mesa/src/compiler/nir/nir_opt_find_array_copies.c

616 lines
20 KiB
C
Raw Normal View History

/*
* Copyright © 2018 Intel Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
#include "nir.h"
#include "nir_builder.h"
#include "nir_deref.h"
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
struct match_node {
/* Note: these fields are only valid for leaf nodes */
unsigned next_array_idx;
int src_wildcard_idx;
nir_deref_path first_src_path;
/* The index of the first read of the source path that's part of the copy
* we're matching. If the last write to the source path is after this, we
* would get a different result from reading it at the end and we can't
* emit the copy.
*/
unsigned first_src_read;
/* The last time there was a write to this node. */
unsigned last_overwritten;
/* The last time there was a write to this node which successfully advanced
* next_array_idx. This helps us catch any intervening aliased writes.
*/
unsigned last_successful_write;
unsigned num_children;
struct match_node *children[];
};
struct match_state {
/* Map from nir_variable * -> match_node */
struct hash_table *table;
unsigned cur_instr;
nir_builder builder;
void *dead_ctx;
};
static struct match_node *
create_match_node(const struct glsl_type *type, struct match_state *state)
{
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
unsigned num_children = 0;
if (glsl_type_is_array_or_matrix(type)) {
/* One for wildcards */
num_children = glsl_get_length(type) + 1;
} else if (glsl_type_is_struct_or_ifc(type)) {
num_children = glsl_get_length(type);
}
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
struct match_node *node = rzalloc_size(state->dead_ctx,
sizeof(struct match_node) +
num_children * sizeof(struct match_node *));
node->num_children = num_children;
node->src_wildcard_idx = -1;
node->first_src_read = UINT32_MAX;
return node;
}
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
static struct match_node *
node_for_deref(nir_deref_instr *instr, struct match_node *parent,
struct match_state *state)
{
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
unsigned idx;
switch (instr->deref_type) {
case nir_deref_type_var: {
struct hash_entry *entry = _mesa_hash_table_search(state->table, instr->var);
if (entry) {
return entry->data;
} else {
struct match_node *node = create_match_node(instr->type, state);
_mesa_hash_table_insert(state->table, instr->var, node);
return node;
}
}
case nir_deref_type_array_wildcard:
idx = parent->num_children - 1;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
break;
case nir_deref_type_array:
if (nir_src_is_const(instr->arr.index)) {
idx = nir_src_as_uint(instr->arr.index);
assert(idx < parent->num_children - 1);
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
} else {
idx = parent->num_children - 1;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
}
break;
case nir_deref_type_struct:
idx = instr->strct.index;
break;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
default:
unreachable("bad deref type");
}
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
assert(idx < parent->num_children);
if (parent->children[idx]) {
return parent->children[idx];
} else {
struct match_node *node = create_match_node(instr->type, state);
parent->children[idx] = node;
return node;
}
}
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
static struct match_node *
node_for_wildcard(const struct glsl_type *type, struct match_node *parent,
struct match_state *state)
{
assert(glsl_type_is_array_or_matrix(type));
unsigned idx = glsl_get_length(type);
if (parent->children[idx]) {
return parent->children[idx];
} else {
struct match_node *node =
create_match_node(glsl_get_array_element(type), state);
parent->children[idx] = node;
return node;
}
}
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
static struct match_node *
node_for_path(nir_deref_path *path, struct match_state *state)
{
struct match_node *node = NULL;
for (nir_deref_instr **instr = path->path; *instr; instr++)
node = node_for_deref(*instr, node, state);
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
return node;
}
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
static struct match_node *
node_for_path_with_wildcard(nir_deref_path *path, unsigned wildcard_idx,
struct match_state *state)
{
struct match_node *node = NULL;
unsigned idx = 0;
for (nir_deref_instr **instr = path->path; *instr; instr++, idx++) {
if (idx == wildcard_idx)
node = node_for_wildcard((*(instr - 1))->type, node, state);
else
node = node_for_deref(*instr, node, state);
}
return node;
}
typedef void (*match_cb)(struct match_node *, struct match_state *);
static void
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
_foreach_aliasing(nir_deref_instr **deref, match_cb cb,
struct match_node *node, struct match_state *state)
{
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
if (*deref == NULL) {
cb(node, state);
return;
}
switch ((*deref)->deref_type) {
case nir_deref_type_struct: {
struct match_node *child = node->children[(*deref)->strct.index];
if (child)
_foreach_aliasing(deref + 1, cb, child, state);
return;
}
case nir_deref_type_array:
case nir_deref_type_array_wildcard: {
if ((*deref)->deref_type == nir_deref_type_array_wildcard ||
!nir_src_is_const((*deref)->arr.index)) {
/* This access may touch any index, so we have to visit all of
* them.
*/
for (unsigned i = 0; i < node->num_children; i++) {
if (node->children[i])
_foreach_aliasing(deref + 1, cb, node->children[i], state);
}
} else {
/* Visit the wildcard entry if any */
if (node->children[node->num_children - 1]) {
_foreach_aliasing(deref + 1, cb,
node->children[node->num_children - 1], state);
}
unsigned index = nir_src_as_uint((*deref)->arr.index);
/* Check that the index is in-bounds */
if (index < node->num_children - 1 && node->children[index])
_foreach_aliasing(deref + 1, cb, node->children[index], state);
}
return;
}
default:
unreachable("bad deref type");
}
}
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
/* Given a deref path, find all the leaf deref nodes that alias it. */
static void
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
foreach_aliasing_node(nir_deref_path *path,
match_cb cb,
struct match_state *state)
{
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
assert(path->path[0]->deref_type == nir_deref_type_var);
struct hash_entry *entry = _mesa_hash_table_search(state->table,
path->path[0]->var);
if (entry)
_foreach_aliasing(&path->path[1], cb, entry->data, state);
}
static nir_deref_instr *
build_wildcard_deref(nir_builder *b, nir_deref_path *path,
unsigned wildcard_idx)
{
assert(path->path[wildcard_idx]->deref_type == nir_deref_type_array);
nir_deref_instr *tail =
nir_build_deref_array_wildcard(b, path->path[wildcard_idx - 1]);
for (unsigned i = wildcard_idx + 1; path->path[i]; i++)
tail = nir_build_deref_follower(b, tail, path->path[i]);
return tail;
}
static void
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
clobber(struct match_node *node, struct match_state *state)
{
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
node->last_overwritten = state->cur_instr;
}
static bool
try_match_deref(nir_deref_path *base_path, int *path_array_idx,
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
nir_deref_path *deref_path, int arr_idx)
{
for (int i = 0; ; i++) {
nir_deref_instr *b = base_path->path[i];
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
nir_deref_instr *d = deref_path->path[i];
/* They have to be the same length */
if ((b == NULL) != (d == NULL))
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
return false;
if (b == NULL)
break;
/* This can happen if one is a deref_array and the other a wildcard */
if (b->deref_type != d->deref_type)
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
return false;;
switch (b->deref_type) {
case nir_deref_type_var:
if (b->var != d->var)
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
return false;
continue;
case nir_deref_type_array:
assert(b->arr.index.is_ssa && d->arr.index.is_ssa);
const bool const_b_idx = nir_src_is_const(b->arr.index);
const bool const_d_idx = nir_src_is_const(d->arr.index);
const unsigned b_idx = const_b_idx ? nir_src_as_uint(b->arr.index) : 0;
const unsigned d_idx = const_d_idx ? nir_src_as_uint(d->arr.index) : 0;
/* If we don't have an index into the path yet or if this entry in
* the path is at the array index, see if this is a candidate. We're
* looking for an index which is zero in the base deref and arr_idx
* in the search deref.
*/
if ((*path_array_idx < 0 || *path_array_idx == i) &&
const_b_idx && b_idx == 0 &&
const_d_idx && d_idx == arr_idx) {
*path_array_idx = i;
continue;
}
/* We're at the array index but not a candidate */
if (*path_array_idx == i)
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
return false;
/* If we're not the path array index, we must match exactly. We
* could probably just compare SSA values and trust in copy
* propagation but doing it ourselves means this pass can run a bit
* earlier.
*/
if (b->arr.index.ssa == d->arr.index.ssa ||
(const_b_idx && const_d_idx && b_idx == d_idx))
continue;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
return false;
case nir_deref_type_array_wildcard:
continue;
case nir_deref_type_struct:
if (b->strct.index != d->strct.index)
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
return false;
continue;
default:
unreachable("Invalid deref type in a path");
}
}
/* If we got here without failing, we've matched. However, it isn't an
* array match unless we found an altered array index.
*/
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
return *path_array_idx > 0;
}
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
static void
handle_read(nir_deref_instr *src, struct match_state *state)
{
/* We only need to create an entry for sources that might be used to form
* an array copy. Hence no indirects or indexing into a vector.
*/
if (nir_deref_instr_has_indirect(src) ||
nir_deref_instr_is_known_out_of_bounds(src) ||
(src->deref_type == nir_deref_type_array &&
glsl_type_is_vector(nir_src_as_deref(src->parent)->type)))
return;
nir_deref_path src_path;
nir_deref_path_init(&src_path, src, state->dead_ctx);
/* Create a node for this source if it doesn't exist. The point of this is
* to know which nodes aliasing a given store we actually need to care
* about, to avoid creating an excessive amount of nodes.
*/
node_for_path(&src_path, state);
}
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
/* The core implementation, which is used for both copies and writes. Return
* true if a copy is created.
*/
static bool
handle_write(nir_deref_instr *dst, nir_deref_instr *src,
unsigned write_index, unsigned read_index,
struct match_state *state)
{
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
nir_builder *b = &state->builder;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
nir_deref_path dst_path;
nir_deref_path_init(&dst_path, dst, state->dead_ctx);
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
unsigned idx = 0;
for (nir_deref_instr **instr = dst_path.path; *instr; instr++, idx++) {
if ((*instr)->deref_type != nir_deref_type_array)
continue;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
/* Get the entry where the index is replaced by a wildcard, so that we
* hopefully can keep matching an array copy.
*/
struct match_node *dst_node =
node_for_path_with_wildcard(&dst_path, idx, state);
if (!src)
goto reset;
if (nir_src_as_uint((*instr)->arr.index) != dst_node->next_array_idx)
goto reset;
if (dst_node->next_array_idx == 0) {
/* At this point there may be multiple source indices which are zero,
* so we can't pin down the actual source index. Just store it and
* move on.
*/
nir_deref_path_init(&dst_node->first_src_path, src, state->dead_ctx);
} else {
nir_deref_path src_path;
nir_deref_path_init(&src_path, src, state->dead_ctx);
bool result = try_match_deref(&dst_node->first_src_path,
&dst_node->src_wildcard_idx,
&src_path, dst_node->next_array_idx);
nir_deref_path_finish(&src_path);
if (!result)
goto reset;
}
/* Check if an aliasing write clobbered the array after the last normal
* write. For example, with a sequence like this:
*
* dst[0][*] = src[0][*];
* dst[0][0] = 0; // invalidates the array copy dst[*][*] = src[*][*]
* dst[1][*] = src[1][*];
*
* Note that the second write wouldn't reset the entry for dst[*][*]
* by itself, but it'll be caught by this check when processing the
* third copy.
*/
if (dst_node->last_successful_write < dst_node->last_overwritten)
goto reset;
dst_node->last_successful_write = write_index;
/* In this case we've successfully processed an array element. Check if
* this is the last, so that we can emit an array copy.
*/
dst_node->next_array_idx++;
dst_node->first_src_read = MIN2(dst_node->first_src_read, read_index);
if (dst_node->next_array_idx > 1 &&
dst_node->next_array_idx == glsl_get_length((*(instr - 1))->type)) {
/* Make sure that nothing was overwritten. */
struct match_node *src_node =
node_for_path_with_wildcard(&dst_node->first_src_path,
dst_node->src_wildcard_idx,
state);
if (src_node->last_overwritten <= dst_node->first_src_read) {
nir_copy_deref(b, build_wildcard_deref(b, &dst_path, idx),
build_wildcard_deref(b, &dst_node->first_src_path,
dst_node->src_wildcard_idx));
foreach_aliasing_node(&dst_path, clobber, state);
return true;
}
} else {
continue;
}
reset:
dst_node->next_array_idx = 0;
dst_node->src_wildcard_idx = -1;
dst_node->last_successful_write = 0;
dst_node->first_src_read = UINT32_MAX;
}
/* Mark everything aliasing dst_path as clobbered. This needs to happen
* last since in the loop above we need to know what last clobbered
* dst_node and this overwrites that.
*/
foreach_aliasing_node(&dst_path, clobber, state);
return false;
}
static bool
opt_find_array_copies_block(nir_builder *b, nir_block *block,
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
struct match_state *state)
{
bool progress = false;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
unsigned next_index = 0;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
_mesa_hash_table_clear(state->table, NULL);
nir_foreach_instr(instr, block) {
if (instr->type != nir_instr_type_intrinsic)
continue;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
/* Index the instructions before we do anything else. */
instr->index = next_index++;
/* Save the index of this instruction */
state->cur_instr = instr->index;
nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
if (intrin->intrinsic == nir_intrinsic_load_deref) {
handle_read(nir_src_as_deref(intrin->src[0]), state);
continue;
}
if (intrin->intrinsic != nir_intrinsic_copy_deref &&
intrin->intrinsic != nir_intrinsic_store_deref)
continue;
nir_deref_instr *dst_deref = nir_src_as_deref(intrin->src[0]);
/* The destination must be local. If we see a non-local store, we
* continue on because it won't affect local stores or read-only
* variables.
*/
if (dst_deref->mode != nir_var_function_temp)
continue;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
/* If there are any known out-of-bounds writes, then we can just skip
* this write as it's undefined and won't contribute to building up an
* array copy anyways.
*/
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
if (nir_deref_instr_is_known_out_of_bounds(dst_deref))
continue;
nir_deref_instr *src_deref;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
unsigned load_index = 0;
if (intrin->intrinsic == nir_intrinsic_copy_deref) {
src_deref = nir_src_as_deref(intrin->src[1]);
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
load_index = intrin->instr.index;
} else {
assert(intrin->intrinsic == nir_intrinsic_store_deref);
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
nir_intrinsic_instr *load = nir_src_as_intrinsic(intrin->src[1]);
if (load == NULL || load->intrinsic != nir_intrinsic_load_deref) {
src_deref = NULL;
} else {
src_deref = nir_src_as_deref(load->src[0]);
load_index = load->instr.index;
}
if (nir_intrinsic_write_mask(intrin) !=
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
(1 << glsl_get_components(dst_deref->type)) - 1) {
src_deref = NULL;
}
}
/* The source must be either local or something that's guaranteed to be
* read-only.
*/
const nir_variable_mode read_only_modes =
nir_var_shader_in | nir_var_uniform | nir_var_system_value;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
if (src_deref &&
!(src_deref->mode & (nir_var_function_temp | read_only_modes))) {
src_deref = NULL;
}
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
/* There must be no indirects in the source or destination and no known
* out-of-bounds accesses in the source, and the copy must be fully
* qualified, or else we can't build up the array copy. We handled
* out-of-bounds accesses to the dest above. The types must match, since
* copy_deref currently can't bitcast mismatched deref types.
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
*/
if (src_deref &&
(nir_deref_instr_has_indirect(src_deref) ||
nir_deref_instr_is_known_out_of_bounds(src_deref) ||
nir_deref_instr_has_indirect(dst_deref) ||
!glsl_type_is_vector_or_scalar(src_deref->type) ||
glsl_get_bare_type(src_deref->type) !=
glsl_get_bare_type(dst_deref->type))) {
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
src_deref = NULL;
}
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
state->builder.cursor = nir_after_instr(instr);
progress |= handle_write(dst_deref, src_deref, instr->index,
load_index, state);
}
return progress;
}
static bool
opt_find_array_copies_impl(nir_function_impl *impl)
{
nir_builder b;
nir_builder_init(&b, impl);
bool progress = false;
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
struct match_state s;
s.dead_ctx = ralloc_context(NULL);
s.table = _mesa_pointer_hash_table_create(s.dead_ctx);
nir_builder_init(&s.builder, impl);
nir_foreach_block(block, impl) {
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
if (opt_find_array_copies_block(&b, block, &s))
progress = true;
}
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
ralloc_free(s.dead_ctx);
if (progress) {
nir_metadata_preserve(impl, nir_metadata_block_index |
nir_metadata_dominance);
}
return progress;
}
/**
* This peephole optimization looks for a series of load/store_deref or
* copy_deref instructions that copy an array from one variable to another and
* turns it into a copy_deref that copies the entire array. The pattern it
* looks for is extremely specific but it's good enough to pick up on the
* input array copies in DXVK and should also be able to pick up the sequence
* generated by spirv_to_nir for a OpLoad of a large composite followed by
* OpStore.
*
nir/find_array_copies: Handle wildcards and overlapping copies This commit rewrites opt_find_array_copies to be able to handle an array copy sequence with other intervening operations in between. In particular, this handles the case where we OpLoad an array of structs and then OpStore it, which generates code like: foo[0].a = bar[0].a foo[0].b = bar[0].b foo[1].a = bar[1].a foo[1].b = bar[1].b ... that wasn't recognized by the previous pass. In order to correctly handle copying arrays of arrays, and in particular to correctly handle copies involving wildcards, we need to use a tree structure similar to lower_vars_to_ssa so that we can walk all the partial array copies invalidated by a particular write, including ones where one of the common indices is a wildcard. I actually think that when factoring in the needed hashing/comparing code, a hash table based approach wouldn't be a lot smaller anyways. All of the changes come from tessellation control shaders in Strange Brigade, where we're able to remove the DXVK-inserted copy at the beginning of the shader. These are the result for radv: Totals from affected shaders: SGPRS: 4576 -> 4576 (0.00 %) VGPRS: 13784 -> 5560 (-59.66 %) Spilled SGPRs: 0 -> 0 (0.00 %) Spilled VGPRs: 0 -> 0 (0.00 %) Private memory VGPRs: 0 -> 0 (0.00 %) Scratch size: 8696 -> 6876 (-20.93 %) dwords per thread Code Size: 329940 -> 263268 (-20.21 %) bytes LDS: 0 -> 0 (0.00 %) blocks Max Waves: 330 -> 898 (172.12 %) Wait states: 0 -> 0 (0.00 %) Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2019-06-18 12:12:49 +02:00
* TODO: Support out-of-order copies.
*/
bool
nir_opt_find_array_copies(nir_shader *shader)
{
bool progress = false;
nir_foreach_function(function, shader) {
if (function->impl && opt_find_array_copies_impl(function->impl))
progress = true;
}
return progress;
}