diff options
Diffstat (limited to 'SingleSource/UnitTests/Vector/AVX512F/fma_addsub.c')
| -rw-r--r-- | SingleSource/UnitTests/Vector/AVX512F/fma_addsub.c | 411 |
1 files changed, 411 insertions, 0 deletions
diff --git a/SingleSource/UnitTests/Vector/AVX512F/fma_addsub.c b/SingleSource/UnitTests/Vector/AVX512F/fma_addsub.c new file mode 100644 index 00000000..c848e8bf --- /dev/null +++ b/SingleSource/UnitTests/Vector/AVX512F/fma_addsub.c @@ -0,0 +1,411 @@ +/* + * Test addsub and subadd instructions. + * Here we check for _mm512_[mask|mask3]_[fmaddsub|fmsubadd]_[round] + * intrinsics. + */ +#include "m512_test_util.h" +#include <stdio.h> +#include <string.h> + +int verbose = 0; + +__m512 f1; +__m512 f2; +__m512 f3; +__m512 f4; +__m512 f5; + +__m512d d1; +__m512d d2; +__m512d d3; +__m512d d4; +__m512d d5; + +typedef enum { + FMA_132, + FMA_231, + FMA_213, +} Fma_order; + +volatile int vol = 0; /* To prevent optimizations */ + +void NOINLINE init() { + int i; + V512 *pf1 = (V512 *)&f1; + V512 *pf2 = (V512 *)&f2; + V512 *pf3 = (V512 *)&f3; + V512 *pd1 = (V512 *)&d1; + V512 *pd2 = (V512 *)&d2; + V512 *pd3 = (V512 *)&d3; + + for (i = 0; i < 16; i++) { + pf1->f32[i] = 17 + ((i & 1) ? 1 : -1) * i + vol; + pf2->f32[i] = -(100 + ((i & 3) == 3 ? 1 : -1) * i + vol); + pf3->f32[i] = 400 + ((i & 1) ? -1 : 1) * i + vol; + } + + for (i = 0; i < 8; i++) { + pd1->f64[i] = pf1->f32[i]; + pd2->f64[i] = -pf2->f32[i]; + pd3->f64[i] = -pf3->f32[i]; + } +} + +void NOINLINE check_equal32(void *vgot, void *vexpected, void *vexpected_orig, + int mask, char *banner) { + int i; + V512 *got = (V512 *)vgot; + V512 *expected = (V512 *)vexpected; + V512 *orig = (V512 *)vexpected_orig; + + for (i = 0; i < 16; i++) { + int ans = (mask & (1 << i)) ? expected->s32[i] : orig->s32[i]; + if (got->s32[i] != ans) { + printf("ERROR: %s failed -- 0x%0.8x != 0x%0.8x at element [%d]\n", + banner ? banner : "", got->s32[i], ans, i); + n_errs++; + break; + } + } +} + +void NOINLINE check_equal64(void *vgot, void *vexpected, void *vexpected_orig, + int mask, char *banner) { + int i; + V512 *got = (V512 *)vgot; + V512 *expected = (V512 *)vexpected; + V512 *orig = (V512 *)vexpected_orig; + + for (i = 0; i < 8; i++) { + __int64 ans = (mask & (1 << i)) ? expected->s64[i] : orig->s64[i]; + if (got->s64[i] != ans) { + printf("ERROR: %s failed -- %0.16" PRIx64 " != %0.16" PRIx64 + " at element [%d]\n", + banner ? banner : "", got->s64[i], ans, i); + n_errs++; + break; + } + } +} + +void NOINLINE emulate_fmaddsub_ps(void *presult, const void *p1, int mask, + const void *p2, const void *p3, + Fma_order order) { + int i; + V512 *result = (V512 *)presult; + V512 *v1 = (V512 *)p1; + V512 *v2 = (V512 *)p2; + V512 *v3 = (V512 *)p3; + for (i = 0; i < 16; i++) { + + if (((1 << i) & mask) == 0) { + result->u32[i] = v1->u32[i]; + continue; + } + + switch (order) { + case FMA_132: + result->f32[i] = + v1->f32[i] * v3->f32[i] + ((i % 2 == 0) ? -v2->f32[i] : v2->f32[i]); + break; + + case FMA_231: + result->f32[i] = + v2->f32[i] * v3->f32[i] + ((i % 2 == 0) ? -v1->f32[i] : v1->f32[i]); + break; + + case FMA_213: + result->f32[i] = + v2->f32[i] * v1->f32[i] + ((i % 2 == 0) ? -v3->f32[i] : v3->f32[i]); + break; + + default: + printf("ERROR -- bad fma order %d\n", (int)order); + n_errs++; + return; + } + } +} + +void NOINLINE emulate_fmsubadd_ps(void *presult, const void *p1, int mask, + const void *p2, const void *p3, + Fma_order order) { + int i; + V512 *result = (V512 *)presult; + V512 *v1 = (V512 *)p1; + V512 *v2 = (V512 *)p2; + V512 *v3 = (V512 *)p3; + + for (i = 0; i < 16; i++) { + + if (((1 << i) & mask) == 0) { + result->u32[i] = v1->u32[i]; + continue; + } + + switch (order) { + case FMA_132: + result->f32[i] = + v1->f32[i] * v3->f32[i] + ((i % 2 == 0) ? v2->f32[i] : -v2->f32[i]); + break; + + case FMA_231: + result->f32[i] = + v2->f32[i] * v3->f32[i] + ((i % 2 == 0) ? v1->f32[i] : -v1->f32[i]); + break; + + case FMA_213: + result->f32[i] = + v2->f32[i] * v1->f32[i] + ((i % 2 == 0) ? v3->f32[i] : -v3->f32[i]); + break; + + default: + printf("ERROR -- bad fma order %d\n", (int)order); + n_errs++; + return; + } + } +} + +void NOINLINE emulate_fmaddsub_pd(void *presult, const void *p1, int mask, + const void *p2, const void *p3, + Fma_order order) { + int i; + V512 *result = (V512 *)presult; + V512 *v1 = (V512 *)p1; + V512 *v2 = (V512 *)p2; + V512 *v3 = (V512 *)p3; + + for (i = 0; i < 8; i++) { + + if (((1 << i) & mask) == 0) { + result->u64[i] = v1->u64[i]; + continue; + } + + switch (order) { + case FMA_132: + result->f64[i] = + v1->f64[i] * v3->f64[i] + ((i % 2 == 0) ? -v2->f64[i] : v2->f64[i]); + break; + + case FMA_231: + result->f64[i] = + v2->f64[i] * v3->f64[i] + ((i % 2 == 0) ? -v1->f64[i] : v1->f64[i]); + break; + + case FMA_213: + result->f64[i] = + v2->f64[i] * v1->f64[i] + ((i % 2 == 0) ? -v3->f64[i] : v3->f64[i]); + break; + + default: + printf("ERROR -- bad fma order %d\n", (int)order); + n_errs++; + return; + } + } +} + +void NOINLINE emulate_fmsubadd_pd(void *presult, const void *p1, int mask, + const void *p2, const void *p3, + Fma_order order) { + int i; + V512 *result = (V512 *)presult; + V512 *v1 = (V512 *)p1; + V512 *v2 = (V512 *)p2; + V512 *v3 = (V512 *)p3; + + for (i = 0; i < 8; i++) { + + if (((1 << i) & mask) == 0) { + result->u64[i] = v1->u64[i]; + continue; + } + + switch (order) { + case FMA_132: + result->f64[i] = + v1->f64[i] * v3->f64[i] + ((i % 2 == 0) ? v2->f64[i] : -v2->f64[i]); + break; + + case FMA_231: + result->f64[i] = + v2->f64[i] * v3->f64[i] + ((i % 2 == 0) ? v1->f64[i] : -v1->f64[i]); + break; + + case FMA_213: + result->f64[i] = + v2->f64[i] * v1->f64[i] + ((i % 2 == 0) ? v3->f64[i] : -v3->f64[i]); + break; + + default: + printf("ERROR -- bad fma order %d\n", (int)order); + n_errs++; + return; + } + } +} + +void NOINLINE do_fmaddsub_ps() { + f4 = _mm512_fmaddsub_ps(f1, f2, f3); + emulate_fmaddsub_ps(&f5, &f1, 0xffff, &f3, &f2, FMA_132); + check_equal32(&f4, &f5, (void *)0, 0xffff, "_mm512_fmaddsub_ps"); + + f4 = _mm512_mask_fmaddsub_ps(f1, 0x79fa, f2, f3); + emulate_fmaddsub_ps(&f5, &f1, 0x79fa, &f3, &f2, FMA_132); + check_equal32(&f4, &f5, &f1, 0x79fa, "_mm512_mask_fmaddsub_ps"); + + f4 = _mm512_mask3_fmaddsub_ps(f1, f2, f3, 0x563a); + emulate_fmaddsub_ps(&f5, &f3, 0x563a, &f1, &f2, FMA_231); + check_equal32(&f4, &f5, &f3, 0x563a, "_mm512_mask3_fmaddsub_ps"); + + /* + * Employ rounding modes. + * Our FP inputs are all integer values, so there's no need for any + * special emulation routine. + */ + + f4 = _mm512_fmaddsub_round_ps(f1, f2, f3, + _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC); + emulate_fmaddsub_ps(&f5, &f1, 0xffff, &f3, &f2, FMA_132); + check_equal32(&f4, &f5, (void *)0, 0xffff, "_mm512_fmaddsub_round_ps"); + + f4 = _mm512_mask_fmaddsub_round_ps(f1, 0x79fa, f2, f3, + _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); + emulate_fmaddsub_ps(&f5, &f1, 0x79fa, &f3, &f2, FMA_132); + check_equal32(&f4, &f5, &f1, 0x79fa, "_mm512_mask_fmaddsub_round_ps"); + + f4 = _mm512_mask3_fmaddsub_round_ps( + f1, f2, f3, 0x563a, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); + emulate_fmaddsub_ps(&f5, &f3, 0x563a, &f1, &f2, FMA_231); + check_equal32(&f4, &f5, &f3, 0x563a, "_mm512_mask3_fmaddsub_round_ps"); +} + +void NOINLINE do_fmaddsub_pd() { + d4 = _mm512_fmaddsub_pd(d1, d2, d3); + emulate_fmaddsub_pd(&d5, &d1, 0xffff, &d3, &d2, FMA_132); + check_equal64(&d4, &d5, (void *)0, 0xffff, "_mm512_fmaddsub_pd"); + + d4 = _mm512_mask_fmaddsub_pd(d1, 0xfa, d2, d3); + emulate_fmaddsub_pd(&d5, &d1, 0xfa, &d3, &d2, FMA_132); + check_equal64(&d4, &d5, &d1, 0xfa, "_mm512_mask_fmaddsub_pd"); + + d4 = _mm512_mask3_fmaddsub_pd(d1, d2, d3, 0x56); + emulate_fmaddsub_pd(&d5, &d3, 0x56, &d1, &d2, FMA_231); + check_equal64(&d4, &d5, &d3, 0x56, "_mm512_mask3_fmaddsub_pd"); + + /* + * Employ rounding modes. + * Our FP inputs are all integer values, so there's no need for any + * special emulation routine. + */ + + d4 = _mm512_fmaddsub_round_pd(d1, d2, d3, + _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC); + emulate_fmaddsub_pd(&d5, &d1, 0xffff, &d3, &d2, FMA_132); + check_equal64(&d4, &d5, (void *)0, 0xffff, "_mm512_fmaddsub_round_pd"); + + d4 = _mm512_mask_fmaddsub_round_pd(d1, 0x79, d2, d3, + _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); + emulate_fmaddsub_pd(&d5, &d1, 0x79, &d3, &d2, FMA_132); + check_equal64(&d4, &d5, &d1, 0x79, "_mm512_mask_fmaddsub_round_pd"); + + d4 = _mm512_mask3_fmaddsub_round_pd( + d1, d2, d3, 0x63, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); + emulate_fmaddsub_pd(&d5, &d3, 0x63, &d1, &d2, FMA_231); + check_equal64(&d4, &d5, &d3, 0x63, "_mm512_mask3_fmaddsub_round_pd"); +} + +void NOINLINE do_fmsubadd_ps() { + f4 = _mm512_fmsubadd_ps(f1, f2, f3); + emulate_fmsubadd_ps(&f5, &f1, 0xffff, &f3, &f2, FMA_132); + check_equal32(&f4, &f5, (void *)0, 0xffff, "_mm512_fmsubadd_ps"); + + f4 = _mm512_mask_fmsubadd_ps(f1, 0x79fa, f2, f3); + emulate_fmsubadd_ps(&f5, &f1, 0x79fa, &f3, &f2, FMA_132); + check_equal32(&f4, &f5, &f1, 0x79fa, "_mm512_mask_fmsubadd_ps"); + + f4 = _mm512_mask3_fmsubadd_ps(f1, f2, f3, 0x563a); + emulate_fmsubadd_ps(&f5, &f3, 0x563a, &f1, &f2, FMA_231); + check_equal32(&f4, &f5, &f3, 0x563a, "_mm512_mask3_fmsubadd_ps"); + + /* + * Employ rounding modes. + * Our FP inputs are all integer values, so there's no need for any + * special emulation routine. + */ + + f4 = _mm512_fmsubadd_round_ps(f1, f2, f3, + _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC); + emulate_fmsubadd_ps(&f5, &f1, 0xffff, &f3, &f2, FMA_132); + check_equal32(&f4, &f5, (void *)0, 0xffff, "_mm512_fmsubadd_round_ps"); + + f4 = _mm512_mask_fmsubadd_round_ps(f1, 0x79fa, f2, f3, + _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); + emulate_fmsubadd_ps(&f5, &f1, 0x79fa, &f3, &f2, FMA_132); + check_equal32(&f4, &f5, &f1, 0x79fa, "_mm512_mask_fmsubadd_round_ps"); + + f4 = _mm512_mask3_fmsubadd_round_ps( + f1, f2, f3, 0x563a, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); + emulate_fmsubadd_ps(&f5, &f3, 0x563a, &f1, &f2, FMA_231); + check_equal32(&f4, &f5, &f3, 0x563a, "_mm512_mask3_fmsubadd_round_ps"); +} + +void NOINLINE do_fmsubadd_pd() { + d4 = _mm512_fmsubadd_pd(d1, d2, d3); + emulate_fmsubadd_pd(&d5, &d1, 0xffff, &d3, &d2, FMA_132); + check_equal64(&d4, &d5, (void *)0, 0xffff, "_mm512_fmsubadd_pd"); + + d4 = _mm512_mask_fmsubadd_pd(d1, 0xfa, d2, d3); + emulate_fmsubadd_pd(&d5, &d1, 0xfa, &d3, &d2, FMA_132); + check_equal64(&d4, &d5, &d1, 0xfa, "_mm512_mask_fmsubadd_pd"); + + d4 = _mm512_mask3_fmsubadd_pd(d1, d2, d3, 0x56); + emulate_fmsubadd_pd(&d5, &d3, 0x56, &d1, &d2, FMA_231); + check_equal64(&d4, &d5, &d3, 0x56, "_mm512_mask3_fmsubadd_pd"); + + /* + * Employ rounding modes. + * Our FP inputs are all integer values, so there's no need for any + * special emulation routine. + */ + + d4 = _mm512_fmsubadd_round_pd(d1, d2, d3, + _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC); + emulate_fmsubadd_pd(&d5, &d1, 0xffff, &d3, &d2, FMA_132); + check_equal64(&d4, &d5, (void *)0, 0xffff, "_mm512_fmsubadd_round_pd"); + + d4 = _mm512_mask_fmsubadd_round_pd(d1, 0x79, d2, d3, + _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); + emulate_fmsubadd_pd(&d5, &d1, 0x79, &d3, &d2, FMA_132); + check_equal64(&d4, &d5, &d1, 0x79, "_mm512_mask_fmsubadd_round_pd"); + + d4 = _mm512_mask3_fmsubadd_round_pd( + d1, d2, d3, 0x63, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); + emulate_fmsubadd_pd(&d5, &d3, 0x63, &d1, &d2, FMA_231); + check_equal64(&d4, &d5, &d3, 0x63, "_mm512_mask3_fmsubadd_round_pd"); +} + +int main(int argc, char *argv[]) { + if (argc > 1 && argv[1][0] == '-' && argv[1][1] == 'v' && + argv[1][2] == '\0') { + verbose = 1; + } + + init(); + + do_fmaddsub_ps(); + do_fmaddsub_pd(); + + do_fmsubadd_ps(); + do_fmsubadd_pd(); + + if (n_errs != 0) { + printf("FAILED\n"); + return 1; + } + + printf("PASSED\n"); + return 0; +} |