fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-23 22:00:13 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

radv: handle unaligned vertex fetches on GFX6/GFX10

Browse source 
The Vulkan spec doesn't have any words for vertex attributes alignment.

Fixes a test failure on GFX6 and a GPU hang on GFX10 with:
dEQP-VK.spirv_assembly.instruction.spirv1p4.entrypoint.tess_con_pc_entry_point

vkpipeline-db results on GFX10:
Totals from affected shaders:
SGPRS: 463772 -> 472972 (1.98 %)
VGPRS: 343208 -> 343752 (0.16 %)
Spilled SGPRs: 323 -> 336 (4.02 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Code Size: 13806200 -> 14164472 (2.60 %) bytes
Max Waves: 84021 -> 83755 (-0.32 %)

Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/2161
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This commit is contained in:
Samuel Pitoiset 2019-11-29 15:12:30 +01:00
parent bd888bc1d6
commit b37c91c12e
1 changed files with 87 additions and 48 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

135
src/amd/vulkan/radv_nir_to_llvm.c
View file

@ -1280,34 +1280,28 @@ adjust_vertex_fetch_alpha(struct radv_shader_context *ctx,
return LLVMBuildBitCast(ctx->ac.builder, alpha, ctx->ac.i32, ""); return LLVMBuildBitCast(ctx->ac.builder, alpha, ctx->ac.i32, "");
} }
static unsigned static const struct vertex_format_info {
get_num_channels_from_data_format(unsigned data_format) uint8_t vertex_byte_size;
{ uint8_t num_channels;
switch (data_format) { uint8_t chan_byte_size;
case V_008F0C_BUF_DATA_FORMAT_8: uint8_t chan_format;
case V_008F0C_BUF_DATA_FORMAT_16: } vertex_format_table[] = {
case V_008F0C_BUF_DATA_FORMAT_32: { 0, 4, 0, V_008F0C_BUF_DATA_FORMAT_INVALID }, /* BUF_DATA_FORMAT_INVALID */
return 1; { 1, 1, 1, V_008F0C_BUF_DATA_FORMAT_8 }, /* BUF_DATA_FORMAT_8 */
case V_008F0C_BUF_DATA_FORMAT_8_8: { 2, 1, 2, V_008F0C_BUF_DATA_FORMAT_16 }, /* BUF_DATA_FORMAT_16 */
case V_008F0C_BUF_DATA_FORMAT_16_16: { 2, 2, 1, V_008F0C_BUF_DATA_FORMAT_8 }, /* BUF_DATA_FORMAT_8_8 */
case V_008F0C_BUF_DATA_FORMAT_32_32: { 4, 1, 4, V_008F0C_BUF_DATA_FORMAT_32 }, /* BUF_DATA_FORMAT_32 */
return 2; { 4, 2, 2, V_008F0C_BUF_DATA_FORMAT_16 }, /* BUF_DATA_FORMAT_16_16 */
case V_008F0C_BUF_DATA_FORMAT_10_11_11: { 4, 3, 0, V_008F0C_BUF_DATA_FORMAT_10_11_11 }, /* BUF_DATA_FORMAT_10_11_11 */
case V_008F0C_BUF_DATA_FORMAT_11_11_10: { 4, 3, 0, V_008F0C_BUF_DATA_FORMAT_11_11_10 }, /* BUF_DATA_FORMAT_11_11_10 */
case V_008F0C_BUF_DATA_FORMAT_32_32_32: { 4, 4, 0, V_008F0C_BUF_DATA_FORMAT_10_10_10_2 }, /* BUF_DATA_FORMAT_10_10_10_2 */
return 3; { 4, 4, 0, V_008F0C_BUF_DATA_FORMAT_2_10_10_10 }, /* BUF_DATA_FORMAT_2_10_10_10 */
case V_008F0C_BUF_DATA_FORMAT_8_8_8_8: { 4, 4, 1, V_008F0C_BUF_DATA_FORMAT_8 }, /* BUF_DATA_FORMAT_8_8_8_8 */
case V_008F0C_BUF_DATA_FORMAT_10_10_10_2: { 8, 2, 4, V_008F0C_BUF_DATA_FORMAT_32 }, /* BUF_DATA_FORMAT_32_32 */
case V_008F0C_BUF_DATA_FORMAT_2_10_10_10: { 8, 4, 2, V_008F0C_BUF_DATA_FORMAT_16 }, /* BUF_DATA_FORMAT_16_16_16_16 */
case V_008F0C_BUF_DATA_FORMAT_16_16_16_16: { 12, 3, 4, V_008F0C_BUF_DATA_FORMAT_32 }, /* BUF_DATA_FORMAT_32_32_32 */
case V_008F0C_BUF_DATA_FORMAT_32_32_32_32: { 16, 4, 4, V_008F0C_BUF_DATA_FORMAT_32 }, /* BUF_DATA_FORMAT_32_32_32_32 */
return 4; };
default:
break;
}
return 4;
}
static LLVMValueRef static LLVMValueRef
radv_fixup_vertex_input_fetches(struct radv_shader_context *ctx, radv_fixup_vertex_input_fetches(struct radv_shader_context *ctx,
@ -1393,11 +1387,13 @@ handle_vs_input_decl(struct radv_shader_context *ctx,
ctx->args->ac.base_vertex), ""); ctx->args->ac.base_vertex), "");
} }
assert(data_format < ARRAY_SIZE(vertex_format_table));
const struct vertex_format_info *vtx_info = &vertex_format_table[data_format];
/* Adjust the number of channels to load based on the vertex /* Adjust the number of channels to load based on the vertex
* attribute format. * attribute format.
*/ */
unsigned num_format_channels = get_num_channels_from_data_format(data_format); unsigned num_channels = MIN2(num_input_channels, vtx_info->num_channels);
unsigned num_channels = MIN2(num_input_channels, num_format_channels);
unsigned attrib_binding = ctx->args->options->key.vs.vertex_attribute_bindings[attrib_index]; unsigned attrib_binding = ctx->args->options->key.vs.vertex_attribute_bindings[attrib_index];
unsigned attrib_offset = ctx->args->options->key.vs.vertex_attribute_offsets[attrib_index]; unsigned attrib_offset = ctx->args->options->key.vs.vertex_attribute_offsets[attrib_index];
unsigned attrib_stride = ctx->args->options->key.vs.vertex_attribute_strides[attrib_index]; unsigned attrib_stride = ctx->args->options->key.vs.vertex_attribute_strides[attrib_index];
@ -1409,27 +1405,70 @@ handle_vs_input_decl(struct radv_shader_context *ctx,
num_channels = MAX2(num_channels, 3); num_channels = MAX2(num_channels, 3);
} }
if (attrib_stride != 0 && attrib_offset > attrib_stride) {
LLVMValueRef buffer_offset =
LLVMConstInt(ctx->ac.i32,
attrib_offset / attrib_stride, false);
buffer_index = LLVMBuildAdd(ctx->ac.builder,
buffer_index,
buffer_offset, "");
attrib_offset = attrib_offset % attrib_stride;
}
t_offset = LLVMConstInt(ctx->ac.i32, attrib_binding, false); t_offset = LLVMConstInt(ctx->ac.i32, attrib_binding, false);
t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_ptr, t_offset); t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_ptr, t_offset);
input = ac_build_struct_tbuffer_load(&ctx->ac, t_list, /* Perform per-channel vertex fetch operations if unaligned
buffer_index, * access are detected. Only GFX6 and GFX10 are affected.
LLVMConstInt(ctx->ac.i32, attrib_offset, false), */
ctx->ac.i32_0, ctx->ac.i32_0, bool unaligned_vertex_fetches = false;
num_channels, if ((ctx->ac.chip_class == GFX6 || ctx->ac.chip_class == GFX10) &&
data_format, num_format, 0, true); vtx_info->chan_format != data_format &&
((attrib_offset % vtx_info->vertex_byte_size) ||
(attrib_stride % vtx_info->vertex_byte_size)))
unaligned_vertex_fetches = true;
if (unaligned_vertex_fetches) {
unsigned chan_format = vtx_info->chan_format;
LLVMValueRef values[4];
assert(ctx->ac.chip_class == GFX6 ||
ctx->ac.chip_class == GFX10);
for (unsigned chan = 0; chan < num_channels; chan++) {
unsigned chan_offset = attrib_offset + chan * vtx_info->chan_byte_size;
LLVMValueRef chan_index = buffer_index;
if (attrib_stride != 0 && chan_offset > attrib_stride) {
LLVMValueRef buffer_offset =
LLVMConstInt(ctx->ac.i32,
chan_offset / attrib_stride, false);
chan_index = LLVMBuildAdd(ctx->ac.builder,
buffer_index,
buffer_offset, "");
chan_offset = chan_offset % attrib_stride;
}
values[chan] = ac_build_struct_tbuffer_load(&ctx->ac, t_list,
chan_index,
LLVMConstInt(ctx->ac.i32, chan_offset, false),
ctx->ac.i32_0, ctx->ac.i32_0, 1,
chan_format, num_format, 0, true);
}
input = ac_build_gather_values(&ctx->ac, values, num_channels);
} else {
if (attrib_stride != 0 && attrib_offset > attrib_stride) {
LLVMValueRef buffer_offset =
LLVMConstInt(ctx->ac.i32,
attrib_offset / attrib_stride, false);
buffer_index = LLVMBuildAdd(ctx->ac.builder,
buffer_index,
buffer_offset, "");
attrib_offset = attrib_offset % attrib_stride;
}
input = ac_build_struct_tbuffer_load(&ctx->ac, t_list,
buffer_index,
LLVMConstInt(ctx->ac.i32, attrib_offset, false),
ctx->ac.i32_0, ctx->ac.i32_0,
num_channels,
data_format, num_format, 0, true);
}
if (ctx->args->options->key.vs.post_shuffle & (1 << attrib_index)) { if (ctx->args->options->key.vs.post_shuffle & (1 << attrib_index)) {
LLVMValueRef c[4]; LLVMValueRef c[4];