Commit 8b5b271993 for aom
commit 8b5b271993f1a01d438f3a377b395da782931148
Author: Jerome Jiang <jianj@google.com>
Date: Thu Nov 19 11:12:41 2026 -0500
cost_test: Parameterize batch tests under C and AVX2
- Move batch correctness and speed tests from TEST_F to TEST_P
(CostTokensFromCdfBatchTest) instantiated under C and AVX2
prefixes so test_libaom's negative AVX2 filter skips AVX2 runs
on CPUs/emulators without AVX2 (e.g. x86 Valgrind).
- Wrap batch calls in API_REGISTER_STATE_CHECK() in correctness
checks and verify compilation with HAVE_AVX2=0.
TAG=agy
CONV=cf0b23f9-8b4f-428b-9b8e-d17a5614d8f8
Change-Id: I9f4d2a7813d4e529d95247765774c7eec0bb23c0
diff --git a/test/cost_test.cc b/test/cost_test.cc
index c297a3babf..e865fdfb12 100644
--- a/test/cost_test.cc
+++ b/test/cost_test.cc
@@ -10,6 +10,7 @@
*/
#include <algorithm>
+#include <tuple>
#include <vector>
#include "config/aom_config.h"
@@ -20,14 +21,17 @@
#include "av1/common/entropy.h"
#include "av1/encoder/cost.h"
#include "test/acm_random.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
#include "gtest/gtest.h"
namespace {
using libaom_test::ACMRandom;
-using Cdf4BatchFn = void (*)(int *costs, int costs_stride, const uint16_t *cdf,
- int cdf_stride, int n);
+using CdfBatchFn = void (*)(int *costs, int costs_stride, const uint16_t *cdf,
+ int cdf_stride, int n);
+using CdfBatchTestParams = std::tuple<CdfBatchFn, CdfBatchFn>;
// Fixture that fills av1_cost_lut_low once before exercising the cost helpers.
class CostTokensFromCdf4Test : public ::testing::Test {
@@ -35,6 +39,12 @@ class CostTokensFromCdf4Test : public ::testing::Test {
void SetUp() override { av1_init_cost_luts(); }
};
+class CostTokensFromCdfBatchTest
+ : public ::testing::TestWithParam<CdfBatchTestParams> {
+ protected:
+ void SetUp() override { av1_init_cost_luts(); }
+};
+
// Builds a valid stored 4-symbol CDF from monotonic cumulative probs.
void BuildCdf4(ACMRandom *rnd, aom_cdf_prob cdf[5]) {
uint16_t v[3];
@@ -101,7 +111,7 @@ TEST_F(CostTokensFromCdf4Test, MatchesGenericEdge) {
// filled entry against the generic scalar path. Exercises the SIMD main loop
// plus every tail length (n % 8) and both output strides (8 = base_cost,
// 4 = br_rate).
-void CheckBatchAgainstGeneric(Cdf4BatchFn fn) {
+void CheckBatchAgainstGeneric(CdfBatchFn fn) {
ACMRandom rnd(ACMRandom::DeterministicSeed());
const int kCdfStride = CDF_SIZE(4); // 5
for (const int costs_stride : { 4, 8 }) {
@@ -114,7 +124,8 @@ void CheckBatchAgainstGeneric(Cdf4BatchFn fn) {
for (int k = 0; k < n; ++k)
av1_cost_tokens_from_cdf(&ref[k * 4], &cdf[k * kCdfStride], nullptr);
- fn(out.data(), costs_stride, cdf.data(), kCdfStride, n);
+ API_REGISTER_STATE_CHECK(
+ fn(out.data(), costs_stride, cdf.data(), kCdfStride, n));
for (int k = 0; k < n; ++k)
for (int i = 0; i < 4; ++i)
@@ -125,18 +136,8 @@ void CheckBatchAgainstGeneric(Cdf4BatchFn fn) {
}
}
-TEST_F(CostTokensFromCdf4Test, BatchCMatchesGeneric) {
- CheckBatchAgainstGeneric(av1_cost_tokens_from_cdf_4_batch_c);
-}
-
-#if HAVE_AVX2
-TEST_F(CostTokensFromCdf4Test, BatchAVX2MatchesGeneric) {
- CheckBatchAgainstGeneric(av1_cost_tokens_from_cdf_4_batch_avx2);
-}
-#endif // HAVE_AVX2
-
// Same as CheckBatchAgainstGeneric but for the N=2 batch primitive.
-void CheckBatch2AgainstGeneric(Cdf4BatchFn fn) {
+void CheckBatch2AgainstGeneric(CdfBatchFn fn) {
ACMRandom rnd(ACMRandom::DeterministicSeed());
const int kCdfStride = CDF_SIZE(2); // 3
for (const int costs_stride : { 2, 4 }) {
@@ -149,7 +150,8 @@ void CheckBatch2AgainstGeneric(Cdf4BatchFn fn) {
for (int k = 0; k < n; ++k)
av1_cost_tokens_from_cdf(&ref[k * 2], &cdf[k * kCdfStride], nullptr);
- fn(out.data(), costs_stride, cdf.data(), kCdfStride, n);
+ API_REGISTER_STATE_CHECK(
+ fn(out.data(), costs_stride, cdf.data(), kCdfStride, n));
for (int k = 0; k < n; ++k)
for (int i = 0; i < 2; ++i)
@@ -160,21 +162,10 @@ void CheckBatch2AgainstGeneric(Cdf4BatchFn fn) {
}
}
-TEST_F(CostTokensFromCdf4Test, Batch2CMatchesGeneric) {
- CheckBatch2AgainstGeneric(av1_cost_tokens_from_cdf_2_batch_c);
-}
-
-#if HAVE_AVX2
-TEST_F(CostTokensFromCdf4Test, Batch2AVX2MatchesGeneric) {
- CheckBatch2AgainstGeneric(av1_cost_tokens_from_cdf_2_batch_avx2);
-}
-#endif // HAVE_AVX2
-
-#if HAVE_AVX2
-// Times the C vs AVX2 batch primitive on the real av1_fill_coeff_costs call
+// Times the C vs target batch primitive on the real av1_fill_coeff_costs call
// shapes and prints ns/call and the speedup. DISABLED_ so it runs only on
// request (--gtest_also_run_disabled_tests).
-void RunBatchSpeed(const char *name, Cdf4BatchFn c_fn, Cdf4BatchFn simd_fn,
+void RunBatchSpeed(const char *name, CdfBatchFn c_fn, CdfBatchFn simd_fn,
int cdf_stride, int syms, const char *label, int n,
int costs_stride) {
ACMRandom rnd(ACMRandom::DeterministicSeed());
@@ -187,7 +178,7 @@ void RunBatchSpeed(const char *name, Cdf4BatchFn c_fn, Cdf4BatchFn simd_fn,
}
std::vector<int> out(n * costs_stride);
- const Cdf4BatchFn funcs[2] = { c_fn, simd_fn };
+ const CdfBatchFn funcs[2] = { c_fn, simd_fn };
const int kLoops = 1000000;
double ns[2];
for (int i = 0; i < 2; ++i) {
@@ -199,27 +190,42 @@ void RunBatchSpeed(const char *name, Cdf4BatchFn c_fn, Cdf4BatchFn simd_fn,
ns[i] =
1000.0 * static_cast<double>(aom_usec_timer_elapsed(&timer)) / kLoops;
}
- printf("%s %-9s n=%2d stride=%d: c %7.1f ns avx2 %7.1f ns (%4.2fx)\n", name,
+ printf("%s %-9s n=%2d stride=%d: c %7.1f ns tst %7.1f ns (%4.2fx)\n", name,
label, n, costs_stride, ns[0], ns[1], ns[0] / ns[1]);
}
-TEST_F(CostTokensFromCdf4Test, DISABLED_Batch4Speed) {
- RunBatchSpeed("cdf_4_batch", av1_cost_tokens_from_cdf_4_batch_c,
- av1_cost_tokens_from_cdf_4_batch_avx2, CDF_SIZE(4), 4,
- "base_cost", SIG_COEF_CONTEXTS, 8);
- RunBatchSpeed("cdf_4_batch", av1_cost_tokens_from_cdf_4_batch_c,
- av1_cost_tokens_from_cdf_4_batch_avx2, CDF_SIZE(4), 4,
- "br_rate", LEVEL_CONTEXTS, BR_CDF_SIZE);
+TEST_P(CostTokensFromCdfBatchTest, Batch4MatchesGeneric) {
+ CheckBatchAgainstGeneric(GET_PARAM(0));
+}
+
+TEST_P(CostTokensFromCdfBatchTest, Batch2MatchesGeneric) {
+ CheckBatch2AgainstGeneric(GET_PARAM(1));
+}
+
+TEST_P(CostTokensFromCdfBatchTest, DISABLED_Batch4Speed) {
+ RunBatchSpeed("cdf_4_batch", av1_cost_tokens_from_cdf_4_batch_c, GET_PARAM(0),
+ CDF_SIZE(4), 4, "base_cost", SIG_COEF_CONTEXTS, 8);
+ RunBatchSpeed("cdf_4_batch", av1_cost_tokens_from_cdf_4_batch_c, GET_PARAM(0),
+ CDF_SIZE(4), 4, "br_rate", LEVEL_CONTEXTS, BR_CDF_SIZE);
}
-TEST_F(CostTokensFromCdf4Test, DISABLED_Batch2Speed) {
- RunBatchSpeed("cdf_2_batch", av1_cost_tokens_from_cdf_2_batch_c,
- av1_cost_tokens_from_cdf_2_batch_avx2, CDF_SIZE(2), 2,
- "txb_skip", TXB_SKIP_CONTEXTS, 2);
- RunBatchSpeed("cdf_2_batch", av1_cost_tokens_from_cdf_2_batch_c,
- av1_cost_tokens_from_cdf_2_batch_avx2, CDF_SIZE(2), 2,
- "eob_extra", EOB_COEF_CONTEXTS, 2);
+TEST_P(CostTokensFromCdfBatchTest, DISABLED_Batch2Speed) {
+ RunBatchSpeed("cdf_2_batch", av1_cost_tokens_from_cdf_2_batch_c, GET_PARAM(1),
+ CDF_SIZE(2), 2, "txb_skip", TXB_SKIP_CONTEXTS, 2);
+ RunBatchSpeed("cdf_2_batch", av1_cost_tokens_from_cdf_2_batch_c, GET_PARAM(1),
+ CDF_SIZE(2), 2, "eob_extra", EOB_COEF_CONTEXTS, 2);
}
+
+INSTANTIATE_TEST_SUITE_P(
+ C, CostTokensFromCdfBatchTest,
+ ::testing::Values(CdfBatchTestParams(av1_cost_tokens_from_cdf_4_batch_c,
+ av1_cost_tokens_from_cdf_2_batch_c)));
+
+#if HAVE_AVX2
+INSTANTIATE_TEST_SUITE_P(AVX2, CostTokensFromCdfBatchTest,
+ ::testing::Values(CdfBatchTestParams(
+ av1_cost_tokens_from_cdf_4_batch_avx2,
+ av1_cost_tokens_from_cdf_2_batch_avx2)));
#endif // HAVE_AVX2
} // namespace