@samitouri / QOSamiQemu / commits / 7646bc0b72

target/hexagon: add v68 HVX IEEE float compare insns

Add HVX IEEE floating-point compare instructions: - V6_vgthf, V6_vgtsf: greater-than compare - V6_vgthf_and, V6_vgtsf_and: greater-than with predicate-and - V6_vgthf_or, V6_vgtsf_or: greater-than with predicate-or - V6_vgthf_xor, V6_vgtsf_xor: greater-than with predicate-xor Reviewed-by: Taylor Simpson <ltaylorsimpson@gmail.com> Signed-off-by: Matheus Tavares Bernardino <matheus.bernardino@oss.qualcomm.com> Reviewed-by: Brian Cain <brian.cain@oss.qualcomm.com> Link: https://lore.kernel.org/qemu-devel/96e27588ad5a7d47ba9f56f1d547729a19b691c5.1776339451.git.matheus.bernardino@oss.qualcomm.com Signed-off-by: Brian Cain <brian.cain@oss.qualcomm.com>

Matheus Tavares Bernardino committed Apr 16, 2026 at 04:38 UTC 7646bc0b72a78960d497371036d5841dc590412b
5 files changed +128 -1
target/hexagon/imported/mmvec/encode_ext.def
+10
@@ -858,4 +858,14 @@ DEF_ENC(V6_vconv_w_sf,"00011110--0--101PP1uuuuu001ddddd")
858 DEF_ENC(V6_vconv_hf_h,"00011110--0--101PP1uuuuu100ddddd")
859 DEF_ENC(V6_vconv_h_hf,"00011110--0--101PP1uuuuu010ddddd")
860
861 +/* IEEE FP compare instructions */
862 +DEF_ENC(V6_vgtsf,"00011100100vvvvvPP1uuuuu011100dd")
863 +DEF_ENC(V6_vgthf,"00011100100vvvvvPP1uuuuu011101dd")
864 +DEF_ENC(V6_vgtsf_and,"00011100100vvvvvPP1uuuuu110010xx")
865 +DEF_ENC(V6_vgthf_and,"00011100100vvvvvPP1uuuuu110011xx")
866 +DEF_ENC(V6_vgtsf_or,"00011100100vvvvvPP1uuuuu001100xx")
867 +DEF_ENC(V6_vgthf_or,"00011100100vvvvvPP1uuuuu001101xx")
868 +DEF_ENC(V6_vgtsf_xor,"00011100100vvvvvPP1uuuuu111010xx")
869 +DEF_ENC(V6_vgthf_xor,"00011100100vvvvvPP1uuuuu111011xx")
870 +
871 #endif /* NO MMVEC */
target/hexagon/imported/mmvec/ext.idef
+61
@@ -3143,6 +3143,67 @@ ITERATOR_INSN_SHIFT_SLOT_FLT(16, vconv_hf_h,"Vd32.hf=Vu32.h",
3143 "Vector conversion of int hw format to hf16",
3144 VdV.hf[i] = float16_val(int16_to_float16(VuV.h[i], &env->hvx_fp_status)))
3145
3146 +/******************************************************************************
3147 + * IEEE FP compare instructions
3148 + ******************************************************************************/
3149 +
3150 +#define VCMPGT_SF(DEST, ASRC, ASRCOP, CMP, N, SRC, MASK, WIDTH) \
3151 +{ \
3152 + for (fHIDE(int) i = 0; i < fVBYTES(); i += WIDTH) { \
3153 + fHIDE(int) VAL = fCMPGT_SF(VuV.SRC[i/WIDTH],VvV.SRC[i/WIDTH]) ? MASK : 0; \
3154 + fSETQBITS(DEST,WIDTH,MASK,i,ASRC ASRCOP VAL); \
3155 + } \
3156 +}
3157 +
3158 +#define VCMPGT_HF(DEST, ASRC, ASRCOP, CMP, N, SRC, MASK, WIDTH) \
3159 +{ \
3160 + for (fHIDE(int) i = 0; i < fVBYTES(); i += WIDTH) { \
3161 + fHIDE(int) VAL = fCMPGT_HF(VuV.SRC[i/WIDTH],VvV.SRC[i/WIDTH]) ? MASK : 0; \
3162 + fSETQBITS(DEST,WIDTH,MASK,i,ASRC ASRCOP VAL); \
3163 + } \
3164 +}
3165 +
3166 +/* Vector SF compare */
3167 +#define MMVEC_CMPGT_SF(TYPE,TYPE2,DESCR,N,MASK,WIDTH,SRC) \
3168 + EXTINSN(V6_vgt##TYPE##_and, "Qx4&=vcmp.gt(Vu32." TYPE2 ",Vv32." TYPE2 ")", \
3169 + ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VA,A_CVI_VA_2SRC,A_HVX_FLT), \
3170 + DESCR" greater than with predicate-and", \
3171 + VCMPGT_SF(QxV, fGETQBITS(QxV,WIDTH,MASK,i), &, ">", N, SRC, MASK, WIDTH)) \
3172 + EXTINSN(V6_vgt##TYPE##_xor, "Qx4^=vcmp.gt(Vu32." TYPE2 ",Vv32." TYPE2 ")", \
3173 + ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VA,A_CVI_VA_2SRC,A_HVX_FLT), \
3174 + DESCR" greater than with predicate-xor", \
3175 + VCMPGT_SF(QxV, fGETQBITS(QxV,WIDTH,MASK,i), ^, ">", N, SRC, MASK, WIDTH)) \
3176 + EXTINSN(V6_vgt##TYPE##_or, "Qx4|=vcmp.gt(Vu32." TYPE2 ",Vv32." TYPE2 ")", \
3177 + ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VA,A_CVI_VA_2SRC,A_HVX_FLT), \
3178 + DESCR" greater than with predicate-or", \
3179 + VCMPGT_SF(QxV, fGETQBITS(QxV,WIDTH,MASK,i), |, ">", N, SRC, MASK, WIDTH)) \
3180 + EXTINSN(V6_vgt##TYPE, "Qd4=vcmp.gt(Vu32." TYPE2 ",Vv32." TYPE2 ")", \
3181 + ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VA,A_CVI_VA_2SRC,A_HVX_FLT), \
3182 + DESCR" greater than", \
3183 + VCMPGT_SF(QdV, , , ">", N, SRC, MASK, WIDTH))
3184 +
3185 +/* Vector HF compare */
3186 +#define MMVEC_CMPGT_HF(TYPE,TYPE2,DESCR,N,MASK,WIDTH,SRC) \
3187 + EXTINSN(V6_vgt##TYPE##_and, "Qx4&=vcmp.gt(Vu32." TYPE2 ",Vv32." TYPE2 ")", \
3188 + ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VA,A_CVI_VA_2SRC,A_HVX_FLT), \
3189 + DESCR" greater than with predicate-and", \
3190 + VCMPGT_HF(QxV, fGETQBITS(QxV,WIDTH,MASK,i), &, ">", N, SRC, MASK, WIDTH)) \
3191 + EXTINSN(V6_vgt##TYPE##_xor, "Qx4^=vcmp.gt(Vu32." TYPE2 ",Vv32." TYPE2 ")", \
3192 + ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VA,A_CVI_VA_2SRC,A_HVX_FLT), \
3193 + DESCR" greater than with predicate-xor", \
3194 + VCMPGT_HF(QxV, fGETQBITS(QxV,WIDTH,MASK,i), ^, ">", N, SRC, MASK, WIDTH)) \
3195 + EXTINSN(V6_vgt##TYPE##_or, "Qx4|=vcmp.gt(Vu32." TYPE2 ",Vv32." TYPE2 ")", \
3196 + ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VA,A_CVI_VA_2SRC,A_HVX_FLT), \
3197 + DESCR" greater than with predicate-or", \
3198 + VCMPGT_HF(QxV, fGETQBITS(QxV,WIDTH,MASK,i), |, ">", N, SRC, MASK, WIDTH)) \
3199 + EXTINSN(V6_vgt##TYPE, "Qd4=vcmp.gt(Vu32." TYPE2 ",Vv32." TYPE2 ")", \
3200 + ATTRIBS(A_EXTENSION,A_CVI,A_CVI_VA,A_CVI_VA_2SRC,A_HVX_FLT), \
3201 + DESCR" greater than", \
3202 + VCMPGT_HF(QdV, , , ">", N, SRC, MASK, WIDTH))
3203 +
3204 +MMVEC_CMPGT_SF(sf,"sf","Vector sf Compare ", fVELEM(32), 0xF, 4, sf)
3205 +MMVEC_CMPGT_HF(hf,"hf","Vector hf Compare ", fVELEM(16), 0x3, 2, hf)
3206 +
3207 /******************************************************************************
3208 DEBUG Vector/Register Printing
3209 ******************************************************************************/
target/hexagon/mmvec/hvx_ieee_fp.c
+48
@@ -87,3 +87,51 @@ int16_t conv_h_hf(float16 a, float_status *fp_status)
87 }
88 return float16_to_int16_round_to_zero(a, fp_status);
89 }
90 +
91 +/*
92 + * Returns true if f1 > f2, where at least one of the elements is guaranteed
93 + * to be NaN.
94 + * Up to v73, Hexagon HVX IEEE FP follows this order:
95 + * QNaN > SNaN > +Inf > numbers > -Inf > SNaN_neg > QNaN_neg
96 + */
97 +static bool float32_nan_compare(float32 f1, float32 f2, float_status *fp_status)
98 +{
99 + /* opposite signs case */
100 + if (float32_is_neg(f1) != float32_is_neg(f2)) {
101 + return !float32_is_neg(f1);
102 + }
103 +
104 + /* same sign case */
105 + bool result = (float32_is_any_nan(f1) && !float32_is_any_nan(f2)) ||
106 + (float32_is_quiet_nan(f1, fp_status) && !float32_is_quiet_nan(f2, fp_status));
107 + return float32_is_neg(f1) ? !result : result;
108 +}
109 +
110 +static bool float16_nan_compare(float16 f1, float16 f2, float_status *fp_status)
111 +{
112 + /* opposite signs case */
113 + if (float16_is_neg(f1) != float16_is_neg(f2)) {
114 + return !float16_is_neg(f1);
115 + }
116 +
117 + /* same sign case */
118 + bool result = (float16_is_any_nan(f1) && !float16_is_any_nan(f2)) ||
119 + (float16_is_quiet_nan(f1, fp_status) && !float16_is_quiet_nan(f2, fp_status));
120 + return float16_is_neg(f1) ? !result : result;
121 +}
122 +
123 +uint32_t cmpgt_sf(float32 a1, float32 a2, float_status *fp_status)
124 +{
125 + if (float32_is_any_nan(a1) || float32_is_any_nan(a2)) {
126 + return float32_nan_compare(a1, a2, fp_status);
127 + }
128 + return float32_compare(a1, a2, fp_status) == float_relation_greater;
129 +}
130 +
131 +uint16_t cmpgt_hf(float16 a1, float16 a2, float_status *fp_status)
132 +{
133 + if (float16_is_any_nan(a1) || float16_is_any_nan(a2)) {
134 + return float16_nan_compare(a1, a2, fp_status);
135 + }
136 + return float16_compare(a1, a2, fp_status) == float_relation_greater;
137 +}
target/hexagon/mmvec/hvx_ieee_fp.h
+4
@@ -25,4 +25,8 @@ float16 qf_min_hf(float16 a1, float16 a2, float_status *fp_status);
25 int32_t conv_w_sf(float32 a, float_status *fp_status);
26 int16_t conv_h_hf(float16 a, float_status *fp_status);
27
28 +/* IEEE - FP compare instructions */
29 +uint32_t cmpgt_sf(float32 a1, float32 a2, float_status *fp_status);
30 +uint16_t cmpgt_hf(float16 a1, float16 a2, float_status *fp_status);
31 +
32 #endif
target/hexagon/mmvec/macros.h
+5 -1
@@ -354,6 +354,10 @@
354 do { \
355 COE = (sextract32(VAL, 24 + 2 * POS, 2) << 8) | \
356 extract32(VAL, POS * 8, 8); \
357 - } while (0);
357 + } while (0) \
358 + ;
359 +
360 +#define fCMPGT_SF(A, B) cmpgt_sf(A, B, &env->hvx_fp_status)
361 +#define fCMPGT_HF(A, B) cmpgt_hf(A, B, &env->hvx_fp_status)
362
363 #endif