- chan = vir_FMOV(c, chan);
- vir_set_unpack(c->defs[chan.index], 0, unpack);
- } else {
- /* If we're unpacking the low field, shift it
- * up to the top first.
- */
- if ((i & 1) == 0) {
- chan = vir_SHL(c, chan,
- vir_uniform_ui(c, 16));
- }
-
- /* Do proper sign extension to a 32-bit int. */
- if (nir_alu_type_get_base_type(instr->dest_type) ==
- nir_type_int) {
- chan = vir_ASR(c, chan,
- vir_uniform_ui(c, 16));
- } else {
- chan = vir_SHR(c, chan,
- vir_uniform_ui(c, 16));
- }
- }
- } else {
- chan = vir_MOV(c, return_values[i]);
+ p2_unpacked.op = v3d40_image_load_store_tmu_op(instr);
+
+ /* If we were able to replace atomic_add for an inc/dec, then we
+ * need/can to do things slightly different, like not loading the
+ * amount to add/sub, as that is implicit.
+ */
+ bool atomic_add_replaced = (instr->intrinsic == nir_intrinsic_image_deref_atomic_add &&
+ (p2_unpacked.op == V3D_TMU_OP_WRITE_AND_READ_INC ||
+ p2_unpacked.op == V3D_TMU_OP_WRITE_OR_READ_DEC));
+
+ bool is_1d = false;
+ switch (glsl_get_sampler_dim(sampler_type)) {
+ case GLSL_SAMPLER_DIM_1D:
+ is_1d = true;
+ break;
+ case GLSL_SAMPLER_DIM_BUF:
+ break;
+ case GLSL_SAMPLER_DIM_2D:
+ case GLSL_SAMPLER_DIM_RECT:
+ vir_TMU_WRITE(c, V3D_QPU_WADDR_TMUT,
+ ntq_get_src(c, instr->src[1], 1), &tmu_writes);
+ break;
+ case GLSL_SAMPLER_DIM_3D:
+ case GLSL_SAMPLER_DIM_CUBE:
+ vir_TMU_WRITE(c, V3D_QPU_WADDR_TMUT,
+ ntq_get_src(c, instr->src[1], 1), &tmu_writes);
+ vir_TMU_WRITE(c, V3D_QPU_WADDR_TMUR,
+ ntq_get_src(c, instr->src[1], 2), &tmu_writes);
+ break;
+ default:
+ unreachable("bad image sampler dim");
+ }
+
+ if (glsl_sampler_type_is_array(sampler_type)) {
+ vir_TMU_WRITE(c, V3D_QPU_WADDR_TMUI,
+ ntq_get_src(c, instr->src[1],
+ is_1d ? 1 : 2), &tmu_writes);
+ }
+
+ /* Limit the number of channels returned to both how many the NIR
+ * instruction writes and how many the instruction could produce.
+ */
+ uint32_t instr_return_channels = nir_intrinsic_dest_components(instr);
+ if (!p1_unpacked.output_type_32_bit)
+ instr_return_channels = (instr_return_channels + 1) / 2;
+
+ p0_unpacked.return_words_of_texture_data =
+ (1 << instr_return_channels) - 1;
+
+ uint32_t p0_packed;
+ V3D41_TMU_CONFIG_PARAMETER_0_pack(NULL,
+ (uint8_t *)&p0_packed,
+ &p0_unpacked);
+
+ uint32_t p1_packed;
+ V3D41_TMU_CONFIG_PARAMETER_1_pack(NULL,
+ (uint8_t *)&p1_packed,
+ &p1_unpacked);
+
+ uint32_t p2_packed;
+ V3D41_TMU_CONFIG_PARAMETER_2_pack(NULL,
+ (uint8_t *)&p2_packed,
+ &p2_unpacked);
+
+ /* Load unit number into the high bits of the texture or sampler
+ * address field, which will be be used by the driver to decide which
+ * texture to put in the actual address field.
+ */
+ p0_packed |= unit << 24;
+
+ vir_WRTMUC(c, QUNIFORM_IMAGE_TMU_CONFIG_P0, p0_packed);
+ if (memcmp(&p1_unpacked, &p1_unpacked_default, sizeof(p1_unpacked)) != 0)
+ vir_WRTMUC(c, QUNIFORM_CONSTANT, p1_packed);
+ if (memcmp(&p2_unpacked, &p2_unpacked_default, sizeof(p2_unpacked)) != 0)
+ vir_WRTMUC(c, QUNIFORM_CONSTANT, p2_packed);
+
+ /* Emit the data writes for atomics or image store. */
+ if (instr->intrinsic != nir_intrinsic_image_deref_load &&
+ !atomic_add_replaced) {
+ /* Vector for stores, or first atomic argument */
+ struct qreg src[4];
+ for (int i = 0; i < nir_intrinsic_src_components(instr, 3); i++) {
+ src[i] = ntq_get_src(c, instr->src[3], i);
+ vir_TMU_WRITE(c, V3D_QPU_WADDR_TMUD, src[i],
+ &tmu_writes);
+ }
+
+ /* Second atomic argument */
+ if (instr->intrinsic ==
+ nir_intrinsic_image_deref_atomic_comp_swap) {
+ vir_TMU_WRITE(c, V3D_QPU_WADDR_TMUD,
+ ntq_get_src(c, instr->src[4], 0),
+ &tmu_writes);