KonstantinSeurer
/
mesa
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

radv: Implement software emulation for intersect_ray 

Signed-off-by: Joshua Ashton <joshua@froggi.es>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12974>
This commit is contained in:
Joshua Ashton 2021-09-21 15:29:59 +01:00 committed by Marge Bot
parent a0f51921a6
commit 548382de42
1 changed files with 295 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

299
src/amd/vulkan/radv_pipeline_rt.c
View File

@ -28,6 +28,7 @@
#include "nir/nir.h"
#include "nir/nir_builder.h"
#include "nir/nir_builtin_builder.h"
static VkRayTracingPipelineCreateInfoKHR
radv_create_merged_rt_create_info(const VkRayTracingPipelineCreateInfoKHR *pCreateInfo)
@ -1517,6 +1518,277 @@ insert_traversal_aabb_case(struct radv_device *device,
nir_pop_if(b, NULL);
}
static void
nir_sort_hit_pair(nir_builder *b, nir_variable *var_distances, nir_variable *var_indices, uint32_t chan_1, uint32_t chan_2)
{
nir_ssa_def *ssa_distances = nir_load_var(b, var_distances);
nir_ssa_def *ssa_indices = nir_load_var(b, var_indices);
/* if (distances[chan_2] < distances[chan_1]) { */
nir_push_if(b, nir_flt(b, nir_channel(b, ssa_distances, chan_2), nir_channel(b, ssa_distances, chan_1)));
{
/* swap(distances[chan_2], distances[chan_1]); */
nir_ssa_def *new_distances[4] = {nir_ssa_undef(b, 1, 32), nir_ssa_undef(b, 1, 32), nir_ssa_undef(b, 1, 32), nir_ssa_undef(b, 1, 32)};
nir_ssa_def *new_indices[4] = {nir_ssa_undef(b, 1, 32), nir_ssa_undef(b, 1, 32), nir_ssa_undef(b, 1, 32), nir_ssa_undef(b, 1, 32)};
new_distances[chan_2] = nir_channel(b, ssa_distances, chan_1);
new_distances[chan_1] = nir_channel(b, ssa_distances, chan_2);
new_indices[chan_2] = nir_channel(b, ssa_indices, chan_1);
new_indices[chan_1] = nir_channel(b, ssa_indices, chan_2);
nir_store_var(b, var_distances, nir_vec(b, new_distances, 4), (1u << chan_1) | (1u << chan_2));
nir_store_var(b, var_indices, nir_vec(b, new_indices, 4), (1u << chan_1) | (1u << chan_2));
}
/* } */
nir_pop_if(b, NULL);
}
static nir_ssa_def *
intersect_ray_amd_software_box(struct radv_device *device,
nir_builder *b, nir_ssa_def *bvh_node,
nir_ssa_def *ray_tmax, nir_ssa_def *origin,
nir_ssa_def *dir, nir_ssa_def *inv_dir)
{
const struct glsl_type *vec4_type = glsl_vector_type(GLSL_TYPE_FLOAT, 4);
const struct glsl_type *uvec4_type = glsl_vector_type(GLSL_TYPE_UINT, 4);
nir_ssa_def *node_addr = build_node_to_addr(device, b, bvh_node);
/* vec4 distances = vec4(INF, INF, INF, INF); */
nir_variable *distances = nir_variable_create(b->shader, nir_var_shader_temp, vec4_type, "distances");
nir_store_var(b, distances, nir_imm_vec4(b, INFINITY, INFINITY, INFINITY, INFINITY), 0xf);
/* uvec4 child_indices = uvec4(0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff); */
nir_variable *child_indices = nir_variable_create(b->shader, nir_var_shader_temp, uvec4_type, "child_indices");
nir_store_var(b, child_indices, nir_imm_ivec4(b, 0xffffffffu, 0xffffffffu, 0xffffffffu, 0xffffffffu), 0xf);
/* Need to remove infinities here because otherwise we get nasty NaN propogation
* if the direction has 0s in it. */
/* inv_dir = clamp(inv_dir, -FLT_MAX, FLT_MAX); */
inv_dir = nir_fclamp(b, inv_dir, nir_imm_float(b, -FLT_MAX), nir_imm_float(b, FLT_MAX));
for (int i = 0; i < 4; i++) {
const uint32_t child_offset = offsetof(struct radv_bvh_box32_node, children[i]);
const uint32_t coord_offsets[2] = {
offsetof(struct radv_bvh_box32_node, coords[i][0][0]),
offsetof(struct radv_bvh_box32_node, coords[i][1][0]),
};
/* node->children[i] -> uint */
nir_ssa_def *child_index = nir_build_load_global(b, 1, 32, nir_iadd(b, node_addr, nir_imm_int64(b, child_offset)), .align_mul = 64, .align_offset = child_offset % 64 );
/* node->coords[i][0], node->coords[i][1] -> vec3 */
nir_ssa_def *node_coords[2] = {
nir_build_load_global(b, 3, 32, nir_iadd(b, node_addr, nir_imm_int64(b, coord_offsets[0])), .align_mul = 64, .align_offset = coord_offsets[0] % 64 ),
nir_build_load_global(b, 3, 32, nir_iadd(b, node_addr, nir_imm_int64(b, coord_offsets[1])), .align_mul = 64, .align_offset = coord_offsets[1] % 64 ),
};
/* If x of the aabb min is NaN, then this is an inactive aabb.
* We don't need to care about any other components being NaN as that is UB.
* https://www.khronos.org/registry/vulkan/specs/1.2-extensions/html/chap36.html#VkAabbPositionsKHR */
nir_ssa_def *min_x = nir_channel(b, node_coords[0], 0);
nir_ssa_def *min_x_is_not_nan = nir_inot(b, nir_fneu(b, min_x, min_x)); /* NaN != NaN -> true */
/* vec3 bound0 = (node->coords[i][0] - origin) * inv_dir; */
nir_ssa_def *bound0 = nir_fmul(b, nir_fsub(b, node_coords[0], origin), inv_dir);
/* vec3 bound1 = (node->coords[i][1] - origin) * inv_dir; */
nir_ssa_def *bound1 = nir_fmul(b, nir_fsub(b, node_coords[1], origin), inv_dir);
/* float tmin = max(max(min(bound0.x, bound1.x), min(bound0.y, bound1.y)), min(bound0.z, bound1.z)); */
nir_ssa_def *tmin = nir_fmax(b, nir_fmax(b,
nir_fmin(b, nir_channel(b, bound0, 0), nir_channel(b, bound1, 0)),
nir_fmin(b, nir_channel(b, bound0, 1), nir_channel(b, bound1, 1))),
nir_fmin(b, nir_channel(b, bound0, 2), nir_channel(b, bound1, 2)));
/* float tmax = min(min(max(bound0.x, bound1.x), max(bound0.y, bound1.y)), max(bound0.z, bound1.z)); */
nir_ssa_def *tmax = nir_fmin(b, nir_fmin(b,
nir_fmax(b, nir_channel(b, bound0, 0), nir_channel(b, bound1, 0)),
nir_fmax(b, nir_channel(b, bound0, 1), nir_channel(b, bound1, 1))),
nir_fmax(b, nir_channel(b, bound0, 2), nir_channel(b, bound1, 2)));
/* if (!isnan(node->coords[i][0].x) && tmax >= max(0.0f, tmin) && tmin < ray_tmax) { */
nir_push_if(b,
nir_iand(b,
min_x_is_not_nan,
nir_iand(b,
nir_fge(b, tmax, nir_fmax(b, nir_imm_float(b, 0.0f), tmin)),
nir_flt(b, tmin, ray_tmax))));
{
/* child_indices[i] = node->children[i]; */
nir_ssa_def *new_child_indices[4] = {child_index, child_index, child_index, child_index};
nir_store_var(b, child_indices, nir_vec(b, new_child_indices, 4), 1u << i);
/* distances[i] = tmin; */
nir_ssa_def *new_distances[4] = {tmin, tmin, tmin, tmin};
nir_store_var(b, distances, nir_vec(b, new_distances, 4), 1u << i);
}
/* } */
nir_pop_if(b, NULL);
}
/* Sort our distances with a sorting network. */
nir_sort_hit_pair(b, distances, child_indices, 0, 1);
nir_sort_hit_pair(b, distances, child_indices, 2, 3);
nir_sort_hit_pair(b, distances, child_indices, 0, 2);
nir_sort_hit_pair(b, distances, child_indices, 1, 3);
nir_sort_hit_pair(b, distances, child_indices, 1, 2);
return nir_load_var(b, child_indices);
}
static nir_ssa_def *
intersect_ray_amd_software_tri(struct radv_device *device,
nir_builder *b, nir_ssa_def *bvh_node,
nir_ssa_def *ray_tmax, nir_ssa_def *origin,
nir_ssa_def *dir, nir_ssa_def *inv_dir)
{
const struct glsl_type *vec4_type = glsl_vector_type(GLSL_TYPE_FLOAT, 4);
nir_ssa_def *node_addr = build_node_to_addr(device, b, bvh_node);
const uint32_t coord_offsets[3] = {
offsetof(struct radv_bvh_triangle_node, coords[0]),
offsetof(struct radv_bvh_triangle_node, coords[1]),
offsetof(struct radv_bvh_triangle_node, coords[2]),
};
/* node->coords[0], node->coords[1], node->coords[2] -> vec3 */
nir_ssa_def *node_coords[3] = {
nir_build_load_global(b, 3, 32, nir_iadd(b, node_addr, nir_imm_int64(b, coord_offsets[0])), .align_mul = 64, .align_offset = coord_offsets[0] % 64 ),
nir_build_load_global(b, 3, 32, nir_iadd(b, node_addr, nir_imm_int64(b, coord_offsets[1])), .align_mul = 64, .align_offset = coord_offsets[1] % 64 ),
nir_build_load_global(b, 3, 32, nir_iadd(b, node_addr, nir_imm_int64(b, coord_offsets[2])), .align_mul = 64, .align_offset = coord_offsets[2] % 64 ),
};
nir_variable *result = nir_variable_create(b->shader, nir_var_shader_temp, vec4_type, "result");
nir_store_var(b, result, nir_imm_vec4(b, INFINITY, 1.0f, 0.0f, 0.0f), 0xf);
/* Based on watertight Ray/Triangle intersection from
* http://jcgt.org/published/0002/01/05/paper.pdf */
/* Calculate the dimension where the ray direction is largest */
nir_ssa_def *abs_dir = nir_fabs(b, dir);
nir_ssa_def *abs_dirs[3] = {
nir_channel(b, abs_dir, 0),
nir_channel(b, abs_dir, 1),
nir_channel(b, abs_dir, 2),
};
/* Find index of greatest value of abs_dir and put that as kz. */
nir_ssa_def *kz = nir_bcsel(b, nir_fge(b, abs_dirs[0], abs_dirs[1]),
nir_bcsel(b, nir_fge(b, abs_dirs[0], abs_dirs[2]),
nir_imm_int(b, 0), nir_imm_int(b, 2)),
nir_bcsel(b, nir_fge(b, abs_dirs[1], abs_dirs[2]),
nir_imm_int(b, 1), nir_imm_int(b, 2)));
nir_ssa_def *kx = nir_imod(b, nir_iadd(b, kz, nir_imm_int(b, 1)), nir_imm_int(b, 3));
nir_ssa_def *ky = nir_imod(b, nir_iadd(b, kx, nir_imm_int(b, 1)), nir_imm_int(b, 3));
nir_ssa_def *k_indices[3] = { kx, ky, kz };
nir_ssa_def *k = nir_vec(b, k_indices, 3);
/* Swap kx and ky dimensions to preseve winding order */
unsigned swap_xy_swizzle[4] = {1, 0, 2, 3};
k = nir_bcsel(b,
nir_flt(b, nir_vector_extract(b, dir, kz), nir_imm_float(b, 0.0f)),
nir_swizzle(b, k, swap_xy_swizzle, 3),
k);
kx = nir_channel(b, k, 0);
ky = nir_channel(b, k, 1);
kz = nir_channel(b, k, 2);
/* Calculate shear constants */
nir_ssa_def *sz = nir_frcp(b, nir_vector_extract(b, dir, kz));
nir_ssa_def *sx = nir_fmul(b, nir_vector_extract(b, dir, kx), sz);
nir_ssa_def *sy = nir_fmul(b, nir_vector_extract(b, dir, ky), sz);
/* Calculate vertices relative to ray origin */
nir_ssa_def *v_a = nir_fsub(b, node_coords[0], origin);
nir_ssa_def *v_b = nir_fsub(b, node_coords[1], origin);
nir_ssa_def *v_c = nir_fsub(b, node_coords[2], origin);
/* Perform shear and scale */
nir_ssa_def *ax = nir_fsub(b, nir_vector_extract(b, v_a, kx), nir_fmul(b, sx, nir_vector_extract(b, v_a, kz)));
nir_ssa_def *ay = nir_fsub(b, nir_vector_extract(b, v_a, ky), nir_fmul(b, sy, nir_vector_extract(b, v_a, kz)));
nir_ssa_def *bx = nir_fsub(b, nir_vector_extract(b, v_b, kx), nir_fmul(b, sx, nir_vector_extract(b, v_b, kz)));
nir_ssa_def *by = nir_fsub(b, nir_vector_extract(b, v_b, ky), nir_fmul(b, sy, nir_vector_extract(b, v_b, kz)));
nir_ssa_def *cx = nir_fsub(b, nir_vector_extract(b, v_c, kx), nir_fmul(b, sx, nir_vector_extract(b, v_c, kz)));
nir_ssa_def *cy = nir_fsub(b, nir_vector_extract(b, v_c, ky), nir_fmul(b, sy, nir_vector_extract(b, v_c, kz)));
nir_ssa_def *u = nir_fsub(b, nir_fmul(b, cx, by), nir_fmul(b, cy, bx));
nir_ssa_def *v = nir_fsub(b, nir_fmul(b, ax, cy), nir_fmul(b, ay, cx));
nir_ssa_def *w = nir_fsub(b, nir_fmul(b, bx, ay), nir_fmul(b, by, ax));
nir_variable *u_var = nir_variable_create(b->shader, nir_var_shader_temp, glsl_float_type(), "u");
nir_variable *v_var = nir_variable_create(b->shader, nir_var_shader_temp, glsl_float_type(), "v");
nir_variable *w_var = nir_variable_create(b->shader, nir_var_shader_temp, glsl_float_type(), "w");
nir_store_var(b, u_var, u, 0x1);
nir_store_var(b, v_var, v, 0x1);
nir_store_var(b, w_var, w, 0x1);
/* Fallback to testing edges with double precision...
*
* The Vulkan spec states it only needs single precision watertightness
* but we fail dEQP-VK.ray_tracing_pipeline.watertightness.closedFan2.1024 with
* failures = 1 without doing this. :( */
nir_ssa_def *cond_retest = nir_ior(b, nir_ior(b,
nir_feq(b, u, nir_imm_float(b, 0.0f)),
nir_feq(b, v, nir_imm_float(b, 0.0f))),
nir_feq(b, w, nir_imm_float(b, 0.0f)));
nir_push_if(b, cond_retest);
{
ax = nir_f2f64(b, ax); ay = nir_f2f64(b, ay);
bx = nir_f2f64(b, bx); by = nir_f2f64(b, by);
cx = nir_f2f64(b, cx); cy = nir_f2f64(b, cy);
nir_store_var(b, u_var, nir_f2f32(b, nir_fsub(b, nir_fmul(b, cx, by), nir_fmul(b, cy, bx))), 0x1);
nir_store_var(b, v_var, nir_f2f32(b, nir_fsub(b, nir_fmul(b, ax, cy), nir_fmul(b, ay, cx))), 0x1);
nir_store_var(b, w_var, nir_f2f32(b, nir_fsub(b, nir_fmul(b, bx, ay), nir_fmul(b, by, ax))), 0x1);
}
nir_pop_if(b, NULL);
u = nir_load_var(b, u_var);
v = nir_load_var(b, v_var);
w = nir_load_var(b, w_var);
/* Perform edge tests. */
nir_ssa_def *cond_back = nir_ior(b, nir_ior(b,
nir_flt(b, u, nir_imm_float(b, 0.0f)),
nir_flt(b, v, nir_imm_float(b, 0.0f))),
nir_flt(b, w, nir_imm_float(b, 0.0f)));
nir_ssa_def *cond_front = nir_ior(b, nir_ior(b,
nir_flt(b, nir_imm_float(b, 0.0f), u),
nir_flt(b, nir_imm_float(b, 0.0f), v)),
nir_flt(b, nir_imm_float(b, 0.0f), w));
nir_ssa_def *cond = nir_inot(b, nir_iand(b, cond_back, cond_front));
nir_push_if(b, cond);
{
nir_ssa_def *det = nir_fadd(b, u, nir_fadd(b, v, w));
nir_ssa_def *az = nir_fmul(b, sz, nir_vector_extract(b, v_a, kz));
nir_ssa_def *bz = nir_fmul(b, sz, nir_vector_extract(b, v_b, kz));
nir_ssa_def *cz = nir_fmul(b, sz, nir_vector_extract(b, v_c, kz));
nir_ssa_def *t = nir_fadd(b, nir_fadd(b, nir_fmul(b, u, az), nir_fmul(b, v, bz)), nir_fmul(b, w, cz));
nir_ssa_def *t_signed = nir_fmul(b, nir_fsign(b, det), t);
nir_ssa_def *det_cond_front = nir_inot(b, nir_flt(b, t_signed, nir_imm_float(b, 0.0f)));
nir_push_if(b, det_cond_front);
{
nir_ssa_def *indices[4] = {
t, det,
v, w
};
nir_store_var(b, result, nir_vec(b, indices, 4), 0xf);
}
nir_pop_if(b, NULL);
}
nir_pop_if(b, NULL);
return nir_load_var(b, result);
}
static void
insert_traversal(struct radv_device *device, const VkRayTracingPipelineCreateInfoKHR *pCreateInfo,
nir_builder *b, const struct rt_variables *vars)
@ -1604,10 +1876,13 @@ insert_traversal(struct radv_device *device, const VkRayTracingPipelineCreateInf
nir_ssa_def *bvh_node_type = nir_iand(b, bvh_node, nir_imm_int(b, 7));
bvh_node = nir_iadd(b, nir_load_var(b, trav_vars.bvh_base), nir_u2u(b, bvh_node, 64));
nir_ssa_def *result = nir_bvh64_intersect_ray_amd(
b, 32, desc, nir_unpack_64_2x32(b, bvh_node), nir_load_var(b, vars->tmax),
nir_load_var(b, trav_vars.origin), nir_load_var(b, trav_vars.dir),
nir_load_var(b, trav_vars.inv_dir));
nir_ssa_def *intrinsic_result = NULL;
if (device->physical_device->rad_info.chip_class >= GFX10_3) {
intrinsic_result = nir_bvh64_intersect_ray_amd(
b, 32, desc, nir_unpack_64_2x32(b, bvh_node), nir_load_var(b, vars->tmax),
nir_load_var(b, trav_vars.origin), nir_load_var(b, trav_vars.dir),
nir_load_var(b, trav_vars.inv_dir));
}
nir_push_if(b, nir_ine(b, nir_iand(b, bvh_node_type, nir_imm_int(b, 4)), nir_imm_int(b, 0)));
{
@ -1681,6 +1956,14 @@ insert_traversal(struct radv_device *device, const VkRayTracingPipelineCreateInf
nir_push_else(b, NULL);
{
/* box */
nir_ssa_def *result = intrinsic_result;
if (!result) {
/* If we didn't run the intrinsic cause the hardware didn't support it,
* emulate ray/box intersection here */
result = intersect_ray_amd_software_box(device,
b, bvh_node, nir_load_var(b, vars->tmax), nir_load_var(b, trav_vars.origin),
nir_load_var(b, trav_vars.dir), nir_load_var(b, trav_vars.inv_dir));
}
for (unsigned i = 4; i-- > 0; ) {
nir_ssa_def *new_node = nir_vector_extract(b, result, nir_imm_int(b, i));
@ -1700,6 +1983,14 @@ insert_traversal(struct radv_device *device, const VkRayTracingPipelineCreateInf
}
nir_push_else(b, NULL);
{
nir_ssa_def *result = intrinsic_result;
if (!result) {
/* If we didn't run the intrinsic cause the hardware didn't support it,
* emulate ray/tri intersection here */
result = intersect_ray_amd_software_tri(device,
b, bvh_node, nir_load_var(b, vars->tmax), nir_load_var(b, trav_vars.origin),
nir_load_var(b, trav_vars.dir), nir_load_var(b, trav_vars.inv_dir));
}
insert_traversal_triangle_case(device, pCreateInfo, b, result, vars, &trav_vars, bvh_node);
}
nir_pop_if(b, NULL);