#endif /* __OPTIMIZE__ */
#endif /* __x86_64__ */
+/* Intrinsics vcvttsh2si, vcvttsh2us. */
+extern __inline int
+__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
+_mm_cvttsh_i32 (__m128h __A)
+{
+ return (int)
+ __builtin_ia32_vcvttsh2si32_round (__A, _MM_FROUND_CUR_DIRECTION);
+}
+
+extern __inline unsigned
+__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
+_mm_cvttsh_u32 (__m128h __A)
+{
+ return (int)
+ __builtin_ia32_vcvttsh2usi32_round (__A, _MM_FROUND_CUR_DIRECTION);
+}
+
+#ifdef __OPTIMIZE__
+extern __inline int
+__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
+_mm_cvtt_roundsh_i32 (__m128h __A, const int __R)
+{
+ return (int) __builtin_ia32_vcvttsh2si32_round (__A, __R);
+}
+
+extern __inline unsigned
+__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
+_mm_cvtt_roundsh_u32 (__m128h __A, const int __R)
+{
+ return (int) __builtin_ia32_vcvttsh2usi32_round (__A, __R);
+}
+
+#else
+#define _mm_cvtt_roundsh_i32(A, B) \
+ ((int)__builtin_ia32_vcvttsh2si32_round ((A), (B)))
+#define _mm_cvtt_roundsh_u32(A, B) \
+ ((int)__builtin_ia32_vcvttsh2usi32_round ((A), (B)))
+
+#endif /* __OPTIMIZE__ */
+
+#ifdef __x86_64__
+extern __inline long long
+__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
+_mm_cvttsh_i64 (__m128h __A)
+{
+ return (long long)
+ __builtin_ia32_vcvttsh2si64_round (__A, _MM_FROUND_CUR_DIRECTION);
+}
+
+extern __inline unsigned long long
+__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
+_mm_cvttsh_u64 (__m128h __A)
+{
+ return (long long)
+ __builtin_ia32_vcvttsh2usi64_round (__A, _MM_FROUND_CUR_DIRECTION);
+}
+
+#ifdef __OPTIMIZE__
+extern __inline long long
+__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
+_mm_cvtt_roundsh_i64 (__m128h __A, const int __R)
+{
+ return (long long) __builtin_ia32_vcvttsh2si64_round (__A, __R);
+}
+
+extern __inline unsigned long long
+__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
+_mm_cvtt_roundsh_u64 (__m128h __A, const int __R)
+{
+ return (long long) __builtin_ia32_vcvttsh2usi64_round (__A, __R);
+}
+
+#else
+#define _mm_cvtt_roundsh_i64(A, B) \
+ ((long long)__builtin_ia32_vcvttsh2si64_round ((A), (B)))
+#define _mm_cvtt_roundsh_u64(A, B) \
+ ((long long)__builtin_ia32_vcvttsh2usi64_round ((A), (B)))
+
+#endif /* __OPTIMIZE__ */
+#endif /* __x86_64__ */
+
/* Intrinsics vcvtsi2sh, vcvtusi2sh. */
extern __inline __m128h
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
BDESC (OPTION_MASK_ISA_64BIT, OPTION_MASK_ISA2_AVX512FP16, CODE_FOR_avx512fp16_vcvtsh2siq_round, "__builtin_ia32_vcvtsh2si64_round", IX86_BUILTIN_VCVTSH2SI64_ROUND, UNKNOWN, (int) INT64_FTYPE_V8HF_INT)
BDESC (0, OPTION_MASK_ISA2_AVX512FP16, CODE_FOR_avx512fp16_vcvtsh2usi_round, "__builtin_ia32_vcvtsh2usi32_round", IX86_BUILTIN_VCVTSH2USI32_ROUND, UNKNOWN, (int) UINT_FTYPE_V8HF_INT)
BDESC (OPTION_MASK_ISA_64BIT, OPTION_MASK_ISA2_AVX512FP16, CODE_FOR_avx512fp16_vcvtsh2usiq_round, "__builtin_ia32_vcvtsh2usi64_round", IX86_BUILTIN_VCVTSH2USI64_ROUND, UNKNOWN, (int) UINT64_FTYPE_V8HF_INT)
+BDESC (0, OPTION_MASK_ISA2_AVX512FP16, CODE_FOR_avx512fp16_fix_truncsi2_round, "__builtin_ia32_vcvttsh2si32_round", IX86_BUILTIN_VCVTTSH2SI32_ROUND, UNKNOWN, (int) INT_FTYPE_V8HF_INT)
+BDESC (OPTION_MASK_ISA_64BIT, OPTION_MASK_ISA2_AVX512FP16, CODE_FOR_avx512fp16_fix_truncdi2_round, "__builtin_ia32_vcvttsh2si64_round", IX86_BUILTIN_VCVTTSH2SI64_ROUND, UNKNOWN, (int) INT64_FTYPE_V8HF_INT)
+BDESC (0, OPTION_MASK_ISA2_AVX512FP16, CODE_FOR_avx512fp16_fixuns_truncsi2_round, "__builtin_ia32_vcvttsh2usi32_round", IX86_BUILTIN_VCVTTSH2USI32_ROUND, UNKNOWN, (int) UINT_FTYPE_V8HF_INT)
+BDESC (OPTION_MASK_ISA_64BIT, OPTION_MASK_ISA2_AVX512FP16, CODE_FOR_avx512fp16_fixuns_truncdi2_round, "__builtin_ia32_vcvttsh2usi64_round", IX86_BUILTIN_VCVTTSH2USI64_ROUND, UNKNOWN, (int) UINT64_FTYPE_V8HF_INT)
BDESC (0, OPTION_MASK_ISA2_AVX512FP16, CODE_FOR_avx512fp16_vcvtsi2sh_round, "__builtin_ia32_vcvtsi2sh32_round", IX86_BUILTIN_VCVTSI2SH32_ROUND, UNKNOWN, (int) V8HF_FTYPE_V8HF_INT_INT)
BDESC (OPTION_MASK_ISA_64BIT, OPTION_MASK_ISA2_AVX512FP16, CODE_FOR_avx512fp16_vcvtsi2shq_round, "__builtin_ia32_vcvtsi2sh64_round", IX86_BUILTIN_VCVTSI2SH64_ROUND, UNKNOWN, (int) V8HF_FTYPE_V8HF_INT64_INT)
BDESC (0, OPTION_MASK_ISA2_AVX512FP16, CODE_FOR_avx512fp16_vcvtusi2sh_round, "__builtin_ia32_vcvtusi2sh32_round", IX86_BUILTIN_VCVTUSI2SH32_ROUND, UNKNOWN, (int) V8HF_FTYPE_V8HF_UINT_INT)
(set_attr "prefix" "evex")
(set_attr "mode" "TI")])
+(define_insn "avx512fp16_fix<fixunssuffix>_trunc<mode>2<round_saeonly_name>"
+ [(set (match_operand:SWI48 0 "register_operand" "=r")
+ (any_fix:SWI48
+ (vec_select:HF
+ (match_operand:V8HF 1 "register_operand" "v")
+ (parallel [(const_int 0)]))))]
+ "TARGET_AVX512FP16"
+ "%vcvttsh2<fixsuffix>si\t{<round_saeonly_op2>%1, %0|%0, %k1<round_saeonly_op2>}"
+ [(set_attr "type" "sseicvt")
+ (set_attr "prefix" "evex")
+ (set_attr "mode" "<MODE>")])
+
+(define_insn "avx512fp16_fix<fixunssuffix>_trunc<mode>2_mem"
+ [(set (match_operand:SWI48 0 "register_operand" "=r")
+ (any_fix:SWI48
+ (match_operand:HF 1 "memory_operand" "vm")))]
+ "TARGET_AVX512FP16"
+ "%vcvttsh2<fixsuffix>si\t{%1, %0|%0, %1}"
+ [(set_attr "type" "sseicvt")
+ (set_attr "prefix" "evex")
+ (set_attr "mode" "<MODE>")])
+
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;
;; Parallel single-precision floating point conversion operations
#define __builtin_ia32_vcvtsh2si64_round(A, B) __builtin_ia32_vcvtsh2si64_round(A, 8)
#define __builtin_ia32_vcvtsh2usi32_round(A, B) __builtin_ia32_vcvtsh2usi32_round(A, 8)
#define __builtin_ia32_vcvtsh2usi64_round(A, B) __builtin_ia32_vcvtsh2usi64_round(A, 8)
+#define __builtin_ia32_vcvttsh2si32_round(A, B) __builtin_ia32_vcvttsh2si32_round(A, 8)
+#define __builtin_ia32_vcvttsh2si64_round(A, B) __builtin_ia32_vcvttsh2si64_round(A, 8)
+#define __builtin_ia32_vcvttsh2usi32_round(A, B) __builtin_ia32_vcvttsh2usi32_round(A, 8)
+#define __builtin_ia32_vcvttsh2usi64_round(A, B) __builtin_ia32_vcvttsh2usi64_round(A, 8)
#define __builtin_ia32_vcvtsi2sh32_round(A, B, C) __builtin_ia32_vcvtsi2sh32_round(A, B, 8)
#define __builtin_ia32_vcvtsi2sh64_round(A, B, C) __builtin_ia32_vcvtsi2sh64_round(A, B, 8)
#define __builtin_ia32_vcvtusi2sh32_round(A, B, C) __builtin_ia32_vcvtusi2sh32_round(A, B, 8)
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-mavx512fp16 -O2" } */
+/* { dg-final { scan-assembler-times "vcvttsh2si\[ \\t\]+\[^\{\n\]*%xmm\[0-9\]+\[^\n\r]*%eax" 1 } } */
+/* { dg-final { scan-assembler-times "vcvttsh2si\[ \\t\]+\{sae\}\[^\{\n\]*%xmm\[0-9\]+\[^\n\r]*%eax" 1 } } */
+
+#include <immintrin.h>
+
+volatile __m128h x1;
+volatile int res1;
+
+void extern
+avx512f_test (void)
+{
+ res1 = _mm_cvttsh_i32 (x1);
+ res1 = _mm_cvtt_roundsh_i32 (x1, 8);
+}
--- /dev/null
+/* { dg-do run { target avx512fp16 } } */
+/* { dg-options "-O2 -mavx512fp16 -mavx512dq" } */
+
+
+#define AVX512FP16
+#include "avx512fp16-helper.h"
+
+#define N_ELEMS 2
+
+void NOINLINE
+emulate_cvtph2_d(V512 * dest, V512 op1,
+ __mmask32 k, int zero_mask)
+{
+ V512 v1, v2, v3, v4, v5, v6, v7, v8;
+ int i;
+ __mmask16 m1, m2;
+
+ m1 = k & 0xffff;
+
+ unpack_ph_2twops(op1, &v1, &v2);
+
+ for (i = 0; i < 16; i++) {
+ if (((1 << i) & m1) == 0) {
+ if (zero_mask) {
+ v5.u32[i] = 0;
+ }
+ else {
+ v5.u32[i] = dest->u32[i];
+ }
+ }
+ else {
+ v5.u32[i] = v1.f32[i];
+
+ }
+ }
+ *dest = v5;
+}
+
+void
+test_512 (void)
+{
+ V512 res;
+ V512 exp;
+
+ init_src();
+ emulate_cvtph2_d(&exp, src1, NET_MASK, 0);
+ res.i32[0] = _mm_cvtt_roundsh_i32(src1.xmmh[0], _ROUND_NINT);
+ check_results(&res, &exp, N_ELEMS, "_mm_cvtt_roundsh_i32");
+
+ if (n_errs != 0) {
+ abort ();
+ }
+}
+
--- /dev/null
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-mavx512fp16 -O2" } */
+/* { dg-final { scan-assembler-times "vcvttsh2si\[ \\t\]+\[^\{\n\]*%xmm\[0-9\]+\[^\n\r]*%rax" 1 } } */
+/* { dg-final { scan-assembler-times "vcvttsh2si\[ \\t\]+\{sae\}\[^\{\n\]*%xmm\[0-9\]+\[^\n\r]*%rax" 1 } } */
+
+#include <immintrin.h>
+
+volatile __m128h x1;
+volatile long long res2;
+
+void extern
+avx512f_test (void)
+{
+ res2 = _mm_cvttsh_i64 (x1);
+ res2 = _mm_cvtt_roundsh_i64 (x1, 8);
+}
--- /dev/null
+/* { dg-do run { target { { ! ia32 } && avx512fp16 } } } */
+/* { dg-options "-O2 -mavx512fp16 -mavx512dq" } */
+
+#define AVX512FP16
+#include "avx512fp16-helper.h"
+
+#define N_ELEMS 4
+
+void NOINLINE
+emulate_cvtph2_q(V512 * dest, V512 op1,
+ __mmask32 k, int zero_mask)
+{
+ V512 v1, v2, v3, v4, v5, v6, v7, v8;
+ int i;
+ __mmask16 m1, m2;
+
+ m1 = k & 0xffff;
+
+ unpack_ph_2twops(op1, &v1, &v2);
+
+ for (i = 0; i < 8; i++) {
+ if (((1 << i) & m1) == 0) {
+ if (zero_mask) {
+ v5.u64[i] = 0;
+ }
+ else {
+ v5.u64[i] = dest->u64[i];
+ }
+ }
+ else {
+ v5.u64[i] = v1.f32[i];
+ }
+ }
+ *dest = v5;
+}
+
+void
+test_512 (void)
+{
+ V512 res;
+ V512 exp;
+
+ init_src();
+ emulate_cvtph2_q(&exp, src1, NET_MASK, 0);
+ res.s64[0] = _mm_cvtt_roundsh_i64(src1.xmmh[0], _ROUND_NINT);
+ check_results(&res, &exp, N_ELEMS, "_mm_cvtt_roundsh_i64");
+
+ if (n_errs != 0) {
+ abort ();
+ }
+}
+
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-mavx512fp16 -O2" } */
+/* { dg-final { scan-assembler-times "vcvttsh2usi\[ \\t\]+\[^\{\n\]*%xmm\[0-9\]+\[^\n\r]*%eax" 1 } } */
+/* { dg-final { scan-assembler-times "vcvttsh2usi\[ \\t\]+\{sae\}\[^\{\n\]*%xmm\[0-9\]+\[^\n\r]*%eax" 1 } } */
+
+#include <immintrin.h>
+
+volatile __m128h x1;
+volatile unsigned int res1;
+
+void extern
+avx512f_test (void)
+{
+ res1 = _mm_cvttsh_u32 (x1);
+ res1 = _mm_cvtt_roundsh_u32 (x1, 8);
+}
--- /dev/null
+/* { dg-do run { target avx512fp16 } } */
+/* { dg-options "-O2 -mavx512fp16 -mavx512dq" } */
+
+
+#define AVX512FP16
+#include "avx512fp16-helper.h"
+
+#define N_ELEMS 2
+
+void NOINLINE
+emulate_cvtph2_d(V512 * dest, V512 op1,
+ __mmask32 k, int zero_mask)
+{
+ V512 v1, v2, v3, v4, v5, v6, v7, v8;
+ int i;
+ __mmask16 m1, m2;
+
+ m1 = k & 0xffff;
+
+ unpack_ph_2twops(op1, &v1, &v2);
+
+ for (i = 0; i < 16; i++) {
+ if (((1 << i) & m1) == 0) {
+ if (zero_mask) {
+ v5.u32[i] = 0;
+ }
+ else {
+ v5.u32[i] = dest->u32[i];
+ }
+ }
+ else {
+ v5.u32[i] = v1.f32[i];
+
+ }
+ }
+ *dest = v5;
+}
+
+void
+test_512 (void)
+{
+ V512 res;
+ V512 exp;
+
+ init_src();
+ emulate_cvtph2_d(&exp, src1, NET_MASK, 0);
+ res.u32[0] = _mm_cvtt_roundsh_i32(src1.xmmh[0], _ROUND_NINT);
+ check_results(&res, &exp, N_ELEMS, "_mm_cvtt_roundsh_u32");
+
+ if (n_errs != 0) {
+ abort ();
+ }
+}
+
--- /dev/null
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-mavx512fp16 -O2 " } */
+/* { dg-final { scan-assembler-times "vcvttsh2usi\[ \\t\]+\[^\{\n\]*%xmm\[0-9\]+\[^\n\r]*%rax" 1 } } */
+/* { dg-final { scan-assembler-times "vcvttsh2usi\[ \\t\]+\{sae\}\[^\{\n\]*%xmm\[0-9\]+\[^\n\r]*%rax" 1 } } */
+
+#include <immintrin.h>
+
+volatile __m128h x1;
+volatile unsigned long long res2;
+
+void extern
+avx512f_test (void)
+{
+ res2 = _mm_cvttsh_u64 (x1);
+ res2 = _mm_cvtt_roundsh_u64 (x1, 8);
+}
--- /dev/null
+/* { dg-do run { target { { ! ia32 } && avx512fp16 } } } */
+/* { dg-options "-O2 -mavx512fp16 -mavx512dq" } */
+
+
+#define AVX512FP16
+#include "avx512fp16-helper.h"
+
+#define N_ELEMS 4
+
+void NOINLINE
+emulate_cvtph2_q(V512 * dest, V512 op1,
+ __mmask32 k, int zero_mask)
+{
+ V512 v1, v2, v3, v4, v5, v6, v7, v8;
+ int i;
+ __mmask16 m1, m2;
+
+ m1 = k & 0xffff;
+
+ unpack_ph_2twops(op1, &v1, &v2);
+
+ for (i = 0; i < 8; i++) {
+ if (((1 << i) & m1) == 0) {
+ if (zero_mask) {
+ v5.u64[i] = 0;
+ }
+ else {
+ v5.u64[i] = dest->u64[i];
+ }
+ }
+ else {
+ v5.u64[i] = v1.f32[i];
+ }
+ }
+ *dest = v5;
+}
+
+void
+test_512 (void)
+{
+ V512 res;
+ V512 exp;
+
+ init_src();
+ emulate_cvtph2_q(&exp, src1, NET_MASK, 0);
+ res.u64[0] = _mm_cvtt_roundsh_i64(src1.xmmh[0], _ROUND_NINT);
+ check_results(&res, &exp, 4, "_mm_cvtt_roundsh_u64");
+
+ if (n_errs != 0) {
+ abort ();
+ }
+}
+
#define __builtin_ia32_vcvtsh2si64_round(A, B) __builtin_ia32_vcvtsh2si64_round(A, 8)
#define __builtin_ia32_vcvtsh2usi32_round(A, B) __builtin_ia32_vcvtsh2usi32_round(A, 8)
#define __builtin_ia32_vcvtsh2usi64_round(A, B) __builtin_ia32_vcvtsh2usi64_round(A, 8)
+#define __builtin_ia32_vcvttsh2si32_round(A, B) __builtin_ia32_vcvttsh2si32_round(A, 8)
+#define __builtin_ia32_vcvttsh2si64_round(A, B) __builtin_ia32_vcvttsh2si64_round(A, 8)
+#define __builtin_ia32_vcvttsh2usi32_round(A, B) __builtin_ia32_vcvttsh2usi32_round(A, 8)
+#define __builtin_ia32_vcvttsh2usi64_round(A, B) __builtin_ia32_vcvttsh2usi64_round(A, 8)
#define __builtin_ia32_vcvtsi2sh32_round(A, B, C) __builtin_ia32_vcvtsi2sh32_round(A, B, 8)
#define __builtin_ia32_vcvtsi2sh64_round(A, B, C) __builtin_ia32_vcvtsi2sh64_round(A, B, 8)
#define __builtin_ia32_vcvtusi2sh32_round(A, B, C) __builtin_ia32_vcvtusi2sh32_round(A, B, 8)
test_1 (_mm512_cvt_roundepu64_ph, __m128h, __m512i, 8)
test_1 (_mm_cvt_roundsh_i32, int, __m128h, 8)
test_1 (_mm_cvt_roundsh_u32, unsigned, __m128h, 8)
+test_1 (_mm_cvtt_roundsh_i32, int, __m128h, 8)
+test_1 (_mm_cvtt_roundsh_u32, unsigned, __m128h, 8)
#ifdef __x86_64__
test_1 (_mm_cvt_roundsh_i64, long long, __m128h, 8)
test_1 (_mm_cvt_roundsh_u64, unsigned long long, __m128h, 8)
+test_1 (_mm_cvtt_roundsh_i64, long long, __m128h, 8)
+test_1 (_mm_cvtt_roundsh_u64, unsigned long long, __m128h, 8)
test_2 (_mm_cvt_roundi64_sh, __m128h, __m128h, long long, 8)
test_2 (_mm_cvt_roundu64_sh, __m128h, __m128h, unsigned long long, 8)
#endif
test_1 (_mm512_cvt_roundepu64_ph, __m128h, __m512i, 8)
test_1 (_mm_cvt_roundsh_i32, int, __m128h, 8)
test_1 (_mm_cvt_roundsh_u32, unsigned, __m128h, 8)
+test_1 (_mm_cvtt_roundsh_i32, int, __m128h, 8)
+test_1 (_mm_cvtt_roundsh_u32, unsigned, __m128h, 8)
#ifdef __x86_64__
test_1 (_mm_cvt_roundsh_i64, long long, __m128h, 8)
test_1 (_mm_cvt_roundsh_u64, unsigned long long, __m128h, 8)
+test_1 (_mm_cvtt_roundsh_i64, long long, __m128h, 8)
+test_1 (_mm_cvtt_roundsh_u64, unsigned long long, __m128h, 8)
test_2 (_mm_cvt_roundi64_sh, __m128h, __m128h, long long, 8)
test_2 (_mm_cvt_roundu64_sh, __m128h, __m128h, unsigned long long, 8)
#endif
#define __builtin_ia32_vcvtsh2si64_round(A, B) __builtin_ia32_vcvtsh2si64_round(A, 8)
#define __builtin_ia32_vcvtsh2usi32_round(A, B) __builtin_ia32_vcvtsh2usi32_round(A, 8)
#define __builtin_ia32_vcvtsh2usi64_round(A, B) __builtin_ia32_vcvtsh2usi64_round(A, 8)
+#define __builtin_ia32_vcvttsh2si32_round(A, B) __builtin_ia32_vcvttsh2si32_round(A, 8)
+#define __builtin_ia32_vcvttsh2si64_round(A, B) __builtin_ia32_vcvttsh2si64_round(A, 8)
+#define __builtin_ia32_vcvttsh2usi32_round(A, B) __builtin_ia32_vcvttsh2usi32_round(A, 8)
+#define __builtin_ia32_vcvttsh2usi64_round(A, B) __builtin_ia32_vcvttsh2usi64_round(A, 8)
#define __builtin_ia32_vcvtsi2sh32_round(A, B, C) __builtin_ia32_vcvtsi2sh32_round(A, B, 8)
#define __builtin_ia32_vcvtsi2sh64_round(A, B, C) __builtin_ia32_vcvtsi2sh64_round(A, B, 8)
#define __builtin_ia32_vcvtusi2sh32_round(A, B, C) __builtin_ia32_vcvtusi2sh32_round(A, B, 8)
projects / thirdparty / gcc.git / commitdiff
