Changeset 4795 in CLRX

Timestamp:

Jul 17, 2019, 12:07:02 PM (5 months ago)

Author:

matszpk

Message:

CLRadeonExtender: GCNAsm: Add support for 1/(2*PI) promotion for GFX10. Add SOP2 instructions testcases for GFX10.

Location:

CLRadeonExtender/trunk

Files:

• amdasm/GCNAsmEncode1.cpp (modified) (1 diff)
• tests/amdasm/GCNAsmOpc15.cpp (modified) (2 diffs)

CLRadeonExtender/trunk/amdasm/GCNAsmEncode1.cpp

@@ -51 +51 @@
     else if (!src0Op.range.isRegVar() &&
             ((src0Op.range.start>=240 && src0Op.range.start<248) ||
-             ((arch&ARCH_GCN_1_2_4)!=0 && src0Op.range.start==248)))
+             ((arch&ARCH_GCN_1_2_4_5)!=0 && src0Op.range.start==248)))
     {
         // floating point immediates to literal

CLRadeonExtender/trunk/tests/amdasm/GCNAsmOpc15.cpp

@@ -48 +48 @@
     { "    s_add_u32  vcc[1:1], s4, s61", 0x806b3d04U, 0, false, true, "" },
     { "    s_add_u32  vcc[1], s4, s61", 0x806b3d04U, 0, false, true, "" },
+    { "    s_add_u32  vcc[0:0], s4, s61", 0x806a3d04U, 0, false, true, "" },
+    { "    s_add_u32  vcc[0], s4, s61", 0x806a3d04U, 0, false, true, "" },
+    { "    s_add_u32  vcc_lo, s4, s61", 0x806a3d04U, 0, false, true, "" },
+    { "    s_add_u32  vcc_hi, s4, s61", 0x806b3d04U, 0, false, true, "" },
     { "    s_add_u32  ttmp0, s4, s61", 0x806c3d04U, 0, false, true, "" },
     { "    s_add_u32  ttmp1, s4, s61", 0x806d3d04U, 0, false, true, "" },
@@ -125 +129 @@
     { "    s_xor_b64 s[22:23], execz, s[62:63]\n", 0x89963efcU, 0, false, true, "" },
     { "    s_xor_b64 s[22:23], scc, s[62:63]\n", 0x89963efdU, 0, false, true, "" },
+    /* SOP2 encodings */
+    { "    s_sub_u32  s21, s4, s61", 0x80953d04U, 0, false, true, "" },
+    { "    s_sub_u32_e32  s21, s4, s61", 0x80953d04U, 0, false, true, "" },
+    { "    s_add_i32  s21, s4, s61", 0x81153d04U, 0, false, true, "" },
+    { "    s_sub_i32  s21, s4, s61", 0x81953d04U, 0, false, true, "" },
+    { "    s_addc_u32  s21, s4, s61", 0x82153d04U, 0, false, true, "" },
+    { "    s_subb_u32  s21, s4, s61", 0x82953d04U, 0, false, true, "" },
+    { "    s_min_i32  s21, s4, s61", 0x83153d04U, 0, false, true, "" },
+    { "    s_min_u32  s21, s4, s61", 0x83953d04U, 0, false, true, "" },
+    { "    s_max_i32  s21, s4, s61", 0x84153d04U, 0, false, true, "" },
+    { "    s_max_u32  s21, s4, s61", 0x84953d04U, 0, false, true, "" },
+    { "    s_cselect_b32  s21, s4, s61", 0x85153d04U, 0, false, true, "" },
+    { "    s_cselect_b64  s[20:21], s[4:5], s[62:63]", 0x85943e04U, 0, false, true, "" },
+    { "    s_and_b32  s21, s4, s61", 0x87153d04U, 0, false, true, "" },
+    { "    s_and_b64  s[20:21], s[4:5], s[62:63]", 0x87943e04U, 0, false, true, "" },
+    { "    s_or_b32  s21, s4, s61", 0x88153d04U, 0, false, true, "" },
+    { "    s_or_b64  s[20:21], s[4:5], s[62:63]", 0x88943e04U, 0, false, true, "" },
+    { "    s_xor_b32  s21, s4, s61", 0x89153d04U, 0, false, true, "" },
+    { "    s_xor_b64  s[20:21], s[4:5], s[62:63]", 0x89943e04U, 0, false, true, "" },
+    { "    s_andn2_b32  s21, s4, s61", 0x8a153d04U, 0, false, true, "" },
+    { "    s_andn2_b64  s[20:21], s[4:5], s[62:63]", 0x8a943e04U, 0, false, true, "" },
+    { "    s_orn2_b32  s21, s4, s61", 0x8b153d04U, 0, false, true, "" },
+    { "    s_orn2_b64  s[20:21], s[4:5], s[62:63]", 0x8b943e04U, 0, false, true, "" },
+    { "    s_nand_b32  s21, s4, s61", 0x8c153d04U, 0, false, true, "" },
+    { "    s_nand_b64  s[20:21], s[4:5], s[62:63]", 0x8c943e04U, 0, false, true, "" },
+    { "    s_nor_b32  s21, s4, s61", 0x8d153d04U, 0, false, true, "" },
+    { "    s_nor_b64  s[20:21], s[4:5], s[62:63]", 0x8d943e04U, 0, false, true, "" },
+    { "    s_xnor_b32  s21, s4, s61", 0x8e153d04U, 0, false, true, "" },
+    { "    s_xnor_b64  s[20:21], s[4:5], s[62:63]", 0x8e943e04U, 0, false, true, "" },
+    { "    s_lshl_b32  s21, s4, s61", 0x8f153d04U, 0, false, true, "" },
+    { "    s_lshl_b64  s[20:21], s[4:5], s61", 0x8f943d04U, 0, false, true, "" },
+    { "    s_lshr_b32  s21, s4, s61", 0x90153d04U, 0, false, true, "" },
+    { "    s_lshr_b64  s[20:21], s[4:5], s61", 0x90943d04U, 0, false, true, "" },
+    { "    s_ashr_i32  s21, s4, s61", 0x91153d04U, 0, false, true, "" },
+    { "    s_ashr_i64  s[20:21], s[4:5], s61", 0x91943d04U, 0, false, true, "" },
+    { "    s_bfm_b32  s21, s4, s61", 0x92153d04U, 0, false, true, "" },
+    { "    s_bfm_b64  s[20:21], s4, s62", 0x92943e04U, 0, false, true, "" },
+    { "    s_mul_i32  s21, s4, s61", 0x93153d04U, 0, false, true, "" },
+    { "    s_bfe_u32  s21, s4, s61", 0x93953d04U, 0, false, true, "" },
+    { "    s_bfe_i32  s21, s4, s61", 0x94153d04U, 0, false, true, "" },
+    { "    s_bfe_u64  s[20:21], s[4:5], s61", 0x94943d04U, 0, false, true, "" },
+    { "    s_bfe_i64  s[20:21], s[4:5], s61", 0x95143d04U, 0, false, true, "" },
+    { "    s_cbranch_g_fork  s[4:5], s[62:63]", 0, 0, false, false,
+        "test.s:1:5: Error: Unknown instruction\n" },
+    { "    s_absdiff_i32  s21, s4, s61", 0x96153d04U, 0, false, true, "" },
+    { "        s_lshl1_add_u32 s21, s4, s61\n", 0x97153d04U, 0, false, true, "" },
+    { "        s_lshl2_add_u32 s21, s4, s61\n", 0x97953d04U, 0, false, true, "" },
+    { "        s_lshl3_add_u32 s21, s4, s61\n", 0x98153d04U, 0, false, true, "" },
+    { "        s_lshl4_add_u32 s21, s4, s61\n", 0x98953d04U, 0, false, true, "" },
+    { "        s_pack_ll_b32_b16 s21, s4, s61\n", 0x99153d04U, 0, false, true, "" },
+    { "        s_pack_lh_b32_b16 s21, s4, s61\n", 0x99953d04U, 0, false, true, "" },
+    { "        s_pack_hh_b32_b16 s21, s4, s61\n", 0x9a153d04U, 0, false, true, "" },
+    { "        s_mul_hi_u32    s21, s4, s61\n", 0x9a953d04U, 0, false, true, "" },
+    { "        s_mul_hi_i32    s21, s4, s61\n", 0x9b153d04U, 0, false, true, "" },
     { nullptr, 0, 0, false, false, 0 }
 };