}
} else {
assert(type->is_scalar() || type->is_vector());
- this->outputs[*location] = *reg;
- this->output_components[*location] = type->vector_elements;
- *reg = offset(*reg, bld, 4);
- (*location)++;
+ unsigned num_elements = type->vector_elements;
+ if (type->is_double())
+ num_elements *= 2;
+ for (unsigned count = 0; count < num_elements; count += 4) {
+ this->outputs[*location] = *reg;
+ this->output_components[*location] = MIN2(4, num_elements - count);
+ *reg = offset(*reg, bld, 4);
+ (*location)++;
+ }
}
}
brw_imm_ud(4 * REG_SIZE));
}
- if (indirect_offset.file == BAD_FILE) {
- /* Constant indexing - use global offset. */
- inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, dst, icp_handle);
- inst->offset = imm_offset;
- inst->mlen = 1;
- inst->base_mrf = -1;
- inst->regs_written = instr->num_components;
- } else {
- /* Indirect indexing - use per-slot offsets as well. */
- const fs_reg srcs[] = { icp_handle, indirect_offset };
- fs_reg payload = bld.vgrf(BRW_REGISTER_TYPE_UD, 2);
- bld.LOAD_PAYLOAD(payload, srcs, ARRAY_SIZE(srcs), 0);
+ /* We can only read two double components with each URB read, so
+ * we send two read messages in that case, each one loading up to
+ * two double components.
+ */
+ unsigned num_iterations = 1;
+ unsigned num_components = instr->num_components;
+ fs_reg orig_dst = dst;
+ if (type_sz(dst.type) == 8) {
+ if (instr->num_components > 2) {
+ num_iterations = 2;
+ num_components = 2;
+ }
- inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, dst, payload);
- inst->offset = imm_offset;
- inst->base_mrf = -1;
- inst->mlen = 2;
- inst->regs_written = instr->num_components;
+ fs_reg tmp = fs_reg(VGRF, alloc.allocate(4), dst.type);
+ dst = tmp;
}
- /* Copy the temporary to the destination to deal with writemasking.
- *
- * Also attempt to deal with gl_PointSize being in the .w component.
- */
- if (inst->offset == 0 && indirect_offset.file == BAD_FILE) {
- inst->dst = bld.vgrf(dst.type, 4);
- inst->regs_written = 4;
- bld.MOV(dst, offset(inst->dst, bld, 3));
+ for (unsigned iter = 0; iter < num_iterations; iter++) {
+ if (indirect_offset.file == BAD_FILE) {
+ /* Constant indexing - use global offset. */
+ inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, dst, icp_handle);
+ inst->offset = imm_offset;
+ inst->mlen = 1;
+ inst->base_mrf = -1;
+ } else {
+ /* Indirect indexing - use per-slot offsets as well. */
+ const fs_reg srcs[] = { icp_handle, indirect_offset };
+ fs_reg payload = bld.vgrf(BRW_REGISTER_TYPE_UD, 2);
+ bld.LOAD_PAYLOAD(payload, srcs, ARRAY_SIZE(srcs), 0);
+
+ inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, dst, payload);
+ inst->offset = imm_offset;
+ inst->base_mrf = -1;
+ inst->mlen = 2;
+ }
+ inst->regs_written = num_components * type_sz(dst.type) / 4;
+
+ /* If we are reading 64-bit data using 32-bit read messages we need
+ * build proper 64-bit data elements by shuffling the low and high
+ * 32-bit components around like we do for other things like UBOs
+ * or SSBOs.
+ */
+ if (type_sz(dst.type) == 8) {
+ shuffle_32bit_load_result_to_64bit_data(
+ bld, dst, retype(dst, BRW_REGISTER_TYPE_F), num_components);
+
+ for (unsigned c = 0; c < num_components; c++) {
+ bld.MOV(offset(orig_dst, bld, iter * 2 + c),
+ offset(dst, bld, c));
+ }
+ }
+
+ /* Copy the temporary to the destination to deal with writemasking.
+ *
+ * Also attempt to deal with gl_PointSize being in the .w component.
+ */
+ if (inst->offset == 0 && indirect_offset.file == BAD_FILE) {
+ assert(type_sz(dst.type) < 8);
+ inst->dst = bld.vgrf(dst.type, 4);
+ inst->regs_written = 4;
+ bld.MOV(dst, offset(inst->dst, bld, 3));
+ }
+
+ /* If we are loading double data and we need a second read message
+ * adjust the write offset
+ */
+ if (num_iterations > 1) {
+ num_components = instr->num_components - 2;
+ if (indirect_offset.file == BAD_FILE) {
+ imm_offset++;
+ } else {
+ fs_reg new_indirect = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
+ bld.ADD(new_indirect, indirect_offset, brw_imm_ud(1u));
+ indirect_offset = new_indirect;
+ }
+ }
}
break;
}
case nir_intrinsic_store_output:
case nir_intrinsic_store_per_vertex_output: {
fs_reg value = get_nir_src(instr->src[0]);
+ bool is_64bit = (instr->src[0].is_ssa ?
+ instr->src[0].ssa->bit_size : instr->src[0].reg.reg->bit_size) == 64;
fs_reg indirect_offset = get_indirect_offset(instr);
unsigned imm_offset = instr->const_index[0];
unsigned swiz = BRW_SWIZZLE_XYZW;
unsigned num_components = _mesa_fls(mask);
enum opcode opcode;
- if (mask != WRITEMASK_XYZW) {
- srcs[header_regs++] = brw_imm_ud(mask << 16);
- opcode = indirect_offset.file != BAD_FILE ?
- SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT :
- SHADER_OPCODE_URB_WRITE_SIMD8_MASKED;
- } else {
- opcode = indirect_offset.file != BAD_FILE ?
- SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT :
- SHADER_OPCODE_URB_WRITE_SIMD8;
+ /* We can only pack two 64-bit components in a single message, so send
+ * 2 messages if we have more components
+ */
+ unsigned num_iterations = 1;
+ unsigned iter_components = num_components;
+ if (is_64bit && instr->num_components > 2) {
+ num_iterations = 2;
+ iter_components = 2;
}
- for (unsigned i = 0; i < num_components; i++) {
- if (mask & (1 << i))
- srcs[header_regs + i] = offset(value, bld, BRW_GET_SWZ(swiz, i));
- }
+ /* 64-bit data needs to me shuffled before we can write it to the URB.
+ * We will use this temporary to shuffle the components in each
+ * iteration.
+ */
+ fs_reg tmp =
+ fs_reg(VGRF, alloc.allocate(2 * iter_components), value.type);
+
+ for (unsigned iter = 0; iter < num_iterations; iter++) {
+ if (!is_64bit && mask != WRITEMASK_XYZW) {
+ srcs[header_regs++] = brw_imm_ud(mask << 16);
+ opcode = indirect_offset.file != BAD_FILE ?
+ SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT :
+ SHADER_OPCODE_URB_WRITE_SIMD8_MASKED;
+ } else if (is_64bit && ((mask & WRITEMASK_XY) != WRITEMASK_XY)) {
+ /* Expand the 64-bit mask to 32-bit channels. We only handle
+ * two channels in each iteration, so we only care about X/Y.
+ */
+ unsigned mask32 = 0;
+ if (mask & WRITEMASK_X)
+ mask32 |= WRITEMASK_XY;
+ if (mask & WRITEMASK_Y)
+ mask32 |= WRITEMASK_ZW;
+
+ /* If the mask does not include any of the channels X or Y there
+ * is nothing to do in this iteration. Move on to the next couple
+ * of 64-bit channels.
+ */
+ if (!mask32) {
+ mask >>= 2;
+ imm_offset++;
+ continue;
+ }
- unsigned mlen = header_regs + num_components;
+ srcs[header_regs++] = brw_imm_ud(mask32 << 16);
+ opcode = indirect_offset.file != BAD_FILE ?
+ SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT :
+ SHADER_OPCODE_URB_WRITE_SIMD8_MASKED;
+ } else {
+ opcode = indirect_offset.file != BAD_FILE ?
+ SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT :
+ SHADER_OPCODE_URB_WRITE_SIMD8;
+ }
- fs_reg payload =
- bld.vgrf(BRW_REGISTER_TYPE_UD, mlen);
- bld.LOAD_PAYLOAD(payload, srcs, mlen, header_regs);
+ for (unsigned i = 0; i < iter_components; i++) {
+ if (!(mask & (1 << i)))
+ continue;
- fs_inst *inst = bld.emit(opcode, bld.null_reg_ud(), payload);
- inst->offset = imm_offset;
- inst->mlen = mlen;
- inst->base_mrf = -1;
+ if (!is_64bit) {
+ srcs[header_regs + i] = offset(value, bld, BRW_GET_SWZ(swiz, i));
+ } else {
+ /* We need to shuffle the 64-bit data to match the layout
+ * expected by our 32-bit URB write messages. We use a temporary
+ * for that.
+ */
+ unsigned channel = BRW_GET_SWZ(swiz, iter * 2 + i);
+ shuffle_64bit_data_for_32bit_write(bld,
+ retype(offset(tmp, bld, 2 * i), BRW_REGISTER_TYPE_F),
+ retype(offset(value, bld, 2 * channel), BRW_REGISTER_TYPE_DF),
+ 1);
+
+ /* Now copy the data to the destination */
+ fs_reg dest = fs_reg(VGRF, alloc.allocate(2), value.type);
+ unsigned idx = 2 * i;
+ bld.MOV(dest, offset(tmp, bld, idx));
+ bld.MOV(offset(dest, bld, 1), offset(tmp, bld, idx + 1));
+ srcs[header_regs + idx] = dest;
+ srcs[header_regs + idx + 1] = offset(dest, bld, 1);
+ }
+ }
+
+ unsigned mlen =
+ header_regs + (is_64bit ? 2 * iter_components : iter_components);
+ fs_reg payload =
+ bld.vgrf(BRW_REGISTER_TYPE_UD, mlen);
+ bld.LOAD_PAYLOAD(payload, srcs, mlen, header_regs);
+
+ fs_inst *inst = bld.emit(opcode, bld.null_reg_ud(), payload);
+ inst->offset = imm_offset;
+ inst->mlen = mlen;
+ inst->base_mrf = -1;
+
+ /* If this is a 64-bit attribute, select the next two 64-bit channels
+ * to be handled in the next iteration.
+ */
+ if (is_64bit) {
+ mask >>= 2;
+ imm_offset++;
+ }
+ }
break;
}
if (imm_offset < max_push_slots) {
fs_reg src = fs_reg(ATTR, imm_offset / 2, dest.type);
for (int i = 0; i < instr->num_components; i++) {
- bld.MOV(offset(dest, bld, i),
- component(src, 4 * (imm_offset % 2) + i));
+ unsigned comp = 16 / type_sz(dest.type) * (imm_offset % 2) + i;
+ bld.MOV(offset(dest, bld, i), component(src, comp));
}
tes_prog_data->base.urb_read_length =
MAX2(tes_prog_data->base.urb_read_length,
/* Writemask */
unsigned writemask = instr->const_index[1];
+ /* get_nir_src() retypes to integer. Be wary of 64-bit types though
+ * since the untyped writes below operate in units of 32-bits, which
+ * means that we need to write twice as many components each time.
+ * Also, we have to suffle 64-bit data to be in the appropriate layout
+ * expected by our 32-bit write messages.
+ */
+ unsigned type_size = 4;
+ unsigned bit_size = instr->src[0].is_ssa ?
+ instr->src[0].ssa->bit_size : instr->src[0].reg.reg->bit_size;
+ if (bit_size == 64) {
+ type_size = 8;
+ fs_reg tmp =
+ fs_reg(VGRF, alloc.allocate(alloc.sizes[val_reg.nr]), val_reg.type);
+ shuffle_64bit_data_for_32bit_write(
+ bld,
+ retype(tmp, BRW_REGISTER_TYPE_F),
+ retype(val_reg, BRW_REGISTER_TYPE_DF),
+ instr->num_components);
+ val_reg = tmp;
+ }
+
+ unsigned type_slots = type_size / 4;
+
/* Combine groups of consecutive enabled channels in one write
* message. We use ffs to find the first enabled channel and then ffs on
* the bit-inverse, down-shifted writemask to determine the length of
while (writemask) {
unsigned first_component = ffs(writemask) - 1;
unsigned length = ffs(~(writemask >> first_component)) - 1;
- fs_reg offset_reg;
+ /* We can't write more than 2 64-bit components at once. Limit the
+ * length of the write to what we can do and let the next iteration
+ * handle the rest
+ */
+ if (type_size > 4)
+ length = MIN2(2, length);
+
+ fs_reg offset_reg;
nir_const_value *const_offset = nir_src_as_const_value(instr->src[1]);
if (const_offset) {
offset_reg = brw_imm_ud(instr->const_index[0] + const_offset->u32[0] +
- 4 * first_component);
+ type_size * first_component);
} else {
offset_reg = vgrf(glsl_type::uint_type);
bld.ADD(offset_reg,
retype(get_nir_src(instr->src[1]), BRW_REGISTER_TYPE_UD),
- brw_imm_ud(instr->const_index[0] + 4 * first_component));
+ brw_imm_ud(instr->const_index[0] + type_size * first_component));
}
emit_untyped_write(bld, surf_index, offset_reg,
- offset(val_reg, bld, first_component),
- 1 /* dims */, length,
+ offset(val_reg, bld, first_component * type_slots),
+ 1 /* dims */, length * type_slots,
BRW_PREDICATE_NONE);
/* Clear the bits in the writemask that we just wrote, then try
for (unsigned j = 0; j < instr->num_components; j++) {
bld.MOV(offset(dest, bld, j), offset(src, bld, j));
}
+
+ if (type_sz(src.type) == 8) {
+ shuffle_32bit_load_result_to_64bit_data(bld,
+ dest,
+ retype(dest, BRW_REGISTER_TYPE_F),
+ instr->num_components);
+ }
+
break;
}
/* Writemask */
unsigned writemask = instr->const_index[0];
+ /* get_nir_src() retypes to integer. Be wary of 64-bit types though
+ * since the untyped writes below operate in units of 32-bits, which
+ * means that we need to write twice as many components each time.
+ * Also, we have to suffle 64-bit data to be in the appropriate layout
+ * expected by our 32-bit write messages.
+ */
+ unsigned type_size = 4;
+ unsigned bit_size = instr->src[0].is_ssa ?
+ instr->src[0].ssa->bit_size : instr->src[0].reg.reg->bit_size;
+ if (bit_size == 64) {
+ type_size = 8;
+ fs_reg tmp =
+ fs_reg(VGRF, alloc.allocate(alloc.sizes[val_reg.nr]), val_reg.type);
+ shuffle_64bit_data_for_32bit_write(bld,
+ retype(tmp, BRW_REGISTER_TYPE_F),
+ retype(val_reg, BRW_REGISTER_TYPE_DF),
+ instr->num_components);
+ val_reg = tmp;
+ }
+
+ unsigned type_slots = type_size / 4;
+
/* Combine groups of consecutive enabled channels in one write
* message. We use ffs to find the first enabled channel and then ffs on
* the bit-inverse, down-shifted writemask to determine the length of
unsigned first_component = ffs(writemask) - 1;
unsigned length = ffs(~(writemask >> first_component)) - 1;
+ /* We can't write more than 2 64-bit components at once. Limit the
+ * length of the write to what we can do and let the next iteration
+ * handle the rest
+ */
+ if (type_size > 4)
+ length = MIN2(2, length);
+
fs_reg offset_reg;
nir_const_value *const_offset = nir_src_as_const_value(instr->src[2]);
if (const_offset) {
- offset_reg = brw_imm_ud(const_offset->u32[0] + 4 * first_component);
+ offset_reg = brw_imm_ud(const_offset->u32[0] +
+ type_size * first_component);
} else {
offset_reg = vgrf(glsl_type::uint_type);
bld.ADD(offset_reg,
retype(get_nir_src(instr->src[2]), BRW_REGISTER_TYPE_UD),
- brw_imm_ud(4 * first_component));
+ brw_imm_ud(type_size * first_component));
}
+
emit_untyped_write(bld, surf_index, offset_reg,
- offset(val_reg, bld, first_component),
- 1 /* dims */, length,
+ offset(val_reg, bld, first_component * type_slots),
+ 1 /* dims */, length * type_slots,
BRW_PREDICATE_NONE);
/* Clear the bits in the writemask that we just wrote, then try
assert(const_offset && "Indirect output stores not allowed");
new_dest = offset(new_dest, bld, const_offset->u32[0]);
- for (unsigned j = 0; j < instr->num_components; j++) {
+ unsigned num_components = instr->num_components;
+ unsigned bit_size = instr->src[0].is_ssa ?
+ instr->src[0].ssa->bit_size : instr->src[0].reg.reg->bit_size;
+ if (bit_size == 64) {
+ fs_reg tmp =
+ fs_reg(VGRF, alloc.allocate(2 * num_components),
+ BRW_REGISTER_TYPE_F);
+ shuffle_64bit_data_for_32bit_write(
+ bld, tmp, retype(src, BRW_REGISTER_TYPE_DF), num_components);
+ src = retype(tmp, src.type);
+ num_components *= 2;
+ }
+
+ for (unsigned j = 0; j < num_components; j++) {
bld.MOV(offset(new_dest, bld, j), offset(src, bld, j));
}
break;