aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorMarek Olšák <[email protected]>2018-11-19 21:53:55 -0500
committerMarek Olšák <[email protected]>2018-11-28 20:20:27 -0500
commit709905cbb683d5b0b9fdbb82c29165aba0149706 (patch)
tree8093319cb15114cc73097f78d21c944b624388aa
parent648dc52367c666817f04cc2c2c4f8f853d4e9ca9 (diff)
radeonsi: fix is_oneway_access_only for image stores
We need to look at the Dst for image stores.
-rw-r--r--src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c49
1 files changed, 37 insertions, 12 deletions
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
index 2ba3f251ff8..81df73ea9b1 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
@@ -402,6 +402,32 @@ static bool is_oneway_access_only(const struct tgsi_full_instruction *inst,
unsigned shader_buffers_reverse_access_mask,
unsigned images_reverse_access_mask)
{
+ enum tgsi_file_type resource_file;
+ unsigned resource_index;
+ bool resource_indirect;
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_STORE) {
+ resource_file = inst->Dst[0].Register.File;
+ resource_index = inst->Dst[0].Register.Index;
+ resource_indirect = inst->Dst[0].Register.Indirect;
+ } else {
+ resource_file = inst->Src[0].Register.File;
+ resource_index = inst->Src[0].Register.Index;
+ resource_indirect = inst->Src[0].Register.Indirect;
+ }
+
+ assert(resource_file == TGSI_FILE_BUFFER ||
+ resource_file == TGSI_FILE_IMAGE ||
+ /* bindless image */
+ resource_file == TGSI_FILE_INPUT ||
+ resource_file == TGSI_FILE_OUTPUT ||
+ resource_file == TGSI_FILE_CONSTANT ||
+ resource_file == TGSI_FILE_TEMPORARY ||
+ resource_file == TGSI_FILE_IMMEDIATE);
+
+ assert(resource_file != TGSI_FILE_BUFFER ||
+ inst->Memory.Texture == TGSI_TEXTURE_BUFFER);
+
/* RESTRICT means NOALIAS.
* If there are no writes, we can assume the accessed memory is read-only.
* If there are no reads, we can assume the accessed memory is write-only.
@@ -409,7 +435,7 @@ static bool is_oneway_access_only(const struct tgsi_full_instruction *inst,
if (inst->Memory.Qualifier & TGSI_MEMORY_RESTRICT) {
unsigned reverse_access_mask;
- if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
+ if (resource_file == TGSI_FILE_BUFFER) {
reverse_access_mask = shader_buffers_reverse_access_mask;
} else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
reverse_access_mask = info->images_buffers &
@@ -419,12 +445,12 @@ static bool is_oneway_access_only(const struct tgsi_full_instruction *inst,
images_reverse_access_mask;
}
- if (inst->Src[0].Register.Indirect) {
+ if (resource_indirect) {
if (!reverse_access_mask)
return true;
} else {
if (!(reverse_access_mask &
- (1u << inst->Src[0].Register.Index)))
+ (1u << resource_index)))
return true;
}
}
@@ -437,10 +463,8 @@ static bool is_oneway_access_only(const struct tgsi_full_instruction *inst,
* Same for the case when there are no writes/reads for non-buffer
* images.
*/
- if (inst->Src[0].Register.File == TGSI_FILE_BUFFER ||
- (inst->Memory.Texture == TGSI_TEXTURE_BUFFER &&
- (inst->Src[0].Register.File == TGSI_FILE_IMAGE ||
- tgsi_is_bindless_image_file(inst->Src[0].Register.File)))) {
+ if (resource_file == TGSI_FILE_BUFFER ||
+ inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
if (!shader_buffers_reverse_access_mask &&
!(info->images_buffers & images_reverse_access_mask))
return true;
@@ -650,6 +674,12 @@ static void store_emit(
struct tgsi_full_src_register resource_reg =
tgsi_full_src_register_from_dst(&inst->Dst[0]);
unsigned target = inst->Memory.Texture;
+
+ if (inst->Dst[0].Register.File == TGSI_FILE_MEMORY) {
+ store_emit_memory(ctx, emit_data);
+ return;
+ }
+
bool writeonly_memory = is_oneway_access_only(inst, info,
info->shader_buffers_load |
info->shader_buffers_atomic,
@@ -662,11 +692,6 @@ static void store_emit(
LLVMValueRef voffset = ctx->i32_0;
struct ac_image_args args = {};
- if (inst->Dst[0].Register.File == TGSI_FILE_MEMORY) {
- store_emit_memory(ctx, emit_data);
- return;
- }
-
for (unsigned chan = 0; chan < 4; ++chan)
chans[chan] = lp_build_emit_fetch(bld_base, inst, 1, chan);