all : be more strict about converting float to double (#458)

* Be more strict about converting float to double

* Test equivalence of round, SILU implementations

Test module is commented out in CMakeLists.txt because the tests may
take a long time, depending on how much the compiler optimizes.

* Fix softmax in perplexity.cpp

* all : prefer float over double where appropriate

* perplexity : add <cmath>

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

2 files changed, 54 insertions, 0 deletions

diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index b44d7fe..157d733 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -5,5 +5,6 @@ function(llama_add_test source)
     add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
 endfunction()
 
+# llama_add_test(test-double-float.c) # SLOW
 llama_add_test(test-quantize.c)
 llama_add_test(test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab.bin)
diff --git a/tests/test-double-float.c b/tests/test-double-float.c
new file mode 100644
index 0000000..89dafc9
--- /dev/null
+++ b/tests/test-double-float.c
@@ -0,0 +1,53 @@
+// These tests may take a long time!
+// They are to prove that conversion from double to float of various functions in ggml.c doesn't affect the result.
+// This is done by checking all finite (non-NaN, non-infinite) floats.
+
+#undef NDEBUG
+#include <assert.h>
+#include <immintrin.h>
+#include <math.h>
+#include <stdint.h>
+
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wdouble-promotion"
+
+// ggml.c::quantize_row_q4_0_reference
+inline static uint8_t round_orig(float v0) { return ((int8_t) (round(v0))) + 8; }
+
+// ggml.c::ggml_silu_f32
+inline static float silu_orig(float x) {
+    return x/(1.0 + exp(-x));
+}
+
+#pragma GCC diagnostic pop
+
+// ggml.c::quantize_row_q4_0_reference
+inline static uint8_t round_float(float v0) { return (int8_t)roundf(v0) + 8; }
+
+// ggml.c::ggml_silu_f32
+inline static float silu_float(float x) {
+    return x/(1.0f + expf(-x));
+}
+
+int main(void) {
+    uint32_t x = UINT32_MAX;
+    do {
+        float f = *(float *)&x;
+        assert(!isfinite(f) || (round_orig(f) == round_float(f)));
+    } while (x--);
+
+#ifdef __F16C__
+    // GELU and SILU implementations are used with a FP16 lookup table.
+    // The original and float-only results are not equal for all inputs after converting to FP16.
+    // GELU is an approximation anyway (tanh), not tested here.
+    // For SILU, verify that the results are at least the closest floating point numbers, if the FP16 values don't match.
+    for (x = 0; x <= UINT16_MAX; x++) {
+        float f = _cvtsh_ss(x);
+        const float so = silu_orig(f);
+        const float sf = silu_float(f);
+        assert(   (_cvtss_sh(so, 0) == _cvtss_sh(sf, 0))
+               || (nextafterf(so, sf) == sf)
+               || (nextafterf(sf, so) == so));
+    }
+#endif
+}