vc4: Add support for SNE/SEQ/SGE/SLT.

6 files changed, 96 insertions, 11 deletions

diff --git a/src/gallium/drivers/vc4/vc4_program.c b/src/gallium/drivers/vc4/vc4_program.c
index 4755ea0db90..4742c54772a 100644
--- a/src/gallium/drivers/vc4/vc4_program.c
+++ b/src/gallium/drivers/vc4/vc4_program.c
@@ -254,6 +254,10 @@ emit_tgsi_instruction(struct tgsi_to_qir *trans,
                 [TGSI_OPCODE_MIN] = { QOP_FMIN, tgsi_to_qir_alu },
                 [TGSI_OPCODE_MAX] = { QOP_FMAX, tgsi_to_qir_alu },
                 [TGSI_OPCODE_RSQ] = { QOP_RSQ, tgsi_to_qir_alu },
+                [TGSI_OPCODE_SEQ] = { QOP_SEQ, tgsi_to_qir_alu },
+                [TGSI_OPCODE_SNE] = { QOP_SNE, tgsi_to_qir_alu },
+                [TGSI_OPCODE_SGE] = { QOP_SGE, tgsi_to_qir_alu },
+                [TGSI_OPCODE_SLT] = { QOP_SLT, tgsi_to_qir_alu },
                 [TGSI_OPCODE_MAD] = { 0, tgsi_to_qir_mad },
                 [TGSI_OPCODE_DP2] = { 0, tgsi_to_qir_dp2 },
                 [TGSI_OPCODE_DP3] = { 0, tgsi_to_qir_dp3 },
diff --git a/src/gallium/drivers/vc4/vc4_qir.c b/src/gallium/drivers/vc4/vc4_qir.c
index a4bb6cd1fd1..4ee1f018fc5 100644
--- a/src/gallium/drivers/vc4/vc4_qir.c
+++ b/src/gallium/drivers/vc4/vc4_qir.c
@@ -43,6 +43,12 @@ static const struct qir_op_info qir_op_info[] = {
         [QOP_FMAX] = { "fmax", 1, 2 },
         [QOP_FMINABS] = { "fminabs", 1, 2 },
         [QOP_FMAXABS] = { "fmaxabs", 1, 2 },
+
+        [QOP_SEQ] = { "seq", 1, 2 },
+        [QOP_SNE] = { "sne", 1, 2 },
+        [QOP_SGE] = { "sge", 1, 2 },
+        [QOP_SLT] = { "slt", 1, 2 },
+
         [QOP_FTOI] = { "ftoi", 1, 1 },
         [QOP_RCP] = { "rcp", 1, 1 },
         [QOP_RSQ] = { "rsq", 1, 1 },
diff --git a/src/gallium/drivers/vc4/vc4_qir.h b/src/gallium/drivers/vc4/vc4_qir.h
index ae9e1796b90..4263adcddd1 100644
--- a/src/gallium/drivers/vc4/vc4_qir.h
+++ b/src/gallium/drivers/vc4/vc4_qir.h
@@ -50,6 +50,12 @@ enum qop {
         QOP_FMAX,
         QOP_FMINABS,
         QOP_FMAXABS,
+
+        QOP_SEQ,
+        QOP_SNE,
+        QOP_SGE,
+        QOP_SLT,
+
         QOP_FTOI,
         QOP_RCP,
         QOP_RSQ,
diff --git a/src/gallium/drivers/vc4/vc4_qpu_defines.h b/src/gallium/drivers/vc4/vc4_qpu_defines.h
index 13c940c0f8e..bdd5d94708f 100644
--- a/src/gallium/drivers/vc4/vc4_qpu_defines.h
+++ b/src/gallium/drivers/vc4/vc4_qpu_defines.h
@@ -223,6 +223,8 @@ enum qpu_pack_a {
 #define QPU_COND_MUL_SHIFT              46
 #define QPU_COND_MUL_MASK               QPU_MASK(48, 46)
 
+#define QPU_SF                          ((uint64_t)1 << 45)
+
 #define QPU_WADDR_ADD_SHIFT             38
 #define QPU_WADDR_ADD_MASK              QPU_MASK(43, 38)
 #define QPU_WADDR_MUL_SHIFT             32
diff --git a/src/gallium/drivers/vc4/vc4_qpu_disasm.c b/src/gallium/drivers/vc4/vc4_qpu_disasm.c
index 0aea2970f68..4ec6d9657b7 100644
--- a/src/gallium/drivers/vc4/vc4_qpu_disasm.c
+++ b/src/gallium/drivers/vc4/vc4_qpu_disasm.c
@@ -199,6 +199,17 @@ static const char *qpu_pack_a[] = {
         [QPU_PACK_A_8D_SAT] = ".8d.sat",
 };
 
+static const char *qpu_condflags[] = {
+        [QPU_COND_NEVER] = ".never",
+        [QPU_COND_ALWAYS] = "",
+        [QPU_COND_ZS] = ".zs",
+        [QPU_COND_ZC] = ".zc",
+        [QPU_COND_NS] = ".ns",
+        [QPU_COND_NC] = ".nc",
+        [QPU_COND_CS] = ".cs",
+        [QPU_COND_CC] = ".cc",
+};
+
 #define DESC(array, index)                                        \
         ((index > ARRAY_SIZE(array) || !(array)[index]) ?         \
          "???" : (array)[index])
@@ -282,11 +293,15 @@ static void
 print_add_op(uint64_t inst)
 {
         uint32_t op_add = QPU_GET_FIELD(inst, QPU_OP_ADD);
+        uint32_t cond = QPU_GET_FIELD(inst, QPU_COND_ADD);
         bool is_mov = (op_add == QPU_A_OR &&
                        QPU_GET_FIELD(inst, QPU_ADD_A) ==
                        QPU_GET_FIELD(inst, QPU_ADD_B));
 
-        fprintf(stderr, "%s ", is_mov ? "mov" : DESC(qpu_add_opcodes, op_add));
+        fprintf(stderr, "%s%s%s ",
+                is_mov ? "mov" : DESC(qpu_add_opcodes, op_add),
+                ((inst & QPU_SF) && op_add != QPU_A_NOP) ? ".sf" : "",
+                op_add != QPU_A_NOP ? DESC(qpu_condflags, cond) : "");
 
         print_alu_dst(inst, false);
         fprintf(stderr, ", ");
@@ -303,12 +318,17 @@ print_add_op(uint64_t inst)
 static void
 print_mul_op(uint64_t inst)
 {
+        uint32_t op_add = QPU_GET_FIELD(inst, QPU_OP_ADD);
         uint32_t op_mul = QPU_GET_FIELD(inst, QPU_OP_MUL);
+        uint32_t cond = QPU_GET_FIELD(inst, QPU_COND_MUL);
         bool is_mov = (op_mul == QPU_M_V8MIN &&
                        QPU_GET_FIELD(inst, QPU_MUL_A) ==
                        QPU_GET_FIELD(inst, QPU_MUL_B));
 
-        fprintf(stderr, "%s ", is_mov ? "mov" : DESC(qpu_mul_opcodes, op_mul));
+        fprintf(stderr, "%s%s%s ",
+                is_mov ? "mov" : DESC(qpu_mul_opcodes, op_mul),
+                ((inst & QPU_SF) && op_add == QPU_A_NOP) ? ".sf" : "",
+                op_mul != QPU_M_NOP ? DESC(qpu_condflags, cond) : "");
 
         print_alu_dst(inst, true);
         fprintf(stderr, ", ");
@@ -325,12 +345,18 @@ static void
 print_load_imm(uint64_t inst)
 {
         uint32_t imm = inst;
+        uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
+        uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
+        uint32_t cond_add = QPU_GET_FIELD(inst, QPU_COND_ADD);
+        uint32_t cond_mul = QPU_GET_FIELD(inst, QPU_COND_MUL);
 
         fprintf(stderr, "load_imm ");
         print_alu_dst(inst, false);
-        fprintf(stderr, ", ");
+        fprintf(stderr, "%s, ", (waddr_add != QPU_W_NOP ?
+                                 DESC(qpu_condflags, cond_add) : ""));
         print_alu_dst(inst, true);
-        fprintf(stderr, ", ");
+        fprintf(stderr, "%s, ", (waddr_mul != QPU_W_NOP ?
+                                 DESC(qpu_condflags, cond_mul) : ""));
         fprintf(stderr, "0x%08x (%f)", imm, uif(imm));
 }
 
diff --git a/src/gallium/drivers/vc4/vc4_qpu_emit.c b/src/gallium/drivers/vc4/vc4_qpu_emit.c
index 330876734d1..d0f7f894182 100644
--- a/src/gallium/drivers/vc4/vc4_qpu_emit.c
+++ b/src/gallium/drivers/vc4/vc4_qpu_emit.c
@@ -40,6 +40,28 @@ vc4_dump_program(struct qcompile *c)
         }
 }
 
+/**
+ * This is used to resolve the fact that we might register-allocate two
+ * different operands of an instruction to the same physical register file
+ * even though instructions have only one field for the register file source
+ * address.
+ *
+ * In that case, we need to move one to a temporary that can be used in the
+ * instruction, instead.
+ */
+static void
+fixup_raddr_conflict(uint64_t *insts, uint32_t *ni,
+               struct qpu_reg src0, struct qpu_reg *src1)
+{
+        if ((src0.mux == QPU_MUX_A || src0.mux == QPU_MUX_B) &&
+            (src1->mux == QPU_MUX_A || src1->mux == QPU_MUX_B) &&
+            src0.addr != src1->addr) {
+                insts[(*ni)++] = qpu_inst(qpu_a_MOV(qpu_r5(), *src1),
+                                          qpu_m_NOP());
+                *src1 = qpu_r5();
+        }
+}
+
 void
 vc4_generate_code(struct qcompile *c)
 {
@@ -110,6 +132,13 @@ vc4_generate_code(struct qcompile *c)
                         M(FMUL),
                 };
 
+                static const uint32_t compareflags[] = {
+                        [QOP_SEQ - QOP_SEQ] = QPU_COND_ZS,
+                        [QOP_SNE - QOP_SEQ] = QPU_COND_ZC,
+                        [QOP_SLT - QOP_SEQ] = QPU_COND_NS,
+                        [QOP_SGE - QOP_SEQ] = QPU_COND_NC,
+                };
+
                 struct qpu_reg src[4];
                 for (int i = 0; i < qir_get_op_nsrc(qinst->op); i++) {
                         int index = qinst->src[i].index;
@@ -184,6 +213,24 @@ vc4_generate_code(struct qcompile *c)
                         }
                         break;
 
+                case QOP_SEQ:
+                case QOP_SNE:
+                case QOP_SGE:
+                case QOP_SLT:
+                        fixup_raddr_conflict(insts, &ni, src[0], &src[1]);
+                        insts[ni++] = qpu_inst(qpu_a_SUB(qpu_ra(QPU_W_NOP),
+                                                         src[0], src[1]),
+                                               qpu_m_NOP());
+                        insts[ni - 1] |= QPU_SF;
+
+                        insts[ni++] = qpu_load_imm_f(dst, 0.0);
+                        insts[ni++] = qpu_load_imm_f(dst, 1.0);
+                        insts[ni - 1] = ((insts[ni - 1] & ~QPU_COND_ADD_MASK)
+                                         | QPU_SET_FIELD(compareflags[qinst->op - QOP_SEQ],
+                                                         QPU_COND_ADD));
+
+                        break;
+
                 case QOP_VPM_WRITE:
                         insts[ni++] = qpu_inst(qpu_a_MOV(qpu_ra(QPU_W_VPM),
                                                          src[0]),
@@ -274,13 +321,7 @@ vc4_generate_code(struct qcompile *c)
                         if (qir_get_op_nsrc(qinst->op) == 1)
                                 src[1] = src[0];
 
-                        if ((src[0].mux == QPU_MUX_A || src[0].mux == QPU_MUX_B) &&
-                            (src[1].mux == QPU_MUX_A || src[1].mux == QPU_MUX_B) &&
-                            src[0].addr != src[1].addr) {
-                                insts[ni++] = qpu_inst(qpu_a_MOV(qpu_r5(), src[1]),
-                                                       qpu_m_NOP());
-                                src[1] = qpu_r5();
-                        }
+                        fixup_raddr_conflict(insts, &ni, src[0], &src[1]);
 
                         if (translate[qinst->op].is_mul) {
                                 insts[ni++] = qpu_inst(qpu_a_NOP(),