fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-05 00:58:05 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

radv,aco: use pipe_format for dynamic vertex input state

Browse source 
Also prepare for 64-bit and R8G8B8/R16G16B16 with the addition of
radv_vs_input_state::nontrivial_formats.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/5021
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17894>
This commit is contained in:
Rhys Perry 2022-08-03 14:38:16 +01:00 • committed by Marge Bot
parent c06a5a5ebd
commit 831257bdce
4 changed files with 50 additions and 104 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
src/amd/compiler/aco_instruction_selection.cpp
View file

@ -12221,6 +12221,9 @@ select_vs_prolog(Program* program, const struct aco_vs_prolog_key* key, ac_shade
num_vgprs++; /* make space for nontrivial_tmp_vgpr1 */
unsigned num_sgprs = 0;
const struct ac_vtx_format_info* vtx_info_table =
ac_get_vtx_format_info_table(GFX8, CHIP_POLARIS10);
for (unsigned loc = 0; loc < key->num_attributes;) {
unsigned num_descs =
load_vb_descs(bld, desc, Operand(vertex_buffers, s2), loc, key->num_attributes - loc);
@ -12285,9 +12288,12 @@ select_vs_prolog(Program* program, const struct aco_vs_prolog_key* key, ac_shade
/* perform load */
PhysReg cur_desc = desc.advance(i * 16);
if ((key->misaligned_mask & (1u << loc))) {
unsigned dfmt = key->state.formats[loc] & 0xf;
unsigned nfmt = key->state.formats[loc] >> 4;
const struct ac_data_format_info* vtx_info = ac_get_data_format_info(dfmt);
const struct ac_vtx_format_info* vtx_info = &vtx_info_table[key->state.formats[loc]];
assert(vtx_info->has_hw_format & 0x1);
unsigned dfmt = vtx_info->hw_format[0] & 0xf;
unsigned nfmt = vtx_info->hw_format[0] >> 4;
for (unsigned j = 0; j < vtx_info->num_channels; j++) {
bool post_shuffle = key->state.post_shuffle & (1u << loc);
unsigned offset = vtx_info->chan_byte_size * (post_shuffle && j < 3 ? 2 - j : j);
@ -12297,14 +12303,14 @@ select_vs_prolog(Program* program, const struct aco_vs_prolog_key* key, ac_shade
* MTBUF can hang, but MUBUF doesn't (probably gives garbage, but GL CTS doesn't
* care).
*/
if (vtx_info->chan_format == V_008F0C_BUF_DATA_FORMAT_32)
if (dfmt == V_008F0C_BUF_DATA_FORMAT_32)
bld.mubuf(aco_opcode::buffer_load_dword, Definition(dest.advance(j * 4u), v1),
Operand(cur_desc, s4), fetch_index, Operand::c32(0u), offset, false,
false, true);
else
bld.mtbuf(aco_opcode::tbuffer_load_format_x, Definition(dest.advance(j * 4u), v1),
Operand(cur_desc, s4), fetch_index, Operand::c32(0u),
vtx_info->chan_format, nfmt, offset, false, true);
Operand(cur_desc, s4), fetch_index, Operand::c32(0u), dfmt, nfmt,
offset, false, true);
}
uint32_t one =
nfmt == V_008F0C_BUF_NUM_FORMAT_UINT || nfmt == V_008F0C_BUF_NUM_FORMAT_SINT

124
src/amd/vulkan/radv_cmd_buffer.c
View file

@ -308,8 +308,6 @@ radv_destroy_cmd_buffer(struct radv_cmd_buffer *cmd_buffer)
{
list_del(&cmd_buffer->pool_link);
util_dynarray_fini(&cmd_buffer->cached_vertex_formats);
list_for_each_entry_safe(struct radv_cmd_buffer_upload, up, &cmd_buffer->upload.list, list)
{
cmd_buffer->device->ws->buffer_destroy(cmd_buffer->device->ws, up->upload_bo);
@ -380,8 +378,6 @@ radv_create_cmd_buffer(struct radv_device *device, struct radv_cmd_pool *pool,
vk_object_base_init(&device->vk, &cmd_buffer->meta_push_descriptors.base,
VK_OBJECT_TYPE_DESCRIPTOR_SET);
util_dynarray_init(&cmd_buffer->cached_vertex_formats, NULL);
for (unsigned i = 0; i < MAX_BIND_POINTS; i++)
vk_object_base_init(&device->vk, &cmd_buffer->descriptors[i].push_set.set.base,
VK_OBJECT_TYPE_DESCRIPTOR_SET);
@ -2964,6 +2960,7 @@ lookup_vs_prolog(struct radv_cmd_buffer *cmd_buffer, struct radv_shader *vs_shad
cmd_buffer->state.vbo_misaligned_mask = misaligned_mask;
cmd_buffer->state.vbo_misaligned_mask_invalid &= ~attribute_mask;
}
misaligned_mask |= state->nontrivial_formats;
/* try to use a pre-compiled prolog first */
struct radv_shader_part *prolog = NULL;
@ -3540,39 +3537,6 @@ radv_flush_constants(struct radv_cmd_buffer *cmd_buffer, VkShaderStageFlags stag
cmd_buffer->push_constant_stages |= dirty_stages;
}
enum radv_dst_sel {
DST_SEL_0001 = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_0) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_0) |
S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_0) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_1),
DST_SEL_X001 = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_0) |
S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_0) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_1),
DST_SEL_XY01 = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_0) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_1),
DST_SEL_XYZ1 = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_1),
DST_SEL_XYZW = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W),
DST_SEL_ZYXW = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W),
};
static const uint32_t data_format_dst_sel[] = {
[V_008F0C_BUF_DATA_FORMAT_INVALID] = DST_SEL_0001,
[V_008F0C_BUF_DATA_FORMAT_8] = DST_SEL_X001,
[V_008F0C_BUF_DATA_FORMAT_16] = DST_SEL_X001,
[V_008F0C_BUF_DATA_FORMAT_8_8] = DST_SEL_XY01,
[V_008F0C_BUF_DATA_FORMAT_32] = DST_SEL_X001,
[V_008F0C_BUF_DATA_FORMAT_16_16] = DST_SEL_XY01,
[V_008F0C_BUF_DATA_FORMAT_10_11_11] = DST_SEL_XYZ1,
[V_008F0C_BUF_DATA_FORMAT_11_11_10] = DST_SEL_XYZ1,
[V_008F0C_BUF_DATA_FORMAT_10_10_10_2] = DST_SEL_XYZW,
[V_008F0C_BUF_DATA_FORMAT_2_10_10_10] = DST_SEL_XYZW,
[V_008F0C_BUF_DATA_FORMAT_8_8_8_8] = DST_SEL_XYZW,
[V_008F0C_BUF_DATA_FORMAT_32_32] = DST_SEL_XY01,
[V_008F0C_BUF_DATA_FORMAT_16_16_16_16] = DST_SEL_XYZW,
[V_008F0C_BUF_DATA_FORMAT_32_32_32] = DST_SEL_XYZ1,
[V_008F0C_BUF_DATA_FORMAT_32_32_32_32] = DST_SEL_XYZW,
};
void
radv_write_vertex_descriptors(const struct radv_cmd_buffer *cmd_buffer,
const struct radv_graphics_pipeline *pipeline,
@ -3580,6 +3544,7 @@ radv_write_vertex_descriptors(const struct radv_cmd_buffer *cmd_buffer,
{
struct radv_shader *vs_shader = radv_get_shader(&pipeline->base, MESA_SHADER_VERTEX);
enum amd_gfx_level chip = cmd_buffer->device->physical_device->rad_info.gfx_level;
enum radeon_family family = cmd_buffer->device->physical_device->rad_info.family;
unsigned desc_index = 0;
uint32_t mask = pipeline->vb_desc_usage_mask;
uint64_t va;
@ -3587,6 +3552,9 @@ radv_write_vertex_descriptors(const struct radv_cmd_buffer *cmd_buffer,
vs_shader->info.vs.dynamic_inputs ? &cmd_buffer->state.dynamic_vs_input : NULL;
assert(!vs_state || pipeline->use_per_attribute_vb_descs);
const struct ac_vtx_format_info *vtx_info_table =
vs_state ? ac_get_vtx_format_info_table(chip, family) : NULL;
while (mask) {
unsigned i = u_bit_scan(&mask);
uint32_t *desc = &((uint32_t *)vb_ptr)[desc_index++ * 4];
@ -3598,23 +3566,25 @@ radv_write_vertex_descriptors(const struct radv_cmd_buffer *cmd_buffer,
unsigned num_records;
unsigned stride;
if (vs_state) {
unsigned format = vs_state->formats[i];
unsigned dfmt = format & 0xf;
unsigned nfmt = (format >> 4) & 0x7;
if (vs_state && !(vs_state->nontrivial_formats & BITFIELD_BIT(i))) {
const struct ac_vtx_format_info *vtx_info = &vtx_info_table[vs_state->formats[i]];
unsigned hw_format = vtx_info->hw_format[vtx_info->num_channels - 1];
rsrc_word3 = vs_state->post_shuffle & (1u << i) ? DST_SEL_ZYXW : data_format_dst_sel[dfmt];
if (chip >= GFX10)
rsrc_word3 |= S_008F0C_FORMAT(ac_get_tbuffer_format(chip, dfmt, nfmt));
else
rsrc_word3 |= S_008F0C_NUM_FORMAT(nfmt) | S_008F0C_DATA_FORMAT(dfmt);
if (chip >= GFX10) {
rsrc_word3 = vtx_info->dst_sel | S_008F0C_FORMAT(hw_format);
} else {
rsrc_word3 = vtx_info->dst_sel | S_008F0C_NUM_FORMAT((hw_format >> 4) & 0x7) |
S_008F0C_DATA_FORMAT(hw_format & 0xf);
}
} else {
rsrc_word3 = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
if (chip >= GFX10)
rsrc_word3 = DST_SEL_XYZW | S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_UINT);
rsrc_word3 |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_UINT);
else
rsrc_word3 = DST_SEL_XYZW | S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_UINT) |
S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
rsrc_word3 |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_UINT) |
S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
}
if (pipeline->dynamic_states & (RADV_DYNAMIC_VERTEX_INPUT_BINDING_STRIDE |
@ -5934,6 +5904,9 @@ radv_CmdSetVertexInputEXT(VkCommandBuffer commandBuffer, uint32_t vertexBindingD
state->bindings_match_attrib = true;
enum amd_gfx_level chip = cmd_buffer->device->physical_device->rad_info.gfx_level;
enum radeon_family family = cmd_buffer->device->physical_device->rad_info.family;
const struct ac_vtx_format_info *vtx_info_table = ac_get_vtx_format_info_table(chip, family);
for (unsigned i = 0; i < vertexAttributeDescriptionCount; i++) {
const VkVertexInputAttributeDescription2EXT *attrib = &pVertexAttributeDescriptions[i];
const VkVertexInputBindingDescription2EXT *binding = bindings[attrib->binding];
@ -5955,50 +5928,27 @@ radv_CmdSetVertexInputEXT(VkCommandBuffer commandBuffer, uint32_t vertexBindingD
cmd_buffer->vertex_bindings[attrib->binding].stride = binding->stride;
state->offsets[loc] = attrib->offset;
struct dynamic_vertex_format_cache *found = NULL;
util_dynarray_foreach(&cmd_buffer->cached_vertex_formats,
struct dynamic_vertex_format_cache,
vf) {
if (vf->format == attrib->format) {
found = vf;
break;
}
}
if (!found) {
unsigned nfmt, dfmt;
bool post_shuffle;
enum ac_vs_input_alpha_adjust alpha_adjust;
const struct util_format_description *format_desc = vk_format_description(attrib->format);
enum pipe_format format = vk_format_to_pipe_format(attrib->format);
const struct ac_vtx_format_info *vtx_info = &vtx_info_table[format];
found = util_dynarray_grow(&cmd_buffer->cached_vertex_formats,
struct dynamic_vertex_format_cache, 1);
radv_translate_vertex_format(cmd_buffer->device->physical_device, attrib->format, format_desc,
&dfmt, &nfmt, &post_shuffle, &alpha_adjust);
found->format = attrib->format;
found->hw_fmt = dfmt | (nfmt << 4);
const uint8_t format_align_req_minus_1 = format_desc->channel[0].size >= 32 ? 3 :
(format_desc->block.bits / 8u - 1);
found->fmt_align_req_minus_1 = format_align_req_minus_1;
found->fmt_size = format_desc->block.bits / 8u;
found->post_shuffle = post_shuffle;
found->alpha_adjust_lo = alpha_adjust & 0x1;
found->alpha_adjust_hi = (alpha_adjust >> 1) & 0x1;
}
state->formats[loc] = format;
uint8_t align_req_minus_1 = vtx_info->chan_byte_size >= 4 ? 3 : (vtx_info->element_size - 1);
state->format_align_req_minus_1[loc] = align_req_minus_1;
state->format_sizes[loc] = vtx_info->element_size;
state->alpha_adjust_lo |= (vtx_info->alpha_adjust & 0x1) << loc;
state->alpha_adjust_hi |= (vtx_info->alpha_adjust >> 1) << loc;
if (G_008F0C_DST_SEL_X(vtx_info->dst_sel) == V_008F0C_SQ_SEL_Z)
state->post_shuffle |= BITFIELD_BIT(loc);
state->formats[loc] = found->hw_fmt;
state->format_align_req_minus_1[loc] = found->fmt_align_req_minus_1;
state->format_sizes[loc] = found->fmt_size;
state->alpha_adjust_lo |= found->alpha_adjust_lo << loc;
state->alpha_adjust_hi |= found->alpha_adjust_hi << loc;
if (found->post_shuffle)
state->post_shuffle |= 1u << loc;
if (!(vtx_info->has_hw_format & BITFIELD_BIT(vtx_info->num_channels - 1)))
state->nontrivial_formats |= BITFIELD_BIT(loc);
if ((chip == GFX6 || chip >= GFX10) &&
cmd_buffer->state.vbo_bound_mask & BITFIELD_BIT(attrib->binding)) {
if (binding->stride & found->fmt_align_req_minus_1) {
if (binding->stride & align_req_minus_1) {
cmd_buffer->state.vbo_misaligned_mask |= BITFIELD_BIT(loc);
} else if ((cmd_buffer->vertex_bindings[attrib->binding].offset + state->offsets[loc]) &
found->fmt_align_req_minus_1) {
align_req_minus_1) {
cmd_buffer->state.vbo_misaligned_mask |= BITFIELD_BIT(loc);
}
}

11
src/amd/vulkan/radv_private.h
View file

@ -1568,16 +1568,6 @@ enum radv_cmd_buffer_status {
RADV_CMD_BUFFER_STATUS_PENDING,
};
struct dynamic_vertex_format_cache {
VkFormat format;
uint8_t hw_fmt;
uint8_t fmt_align_req_minus_1;
uint8_t fmt_size;
bool post_shuffle;
bool alpha_adjust_lo;
bool alpha_adjust_hi;
};
struct radv_cmd_buffer {
struct vk_command_buffer vk;
@ -1586,7 +1576,6 @@ struct radv_cmd_buffer {
struct radv_cmd_pool *pool;
struct list_head pool_link;
struct util_dynarray cached_vertex_formats;
VkCommandBufferUsageFlags usage_flags;
enum radv_cmd_buffer_status status;
struct radeon_cmdbuf *cs;

1
src/amd/vulkan/radv_shader.h
View file

@ -374,6 +374,7 @@ struct radv_vs_input_state {
*/
uint32_t alpha_adjust_lo;
uint32_t alpha_adjust_hi;
uint32_t nontrivial_formats;
uint8_t bindings[MAX_VERTEX_ATTRIBS];
uint32_t divisors[MAX_VERTEX_ATTRIBS];