256_broadcastsd_pd(__m128d __X) { return (__m256d)__builtin_shufflevector((__v2df)__X, (__v2df)__X, 0, 0, 0, 0); } +/// Broadcasts the 128-bit integer data from \a __X to both the lower and +/// upper halves of the 256-bit result. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VBROADCASTI128 instruction. +/// +/// \param __X +/// A 128-bit integer vector to be broadcast. +/// \returns A 256-bit integer vector containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_broadcastsi128_si256(__m128i __X) { @@ -1637,87 +2231,410 @@ _mm256_broadcastsi128_si256(__m128i __X) #define _mm_broadcastsi128_si256(X) _mm256_broadcastsi128_si256(X) +/// Merges 32-bit integer elements from either of the two 128-bit vectors of +/// [4 x i32] in \a V1 or \a V2 to the result's 128-bit vector of [4 x i32], +/// as specified by the immediate integer operand \a M. +/// +/// \code{.operation} +/// FOR i := 0 TO 3 +/// j := i*32 +/// IF M[i] == 0 +/// result[31+j:j] := V1[31+j:j] +/// ELSE +/// result[31+j:j] := V2[32+j:j] +/// FI +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// \code +/// __m128i _mm_blend_epi32(__m128i V1, __m128i V2, const int M); +/// \endcode +/// +/// This intrinsic corresponds to the \c VPBLENDDD instruction. +/// +/// \param V1 +/// A 128-bit vector of [4 x i32] containing source values. +/// \param V2 +/// A 128-bit vector of [4 x i32] containing source values. +/// \param M +/// An immediate 8-bit integer operand, with bits [3:0] specifying the +/// source for each element of the result. The position of the mask bit +/// corresponds to the index of a copied value. When a mask bit is 0, the +/// element is copied from \a V1; otherwise, it is copied from \a V2. +/// \returns A 128-bit vector of [4 x i32] containing the result. #define _mm_blend_epi32(V1, V2, M) \ ((__m128i)__builtin_ia32_pblendd128((__v4si)(__m128i)(V1), \ (__v4si)(__m128i)(V2), (int)(M))) +/// Merges 32-bit integer elements from either of the two 256-bit vectors of +/// [8 x i32] in \a V1 or \a V2 to return a 256-bit vector of [8 x i32], +/// as specified by the immediate integer operand \a M. +/// +/// \code{.operation} +/// FOR i := 0 TO 7 +/// j := i*32 +/// IF M[i] == 0 +/// result[31+j:j] := V1[31+j:j] +/// ELSE +/// result[31+j:j] := V2[32+j:j] +/// FI +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// \code +/// __m256i _mm256_blend_epi32(__m256i V1, __m256i V2, const int M); +/// \endcode +/// +/// This intrinsic corresponds to the \c VPBLENDDD instruction. +/// +/// \param V1 +/// A 256-bit vector of [8 x i32] containing source values. +/// \param V2 +/// A 256-bit vector of [8 x i32] containing source values. +/// \param M +/// An immediate 8-bit integer operand, with bits [7:0] specifying the +/// source for each element of the result. The position of the mask bit +/// corresponds to the index of a copied value. When a mask bit is 0, the +/// element is copied from \a V1; otherwise, it is is copied from \a V2. +/// \returns A 256-bit vector of [8 x i32] containing the result. #define _mm256_blend_epi32(V1, V2, M) \ ((__m256i)__builtin_ia32_pblendd256((__v8si)(__m256i)(V1), \ (__v8si)(__m256i)(V2), (int)(M))) +/// Broadcasts the low byte from the 128-bit integer vector in \a __X to all +/// bytes of the 256-bit result. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPBROADCASTB instruction. +/// +/// \param __X +/// A 128-bit integer vector whose low byte will be broadcast. +/// \returns A 256-bit integer vector containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_broadcastb_epi8(__m128i __X) { return (__m256i)__builtin_shufflevector((__v16qi)__X, (__v16qi)__X, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); } +/// Broadcasts the low element from the 128-bit vector of [8 x i16] in \a __X +/// to all elements of the result's 256-bit vector of [16 x i16]. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPBROADCASTW instruction. +/// +/// \param __X +/// A 128-bit vector of [8 x i16] whose low element will be broadcast. +/// \returns A 256-bit vector of [16 x i16] containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_broadcastw_epi16(__m128i __X) { return (__m256i)__builtin_shufflevector((__v8hi)__X, (__v8hi)__X, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); } +/// Broadcasts the low element from the 128-bit vector of [4 x i32] in \a __X +/// to all elements of the result's 256-bit vector of [8 x i32]. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPBROADCASTD instruction. +/// +/// \param __X +/// A 128-bit vector of [4 x i32] whose low element will be broadcast. +/// \returns A 256-bit vector of [8 x i32] containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_broadcastd_epi32(__m128i __X) { return (__m256i)__builtin_shufflevector((__v4si)__X, (__v4si)__X, 0, 0, 0, 0, 0, 0, 0, 0); } +/// Broadcasts the low element from the 128-bit vector of [2 x i64] in \a __X +/// to all elements of the result's 256-bit vector of [4 x i64]. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPBROADCASTQ instruction. +/// +/// \param __X +/// A 128-bit vector of [2 x i64] whose low element will be broadcast. +/// \returns A 256-bit vector of [4 x i64] containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_broadcastq_epi64(__m128i __X) { return (__m256i)__builtin_shufflevector((__v2di)__X, (__v2di)__X, 0, 0, 0, 0); } +/// Broadcasts the low byte from the 128-bit integer vector in \a __X to all +/// bytes of the 128-bit result. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPBROADCASTB instruction. +/// +/// \param __X +/// A 128-bit integer vector whose low byte will be broadcast. +/// \returns A 128-bit integer vector containing the result. static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_broadcastb_epi8(__m128i __X) { return (__m128i)__builtin_shufflevector((__v16qi)__X, (__v16qi)__X, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); } +/// Broadcasts the low element from the 128-bit vector of [8 x i16] in +/// \a __X to all elements of the result's 128-bit vector of [8 x i16]. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPBROADCASTW instruction. +/// +/// \param __X +/// A 128-bit vector of [8 x i16] whose low element will be broadcast. +/// \returns A 128-bit vector of [8 x i16] containing the result. static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_broadcastw_epi16(__m128i __X) { return (__m128i)__builtin_shufflevector((__v8hi)__X, (__v8hi)__X, 0, 0, 0, 0, 0, 0, 0, 0); } - +/// Broadcasts the low element from the 128-bit vector of [4 x i32] in \a __X +/// to all elements of the result's vector of [4 x i32]. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPBROADCASTD instruction. +/// +/// \param __X +/// A 128-bit vector of [4 x i32] whose low element will be broadcast. +/// \returns A 128-bit vector of [4 x i32] containing the result. static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_broadcastd_epi32(__m128i __X) { return (__m128i)__builtin_shufflevector((__v4si)__X, (__v4si)__X, 0, 0, 0, 0); } +/// Broadcasts the low element from the 128-bit vector of [2 x i64] in \a __X +/// to both elements of the result's 128-bit vector of [2 x i64]. +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPBROADCASTQ instruction. +/// +/// \param __X +/// A 128-bit vector of [2 x i64] whose low element will be broadcast. +/// \returns A 128-bit vector of [2 x i64] containing the result. static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_broadcastq_epi64(__m128i __X) { return (__m128i)__builtin_shufflevector((__v2di)__X, (__v2di)__X, 0, 0); } +/// Sets the result's 256-bit vector of [8 x i32] to copies of elements of the +/// 256-bit vector of [8 x i32] in \a __a as specified by indexes in the +/// elements of the 256-bit vector of [8 x i32] in \a __b. +/// +/// \code{.operation} +/// FOR i := 0 TO 7 +/// j := i*32 +/// k := __b[j+2:j] * 32 +/// result[j+31:j] := __a[k+31:k] +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPERMD instruction. +/// +/// \param __a +/// A 256-bit vector of [8 x i32] containing the source values. +/// \param __b +/// A 256-bit vector of [8 x i32] containing indexes of values to use from +/// \a __a. +/// \returns A 256-bit vector of [8 x i32] containing the result. static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_permutevar8x32_epi32(__m256i __a, __m256i __b) { return (__m256i)__builtin_ia32_permvarsi256((__v8si)__a, (__v8si)__b); } +/// Sets the result's 256-bit vector of [4 x double] to copies of elements of +/// the 256-bit vector of [4 x double] in \a V as specified by the +/// immediate value \a M. +/// +/// \code{.operation} +/// FOR i := 0 TO 3 +/// j := i*64 +/// k := (M >> i*2)[1:0] * 64 +/// result[j+63:j] := V[k+63:k] +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// \code +/// __m256d _mm256_permute4x64_pd(__m256d V, const int M); +/// \endcode +/// +/// This intrinsic corresponds to the \c VPERMPD instruction. +/// +/// \param V +/// A 256-bit vector of [4 x double] containing the source values. +/// \param M +/// An immediate 8-bit value specifying which elements to copy from \a V. +/// \a M[1:0] specifies the index in \a a for element 0 of the result, +/// \a M[3:2] specifies the index for element 1, and so forth. +/// \returns A 256-bit vector of [4 x double] containing the result. #define _mm256_permute4x64_pd(V, M) \ ((__m256d)__builtin_ia32_permdf256((__v4df)(__m256d)(V), (int)(M))) +/// Sets the result's 256-bit vector of [8 x float] to copies of elements of +/// the 256-bit vector of [8 x float] in \a __a as specified by indexes in +/// the elements of the 256-bit vector of [8 x i32] in \a __b. +/// +/// \code{.operation} +/// FOR i := 0 TO 7 +/// j := i*32 +/// k := __b[j+2:j] * 32 +/// result[j+31:j] := __a[k+31:k] +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// This intrinsic corresponds to the \c VPERMPS instruction. +/// +/// \param __a +/// A 256-bit vector of [8 x float] containing the source values. +/// \param __b +/// A 256-bit vector of [8 x i32] containing indexes of values to use from +/// \a __a. +/// \returns A 256-bit vector of [8 x float] containing the result. static __inline__ __m256 __DEFAULT_FN_ATTRS256 _mm256_permutevar8x32_ps(__m256 __a, __m256i __b) { return (__m256)__builtin_ia32_permvarsf256((__v8sf)__a, (__v8si)__b); } +/// Sets the result's 256-bit vector of [4 x i64] result to copies of elements +/// of the 256-bit vector of [4 x i64] in \a V as specified by the +/// immediate value \a M. +/// +/// \code{.operation} +/// FOR i := 0 TO 3 +/// j := i*64 +/// k := (M >> i*2)[1:0] * 64 +/// result[j+63:j] := V[k+63:k] +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// \code +/// __m256i _mm256_permute4x64_epi64(__m256i V, const int M); +/// \endcode +/// +/// This intrinsic corresponds to the \c VPERMQ instruction. +/// +/// \param V +/// A 256-bit vector of [4 x i64] containing the source values. +/// \param M +/// An immediate 8-bit value specifying which elements to copy from \a V. +/// \a M[1:0] specifies the index in \a a for element 0 of the result, +/// \a M[3:2] specifies the index for element 1, and so forth. +/// \returns A 256-bit vector of [4 x i64] containing the result. #define _mm256_permute4x64_epi64(V, M) \ ((__m256i)__builtin_ia32_permdi256((__v4di)(__m256i)(V), (int)(M))) +/// Sets each half of the 256-bit result either to zero or to one of the +/// four possible 128-bit halves of the 256-bit vectors \a V1 and \a V2, +/// as specified by the immediate value \a M. +/// +/// \code{.operation} +/// FOR i := 0 TO 1 +/// j := i*128 +/// k := M >> (i*4) +/// IF k[3] == 0 +/// CASE (k[1:0]) OF +/// 0: result[127+j:j] := V1[127:0] +/// 1: result[127+j:j] := V1[255:128] +/// 2: result[127+j:j] := V2[127:0] +/// 3: result[127+j:j] := V2[255:128] +/// ESAC +/// ELSE +/// result[127+j:j] := 0 +/// FI +/// ENDFOR +/// \endcode +/// +/// \headerfile <immintrin.h> +/// +/// \code +/// __m256i _mm256_permute2x128_si256(__m256i V1, __m256i V2, const int M); +/// \endcode +/// +/// This intrinsic corresponds to the \c VPERM2I128 instruction. +/// +/// \param V1 +/// A 256-bit integer vector containing source values. +/// \param V2 +/// A 256-bit integer vector containing source values. +/// \param M +/// An immediate value specifying how to form the result. Bits [3:0] +/// control the lower half of the result, bits [7:4] control the upper half. +/// Within each 4-bit control value, if bit 3 is 1, the result is zero, +/// otherwise bits [1:0] determine the source as follows. \n +/// 0: the lower half of \a V1 \n +/// 1: the upper half of \a V1 \n +/// 2: the lower half of \a V2 \n +/// 3: the upper half of \a V2 +/// \returns A 256-bit integer vector containing the result. #define _mm256_permute2x128_si256(V1, V2, M) \ ((__m256i)__builtin_ia32_permti256((__m256i)(V1), (__m256i)(V2), (int)(M))) +/// Extracts half of the 256-bit vector \a V to the 128-bit result. If bit 0 +/// of the immediate \a M is zero, extracts the lower half of the result; +/// otherwise, extracts the upper half. +/// +/// \headerfile <immintrin.h> +/// +/// \code +/// __m128i _mm256_extracti128_si256(__m256i V, const int M); +/// \endcode +/// +/// This intrinsic corresponds to the \c VEXTRACTI128 instruction. +/// +/// \param V +/// A 256-bit integer vector containing the source values. +/// \param M +/// An immediate value specifying which half of \a V to extract. +/// \returns A 128-bit integer vector containing the result. #define _mm256_extracti128_si256(V, M) \ ((__m128i)__builtin_ia32_extract128i256((__v4di)(__m256i)(V), (int)(M))) +/// Copies the 256-bit vector \a V1 to the result, then overwrites half of the +/// result with the 128-bit vector \a V2. If bit 0 of the immediate \a M +/// is zero, overwrites the lower half of the result; otherwise, +/// overwrites the upper half. +/// +/// \headerfile <immintrin.h> +/// +/// \code +/// __m256i _mm256_inserti128_si256(__m256i V1, __m128i V2, const int M); +/// \endcode +/// +/// This intrinsic corresponds to the \c VINSERTI128 instruction. +/// +/// \param V1 +/// A 256-bit integer vector containing a source value. +/// \param V2 +/// A 128-bit integer vector containing a source value. +/// \param M +/// An immediate value specifying where to put \a V2 in the result. +/// \returns A 256-bit integer vector containing the result. #define _mm256_inserti128_si256(V1, V2, M) \ ((__m256i)__builtin_ia32_insert128i256((__v4di)(__m256i)(V1), \ (__v2di)(__m128i)(V2), (int)(M))) _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits