From f27a97328339a9e640971d004494cf515d5fdabf Mon Sep 17 00:00:00 2001 From: Konstantin Seurer Date: Wed, 24 Jan 2024 17:34:05 +0100 Subject: [PATCH] radv/rt: Use doubles inside intersect_ray_amd_software_tri Increasing precision lets us avoid duplicate hits on shared edges. The previous biasing WA made a bunch of tests fail. Reviewed-by: Friedrich Vock Part-of: --- src/amd/vulkan/nir/radv_nir_rt_common.c | 33 ++++++++++++++----------- 1 file changed, 18 insertions(+), 15 deletions(-) diff --git a/src/amd/vulkan/nir/radv_nir_rt_common.c b/src/amd/vulkan/nir/radv_nir_rt_common.c index 9ae88c44ca2..f0fccf049f3 100644 --- a/src/amd/vulkan/nir/radv_nir_rt_common.c +++ b/src/amd/vulkan/nir/radv_nir_rt_common.c @@ -227,6 +227,13 @@ intersect_ray_amd_software_tri(struct radv_device *device, nir_builder *b, nir_d nir_def *cx = nir_fsub(b, nir_vector_extract(b, v_c, kx), nir_fmul(b, sx, nir_vector_extract(b, v_c, kz))); nir_def *cy = nir_fsub(b, nir_vector_extract(b, v_c, ky), nir_fmul(b, sy, nir_vector_extract(b, v_c, kz))); + 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_def *u = nir_fsub(b, nir_fmul(b, cx, by), nir_fmul(b, cy, bx)); nir_def *v = nir_fsub(b, nir_fmul(b, ax, cy), nir_fmul(b, ay, cx)); nir_def *w = nir_fsub(b, nir_fmul(b, bx, ay), nir_fmul(b, by, ax)); @@ -244,6 +251,12 @@ intersect_ray_amd_software_tri(struct radv_device *device, nir_builder *b, nir_d { nir_def *det = nir_fadd(b, u, nir_fadd(b, v, w)); + sz = nir_f2f64(b, sz); + + v_a = nir_f2f64(b, v_a); + v_b = nir_f2f64(b, v_b); + v_c = nir_f2f64(b, v_c); + nir_def *az = nir_fmul(b, sz, nir_vector_extract(b, v_a, kz)); nir_def *bz = nir_fmul(b, sz, nir_vector_extract(b, v_b, kz)); nir_def *cz = nir_fmul(b, sz, nir_vector_extract(b, v_c, kz)); @@ -256,7 +269,11 @@ intersect_ray_amd_software_tri(struct radv_device *device, nir_builder *b, nir_d nir_push_if(b, det_cond_front); { - nir_def *indices[4] = {t, det, v, w}; + t = nir_f2f32(b, nir_fdiv(b, t, det)); + v = nir_f2f32(b, nir_fdiv(b, v, det)); + w = nir_f2f32(b, nir_fdiv(b, w, det)); + + nir_def *indices[4] = {t, nir_imm_float(b, 1.0), v, w}; nir_store_var(b, result, nir_vec(b, indices, 4), 0xf); } nir_pop_if(b, NULL); @@ -399,20 +416,6 @@ insert_traversal_triangle_case(struct radv_device *device, nir_builder *b, const nir_def *divs[2] = {div, div}; intersection.barycentrics = nir_fdiv(b, nir_channels(b, result, 0xc), nir_vec(b, divs, 2)); - nir_def *hit_t = intersection.t; - /* t values within 10 ULP of the current hit t are most likely duplicate hits along shared edges, which - * might occur with emulated RT. The Vulkan spec discourages double-hits along shared-edges, so reject them - * here by subtracting 10 ULP from t. - */ - if (radv_emulate_rt(device->physical_device)) { - nir_def *abs_t = nir_fabs(b, hit_t); - nir_def *sign_t = nir_fsign(b, hit_t); - - nir_def *tm1 = nir_iadd(b, hit_t, nir_imul_imm(b, nir_f2i32(b, sign_t), -10)); - nir_def *tm2 = nir_fmul(b, nir_isub_imm(b, 10, abs_t), nir_fneg(b, sign_t)); - intersection.t = nir_bcsel(b, nir_ige_imm(b, abs_t, 10), tm1, tm2); - } - args->triangle_cb(b, &intersection, args, ray_flags); } nir_pop_if(b, NULL);