summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--src/intel/compiler/brw_fs.h5
-rw-r--r--src/intel/compiler/brw_fs_nir.cpp497
2 files changed, 139 insertions, 363 deletions
diff --git a/src/intel/compiler/brw_fs.h b/src/intel/compiler/brw_fs.h
index 020d42356eb..1dde4c9c7d2 100644
--- a/src/intel/compiler/brw_fs.h
+++ b/src/intel/compiler/brw_fs.h
@@ -572,11 +572,6 @@ void shuffle_from_32bit_read(const brw::fs_builder &bld,
uint32_t first_component,
uint32_t components);
-fs_reg shuffle_for_32bit_write(const brw::fs_builder &bld,
- const fs_reg &src,
- uint32_t first_component,
- uint32_t components);
-
fs_reg setup_imm_df(const brw::fs_builder &bld,
double v);
diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp
index ee3634b7b99..b56d7156954 100644
--- a/src/intel/compiler/brw_fs_nir.cpp
+++ b/src/intel/compiler/brw_fs_nir.cpp
@@ -2352,13 +2352,12 @@ fs_visitor::emit_gs_input_load(const fs_reg &dst,
unsigned num_components,
unsigned first_component)
{
+ assert(type_sz(dst.type) == 4);
struct brw_gs_prog_data *gs_prog_data = brw_gs_prog_data(prog_data);
const unsigned push_reg_count = gs_prog_data->base.urb_read_length * 8;
/* TODO: figure out push input layout for invocations == 1 */
- /* TODO: make this work with 64-bit inputs */
if (gs_prog_data->invocations == 1 &&
- type_sz(dst.type) <= 4 &&
nir_src_is_const(offset_src) && nir_src_is_const(vertex_src) &&
4 * (base_offset + nir_src_as_uint(offset_src)) < push_reg_count) {
int imm_offset = (base_offset + nir_src_as_uint(offset_src)) * 4 +
@@ -2452,87 +2451,50 @@ fs_visitor::emit_gs_input_load(const fs_reg &dst,
}
fs_inst *inst;
-
- fs_reg tmp_dst = dst;
fs_reg indirect_offset = get_nir_src(offset_src);
- unsigned num_iterations = 1;
- unsigned orig_num_components = num_components;
-
- if (type_sz(dst.type) == 8) {
- if (num_components > 2) {
- num_iterations = 2;
- num_components = 2;
- }
- fs_reg tmp = fs_reg(VGRF, alloc.allocate(4), dst.type);
- tmp_dst = tmp;
- first_component = first_component / 2;
- }
- for (unsigned iter = 0; iter < num_iterations; iter++) {
- if (nir_src_is_const(offset_src)) {
- /* Constant indexing - use global offset. */
- if (first_component != 0) {
- unsigned read_components = num_components + first_component;
- fs_reg tmp = bld.vgrf(dst.type, read_components);
- inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, tmp, icp_handle);
- inst->size_written = read_components *
- tmp.component_size(inst->exec_size);
- for (unsigned i = 0; i < num_components; i++) {
- bld.MOV(offset(tmp_dst, bld, i),
- offset(tmp, bld, i + first_component));
- }
- } else {
- inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, tmp_dst,
- icp_handle);
- inst->size_written = num_components *
- tmp_dst.component_size(inst->exec_size);
- }
- inst->offset = base_offset + nir_src_as_uint(offset_src);
- inst->mlen = 1;
- } else {
- /* Indirect indexing - use per-slot offsets as well. */
- const fs_reg srcs[] = { icp_handle, indirect_offset };
+ if (nir_src_is_const(offset_src)) {
+ /* Constant indexing - use global offset. */
+ if (first_component != 0) {
unsigned read_components = num_components + first_component;
fs_reg tmp = bld.vgrf(dst.type, read_components);
- fs_reg payload = bld.vgrf(BRW_REGISTER_TYPE_UD, 2);
- bld.LOAD_PAYLOAD(payload, srcs, ARRAY_SIZE(srcs), 0);
- if (first_component != 0) {
- inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, tmp,
- payload);
- inst->size_written = read_components *
- tmp.component_size(inst->exec_size);
- for (unsigned i = 0; i < num_components; i++) {
- bld.MOV(offset(tmp_dst, bld, i),
- offset(tmp, bld, i + first_component));
- }
- } else {
- inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, tmp_dst,
- payload);
- inst->size_written = num_components *
- tmp_dst.component_size(inst->exec_size);
+ inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, tmp, icp_handle);
+ inst->size_written = read_components *
+ tmp.component_size(inst->exec_size);
+ for (unsigned i = 0; i < num_components; i++) {
+ bld.MOV(offset(dst, bld, i),
+ offset(tmp, bld, i + first_component));
}
- inst->offset = base_offset;
- inst->mlen = 2;
- }
-
- if (type_sz(dst.type) == 8) {
- shuffle_from_32bit_read(bld,
- offset(dst, bld, iter * 2),
- retype(tmp_dst, BRW_REGISTER_TYPE_D),
- 0,
- num_components);
+ } else {
+ inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, dst, icp_handle);
+ inst->size_written = num_components *
+ dst.component_size(inst->exec_size);
}
-
- if (num_iterations > 1) {
- num_components = orig_num_components - 2;
- if(nir_src_is_const(offset_src)) {
- base_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;
+ inst->offset = base_offset + nir_src_as_uint(offset_src);
+ inst->mlen = 1;
+ } else {
+ /* Indirect indexing - use per-slot offsets as well. */
+ const fs_reg srcs[] = { icp_handle, indirect_offset };
+ unsigned read_components = num_components + first_component;
+ fs_reg tmp = bld.vgrf(dst.type, read_components);
+ fs_reg payload = bld.vgrf(BRW_REGISTER_TYPE_UD, 2);
+ bld.LOAD_PAYLOAD(payload, srcs, ARRAY_SIZE(srcs), 0);
+ if (first_component != 0) {
+ inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, tmp,
+ payload);
+ inst->size_written = read_components *
+ tmp.component_size(inst->exec_size);
+ for (unsigned i = 0; i < num_components; i++) {
+ bld.MOV(offset(dst, bld, i),
+ offset(tmp, bld, i + first_component));
}
+ } else {
+ inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, dst, payload);
+ inst->size_written = num_components *
+ dst.component_size(inst->exec_size);
}
+ inst->offset = base_offset;
+ inst->mlen = 2;
}
}
@@ -2569,20 +2531,13 @@ fs_visitor::nir_emit_vs_intrinsic(const fs_builder &bld,
unreachable("should be lowered by nir_lower_system_values()");
case nir_intrinsic_load_input: {
+ assert(nir_dest_bit_size(instr->dest) == 32);
fs_reg src = fs_reg(ATTR, nir_intrinsic_base(instr) * 4, dest.type);
- unsigned first_component = nir_intrinsic_component(instr);
- unsigned num_components = instr->num_components;
-
+ src = offset(src, bld, nir_intrinsic_component(instr));
src = offset(src, bld, nir_src_as_uint(instr->src[0]));
- if (type_sz(dest.type) == 8)
- first_component /= 2;
-
- /* For 16-bit support maybe a temporary will be needed to copy from
- * the ATTR file.
- */
- shuffle_from_32bit_read(bld, dest, retype(src, BRW_REGISTER_TYPE_D),
- first_component, num_components);
+ for (unsigned i = 0; i < instr->num_components; i++)
+ bld.MOV(offset(dest, bld, i), offset(src, bld, i));
break;
}
@@ -2781,6 +2736,7 @@ fs_visitor::nir_emit_tcs_intrinsic(const fs_builder &bld,
break;
case nir_intrinsic_load_per_vertex_input: {
+ assert(nir_dest_bit_size(instr->dest) == 32);
fs_reg indirect_offset = get_indirect_offset(instr);
unsigned imm_offset = instr->const_index[0];
fs_inst *inst;
@@ -2793,97 +2749,64 @@ fs_visitor::nir_emit_tcs_intrinsic(const fs_builder &bld,
* 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;
unsigned first_component = nir_intrinsic_component(instr);
- fs_reg orig_dst = dst;
- if (type_sz(dst.type) == 8) {
- first_component = first_component / 2;
- if (instr->num_components > 2) {
- num_iterations = 2;
- num_components = 2;
- }
- fs_reg tmp = fs_reg(VGRF, alloc.allocate(4), dst.type);
- dst = tmp;
- }
-
- for (unsigned iter = 0; iter < num_iterations; iter++) {
- if (indirect_offset.file == BAD_FILE) {
- /* Constant indexing - use global offset. */
- if (first_component != 0) {
- unsigned read_components = num_components + first_component;
- fs_reg tmp = bld.vgrf(dst.type, read_components);
- inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, tmp, icp_handle);
- for (unsigned i = 0; i < num_components; i++) {
- bld.MOV(offset(dst, bld, i),
- offset(tmp, bld, i + first_component));
- }
- } else {
- inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, dst, icp_handle);
+ if (indirect_offset.file == BAD_FILE) {
+ /* Constant indexing - use global offset. */
+ if (first_component != 0) {
+ unsigned read_components = num_components + first_component;
+ fs_reg tmp = bld.vgrf(dst.type, read_components);
+ inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, tmp, icp_handle);
+ for (unsigned i = 0; i < num_components; i++) {
+ bld.MOV(offset(dst, bld, i),
+ offset(tmp, bld, i + first_component));
}
- inst->offset = imm_offset;
- inst->mlen = 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);
- if (first_component != 0) {
- unsigned read_components = num_components + first_component;
- fs_reg tmp = bld.vgrf(dst.type, read_components);
- inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, tmp,
- payload);
- for (unsigned i = 0; i < num_components; i++) {
- bld.MOV(offset(dst, bld, i),
- offset(tmp, bld, i + first_component));
- }
- } else {
- inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, dst,
- payload);
- }
- inst->offset = imm_offset;
- inst->mlen = 2;
+ inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, dst, icp_handle);
}
- inst->size_written = (num_components + first_component) *
- inst->dst.component_size(inst->exec_size);
-
- /* 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_from_32bit_read(bld,
- offset(orig_dst, bld, iter * 2),
- retype(dst, BRW_REGISTER_TYPE_D),
- 0, num_components);
- }
-
- /* 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->size_written = 4 * REG_SIZE;
- bld.MOV(dst, offset(inst->dst, bld, 3));
+ inst->offset = imm_offset;
+ inst->mlen = 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);
+ if (first_component != 0) {
+ unsigned read_components = num_components + first_component;
+ fs_reg tmp = bld.vgrf(dst.type, read_components);
+ inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, tmp,
+ payload);
+ for (unsigned i = 0; i < num_components; i++) {
+ bld.MOV(offset(dst, bld, i),
+ offset(tmp, bld, i + first_component));
+ }
+ } else {
+ inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, dst,
+ payload);
}
+ inst->offset = imm_offset;
+ inst->mlen = 2;
+ }
+ inst->size_written = (num_components + first_component) *
+ inst->dst.component_size(inst->exec_size);
- /* 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;
- imm_offset++;
- }
+ /* 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) == 4);
+ inst->dst = bld.vgrf(dst.type, 4);
+ inst->size_written = 4 * REG_SIZE;
+ bld.MOV(dst, offset(inst->dst, bld, 3));
}
break;
}
case nir_intrinsic_load_output:
case nir_intrinsic_load_per_vertex_output: {
+ assert(nir_dest_bit_size(instr->dest) == 32);
fs_reg indirect_offset = get_indirect_offset(instr);
unsigned imm_offset = instr->const_index[0];
unsigned first_component = nir_intrinsic_component(instr);
@@ -2947,9 +2870,8 @@ fs_visitor::nir_emit_tcs_intrinsic(const fs_builder &bld,
case nir_intrinsic_store_output:
case nir_intrinsic_store_per_vertex_output: {
+ assert(nir_src_bit_size(instr->src[0]) == 32);
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 mask = instr->const_index[1];
@@ -2972,94 +2894,35 @@ fs_visitor::nir_emit_tcs_intrinsic(const fs_builder &bld,
/* 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;
unsigned first_component = nir_intrinsic_component(instr);
- if (is_64bit) {
- first_component = first_component / 2;
- if (instr->num_components > 2) {
- num_iterations = 2;
- iter_components = 2;
- }
- }
-
mask = mask << first_component;
- 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;
- }
-
- 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;
- }
-
- for (unsigned i = 0; i < iter_components; i++) {
- if (!(mask & (1 << (i + first_component))))
- continue;
-
- if (!is_64bit) {
- srcs[header_regs + i + first_component] = offset(value, bld, 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 = iter * 2 + i;
- fs_reg dest = shuffle_for_32bit_write(bld, value, channel, 1);
+ 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;
+ }
- srcs[header_regs + (i + first_component) * 2] = dest;
- srcs[header_regs + (i + first_component) * 2 + 1] =
- offset(dest, bld, 1);
- }
- }
+ for (unsigned i = 0; i < num_components; i++) {
+ if (!(mask & (1 << (i + first_component))))
+ continue;
- unsigned mlen =
- header_regs + (is_64bit ? 2 * iter_components : iter_components) +
- (is_64bit ? 2 * first_component : first_component);
- fs_reg payload =
- bld.vgrf(BRW_REGISTER_TYPE_UD, mlen);
- bld.LOAD_PAYLOAD(payload, srcs, mlen, header_regs);
+ srcs[header_regs + i + first_component] = offset(value, bld, i);
+ }
- fs_inst *inst = bld.emit(opcode, bld.null_reg_ud(), payload);
- inst->offset = imm_offset;
- inst->mlen = mlen;
+ unsigned mlen = header_regs + num_components + first_component;
+ fs_reg payload =
+ bld.vgrf(BRW_REGISTER_TYPE_UD, mlen);
+ bld.LOAD_PAYLOAD(payload, srcs, mlen, header_regs);
- /* 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++;
- }
- }
+ fs_inst *inst = bld.emit(opcode, bld.null_reg_ud(), payload);
+ inst->offset = imm_offset;
+ inst->mlen = mlen;
break;
}
@@ -3093,35 +2956,27 @@ fs_visitor::nir_emit_tes_intrinsic(const fs_builder &bld,
case nir_intrinsic_load_input:
case nir_intrinsic_load_per_vertex_input: {
+ assert(nir_dest_bit_size(instr->dest) == 32);
fs_reg indirect_offset = get_indirect_offset(instr);
unsigned imm_offset = instr->const_index[0];
unsigned first_component = nir_intrinsic_component(instr);
- if (type_sz(dest.type) == 8) {
- first_component = first_component / 2;
- }
-
fs_inst *inst;
if (indirect_offset.file == BAD_FILE) {
/* Arbitrarily only push up to 32 vec4 slots worth of data,
* which is 16 registers (since each holds 2 vec4 slots).
*/
- unsigned slot_count = 1;
- if (type_sz(dest.type) == 8 && instr->num_components > 2)
- slot_count++;
-
const unsigned max_push_slots = 32;
- if (imm_offset + slot_count <= max_push_slots) {
+ 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++) {
- unsigned comp = 16 / type_sz(dest.type) * (imm_offset % 2) +
- i + first_component;
+ unsigned comp = 4 * (imm_offset % 2) + i + first_component;
bld.MOV(offset(dest, bld, i), component(src, comp));
}
tes_prog_data->base.urb_read_length =
MAX2(tes_prog_data->base.urb_read_length,
- DIV_ROUND_UP(imm_offset + slot_count, 2));
+ (imm_offset / 2) + 1);
} else {
/* Replicate the patch handle to all enabled channels */
const fs_reg srcs[] = {
@@ -3156,65 +3011,32 @@ fs_visitor::nir_emit_tes_intrinsic(const fs_builder &bld,
* 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_dest = dest;
- if (type_sz(dest.type) == 8) {
- if (instr->num_components > 2) {
- num_iterations = 2;
- num_components = 2;
- }
- fs_reg tmp = fs_reg(VGRF, alloc.allocate(4), dest.type);
- dest = tmp;
- }
-
- for (unsigned iter = 0; iter < num_iterations; iter++) {
- const fs_reg srcs[] = {
- retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD),
- indirect_offset
- };
- fs_reg payload = bld.vgrf(BRW_REGISTER_TYPE_UD, 2);
- bld.LOAD_PAYLOAD(payload, srcs, ARRAY_SIZE(srcs), 0);
-
- if (first_component != 0) {
- unsigned read_components =
- num_components + first_component;
- fs_reg tmp = bld.vgrf(dest.type, read_components);
- inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, tmp,
- payload);
- for (unsigned i = 0; i < num_components; i++) {
- bld.MOV(offset(dest, bld, i),
- offset(tmp, bld, i + first_component));
- }
- } else {
- inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, dest,
- payload);
- }
- inst->mlen = 2;
- inst->offset = imm_offset;
- inst->size_written = (num_components + first_component) *
- inst->dst.component_size(inst->exec_size);
-
- /* 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(dest.type) == 8) {
- shuffle_from_32bit_read(bld,
- offset(orig_dest, bld, iter * 2),
- retype(dest, BRW_REGISTER_TYPE_D),
- 0, num_components);
- }
+ const fs_reg srcs[] = {
+ retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD),
+ indirect_offset
+ };
+ fs_reg payload = bld.vgrf(BRW_REGISTER_TYPE_UD, 2);
+ bld.LOAD_PAYLOAD(payload, srcs, ARRAY_SIZE(srcs), 0);
- /* If we are loading double data and we need a second read message
- * adjust the offset
- */
- if (num_iterations > 1) {
- num_components = instr->num_components - 2;
- imm_offset++;
+ if (first_component != 0) {
+ unsigned read_components =
+ num_components + first_component;
+ fs_reg tmp = bld.vgrf(dest.type, read_components);
+ inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, tmp,
+ payload);
+ for (unsigned i = 0; i < num_components; i++) {
+ bld.MOV(offset(dest, bld, i),
+ offset(tmp, bld, i + first_component));
}
+ } else {
+ inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, dest,
+ payload);
}
+ inst->mlen = 2;
+ inst->offset = imm_offset;
+ inst->size_written = (num_components + first_component) *
+ inst->dst.component_size(inst->exec_size);
}
break;
}
@@ -3641,11 +3463,10 @@ fs_visitor::nir_emit_fs_intrinsic(const fs_builder &bld,
case nir_intrinsic_load_input: {
/* load_input is only used for flat inputs */
+ assert(nir_dest_bit_size(instr->dest) == 32);
unsigned base = nir_intrinsic_base(instr);
unsigned comp = nir_intrinsic_component(instr);
unsigned num_components = instr->num_components;
- fs_reg orig_dest = dest;
- enum brw_reg_type type = dest.type;
/* Special case fields in the VUE header */
if (base == VARYING_SLOT_LAYER)
@@ -3653,24 +3474,9 @@ fs_visitor::nir_emit_fs_intrinsic(const fs_builder &bld,
else if (base == VARYING_SLOT_VIEWPORT)
comp = 2;
- if (nir_dest_bit_size(instr->dest) == 64) {
- /* const_index is in 32-bit type size units that could not be aligned
- * with DF. We need to read the double vector as if it was a float
- * vector of twice the number of components to fetch the right data.
- */
- type = BRW_REGISTER_TYPE_F;
- num_components *= 2;
- dest = bld.vgrf(type, num_components);
- }
-
for (unsigned int i = 0; i < num_components; i++) {
- bld.MOV(offset(retype(dest, type), bld, i),
- retype(component(interp_reg(base, comp + i), 3), type));
- }
-
- if (nir_dest_bit_size(instr->dest) == 64) {
- shuffle_from_32bit_read(bld, orig_dest, dest, 0,
- instr->num_components);
+ bld.MOV(offset(dest, bld, i),
+ retype(component(interp_reg(base, comp + i), 3), dest.type));
}
break;
}
@@ -4799,15 +4605,12 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
}
case nir_intrinsic_store_output: {
+ assert(nir_src_bit_size(instr->src[0]) == 32);
fs_reg src = get_nir_src(instr->src[0]);
unsigned store_offset = nir_src_as_uint(instr->src[1]);
unsigned num_components = instr->num_components;
unsigned first_component = nir_intrinsic_component(instr);
- if (nir_src_bit_size(instr->src[0]) == 64) {
- src = shuffle_for_32bit_write(bld, src, 0, num_components);
- num_components *= 2;
- }
fs_reg new_dest = retype(offset(outputs[instr->const_index[0]], bld,
4 * store_offset), src.type);
@@ -5928,28 +5731,6 @@ shuffle_from_32bit_read(const fs_builder &bld,
}
fs_reg
-shuffle_for_32bit_write(const fs_builder &bld,
- const fs_reg &src,
- uint32_t first_component,
- uint32_t components)
-{
- fs_reg dst = bld.vgrf(BRW_REGISTER_TYPE_D,
- DIV_ROUND_UP (components * type_sz(src.type), 4));
- /* This function takes components in units of the source type while
- * shuffle_src_to_dst takes components in units of the smallest type
- */
- if (type_sz(src.type) > 4) {
- assert(type_sz(src.type) == 8);
- first_component *= 2;
- components *= 2;
- }
-
- shuffle_src_to_dst(bld, dst, src, first_component, components);
-
- return dst;
-}
-
-fs_reg
setup_imm_df(const fs_builder &bld, double v)
{
const struct gen_device_info *devinfo = bld.shader->devinfo;
generated by cgit v1.2.3 (git 2.39.1) at 2024-12-21 15:11:44 +0000