jay: Add a GPR_FROM_UGPRS opcode

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41688>

src/intel/compiler/jay/jay_builder_opcodes.h.py

@@ -58,7 +58,7 @@ TEMPLATE = """
     multi_type = len(op.types) > 1
     info_size = f'sizeof(jay_{op.name}_info)' if op.extra_struct else '0'
     operands = ["dst"] + [f"src{i}" for i in range(num_srcs)]
-    if num_srcs > 0:
+    if num_srcs > 0 and op.name != 'gpr_from_ugprs':
         uniform = " && " .join([f"jay_is_uniform(src{i})" for i in range(num_srcs)])
         reg_file = f"({uniform}) ? UGPR : GPR"
     else:

src/intel/compiler/jay/jay_ir.h

@@ -691,6 +691,9 @@ jay_src_type(const jay_inst *I, unsigned s)
    if (I->op == JAY_OPCODE_CVT)
       return jay_cvt_src_type(I);
 
+   if (I->op == JAY_OPCODE_GPR_FROM_UGPRS)
+      return jay_gpr_from_ugprs_src_type(I);
+
    /* 16-bit operand */
    if (I->op == JAY_OPCODE_MUL_32X16 && s == 1)
       return jay_type_resize(I->type, jay_type_size_bits(I->type) / 2);

src/intel/compiler/jay/jay_opcodes.py

@@ -161,6 +161,16 @@ op('deswizzle_even', 1, 'f32', 0, ['bool src_hi'])
 op('lane_id_8', 0, 'u16')
 op('lane_id_expand', 1, 'u16', 0, ['unsigned width'])
 
+# Fill a scalar GPR from a contiguous UGPR[16] range containing words or bytes.
+# src_type can be either U8 or U16 (only).  For U8, stride can be 1 or 2, and
+# index can be either 0 or 1.  For U16, both stride and index must be 0.
+op('gpr_from_ugprs', 1, 'u32', 0, [
+    'enum jay_type src_type',
+    'uint8_t stride',
+    'uint8_t index',
+    'uint8_t pad',
+])
+
 # Sample ID calculation
 op('extract_byte_per_8lanes', 2, 'u32')
 op('shr_odd_subspans_by_4', 1, 'u16')

src/intel/compiler/jay/jay_to_binary.c

@@ -484,6 +484,12 @@ emit(struct brw_codegen *p,
               brw_imm_uw(jay_lane_id_expand_width(I)));
       break;
 
+   case JAY_OPCODE_GPR_FROM_UGPRS:
+      brw_MOV(p, dst,
+              byte_offset(stride(SRC(0), jay_gpr_from_ugprs_stride(I), 1, 0),
+                          jay_gpr_from_ugprs_index(I)));
+      break;
+
    case JAY_OPCODE_EXTRACT_BYTE_PER_8LANES:
       brw_MOV(p, dst, stride(retype(SRC(simd_offs), BRW_TYPE_UB), 1, 8, 0));
       break;

src/intel/compiler/jay/jay_validate.c

@@ -85,6 +85,10 @@ get_src_words(struct validate_state *validate, jay_inst *I, unsigned s)
       return 4;
    }
 
+   if (I->op == JAY_OPCODE_GPR_FROM_UGPRS) {
+      return jay_ugpr_per_grf(validate->func->shader);
+   }
+
    bool vectorized = I->dst.file == UGPR &&
                      jay_num_values(I->dst) > jay_type_vector_length(I->type) &&
                      I->op != JAY_OPCODE_SEND &&
@@ -252,6 +256,14 @@ validate_inst(struct validate_state *validate, jay_inst *I)
       CHECK(jay_is_flag(I->src[2]) && "SEL src[2] (selector) must be a flag");
    } else if (I->op == JAY_OPCODE_SYNC) {
       CHECK(validate->post_ra && "SYNC does not exist while scheduling");
+   } else if (I->op == JAY_OPCODE_GPR_FROM_UGPRS) {
+      enum jay_type src_type = jay_gpr_from_ugprs_src_type(I);
+      CHECK(I->dst.file == GPR);
+      CHECK(I->src[0].file == UGPR);
+      CHECK(jay_num_values(I->src[0]) == 16);
+      CHECK(src_type == JAY_TYPE_U8 || src_type == JAY_TYPE_U16);
+      CHECK(jay_gpr_from_ugprs_stride(I) <= 16 / jay_type_size_bits(src_type));
+      CHECK(jay_gpr_from_ugprs_index(I) < 16 / jay_type_size_bits(src_type));
    }
 }