From 548382de42f118f2686c6ebbb40f8770839a3183 Mon Sep 17 00:00:00 2001 From: Joshua Ashton Date: Tue, 21 Sep 2021 15:29:59 +0100 Subject: [PATCH] radv: Implement software emulation for intersect_ray Signed-off-by: Joshua Ashton Reviewed-by: Bas Nieuwenhuizen Part-of: --- src/amd/vulkan/radv_pipeline_rt.c | 299 +++++++++++++++++++++++++++++- 1 file changed, 295 insertions(+), 4 deletions(-) diff --git a/src/amd/vulkan/radv_pipeline_rt.c b/src/amd/vulkan/radv_pipeline_rt.c index e38469d7e14..d46e74b8dbd 100644 --- a/src/amd/vulkan/radv_pipeline_rt.c +++ b/src/amd/vulkan/radv_pipeline_rt.c @@ -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);