feature/adm: fix undefined rounding constant in adm_cm for frame widths 17 to 32

libvmaf/src/feature/integer_adm.c

@@ -1560,11 +1560,13 @@ float adm_cm(AdmBuffer *buf, int w, int h, int src_stride, int csf_a_stride,
     const int32_t add_shift_xvsq = 268435456;
     const int32_t add_shift_xdsq = 536870912;
 
+    /* shift_xhcub and shift_xvcub are 0 for a band of 9 to 16 samples, where
+     * pow(2, shift - 1) is inf and its conversion to uint32_t is undefined. */
     const uint32_t shift_xhcub = (uint32_t)ceil(log2(w) - 4);
-    const uint32_t add_shift_xhcub = (uint32_t)pow(2, (shift_xhcub - 1));
+    const uint32_t add_shift_xhcub = shift_xhcub ? 1u << (shift_xhcub - 1) : 0;
 
     const uint32_t shift_xvcub = (uint32_t)ceil(log2(w) - 4);
-    const uint32_t add_shift_xvcub = (uint32_t)pow(2, (shift_xvcub - 1));
+    const uint32_t add_shift_xvcub = shift_xvcub ? 1u << (shift_xvcub - 1) : 0;
 
     const uint32_t shift_xdcub = (uint32_t)ceil(log2(w) - 3);
     const uint32_t add_shift_xdcub = (uint32_t)pow(2, (shift_xdcub - 1));

libvmaf/src/feature/x86/adm_avx2.c

@@ -2017,11 +2017,13 @@ float adm_cm_avx2(AdmBuffer *buf, int w, int h, int src_stride, int csf_a_stride
     const int32_t add_shift_xvsq = 268435456;
     const int32_t add_shift_xdsq = 536870912;
 
+    /* shift_xhcub and shift_xvcub are 0 for a band of 9 to 16 samples, where
+     * pow(2, shift - 1) is inf and its conversion to uint32_t is undefined. */
     const uint32_t shift_xhcub = (uint32_t)ceil(log2(w) - 4);
-    const uint32_t add_shift_xhcub = (uint32_t)pow(2, (shift_xhcub - 1));
+    const uint32_t add_shift_xhcub = shift_xhcub ? 1u << (shift_xhcub - 1) : 0;
 
     const uint32_t shift_xvcub = (uint32_t)ceil(log2(w) - 4);
-    const uint32_t add_shift_xvcub = (uint32_t)pow(2, (shift_xvcub - 1));
+    const uint32_t add_shift_xvcub = shift_xvcub ? 1u << (shift_xvcub - 1) : 0;
 
     const uint32_t shift_xdcub = (uint32_t)ceil(log2(w) - 3);
     const uint32_t add_shift_xdcub = (uint32_t)pow(2, (shift_xdcub - 1));

libvmaf/src/feature/x86/adm_avx512.c

@@ -1672,11 +1672,13 @@ float adm_cm_avx512(AdmBuffer *buf, int w, int h, int src_stride, int csf_a_stri
     const int32_t add_shift_xvsq = 268435456;
     const int32_t add_shift_xdsq = 536870912;
 
+    /* shift_xhcub and shift_xvcub are 0 for a band of 9 to 16 samples, where
+     * pow(2, shift - 1) is inf and its conversion to uint32_t is undefined. */
     const uint32_t shift_xhcub = (uint32_t)ceil(log2(w) - 4);
-    const uint32_t add_shift_xhcub = (uint32_t)pow(2, (shift_xhcub - 1));
+    const uint32_t add_shift_xhcub = shift_xhcub ? 1u << (shift_xhcub - 1) : 0;
 
     const uint32_t shift_xvcub = (uint32_t)ceil(log2(w) - 4);
-    const uint32_t add_shift_xvcub = (uint32_t)pow(2, (shift_xvcub - 1));
+    const uint32_t add_shift_xvcub = shift_xvcub ? 1u << (shift_xvcub - 1) : 0;
 
     const uint32_t shift_xdcub = (uint32_t)ceil(log2(w) - 3);
     const uint32_t add_shift_xdcub = (uint32_t)pow(2, (shift_xdcub - 1));

libvmaf/test/meson.build

@@ -162,6 +162,21 @@ test_conversion_policy = executable('test_conversion_policy',
     include_directories : [libvmaf_inc, test_inc, include_directories('../src/')],
 )
 
+test_adm_cm_small_width = executable('test_adm_cm_small_width',
+    ['test.c', 'test_adm_cm_small_width.c', '../src/mem.c', '../src/picture.c', '../src/ref.c',
+     '../src/dict.c', '../src/opt.c', '../src/log.c', '../src/predict.c',
+     '../src/metadata_handler.c'],
+    include_directories : [libvmaf_inc, test_inc, include_directories('../src/')],
+    dependencies : [math_lib, stdatomic_dependency, thread_lib, cuda_dependency],
+    objects : [
+      common_cuda_objects,
+      platform_specific_cpu_objects,
+      libvmaf_feature_static_lib.extract_all_objects(recursive: true),
+      libvmaf_cpu_static_lib.extract_all_objects(recursive: true),
+      libsvm_static_lib.extract_all_objects(recursive: true),
+    ]
+)
+
 test_motion_blend = executable('test_motion_blend',
     ['test.c', 'test_motion_blend.c'],
     include_directories : [libvmaf_inc, test_inc, include_directories('../src/')],
@@ -243,6 +258,7 @@ test('test_framesync', test_framesync)
 test('test_propagate_metadata', test_propagate_metadata)
 test('test_adm_csf', test_adm_csf)
 test('test_barten_csf', test_barten_csf)
+test('test_adm_cm_small_width', test_adm_cm_small_width)
 test('test_motion_blend', test_motion_blend)
 test('test_conversion_policy', test_conversion_policy)
 

libvmaf/test/test_adm_cm_small_width.c

@@ -0,0 +1,133 @@
+/**
+ *
+ *  Copyright 2026 Lusoris
+ *
+ *     Licensed under the BSD+Patent License (the "License");
+ *     you may not use this file except in compliance with the License.
+ *     You may obtain a copy of the License at
+ *
+ *         https://opensource.org/licenses/BSDplusPatent
+ *
+ *     Unless required by applicable law or agreed to in writing, software
+ *     distributed under the License is distributed on an "AS IS" BASIS,
+ *     WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ *     See the License for the specific language governing permissions and
+ *     limitations under the License.
+ *
+ */
+
+#include <math.h>
+#include <stdint.h>
+#include <stdlib.h>
+
+#include "test.h"
+#include "config.h"
+#include "cpu.h"
+#include "feature/integer_adm.h"
+
+#if ARCH_X86
+#include "feature/x86/adm_avx2.h"
+#if HAVE_AVX512
+#include "feature/x86/adm_avx512.h"
+#endif
+#endif
+
+typedef float (*adm_cm_fn)(AdmBuffer *buf, int w, int h, int src_stride,
+                           int csf_a_stride, double adm_norm_view_dist,
+                           int adm_ref_display_height, int adm_csf_mode,
+                           double adm_csf_scale, double adm_csf_diag_scale,
+                           double adm_noise_weight, bool measure_aim);
+
+#if ARCH_X86
+static uint32_t lcg_state;
+
+static int16_t lcg_sample(void)
+{
+    lcg_state = lcg_state * 1664525u + 1013904223u;
+    return (int16_t)((lcg_state >> 8) % 801u) - 400;
+}
+
+static void fill_band(adm_dwt_band_t *band, int rows, int stride)
+{
+    for (int i = 0; i < rows * stride; i++) {
+        band->band_h[i] = lcg_sample();
+        band->band_v[i] = lcg_sample();
+        band->band_d[i] = lcg_sample();
+    }
+}
+
+/* As adm_csf() leaves it: csf_f holds 1/30 of the magnitude of csf_a. */
+static void fill_csf_f(adm_dwt_band_t *csf_f, const adm_dwt_band_t *csf_a,
+                       int rows, int stride)
+{
+    for (int i = 0; i < rows * stride; i++) {
+        csf_f->band_h[i] = (int16_t)(abs(csf_a->band_h[i]) / 30);
+        csf_f->band_v[i] = (int16_t)(abs(csf_a->band_v[i]) / 30);
+        csf_f->band_d[i] = (int16_t)(abs(csf_a->band_d[i]) / 30);
+    }
+}
+
+static float run_adm_cm(adm_cm_fn fn, int w, int h)
+{
+    AdmBuffer buf;
+    if (adm_buffer_alloc(&buf, w, h)) return NAN;
+
+    const int stride = (int)(buf.ind_size_x >> 2);
+    const int w_half = (w + 1) / 2, h_half = (h + 1) / 2;
+
+    lcg_state = (uint32_t)(w * 4099 + h);
+    fill_band(&buf.decouple_r, h_half, stride);
+    fill_band(&buf.csf_a, h_half, stride);
+    fill_csf_f(&buf.csf_f, &buf.csf_a, h_half, stride);
+
+    const float score =
+        fn(&buf, w_half, h_half, stride, stride, DEFAULT_ADM_NORM_VIEW_DIST,
+           DEFAULT_ADM_REF_DISPLAY_HEIGHT, DEFAULT_ADM_CSF_MODE,
+           DEFAULT_ADM_CSF_SCALE, DEFAULT_ADM_CSF_DIAG_SCALE,
+           DEFAULT_ADM_NOISE_WEIGHT, false);
+
+    adm_buffer_free(&buf);
+    return score;
+}
+
+static int matches_scalar(adm_cm_fn fn, const char *name)
+{
+    /* Frame widths 17 to 32 give a scale 0 band of 9 to 16 samples, for which
+     * the cube shift in adm_cm() is 0. 40 and 64 are controls. */
+    static const int widths[] = { 17, 18, 20, 23, 24, 27, 31, 32, 40, 64 };
+
+    for (unsigned i = 0; i < sizeof(widths) / sizeof(widths[0]); i++) {
+        const int w = widths[i], h = 36;
+        const float expected = run_adm_cm(adm_cm, w, h);
+        const float actual = run_adm_cm(fn, w, h);
+        if (!(fabsf(expected - actual) <= 1e-6f * fabsf(expected))) {
+            fprintf(stderr, "%s, frame width %d: %.9g, scalar %.9g\n",
+                    name, w, actual, expected);
+            return 0;
+        }
+    }
+    return 1;
+}
+#endif
+
+static char *test_adm_cm_small_width_matches_scalar()
+{
+#if ARCH_X86
+    const unsigned flags = vmaf_get_cpu_flags();
+
+    if (flags & VMAF_X86_CPU_FLAG_AVX2)
+        mu_assert("adm_cm_avx2 differs from adm_cm", matches_scalar(adm_cm_avx2, "adm_cm_avx2"));
+#if HAVE_AVX512
+    if (flags & VMAF_X86_CPU_FLAG_AVX512)
+        mu_assert("adm_cm_avx512 differs from adm_cm", matches_scalar(adm_cm_avx512, "adm_cm_avx512"));
+#endif
+#endif
+    return NULL;
+}
+
+char *run_tests()
+{
+    vmaf_init_cpu();
+    mu_run_test(test_adm_cm_small_width_matches_scalar);
+    return NULL;
+}